Quat Book

Quaternion algebras
John Voight
[email protected]
v.0.9.28
May 20, 2021
Preface
Goal
Quaternion algebras sit prominently at the intersection of many mathematical subjects.
They capture essential features of noncommutative ring theory, number theory, 𝐾-
theory, group theory, geometric topology, Lie theory, functions of a complex variable,
spectral theory of Riemannian manifolds, arithmetic geometry, representation theory,
the Langlands program—and the list goes on. Quaternion algebras are especially
fruitful to study because they often reflect some of the general aspects of these subjects,
while at the same time they remain amenable to concrete argumentation. Moreover,
quaternions often encapsulate unique features that are absent from the general theory
(even as they provide motivation for it).
With this in mind, the main goal in writing this text is to introduce a large subset
of the above topics to graduate students interested in algebra, geometry, and number
theory. To get the most out of reading this text, readers will likely want to have
been exposed to some algebraic number theory, commutative algebra (e.g., module
theory, localization, and tensor products), as well as the fundamentals of linear algebra,
topology, and complex analysis. For certain sections, further experience with objects in
differential geometry or arithmetic geometry (e.g., Riemannian manifolds and elliptic
curves), may be useful. With these prerequisites in mind, I have endeavored to present
the material in the simplest, motivated version—full of rich interconnections and
illustrative examples—so even if the reader is missing a piece of background, it can be
quickly filled in.
Unfortunately, this text only scratches the surface of most of the topics covered
in the book! In particular, some appearances of quaternion algebras in arithmetic
geometry that are dear to me are absent, as they would substantially extend the length
and scope of this already long book. I hope that the presentation herein will serve as a
foundation upon which a detailed and more specialized treatment of these topics will
be possible.
I have tried to maximize exposition of ideas and minimize technicality: sometimes
I allow a quick and dirty proof, but sometimes the “right level of generality” (where
things can be seen most clearly) is pretty abstract. So my efforts have resulted in a level
of exposition that is occasionally uneven jumping between sections. I consider this a
feature of the book, and I hope that the reader will agree and feel free to skip around
(see How to use this book below). I tried to “reboot” at the beginning of each part
and again at the beginning of each chapter, to refresh our motivation. For researchers
i
ii
working with quaternion algebras, I have tried to collect results otherwise scattered
in the literature and to provide some clarifications, corrections, and complete proofs
in the hopes that this text will provide a convenient reference. In order to provide
these features, to the extent possible I have opted for an organizational pattern that is
“horizontal” rather than “vertical”: the text has many chapters, each representing a
different slice of the theory.
I tried to compactify the text as much as possible, without sacrificing completeness.
There were a few occasions when I thought a topic could use further elaboration or has
evolved from the existing literature, but did not want to overburden the text; I collected
these in a supplementary text Quaternion algebras companion, available at the website
for the text at http://quatalg.org.
As usual, each chapter also contains a number of exercises at the end, ranging
from checking basic facts used in a proof to more difficult problems that stretch the
reader. Exercises that are used in the text are marked by I. For a subset of exercises
(including many of those marked with I), there are hints, comments, or a complete
solution available online.
How to use this book

With apologies to Whitman, this book is large, it contains multitudes—and hopefully,
it does not contradict itself!
There is no obligation to read the book linearly cover to cover, and the reader is
encouraged to find their own path, such as one the following.
1. For an introductory survey course on quaternion algebras, read just the introduc-
tory sections in each chapter, those labelled with ⊲, and supplement with sections
from the text when interested. These introductions usually contain motivation
and a summary of the results in the rest of the chapter, and I often restrict the
level of generality or make simplifying hypotheses so that the main ideas are
made plain. The reader who wants to quickly and gently grab hold of the basic
concepts may digest the book in this way. The instructor may desire to fill in
some further statements or proofs to make for a one semester course: chapters
1, 2, 11, 25, and 35 could be fruitfully read in their entirety.
2. For a mini-course in noncommutative algebra with emphasis on quaternion
algebras, read just part I. Such an early graduate course would have minimal
prerequisites and in a semester could be executed at a considered pace; it would
provide the foundation for further study in many possible directions.
3. For quaternion algebras and algebraic number theory, read parts I and II. This
course would be a nice second-semester addition following a standard first-
semester course in algebraic number theory, suitable for graduate students in
algebra and number theory who are motivated to study quaternion algebras as
“noncommutative quadratic fields”. For a lighter course, chapters 6, 20, and
21 could be skipped, and the instructor may opt to cover only the introductory
section of a chapter for reasons of time and interest. To reinforce concepts from
algebraic number theory, special emphasis could be placed on chapter 13 (where
iii
local division algebras are treated like local fields) and 18 (where maximal orders
are treated like noncommutative Dedekind domains).
There are also more specialized options, beginning with the introductory sections
in part I and continuing as follows.
4. For quaternion algebras and analytic number theory, continue with the intro-
ductory sections in part II (just chapters 9–17), and then cover part III (at least
through Chapter 29). This course could follow a first-semester course in an-
alytic number theory, enriching students’ understanding of zeta functions and
𝐿-functions (roughly speaking, beginning the move from GL1 to GL2 ). The
additional prerequisite of real analysis (measure theory) is recommended. Op-
tionally, this course could break after chapter 26 to avoid adeles, and perhaps
resume in an advanced topics course with the remaining chapters.
5. For quaternionic applications to geometry (specifically, hyperbolic geometry
and low-dimensional topology), continue with the introductory sections in part
II (through chapter 14), and then cover part IV (optionally skipping Chapter 32).
6. For an advanced course on quaternion algebras and arithmetic geometry, con-
tinue with part II, the introductory sections in part IV, and part V. Chapter 41
could be read immediately after part II. This path is probably most appropriate
for an advanced course for students with some familiarity with modular forms
and some hyperbolic geometry, and chapter 42 is probably only meaningful for
students with a background in elliptic curves (though the relevant concepts are
reviewed at the start).
7. Finally, for the reader who is studying quaternion algebras with an eye to appli-
cations with supersingular elliptic curves, the reader may follow chapters 2–4,
9–10, 13–14, 16 –17, 23, then the main event in chapter 42. For further reading
on quaternion orders and ternary quadratic forms, I suggest chapters 5, 22, and
24.
Sections of the text that are more advanced (requiring more background) or those
may be omitted are labeled with ∗. The final chapter (Chapter 43) is necessarily
more advanced, and additional prerequisites in algebraic and arithmetic geometry are
indicated.
It is a unique feature of quaternion algebras that topics overlap and fold together like
this, and so I hope the reader will forgive the length of the book. The reader may find
the symbol definition list at the end to help in identifying unfamiliar notation. Finally,
to ease in location I have chosen to number all objects (theorem-like environments,
equations, and figures) consecutively.
Companion reading
Several general texts can serve as companion reading for this monograph:
• The lecture notes of Vignéras [Vig80a] have been an essential reference for the
arithmetic of quaternion algebras since their publication.
iv
• The seminal text by Reiner [Rei2003] on maximal orders treats many introduc-
tory topics that overlap this text.
• The book of Maclachlan–Reid [MR2003] gives an introduction to quaternion
algebras with application to the geometry of 3-manifolds.
• The book by Deuring [Deu68] (in German) develops the theory of algebras over
fields, culminating in the treatment of zeta functions of division algebras over
the rationals, and may be of historical interest as well.
• Finally, Pizer [Piz76a] and Alsina–Bayer [AB2004] present arithmetic and al-
gorithmic aspects of quaternion algebras over Q.
Acknowledgements
This book began as notes from a course offered at McGill University in the Winter 2010
semester, entitled Computational aspects of quaternion algebras and Shimura curves. I
would like to thank the members of my Math 727 class for their invaluable discussions
and corrections: Dylan Attwell-Duval, Xander Faber, Luis Finotti, Andrew Fiori,
Cameron Franc, Adam Logan, Marc Masdeu, Jungbae Nam, Aurel Page, Jim Parks,
Victoria de Quehen, Rishikesh, Shahab Shahabi, and Luiz Takei. This course was part
of the special thematic semester Number Theory as Applied and Experimental Science
organized by Henri Darmon, Eyal Goren, Andrew Granville, and Mike Rubinstein
at the Centre de Recherche Mathématiques (CRM) in Montréal, Québec, and the
extended visit was made possible by the generosity of Dominico Grasso, dean of the
College of Engineering and Mathematical Sciences, and Jim Burgmeier, chair of the
Department of Mathematics and Statistics, at the University of Vermont. I am very
grateful for their support.
After a long hiatus, the writing continued while the author was on sabbatical
at the University of California, Berkeley. Several students attended these lectures
and gave helpful feedback: Watson Ladd, Andrew Niles, Shelly Manber, Eugenia
Ros, u, Emmanuel Tsukerman, Victoria Wood, and Alex Youcis. Thanks to Ken Ribet
for sponsoring my visit and helping to organize these lectures. My sabbatical from
Dartmouth College for the Fall 2013 and Winter 2014 quarters was made possible by
the efforts of Associate Dean David Kotz, and I thank him for his support. During this
sabbatical, further progress on the text was made in preparation for a minicourse on
Brandt modules as part of Minicourses on Algebraic and Explicit Methods in Number
Theory, organized by Cécile Armana and Christophe Delaunay at the Laboratoire de
Mathématiques de Besançon in Salins-les-Bains, France.
Thanks also go to the patient participants in my Math 125 Quaternion algebras
class at Dartmouth in Spring 2014 (Daryl Deford, Tim Dwyer, Zeb Engberg, Michael
Firrisa, Jeff Hein, Nathan McNew, Jacob Richey, Tom Shemanske, Scott Smedinghoff,
and David Webb) and my Math 125 Geometry of discrete groups in Summer 2015
(Angelica Babei, Ben Breen, Sara Chari, Melanie Dennis, James Drain, Tim Dwyer,
Jon Epstein, David Freund, Sam Kater, Michael Musty, Nicolas Petit, Sam Schiavone,
Scott Smedinghoff, and Everett Sullivan). I am also grateful for the feedback given
during my course at the Institute for Computational and Experimental Research in
Mathematics (ICERM) at Brown University in Fall 2015 as part of the Semester
v
Program on Computational Aspects of the Langlands Program: in particular, thanks

to Henri Cohen, Edgar Costa, David Farmer, Winnie Li, David Roberts, and Drew
Sutherland for their comments and questions.
During the preparation of this monograph, I was supported by the NSF Divi-
sion of Mathematical Sciences Awards for Quaternion algebras, Shimura curves,
and modular forms: Algorithms and arithmetic (DMS-0901971, July 2009–August
2011) and CAREER: Explicit methods in arithmetic geometry (DMS-1151047, July
2012–July 2018), and by a Simons Collaborations in Mathematics and the Physical
Sciences Award for Arithmetic geometry, number theory, and computation (Septem-
ber 2017–August 2021, with Jennifer Balakrishnan, Noam Elkies, Brendan Hassett,
Bjorn Poonen, and Drew Sutherland). The Open Access publication of this book
was made possible in part by generous support from Dartmouth College via Faculty
Research and Professional Development funds, a John M. Manley Huntington Award,
and Dartmouth Library Open Access Publishing.
It is somehow fitting that I would find myself writing this text while a faculty
member at Dartmouth College: the story of the quaternions is interwoven with the
history of mathematics at Dartmouth. The only mathematical output by a Dartmouth
professor in the 19th century was by Arthur Sherburne Hardy, Ph.D., the author of an
1881 text on quaternions entitled Elements of quaternions [Har1881]. Brown describes
it as
an adequate, if not inspiring text. It was something for Dartmouth to offer

a course in such an abstruse field, and the course was actually given a
few times when a student and an instructor could be found simultaneously
[Bro61, p. 2].
I can only hope that this book will receive better reviews!
On a more personal note, I have benefited from the insight of incredible teachers
and collaborators throughout my mathematical life, and their positive impact can
hopefully be seen in this manuscript: among many, I would like to give special thanks
to Pete L. Clark, Bas Edixhoven, Benedict Gross, Hendrik Lenstra, and Bjorn Poonen
for their inspiration and guidance over the years.
Many thanks go to the others who offered helpful comments, corrections, answered
my questions, and provided feedback: Konstantin Ardakov, Sarah Arpin, Asher Auel,
Andreas Bächle, John Boxall, Ben Breen, Juliusz Brzezinski, Billy Chan, Pierre
Clare, France Dacar, Lassina Dembélé, Juanita Duque Rosero, Kirsten Eisentraeger,
John Enns, Zak Evans, Steven Galbraith, Darĳ Grinberg, Joseph Gunther, Jonathan
Hanke, Thorsten Herrig, Aleksander Horawa, Will Jagy, Ariyan Javanpeykar, Bo-
Gwang Jeon, Chan-Ho Kim, Robert Kucharczyk, Joel Laity, Dion Leĳnse, Patrick
Lenning, Chao Li, Wen-Wei Li, Benjamin Linowitz, Travis Morrison, Ariel Pacetti,
Lorenz Panny, Jennifer Park, Carl Pomerance, Jose Pujol, Matthieu Rambaud, John
Riccardi, Sam Roven, Nithi Rungtanapirom, Ciaran Schembri, sibilant, Jana Sotakova,
Tom Shemanske, Daniel Smertnig, Jim Stankewicz, Kate Stange, Drew Sutherland,
Nicole Sutherland, Jacob Swenberg, Anne van Weerden, Annalena Wernz, Mckenzie
West, Dana Williams, Jiangwei Xue, and Vitthal Yellambalse. I am indebted to C.
Emily I. Redelmeier for fabulous TeXnical assistance with figures. Special thanks to the
vi
anonymous referees for their constructive feedback and to Sara Chari, Peter Landweber,
Joe Quinn, and Grant Molnar for a careful reading. The errors and omissions that
remain are, of course, my own: please contact me at [email protected] if you find
mistakes or have other suggestions (e.g., additional references, alternate proofs, further
exercises or applications). Collected errata and a corrected version will be posted on
the webpage for the text (http://quatalg.org).
I am profoundly grateful to those who lent encouragement at various times during
the writing of this book, when the going was tough: Srinath Baba, Chantal David,
Matthew Greenberg and Kristina Loeschner, Laurie Johnson at the National Center
for Faculty Development and Diversity (NCFDD), David Michaels, Tom Shemanske,
and my mother Connie Voight. Thank you all! I am also grateful to Marc Strauss and
Loretta Bartolini, editors at Springer, for their professional guidance. Finally, I would
like to offer my deepest gratitude to my partner Brian Kennedy: this book would not
have been possible without his patience, equanimity, and wit. As they say, happiness
is a journey, not a destination; the same I think is true for happy mathematics, and I
am blessed to have had both personal and professional companionship on this one.
Contents
Contents vii
1 Introduction 1
1.1 Hamilton’s quaternions . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Algebra after the quaternions . . . . . . . . . . . . . . . . . . . . . 6
1.3 Quadratic forms and arithmetic . . . . . . . . . . . . . . . . . . . . 10
1.4 Modular forms and geometry . . . . . . . . . . . . . . . . . . . . . 11
1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
I Algebra 19
2 Beginnings 21
2.1 ⊲ Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 ⊲ Quaternion algebras . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 ⊲ Matrix representations . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4 ⊲ Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3 Involutions 35
3.1 ⊲ Conjugation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2 Involutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3 Reduced trace and reduced norm . . . . . . . . . . . . . . . . . . . 38
3.4 Uniqueness and degree . . . . . . . . . . . . . . . . . . . . . . . . 39
3.5 Quaternion algebras . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4 Quadratic forms 47
4.1 ⊲ Reduced norm as quadratic form . . . . . . . . . . . . . . . . . . 47
4.2 Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.3 Discriminants, nondegeneracy . . . . . . . . . . . . . . . . . . . . . 52
4.4 Nondegenerate standard involutions . . . . . . . . . . . . . . . . . . 54
4.5 Special orthogonal groups . . . . . . . . . . . . . . . . . . . . . . . 55
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
vii
viii CONTENTS
5 Ternary quadratic forms 63

5.1 ⊲ Reduced norm as quadratic form . . . . . . . . . . . . . . . . . . 63
5.2 Isomorphism classes of quaternion algebras . . . . . . . . . . . . . 64
5.3 Clifford algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.4 Splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.5 Conics, embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.6 Orientations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6 Characteristic 2 83
6.1 Separability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.3 ∗ Quadratic forms . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.4 ∗ Characterizing quaternion algebras . . . . . . . . . . . . . . . . . 87
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7 Simple algebras 93
7.1 ⊲ Motivation and summary . . . . . . . . . . . . . . . . . . . . . . 93
7.2 Simple modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.3 Wedderburn–Artin . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.4 Jacobson radical . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.5 Central simple algebras . . . . . . . . . . . . . . . . . . . . . . . . 103
7.7 The Skolem–Noether theorem . . . . . . . . . . . . . . . . . . . . . 106
7.8 Reduced trace and norm, universality . . . . . . . . . . . . . . . . . 110
7.9 Separable algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
8 Simple algebras and involutions 121

8.1 ⊲ The Brauer group and involutions . . . . . . . . . . . . . . . . . . 121
8.2 Biquaternion algebras . . . . . . . . . . . . . . . . . . . . . . . . . 122
8.3 Brauer group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.4 Positive involutions . . . . . . . . . . . . . . . . . . . . . . . . . . 126
8.5 ∗ Endomorphism algebras of abelian varieties . . . . . . . . . . . . 129
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
II Arithmetic 133
9 Lattices and integral quadratic forms 135

9.1 ⊲ Integral structures . . . . . . . . . . . . . . . . . . . . . . . . . . 135
9.2 Bits of commutative algebra . . . . . . . . . . . . . . . . . . . . . . 136
9.3 Lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
9.4 Localizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
9.5 Completions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
9.6 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
CONTENTS ix
9.7 Quadratic forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

9.8 Normalized form . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
10 Orders 151
10.1 ⊲ Lattices with multiplication . . . . . . . . . . . . . . . . . . . . . 151
10.2 Orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
10.3 Integrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
10.4 Maximal orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
10.5 Orders in a matrix ring . . . . . . . . . . . . . . . . . . . . . . . . 156
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
11 The Hurwitz order 159

11.1 ⊲ The Hurwitz order . . . . . . . . . . . . . . . . . . . . . . . . . . 159
11.2 ⊲ Hurwitz units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
11.3 ⊲ Euclidean algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 163
11.4 ⊲ Unique factorization . . . . . . . . . . . . . . . . . . . . . . . . . 164
11.5 Finite quaternionic unit groups . . . . . . . . . . . . . . . . . . . . 166
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
12 Ternary quadratic forms over local fields 175

12.1 ⊲ The p-adic numbers and local quaternion algebras . . . . . . . . . 175
12.2 Local fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
12.3 Classification via quadratic forms . . . . . . . . . . . . . . . . . . . 183
12.4 Hilbert symbol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
13 Quaternion algebras over local fields 195

13.1 Extending the valuation . . . . . . . . . . . . . . . . . . . . . . . . 195
13.2 Valuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
13.3 Classification via extensions of valuations . . . . . . . . . . . . . . 199
13.4 Consequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
13.5 Some topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
14 Quaternion algebras over global fields 209

14.1 ⊲ Ramification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
14.2 ⊲ Hilbert reciprocity over the rationals . . . . . . . . . . . . . . . . 211
14.3 ⊲ Hasse–Minkowski theorem over the rationals . . . . . . . . . . . . 215
14.4 Global fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
14.5 Ramification and discriminant . . . . . . . . . . . . . . . . . . . . . 222
14.6 Quaternion algebras over global fields . . . . . . . . . . . . . . . . 223
14.7 Theorems on norms . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
15 Discriminants 233
x CONTENTS
15.1 ⊲ Discriminantal notions . . . . . . . . . . . . . . . . . . . . . . . . 233

15.2 Discriminant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
15.3 Quadratic forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
15.4 Reduced discriminant . . . . . . . . . . . . . . . . . . . . . . . . . 238
15.5 Maximal orders and discriminants . . . . . . . . . . . . . . . . . . 240
15.6 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
16 Quaternion ideals and invertibility 249

16.1 ⊲ Quaternion ideals . . . . . . . . . . . . . . . . . . . . . . . . . . 249
16.2 Locally principal, compatible lattices . . . . . . . . . . . . . . . . . 251
16.3 Reduced norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
16.4 Algebra and absolute norm . . . . . . . . . . . . . . . . . . . . . . 255
16.5 Invertible lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
16.6 Invertibility with a standard involution . . . . . . . . . . . . . . . . 260
16.7 One-sided invertibility . . . . . . . . . . . . . . . . . . . . . . . . . 263
16.8 Invertibility and the codifferent . . . . . . . . . . . . . . . . . . . . 265
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
17 Classes of quaternion ideals 269

17.1 ⊲ Ideal classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
17.2 Matrix ring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
17.3 Classes of lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
17.4 Types of orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
17.5 ⊲ Finiteness of the class set: over the integers . . . . . . . . . . . . . 276
17.6 ⊲ Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
17.7 Finiteness of the class set: over number rings . . . . . . . . . . . . . 279
17.8 Eichler’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
18 Picard group 289

18.1 ⊲ Noncommutative Dedekind domains . . . . . . . . . . . . . . . . 289
18.2 Prime ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
18.3 Invertibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
18.4 Picard group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
18.5 Classes of two-sided ideals . . . . . . . . . . . . . . . . . . . . . . 297
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
19 Brandt groupoids 301

19.1 ⊲ Composition laws and ideal multiplication . . . . . . . . . . . . . 301
19.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
19.3 Groupoid structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
19.4 Brandt groupoid . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
19.5 Brandt class groupoid . . . . . . . . . . . . . . . . . . . . . . . . . 308
19.6 Quadratic forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
CONTENTS xi
20 Integral representation theory 315

20.1 ⊲ Projectivity, invertibility, and representation theory . . . . . . . . . 315
20.2 Projective modules . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
20.3 Projective modules and invertible lattices . . . . . . . . . . . . . . . 318
20.4 Jacobson radical . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
20.5 Local Jacobson radical . . . . . . . . . . . . . . . . . . . . . . . . . 322
20.6 Integral representation theory . . . . . . . . . . . . . . . . . . . . . 323
20.7 ∗ Stable class group and cancellation . . . . . . . . . . . . . . . . . 325
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
21 Hereditary and extremal orders 333

21.1 ⊲ Hereditary and extremal orders . . . . . . . . . . . . . . . . . . . 333
21.2 Extremal orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
21.3 ∗ Explicit description of extremal orders . . . . . . . . . . . . . . . 336
21.4 Hereditary orders . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
21.5 ∗ Classification of local hereditary orders . . . . . . . . . . . . . . . 340
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
22 Ternary quadratic forms 343

22.1 ⊲ Quaternion orders and ternary quadratic forms . . . . . . . . . . . 343
22.2 Even Clifford algebras . . . . . . . . . . . . . . . . . . . . . . . . . 346
22.3 Even Clifford algebra of a ternary quadratic module . . . . . . . . . 349
22.4 Over a PID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
22.5 Twisting and final bĳection . . . . . . . . . . . . . . . . . . . . . . 358
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
23 Quaternion orders 365

23.1 ⊲ Highlights of quaternion orders . . . . . . . . . . . . . . . . . . . 365
23.2 Maximal orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
23.3 Hereditary orders . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
23.4 Eichler orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
23.5 Bruhat–Tits tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
24 Quaternion orders: second meeting 381

24.1 ⊲ Advanced quaternion orders . . . . . . . . . . . . . . . . . . . . . 381
24.2 Gorenstein orders . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
24.3 Eichler symbol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
24.4 Chains of orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391
24.5 Bass and basic orders . . . . . . . . . . . . . . . . . . . . . . . . . 394
24.6 Tree of odd Bass orders . . . . . . . . . . . . . . . . . . . . . . . . 397
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
xii CONTENTS
III Analysis 401
25 The Eichler mass formula 403

25.1 ⊲ Weighted class number formula . . . . . . . . . . . . . . . . . . . 403
25.2 ⊲ Imaginary quadratic class number formula . . . . . . . . . . . . . 404
25.3 ⊲ Eichler mass formula: over the rationals . . . . . . . . . . . . . . 408
25.4 Class number one and type number one . . . . . . . . . . . . . . . . 411
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
26 Classical zeta functions 419

26.1 ⊲ Eichler mass formula . . . . . . . . . . . . . . . . . . . . . . . . 419
26.2 Analytic class number formula . . . . . . . . . . . . . . . . . . . . 421
26.3 Classical zeta functions of quaternion algebras . . . . . . . . . . . . 424
26.4 Counting ideals in a maximal order . . . . . . . . . . . . . . . . . . 426
26.5 Eichler mass formula: maximal orders . . . . . . . . . . . . . . . . 429
26.6 Eichler mass formula: general case . . . . . . . . . . . . . . . . . . 432
26.7 Class number one . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
26.8 Functional equation and classification . . . . . . . . . . . . . . . . . 435
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
27 Adelic framework 445

27.1 ⊲ The rational adele ring . . . . . . . . . . . . . . . . . . . . . . . . 445
27.2 ⊲ The rational idele group . . . . . . . . . . . . . . . . . . . . . . . 448
27.3 ⊲ Rational quaternionic adeles and ideles . . . . . . . . . . . . . . . 450
27.4 Adeles and ideles . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
27.5 Class field theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
27.6 Noncommutative adeles . . . . . . . . . . . . . . . . . . . . . . . . 456
27.7 Reduced norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
28 Strong approximation 465

28.1 ⊲ Beginnings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
28.2 ⊲ Strong approximation for SL2 (Q) . . . . . . . . . . . . . . . . . . 467
28.3 Elementary matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 470
28.4 Strong approximation and the ideal class set . . . . . . . . . . . . . 471
28.5 Statement and first applications . . . . . . . . . . . . . . . . . . . . 473
28.6 Further applications . . . . . . . . . . . . . . . . . . . . . . . . . . 476
28.7 First proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 478
28.8 Second proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482
28.9 ∗ Normalizer groups . . . . . . . . . . . . . . . . . . . . . . . . . . 483
28.10 ∗ Stable class group . . . . . . . . . . . . . . . . . . . . . . . . . . 487
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
29 Idelic zeta functions 491

29.1 ⊲ Poisson summation and the Riemann zeta function . . . . . . . . . 491
29.2 ⊲ Idelic zeta functions, after Tate . . . . . . . . . . . . . . . . . . . 493
CONTENTS xiii
29.3 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496

29.4 Modulus and Fourier inversion . . . . . . . . . . . . . . . . . . . . 498
29.5 Archimedean measures and zeta functions . . . . . . . . . . . . . . 500
29.6 Local measures: commutative nonarchimedean case . . . . . . . . . 503
29.7 Local zeta functions: nonarchimedean case . . . . . . . . . . . . . . 505
29.8 Idelic zeta functions . . . . . . . . . . . . . . . . . . . . . . . . . . 509
29.9 Convergence and residue . . . . . . . . . . . . . . . . . . . . . . . 514
29.10 Main theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516
29.11 Tamagawa numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 523
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
30 Optimal embeddings 529

30.1 ⊲ Representation numbers . . . . . . . . . . . . . . . . . . . . . . . 529
30.2 Sums of three squares . . . . . . . . . . . . . . . . . . . . . . . . . 531
30.3 Optimal embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . 533
30.4 Counting embeddings, idelically: the trace formula . . . . . . . . . . 536
30.5 Local embedding numbers: maximal orders . . . . . . . . . . . . . 539
30.6 ∗ Local embedding numbers: Eichler orders . . . . . . . . . . . . . 542
30.7 Global embedding numbers . . . . . . . . . . . . . . . . . . . . . . 547
30.8 Class number formula . . . . . . . . . . . . . . . . . . . . . . . . . 548
30.9 Type number formula . . . . . . . . . . . . . . . . . . . . . . . . . 551
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
31 Selectivity 557
31.1 Selective orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
31.2 Selectivity conditions . . . . . . . . . . . . . . . . . . . . . . . . . 560
31.3 ∗ Selectivity setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 561
31.4 ∗ Outer selectivity inequalities . . . . . . . . . . . . . . . . . . . . 564
31.5 ∗ Middle selectivity equality . . . . . . . . . . . . . . . . . . . . . . 565
31.6 ∗ Optimal selectivity conclusion . . . . . . . . . . . . . . . . . . . . 567
31.7 ∗ Selectivity, without optimality . . . . . . . . . . . . . . . . . . . . 568
31.8 ∗ Isospectral, nonisometric manifolds . . . . . . . . . . . . . . . . . 570
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573
IV Geometry and topology 575
32 Unit groups 577

32.1 ⊲ Quaternion unit groups . . . . . . . . . . . . . . . . . . . . . . . 577
32.2 Structure of units . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578
32.3 Units in definite quaternion orders . . . . . . . . . . . . . . . . . . 580
32.4 Finite subgroups of quaternion unit groups . . . . . . . . . . . . . . 582
32.5 Cyclic subgroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . 583
32.6 ∗ Dihedral subgroups . . . . . . . . . . . . . . . . . . . . . . . . . 585
32.7 ∗ Exceptional subgroups . . . . . . . . . . . . . . . . . . . . . . . . 587
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 588
xiv CONTENTS
33 Hyperbolic plane 591

33.1 ⊲ The beginnings of hyperbolic geometry . . . . . . . . . . . . . . . 591
33.2 Geodesic spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
33.3 Upper half-plane . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594
33.4 Classification of isometries . . . . . . . . . . . . . . . . . . . . . . 597
33.5 Geodesics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
33.6 Hyperbolic area and the Gauss–Bonnet formula . . . . . . . . . . . 602
33.7 Unit disc and Lorentz models . . . . . . . . . . . . . . . . . . . . . 604
33.8 Riemannian geometry . . . . . . . . . . . . . . . . . . . . . . . . . 607
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610
34 Discrete group actions 615

34.1 Topological group actions . . . . . . . . . . . . . . . . . . . . . . . 615
34.2 ⊲ Summary of results . . . . . . . . . . . . . . . . . . . . . . . . . 618
34.3 Covering space and wandering actions . . . . . . . . . . . . . . . . 618
34.4 Hausdorff quotients and proper group actions . . . . . . . . . . . . . 620
34.5 Proper actions on a locally compact space . . . . . . . . . . . . . . 623
34.6 Symmetric space model . . . . . . . . . . . . . . . . . . . . . . . . 625
34.7 Fuchsian groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . 626
34.8 Riemann uniformization and orbifolds . . . . . . . . . . . . . . . . 628
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 631
35 Classical modular group 635

35.1 ⊲ The fundamental set . . . . . . . . . . . . . . . . . . . . . . . . . 635
35.2 Binary quadratic forms . . . . . . . . . . . . . . . . . . . . . . . . 641
35.3 Moduli of lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
35.4 Congruence subgroups . . . . . . . . . . . . . . . . . . . . . . . . 644
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 646
36 Hyperbolic space 649

36.1 Hyperbolic space . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
36.2 Isometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
36.3 Unit ball, Lorentz, and symmetric space models . . . . . . . . . . . 655
36.4 Bianchi groups and Kleinian groups . . . . . . . . . . . . . . . . . . 657
36.5 Hyperbolic volume . . . . . . . . . . . . . . . . . . . . . . . . . . 658
36.6 Picard modular group . . . . . . . . . . . . . . . . . . . . . . . . . 662
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 666
37 Fundamental domains 671

37.1 ⊲ Dirichlet domains for Fuchsian groups . . . . . . . . . . . . . . . 671
37.2 ⊲ Ford domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 674
37.3 ⊲ Generators and relations . . . . . . . . . . . . . . . . . . . . . . . 676
37.4 Dirichlet domains . . . . . . . . . . . . . . . . . . . . . . . . . . . 681
37.5 Hyperbolic Dirichlet domains . . . . . . . . . . . . . . . . . . . . . 685
37.6 Poincare’s polyhedron theorem . . . . . . . . . . . . . . . . . . . . 686
37.7 Signature of a Fuchsian group . . . . . . . . . . . . . . . . . . . . . 689
CONTENTS xv
37.8 The (6, 4, 2)-triangle group . . . . . . . . . . . . . . . . . . . . . . 690

37.9 Unit group for discriminant 6 . . . . . . . . . . . . . . . . . . . . . 693
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 698
38 Quaternionic arithmetic groups 701

38.1 ⊲ Rational quaternion groups . . . . . . . . . . . . . . . . . . . . . 701
38.2 Isometries from quaternionic groups . . . . . . . . . . . . . . . . . 703
38.3 Discreteness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705
38.4 Compactness and finite generation . . . . . . . . . . . . . . . . . . 707
38.5 ∗ Arithmetic groups, more generally . . . . . . . . . . . . . . . . . 709
38.6 ∗ Modular curves, seen idelically . . . . . . . . . . . . . . . . . . . 710
38.7 ∗ Double cosets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 712
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
39 Volume formula 717

39.1 ⊲ Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 717
39.2 Volume setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 720
39.3 Volume derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . 722
39.4 Genus formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 723
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 727
V Arithmetic geometry 729
40 Classical modular forms 731

40.1 ⊲ Functions on lattices . . . . . . . . . . . . . . . . . . . . . . . . . 731
40.2 ⊲ Eisenstein series as modular forms . . . . . . . . . . . . . . . . . 735
40.3 ⊲ Classical modular forms . . . . . . . . . . . . . . . . . . . . . . . 738
40.4 Theta series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 741
40.5 Hecke operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 742
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 744
41 Brandt matrices 747

41.1 ⊲ Brandt matrices, neighbors, and modular forms . . . . . . . . . . . 747
41.2 Brandt matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 751
41.3 Commutativity of Brandt matrices . . . . . . . . . . . . . . . . . . 754
41.4 Semisimplicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 757
41.5 Eichler trace formula . . . . . . . . . . . . . . . . . . . . . . . . . 759
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
42 Supersingular elliptic curves 767

42.1 Supersingular elliptic curves . . . . . . . . . . . . . . . . . . . . . . 767
42.2 Supersingular isogenies . . . . . . . . . . . . . . . . . . . . . . . . 771
42.3 Equivalence of categories . . . . . . . . . . . . . . . . . . . . . . . 775
42.4 Supersingular endomorphism rings . . . . . . . . . . . . . . . . . . 777
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 780
xvi CONTENTS
43 Abelian surfaces with QM 783

43.1 ⊲ QM abelian surfaces . . . . . . . . . . . . . . . . . . . . . . . . . 783
43.2 ⊲ QM by discriminant 6 . . . . . . . . . . . . . . . . . . . . . . . . 786
43.3 Genus 2 curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 790
43.4 Complex abelian varieties . . . . . . . . . . . . . . . . . . . . . . . 793
43.5 Complex abelian surfaces . . . . . . . . . . . . . . . . . . . . . . . 798
43.6 Abelian surfaces with QM . . . . . . . . . . . . . . . . . . . . . . . 801
43.7 Real points, CM points . . . . . . . . . . . . . . . . . . . . . . . . 807
43.8 ∗ Canonical models . . . . . . . . . . . . . . . . . . . . . . . . . . 809
43.9 ∗ Modular forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 810
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 812
Symbol Definition List 815
Bibliography 823
Index 849
Chapter 1
Introduction
We begin following the historical arc of quaternion algebras and tracing their impact
on the development of mathematics. Our account is selective: for further overview,
see Lam [Lam2003] and Lewis [Lew2006a].
1.1 Hamilton’s quaternions

In perhaps the “most famous act of mathematical vandalism”, on October 16, 1843, Sir
William Rowan Hamilton (1805–1865, Figure 1.1.2) carved the following equations
into the Brougham Bridge (now Broom Bridge) in Dublin:
𝑖 2 = 𝑗 2 = 𝑘 2 = 𝑖 𝑗 𝑘 = −1. (1.1.1)
His discovery was a defining moment in the history of algebra (Figure 1.1.3).
Figure 1.1.2: William Rowan Hamilton

(public domain; scan by Wellesley College Library)
1
2 CHAPTER 1. INTRODUCTION
For at least ten years (on and off), Hamilton had been attempting to model (real)
three-dimensional space with a structure like the complex numbers, whose addition
and multiplication occur in two-dimensional space. Just like the complex numbers had
a “real” and “imaginary” part, so too did Hamilton hope to find an algebraic system
whose elements had a “real” and two-dimensional “imaginary” part. In the early part
of the month of October 1843, his sons Archibald Henry and William Edwin Hamilton,
while still very young, would ask their father at breakfast [Ham67, p. xv]: “Well, papa,
can you multiply triplets?” To which Hamilton would reply, “with a sad shake of the
head, ‘No, I can only add and subtract them’” [Ham67, p. xv]. For a history of the
“multiplying triplets” problem—the nonexistence of division algebra over the reals of
dimension 3—see May [May66, p. 290].
Figure 1.1.3: William Rowan Hamilton, a sand sculpture by Daniel Doyle,

part of the 2012 Dublin castle exhibition, Irish Science
(reproduced with permission)
Then, on the dramatic day in 1843, Hamilton’s had a flash of insight [Ham67,
p. xv–xvi], which he described in a letter to Archibald (written in 1865):
On the 16th day of [October]—which happened to be a Monday, and a
Council day of the Royal Irish Academy—I was walking in to attend and
preside, and your mother was walking with me, along the Royal Canal, to
which she had perhaps driven; and although she talked with me now and
then, yet an under-current of thought was going on in my mind, which
gave at last a result, whereof it is not too much to say that I felt at once
the importance. An electric circuit seemed to close; and a spark flashed
forth, the herald (as I foresaw, immediately) of many long years to come
of definitely directed thought and work, by myself if spared, and at all
1.1. HAMILTON’S QUATERNIONS 3
events on the part of others, if I should even be allowed to live long

enough distinctly to communicate the discovery. Nor could I resist the
impulse—unphilosophical as it may have been—to cut with a knife on a
stone of Brougham Bridge, as we passed it, the fundamental formula with
the symbols, 𝑖, 𝑗, 𝑘; namely,
𝑖 2 = 𝑗 2 = 𝑘 2 = 𝑖 𝑗 𝑘 = −1
which contains the Solution of the Problem, but of course, as an inscrip-

tion, has long since mouldered away.
In this moment, Hamilton realized that he needed a fourth dimension; he later coined
the term quaternions for the real space spanned by the elements 1, 𝑖, 𝑗, 𝑘, subject to his
multiplication laws. He presented his theory of quaternions to the Royal Irish Academy
in a paper entitled “On a new Species of Imaginary Quantities connected with a theory
of Quaternions” [Ham1843]. Today, we denote this algebra H := R + R𝑖 + R 𝑗 + R𝑘 and
call H the ring of Hamilton quaternions in his honor.
This charming story of quaternionic discovery remains in the popular conscious-
ness, and to commemorate Hamilton’s discovery of the quaternions, there is an annual
“Hamilton walk” in Dublin [ÓCa2010]. Although his carvings have long since worn
away, a plaque on the bridge now commemorates this significant event in mathematical
history (Figure 1.1.4).
Figure 1.1.4: The Broom Bridge plaque (author’s photo)
For more on the history of Hamilton’s discovery, see the extensive and detailed
accounts of Dickson [Dic19] and Van der Waerden [vdW76]. There are also three
main biographies written about the life of William Rowan Hamilton, a man sometimes
referred to as “Ireland’s greatest mathematician”: by Graves [Grav1882, Grav1885,
Grav1889] in three volumes, Hankins [Hankin80], and O’Donnell [O’Do83]. Numer-
ous other shorter biographies have been written [DM89, Lanc67, ÓCa2000]. (Certain
aspects of Hamilton’s private life deserve a more positive portrayal, however: see Van
Weerden–Wepster [WW2018].)
160 BLBHENTS OF QUATBRNIOliS. (BOOK. U.
the law of i, j, Aagree with usual and algebraic law: namely~

in the A11ociative Property of Multiplication ; or in the pro-
perty that the new symbols always obey the a1sociative .for-
mula (comp. 9),
'.«A ... '" .~.
whichever of them. may be substituted for c, for "• and for :\ ;
in virtue of which equality of values we may omit tlte point, in
any such symbol of a terMry product (whether of equal or of
unequal factoni), and write it simply as c«A. In particular
we have thus,
i jl: i o i - i'l - - J ;
D ij.A=l:./e-l:'=-1;
or briefly,
ij.4=-1.
We may, therefore, by 182, establish the following important
Formula:
I"'J-i' ... "' = ijk .. - 1 ; (A)
to which we shall occasionally refer, as to "Formula A," and
which we shall find to contain (virtually) an the laws of the
•ymbou ijk, and therefore to be a •u.fficient symbolical basil
for the whole Calculu1 of Quaternioru :• because it will be
shown that every quaternion can be reduced to the Quadrino-
mial Form,
q=w+i~+j!J+kz,
where w, ~. y, z compose a '1Jitem offour scalars, while i, j, l
are the BIUDe tlcree rirfkt versor1 as above.
(1.) A direct proof of the equation, ijl =-1, may be derived from thedeftDitlODs
of the symbols in Art. 181. In fact, we have only to remember that thoee deftnl-
tions were- to give,
• Thia formula (A),.,.. accordingly made the bcui• of that Calculns In the lint
communication on the subject, by tbe preeent writer, to the Royal lriah Aeademy in
1848 ; and the !etten, i, j, I, coatinued to be, for some time, the ore~, pHtlliGr .,..
hl• of the c:aleulus In question. But it was gradually found to be nsefa1 to Incor-
porate with these a few other aotatiolu (auch as K and U, &c.), for rep~ntlag
OportJtitnU 011 QuleMiiOfl•. It was aleo thought to be Instructive to eetablilh the
priAeiplu or that Calculus, on a more !lft't11Mcal (or leu exeluaively ·~
f'*"datiora tban at first ; which was accordingly af'lerwards done, in the volume en-
titled: L«nru 011 QruJtmaiou (Dublin, 1858); and I• again attempted in the pre·
eent work, although with many dilfereneea In the adopted plara of expoeltion 1 and in
the applictJtioro• brought forward, or suppreaeed.
Figure 1.1.5: A page from Hamilton’s Elements of quaternions [Ham1866]

(public domain)DUtized by Coogle
There are several precursors to Hamilton’s discovery that bear mentioning. First,
the quaternion multiplication laws are already implicit in the four-square identity of
1.1. HAMILTON’S QUATERNIONS 5
Leonhard Euler (1707–1783):
(𝑎 21 + 𝑎 22 + 𝑎 23 + 𝑎 24 ) (𝑏 21 + 𝑏 22 + 𝑏 23 + 𝑏 24 ) = 𝑐21 + 𝑐22 + 𝑐23 + 𝑐24 =

(𝑎 1 𝑏 1 − 𝑎 2 𝑏 2 − 𝑎 3 𝑏 3 − 𝑎 4 𝑏 4 ) 2 + (𝑎 1 𝑏 2 + 𝑎 2 𝑏 1 + 𝑎 3 𝑏 4 − 𝑎 4 𝑏 3 ) 2 (1.1.6)
2 2
+ (𝑎 1 𝑏 3 − 𝑎 2 𝑏 4 + 𝑎 3 𝑏 1 + 𝑎 4 𝑏 2 ) + (𝑎 1 𝑏 4 + 𝑎 2 𝑏 3 − 𝑎 3 𝑏 2 + 𝑎 4 𝑏 1 ) .
Indeed, the full multiplication law for quaternions reads precisely
(𝑎 1 + 𝑎 2𝑖 + 𝑎 3 𝑗 + 𝑎 4 𝑘) (𝑏 1 + 𝑏 2𝑖 + 𝑏 3 𝑗 + 𝑏 4 𝑘) = 𝑐 1 + 𝑐 2 𝑖 + 𝑐 3 𝑗 + 𝑐 4 𝑘
with 𝑐 1 , 𝑐 2 , 𝑐 3 , 𝑐 4 as defined in (1.1.6); the four-square identity corresponds to taking

a norm on both sides.
It was perhaps Carl Friedrich Gauss (1777–1855) who first observed this connec-
tion. In a note dated around 1819 [Gau00], he interpreted the formula (1.1.6) as a way
of composing real quadruples: to the quadruples (𝑎 1 , 𝑎 2 , 𝑎 3 , 𝑎 4 ) and (𝑏 1 , 𝑏 2 , 𝑏 3 , 𝑏 4 )
in R4 , he defined the composite tuple (𝑐 1 , 𝑐 2 , 𝑐 3 , 𝑐 4 ) and noted the noncommutativity
of this operation. Gauss elected not to publish these findings (as he chose not to do
with many of his discoveries). In letters to De Morgan [Grav1885, Grav1889, p. 330,
p. 490], Hamilton attacks the allegation that Gauss had discovered quaternions first.
Finally, Olinde Rodrigues (1795–1851) (of the Rodrigues formula for Legendre
polynomials) gave a formula for the angle and axis of a rotation in R3 obtained
from two successive rotations—essentially giving a different parametrization of the
quaternions—but had left mathematics for banking long before the publication of his
paper [Rod1840]. The story of Rodrigues and the quaternions is given by Altmann
[Alt89] and Pujol [Puj2012], and the fuller story of his life is recounted by Altmann–
Ortiz [AO2005]. See also the description by Pujol [Puj2014] of Hamilton’s derivation
of the relation between rotations and quaternions from 1847, set in historical context.
In any case, the quaternions consumed the rest of Hamilton’s academic life and
resulted in the publication of two bulky treatises [Ham1853, Ham1866] (see also the
review [Ham1899]). Hamilton’s mathematical writing over these years, an example of
which can be found in Figure 1.1.5, was at times opaque; nevertheless, many physicists
used quaternions extensively and for a long time in the mid-19th century, quaternions
were an essential notion in physics.
Other figures contemporaneous with Hamilton were also developing vectorial
systems, most notably Hermann Grassmann (1809–1877) [Gras1862]. The modern
notion of vectors was developed by Willard Gibbs (1839–1903) and Oliver Heaviside
(1850–1925), independently. In 1881 and 1884 (in two halves), Gibbs introduced in
a pamphlet Elements of Vector Analysis the now standard vector notation of the cross
product and dot product, with the splendid equality
𝑣𝑤 = −𝑣 · 𝑤 + 𝑣 × 𝑤 (1.1.7)
for 𝑣, 𝑤 ∈ R𝑖 + R 𝑗 + R𝑘 ⊂ H relating quaternionic multiplication on the left to dot

and cross products on the right. (The equality (1.1.7) also appears in Hamilton’s
work, but in different notation.) Gibbs did not consider the quaternion product to be
a “fundamental notion in vector analysis” [Gib1891, p. 512], and argued for a vector
analysis that would apply in arbitrary dimension; on the relationship between these
works, Gibbs wrote after learning of the work of Grassmann: “I saw that the methods
wh[ich] I was using, while nearly those of Hamilton, were almost exactly those of
Grassmann” [Whe62, p. 108]. For more on the history of quaternionic and vector
calculus, see Crowe [Cro64] and Simons [Sim2010].
The rivalry between physical notations flared into a war in the latter part of the 19th
century between the ‘quaternionists’ and the ‘vectorists’, and for some the preference
of one system versus the other became an almost partisan split. On the side of
quaternions, James Clerk Maxwell (1831–1879), who derived the equations which
describe electromagnetic fields, wrote [Max1869, p. 226]:
The invention of the calculus of quaternions is a step towards the knowl-
edge of quantities related to space which can only be compared, for its
importance, with the invention of triple coordinates by Descartes. The
ideas of this calculus, as distinguished from its operations and symbols,
are fitted to be of the greatest use in all parts of science.
And Peter Tait (1831–1901), Hamilton’s “chief disciple” [Hankin80, p. 316], wrote in
1890 [Tai1890] decrying notation and attacking Willard Gibbs (1839–1903):
It is disappointing to find how little progress has recently been made
with the development of Quaternions. One cause, which has been spe-
cially active in France, is that workers at the subject have been more
intent on modifying the notation, or the mode of presentation of the
fundamental principles, than on extending the applications of the Calcu-
lus. . . . Even Prof. Willard Gibbs must be ranked as one the retarders of
quaternions progress, in virtue of his pamphlet on Vector Analysis, a sort
of hermaphrodite monster, compounded of the notation of Hamilton and
Grassman.
Game on! On the vectorist side, Lord Kelvin (a.k.a. William Thomson, who formulated
the laws of thermodynamics), said in an 1892 letter to R. B. Hayward about his textbook
in algebra (quoted in Thompson [Tho10, p. 1070]):
Quaternions came from Hamilton after his really good work had been
done; and, though beautifully ingenious, have been an unmixed evil to
those who have touched them in any way, including Clerk Maxwell.
(There is also a rompous fictionalized account by Pynchon in his tome Against the Day
[Pyn2006].) Ultimately, the superiority and generality of vector notation carried the
day, and only certain useful fragments of Hamilton’s quaternionic notation—e.g., the
“right-hand rule” 𝑖 × 𝑗 = 𝑘 in multivariable calculus—remain in modern usage.
1.2 Algebra after the quaternions

The debut of Hamilton’s quaternions was met with some resistance in the mathematical
world: it proposed a system of “numbers” that did not satisfy the usual commutative
rule of multiplication. Quaternions predated even the notion of matrices, introduced in
1.2. ALGEBRA AFTER THE QUATERNIONS 7
1855 by Arthur Cayley (1821–1895). Hamilton’s bold proposal of a noncommutative

multiplication law was the harbinger of a burgeoning array of algebraic structures. In
the words of J.J. Sylvester [Syl1883, pp. 271–272]:
In Quaternions (which, as will presently be seen, are but the simplest

order of matrices viewed under a particular aspect) the example had been
given of Algebra released from the yoke of the commutative principle
of multiplication—an emancipation somewhat akin to Lobachevsky’s of
Geometry from Euclid’s noted empirical axiom; and later on, the Peirces,
father and son (but subsequently to 1858) had prefigured the universal-
ization of Hamilton’s theory, and had emitted an opinion to the effect that
probably all systems of algebraical symbols subject to the associative law
of multiplication would be eventually found to be identical with linear
transformations of schemata susceptible of matriculate representation.
So with the introduction of the quaternions, the floodgates of algebraic possibility had
been opened. See Happel [Hap80] for an overview of the early development of algebra
following Hamilton’s quaternions, as well as the more general history given by Van
der Waerden [vdW85, Chapters 10–11].
The day after his discovery, Hamilton sent a letter [Ham1844] describing the
quaternions to his friend John T. Graves (1806–1870). Graves replied on October 26,
1843, with his compliments, but added:
There is still something in the system which gravels me. I have not yet
any clear views as to the extent to which we are at liberty arbitrarily to
create imaginaries, and to endow them with supernatural properties. . . .
If with your alchemy you can make three pounds of gold, why should you
stop there?
Following through on this invitation, on December 26, 1843, Graves wrote to Hamilton
that he had successfully generalized the quaternions to the “octaves”, now called
octonions O, an algebra in eight dimensions, with which he was able to prove that the
product of two sums of eight perfect squares is another sum of eight perfect squares,
a formula generalizing (1.1.6). In fact, Hamilton first invented the term associative in
1844, around the time of his correspondence with Graves. Unfortunately for Graves, the
octonions were discovered independently and published in 1845 by Cayley [Cay1845b],
who often is credited for their discovery. (Even worse, the eight squares identity was
also previously discovered by C. F. Degen.) For a more complete account of this
story and the relationships between quaternions and octonions, see the survey article
by Baez [Bae2002], the article by Van der Blĳ [vdB60], and the delightful book by
Conway–Smith [CSm2003].
Cayley also studied quaternions themselves [Cay1845a] and was able to reinterpret
them as arising from a doubling process, also called the Cayley–Dickson construction,
which starting from R produces C then H then O, taking the ordered, commutative,
associative algebra R and progressively deleting one adjective at a time. So algebras
were first studied over the real and complex numbers and were accordingly called
hypercomplex numbers in the late 19th and early 20th century. And this theory
flourished. Hamilton himself considered the algebra over C defined by his famous
equations (1.1.1), calling them biquaternions. In 1878, Ferdinand Frobenius (1849–
1917) proved that the only finite-dimensional associative real division algebras are R,
C, and H [Fro1878]. This result was also proven independently by C.S. Peirce, the
son of Benjamin Peirce, below. Adolf Hurwitz (1859–1919) later showed that the
only normed finite-dimensional not-necessarily-associative real division algebras are
R, C, H, and O. (The same statement is true without the condition that the algebra
be normed, but currently the proofs use topology, not algebra! Bott–Milnor [BM58]
and Kervaire [Ker58] proved that the (𝑛 − 1)-dimensional sphere {𝑥 ∈ R𝑛 : k𝑥k 2 = 1}
has trivial tangent bundle if and only if there is an 𝑛-dimensional not-necessarily-
associative real division algebra if and only if 𝑛 = 1, 2, 4, 8. The solution to the
Hopf invariant one problem by Adams also implies this result; an elegant and concise
proof using 𝐾-theory, Adams operations, and elementary number theory was given
by Adams–Atiyah [AA66]. See Hirzebruch [Hir91] or Ranicki [Ran2011] for a more
complete account.)
In another attempt to seek a generalization of the quaternions to higher dimension,
William Clifford (1845–1879) developed a way to build algebras from quadratic forms
in 1876 [Cli1878]. Clifford constructed what we now call a Clifford algebra 𝐶 (𝑉)
associated to 𝑉 = R𝑛 (with the standard Euclidean norm); it is an algebra of dimension
2𝑛 containing 𝑉 with multiplication induced from the relation 𝑥 2 = −k𝑥k 2 for all 𝑥 ∈ 𝑉.
We have 𝐶 (R1 ) = C and 𝐶 (R2 ) = H, so the Hamilton quaternions arise as a Clifford
algebra—but 𝐶 (R3 ) is not the octonions. The theory of Clifford algebras tightly
connects the theory of quadratic forms and the theory of normed division algebras
and its impact extends in many mathematical directions. For more on the history of
Clifford algebras, see Diek–Kantowski [DK95].
A further physically motivated generalization was pursued by Alexander Macfar-
lane (1851–1913): he developed a theory of what he called hyperbolic quaternions
[Macf00] (a revised version of an earlier, nonassociative attempt [Macf1891]), with
the multiplication laws
𝑖 2 = 𝑗 2 = 𝑘 2 = 1,
√ √ √ (1.2.1)
𝑖 𝑗 = −1𝑘 = − 𝑗𝑖, 𝑗 𝑘 = −1𝑖 = −𝑘 𝑗, 𝑘𝑖 = −1 𝑗 = −𝑖𝑘.
√
Thought of as an algebra over C = R( −1), Macfarlane’s hyperbolic quaternions
are isomorphic to Hamilton’s biquaternions (and therefore isomorphic to M2 (C)).
Moreover, the restriction of the norm to the real span of the basis 1, 𝑖, 𝑗, 𝑘 in Mac-
farlane’s algebra is a quadratic form of signature (1, 3): this gives a quaternionic
version of space-time, something also known as Minkowski space (but with Macfar-
lane’s construction predating that of Minkowski). For more on the history and further
connections, see Crowe [Cro64].
Around this time, other types of algebras over the real numbers were also being
investigated, the most significant of which were Lie algebras. In the seminal work
of Sophus Lie (1842–1899), group actions on manifolds were understood by looking
at this action infinitesimally; one thereby obtains a Lie algebra of vector fields that
determines the local group action. The simplest nontrivial example of a Lie algebra
is the cross product of two vectors, related to quaternion multiplication in (1.1.7): it
1.2. ALGEBRA AFTER THE QUATERNIONS 9
defines, a linear, alternating, but nonassociative binary operation on R3 that satisfies

the Jacobi identity emblematized by
𝑖 × ( 𝑗 × 𝑘) + 𝑘 × (𝑖 × 𝑗) + 𝑗 × (𝑘 × 𝑖) = 0. (1.2.2)
The Lie algebra “linearizes” the group action and is therefore more accessible. Wilhelm
Killing (1847–1923) initiated the study of the classification of Lie algebras in a series
of papers [Kil1888], and this work was completed by Élie Cartan (1869–1951). We
refer to Hawkins [Haw2000] for a description of this rich series of developments.
In this way, the study of division algebras gradually evolved, independent of
physical interpretations. Benjamin Peirce (1809–1880) in 1870 developed what he
called linear associative algebras [Pei1882]; he provided a decomposition of an algebra
relative to an idempotent (his terminology). The first definition of an algebra over an
arbitrary field seems to have been given by Leonard E. Dickson (1874–1954) [Dic03]:
at first he still called the resulting object a system of complex numbers and only later
adopted the name linear algebra.
The notion of a simple algebra had been discovered by Cartan, and Theodor Molien
(1861–1941) had earlier shown in his terminology that every simple algebra over the
complex numbers is a matrix algebra [Mol1893]. But it was Joseph Henry Maclagan
Wedderburn (1882–1948) who was the first to find meaning in the structure of simple
algebras over an arbitrary field, in many ways leading the way forward. The jewel
of his 1908 paper [Wed08] is still foundational in the structure theory of algebras: a
simple algebra (finite-dimensional over a field) is isomorphic to a matrix ring over a
division ring. Wedderburn also proved that a finite division ring is a field, a result that
like his structure theorem has inspired much mathematics. For more on the legacy of
Wedderburn, see Artin [Art50].
In the early 1900s, Dickson was the first to consider quaternion algebras over
a general field [Dic12, (8), p. 65]. He began by considering more generally those
algebras in which every element satisfies a quadratic equation [Dic12], exhibited a
diagonalized basis for such an algebra, and considered when such an algebra can be
a division algebra. This led him to multiplication laws for what he later called a
generalized quaternion algebra [Dic14, Dic23], with multiplication laws
𝑖 2 = 𝑎, 𝑗 2 = 𝑏, 𝑘 2 = −𝑎𝑏,
(1.2.3)
𝑖 𝑗 = 𝑘 = − 𝑗𝑖, 𝑖𝑘 = 𝑎 𝑗 = −𝑘𝑖, 𝑘 𝑗 = 𝑏𝑖 = − 𝑗 𝑘
with 𝑎, 𝑏 nonzero elements in the base field. (To keep track of these, it is helpful
to write 𝑖, 𝑗, 𝑘 around a circle clockwise.) Today, we no longer employ the adjective
“generalized”—over fields other than R, there is no reason to privilege the Hamiltonian
quaternions—and we can reinterpret this vein of Dickson’s work as showing that every
4-dimensional central simple algebra is a quaternion algebra (a statement that holds
even over a field 𝐹 with char 𝐹 = 2). See Fenster [Fen98] for a summary of Dickson’s
work in algebra, and Lewis [Lew2006b] for a broad survey of the role of involutions
and anti-automorphisms in the classification of algebras.
1.3 Quadratic forms and arithmetic

Hamilton’s quaternions also fused a link between quadratic forms and arithmetic,
phrased in the language of noncommutative algebra. Indeed, part of Dickson’s interest
in quaternion algebras stemmed from earlier work of Hurwitz [Hur1898], alluded to
above. Hurwitz had asked for generalizations of the composition laws arising from
sum of squares laws like that of Euler (1.1.6) for four squares and Cayley for eight
squares: for which 𝑛 does there exist an identity
(𝑎 21 + · · · + 𝑎 2𝑛 ) (𝑏 21 + · · · + 𝑏 2𝑛 ) = 𝑐21 + · · · + 𝑐2𝑛
with each 𝑐 𝑖 bilinear in the variables 𝑎 and 𝑏? He then proved [Hur1898] that over
a field where 2 is invertible, these identities exist only for 𝑛 = 1, 2, 4, 8 variables (so
in particular, there is no formula expressing the product of two sums of 16 squares as
the sum of 16 squares). As Dickson [Dic19] further explained, this result of Hurwitz
is intimately tied to the theory of algebras. For more on compositions of quadratic
forms and their history, including theorems of Hurwitz–Radon and Pfister, see Shapiro
[Sha90].
Thinking along similar lines, Hurwitz gave a new proof of the four-square theorem
of Lagrange, that every positive integer is the sum of four integer squares: he first
wrote about this in 1896 on quaternionic number theory (“Über die Zahlentheorie
der Quaternionen”) [Hur1896], then published a short book on the subject in 1919
[Hur19]. To this end, Hurwitz considered Hamilton’s equations over the rational
numbers and said that a quaternion 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 with 𝑡, 𝑥, 𝑦, 𝑧 ∈ Q was an integer
if 𝑡, 𝑥, 𝑦, 𝑧 all belonged to Z or all to 12 + Z, conditions for the quaternion to satisfy
a quadratic polynomial with integer coefficients. Hurwitz showed that his ring of
integer quaternions, today called the Hurwitz order, admits a generalization of the
Euclidean algorithm and thereby a factorization theory. He then applied this to count
the number of ways of representing an integer as the sum of four squares, a result
due to Jacobi. The notion of integral quaternions was also explored in the 1920s
by Venkov [Ven22, Ven29] and the 1930s by Albert [Alb34]. Dickson considered
further questions of representing positive integers by integral quaternary quadratic
forms [Dic19, Dic23, Dic24] in the same vein.
So by the end of the 1920s, quaternion algebras were used to study quadratic
forms in a kind of noncommutative algebraic number theory [Lat26, Gri28]. It was
known that a (generalized) quaternion algebra (1.2.3) was semisimple in the sense of
Wedderburn, and thus it was either a division algebra or a full matrix algebra over
the ground field. Indeed, a quaternion algebra is a matrix algebra if and only if a
certain ternary quadratic form has a nontrivial zero, and over the rational numbers this
problem was already studied by Legendre. Helmut Hasse (1898–1979) reformulated
Legendre’s conditions: a quadratic form has a nontrivial zero over the rationals if
and only if it has a nontrivial zero over the real numbers and Hensel’s field of 𝑝-adic
numbers for all odd primes 𝑝. This result paved the way for many further advances,
and it is now known as the Hasse principle or the local-global principle for quadratic
forms. For an overview of this history, see Scharlau [Scha2009, §1].
Further deep results in number theory were soon to follow. Dickson [Dic14] had
defined cyclic algebras, reflecting many properties of quaternion algebras, and in 1929
1.4. MODULAR FORMS AND GEOMETRY 11
lectures Emmy Noether (1882–1935) considered the even more general crossed product
algebras. Not very long after, in a volume dedicated to Hensel’s seventieth birthday,
Richard Brauer (1901–1977), Hasse, and Noether proved a fundamental theorem for
the structure theory of algebras over number fields [BHN31]: every central division
algebra over a number field is a cyclic algebra. This crucial statement had profound
implications for class field theory, the classification of abelian extensions of a number
field, with a central role played by the Brauer group of a number field, a group
encoding its division algebras. For a detailed history and discussion of these lines,
see Fenster–Schwermer [FS2007], Roquette [Roq2006], and the history of class field
theory summarized by Hasse himself [Hass67].
At the same time, Abraham Adrian Albert (1905–1972), a doctoral student of Dick-
son, was working on the structure of division algebras and algebras with involution,
and he had written a full book on the subject [Alb39] collecting his work in the area,
published in 1939. Albert had examined the tensor product of two quaternion algebras,
called a biquaternion algebra (not to be confused with Hamilton’s biquaternions), and
he characterized when such an algebra was a division algebra in terms of a senary (six
variable) quadratic form. Albert’s classification of algebras with involution was moti-
vated by understanding possible endomorphism algebras of abelian varieties, viewed
as multiplier rings of Riemann matrices and equipped with the Rosati involution: a
consequence of this classification is that quaternion algebras are the only noncommu-
tative endomorphism algebras of simple abelian varieties. He also proved that a central
simple algebra admits an involution if and only if the algebra is isomorphic to its oppo-
site algebra (equivalently, it has order at most 2 in the Brauer group). For a biography
of Albert and a survey of his work, see Jacobson [Jacn74]. Roquette argues convinc-
ingly [Roq2006, §8] that because of Albert’s contributions to its proof (for example,
his work with Hasse [AH32]), we should refer to the Albert–Brauer–Hasse–Noether
theorem in the previous paragraph.
1.4 Modular forms and geometry

Quaternion algebras also played a formative role in what began as a subfield of complex
analysis and ordinary differential equations and then branched into the theory of
modular forms—and ultimately became a central area of modern number theory.
Returning to a thread from the previous section, the subject of representing numbers
as the sum of four squares saw considerable interest in the 17th and 18th centuries
[Dic71, Chapter VIII]. Carl Jacobi (1804–1851) approached the subject from the
analytic point of view of theta functions, the basic building blocks for elliptic functions;
these were first studied in connection with the problem of the arc length of an ellipse,
going back to Abel. Jacobi studied the series
∞
∑︁
𝜃 (𝜏) := exp(2𝜋𝑖𝑛2 𝜏) = 1 + 2𝑞 + 2𝑞 4 + 2𝑞 9 + . . . (1.4.1)
𝑛=−∞
where 𝜏 is a complex number with positive imaginary part and 𝑞 = exp(2𝜋𝑖𝜏). Jacobi
proved the remarkable identity

∑︁ ∞
∑︁
2 +𝑏 2 +𝑐 2 +𝑑 2
𝜃 (𝜏) 4 = 𝑞𝑎 =1+8 𝜎 ∗ (𝑛)𝑞 𝑛 , (1.4.2)
𝑎,𝑏,𝑐,𝑑 ∈Z 𝑛=1
where 𝜎 ∗ (𝑛) := 4-𝑑 |𝑛 𝑑 is the sum of divisors of 𝑛 not divisible by 4. In this way,
Í
Jacobi gave an explicit formula for the number of ways of expressing a number as the
sum of four squares. For a bit of history and an elementary derivation in the style of
Gauss and Jacobi, see Ewell [Ewe82].
As a Fourier series, the Jacobi theta function 𝜃 (1.4.1) visibly satisfies 𝜃 (𝜏 + 1) =
𝜃 (𝜏). Moreover, owing to its symmetric description, Jacobi showed using Poisson
summation that 𝜃 also satisfies the transformation formula
√︁
𝜃 (−1/𝜏) = 𝜏/𝑖 𝜃 (𝜏). (1.4.3)
Felix Klein (1849–1925) saw geometry in formulas like (1.4.3). In his Erlangen
Program (1872), he recast 19th century geometry in terms of the underlying group
of symmetries, unifying Euclidean and non-Euclidean formulations. Turning then to
hyperbolic geometry, he studied the modular group SL2 (Z) acting by linear fractional
transformations on the upper half-plane, and interpreted transformation formulas for
elliptic functions: in particular, Klein defined his absolute invariant 𝐽 (𝜏) [Kle1878],
a function invariant under the modular group. Together with his student Robert Fricke
(1861–1930), this led to four volumes [FK1890-2, FK1897, FK12] on elliptic modular
functions and automorphic functions, combining brilliant advances in group theory,
number theory, geometry, and invariant theory (Figure 1.4.4).
Figure 1.4.4: The (2, 3, 7)-tiling by Fricke and Klein [FK1890-2]

(public domain)
1.4. MODULAR FORMS AND GEOMETRY 13
At the same time, Henri Poincaré (1854–1912) brought in the theory of linear
differential equations—and a different, group-theoretic approach. In correspondence
with Fuchs in 1880 on hypergeometric differential equations, he writes about the
beginnings of his discovery of a new class of analytic functions [Gray2000, p.177]:
They present the greatest analogy with elliptic functions, and can be
represented as the quotient of two infinite series in infinitely many ways.
Amongst those series are those which are entire series playing the role of
Theta functions. These converge in a certain circle and do not exist outside
it, as thus does the Fuchsian function itself. Besides these functions there
are others which play the same role as the zeta functions in the theory
of elliptic functions, and by means of which I solve linear differential
equations of arbitrary orders with rational coefficients whenever there are
only two finite singular points and the roots of the three determinantal
equations are commensurable.
As he reminisced later in his Science et Méthode [Poi1908, p. 53]:
I then undertook to study some arithmetical questions without any great
result appearing and without expecting that this could have the least con-
nection with my previous researches. Disgusted with my lack of success,
I went to spend some days at the sea-side and thought of quite different
things. One day, walking along the cliff, the idea came to me, always with
the same characteristics of brevity, suddenness, and immediate certainty,
that the arithmetical transformations of ternary indefinite quadratic forms
were identical with those of non-Euclidean geometry.
In other words, like Klein, Poincaré launched a program to study complex analytic
functions defined on the unit disc that are invariant with respect to a discrete group
of matrix transformations that preserve a rational indefinite ternary quadratic form.
Today, such groups are called arithmetic Fuchsian groups, and we study them as unit
groups of quaternion algebras. To read more on the history of differential equations
in the time of Riemann and Poincaré, see the history by Gray [Gray2000], as well as
Gray’s scientific biography of Poincaré [Gray2013].
In the context of these profound analytic discoveries, Erich Hecke (1887–1947)
began his study of modular forms. He studied the Dedekind zeta function, a gener-
alization of Riemann’s zeta function to number fields, and established its functional
equation using theta functions. In the study of similarly defined analytic functions
arising from modular forms, he was led to define the “averaging” operators acting on
spaces of modular forms that now bear his name. In this way, he could interpret the
Fourier coefficients 𝑎(𝑛) of a Hecke eigenform (normalized, weight 2) as eigenvalues
of his operators: he proved that they satisfy a relation of the form
∑︁
𝑎(𝑚)𝑎(𝑛) = 𝑎(𝑚𝑛/𝑑 2 )𝑑 (1.4.5)
𝑑 |gcd(𝑚,𝑛)
and consequently a two-term recursion relation. He thereby showed that the Dirichlet
𝐿-series of an eigenform, defined via Mellin transform, has an Euler product, analytic
continuation, and functional equation.
Hecke went further, and connected the analytic theory of modular forms and his
operators to the arithmetic theory of quadratic forms. In 1935–1936, he found that
for certain systems of quaternary quadratic forms, the number of representations of
integers by the system satisfied the recursion (1.4.5), in analogy with binary quadratic
forms. He published a conjecture on this subject in 1940 [Hec40, Satz 53, p. 100]:
that the weighted representation numbers satisfy the Hecke recursion, connecting
coefficients to operators on theta series, and further that the columns in a composition
table always result in linearly independent theta series. He verified the conjecture up
to prime level 𝑞 < 37, but was not able to prove this recursion using his methods of
complex analysis (see his letter [Bra41, Footnote 1]).
The arithmetic part of these conjectures was investigated by Heinrich Brandt
(1886–1954) in the quaternionic context—and so the weave of our narrative is further
tightly sewn. Preceding Hecke’s work, and inspired by Gauss composition of binary
quadratic forms as the product of classes of ideals in a quadratic field, Brandt had
earlier considered a generalization to quaternary quadratic forms and the product of
classes of ideals in a quaternion algebra [Bra28]: he was only able to define a partially
defined product, and so he coined the term groupoid for such a structure [Bra40]. He
then considered the combinatorial problem of counting the ways of factoring an ideal
into prime ideals, according to their classes. In this way, he recorded these counts in
a matrix 𝑇 (𝑛) for each positive integer 𝑛, and he proved strikingly (sketched in 1941
[Bra41], dated 1939, and proved completely in 1943 [Bra43]) that the matrices 𝑇 (𝑛)
satisfy Hecke’s recursion (1.4.5). To read more on the life and work of Brandt, see
Hoehnke–Knus [HK2004]. Today we call the matrices 𝑇 (𝑛) Brandt matrices, and
for certain purposes, they are still the most convenient way to get ahold of spaces of
modular forms.
Martin Eichler (1912–1992) wrote his thesis [Eic36] under the supervision of
Brandt on quaternion orders over the integers, in particular studying the orders that
now bear his name. Later he continued the grand synthesis of modular forms, quadratic
forms, and quaternion algebras, viewing in generality the orthogonal group of a
quadratic form as acting via automorphic transformations [Eic53]. In this vein, he
formulated his basis problem (arising from the conjecture of Hecke) which sought
to understand explicitly the span of quaternionic theta series among classical modu-
lar forms, giving a correspondence between systems of Hecke eigenvalues appearing
in the quaternionic and classical context. He answered the basis problem in affir-
mative for the case of prime level in 1955 [Eic56a] and then for squarefree level
[Eic56b, Eic58, Eic73]. For more on Eichler’s basis problem and its history, see
Hĳikata–Pizer–Shemanske [HPS89a].
Having come to recent history, our account now becomes much more abbreviated:
we provide further commentary in situ in remarks in the rest of this text, and we
conclude with just a few highlights. In the 1950s and 1960s, there was subtantial
work done in understanding zeta functions of certain varieties arising from quaternion
algebras over totally real number fields. For example, Eichler’s correspondence was
generalized to totally real fields by Shimizu [Shz65]. Shimura embarked on a deep and
systematic study of arithmetic groups obtained from indefinite quaternion algebras over
totally real fields, including both the arithmetic Fuchsian groups of Poincaré, Fricke,
and Klein, and the generalization of the modular group to totally real fields studied
1.5. CONCLUSION 15
by Hilbert. In addition to understanding their zeta functions, he also formulated a

general theory of complex multiplication in terms of automorphic functions; as a
consequence, he found the corresponding arithmetic quotients can be defined as an
algebraic variety with equations defined over a number field—and so today we refer to
quaternionic Shimura varieties. For an overview of Shimura’s work, see his lectures
at the International Congress of Mathematicians in 1978 [Shi80]. As it turns out,
quaternion algebras over number fields also give rise to arithmetic manifolds that are
not algebraic varieties, and they are quite important in the areas of spectral theory,
low-dimensional geometry, and topology—in particular, in Thurston’s geometrization
program for hyperbolic 3-manifolds and in classifying knots and links.
Just as the Hecke operators determine the coefficients of classical modular forms
and Dirichlet 𝐿-series, they may be vastly generalized, replacing modular groups
by other algebraic groups, such as the group of units in a central simple algebra or
the orthogonal group of a quadratic form. Understanding the theory of automorphic
forms in this context is a program that continues today: formulated in the language
of automorphic representations, and seen as a nonabelian generalization of class field
theory, Langlands initiated this program in a letter to Weil in January 1967. It
is indeed fitting that an early success of the Langlands program [Gel84, B+2003]
would be on the subject of quaternion algebras: a generalization of the Eichler–
Shimizu correspondence to encompass arbitrary quaternion algebras over number
fields was achieved in foundational work by Jacquet–Langlands [JL70] in 1970. For
more on the modern arithmetic history of modular forms, see Edixhoven–Van der Geer–
Moonen [EGM2008]; Alsina–Bayer [AB2004, Appendices B–C] also give references
for further applications of quaternion algebras in arithmetic geometry (in particular,
of Shimura curves).
1.5 Conclusion
We have seen how quaternion algebras have threaded mathematical history through to
the present day, weaving together advances in algebra, quadratic forms, number theory,
geometry, and modular forms. And although our history ends here, the story does not!
Quaternion algebras continue to arise in unexpected ways. In the arithmetic setting,
quaternion orders arise as endomorphism rings of supersingular elliptic curves and
have been used in proposed post-quantum cryptosystems and digital signature schemes
(see for example the overview by Galbraith–Vercauteren [GV2018]). In the field of
quantum computation, Parzanchevski–Sarnak [PS2018] have proposed Super-Golden-
Gates built from certain special quaternion algebras and their arithmetic groups that
would give efficient 1-qubit quantum gates. In coding theory, lattices in quaternion
algebras (and more generally central simple algebras over number fields) yield space-
time codes that achieve high spectral efficiency on wireless channels with two transmit
antennas, currently part of certain IEEE standards [BO2013].
Quaternions have also seen a revival in computer graphics, modeling, and anima-
tion [HFK94, Sho85]. Indeed, a rotation in R3 about an axis through the origin can be
represented by a 3 × 3 orthogonal matrix with determinant 1, conveniently encoded in
Euler angles. However, the matrix representation is redundant, as there are only three
degrees of freedom in such a rotation. Moreover, to compose two rotations requires the
product of the two corresponding matrices, which requires 27 multiplications and 18
additions in R. Quaternions, on the other hand, represent this rotation with a 4-tuple,
and multiplication of two quaternions takes only 16 multiplications and 12 additions
in R (if done naively). In computer games, quaternion interpolation provides a way
to smoothly interpolate between orientations in space—something crucial for fighting
Nazi zombies. Quaternions are also vital for attitude control of aircraft and spacecraft
[Hans2006]: they avoid the ambiguity that can arise when two rotation axes align,
leading to a potentially disastrous loss of control called gimbal lock.
In quantum physics, quaternions yield elegant expressions for Lorentz transforma-
tions, the basis of the modern theory of relativity [Gir83]. Some physicists are now
hoping to find deeper understanding of these principles of quantum physics in terms
of quaternions. And so, although much of Hamilton’s quaternionic physics fell out of
favor long ago, we have come full circle in our elongated historical arc. The enduring
role of quaternion algebras as a catalyst for a vast range of mathematical research
promises rewards for many years to come.
Exercises
1. Hamilton sought a multiplication ∗ : R3 × R3 → R3 that preserves length:
k𝑣k 2 · k𝑤k 2 = k𝑣 ∗ 𝑤k 2
for 𝑣, 𝑤 ∈ R3 . Expanding out in terms of coordinates, such a multiplication

would imply that the product of the sum of three squares over R is again the sum
of three squares in R. (Such a law holds for the sum of four squares (1.1.6).)
Show that such a formula for three squares is impossible as an identity in the
polynomial ring in 6 variables over Z. [Hint: Find a natural number that is the
product of two sums of three squares which is not itself the sum of three squares.]
2. Hamilton originally sought an associative multiplication law on
𝐷 := R + R𝑖 + R 𝑗 ' R3
where 𝑖 2 = −1 and every nonzero element of 𝐷 has a (two-sided) inverse. Show

this cannot happen in two (not really different) ways.
(a) If 𝑖 𝑗 = 𝑎 + 𝑏𝑖 + 𝑐 𝑗 with 𝑎, 𝑏, 𝑐 ∈ R, multiply on the left by 𝑖 and derive a
contradiction.
(b) Show that 𝐷 is a (left) C-vector space, so 𝐷 has even dimension as an
R-vector space, a contradiction.
3. Show that there is no way to give R3 the structure of a ring (with 1) in which
multiplication respects scalar multiplication by R, i.e.,
𝑥 · (𝑐𝑦) = 𝑐(𝑥 · 𝑦) = (𝑐𝑥) · 𝑦 for all 𝑐 ∈ R and 𝑥, 𝑦 ∈ R3
and every nonzero element has a (two-sided) inverse, as follows.

1.5. CONCLUSION 17
(a) Suppose 𝐵 := R3 is equipped with a multiplication law that respects scalar

multiplication. Show that left multiplication by 𝛼 ∈ 𝐵 is R-linear and 𝛼
satisfies the characteristic polynomial of this linear map, a polynomial of
degree 3.
(b) Now suppose that every nonzero 𝛼 ∈ 𝐵 has an inverse. By consideration
of eigenvalues or the minimal polynomial, derive a contradiction. [Hint:
show that the characteristic polynomial has a real eigenvalue, or that
every 𝛼 ∈ 𝐵 satisfies a (minimal) polynomial of degree 1, and derive a
contradiction from either statement.]
Part I
Algebra
19
Chapter 2
Beginnings
In this chapter, we define quaternion algebras over fields by giving a multiplication

table, following Hamilton; we then consider the classical application of understanding
rotations in R3 .
2.1 ⊲ Conventions
Throughout this text (unless otherwise stated), we let 𝐹 be a (commutative) field with
algebraic closure 𝐹 al .
When 𝐺 is a group, and 𝐻 ⊆ 𝐺 is a subset, we write 𝐻 ≤ 𝐺 when 𝐻 is a subgroup
and 𝐻 E 𝐺 when 𝐻 is a normal subgroup; if 𝐺 is abelian (written multiplicatively),
we write 𝐺 𝑛 := {𝑔 𝑛 : 𝑔 ∈ 𝐺} ≤ 𝐺 for the subgroup of 𝑛th powers for 𝑛 ∈ Z>0 .
We suppose throughout that all rings are associative, not necessarily commutative,
with multiplicative identity 1, and that ring homomorphisms preserve 1. In particular, a
subring of a ring has the same 1. For a ring 𝐴, we write 𝐴× for the multiplicative group
of units of 𝐴. An algebra over the field 𝐹 is a ring 𝐵 equipped with a homomorphism
𝐹 → 𝐵 such that the image of 𝐹 lies in the center 𝑍 (𝐵) of 𝐵, defined by
𝑍 (𝐵) := {𝛼 ∈ 𝐵 : 𝛼𝛽 = 𝛽𝛼 for all 𝛽 ∈ 𝐵}; (2.1.1)
if 𝑍 (𝐵) = 𝐹, we say 𝐵 is central (as an 𝐹-algebra). We write M𝑛 (𝐹) for the 𝐹-algebra
of 𝑛 × 𝑛-matrices with entries in 𝐹.
One may profitably think of an 𝐹-algebra as being an 𝐹-vector space that is also
compatibly a ring. If the 𝐹-algebra 𝐵 is not the zero ring, then its structure map 𝐹 → 𝐵
is necessarily injective (since 1 maps to 1) and we identify 𝐹 with its image; keeping
track of the structure map just litters notation. The dimension dim𝐹 𝐵 of an 𝐹-algebra
𝐵 is its dimension as an 𝐹-vector space.
A homomorphism of 𝐹-algebras is a ring homomorphism which restricts to
the identity on 𝐹. An 𝐹-algebra homomorphism is necessarily 𝐹-linear. An 𝐹-
algebra homomorphism 𝐵 → 𝐵 is called an endomorphism. By convention (and
as usual for functions), endomorphisms act on the left. An invertible 𝐹-algebra
homomorphism 𝐵 −→ ∼ 𝐵 0 is called an isomorphism, and an invertible endomorphism
is an automorphism.
21
22 CHAPTER 2. BEGINNINGS
The set of automorphisms of 𝐵 forms a group, which we write as Aut(𝐵)—

these maps are necessarily 𝐹-linear, but we do not include this in the notation. We
reserve the notation End𝐹 (𝑉) for the ring of 𝐹-linear endomorphisms of the 𝐹-vector
space 𝑉, and Aut𝐹 (𝑉) for the group of 𝐹-linear automorphisms of 𝑉; in particular,
End𝐹 (𝐵) ∼ M𝑛 (𝐹) if 𝑛 = dim𝐹 𝐵.
Remark 2.1.2. Throughout, whenever we define a homomorphism of objects, we adopt
the (categorical) convention extending this to the terms endomorphism (homomor-
phism with equal domain and codomain), isomorphism (invertible homomorphism),
and automorphism (invertible endomorphism).
A division ring (also called a skew field) is a ring 𝐷 in which every nonzero
element has a (two-sided) inverse, i.e., 𝐷 r {0} is a group under multiplication. A
division algebra is an algebra that is a division ring.
2.2 ⊲ Quaternion algebras

In this section, we define quaternion algebras in a direct way, via generators and
relations. Throughout the rest of this chapter, suppose that char 𝐹 ≠ 2; the case
char 𝐹 = 2 is treated in Chapter 6.
Definition 2.2.1. An algebra 𝐵 over 𝐹 is a quaternion algebra if there exist 𝑖, 𝑗 ∈ 𝐵
such that 1, 𝑖, 𝑗, 𝑖 𝑗 is an 𝐹-basis for 𝐵 and
𝑖 2 = 𝑎, 𝑗 2 = 𝑏, and 𝑗𝑖 = −𝑖 𝑗 (2.2.2)
for some 𝑎, 𝑏 ∈ 𝐹 × .
The entire multiplication table for a quaternion algebra is determined by the mul-
tiplication rules (2.2.2), linearity, and associativity: for example,
(𝑖 𝑗) 2 = (𝑖 𝑗) (𝑖 𝑗) = 𝑖( 𝑗𝑖) 𝑗 = 𝑖(−𝑖 𝑗) 𝑗 = −(𝑖 2 ) ( 𝑗 2 ) = −𝑎𝑏
and 𝑗 (𝑖 𝑗) = (−𝑖 𝑗) 𝑗 = −𝑏𝑖. Conversely, given 𝑎, 𝑏 ∈ 𝐹 × , one can write down the unique
possible associative multiplication table on the basis 1, 𝑖, 𝑗, 𝑘 compatible with (2.2.2),
and then verify independently
that it is associative (Exercise 2.1). Accordingly, for
× 𝑎, 𝑏
𝑎, 𝑏 ∈ 𝐹 , we define to be the quaternion algebra over 𝐹 with 𝐹-basis 1, 𝑖, 𝑗, 𝑖 𝑗
𝐹
subject to the multiplication (2.2.2); we will also write (𝑎, 𝑏 | 𝐹) when convenient for
formatting. By definition, we have dim𝐹 (𝑎, 𝑏 | 𝐹) = 4.
𝑎, 𝑏 𝑏, 𝑎
The map which interchanges 𝑖 and 𝑗 gives an isomorphism ' ,
𝐹 𝐹
so Definition 2.2.1 is symmetric in 𝑎, 𝑏. The elements 𝑎, 𝑏 are far from unique in
determining the isomorphism class of a quaternion algebra: see Exercise 2.4.
If 𝐾 ⊇ 𝐹 is a field extension of 𝐹, then there is a canonical isomorphism

𝑎, 𝑏 𝑎, 𝑏
⊗𝐹 𝐾 '
𝐹 𝐾
extending scalars (same basis, but now spanning a 𝐾-vector space), so Definition 2.2.1
behaves well with respect to inclusion of fields.
2.2. ⊲ QUATERNION ALGEBRAS 23

−1, −1
Example 2.2.3. The R-algebra H := is the ring of quaternions over the real
R
numbers, discovered by Hamilton; we call H the ring of (real) Hamiltonians (also
known as Hamilton’s quaternions).
Example 2.2.4. The ring M2 (𝐹) of 2×2-matrices

with coefficients in 𝐹 is a quaternion
1, 1 ∼
algebra over 𝐹: there is an isomorphism −→ M2 (𝐹) of 𝐹-algebras induced by
𝐹

1 0 0 1
𝑖 ↦→ , 𝑗 ↦→ .
0 −1 1 0
If 𝐹 = 𝐹 al is algebraically closed and 𝐵 is a quaternion algebra over 𝐹, then

necessarily 𝐵 ' M2 (𝐹) (Exercise 2.4). Consequently, every quaternion algebra 𝐵 over
𝐹 has 𝐵 ⊗𝐹 𝐹 al ' M2 (𝐹 al ).
A quaternion algebra 𝐵 is generated by the elements 𝑖, 𝑗 by definition (2.2.2). How-

ever, exhibiting an algebra by generators and relations (instead of by a multiplication
table) can be a bit subtle, as the dimension of such an algebra is not a priori clear. But
working with presentations is quite useful; and at least for quaternion algebras, we can
think in these terms as follows.
Lemma 2.2.5. An 𝐹-algebra 𝐵 is a quaternion algebra if and only if there exist

nonzero 𝑖, 𝑗 ∈ 𝐵 that generate 𝐵 as an 𝐹-algebra and satisfy
𝑖 2 = 𝑎, 𝑗 2 = 𝑏, and 𝑖 𝑗 = − 𝑗𝑖 (2.2.6)
with 𝑎, 𝑏 ∈ 𝐹 × .
In other words, once the relations (2.2.6) are satisfied for generators 𝑖, 𝑗, then
automatically 𝐵 has dimension 4 as an 𝐹-vector space, with 𝐹-basis 1, 𝑖, 𝑗, 𝑖 𝑗.
Proof. It is necessary and sufficient to prove that the elements 1, 𝑖, 𝑗, 𝑖 𝑗 are linearly
independent. Suppose that 𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 = 0 with 𝑡, 𝑥, 𝑦, 𝑧 ∈ 𝐹. Using the
relations given, we compute that
0 = 𝑖(𝛼𝑖 + 𝑖𝛼) = 2𝑎(𝑡 + 𝑥𝑖).
Since char 𝐹 ≠ 2 and 𝑎 ≠ 0, we conclude that 𝑡 + 𝑥𝑖 = 0. Repeating with 𝑗 and 𝑖 𝑗, we

similarly find that 𝑡 + 𝑦 𝑗 = 𝑡 + 𝑧𝑖 𝑗 = 0. Thus
𝛼 − (𝑡 + 𝑥𝑖) − (𝑡 + 𝑦 𝑗) − (𝑡 + 𝑧𝑖 𝑗) = −2𝑡 = 0.
Since 𝑖, 𝑗 are nonzero, 𝐵 is not the zero ring, so 1 ≠ 0; thus 𝑡 = 0 and so 𝑥𝑖 = 𝑦 𝑗 =

𝑧𝑖 𝑗 = 0. Finally, if 𝑥 ≠ 0, then 𝑖 = 0 so 𝑖 2 = 0 = 𝑎, impossible; hence 𝑥 = 0. Similarly,
𝑦 = 𝑧 = 0.
Accordingly, we will call elements 𝑖, 𝑗 ∈ 𝐵 satisfying (2.2.6) standard generators

for a quaternion algebra 𝐵.
Remark 2.2.7. Invertibility of both 𝑎 and 𝑏 in 𝐹 is needed for Lemma 2.2.5: the
commutative algebra 𝐵 = 𝐹 [𝑖, 𝑗]/(𝑖, 𝑗) 2 is generated by the elements 𝑖, 𝑗 satisfying
𝑖 2 = 𝑗 2 = 𝑖 𝑗 = − 𝑗𝑖 = 0 but 𝐵 is not a quaternion algebra.
Remark 2.2.8. In light of Lemma 2.2.5, we will often drop the symbol 𝑘 = 𝑖 𝑗 and
reserve it for other use. (In particular, in later sections we will want 𝑘 to represent
other quaternion elements.) If we wish to use this abbreviation, we will assign 𝑘 := 𝑖 𝑗.
2.3 ⊲ Matrix representations

Every quaternion algebra can be viewed as a subalgebra of 2 × 2-matrices over an at
most quadratic extension; this is sometimes taken to be the definition!

𝑎, 𝑏 √
Proposition 2.3.1. Let 𝐵 := be a quaternion algebra over 𝐹 and let 𝐹 ( 𝑎) be
𝐹
√ √
a splitting field over 𝐹 for the polynomial 𝑥 2 − 𝑎, with root 𝑎 ∈ 𝐹 ( 𝑎). Then the
map
√
𝜆 : 𝐵 → M2 (𝐹 ( 𝑎))
√
𝑎 0
√ 0 𝑏
𝑖, 𝑗 ↦→ ,
0 − 𝑎 1 0 (2.3.2)
√ √
𝑡 + 𝑥 √𝑎 𝑏(𝑦 + 𝑧√ 𝑎)
𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 ↦→
𝑦−𝑧 𝑎 𝑡−𝑥 𝑎
is an injective 𝐹-algebra homomorphism and an isomorphism onto its image.
Proof. Injectivity follows by checking ker 𝜆 = {0} on matrix entries, and the homo-
morphism property can be verified directly, checking the multiplication table (Exercise
2.10).
Remark 2.3.3. Proposition 2.3.1 can be turned around to assert the existence of quater-
nion algebras: one can check that the set
√ √
√

𝑡 + 𝑥 √𝑎 𝑏(𝑦 + 𝑧√ 𝑎)
: 𝑡, 𝑥, 𝑦, 𝑧 ∈ 𝐹 ⊆ M2 (𝐹 ( 𝑎))
𝑦−𝑧 𝑎 𝑡−𝑥 𝑎
is an 𝐹-vector subspace of dimension 4, closed under multiplication, with the matrices

𝜆(𝑖), 𝜆( 𝑗) satisfying the defining relations (2.2.2).
√
2.3.4. If 𝑎 ∉ 𝐹 ×2 , then 𝐾 = 𝐹 ( 𝑎) ⊇ 𝐹 is a quadratic extension of 𝐹. Let Gal(𝐾 | 𝐹) =
Aut𝐹 (𝐾) ' Z/2Z be the Galois group of 𝐾 over 𝐹 and let 𝜎 ∈ Gal(𝐾 | 𝐹) be the
nontrivial element. Then we can rewrite the image 𝜆(𝐵) in (2.3.2) as

𝑢 𝑏𝑣
𝜆(𝐵) = : 𝑢, 𝑣 ∈ 𝐾 ⊂ M2 (𝐾). (2.3.5)
𝜎(𝑣) 𝜎(𝑢)
2.3. ⊲ MATRIX REPRESENTATIONS 25
Corollary 2.3.6. We have an isomorphism

1, 𝑏 ∼
−→ M2 (𝐹)
𝐹
(2.3.7)
1 0 0 𝑏
𝑖, 𝑗 ↦→ ,
0 −1 1 0
Proof. Specializing Proposition 2.3.1, we see the map is an injective 𝐹-algebra homo-
morphism, so since dim𝐹 𝐵 = dim𝐹 𝑀2 (𝐹) = 4, the map is also surjective.
The provenance of the map (2.3.2) is itself important, so we now pursue another
(more natural) proof of Proposition 2.3.1.
2.3.8. Let
𝐾 := 𝐹 [𝑖] = 𝐹 ⊕ 𝐹𝑖 ' 𝐹 [𝑥]/(𝑥 2 − 𝑎)
be the (commutative) 𝐹-algebra
√ generated by 𝑖. Suppose first that 𝐾 is a field (so
𝑎 ∉ 𝐹 ×2 ): then 𝐾 ' 𝐹 ( 𝑎) is a quadratic field extension of 𝐹. The algebra 𝐵 has the
structure of a right 𝐾-vector space of dimension 2, with basis 1, 𝑗: explicitly,
𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 = (𝑡 + 𝑥𝑖) + 𝑗 (𝑦 − 𝑧𝑖) ∈ 𝐾 ⊕ 𝑗 𝐾
for all 𝛼 ∈ 𝐵, so 𝐵 = 𝐾 ⊕ 𝑗 𝐾. We then define the left regular representation of 𝐵

over 𝐾 by
𝜆 : 𝐵 → End𝐾 (𝐵)
(2.3.9)
𝛼 ↦→ (𝜆 𝛼 : 𝛽 ↦→ 𝛼𝛽).
Each map 𝜆 𝛼 is indeed a 𝐾-linear endomorphism in 𝐵 (considered as a right 𝐾-vector
space) by associativity in 𝐵: for all 𝛼, 𝛽 ∈ 𝐵 and 𝑤 ∈ 𝐾,
𝜆 𝛼 (𝛽𝑤) = 𝛼(𝛽𝑤) = (𝛼𝛽)𝑤 = 𝜆 𝛼 (𝛽)𝑤.
Similarly, 𝜆 is an 𝐹-algebra homomorphism: for all 𝛼, 𝛽, 𝜈 ∈ 𝐵
𝜆 𝛼𝛽 (𝜈) = (𝛼𝛽)𝜈 = 𝛼(𝛽(𝜈)) = (𝜆 𝛼 𝜆 𝛽 ) (𝜈)
reading functions from right to left as usual. The map 𝜆 is injective (𝜆 is a faithful
representation) since 𝜆 𝛼 = 0 implies 𝜆 𝛼 (1) = 𝛼 = 0.
In the basis 1, 𝑗 we have End𝐾 (𝐵) ' M2 (𝐾), and 𝜆 is given by

𝑖 0 0 𝑏
𝑖 ↦→ 𝜆𝑖 = , 𝑗 ↦→ 𝜆 𝑗 = ; (2.3.10)
0 −𝑖 1 0
these matrices act on column vectors on the left. We then recognize the map 𝜆 given
in (2.3.2).
If 𝐾 is not a field, then 𝐾 ' 𝐹 × 𝐹, and we repeat the above argument but with
√ free module of rank 2 over 𝐾; then projecting onto one of the factors (choosing
𝐵 a
𝑎 ∈ 𝐹) gives the map 𝜆, which is still injective and therefore induces an 𝐹-algebra
isomorphism 𝐵 ' M2 (𝐹).
Remark 2.3.11. In Proposition 2.3.1, 𝐵 acts on columns on the left; if instead, one
wishes to have 𝐵 act on the right on rows, give 𝐵 the structure of a left 𝐾-vector space
and define accordingly the right regular representation instead (taking care about the
order of multiplication).
2.3.12. In some circumstances, it can be notationally convenient to consider variants

of the injection (2.3.2): for example
√
𝐵 → M2 (𝐹 ( 𝑎))
√ √
(2.3.13)

𝑡 + 𝑥 √𝑎 𝑦 + 𝑧√ 𝑎
𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 ↦→
𝑏(𝑦 − 𝑧 𝑎) 𝑡−𝑥 𝑎

1 0
is obtained by taking the basis 1, 𝑏 −1 𝑗, equivalently postcomposing by . See
0 𝑏
also Exercise 2.12.
Remark 2.3.14. The left regular representation 2.3.8 is not the only way to embed 𝐵
as a subalgebra of 2 × 2-matrices. Indeed, the “splitting” of quaternion algebras in
this way, in particular the question of whether or not 𝐵 ' M2 (𝐹), is a theme that will
reappear throughout this text. For a preview, see Main Theorem 5.4.4.
2.3.15. Thinking of a quaternion algebra as in 2.3.8 as a right 𝐾-vector space suggests

notation for quaternion algebras that is also useful: for a peek, see 6.1.5.
2.4 ⊲ Rotations
To conclude this chapter, we return to Hamilton’s original design: quaternions model
rotations in 3-dimensional space. This development is not only historically important
but it also previews many aspects of the general theory of quaternion algebras over
fields. In this section, we follow Hamilton and take 𝑘 := 𝑖 𝑗.
Proposition 2.3.1 provides an R-algebra embedding
𝜆 : H ↩→ EndC (H) ' M2 (C)

(2.4.1)

𝑡 + 𝑥𝑖 −𝑦 − 𝑧𝑖 𝑢 −𝑣
𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 = 𝑢 + 𝑗 𝑣 ↦→ =
𝑦 − 𝑧𝑖 𝑡 − 𝑥𝑖 𝑣 𝑢
where 𝑢 := 𝑡 + 𝑥𝑖 and 𝑣 := 𝑦 + 𝑧𝑖 and denotes complex conjugation. (The abuse of

notation, taking 𝑖 ∈ H as well as 𝑖 ∈ C is harmless: we may think of C ⊂ H.) We have

𝑢 −𝑣
det = |𝑢| 2 + |𝑣| 2 = 𝑡 2 + 𝑥 2 + 𝑦 2 + 𝑧2 ,
𝑣 𝑢

𝑢 𝑣
thus H× = H r {0}. If preferred, see (2.3.13) to obtain matrices of the form
−𝑣 𝑢
instead.
2.4. ⊲ ROTATIONS 27
2.4.2. We define the subgroup of unit Hamiltonians as
H1 := {𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 ∈ H : 𝑡 2 + 𝑥 2 + 𝑦 2 + 𝑧2 = 1}.
(In some contexts, one also writes GL1 (H) = H× and SL1 (H) = H1 .)
As a set, the unit Hamiltonians are naturally identified with the 3-sphere in R4 . As
groups, we have an isomorphism H1 ' SU(2) with the special unitary group of rank
2, defined by
SU(𝑛) := { 𝐴 ∈ SL𝑛 (C) : 𝐴∗ = 𝐴−1 } = { 𝐴 ∈ SL𝑛 (C) : 𝐽 𝐴 = 𝐴𝐽} (2.4.3)

t 0 −1
where 𝐴∗ = 𝐴 is the (complex) conjugate transpose of 𝐴 and 𝐽 := is the
1 0
1
image of 𝑗 ∈ H .
Definition 2.4.4. Let 𝛼 ∈ H. We say 𝛼 is real if 𝛼 ∈ R, and we say 𝛼 is pure (or

imaginary) if 𝛼 ∈ R𝑖 + R 𝑗 + R𝑘.
2.4.5. Just as for the complex numbers, every element of H is the sum of its real
part and its pure (imaginary) part. And just like complex conjugation, we define a
(quaternion) conjugation map
:H→H
(2.4.6)
𝛼 = 𝑡 + (𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘) ↦→ 𝛼 = 𝑡 − (𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘)
by negating the imaginary part. We compute that
𝛼 + 𝛼 = tr(𝜆(𝛼)) = 2𝑡
(2.4.7)
k𝛼k 2 := det(𝜆(𝛼)) = 𝛼𝛼 = 𝛼𝛼 = 𝑡 2 + 𝑥 2 + 𝑦 2 + 𝑧 2 .
The notation k k 2 is used to indicate that it agrees the usual square norm on H ' R4 .
The conjugate transpose map on M2 (C) restricts to quaternion conjugation on the
image of H in (2.4.1), also known as adjugation

𝑢 −𝑣 𝑢 𝑣
𝛼= ↦→ 𝜆(𝛼) = .
𝑣 𝑢 −𝑣 𝑢
Thus the elements 𝛼 ∈ H such that 𝜆(𝛼) = 𝜆(𝛼) (i.e., 𝐴∗ = 𝐴, and we say 𝐴 is
Hermitian), are exactly the scalar (real) matrices; and those that are skew-Hermitian,
i.e., 𝐴∗ = −𝐴, are exactly the pure quaternions. The conjugation map plays a crucial
role for quaternion algebras and is the subject of the next chapter (Chapter 3), where
to avoid confusion with other notions of conjugation we refer to it as the standard
involution.
2.4.8. Let
H0 := {𝑣 = 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 ∈ H : 𝑥, 𝑦, 𝑧 ∈ R} ' R3
be the set of pure Hamiltonians, the three-dimensional real space on which we will soon
see that the (unit) Hamiltonians act by rotations. (The reader should not confuse 𝑣 ∈ H0
with 𝑣 the entry of a 2 × 2-matrix in a local instantiation above.) For 𝑣 ∈ H0 ' R3 ,
k𝑣k 2 = 𝑥 2 + 𝑦 2 + 𝑧2 = det(𝜆(𝑣)), (2.4.9)
and from (2.4.1),

H0 = {𝑣 ∈ H : tr(𝜆(𝑣)) = 𝑣 + 𝑣 = 0}.
We again see that 𝑣 = −𝑣 for 𝑣 ∈ H0 .
The set H0 is not closed under multiplication: if 𝑣, 𝑤 ∈ H0 , then
𝑣𝑤 = −𝑣 · 𝑤 + 𝑣 × 𝑤 (2.4.10)
where 𝑣 · 𝑤 is the dot product on R3 and 𝑣 × 𝑤 ∈ H0 is the cross product, defined as

the determinant
𝑖 𝑗 𝑘
𝑣 × 𝑤 = det 𝑣 1 𝑣 2 𝑣 3 ® (2.4.11)
© ª
𝑤
« 1 𝑤 2 𝑤 3¬
where 𝑣 = 𝑣 1𝑖 + 𝑣 2 𝑗 + 𝑣 3 𝑘 and 𝑤 = 𝑤 1𝑖 + 𝑤 2 𝑗 + 𝑤 3 𝑘, so
𝑣 · 𝑤 = 𝑣1 𝑤1 + 𝑣2 𝑤2 + 𝑣3 𝑤3
and
𝑣 × 𝑤 = (𝑣 2 𝑤 3 − 𝑣 3 𝑤 2 )𝑖 + (𝑣 3 𝑤 1 − 𝑣 1 𝑤 3 ) 𝑗 + (𝑣 1 𝑤 2 − 𝑣 2 𝑤 1 )𝑘.
The formula (2.4.10) is striking: it contains three different kinds of ‘multiplications’!
Lemma 2.4.12. For all 𝑣, 𝑤 ∈ H0 , the following statements hold.
(a) 𝑣𝑤 ∈ H0 if and only if 𝑣, 𝑤 are orthogonal.

(b) 𝑣 2 = −k𝑣k 2 ∈ R.
(c) 𝑤𝑣 = −𝑣𝑤 if and only if 𝑣, 𝑤 are orthogonal.
Proof. Apply (2.4.10).
2.4.13. The group H1 acts on our three-dimensional space H0 (on the left) by conju-
gation:
H1 H0 → H0
(2.4.14)
𝑣 ↦→ 𝛼𝑣𝛼−1 ;
indeed, tr(𝜆(𝛼𝑣𝛼−1 )) = tr(𝜆(𝑣)) = 0 by properties of the trace, so 𝛼𝑣𝛼−1 ∈ H0 . Or
H0 = {𝑣 ∈ H : 𝑣 2 ∈ R ≤0 }
and this latter set is visibly stable under conjugation. The representation (2.4.14) is
called the adjoint representation.
2.4.15. Let 𝛼 ∈ H1 r {±1}. Then there exists a unique 𝜃 ∈ (0, 𝜋) such that
𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 = cos 𝜃 + (sin 𝜃)𝐼 (𝛼) (2.4.16)
where 𝐼 (𝛼) is pure and k𝐼 (𝛼) k = 1: to be precise, we take 𝜃 such that cos 𝜃 = 𝑡 and
𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘
𝐼 (𝛼) := .
|sin 𝜃|
We call 𝐼 (𝛼) the axis of 𝛼, and observe that 𝐼 (𝛼) 2 = −1.
Remark 2.4.17. In analogy with Euler’s formula, we can write (2.4.16) as
𝛼 = exp(𝐼 (𝛼)𝜃).
We are now prepared to identify this action by quaternions with rotations. As
usual, let
O(𝑛) := { 𝐴 ∈ M𝑛 (R) : 𝐴t = 𝐴−1 }
be the orthogonal group of R𝑛 (preserving the standard inner product), and let
SO(𝑛) := { 𝐴 ∈ O(𝑛) : det( 𝐴) = 1} E O(𝑛)
to be the special orthogonal group of rotations of R𝑛 , a normal subgroup of index 2
fitting into the exact sequence
det
1 → SO(𝑛) → O(𝑛) −−→ {±1} → 1.
Proposition 2.4.18. H1 acts by rotation on H0 ' R3 via conjugation (2.4.14): specif-
ically, 𝛼 acts by rotation through the angle 2𝜃 about the axis 𝐼 (𝛼).
Proof. Let 𝛼 ∈ H1 r {±1}. Then for all 𝑣 ∈ H0 ,
k𝛼𝑣𝛼−1 k 2 = k𝑣k 2
by (2.4.9), so 𝛼 acts by a matrix belonging to O(3).
But we can be more precise. Let 𝑗 0 ∈ H0 be a unit vector orthogonal to 𝑖 0 = 𝐼 (𝛼).
Then (𝑖 0) 2 = ( 𝑗 0) 2 = −1 by Lemma 2.4.12(b) and 𝑗 0𝑖 0 = −𝑖 0 𝑗 0 by Lemma 2.4.12(c),
so without loss of generality we may suppose that 𝐼 (𝛼) = 𝑖 and 𝑗 0 = 𝑗. Thus
𝛼 = 𝑡 + 𝑥𝑖 = cos 𝜃 + (sin 𝜃)𝑖 with 𝑡 2 + 𝑥 2 = cos2 𝜃 + sin2 𝜃 = 1, and 𝛼−1 = 𝑡 − 𝑥𝑖.
We have 𝛼𝑖𝛼−1 = 𝑖, and
𝛼 𝑗 𝛼−1 = (𝑡 + 𝑥𝑖) 𝑗 (𝑡 − 𝑥𝑖) = (𝑡 + 𝑥𝑖) (𝑡 + 𝑥𝑖) 𝑗
(2.4.19)
= ((𝑡 2 − 𝑥 2 ) + 2𝑡𝑥𝑖) 𝑗 = (cos 2𝜃) 𝑗 + (sin 2𝜃)𝑘
by the double angle formula. Consequently,
𝛼𝑘𝛼−1 = 𝑖(𝛼 𝑗 𝛼−1 ) = (− sin 2𝜃) 𝑗 + (cos 2𝜃)𝑘
so the matrix of 𝛼 in the basis 𝑖, 𝑗, 𝑘 is
1 0 0
𝐴 = 0 cos 2𝜃 − sin 2𝜃 ® , (2.4.20)
© ª
«0 sin 2𝜃 cos 2𝜃 ¬
a (counterclockwise) rotation (determinant 1) through the angle 2𝜃 about 𝑖.
Corollary 2.4.21. The action (2.4.13) defines a group homomorphism H1 → SO(3),

fitting into an exact sequence
1 → {±1} → H1 → SO(3) → 1.
Proof. The map H1 → SO(3) is surjective, since every element of SO(3) is rotation
about some axis (Exercise 2.15). If 𝛼 belongs to the kernel, then 𝛼 = cos 𝜃 +(sin 𝜃)𝐼 (𝛼)
must have sin 𝜃 = 0 so 𝛼 = ±1.
2.4.22. The matrix representation of H in section 2.4 extends to a matrix representation

of H ⊗R C, and this representation and its connection to unitary matrices is still used
widely in quantum mechanics. In the embedding with

𝑖 0 0 1 0 𝑖
𝑖 ↦→ , − 𝑗 ↦→ , −𝑘 ↦→
0 −𝑖 −1 0 𝑖 0
whose images are unitary matrices, we multiply by −𝑖 to obtain Hermitian matrices

1 0 0 −𝑖 0 1
𝜎𝑧 := , 𝜎𝑦 := , 𝜎𝑥 :=
0 −1 𝑖 0 1 0
where 𝜎𝑥 , 𝜎𝑦 , 𝜎𝑧 are the famous Pauli spin matrices. Because of this application to
the spin (a kind of angular momentum) of an electron in particle physics, the group
H1 also goes by the name H1 ' Spin(3).
The extra bit of information conveyed by spin can also be seen by the “belt trick”
[Hans2006, Chapter 2].
2.4.23. We conclude with one final observation, returning to the formula (2.4.10).
There is another way to mix the dot product and cross product (2.4.11) in H: we define
the scalar triple product
H×H×H→R
(2.4.24)
(𝑢, 𝑣, 𝑤) ↦→ 𝑢 · (𝑣 × 𝑤).
Amusingly, this gives a way to “multiply” triples of triples! The map (2.4.24) defines
an alternating, trilinear form (Exercise 2.19). If 𝑢, 𝑣, 𝑤 ∈ H0 , then the scalar triple
product is a determinant
𝑢 𝑢2 𝑢3
© 1
𝑢 · (𝑣 × 𝑤) = det 𝑣 1
ª
𝑣2 𝑣3 ®
«𝑤 1 𝑤2 𝑤3¬
and |𝑢 · (𝑣 × 𝑤)| is the volume of a parallelepiped in R3 whose sides are given by

𝑢, 𝑣, 𝑤.
Exercises
Let 𝐹 be a field with char 𝐹 ≠ 2.
I 1. Show that a (not necessarily associative) 𝐹-algebra is associative if and only if

the associative law holds on a basis, and then check that the multiplication table
implied by (2.2.2) is associative.
2. Show that if 𝐵 is an 𝐹-algebra generated by 𝑖, 𝑗 ∈ 𝐵 and 1, 𝑖, 𝑗 are linearly
dependent, then 𝐵 is commutative.

1, 1 ∼
3. Verify directly that the map −→ M2 (𝐹) in Example 2.2.4 is an isomor-
𝐹
phism of 𝐹-algebras.
I 4. Let 𝑎, 𝑏 ∈ 𝐹 × .

𝑎, 𝑏 𝑎, −𝑎𝑏 𝑏, −𝑎𝑏
(a) Show that ' ' .
𝐹 𝐹 𝐹 2
𝑎𝑐 , 𝑏𝑑 2

× 𝑎, 𝑏
(b) Show that if 𝑐, 𝑑 ∈ 𝐹 then ' . Conclude that if
𝐹 𝐹
× ×2
𝐹 /𝐹 is finite, then there are only finitely many isomorphism classes of
quaternion algebras over 𝐹, and ×2 ×
in particular that if 𝐹 = 𝐹 then there is
1, 1
only one isomorphism class ' M2 (𝐹). [The converse is not true,
𝐹
see Exercise 3.16.]
𝑎, 𝑏
(c) Show that if 𝐵 = is a quaternion algebra over R, then 𝐵 ' M2 (R)
R
or 𝐵 ' H, the latter occurring if and only if 𝑎 < 0 and 𝑏 < 0. Conclude
that if 𝐵 is a division quaternion algebra over R, then 𝐵 ' H.
(d) Let 𝐵 be a quaternion algebra over 𝐹. Show that 𝐵 ⊗𝐹 𝐹 al ' M2 (𝐹 al ),
where 𝐹 al is an algebraic closure of 𝐹.
(e) Refine part (d) as follows. A field 𝐾 ⊇ 𝐹 is a splitting field for 𝐵 if
𝐵 ⊗𝐹 𝐾 ' M2 (𝐾). Show that 𝐵 has a splitting field 𝐾 with [𝐾 : 𝐹] ≤ 2.

𝑎, 𝑏
5. Let 𝐵 = be a quaternion algebra over 𝐹. Let 𝑖 0 ∈ 𝐵 r 𝐹 satisfy (𝑖 0) 2 =
𝐹 0 0
𝑎 ,𝑏
𝑎 0 ∈ 𝐹 × . Show that there exists 𝑏 0 ∈ 𝐹 × and an isomorphism 𝐵 '
𝐹
(under which 𝑖 0 maps to the first standard generator).

−1, −1
6. Use the quaternion algebra 𝐵 = , multiplicativity of the determinant,
𝐹
and the left regular representation (2.3.2) to show that if two elements of 𝐹 can
be written as the sum of four squares, then so too can their product (a discovery
of Euler in 1748). [In Chapter 3, this statement will follow immediately from
the multiplicativity of the reduced norm on 𝐵; here, the formula is derived easily
from multiplicativity of the determinant.]
I 7. Let 𝐵 be an 𝐹-algebra. Show that if 𝐵 is a quaternion algebra over 𝐹, then 𝐵 is
central.
I 8. Let 𝐴, 𝐵 be 𝐹-algebras, and let 𝜙 : 𝐴 → 𝐵 be a surjective 𝐹-algebra homomor-
phism. Show that 𝜙 restricts to an 𝐹-algebra homomorphism 𝑍 ( 𝐴) → 𝑍 (𝐵).
I 9. Prove the following partial generalization of Exercise 2.4(b). Let 𝐵 be a finite-

dimensional algebra over 𝐹.
(a) Show that every element 𝛼 ∈ 𝐵 satisfies a unique monic polynomial of
smallest degree with coefficients in 𝐹.
(b) Suppose that 𝐵 = 𝐷 is a division algebra. Show that the minimal poly-
nomial of 𝛼 ∈ 𝐷 is irreducible over 𝐹. Conclude that if 𝐹 = 𝐹 al is
algebraically closed, then 𝐷 = 𝐹.
I 10. Prove Proposition 2.3.1: show directly that the map
√
𝜆 : 𝐵 → M2 (𝐹 ( 𝑎))
√
𝑎 0
√ , 0 𝑏
𝑖, 𝑗 ↦→
0 − 𝑎 1 0
extends uniquely to an injective 𝐹-algebra homomorphism. [Hint: check that
the relations are satisfied.]
11. (a) Show explicitly that every quaternion algebra 𝐵 = (𝑎, 𝑏 | 𝐹) is isomorphic
to an 𝐹-subalgebra of M4 (𝐹) via the left (or right) regular representation
over 𝐹: write down 4 × 4-matrices representing 𝑖 and 𝑗 and verify that
the relations 𝑖 2 = 𝑎, 𝑗 2 = 𝑏, 𝑗𝑖 = −𝑖 𝑗 hold for these matrices. Note the
2 × 2-block structure of these matrices.
(b) With respect to a suitable such embedding in (a) for 𝐵 = H, verify that
the quaternionic conjugation map 𝛼 ↦→ 𝛼 is the matrix transpose, and the
matrix determinant is the square of the norm k𝛼k 2 = 𝛼𝛼.
12. In certain circumstances, one may not want to “play favorites” in the left regular
representation (Proposition 2.3.1) and so involve 𝑖 and 𝑗 on more equal footing.
To this end, show that the map
√ √

𝑎, 𝑏
𝐵= → M2 (𝐹 ( 𝑎, 𝑏))
𝐹
√ √ (2.4.25)
𝑎 0 0 𝑏
𝑖, 𝑗 ↦→ √ , √
0 − 𝑎 𝑏 0
is an injective 𝐹-algebra isomorphism. How is it related to the left regular
representation?
13. Let 𝐵 = (𝑎, 𝑏 | 𝐹) be a quaternion algebra over 𝐹. For a nonzero element
𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 ∈ 𝐵, show that the following are equivalent:
(i) 𝑡 = 0; and
(ii) 𝛼 ∉ 𝐹 and 𝛼2 ∈ 𝐹.
[So the notion of “pure quaternion” is not tethered to a particular basis.]
14. Verify that (2.3.7) is an isomorphism of 𝐹-algebras, and interpret this map as
arising from the left regular representation via a map 𝐵 ↩→ M2 (𝐹 × 𝐹) →
M2 (𝐹).
15. Show that every rotation 𝐴 ∈ SO(3) fixes an axis. [Hint: Consider the eigen-
values of 𝐴.]
16. For 𝑣 ∈ H0 and 𝛽 ∈ H0 r {0}, consider the map 𝑣 ↦→ 𝛽−1 𝑣 𝛽 = −𝛽−1 𝑣 𝛽 ∈ H0 .

Show that this map is the reflection across the plane {𝑤 ∈ H0 : 𝛽𝑤 = 0}. (For
example, taking 𝛽 = 𝑖, the map is 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 ↦→ −𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘.)
I 17. In Corollary 2.4.21, we showed that SU(2) ' H1 has a 2-to-1 map to SO(3),
where H1 acts on H0 ' R3 by conjugation: quaternions model rotations in
three-dimensional space, with spin. Quaternions also model rotations in four-
dimensional space, as follows.
(a) Show that the map
(H1 × H1 ) H → H
(2.4.26)
𝑥 ↦→ 𝛼𝑥 𝛽−1
defines a (left) action of H1 ×H1 on H ' R4 , giving a group homomorphism
𝜙 : H1 × H1 → O(4).
(b) Show that 𝜙 surjects onto SO(4) < O(4). [Hint: If 𝐴 ∈ SO(4) fixes 1 ∈ H,
then 𝐴 restricted to H0 is a rotation and so is given by conjugation. More
generally, if 𝐴1 = 𝛼, consider 𝑥 ↦→ 𝛼−1 𝐴𝑥.]
(c) Show that the kernel of 𝜙 is {±1} embedded diagonally, so there is an exact
sequence
1 → {±1} → SU(2) × SU(2) → SO(4) → 1.
[More generally, the universal cover of SO(𝑛) for 𝑛 ≥ 3 is a double cover called
the spin group Spin(𝑛), and so Corollary 2.4.21 shows that Spin(3) ' SU(2)
and this exercise shows that Spin(4) ' SU(2) × SU(2). For further reading, see
e.g. Fulton–Harris [FH91, Lecture 20].]
18. Let 𝜌𝑢, 𝜃 : R3 → R3 be the counterclockwise rotation by the angle 𝜃 about the
axis 𝑢 ∈ R3 ' H0 , with k𝑢k = 1. Prove Rodrigues’s rotation formula: for all
𝑣 ∈ R3 ,
𝜌𝑢, 𝜃 (𝑣) = (cos 𝜃)𝑣 + (sin 𝜃) (𝑢 × 𝑣) + (1 − cos 𝜃) (𝑢 · 𝑣)𝑢
where 𝑢 × 𝑣 and 𝑢 · 𝑣 are the cross and dot product, respectively.

19. Verify that the map (2.4.24) is a trilinear alternating form on H, i.e., show the
form is linear when any two of the three arguments are fixed and zero when two
argument are equal.
20. Let 𝐵 be a quaternion algebra over 𝐹 and let M2 (𝐵) be the ring of 2 × 2-
matrices over 𝐵. (Be careful in the definition of matrix multiplication: 𝐵 is
noncommutative!) Consider the Cayley determinant:
Cdet : M2 (𝐵) → 𝐵

𝛼 𝛽
Cdet = 𝛼𝛿 − 𝛾 𝛽
𝛾 𝛿
(a) Show that Cdet is 𝐹-multilinear in the rows and columns of the matrix.
(b) Show that Cdet is not left 𝐵-multilinear in the rows of the matrix.
(c) Give an example showing that Cdet is not multiplicative.
(d) Find a matrix 𝐴 ∈ M2 (H) that is invertible (i.e., having a two-sided inverse)
but has Cdet( 𝐴) = 0. Then find such an 𝐴 with the further property that
its transpose has nonzero determinant but is not invertible.
[Moral: be careful with matrix rings over noncommutative rings! For more on
quaternionic determinants, including the Dieudonné determinant, see Aslaksen
[Asl96].]
Chapter 3
Involutions
In this chapter, we define the standard involution on a quaternion algebra. In this

way, we characterize division quaternion algebras as noncommutative division rings
equipped with a standard involution.
3.1 ⊲ Conjugation
The quaternion conjugation map (2.4.6) defined on the Hamiltonians H arises naturally
from the notion of real and pure (imaginary) parts, as defined by Hamilton. This
involution has a natural generalization to a quaternion algebra 𝐵 = (𝑎, 𝑏 | 𝐹) over a
field 𝐹 with char 𝐹 ≠ 2: we define
:𝐵→𝐵
𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 ↦→ 𝛼 = 𝑡 − (𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗)
Multiplying out, we then verify that
𝛼𝛼 = 𝛼𝛼 = 𝑡 2 − 𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧 2 ∈ 𝐹.
The way in which the cross terms cancel, because the basis elements 𝑖, 𝑗, 𝑖 𝑗 skew
commute, is a calculation that never fails to enchant!
But this definition seems to depend on a basis: it is not intrinsically defined.
What properties characterize it? Is it unique? We are looking for a good definition
of conjugation : 𝐵 → 𝐵 on an 𝐹-algebra 𝐵: we will call such a map a standard
involution.
The involutions we consider should have basic linearity properties: they are 𝐹-
linear (with 1 = 1, so they act as the identity on 𝐹) and have order 2 as an 𝐹-linear map.
An involution should also respect the multiplication structure on 𝐵, but we should not
require that it be an 𝐹-algebra isomorphism: instead, like the inverse map (or transpose
map) reverses order of multiplication, we ask that 𝛼𝛽 = 𝛽 𝛼 for all 𝛼 ∈ 𝐵. Finally, we
want the standard involution to give rise to a trace and norm (a measure of size), which
is to say, we want 𝛼 + 𝛼 ∈ 𝐹 and 𝛼𝛼 = 𝛼𝛼 ∈ 𝐹 for all 𝛼 ∈ 𝐵. The precise definition
is given in Definition 3.2.1, and the defining properties are rigid: if an algebra 𝐵 has a
standard involution, then it is necessarily unique (Corollary 3.4.4).
35
36 CHAPTER 3. INVOLUTIONS
The existence of a standard involution on 𝐵 implies that every element of 𝐵

satisfies a quadratic equation: by direct substitution, we see that 𝛼 ∈ 𝐵 is a root of the
polynomial 𝑥 2 − 𝑡𝑥 + 𝑛 ∈ 𝐹 [𝑥] where 𝑡 := 𝛼 + 𝛼 and 𝑛 := 𝛼𝛼 = 𝛼𝛼, since then
𝛼2 − (𝛼 + 𝛼)𝛼 + 𝛼𝛼 = 0
identically. Accordingly, we define the reduced trace trd : 𝐵 → 𝐹 by trd(𝛼) = 𝛼 + 𝛼

and reduced norm nrd : 𝐵 → 𝐹 by nrd(𝛼) = 𝛼𝛼. We observe that trd is 𝐹-linear and
nrd is multiplicative on 𝐵× .
Motivated by this setting, we say that 𝐵 has degree 2 if every element 𝛼 ∈ 𝐵
satisfies a (monic) polynomial in 𝐹 [𝑥] of degree 2 and, to avoid trivialities, that 𝐵 ≠ 𝐹
(or equivalently, at least one element of 𝐵 satisfies no polynomial of degree 1). The
final result of this section is the following theorem (see Theorem 3.5.1).
Theorem 3.1.1. Let 𝐵 be a division 𝐹-algebra of degree 2 over a field 𝐹 with char 𝐹 ≠
2. Then either 𝐵 = 𝐾 is a quadratic field extension of 𝐹 or 𝐵 is a division quaternion
algebra over 𝐹.
As a consequence, division quaternion algebras are characterized as noncommu-

tative division algebras with a standard involution, when char 𝐹 ≠ 2.
3.2 Involutions
Throughout this chapter, let 𝐵 be an 𝐹-algebra. For the moment, we allow 𝐹 to be of
arbitrary characteristic. We begin by defining involutions on 𝐵.
Definition 3.2.1. An involution : 𝐵 → 𝐵 is an 𝐹-linear map which satisfies:
(i) 1 = 1;
(ii) 𝛼 = 𝛼 for all 𝛼 ∈ 𝐵; and
(iii) 𝛼𝛽 = 𝛽 𝛼 for all 𝛼, 𝛽 ∈ 𝐵 (the map is an anti-automorphism).
3.2.2. We define the opposite algebra of 𝐵 by letting 𝐵op = 𝐵 as 𝐹-vector spaces but
with multiplication 𝛼 ·op 𝛽 = 𝛽 · 𝛼 for 𝛼, 𝛽 ∈ 𝐵.
One can then equivalently define an involution to be an 𝐹-algebra isomorphism
∼ 𝐵op whose underlying 𝐹-linear map has order at most 2.
𝐵 −→
Remark 3.2.3. What we have defined to be an involution is known in other contexts as

an involution of the first kind. An involution of the second kind is a map which acts
nontrivially when restricted to 𝐹, and hence is not 𝐹-linear; although these involutions
are interesting in other contexts, they will not figure in our discussion (and anyway
one can consider such an algebra over the fixed field of the involution).
Definition 3.2.4. An involution is standard if 𝛼𝛼 ∈ 𝐹 for all 𝛼 ∈ 𝐵.
Remark 3.2.5. Standard involutions go by many other names. The terminology stan-
dard is employed because conjugation on a quaternion algebra is the “standard” ex-
ample of such an involution. Other authors call the standard involution the main
3.2. INVOLUTIONS 37
involution for quaternion algebras, but then find situations where the “main” involu-
tion is not standard by our definition. The standard involution is also called conjugation
on 𝐵, but this can be confused with conjugation by an element in 𝐵× . We will see in
Corollary 3.4.4 that a standard involution is unique, so it is also called the canonical
involution; however, there are other circumstances where involutions can be defined
canonically that are not standard (like the map induced by 𝑔 ↦→ 𝑔 −1 on the group ring
𝐹 [𝐺]).
3.2.6. If is a standard involution, so that 𝛼𝛼 ∈ 𝐹 for all 𝛼 ∈ 𝐵, then
(𝛼 + 1) (𝛼 + 1) = (𝛼 + 1) (𝛼 + 1) = 𝛼𝛼 + 𝛼 + 𝛼 + 1 ∈ 𝐹
and hence 𝛼 + 𝛼 ∈ 𝐹 for all 𝛼 ∈ 𝐵 as well; it then also follows that 𝛼𝛼 = 𝛼𝛼, since
(𝛼 + 𝛼)𝛼 = 𝛼(𝛼 + 𝛼).
Example 3.2.7. The identity map is a standard involution on 𝐵 = 𝐹 as an 𝐹-algebra.

The R-algebra C has a standard involution, namely, complex conjugation.
Example 3.2.8. The adjugate map

𝑎 𝑏 † 𝑑 −𝑏
𝐴= ↦→ 𝐴 =
𝑐 𝑑 −𝑐 𝑎
is a standard involution on M2 (𝐹) since 𝐴𝐴† = 𝐴† 𝐴 = 𝑎𝑑 − 𝑏𝑐 = det 𝐴 ∈ 𝐹.

Matrix transpose is an involution on M𝑛 (𝐹) but is a standard involution (if and)
only if 𝑛 = 1.
3.2.9. Suppose char 𝐹 ≠ 2 and let 𝐵 = (𝑎, 𝑏 | 𝐹). Then the map
:𝐵→𝐵
𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 ↦→ 𝛼 = 𝑡 − 𝑥𝑖 − 𝑦 𝑗 − 𝑧𝑖 𝑗
defines a standard involution on 𝐵 and 𝛼 = 2𝑡 − 𝛼. The map is 𝐹-linear with 1 = 1 and

𝛼 = 𝛼, so properties (i) and (ii) hold. By 𝐹-linearity, it is enough to check property
(iii) on a basis (Exercise 3.1), and we verify for instance that
𝑖 𝑗 = 𝑖 𝑗 = −𝑖 𝑗 = 𝑗𝑖 = (− 𝑗) (−𝑖) = 𝑗 𝑖
(see Exercise 3.3). Finally, the involution is standard because
(𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗) (𝑡 − 𝑥𝑖 − 𝑦 𝑗 − 𝑧𝑖 𝑗) = 𝑡 2 − 𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧2 ∈ 𝐹. (3.2.10)
Remark 3.2.11. Algebras with involution play an important role in analysis, in particu-
lar Banach algebras with involution and 𝐶 ∗ -algebras (generally of infinite dimension).
A good reference is the text by Dixmier [Dix77] (or the more introductory book by
Conway [Con2012]).
3.3 Reduced trace and reduced norm

Let : 𝐵 → 𝐵 be a standard involution on 𝐵. We define the reduced trace on 𝐵 by
trd : 𝐵 → 𝐹
(3.3.1)
𝛼 ↦→ 𝛼 + 𝛼
and similarly the reduced norm
nrd : 𝐵 → 𝐹
(3.3.2)
𝛼 ↦→ 𝛼𝛼.
Example 3.3.3. For 𝐵 = M2 (𝐹), equipped with the adjugate map as a standard
involution as in Example 3.2.8, the reduced trace is the usual matrix trace and the
reduced norm is the determinant.
3.3.4. The reduced trace trd is an 𝐹-linear map, since this is true for the standard
involution:
trd(𝛼 + 𝛽) = (𝛼 + 𝛽) + (𝛼 + 𝛽) = (𝛼 + 𝛼) + (𝛽 + 𝛽) = trd(𝛼) + trd(𝛽)
for 𝛼, 𝛽 ∈ 𝐵. The reduced norm nrd is multiplicative, since
nrd(𝛼𝛽) = (𝛼𝛽)(𝛼𝛽) = 𝛼𝛽𝛽 𝛼 = 𝛼 nrd(𝛽)𝛼 = nrd(𝛼) nrd(𝛽)
for all 𝛼, 𝛽 ∈ 𝐵.
It will be convenient to write
𝐵0 := {𝛼 ∈ 𝐵 : trd(𝛼) = 0}
(3.3.5)
𝐵1 := {𝛼 ∈ 𝐵× : nrd(𝛼) = 1}
for the 𝐹-subspace 𝐵0 ⊆ 𝐵 of elements of reduced trace 0 and for the subgroup
𝐵1 ≤ 𝐵× of elements of reduced norm 1. We observe that 𝐵1 E 𝐵× is normal, by
multiplicativity, indeed we have an exact sequence of groups
nrd
1 → 𝐵1 → 𝐵× −−→ 𝐹 ×
(noting that the reduced norm map need not be surjective).
Lemma 3.3.6. If 𝐵 is not the zero ring, then 𝛼 ∈ 𝐵 is a unit (has a two-sided inverse)
if and only if nrd(𝛼) ≠ 0.
Proof. Exercise 3.5.
Lemma 3.3.7. For all 𝛼, 𝛽 ∈ 𝐵, we have trd(𝛽𝛼) = trd(𝛼𝛽).

3.4. UNIQUENESS AND DEGREE 39
Proof. We have
trd(𝛼𝛽) = trd(𝛼(trd(𝛽) − 𝛽)) = trd(𝛼) trd(𝛽) − trd(𝛼𝛽)
and so
trd(𝛼𝛽) = trd(𝛼𝛽) = trd(𝛽𝛼) = trd(𝛼) trd(𝛽) − trd(𝛽𝛼)
therefore trd(𝛼𝛽) = trd(𝛽𝛼).
Remark 3.3.8. The maps trd and nrd are called reduced for the following reason.
Let 𝐴 be a finite-dimensional 𝐹-algebra, and consider the left regular representation
𝜆 : 𝐴 ↩→ End𝐹 ( 𝐴) given by left multiplication in 𝐴 (cf. Proposition 2.3.1, but over 𝐹).
We then have a (left) trace map Tr : 𝐴 → 𝐹 and (left) norm map Nm : 𝐴 → 𝐹 given
by mapping 𝛼 ∈ 𝐵 to the trace and determinant of the endomorphism 𝜆 𝛼 ∈ End𝐹 ( 𝐴).
When 𝐴 = M2 (𝐹), a direct calculation (Exercise 3.13) reveals that
Tr(𝛼) = 2 trd(𝛼) = 2 tr(𝛼)
(algebra trace, reduced trace, and matrix trace, respectively; there is no difference
between left and right), and
Nm(𝛼) = nrd(𝛼) 2 = det(𝛼) 2
for all 𝛼 ∈ 𝐴, whence the name reduced. (To preview the language of chapter 7, this
calculation can be efficiently summarized: as a left 𝐴-module, 𝐴 is the sum of two
simple 𝐴-modules—acting on the columns of a matrix—and the reduced trace and
reduced norm represent ‘half’ of this action.)
3.3.9. Since
𝛼2 − (𝛼 + 𝛼)𝛼 + 𝛼𝛼 = 0 (3.3.10)
identically we see that 𝛼 ∈ 𝐵 is a root of the polynomial
𝑥 2 − trd(𝛼)𝑥 + nrd(𝛼) ∈ 𝐹 [𝑥] (3.3.11)
which we call the reduced characteristic polynomial of 𝛼. The fact that 𝛼 satisfies its
reduced characteristic polynomial is the reduced Cayley–Hamilton theorem for an
algebra with standard involution. When 𝛼 ∉ 𝐹, the reduced characteristic polynomial
of 𝛼 is its minimal polynomial, since if 𝛼 satisfies a polynomial of degree 1 then 𝛼 ∈ 𝐹.
3.4 Uniqueness and degree

Definition 3.4.1. An 𝐹-algebra 𝐾 with dim𝐹 𝐾 = 2 is called a quadratic algebra.
Lemma 3.4.2. Let 𝐾 be a quadratic 𝐹-algebra. Then 𝐾 is commutative and has a

unique standard involution.
Proof. Let 𝛼 ∈ 𝐾 r 𝐹. Then 𝐾 = 𝐹 ⊕ 𝐹𝛼 = 𝐹 [𝛼], so in particular 𝐾 is commutative.

Then 𝛼2 = 𝑡𝛼 − 𝑛 for unique 𝑡, 𝑛 ∈ 𝐹, since 1, 𝛼 is a basis for 𝐾.
If : 𝐾 → 𝐾 is any standard involution, then from (3.3.10) and uniqueness we
conclude 𝑡 = 𝛼 + 𝛼 (and 𝑛 = 𝛼𝛼), and so any involution must have 𝛼 = 𝑡 − 𝛼. On
the other hand, there is a unique standard involution 𝑥 : 𝐵 → 𝐵 with 𝛼 = 𝑡 − 𝛼: the
verification is straightforward (see Exercise 3.2).
Example 3.4.3. The reduced trace and norm on a quadratic algebra are precisely the
usual algebra trace and norm. If char 𝐹 ≠ 2 and 𝐾 ⊇ 𝐹 is a quadratic field extension
of 𝐹, then the standard involution is just the nontrivial element of Gal(𝐾 | 𝐹).
Corollary 3.4.4. If 𝐵 has a standard involution, then this involution is unique.
Proof. For any 𝛼 ∈ 𝐵r𝐹, we have from (3.3.10) that dim𝐹 𝐹 [𝛼] = 2, so the restriction
of the standard involution to 𝐹 [𝛼] is unique. Therefore the standard involution on 𝐵
is itself unique.
We have seen that the equation (3.3.10), implying that if 𝐵 has a standard involution
then every 𝛼 ∈ 𝐵 satisfies a quadratic equation, has figured prominently in the above
proofs. To further clarify the relationship between these two notions, we make the
following definition.
Definition 3.4.5. The degree of 𝐵 is the smallest 𝑚 ∈ Z ≥0 such that every element
𝛼 ∈ 𝐵 satisfies a monic polynomial 𝑓 (𝑥) ∈ 𝐹 [𝑥] of degree 𝑚, if such an integer exists;
otherwise, we say 𝐵 has degree ∞.
3.4.6. If 𝐵 has finite dimension 𝑛 = dim𝐹 𝐵 < ∞, then every element of 𝐵 satisfies
a polynomial of degree at most 𝑛: if 𝛼 ∈ 𝐵 then the elements 1, 𝛼, . . . , 𝛼 𝑛 are
linearly dependent over 𝐹. Consequently, every finite-dimensional 𝐹-algebra has a
(well-defined) integer degree, at most 𝑛.
Example 3.4.7. By convention, we interpret Definition 3.4.5 as defining the degree

of the zero ring to be 0 (since 1 = 0, the element 0 satisfies the monic polynomial
0𝑥)—whatever!
If 𝐵 has degree 1, then 𝐵 = 𝐹. If 𝐵 has a standard involution, then either 𝐵 = 𝐹 or
𝐵 has degree 2 by (3.3.11).
3.5 Quaternion algebras

We are now ready to characterize division algebras of degree 2 when char 𝐹 ≠ 2. (For
the case char 𝐹 = 2, see Chapter 6.)
Theorem 3.5.1. Suppose char 𝐹 ≠ 2 and let 𝐵 be a division 𝐹-algebra. Then 𝐵 has
degree at most 2 if and only if one of the following hold:
(i) 𝐵 = 𝐹;
(ii) 𝐵 = 𝐾 is a quadratic field extension of 𝐹; or
(iii) 𝐵 is a division quaternion algebra over 𝐹.
3.5. QUATERNION ALGEBRAS 41
Proof. From Example 3.4.7, we may suppose that 𝐵 ≠ 𝐹 and 𝐵 has degree 2.
Let 𝑖 ∈ 𝐵 r 𝐹. Then 𝐹 [𝑖] = 𝐾 is a (commutative) quadratic 𝐹-subalgebra of the
division ring 𝐵, so 𝐾 = 𝐹 (𝑖) is a field. If 𝐾 = 𝐵, we are done. Completing the square
(since char 𝐹 ≠ 2), we may suppose that 𝑖 2 = 𝑎 ∈ 𝐹 × .
Let 𝜙 : 𝐵 → 𝐵 be the map given by conjugation by 𝑖, i.e., 𝜙(𝛼) = 𝑖 −1 𝛼𝑖. Then 𝜙
is a 𝐾-linear endomorphism of 𝐵, thought of as a (left) 𝐾-vector space, and 𝜙2 is the
identity on 𝐵. Therefore 𝜙 is diagonalizable, and we may decompose 𝐵 = 𝐵+ ⊕ 𝐵−
into eigenspaces for 𝜙: explicitly, we can always write
𝛼 + 𝜙(𝛼) 𝛼 − 𝜙(𝛼)
𝛼= + ∈ 𝐵+ ⊕ 𝐵− .
2 2
We now prove dim𝐾 𝐵+ = 1. Let 𝛼 ∈ 𝐵+ . Then 𝐿 = 𝐹 (𝛼, 𝑖) is a field. Since
char 𝐹 ≠ 2, and 𝐿 is a compositum of quadratic extensions of 𝐹, the primitive element
theorem implies that 𝐿 = 𝐹 (𝛽) for some 𝛽 ∈ 𝐿. But by hypothesis 𝛽 satisfies a
quadratic equation so dim𝐹 𝐿 = 2 and hence 𝐿 = 𝐾. (For an alternative direct proof
of this claim, see Exercise 3.10.)
If 𝐵 = 𝐵+ = 𝐾, we are done. So suppose 𝐵− ≠ {0}. We will prove that
dim𝐾 𝐵− = 1. If 0 ≠ 𝑗 ∈ 𝐵− then 𝑖 −1 𝑗𝑖 = − 𝑗, so 𝑖 = − 𝑗 −1𝑖 𝑗 and hence all elements of
𝐵− conjugate 𝑖 to −𝑖. Thus if 0 ≠ 𝑗1 , 𝑗2 ∈ 𝐵− then 𝑗 1 𝑗 2 centralizes 𝑖 and 𝑗 1 𝑗 2 ∈ 𝐵+ = 𝐾.
Thus any two nonzero elements of 𝐵− are 𝐾-multiples of each other.
Finally, let 𝑗 ∈ 𝐵− r {0}; then 𝐵 = 𝐵+ ⊕ 𝐵− = 𝐾 ⊕ 𝐾 𝑗 so 𝐵 has 𝐹-basis 1, 𝑖, 𝑗, 𝑖 𝑗
and 𝑗𝑖 = −𝑖 𝑗. We claim that trd( 𝑗) = 0: indeed, both 𝑗 and 𝑖 −1 𝑗𝑖 = − 𝑗 satisfy
the same reduced characteristic (or minimal) polynomial of degree 2, so trd( 𝑗) =
trd(− 𝑗) = − trd( 𝑗) so trd( 𝑗) = 0. Thus 𝑗 2 = 𝑏 ∈ 𝐹 × , and 𝐵 is a quaternion algebra by
definition.
Remark 3.5.2. We need not assume in Theorem 3.5.1 that 𝐵 is finite-dimensional;

somehow, it is a consequence, and every division algebra over 𝐹 (with char 𝐹 ≠ 2) of
degree ≤ 2 is finite-dimensional.
There are algebras of arbitary (finite or infinite) dimension over 𝐹 of degree 2: see
Exercise 3.15. Also, a boolean ring (see Exercise 3.12) has degree 2 as an F2 -algebra,
and there are such rings of arbitrary dimension over F2 . Such algebras are quite far
from being division rings, of course.
Remark 3.5.3. The proof of Theorem 3.5.1 has quite a bit of history, discussed by van
Praag [vPr2002] (along with several proofs). See Lam [Lam2005, Theorem III.5.1] for
a parallel proof of Theorem 3.5.1. Moore [Moore35, Theorem 14.4] in 1915 studied
algebra of matrices over skew fields and in particular the role of involutions, and gives
an elementary proof of this theorem (with the assumption char 𝐹 ≠ 2). Dieudonné
[Die48, Die53] gave another proof that relies on structure theory for finite-dimensional
division algebras.
Corollary 3.5.4. Let 𝐵 be a division 𝐹-algebra with char 𝐹 ≠ 2. Then 𝐵 has degree
at most 2 if and only if 𝐵 has a standard involution.
Proof. In each of the cases (i)–(iii), 𝐵 has a standard involution; and conversely if 𝐵
has a standard involution, then 𝐵 has degree at most 2 (Example 3.4.7).
Remark 3.5.5. The statement of Corollary 3.5.4 holds more generally—even if 𝐵 is

not necessarily a division ring—as follows. Let 𝐵 be an 𝐹-algebra with char 𝐹 ≠ 2.
Then 𝐵 has a standard involution if and only if 𝐵 has degree at most 2 [Voi2011b].
However, this is no longer true in characteristic 2 (Exercise 3.12).
Corollary 3.5.6. Let 𝐵 be a division 𝐹-algebra with char 𝐹 ≠ 2. Then the following
are equivalent:
(i) 𝐵 is a quaternion algebra;

(ii) 𝐵 is noncommutative and has degree 2; and
(iii) 𝐵 is central and has degree 2.
Definition 3.5.7. An 𝐹-algebra 𝐵 is algebraic if every 𝛼 ∈ 𝐵 is algebraic over 𝐹 (i.e.,

𝛼 satisfies a polynomial with coefficients in 𝐹).
If 𝐵 has finite degree (such as when dim𝐹 𝐵 = 𝑛 < ∞), then 𝐵 is algebraic.
Corollary 3.5.8 (Frobenius). Let 𝐵 be an algebraic division algebra over R. Then

either 𝐵 = R or 𝐵 ' C or 𝐵 ' H as R-algebras.
Proof. If 𝛼 ∈ 𝐵 r R then R(𝛼) ' C, so 𝛼 satisfies a polynomial of degree 2. Thus if

𝐵 ≠ R then 𝐵 has degree 2 and either 𝐵 ' C or 𝐵 is a division quaternion algebra over
R, and hence 𝐵 ' H by Exercise 2.4(c).
Example 3.5.9. Division algebras over R of infinite dimension abound. Transcen-

dental field extensions of R, such as the function field R(𝑥) or the Laurent series
field R((𝑥)), are examples of infinite-dimensional division algebras over R. Also, the
free algebra in two (noncommuting) variables is a subring of a division ring 𝐵 (its
“noncommutative ring of fractions”) with center R and of infinite dimension over R.
Remark 3.5.10. The theorem of Frobenius (Corollary 3.5.8) extends directly to fields
𝐹 akin to R, as follows. A field is formally real if −1 cannot be expressed in 𝐹 as a
sum of squares and real closed if 𝐹 is formally real and has no formally real proper
algebraic extension. The real numbers R and the field of all real algebraic numbers are
real closed. A real closed field has characteristic zero, is totally ordered, and contains
a square root of each nonnegative element; the field obtained from 𝐹 by adjoining a
root of the irreducible polynomial 𝑥 2 + 1 is algebraically closed. For these statements,
see Rajwade [Raj93, Chapter 15]. Every √ finite-dimensional division algebra over a
real closed field 𝐹 is either 𝐹 or 𝐾 = 𝐹 ( −1) or 𝐵 = (−1, −1 | 𝐹).
Remark 3.5.11. Algebras of dimension 3, sitting somehow between quadratic ex-
tensions and quaternion algebras, can be characterized in a similar way. If 𝐵 is an
R-algebra of dimension 3, then either 𝐵 is commutative or 𝐵 has a standard involution
and is isomorphic to the subring of upper triangular matrices in M2 (R). A similar
statement holds for free 𝑅-algebras of rank 3 over a (commutative) domain 𝑅; see
Levin [Lev2013].
Exercises
Throughout these exercises, let 𝐹 be a field.
I 1. Let 𝐵 be an 𝐹-algebra and let : 𝐵 → 𝐵 be an 𝐹-linear map with 1 = 1. Show
that is an involution if and only if (ii)–(iii) in Definition 3.2.1 hold for a basis
of 𝐵 (as an 𝐹-vector space).
I 2. Let 𝐾 = 𝐹 [𝛼] be a quadratic 𝐹-algebra, with 𝛼2 = 𝑡𝛼 − 𝑛 for (unique) 𝑡, 𝑛 ∈ 𝐹.
Extending linearly, show that there is a unique standard involution : 𝐾 → 𝐾
with the property that 𝛼 = 𝑡 − 𝛼, and show that
trd(𝑥 + 𝑦𝛼) = 2𝑥 + 𝑡𝑦
nrd(𝑥 + 𝑦𝛼) = 𝑥 2 + 𝑡𝑥𝑦 + 𝑛𝑦 2
for all 𝑥 + 𝑦𝛼 ∈ 𝐹 [𝛼].
I 3. Verify that the map in Example 3.2.9 is a standard involution.
4. Determine the standard involution on 𝐾 = 𝐹 × 𝐹 (with 𝐹 ↩→ 𝐾 under the
diagonal map).
I 5. Let 𝐵 be an 𝐹-algebra with a standard involution. Show that 0 ≠ 𝛼 ∈ 𝐵 is a left
zerodivisor if and only if 𝛼 is a right zerodivisor if and only if nrd(𝛼) = 0. In
particular, if 𝐵 is not the zero ring, then 𝛼 ∈ 𝐵 is (left and right) invertible if
and only if nrd(𝛼) ≠ 0.
6. Suppose char 𝐹 ≠ 2, let 𝐵 be a division quaternion algebra over 𝐹, and let
𝐾1 , 𝐾2 ⊆ 𝐵 be subfields with 𝐾1 ∩ 𝐾2 = 𝐹. Show that the 𝐹-subalgebra of 𝐵
generated by 𝐾1 and 𝐾2 is equal to 𝐵. Conclude that if 1, 𝛼, 𝛽 ∈ 𝐵 are 𝐹-linearly
independent, then 1, 𝛼, 𝛽, 𝛼𝛽 are an 𝐹-basis for 𝐵. [Hint: use the involution.]
By way of counterexample, show that these results need not hold for 𝐵 = M2 (𝐹).
7. Show that 𝐵 = M𝑛 (𝐹) has a standard involution if and only if 𝑛 ≤ 2.
8. Let 𝐺 be a finite group. Show that the 𝐹-linear mapÉ induced by 𝑔 ↦→ 𝑔 −1
for 𝑔 ∈ 𝐺 is an involution on the group ring 𝐹 [𝐺] = 𝑔 ∈𝐺 𝐹𝑔. Determine
necessary and sufficient conditions for this map to be a standard involution.
9. Let 𝐵 be an 𝐹-algebra with a standard involution : 𝐵 → 𝐵. In this exercise,
we examine when is the identity map.
(a) Show that if char 𝐹 ≠ 2, then 𝑥 ∈ 𝐵 satisfies 𝑥 = 𝑥 if and only 𝑥 ∈ 𝐹.
(b) Suppose that dim𝐹 𝐵 < ∞. Show that the identity map is a standard
involution on 𝐵 if and only if (i) 𝐵 = 𝐹 or (ii) char 𝐹 = 2 and 𝐵 is a
quotient of the commutative ring 𝐹 [𝑥1 , . . . , 𝑥 𝑛 ]/(𝑥12 − 𝑎 1 , . . . , 𝑥 𝑛2 − 𝑎 𝑛 )
with 𝑎 𝑖 ∈ 𝐹.
10. Let 𝐾 ⊇ 𝐹 be a field which has degree 𝑚 as an 𝐹-algebra in the sense of
Definition 3.4.5. Suppose that char 𝐹 - 𝑚. Show that [𝐾 : 𝐹] = 𝑚, i.e., 𝐾 has
degree 𝑚 in the usual sense. (What happens when char 𝐹 | 𝑚?)
11. Let 𝐵 be an 𝐹-algebra with standard involution. Suppose that 𝜙 : 𝐵 −→∼ 𝐵 is
an 𝐹-algebra automorphism. Show for 𝛼 ∈ 𝐵 that 𝜙(𝛼) = 𝜙(𝛼), and therefore
that trd(𝜙(𝛼)) = trd(𝛼) and nrd(𝜙(𝛼)) = nrd(𝛼). [Hint: consider the map
𝛼 ↦→ 𝜙−1 (𝜙(𝛼)).]
12. In this exercise, we explore further the relationship between algebras of degree
2 and those with standard involutions (Remark 3.5.5).
(a) Suppose char 𝐹 ≠ 2 and let 𝐵 be a finite-dimensional 𝐹-algebra. Show
that 𝐵 has a standard involution if and only if deg𝐹 𝐵 ≤ 2.
(b) Let 𝐹 = F2 and let 𝐵 be a Boolean ring, a ring such that 𝑥 2 = 𝑥 for all
𝑥 ∈ 𝐵. (Verify that 2 = 0 in 𝐵, so 𝐵 is an F2 -algebra.) Prove that 𝐵 does not
have a standard involution unless 𝐵 = F2 or 𝐵 = F2 × F2 , but nevertheless
any Boolean ring has degree at most 2.
I 13. Let 𝐵 = M𝑛 (𝐹), and consider the map 𝜆 : 𝐵 ↩→ End𝐹 (𝐵) by 𝛼 ↦→ 𝜆 𝛼 defined
by left-multiplication in 𝐵. Show that for all 𝛼 ∈ M𝑛 (𝐹), the characteristic
polynomial of 𝜆 𝛼 is the 𝑛th power of the usual characteristic polynomial of 𝛼.
Conclude when 𝑛 = 2 that tr(𝛼) = 2 trd( 𝐴) and det(𝛼) = nrd(𝛼) 2 .
14. Considering a slightly different take on the previous exercise: let 𝐵 be a quater-
nion algebra over 𝐹. Show that the characteristic polynomial of left multipli-
cation by 𝛼 ∈ 𝐵 is equal to that of right multiplication and is the square of the
reduced characteristic polynomial. [Hint: if a direct approach is too cumber-
some, consider applying the previous exercise and the left regular representation
as in 2.3.8.]
15. Let 𝑉 be an 𝐹-vector space and let 𝑡 : 𝑉 → 𝐹 be an 𝐹-linear map. Let 𝐵 = 𝐹 ⊕𝑉
and define the binary operation 𝑥 · 𝑦 = 𝑡 (𝑥)𝑦 for 𝑥, 𝑦 ∈ 𝑉. Show that · induces
a multiplication on 𝐵, and that the map 𝑥 ↦→ 𝑥 = 𝑡 (𝑥) − 𝑥 for 𝑥 ∈ 𝑉 induces
a standard involution on 𝐵. [Such an algebra is called an exceptional algebra
[GrLu2009, Voi2011b].] Conclude that there exists a central 𝐹-algebra 𝐵 with
a standard involution in any dimension 𝑛 = dim𝐹 𝐵 ≥ 3.
I 16. In this exercise, we mimic the proof of Theorem 3.5.1 to prove that a quaternion
algebra over a finite field of odd cardinality is not a division ring, a special case
of Wedderburn’s little theorem: a finite division ring is a field.
Assume for purposes of contradiction that 𝐵 is a division quaternion algebra
over 𝐹 = F𝑞 with 𝑞 odd.
(a) Let 𝑖 ∈ 𝐵 r 𝐹. Show that the centralizer 𝐶 𝐵× (𝑖) = {𝛼 ∈ 𝐵× : 𝑖𝛼 = 𝛼𝑖} of 𝑖
in 𝐵× satisfies 𝐶 𝐵× (𝑖) = 𝐹 (𝑖) × .
(b) Conclude that any noncentral conjugacy class in 𝐵× has order 𝑞 2 + 1.
(c) Derive a contradiction from the class equation 𝑞 4 − 1 = 𝑞 − 1 + 𝑚(𝑞 2 + 1)
(where 𝑚 ∈ Z).
[For the case 𝑞 even, see Exercise 6.16; for fun, the eager reader may wish to
prove Weddernburn’s little theorem for 𝐹 = F2 directly.]
17. Derive Euler’s identity (1.1.6) that the product of the sum of four squares is
again the sum of four squares as follows. Let 𝐹 = Q(𝑥1 , . . . , 𝑥 4 , 𝑦 1 , . . . , 𝑦 4 ) be a
function field over Q in 8 variables and consider the quaternion algebra (−1, −1 |
𝐹). Show (by an explicit universal formula) that if 𝑅 is any commutative ring
and 𝑥, 𝑦 ∈ 𝑅 are the sum of four squares in 𝑅, then 𝑥𝑦 is the sum of four squares
in 𝑅.
18. Suppose char 𝐹 ≠ 2. For an 𝐹-algebra 𝐵, let
𝑉 (𝐵) = {𝛼 ∈ 𝐵 r 𝐹 : 𝛼2 ∈ 𝐹} ∪ {0}.
Let 𝐵 be a division ring. Show that 𝑉 (𝐵) is a vector space (closed under
addition) if and only if 𝐵 = 𝐹 or 𝐵 = 𝐾 is a quadratic field extension of 𝐹 or 𝐵
is a quaternion algebra over 𝐹.
19. Let 𝐵 be an 𝐹-algebra with 𝐹-basis 𝑒 1 , 𝑒 2 , . . . , 𝑒 𝑛 . Let : 𝐵 → 𝐵 be an
involution. Show that is standard if and only if
𝑒 𝑖 𝑒 𝑖 ∈ 𝐹 and (𝑒 𝑖 + 𝑒 𝑗 ) (𝑒 𝑖 + 𝑒 𝑗 ) ∈ 𝐹 for all 𝑖, 𝑗 = 1, . . . , 𝑛.

Chapter 4
Quadratic forms
Quaternion algebras, as algebras equipped with a standard involution, are intrinsically

related to quadratic forms. We develop this connection in the next two chapters.
4.1 ⊲ Reduced norm as quadratic form

Let 𝐹 be a field with char 𝐹 ≠ 2 and let 𝐵 = (𝑎, 𝑏 | 𝐹) be a quaternion algebra over
𝐹. We have seen (3.2.9) that 𝐵 has a unique standard involution and consequently a
reduced norm map, with
nrd(𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗) = 𝑡 2 − 𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧 2 (4.1.1)
for 𝑡, 𝑥, 𝑦, 𝑧 ∈ 𝐹. The reduced norm therefore defines a quadratic form, a homoge-

neous polynomial of degree 2 in 𝐹 [𝑡, 𝑥, 𝑦, 𝑧] (thought of as a function of the coefficients
of an element respect to the basis 1, 𝑖, 𝑗, 𝑖 𝑗). It should come as no surprise, then, that
the structure of the quaternion algebra 𝐵 is related to properties of the quadratic form
nrd.
Let 𝑄 : 𝑉 → 𝐹 be a quadratic form. Then 𝑄 can be diagonalized by a change of
variables: there is a basis 𝑒 1 , . . . , 𝑒 𝑛 of 𝑉 such that
𝑄(𝑥 1 𝑒 1 + · · · + 𝑥 𝑛 𝑒 𝑛 ) = 𝑄(𝑥1 , . . . , 𝑥 𝑛 ) = 𝑎 1 𝑥12 + · · · + 𝑎 𝑛 𝑥 𝑛2
with 𝑎 𝑖 ∈ 𝐹. We define the discriminant of 𝑄 to be the (well-defined) product

disc(𝑄) := 𝑎 1 · · · 𝑎 𝑛 /2𝑛 ∈ 𝐹/𝐹 ×2 . (The factor 2𝑛 is for consistency with more general
notions; it is harmless if a bit annoying.) We say that a quadratic form is nondegenerate
if its discriminant is nonzero. The reduced norm quadratic form (4.1.1) is already
diagonal in the basis 1, 𝑖, 𝑗, 𝑖 𝑗, and it is nondegenerate because 𝑎, 𝑏 ≠ 0.
A similarity from 𝑄 to another quadratic form 𝑄 0 : 𝑉 0 → 𝐹 is a pair ( 𝑓 , 𝑢) where
∼
𝑓:𝑉 − → 𝑉 0 is an 𝐹-linear isomorphism and 𝑢 ∈ 𝐹 × satisfy 𝑄 0 ( 𝑓 (𝑥)) = 𝑢𝑄(𝑥) for all
𝑥 ∈ 𝑉. An isometry is a similarity with 𝑢 = 1. The orthogonal group of 𝑄 is the
group of self-isometries of 𝑄, i.e.,
O(𝑄) (𝐹) := { 𝑓 ∈ Aut𝐹 (𝑉) : 𝑄( 𝑓 (𝑥)) = 𝑄(𝑥) for all 𝑥 ∈ 𝑉 }.
47
48 CHAPTER 4. QUADRATIC FORMS
An isometry 𝑓 ∈ O(𝑄) (𝐹) is special if det 𝑓 = 1, and the special orthogonal group
of 𝑄 is the group of special isometries of 𝑄.
More generally, we have seen that any algebra with a standard involution has a
quadratic form nrd. We say that the standard involution is nondegenerate whenever
the quadratic form nrd is so. Generalizing Theorem 3.1.1, we prove the following (see
Main Theorem 4.4.1 for the proof).
Main Theorem 4.1.2. Let 𝐵 be an 𝐹-algebra. Then 𝐵 has a nondegenerate standard

involution if and only if one of the following holds:
(i) 𝐵 = 𝐹;
(ii) 𝐵 = 𝐾 has dim𝐹 𝐾 = 2 and either 𝐾 ' 𝐹 × 𝐹 or 𝐾 is a field; or
(iii) 𝐵 is a quaternion algebra over 𝐹.
This theorem gives another way of characterizing quaternion algebras: they are
noncommutative algebras with a nondegenerate standard involution.
In Section 2.4, we saw that the unit Hamiltonians H1 act on the pure Hamiltonians
0
H (Section 2.4) by rotations: the standard Euclidean quadratic form (sum of squares)
is preserved by conjugation. This generalizes in a natural way to an arbitrary field,
and so we can understand the group of linear transformations that preserve a ternary
(or quaternary) form in terms of the unit group of a quaternion algebra 𝐵 (Proposition
4.5.10): there is an exact sequence
1 → 𝐹 × → 𝐵× → SO(nrd | 𝐵0 ) (𝐹) → 1
where SO(𝑄) (𝐹) is the group of special (or oriented) isometries of the quadratic form
𝑄 and 𝐵0 := {𝛼 ∈ 𝐵 : trd(𝛼) = 0}.
4.2 Basic definitions

In this section, we summarize basic definitions and notation for quadratic forms over
fields. The “Bible for all quadratic form practitioners” (according to the MathSciNet
review by K. Szymiczek) is the book by Lam [Lam2005]; in particular, Lam gives
a very readable account of the relationship between quadratic forms and quaternion
algebras over 𝐹 when char 𝐹 ≠ 2 [Lam2005, Sections III.1–III.2] and many other
topics in the algebraic theory of quadratic forms. Also recommended are the books by
Cassels [Cas78], O’Meara [O’Me73], and Scharlau [Scha85], as well as the book by
Grove [Grov2002], who treats quadratic forms from a geometric point of view in terms
of the orthogonal group. For reference and further inspiration, see also the hugely
influential book by Eichler [Eic53].
Let 𝐹 be a field. (For now, we allow char 𝐹 to be arbitrary.)
Definition 4.2.1. A quadratic form 𝑄 is a map 𝑄 : 𝑉 → 𝐹 on an 𝐹-vector space 𝑉

satisfying:
(i) 𝑄(𝑎𝑥) = 𝑎 2 𝑄(𝑥) for all 𝑎 ∈ 𝐹 and 𝑥 ∈ 𝑉; and

4.2. BASIC DEFINITIONS 49
(ii) The map 𝑇 : 𝑉 × 𝑉 → 𝐹 defined by

𝑇 (𝑥, 𝑦) = 𝑄(𝑥 + 𝑦) − 𝑄(𝑥) − 𝑄(𝑦)
is 𝐹-bilinear.
We call the pair (𝑉, 𝑄) a quadratic space and 𝑇 the associated bilinear form.
We will often abbreviate a quadratic space (𝑉, 𝑄) by simply 𝑉. If 𝑄 is a quadratic
form then the associated bilinear form 𝑇 is symmetric, satisfying 𝑇 (𝑥, 𝑦) = 𝑇 (𝑦, 𝑥)
for all 𝑥, 𝑦 ∈ 𝑉; in particular, 𝑇 (𝑥, 𝑥) = 2𝑄(𝑥) for all 𝑥 ∈ 𝑉, so when char 𝐹 ≠ 2 we
recover the quadratic form from the symmetric bilinear form.
For the remainder of this section, let 𝑄 : 𝑉 → 𝐹 be a quadratic form with associated
bilinear form 𝑇.
4.2.2. Suppose dim𝐹 𝑉 = 𝑛 < ∞. Let 𝑒 1 , . . . , 𝑒 𝑛 be a basis for 𝑉, giving an isomor-
phism 𝑉 ' 𝐹 𝑛 . Then 𝑄 can be written
∑︁ ∑︁
𝑄(𝑥1 𝑒 1 + · · · + 𝑥 𝑛 𝑒 𝑛 ) = 𝑄(𝑒 𝑖 )𝑥𝑖2 + 𝑇 (𝑒 𝑖 , 𝑒 𝑗 )𝑥 𝑖 𝑥 𝑗 ∈ 𝐹 [𝑥1 , . . . , 𝑥 𝑛 ]
𝑖 𝑖< 𝑗
as a homogeneous polynomial of degree 2.

The Gram matrix of 𝑄 in the basis 𝑒 𝑖 is the (symmetric) matrix
[𝑇] := (𝑇 (𝑒 𝑖 , 𝑒 𝑗 ))𝑖, 𝑗 ∈ M𝑛 (𝐹).
We then have 𝑇 (𝑥, 𝑦) = 𝑥 t [𝑇] 𝑦 for 𝑥, 𝑦 ∈ 𝑉 ' 𝐹 𝑛 as column vectors. Under a change
of basis 𝐴 ∈ GL𝑛 (𝐹) with 𝑒 𝑖0 = 𝐴𝑒 𝑖 , the Gram matrix [𝑇] 0 in the basis 𝑒 𝑖0 has
[𝑇] 0 = 𝐴t [𝑇] 𝐴. (4.2.3)
Definition 4.2.4. A similarity of quadratic forms from 𝑄 : 𝑉 → 𝐹 to 𝑄 0 : 𝑉 0 → 𝐹
∼ 𝑉 0 is an 𝐹-linear isomorphism and 𝑢 ∈ 𝐹 × satisfy
is a pair ( 𝑓 , 𝑢) where 𝑓 : 𝑉 −→
0
𝑄 ( 𝑓 (𝑥)) = 𝑢𝑄(𝑥) for all 𝑥 ∈ 𝑉, i.e., such that the diagram
𝑉
𝑄
/𝐹
o 𝑓 o 𝑢 (4.2.5)
𝑄0
𝑉0 /𝐹
commutes. In a similarity ( 𝑓 , 𝑢), the scalar 𝑢 is called the similitude factor of the
similarity. An isometry of quadratic forms (or isomorphism of quadratic spaces) is a
similarity with similitude factor 𝑢 = 1; we write in this case 𝑄 ' 𝑄 0.
Definition 4.2.6. The general orthogonal group (or similarity group) of the quadratic
form 𝑄 is the group of self-similarities of 𝑄 under composition
GO(𝑄) (𝐹) := {( 𝑓 , 𝑢) ∈ Aut𝐹 (𝑉) × 𝐹 × : 𝑄( 𝑓 (𝑥)) = 𝑢𝑄(𝑥) for all 𝑥 ∈ 𝑉 };
the orthogonal group of 𝑄 is the group of self-isometries of 𝑄, i.e.,
O(𝑄) (𝐹) := { 𝑓 ∈ Aut𝐹 (𝑉) : 𝑄( 𝑓 (𝑥)) = 𝑄(𝑥) for all 𝑥 ∈ 𝑉 }.
Remark 4.2.7. A similarity allows isomorphisms of the target 𝐹 (as a one-dimensional

𝐹-vector space). The notion of isometry comes from the connection with measuring
lengths, when working with the usual Euclidean norm form on a vector space over R:
similarity allows these lengths to scale uniformly (e.g., similar triangles).
There is a canonical exact sequence
1 → O(𝑄) (𝐹) → GO(𝑄) (𝐹) → 𝐹 ×

(4.2.8)
( 𝑓 , 𝑢) ↦→ 𝑢
realizing O(𝑄) (𝐹) ≤ GO(𝑄) (𝐹) as the subgroup of self-similarities with similitude
factor 𝑢 = 1.
4.2.9. Returning to 4.2.2, suppose dim𝐹 𝑉 = 𝑛 < ∞ and char 𝐹 ≠ 2. Then one can
understand the orthogonal group of 𝑄 quite concretely in matrix terms as follows.
Choose a basis 𝑒 1 , . . . , 𝑒 𝑛 for 𝑉 and let [𝑇] be the Gram matrix of 𝑄 with respect to
this basis, so that 2𝑄(𝑥) = 𝑥 t [𝑇]𝑥 for all 𝑥 ∈ 𝑉 ' 𝐹 𝑛 . Then Aut𝐹 (𝑉) ' GL𝑛 (𝐹) and
𝐴 ∈ GL𝑛 (𝐹) belongs to O(𝑄) if and only if
( 𝐴𝑥) t [𝑇] ( 𝐴𝑥) = 𝑥 t ( 𝐴t [𝑇] 𝐴)𝑥 = 𝑥 t [𝑇]𝑥
for all 𝑥 ∈ 𝑉, and therefore
O(𝑄) (𝐹) = { 𝐴 ∈ GL𝑛 (𝐹) : 𝐴t [𝑇] 𝐴 = [𝑇]} (4.2.10)
and
GO(𝑄) (𝐹) = {( 𝐴, 𝑢) ∈ GL𝑛 (𝐹) × 𝐹 × : 𝐴t [𝑇] 𝐴 = 𝑢[𝑇]} (4.2.11)
From now on, let 𝑄 : 𝑉 → 𝐹 be a quadratic form and let 𝑇 : 𝑉 × 𝑉 → 𝐹 be the

symmetric bilinear form associated to 𝑄.
Definition 4.2.12. Let 𝑥, 𝑦 ∈ 𝑉. We say that 𝑥 is orthogonal to 𝑦 (with respect to 𝑄)

if 𝑇 (𝑥, 𝑦) = 0.
Since 𝑇 is symmetric, 𝑥 is orthogonal to 𝑦 if and only if 𝑦 is orthogonal to 𝑥 for

𝑥, 𝑦 ∈ 𝑉, and so we simply say 𝑥, 𝑦 are orthogonal. If 𝑆 ⊆ 𝑉 is a subset, we write
𝑆 ⊥ := {𝑥 ∈ 𝑉 : 𝑇 (𝑣, 𝑥) = 0 for all 𝑣 ∈ 𝑆}
for the subspace of 𝑉 which is orthogonal to (the span of) 𝑆.
4.2.13. Let 𝐵 be an algebra over 𝐹 with a standard involution. Then nrd : 𝐵 → 𝐹 is a

quadratic form on 𝐵. Indeed, nrd(𝑎𝛼) = 𝑎 2 𝛼 for all 𝛼 ∈ 𝐵, and the map 𝑇 given by
𝑇 (𝛼, 𝛽) = (𝛼 + 𝛽) (𝛼 + 𝛽) − 𝛼𝛼 − 𝛽𝛽 = 𝛼𝛽 + 𝛽𝛼 = 𝛼𝛽 + 𝛼𝛽 = trd(𝛼𝛽) (4.2.14)
for 𝛼, 𝛽 ∈ 𝐵 is bilinear, and
𝑇 (𝛼, 𝛽) = trd(𝛼𝛽) = trd(𝛼(trd(𝛽) − 𝛽)) = trd(𝛼) trd(𝛽) − trd(𝛼𝛽). (4.2.15)

4.2. BASIC DEFINITIONS 51
So 𝛼, 𝛽 ∈ 𝐵 are orthogonal with respect to nrd if and only if

trd(𝛼𝛽) = 𝛼𝛽 + 𝛽𝛼 = 0
if and only if
trd(𝛼𝛽) = trd(𝛼) trd(𝛽).
Thus 1 and 𝛼 ∈ 𝐵 are orthogonal if and only if trd(𝛼) = 0 if and only if 𝛼2 = − nrd(𝛼).
Moreover, rearranging (4.2.14),
𝛼𝛽 + 𝛽𝛼 = trd(𝛽)𝛼 + trd(𝛼) 𝛽 − 𝑇 (𝛼, 𝛽). (4.2.16)
In particular, if 1, 𝛼, 𝛽 ∈ 𝐵 are linearly independent over 𝐹, then by (4.2.16) they are
pairwise orthogonal if and only if 𝛽𝛼 = −𝛼𝛽.
In this way, we see that the multiplication law in 𝐵 is governed in a fundamental
way by the reduced norm quadratic form.
Definition 4.2.17. Let 𝑄 : 𝑉 → 𝐹 be a quadratic form. We say that 𝑄 represents an
element 𝑎 ∈ 𝐹 if there exists 𝑥 ∈ 𝑉 such that 𝑄(𝑥) = 𝑎. A quadratic form is universal
if it represents every element of 𝐹.
Definition 4.2.18. A quadratic form 𝑄 (or a quadratic space 𝑉) is isotropic if 𝑄
represents 0 nontrivially (there exists 0 ≠ 𝑥 ∈ 𝑉 such that 𝑄(𝑥) = 0) and otherwise 𝑄
is anisotropic.
Remark 4.2.19. The terminology isotropic is as least as old as Eichler [Eic53, p. 3],
and perhaps it goes back to Witt. The word can be used to mean “having properties that
are identical in all directions”, and so the motivation for this language possibly comes
from physics: the second fundamental form associated to a parametrized surface
𝑧 = 𝑓 (𝑥, 𝑦) in R3 is a quadratic form, and (roughly speaking) this quadratic form
defines the curvature at a given point. In this sense, if the quadratic form vanishes,
then the curvature is zero, and things look the same in all directions.
4.2.20. Let 𝑄 0 : 𝑉 0 → 𝐹 be another quadratic form. We define the orthogonal direct
sum
𝑄 𝑄0 : 𝑉 ⊕ 𝑉 0 → 𝐹
(𝑄 𝑄 0) (𝑥 + 𝑥 0) = 𝑄(𝑥) + 𝑄(𝑥 0)
where 𝑥 ∈ 𝑉 and 𝑥 0 ∈ 𝑉 0; the associated bilinear form 𝑇 𝑇 0 has
(𝑇 𝑇 0) (𝑥 + 𝑥 0, 𝑦 + 𝑦 0) = 𝑇 (𝑥, 𝑦) + 𝑇 (𝑥 0, 𝑦 0)
for all 𝑥, 𝑦 ∈ 𝑉 and 𝑥 0, 𝑦 0 ∈ 𝑉 0. By definition, under the natural inclusion of 𝑉, 𝑉 0 ⊆
𝑉 ⊕ 𝑉 0, we have 𝑉 0 ⊆ 𝑉 ⊥ (and 𝑉 ⊆ (𝑉 0) ⊥ ).
4.2.21. For 𝑎 ∈ 𝐹, we write h𝑎i for the quadratic form 𝑎𝑥 2 on 𝐹. More generally, for
𝑎 1 , . . . , 𝑎 𝑛 ∈ 𝐹, we write
h𝑎 1 i · · · h𝑎 𝑛 i := h𝑎 1 , . . . , 𝑎 𝑛 i
for the quadratic form on 𝐹 𝑛 defined by 𝑄(𝑥1 , . . . , 𝑥 𝑛 ) = 𝑎 1 𝑥12 + · · · + 𝑎 𝑛 𝑥 𝑛2 .
To conclude this introduction, we state an important result due originally to Witt

which governs the decomposition of quadratic spaces into orthogonal sums up to
isometry.
Theorem 4.2.22. Let 𝑉 ' 𝑉 0 be isometric quadratic spaces with orthogonal decom-
positions 𝑉 ' 𝑊1 𝑊2 and 𝑉 0 ' 𝑊10 𝑊20 .
(a) If 𝑊1 ' 𝑊10 , then 𝑊2 ' 𝑊20 .

(b) If 𝑔 : 𝑊1 −→∼ 𝑊 0 is an isometry, then there exists an isometry 𝑓 : 𝑉 −→
∼ 𝑉 0 such
1
0
that 𝑓 | 𝑊1 = 𝑔 and 𝑓 (𝑊2 ) = 𝑊2 .
Proof. The proof is requested in Exercise 4.16. For a proof and the equivalence be-
tween Witt cancellation (part (a)) and Witt extension (part (b)), see Lam [Lam2005,
Proof of Theorem I.4.2, p. 14], Scharlau [Scha85, Theorem 1.5.3], or O’Meara
[O’Me73, Theorem 42:17].
Theorem 4.2.22(a) is called Witt cancellation and 4.2.22(b) is called Witt exten-
sion.
4.3 Discriminants, nondegeneracy

For the remainder of this chapter, we suppose that char 𝐹 ≠ 2. (We take up the case
char 𝐹 = 2 in section 6.3.) Throughout, let 𝑄 : 𝑉 → 𝐹 be a quadratic form with
dim𝐹 𝑉 = 𝑛 < ∞ and associated symmetric bilinear form 𝑇.
The following result (proven by induction) is a standard application of Gram–
Schmidt orthogonalization (Exercise 4.1); working with a quadratic form as a polyno-
mial, this procedure can be thought of as iteratively completing the square.
Lemma 4.3.1. There exists a basis of 𝑉 such that 𝑄 ' h𝑎 1 , . . . , 𝑎 𝑛 i with 𝑎 𝑖 ∈ 𝐹.
A form presented with a basis as in Lemma 4.3.1 is called normalized (or diag-
onal). For a diagonal quadratic form 𝑄, the associated Gram matrix [𝑇] is diagonal
with entries 2𝑎 1 , . . . , 2𝑎 𝑛 .
4.3.2. The determinant det( [𝑇]) of a Gram matrix for 𝑄 depends on a choice of
basis for 𝑉, but by (4.2.3), a change of basis matrix 𝐴 ∈ GL𝑛 (𝐹) operates on [𝑇] by
𝐴t [𝑇] 𝐴, and det( 𝐴t [𝑇] 𝐴) = det( 𝐴) 2 det( [𝑇]), so we obtain a well-defined element
det(𝑇) ∈ 𝐹/𝐹 ×2 independent of the choice of basis.
Definition 4.3.3. The discriminant of 𝑄 is
disc(𝑄) := 2−𝑛 det(𝑇) ∈ 𝐹/𝐹 ×2 .
The signed discriminant of 𝑄 is
sgndisc(𝑄) := (−1) 𝑛(𝑛−1)/2 disc(𝑄) ∈ 𝐹/𝐹 ×2 .

4.3. DISCRIMINANTS, NONDEGENERACY 53
When it will cause no confusion, we will represent the class of the discriminant in
𝐹/𝐹 ×2 simply by a representative element in 𝐹.
Remark 4.3.4. The extra factor 2−𝑛 is harmless since char 𝐹 ≠ 2, and it allows us to
naturally cancel certain factors 2 that appear whether we are in even or odd dimension—
it will be essential when we consider the case char 𝐹 = 2 (see 6.3.1). The distinction
between even and odd dimensional quadratic spaces is not arbitrary: indeed, this
distinction is pervasive, even down to the classification of semisimple Lie algebras.
Example 4.3.5. We have disc(h𝑎 1 , . . . , 𝑎 𝑛 i) = 𝑎 1 · · · 𝑎 𝑛 for 𝑎 𝑖 ∈ 𝐹.
Definition 4.3.6. The bilinear form 𝑇 : 𝑉 × 𝑉 → 𝐹 is nondegenerate if for all

𝑥 ∈ 𝑉 r {0}, the linear functional 𝑇𝑥 : 𝑉 → 𝐹 defined by 𝑇𝑥 (𝑦) = 𝑇 (𝑥, 𝑦) is nonzero,
i.e., there exists 𝑦 ∈ 𝑉 such that 𝑇 (𝑥, 𝑦) ≠ 0. We say that 𝑄 (or 𝑉) is nondegenerate
if the associated bilinear form 𝑇 is nondegenerate.
4.3.7. The bilinear form 𝑇 induces a map
𝑉 → Hom(𝑉, 𝐹)
𝑥 ↦→ (𝑦 ↦→ 𝑇 (𝑥, 𝑦))
and 𝑇 is nondegenerate if and only if this map is injective (and hence an isomorphism) if
and only if det(𝑇) ≠ 0. Put another way, 𝑄 is nondegenerate if and only if disc(𝑄) ≠ 0,
and so a diagonal form h𝑎 1 , . . . , 𝑎 𝑛 i is nondegenerate if and only if 𝑎 𝑖 ≠ 0 for all 𝑖.
Example 4.3.8. Let 𝐵 = (𝑎, 𝑏 | 𝐹) be a quaternion algebra. Then by 3.2.9, the

quadratic form nrd : 𝐵 → 𝐹 is normalized with respect to the basis 1, 𝑖, 𝑗, 𝑖 𝑗. Indeed,
nrd ' h1, −𝑎, −𝑏, 𝑎𝑏i.
We have disc(nrd) = (𝑎𝑏) 2 ≠ 0, so nrd is nondegenerate.
If 𝐵 is an 𝐹-algebra with a standard involution, then the reduced norm defines a

quadratic form on 𝐵, and we say that the standard involution is nondegenerate if nrd
is nondegenerate.
4.3.9. One can often restrict to the case where a quadratic form 𝑄 is nondegenerate
by splitting off the radical, as follows. We define the radical of 𝑄 to be
rad(𝑄) := 𝑉 ⊥ = {𝑥 ∈ 𝑉 : 𝑇 (𝑥, 𝑦) = 0 for all 𝑦 ∈ 𝑉 }.
The radical rad(𝑄) ⊆ 𝑉 is a subspace, so completing a basis of rad(𝑄) to 𝑉 we can

write (noncanonically) 𝑉 = rad(𝑄) 𝑊, as the direct sum is an orthogonal direct sum
by definition of the radical. In this decomposition, 𝑄| rad(𝑄) is identically zero and
𝑄| 𝑊 is nondegenerate.
4.4 Nondegenerate standard involutions

In this section, we follow Theorem 3.5.1 with a characterization of quaternion algebras
beyond division algebras.
Main Theorem 4.4.1. Suppose char 𝐹 ≠ 2 and let 𝐵 be an 𝐹-algebra. Then 𝐵 has a
nondegenerate standard involution if and only if one of the following holds:
(i) 𝐵 = 𝐹;
(ii) 𝐵 = 𝐾 is a quadratic 𝐹-algebra and either 𝐾 ' 𝐹 × 𝐹 or 𝐾 is a field; or
Case (ii) in Main Theorem 4.4.1 is equivalent to requiring that 𝐾 be a quadratic
𝐹-algebra that is reduced (has no nonzero nilpotent elements).
Remark 4.4.2. By Exercise 3.15, there exist 𝐹-algebras with standard involution having
arbitrary dimension, so it is remarkable that the additional requirement that the standard
involution be nondegenerate gives such a tidy result.
Proof of Main Theorem 4.4.1. If 𝐵 = 𝐹, then the standard involution is the identity
and nrd is nondegenerate. If dim𝐹 𝐾 = 2, then after completing the square we may
write 𝐾 ' 𝐹 [𝑥]/(𝑥 2 − 𝑎) and in the basis 1, 𝑥 we find nrd ' h1, 𝑎i. By Example 4.3.5,
nrd is nondegenerate if and only if 𝑎 ∈ 𝐹 × if and only if 𝐾 is a quadratic field extension
of 𝐹 or 𝐾 ' 𝐹 × 𝐹.
Suppose that dim𝐹 𝐵 > 2. Let 1, 𝑖, 𝑗 be a part of a normalized basis for 𝐵 with
respect to the quadratic form nrd. Then 𝑇 (1, 𝑖) = trd(𝑖) = 0, so 𝑖 2 = 𝑎 ∈ 𝐹 × , since
nrd is nondegenerate. Note in particular that 𝑖 = −𝑖. Similarly 𝑗 2 = 𝑏 ∈ 𝐹 × , and
by (4.2.16) we have trd(𝑖 𝑗) = 𝑖 𝑗 + 𝑗𝑖 = 0. We have 𝑇 (1, 𝑖 𝑗) = trd(𝑖 𝑗) = 0, and
𝑇 (𝑖 𝑗, 𝑖) = trd(𝑖(𝑖 𝑗)) = −𝑎 trd( 𝑗) = 0 and similarly 𝑇 (𝑖 𝑗, 𝑗) = 0, hence 𝑖 𝑗 ∈ {1, 𝑖, 𝑗 }⊥ .
If 𝑖 𝑗 = 0 then 𝑖(𝑖 𝑗) = 𝑎 𝑗 = 0 so 𝑗 = 0, a contradiction. Since nrd is nondegenerate, it
follows then that the set 1, 𝑖, 𝑗, 𝑖 𝑗 is linearly independent.
Therefore, the subalgebra 𝐴 of 𝐵 generated by 𝑖, 𝑗 satisfies 𝐴 ' (𝑎, 𝑏 | 𝐹), and if
dim𝐹 𝐵 = 4 we are done. So let 𝑘 ∈ 𝐴⊥ ; then trd(𝑘) = 0 and 𝑘 2 = 𝑐 ∈ 𝐹 × . Thus
𝑘 ∈ 𝐵× , with 𝑘 −1 = 𝑐−1 𝑘. By 4.2.13 we have 𝑘𝛼 = 𝛼𝑘 for any 𝛼 ∈ 𝐴 since 𝑘 = −𝑘.
But then
𝑘 (𝑖 𝑗) = (𝑖 𝑗)𝑘 = 𝑗 𝑖𝑘 = 𝑗 𝑘𝑖 = 𝑘 ( 𝑗𝑖). (4.4.3)
But 𝑘 ∈ 𝐵× so 𝑖 𝑗 = 𝑗𝑖 = −𝑖 𝑗, and this is a contradiction.
Main Theorem 4.4.1 has the following corollaries.

Corollary 4.4.4. Let 𝐵 be an 𝐹-algebra with char 𝐹 ≠ 2. Then 𝐵 is a quaternion alge-
bra if and only if 𝐵 is noncommutative and has a nondegenerate standard involution.
Proof. Immediate.
Corollary 4.4.5. Let 𝐵 have a nondegenerate standard involution, and suppose that
𝐾 ⊆ 𝐵 is a commutative 𝐹-subalgebra such that the restriction of the standard
involution is nondegenerate. Then dim𝐹 𝐾 ≤ 2. Moreover, if 𝐾 ≠ 𝐹, then the
centralizer of 𝐾 × in 𝐵× is 𝐾 × .
4.5. SPECIAL ORTHOGONAL GROUPS 55
Proof. The first statement is immediate; the second follows by considering the algebra
generated by the centralizer.
Remark 4.4.6. Algebras with involutions come from quadratic forms, and the re-
sults of this chapter are just one special case of a much more general theory. More
precisely, there is a natural bĳection between the set of isomorphism classes of finite-
dimensional simple 𝐹-algebras equipped with an 𝐹-linear involution and the set of
similarity classes of nondegenerate quadratic forms on finite-dimensional 𝐹-vector
spaces. More generally, for involutions that act nontrivially on the base field, one
looks at Hermitian forms. Consequently, there are three broad types of involutions
on central simple algebras, depending on the associated quadratic or Hermitian form:
orthogonal, symplectic, and unitary. Accordingly, algebras with involutions can be
classified by the invariants of the associated form. This connection is the subject
of the tome by Knus–Merkurjev–Rost–Tignol [KMRT98]. In this way the theory of
quadratic forms belongs to the theory of algebras with involution, which in turn is a
part of the theory of linear algebraic groups, as expounded by Weil [Weil60]: see the
survey by Tignol [Tig98] for an overview and further references.
4.5 Special orthogonal groups

In this section, we revisit the original motivation of Hamilton (Section 2.4) in a more
general context, relating quaternions to the orthogonal group of a quadratic form. We
retain our running hypothesis that char 𝐹 ≠ 2 and 𝑄 : 𝑉 → 𝐹 is a nondegenerate
quadratic form with dim𝐹 𝑉 = 𝑛 < ∞.
Definition 4.5.1. An isometry 𝑓 ∈ O(𝑄) (𝐹) is special (or proper) if det 𝑓 = 1. The
special orthogonal group of 𝑄 is the group of special isometries of 𝑄:
SO(𝑄) (𝐹) := { 𝑓 ∈ O(𝑄) (𝐹) : det( 𝑓 ) = 1}.
The condition “det 𝑓 = 1” is well-defined, independent of the choice of 𝐹-basis
of 𝑉; having chosen a basis of 𝑉 so that O(𝑄) (𝐹) ≤ GL𝑛 (𝐹), we have SO(𝑄) (𝐹) =
O(𝑄) (𝐹) ∩ SL𝑛 (𝐹).
4.5.2. Suppose that 𝑉 = 𝐹 𝑛 and let 𝑓 ∈ O(𝑄) be a self-isometry of 𝑄, represented in
the standard basis by 𝐴 ∈ GL𝑛 (𝐹). Taking determinants in (4.2.10) we conclude that
det( 𝐴) 2 = 1 so det( 𝐴) = ±1. The determinant is surjective (see Exercise 4.15), so we
have an exact sequence
det
1 → SO(𝑄) (𝐹) → O(𝑄) (𝐹) −−→ {±1} → 1.
If 𝑛 is odd, then either 𝑓 or − 𝑓 is special, so the sequence splits and
O(𝑄) (𝐹) ' {±1} × SO(𝑄) (𝐹). (4.5.3)
4.5.4. Similarly, if ( 𝑓 , 𝑢) ∈ GO(𝑄) (𝐹) then from (4.2.11) we get 𝑢 −𝑛 det( 𝑓 ) 2 = 1. If
𝑛 = 2𝑚 is even, then 𝑢 −𝑚 det( 𝑓 ) = ±1, and we define the general special orthogonal
group (or special similarity group) of 𝑄 to be
GSO(𝑄) (𝐹) := {( 𝑓 , 𝑢) ∈ GO(𝑄) (𝐹) : 𝑢 −𝑚 det( 𝑓 ) = 1}
giving an exact sequence
1 → GSO(𝑄) (𝐹) → GO(𝑄) (𝐹) → {±1} → 1.
If 𝑛 is odd, we have little choice other than to define GSO(𝑄) (𝐹) := GO(𝑄) (𝐹).
Example 4.5.5. If 𝑉 = R𝑛 and 𝑄 is the usual Euclidean norm on 𝑉, then
O(𝑄) (R) = O(𝑛) = { 𝐴 ∈ GL𝑛 (R) : 𝐴𝐴t = 1}
is the group of linear maps preserving length (but not necessarily orientation), whereas
SO(𝑄) (R) is the usual group of rotations of 𝑉 (preserving orientation). Similarly,
GSO(𝑄) (R) consists of orientation-preserving similarities, preserving orientation but
allowing a constant scaling.
In particular, if 𝑛 = 2 then O(2) := O(𝑄) (R) contains

cos 𝜃 sin 𝜃
:=
SO(2) SO(𝑄) (R) = : 𝜃 ∈ R ' R/(2𝜋Z) ' S1
− sin 𝜃 cos 𝜃
(the circle group) with index 2, with a reflection in any line through the origin repre-
senting a nontrivial coset of SO(2) ≤ O(2).
4.5.6. More generally, we may define reflections in O(𝑄) (𝐹) as follows. For 𝑥 ∈ 𝑉
anisotropic (so 𝑄(𝑥) ≠ 0), we define the reflection in 𝑥 to be
𝜏𝑥 : 𝑉 → 𝑉
𝑇 (𝑣, 𝑥)
𝜏𝑥 (𝑣) = 𝑣 − 𝑥.
𝑄(𝑥)
We have 𝜏𝑥 (𝑥) = 𝑥 − 2𝑥 = −𝑥, and

𝑇 (𝑣, 𝑥) 𝑇 (𝑣, 𝑥)
𝑄(𝜏𝑥 (𝑣)) = 𝑄(𝑣) + 𝑄 − 𝑥 + 𝑇 𝑣, − 𝑥
𝑄(𝑥) 𝑄(𝑥)
𝑇 (𝑣, 𝑥) 2 𝑇 (𝑣, 𝑥)
= 𝑄(𝑣) + 𝑄(𝑥) − 𝑇 (𝑣, 𝑥) = 𝑄(𝑣)
𝑄(𝑥) 2 𝑄(𝑥)
so 𝜏𝑥 ∈ O(𝑄) (𝐹) r SO(𝑄) (𝐹).
By a classical theorem of Cartan and Dieudonné, the orthogonal group is generated

by reflections.
Theorem 4.5.7 (Cartan–Dieudonné). Let (𝑉, 𝑄) be a nondegenerate quadratic space

with dim𝐹 𝑉 = 𝑛. Then every isometry 𝑓 ∈ O(𝑄) (𝐹) is a product of at most 𝑛
reflections.
Proof. See Lam [Lam2005, §I.7], O’Meara [O’Me73, §43B], or Scharlau [Scha85,
Theorem 1.5.4]. The proof is by induction on 𝑛, carefully recording the effect of a
reflection in an anisotropic vector.
Since reflections have determinant −1, an isometry 𝑓 is special (and 𝑓 ∈ SO(𝑄) (𝐹))
if and only if it is the product of an even number of reflections.
4.5.8. Now let 𝐵 be a quaternion algebra over 𝐹, and recall (3.3.5) that we have defined
𝐵0 := {𝑣 ∈ 𝐵 : trd(𝐵) = 0}.
Then there is a (left) action

𝐵× 𝐵0 → 𝐵0
(4.5.9)
𝛼 · 𝑣 = 𝛼𝑣𝛼−1 .
since trd(𝛼𝑣𝛼−1 ) = trd(𝑣) = 0. Moreover, 𝐵× acts on 𝑉 by isometries with respect to
the quadratic form 𝑄 = nrd | 𝐵0 : 𝑉 → 𝐹, since nrd(𝛼𝑣𝛼−1 ) = nrd(𝑣) for all 𝛼 ∈ 𝐵 and
𝑣 ∈ 𝑉.
Proposition 4.5.10. Let 𝐵 be a quaternion algebra over 𝐹. Then the action (4.5.9)
induces an exact sequence
1 → 𝐹 × → 𝐵× → SO(nrd | 𝐵0 ) (𝐹) → 1. (4.5.11)
If further nrd(𝐵× ) = 𝐹 ×2 , then
1 → {±1} → 𝐵1 → SO(nrd | 𝐵0 ) (𝐹) → 1,
where 𝐵1 := {𝛼 ∈ 𝐵 : nrd(𝛼) = 1}.
Proof. Let 𝑄 = nrd | 𝐵0 . We saw in 4.5.8 that the action of 𝐵× is by isometries, so

lands in O(𝑄) (𝐹). By the Cartan–Dieudonné theorem (Theorem 4.5.7, the weak
version of Exercise 4.17 suffices), every isometry is the product of reflections, and by
determinants an isometry is special if and only if it is the product of an even number
of reflections. A reflection in 𝑥 ∈ 𝑉 = 𝐵0 with 𝑄(𝑥) = nrd(𝑥) ≠ 0 is of the form
𝑇 (𝑣, 𝑥) trd(𝑣𝑥)
𝜏𝑥 (𝑣) = 𝑣 − 𝑥=𝑣− 𝑥
𝑄(𝑥) nrd(𝑥) (4.5.12)
= 𝑣 − (𝑣𝑥 + 𝑥𝑣)𝑥 −1 = −𝑥𝑣 𝑥 −1 = 𝑥𝑣𝑥 −1 ,
the final equality from 𝑥 = −𝑥 as 𝑥 ∈ 𝐵0 . The product of two such reflections is thus
of the form 𝑣 ↦→ 𝛼𝑣𝛼−1 with 𝛼 ∈ 𝐵× . Therefore 𝐵× acts by special isometries, and
every special isometry so arises: the map 𝐵× → O(𝑄) (𝐹) surjects onto SO(𝑄) (𝐹).
The kernel of the action is given by those 𝛼 ∈ 𝐵× with 𝛼𝑣𝛼−1 = 𝑣 for all 𝑣 ∈ 𝐵0 , i.e.,
𝛼 ∈ 𝑍 (𝐵× ) = 𝐹 × .
The second statement follows directly by writing 𝐵× = 𝐵1 𝐹 × .
Example 4.5.13. If 𝐵 ' M2 (𝐹), then nrd = det, so det0 ' h1, −1, −1i and (4.5.11)
yields the isomorphism PGL2 (𝐹) ' SO(h1, −1, −1i) (𝐹).
Example 4.5.14. If 𝐹 = R and 𝐵 = H, then det(H) = R>0 = R×2 , and the second
exact sequence is Hamilton’s (Section 2.4).
To conclude, we pass from three variables to four variables.
4.5.15. In Exercise 2.17, we showed that there is an exact sequence
1 → {±1} → H1 × H1 → SO(4) → 1
with H1 × H1 acting on H ' R4 by 𝑣 ↦→ 𝛼𝑣 𝛽−1 = 𝛼𝑣 𝛽 for 𝛼, 𝛽 ∈ H1 .

More generally, let 𝐵 be a quaternion algebra over 𝐹. Then there is a left action of
𝐵× × 𝐵× on 𝐵:
𝐵× × 𝐵× 𝐵 → 𝐵
(4.5.16)
(𝛼, 𝛽) · 𝑣 = 𝛼𝑣 𝛽−1 .
This action is by similarities, since if 𝑎 = nrd(𝛼) and 𝑏 = nrd(𝛽), then
𝑎
nrd(𝛼𝑣 𝛽−1 ) = nrd(𝛼) nrd(𝑣) nrd(𝛽−1 ) = nrd(𝑣)
𝑏
for all 𝑣 ∈ 𝑉, with similitude factor 𝑢 = 𝑎/𝑏. In particular, if nrd(𝛼) = nrd(𝛽), then
the action is by isometries.
Proposition 4.5.17. With notation as in 4.5.15, the left action (4.5.16) induces exact
sequences
1 → 𝐹 × → 𝐵× × 𝐵× → GSO(nrd) (𝐹) → 1
(4.5.18)
𝑎 ↦→ (𝑎, 𝑎)
and
1 → 𝐹 × → {(𝛼, 𝛽) ∈ 𝐵× × 𝐵× : nrd(𝛼) = nrd(𝛽)} → SO(nrd) (𝐹) → 1.
If further nrd(𝐵× ) = 𝐹 ×2 , then the sequence
1 → {±1} → 𝐵1 × 𝐵1 → SO(nrd) (𝐹) → 1
is exact.
Proof. For the first statement, we first show that the kernel of the action is the diagonally
embedded 𝐹 × . Suppose that 𝛼𝑣 𝛽−1 = 𝑣 for all 𝑣 ∈ 𝐵; taking 𝑣 = 1 shows 𝛽 = 𝛼, and
then we conclude that 𝛼𝑣 = 𝑣𝛼 for all 𝑣 ∈ 𝐵 so 𝛼 ∈ 𝑍 (𝐵) = 𝐹.
Next, the map 𝐵× × 𝐵× → GSO(nrd) (𝐹) is surjective. If 𝑓 ∈ GSO(nrd) (𝐹)
then nrd( 𝑓 (𝑥)) = 𝑢 nrd(𝑥) for all 𝑥 ∈ 𝐵, so in particular 𝑢 ∈ nrd(𝐵× ). Every such
similitude factor occurs, since the similitude factor of (𝛼, 1) is nrd(𝛼). So it suffices
to show that the map
{(𝛼, 𝛽) ∈ 𝐵× × 𝐵× : nrd(𝛼) = nrd(𝛽)} → SO(nrd) (𝐹)
is surjective. We again appeal to the Cartan–Dieudonné theorem; by the same com-

putation as in (4.5.12), we calculate that a reflection in 𝑥 ∈ 𝐵× is of the form
𝜏𝑥 (𝑣) = −𝑥𝑣 𝑥 −1 .
The product of two reflections for 𝑥, 𝑦 ∈ 𝐵× is thus of the form

−1
𝑣 ↦→ −𝑦(−𝑥𝑥𝑥 −1 )𝑦 −1 = (𝑦𝑥 −1 )𝑣(𝑦𝑥 −1 ) = 𝛼𝑣 𝛽−1
where 𝛼 = 𝑦𝑥 −1 and 𝛽 = 𝛼, and in particular the action is by special similarities. We

conclude that (4.5.18) and the second sequence are both exact.
The final statement again follows by writing 𝐵× = 𝐵1 𝐹 × , and seeing the kernel as
×
𝐹 ∩ 𝐵1 = {±1}.
Example 4.5.19. When 𝐵 = M2 (𝐹), then nrd(𝐵× ) = det(GL2 (𝐹)) = 𝐹 × , giving the
exact sequence
1 → GL1 (𝐹) → GL2 (𝐹) × GL2 (𝐹) → GSO(det) (𝐹) → 1.
Exercises
Let 𝐹 be a field with char 𝐹 ≠ 2.
I 1. Give an algorithmic proof that every finite-dimensional quadratic space has a

normalized basis (Lemma 4.3.1).
2. Let 𝐹 = R and let
∞
∑︁
𝑉 = (𝑎 𝑛 ) 𝑛 : 𝑎 𝑛 ∈ R for all 𝑛 ≥ 0 and 𝑎 2𝑛 converges .
𝑛=0
Show that 𝑉 is an R-vector space, and the map 𝑄 : 𝑉 → R by 𝑄((𝑎 𝑛 ) 𝑛 ) =

Í ∞ 2
𝑛=0 𝑎 𝑛 is a quadratic form, and so 𝑉 is an example of an infinite-dimensional
quadratic space. [This example generalizes to the context of Hilbert spaces.]
3. Let 𝐵 be a quaternion algebra over 𝐹. Let 𝑁 : 𝐵 → 𝐹 and Δ : 𝐵 → 𝐹 be
defined by 𝑁 (𝛼) = trd(𝛼2 ) and Δ(𝛼) = trd(𝛼) 2 − 4 nrd(𝛼). Show that 𝑁, Δ are
quadratic forms on 𝐵, describe their associated bilinear forms, and compute a
normalized form (and basis) for each.
4. Generalize Exercise 2.15 as follows. Let 𝑄 : 𝑉 → 𝐹 be a quadratic form with
dim𝐹 𝑉 = 𝑛 and let 𝛾 ∈ O(𝑄).
a) If 𝑛 is odd and det 𝛾 = 1, then 𝛾 has a nonzero fixed vector (and therefore
restricts to the identity on a one-dimensional subspace of 𝑉).
b) If 𝑛 is even and det 𝛾 = −1, then 𝛾 has both eigenvalues −1 and 1.
5. Generalizing part of Exercise 4.3, let 𝐵 be an 𝐹-algebra with a standard involu-
tion. Show that the discriminant form
Δ: 𝐵 → 𝐹
Δ(𝛼) = trd(𝛼) 2 − 4 nrd(𝛼)
is a quadratic form.
6. Let 𝑄 : 𝑉 → 𝐹 be a quadratic form with dim𝐹 𝑉 < ∞ and associated bilinear

form 𝑇. The map
𝑉 → Hom𝐹 (𝑉, 𝐹)
𝑥 ↦→ (𝑦 ↦→ 𝑇 (𝑥, 𝑦))
is 𝐹-linear. Show that 𝑄 is nondegenerate if and only if this map is an isomor-

phism.
7. Write out the action (4.5.9) explicitly, as follows. Let 𝐵 = (𝑎, 𝑏 | 𝐹) and let
𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗.
(a) Show that the matrix of the action 𝑣 ↦→ 𝛼𝑣𝛼−1 in the 𝐹-basis 𝛽 = {𝑖, 𝑗, 𝑘 }
for 𝐵0 is [𝛼] where nrd(𝛼) [𝛼] is equal to
𝑡 2 − 𝑎𝑥 2 + 𝑏𝑦 2 − 𝑎𝑏𝑧 2 −2𝑎(𝑡𝑧 + 𝑥𝑦) −2(𝑡𝑦 + 𝑎𝑥𝑧)

2𝑏(𝑡𝑧 − 𝑥𝑦) 𝑡 2 + 𝑎𝑥 2 − 𝑏𝑦 2 − 𝑎𝑏𝑧2 2(𝑡𝑥 − 𝑏𝑦𝑧)
© ª
®
« 2𝑏(𝑎𝑥𝑧 − 𝑡𝑦) 2𝑎(𝑡𝑥 + 𝑏𝑦𝑧) 𝑡 2 + 𝑎𝑥 2 + 𝑏𝑦 2 + 𝑎𝑏𝑧 2 ¬
and nrd(𝛼) = 𝑡 2 − 𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧 2 .

(b) Let 𝑄 = nrd | 𝐵0 and let 𝑇 be the associated bilinear form. Show that
the Gram matrix [𝑇] in the basis 𝛽 is the diagonal matrix with entries
−2𝑎, −2𝑏, 2𝑎𝑏. Then confirm by direct calculation that
[𝛼] ∈ SO(𝑄) (𝐹) = { 𝐴 ∈ SL3 (𝐹) : 𝐴[𝑇] 𝐴t = [𝑇]}.
8. In this exercise, we prove the chain lemma. Let 𝐵 := (𝑎, 𝑏 | 𝐹) be a quaternion

algebra.
(a) Show that if 𝑖 0 is orthogonal to 1, 𝑗, then (𝑖 0) 2 = 𝑎 0 ∈ 𝐹 × and 𝑖 0, 𝑗 are
standard generators for 𝐵, so 𝐵 ' (𝑎 0, 𝑏 | 𝐹).
(b) Let 𝐵 0 := (𝑎 0, 𝑏 0 | 𝐹), and suppose that 𝐵 is isomorphic to 𝐵 0. Show that
there exists 𝑐 ∈ 𝐹 × such that
𝑐, 𝑏 0
0 0
𝑎, 𝑏 𝑐, 𝑏 𝑎 ,𝑏
𝐵= ' ' ' .
𝐹 𝐹 𝐹 𝐹
∼ 𝐵 be the isomorphism, and take an element orthogonal
[Hint: let 𝜙 : 𝐵 0 −→
0
to 1, 𝑗, 𝜙( 𝑗 ).]
9. In this exercise, we develop some of the notions mentioned in Remark 3.3.8 in
the context of quadratic forms.
Let 𝐵 be a finite-dimensional 𝐹-algebra (not necessarily a quaternion algebra),
and let Tr : 𝐵 → 𝐹 be the left algebra trace (the trace of the endomorphism
given by left multiplication).
(a) Show that the map 𝐵 → 𝐹 defined by 𝑥 ↦→ Tr(𝑥 2 ) is a quadratic form on
𝐵; this form is called the (left) trace form on 𝐵.
(b) Compute the trace form of 𝐴 × 𝐵 and 𝐴 ⊗𝐹 𝐵 in terms of the trace form of
𝐴 and 𝐵.
(c) Show that if 𝐾 ⊇ 𝐹 is a inseparable field extension of finite degree, then the
trace form on 𝐾 (as an 𝐹-algebra) is identically zero. On the other hand,
show that if 𝐾/𝐹 is a finite separable field extension (with char 𝐹 ≠ 2)
then the trace form is nondegenerate.
√
(d) Compute the trace form on Q( 5) and Q(𝛼) where 𝛼 = 2 cos(2𝜋/7), so
that 𝛼3 + 𝛼2 − 2𝛼 − 1 = 0.
I 10. Let 𝑄 : 𝑉 → 𝐹 and 𝑄 0 : 𝑉 0 → 𝐹 be quadratic forms over 𝐹 with dim𝐹 𝑉 =
dim𝐹 𝑉 0 = 𝑛 < ∞, and let 𝑇, 𝑇 0 be the associated bilinear forms. Suppose
that there is a similarity 𝑄 ∼ 𝑄 0 with similitude factor 𝑢 ∈ 𝐹 × . Show that
det 𝑇 0 = 𝑢 𝑛 det 𝑇 ∈ 𝐹/𝐹 ×2 .
11. Let 𝑄 : 𝑉 → 𝐹 be a nondegenerate quadratic form with dim𝐹 𝑉 = 𝑛 < ∞.
(a) A subspace 𝑊 ⊆ 𝑉 is totally isotropic if 𝑄| 𝑊 = 0 is identically zero.

The Witt index 𝜈(𝑄) of 𝑄 is the maximal dimension of a totally isotropic
subspace. Show that if 𝜈(𝑄) = 𝑚 then 2𝑚 ≤ 𝑛.
(b) A Pfister form is a form in 2𝑚 variables defined inductively by hh𝑎ii =
h1, −𝑎i and
hh𝑎 1 , . . . , 𝑎 𝑚−1 , 𝑎 𝑚 ii = hh𝑎 1 , . . . , 𝑎 𝑚−1 ii −𝑎 𝑚 hh𝑎 1 , . . . , 𝑎 𝑚−1 ii.

𝑎, 𝑏
Show that the reduced norm nrd on is the Pfister form hh𝑎, 𝑏ii.
𝐹
(c) The hyperbolic plane is the quadratic form 𝐻 : 𝐹 2 → 𝐹 with 𝐻 (𝑥, 𝑦) =
𝑥𝑦. A quadratic form 𝑄 is a hyperbolic plane if 𝑄 ' 𝐻. A quadratic form
𝑄 is totally hyperbolic if 𝑄 ' 𝐻 · · · 𝐻 where 𝐻 is a hyperbolic plane.
Show that if 𝑄 is an isotropic Pfister form, then 𝑄 is totally hyperbolic.
(d) Suppose that 𝑄 is an isotropic Pfister form with 𝑛 ≥ 4. Let 𝑊 ⊂ 𝑉 be
a subspace of dimension 𝑛 − 1. Show that 𝑄| 𝑊 is isotropic. [This gives
another proof of Main Theorem 5.4.4 (iii) ⇒ (iv).]
12. (a) Let 𝐵 be a quaternion algebra over 𝐹. Show that the reduced norm is
the unique nonzero quadratic form 𝑄 on 𝐵 that is multiplicative, i.e.,
𝑄(𝛼𝛽) = 𝑄(𝛼)𝑄(𝛽) for all 𝛼, 𝛽 ∈ 𝐵.
(b) Show that (a) does not necessarily hold more generally, for 𝐵 an algebra
with a standard involution. [Hint: consider upper triangular matrices.]
I 13. In this exercise, we pursue some geometric notions for readers some background
in algebraic geometry (at the level of Hartshorne [Har77, Chapter 1]).
Let 𝑄 be nonzero quadratic form on 𝑉 with dim𝐹 𝑉 = 𝑛. The vanishing locus
of 𝑄(𝑥) = 0 defines a projective variety 𝑋 ⊆ P(𝑉) ' P𝑛 of degree 2 called a
quadric. Show that the quadratic form 𝑄 is nondegenerate if and only if the
projective variety 𝑋 is nonsingular. [For this reason, a nondegenerate quadratic
form is also synonymously called nonsingular.]
14. In this exercise, we work out from scratch Example 4.5.13: we translate the
results on rotations in section 2.4 to 𝐵 = M2 (R), but with respect to a different
measure of ‘length’.
Let

𝑥 𝑦
M2 (R) 0 = {𝑣 ∈ M2 (R) : tr(𝑣) = 0} = : 𝑥, 𝑦, 𝑧 ∈ R .
𝑧 −𝑥
For 𝑣 ∈ M2 (R) 0 , we have det(𝑣) = −𝑥 2 − 𝑦𝑧. Show that the group
M2 (R) 1 = SL2 (R) = {𝛼 ∈ M2 (R) : det(𝛼) = 1}
acts linearly on M2 (R) 0 by conjugation (the adjoint representation) preserving

the determinant, giving rise to an exact sequence
1 → {±1} → SL2 (R) → SO(det) → 1.
I 15. Let 𝑄 : 𝑉 → 𝐹 be a quadratic form with 𝑉 finite-dimensional over 𝐹. Show

that SO(𝑄) ≤ O(𝑄) is a (normal) subgroup of index 2. What can you say about
GSO(𝑄) ≤ GO(𝑄)?
I 16. In this exercise, we prove Theorem 4.2.22. Let 𝑄 : 𝑉 → 𝐹 be a quadratic form
with dim𝐹 𝑉 < ∞ and let 𝑇 be its associated bilinear form.
(a) Let 𝑣 ∈ 𝑉 be anisotropic. Define the reflection along 𝑣 by
𝜏𝑣 : 𝑉 → 𝑉
𝑇 (𝑣, 𝑥)
𝜏𝑣 (𝑥) = 𝑥 − 𝑣.
𝑄(𝑣)
Observe that 𝜏𝑣 is 𝐹-linear, and then show that 𝜏𝑣 ∈ O(𝑉) with det 𝜏𝑣 = −1.
[Hint: extend 𝑣 to a basis of the orthogonal complement of 𝑉.] Why is 𝜏𝑣
called a reflection?
(b) If 𝑥, 𝑦 ∈ 𝑉 are anisotropic with 𝑄(𝑥) = 𝑄(𝑦), show that there exists
𝑓 ∈ O(𝑉) such that 𝑓 (𝑥) = 𝑦. [Hint: reflect along either 𝑣 = 𝑥 + 𝑦 or
𝑣 = 𝑥 − 𝑦 as at least one is anisotropic, in the former case postcomposing
with reflection along 𝑥.]
(c) Let 𝑄 0 : 𝑉 0 → 𝐹 be another quadratic form, and let 𝑓 : 𝑉 −→ ∼ 𝑉 0 be an
⊥
isometry. For 𝑊 ⊆ 𝑉, show that 𝑓 (𝑊 ) = 𝑓 (𝑊) . ⊥
(d) Prove Theorem 4.2.22(a). [Hint: reduce to the case where dim𝐹 𝑊1 =
dim𝐹 𝑊10 = 1; apply parts (b) and (c).]
(e) Prove Theorem 4.2.22(b). [Hint: compare the isometry 𝑉 ' 𝑉 0 with the
isometry 𝑔.]
I 17. Prove the following weakened version of the Cartan–Dieudonné theorem (The-
orem 4.5.7): Let (𝑉, 𝑄) be a nondegenerate quadratic space with dim𝐹 𝑉 = 𝑛.
Show that every isometry 𝑓 ∈ O(𝑄) (𝐹) is a product of at most 2𝑛−1 reflections.
[Hint: in the proof of Exercise 4.16(b), note that 𝑓 can be taken to be a product
of at most 2 reflections, and finish by induction.]
Chapter 5
Ternary quadratic forms and

quaternion algebras
Continuing our treatment of quadratic forms, in this chapter we connect quaternion

algebras to ternary quadratic forms.
5.1 ⊲ Reduced norm as quadratic form

Let 𝐹 be a field with char 𝐹 ≠ 2 and let 𝐵 = (𝑎, 𝑏 | 𝐹) be a quaternion algebra over
𝐹. We saw in the previous chapter (4.1.1) that the reduced norm defines a quadratic
form. But always have scalar norms nrd(𝑡) = 𝑡 2 for 𝑡 ∈ 𝐹, so the form carries the same
information when restricted to the space of pure quaternions
𝐵0 := {𝛼 ∈ 𝐵 : trd(𝛼) = 0}
with basis 𝑖, 𝑗, 𝑖 𝑗. This quadratic form restricted to 𝐵0 is
nrd(𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗) = −𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧 2
with discriminant (−𝑎) (−𝑏) (𝑎𝑏) = (𝑎𝑏) 2 , so the trivial class in 𝐹 × /𝐹 ×2 .

We might now try to classify quaternion algebras over 𝐹 up to isomorphism in
terms of this quadratic form. Recall as in the previous chapters that for morphisms
between quadratic forms, one allows either isometries, an invertible change of basis
preserving the quadratic form, or similarities, which allow a rescaling of the quadratic
form by a nonzero element of 𝐹. Our main result is as follows (Corollary 5.2.6).
Theorem 5.1.1. The map 𝐵 ↦→ nrd | 𝐵0 induces a bĳection:

( Ternary quadratic forms over 𝐹 )
Quaternion algebras over 𝐹
↔ with discriminant 1 ∈ 𝐹 × /𝐹 ×2
up to isomorphism
up to isometry
( Nondegenerate ternary )
↔ quadratic forms over 𝐹
up to similarity
63
64 CHAPTER 5. TERNARY QUADRATIC FORMS
The map 𝐵 ↦→ nrd | 𝐵0 in Theorem 5.1.1 has inverse defined by the even Clifford
algebra (see section 5.3). The similarity class of a nondegenerate ternary quadratic
form cuts out a well-defined plane conic 𝐶 ⊆ P2 over 𝐹, so one also has a bĳection
between isomorphism classes of quaternion algebras over 𝐹 and isomorphism classes
of conics over 𝐹. Finally, keeping track of an orientation allows one to fully upgrade
this bĳection to an equivalence of categories (Theorem 5.6.8).
The classification of quaternion algebras over 𝐹 is now rephrased in terms of
quadratic forms, and a more detailed description depends on the field 𝐹. In this vein,
the most basic question we can ask about a quaternion algebra 𝐵 is if it is isomorphic
to the matrix ring 𝐵 ' M2 (𝐹): if so, we say that 𝐵 is split over 𝐹. For example, every
quaternion algebra over C (or an algebraically closed field) is split, and a quaternion
algebra (𝑎, 𝑏 | R) is split if and only if 𝑎 > 0 or 𝑏 > 0.
Ultimately, we will identify six equivalent ways (Main Theorem 5.4.4) to check if
a quaternion algebra 𝐵 is split; in light of Theorem 5.1.1, we isolate the following.
Proposition 5.1.2. 𝐵 is split if and only if the quadratic form nrd | 𝐵0 represents 0
nontrivially.
In later chapters, we will return to this classification problem, gradually increasing

the “arithmetic complexity” of the field 𝐹.
5.2 Isomorphism classes of quaternion algebras

In Section 2.4, we found that the unit Hamiltonians act by conjugation on the pure
quaternions H0 ' R3 as rotations, preserving the standard inner product. In this
section, we return to this theme for a general quaternion algebra, and we characterize
isomorphism classes of quaternion algebras in terms of isometry classes of ternary
quadratic forms.
Throughout this chapter, let 𝐹 be a field with char 𝐹 ≠ 2, and let 𝐵 = (𝑎, 𝑏 | 𝐹) be
a quaternion algebra over 𝐹.
Definition 5.2.1. 𝛼 ∈ 𝐵 is scalar if 𝛼 ∈ 𝐹 and pure if trd(𝛼) = 0.
5.2.2. Recalling (3.3.5), we have the 𝐹-vector space of pure (trace 0) elements of 𝐵
given by 𝐵0 = {1}⊥ . The standard involution restricted to 𝐵0 is given by 𝛼 = −𝛼 for
𝛼 ∈ 𝐵0 , so equivalently 𝐵0 is the −1-eigenspace for . We have 𝐵0 = 𝐹𝑖 ⊕ 𝐹 𝑗 ⊕ 𝐹𝑖 𝑗
and in this basis
nrd | 𝐵0 ' h−𝑎, −𝑏, 𝑎𝑏i (5.2.3)
so that disc(nrd | 𝐵0 ) = (𝑎𝑏) 2 = 1 ∈ 𝐹 × /𝐹 ×2 (cf. Example 4.3.8).
Proposition 5.2.4. Let 𝐵, 𝐵 0 be quaternion algebras over 𝐹. Then the following are
equivalent:
(i) 𝐵 ' 𝐵 0 are isomorphic as 𝐹-algebras;

(ii) 𝐵 ' (𝐵 0) op are isomorphic as 𝐹-algebras;
(iii) 𝐵 ' 𝐵 0 are isometric as quadratic spaces; and
(iv) 𝐵0 ' (𝐵 0) 0 are isometric as quadratic spaces.
5.2. ISOMORPHISM CLASSES OF QUATERNION ALGEBRAS 65
∼ (𝐵 0) 0 is an isometry, then 𝑓 extends uniquely to either an isomorphism

If 𝑓 : 𝐵0 −→
∼
𝑓 : 𝐵 −→ 𝐵 0 or an isomorphism 𝑓 : 𝐵 −→∼ (𝐵 0) op of 𝐹-algebras.
Proof. We follow Lam [Lam2005, Theorem III.2.5]. The equivalence (i) ⇔ (ii)
follows from postcomposing with the standard involution : 𝐵 0 −→ ∼ (𝐵 0) op .
The implication (i) ⇒ (iii) follows from the fact that the standard involution on an
algebra is unique and the reduced norm is determined by this standard involution, so
the reduced norm on 𝐵 is identified with the reduced norm on 𝐵 0.
The implication (iii) ⇒ (iv) follows from Witt cancellation (Theorem 4.2.22); and
(iv) ⇒ (iii) is immediate, since 𝐵 = h1i 𝐵0 and 𝐵 0 = h1i (𝐵 0) 0 so the isometry
extends by mapping 1 ↦→ 1. (Or use Witt extension, Theorem 4.2.22(b).)
So finally we prove (iv) ⇒ (i). Let 𝑓 : 𝐵0 → (𝐵 0) 0 be an isometry of quadratic
spaces. Suppose 𝐵 ' (𝑎, 𝑏 | 𝐹). Since 𝑓 is an isometry, nrd( 𝑓 (𝑖)) = nrd(𝑖) = −𝑎 and
nrd( 𝑓 (𝑖)) = 𝑓 (𝑖) 𝑓 (𝑖) = − 𝑓 (𝑖) 2
so 𝑓 (𝑖) 2 = 𝑎. Similarly 𝑓 ( 𝑗) 2 = 𝑏. Finally, 𝑗𝑖 = −𝑖 𝑗 since 𝑖, 𝑗 are orthogonal (as in

the proof of Main Theorem 4.4.1), but then 𝑓 (𝑖), 𝑓 ( 𝑗) are orthogonal as well and so
𝑓 ( 𝑗) 𝑓 (𝑖) = − 𝑓 (𝑖) 𝑓 ( 𝑗).
Similarly, we know that 𝑖 𝑗 is orthogonal to 𝑖, 𝑗, thus 𝑓 (𝑖 𝑗) is orthogonal to both
𝑓 (𝑖) and 𝑓 ( 𝑗) and so 𝑓 (𝑖 𝑗) = 𝑢 𝑓 (𝑖) 𝑓 ( 𝑗) for some 𝑢 ∈ 𝐹 × ; taking reduced norms gives
nrd(𝑖 𝑗) = 𝑢 2 nrd(𝑖) nrd( 𝑗) so 𝑢 2 = 1 thus 𝑢 = ±1. If 𝑢 = 1, then 𝑓 (𝑖 𝑗) = 𝑓 (𝑖) 𝑓 ( 𝑗),
and 𝑓 extends via 𝑓 (1) = 1 to an 𝐹-algebra isomorphism 𝐵 −→ ∼ 𝐵 0. Otherwise,
𝑢 = −1 and 𝑓 (𝑖 𝑗) = − 𝑓 (𝑖) 𝑓 ( 𝑗) = 𝑓 ( 𝑗) 𝑓 (𝑖), in which case 𝑓 extends to an 𝐹-algebra
anti-isomorphism, or equivalently an 𝐹-algebra isomorphism 𝐵 −→ ∼ (𝐵 0) op ; but then
postcomposing with the standard involution we obtain an 𝐹-algebra isomorphism
𝐵 −→ ∼ 𝐵 0.
Main Theorem 5.2.5. Let 𝐹 be a field with char 𝐹 ≠ 2. Then the functor 𝐵 ↦→ nrd | 𝐵0
yields an equivalence of categories between
Quaternion algebras over 𝐹,

under 𝐹-algebra isomorphisms and anti-isomorphisms
and
Ternary quadratic forms over 𝐹 with discriminant 1 ∈ 𝐹 × /𝐹 ×2 ,

under isometries.
Proof. The association 𝐵 ↦→ nrd | 𝐵0 gives a functor from quaternion algebras to

nondegenerate ternary quadratic forms with discriminant 1, by 5.2.2; the map sends
isomorphisms and anti-isomorphisms to isometries and vice versa by Proposition
5.2.4. Therefore the functor is fully faithful. To conclude, we show that the functor
is essentially surjective. Let 𝑉 be a nondegenerate ternary quadratic space with
discriminant 1 ∈ 𝐹 × /𝐹 ×2 . Choose a normalized basis for 𝑉, so that 𝑄 ' h−𝑎, −𝑏, 𝑐i
with 𝑎, 𝑏, 𝑐 ∈ 𝐹 × . By hypothesis, we have disc(𝑄) = 𝑎𝑏𝑐 ∈ 𝐹 ×2 , so applying the
isometry rescaling the third basis vector we may suppose 𝑐 = 𝑎𝑏. We then associate to
𝑉 the isomorphism class of the quaternion algebra (𝑎, 𝑏 | 𝐹). The result follows.
Corollary 5.2.6. The map 𝐵 ↦→ nrd | 𝐵0 yields a bĳection
( Ternary quadratic forms over 𝐹 )

Quaternion algebras over 𝐹
↔ with discriminant 1 ∈ 𝐹 × /𝐹 ×2
up to isomorphism
up to isometry
up to similarity
that is functorial with respect to 𝐹.
By the expression functorial with respect to 𝐹, we mean that this bĳection respects
(is compatible with) field extensions: explicitly, if 𝐹 ↩→ 𝐾 is an inclusion of fields,
and 𝐵 is a quaternion algebra with associated ternary quadratic form 𝑄 : 𝐵0 → 𝐹,
then the quaternion algebra 𝐵 𝐾 = 𝐵 ⊗𝐹 𝐾 has associated ternary quadratic form
𝑄 𝐾 : 𝐵0𝐾 = 𝐵0 ⊗𝐹 𝐾 → 𝐾.
Proof of Corollary 5.2.6. Functoriality boils down to the fact that
(𝐵 𝐾 ) 0 = (𝐵 ⊗𝐹 𝐾) 0 = 𝐵0 ⊗𝐹 𝐾
for 𝐹 ↩→ 𝐾 an inclusion of fields. The first bĳection is an immediate consequence of

Main Theorem 5.2.5. We do not need anti-isomorphisms once we restrict to classes,
since if there is an anti-isomorphism 𝐵 −→ ∼ 𝐵 0 then composing with the standard
involution gives a straight up isomorphism.
Next, we examine the natural map from isometry classes to similarity classes and
show it is surjective. Every nondegenerate ternary quadratic form (or any quadratic
form in odd dimension) is similar to a unique isometry class of quadratic forms with
trivial discriminant: if 𝑄 = h𝑎, 𝑏, 𝑐i with 𝑎, 𝑏, 𝑐 ∈ 𝐹 × , then disc(h𝑎, 𝑏, 𝑐i) = 𝑎𝑏𝑐 and
𝑄 = h𝑎, 𝑏, 𝑐i ∼ 𝑎𝑏𝑐h𝑎, 𝑏, 𝑐i = h𝑎 2 𝑏𝑐, 𝑎𝑏 2 𝑐, 𝑎𝑏𝑐2 i ' h𝑏𝑐, 𝑎𝑐, 𝑎𝑏i
and disc(h𝑏𝑐, 𝑎𝑐, 𝑎𝑏i) = (𝑎𝑏𝑐) 2 = 1 ∈ 𝐹 × /𝐹 ×2 . Therefore the map is surjective.

To conclude, we show this map is injective. Suppose that 𝑄, 𝑄 0 are forms of
discriminant 1, so det 𝑇, det 𝑇 0 ∈ 𝐹 ×2 . Suppose there is a similarity 𝑄 ∼ 𝑄 0, so
𝑄 0 ( 𝑓 (𝑥)) = 𝑢𝑄(𝑥) for some 𝑓 : 𝑉 → 𝑉 0 and 𝑢 ∈ 𝐹 × ; we show in fact that 𝑄 ' 𝑄 0 are
isometric. By Exercise 4.10, we have det 𝑇 0 = 𝑢 3 det 𝑇, and 𝑢 = 𝑐2 ∈ 𝐹 ×2 . Therefore
𝑄 0 (𝑐−1 𝑓 (𝑥)) = 𝑐−2 𝑄 0 ( 𝑓 (𝑥)) = 𝑢 −1 𝑄 0 ( 𝑓 (𝑥)) = 𝑄(𝑥)
∼ 𝑉 0 is the sought after isometry.

and 𝑐−1 𝑓 : 𝑉 −→
Remark 5.2.7. We will refine Main Theorem 5.2.5 in section 5.6 by restricting the
isometries to those that preserve orientation.
5.3. CLIFFORD ALGEBRAS 67
5.3 Clifford algebras

In this section, we define a functorial inverse to 𝐵 ↦→ nrd | 𝐵0 = 𝑄 in Main Theorem
5.2.5: this is the even Clifford algebra of 𝑄. The Clifford algebra is useful in many
contexts, so we define it more generally. Loosely speaking, the Clifford algebra of a
quadratic form 𝑄 is the algebra generated by 𝑉 subject to the condition 𝑥 2 = 𝑄(𝑥)
for all 𝑥 ∈ 𝑉, so the multiplication on the Clifford algebra is induced by the quadratic
form.
Let 𝑄 : 𝑉 → 𝐹 be a quadratic form with dim𝐹 𝑉 = 𝑛 < ∞; in this section, we
pause our assumption and allow 𝐹 of arbitrary characteristic.
Proposition 5.3.1. There exists an 𝐹-algebra Clf 𝑄 with the following properties:
(i) There is an 𝐹-linear map 𝜄 : 𝑉 → Clf 𝑄 such that 𝜄(𝑥) 2 = 𝑄(𝑥) for all 𝑥 ∈ 𝑉;
and
(ii) Clf 𝑄 has the following universal property: if 𝐴 is an 𝐹-algebra and 𝜄 𝐴 : 𝑉 → 𝐴
is a map such that 𝜄 𝐴 (𝑥) 2 = 𝑄(𝑥) for all 𝑥 ∈ 𝑉, then there exists a unique 𝐹-
algebra homomorphism 𝜙 : Clf 𝑄 → 𝐴 such that the diagram
𝑉
𝜄 / Clf 𝑄
𝜙
!
𝜄𝐴
𝐴
commutes.
The pair (Clf 𝑄, 𝜄) is unique up to unique isomorphism.
The algebra Clf 𝑄 in Proposition 5.3.1 is called the Clifford algebra of 𝑄.
Proof. Let
∞
Ê
Ten 𝑉 := 𝑉 ⊗𝑑 (5.3.2)
𝑑=0
where
𝑉 ⊗𝑑 := 𝑉 ⊗ · · · ⊗ 𝑉 and 𝑉 ⊗0 := 𝐹,
| {z }
𝑑
so that
Ten 𝑉 = 𝐹 ⊕ 𝑉 ⊕ (𝑉 ⊗ 𝑉) ⊕ . . . .
Then Ten 𝑉 has a multiplication given by tensor product: for 𝑥 ∈ 𝑉 ⊗𝑑 and 𝑦 ∈ 𝑉 ⊗𝑒 we
define
𝑥 · 𝑦 = 𝑥 ⊗ 𝑦 ∈ 𝑉 ⊗ (𝑑+𝑒)
(concatenate, and possibly distribute, tensors). In this manner, Ten 𝑉 has the structure
of an 𝐹-algebra, and we call Ten 𝑉 the tensor algebra of 𝑉.
Let
𝐼 (𝑄) = h𝑥 ⊗ 𝑥 − 𝑄(𝑥) : 𝑥 ∈ 𝑉i ⊆ Ten 𝑉 (5.3.3)
be the two-sided ideal generated the elements 𝑥 ⊗ 𝑥 − 𝑄(𝑥) for all 𝑥 ∈ 𝑉. Let
Clf 𝑄 = Ten 𝑉/𝐼 (𝑄). (5.3.4)
The algebra Clf 𝑄 by construction satisfies (i). And if 𝜄 𝐴 : 𝑉 → 𝐴 is as in (ii), then

the map 𝜄(𝑥) ↦→ 𝜄 𝐴 (𝑥) for 𝑥 ∈ 𝑉 extends to a unique 𝐹-algebra map Ten 𝑉 → 𝐴; since
further 𝜄 𝐴 (𝑥) 2 = 𝑄(𝑥) for all 𝑥 ∈ 𝑉, this algebra map factors through 𝜙 : Clf 𝑄 → 𝐴.
By abstract nonsense (taking 𝐴 = Clf 𝑄), we see that any other algebra having the
same property as Clf (𝑄) is uniquely isomorphic to it, i.e., Clf 𝑄 is unique up to unique
isomorphism.
Example 5.3.5. If 𝑄 : 𝐹 → 𝐹 is the quadratic form 𝑄(𝑥) = 𝑎𝑥 2 with 𝑎 ∈ 𝐹, then

Clf (𝐹) ' 𝐹 [𝑥]/(𝑥 2 − 𝑎) (Exercise 5.6).
É𝑛 Ó 𝑑
Example 5.3.6. In the extreme case where 𝑄 = 0 identically, Clf 𝑄 ' 𝑑=0 𝑉 is
canonically identified with the exterior algebra on 𝑉.
5.3.7. Let 𝑥, 𝑦 ∈ 𝑉. Then in Clf 𝑄, we have
(𝑥 + 𝑦) ⊗ (𝑥 + 𝑦) − 𝑥 ⊗ 𝑥 − 𝑦 ⊗ 𝑦 = 𝑄(𝑥 + 𝑦) − 𝑄(𝑥) − 𝑄(𝑦)

(5.3.8)
𝑥 ⊗ 𝑦 + 𝑦 ⊗ 𝑥 = 𝑇 (𝑥, 𝑦).
In particular, 𝑥, 𝑦 are orthogonal if and only if 𝑥 ⊗ 𝑦 = −𝑦 ⊗ 𝑥.

5.3.9. Let 𝑒 1 , . . . , 𝑒 𝑛 be an 𝐹-basis for 𝑉. Then finite tensors on these elements are
an 𝐹-basis for Ten 𝑉. In Clf 𝑄, by 5.3.7 we have 𝑒 𝑖 ⊗ 𝑒 𝑖 = 𝑄(𝑒 𝑖 ) and 𝑒 𝑗 ⊗ 𝑒 𝑖 =
𝑇 (𝑒 𝑖 , 𝑒 𝑗 ) − 𝑒 𝑖 ⊗ 𝑒 𝑗 , so an 𝐹-spanning set for Clf 𝑄 is given by 𝑒 𝑖1 ⊗ · · · ⊗ 𝑒 𝑖𝑑 with
1 ≤ 𝑖1 < 𝑖2 < · · · < 𝑖 𝑑 ≤ 𝑛 (including 1 arising from the empty tensor product), and
so
𝑛
∑︁ 𝑛
dim𝐹 Clf(𝑄) ≤ = 2𝑛 . (5.3.10)
𝑑=0
𝑑
It is customary to abbreviate 𝑒 𝑖1 ⊗ · · · ⊗ 𝑒 𝑖𝑑 = 𝑒 𝑖1 · · · 𝑒 𝑖𝑑 .
Example 5.3.11. If 𝑄 ' h𝑎 1 , . . . , 𝑎 𝑛 i is diagonal in the basis 𝑒 𝑖 , then
(𝑒 𝑖1 · · · 𝑒 𝑖𝑑 ) 2 = sgn(𝑖 1 . . . 𝑖 𝑑 )𝑒 2𝑖1 (𝑒 𝑖2 · · · 𝑒 𝑖𝑑 ) 2 = · · · = (−1) 𝑑 (𝑑−1)/2 𝑎 𝑖1 · · · 𝑎 𝑖𝑑 .
Example 5.3.12. Suppose char 𝐹 ≠ 2 and let 𝑄 : 𝐹 2 → 𝐹 be the quadratic form

𝑄(𝑥) = h𝑎, 𝑏i. Then by a direct calculation using 5.3.9, we find
Clf 𝑄 = 𝐹 ⊕ 𝐹𝑒 1 ⊕ 𝐹𝑒 2 ⊕ 𝐹𝑒 1 𝑒 2 (5.3.13)
with multiplication 𝑒 21 = 𝑎 and 𝑒 22 = 𝑏 and 𝑒 2 𝑒 1 = −𝑒 1 𝑒 2 , i.e., with 𝑖 := 𝑒 1 and 𝑗 := 𝑒 2

𝑎, 𝑏
we have identified Clf 𝑄 ' when 𝑎, 𝑏 ≠ 0.
𝐹
Example 5.3.12 generalizes as follows.
Lemma 5.3.14. The map 𝜄 : 𝑉 → Clf 𝑄 is injective, and dim𝐹 Clf(𝑄) = 2𝑛 .
5.3. CLIFFORD ALGEBRAS 69
Proof. By Lemma 4.3.1, we may choose a basis 𝑒 1 , . . . , 𝑒 𝑛 for 𝑉 in which 𝑄 '

h𝑎 1 , · · · , 𝑎 𝑛 i is diagonal. Let 𝐴 be the 𝐹-vector space with basis the symbols 𝑧𝑖1 · · · 𝑧𝑖𝑑
for 1 ≤ 𝑖1 < 𝑖 2 < · · · < 𝑖 𝑑 ≤ 𝑛. In the same way as what was considered for the
relations (2.2.2), we verify directly that there is a unique, associative multiplication
on 𝐴 such that 𝑧 2𝑖 = 𝑎 𝑖 and 𝑧 𝑗 𝑧𝑖 = −𝑧𝑖 𝑧 𝑗 . (Alternatively, this can be viewed as a
graded tensor product; see Exercise 5.21.) The map 𝜄 𝐴 : 𝑉 → 𝐴 by 𝑒 𝑖 ↦→ 𝑧𝑖 has
𝜄 𝐴 (𝑥) 2 = 𝜄 𝐴 ( 𝑖 𝑥 𝑖 𝑒 𝑖 ) 2 = 𝑎 1 𝑥12 + · · · + 𝑎 𝑛 𝑥 𝑛2 = 𝑄(𝑥), so by the universal property of
Í
Clf(𝑄), there exists a unique 𝐹-algebra homomorphism 𝜙 : Clf (𝑄) → 𝐴 such that
𝜙𝜄 = 𝜄 𝐴. Since the elements 𝑧𝑖1 · · · 𝑧 𝑖𝑑 are 𝐹-linearly independent in 𝐴, so too are their
preimages 𝑒 𝑖1 · · · 𝑒 𝑖𝑑 in Clf 𝑄, so the spanning set given in 5.3.9 is in fact a basis and
𝜙 is an isomorphism.
As it will cause no confusion, we may identify 𝑉 with its image 𝜄(𝑉) ↩→ Clf 𝑄.
5.3.15. The reversal map, given by
rev : Clf 𝑄 → Clf 𝑄

(5.3.16)
𝑥1 ⊗ · · · ⊗ 𝑥𝑟 ↦→ 𝑥𝑟 ⊗ · · · ⊗ 𝑥1
on pure tensors (and extended 𝐹-linearly) is well-defined, as it maps the ideal 𝐼 (𝑄) to
itself, and so it defines an involution on Clf 𝑄 that we call the reversal involution.
Lemma 5.3.17. The association 𝑄 ↦→ Clf 𝑄 induces a faithful functor from the
category of
quadratic forms over 𝐹, under isometries
to the category of
finite-dimensional 𝐹-algebras with involution, under isomorphisms.
Proof. Let 𝑄 0 : 𝑉 0 → 𝐹 be another quadratic form and let 𝑓 : 𝑉 → 𝑉 0 be an isometry.

Then 𝑓 induces an 𝐹-algebra map Ten 𝑉 → Ten(𝑉 0) and
𝑓 (𝑥 ⊗ 𝑥 − 𝑄(𝑥)) = 𝑓 (𝑥) ⊗ 𝑓 (𝑥) − 𝑄(𝑥) = 𝑓 (𝑥) ⊗ 𝑓 (𝑥) − 𝑄 0 ( 𝑓 (𝑥))
so 𝑓 also induces an 𝐹-algebra map Clf 𝑄 → Clf(𝑄 0). Repeating with the inverse
map, and applying the universal property, we see that these maps are inverse, so define
isomorphisms. The functor is faithful because 𝑉 ⊂ Clf 𝑄, so if 𝑓 : 𝑉 −→ ∼ 𝑉 acts as the
identity on Clf 𝑄 then it acts as the identity on 𝑉, so 𝑓 itself is the identity. (This can
be rephrased in terms of the universal property: see Exercise 5.13.)
5.3.18. The tensor algebra Ten 𝑉 has a natural Z ≥0 grading by degree, and by con-
struction (5.3.4), the quotient Clf 𝑄 = Ten 𝑉/𝐼 (𝑄) retains a Z/2Z-grading
Clf 𝑄 = Clf 0 𝑄 ⊕ Clf 1 𝑄
where Clf 0 𝑄 ⊆ Clf 𝑄 is the 𝐹-subalgebra of terms of even degree and Clf 1 𝑄 the
Clf 0 𝑄-bimodule of terms with odd degree. The reversal involution 5.3.15 preserves
Clf 0 𝑄 and so descends to an involution on Clf 0 𝑄.
We call Clf 0 𝑄 the even Clifford algebra and Clf 1 𝑄 the odd Clifford bimodule
of 𝑄. The former admits the following direct construction: let
∞
Ê
Ten0 𝑉 := 𝑉 ⊗2𝑑
𝑑=0
and let 𝐼 0 (𝑄) := 𝐼 (𝑄) ∩ Ten0 𝑉; then Clf 0 𝑄 ' Ten0 𝑉/𝐼 0 (𝑄).
5.3.19. Referring to 5.3.9, the elements 𝑒 1 𝑒 2 , . . . , 𝑒 𝑛−1 𝑒 𝑛 generate Clf 0 𝑄 as an 𝐹-

algebra, and Clf 0 𝑄 has basis 𝑒 𝑖1 · · · 𝑒 𝑖𝑑 where 𝑑 is even (including the empty product
1), so dim𝐹 Clf 0 (𝑄) = 2𝑛−1 .
Example 5.3.20. Continuing Example 5.3.12, we see that the reversal involution fixes
𝑖, 𝑗 and acts as the standard involution on Clf 0 𝑄. So the algebra Clf 𝑄 is not just a
quaternion algebra, but one retaining a Z/2Z-grading.
Lemma 5.3.21. The association 𝑄 ↦→ Clf 0 𝑄 defines a functor from the category of
quadratic forms over 𝐹, under similarities
to the category of
finite-dimensional 𝐹-algebras with involution, under isomorphisms.
Proof. Let 𝑄 0 : 𝑉 0 → 𝐹 be another quadratic form and let ( 𝑓 , 𝑢) be a similarity, with

𝑓 : 𝑉 → 𝑉 0 and 𝑢 ∈ 𝐹 × , so that 𝑢𝑄(𝑥) = 𝑄 0 ( 𝑓 (𝑥)) for all 𝑥 ∈ 𝑉. We modify the proof
in Lemma 5.3.17: we define a map
Ten0 𝑉 → Ten0 (𝑉 0)
𝑥 1 ⊗ · · · ⊗ 𝑥 𝑑 ↦→ (𝑢 −1 ) 𝑑/2 𝑓 (𝑥 1 ) ⊗ · · · ⊗ 𝑓 (𝑥 𝑑 ).
Then under this map, we have
𝑥 ⊗ 𝑥 − 𝑄(𝑥) ↦→ 𝑢 −1 ( 𝑓 (𝑥) ⊗ 𝑓 (𝑥)) − 𝑄(𝑥) = 𝑢 −1 ( 𝑓 (𝑥) ⊗ 𝑓 (𝑥) − 𝑄 0 ( 𝑓 (𝑥)))
so 𝐼 0 (𝑄) maps to 𝐼 0 (𝑄 0), and the induced map Clf 0 𝑄 → Clf 0 (𝑄 0) is an 𝐹-algebra
isomorphism.
5.3.22. Note that unlike the Clifford functor, the even Clifford functor need not be
faithful: for example, the map −1 : 𝐹 2 → 𝐹 2 has 𝑒 1 𝑒 2 ↦→ (−𝑒 1 ) (−𝑒 2 ) = 𝑒 1 𝑒 2 so acts
by the identity on Clf 0 𝑄.
We now come to the important immediate application.
5.3.23. Suppose that char 𝐹 ≠ 2 and let 𝑄(𝑥) = h𝑎, 𝑏, 𝑐i be a nondegenerate ternary
quadratic form. Then the even Clifford algebra Clf 0 𝑄 is given by
Clf 0 𝑄 = 𝐹 ⊕ 𝐹𝑖 ⊕ 𝐹 𝑗 ⊕ 𝐹𝑖 𝑗
5.4. SPLITTING 71
where 𝑖 = 𝑒 1 𝑒 2 , 𝑗 = 𝑒 2 𝑒 3 , subject to the multiplication
𝑖 2 = −𝑎𝑏, 𝑗 2 = −𝑏𝑐, 𝑖 𝑗 + 𝑗𝑖 = 0.
So
0 −𝑎𝑏, −𝑏𝑐
Clf 𝑄 ' .
𝐹
Letting 𝑘 = 𝑒 3 𝑒 1 , we obtain symmetrically with the other two pairs of generators 𝑗, 𝑘
or 𝑘, 𝑖 that
0 −𝑏𝑐, −𝑎𝑐 −𝑎𝑐, −𝑎𝑏
Clf 𝑄 ' ' .
𝐹 𝐹
The reversal involution is the standard involution on Clf 0 𝑄. Letting 𝐵 = Clf 0 𝑄,
nrd | 𝐵0 = h𝑎𝑏, 𝑏𝑐, 𝑎𝑐i ' h𝑎𝑏𝑐2 , 𝑎 2 𝑏𝑐, 𝑎𝑏 2 𝑐i = 𝑎𝑏𝑐h𝑎, 𝑏, 𝑐i.
So if disc 𝑄(𝑥) = 𝑎𝑏𝑐 ∈ 𝐹 ×2 , then nrd | 𝐵0 is isometric to 𝑄. In a similar way, if

𝑎, 𝑏
𝐵= , then in Main Theorem 5.2.5 we associate 𝑄 = nrd 𝐵0 = h−𝑎, −𝑏, 𝑎𝑏i, and
𝐹
−𝑎𝑏, 𝑎𝑏 2

0 𝑎, 𝑏
Clf 𝑄 ' ' . (5.3.24)
𝐹 𝐹
This gives another tidy proof of the bĳection in Corollary 5.2.6.
Remark 5.3.25. The even Clifford map does not furnish an equivalence of categories
for the same reason as in 5.3.22; one way to deal with this issue is to restrict the
isometries to those that preserve orientation: we carry this out in section 5.6.
5.4 Splitting
The moral of Main Theorem 5.2.5 is that the problem of classifying quaternion algebras
depends on the theory of ternary quadratic forms over that field (and vice versa). We
now pursue the first consequence of this moral, and we characterize the matrix ring
among quaternion algebras. Suppose that char 𝐹 ≠ 2, but still 𝑄 : 𝑉 → 𝐹 a quadratic
form with dim𝐹 𝑉 < ∞.
Definition 5.4.1. The hyperbolic plane is the quadratic form 𝐻 : 𝐹 2 → 𝐹 defined by

𝐻 (𝑥, 𝑦) = 𝑥𝑦. A quadratic form is a hyperbolic plane if it is isometric to 𝐻.
A hyperbolic plane 𝐻 is universal, its associated bilinear form has Gram matrix
0 1
in the standard basis, and 𝐻 has normalized form 𝐻 ' h1, −1i.
1 0
Lemma 5.4.2. Suppose 𝑄 is nondegenerate. Then 𝑄 is isotropic if and only if there

exists an isometry 𝑄 ' 𝐻 𝑄 0 with 𝑄 0 nondegenerate and 𝐻 a hyperbolic plane.
Proof. For the implication (⇐), we have an isotropic vector from either one of the
two basis vectors. For the implication (⇒), let 𝑥 ∈ 𝑉 be isotropic, so 𝑥 ≠ 0 and
satisfy 𝑄(𝑥) = 0. Since 𝑄 is nondegenerate, there exists 𝑦 ∈ 𝑉 such that 𝑇 (𝑥, 𝑦) ≠ 0;
rescaling 𝑦, we may assume 𝑇 (𝑥, 𝑦) = 1. If 𝑇 (𝑦, 𝑦) = 2𝑄(𝑦, 𝑦) ≠ 0, replacing
𝑦 ← 𝑦 − 2𝑥/𝑇 (𝑦, 𝑦) gives 𝑇 (𝑦, 𝑦) = 0. Thus 𝑄 restricted to 𝐹𝑥 + 𝐹 𝑦 is isometric
to 𝐻, and in particular is nondegenerate. Therefore letting 𝑉 0 := (𝐹𝑥 + 𝐹 𝑦) ⊥ and
𝑄 0 := 𝑄| 𝑉 0 , we have 𝑉 ' (𝐹𝑥 + 𝐹 𝑦) 𝑉 0 and 𝑄 ' 𝐻 𝑄 0.
Lemma 5.4.3. Suppose 𝑄 is nondegenerate and let 𝑎 ∈ 𝐹 × . Then the following are
equivalent:
(i) 𝑄 represents 𝑎;
(ii) 𝑄 ' h𝑎i 𝑄 0 for some nondegenerate form 𝑄 0; and
(iii) h−𝑎i 𝑄 is isotropic.
Proof. For (i) ⇒ (ii), we take 𝑄 0 = 𝑄| 𝑊 and 𝑊 = {𝑣}⊥ ⊂ 𝑉 where 𝑄(𝑣) = 𝑎. For (ii)
⇒ (iii), we note that h−𝑎i 𝑄 ' h𝑎, −𝑎i 𝑄 0 is isotropic. For (iii) ⇒ (i), suppose
(h−𝑎i 𝑄) (𝑣) = 0, so 𝑄(𝑣) = 𝑎𝑥 2 for some 𝑥 ∈ 𝐹. If 𝑥 = 0, then 𝑄 is isotropic and
by Lemma 5.4.2 represents 𝑎; if 𝑥 ≠ 0, then by homogeneity 𝑄(𝑣/𝑥) = 𝑎 and again 𝑄
represents 𝑎.
We now come to a main result.

𝑎, 𝑏
Main Theorem 5.4.4. Let 𝐵 = be a quaternion algebra over 𝐹 (with char 𝐹 ≠
𝐹
2). Then the following are equivalent:

1, 1
(i) 𝐵 ' ' M2 (𝐹);
𝐹
(ii) 𝐵 is not a division ring;
(iii) The quadratic form nrd ' h1, −𝑎, −𝑏, 𝑎𝑏i is isotropic;
(iv) The quadratic form nrd | 𝐵0 ' h−𝑎, −𝑏, 𝑎𝑏i is isotropic;
(v) The binary form h𝑎, 𝑏i represents 1;
(vi) 𝑏 ∈ Nm𝐾 |𝐹 (𝐾 × ) where 𝐾 = 𝐹 [𝑖]; √ and
(vi0) 𝑏 ∈ Nm𝐾 |𝐹 (𝐾 × ) where 𝐾 = 𝐹 ( 𝑎).
Condition (vi) holds if and only if there exist 𝑥, 𝑦 ∈ 𝐹 such that 𝑥 2 − 𝑎𝑦 2 = 𝑏; if
√ not a field then 𝐾 ' 𝐹 × 𝐹 and
𝐾 is Nm𝐾 |𝐹 (𝐾 × ) = 𝐹 × . In condition (vi0), we take
𝐹 ( 𝑎) to be a splitting field for 𝑥 − 𝑎 over 𝐹, so equal to 𝐹 if 𝑎 ∈ 𝐹 × . (Depending
2
on the circumstances, one of these formulations may be more natural than the other.)
Proof. We follow Lam [Lam2005, Theorem 2.7]. The isomorphism (1, 1 | 𝐹) '
M2 (𝐹) in (i) follows from Example 2.2.4. The implication (i) ⇒ (ii) is clear. The
equivalence (ii) ⇔ (iii) follows from the fact that 𝛼 ∈ 𝐵× if and only if nrd(𝛼) ∈ 𝐹 ×
(Exercise 3.5).
We now prove (iii) ⇒ (iv). Let 0 ≠ 𝛼 ∈ 𝐵 be such that nrd(𝛼) = 0. If trd(𝛼) = 0,
then we are done. Otherwise, trd(𝛼) ≠ 0. Let 𝛽 be orthogonal to 1, 𝛼, so that
trd(𝛼𝛽) = 0. We cannot have both 𝛼𝛽 = 0 and 𝛼𝛽 = (trd(𝛼) − 𝛼) 𝛽 = 0, so we may
suppose 𝛼𝛽 ≠ 0. But then nrd(𝛼𝛽) = nrd(𝛼) nrd(𝛽) = 0 as desired.
5.4. SPLITTING 73
To complete the equivalence of the first four we prove (iv) ⇒ (i). Let 𝛽 ∈ 𝐵0
satisfy nrd(𝛽) = 0. Since nrd | 𝐵0 is nondegenerate, there exists 0 ≠ 𝛼 ∈ 𝐵0 such
that trd(𝛼𝛽) ≠ 0. Therefore, the restriction of nrd to 𝐹𝛼 ⊕ 𝐹 𝛽 is nondegenerate
and isotropic. By Lemma 5.4.2, we conclude there exists a basis for 𝐵0 such that
nrd | 𝐵0 ' h1, −1i h𝑐i = h1, −1, 𝑐i; but disc(nrd | 𝐵0 ) = −𝑐 ∈ 𝐹 ×2 by 5.2.2; rescaling,
we may suppose 𝑐 = −1. But then by Proposition 5.2.4 we have 𝐵 ' (1, 1 | 𝐹).
Now we show (iv) ⇒ (v). For 𝛼 ∈ 𝐵0 ,
nrd(𝛼) = nrd(𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗) = −𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧 2
as in 5.2.2. Suppose nrd(𝛼) = 0. If 𝑧 = 0, then the binary form h𝑎, 𝑏i is isotropic so

is a hyperbolic plane by Lemma 5.4.2 and thus represents 1. If 𝑧 ≠ 0 then
2 2
𝑦 𝑥
𝑎 +𝑏 = 1.
𝑎𝑧 𝑏𝑧
Next we prove (v) ⇒ (vi). If 𝑎 ∈ 𝐹 ×2 then 𝐾 ' 𝐹 × 𝐹 and Nm𝐾 |𝐹 (𝐾 × ) = 𝐹 × 3 𝑏.

If 𝑎 ∉ 𝐹 ×2 , then given 𝑎𝑥 2 + 𝑏𝑦 2 = 1 we must have 𝑦 ≠ 0 so
2 2 √
1 𝑥 1−𝑥 𝑎
−𝑎 = Nm𝐾 |𝐹 = 𝑏.
𝑦 𝑦 𝑦
In the equivalence (vi) ⇔ (vi0), the two statements are identical if 𝑎 ∉ 𝐹 ×2 and
both automatically satisfied if 𝑎 ∈ 𝐹 ×2 .
To conclude, we prove (vi) ⇒ (iii). If 𝑏 = 𝑥 2 − 𝑎𝑦 2 ∈ Nm𝐾 |𝐹 (𝐾 × ), then 𝛼 =
𝑥 + 𝑦𝑖 + 𝑗 ≠ 0 has nrd(𝛼) = 𝑥 2 − 𝑎𝑦 2 − 𝑏 = 0.
We give a name to the equivalent conditions in Main Theorem 5.4.4.
Definition 5.4.5. A quaternion algebra 𝐵 over 𝐹 is split if 𝐵 ' M2 (𝐹). A field 𝐾

containing 𝐹 is a splitting field for 𝐵 if 𝐵 ⊗𝐹 𝐾 is split.
Example 5.4.6. The fundamental example of a splitting field for a quaternion algebra
is that C splits the real Hamiltonians H: we have H ⊗R C ' M2 (C) as in (2.4.1).
Lemma 5.4.7. Let 𝐾 ⊃ 𝐹 be a quadratic extension of fields. Then 𝐾 is a splitting

field for 𝐵 if and only if there is an injective 𝐹-algebra homomorphism 𝐾 ↩→ 𝐵.
√
Proof.√First, suppose 𝜄 : 𝐾 ↩→ 𝐵.√We may suppose that 𝐾 = 𝐹 ( 𝑑) with 𝑑 ∈ 𝐹 × . Let
𝜇 = 𝜄( 𝑑), so 𝜇2 = 𝑑. Then 1 ⊗ 𝑑 − 𝜇 ⊗ 1 is a zerodivisor in 𝐵 ⊗𝐹 𝐾:
√ √
(1 ⊗ 𝑑 − 𝜇 ⊗ 1) (1 ⊗ 𝑑 + 𝜇 ⊗ 1) = 1 ⊗ 𝑑 − 𝑑 ⊗ 1 = 0.
By Main Theorem 5.4.4, we conclude that 𝐵 ⊗𝐹 𝐾 ' M2 (𝐾).

Next we prove the converse. If 𝐵 ' M2 (𝐹) already, then any quadratic field 𝐾
embeds in 𝐵 (take a matrix in rational normal form) and 𝐵 ⊗𝐹 𝐾 ' M2 (𝐾) for
√ any 𝐾.
So by Main Theorem 5.4.4, we may suppose 𝐵 is a division ring. Let 𝐾 = 𝐹 ( 𝑑). We
have 𝐵 ⊗𝐹 𝐾 ' M2 (𝐾) if and only if h−𝑎, −𝑏, 𝑎𝑏i is isotropic over 𝐾, which is to say
there exist 𝑥, 𝑦, 𝑧, 𝑢, 𝑣, 𝑤 ∈ 𝐹 such that
√ √ √
−𝑎(𝑥 + 𝑢 𝑑) 2 − 𝑏(𝑦 + 𝑣 𝑑) 2 + 𝑎𝑏(𝑧 + 𝑤 𝑑) 2 = 0. (5.4.8)
Let 𝛼 = 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 and 𝛽 = 𝑢𝑖 + 𝑣 𝑗 + 𝑤𝑖 𝑗. Then trd(𝛼) = trd(𝛽) = 0. Expansion

of (5.4.8) (Exercise 5.14) shows that 𝛼 is orthogonal to 𝛽, so trd(𝛼𝛽) = 0, and that
nrd(𝛼) + 𝑑 nrd(𝛽) = 0. Since 𝐵 is a division ring, if nrd(𝛽) = 𝑐 = 0 then 𝛽 = 0
so nrd(𝛼) = 0 as well and 𝛼 = 0, a contradiction. So nrd(𝛽) ≠ 0, and the element
𝛾 = 𝛼𝛽−1 = 𝑐−1 𝛼𝛽 ∈ 𝐵 has nrd(𝛾) = −𝑑 and trd(𝛾) = 𝑐−1 trd(𝛼𝛽) = 0 so 𝛾 2 = 𝑑 as
desired.

𝑎, 𝑏 1, 𝑏
Example 5.4.9. If 𝐵 = , then either 𝑎 ∈ 𝐹 ×2 and 𝐵 ' ' M2 (𝐹) is split,
𝐹 𝐹
√
or 𝑎 ∉ 𝐹 ×2 and 𝐾 = 𝐹 ( 𝑎) splits 𝐵.
Example 5.4.10. Let 𝑝 be an odd prime and let 𝑎 be a quadratic nonresidue modulo
𝑎, 𝑝
𝑝. We claim that is a division quaternion algebra over Q. By Main Theorem
Q
5.4.4, it suffices to show that the quadratic form h1, −𝑎, −𝑝, 𝑎 𝑝i is anisotropic. So
suppose that 𝑡 2 − 𝑎𝑥 2 = 𝑝(𝑦 2 − 𝑎𝑧2 ) with 𝑡, 𝑥, 𝑦, 𝑧 ∈ Q not all zero. The equation is
homogeneous, so we can multiply through by a common denominator and suppose
that 𝑡, 𝑥, 𝑦, 𝑧 ∈ Z with gcd(𝑡, 𝑥, 𝑦, 𝑧) = 1. Reducing modulo 𝑝 we find 𝑡 2 ≡ 𝑎𝑥 2
(mod 𝑝); since 𝑎 is a quadratic nonresidue, we must have 𝑡 ≡ 𝑥 ≡ 0 (mod 𝑝).
Plugging back in and cancelling a factor of 𝑝 we find 𝑦 2 ≡ 𝑎𝑧 2 ≡ 0 (mod 𝑝), and again
𝑦 ≡ 𝑧 ≡ 0 (mod 𝑝), a contradiction.
5.5 Conics, embeddings

Following Main Theorem 5.2.5, we are led to consider the zero locus of the quadratic
form nrd | 𝐵0 up to scaling; this gives a geometric way to view the preceding results.
Definition 5.5.1. A conic 𝐶 ⊂ P2 over 𝐹 is a nonsingular projective plane curve of

degree 2. An isomorphism of conics 𝐶, 𝐶 0 over 𝐹 is an element 𝑓 ∈ PGL3 (𝐹) =
∼ 𝐶 0.
Aut(P2 ) (𝐹) that induces an isomorphism of curves 𝑓 : 𝐶 −→
If we identify
P(𝐵0 ) := (𝐵0 r {0})/𝐹 × ' P2 (𝐹)
with (the points of) the projective plane over 𝐹, then the vanishing locus 𝐶 = 𝑉 (𝑄) of
𝑄 = nrd | 𝐵0 defines a conic over 𝐹: if we take the basis 𝑖, 𝑗, 𝑖 𝑗 for 𝐵0 , then the conic
𝐶 is defined by the vanishing of the equation
𝑄(𝑥, 𝑦, 𝑧) = nrd(𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗) = −𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧2 = 0.
Here, nondegeneracy of the quadratic form is equivalent to the nonsingularity of the

associated plane curve (Exercise 4.13).
The following corollary is then simply a rephrasing of Main Theorem 5.2.5.
5.6. ORIENTATIONS 75
Corollary 5.5.2. The map 𝐵 ↦→ 𝐶 = 𝑉 (nrd | 𝐵0 ) yields a bĳection

Quaternion algebras over 𝐹 Conics over 𝐹
↔
up to isomorphism up to isomorphism
that is functorial with respect to 𝐹.
Main Theorem 5.4.4 also extends to this context.
Theorem 5.5.3. The following are equivalent:
(i) 𝐵 ' M2 (𝐹);

(vii) The conic 𝐶 associated to 𝐵 has an 𝐹-rational point.
By Lemma 5.4.7, a quadratic field 𝐾 over 𝐹 embeds in 𝐵 if and only if the ternary
quadratic form nrd | 𝐵0 represents 0 over 𝐾. We can also rephrase this in terms of the
values represented by nrd | 𝐵0 .
Lemma 5.5.4. Let 𝐾 be a quadratic extension of 𝐹 of discriminant 𝑑. Then 𝐾 ↩→ 𝐵

if and only if nrd | 𝐵0 represents −𝑑 over 𝐹.
√
Proof. Write 𝐾 = 𝐹 ( 𝑑). Then 𝐾 ↩→ 𝐵 if and only if there exists 𝛼 ∈ 𝐵 such
that 𝛼2 = 𝑑 if and only if there exists 𝛼 ∈ 𝐵 with trd(𝛼) = 0 and nrd(𝛼) = −𝑑, as
claimed.
Remark 5.5.5. Two conics over 𝐹 are isomorphic (as plane curves) if and only if their
function fields are isomorphic (Exercise 5.23).
5.6 Orientations
To conclude, we show that the notion of orientation underlying the definition of special
isometries (as in Example 4.5.5) extends more generally to isometries between two
different quadratic spaces by keeping track of one bit of extra information, refining
Main Theorem 5.2.5. We follow Knus–Murkurjev–Rost–Tignol [KMRT98, Theorem
15.2]. We retain our hypothesis that char 𝐹 ≠ 2.
Let 𝑄 : 𝑉 → 𝐹 be a quadratic space with dim𝐹 𝑉 = 𝑛 odd.
Lemma 5.6.1. Suppose 𝑄 has signed discriminant sgndisc 𝑄 = 𝑑 ∈ 𝐹 × /𝐹 ×2 . Let

𝐴 := Clf 𝑄 be the Clifford algebra of 𝑄, and let 𝐾 = 𝑍 ( 𝐴) be the center of 𝐴. Then
𝐾 ' 𝐹 [𝑥]/(𝑥 2 − 𝑑).
The signed discriminant gives a simpler statement; one could equally well work
with the usual discriminant and keep track of the sign.
Proof. We do the case 𝑛 = 3. We may suppose 𝑉 ' 𝐹 3 with standard basis 𝑒 1 , 𝑒 2 , 𝑒 3

and that 𝑄 ' h𝑎, 𝑏, 𝑐i is diagonal, with sgndisc(𝑄) = −𝑎𝑏𝑐 = 𝑑. We have the relation
𝑒 𝑖 𝑒 𝑗 = −𝑒 𝑗 𝑒 𝑖 for 𝑖 ≠ 𝑗; for all 𝑖 = 1, 2, 3, conjugation by 𝑒 𝑖 acts by −1 on 𝑒 𝑗 and 𝑒 𝑖 𝑒 𝑗
for 𝑗 ≠ 𝑖. This implies 𝑍 ( 𝐴) ⊆ 𝐹 + 𝐹𝑒 1 𝑒 2 𝑒 3 . Let 𝛿 := 𝑒 1 𝑒 2 𝑒 3 = 𝑒 2 𝑒 3 𝑒 1 = 𝑒 3 𝑒 1 𝑒 2 ;

then 𝛿𝑒 𝑖 = 𝑒 𝑖 𝛿 for 𝑖 = 1, 2, 3, so 𝑍 ( 𝐴) = 𝐹 [𝛿]. We compute
𝛿2 = (𝑒 1 𝑒 2 𝑒 3 ) (𝑒 1 𝑒 2 𝑒 3 ) = 𝑒 21 (𝑒 2 𝑒 3 ) (𝑒 2 𝑒 3 ) = −𝑎𝑏𝑐 = sgndisc(𝑄) = 𝑑. (5.6.2)
Therefore 𝐾 ' 𝐹 [𝑥]/(𝑥 2 − 𝑑).

The general case is requested in Exercise 5.19: with a basis 𝑒 1 , . . . , 𝑒 𝑛 for 𝑉, the
center is generated over 𝐹 by 𝛿 = 𝑒 1 · · · 𝑒 𝑛 .
From now on, suppose sgndisc(𝑄) = 𝑑 = 1.

Definition 5.6.3. An orientation of 𝑄 is a choice of 𝜁 ∈ 𝑍 (Clf 𝑄) r 𝐹 with 𝜁 2 = 1.
5.6.4. 𝑄 has exactly two choices of orientation 𝜁, differing by sign, by Lemma 5.6.1:
under an isomorphism 𝑍 (Clf 𝑄) ' 𝐹 × 𝐹, the two orientations are (−1, 1) and (1, −1).
More intrinsically, given an orientation 𝜁, we have a projection 𝐾 → 𝐾/(𝜁 − 1) ' 𝐹,
and conversely given a projection 𝜋 : 𝐾 → 𝐹, there is a unique orientation 𝜁 with
𝜋(𝜁) = 1 (the other maps to −1, by 𝐹-linearity).
Definition 5.6.5. Let 𝜁, 𝜁 0 be orientations on 𝑄, 𝑄 0. An isometry 𝑓 : 𝑉 → 𝑉 0 is
oriented (with respect to 𝜁, 𝜁 0) if in the induced map 𝑓 : 𝑍 (Clf 𝑄) → 𝑍 (Clf 𝑄 0) we
have 𝑓 (𝜁) = 𝜁 0.
5.6.6. An oriented isometry is the same as a special isometry (Definition 4.5.1) when
𝑉 ' 𝐹 𝑛 (𝑛 still odd), as follows. Let 𝐴 = Clf 𝑄. Let 𝑒 1 , . . . , 𝑒 𝑛 be a basis for 𝑉 adapted
as in the proof of Lemma 5.6.1 and 𝛿 = 𝑒 1 . . . 𝑒 𝑛 . Then 𝑍 ( 𝐴) is generated by 𝛿 and
𝛿2 = 1. If 𝑓 ∈ O(𝑄) (𝐹), then 𝑓 (𝛿) = (det 𝑓 )𝛿, so 𝜁 = ±𝛿 is preserved if and only if
det( 𝑓 ) = 1, and this is independent of the choice of orientation.
So we define the oriented or special orthogonal group of a quadratic space by
choosing an orientation and letting
SO(𝑄) (𝐹) := { 𝑓 ∈ O(𝑄) (𝐹) : 𝑓 is oriented};
the resulting group is independent of the choice, and we recover the same group as in
Definition 4.5.1.

𝑎, 𝑏
5.6.7. Let 𝐵 = be a quaternion algebra over 𝐹. In previous sections, we took
𝐹
nrd | 𝐵0 : 𝐵0 → 𝐹, a nondegenerate ternary quadratic space of discriminant 1. Since
we are working with the signed discriminant, we take instead − nrd | 𝐵0 : 𝐵0 → 𝐹 with
sgndisc(− nrd | 𝐵0 ) = 1; this map has a nice description as the squaring map, since
𝛼2 = − nrd(𝛼) for 𝛼 ∈ 𝐵0 .
We claim that 𝐵0 has a canonical orientation. We have an inclusion 𝜄 : 𝐵0 ↩→ 𝐵
with 𝜄(𝑥) 2 = − nrd(𝑥) for all 𝑥 ∈ 𝐵0 . By the universal property of Clifford algebras,
we get an 𝐹-algebra homomorphism 𝜙 : Clf (𝐵0 ) → 𝐵. We see that 𝜙 is surjective
so it induces an 𝐹-algebra map 𝜋 : 𝑍 (Clf(𝐵0 )) → 𝑍 (𝐵) = 𝐹 (Exercise 2.8). This
defines a unique orientation 𝜁 𝐵 = 𝜁 with 𝜁 − 1 ∈ ker 𝜋, by 5.6.4.
Explicitly, let 𝑖, 𝑗, 𝑘 be the standard basis for 𝐵 with 𝑘 = 𝑖 𝑗. Then nrd(𝑘) = 𝑎𝑏,
and 𝑖, 𝑗, 𝑘 is a basis for 𝐵0 . Let 𝜁 = 𝑖 𝑗 𝑘 −1 = −𝑖 𝑗 𝑘/(𝑎𝑏) ∈ 𝑍 (Clf(𝐵0 )). Then
𝛿2 = −𝑎𝑏(−𝑎𝑏)/(𝑎𝑏) 2 = 1 as in (5.6.2). Multiplying out in 𝐵, we get 𝜙(𝜁) = 1 ∈ 𝐵,

so 𝜁 is the same orientation as in the previous paragraph.
The following theorem then refines Main Theorem 5.2.5.
Theorem 5.6.8. Let 𝐹 be a field with char 𝐹 ≠ 2. Then the functors
(𝑄, 𝜁) ↦→ Clf 0 𝑄
(− nrd | 𝐵0 , 𝜁 𝐵 ) ← 𝐵
yield an equivalence of categories between
Oriented ternary quadratic forms over 𝐹 with signed discriminant 1 ∈ 𝐹 × /𝐹 ×2 ,

under oriented isometries.
and
Quaternion algebras over 𝐹, under 𝐹-algebra isomorphisms.
Proof. Let 𝐵 be a quaternion algebra. As in 5.6.7, the inclusion 𝜄 : 𝐵0 ↩→ 𝐵 gives

an 𝐹-algebra homomorphism Clf(− nrd | 𝐵0 ) → 𝐵 which restricts to a canonical 𝐹-
algebra homomorphism Clf 0 (− nrd | 𝐵0 ) → 𝐵. In fact, in coordinates, this map is
the isomorphism (5.3.24): choosing the standard basis 𝑖, 𝑗, 𝑘 for 𝐵 = (𝑎, 𝑏 | 𝐹), and
letting 𝑒 1 = 𝑖, 𝑒 2 = 𝑗, 𝑒 3 = 𝑘, we have

Clf 0 (− nrd | 𝐵0 ) = Clf 0 (h𝑎, 𝑏, −𝑎𝑏i) =
𝐹
with the standard generators 𝑖 0 := 𝑒 1 𝑒 2 = 𝑖 𝑗 and 𝑗0 := 𝑒 2 𝑒 3 = 𝑗 𝑘. We define the

isomorphism

𝑎, 𝑏
→
𝐹 𝐹
𝑖 0 , 𝑗0 ↦→ 𝑖 𝑗, 𝑗 𝑘.
∼ 𝐵 yields a natural isomor-
Therefore, the canonical isomorphism Clf 0 (− nrd | 𝐵0 ) −→
phism between these composed functors and the identity functor, giving an equivalence
of categories.
Conversely, let (𝑄, 𝜁) be an oriented ternary quadratic space, let 𝐵 = Clf 0 𝑄, and
consider (− nrd | 𝐵0 , 𝜁 𝐵 ). We define a natural oriented isometry between these two
spaces. We have a natural inclusion 𝑉 ↩→ Clf 𝑄, and we define the linear map
𝑚𝜁 : 𝑉 → 𝐵
𝑣 ↦→ 𝑣𝜁;
since 𝑣, 𝜁 ∈ Clf 1 𝑄, we have 𝑣𝜁 ∈ Clf 0 𝑄 = 𝐵. We now show that 𝑚 𝜁 induces an

oriented isometry 𝑚 𝜁 : 𝑉 → 𝐵0 . To do so, we let 𝑉 ' 𝐹 3 by choosing an orthogonal
basis 𝑒 1 , 𝑒 2 , 𝑒 3 in which 𝑄 ' h𝑎, 𝑏, 𝑐i and −𝑎𝑏𝑐 = 1. We identify 𝐵 ' (−𝑎𝑏, −𝑏𝑐 | 𝐹)
as in 5.3.23, with 𝑖 = 𝑒 1 𝑒 2 and 𝑗 = 𝑒 2 𝑒 3 , and we let 𝑘 = 𝑖 𝑗 = −𝑏𝑒 3 𝑒 1 so 𝑘 2 =

𝑏 2 (−𝑎𝑐) = 𝑏. Then 𝜁 = 𝜖 𝑒 1 𝑒 2 𝑒 3 with 𝜖 = ±1, and
𝜖𝑚 𝜁 (𝑒 1 ) = 𝑒 1 (𝑒 1 𝑒 2 𝑒 3 ) = 𝑎𝑒 2 𝑒 3 = 𝑎 𝑗
𝜖𝑚 𝜁 (𝑒 2 ) = 𝑒 2 (𝑒 1 𝑒 2 𝑒 3 ) = −𝑏𝑒 1 𝑒 3 = −𝑘 (5.6.9)
𝜖𝑚 𝜁 (𝑒 3 ) = 𝑒 3 (𝑒 1 𝑒 2 𝑒 3 ) = 𝑐𝑖;
so in particular 𝑚 𝜁 (𝑉) ⊆ 𝐵0 . The map is an isometry, because
− nrd(𝑚 𝜁 (𝑣)) = − nrd(𝑣𝜁) = (𝑣𝜁) 2 = 𝑣 2 = − nrd(𝑣) (5.6.10)
since 𝜁 2 = 1 and 𝜁 is central. Finally, the map is oriented:
𝑚 𝜁 (𝜁) = 𝑚 𝜁 (𝜖 𝑒 1 𝑒 2 𝑒 3 ) = 𝜖 (𝑒 1 𝜁) (𝑒 2 𝜁) (𝑒 3 𝜁)
= 𝜖 (𝜖 𝑎 𝑗) (−𝜖 𝑘) (𝜖 𝑐𝑖) = (−𝑎𝑐) (𝑖 𝑗 𝑘) = (−𝑎𝑏𝑐)𝑖 𝑗 𝑘 −1 = 𝑖 𝑗 𝑘 −1 = 𝜁 𝐵 .
This natural oriented isometry gives a natural transformation between these composed
functors and the identity functor, and the statement follows.
Remark 5.6.11. Theorem 5.6.8 can be seen as a manifestation of the isomorphism of

Dynkin diagrams 𝐴1 ' 𝐵1 (consisting of a single node •), corresponding to the iso-
morphism of Lie algebras sl2 ' so3 . This is just one of the (finitely many) exceptional
isomorphisms—the others are just as beautiful, with deep implications, and the reader
is encouraged to read the bible by Knus–Merkurjev–Rost–Tignol [KMRT98, §15].
We record the following important consequence.
Corollary 5.6.12. We have Aut(𝐵) ' 𝐵× /𝐹 × .
Proof. We take stabilizers of objects on both sides of the equivalence of categories in

Theorem 5.6.8; we find Aut(𝐵) ' SO(𝑄) (𝐹) if 𝐵 corresponds to 𝑄. But by Proposition
4.5.10, there is an isomorphism 𝐵× /𝐹 × ' SO(𝑄) (𝐹), and the result follows.
Remark 5.6.13. We will return to Corollary 5.6.12 in the Skolem–Noether theorem in

section 7.7, generalizing to the context of embeddings into a simple algebra.
To conclude, we extend the notion of oriented isometry to similarities.
5.6.14. Let 𝜁, 𝜁 0 be orientations on quadratic spaces 𝑉, 𝑉 0 and suppose dim 𝑉 =

dim 𝑉 0 =Ó𝑛 = 2𝑚 Then a similarity ( 𝑓 , 𝑢) from 𝑉 to 𝑉 0 induces an 𝐹-linear
Ó𝑛is even. Ó
−𝑚 (𝑉) → 𝑛 (𝑉 0), and we say ( 𝑓 , 𝑢) is oriented if the map 𝑢 −𝑚 𝑛 𝑓
Ó
map 𝑢 𝑛 𝑓:
preserves orientations. We define
GSO(𝑄) (𝐹) := {( 𝑓 , 𝑢) ∈ GO(𝑄) (𝐹) : ( 𝑓 , 𝑢) is oriented}
and recover the same group as in 4.5.4. If 𝑛 is odd, we declare that every similarity is
oriented and let GSO(𝑄) (𝐹) := GO(𝑄) (𝐹).
Exercises
Throughout, let 𝐹 be a field with char 𝐹 ≠ 2.
1. Let 𝐵, 𝐵 0 be quaternion algebras over 𝐹. Show that if the quadratic forms nrd 𝐵
and nrd 𝐵0 are similar, then they are isometric.
2. Consider the hyperbolic quaternions 𝐻Mac of Macfarlane (1.2.1).
(a) Show that 𝐻Mac is the Clifford algebra of h1, 1, 1i over R.√
(b) Show that 𝐻Mac is isomorphic as an algebra over√C = R( −1) to the even
Clifford algebra of the ternary quadratic form − −1h1, 1, 1i.
3. Prove the implication (vi) ⇒ (v) of Main Theorem 5.4.4 directly.
4. Use Main Theorem 5.4.4(vi) to give another proof that there is no division
quaternion algebra 𝐵 over a finite field 𝐹 = F𝑞 (with 𝑞 odd).
5. (a) Show that the quadratic form 𝑄(𝑥, 𝑦, 𝑧) = 𝑥 2 + 𝑦 2 + 𝑧2 is isotropic over
F 𝑝 for all odd primes 𝑝. Conclude that (−1, −1 | F 𝑝 ) ' M2 (F 𝑝 ). [Hint:
count squares and nonsquares.]
(b) More generally, show that every ternary quadratic form over a finite field
F𝑞 (with 𝑞 odd) is isotropic. [Hint: Reduce to the case of finding a solution
to 𝑦 2 = 𝑓 (𝑥) where 𝑓 is a polynomial of degree 2.] Use Main Theorem
5.4.4(iv) to give yet another proof that there is no division quaternion
algebra 𝐵 over F𝑞 .
(c) Show that over a finite field F𝑞 with 𝑞 odd, there is a unique anisotropic
binary quadratic form up to isometry.
6. Show that if 𝑄 : 𝐹 → 𝐹 is the quadratic form 𝑄(𝑥) = 𝑎𝑥 2 with 𝑎 ∈ 𝐹, then
Clf (𝐹) ' 𝐹 [𝑥]/(𝑥 2 − 𝑎).

−1, 26
7. Show that (−1, 26)Q = 1, i.e., ' M2 (Q).
Q

−1, 𝑝
8. Let 𝑝 be prime. Show that ' M2 (Q) if and only if 𝑝 = 2 or 𝑝 ≡ 1
Q
(mod 4).
9. Show that

−2, −3 −1, −1 −2, −5 −1, −1
' but that ; .
Q Q Q Q
10. Let 𝐵 = (𝑎, 𝑏 | 𝐹) be a quaternion algebra over 𝐹. Give a constructive
(algorithmic) proof of the implication (iv) ⇒ (i) in Main Theorem 5.4.4, as
follows.
Let 𝜖 = 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 ∈ 𝐵 satisfy nrd(𝜖) = −𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧2 = −𝜖 2 = 0.
(a) Show that there exists 𝑘 ∈ {𝑖, 𝑗, 𝑖 𝑗 } such that trd(𝜖 𝑘) = 𝑠 ≠ 0.
(b) Let 𝑡 := trd(𝑘) and 𝑛 := nrd(𝑘), and let 𝜖 0 := 𝑠−1 𝜖. Let
𝑖 0 := 𝜖 0 𝑘 − (𝑘 + 𝑡)𝜖 0
𝑗 0 := 𝑘 + (−𝑡𝑘 + 𝑛 + 1)𝜖 0 .
Show that 𝑖 0, 𝑗 0 generate 𝐵 as an 𝐹-algebra, and that (𝑖 0) 2 = ( 𝑗 0) 2 = 1 and

𝑗 0𝑖 0 = −𝑖 0 𝑗 0. Conclude that 𝐵 ' M2 (𝐹).
(c) Show that 𝐼 := 𝐹𝜖 0 + 𝐹 𝑘𝜖 0 is a left ideal of 𝐵 with dim𝐹 𝐼 = 2, and interpret
(b) as arising from the left multiplication map 𝐵 → End𝐹 (𝐼) ' M2 (𝐹).
11. Let 𝐵 be a quaternion algebra over 𝐹. Let 𝑄 be the reduced norm on 𝐵, and for
clarity write 𝑒 0 = 1, 𝑒 1 = 𝑖, 𝑒 2 = 𝑗, 𝑒 3 = 𝑘 as a basis for the domain of 𝑄.
(a) Let 𝐶 0 = Clf 0 𝑄 be the even Clifford algebra of the reduced norm 𝑄. Show
that 𝑍 (𝐶 0 ) ' 𝐹 × 𝐹. [Hint: 𝑍 (𝐶 0 ) is generated by 𝑒 0 𝑒 1 𝑒 2 𝑒 3 .]
(b) Show that 𝐶 0 ' 𝐵 × 𝐵op (' 𝐵 × 𝐵) as 𝐹-algebras.
(c) Prove that if 𝐵 0 is a quaternion algebra over 𝐹 then 𝐵 ' 𝐵 0 are isomorphic
as 𝐹-algebras if and only if the reduced norms 𝑄 ∼ 𝑄 0 are similar as
quadratic spaces.
12. Let 𝑄 : 𝑉 → 𝐹 be a nondegenerate quadratic form. Show that the reversal map
: Clf 0 𝑄 → Clf 0 𝑄 on the Clifford algebra has the property that 𝑥𝑥 ∈ 𝐹 for all
pure tensors 𝑥 = 𝑒 1 𝑒 2 · · · 𝑒 𝑑 , but defines a standard involution on Clf 𝑄 if and
only if 𝑉 = {0} and on Clf 0 𝑄 if and only if dim𝐹 𝑉 ≤ 3.
13. Give another proof of Lemma 5.3.17 using the universal property of the Clifford
algebra.
I 14. Expand (5.4.8) and prove as a consequence that if 𝛼 = 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 and
𝛽 = 𝑢𝑖 + 𝑣 𝑗 + 𝑤𝑖 𝑗, then trd(𝛼𝛽) = 0 (so 𝛼 is orthogonal to 𝛽) and moreover
nrd(𝛼) + 𝑑 nrd(𝛽) = 0.
15. Show that the Hilbert symbol is Galois equivariant, in the following sense: for all
field automorphisms 𝜎 ∈ Aut(𝐹) and all 𝑎, 𝑏 ∈ 𝐹 × , we have (𝜎(𝑎), 𝜎(𝑏)) 𝐹 =
(𝑎, 𝑏) 𝐹 .
16. Let 𝑎, 𝑏, 𝑏 0 ∈ 𝐹 × . Show that there exists an 𝐹-linear isomorphism 𝜙 : (𝑎, 𝑏 |
∼ (𝑎, 𝑏 0 | 𝐹) with 𝜙(𝑖) = 𝑖 0 if and only if 𝑏/𝑏 0 ∈ Nm
𝐹) −→ ×
√ 𝐾 |𝐹 (𝐾 ) where
𝐾 = 𝐹 ( 𝑎). [More generally, see Corollary 7.7.6.]
17. Let 𝑎 ∈ Q× r Q×2 . Show that there are infinitely many distinct isomorphism
classes of conics 𝑥 2 − 𝑎𝑦 2 = 𝑏𝑧 2 for 𝑏 ∈ Q× .
18. Let 𝐾 = 𝐹 (𝑎, 𝑏) with 𝑎, 𝑏 algebraically independent,
transcendental elements.
𝑎, 𝑏
Show that the generic quaternion algebra is a division algebra. [Hint:
𝐾
show the associated ternary quadratic form is anisotropic.]
I 19. Prove Lemma 5.6.1 for general odd 𝑛 as follows.
(a) For a subset 𝐼 = {𝑖1 , . . . , 𝑖𝑟 } ⊆ {1, . . . , 𝑛}, let 𝑒 𝐼 = 𝑒 𝑖1 · · · 𝑒 𝑖𝑟 with 𝑖 1 <

· · · < 𝑖𝑟 . Then for subsets 𝐼, 𝐽 ⊆ {1, . . . , 𝑛}, show that
𝑒 𝐼 𝑒 𝐽 = 𝑒 𝐽 𝑒 𝐼 (−1) #𝐼 ·#𝐽 −#(𝐼 ∩𝐽 ) .
(b) Show that 𝑍 (Clf 𝑄) = 𝐹 [𝛿] ' 𝐹 [𝑥]/(𝑥 2 − 𝑑) where 𝛿 = 𝑒 1 𝑒 2 . . . 𝑒 𝑛 and

𝑑 = sgndisc(𝑄). [Hint: Argue on bases and choose #𝐽 = 2 with 𝐼 ∩ 𝐽 = 1.]
20. Let 𝑄 : 𝑉 → 𝐹 be a quadratic form. Show that the even Clifford algebra Clf 0 𝑄
with its map 𝜄 : 𝑉 ⊗ 𝑉 → Clf 0 𝑄 has the following universal property: if 𝐴 is an
𝐹-algebra and 𝜄 𝐴 : 𝑉 ⊗ 𝑉 → 𝐴 is an 𝐹-linear map such that
(i) 𝜄 𝐴 (𝑥 ⊗ 𝑥) = 𝑄(𝑥) for all 𝑥 ∈ 𝑉, and
(ii) 𝜄 𝐴 (𝑥 ⊗ 𝑦)𝜄 𝐴 (𝑦 ⊗ 𝑧) = 𝑄(𝑦)𝜄 𝐴 (𝑥 ⊗ 𝑧) for all 𝑥, 𝑦, 𝑧 ∈ 𝑉,
then there exists a unique 𝐹-algebra homomorphism 𝜙 : Clf 0 𝑄 → 𝐴 such that
the diagram
𝜄 /
𝑉 ⊗𝑉 Clf 0 𝑄
𝜙
$
𝜄𝐴
commutes. Conclude that the pair (Clf 0 𝑄, 𝜄) is unique up to unique isomor-

phism.
21. In this exercise, we consider graded tensor products, giving an alternate verifi-
cation of Lemma 5.3.14.
Let 𝐴 = 𝐴0 ⊕ 𝐴1 and 𝐵 = 𝐵0 ⊕ 𝐵1 be finite-dimensional 𝐹-algebras equipped
with a Z/2Z-grading. We define the graded tensor product 𝐴b ⊗ 𝐵 to be the
usual tensor product as an 𝐹-vector space but with multiplication law defined
on simple tensors by
0
(𝑎 ⊗ 𝑏) · (𝑎 0 ⊗ 𝑏 0) = (−1) (deg 𝑎 ) (deg 𝑏) (𝑎𝑎 0 ⊗ 𝑏𝑏 0).
(a) Show that 𝐴b ⊗ 𝐵 is an 𝐹-algebra of dimension (dim𝐹 𝐴) (dim𝐹 𝐵).

(b) Let 𝑄 1 : 𝑉1 → 𝐹 and 𝑄 2 : 𝑉2 → 𝐹, and let 𝑄 := 𝑄 1 𝑄 2 be the orthogonal
direct sum on the quadratic space 𝑉 := 𝑉1 𝑉2 . Show there is a canonical
isomorphism of Clifford algebras
Clf (𝑄) Clf (𝑄 1 ) b

⊗ Clf(𝑄 2 ).
(c) Observe that (b) gives another proof of Lemma 5.3.14.
22. For 𝑖 = 1, 2, let 𝑄 𝑖 : 𝑉𝑖 → 𝐹 be quadratic forms over 𝐹.

(a) Prove that there exists a canonical 𝑅-algebra isomorphism
∼ (Clf 0 (𝑄 ) ⊗ Clf 0 (𝑄 )) ⊕ (Clf 1 (𝑄 ) ⊗ Clf 1 (𝑄 ))
Clf 0 (𝑄 1 𝑄 2 ) −→ 1 2 1 2
where Clf 1 (𝑄 1 ) ⊗ Clf 1 (𝑄 2 ) has multiplication induced from the full Clif-
ford algebras Clf (𝑄 1 ) and Clf (𝑄 2 ).
(b) Prove that there is a Clf 0 (𝑄 1 𝑄 2 )-bimodule isomorphism
∼ (Clf 0 (𝑄 ) ⊗ Clf 1 (𝑄 )) ⊕ (Clf 1 (𝑄 ) ⊗ Clf 0 (𝑄 ))
Clf 1 (𝑄 1 𝑄 2 ) −→ 1 2 1 2
with bimodule structure induced by multiplication in the full Clifford

algebra.
23. In this exercise, we assume background in algebraic curves. Show that two
conics over 𝐹 are isomorphic (as projective plane curves) if and only if their
function fields are isomorphic. [Hint: conics are anticanonically embedded—
the restriction of 𝒪P2 (−1) to the conic is a canonical sheaf—so an isomorphism
of function fields induces an linear isomorphism of conics.]
Chapter 6
Characteristic 2
In this chapter, we extend the results from the previous four chapters to the neglected
case where the base field has characteristic 2. Throughout this chapter, let 𝐹 be a field
with algebraic closure 𝐹 al .
6.1 Separability
To get warmed up, we give a different notation (symbol) for quaternion algebras that
holds in any characteristic and which is convenient for many purposes.
Definition 6.1.1. Let 𝐴 be a commutative, finite-dimensional algebra over 𝐹. We say

𝐴 is separable if
𝐴 ⊗𝐹 𝐹 al ' 𝐹 al × · · · × 𝐹 al ;
otherwise, we say 𝐴 is inseparable.
Example 6.1.2. If 𝐴 ' 𝐹 [𝑥]/( 𝑓 (𝑥)) with 𝑓 (𝑥) ∈ 𝐹 [𝑥], then 𝐴 is separable if and
only if 𝑓 has distinct roots in 𝐹 al .
6.1.3. If char 𝐹 ≠ 2, and 𝐾 is a quadratic 𝐹-algebra, then after completing the square,
we see that the following are equivalent:
(i) 𝐾 is separable;
(ii) 𝐾 ' 𝐹 [𝑥]/(𝑥 2 − 𝑎) with 𝑎 ≠ 0;
(iii) 𝐾 is reduced (𝐾 has no nonzero nilpotent elements);
(iv) 𝐾 is a field or 𝐾 ' 𝐹 × 𝐹.
6.1.4. If char 𝐹 = 2, then a quadratic 𝐹-algebra 𝐾 is separable if and only if
𝐾 ' 𝐹 [𝑥]/(𝑥 2 + 𝑥 + 𝑎)
for some 𝑎 ∈ 𝐹. A quadratic algebra of the form 𝐾 = 𝐹 [𝑥]/(𝑥 2 + 𝑎) with 𝑎 ∈ 𝐹 is

inseparable.
Now we introduce the more general notation.
83
84 CHAPTER 6. CHARACTERISTIC 2
6.1.5. Let 𝐾 be a separable quadratic 𝐹-algebra, and let 𝑏 ∈ 𝐹 × . We denote by

𝐾, 𝑏
:= 𝐾 ⊕ 𝐾 𝑗
𝐹
the 𝐹-algebra with basis 1, 𝑗 as a left 𝐾-vector space and with the multiplication rules
𝑗 2 = 𝑏 and 𝑗 𝛼 = 𝛼 𝑗 for 𝛼 ∈ 𝐾, where is the standard involution on 𝐾 (the nontrivial
element of Gal(𝐾 | 𝐹) if 𝐾 is a field). We will also write (𝐾, 𝑏 | 𝐹) for formatting.
From 6.1.3, if char 𝐹 ≠ 2 then writing 𝐾 ' 𝐹 [𝑥]/(𝑥 2 − 𝑎) we see that

𝐾, 𝑏 𝑎, 𝑏
'
𝐹 𝐹
is a quaternion algebra over 𝐹. The point is that we cannot complete the square in
characteristic 2, so the more general notation gives a characteristic-independent way
to define quaternion algebras. In using this symbol, we are breaking the symmetry
between the standard generators 𝑖, 𝑗, but otherwise have not changed anything about
the definition.

Throughout the rest of this chapter, we suppose that char 𝐹 = 2. (We will occasionally
remind the reader of this supposition, but it is meant to hold throughout.)
Definition 6.2.1. An algebra 𝐵 over 𝐹 (with char 𝐹 = 2) is a quaternion algebra if

there exists an 𝐹-basis 1, 𝑖, 𝑗, 𝑘 for 𝐵 such that
𝑖 2 + 𝑖 = 𝑎, 𝑗 2 = 𝑏, and 𝑘 = 𝑖 𝑗 = 𝑗 (𝑖 + 1) (6.2.2)
with 𝑎 ∈ 𝐹 and 𝑏 ∈ 𝐹 × .
Just as when char 𝐹 ≠ 2, we find that the multiplication table for a quaternion
algebra 𝐵 is determined by the rules (6.2.2), e.g.
𝑗 𝑘 = 𝑗 (𝑖 𝑗) = (𝑖 𝑗 + 𝑗) 𝑗 = 𝑏𝑖 + 𝑏 = 𝑘 𝑗 + 𝑏.

𝑎, 𝑏
We denote by or [𝑎, 𝑏 | 𝐹) the 𝐹-algebra with basis 1, 𝑖, 𝑗, 𝑖 𝑗 subject to the
𝐹
𝑎, 𝑏
multiplication rules (6.2.2). The algebra is not symmetric in 𝑎, 𝑏 (explaining
𝐹
the choice of notation), but it is still functorial in the field 𝐹.
If we let 𝐾 = 𝐹 [𝑖] ' 𝐹 [𝑥]/(𝑥 2 + 𝑥 + 𝑎), then

𝑎, 𝑏 𝐾, 𝑏
'
𝐹 𝐹
and our notation extends that of Section 6.1.

6.3. ∗ QUADRATIC FORMS 85
Example 6.2.3. The ring M2 (𝐹) of 2 × 2-matrices with coefficients in 𝐹 is again a

quaternion algebra over 𝐹, via the isomorphism

1, 1 ∼
−→ M2 (𝐹)
𝐹

0 1 0 1
𝑖, 𝑗 ↦→ , .
1 1 1 0
Lemma 6.2.4. An 𝐹-algebra 𝐵 is a quaternion algebra if and only if there exist

𝐹-algebra generators 𝑖, 𝑗 ∈ 𝐵 satisfying
𝑖 2 + 𝑖 = 𝑎, 𝑗 2 = 𝑏, and 𝑖 𝑗 = 𝑗 (𝑖 + 1). (6.2.5)
Proof. Proven the same way as Lemma 2.2.5.
6.2.6. Let 𝐵 = [𝑎, 𝑏 | 𝐹) be a quaternion algebra over 𝐹. Then 𝐵 has a (unique)

standard involution : 𝐵 → 𝐵 given by
𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 ↦→ 𝛼 = 𝑥 + 𝛼 = (𝑡 + 𝑥) + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗
since
𝛼𝛼 = (𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗) ((𝑡 + 𝑥) + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗)
(6.2.7)
= 𝑡 2 + 𝑡𝑥 + 𝑎𝑥 2 + 𝑏𝑦 2 + 𝑏𝑦𝑧 + 𝑎𝑏𝑧2 ∈ 𝐹.
Consequently, one has a reduced trace and reduced norm on 𝐵 as in Chapter 3.
We now state a version of Theorem 3.5.1 in characteristic 2; the proof is similar

and is left as an exercise.
Theorem 6.2.8. Let 𝐵 be a division 𝐹-algebra with a standard involution that is not
the identity. Then either 𝐵 is a separable quadratic field extension of 𝐹 or 𝐵 is a
quaternion algebra over 𝐹.
Proof. Exercise 6.9. (This theorem is also implied by Theorem 6.4.1.)
6.3 ∗ Quadratic forms

We now turn to the theory of quadratic forms over 𝐹 with char 𝐹 = 2. The basic
definitions from section 4.2 apply. For further reference, Grove [Grov2002, Chapters
12–14] treats quadratic forms in characteristic 2, and the book by Elman–Karpenko–
Merkurjev [EKM2008, Chapters I–II] discusses bilinear forms and quadratic forms in
all characteristics.
Let 𝑄 : 𝑉 → 𝐹 be a quadratic form with dim𝐹 𝑉 = 𝑛 < ∞ and associated bilinear
form 𝑇. Then 𝑇 (𝑥, 𝑥) = 2𝑄(𝑥) = 0 for all 𝑥 ∈ 𝑉, so one cannot recover the quadratic
form from the symmetric (equivalently, alternating) bilinear form.
6.3.1. We begin with the definition of the discriminant. When 𝑛 is even, we sim-
ply define disc(𝑄) = det(𝑇) ∈ 𝐹/𝐹 ×2 —this is equivalent to Definition 4.3.3 when
char 𝐹 ≠ 2, having absorbed the square power of 2.
When 𝑛 is odd, the symmetric matrix 𝑇 always has determinant 0 (Exercise 6.8);
we need to “divide this by 2”. So instead we work with a generic quadratic form, as
follows. Consider the quadratic form
∑︁
𝑄 univ (𝑥1 , . . . , 𝑥 𝑛 ) := 𝑎 𝑖 𝑗 𝑥𝑖 𝑥 𝑗 = 𝑎 11 𝑥12 + 𝑎 12 𝑥1 𝑥 2 + · · · + 𝑎 𝑛𝑛 𝑥 𝑛2 (6.3.2)
1≤𝑖 ≤ 𝑗 ≤𝑛
over the field 𝐹 univ := Q(𝑎 𝑖 𝑗 )𝑖, 𝑗=1,...,𝑛 (now of characteristic zero!) with 𝑎 𝑖 𝑗 transcen-
dental elements. We compute its universal determinant
2𝑎 11 𝑎 12 ··· 𝑎 1𝑛
2𝑎 22 ···
© ª
𝑎 12 𝑎 2𝑛
det( [𝑇 univ ]) = det .
®
.. .. .. ® ∈ 2Z[𝑎 𝑖 𝑗 ] 𝑖, 𝑗 (6.3.3)
.. . ®
. . ®
« 𝑎 1𝑛 𝑎 2𝑛 ··· 2𝑎 𝑛𝑛 ¬
as a polynomial with integer coefficients. We claim all of these coefficients are even:
indeed, reducing modulo 2 and computing the determinant over F2 (𝑎 𝑖 𝑗 )𝑖, 𝑗 , we recall
that the determinant of an alternating matrix of odd size is zero (over any field).
Therefore, we may let
𝛿(𝑎 11 , . . . , 𝑎 𝑛𝑛 ) := det(𝑇 univ )/2 ∈ Z[𝑎 𝑖 𝑗 ] 𝑖, 𝑗 (6.3.4)
be the universal (half-)discriminant. We then define
disc(𝑄) := 𝛿(𝑄(𝑒 1 ), 𝑇 (𝑒 1 , 𝑒 2 ), . . . , 𝑄(𝑒 𝑛 )) ∈ 𝐹/𝐹 ×2
by specialization. Repeating the argument in 4.3.2, if 𝑡𝑖 𝑗 ∈ 𝐹 and 𝑒 𝑖0 :=

Í
𝑗 𝑡𝑖 𝑗 𝑒𝑖 then
𝛿(𝑄(𝑒 10 ), 𝑇 (𝑒 10 , 𝑒 20 ), . . . , 𝑄(𝑒 𝑛0 )) = 𝛿(𝑄(𝑒 1 ), 𝑇 (𝑒 1 , 𝑒 2 ), . . . , 𝑄(𝑒 𝑛 )) det(𝑡𝑖 𝑗 ) 2
(verified universally!) so disc(𝑄) is well-defined. Moreover, this definition agrees

with Definition 4.3.3 when char 𝐹 ≠ 2.
Example 6.3.5. For example, disc(h𝑎i) = 𝑎 for 𝑎 ∈ 𝐹, and if
𝑄(𝑥, 𝑦, 𝑧) = 𝑎𝑥 2 + 𝑏𝑦 2 + 𝑐𝑧2 + 𝑢𝑦𝑧 + 𝑣𝑥𝑧 + 𝑤𝑥𝑦
with 𝑎, 𝑏, 𝑐, 𝑢, 𝑣, 𝑤 ∈ 𝐹, then
disc(𝑄) = 4𝑎𝑏𝑐 + 𝑢𝑣𝑤 − 𝑎𝑢 2 − 𝑏𝑣 2 − 𝑐𝑤 2
in all characteristics.
Definition 6.3.6. We say 𝑄 is nondegenerate if disc(𝑄) ≠ 0.
Next, even though not every quadratic form over 𝐹 can be diagonalized, so we will
also make use of one extra form: for 𝑎, 𝑏 ∈ 𝐹, we write [𝑎, 𝑏] for the quadratic form
𝑎𝑥 2 + 𝑎𝑥𝑦 + 𝑏𝑦 2 on 𝐹 2 .
6.4. ∗ CHARACTERIZING QUATERNION ALGEBRAS 87
Lemma 6.3.7. There exists a basis of 𝑉 such that
𝑄 ' [𝑎 1 , 𝑏 1 ] · · · [𝑎 𝑚 , 𝑏 𝑚 ] h𝑐 1 , . . . , 𝑐𝑟 i (6.3.8)
with 𝑎 𝑖 , 𝑏 𝑖 , 𝑐 𝑗 ∈ 𝐹.
We say that a quadratic form 𝑄 is normalized if 𝑄 is presented with a basis as in

(6.3.8).
Example 6.3.9. The quadratic forms 𝑄(𝑥, 𝑦, 𝑧) = 𝑥 2 +𝑦𝑧+𝑧2 and 𝑄(𝑥, 𝑦, 𝑧) = 𝑥 2 +𝑦 2 +𝑧2
are normalized over F2 , but the quadratic form 𝑄(𝑥, 𝑦, 𝑧) = 𝑥𝑧 + 𝑦𝑧 + 𝑧 2 is not.
Example 6.3.10. For a normalized quadratic form as in (6.3.8),
disc(𝑄) = disc( [𝑎 1 , 𝑏 1 ] · · · [𝑎 𝑚 , 𝑏 𝑚 ]) disc(h𝑐 1 , . . . , 𝑐𝑟 i)

= (𝑎 1 · · · 𝑎 𝑚 ) 2 disc(h𝑐 1 , . . . , 𝑐𝑟 i).
In 𝐹/𝐹 ×2 , we have


 0, if 𝑟 ≥ 2;


disc(h𝑐 1 , . . . , 𝑐𝑟 i) = 𝑐 1 , if 𝑟 = 1;

 1,

if 𝑟 = 0.

Therefore, 𝑄 is nondegenerate if and only if 𝑎 1 · · · 𝑎 𝑚 𝑐 1 · · · 𝑐𝑟 ≠ 0 and 𝑟 ≤ 1.

𝑎, 𝑏
Example 6.3.11. Let 𝐵 = be a quaternion algebra. Then 1, 𝑖, 𝑗, 𝑖 𝑗 is a normal-
𝐹
ized basis for 𝐵, and by (6.2.7),
nrd ' [1, 𝑎] [𝑏, 𝑎𝑏],
so disc(nrd) = 𝑏 2 so nrd is nondegenerate.
6.4 ∗ Characterizing quaternion algebras

We now consider the characterization of quaternion algebras as those equipped with
a nondegenerate standard involution (revisiting Main Theorem 4.4.1, but now with
char 𝐹 = 2).
Theorem 6.4.1. Let 𝐵 be an 𝐹-algebra (with char 𝐹 = 2). Then 𝐵 has a nondegenerate
standard involution if and only if one of the following holds:
(i) 𝐵 = 𝐹;
(ii) 𝐵 = 𝐾 is a separable quadratic 𝐹-algebra; or
Proof. If 𝐵 = 𝐹, then the standard involution is the identity, and nrd is nondegenerate
on 𝐹 because the reduced (half-)discriminant of the quadratic form nrd(𝑥) = 𝑥 2 is 1.
If dim𝐹 𝐵 = 2, then 𝐵 = 𝐾 has a unique standard involution (Lemma 3.4.2). By
6.1.4, we see that the involution is nondegenerate if and only if 𝐾 is separable.
So suppose dim𝐹 𝐵 > 2. Since 𝐵 has a nondegenerate standard involution, there
exists an element 𝑖 ∈ 𝐵 such that 𝑇 (𝑖, 1) = trd(𝑖) ≠ 0. We have 𝑖 ∉ 𝐹 since
trd(𝐹) = {0}. Rescaling we may suppose trd(𝑖) = 1, whence 𝑖 2 = 𝑖 + 𝑎 for some 𝑎 ∈ 𝐹,
and nrd | 𝐹 +𝐹 𝑖 = [1, 𝑎]. (We have started the proof of Lemma 6.3.7, and 1, 𝑖 is part of
a normalized basis, in this special case.)
By nondegeneracy, there exists 𝑗 ∈ {1, 𝑖}⊥ such that nrd( 𝑗) = 𝑏 ≠ 0. Thus
trd( 𝑗) = 0 so 𝑗 = 𝑗 and 𝑗 2 = 𝑏 ∈ 𝐹 × . Furthermore,
0 = trd(𝑖 𝑗) = 𝑖 𝑗 + 𝑗𝑖 = 𝑖 𝑗 + 𝑗 (𝑖 + 1)
so 𝑖 𝑗 = 𝑗 (𝑖 + 1). Therefore 𝑖, 𝑗 generate an 𝐹-subalgebra 𝐴 ' [𝑎, 𝑏 | 𝐹).

The conclusion of the proof follows exactly as in (4.4.3): if 𝑘 ∈ {1, 𝑖, 𝑗, 𝑖 𝑗 }⊥ then
𝑘 (𝑖 𝑗) = 𝑘 ( 𝑗𝑖), a contradiction.
Corollary 6.4.2. Let 𝐵 be a quaternion algebra over 𝐹, and suppose that 𝐾 ⊆ 𝐵 is a

commutative separable 𝐹-subalgebra. Then dim𝐹 𝐾 ≤ 2. Moreover, if 𝐾 ≠ 𝐹, then
the centralizer of 𝐾 × in 𝐵× is again 𝐾 × .
Next, we characterize isomorphism classes of quaternion algebras in characteristic

2 in the language of quadratic forms.
6.4.3. Let 𝐵 be a quaternion algebra over 𝐹. We again define
𝐵0 := {𝛼 ∈ 𝐵 : trd(𝛼) = 0} = {1}⊥ . (6.4.4)
But now 𝐵0 = 𝐹 ⊕ 𝐹 𝑗 ⊕ 𝐹 𝑘 and in this basis
nrd(𝑥 + 𝑦 𝑗 + 𝑧𝑖 𝑗) = 𝑥 2 + 𝑏𝑦 2 + 𝑏𝑦𝑧 + 𝑎𝑏𝑧2 (6.4.5)
so nrd | 𝐵0 ' h1i [𝑏, 𝑎𝑏]. The discriminant is therefore
disc(nrd | 𝐵0 ) = 𝑏 2 = 1 ∈ 𝐹 × /𝐹 ×2 . (6.4.6)
Theorem 6.4.7. Let 𝐹 be a field with char 𝐹 = 2. Then the functor 𝐵 ↦→ nrd | 𝐵0 yields
an equivalence of categories between
Quaternion algebras over 𝐹,

under 𝐹-algebra isomorphisms
and
Ternary quadratic forms over 𝐹 with discriminant 1 ∈ 𝐹 × /𝐹 ×2 ,

under isometries.
Proof. We argue as in Theorem 5.6.8 but with char 𝐹 = 2. The argument here is
easier, because all sign issues go away and there is no orientation to chase: by Exercise
6.12, there is a unique 𝜁 ∈ Clf 1 𝑄 r 𝐹 such that 𝜁 2 = 1. The inclusion 𝜄 : 𝐵0 ↩→ 𝐵
induces a surjective 𝐹-algebra homomorphism Clf 0 (nrd | 𝐵0 ) → 𝐵, so by dimensions
it is an isomorphism; this gives one natural transformation. In the other direction, the
map 𝑚 𝜁 : 𝑉 → 𝐵0 by 𝑣 ↦→ 𝑣𝜁 is again an isometry by (5.6.10), giving the other.
Here is a second direct proof. By 6.4.3, the quadratic form nrd | 𝐵0 has discriminant
1. To show the functor is essentially surjective, let 𝑄 : 𝑉 → 𝐹 be a ternary quadratic
form with discriminant 1 ∈ 𝐹 × /𝐹 ×2 . Then 𝑄 ' h𝑢i [𝑏, 𝑐] for some 𝑢, 𝑏, 𝑐 ∈ 𝐹. We
have disc(𝑄) = 𝑢𝑏 2 = 1 ∈ 𝐹 ×2 so 𝑏 ∈ 𝐹 × and 𝑢 ∈ 𝐹 ×2 . Rescaling the first variable,
we obtain 𝑄 ∼ h1i [𝑏, 𝑐]. Thus by 6.4.3, 𝑄 arises up to isometry from the quaternion
𝑎, 𝑏
algebra with 𝑎 = 𝑐𝑏 −1 .
𝐹
For morphisms, we argue as in the proof of Proposition 5.2.4 but with char 𝐹 = 2. In
one direction, an 𝐹-algebra isomorphism 𝐵 −→ ∼ 𝐵 0 induces an isometry 𝐵0 −→ ∼ (𝐵 0) 0
by uniqueness of the standard involution. Conversely, let 𝑓 : 𝐵 → (𝐵 ) 0 be an
0 0

𝑎, 𝑏
isometry. Let 𝐵 ' . Extend 𝑓 to an 𝐹-linear map 𝐵 → 𝐵 0 by mapping
𝐹
𝑖 ↦→ 𝑏 −1 𝑓 (𝑖 𝑗) 𝑓 ( 𝑗). The map 𝑓 preserves 1: it maps 𝐹 to 𝐹 by Exercise 6.15, since
𝐹 = (𝐵0 ) ⊥ = ((𝐵 0) 0 ) ⊥ , and 1 = nrd(1) = nrd( 𝑓 (1)) = 𝑓 (1) 2 so 𝑓 (1) = 1. We have
𝑓 ( 𝑗) 2 = nrd( 𝑓 ( 𝑗)) = nrd( 𝑗) = 𝑏 and similarly 𝑓 (𝑖 𝑗) 2 = 𝑎𝑏 since 𝑗, 𝑖 𝑗 ∈ 𝐵0 . Thus
1 = trd(𝑖) = 𝑏 −1 trd((𝑖 𝑗) 𝑗) = 𝑏 −1𝑇 (𝑖 𝑗, 𝑗) =
= 𝑏 −1𝑇 ( 𝑓 (𝑖 𝑗), 𝑓 ( 𝑗)) = trd(𝑏 −1 𝑓 (𝑖 𝑗) 𝑓 ( 𝑗)) = trd( 𝑓 (𝑖))
and similarly nrd( 𝑓 (𝑖)) = nrd(𝑖) = 𝑎, thus 𝑓 (𝑖) 2 + 𝑓 (𝑖) + 𝑎 = 0. Finally,
𝑓 (𝑖) 𝑓 ( 𝑗) = 𝑏 −1 𝑓 (𝑖 𝑗) 𝑓 ( 𝑗) 2 = 𝑓 (𝑖 𝑗)
and
𝑓 ( 𝑗) 𝑓 (𝑖) = 𝑏 −1 𝑓 ( 𝑗) 𝑓 (𝑖 𝑗) 𝑓 ( 𝑗) = 𝑏 −1 𝑓 ( 𝑗) ( 𝑓 ( 𝑗) 𝑓 (𝑖 𝑗) + 𝑇 ( 𝑓 ( 𝑗), 𝑓 (𝑖 𝑗)))
= 𝑓 (𝑖 𝑗) + 𝑓 ( 𝑗) = ( 𝑓 (𝑖) + 1) 𝑓 ( 𝑗) = 𝑓 (𝑖) 𝑓 ( 𝑗)
so 𝑓 is an isomorphism of 𝐹-algebras. Therefore the functor is full and faithful,
yielding an equivalence of categories.
Corollary 6.4.8. The maps 𝐵 ↦→ 𝑄 = nrd | 𝐵0 ↦→ 𝐶 = 𝑉 (𝑄) yield bĳections

 Nondegenerate ternary 

Quaternion algebras over 𝐹 quadratic forms over 𝐹


 


↔
up to isomorphism 
 with discriminant 1 ∈ 𝐹 × /𝐹 ×2 

up to isometry
 
 
up to similarity

Conics over 𝐹
↔
up to isomorphism
that are functorial with respect to 𝐹.
Proof. The remaining parts of the bĳection follow as in the proof of Corollary 5.2.6.

We now turn to identifying the matrix ring in characteristic 2.

Definition 6.4.9. A quadratic form 𝐻 : 𝑉 → 𝐹 is a hyperbolic plane if 𝐻 ' [1, 0].
Recall that [1, 0] : 𝐹 2 → 𝐹 is given by the quadratic form 𝑥 2 + 𝑥𝑦 = 𝑥(𝑥 + 𝑦),
visibly isometric to the quadratic form 𝑥𝑦. Definition 6.4.9 agrees with Definition
5.4.1 after a change of basis.
Lemma 6.4.10. If 𝑄 is nondegenerate and isotropic then 𝑄 ' 𝐻 𝑄 0 with 𝐻 a
hyperbolic plane.
Proof. We repeat the proof of Lemma 5.4.2.
We may again characterize division quaternion algebras by examination of the

reduced norm as a quadratic form as in Main Theorem 5.4.4 and Theorem 5.5.3.

𝑎, 𝑏
Theorem 6.4.11. Let 𝐵 = (with char 𝐹 = 2). Then the following are equivalent:
𝐹

1, 1
(i) 𝐵 ' ' M2 (𝐹);
𝐹
(iii) The quadratic form nrd is isotropic;
(iv) The quadratic form nrd | 𝐵0 is isotropic;
(v) The binary form [1, 𝑎] represents 𝑏;
(vi) 𝑏 ∈ Nm𝐾 |𝐹 (𝐾 × ) where 𝐾 = 𝐹 [𝑖]; and
(vii) The conic 𝐶 := 𝑉 (nrd | 𝐵0 ) ⊂ P2 has an 𝐹-rational point.
Proof. Only condition (v) requires significant modification in the case char 𝐹 = 2; see
Exercise 6.13.
Lemma 6.4.12. Let 𝐾 ⊃ 𝐹 be a quadratic extension of fields. Then 𝐾 is a splitting

field for 𝐵 if and only if there is an injective 𝐹-algebra homomorphism 𝐾 ↩→ 𝐵.
Proof. If 𝜄 : 𝐾 ↩→ 𝐵 and 𝐾 = 𝐹 (𝛼), then 1 ⊗ 𝛼 − 𝜄(𝛼) ⊗ 1 is a zerodivisor in 𝐵 ⊗𝐹 𝐾,

since
(1 ⊗ 𝛼 − 𝜄(𝛼) ⊗ 1) (1 ⊗ 𝛼 − 𝜄(𝛼) ⊗ 1) = 0, (6.4.13)
and so 𝐵 ⊗𝐹 𝐾 ' M2 (𝐾) and 𝐾 is a splitting field.
Conversely, let 𝐾 = 𝐹 (𝛼) and suppose 𝐵 ⊗𝐹 𝐾 ' M2 (𝐾). If 𝐵 ' M2 (𝐹), we can
take the embedding mapping
𝛼 to a matrix with the same rational canonical form. So
𝑎, 𝑏
we suppose that 𝐵 = is a division ring. By Theorem 6.4.11(iv) (over 𝐾) and
𝐹
6.4.3, there exist 𝑥, 𝑦, 𝑧, 𝑢, 𝑣, 𝑤 ∈ 𝐹 not all zero such that
(𝑥 + 𝑢𝛼) 2 + 𝑏(𝑦 + 𝑣𝛼) 2 + 𝑏(𝑦 + 𝑣𝛼) (𝑧 + 𝑤𝛼) + 𝑎𝑏(𝑧 + 𝑤𝛼) 2 = 0; (6.4.14)

expanding and rewriting into the powers of 𝛼 gives
(𝑢 2 + 𝑏𝑣 2 + 𝑏𝑣𝑤 + 𝑎𝑏𝑤 2 )𝛼2 + (𝑣𝑧 + 𝑤𝑦)𝑏𝛼 + (𝑥 2 + 𝑏𝑦 2 + 𝑏𝑦𝑧 + 𝑎𝑏𝑧2 ) = 0. (6.4.15)
Let 𝛽 := 𝑥 + 𝑦 𝑗 + 𝑧𝑖 𝑗 and 𝛾 := 𝑢 + 𝑣 𝑗 + 𝑤𝑖 𝑗. Then 𝛾 ∈ 𝐵× , since 𝛾 = 0 implies

nrd(𝛽) = 0 and yet 𝐵 is a division ring. Then the equation (6.4.15) can be written
nrd(𝛾)𝛼2 + trd(𝛽𝛾)𝛼 + nrd(𝛽) = 0;
then a direct calculation shows that the element
𝜇 = 𝛽𝛾 −1 = nrd(𝛾) −1 𝛽𝛾
satisfies the same equation as (6.4.15) in the variable 𝛼, so there is an embedding

𝐾 ↩→ 𝐵 defined by 𝛼 ↦→ 𝜇.
Exercises
Throughout these exercises, we let 𝐹 be a field (of any characteristic, unless specified).
1. Recall the primitive element theorem from Galois theory: if 𝐾 ⊇ 𝐹 is a separable

field extension of finite degree, then there exists 𝛼 ∈ 𝐾 such that 𝐾 = 𝐹 (𝛼)—and
hence 𝐾 ' 𝐹 [𝑥]/( 𝑓 (𝑥)) where 𝑓 (𝑥) ∈ 𝐹 [𝑥] is the minimal polynomial of 𝛼.
Extend this theorem to algebras as follows. Let 𝐵 be a separable, commutative,
finite-dimensional 𝐹-algebra. Show that 𝐵 ' 𝐹 [𝑥]/( 𝑓 (𝑥)) for some 𝑓 (𝑥) ∈
𝐹 [𝑥].
I 2. Let 𝐵 be a quaternion algebra over 𝐹 and let 𝐾 ⊂ 𝐵 be a separable
quadratic
× 𝐾, 𝑏
𝐹-algebra. Show that there exists 𝑏 ∈ 𝐹 such that 𝐵 ' (as in 6.1.5).
𝐹
3. Let 𝐹 sep be a separable closure of 𝐹 and let 𝐵 be a quaternion algebra over 𝐹.
Show that 𝐵 ⊗𝐹 𝐹 sep ' M2 (𝐹 sep ). [More generally, see Exercise 7.23.]

𝐾, 𝑏
I 4. Let 𝐾 be a separable quadratic 𝐹-algebra and let 𝑢, 𝑏 ∈ 𝐹 × . Show that '
𝐹
𝐾, 𝑢𝑏
if and only if 𝑢 ∈ nrd(𝐾 × ) = Nm𝐾 |𝐹 (𝐾 × ).
𝐹
5. Let 𝐵 be a quaternion algebra over 𝐹, and let 𝐾0 ⊇ 𝐹 be a quadratic field. Prove
that there exists a separable extension 𝐾 ⊇ 𝐹 linearly disjoint from 𝐾0 over 𝐹
(i.e., 𝐾 ⊗𝐹 𝐾0 is a domain) such that 𝐾 splits 𝐹.
6. Suppose char 𝐹 = 2 and let 𝑎 ∈ 𝐹 and 𝑏 ∈ 𝐹 × .

𝑎, 𝑏 𝑎, 𝑎𝑏
(a) Show that ' if 𝑎 ≠ 0.
𝐹 𝐹
𝑎 + (𝑡 + 𝑡 2 ), 𝑏𝑢 2

𝑎, 𝑏
(b) Show that if 𝑡 ∈ 𝐹 and 𝑢 ∈ 𝐹 × , then ' .
𝐹 𝐹

𝑎, 𝑏
7. Let char 𝐹 = 2 and let 𝐵 = be a quaternion algebra over 𝐹. Compute the
𝐹
left regular representation 𝜆 : 𝐵 → End𝐾 (𝐵) ' M2 (𝐾) where 𝐾 = 𝐹 [𝑖] as in
2.3.8.
I 8. Suppose char 𝐹 = 2. Let 𝑀 ∈ M𝑛 (𝐹) be a symmetric matrix with 𝑛 odd, and
suppose that all diagonal entries of 𝑀 are zero. Show that det 𝑀 = 0.
I 9. Let char 𝐹 = 2 and let 𝐵 be a division 𝐹-algebra with a standard involution.
Prove that either the standard involution is the identity (and so 𝐵 is classified by
Exercise 3.9), or that the conclusion of Theorem 3.5.1 holds for 𝐵: namely, that
either 𝐵 = 𝐾 is a separable quadratic field extension of 𝐹 or that 𝐵 is a quaternion
algebra over 𝐹. [Hint: Replace conjugation by 𝑖 by the map 𝜙(𝑥) = 𝑖𝑥 + 𝑥𝑖, and
show that 𝜙2 = 𝜙. Then diagonalize and proceed as in the case char 𝐹 ≠ 2.]
I 10. Let char 𝐹 = 2. Show that the even Clifford algebra Clf 0 𝑄 of a nondegenerate
ternary quadratic form 𝑄 : 𝑉 → 𝐹 is a quaternion algebra over 𝐹.
I 11. Prove Lemma 6.3.7, that every quadratic form over 𝐹 with char 𝐹 = 2 has a
normalized basis.
I 12. Let char 𝐹 = 2 and let 𝑄 : 𝑉 → 𝐹 be a quadratic form over 𝐹 with discriminant
𝑑 ∈ 𝐹 × /𝐹 ×2 and dim𝐹 𝑉 = 𝑛 odd. Show that 𝑍 (Clf 𝑄) ' 𝐹 [𝑥]/(𝑥 2 − 𝑑) and
that there is a unique 𝜁 ∈ 𝑍 (Clf 𝑄) ∩ Clf 1 𝑄 such that 𝜁 2 = 1.
I 13. Prove Theorem 6.4.11.
I 14. Let 𝑄 := 𝑄 0 𝑄 00 be an orthogonal sum of two anisotropic quadratic forms over
𝐹 (with 𝐹 of arbitrary characteristic). Show that 𝑄 is isotropic if and only if
there exists 𝑐 ∈ 𝐹 × that is represented by both 𝑄 0 and −𝑄 00.
I 15. Let 𝐵 be a quaternion algebra over 𝐹 (with 𝐹 of arbitrary characteristic). Show
that 𝐹 = (𝐵0 ) ⊥ .
I 16. Prove Wedderburn’s little theorem in the following special case: a quaternion
algebra over a finite field with even cardinality is not a division ring. [Hint: See
Exercise 3.16.]
Chapter 7
Simple algebras
In this chapter, we return to the characterization of quaternion algebras. We initially

defined quaternion algebras in terms of generators and relations in Chapter 2; in the
chapters that followed, we showed that quaternion algebras are equivalently noncom-
mutative algebras with a nondegenerate standard involution. Here, we pursue another
approach, and we characterize quaternion algebras in a different way, as central simple
algebras of dimension 4.
7.1 ⊲ Motivation and summary

Consider now the “simplest” sorts of algebras. Like the primes among the integers or
the finite simple groups among finite groups, it is natural to seek algebras that cannot
be “broken down” any further. Accordingly, we say that a ring 𝐴 is simple if it has no
nontrivial two-sided (bilateral) ideals, i.e., the only two-sided ideals are {0} and 𝐴. To
show the power of this notion, consider this: if 𝜙 : 𝐴 → 𝐴 0 is a ring homomorphism
and 𝐴 is simple, then 𝜙 is either injective or the zero map (since ker 𝜙 ⊆ 𝐵 is a
two-sided ideal).
A division ring 𝐴 is simple, since every nonzero element is a unit and therefore
every nonzero ideal (left, right, or two-sided) contains 1 so is equal to 𝐴. In particular,
a field is a simple ring, and a commutative ring is simple if and only if it is a field.
The matrix ring M𝑛 (𝐹) over a field 𝐹 is also simple, something that can be checked
directly by multiplying by matrix units (Exercise 7.5).
Moreover, quaternion algebras are simple. The shortest proof of this statement,
given what we have done so far, is to employ Main Theorem 5.4.4 (and Theorem 6.4.11
in characteristic 2): a quaternion algebra 𝐵 over 𝐹 is either isomorphic to M2 (𝐹) or is
a division ring, and in either case is simple. One can also prove this directly (Exercise
7.1).
Although the primes are quite mysterious and the classification of finite simple
groups is a monumental achievement in group theory, the situation for algebras is quite
simple, indeed! Our first main result is as follows (Main Theorem 7.3.10).
93
94 CHAPTER 7. SIMPLE ALGEBRAS
Theorem 7.1.1 (Wedderburn–Artin). Let 𝐹 be a field and 𝐵 be a finite-dimensional

𝐹-algebra. Then 𝐵 is simple if and only if 𝐵 ' M𝑛 (𝐷) where 𝑛 ≥ 1 and 𝐷 is a
finite-dimensional division 𝐹-algebra.
As a corollary of Theorem 7.1.1, we give another characterization of quaternion
algebras.
Corollary 7.1.2. Let 𝐵 be an 𝐹-algebra. Then the following are equivalent:
(ii) 𝐵 ⊗𝐹 𝐹 al ' M2 (𝐹 al ), where 𝐹 al is an algebraic closure of 𝐹; and
(iii) 𝐵 is a central simple algebra of dimension dim𝐹 𝐵 = 4.
Moreover, a central simple algebra 𝐵 of dimension dim𝐹 𝐵 = 4 is either a division
algebra or has 𝐵 ' M2 (𝐹).
This corollary has the neat consequence that a division algebra 𝐵 over 𝐹 is a
quaternion algebra over 𝐹 if and only if it is central of dimension dim𝐹 𝐵 = 4.
For the reader in a hurry, we now give a proof of this corollary without invoking
the Wedderburn–Artin theorem; this proof also serves as a preview of some of the
ideas that go into the theorem.
Proof of Corollary 7.1.2. The statement (i) ⇒ (ii) was proven in Exercise 2.4(d).
To prove (ii) ⇒ (iii), suppose 𝐵 is an algebra with 𝐵al := 𝐵 ⊗𝐹 𝐹 al ' M2 (𝐹 al ). The
𝐹 -algebra 𝐵al is central simple, from above. Thus 𝑍 (𝐵) = 𝑍 (𝐵al ) ∩ 𝐵 = 𝐹. And if 𝐼
al
is a two-sided ideal of 𝐵 then 𝐼 al := 𝐼 ⊗𝐹 𝐹 al is a two-sided ideal of 𝐵al , so 𝐼 al = {0}

or 𝐼 al = 𝐵al is trivial, whence 𝐼 = 𝐼 al ∩ 𝐹 is trivial. Finally, dim𝐹 𝐵 = dim𝐹 al 𝐵al = 4.
Finally, we prove (iii) ⇒ (i). Let 𝐵 a central simple 𝐹-algebra of dimension 4.
If 𝐵 is a division algebra we are done; so suppose not. Then 𝐵 has a nontrivial
left ideal (e.g., one generated by a nonunit); let {0} ( 𝐼 ( 𝐵 be a nontrivial left
ideal with 0 < 𝑚 = dim𝐹 𝐼 minimal. Then there is a nonzero homomorphism
𝐵 → End𝐹 (𝐼) ' M𝑚 (𝐹) which is injective, since 𝐵 is simple. By dimension, we
cannot have 𝑚 = 1; if 𝑚 = 2, then 𝐵 ' M2 (𝐹) and we are done. So suppose 𝑚 = 3.
Then by minimality, every nontrivial left ideal of 𝐵 has dimension 3. But for any
𝛼 ∈ 𝐵, we have that 𝐼𝛼 is a left ideal, so the left ideal 𝐼 ∩ 𝐼𝛼 is either {0} or 𝐼; in
either case, 𝐼𝛼 ⊆ 𝐼 and 𝐼 is a right ideal as well. But this contradicts the fact that 𝐵 is
simple.
The Wedderburn–Artin theorem is an important structural result used throughout
mathematics, so we give in this chapter a self-contained account of its proof. More
generally, it will be convenient to work with semisimple algebras, finite direct products
of simple algebras. When treating ideals of an algebra we would be remiss if we did
not discuss more generally modules over the algebra, and the notions of simple and
semisimple module are natural concepts in linear algebra and representation theory:
a semisimple module is one that is a direct sum of simple modules (“completely
reducible”), analogous to a semisimple operator where every invariant subspace has
an invariant complement (e.g., a diagonalizable matrix).
The second important result in this chapter is a theorem that concerns the simple
subalgebras of a simple algebra, as follows (Main Theorem 7.7.1).
7.2. SIMPLE MODULES 95
Theorem 7.1.3 (Skolem–Noether). Let 𝐴, 𝐵 be simple 𝐹-algebras and suppose that

𝐵 is central. Suppose that 𝑓 , 𝑔 : 𝐴 → 𝐵 are homomorphisms. Then there exists 𝛽 ∈ 𝐵
such that 𝑓 (𝛼) = 𝛽−1 𝑔(𝛼) 𝛽 for all 𝛼 ∈ 𝐴.
Corollary 7.1.4. Every 𝐹-algebra automorphism of a simple 𝐹-algebra 𝐵 is inner,

i.e., Aut(𝐵) ' 𝐵× /𝐹 × .
Just as above, for our quaternionic purposes, we can give a direct proof.
Corollary 7.1.5. Let 𝐵 be a quaternion algebra over 𝐹 and let 𝐾1 , 𝐾2 ⊂ 𝐵 be quadratic

subfields. Suppose that 𝜙 : 𝐾1 −→∼ 𝐾 is an isomorphism of 𝐹-algebras. Then 𝜙 lifts to
2
an inner automorphism of 𝐵, i.e., there exists 𝛽 ∈ 𝐵 such that 𝛼2 = 𝜙(𝛼1 ) = 𝛽−1 𝛼1 𝛽
for all 𝛼1 ∈ 𝐾1 . In particular, 𝐾2 = 𝛽−1 𝐾1 𝛽.
Proof. Write 𝐾1 = 𝐹 (𝛼1 ) with 𝛼1 ∈ 𝐵 and let 𝛼2 = 𝜙(𝛼1 ) ∈ 𝐾2 ⊂ 𝐵, so 𝐾2 = 𝐹 (𝛼2 ).

We want to find 𝛽 ∈ 𝐵× such that 𝛼2 = 𝛽−1 𝛼1 𝛽. In the special case 𝐵 ' M2 (𝐹),
then 𝛼1 , 𝛼2 ∈ M2 (𝐹) satisfy the same irreducible characteristic polynomial, so by the
theory of rational canonical forms, 𝛼2 = 𝛽−1 𝛼1 𝛽 where 𝛽 ∈ 𝐵× ' GL2 (𝐹) as desired.
Suppose then that 𝐵 is a division ring. Then the set
𝑊 = {𝛽 ∈ 𝐵 : 𝛽𝛼2 = 𝛼1 𝛽} (7.1.6)
is an 𝐹-vector subspace of 𝐵. Let 𝐹 sep be a separable closure of 𝐹. (Or, apply Exercise

6.5 and work over a splitting field 𝐾 linearly disjoint from 𝐾1 ' 𝐾2 .) Then we have
𝐵 ⊗𝐹 𝐹 sep ' M2 (𝐹 sep ), and the common characteristic polynomial of 𝛼1 , 𝛼2 either
remains irreducible over 𝐹 sep (if 𝐾 ⊃ 𝐹 is inseparable) or splits as the product of two
linear factors with distinct roots. In either case, the theory of rational canonical forms
again applies, and there exists 𝛽 ∈ 𝐵 ⊗𝐹 𝐹 sep ' GL2 (𝐹 sep ) that will do; but then by
linear algebra dim𝐹 sep 𝑊 ⊗𝐹 𝐹 sep = dim𝐹 𝑊 > 0, so there exists 𝛽 ∈ 𝐵 \ {0} = 𝐵×
with the desired property.
As shown in the above proof, Corollary 7.1.5 can be seen as a general reformulation
of the rational canonical form from linear algebra.
7.2 Simple modules

Basic references for this section include Drozd–Kirichenko [DK94, §1–4], Curtis–
Reiner [CR81, §3], Lam [Lam2001, §2–3], and Farb–Dennis [FD93, Part I]. An
elementary approach to the Weddernburn–Artin theorem is given by Brešar [Bre2010].
An overview of the subject of associative algebras is given by Pierce [Pie82] and
Jacobson [Jacn2009].
Throughout this chapter, let 𝐵 be a finite-dimensional 𝐹-algebra.
To understand the algebra 𝐵, we look at its representations. A representation of
𝐵 (over 𝐹) is a vector space 𝑉 over 𝐹 together with an 𝐹-algebra homomorphism 𝐵 →
End𝐹 (𝑉). Equivalently, a representation is given by a left (or right) 𝐵-module 𝑉: this is
almost a tautology. Although one can define infinite-dimensional representations, they
will not interest us here, and we suppose throughout that dim𝐹 𝑉 < ∞, or equivalently
that 𝑉 is a finitely generated (left or right) 𝐵-module. If we choose a basis for 𝑉, we

obtain an isomorphism End𝐹 (𝑉) ' M𝑛 (𝐹) where 𝑛 = dim𝐹 𝑉, so a representation is
just a homomorphic way of thinking of the algebra 𝐵 as an algebra of matrices.
Example 7.2.1. The space of column vectors 𝐹 𝑛 is a left M𝑛 (𝐹)-module; the space
of row vectors is a right M𝑛 (𝐹)-module.
Example 7.2.2. 𝐵 is itself a left 𝐵-module, giving rise to the left regular represen-
tation 𝐵 → End𝐹 (𝐵) over 𝐹 (cf. Remark 3.3.8).
Example 7.2.3. Let 𝐺 be a finite group. Then a representation of 𝐹 [𝐺] (is the same
as an 𝐹 [𝐺]-module which) is the same as a homomorphism 𝐺 → GL(𝑉), where 𝑉 is
an 𝐹-vector space (Exercise 3.8).
Definition 7.2.4. Let 𝑉 be a left 𝐵-module. We say 𝑉 is simple (or irreducible) if
𝑉 ≠ {0} and the only 𝐵-submodules of 𝑉 are {0} and 𝑉.
We say 𝑉 is indecomposable if 𝑉 cannot be written as 𝑉 = 𝑉1 ⊕𝑉2 with 𝑉1 , 𝑉2 ≠ {0}
left 𝐵-modules.
A simple module is indecomposable, but the converse need not hold, and this is a
central point of difficulty in understanding representations.

𝑎 𝑏
Example 7.2.5. If 𝐵 = : 𝑎, 𝑏, 𝑐 ∈ 𝐹 ⊆ M2 (𝐹), then the space 𝑉 = 𝐹 2 of
0 𝑐

1
column vectors is not simple, since the subspace spanned by is a 𝐵-submodule;
0
nevertheless, 𝑉 is indecomposable (Exercise 7.4).
The importance of simple modules is analogous to that of simple groups. Arguing
by induction on the dimension of 𝑉, we have the following lemma analogous to the
Jordan–Hölder theorem on composition series.
Lemma 7.2.6. A (finite-dimensional) left 𝐵-module 𝑉 admits a filtration
𝑉 = 𝑉0 ) 𝑉1 ) 𝑉2 ) · · · ) 𝑉𝑟 = {0}
such that 𝑉𝑖 /𝑉𝑖+1 is simple for each 𝑖.

This filtration is not unique, but up to isomorphism and permutation, the quotients
𝑉𝑖 /𝑉𝑖+1 are unique.
Lemma 7.2.7. If 𝐼 is a maximal left ideal of 𝐵, then 𝐵/𝐼 is a simple 𝐵-module.
Conversely, if 𝑉 is a simple 𝐵-module, then 𝑉 ' 𝐵/𝐼 for a maximal left ideal 𝐼: more
precisely, for any 𝑥 ∈ 𝑉 r {0}, we may take
𝐼 = ann(𝑥) := {𝛼 ∈ 𝐵 : 𝛼𝑥 = 0}.
Proof. For the first statement, a submodule of 𝐵/𝐼 corresponds to a left ideal containing
𝐼, so 𝐵/𝐼 is simple if and only if 𝐼 is maximal. Conversely, letting 𝑥 ∈ 𝑉 r {0} we
have {0} ≠ 𝐵𝑥 ⊆ 𝑉 a 𝐵-submodule and so 𝐵𝑥 = 𝑉; and consequently 𝑉 ' 𝐵/𝐼 where
𝐼 = ann(𝑥) and again 𝐼 is a maximal left ideal.
7.2. SIMPLE MODULES 97
Having defined the notion of simplicity for modules, we now consider simplicity
of the algebra 𝐵.
Definition 7.2.8. An 𝐹-algebra 𝐵 is simple if the only two-sided ideals of 𝐵 are {0}
and 𝐵.
Equivalently, 𝐵 is simple if and only if any 𝐹-algebra (or even ring) homomorphism
𝐵 → 𝐴 is either injective or the zero map.
Example 7.2.9. A division 𝐹-algebra 𝐷 is simple. In fact, the 𝐹-algebra M𝑛 (𝐷) is

simple for any division 𝐹-algebra 𝐷 (Exercise 7.5), and in particular M𝑛 (𝐹) is simple.
Example 7.2.10. Let 𝐹 al be an algebraic closure of 𝐹. If 𝐵 ⊗𝐹 𝐹 al is simple, then 𝐵

is simple. The association 𝐼 ↦→ 𝐼 ⊗𝐹 𝐹 al is an injective map from the set of two-sided
ideals of 𝐵 to the set of two-sided ideals of 𝐵 ⊗𝐹 𝐹 al .
7.2.11. If 𝐵 is a quaternion algebra over 𝐹, then 𝐵 is simple. We have 𝐵 ⊗𝐹 𝐹 al '

M2 (𝐹 al ), which is simple by Example 7.2.9, and 𝐵 is simple by Example 7.2.10.
Example 7.2.9 shows that algebras of the form M𝑛 (𝐷) with 𝐷 a division 𝐹-algebra
yield a large class of simple 𝐹-algebras. In fact, these are all such algebras, a fact we
will now prove. First, a few preliminary results.
Lemma 7.2.12 (Schur). Let 𝐵 be an 𝐹-algebra. Let 𝑉1 , 𝑉2 be simple 𝐵-modules. Then

any homomorphism 𝜙 : 𝑉1 → 𝑉2 of 𝐵-modules is either zero or an isomorphism.
Proof. We have that ker 𝜙 and img 𝜙 are 𝐵-submodules of 𝑉1 and 𝑉2 , respectively, so
either 𝜙 = 0 or ker 𝜙 = {0} and img 𝜙 = 𝑉2 , hence 𝑉1 ' 𝑉2 .
Corollary 7.2.13. If 𝑉 is a simple 𝐵-module, then End 𝐵 (𝑉) is a division ring.
7.2.14. Let 𝐵 be an 𝐹-algebra and consider 𝐵 as a left 𝐵-module. Then there is a map
𝜌 : 𝐵op → End 𝐵 (𝐵)

𝛼 ↦→ (𝜌 𝛼 : 𝛽 ↦→ 𝛽𝛼),
where 𝐵op is the opposite algebra of 𝐵 defined in 3.2.2. The map 𝜌 is injective since
𝜌 𝛼 = 0 implies 𝜌 𝛼 (1) = 𝛼 = 0; it is also surjective, since if 𝜙 ∈ End 𝐵 (𝐵) then letting
𝛼 = 𝜙(1) we have 𝜙(𝛽) = 𝛽𝜙(1) = 𝛽𝛼 for all 𝛽 ∈ 𝐵. Finally, it is an 𝐹-algebra
homomorphism, since
𝜌 𝛼𝛽 (𝜇) = 𝜇(𝛼𝛽) = (𝜇𝛼) 𝛽 = (𝜌 𝛽 ◦ 𝜌 𝛼 ) (𝜇),
and therefore 𝜌 is an isomorphism of 𝐹-algebras.

One lesson here is that a left module has endomorphisms that act naturally on the
right; but the more common convention is that endomorphisms also act on the left. In
order to make this compatible, the opposite algebra intervenes.
7.2.15. More generally, the decomposition of modules is determined by idempotent

endomorphisms as follows. Let 𝑉 be a left 𝐵-module. Then 𝑉 is indecomposable if and
only if End 𝐵 (𝑉) has no nontrivial idempotents: that is to say, if 𝑒 ∈ End 𝐵 (𝑉) satisfies
𝑒 2 = 𝑒, then 𝑒 ∈ {0, 1}. Given a nontrivial idempotent, we can write 𝑉 = 𝑒𝑉 ⊕ (1−𝑒)𝑉,
and conversely if 𝑉 = 𝑉1 ⊕ 𝑉2 then the projection 𝑉 → 𝑉1 ⊆ 𝑉 gives an idempotent.
7.2.16. Many theorems of linear algebra hold equally well over division rings as they
do over fields, as long as one is careful about the direction of scalar multiplication. For
example, let 𝐷 be a division 𝐹-algebra and let 𝑉 be a left 𝐷-module. Then 𝑉 ' 𝐷 𝑛
is free, and choice of basis for 𝑉 gives an isomorphism End𝐷 (𝑉) ' M𝑛 (𝐷 op ). When
𝑛 = 1, this becomes End𝐷 (𝐷) ' 𝐷 op , as in 7.2.14.
Lemma 7.2.17. Let 𝐵 be a (finite-dimensional) simple 𝐹-algebra. Then there exists a

simple left 𝐵-module which is unique up to isomorphism.
Proof. Since 𝐵 is finite-dimensional over 𝐹, there is a nonzero left ideal 𝐼 of 𝐵 of

minimal dimension, and such an ideal 𝐼 is necessarily simple. Moreover, if 𝜈 ∈ 𝐼 is
nonzero then 𝐵𝜈 = 𝐼, since 𝐵𝜈 ⊆ 𝐼 is nonzero and 𝐼 is simple. Let 𝐼 = 𝐵𝜈 with 𝜈 ∈ 𝐼.
Now let 𝑉 be any simple 𝐵-module; we will show 𝐼 ' 𝑉 as 𝐵-modules. Since 𝐵
is simple, the natural map 𝐵 → End𝐹 (𝑉) is injective (since it is nonzero). Therefore,
there exists 𝑥 ∈ 𝑉 such that 𝜈𝑥 ≠ 0, so 𝐼𝑥 ≠ {0}. Thus, the map 𝐼 → 𝑉 by 𝛽 ↦→ 𝛽𝑥 is
a nonzero 𝐵-module homomorphism, so it is an isomorphism by Schur’s lemma.
Example 7.2.18. The unique simple left M𝑛 (𝐹)-module (up to isomorphism) is the
space 𝐹 𝑛 of column vectors (Example 7.2.1).
7.2.19. Every algebra can be decomposed according to its idempotents 7.2.15. Let 𝐵
be a finite-dimensional 𝐹-algebra. Then we can write 𝐵 = 𝐼1 ⊕ · · · ⊕ 𝐼𝑟 as a direct
sum of indecomposable left 𝐵-modules: this follows by induction, as the decomposing
procedure must stop because each factor is a finite-dimensional 𝐹-vector space. This
means we may write
1 = 𝑒1 + · · · + 𝑒𝑟
Í
with 𝑒 𝑖 ∈ 𝐼𝑖 . For each 𝛼 ∈ 𝐼𝑖 we have 𝛼 = 𝑖 𝛼𝑒 𝑖 whence 𝛼𝑒 𝑖 = 𝛼 and 𝛼𝑒 𝑗 = 0 for
𝑗 ≠ 𝑖, which implies that
𝑒 2𝑖 = 𝑒 𝑖 , 𝑒𝑖 𝑒 𝑗 = 0 for 𝑗 ≠ 𝑖, and 𝐼𝑖 = 𝐵𝑒 𝑖 .
Thus each 𝑒 𝑖 is idempotent; we call {𝑒 1 , . . . , 𝑒𝑟 } a complete set of primitive orthog-

onal idempotents: the orthogonal is because 𝑒 𝑖 𝑒 𝑗 = 0 for 𝑗 ≠ 𝑖, and the primitive is
because each 𝑒 𝑖 is not the sum of two other orthogonal idempotents (by 7.2.15).
Remark 7.2.20. The tight connection between 𝐹 and M𝑛 (𝐹) is encoded in the fact that
the two rings are Morita equivalent: there is an equivalence of categories between
𝐹-vector spaces and left M𝑛 (𝐹)-modules. For more on this rich subject, see Lam
[Lam99, §18], Reiner [Rei2003, Chapter 4], and Curtis–Reiner [CR81, §35].
7.3. WEDDERBURN–ARTIN 99
7.3 Semisimple modules and the Wedderburn–Artin theorem

We continue our assumptions that 𝐵 is a finite-dimensional 𝐹-algebra and a 𝐵-module
𝑉 is finite-dimensional.
Definition 7.3.1. A 𝐵-module 𝑉 is semisimple (or completely É reducible) if 𝑉 is

isomorphic to a (finite) direct sum of simple 𝐵-modules 𝑉 ' 𝑖 𝑉𝑖 .
𝐵 is a semisimple 𝐹-algebra if 𝐵 is semisimple as a left 𝐵-module.
Remark 7.3.2. More precisely, we have defined the notion of left semisimple and
could equally well define right semisimple; below we will see that these two notions
are the same.
Example 7.3.3. If 𝐵 = 𝐹, then simple 𝐹-modules are one-dimensional vector spaces,

and as 𝐹 is simple these are the only ones. Every 𝐹-vector space has a basis and so is
the direct sum of one-dimensional subspaces, thus every 𝐹-module is semisimple.
Example 7.3.4. A finite-dimensional commutative 𝐹-algebra 𝐵 is semisimple if and

only if 𝐵 is the product of field extensions of 𝐹, i.e., 𝐵 ' 𝐾1 × · · · × 𝐾𝑟 with 𝐾𝑖 ⊇ 𝐹 a
finite extension of fields.
Lemma 7.3.5. The following statements hold.
(a) A 𝐵-module 𝑉 is semisimple if and only if it is the sum of simple 𝐵-modules.

(b) A submodule or a quotient module of a semisimple 𝐵-module is semisimple.
(c) If 𝐵 is a semisimple 𝐹-algebra, then every 𝐵-module is semisimple.
Í
Proof. For (a), let 𝑉 = 𝑖 𝑉𝑖 be the sum of simple 𝐵-modules. Since 𝑉 is finite-
dimensional, we can rewrite it as an irredundant finite sum; and then since each 𝑉𝑖 is
simple, the intersection of any two distinct summands is {0}, so the sum is direct.
For (b), let 𝑊 ⊆ 𝑉 be a submodule of the semisimple 𝐵-module 𝑉. Every 𝑥Í∈ 𝑊
with 𝑥 ≠ 0 is contained in a simple 𝐵-submodule of 𝑊 by minimality, so 𝑊 = 𝑖 𝑊𝑖
is a sum of simple 𝐵-modules. The result now follows from (a) for submodules. For
quotient modules, suppose 𝜙 : 𝑉 → 𝑍 is a surjective 𝐵-module homomorphism; then
𝜙−1 (𝑍) ⊆ 𝑉 is a 𝐵-submodule, and 𝜙−1 (𝑍) = 𝑖 𝑊𝑖 is a sum of simple 𝐵-modules,
Í
Í
and hence by Schur’s lemma 𝑍 = 𝑖 𝜙(𝑊𝑖 ) is semisimple.
For (c), let 𝑉 be a 𝐵-module. Since 𝑉 is finitely generated as a 𝐵-module, there is a
surjective 𝐵-module homomorphism 𝐵𝑟 → 𝑉 for some 𝑟 ≥ 1. Since 𝐵𝑟 is semisimple,
so too is 𝑉 by (b).
Remark 7.3.6. Doing linear algebra with semisimple modules mirrors very closely
linear algebra over a field. We have already seen that every submodule and quotient
module of a semisimple module is again semisimple. Moreover, every module homo-
morphism 𝑉 → 𝑊 with 𝑉 semisimple splits, and every submodule of a semisimple
module is a direct summand. The extent to which this fails over other rings concerns
the structure of projective modules; we take this up in Chapter 20.
Lemma 7.3.7. If 𝐵 is a simple 𝐹-algebra, then 𝐵 is a semisimple 𝐹-algebra.

Proof. Let 𝐼 ⊆ 𝐵 be a minimal nonzero left ideal, the unique simple left 𝐵-module up
to isomorphism as in Lemma 7.2.17. For all 𝛼 ∈ 𝐵, the left ideal 𝐼𝛼 is a homomorphic
Í
image of 𝐼, so by Schur’s lemma, either 𝐼𝛼 = {0} or 𝐼𝛼 is simple. Let 𝐴 := 𝛼∈𝐵 𝐼𝛼.
Then 𝐴 is a nonzero two-sided (!) ideal of 𝐵, so since 𝐵 is simple, we conclude
𝐴 = 𝐵. Thus 𝐵 is the sum of simple 𝐵-modules, and the result follows from Lemma
7.3.5(a).
Corollary 7.3.8. A (finite) direct product of simple 𝐹-algebras is a semisimple 𝐹-

algebra.
Proof. If 𝐵 ' 𝐵1 × É · · · × 𝐵𝑟 with each 𝐵𝑖 simple, then by Lemma 7.3.7, each 𝐵𝑖 is

semisimple so 𝐵𝑖 = 𝑗 𝐼𝑖 𝑗 is the direct sum of simple 𝐵𝑖 -modules 𝐼𝑖 𝑗 . Each 𝐼𝑖 𝑗 has
the natural structure
É of a 𝐵-module (extending by zero), and with this structure it is
simple, and 𝐵 = 𝑖, 𝑗 𝑖 𝑗 is semisimple.
𝐼
The converse of Corollary 7.3.8 is true and is proven as Corollary 7.3.14, a conse-
quence of the Wedderburn–Artin theorem.
In analogy to 7.2.16, we have the following corollary.
Corollary 7.3.9. Let 𝐵 be a simple 𝐹-algebra and let 𝑉 be a left 𝐵-module. Then
𝑉 ' 𝐼 ⊕𝑛 for some 𝑛 ≥ 1, where 𝐼 is a simple left 𝐵-module. In particular, two left
𝐵-modules 𝑉1 , 𝑉2 are isomorphic if and only if dim𝐹 𝑉1 = dim𝐹 𝑉2 .
Proof. Since 𝐵 is simple, 𝐵 is semisimple by Lemma 7.3.7, and 𝑉 is semisimple by

Lemma 7.3.5. But by Lemma 7.2.17, there is a unique simple left 𝐵-module 𝐼, and
the result follows.
In other words, this corollary says that if 𝐵 is simple then every left 𝐵-module 𝑉
is free over 𝐵, so has a left basis over 𝐵; if we define the rank of a left 𝐵-module 𝑉
to be cardinality of this basis (the integer 𝑛 such that 𝑉 ' 𝐼 ⊕𝑛 as in Corollary 7.3.9),
then two such modules are isomorphic if and only if they have the same rank.
We now come to one of the main results of this chapter.
Main Theorem 7.3.10 (Wedderburn–Artin). Let 𝐵 be a finite-dimensional 𝐹-algebra.

Then 𝐵 is semisimple if and only if there exist integers 𝑛1 , . . . , 𝑛𝑟 and division algebras
𝐷 1 , . . . , 𝐷 𝑟 such that
𝐵 ' M𝑛1 (𝐷 1 ) × · · · × M𝑛𝑟 (𝐷 𝑟 ).
Such a decomposition is unique up to permuting the integers 𝑛1 , . . . , 𝑛𝑟 and applying
an isomorphism to the division rings 𝐷 1 , . . . , 𝐷 𝑟 .
Î
Proof. If 𝐵 ' 𝑖 𝑀𝑛𝑖 (𝐷 𝑖 ), then each factor 𝑀𝑛𝑖 (𝐷 𝑖 ) is a simple 𝐹-algebra by Example
7.2.9, so by Corollary 7.3.8, 𝐵 is semisimple.
So suppose 𝐵 is semisimple. Then we can write 𝐵 as a left 𝐵-module as the direct
sum 𝐵 ' 𝐼1⊕𝑛1 ⊕ · · · ⊕ 𝐼𝑟⊕𝑛𝑟 of simple 𝐵-modules 𝐼1 , . . . , 𝐼𝑟 , grouped up to isomorphism.
We have End 𝐵 (𝐵) ' 𝐵op by 7.2.14. By Schur’s lemma,
Ê
End 𝐵 𝐼𝑖⊕𝑛𝑖 ;

End 𝐵 (𝐵) '
𝑖
7.4. JACOBSON RADICAL 101
by 7.2.16,
End 𝐵 𝐼𝑖⊕𝑛𝑖 ' M𝑛𝑖 (𝐷 𝑖 )
where 𝐷 𝑖 = End 𝐵 (𝐼𝑖 ) is a division ring. So

op op
𝐵 ' End 𝐵 (𝐵) op ' M𝑛1 (𝐷 1 ) × · · · × M𝑛𝑟 (𝐷 𝑟 ).
The statements about uniqueness are then clear.
Remark 7.3.11. Main Theorem 7.3.10 as it is stated was originally proven by Wedder-
burn [Wed08], and so is sometimes called Wedderburn’s theorem. However, this term
may also apply to the theorem of Wedderburn that a finite division ring is a field; and
Artin generalized Main Theorem 7.3.10 to rings where the ascending and descending
chain condition holds for left ideals [Art26]. We follow the common convention by
referring to Main Theorem 7.3.10 as the Wedderburn–Artin theorem.
Corollary 7.3.12. Let 𝐵 be a simple 𝐹-algebra. Then 𝐵 ' M𝑛 (𝐷) for a unique
𝑛 ∈ Z ≥1 and a division algebra 𝐷 unique up to isomorphism.
Example 7.3.13. Let 𝐵 be a division 𝐹-algebra. Then 𝑉 = 𝐵 is a simple 𝐵-module,

and in Corollary 7.3.12 we have 𝐷 = End 𝐵 (𝐵) = 𝐵op , and the Wedderburn–Artin
isomorphism is just 𝐵 ' M1 ((𝐵op ) op ).
Corollary 7.3.14. An 𝐹-algebra 𝐵 is semisimple if and only if 𝐵 is the direct product

of simple 𝐹-algebras.
Proof. Immediate from the Wedderburn–Artin theorem, as each factor M𝑛𝑖 (𝐷 𝑖 ) is

simple.
7.4 Jacobson radical

We now consider an important criterion for establishing the semisimplicity of an
𝐹-algebra. Let 𝐵 be a finite-dimensional 𝐹-algebra.
Definition 7.4.1. The Jacobson radical rad 𝐵 of 𝐵 is the intersection of all maximal
left ideals of 𝐵.
We will in Corollary 7.4.6 see that this definition has left-right symmetry. Before
doing so, we see right away the importance of the Jacobson radical in the following
lemma.
Lemma 7.4.2. 𝐵 is semisimple if and only if rad 𝐵 = {0}.
Proof. First, suppose 𝐵 is semisimple. Then 𝐵 as a left 𝐵-module is isomorphic to

the direct sum of simple left ideals of 𝐵. Suppose rad 𝐵 ≠ {0}; then rad 𝐵 contains
a minimal, hence simple, nonzero left ideal 𝐼 ⊆ 𝐵. Then 𝐵 = 𝐼 ⊕ 𝐼 0 for some 𝐵-
submodule 𝐼 0 and 𝐵/𝐼 0 ' 𝐼 so 𝐼 0 is a maximal left ideal. Therefore rad 𝐵 ⊆ 𝐼 0, but
then rad 𝐵 ∩ 𝐼 = {0}, a contradiction.
Conversely, suppose rad 𝐵 = {0}. Suppose 𝐵 is not semisimple. Let 𝐼1 be a

minimal left ideal of 𝐵. Since 𝐼1 ≠ {0} = rad 𝐵, there exists a maximal left ideal
𝐽1 not containing 𝐼1 , so 𝐼1 ∩ 𝐽1 = {0} and 𝐵 = 𝐼1 ⊕ 𝐽1 . Since 𝐵 is not semisimple,
𝐽1 ≠ {0}, and there exists a minimal left ideal 𝐼2 ( 𝐽1 ⊆ 𝐵. Continuing in this fashion,
we obtain a descending chain 𝐽1 ) 𝐽2 ) . . . , a contradiction.
Corollary 7.4.3. 𝐵/rad 𝐵 is semisimple.
Proof. Let 𝐽 = rad 𝐵. Under the natural map 𝐵 → 𝐵/𝐽, the intersection of all maximal
left ideals of 𝐵/rad 𝐵 corresponds to the intersection of all maximal left ideals of 𝐵
containing 𝐽; but rad 𝐵 is the intersection thereof, so rad(𝐵/𝐽) = {0} and by Lemma
7.4.2, 𝐵/𝐽 is semisimple.
We now characterize the Jacobson radical in several ways.

7.4.4. For a left 𝐵-module 𝑉, define its annihilator by
ann 𝑉 := {𝛼 ∈ 𝐵 : 𝛼𝑉 = 0}.
Every annihilator ann 𝑉 is a two-sided ideal of 𝐵: if 𝛼 ∈ ann(𝑉) and 𝛽 ∈ 𝐵, then

𝛼𝛽𝑉 ⊆ 𝛼𝑉 = {0} so 𝛼𝛽 ∈ ann(𝑉).
Lemma 7.4.5. The Jacobson radical is equal to the intersection of the annihilators of
Ñ
all simple left 𝐵-modules: i.e., we have rad 𝐵 = 𝑉 ann 𝑉, the intersection taken over
all simple left 𝐵-modules. Moreover, if 𝛼 ∈ rad 𝐵, then 1 − 𝛼 ∈ 𝐵× .
Ñ
Proof. We begin with the containment (⊇). Let 𝛼 ∈ 𝑉 ann 𝑉 and let 𝐼 be a maximal
left ideal. Then 𝑉 = 𝐵/𝐼 is a simple left 𝐵-module, so 𝛼 ∈ ann(𝐵/𝐼) whence 𝛼𝐵 ⊆ 𝐼
and 𝛼 ∈ 𝐼.
The containment (⊆) follows with a bit more work. Let 𝛼 ∈ rad 𝐵, and let 𝑉 be a
simple left 𝐵-module. Assume for purposes of contradiction that 𝑥 ∈ 𝑉 has 𝛼𝑥 ≠ 0.
Then as in Lemma 7.2.7, 𝑉 = 𝐵(𝛼𝑥) so 𝑥 = 𝛽𝛼𝑥 for some 𝛽 ∈ 𝐵 and (1 − 𝛽𝛼)𝑥 = 0.
Let 𝐼 be a maximal left ideal containing 1 − 𝛽𝛼. Since 𝛼 ∈ rad 𝐵, we have 𝛼 ∈ 𝐼, and
thus 1 = (1 − 𝛽𝛼) + 𝛽𝛼 ∈ 𝐼, a contradiction. Thus 𝛼𝑉 = {0} and 𝛼 ∈ ann 𝑉.
The final statement follows along similar lines as the previous paragraph (Exercise
7.9).
Corollary 7.4.6. The Jacobson radical rad 𝐵 is a two-sided ideal of 𝐵.
Proof. The statement follows by combining 7.4.4 and Lemma 7.4.5: rad 𝐵 is the
intersection of two-sided ideals and so is itself a two-sided ideal.
Example√︁ 7.4.7. If 𝐵 is commutative (and still a finite-dimensional 𝐹-algebra), then

rad 𝐵 = (0) is the nilradical of 𝐵, the set of all nilpotent elements of 𝐵.
A two-sided ideal 𝐽 ⊆ 𝐵 is nilpotent if 𝐽 𝑛 = {0} for some 𝑛 ≥ 1, i.e., every product
of 𝑛 elements from 𝐽 is zero. Every element of a nilpotent ideal is itself nilpotent.
Lemma 7.4.8. 𝐽 = rad 𝐵 contains every nilpotent two-sided ideal, and 𝐽 itself is
nilpotent.
7.5. CENTRAL SIMPLE ALGEBRAS 103
Proof. If 𝐼 ⊆ 𝐵 is a nilpotent two-sided ideal, then 𝐼 + 𝐽 is a nilpotent two-sided ideal

of 𝐵/𝐽; but rad(𝐵/𝐽) = {0} by Corollary 7.4.3, so 𝐵/𝐽 is the direct product of simple
algebras (Corollary 7.3.14) and therefore has no nonzero nilpotent two-sided ideals.
Therefore 𝐼 ⊆ 𝐼 + 𝐽 ⊆ 𝐽.
Now we prove that 𝐽 is nilpotent. Consider the descending chain
𝐵 ⊃ 𝐽 ⊇ 𝐽2 ⊇ . . . .
There exists 𝑛 ∈ Z ≥1 such that 𝐽 𝑛 = 𝐽 2𝑛 . We claim that 𝐽 𝑛 = {0}. Assume for the
purposes of contradiction that 𝐼 ⊆ 𝐽 𝑛 is a minimal left ideal such that 𝐽 𝑛 𝐼 ≠ {0}. Let
𝛼 ∈ 𝐼 be such that 𝐽 𝑛 𝛼 ≠ {0}; by minimality 𝐽 𝑛 𝛼 = 𝐼, so 𝛼 = 𝜂𝛼 for some 𝜂 ∈ 𝐽 𝑛 ,
thus (1 − 𝜂)𝛼 = 0. But 𝜂 ∈ 𝐽 𝑛 ⊆ 𝐽 = rad 𝐵. By Lemma 7.4.5, 1 − 𝜂 ∈ 𝐵× is a unit
hence 𝛼 = 0, a contradiction.
Example 7.4.9. Suppose 𝐵 has a standard involution. Then by Lemma 7.4.8 and the
fact that 𝐵 has degree 2, we conclude that rad 𝐵 ⊆ {𝜖 ∈ 𝐵 : 𝜖 2 = 0}. If char 𝐹 ≠ 2
and we define rad(nrd) as in 4.3.9 for the quadratic form nrd, then rad(nrd) = rad 𝐵
(Exercise 7.20).
Corollary 7.4.10. The Jacobson radical rad 𝐵 is the intersection of all maximal right
ideals of 𝐵.
Proof. Lemma 7.4.8 gives a left-right symmetric characterization of the Jacobson

radical, so rad 𝐵 = rad 𝐵op . There is a bĳection between simple left 𝐵-modules and
simple right 𝐵op -modules, and the result follows.
7.5 Central simple algebras

For more on central simple algebras (and in particular division algebras), see e.g.
Saltman [Sal99] or Draxl [Dra83].
Recall (2.1.1) that the center of 𝐵 is defined as
𝑍 (𝐵) := {𝛼 ∈ 𝐵 : 𝛼𝛽 = 𝛽𝛼 for all 𝛼 ∈ 𝐵}.
Remark 7.5.1. An 𝐹-algebra 𝐵 is a central 𝑍 (𝐵)-algebra when 𝑍 (𝐵) is a field. (Under

a more general definition of algebra, every algebra is an algebra over its center.)
Example 7.5.2. The center 𝑍 (𝐵) of a simple 𝐹-algebra is a field, since it is a simple
commutative 𝐹-algebra. One reaches the same conclusion by applying Corollary
7.3.12 together with 𝑍 (M𝑛 (𝐷)) = 𝑍 (𝐷) (Exercise 7.5).
The category of central simple algebras is closed under tensor product, as follows.
Proposition 7.5.3. Let 𝐴, 𝐵 be 𝐹-algebras and suppose that 𝐵 is central.
(a) The center of 𝐴 ⊗𝐹 𝐵 is the image of 𝑍 ( 𝐴) ↩→ 𝐴 ⊗𝐹 𝐵 under 𝑧 ↦→ 𝑧 ⊗ 1.

(b) Suppose that 𝐴, 𝐵 are simple. Then 𝐴 ⊗𝐹 𝐵 is simple.
Í
Proof. First, centrality in part (a). Suppose that 𝛾 = 𝑖 𝛼𝑖 ⊗ 𝛽𝑖 ∈ 𝑍 ( 𝐴 ⊗ 𝐵) (a finite
sum). By rewriting the tensor, without loss of generality, we may suppose that 𝛼𝑖
are linearly independent over 𝐹. Í Then by properties of tensor products, the elements
𝛽𝑖 ∈ 𝐵 in the representation 𝛾 = 𝑖 𝛼𝑖 ⊗ 𝛽𝑖 are unique. But then for all 𝛽 ∈ 𝐵,
! !
∑︁ ∑︁ ∑︁ ∑︁
(𝛼𝑖 ⊗ 𝛽𝛽𝑖 ) = (1 ⊗ 𝛽) 𝛼𝑖 ⊗ 𝛽 𝑖 = 𝛼𝑖 ⊗ 𝛽𝑖 (1 ⊗ 𝛽) = (𝛼𝑖 ⊗ 𝛽𝑖 𝛽)
𝑖 𝑖 𝑖 𝑖
so 𝛽𝛽𝑖 = 𝛽𝑖 𝛽 for each 𝑖; thus 𝛽𝑖 = 𝑏 𝑖 ∈ 𝑍 (𝐵) = 𝐹. Hence

!
∑︁ ∑︁ ∑︁
𝛾= 𝛼𝑖 ⊗ 𝑏 𝑖 = 𝛼𝑖 𝑏 𝑖 ⊗ 1 = 𝛼𝑖 𝑏 𝑖 ⊗ 1;
𝑖 𝑖 𝑖
Í
since 𝛼 ⊗ 1 also commutes with 𝛾 for all 𝛼 ∈ 𝐴, we have 𝑖 𝛼𝑖 𝑏 𝑖 ∈ 𝑍 ( 𝐴). Thus
𝛾 ∈ 𝑍 ( 𝐴) ⊗ 𝐹 = 𝑍 ( 𝐴).
Next,
Í𝑚simplicity in part (b). Let 𝐼 be a nontrivial two-sided ideal in 𝐴 ⊗ 𝐵, and
let 𝛾 = 𝑖=1 𝛼𝑖 ⊗ 𝛽𝑖 ∈ 𝐼 r {0}. Without loss of generality, we may suppose 𝛽1 ≠ 0.
Then 𝐵𝛽1 𝐵 = 𝐵 since 𝐵 is simple; multiplying on the left and right by elements of
𝐵 ⊆ 𝐴 ⊗ 𝐵, we may suppose further that 𝛽1 = 1. Let 𝛾 ∈ 𝐼 r{0} be such an element that
is minimal with respect to 𝑚; then in particular the elements 𝛽𝑖 are linearly independent
over 𝐹. Now for each 𝛽 ∈ 𝐵,
𝑚
∑︁
(1 ⊗ 𝛽)𝛾 − 𝛾(1 ⊗ 𝛽) = (𝛼𝑖 ⊗ (𝛽𝛽𝑖 − 𝛽𝑖 𝛽)) ∈ 𝐼;
𝑖=2
but by minimality of 𝑚, the right-hand side is zero, so 𝛽𝛽𝑖 = 𝛽𝑖 𝛽 for all 𝑖. Hence
𝛽𝑖 ∈ 𝑍 (𝐵) = 𝐹 for all 𝑖 and as above 𝛾 = 𝛼 ⊗ 1 for some 0 ≠ 𝛼 ∈ 𝐴. But then
𝐼 ⊇ ( 𝐴 ⊗ 1) (𝛼 ⊗ 1) ( 𝐴 ⊗ 1) = ( 𝐴𝛼𝐴) ⊗ 1 = 𝐴 ⊗ 1
since 𝐴 is simple, so 𝐼 ⊇ ( 𝐴 ⊗ 1) (1 ⊗ 𝐵) = 𝐴 ⊗ 𝐵, and thus 𝐼 = 𝐴 ⊗ 𝐵 and 𝐴 ⊗ 𝐵 is

simple.
Lemma 7.5.4. If 𝐵 is a finite-dimensional algebra over 𝐹, then 𝐵 is a central simple

𝐹-algebra if and only if the map
𝜙 : 𝐵 ⊗𝐹 𝐵op −→ ∼ End (𝐵)

𝐹
Í Í
𝑖 𝛼𝑖 ⊗ 𝛽 𝑖 →
↦ (𝜇 →
↦ 𝑖 𝛼𝑖 𝜇𝛽𝑖 )
is an isomorphism.
Proof. First, the implication (⇒). Just as in 7.2.14, 𝜙 is a nonzero 𝐹-algebra ho-
momorphism. By Proposition 7.5.3, 𝐵 ⊗𝐹 𝐵op is simple, so 𝜙 is injective. Since
dim𝐹 (𝐵 ⊗𝐹 𝐵op ) = dim𝐹 End𝐹 (𝐵) = (dim𝐹 𝐵) 2 , 𝜙 is an isomorphism.
Now the converse (⇐); suppose 𝜙 is an isomorphism. If 𝐼 is an ideal of 𝐵 then
𝜙(𝐼 ⊗ 𝐵op ) ⊆ End𝐹 (𝐵) is an ideal; but End𝐹 (𝐵) is simple over 𝐹, therefore 𝐼 is trivial.
And if 𝛼 ∈ 𝑍 (𝐵) then 𝜙(𝛼 ⊗ 1) ∈ 𝑍 (End𝐹 (𝐵)) = 𝐹, so 𝛼 ∈ 𝐹.
7.5.5. Among central simple algebras over a field, quaternion algebras have an espe-
cially nice presentation because the quadratic norm form can be put into a standard
form (indeed, diagonalized in characteristic not 2). More generally, one may look at
algebras with a similarly nice presentation, as follows.
Let 𝐹 be a field, let 𝐾 ⊃ 𝐹 be a cyclic extension of 𝐹 of degree 𝑛 = [𝐾 : 𝐹], let
𝜎 ∈ Gal(𝐾 | 𝐹) be a generator, and let 𝑏 ∈ 𝐹 × . For example, if 𝐹 contains a primitive
√
𝑛th√root of √unity 𝜁 ∈ 𝐹 × , and 𝑎 ∈ 𝐹 × r 𝐹 ×𝑛 , then we may take 𝐾 = 𝐹 ( 𝑛 𝑎) and
𝜎( 𝑛 𝑎) = 𝜁 𝑛 𝑎. We then define the cyclic algebra

𝐾, 𝜎, 𝑏
= 𝐾 ⊕ 𝐾 𝑗 ⊕ · · · ⊕ 𝐾 𝑗 𝑛−1
𝐹
to be the left 𝐾-vector space with basis 1, 𝑗, . . . , 𝑗 𝑛−1 and with multiplication 𝑗 𝑛 = 𝑏
and 𝑗 𝛼 = 𝜎(𝛼) 𝑗 for 𝛼 ∈ 𝐾. The definition of a cyclic algebra generalizes that
of 6.1.5, where there is only one choice for the generator 𝜎. A cyclic algebra is a
central simple algebra over 𝐹 of dimension 𝑛2 , and indeed (𝐾, 𝜎, 𝑏 | 𝐾) ' M𝑛 (𝐾).
(See Exercise 7.11.) More generally, we may relax the condition that 𝐺 be cyclic:
there is an analogous construction for any finite Galois extension, yielding a central
simple algebra called a crossed product algebra (and giving an interpretation to a
second cohomology group): see Reiner [Rei2003, §29–30]. There are significant
open problems relating cyclic algebras and crossed products to central simple algebras
in general [ABGV2006].
It is a consequence of the main theorem of class field theory that if 𝐹 is a global
field then every (finite-dimensional) central simple algebra over 𝐹 is isomorphic to a
cyclic algebra.
Remark 7.5.6. The theory of central simple algebras and Brauer groups extends to one
over commutative rings (or even schemes), and this becomes the theory of Azumaya
algebras: see Saltman [Sal99, §2].

Having set the stage, we are now ready to prove the following final characterizations
of quaternion algebras.
Proposition 7.6.1. Let 𝐵 be an 𝐹-algebra. Then the following are equivalent:

(ii) 𝐵 is a central simple 𝐹-algebra with dim𝐹 𝐵 = 4;
(iii) 𝐵 is a central semisimple 𝐹-algebra with dim𝐹 𝐵 = 4; and
(iv) 𝐵 ⊗𝐹 𝐹 al ' M2 (𝐹 al ), where 𝐹 al is an algebraic closure of 𝐹.
Proof. First, (i) ⇒ (ii): if 𝐵 is a quaternion algebra, then 𝐵 is central simple (7.2.11).
The equivalence (ii) ⇔ (iii) follows from the Wedderburn–Artin theorem:
𝑟
∑︁
1 = dim 𝑍 (𝐵) = dim𝐹 𝑍 (𝐷 𝑖 ) ≥ 𝑟
𝑖=1
so 𝑟 = 1.
Next we prove (ii) ⇒ (iv). If 𝐵 is central simple, then 𝐵 ⊗𝐹 𝐹 al is a central simple
𝐹 -algebra by Proposition 7.5.3. But by Exercise 2.9, the only division 𝐹 al -algebra
al
is 𝐹 al , so by the Wedderburn–Artin theorem, 𝐵 ⊗𝐹 𝐹 al ' M𝑛 (𝐹 al ); by dimensions,

𝑛 = 2.
It remains to prove (iv) ⇒ (i). So suppose 𝐵 ⊗𝐹 𝐹 al ' M2 (𝐹 al ). Then 𝐵 is
simple by Example 7.2.10 and dim𝐹 𝐵 = 4. By the Wedderburn–Artin theorem
(Corollary 7.3.12), we have 𝐵 ' M𝑛 (𝐷) with 𝑛 ∈ Z ≥1 and 𝐷 a division ring. Since
4 = dim𝐹 𝐵 = 𝑛2 dim𝐹 𝐷, either 𝑛 = 2 and 𝐵 ' M2 (𝐹), or 𝑛 = 1 and 𝐵 is a division
ring.
In this latter case, the result will follow from Theorem 3.5.1 (and Theorem 6.2.8
for the case char 𝐹 = 2) if we show that 𝐵 has degree 2. But for any 𝛼 ∈ 𝐵 we have
that 𝛼 ∈ 𝐵 ⊗𝐹 𝐹 al ' M2 (𝐹 al ) satisfies its characteristic polynomial of degree 2, so
that 1, 𝛼, 𝛼2 are linearly dependent over 𝐹 al and hence linearly dependent over 𝐹, by
linear algebra.
Inspired by the proof of this result, we reconsider and reprove our splitting criterion
for quaternion algebras.
Proposition 7.6.2. Let 𝐵 be a quaternion algebra over 𝐹. Then the following are
equivalent:
(i) 𝐵 ' M2 (𝐹);

(iii) There exists 0 ≠ 𝜖 ∈ 𝐵 such that 𝜖 2 = 0;
(iv) 𝐵 has a nontrivial left ideal 𝐼 ⊆ 𝐵;
Proof. The equivalence (i) ⇔ (ii) follows from the Wedderburn–Artin theorem (also
proved in Main Theorem 5.4.4 and Theorem 6.4.11). The implications (i) ⇒ (iii) ⇒
(ii) and (i) ⇒ (iv) ⇒ (ii) are clear.
7.6.3. We showed in Lemma 7.2.17 that a simple algebra 𝐵 has a unique simple left
𝐵-module 𝐼 up to isomorphism, obtained as a minimal nonzero left ideal. If 𝐵 is
a quaternion algebra, this simple module 𝐼 can be readily identified using the above
proposition. If 𝐵 is a division ring, then necessarily 𝐼 = 𝐵. Otherwise, 𝐵 ' M2 (𝐹),
and then 𝐼 ' 𝐹 2 , and the map 𝐵 → End𝐹 (𝐼) given by left matrix multiplication is an
isomorphism.
7.7 The Skolem–Noether theorem

We conclude this chapter with a fundamental result that characterizes the automor-
phisms of a simple algebra—and much more.
Main Theorem 7.7.1 (Skolem–Noether). Let 𝐴, 𝐵 be simple 𝐹-algebras and suppose

that 𝐵 is central. Suppose that 𝑓 , 𝑔 : 𝐴 → 𝐵 are homomorphisms. Then there exists
𝛽 ∈ 𝐵× such that 𝑓 (𝛼) = 𝛽−1 𝑔(𝛼) 𝛽 for all 𝛼 ∈ 𝐴.
7.7. THE SKOLEM–NOETHER THEOREM 107
Proof. By Corollary 7.3.12, we have 𝐵 ' End𝐷 (𝑉) ' M𝑛 (𝐷 op ) where 𝑉 is a simple
𝐵-module and 𝐷 = End 𝐵 (𝑉) is a central 𝐹-algebra. Now the maps 𝑓 , 𝑔 give 𝑉 the
structure of an 𝐴-module in two ways. The 𝐴-module structure commutes with the
𝐷-module structure since 𝐵 ' End𝐷 (𝑉). So 𝑉 has two 𝐴 ⊗𝐹 𝐷-module structures via
𝑓 and 𝑔.
By Proposition 7.5.3, since 𝐷 is central over 𝐹, we have that 𝐴 ⊗𝐹 𝐷 is a simple 𝐹-
algebra. By Corollary 7.3.9 and a dimension count, the two 𝐴 ⊗𝐹 𝐷-module structures
on 𝑉 are isomorphic. Thus, there exists an isomorphism 𝛽 : 𝑉 → 𝑉 of 𝐴⊗𝐹 𝐷-modules;
i.e. 𝛽( 𝑓 (𝛼)𝑥) = 𝑔(𝛼) 𝛽(𝑥) for all 𝛼 ∈ 𝐴 and 𝑥 ∈ 𝑉, and 𝛽(𝛿𝑥) = 𝛿𝛽(𝑥) for all 𝛿 ∈ 𝐷
and 𝑥 ∈ 𝑉. We have 𝛽 ∈ End𝐷 (𝑉) ' 𝐵 and so we can write 𝛽 𝑓 (𝛼) 𝛽−1 = 𝑔(𝛼) for all
𝛼 ∈ 𝐴, as claimed.
The following corollaries are immediate consequences (special cases) of the

Skolem–Noether theorem.
Corollary 7.7.2. If 𝐴1 , 𝐴2 are simple 𝐹-subalgebras of a central simple 𝐹-algebra
∼ 𝐴 is an isomorphism of 𝐹-algebras, then 𝜙 is induced by an inner
𝐵 and 𝜙 : 𝐴1 −→ 2
automorphism of 𝐵.
Proof. Let 𝜄𝑖 : 𝐴𝑖 ↩→ 𝐵 be the natural inclusions, and apply Main Theorem 7.7.1 to
𝑓 := 𝜄1 and 𝑔 := 𝜄2 ◦ 𝜙: we conclude there exists 𝛽 ∈ 𝐵× such that 𝜄1 (𝛼) = 𝛼 =
𝛽−1 𝜄2 (𝜙(𝛼)) 𝛽 or equivalently 𝜙(𝛼) = 𝛽𝛼𝛽−1 for all 𝛼 ∈ 𝐴1 , as desired.
Corollary 7.7.3. If 𝐵 is a central simple 𝐹-algebra and 𝛼1 , 𝛼2 ∈ 𝐵× , then 𝛼1 , 𝛼2 have

the same irreducible minimal polynomial over 𝐹 if and only if there exists 𝛽 ∈ 𝐵× such
that 𝛼2 = 𝛽−1 𝛼1 𝛽.
Proof. The implication (⇐) is immediate. Conversely (⇒), let 𝐴𝑖 := 𝐹 [𝛼𝑖 ] '
𝐹 [𝑥]/( 𝑓𝑖 (𝑥)) where 𝑓𝑖 (𝑥) ∈ 𝐹 [𝑥] are minimal polynomials over 𝐹. Since these
polynomials are irreducible, 𝐴𝑖 is a field hence simple, so Corollary 7.7.2 gives the
result.
Corollary 7.7.4. The group of 𝐹-algebra automorphisms of a central simple algebra

𝐵 is Aut(𝐵) ' 𝐵× /𝐹 × .
Proof. Taking 𝐴 = 𝐵 in Main Theorem 7.7.1, we conclude that every automorphism

of 𝐵 as an 𝐹-algebra is inner, and an inner automorphism is trivial if and only if it is
conjugation by an element of the center 𝐹 × .
Example 7.7.5. By Corollary 7.7.4, we have a canonical isomorphism of groups
Aut(M𝑛 (𝐹)) GL𝑛 (𝐹)/𝐹 × = PGL𝑛 (𝐹).
As a final application, we extend the splitting criterion of Main Theorem 5.4.4(i)

⇔ (vi) to detect isomorphism classes of quaternion algebras.
Corollary 7.7.6. Let 𝐾 ⊇ 𝐹 be a separable quadratic 𝐹-algebra, and let 𝑏, 𝑏 0 ∈ 𝐹 × .
Then
𝐾, 𝑏 0

𝐾, 𝑏
' ⇔ 𝑏/𝑏 0 ∈ Nm𝐾 |𝐹 (𝐾 × ).
𝐹 𝐹
Taking 𝑏 0 = 1, we recover the previous splitting criteria.
Proof. For the implication (⇐), if 𝑏 0/𝑏 = Nm𝐾 |𝐹 (𝛼) with 𝛼 ∈ 𝐾 × , then an isomor-
phism is furnished as left 𝐾-vector spaces by sending 𝑗 ↦→ 𝛼 𝑗.
∼ 𝐵 0 := (𝐾, 𝑏 0 | 𝐹) be an isomorphism
For the implication (⇒), let 𝜙 : (𝐾, 𝑏 | 𝐹) −→
of 𝐹-algebras. If 𝐾 ' 𝐹 × 𝐹 is not a field, then Nm𝐾 |𝐹 (𝐾 × ) = 𝐹 × and the result holds.
So suppose 𝐾 is a field. Then 𝜙(𝐾) ⊆ 𝐵 0 isomorphic to 𝐾 as an 𝐹-algebra, but need
not be the designated one in 𝐵’; however, by the Skolem–Noether theorem, we may
postcompose 𝜙 with an automorphism that sends 𝜙(𝐾) to the designated one, i.e., we
may suppose that 𝜙 is a 𝐾-linear map (taking the algebras as left 𝐾-vector spaces).
Let 𝜙( 𝑗) = 𝛼 + 𝛽 𝑗 0 with 𝛼, 𝛽 ∈ 𝐾. Then
{0} = trd(𝐾 𝑗) = trd(𝜙(𝐾 𝑗)) = trd(𝐾 𝜙( 𝑗)) = trd(𝐾𝛼)
and thus 𝛼 = 0 since 𝐾 is separable. Consequently,
−𝑏 = nrd( 𝑗) = nrd(𝜙( 𝑗)) = − Nm𝐾 |𝐹 (𝛽)𝑏 0
and so 𝑏/𝑏 0 = Nm𝐾 |𝐹 (𝛽) as desired.
In the remainder of this section, we prove an important consequence of the Skolem–

Noether theorem that compares centralizers of subalgebras to dimensions.
Definition 7.7.7. Let 𝐴 be an 𝐹-subalgebra of 𝐵. Let
𝐶 𝐵 ( 𝐴) := {𝛽 ∈ 𝐵 : 𝛼𝛽 = 𝛽𝛼 for all 𝛼 ∈ 𝐴}
be the centralizer of 𝐴 in 𝐵.
The centralizer 𝐶 𝐵 ( 𝐴) is an 𝐹-subalgebra of 𝐵.
Proposition 7.7.8. Let 𝐵 be a central simple 𝐹-algebra and let 𝐴 ⊆ 𝐵 a simple

𝐹-subalgebra. Then the following statements hold:
(a) 𝐶 𝐵 ( 𝐴) is a simple 𝐹-algebra.

(b) dim𝐹 𝐵 = dim𝐹 𝐴 · dim𝐹 𝐶 𝐵 ( 𝐴).
(c) 𝐶 𝐵 (𝐶 𝐵 ( 𝐴)) = 𝐴.
Part (c) of this proposition is called the double centralizer property.
Proof. First, part (a). We interpret the centralizer as arising from certain kinds of
endomorphisms. We have that 𝐵 is a left 𝐴⊗ 𝐵op module by the action (𝛼⊗ 𝛽) · 𝜇 = 𝛼𝜇𝛽
for 𝛼 ⊗ 𝛽 ∈ 𝐴 ⊗ 𝐵op and 𝜇 ∈ 𝐵. We claim that
𝐶 𝐵 ( 𝐴) = End 𝐴⊗𝐵op (𝐵). (7.7.9)
Any 𝜙 ∈ End 𝐴⊗𝐵op (𝐵) is left multiplication by an element of 𝐵: if 𝛾 = 𝜙(1), then

𝜙(𝜇) = 𝜙(1)𝜇 = 𝛾𝜇 by 1 ⊗ 𝐵op -linearity. Now the equality
𝛾𝛼 = 𝜙(𝛼) = 𝛼𝜙(1) = 𝛼𝛾
7.7. THE SKOLEM–NOETHER THEOREM 109
shows that multiplication by 𝛾 is 𝐴 ⊗ 1-linear if and only if 𝛾 ∈ 𝐶 𝐵 ( 𝐴), proving (7.7.9).

By Proposition 7.5.3, the algebra 𝐴 ⊗ 𝐵op is simple. By the Wedderburn–Artin
theorem, 𝐴 ⊗ 𝐵op ' M𝑛 (𝐷) for some 𝑛 ≥ 1 and division 𝐹-algebra 𝐷. Since M𝑛 (𝐷)
is simple, its unique simple left 𝐷-module is 𝑉 = 𝐷 𝑛 , and EndM𝑛 (𝐷) (𝑉) ' 𝐷 op . In
particular, 𝐵 ' 𝑉 𝑟 for some 𝑟 ≥ 1 as an 𝐴 ⊗ 𝐵op -module. So
𝐶 𝐵 ( 𝐴) = End 𝐴⊗𝐵op (𝐵) ' EndM𝑛 (𝐷) (𝑉 𝑟 ) ' M𝑟 (EndM𝑛 (𝐷) (𝑉)) ' M𝑟 (𝐷 op ).
Thus 𝐶 𝐵 ( 𝐴) is simple.
For part (b),
dim𝐹 𝐶 𝐵 ( 𝐴) = dim𝐹 M𝑟 (𝐷 op ) = 𝑟 2 dim𝐹 𝐷
and
dim𝐹 ( 𝐴 ⊗ 𝐵op ) = dim𝐹 𝐴 · dim𝐹 𝐵 = 𝑛2 dim𝐹 𝐷
and finally
dim𝐹 𝐵 = dim𝐹 𝑉 𝑟 = 𝑟 dim𝐹 𝐷 𝑛 = 𝑟𝑛 dim𝐹 𝐷;
putting these together gives dim𝐹 𝐴 · dim𝐹 𝐶 𝐵 ( 𝐴) = 𝑟𝑛 dim𝐹 𝐷 = dim𝐹 𝐵.
Finally, part (c) follows from (a) and (b):
dim𝐹 𝐵 = dim𝐹 𝐶 𝐵 ( 𝐴) · dim𝐹 𝐶 𝐵 (𝐶 𝐵 ( 𝐴)) = dim𝐹 𝐴 · dim𝐹 𝐶 𝐵 ( 𝐴)
so dim𝐹 𝐴 = dim𝐹 𝐶 𝐵 (𝐶 𝐵 ( 𝐴)) and 𝐴 ⊆ 𝐶 𝐵 (𝐶 𝐵 ( 𝐴)), therefore equality holds.
Example 7.7.10. We always have the two extremes 𝐴 = 𝐹 and 𝐴 = 𝐵, with 𝐶 𝐵 (𝐹) = 𝐵
and 𝐶 𝐵 (𝐵) = 𝐹, accordingly.
We note the following structurally crucial corollary of Proposition 7.7.8.
Corollary 7.7.11. Let 𝐵 be a central simple 𝐹-algebra and let 𝐾 be a maximal subfield.
Then [𝐵 : 𝐹] = [𝐾 : 𝐹] 2 .
Proof. If 𝐾 is maximal, then 𝐶 𝐵 (𝐾) = 𝐾 and [𝐵 : 𝐹] = [𝐾 : 𝐹] 2 .
Corollary 7.7.11 generalizes the comparatively easier statement for quaternion

algebras: the maximal subfields of a quaternion algebra are quadratic. Returning now
to quaternion algebras, we conclude with a nice package of consequences of the above
results concerning embeddings of quadratic fields into quaternion algebras.
7.7.12. Let 𝐵 be a quaternion algebra over 𝐹 and let 𝐾 ⊆ 𝐵 be a quadratic separable
𝐹-subalgebra. Then the set of all embeddings of 𝐾 in 𝐵 is naturally identified with the
set 𝐾 × \𝐵× , as follows.
By the Skolem–Noether theorem (Corollary 7.1.5, and Exercise 7.10 for the case
𝐾 ' 𝐹 × 𝐹), if 𝜙 : 𝐾 ↩→ 𝐵 is another embedding, then there exists 𝛽 ∈ 𝐵× such that
𝜙(𝛼) = 𝛽−1 𝛼𝛽 for all 𝛼 ∈ 𝐾, and conversely. Such a conjugate embedding is the
identity if and only if 𝛽 centralizes 𝐾. By Corollary 4.4.5, and Corollary 6.4.2 for
characteristic 2, the centralizer of 𝐾 × in 𝐵× is 𝐾 × . Therefore, the set of embeddings
of 𝐾 in 𝐵 is naturally identified with the set 𝐾 × \𝐵× , with 𝐾 × acting on the left.
7.8 Reduced trace and norm, universality

In this last section, we consider notions of reduced trace and reduced norm in the
context of semisimple algebras. Thr
7.8.1. Let 𝐵 be a (finite-dimensional) central simple algebra over 𝐹, and let 𝐹 sep denote
a separable closure of 𝐹. By Exercise 7.23, we have an 𝐹-algebra homomorphism
𝜙 : 𝐵 ⊗𝐹 𝐹 sep ' M𝑛 (𝐹 sep )
for some 𝑛 ≥ 1. By the Skolem–Noether theorem (Main Theorem 7.7.1), for any
other isomorphism 𝜙 0 : 𝐵 ⊗𝐹 𝐹 sep ' M𝑛 (𝐹 sep ), there exists 𝑀 ∈ GL𝑛 (𝐹 sep ) such that
𝜙 0 (𝛼) = 𝑀 𝜙(𝛼) 𝑀 −1 , so the characteristic polynomial of an element of 𝐵 ⊗𝐹 𝐹 sep is
independent of the choice of 𝜄. In particular, from the canonical embedding 𝜄 : 𝐵 ↩→
𝐵 ⊗𝐹 𝐹 sep by 𝛼 ↦→ 𝛼 ⊗ 1, we define the reduced characteristic polynomial of 𝛼 ∈ 𝐵
to be the characteristic polynomial of (𝜙𝜄) (𝛼) as an element of 𝐹 sep [𝑇] and similarly
the reduced trace and reduced norm of 𝛼 to be the trace and determinant of (𝜙𝜄) (𝛼)
as elements of 𝐹 sep .
In fact, the reduced characteristic polynomial descends to 𝐹, as follows. The
absolute Galois group Gal𝐹 := Gal(𝐹 sep | 𝐹) acts on 𝐵 ⊗𝐹 𝐹 sep ' M𝑛 (𝐹 sep ) by
𝜎(𝛼 ⊗ 𝑎) = 𝛼 ⊗ 𝜎(𝑎)
for 𝜎 ∈ Gal𝐹 , 𝛼 ∈ 𝐵, and 𝑎 ∈ 𝐹 sep . Let 𝜎 ∈ Gal𝐹 . Since 𝜎(𝛼 ⊗ 1) = 𝛼 ⊗ 𝜎(1) =
𝛼 ⊗ 1, the reduced characteristic polynomials of 𝜄(𝛼) and 𝜎(𝜄(𝛼)) are the same. By
comparison (see e.g. Reiner [Rei2003, Theorem 9.3]), if
𝑓 (𝛼; 𝑇) = det(𝑇 − 𝜄(𝛼)) = 𝑇 𝑛 + 𝑎 𝑛−1𝑇 𝑛−1 + · · · + 𝑎 0
is the reduced characteristic polynomial of 𝜄(𝛼), then the reduced characteristic poly-
nomial of (𝜎(𝜄)) (𝛼) is
𝜎( 𝑓 ) (𝛼; 𝑇) = det(𝑇 − 𝜎(𝜄(𝛼))) = 𝑇 𝑛 + 𝜎(𝑎 𝑛−1 )𝑇 𝑛−1 + · · · + 𝜎(𝑎 0 ).
And then since 𝑓 (𝛼; 𝑇) = 𝜎( 𝑓 ) (𝛼; 𝑇) for all 𝜎 ∈ Gal𝐹 , by the fundamental theorem
of Galois theory, 𝑓 (𝛼; 𝑇) ∈ 𝐹 [𝑇]. Therefore, the reduced norm and reduced trace also
belong to 𝐹.
Alternatively, we may argue as follows. The characteristic polynomial of left
multiplication by 𝛼 on 𝐵 is the same as left multiplication by (𝜙𝜄) (𝛼) on M𝑛 (𝐹 sep )
(by extension of basis), and the latter is the 𝑛th power of the reduced characteristic
polynomial by Exercise 3.13. Finally, if 𝑓 (𝑇) ∈ 𝐹 sep [𝑇] has 𝑓 (𝑇) 𝑛 ∈ 𝐹 [𝑇] then in
fact 𝑓 (𝑇) ∈ 𝐹 [𝑇]: see Exercise 7.24.
These definitions extend to a general semisimple algebra over 𝐹, but to do so it
is convenient to give an alternate approach that avoids going to the separable closure
and works in even more generality using universal elements; for more, see Garibaldi
[Gar2004].
Let 𝐵 be a (finite-dimensional) 𝐹-algebra with 𝑛 := dim𝐹 𝐵, and choose a basis
𝑒 1 , . . . , 𝑒 𝑛 for 𝐵. Let 𝐹 (𝑥1 , . . . , 𝑥 𝑛 ) be a pure transcendental field extension of tran-
scendence degree 𝑛, and let 𝜉 := 𝑥1 𝑒 1 + · · · + 𝑥 𝑛 𝑒 𝑛 ∈ 𝐵 ⊗𝐹 𝐹 (𝑥1 , . . . , 𝑥 𝑛 ); we call 𝜉
the universal element of 𝐵 in the given basis.
7.8. REDUCED TRACE AND NORM, UNIVERSALITY 111
Definition 7.8.2. The universal minimal polynomial of 𝐵 (in the basis 𝑒 1 , . . . , 𝑒 𝑛 ) is

the minimal polynomial 𝑚 𝐵 (𝜉; 𝑇) of 𝜉 over 𝐹 (𝑥1 , . . . , 𝑥 𝑛 ).
For 𝛼 = 𝑎 1 𝑒 1 + · · · + 𝑎 𝑛 𝑒 𝑛 ∈ 𝐵 (with 𝑎 𝑖 ∈ 𝐹), the polynomial obtained from
𝑚 𝐵 (𝜉; 𝑇) by the substitution 𝑥𝑖 ← 𝑎 𝑖 is called the specialization of 𝑚 𝐵 (𝜉; 𝑇) at 𝛼.
The following example will hopefully illustrate the role of this notion.

𝑎, 𝑏
Example 7.8.3. For char 𝐹 ≠ 2 and 𝐵 = , in the basis 1, 𝑖, 𝑗, 𝑖 𝑗 we have
𝐹
𝜉 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 (substituting 𝑡, 𝑥, 𝑦, 𝑧 for 𝑥1 , . . . , 𝑥 4 ). We claim that the universal
minimal polynomial is
𝑚 𝐵 (𝜉; 𝑇) = 𝑇 2 − 2𝑡𝑇 + (𝑡 2 − 𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧 2 ).
Indeed,
that 𝜉 satisfies 𝑚 𝐵 (𝜉; 𝜉) = 0 by considering 𝜉 ∈ 𝐵 ⊗𝐹 𝐹 (𝑡, 𝑥, 𝑦, 𝑧) =
we verify
𝑎, 𝑏
and computing that trd(𝜉) = 2𝑡 and nrd(𝜉) = 𝑡 2 − 𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧 2 ;
𝐹 (𝑡, 𝑥, 𝑦, 𝑧)
and this polynomial is minimal because 𝜉 ∉ 𝐹 (𝑡, 𝑥, 𝑦, 𝑧) does not satisfy a polynomial
of degree 1 over 𝐹 (𝑡, 𝑥, 𝑦, 𝑧).
7.8.4. If 𝐵 ' 𝐵1 × · · · × 𝐵𝑟 , then in a basis for 𝐵 obtained from the union of bases for
the factors 𝐵𝑖 with universal elements 𝜉𝑖 , we have
𝑚 𝐵 (𝜉; 𝑇) = 𝑚 𝐵1 (𝜉1 ; 𝑇) · · · 𝑚 𝐵𝑟 (𝜉𝑟 ; 𝑇).
In the proofs that follow, we abbreviate by using multi-index notation, e.g. writing
𝐹 [𝑥] := 𝐹 [𝑥 1 , . . . , 𝑥 𝑛 ].
Lemma 7.8.5. We have 𝑚 𝐵 (𝜉; 𝑇) ∈ 𝐹 [𝑥1 , . . . , 𝑥 𝑛 ] [𝑇], i.e., the universal minimal
polynomial has coefficients in 𝐹 [𝑥1 , . . . , 𝑥 𝑛 ].
Proof. For part (a), we consider the map given by left multiplication by 𝜉 on 𝐵⊗𝐹 𝐹 (𝑥).
In the basis 𝑒 1 , . . . , 𝑒 𝑛 , almost by construction we find that the matrix of this map has
coefficients in 𝐹 [𝑥] (it is the matrix of linear forms obtained from left multiplication
by 𝑒 𝑖 ). We conclude that 𝜉 satisfies the characteristic polynomial of this matrix,
which is a monic polynomial with coefficients in 𝐹 [𝑥]. Since 𝑚 𝐵 (𝜉; 𝑇) divides this
polynomial (over 𝐹 (𝑥)) by minimality, by Gauss’s lemma we conclude that 𝑚 𝐵 (𝜉; 𝑇) ∈
𝐹 [𝑥] [𝑇].
Proposition 7.8.6. For all 𝛼 ∈ 𝐵, the specialization of 𝑚 𝐵 (𝜉; 𝑇) at 𝛼 is independent of

the choice of basis 𝑒 1 , . . . , 𝑒 𝑛 and is satisfied by the element 𝛼. Moreover, if 𝜙 ∈ Aut(𝐵)
and 𝛼 ∈ 𝐵, then 𝛼 and 𝜙(𝛼) have the same specialized polynomials.
In view of Proposition 7.8.6, we write 𝑚 𝐵 (𝛼; 𝑇) for the specialization of 𝑚 𝐵 (𝜉; 𝑇)

at 𝛼 ∈ 𝐵; from it, we conclude that 𝑚 𝐵 (𝛼; 𝛼) = 0.
Proof. Since 𝑚 𝐵 (𝜉; 𝜉) = 0, by specialization we obtain 𝑚 𝐵 (𝛼; 𝛼) = 0. For the

independence of basis, let 𝑒 10 , . . . , 𝑒 𝑛0 be another 𝐹-basis and 𝜉 0 Í
the corresponding
universal element. Writing 𝑒 𝑖 in the basis 𝑒 𝑖0 allows us to write 𝜉 = 𝑖=1 𝑛
ℓ𝑖 (𝑥)𝑒 𝑖0 where
ℓ𝑖 (𝑥) ∈ 𝐹 [𝑥] are linear forms; moreover, writing 𝛼 = 𝑖 𝑎 𝑖0 𝑒 𝑖0 we have ℓ𝑖 (𝑎) = 𝑎 𝑖0 .

Í
The map 𝑥 𝑖 ↦→ ℓ𝑖 (𝑥) extends to an 𝐹-algebra automorphism 𝜙 of 𝐹 [𝑥] (repeat the
construction with the inverse, and compose) with 𝜙(𝑐) (𝑎) = 𝑐(𝑎 0) for all 𝑐(𝑥) ∈ 𝐹 [𝑥].
We let 𝜙 act on polynomials over 𝐹 [𝑥] by acting on the coefficients; by uniqueness of
minimal polynomials, we have 𝜙(𝑚 𝐵 (𝜉; 𝑇)) = 𝑚 𝐵 (𝜉 0; 𝑇). Therefore, looking at each
coefficient, specializing 𝜙(𝑚 𝐵 (𝜉; 𝑇)) at 𝑎 is the same as specializing 𝑚 𝐵 (𝜉 0; 𝑇) at 𝑎 0,
as claimed.
The second sentence follows by the same argument, as from 𝜙 ∈ Aut(𝐵) we have
a new basis 𝑒 𝑖0 := 𝜙(𝑒 𝑖 ) so the specializations again agree. (This argument replaces
the use of the Skolem–Noether theorem in the special case where 𝐵 is a central simple
algebra.)
Lemma 7.8.7. For any field extension 𝐾 ⊇ 𝐹, we have 𝑚 𝐵 ⊗𝐹 𝐾 (𝜉; 𝑇) = 𝑚 𝐾 (𝜉; 𝑇).
Proof. First, because an 𝐹-basis for 𝐵 is a 𝐾-basis for 𝐵 ⊗𝐹 𝐾, the element 𝜉 (as
the universal element of 𝐵), also serves as a universal element of 𝐵 ⊗𝐹 𝐾. Since
𝐾 (𝑥1 , . . . , 𝑥 𝑛 ) ⊆ 𝐹 (𝑥1 , . . . , 𝑥 𝑛 ), by minimality we have 𝑚 𝐵𝐾 (𝜉; 𝑇) | 𝑚 𝐵 (𝜉; 𝑇). Con-
versely, let 𝐹 (𝑥) [𝜉] ⊆ 𝐵 ⊗𝐹 𝐹 (𝑥) be the subalgebra generated by 𝜉 over 𝐹 (𝑥); then
𝐹 (𝑥) [𝜉] ' 𝐹 (𝑥) [𝑇]/(𝑚 𝐵 (𝜉; 𝑇)). Tensoring with 𝐾 gives
𝐾 (𝑥) [𝜉] ' 𝐾 (𝑥) [𝑇]/(𝑚 𝐵 (𝜉; 𝑇))
as the subalgebra of (𝐵 ⊗𝐹 𝐾) ⊗𝐾 𝐾 (𝑥) generated by 𝜉. Thus 𝑚 𝐵 (𝜉; 𝑇) | 𝑚 𝐵𝐾 (𝜉; 𝑇),

so equality holds.
We conclude by relating this construction to more familiar polynomials.
Lemma 7.8.8. Let 𝛼 ∈ 𝐵. Then the following statements hold.
(a) If 𝐵 = 𝐾 ⊇ 𝐹 is a separable field extension, then 𝑚 𝐾 (𝛼; 𝑇) is the characteristic

polynomial of left multiplication by 𝛼.
(b) If 𝐵 is a central simple 𝐹-algebra, then 𝑚 𝐵 (𝛼; 𝑇) is the reduced characteristic
polynomial of 𝛼.
Proof. For (a), we recall (as in the proof of Lemma 7.8.5 that 𝜉 satisfies the character-
istic polynomial of left multiplication on 𝐾, a polynomial of degree 𝑛 = [𝐾 : 𝐹]; on
the other hand, choosing a primitive element 𝛼, we see the specialization 𝑚 𝐾 (𝛼; 𝑇) is
satisfied by 𝛼 so has degree at least 𝑛, so equality holds and 𝑚 𝐾 (𝛼; 𝑇) is the charac-
teristic polynomial. Therefore the universal minimal polynomial is the characteristic
polynomial, and hence the same is true under every specialization.
For (b), it suffices to prove this when 𝐹 = 𝐹 al is algebraically closed, in which case
𝐵 ' M𝑛 (𝐹); by Proposition 7.8.6, we may assume 𝐵 = M𝑛 (𝐹). By 7.8.1, we want
to show that 𝑚 𝐵 (𝛼; 𝑇) for 𝛼 ∈ M𝑛 (𝐹) is the usual characteristic polynomial. But the
universal element (in a basis of matrix units, or any basis) satisfies its characteristic
polynomial of degree 𝑛, and a nilpotent matrix (with 1s just above the diagonal) has
minimal polynomial 𝑇 𝑛 , so we conclude as in the previous paragraph.
In light of the above, we may make the following definition.

7.9. SEPARABLE ALGEBRAS 113
Definition 7.8.9. Let 𝐵 be a semisimple 𝐹-algebra. For 𝛼 ∈ 𝐵, the reduced charac-

teristic polynomial 𝑓 (𝛼; 𝑇) = 𝑇 𝑛 −𝑐 1𝑇 𝑛−1 +· · ·+ (−1) 𝑛 𝑐 𝑛 ∈ 𝐹 [𝑇] is the specialization
of the universal minimal polynomial 𝑚 𝐵 (𝜉; 𝑇), and the reduced trace and reduced
norm are the coefficients 𝑐 1 , 𝑐 𝑛 , respectively.
Example 7.8.10. For a semisimple algebra 𝐵 ' 𝐵1 × · · · × 𝐵𝑟 , with each 𝐵𝑖 central

simple, we find that the reduced characteristic polynomial is just the product of the
reduced characteristic polynomials on each simple direct factor 𝐵𝑖 ; this is well-defined
(again) by the uniqueness statement in the Wedderburn–Artin theorem (Main Theorem
7.3.10).
Proposition 7.8.11. Let 𝐵 be semisimple. Then the reduced trace trd : 𝐵 → 𝐹 is

𝐹-linear and satisfies trd(𝛼𝛽) = trd(𝛽𝛼) for all 𝛼, 𝛽 ∈ 𝐵; and the reduced norm
nrd : 𝐵 → 𝐹 is multiplicative, satisfying nrd(𝛼𝛽) = nrd(𝛼) nrd(𝛽).
Proof. Consider (again) 𝑉 := 𝐹 (𝑥) [𝜉] ⊆ 𝐵 ⊗𝐹 𝐹 (𝑥) the subalgebra generated over
𝐹 (𝑥) by 𝜉; then 𝜉 acts on 𝑉 ' 𝐹 (𝑥) [𝑇]/(𝑚 𝐵 (𝜉; 𝑇)) by left multiplication with
characteristic polynomial 𝑚 𝐵 (𝜉; 𝑇). It follows that the reduced trace and reduced
norm are the usual trace and determinant in this representation, so the announced
properties follow on specialization.
Remark 7.8.12. It is also possible to define the reduced characteristic polynomial on

a semisimple algebra 𝐵 by writing 𝐵 ' 𝐵1 × · · · × 𝐵𝑟 as a product of simple algebras;
for details, see Reiner [Rei2003, §9].
7.9 Separable algebras

For a (finite-dimensional) 𝐹-algebra, the notions of simple and semisimple are sensitive
to the base field 𝐹 in the sense that these properties need not hold after extending the
base field. Indeed, let 𝐾 ⊇ 𝐹 be a finite extension of fields, so 𝐾 is a simple 𝐹-
algebra. Then 𝐾 ⊗𝐹 𝐹 al is simple only when 𝐾 = 𝐹 and is semisimple if and only if
𝐾 ⊗𝐹 𝐹 al ' 𝐹 al × · · · × 𝐹 al , i.e., 𝐾 is separable over 𝐹.
It is important to have a notion which is stable under base change, as follows.
For further reference, see Drozd–Kirichenko [DK94, §6], Curtis–Reiner [CR81, §7],
Reiner [Rei2003, §7c], or Pierce [Pie82, Chapter 10].
Definition 7.9.1. Let 𝐵 be a finite-dimensional 𝐹-algebra. We say that 𝐵 is a separable

𝐹-algebra if 𝐵 is semisimple and 𝑍 (𝐵) is a separable 𝐹-algebra.
In particular, a separable algebra over a field 𝐹 with char 𝐹 = 0 is just a semisimple

algebra.
7.9.2. For a semisimple algebra 𝐵 ' M𝑛1 (𝐷 1 ) × · · · × M𝑛𝑟 (𝐷 𝑟 ), by Example 7.5.2

we have 𝑍 (𝐵) ' 𝑍 (𝐷 1 ) × · · · × 𝑍 (𝐷 𝑟 ), and 𝐵 is separable if and only if 𝑍 (𝐷 𝑖 ) is
separable for each 𝑖 = 1, . . . , 𝑟.
Lemma 7.9.3. A finite-dimensional simple 𝐹-algebra is a separable algebra over its

center 𝐾.
Proof. The center of 𝐵 is a field 𝐾 = 𝑍 (𝐵) and as a 𝐾-algebra, the center 𝑍 (𝐵) =
𝐾 is certainly separable over 𝐾. (Or use Proposition 7.5.3 and Theorem 7.9.4(iii)
below.)
The notion of separability in this context is quite robust.

Theorem 7.9.4. Let 𝐵 be a finite-dimensional 𝐹-algebra. Then the following are
equivalent:
(i) 𝐵 is separable;
(ii) There exists a finite separable field extension 𝐾 of 𝐹 such that 𝐵 ⊗𝐹 𝐾 '
𝑀𝑛1 (𝐾) × · · · × 𝑀𝑛𝑟 (𝐾) for integers 𝑛1 , . . . , 𝑛𝑟 ≥ 1;
(iii) For every extension 𝐾 ⊇ 𝐹 of fields, the 𝐾-algebra 𝐵 ⊗𝐹 𝐾 is semisimple;
(iv) 𝐵 is semisimple and the bilinear form
𝐵×𝐵→𝐹
(𝛼, 𝛽) ↦→ trd(𝛼𝛽)
is nondegenerate.
Moreover, if char 𝐹 = 0, then these are further equivalent to:
(v) The bilinear form (𝛼, 𝛽) ↦→ Tr 𝐵 |𝐹 (𝛼𝛽) is nondegenerate.
A separable 𝐹-algebra is sometimes called absolutely semisimple, in view of
Theorem 7.9.4(iii).
Proof. First we prove (i) ⇒ (ii). Let 𝐵𝑖 be a simple component of 𝐵; then 𝑍 (𝐵𝑖 ) is
separable over 𝐹. Let 𝐾𝑖 ⊇ 𝐹 be a separable field extension containing 𝑍 (𝐵𝑖 ) that
splits 𝐵𝑖 , so 𝐵𝑖 ⊗ 𝑍 (𝐵𝑖 ) 𝐾𝑖 ' M𝑛𝑖 (𝐾𝑖 ). Let 𝐾 be the compositum of the fields 𝐾𝑖 . Then
𝐾 is separable, and
𝐵𝑖 ⊗𝐹 𝐾 ' M𝑛𝑖 (𝑍 (𝐵𝑖 ) ⊗𝐹 𝐾) ' M𝑛𝑖 (𝐾) × · · · × M𝑛𝑖 (𝐾)
the number of copies equal to [𝑍 (𝐵𝑖 ) : 𝐹]. Î

Next we prove (ii) ⇒ (iii). Suppose 𝐵 ⊗𝐹 𝐾 ' 𝑖 M𝑛𝑖 (𝐾) and let 𝐿 ⊇ 𝐹 be an
extension of fields. Let 𝑀 = 𝐾 𝐿. On the one hand, 𝐵 ⊗𝐹 𝑀 ' (𝐵 ⊗𝐹 𝐾) ⊗𝐾 𝑀 '
Î
𝑖 M𝑛𝑖 (𝑀), so rad 𝐵 ⊗𝐹 𝑀 = {0}; on the other hand, 𝐵 ⊗𝐹 𝑀 ' (𝐵 ⊗𝐹 𝐿) ⊗ 𝐿 𝑀 and
rad(𝐵 ⊗𝐹 𝐿) ⊆ rad(𝐵 ⊗𝐹 𝐿) ⊗ 𝐿 𝑀 = {0}, so 𝐵 ⊗𝐹 𝐿 is semisimple.
For the implication (iii) ⇒ (i), suppose 𝐵 is not separable, and we show that there
exists 𝐾 ⊇ 𝐹 such that 𝐵 ⊗𝐹 𝐾 is not semisimple. If 𝐵 is not semisimple over 𝐹, we
can just take 𝐹 = 𝐾. Otherwise, 𝑍 (𝐵) is not separable as an 𝐹-algebra, and there is a
component of 𝑍 (𝐵) which is an inseparable field extension 𝐾. Then 𝐵 ⊗𝐹 𝐾 contains a
nonzero nilpotent element in its center and this element generates a nonzero nilpotent
ideal, so rad(𝐵 ⊗𝐹 𝐾) ≠ {0} and 𝐵 ⊗𝐹 𝐾 is not semisimple.
The implication (iii) ⇒ (iv) holds for the following reason. We have 𝐵 ⊗𝐹 𝐹 al '
M𝑛1 (𝐹 al ) × · · · × M𝑛𝑟 (𝐹 al ), and the reduced trace pairing on each matrix ring factor is
nondegenerate so the whole pairing is nondegenerate. By linear algebra we conclude
that the bilinear form on 𝐵 is nondegenerate.
The implication (iv) ⇒ (i) holds with char 𝐹 arbitrary: if 𝜖 ∈ rad 𝐵 then 𝛼𝜖 ∈ rad 𝐵
is nilpotent and trd(𝛼𝜖) = 0 for all 𝛼 ∈ 𝐵, and by nondegeneracy 𝜖 = 0.
The final equivalence (iv) ⇔ (v) follows when char 𝐹 = 0 since the algebra trace
pairing on each simple factor is a scalar multiple of the reduced trace pairing.
Exercises
Throughout the exercises, let 𝐹 be a field.

𝑎, 𝑏
I 1. Prove that a quaternion algebra 𝐵 = with char 𝐹 ≠ 2 is simple by a direct
𝐹
calculation, as follows.
(a) Let 𝐼 be a nontrivial two-sided ideal, and let 𝜖 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 ∈ 𝐼. By
considering 𝑖𝜖 − 𝜖𝑖, show that 𝑡 + 𝑥𝑖 ∈ 𝐼.
(b) Arguing symmetrically and taking a linear combination, show that 𝑡 ∈ 𝐼,
and conclude that 𝑡 = 0, whence 𝑥 = 𝑦 = 𝑧 = 0.

𝑎, 𝑏
Modify this argument to show that an algebra 𝐵 = is simple when
𝐹
char 𝐹 = 2. [We proved these statements without separating into cases in
7.2.11.]
2. Let 𝐵 be a quaternion algebra over 𝐹, and let 𝐾 ⊆ 𝐵 be an 𝐹-subalgebra that is
commutative. Show that dim𝐹 𝐾 ≤ 2.
3. Let 𝐵 be a quaternion algebra. Exhibit an explicit isomorphism
∼ M (𝐹).
𝐵 ⊗𝐹 𝐵 −→ 4
[Hint: see Exercise 2.11.]

𝑎 𝑏
I 4. Let 𝐵 = : 𝑎, 𝑏, 𝑐 ∈ 𝐹 ⊆ M2 (𝐹), and 𝑉 = 𝐹 2 be the left 𝐵-module
0 𝑐
of column vectors. Show that 𝑉 is indecomposable, but not simple, as a left
𝐵-module (cf. Example 7.2.5).
I 5. This exercise proves basic but important facts about two-sided ideals in matrix
algebras using matrix units.
(a) Let 𝐷 be a division 𝐹-algebra. Prove that M𝑛 (𝐷) is a simple 𝐹-algebra
with center 𝑍 (𝐷) for all 𝑛 ≥ 1. [Hint: Let 𝐸 𝑖 𝑗 be the matrix with 1 in the
𝑖 𝑗th entry and zeros in all other entries. Show that 𝐸 𝑘𝑖 𝑀 𝐸 𝑗ℓ = 𝑚 𝑖 𝑗 𝐸 𝑘ℓ
where 𝑚 𝑖 𝑗 is the 𝑖 𝑗th entry of 𝑀.]
(b) More generally, let 𝑅 be a ring (associative with 1, but potentially non-
commutative). Show that 𝑍 (M𝑛 (𝑅)) = 𝑍 (𝑅) and that any two-sided ideal
of M𝑛 (𝑅) is of the form M𝑛 (𝐼) ⊆ M𝑛 (𝑅) where 𝐼 is a two-sided ideal of
𝑅.
I 6. Let 𝐹 be a field, let 𝐵 a simple algebra, and let 𝐼 be a left 𝐵-module with
dim𝐹 𝐼 = dim𝐹 𝐵. Show that 𝐼 is isomorphic to 𝐵 as a left 𝐵-module, i.e., there
exists 𝛼 ∈ 𝐼 such 𝐼 = 𝐵𝛼.
7. In this exercise, we consider extensions of the Skolem–Noether theorem.

(a) Let 𝐵 be a quaternion algebra over 𝐹 and let 𝐾1 , 𝐾2 ⊂ 𝐵 be 𝐹-subalgebras
∼ 𝐾 is an isomorphism
(not necessarily subfields). Suppose that 𝜙 : 𝐾1 −→ 2
of 𝐹-algebras. Show that 𝜙 lifts to an inner automorphism of 𝐵. [Hint:
repeat the proof of Corollary 7.1.5.]
(b) Show by example that Corollary 7.7.3 is false if the minimal polynomials
are not supposed to be irreducible. In particular, provide an example
of isomorphic algebras 𝐾1 , 𝐾2 ⊆ 𝐵 that are not isomorphic by an inner
automorphism of 𝐵.
8. Let 𝐵 be a quaternion algebra over 𝐹, and let 𝐾 ⊆ 𝐵 be a separable, quadratic
𝐹-subalgebra. Show that there exists 𝑏 ∈ 𝐹 × such that 𝐵 ' (𝐾 | 𝑏). [Hint: lift
the standard involution on 𝐾 via the Skolem–Noether theorem.]
I 9. Let 𝐵 be a finite-dimensional 𝐹-algebra. Show that if 𝛼 ∈ rad 𝐵, then 1−𝛽𝛼 ∈ 𝐵×
for all 𝛽 ∈ 𝐵. [Hint: if 1 − 𝛽𝛼 is not left invertible then it belongs to a maximal
left ideal; left invertible implies invertible.]
I 10. Extend Corollary 7.1.5 to the case where 𝐾 = 𝐹 × 𝐹 as follows: show that
if 𝐾1 , 𝐾2 ⊆ 𝐵 are 𝐹-subalgebras with 𝐾1 ' 𝐹 × 𝐹, and 𝜙 : 𝐾1 −→ ∼ 𝐾 is an
2
isomorphism of 𝐹-algebras, then 𝜙 lifts to an inner automorphism of 𝐵.
11. Let 𝑛 ∈ Z ≥2 and let 𝐹 be a field with char
𝐹 - 𝑛. Let 𝜁 ∈ 𝐹 be a primitive 𝑛th
𝑎, 𝑏
root of unity. Let 𝑎, 𝑏 ∈ 𝐹 × and let 𝐴 = be the algebra over 𝐹 generated
𝐹, 𝜁
by elements 𝑖, 𝑗 subject to
𝑖 𝑛 = 𝑎, 𝑗 𝑛 = 𝑏, 𝑗𝑖 = 𝜁𝑖 𝑗 .
(a) Show that dim𝐹 𝐴 = 𝑛2 .

(b) Show that 𝐴 is a central simple algebra over 𝐹.
(c) Let 𝐾 = 𝐹 [𝑖] ' 𝐹 [𝑥]/(𝑥 𝑛 − 𝑎). Show that if 𝑏 ∈ Nm𝐾 |𝐹 (𝐾 × ) then
𝐴 ' M𝑛 (𝐹).
[Such algebras are called cyclic algebras or sometimes power norm residue
algebras.]
12. Generalize the statement of Proposition 7.5.3(a) as follows. Let 𝐴, 𝐵 be 𝐹-
algebras, and let 𝐴 0 ⊆ 𝐴 and 𝐵 0 ⊆ 𝐵 be 𝐹-subalgebras. Prove that
𝐶 𝐴⊗𝐵 ( 𝐴 0 ⊗ 𝐵 0) = 𝐶 𝐴 ( 𝐴 0) ⊗ 𝐶 𝐵 (𝐵 0).
13. Let 𝐵 be a finite-dimensional 𝐹-algebra. Show that the following are equivalent:
(i) 𝐵 is separable;
(ii) 𝐵 is semisimple and the center 𝐾 = 𝑍 (𝐵) is separable;
(iii) 𝐵 ⊗𝐹 𝐵op is semisimple.
14. Let 𝐺 ≠ {1} be a finite group. Show that the augmentation ideal, the two-sided
ideal generated by 𝑔 − 1 for 𝑔 ∈ 𝐺, is a nontrivial ideal, and hence 𝐹 [𝐺] is not
simple as an 𝐹-algebra.
15. Let 𝐺 be a finite group of order 𝑛 = #𝐺. Show that 𝐹 [𝐺] is a separable 𝐹-
algebra if and only if char 𝐹 - 𝑛 as follows. [This exercise is known as Maschke’s
theorem.]
(a) Suppose first that char 𝐹 = 0 for a special but quick special case. Compute
Í 𝐹 [𝐺] is separable.
the trace pairing and conclude
(b) If char 𝐹 | 𝑛, show that 𝑁 = 𝑔 ∈𝐺 𝑔 is a nilpotent element in the center of
𝐹 [𝐺], so 𝐹 [𝐺] is not semisimple.
(c) Suppose that char 𝐹 - 𝑛. Let 𝐵 = 𝐹 [𝐺]. Define the map of left 𝐵-modules
by
𝜙 : 𝐵 → 𝐵 ⊗𝐹 𝐵op =: 𝐵e
1 ∑︁
𝜙(1) = 𝑔 ⊗ 𝑔 −1
𝑛 𝑔 ∈𝑔
so that 𝜙(𝛼) = 𝛼𝜙(1) for all 𝛼 ∈ 𝐵. Give 𝐵 the structure of a 𝐵e -algebra

by (𝛼, 𝛼o ) · 𝛽 ↦→ 𝛼𝛽𝛼o . Show that 𝜙 is a homomorphism of 𝐵e -modules,
and that the structure map 𝜓 : 𝐵e → 𝐵 has 𝜓 ◦ 𝜙 = id 𝐵 . Conclude that 𝐵
is separable.
16. Let 𝐵 be an 𝐹-algebra, and let 𝐹 al be an algebraic closure of 𝐹. Show that 𝐵 is
simple if and only if 𝐵 ⊗𝐹 𝐹 al is simple.
17. Let 𝐷 be a (finite-dimensional) division algebra over 𝐹 al . Show that 𝐷 = 𝐹 al .
Conclude that if 𝐵 is a simple algebra over 𝐹 al , then 𝐵 ' M𝑛 (𝐹 al ) for some
𝑛 ≥ 1 and hence is central.
I 18. Let 𝐵 be a (finite-dimensional) 𝐹 algebra, and let 𝐾 ⊇ 𝐹 be a finite separable
extension of fields. Show that rad(𝐵 ⊗𝐹 𝐾) = rad(𝐵) ⊗𝐹 𝐾.
19. Show that if 𝐵 is a semisimple 𝐹-algebra, then so is M𝑛 (𝐵) for any 𝑛 ∈ Z ≥1 .
20. Let 𝐵 be a (finite-dimensional) 𝐹-algebra with standard involution and suppose
char 𝐹 ≠ 2.
(a) Show that rad 𝐵 = rad nrd. Conclude 𝐵 is semisimple if and only if
rad nrd = {0}.
(b) Suppose 𝐵 ≠ 𝐹 and 𝐵 is central. Conclude that 𝐵 is a quaternion algebra
if and only if rad nrd = {0}.
21. Compute the Jacobson radical rad 𝐵 of the 𝐹-algebra 𝐵 with basis 1, 𝑖, 𝑗, 𝑖 𝑗
satisfying
𝑖 2 = 𝑎, 𝑗 2 = 0, and 𝑖 𝑗 = − 𝑗𝑖
for 𝑎 ∈ 𝐹, and compute 𝐵/rad 𝐵. In particular, conclude that such an algebra
is not semisimple, so 𝐵 is not a quaternion algebra. [Hint: restrict to the case
char 𝐹 ≠ 2 first.]
22. Give an example of (finite-dimensional) simple algebras 𝐴, 𝐵 over a field 𝐹 such
that 𝐴 ⊗𝐹 𝐵 is not simple. Then find 𝐴, 𝐵 such that 𝐴 ⊗𝐹 𝐵 is not semisimple.
I 23. In Exercise 7.17, we saw that if 𝐷 is a (finite-dimensional) division algebra over
𝐹 then 𝐷 ⊗𝐹 𝐹 al ' M𝑛 (𝐹 al ) for some 𝑛 ≥ 1. In this exercise, we show the same
is true if we consider the separable closure. (We proved this already in Exercise
6.3 for 𝐷 a quaternion algebra.)
Let 𝐹 be a separably closed field, so every nonconstant separable polynomial
with coefficients in 𝐹 has a root in 𝐹. Let 𝐷 be a finite-dimensional central
division algebra over 𝐹 with char 𝐹 = 𝑝. For purposes of contradiction, assume
that 𝐷 ≠ 𝐹.
(a) Prove that dim𝐹 𝐷 is divisible by 𝑝.
(b) Show that the minimal polynomial of each nonzero 𝑑 ∈ 𝐷 has the form
𝑒
𝑥 𝑝 − 𝑎 for some 𝑎 ∈ 𝐹 and 𝑒 ≥ 0.
∼ M (𝐹 al ). Show that
(c) Choose an 𝐹 al -algebra isomorphism 𝜙 : 𝐷 ⊗𝐹 𝐹 al −→ 𝑛
tr 𝜙(𝑥 ⊗ 1) = 0 for all 𝑥 ∈ 𝐷.
(d) Prove that 𝐷 does not exist.
24. Let 𝐾 ⊇ 𝐹 be a separable (possibly infinite) extension, and let 𝑓 (𝑇) ∈ 𝐾 [𝑇].
Suppose that 𝑓 (𝑇) 𝑛 ∈ 𝐹 [𝑇] for some 𝑛 ∈ Z ≥1 . Show that 𝑓 (𝑇) ∈ 𝐹 [𝑇]. [Hint:
when 𝑝 = char 𝐹 | 𝑛, use the fact that 𝑎 𝑝 ∈ 𝐹 implies 𝑎 ∈ 𝐹.]
25. Let 𝐵 be a finite-dimensional 𝐹-algebra, let 𝛼 ∈ 𝐵, and let 𝑓L (𝛼; 𝑇) and 𝑓R (𝛼; 𝑇)
be the characteristic polynomial of left and right multiplication of 𝛼 on 𝐵,
respectively.
(a) If 𝐵 is semisimple, show that 𝑓L (𝛼; 𝑇) = 𝑓R (𝛼; 𝑇).
(b) Give an example where 𝑓L (𝛼; 𝑇) ≠ 𝑓R (𝛼; 𝑇).
I 26. Use the Skolem–Noether theorem to give another solution to Exercise 6.2: if
𝐾 ⊂ 𝐵 is a separable quadratic 𝐹-algebra then 𝐵 ' (𝐾, 𝑏 | 𝐹) for some 𝑏 ∈ 𝐹 × .
27. Give a direct proof of Corollary 7.7.4. [Hint: Use the fact that there is a unique
simple left 𝐵-module.]
28. Let 𝐵 = (𝐾, 𝑏 | 𝐹) be a quaternion algebra. Show that the subgroup of Aut(𝐵)
that maps 𝐾 ⊆ 𝐵 to itself is isomorphic to the group
𝐾 × /𝐹 × ∪ 𝑗 (𝐾 × /𝐹 × ).
Show that the subgroup of Aut(𝐵) that restricts to the identity on 𝐾 (fixing 𝐾
elementwise) is isomorphic to 𝐾 × /𝐹 × .
I 29. Use the Skolem–Noether theorem and the fact that a finite group cannot be writ-
ten as the union of the conjugates of a proper subgroup to prove Wedderburn’s
little theorem: a finite division ring is a field.
I 30. Let 𝐵 be a quaternion algebra over 𝐹. In this exercise, we show that the
commutator subgroup
[𝐵× , 𝐵× ] = h𝛼𝛽𝛼−1 𝛽−1 : 𝛼, 𝛽 ∈ 𝐵× i
is precisely
[𝐵× , 𝐵× ] = 𝐵1 = {𝛾 ∈ 𝐵× : nrd(𝛾) = 1} = SL1 (𝐵).
(a) Show that [𝐵× , 𝐵× ] ≤ 𝐵1 .

(b) Show that [GL2 (𝐹), GL2 (𝐹)] = SL2 (𝐹) if #𝐹 > 3. [Hint: choose 𝑧 ∈ 𝐹
𝑧 0
such that 𝑧2 − 1 ∈ 𝐹 × , let 𝛾 = , and show that for all 𝑥 ∈ 𝐹 we
0 𝑧 −1
have
1 𝑥(𝑧2 − 1) −1

1 𝑥
= 𝛾,
0 1 0 1
and analogously for the transpose. See 28.3 for a review of elementary
matrices.]
(c) Suppose that 𝐵 is a division algebra. Let 𝛾 ∈ 𝐵1 . Show that there exists
𝛼 ∈ 𝐾 = 𝐹 (𝛾) such that 𝛼𝛼−1 = 𝛾. [Hint: This is a special case of
Hilbert’s theorem 90. Let 𝛼 = 𝛾 + 1 if 𝛾 ≠ −1, and 𝛼 ∈ 𝐵0 r {0} if
𝛾 = −1, with appropriate modifications if char 𝐹 = 2.] Conclude from the
Skolem–Noether theorem that there exists 𝛽 ∈ 𝐵× such that 𝛽𝛼𝛽−1 = 𝛼,
and thus 𝛾 ∈ [𝐵× , 𝐵× ].
31. Show that every ring automorphism of H is inner. (Compare this with ring
automorphisms of C!)
Chapter 8
Simple algebras and involutions
In this chapter, we examine further connections between quaternion algebras, simple

algebras, and involutions.
8.1 ⊲ The Brauer group and involutions

An involution on an 𝐹-algebra 𝐵 induces an isomorphism : 𝐵 −→ ∼ 𝐵op , for example
such an isomorphism is furnished by the standard involution on a quaternion algebra 𝐵.
More generally, if 𝐵1 , 𝐵2 are quaternion algebras, then the tensor product 𝐵1 ⊗𝐹 𝐵2 has
an involution provided by the standard involution on each factor giving an isomorphism
op op
to (𝐵1 ⊗𝐹 𝐵2 ) op ' 𝐵1 ⊗𝐹 𝐵2 —but this involution is no longer a standard involution
(Exercise 8.1). The algebra 𝐵1 ⊗𝐹 𝐵2 is a central simple algebra over 𝐹 called a
biquaternion algebra. In some circumstances, we may have
𝐵1 ⊗𝐹 𝐵2 ' M2 (𝐵3 ) (8.1.1)
where 𝐵3 is again a quaternion algebra, and in other circumstances, we may not;

following Albert, we begin this chapter by studying (8.1.1) and biquaternion algebras
in detail.
To this end, we look at the set of isomorphism classes of central simple algebras
over 𝐹, which is closed under tensor product; if we think that the matrix ring is
something that is ‘no more complicated than its base ring’, it is natural to introduce an
equivalence relation on central simple algebras that identifies a division ring with the
matrix ring (of any rank) over this division ring. More precisely, if 𝐴, 𝐴 0 are central
simple algebras over 𝐹 we say 𝐴, 𝐴 0 are Brauer equivalent if there exist 𝑛, 𝑛 0 ≥ 1 such
that M𝑛 ( 𝐴) ' M𝑛0 ( 𝐴 0). In this way, (8.1.1) reads 𝐵1 ⊗𝐹 𝐵2 ∼ 𝐵3 . The set of Brauer
equivalence classes [ 𝐴] has the structure of a group under tensor product, known as the
Brauer group Br(𝐹) of 𝐹, with identity element [𝐹] and inverse [ 𝐴] −1 = [ 𝐴op ]. The
class [𝐵] ∈ Br(𝐹) of a quaternion algebra 𝐵 is a 2-torsion element, and therefore so is a
biquaternion algebra. In fact, by a striking theorem of Merkurjev, when char 𝐹 ≠ 2, all
2-torsion elements in Br(𝐹) are represented by a tensor product of quaternion algebras
(see section 8.3).
121
122 CHAPTER 8. SIMPLE ALGEBRAS AND INVOLUTIONS
Finally, our interest in involutions in Chapter 3 began with an observation of

Hamilton: the product of a nonzero element with its involute in H is a positive real
number (its norm, or square length). We then proved that the existence of such an
involution characterizes quaternion algebras in an essential way. However, one may
want to relax this setup and instead consider when the product of a nonzero element with
its involute merely has positve trace. Such involutions are called positive involutions
and they arise naturally in algebraic geometry: the Rosati involution is a positive
involution on the endomorphism algebra of an abelian variety, and it is a consequence
that this algebra (over Q) is semisimple, and unsurprisingly quaternion algebras once
again feature prominently (see sections 8.4–8.5).
8.2 Biquaternion algebras

Let 𝐹 be a field. All tensor products in this section will be taken over 𝐹.
8.2.1. Let 𝐵1 , 𝐵2 be quaternion algebras over 𝐹. The tensor product 𝐵1 ⊗ 𝐵2 is a

central simple algebra over 𝐹 of dimension 42 = 16 called a biquaternion algebra.
A biquaternion algebra may be written as a tensor product of two quaternion algebras
in different ways, so the pair is not intrinsic to the biquaternion algebra.
By the Wedderburn–Artin theorem (Main Theorem 7.3.10), we have exactly one
of the three following possibilities for this algebra:
• 𝐵1 ⊗ 𝐵2 is a division algebra;
• 𝐵1 ⊗ 𝐵2 ' M2 (𝐵3 ) where 𝐵3 is a quaternion division algebra over 𝐹; or
• 𝐵1 ⊗ 𝐵2 ' M4 (𝐹).
We could combine the latter two and just say that 𝐵1 ⊗ 𝐵2 ' M2 (𝐵3 ) where 𝐵3 is a
quaternion algebra over 𝐹, since M2 (M2 (𝐹)) ' M4 (𝐹) as 𝐹-algebras.
Example 8.2.2. By Exercise 8.2, when char 𝐹 ≠ 2 we have

𝑎, 𝑏 1 𝑎, 𝑏 2
⊗ ' M2 (𝐵3 )
𝐹 𝐹
𝑎, 𝑏 2

𝑎, 𝑏 1 𝑏 2 𝑎, 𝑏 𝑎, 𝑏
where 𝐵3 = . In particular, ⊗ ' M4 (𝐹), since '
𝐹 𝐹 𝐹 𝐹
M2 (𝐹).
Example 8.2.2 is no accident, as the following proposition indicates.
Proposition 8.2.3 (Albert). The following are equivalent:
(i) There exists a quadratic field extension 𝐾 ⊃ 𝐹 that can be embedded as an

𝐹-algebra in both 𝐵1 and 𝐵2 ;
(ii) 𝐵1 and 𝐵2 have a common quadratic splitting field; and
(iii) 𝐵1 ⊗ 𝐵2 is not a division algebra.
8.2. BIQUATERNION ALGEBRAS 123
Proof. The equivalence (i) ⇔ (ii) follows from Lemma 5.4.7.

For the implication (i) ⇒ (iii), for 𝑖 = 1, 2 let 𝛼𝑖 ∈ 𝐵𝑖 generate 𝐾 so 𝛼𝑖2 = 𝑡𝛼𝑖 − 𝑛
with 𝑡, 𝑛 ∈ 𝐹. Let
𝛽 := 𝛼1 ⊗ 1 − 1 ⊗ 𝛼2 .
Then
𝛽(𝛼1 ⊗ 1 + 1 ⊗ 𝛼2 − 𝑡) = 𝛼12 ⊗ 1 − 1 ⊗ 𝛼22 − 𝑡 𝛽

(8.2.4)
= (𝑡𝛼1 − 𝑛) ⊗ 1 − 1 ⊗ (𝑡𝛼2 − 𝑛) − 𝑡 𝛽 = 0.
Therefore 𝛽 is a zerodivisor and 𝐵1 ⊗ 𝐵2 is not a division algebra.

To finish, we prove (iii) ⇒ (i). We have an embedding
𝐵1 ↩→ 𝐵1 ⊗ 𝐵2
𝛼 ↦→ 𝛼 ⊗ 1
and similarly 𝐵2 ; the images of 𝐵1 and 𝐵2 in 𝐵1 ⊗ 𝐵2 commute. Write 𝐵2 = (𝐾, 𝑏 2 | 𝐹).

Consider (𝐵1 ) 𝐾 = 𝐵1 ⊗ 𝐾 ⊂ 𝐵1 ⊗ 𝐵2 ; then (𝐵1 ) 𝐾 is a quaternion algebra over 𝐾 (with
dim𝐹 (𝐵1 ) 𝐾 = 8). If (𝐵1 ) 𝐾 is not a division algebra, then 𝐾 splits 𝐵1 and 𝐾 ↩→ 𝐵1
and we are done. So suppose that (𝐵1 ) 𝐾 is a division algebra. Then
𝐵1 ⊗ 𝐵2 = (𝐵1 ) 𝐾 + (𝐵1 ) 𝐾 𝑗
is free of rank 2 as a left (𝐵1 ) 𝐾 -module.

Since 𝐵1 ⊗ 𝐵2 ' M2 (𝐵3 ) is not a division algebra, there exists 𝜖 ∈ 𝐵1 ⊗ 𝐵2 nonzero
such that 𝜖 2 = 0. Without loss of generality, we can write 𝜖 = 𝛼1 ⊗ 𝑧 + 𝑗 where 𝛼1 ∈ 𝐵1
and 𝑧 ∈ 𝐾. Then
0 = 𝜖 2 = 𝛼12 ⊗ 𝑧 2 + (𝛼1 ⊗ 𝑧) 𝑗 + (𝛼1 ⊗ 𝑧) 𝑗 + 𝑏 2 . (8.2.5)
From the basis 1, 𝑗 over (𝐵1 ) 𝐾 , if 𝑧 = 𝑡 − 𝑧 with 𝑡 ∈ 𝐹, we conclude that
𝛼1 ⊗ 𝑧 + 𝛼1 ⊗ (𝑡 − 𝑧) = 𝛼1 ⊗ 𝑡 = 0.
Therefore 𝑡 = 0, and 𝑧2 = 𝑐 for some 𝑐 ∈ 𝐹 × . Then from (8.2.5) 𝑐𝛼12 + 𝑏 2 = 0

√
so 𝛼12 = −𝑏 2 /𝑐 and 𝐵1 contains the quadratic field 𝐹 ( −𝑏 2 𝑐). But so does 𝐵2 , as
(𝑧 𝑗) 2 = −𝑏 2 𝑐 as well.
(For an alternate proof, see Jacobson [Jacn2009, Theorem 2.10.3].)
Remark 8.2.6. In view of Proposition 8.2.3, we say that two quaternion algebras
𝐵1 , 𝐵2 over 𝐹 are linked if they contain a common quadratic field extension 𝐾 ⊇ 𝐹.
For further discussion of biquaternion algebras and linkage in characteristic 2 (where
one must treat separable and inseparable extensions differently), see Knus [Knu93],
Lam [Lam2002], or Sah [Sah72]. Garibaldi–Saltman [GS2010] study the subfields of
quaternion algebra over fields with char 𝐹 ≠ 2.
From now on, we suppose that char 𝐹 ≠ 2. (For the case char 𝐹 = 2, see Chapman–
Dolphin–Laghbribi [CDL2015, §6].)
8.2.7. Motivated by Proposition 8.2.3, we consider the quadratic extensions repre-

sented by 𝐵1 and 𝐵2 encoded in the language of quadratic forms (recalling Lemma
5.5.4). Let
𝑉 = {𝛼1 ⊗ 1 − 1 ⊗ 𝛼2 ∈ 𝐵1 ⊗ 𝐵2 : trd(𝛼1 ) = trd(𝛼2 )}.
Then dim𝐹 𝑉 = 6, and we may identify 𝑉 = 𝐵10 ⊗ 1−1 ⊗ 𝐵20 . The reduced norm on each
factor separately defines a quadratic form on 𝑉 by taking the difference: explicitly, if
𝐵1 = (𝑎 1 , 𝑏 1 | 𝐹) and 𝐵2 = (𝑎 2 , 𝑏 2 | 𝐹), then taking the standard bases for 𝐵1 , 𝐵2
𝑄(𝐵1 , 𝐵2 ) ' h−𝑎 1 , −𝑏 1 , 𝑎 1 𝑏 1 i −h−𝑎 2 , −𝑏 2 , 𝑎 2 𝑏 2 i

' h−𝑎 1 , −𝑏 1 , 𝑎 1 𝑏 1 , 𝑎 2 , 𝑏 2 , −𝑎 2 𝑏 2 i.
The quadratic form 𝑄(𝐵1 , 𝐵2 ) : 𝑉 → 𝐹 is called the Albert form of the biquaternion
algebra 𝐵1 ⊗ 𝐵2 .
We then add onto Proposition 8.2.3 as follows.
Proposition 8.2.8 (Albert). Let 𝐵1 ⊗ 𝐵2 be a biquaternion algebra over 𝐹 (with

char 𝐹 ≠ 2) with Albert form 𝑄(𝐵1 , 𝐵2 ). Then the following are equivalent:
(i) 𝐵1 , 𝐵2 have a common quadratic splitting field;

(iv) 𝑄(𝐵1 , 𝐵2 ) is isotropic.
Proof. The implication (ii) ⇒ (iv) follows by construction 8.2.7. To prove (iv) ⇒
(ii), without loss of generality, we may suppose 𝐵1 , 𝐵2 are division algebras; then
an isotropic vector of 𝑄 corresponds√to elements 𝛼1 ∈ 𝐵1 and 𝛼2 ∈ 𝐵2 such that
𝛼12 = 𝛼22 = 𝑐 ∈ 𝐹 × . Therefore 𝐾 = 𝐹 ( 𝑐) is a common quadratic splitting field.
Remark 8.2.9. Albert’s book [Alb39] on algebras still reads well today. The proof of
Proposition 8.2.3 is due to him [Alb72]. (“I discovered this theorem some time ago.
There appears to be some continuing interest in it, and I am therefore publishing it
now.”) Albert used Proposition 8.2.8 to show that

−1, −1 𝑥, 𝑦
𝐵1 = and 𝐵2 =
𝐹 𝐹
over 𝐹 = R(𝑥, 𝑦) have tensor product 𝐵1 ⊗𝐹 𝐵2 a division algebra by verifying that

the Albert form 𝑄(𝐵1 , 𝐵2 ) is anisotropic over 𝐹. See Lam [Lam2005, Example VI.1]
for more details. For the fields of interest in this book (local fields and global fields),
a biquaternion algebra will never be a division algebra—the proof of this fact rests on
classification results for quaternion algebras over these fields, which we will take up
in earnest in Part II.
8.3 Brauer group

Motivated to study the situation where 𝐵1 ⊗ 𝐵2 ' M2 (𝐵3 ) among quaternion algebras
𝐵1 , 𝐵2 , 𝐵3 more generally, we now turn to the Brauer group.
8.3. BRAUER GROUP 125
Let CSA(𝐹) be the set of isomorphism classes of central simple 𝐹-algebras.

The operation of tensor product on CSA(𝐹) defines a commutative binary operation
with identity 𝐹, but inverses are lacking (for dimension reasons). So we define an
equivalence relation ∼ on CSA(𝐹) by
𝐴 ∼ 𝐴 0 if M𝑛0 ( 𝐴) ' M𝑛 ( 𝐴 0) for some 𝑛, 𝑛 0 ≥ 1 (8.3.1)
and we say then that 𝐴, 𝐴 0 are Brauer equivalent. In particular, 𝐴 ∼ M𝑛 ( 𝐴) for all
𝐴 ∈ CSA(𝐹) as needed above.
Lemma 8.3.2. The set of equivalence classes of central simple 𝐹-algebras under the
equivalence relation ∼ has the structure of an abelian group under tensor product,
with identity [𝐹] and inverse [ 𝐴] −1 = [ 𝐴op ].
Proof. By Exercise 8.5, the operation is well-defined: if 𝐴, 𝐴 0 ∈ CSA(𝐹) and 𝐴 0 ∼
𝐴 00 ∈ CSA(𝐹) then 𝐴 ⊗ 𝐴 0 ∼ 𝐴 ⊗ 𝐴 00. To conclude, we need to show that inverses
exist. This is furnished by Lemma 7.5.4: if dim𝐹 𝐴 = 𝑛 and 𝐴op is the opposite algebra
of 𝐴 (3.2.2) then the map
𝐴 ⊗𝐹 𝐴op → End𝐹 ( 𝐴) ' M𝑛 (𝐹)
𝛼 ⊗ 𝛽 ↦→ (𝜇 ↦→ 𝛼𝜇𝛽)
is an isomorphism of 𝐹-algebras, so [ 𝐴] −1 = [ 𝐴op ] provides an inverse to [ 𝐴].

So we make the following definition.
Definition 8.3.3. The Brauer group of 𝐹 is the set Br(𝐹) of Brauer equivalence classes
of central simple 𝐹-algebras (8.3.1) under the group operation of tensor product.
8.3.4. Let 𝐵 be a quaternion algebra over 𝐹. We have 𝐵 ' M2 (𝐹) if and only if
[𝐵] = [𝐹] is the identity. Otherwise, 𝐵 is a division algebra. Then the standard
involution gives an 𝐹-algebra isomorphism 𝐵 −→ ∼ 𝐵op , and hence in Br(𝐹) we have
−1
[𝐵] = [𝐵] and so [𝐵] is an element of order 2. Since Br(𝐹) is abelian, it follows that
biquaternion algebras, or more generally tensor products 𝐵1 ⊗ · · · ⊗ 𝐵𝑡 of quaternion
algebras 𝐵𝑖 , are also elements of order at most 2 in Br(𝐹).
Theorem 8.3.5 (Merkurjev). Let char 𝐹 ≠ 2. Then Br(𝐹) [2] is generated by quater-
nion algebras over 𝐹, i.e., every (finite-dimensional) central division 𝐹-algebra with
involution is Brauer equivalent to a tensor product of quaternion algebras.
Remark 8.3.6. More generally, Merkurjev [Mer82] proved in 1981 that a division
algebra with an involution is Brauer equivalent to a tensor product of quaternion
algebras; more precisely, if 𝐷 is a division 𝐹-algebra with (not necessarily standard)
involution, then there exists 𝑛 ∈ Z ≥1 such that M𝑛 (𝐷) is isomorphic to a tensor
product of quaternion algebras. His theorem, more properly, says that the natural
map 𝐾2 (𝐹) → Br(𝐹) [2] is an isomorphism. (Some care is required in this area:
for example, Amitsur–Rowen–Tignol [ART79] exhibit a division algebra 𝐷 of degree
8 with involution that is not a tensor product of quaternion algebras, but M2 (𝐷) is
a tensor product of quaternion algebras.) For an elementary proof of Merkurjev’s
theorem, see Wadsworth [Wad86].
Remark 8.3.7. Just as quaternion algebras are in correspondence with conics (Corollary
5.5.2), with a quaternion algebra split if and only if the corresponding conic has a
rational point (Theorem 5.5.3), similarly the Brauer group of a field has a geometric
interpretation (see e.g. Serre [Ser79, §X.6]): central simple algebras correspond to
Brauer–Severi varieties—for each degree 𝑛 ≥ 1, both are parametrized by the Galois
cohomology set 𝐻 1 (Gal(𝐹 sep | 𝐹), PGL𝑛 ).
8.4 Positive involutions

We now turn to study algebras with involution more general than a standard involution.
Throughout this section, let 𝐹 ⊆ R be a subfield of R and 𝐵 a finite-dimensional 𝐹-
algebra. We define the trace map Tr : 𝐵 → R by the trace of left multiplication.
Definition 8.4.1. An involution ∗ : 𝐵 → 𝐵 is positive if Tr(𝛼∗ 𝛼) > 0 for all 𝛼 ∈

𝐵 r {0}.
Since the map (𝛼, 𝛽) ↦→ Tr(𝛼∗ 𝛽) is bilinear, an involution ∗ on 𝐵 is positive if and

only if Tr(𝛼∗ 𝛼) > 0 for 𝛼 in a basis for 𝐵 and so is positive if and only if its extension
to 𝐵 ⊗𝐹 R is positive.
Example 8.4.2. The standard involutions on R, C, and H, defined by 𝛼 ↦→ trd(𝛼) − 𝛼,

are positive involutions. The standard involution on R × R is not positive since for
𝛼 = (𝑥1 , 𝑥2 ) ∈ R × R we have Tr(𝛼𝛼) = 2𝑥 1 𝑥 2 . The standard involution on M2 (R) is
also not positive, since for 𝛼 ∈ M2 (R) we have Tr(𝛼𝛼) = 4 det(𝛼).
8.4.3. Let 𝐷 be one of R, C, or H. Let 𝐵 = M𝑛 (𝐷). The standard involution on 𝐷

extends to an involution on 𝐵, acting on coordinates. The conjugate transpose (or,
perhaps better the standard involution transpose) map
∗
:𝐵→𝐵
𝛼 ↦→ 𝛼∗ = 𝛼t
also defines an involution on 𝐵, where t is the transpose map. If 𝛼 = (𝑎 𝑖 𝑗 )𝑖, 𝑗=1,...,𝑛

then
∑︁𝑛
Tr(𝛼∗ 𝛼) = 𝑛(dimR 𝐷) 𝑎 𝑖 𝑗 𝑎 𝑖 𝑗 > 0; (8.4.4)
𝑖, 𝑗=1
thus ∗
is positive, and the norm 𝛼 ↦→ Tr(𝛼∗ 𝛼) is (an integer multiple of) the Frobenius
norm on 𝐵.
We will soon see that every positive involution can be derived from the conjugate
transpose as in 8.4.3. First, we reduce to the case where 𝐵 is a semisimple algebra.
Lemma 8.4.5. Suppose that 𝐵 admits a positive involution ∗ . Then 𝐵 is semisimple.
Proof. We give two proofs. First, we appeal to Theorem 7.9.4: since the trace pairing
is positive definite, it is nondegenerate and immediately 𝐵 is semisimple.
8.4. POSITIVE INVOLUTIONS 127
For a second (more general) proof, let 𝐽 = rad 𝐵 be the Jacobson radical of 𝐵.
By Lemma 7.4.2, 𝐵 is semisimple if and only if rad 𝐵 = {0}, and by Lemma 7.4.8,
𝐽 = rad 𝐵 is nilpotent. Suppose for purposes of contradiction that 𝐽 ≠ {0}. Then there
exists 𝑛 > 0 such that 𝐽 𝑛 ≠ {0} but 𝐽 𝑛+1 = {0}. Let 𝜖 ∈ 𝐽 be such that 𝜖 𝑛 ≠ 0 but
𝜖 𝑛+1 = 0. The involution gives an isomorphism 𝐵 → 𝐵op taking maximal left ideals
to maximal right ideals and therefore by Corollary 7.4.6 we conclude 𝐽 ∗ = 𝐽. Thus
𝜖 𝑛 𝜖 ∗ = 0 so Tr(𝜖 𝑛 (𝜖 ∗ ) 𝑛 ) = Tr(𝜖 𝑛 (𝜖 𝑛 ) ∗ ) = 0, contradicting that ∗ is positive.
8.4.6. Suppose 𝐵 is semisimple with a positive involution ∗ , and let 𝐵𝑖 be a simple

factor of 𝐵. Then ∗ preserves 𝐵𝑖 : for if 𝐵𝑖∗ = 𝐵 𝑗 ≠ 𝐵𝑖 , then 𝐵 𝑗 is a simple factor and
𝐵𝑖 𝐵 𝑗 = 0 so Tr(𝐵𝑖 𝐵𝑖∗ ) = Tr(𝐵𝑖 𝐵 𝑗 ) = {0}, a contradiction.
Putting Lemma 8.4.5 with 8.4.6, we see it is enough to classify positive involutions
on simple R-algebras. By the theorem of Frobenius (Corollary 3.5.8), a simple algebra
over R is isomorphic to M𝑛 (𝐷) with 𝐷 = R, C, H, so 8.4.3 applies.
Proposition 8.4.7. Let 𝐵 ' M𝑛 (𝐷) be a simple R-algebra and let ∗ be the conjugate
transpose involution on 𝐵. Let † : 𝐵 → 𝐵 be another positive involution on 𝐵. Then
there exists an element 𝜇 ∈ 𝐵× with 𝜇∗ = 𝜇 such that
𝛼† = 𝜇−1 𝛼∗ 𝜇
for all 𝛼 ∈ 𝐵.
Proof. First suppose 𝐵 is central over R. Then the involutions † and ∗ give two R-
algebra maps 𝐵 → 𝐵op . By the Skolem–Noether theorem (Main Theorem 7.7.1), there
exists 𝜇 ∈ 𝐵× such that 𝛼† = 𝜇−1 𝛼∗ 𝜇. Since
𝛼 = (𝛼† ) † = (𝜇−1 𝛼∗ 𝜇) † = 𝜇−1 (𝜇−1 𝛼∗ 𝜇) ∗ 𝜇 = (𝜇−1 𝜇∗ )𝛼(𝜇−1 𝜇∗ ) −1 (8.4.8)
for all 𝛼 ∈ 𝐵, we have 𝜇−1 𝜇∗ ∈ 𝑍 (𝐵) = R, so 𝜇∗ = 𝑐𝜇 for some 𝑐 ∈ R. But

(𝜇∗ ) ∗ = 𝜇 = (𝑐𝜇∗ ) ∗ = 𝑐2 𝜇, thus 𝑐 = ±1. But if 𝑐 = −1, then 𝜇 is skew-symmetric so
its top-left entry is 𝜇11 = 0; but then for the matrix unit 𝑒 11 we have
Tr(𝑒 11 𝑒 †11 ) = Tr(𝑒 11 𝜇−1 𝑒 ∗11 𝜇) = Tr(𝜇−1 𝑒 11 𝜇𝑒 11 ) = Tr(𝜇−1 𝜇11 ) = 0, (8.4.9)
a contradiction.
A similar argument holds if 𝐵 has center 𝑍 (𝐵) = C. The restriction of an involution
to 𝑍 (𝐵) is either the identity or complex conjugation; the latter holds for the conjugate
transpose involution, as well as for † : if 𝑧 ∈ 𝑍 (𝐵) then Tr(𝑧𝑧 † ) = 𝑛2 (𝑧𝑧 † ) > 0,
and we must have 𝑧 † = 𝑧. So the map 𝛼 ↦→ (𝛼∗ ) † is a C-linear automorphism, and
again there exists 𝜇 ∈ 𝐵× such that 𝛼† = 𝜇−1 𝛼∗ 𝜇. By the same argument, we have
𝜇∗ = 𝑧𝜇 with 𝑧 ∈ C, but now 𝜇 = (𝜇∗ ) ∗ = 𝑧𝑧𝜇 so | 𝑧| = 1. Let 𝑤 2 = 𝑤/𝑤 = 𝑧; then
(𝑤𝜇) ∗ = 𝑤𝜇∗ = 𝑤𝑧𝜇 = 𝑤𝜇. Replacing 𝜇 by 𝑤𝜇, we may take 𝑧 = 1.
Corollary 8.4.10. The only positive involution on a real division algebra is the stan-
dard involution.
Proof. Apply Proposition 8.4.7 with 𝑛 = 1, noting that 𝜇∗ = 𝜇 = 𝜇 implies 𝜇 ∈ R.

8.4.11. Let 𝜇 ∈ 𝐵× with 𝜇∗ = 𝜇. Then 𝜇 is self-adjoint with respect to the pairing

(𝛼, 𝛽) ↦→ Tr(𝛼∗ 𝛽):
(𝜇𝛼, 𝛽) = Tr((𝜇𝛼) ∗ 𝛽) = Tr(𝛼∗ 𝜇∗ 𝛽) = Tr(𝛼∗ 𝜇𝛽) = (𝛼, 𝜇𝛽).
It follows from the spectral theorem that the R-linear endomorphism of 𝐵 given by
left-multiplication by 𝜇 on 𝐵 as an R-algebra is diagonalizable (with real eigenvalues)
via a symmetric matrix. We say 𝜇 is positive definite (for ∗ ) if all eigenvalues of 𝜇
are positive. The map 𝛼 ↦→ Tr(𝛼∗ 𝜇𝛼) defines a quadratic form on 𝐵, and 𝜇 is positive
definite if and only this quadratic form is positive definite.
Lemma 8.4.12. Let 𝜇∗ = 𝜇. Then the involution 𝛼† = 𝜇−1 𝛼∗ 𝜇 is positive if and only
if either 𝜇 or −𝜇 is positive definite.
Proof. Diagonalize the quadratic form 𝛼 ↦→ Tr(𝛼∗ 𝜇𝛼) to get h𝑎 1 , . . . , 𝑎 𝑚 i in a nor-

malized basis 𝑒 1 , . . . , 𝑒 𝑚 , and suppose without loss of generality that 𝑎 𝑖 = ±1. If all
𝑎 𝑖 = −1, then we can replace 𝜇 with −𝜇 without changing the involution to suppose
they are all +1.
Suppose 𝜇 is not positive, and without loss of generality 𝑎 1 < 0 and 𝑎 2 > 0, then
Tr((𝑒 1 + 𝑒 2 ) ∗ 𝜇(𝑒 1 + 𝑒 2 )) = −1 + 1 = 0, a contradiction. Conversely, if 𝜇 is positive
√
definite, then all eigenvalues are +1. Let 𝜈 = 𝜇 be such that 𝜈 ∗ = 𝜈, and then
Tr(𝛼∗ 𝜇−1 𝛼𝜇) = Tr(𝛼∗ 𝜈 −2 𝛼𝜈 2 ) = Tr((𝜈𝛼∗ 𝜈 −1 ) (𝜈 −1 𝛼𝜈))

(8.4.13)
= Tr((𝜈 −1 𝛼𝜈) ∗ (𝜈 −1 𝛼𝜈)) > 0
for all 𝛼 ∈ 𝐵, so † is positive.
Example 8.4.14. If 𝑛 = 1, and 𝐵 = 𝐷, then the condition 𝜇∗ = 𝜇 implies 𝜇 ∈ R, and

the condition 𝜇 positive implies 𝜇 > 0; rescaling does not affect the involution, so we
can take 𝜇 = 1 and there is a unique positive involution on 𝐷 given by ∗ .

2𝑎 𝑏
Example 8.4.15. Let 𝐵 = M2 (R). Then 𝜇 = is positive definite if and only
𝑏 2𝑐
if 𝑎 > 0 and 𝑏 2 − 4𝑎𝑐 < 0. Combining Proposition 8.4.7 with Lemma 8.4.12, we
see that all positive involutions † on 𝐵 are given by 𝛼† = 𝜇−1 𝛼∗ 𝜇 where 𝜇 is positive
definite.
We can instead relate positive involutions to the standard involution 𝛼 instead of
the transpose; to this end, it is enough to find 𝚥 ∈ 𝐵× = GL2 (R) such that 𝛼 = 𝚥 −1 𝛼∗ 𝚥,
0 1
and the element 𝚥 = does the trick, because
−1 0

0 1 𝑎 𝑐 0 −1 𝑑 −𝑏
= . (8.4.16)
−1 0 𝑏 𝑑 1 0 −𝑐 𝑎

0 1 2𝑎 𝑏 𝑏 2𝑐
From the product = , we conclude that all positive
−1 0 𝑏 2𝑐 −2𝑎 −𝑏
involutions are given by 𝛼† = 𝜇−1 𝛼𝜇 where 𝜇2 ∈ R<0 .
8.5. ∗ ENDOMORPHISM ALGEBRAS OF ABELIAN VARIETIES 129
Remark 8.4.17. Beyond the application to endomorphism algebras, Weil [Weil60] has
given a more general point of view on positive involutions, connecting them to the
classical groups. For more on involutions on finite-dimensional algebras over real
closed fields, see work of Munn [Mun2004].
8.5 ∗ Endomorphism algebras of abelian varieties

We conclude this chapter with an advanced (optional) application: we characterize
endomorphism algebras of (simple) abelian varieties in terms of algebras with involu-
tions. We borrow from the future the notions from section 43.4. Briefly, a complex
torus of dimension 𝑔 is a complex manifold of the form 𝐴 = 𝑉/Λ for 𝑔 ≥ 0, where
Λ ⊂ 𝑉 ' C𝑔 is a lattice (discrete subgroup) and Λ ' Z2𝑔 . A complex abelian variety
is a certain kind of complex torus. A complex abelian variety 𝐴 is simple if 𝐴 has no
abelian subvariety other than {0} and 𝐴.
An endomorphism of 𝐴 is a C-linear map 𝛼 : 𝑉 → 𝑉 such that 𝛼(Λ) ⊆ Λ. Let
End( 𝐴) be the ring (Z-algebra) of endomorphisms of 𝐴.
Proposition 8.5.1. 𝐵 = End( 𝐴) ⊗ Q is a finite-dimensional algebra over Q that admits

a positive involution † : 𝐵 → 𝐵.
Proof. The algebra 𝐵 acts faithfully on Λ ⊗ Q ' Q2𝑔 , so is isomorphic to a subalgebra

of M2𝑔 (Q) hence is finite-dimensional over Q. For positivity, see Proposition 43.4.24
(for the case when 𝐴 is principally polarized).
Remark 8.5.2. The involution † : 𝐵 → 𝐵 is called the Rosati involution (and depends
on a choice of polarization 𝜆 : 𝐴 → 𝐴∨ , where 𝐴∨ is the dual abelian variety).
Now Lemma 8.4.5 and Proposition 8.5.1 imply that 𝐵 is semisimple as a Q-algebra,
with
Ö𝑟
𝐵' M𝑛𝑖 (𝐷 𝑖 )
𝑖=1
where each 𝐷 𝑖 ⊆ 𝐵 is a division algebra. It follows that 𝐴 is isogenous to a product
𝐴1𝑛1 × · · · × 𝐴𝑟𝑛𝑟
where 𝑛1 , . . . , 𝑛𝑟 > 0 and 𝐴1 , . . . , 𝐴𝑟 are simple pairwise nonisogenous abelian sub-

varieties of 𝐴 such that 𝐷 𝑖 = End( 𝐴𝑖 ) ⊗Z Q.
We therefore reduce to the case where 𝐴 is simple, and 𝐷 := End( 𝐴) ⊗ Q is a
division algebra. Let 𝐾 := 𝑍 (𝐷) be the center of 𝐷 and let
𝐾0 := 𝐾 h†i = {𝑎 ∈ 𝐾 : 𝑎 † = 𝑎}
be the subfield of 𝐾 where † acts by the identity.
Lemma 8.5.3. 𝐾0 is a totally real number field, i.e., every embedding 𝐾0 ↩→ C factors
through R, and if † acts nontrivially on 𝐾, then 𝐾 is a CM field, i.e., 𝐾 is a totally
imaginary extension of 𝐾0 .
Proof. The positive involution † restricts to complex conjugation on 𝑍 (𝐷) by Propo-

sition 8.4.7, so for all embeddings 𝐾0 ↩→ C, the image lies in R. For the same reason,
we cannot have † acting nontrivially on 𝐾 and have an embedding 𝐾 ↩→ R.
The following theorem of Albert classifies the possibilities for 𝐷.
Theorem 8.5.4 (Albert). Let 𝐷 be a (finite-dimensional) division algebra over Q with

positive involution † and center 𝐾 = 𝑍 (𝐷), let 𝐾0 := 𝐾 h†i and 𝑛 := [𝐾0 : Q]. Then
𝐾0 is a totally real number field, and one of the four following possibilities holds:
(I) 𝐷 = 𝐾 = 𝐾0 and † is the identity;

(II) 𝐾 = 𝐾0 and 𝐷 is a quaternion algebra over 𝐾0 such that
𝐷 ⊗Q R ' M2 (R) 𝑛 ,
and there exists 𝜇 ∈ 𝐷 × such that 𝜇2 = 𝑑 ∈ 𝐾0× is totally negative and 𝛼† =

𝜇−1 𝛼𝜇 for all 𝛼 ∈ 𝐷;
(III) 𝐾 = 𝐾0 and 𝐷 is a quaternion algebra over 𝐾0 such that
𝐷 ⊗Q R ' H𝑛 ,
and † is the standard involution; or

(IV) 𝐾 ) 𝐾0 and
𝐷 ⊗Q R ' M𝑑 (C) 𝑛
for some 𝑑 ≥ 1, and † extends to the conjugate transpose ∗ on each factor
M𝑑 (C).
Proof. We have assembled many of the tools needed to prove this theorem, and
hopefully motivated its statement sufficiently well—but unfortunately, a proof remains
just out of reach: we require some results about quaternion algebras over number
fields not yet in our grasp. For a proof, see Mumford [Mum70, Application I, §21] or
Birkenhake–Lange [BL2004, §§5.3–5.5].
To connect a few dots as well as we can right now, we give a sketch in the case
where 𝐾 = 𝐾0 for the reader who is willing to flip ahead to Chapter 14. In this case,
𝐷 is a central division algebra over 𝐾 = 𝐾0 and has a 𝐾0 -linear involution giving an
isomorphism 𝐷 −→ ∼ 𝐷 op of 𝐾 -algebras. Looking in the Brauer group Br(𝐾 ), we
0 0
conclude that [𝐷] = [𝐷 op ] = [𝐷] −1 , so [𝐷] ∈ Br(𝐾0 ) has order at most 2. By class
field theory (see Remark 14.6.10), we conclude that either 𝐷 = 𝐾0 or 𝐷 is a (division)
quaternion algebra over 𝐾0 . If 𝐷 = 𝐾0 , weÎare in case (I), so suppose 𝐷 is a quaternion
algebra over 𝐾0 . We have 𝐷 ⊗Q R ' 𝑣 |∞ 𝐷 𝑣 a direct product of 𝑛 quaternion
algebras 𝐷 𝑣 over R indexed by the real places 𝑣 of 𝐾0 . We have 𝐷 𝑣 ' M2 (R) or
𝐷 𝑣 ' H, and our positive involution induces a corresponding positive involution on
each 𝐷 𝑣 . If there exists 𝑣 such that 𝐷 𝑣 ' H, then by Corollary 8.4.10, the positive
involution on 𝐷 𝑣 is the standard involution, so it is so on 𝐷, and then all components
must have 𝐷 𝑣 ' H as the standard involution is not positive on M2 (R)—and we are in
case (II). Otherwise, we are in case (III), with Proposition 8.4.7 and Example 8.4.15
characterizing the positive involution.
8.5. ∗ ENDOMORPHISM ALGEBRAS OF ABELIAN VARIETIES 131
Exercises
Let 𝐹 be a field.
1. Let 𝐵1 , 𝐵2 be quaternion algebras over 𝐹, with standard involution written in

both cases. Let 𝐴 := 𝐵1 ⊗ 𝐵2 .
(a) Show that the map 𝜎 : 𝐴 → 𝐴 defined by 𝛼1 ⊗𝛼2 ↦→ 𝛼1 ⊗𝛼2 for 𝛼1 ∈ 𝐵1 and
𝛼2 ∈ 𝐵2 extends to an involution on 𝐴, but it is not a standard involution.
[Hint: consider sums.]
(b) Suppose that char 𝐹 ≠ 2. Diagonalize 𝐴 = 𝐴+ ⊕ 𝐴− into +1 and −1
eigenspaces for 𝜎. Show that
𝐴+ = 𝐹 ⊕ (𝐵1− ⊗ 𝐵2− ) and 𝐴− = (𝐵1− ⊗ 𝐹) ⊕ (𝐹 ⊗ 𝐵2− ).

𝑎, 𝑏 1 𝑎, 𝑏 2
I 2. Suppose char 𝐹 ≠ 2 and let 𝐵1 := and 𝐵2 := be quaternion
𝐹 𝐹
algebras over 𝐹.
(a) Let 𝐵3 be the 𝐹-span of 1, 𝑖3 := 𝑖1 ⊗ 1, 𝑗3 := 𝑗1 ⊗ 𝑗 2 , and 𝑘 3 := 𝑖3 𝑗3 =
𝑎, 𝑏 1 𝑏 2
𝑖1 𝑗 1 ⊗ 𝑗2 inside 𝐵1 ⊗ 𝐵2 . Show that 𝐵3 ' as 𝐹-algebras.
𝐹
(b) Similarly, let 𝐵4 be the 𝐹-span of 1, 𝑖 4 := 1 ⊗ 𝑗 2 , 𝑗4 := (𝑖1 ⊗ 𝑘 2 )/𝑎, and
𝑏 2 , −𝑏 2
𝑘 4 := 𝑖 4 𝑗 4 . Show that 𝐵4 ' ' M2 (𝐹).
𝐹
(c) Show that
𝐵1 ⊗ 𝐵2 ' 𝐵3 ⊗ 𝐵4 ' M2 (𝐵3 ).
[Hint: Show that 𝐵3 and 𝐵4 are commuting subalgebras, or consider the
map 𝐵3 ⊗ 𝐵4 → 𝐵1 ⊗ 𝐵2 given by multiplication.]
(d) Restore symmetry and repeat (a)–(c) to find algebras 𝐵30 ' 𝐵3 and 𝐵40 '

𝑏 1 , −𝑏 1
with 𝐵1 ⊗ 𝐵2 ' 𝐵30 ⊗ 𝐵40 ' M2 (𝐵30 ).
𝐹
3. Suppose char 𝐹 ≠ 2. Show that 𝐵1 ⊗ 𝐵2 ' M4 (𝐹) if and only if the Albert form
𝑄(𝐵1 , 𝐵2 ) is totally hyperbolic.
4. Let 𝐺 be a finite group. Show that the map induced by 𝑔 ↦→ 𝑔 −1 for 𝑔 ∈ 𝐺
defines an positive involution on R[𝐺]. Then show that this map composed with
coordinatewise complex conjugation defines a positive involution on C[𝐺] (as
an R-algebra).
I 5. Show that if ∼ is the equivalence relation (8.3.1) on CSA(𝐹), then ∼ is compat-
ible with tensor product, i.e., if 𝐴, 𝐴 0 ∈ CSA(𝐹) and 𝐴 0 ∼ 𝐴 00 ∈ CSA(𝐹) then
𝐴 ⊗ 𝐴 0 ∼ 𝐴 ⊗ 𝐴 00.
6. Show that every class in the Brauer group Br(𝐹) contains a unique division
𝐹-algebra, up to isomorphism.
7. Show that Br 𝐹 = {1} if 𝐹 is separably closed, and that Br(R) ' Z/2Z and
Br(F𝑞 ) = {1}.
8. Let 𝐵 ∈ CSA(𝐹) and suppose that 𝐵 has an involution (not necessarily standard).
Show that [𝐵] has order at most 2 in Br 𝐹.
9. Let 𝐾 ⊇ 𝐹 be a field extension. Show that the map 𝐴 ↦→ 𝐴 ⊗𝐹 𝐾 induces
a group homomorphism Br 𝐹 → Br 𝐾. Conclude that the set of isomorphism
classes of central division 𝐹-algebras 𝐷 such that 𝐷 ⊗𝐹 𝐾 ' M𝑛 (𝐾) for some
𝑛 ≥ 1 forms a subgroup of Br 𝐹, called the relative Brauer group Br(𝐾 | 𝐹).
10. In this exercise, we give an example of a central simple algebra of infinite
dimension, called the Weyl algebra.
Suppose char 𝐹 = 0, let 𝐹 [𝑥] be the polynomial ring over 𝐹 in the variable 𝑥.
Inside the enormous algebra End𝐹 𝐹 [𝑥] is the operator 𝑓 (𝑥) ↦→ 𝑥 𝑓 (𝑥), denoted
also 𝑥, and the differentiation operator 𝛿 : 𝐹 [𝑥] → 𝐹 [𝑥]. These two operators
are related by the product rule:
𝛿(𝑥 𝑓 (𝑥)) − 𝑥𝛿( 𝑓 (𝑥)) = 𝑓 (𝑥).
Accordingly, the subalgebra of End𝐹 𝐹 [𝑥] generated by 𝛿, 𝑥 is isomorphic to an

algebra given in terms of generators and relations:
𝑊 := 𝐹 h𝛿, 𝑥i/h𝛿𝑥 − 𝑥𝛿 − 1i,
the quotient of the “noncommutative polynomial ring” in two variables 𝐹 h𝛿, 𝑥i

by the two-sided ideal generated by 𝛿𝑥 − 𝑥𝛿 − 1.
Í𝑛
(a) Show that every element of 𝑊 can be written in the form 𝑖=0 𝑓𝑖 (𝑥)𝛿𝑖
where 𝑓𝑖 (𝑥) ∈ 𝐹 [𝑥] for all 𝑖, i.e., 𝑊 has 𝐹-basis elements 𝑥 𝛿 for 𝑖, 𝑗 ≥ 0.
𝑖 𝑗
(b) Show that 𝑍 (𝑊) = 𝐹.

(c) Let 𝐼 be a two-sided of 𝑊. Show that if there exists nonzero 𝑓 (𝑥) ∈
𝐹 [𝑥] ∩ 𝐼, then 𝐼 = 𝑊. Similarly, show that if 𝛿 𝑛 ∈ 𝐼 for some 𝑛 ≥ 0, then
𝐼 = 𝑊.
(d) Show that 𝑊 is simple. [Hint: argue by induction.]
11. Let 𝐵 be a finite-dimensional R-algebra with positive involution ∗ : 𝐵 → 𝐵. Let
𝑃(𝐵, ∗ ) := {𝜇 ∈ 𝐵 : 𝜇∗ = 𝜇 and 𝜇 is positive definite for ∗ }.
(a) Show that 𝐵× acts on 𝑃(𝐵, ∗ ) by 𝛽 · 𝜇 := 𝛽∗ 𝜇𝛽.

(b) Show that 𝑃(𝐵, ∗ ) is a convex open subset of {𝛼 ∈ 𝐵 : 𝛼∗ = 𝛼}, an
R-vector subspace of 𝐵.
(c) Let 𝜓 : 𝐵 → 𝐵 be an R-algebra automorphism or anti-automorphism.
Show that 𝛼† := 𝜓 −1 (𝜓(𝛼) ∗ ) defines a positive involution for 𝛼 ∈ 𝐵, and
that 𝜓 maps 𝑃(𝐵, † ) bĳectively to 𝑃(𝐵, 𝜓 ).
Part II
Arithmetic
133
Chapter 9
Lattices and integral quadratic forms
In many ways, quaternion algebras are like “noncommutative quadratic field exten-
sions”: this is apparent from their very definition, but also
√ from their description as
wannabe 2 × 2-matrices. Just as the quadratic fields Q( 𝑑) are wonderously rich,
so too are their noncommutative analogues. In this part of the text, we explore these
beginnings of noncommutative algebraic number theory.
In this chapter, we begin with some prerequisites from commutative algebra,
embarking on a study of integral structures and linear algebra over domains.
9.1 ⊲ Integral structures

Just as we find the integers Z inside the rational numbers Q, more generally we want
a robust notion of integrality for possibly noncommutative algebras: this is the theory
of orders over a domain.
We first have to understand the linear algebra aspects of this question. Let 𝑅
be a domain with field of fractions 𝐹 := Frac 𝑅, and let 𝑉 be a finite-dimensional
𝐹-vector space. An 𝑅-lattice in 𝑉 is a finitely generated 𝑅-submodule 𝑀 ⊂ 𝑉 with
𝑀 𝐹 = 𝑉. If 𝑅 is a PID (for example, 𝑅 = Z), then 𝑀 is an 𝑅-lattice if and only if
𝑀 = 𝑅𝑥 1 ⊕ · · · ⊕ 𝑅𝑥 𝑛 where 𝑥 1 , . . . , 𝑥 𝑛 is a basis for 𝑉 as an 𝐹-vector space.
Between 𝑀 and 𝑉 lies intermediate structures, where instead of allowing all
denominators (in the field of fractions), we only allow certain denominators; these are
the localizations of 𝑀. To fix ideas, suppose 𝑅 = Z, so 𝑀 ' Z𝑛 ; we call a Z-lattice
simply a lattice. For a prime 𝑝, we define the localization of Z away from 𝑝 to be
Z ( 𝑝) := {𝑎/𝑏 ∈ Q : 𝑝 - 𝑏} ⊂ Q.
In the localization, we can focus on those aspects of the lattice concentrated at the
prime 𝑝. Extending scalars, 𝑀 ( 𝑝) := 𝑀Z ( 𝑝) ⊆ 𝑉 is a Z ( 𝑝) -lattice in 𝑉, again called the
localization of 𝑀 at 𝑝. These localizations determine the lattice 𝑀 in the following
strong sense (Theorem 9.4.9).
Theorem 9.1.1 (Local-global dictionary for lattices). Let 𝑉 be a finite-dimensional

Q-vector space, and let 𝑀 ⊆ 𝑉 be a lattice. Then the map 𝑁 ↦→ (𝑁 ( 𝑝) ) 𝑝 establishes
135
136 CHAPTER 9. LATTICES AND INTEGRAL QUADRATIC FORMS
a bĳection between lattices 𝑁 and collections of lattices (𝑁 ( 𝑝) ) 𝑝 (indexed by primes

𝑝) where 𝑀 ( 𝑝) = 𝑁 ( 𝑝) for all but finitely many primes 𝑝.
By this theorem, the choice of “reference” lattice 𝑀 is arbitrary. Because of the

importance of this theorem, a property of a lattice that holds if and only if it holds over
every localization is called a local property.
Finally, often vector spaces come equipped with a measure of length, or more
generally a quadratic form; we can restrict these to the lattice 𝑀 ⊆ 𝑉 with 𝑉 a Q-
vector space. More intrinsically, we define a quadratic form 𝑄 : 𝑀 → Z to be a map
satisfying:
(i) 𝑄(𝑎𝑥) = 𝑎 2 𝑄(𝑥) for all 𝑎 ∈ Z and 𝑥 ∈ 𝑀, and

(ii) The associated map 𝑇 : 𝑀 × 𝑀 → Z by 𝑇 (𝑥, 𝑦) = 𝑄(𝑥 + 𝑦) − 𝑄(𝑥) − 𝑄(𝑦) is
(Z-)bilinear.
Condition (i) explains (partly) the ‘quadratic’ nature of the map, and part (ii) is the usual
way relating norms (quadratic forms) to bilinear forms. Choosing a basis 𝑒 1 , . . . , 𝑒 𝑛
for 𝑀 ' Z𝑛 , we may then write
𝑄(𝑥1 𝑒 1 + · · · + 𝑥 𝑛 𝑒 𝑛 ) = 𝑎 11 𝑥12 + 𝑎 12 𝑥1 𝑥2 + · · · + 𝑎 𝑛𝑛 𝑥 𝑛2 ∈ Z[𝑥1 , . . . , 𝑥 𝑛 ] (9.1.2)
as a homogeneous polynomial of degree 2.
9.2 Bits of commutative algebra

We begin with a brief review of some bits of commutative algebra relevant to our
context: we need just enough to do linear algebra over (commutative) domains with
good properties. Good general references for the basic facts from algebra we use
(Dedekind domains, localization, etc.) are Atiyah–Macdonald [AM69], Matsumura
[Mat89, Chapter 8], Curtis–Reiner [CR81, §1, §4], Reiner [Rei2003, Chapter 1], and
Bourbaki [Bou98].
Throughout this chapter, let 𝑅 be a (commutative) noetherian domain with field of
fractions 𝐹 := Frac 𝑅.
9.2.1. An 𝑅-module 𝑃 is projective if it is a direct summand of a free module;

equivalently, 𝑃 is projective if and only if every 𝑅-module surjection 𝑀 → 𝑃 of
𝑅-modules has a section, i.e., an 𝑅-module homomorphism 𝑔 : 𝑃 → 𝑀 such that
𝑓 ◦ 𝑔 = id 𝑃 .
Accordingly, a free 𝑅-module is projective. A projective 𝑅-module 𝑀 is necessar-
ily torsion free over 𝑅, which is to say, if 𝑟𝑥 = 0 with 𝑟 ∈ 𝑅 and 𝑥 ∈ 𝑀, then 𝑟 = 0 or
𝑥 = 0.
9.2.2. A fractional ideal of 𝑅 is a nonzero finitely generated 𝑅-submodule 𝔟 ⊆ 𝐹, or

equivalently, a subset of the form 𝔟 = 𝑑𝔞 where 𝔞 ⊆ 𝑅 is a nonzero ideal and 𝑑 ∈ 𝐹 × .
Two fractional ideals 𝔞, 𝔟 of 𝑅 are isomorphic (as 𝑅-modules) if and only if there exists
𝑐 ∈ 𝐹 × such that 𝔟 = 𝑐𝔞: indeed, given an isomorphism 𝔞 ' 𝔟, we may extend scalars
to 𝐹 to obtain an 𝐹-linear map 𝐹 ' 𝐹, which must be given by 𝑐 ∈ 𝐹 × , and conversely.
9.3. LATTICES 137
A Dedekind domain is a noetherian, integrally closed domain such that every

nonzero prime ideal is maximal.
Example 9.2.3. A field or a PID is a Dedekind domain; in particular, the rings Z and
F 𝑝 [𝑡] are Dedekind domains. If 𝐹 is a finite extension of Q or F 𝑝 (𝑡), then the integral
closure of Z or F 𝑝 [𝑡] in 𝐹 respectively is a Dedekind domain.
9.2.4. Suppose 𝑅 is a Dedekind domain. Then a finitely generated 𝑅-module is
projective if and only if it is torsion free. Moreover, every nonzero ideal 𝔞 of 𝑅 can be
written uniquely as the product of prime ideals (up to reordering). For every fractional
ideal 𝔞 of 𝑅, the set 𝔞−1 := {𝑎 ∈ 𝐹 : 𝑎𝔞 ⊆ 𝑅} is a fractional ideal with 𝔞𝔞−1 = 𝑅.
Therefore the set of fractional ideals of 𝑅 forms a group under multiplication. The
set of principal fractional ideals comprises a subgroup, and we define Cl 𝑅 to be the
quotient, or equivalently the group of isomorphism classes of fractional ideals of 𝑅.
9.3 Lattices
Let 𝑉 be a finite-dimensional 𝐹-vector space.
Definition 9.3.1. An 𝑅-lattice in 𝑉 is a finitely generated 𝑅-submodule 𝑀 ⊆ 𝑉 with
𝑀 𝐹 = 𝑉. We refer to a Z-lattice as a lattice.
The condition that 𝑀 𝐹 = 𝑉 is equivalent to the requirement that 𝑀 contains a
basis for 𝑉 as an 𝐹-vector space.
Example 9.3.2. An 𝑅-lattice in 𝑉 = 𝐹 is the same thing as a fractional ideal of 𝑅.
We will be primarily concerned with projective 𝑅-lattices; if 𝑅 is a Dedekind
domain, then a finitely generated 𝑅-submodule 𝑀 ⊆ 𝑉 is torsion free and hence
automatically projective (9.2.4).
9.3.3. If there is no ambient vector space around, we will also call a finitely generated
torsion free 𝑅-module 𝑀 an 𝑅-lattice: in this case, 𝑀 is a lattice in the 𝐹-vector space
𝑀 ⊗𝑅 𝐹 because the map 𝑀 ↩→ 𝑀 ⊗𝑅 𝐹 is injective (as 𝑀 is torsion free).
Remark 9.3.4. Some authors omit the second condition in the definition of an 𝑅-lattice
and say that 𝑀 is full if 𝑀 𝐹 = 𝑉. We will not encounter 𝑅-lattices that are not full
(and when we do, we call them finitely generated 𝑅-submodules), so we avoid this
added nomenclature.
By definition, an 𝑅-lattice can be thought of an 𝑅-submodule that “allows bounded
denominators”, as follows.
Lemma 9.3.5. Let 𝑀 ⊆ 𝑉 be an 𝑅-lattice and let 𝐽 ⊆ 𝑉 be a finitely generated
𝑅-submodule. Then the following statements hold.
(a) For all 𝑥 ∈ 𝑉, there exists nonzero 𝑟 ∈ 𝑅 such that 𝑟𝑥 ∈ 𝑀.

(b) There exists nonzero 𝑟 ∈ 𝑅 such that 𝑟 𝐽 ⊆ 𝑀.
(c) 𝐽 is an 𝑅-lattice if and only if there exists nonzero 𝑟 ∈ 𝑅 such that 𝑟 𝑀 ⊆ 𝐽 ⊆
𝑟 −1 𝑀.
Proof. First (a). Since 𝐹 𝑀 = 𝑉, the 𝑅-lattice 𝑀 contains an 𝐹-basis 𝑥 1 , . . . , 𝑥 𝑛 for 𝑉,

so in particular 𝑀 ⊇ 𝑅𝑥 1 ⊕ · · · ⊕ 𝑅𝑥 𝑛 . For all 𝑥 ∈ 𝑉, writing 𝑥 in the basis 𝑥 1 , . . . , 𝑥 𝑛
and clearing (finitely many) denominators, we conclude that there exists nonzero 𝑟 ∈ 𝑅
such that 𝑟𝑥 ∈ 𝑀.
For (b), let 𝑦 1 , . . . , 𝑦 𝑚 generate 𝐽 as an 𝑅-module; then for each 𝑖, there exist 𝑟 𝑖 ∈ 𝑅
Î
nonzero such that 𝑟 𝑖 𝑦 𝑖 ∈ 𝑀 hence 𝑟 := 𝑖 𝑟 𝑖 ≠ 0 satisfies 𝑟 𝐽 ⊆ 𝑀, and therefore
𝐽 ⊆ 𝑟 −1 𝑀. For (c), we repeat (b) with 𝑀 interchanged with 𝐽 to find nonzero 𝑠 ∈ 𝑅
such that 𝑠𝑀 ⊆ 𝐽, so then
𝑟 𝑠𝑀 ⊆ 𝑠𝑀 ⊆ 𝐽 ⊆ 𝑟 −1 𝑀 ⊆ (𝑟 𝑠) −1 𝑀.
For the rest of this section, we suppose that 𝑅 is a Dedekind domain and treat lattices
over 𝑅; for further references, see Curtis–Reiner [CR62, §22], O’Meara [O’Me73,
§81], or Fröhlich–Taylor [FT91, §II.4]. It turns out that although not every 𝑅-lattice
has a basis, it can be decomposed as a direct sum, as follows.
Theorem 9.3.6. Let 𝑅 be a Dedekind domain, let 𝑀 ⊆ 𝑉 be an 𝑅-lattice and let
𝑦 1 , . . . , 𝑦 𝑛 be an 𝐹-basis for 𝑉. Then there exist 𝑥 1 , . . . , 𝑥 𝑛 ∈ 𝑀 and fractional ideals
𝔞1 , . . . , 𝔞𝑛 such that
𝑀 = 𝔞1 𝑥 1 ⊕ · · · ⊕ 𝔞𝑛 𝑥 𝑛 (9.3.7)
and 𝑥 𝑗 ∈ 𝐹 𝑦 1 + · · · + 𝐹 𝑦 𝑗 for 𝑗 = 1, . . . , 𝑛.
Accordingly, we say that every 𝑅-lattice 𝑀 is completely decomposable (as a
direct sum of fractional ideals), and we call the elements 𝑥1 , . . . , 𝑥 𝑛 a pseudobasis
for the lattice 𝑀 with respect to the coefficient ideals 𝔞1 , . . . , 𝔞𝑛 . The matrix with
rows 𝑥𝑖 in the basis 𝑦 𝑖 is lower triangular by construction; without loss of generality
(rescaling), we may suppose that the diagonal entries are equal to 1, in which case we
say that the pseudobasis for 𝑀 is given in Hermite normal form.
More generally, if 𝑀 = 𝔞1 𝑥1 + · · · + 𝔞𝑚 𝑥 𝑚 , the sum not necessarily direct, then we
say that the elements 𝑥𝑖 are a pseudogenerating set for 𝑀 with coefficient ideals 𝔞𝑖 .
Proof of Theorem 9.3.6. We argue by induction on 𝑛, the case 𝑛 = 1 corresponding to
the case of a single fractional ideal.
Let 𝑊 := 𝐹 𝑦 1 + · · · + 𝐹 𝑦 𝑛−1 , and let 𝑁 = 𝑀 ∩ 𝑊. Then there is a commutative
diagram
0 /𝑁 /𝑀 / 𝑀/𝑁 /0
(9.3.8)

0 /𝑊 /𝑉 / 𝑉/𝑊 /0
Since 𝑁 = 𝑊 ∩ 𝑀, we have 𝑀/𝑁 ↩→ 𝑉/𝑊, and 𝑉/𝑊 ' 𝐹 projecting onto 𝐹 𝑦 𝑛 .

Since 𝑀/𝑁 is nonzero and finitely generated, by 9.2.4 we conclude 𝑀/𝑁 ' 𝔞 ⊆ 𝐹 is
a fractional ideal, hence projective. Therefore the top exact sequence of 𝑅-modules
splits (the surjection has a section), so there exists 𝑥 ∈ 𝑀 such that 𝑀 = 𝑁 ⊕ 𝔞𝑥 as
𝑅-modules. The result then follows by applying the inductive hypothesis to 𝑁.
An argument generalizing that of Theorem 9.3.6 yields the following [O’Me73,
81:11].
9.4. LOCALIZATIONS 139
Theorem 9.3.9 (Invariant factors). Let 𝑅 be a Dedekind domain and let 𝑀, 𝑁 ⊆ 𝑉 be

𝑅-lattices. Then there exists a common pseudobasis 𝑥 1 , . . . , 𝑥 𝑛 for 𝑀, 𝑁; i.e., there
exists a basis 𝑥1 , . . . , 𝑥 𝑛 for 𝑉 and fractional ideals 𝔞1 , . . . , 𝔞𝑛 and 𝔟1 , . . . , 𝔟𝑛 such that
𝑀 = 𝔞1 𝑥1 ⊕ · · · ⊕ 𝔞𝑛 𝑥 𝑛
𝑁 = 𝔟1 𝑥1 ⊕ · · · ⊕ 𝔟𝑛 𝑥 𝑛
Moreover, letting 𝔡𝑖 := 𝔟𝑖 𝔞−1

𝑖 we may further take 𝔡1 | · · · | 𝔡𝑛 , and then such 𝔡𝑖 are
unique.
The unique fractional ideals 𝔡1 , . . . , 𝔡𝑛 given by Theorem 9.3.9 are called the
invariant factors of 𝑁 relative to 𝑀.
9.3.10. Let 𝑀 ⊆ 𝑉 be an 𝑅-lattice with pseudobasis as in (9.3.7). The class

[𝔞1 · · · 𝔞𝑛 ] ∈ Cl 𝑅 is well-defined (Exercise 9.7) and called the Steinitz class.
In fact, if we do not require that 𝑥 𝑗 ∈ 𝐹 𝑦 1 + · · · + 𝐹 𝑦 𝑖 for 𝑗 = 1, . . . , 𝑛 in Theorem
9.3.6, then we can find a pseudobasis for 𝑀 with 𝔞1 = · · · = 𝔞𝑛−1 = 𝑅, i.e.,
𝑀 = 𝑅𝑥1 ⊕ · · · ⊕ 𝑅𝑥 𝑛−1 ⊕ 𝔞𝑥 𝑛
with [𝔞] the Steinitz class of 𝑀.
9.4 Localizations
Properties of a domain are governed in an important way by its localizations, and
consequently the structure of lattices, orders, and algebras can often be understood by
looking at their localizations (and later, completions).
For a prime ideal 𝔭 ⊆ 𝑅, we denote by
𝑅 (𝔭) := {𝑟/𝑠 ∈ 𝐹 : 𝑠 ∉ 𝔭} ⊆ 𝐹 (9.4.1)
the localization of 𝑅 at 𝔭. (We reserve the simpler subscript notation for the completion,
defined in section 9.5.)
Example 9.4.2. If 𝑅 = Z and 𝔭 = (2), then 𝑅 (2) = {𝑟/𝑠 ∈ Q : 𝑠 is odd} consists of

the subring of rational numbers with odd denominator.
Since 𝑅 is a domain, the map 𝑅 ↩→ 𝑅 (𝔭) is an embedding and we can recover 𝑅 as

an intersection Ù Ù
𝑅= 𝑅 (𝔭) = 𝑅 (𝔪) ⊆ 𝐹 (9.4.3)
𝔭 𝔪
where the intersections are over all prime ideals of 𝑅 and all maximal ideals of 𝑅,
respectively.
Let 𝑉 be a finite-dimensional 𝐹-vector space and let 𝑀 ⊆ 𝑉 be an 𝑅-lattice. For a
prime 𝔭 of 𝑅, let
𝑀 (𝔭) := 𝑀 𝑅 (𝔭) ⊆ 𝑉
be the extension of scalars of 𝑀 over 𝑅 (𝔭) ; identifying 𝑉 = 𝑀 𝐹 ' 𝑀 ⊗𝑅 𝐹 under

multiplication, we could similarly define
𝑀 (𝔭) := 𝑀 ⊗𝑅 𝑅 (𝔭) .
In either lens, 𝑀 (𝔭) is an 𝑅 (𝔭) -lattice in 𝑉. In this way, 𝑀 determines a collection

(𝑀 (𝔭) )𝔭 indexed over the primes 𝔭 of 𝑅.
9.4.4. Returning to 9.2.1, a finitely generated 𝑅-module 𝑀 is projective if and only if
it is locally free, i.e., 𝑀 (𝔭) is free for all prime ideals of 𝑅.
The ability to argue locally and then with free objects is very useful, and so very
often we will restrict our attention to projective (equivalently, locally free) 𝑅-modules.
9.4.5. The localization of a Dedekind domain 𝑅 is a discrete valuation ring (DVR). A
DVR is equivalently a local PID that is not a field. In particular, a DVR is integrally
closed, and every finitely generated module over a DVR is free.
Consequently, if 𝑅 is a Dedekind domain, then every fractional ideal of 𝑅 is
locally principal, i.e., if 𝔞 ⊆ 𝐹 is a fractional ideal, then for all primes 𝔭 of 𝑅 we have
𝔞 (𝔭) = 𝑎 𝔭 𝑅 (𝔭) for some 𝑎 𝔭 ∈ 𝐹 × .
We now prove a version of the equality (9.4.3) for 𝑅-lattices (recall Definition
9.3.1).
Lemma 9.4.6. Let 𝑀 be an 𝑅-lattice in 𝑉. Then
Ù Ù
𝑀= 𝑀 (𝔭) = 𝑀 (𝔪) ⊆ 𝑉
𝔭 𝔪
where the intersection is over all prime (maximal) ideals 𝔭.
Ñ for maximal ideals since 𝑀 (𝔪) ⊆ 𝑀 (𝔭) when-

Proof. It suffices to prove the statement
ever 𝔪 ⊇ 𝔭. The inclusion 𝑀 ⊆ 𝔪 𝑀 (𝔪) is clear. Conversely, let 𝑥 ∈ 𝑉 satisfy
𝑥 ∈ 𝑀 (𝔪) for all maximal ideals 𝔪. Let
(𝑀 : 𝑥) := {𝑟 ∈ 𝑅 : 𝑟𝑥 ∈ 𝑀 }.
Then (𝑀 : 𝑥) is an ideal of 𝑅. For a maximal ideal 𝔪 of 𝑅, since 𝑥 ∈ 𝑀 (𝔪) there exists

0 ≠ 𝑟 𝔪 ∈ 𝑅 r 𝔪 such that 𝑟 𝔪 𝑥 ∈ 𝑀. Thus 𝑟 𝔪 ∈ (𝑀 : 𝑥) and (𝑀 : 𝑥) is not contained
in any maximal ideal of 𝑅. Therefore (𝑀 : 𝑥) = 𝑅 and hence 𝑥 ∈ 𝑀.
Corollary 9.4.7. Let 𝑀, 𝑁 be 𝑅-lattices in 𝑉. Then the following are equivalent:

(i) 𝑀 ⊆ 𝑁;
(ii) 𝑀 (𝔭) ⊆ 𝑁 (𝔭) for all prime ideals 𝔭 of 𝑅; and
(iii) 𝑀 (𝔪) ⊆ 𝑁 (𝔪) for all maximal ideals 𝔪 of 𝑅.
Ñ(i) ⇒ (ii) ⇒ (iii) are direct; for the implication (iii) ⇒ (i), we
Proof. TheÑimplications
have 𝑀 = 𝔭 𝑀 (𝔪) ⊆ 𝔭 𝑁 (𝔪) = 𝑁 by Lemma 9.4.6.
In particular, it follows from Corollary 9.4.7 that 𝑀 = 𝑁 for 𝑅-lattices 𝑀, 𝑁 if and

only if 𝑀 (𝔭) = 𝑁 (𝔭) for all primes 𝔭 of 𝑅.
9.5. COMPLETIONS 141
9.4.8. A property that holds if and only if it holds locally (as in Corollary 9.4.7, for
the property that one lattice is contained in another) is called a local property.
To conclude this section, we characterize in a simple way the conditions under

which a collection (𝑀 (𝔭) )𝔭 of 𝑅 (𝔭) -lattices arise from a global 𝑅-lattice. We will see
that just as a nonzero ideal of 𝑅 can be factored uniquely into a product of prime
ideals, and hence by the data of these primes and their exponents, so too can a lattice
be understood by a finite number of localized lattices, once a “reference” lattice has
been chosen (to specify the local behavior of the lattice at other primes).
Theorem 9.4.9 (Local-global dictionary for lattices). Let 𝑅 be a Dedekind domain,

and let 𝑀 ⊆ 𝑉 be an 𝑅-lattice. Then the map 𝑁 ↦→ (𝑁 (𝔭) )𝔭 establishes a bĳection
between 𝑅-lattices 𝑁 ⊆ 𝑉 and collections of lattices (𝑁 (𝔭) )𝔭 indexed by the primes 𝔭
of 𝑅 satisfying 𝑀 (𝔭) = 𝑁 (𝔭) for all but finitely many primes 𝔭.
In Theorem 9.4.9, the choice of the “reference” lattice 𝑀 is arbitrary: if 𝑀 0 is

0 for all but finitely many primes 𝔭,
another lattice, then by Theorem 9.4.9 𝑀 (𝔭) = 𝑀 (𝔭)
so we get the same set of lattices replacing 𝑀 by 𝑀 0. In particular, any lattice 𝑁 ⊆ 𝑉
agrees with any other one at all but finitely many localizations.
Remark 9.4.10. In Theorem 9.4.9, there is a bit of notational abuse: when we write a
collection (𝑁 (𝔭) )𝔭 , we do not mean to imply that there is (yet) an 𝑅-lattice 𝑁 such that
the localization of 𝑁 at 𝔭 is equal to 𝑁 (𝔭) . This conclusion is what is provided by the
theorem (the statement of surjectivity), so the notational conflict is only temporary.
Proof of Theorem 9.4.9. Let 𝑁 ⊆ 𝑉 be an 𝑅-lattice. Then there exists 0 ≠ 𝑟 ∈ 𝑅 such

that 𝑟 𝑀 ⊆ 𝑁 ⊆ 𝑟 −1 𝑀. But 𝑟 is contained in only finitely many prime (maximal)
ideals of 𝑅, so for all but finitely many primes 𝔭, the element 𝑟 is a unit in 𝑅 (𝔭) and
thus 𝑀 (𝔭) = 𝑁 (𝔭) .
So consider the set of collections (𝑁 (𝔭) )𝔭 of lattices where 𝑁 (𝔭) is an 𝑅 (𝔭) -lattice
for each prime 𝔭 with the property that 𝑀 (𝔭) = 𝑁 (𝔭) for all but finitely many primes 𝔭 of
Ñ
𝑅. Given such a collection, we define 𝑁 = 𝔭 𝑁 (𝔭) ⊆ 𝑉. Then 𝑁 is an 𝑅-submodule
of 𝑉. We show it is an 𝑅-lattice in 𝑉. For each 𝔭 such that 𝑀Î (𝔭) ≠ 𝑁 (𝔭) , there exists
𝑟 𝔭 ∈ 𝑅 such that 𝑟 𝔭 𝑀 (𝔭) ⊆ 𝑁 (𝔭) ⊆ 𝑟 𝔭−1 𝑀 (𝔭) . Therefore, if 𝑟 = 𝔭 𝑟 𝔭 is the product of
these elements, then 𝑟 𝑀 (𝔭) ⊆ 𝑁 ⊆ 𝑟 −1 𝑀 (𝔭) for all primes 𝔭 with 𝑀 (𝔭) ≠ 𝑁 (𝔭) . On the
other hand, if 𝑀 (𝔭) = 𝑁 (𝔭) then already 𝑟 𝑀 (𝔭) ⊆ 𝑀 (𝔭) = 𝑁 (𝔭) ⊆ 𝑟 −1 𝑁 (𝔭) = 𝑟 −1 𝑀 (𝔭) .
Therefore by Corollary 9.4.7, we have 𝑟 𝑀 ⊆ 𝑁 ⊆ 𝑟 −1 𝑀, and so 𝑁 is an 𝑅-lattice.
Ñ
By Lemma 9.4.6, the association (𝑁 (𝔭) )𝔭 ↦→ 𝔭 𝑁 (𝔭) is an inverse to 𝑁Ñ↦→ (𝑁(𝔭) )𝔭 .
Conversely, given a collection (𝑁 (𝔭) )𝔭 , for a nonzero prime 𝔭, we have 𝔮 𝑁𝔮 (𝔭) =
𝑁 (𝔭) since (𝑅𝔮 ) (𝔭) = 𝐹 so (𝑁𝔮 ) (𝔭) = 𝑉 whenever 𝔮 ≠ 𝔭.
9.5 Completions
Next, we briefly define the completion and show that the local-global dictionary holds
in this context as well. (We will consider completions in the context of local fields
more generally starting in chapter 12, so the reader may wish to return to this section
later.) For a general reference on completions (and the induced topology), see e.g.
Atiyah–Macdonald [AM69, Chapter 10], Matsumura [Mat89, Chapter 8], Bourbaki

[Bou98, Chapter III, §3].
To avoid diving too deeply into commutative algebra we suppose that 𝑅 is a DVR,
with maximal ideal 𝔭: for example, we might take the localization of a Dedekind
domain 𝑅 at a prime ideal by 9.4.5. There is a natural system of compatible projection
maps 𝑅/𝔭𝑛+1 → 𝑅/𝔭𝑛 indexed by integers 𝑛 ≥ 1, and we define the completion of 𝑅
at 𝔭 to be the inverse (or projective) limit under this system:
𝑅𝔭 := lim 𝑅/𝔭𝑛
← −−
𝑛
( ∞
Ö
) (9.5.1)
𝑛 𝑛
:= 𝑎 = (𝑎 𝑛 ) 𝑛 ∈ 𝑅/𝔭 : 𝑎 𝑛+1 ≡ 𝑎 𝑛 (mod 𝔭 ) for all 𝑛 ≥ 1 .
𝑛=1
The completion 𝑅𝔭 is again a commutative ring, and we have aÑnatural map 𝑅 → 𝑅𝔭

defined by 𝑎 ↦→ (𝑎) 𝑛 . Since 𝑅 has a discrete valuation we have ∞𝑛=0 𝔭 = {0}, so this
𝑛
map is injective. Moreover, since 𝔭 is maximal, then in fact this inclusion factors via
𝑅 ↩→ 𝑅 (𝔭) ↩→ 𝑅𝔭 inducing isomorphisms (Exercise 9.8)
∼ 𝑅 /𝔭𝑒 𝑅 −→
𝑅/𝔭𝑒 −→ ∼ 𝑅 /𝔭𝑒 𝑅 (9.5.2)
(𝔭) (𝔭) 𝔭 𝔭
for all 𝑒 ≥ 1; in particular, the operation of completion is in a sense ‘stronger’ than

the operation of localization, and the valuation on 𝑅 extends naturally to 𝑅𝔭 , so
𝑅 = 𝐹 ∩ 𝑅𝔭 ⊆ 𝐹𝔭 . However, once local the completion looks rather similar in the
context of lattices, as follows. Let 𝐹𝔭 := 𝐹 ⊗𝑅 𝑅𝔭 and 𝑉𝔭 := 𝑉 ⊗𝐹 𝐹𝔭 .
Lemma 9.5.3. Let 𝑅 be a DVR with maximal ideal 𝔭 ⊆ 𝑅. Then the maps
𝑀 ↦→ 𝑀𝔭 := 𝑀 ⊗𝑅 𝑅𝔭
(9.5.4)
𝑀𝔭 ∩ 𝑉 ← 𝑀𝔭
are mutually inverse bĳections between the set of 𝑅-lattices in 𝑉 and the set of 𝑅𝔭 -
lattices in 𝑉𝔭 .
Proof. Let 𝑀 ⊆ 𝑉 be an 𝑅-lattice. By 9.4.5 we have 𝑀 ' 𝑅 𝑛 free over 𝑅; choose
a basis 𝑀 = 𝑅𝑥 1 ⊕ · · · ⊕ 𝑅𝑥 𝑛 . Then 𝑀𝔭 = 𝑀 ⊗𝑅 𝑅𝔭 ' 𝑅𝔭 𝑥1 ⊕ · · · ⊕ 𝑅𝔭 𝑥 𝑛 . Let
𝑀 0 := 𝑀𝔭 ∩ 𝑉 ⊆ 𝑉𝔭 . Then 𝑥 0 ∈ 𝑀 0 if and only if 𝑥 0 = 𝑎 1 𝑥1 + · · · + 𝑎 𝑛 𝑥 𝑛 with
𝑎 𝑖 ∈ 𝑅𝔭 ∩ 𝐹 = 𝑅, so indeed 𝑀 0 = 𝑀.
Conversely, let 𝑀𝔭 ⊆ 𝑉𝔭 and let 𝑀 0 := 𝑀𝔭 ∩ 𝑉. Then (𝑀 0)𝔭 ⊆ 𝑀𝔭 , and we
prove the opposite inclusion. First, a bit of setup. Let 𝑦 1 , . . . , 𝑦 𝑛 be an 𝐹-basis for
𝑉, and let 𝑁 = 𝑅𝑦 1 ⊕ · · · ⊕ 𝑅𝑦 𝑛 . By Lemma 9.3.5, there exists nonzero 𝑟 ∈ 𝑅𝔭 such
that 𝑟 𝑁𝔭 ⊆ 𝑀𝔭 ⊆ 𝑟 −1 𝑁𝔭 . Choosing an element 𝑠 ∈ 𝑅 with the same valuation as 𝑟,
we have 𝑟/𝑠 ∈ 𝑅𝔭× so in fact may suppose that 𝑟 ∈ 𝑅. Rescaling the basis vectors
𝑦 𝑖 and replacing 𝑟 2 by 𝑟 we may suppose that (𝑟 𝑁)𝔭 = 𝑟 𝑁𝔭 ⊆ 𝑀𝔭 ⊆ 𝑁𝔭 . From the
previous paragraph, we have 𝑟 𝑁𝔭 = 𝑟 𝑁 ⊗𝑅 𝑅𝔭 ⊆ 𝑀𝔭0 . Letting (𝑟) = 𝔭𝑒 and taking
(9.5.2) on each coordinate, we have an isomorphism 𝜑 : 𝑁/𝑟 𝑁 ' 𝑁𝔭 /𝑟 𝑁𝔭 induced
from the natural inclusion 𝑁 ↩→ 𝑁𝔭 . Now to show the inclusion, let 𝑦 ∈ 𝑀𝔭 . Let
𝑥 ∈ 𝑁 ⊆ 𝑉 be such that 𝜑(𝑥 + 𝑟 𝑁) = 𝑦 + 𝑟 𝑁𝔭 ; lifting to 𝑁𝔭 , we find that there exists
𝑧 ∈ 𝑟 𝑁𝔭 ⊆ (𝑀 0)𝔭 ⊆ 𝑀𝔭 such that 𝑥 = 𝑦 + 𝑧 ∈ 𝑀𝔭 ∩ 𝑉, so 𝑦 = 𝑥 − 𝑧 ∈ (𝑀 0)𝔭 .
9.6. INDEX 143
In particular, Lemma 9.5.3 implies that in the local-global dictionary for lattices
over a Dedekind domain 𝑅 (Theorem 9.4.9), we may also work with collections of
𝑅𝔭 -lattices (𝑁𝔭 )𝔭 over the completions at primes.
9.6 Index
Continuing with 𝑅 a noetherian domain, let 𝑀, 𝑁 ⊆ 𝑉 be 𝑅-lattices.
Definition 9.6.1. The 𝑅-index of 𝑁 in 𝑀, written [𝑀 : 𝑁] 𝑅 , is the 𝑅-submodule of

𝐹 generated by the set
{det(𝛿) : 𝛿 ∈ End𝐹 (𝑉) and 𝛿(𝑀) ⊆ 𝑁 }. (9.6.2)
The style of Definition 9.6.1, given by a large generating set (9.6.2), is the replace-
ment for being able to work with given bases; this style will be typical for us in what
follows. The determinants det(𝛿) are meant in the intrinsic sense, but can be computed
as the determinant of a matrix upon choosing a basis for 𝑉.
Lemma 9.6.3. The index [𝑀 : 𝑁] 𝑅 is a nonzero 𝑅-module, and if 𝛼 ∈ Aut𝐹 (𝑉) then
[𝛼𝑀 : 𝑁] = det(𝛼) −1 [𝑀 : 𝑁].
Lemma 9.6.4. If 𝑀, 𝑁 are free (as 𝑅-submodules), then [𝑀 : 𝑁] 𝑅 is a free 𝑅-module

generated by the determinant of any 𝛿 ∈ End𝐹 (𝑉) giving a change of basis from 𝑀 to
𝑁.
Proof. Let 𝑥1 , . . . , 𝑥 𝑛 be an 𝑅-basis for 𝑀, thereby an 𝐹-basis for 𝑉. Let 𝑦 1 , . . . , 𝑦 𝑛

be an 𝑅-basis for 𝑁; then the map 𝑥𝑖 ↦→ 𝑦 𝑖 first extends to an 𝑅-linear isomorphism
𝑀 −→∼ 𝑁 and thereby to an 𝐹-linear map 𝛿 ∈ End (𝑉), and of course 𝛿(𝑀) ⊆ 𝑁
𝐹
by construction, so det(𝛿) ∈ [𝑀 : 𝑁] 𝑅 . Conversely, let 𝛿 0 ∈ End𝐹 (𝑉) be such
that 𝛿 0 (𝑀) ⊆ 𝑁. The map 𝛿 0 𝛿−1 : 𝑁 → 𝑁 is an 𝑅-linear map, so det(𝛿 0 𝛿−1 ) =
det(𝛿 0) det(𝛿) −1 ∈ 𝑅, so det(𝛿 0) ∈ det(𝛿)𝑅.
Example 9.6.5. If 𝑁 = 𝑟 𝑀 with 𝑟 ∈ 𝑅, then [𝑀 : 𝑁] 𝑅 = 𝑟 𝑛 𝑅 where 𝑛 = dim𝐹 𝑉.
Example 9.6.6. If 𝑅 = Z and 𝑁 ⊆ 𝑀, then [𝑀 : 𝑁] Z is the ideal generated by

#(𝑀/𝑁), the usual index taken as abelian groups. In this case, for convenience we
will often identify [𝑀 : 𝑁] Z with its unique positive generator.
Forming the 𝑅-index commutes with localization, as follows.
Lemma 9.6.7. Let 𝔭 be a prime of 𝑅. Then
[𝑀 (𝔭) : 𝑁 (𝔭) ] 𝑅 (𝔭) = ( [𝑀 : 𝑁] 𝑅 ) (𝔭) .

Proof. If 𝛿(𝑀) ⊆ 𝑁, then 𝛿(𝑀 (𝔭) ) ⊆ 𝑁 (𝔭) by 𝑅 (𝔭) -linearity, giving the inclusion
(⊇). For (⊆), let 𝛿 ∈ End𝐹 (𝑉) be such that 𝛿(𝑀 (𝔭) ) ⊆ 𝑁 (𝔭) . For any 𝑥 ∈ 𝑀, we
have 𝛿(𝑥) ∈ 𝛿(𝑀) ⊆ 𝛿(𝑀 (𝔭) ) ⊆ 𝑁 (𝔭) , so there exists 𝑦 ∈ 𝑁 and 𝑠 ∈ 𝑅 r 𝔭 such
that 𝑠𝛿(𝑥) = 𝑦 ∈ 𝑁. Let 𝑥1 , . . . , 𝑥 𝑚 generate 𝑀 as an 𝑅-module, and for each 𝑖,
Î
let 𝑠𝑖 ∈ 𝑅 r 𝔭 be such that 𝑠𝑖 𝛿(𝑥 𝑖 ) ∈ 𝑁. Let 𝑠 := 𝑖 𝑠𝑖 . Then 𝑠𝛿(𝑀) ⊆ 𝑁, so
det(𝑠𝛿) = 𝑠 𝑛 det(𝛿) ∈ [𝑀 : 𝑁] 𝑅 , if 𝑛 := dim𝐹 𝑉. Finally, 𝑠 ∈ 𝑅 ×(𝔭) , we conclude that
det 𝛿 ∈ ( [𝑀 : 𝑁] 𝑅 ) (𝔭) , as desired.
Proposition 9.6.8. Suppose that 𝑀, 𝑁 are projective 𝑅-modules. Then [𝑀 : 𝑁] 𝑅 is a
projective 𝑅-module. Moreover, if 𝑁 ⊆ 𝑀 then [𝑀 : 𝑁] 𝑅 = 𝑅 if and only if 𝑀 = 𝑁.
Proof. Let 𝔭 be a prime of 𝑅 and consider the localization ( [𝑀 : 𝑁] 𝑅 ) (𝔭) at 𝔭. Since
𝑀, 𝑁 are projective 𝑅-modules, they are locally free (9.2.1). By Lemma 9.6.4, the
local index [𝑀 (𝔭) : 𝑁 (𝔭) ] 𝑅 (𝔭) is a principal 𝑅 (𝔭) -ideal. By Lemma 9.6.7, we conclude
that [𝑀 : 𝑁] 𝑅 is locally principal, therefore projective.
The second statement follows in a similar way: we may suppose that 𝑅 is local and
thus 𝑁 ⊆ 𝑀 are free, in which case 𝑀 = 𝑁 if and only if a change of basis matrix
from 𝑁 to 𝑀 has determinant in 𝑅 × .
For Dedekind domains, the 𝑅-index can be described as follows.
Lemma 9.6.9. If 𝑅 is a Dedekind domain and 𝑁 ⊆ 𝑀, then [𝑀 : 𝑁] 𝑅 is the product
of the invariant factors (or elementary divisors) of the torsion 𝑅-module 𝑀/𝑁.
9.7 Quadratic forms

In setting up an integral theory, we will also have need of an extension of the theory
of quadratic forms integrally, generalizing those over fields (Section 4.2). For further
reading on quadratic forms over rings, we suggest the books by O’Meara [O’Me73],
Knus [Knu88], and Scharlau [Scha85].
Definition 9.7.1. A quadratic map is a map 𝑄 : 𝑀 → 𝑁 between 𝑅-modules,
satisfying:
(i) 𝑄(𝑟𝑥) = 𝑟 2 𝑄(𝑥) for all 𝑟 ∈ 𝑅 and 𝑥 ∈ 𝑀; and
(ii) The map 𝑇 : 𝑀 × 𝑀 → 𝑁 defined by
𝑇 (𝑥, 𝑦) = 𝑄(𝑥 + 𝑦) − 𝑄(𝑥) − 𝑄(𝑦)
is 𝑅-bilinear.
The map 𝑇 in (ii) is called the associated bilinear map.
Remark 9.7.2. The bilinearity condition (ii) can be given purely in terms of 𝑄: we
require
𝑄(𝑥 + 𝑦 + 𝑧) = 𝑄(𝑥 + 𝑦) + 𝑄(𝑥 + 𝑧) + 𝑄(𝑦 + 𝑧) − 𝑄(𝑥) − 𝑄(𝑦) − 𝑄(𝑧)
for all 𝑥, 𝑦, 𝑧 ∈ 𝑀.
9.7. QUADRATIC FORMS 145
Definition 9.7.3. A quadratic module over 𝑅 is a quadratic map 𝑄 : 𝑀 → 𝐿 where

𝑀 is a projective 𝑅-module of finite rank and 𝐿 is a projective 𝑅-module of rank 1. A
quadratic form over 𝑅 is a quadratic module with codomain 𝐿 = 𝑅.
A quadratic module 𝑄 : 𝑀 → 𝐿 is free if 𝑀 and 𝐿 are free as 𝑅-modules, and a
quadratic form 𝑄 : 𝑀 → 𝑅 is free if 𝑀 is free as an 𝑅-module.
Example 9.7.4. Let 𝑄 : 𝑉 → 𝐹 be a quadratic form. Let 𝑀 ⊆ 𝑉 be an 𝑅-lattice such
that 𝑄(𝑀) ⊆ 𝐿 where 𝐿 is an invertible 𝑅-module. (When 𝑅 is a Dedekind domain,
we may take 𝐿 = 𝑄(𝑀), see Exercise 9.12.) Then the restriction 𝑄| 𝑀 : 𝑀 → 𝐿 is a
quadratic module over 𝑅.
Conversely, if 𝑄 : 𝑀 → 𝐿 is a quadratic module over 𝑅, then the extension
𝑄 : 𝑀 ⊗𝑅 𝐹 → 𝐿 ⊗𝑅 𝐹 ' 𝐹 is a quadratic form over 𝐹. Moreover, at the slight
cost of some generality (replacing an object by an isomorphic one), by choosing an
isomorphism 𝐿 ⊗𝑅 𝐹 ' 𝐹 we may suppose that 𝑄 takes values in an invertible fractional
ideal 𝔩 ⊆ 𝐹.
Example 9.7.5. If 𝑄 : 𝑀 → 𝐿 is a quadratic module and 𝔞 ⊆ 𝑅 is a projective 𝑅-ideal,
then 𝑄 extends naturally by property (i) to a quadratic module 𝔞𝑀 → 𝔞2 𝐿.
Definition 9.7.6. A similarity between two quadratic modules 𝑄 : 𝑀 → 𝐿 and
𝑄 0 : 𝑀 0 → 𝐿 0 is a pair of 𝑅-module isomorphisms 𝑓 : 𝑀 −→ ∼ 𝑀 0 and ℎ : 𝐿 −→
∼ 𝐿0
such that 𝑄 0 ( 𝑓 (𝑥)) = ℎ(𝑄(𝑥)) for all 𝑥 ∈ 𝑀, i.e., such that the diagram
𝑀
𝑄
/𝐿
o 𝑓 o ℎ (9.7.7)
𝑄0
𝑀0 / 𝐿0
commutes. An isometry between quadratic modules is a similarity with 𝐿 = 𝐿 0 and ℎ

the identity map.
Definition 9.7.8. Let 𝑄 : 𝑀 → 𝐿 be a quadratic module over 𝑅. Then 𝑄 is nonde-
generate if the 𝑅-linear map
𝑇 : 𝑀 → Hom𝑅 (𝑀, 𝐿)
(9.7.9)
𝑥 ↦→ (𝑦 ↦→ 𝑇 (𝑥, 𝑦))
is injective; and 𝑄 is nonsingular (or regular) if the map (9.7.9) is an isomorphism.

Example 9.7.10. If 𝑅 = 𝐹 is a field, then (by linear algebra) 𝑄 is nondegenerate if
and only if 𝑄 is nonsingular.
Example 9.7.11. A quadratic module is nondegenerate if and only if its base extension
𝑄 𝐹 : 𝑀 ⊗𝑅 𝐹 → 𝐿 ⊗𝑅 𝐹 ' 𝐹
is nondegenerate, since the kernel can be detected over 𝐹. Recalling the definition
of discriminant (Definition 4.3.3 for char 𝐹 ≠ 2 and Definition 6.3.1 in general), we
conclude that 𝑄 is nondegenerate if and only if disc 𝑄 𝐹 ≠ 0.
The apparent notion of discriminant of a quadratic module needs some care in

its definition in this generality; it is delayed until section 15.3, where discriminantal
notions are explored in some detail.
Example 9.7.12. Borrowing from the future (see Lemma 15.3.8): if 𝑀 ' 𝑅 𝑛 is free,
then choosing a basis for 𝑀 and computing (half-)discriminant disc 𝑄, we will see that
𝑀 is nonsingular if and only if disc 𝑄 ∈ 𝑅 × .
We now define the notions of genus and classes.
Definition 9.7.13. Let 𝑄 : 𝑀 → 𝐿 be a quadratic module. The genus Gen 𝑄 is the set
of quadratic modules that are locally isometric to 𝑄, i.e., 𝑄 0(𝔭) ∼ 𝑄 (𝔭) for all primes
𝔭 ⊆ 𝑅. The class set Cl 𝑄 is the set of isometry classes in the genus.
We conclude with some comments on the codomain of a quadratic map.
Definition 9.7.14. A quadratic module 𝑄 : 𝑀 → 𝐿 is primitive if 𝑄(𝑀) generates 𝐿
as an 𝑅-module.
9.7.15. If 𝑄 : 𝑅 𝑛 → 𝑅 is a quadratic form, written
∑︁
𝑄(𝑥 1 , . . . , 𝑥 𝑛 ) = 𝑎 𝑖 𝑗 𝑥𝑖 𝑥 𝑗 ∈ 𝑅[𝑥 1 , . . . , 𝑥 𝑛 ],
1≤𝑖 ≤ 𝑗 ≤𝑛
then 𝑄 is primitive if and only if the coefficients 𝑎 𝑖 𝑗 generate the unit ideal 𝑅.
If 𝑅 is a Dedekind domain, then 𝑄(𝑀) ⊆ 𝐿 is again projective (locally at a prime
generated by an element of minimal valuation), so one can always replace 𝑄 : 𝑀 → 𝐿
by 𝑄 : 𝑀 → 𝑄(𝑀) to get a primitive quadratic module; when 𝑅 is a PID, up to
similarity we may divide through by greatest common divisor of the coefficients 𝑎 𝑖 𝑗
in the previous paragraph.
9.7.16. In our admittedly abstract treatment of quadratic modules so far, we have
specifically allowed the codomain of the quadratic map to vary at the same time as the
domain—in particular, we do not ask that they necessarily take values in 𝑅.
Remark 9.7.17. In certain lattice contexts with 𝑅 a Dedekind domain, a quadratic form
with values in a fractional ideal 𝔞 is called an 𝔞-modular quadratic form. Given the
overloaded meanings of the word modular, we do not employ this terminology. In the
geometric context, a quadratic module is called a line-bundle valued quadratic form.
Whatever the terminology, we will see in Chapter 22 that it is important to keep track
of the codomain of the quadratic map just as much as the domain, and in particular we
cannot assume that either is free when 𝑅 is not a PID.
9.8 Normalized form

To conclude this chapter, we discuss an explicit normalized form for quadratic forms.
Let 𝑅 be a local PID; then 𝑅 is either a field or a DVR. In either case, 𝑅 has valuation
𝑣 : 𝑅 → Z ≥0 ∪ {∞} and uniformizer 𝜋; when 𝑅 is a field, we take a trivial valuation
and 𝜋 = 1.
9.8. NORMALIZED FORM 147
Let 𝑄 : 𝑀 → 𝑅 be a quadratic form over 𝑅. Then since 𝑅 is a PID, 𝑀 ' 𝑅 𝑛 is

free. We compute a basis for 𝑀 in which 𝑄 has a particularly nice form, diagonalizing
𝑄 as far as possible. In cases where 2 ∈ 𝑅 × , we can accomplish a full diagonalization;
otherwise, we can at least break up the form as much as possible, as follows. For
𝑎, 𝑏, 𝑐 ∈ 𝑅, the quadratic form 𝑄(𝑥, 𝑦) = 𝑎𝑥 2 + 𝑏𝑥𝑦 + 𝑐𝑦 2 on 𝑅 2 is denoted [𝑎, 𝑏, 𝑐].
Definition 9.8.1. A quadratic form 𝑄 over 𝑅 is atomic if either:
(i) 𝑄 ' h𝑎i for some 𝑎 ∈ 𝑅 × , or

(ii) 2 ∉ 𝑅 × and 𝑄 ' [𝑎, 𝑏, 𝑐] with 𝑎, 𝑏, 𝑐 ∈ 𝑅 satisfying
𝑣(𝑏) < 𝑣(2𝑎) ≤ 𝑣(2𝑐) and 𝑣(𝑎)𝑣(𝑏) = 0.
In case (ii), we necessarily have 𝑣(2) > 0 and 𝑣(𝑏 2 − 4𝑎𝑐) = 2𝑣(𝑏).
Example 9.8.2. Suppose 𝑅 = Z2 is the ring of 2-adic integers, so that 𝑣(𝑥) = ord2 (𝑥)
is the largest power of 2 dividing 𝑥 ∈ Z2 . Recall that Z×2 /Z×2
2 is represented by the
elements ±1, ±5, therefore a quadratic form 𝑄 over Z2 is atomic of type (i) above
if and only if 𝑄(𝑥) ' ±𝑥 2 or 𝑄(𝑥) ' ±5𝑥 2 . For forms of type (ii), the conditions
𝑣(𝑏) < 𝑣(2𝑎) = 𝑣(𝑎) + 1 and 𝑣(𝑎)𝑣(𝑏) = 0 imply 𝑣(𝑏) = 0, and so a quadratic
form 𝑄 over Z2 is atomic of type (ii) if and only if 𝑄(𝑥, 𝑦) ' 𝑎𝑥 2 + 𝑥𝑦 + 𝑐𝑦 2 with
ord2 (𝑎) ≤ ord2 (𝑐). Replacing 𝑥 by 𝑢𝑥 and 𝑦 by 𝑢 −1 𝑦 for 𝑢 ∈ Z×2 we may suppose
𝑎 = ±2𝑡 or 𝑎 = ±5 · 2𝑡 with 𝑡 ≥ 0, and then the atomic representative [𝑎, 1, 𝑐] of the
isomorphism class of 𝑄 is unique.
A quadratic form 𝑄 is decomposable if 𝑄 can be written as the orthogonal sum

of two quadratic forms (𝑄 ' 𝑄 1 𝑄 2 ) and is indecomposable otherwise. It follows
by induction on the rank of 𝑀 that 𝑄 is the orthogonal sum of indecomposable forms.
We will soon give an algorithmic proof of this fact and write each indecomposable
form as a scalar multiple of an atomic form. We begin with the following lemma.
Lemma 9.8.3. An atomic form 𝑄 is indecomposable.
Proof. If 𝑄 is atomic of type (i) then the space underlying 𝑄 has rank 1 and is therefore
indecomposable. Suppose 𝑄 = [𝑎, 𝑏, 𝑐] is atomic of type (ii) and assume for purposes
of contradiction that 𝑄 is decomposable. It follows that if 𝑥, 𝑦 ∈ 𝑀 then 𝑇 (𝑥, 𝑦) ∈ 2𝑅.
Thus we cannot have 𝑣(𝑏) = 0, so 𝑣(𝑎) = 0, and further 𝑣(𝑏) ≥ 𝑣(2) = 𝑣(2𝑎); this
contradicts the fact that 𝑄 is atomic.
Proposition 9.8.4. Let 𝑅 be a local PID and let 𝑄 : 𝑀 → 𝑅 be a quadratic form.

Then there exists a basis of 𝑀 such that the form 𝑄 can be written
𝑄 ' 𝜋 𝑒1 𝑄 1 · · · 𝜋 𝑒𝑛 𝑄 𝑛
where the forms 𝑄 𝑖 are atomic and 0 ≤ 𝑒 1 ≤ · · · ≤ 𝑒 𝑛 ≤ ∞.
In the above proposition, we interpret 𝜋 ∞ = 0. A form as presented in Proposition

9.8.4 is called normalized; this normalized form need not be unique.
Proof. When 𝑅 = 𝐹 is a field with char 𝐹 ≠ 2, we are applying the standard method
of Gram–Schmidt orthogonalization to diagonalize the quadratic form. This argument
can be adapted to the case where 𝑅 = 𝐹 is a field with char 𝐹 = 2, see e.g. Scharlau
[Scha85, §9.4]. For the general case, we make further adaptations to this procedure:
see Voight [Voi2013, Algorithm 3.12] for a constructive (algorithmic) approach.
Exercises
Let 𝑅 be a noetherian domain with field of fractions 𝐹 := Frac 𝑅.
1. Let 𝑉 be a finite-dimensional 𝐹-vector space and let 𝑀, 𝑁 ⊆ 𝑉 be 𝑅-lattices.

Show that 𝑀 + 𝑁 and 𝑀 ∩ 𝑁 are 𝑅-lattices.
I 2. Let 𝐵 be an 𝐹-algebra and let 𝐼 ⊂ 𝐵 be an 𝑅-lattice. Show that there exists a
nonzero 𝑟 ∈ 𝑅 ∩ 𝐼.
3. Give an example of a non-noetherian ring 𝑅 and modules 𝑁 ⊂ 𝑀 such that 𝑀
is finitely generated but 𝑁 is not finitely generated.
4. Let 𝑘 be a field and 𝑅 = 𝑘 [𝑥, 𝑦]. Show that the 𝑅-module (𝑥, 𝑦) is not projective.
5. Let 𝑅 be a Dedekind domain. Show that every ideal of 𝑅 is projective, as
follows. Let 𝔞 ⊆ 𝑅 be a nonzero ideal.
Í𝑛 (The zero ideal is trivially −1 projective.)
Since 𝔞𝔞−1 = 𝑅, we may write 1 = 𝑖=1 𝑎 𝑖 𝑏 𝑖 with 𝑎 𝑖 ∈ 𝔞 and 𝑏 𝑖 ∈ 𝔞 .
Í𝑛
(a) Define the map 𝜙 : 𝑅 𝑛 → 𝔞 by 𝜙(𝑥1 , . . . , 𝑥 𝑛 ) = 𝑖=1 𝑎 𝑖 𝑥𝑖 . Observe that 𝜙
is an 𝑅-module homomorphism, and construct a right inverse 𝜓 to 𝜙, i.e.,
𝜙𝜓 = id𝔞 .
(b) Using (a), show that 𝔞 is a direct summand of 𝑅 𝑛 , so 𝔞 is projective.
6. Let 𝔪 ⊂ 𝑅 be a maximal ideal and let 𝑀 be a finitely generated 𝑅-module. Let
ann𝑅 𝑀 := {𝑟 ∈ 𝑅 : 𝑟𝑥 = 0 for all 𝑥 ∈ 𝑀 }
be the annihilator of 𝑀. Show that 𝑀 (𝔪) = {0} if and only if 𝔪 + ann𝑅 𝑀 = 𝑅.

7. Suppose 𝑅 is a Dedekind domain. Let 𝑉 be a finite-dimensional 𝐹-vector space
and let 𝑀 ⊆ 𝑉 be an 𝑅-lattice. Given a pseudobasis 𝑀 = 𝔞1 𝑥1 ⊕ · · · ⊕ 𝔞𝑛 𝑥 𝑛 as
in (9.3.7), let [𝔞1 · · · 𝔞𝑛 ] ∈ Cl 𝑅. Show that this class (the Steinitz class, 9.3.10)
is well-defined for 𝑀 independent of the choice of pseudobasis.
8. Let 𝑅 be a DVR with maximal ideal 𝔪. Show that if 𝑠 ∉ 𝔪 then 1/𝑠 ∈ 𝑅𝔪 , so
there are natural inclusions
𝑅 ↩→ 𝑅 (𝔪) ↩→ 𝑅𝔪
from the domain into its localization into the completion, inducing isomorphisms
∼ 𝑅 /𝔭𝑒 𝑅 −→
𝑅/𝔭𝑒 −→ ∼ 𝑅 /𝔭𝑒 𝑅
(𝔭) (𝔭) 𝔭 𝔭
for all 𝑒 ≥ 1.
9. Let 𝑉 be a finite-dimensional 𝐹-vector space and let 𝑀, 𝑁 ⊆ 𝑉 be 𝑅-lattices.
9.8. NORMALIZED FORM 149
(a) Show that the index [𝑀 : 𝑁] 𝑅 is a nonzero 𝑅-module. [Hint: use Lemma
9.3.5.]
(b) For 𝛼 ∈ Aut𝐹 (𝑉), show [𝛼𝑀 : 𝑁] = det(𝛼) −1 [𝑀 : 𝑁].
10. Find 𝑅-lattices 𝑀, 𝑁 ⊆ 𝑉 such that [𝑀 : 𝑁] 𝑅 = 𝑅 but 𝑀 ≠ 𝑁.
11. Prove Lemma 9.6.9, as follows. Suppose 𝑅 is a Dedekind domain, and let
𝑁 ⊆ 𝑀 ⊆ 𝑉 be 𝑅-lattices in a finite-dimensional vector space 𝑉 over 𝐹. Prove
that [𝑀 : 𝑁] 𝑅 is the product of the invariant factors (or elementary divisors) of
the torsion 𝑅-module 𝑀/𝑁.
12. Suppose 𝑅 is a Dedekind domain. Let 𝑄 : 𝑉 → 𝐹 be a quadratic form over 𝐹,
let 𝑀 ⊆ 𝑉 be an 𝑅-lattice, and let 𝐿 := 𝑄(𝑀) ⊆ 𝐹 be the 𝑅-submodule of 𝐹
generated by the values of 𝑄. Show that 𝐿 is a fractional 𝑅-ideal.
13. Consider the ternary quadratic form 𝑄(𝑥, 𝑦, 𝑧) = 𝑥𝑦 + 𝑥𝑧 over Z2 . Compute a
normalized form for 𝑄.
14. Consider the following ‘counterexamples’ to Theorem 9.4.9 for more general
integral domains as follows. Let 𝑅 = Q[𝑥, 𝑦] be the polynomial ring in two
variables over Q, so that 𝐹 = Q(𝑥, 𝑦). Let 𝑉 = 𝐹 and 𝐼 = 𝑅.
(a) Show that 𝑦𝑅 has the property that 𝑦𝑅𝔭 ≠ 𝑅𝔭 for infinitely many prime
ideals 𝔭 of 𝑅.
(b) Consider the collection of lattices given by 𝐽𝔭 = 𝑓 (𝑥)𝑅𝔭 if 𝔭 = (𝑦, 𝑓 (𝑥))
where 𝑓 (𝑥) ∈ Q[𝑥] is irreducible and 𝐽𝔭 = 𝑅𝔭 otherwise. Show that
Ñ
𝔭 𝐽𝔭 = (0).
[Instead, to conclude that a collection (𝐽𝔭 )𝔭 of 𝑅𝔭 -lattices arises from a global

𝑅-lattice 𝐽, one needs that the collection forms a sheaf.]
15. In this advanced exercise, we consider generalizations of the notion of lattices to
a geometric context; we assume background in algebraic geometry at the level
of Hartshorne [Har77, Chapter II].
Let 𝑋 be a separated, integral scheme—so for each open 𝑈, the ring 𝒪𝑋 (𝑈) is
a(n integral) domain—and let 𝒪𝑋 be its structure sheaf. Let 𝐹 be the function
field of 𝑋 (so 𝐹 = 𝒪𝑋 ({𝜂}) where 𝜂 is the generic point of 𝑋). Let 𝑉 be a
finite-dimensional 𝐹-vector space.
Define a sheaf of 𝒪𝑋 -lattices in 𝑉 (also called an 𝒪𝑋 -lattice in 𝑉), to be a sheaf
ℳ of 𝒪𝑋 -modules such that for each affine open set 𝑈 ⊆ 𝑋, the set ℳ(𝑈) is an
𝒪𝑋 (𝑈)-lattice in 𝑉. As usual, for 𝑃 ∈ 𝑋 a point, we denote by ℳ( 𝑃) the stalk
of ℳ at 𝑃.
(a) Show that a sheaf of 𝒪𝑋 -lattices in 𝑉 is naturally a subsheaf of the constant

sheaf 𝑉 over 𝑋.
Ð
(b) Let 𝑋 = 𝑖 𝑈𝑖 be an affine open cover of 𝑋, with 𝑈𝑖 = Spec 𝑅𝑖 . Since 𝑋
is separated, each intersection 𝑈𝑖 ∩ 𝑈 𝑗 = Spec 𝑅𝑖 𝑗 is affine, so there are
natural inclusions 𝑅𝑖 , 𝑅 𝑗 ↩→ 𝑅𝑖 𝑗 ⊆ 𝐹 of rings for each 𝑖, 𝑗. Show that a
sheaf of 𝒪𝑋 -lattices is specified uniquely by 𝑅𝑖 -lattices 𝑀𝑖 ⊆ 𝑉 for each 𝑖,
subject to the condition that 𝑀𝑖 𝑅𝑖 𝑗 = 𝑀 𝑗 𝑅𝑖 𝑗 for each 𝑖, 𝑗. [Hint: this is an
easy case of gluing, where isomorphism is replaced by equality in 𝑉.]
(c) Now suppose further that 𝑋 is noetherian, normal, and of dimension ≤ 1

(also called a Dedekind scheme). Then the local rings of 𝑋 at closed
points are DVRs with fraction field 𝐹, and nonempty affine open subsets
of 𝑋 are the complements of finite subsets of closed points and of the form
𝑈 = Spec 𝑅 with 𝑅 an Dedekind domain. (For example, we may take
𝑋 = Spec 𝑅 for 𝑅 a Dedekind domain or 𝑋 a smooth projective integral
curve over a field.)
Extend the local-global dictionary for lattices to 𝑋, in the following way.
Let 𝑈 = Spec 𝑅 ⊆ 𝑋 be a nonempty affine open subset, and let 𝑀 ⊆ 𝑉
be an 𝑅-lattice. Show that the map 𝒩 → (𝒩( 𝑃) ) 𝑃 establishes a bĳection
between 𝒪𝑋 -lattices 𝒩 in 𝑉 and collections of lattices (𝑁 ( 𝑃) ) 𝑃 indexed
by the points 𝑃 ∈ 𝑋, such that for all but finitely many 𝑃 ∈ 𝑈 given by the
prime 𝔭 ⊆ 𝑅, we have 𝑀 (𝔭) = 𝑁 ( 𝑃) ⊆ 𝑉.
Chapter 10
Orders
In this chapter, continuing with a second background installment, we study when

lattices over a domain are closed under a multiplication law: these will be orders, an
integral analogue of algebras over fields.
10.1 ⊲ Lattices with multiplication

We begin with a brief indication of the theory of orders over the integers. Let 𝐵 be a
finite-dimensional Q-algebra. An order O ⊂ 𝐵 is a lattice that is also a subring of 𝐵
(in particular, 1 ∈ O). The property of being an order is a local property for a lattice,
i.e., one may check that it is closed under multiplication in every localization O ( 𝑝) , for
𝑝 prime.
An order is maximal if it is not properly contained
in another order. For example,
𝑎, 𝑏
if we start with the quaternion algebra 𝐵 := with 𝑎, 𝑏 ∈ Z nonzero, then the
Q
lattice
O := Z + Z𝑖 + Z 𝑗 + Z𝑖 𝑗 ⊆ 𝐵
is closed under multiplication, and so defines an order—but it is never a maximal order.
An important construction of lattices comes about as follows: if 𝐼 ⊆ 𝐵 is a lattice,

then
OL (𝐼) := {𝛼 ∈ 𝐵 : 𝛼𝐼 ⊆ 𝐼}
is an order, called the left order of 𝐼; we similarly define the right order.
If O ⊂ 𝐵 is an order and 𝛼 ∈ O, then 𝛼 is integral (over Z), satisfying a monic
polynomial with integer coefficients. If 𝐵 is a quaternion algebra, then 𝛼 ∈ 𝐵 satisfies
its reduced characteristic polynomial of degree 2, and 𝛼 is integral if and only if
trd(𝛼), nrd(𝛼) ∈ Z (Corollary 10.3.6). When 𝐵 = 𝐹 is a number field, the most
important order in 𝐹 is the ring of integers, the set of all integral elements: it is the
unique maximal order.
Unfortunately, this construction does not work in the noncommutative setting: the
set of all integral elements does not form an order. For one thing, if O ⊆ 𝐵 is a
151
152 CHAPTER 10. ORDERS
maximal order and 𝛼 ∈ 𝐵× , then 𝛼O𝛼−1 ⊆ 𝐵 is a maximal order and when 𝐵 is

noncommutative, we may have 𝛼O𝛼−1 ≠ O. But there are more serious problems, as
the following example indicates.

0 1/2 0 0
Example 10.1.1. Let 𝐵 = M2 (Q) and let 𝛼 = and 𝛽 = . Then
0 0 1/2 0
𝛼2 = 𝛽2 = 0, so 𝛼, 𝛽 are integral over 𝑅 = Z, but 𝛼 + 𝛽 and 𝛼𝛽 are not integral since
nrd(𝛼 + 𝛽) = −1/4 and trd(𝛼𝛽) = 1/4. (Such a counterexample does not require the
existence of zerodivisors: see Exercise 10.10.)
Understanding orders in quaternion algebras is a major task of this second part of

the text. In the simplest case 𝐵 = M2 (Q), every maximal order is conjugate (and thus
isomorphic) in 𝐵 to M2 (Z). The reader may wish to skip ahead to Chapter 11 to get
to know the Hurwitz order before returning to study orders more generally.
10.2 Orders
Throughout, let 𝑅 be a domain with field of fractions 𝐹 := Frac(𝑅), and let 𝐵 be
a finite-dimensional 𝐹-algebra. For further reference about orders (as lattices), see
Reiner [Rei2003, Chapter 2] and Curtis–Reiner [CR81, §§23, 26].
Definition 10.2.1. An 𝑅-order O ⊆ 𝐵 is an 𝑅-lattice that is also a subring of 𝐵.
In particular, if O is an 𝑅-order, then since O is a subring we have 1 ∈ O, and since

O is an 𝑅-module we have 𝑅 ⊆ O. We will primarily be concerned with 𝑅-orders that
are projective as 𝑅-modules, and call them projective 𝑅-orders.
10.2.2. An 𝑅-algebra is a ring O equipped with an embedding 𝑅 ↩→ O whose image

lies in the center of O. An 𝑅-order O has the structure of an 𝑅-algebra, and if O is
an 𝑅-algebra that is finitely generated as an 𝑅-module, then O is an 𝑅-order of the
𝐹-algebra 𝐵 = O ⊗𝑅 𝐹.
ExampleÉ10.2.3. The matrix algebra M𝑛 (𝐹) has the 𝑅-order M𝑛 (𝑅). The subring
𝑅[𝐺] = 𝑔 ∈𝐺 𝑅𝑔 is an 𝑅-order in the group ring 𝐹 [𝐺].
Example 10.2.4. Let 𝑎, 𝑏 ∈ 𝑅r{0} and consider the quaternion algebra 𝐵 = (𝑎, 𝑏 | 𝐹).
Then O = 𝑅 ⊕ 𝑅𝑖 ⊕ 𝑅 𝑗 ⊕ 𝑅𝑘 is an 𝑅-order, because it is closed under multiplication
(e.g., 𝑖𝑘 = 𝑖(𝑖 𝑗) = 𝑎 𝑗 ∈ O).
Let 𝐼 ⊆ 𝐵 be an 𝑅-lattice in the 𝐹-algebra 𝐵.
10.2.5. An important construction of orders comes as follows. Let
OL (I) := {𝛼 ∈ 𝐵 : 𝛼𝐼 ⊆ 𝐼}. (10.2.6)
Lemma 10.2.7. OL (I) ⊆ 𝐵 is an 𝑅-order.

10.2. ORDERS 153
Proof. Then OL (𝐼) is an 𝑅-submodule of 𝐵 which is a ring. We show it is also an

𝑅-lattice. For all 𝛼 ∈ 𝐵, by Lemma 9.3.5(b), there exists nonzero 𝑟 ∈ 𝑅 such that
𝑟 (𝛼𝐼) ⊆ 𝐼, hence OL (𝐼)𝐹 = 𝐵. Also by this lemma, there exists nonzero 𝑠 ∈ 𝑅 such
that 𝑠 = 𝑠 · 1 ∈ 𝐼; thus OL (𝐼)𝑠 ⊆ 𝐼 so OL (𝐼) ⊆ 𝑠−1 𝐼. Since 𝑅 is noetherian and 𝑠−1 𝐼 is
an 𝑅-lattice so finitely generated, we conclude that OL (𝐼) is finitely generated and is
thus an 𝑅-lattice.
Definition 10.2.8. The order OL (𝐼) = {𝛼 ∈ 𝐵 : 𝛼𝐼 ⊆ 𝐼} in (10.2.6) is called the left

order of 𝐼. We similarly define the right order of 𝐼 by
OR (𝐼) := {𝛼 ∈ 𝐵 : 𝐼𝛼 ⊆ 𝐼}.
Example 10.2.9. It follows from Lemma 10.2.7 that 𝐵 has an 𝑅-order: the 𝑅-span of
an 𝐹-basis for 𝐵 defines an 𝑅-lattice, so OL (𝐼) is an 𝑅-order. (This is a nice way of
“clearing denominators” from a multiplication table to obtain an order.)
We can read other properties about lattices from their localizations, such as in the
following lemma.
Lemma 10.2.10. Let 𝐵 be a finite-dimensional 𝐹-algebra and let 𝐼 ⊆ 𝐵 be an 𝑅-lattice.

Then the following are equivalent:
(i) 𝐼 is an 𝑅-order;
(ii) 𝐼 (𝔭) is an 𝑅 (𝔭) -order for all primes 𝔭 of 𝑅; and
(iii) 𝐼 (𝔪) is an 𝑅 (𝔪) -order for all maximal ideals 𝔪 of 𝑅.
Proof. For (i) ⇒ (ii) ⇒ (iii), if 𝐼 is an 𝑅-order then 𝐼 (𝔭) is an 𝑅 (𝔭) -order for all primes
𝔭, hence a fortiori for all maximal ideals 𝔪.
To conclude, we prove Ñ (iii) ⇒ (i), and suppose that 𝐼 (𝔪) is an Ñ 𝑅 (𝔪) -order for all
maximal ideals 𝔪. Then Ñ 𝐼
𝔪 (𝔪) = 𝐼 by Lemma 9.4.6. Thus 1 ∈ 𝔪 𝐼 (𝔪) = 𝐼, and for
all 𝛼, 𝛽 ∈ 𝐼 we have 𝛼𝛽 ∈ 𝔪 𝐼 (𝔪) = 𝐼, so 𝐼 is a subring of 𝐵 and hence an order.
Remark 10.2.11. The hypothesis that 𝑅 is noetherian is used in Lemma 10.2.7, but it is
not actually needed; the fact that OL (𝐼) is an order follows by a process often referred
to as noetherian reduction. A basis of 𝐵 yields a multiplication table, consisting
of finitely many elements of 𝐹; moreover, we know that 𝐼 is finitely generated as an
𝑅-module. Writing these generators in terms of a basis we can express these generators
over the basis using finitely many elements of 𝐹. Let 𝑅0 be the subring of 𝑅 generated
by these finitely elements, with field of fractions 𝐹0 , let 𝐵0 be the 𝐹0 -algebra with the
same multiplication table as 𝐵; let 𝐼0 be the 𝑅0 -submodule generated by the generators
for 𝐼 written over 𝑅0 . Then 𝐵 = 𝐵0 ⊗𝐹0 𝐹 and 𝐼 = 𝐼0 ⊗𝑅0 𝑅. But now 𝑅0 is a finitely
generated commutative algebra over its prime ring (the subring generated by 1), so by
the Hilbert basis theorem, 𝑅0 is noetherian. The argument given then shows that 𝐼0 is
finitely generated as an 𝑅0 -module, whence 𝐼 is finitely generated as an 𝑅-module.
Noetherian reduction applies to many results in this text, but non-noetherian rings
are not our primary concern; we retain the noetherian hypothesis for simplicity of
argument and encourage the interested reader to seek generalizations (when they are
possible).
10.3 Integrality
Orders are composed of integral elements, defined as follows. If 𝛼 ∈ 𝐵, we denote by
Í
𝑅[𝛼] = 𝑑 𝑅𝛼 𝑑 the (commutative) 𝑅-subalgebra of 𝐵 generated by 𝛼.
Definition 10.3.1. An element 𝛼 ∈ 𝐵 is integral over 𝑅 if 𝛼 satisfies a monic polyno-

mial with coefficients in 𝑅.
Lemma 10.3.2. For 𝛼 ∈ 𝐵, the following are equivalent:
(i) 𝛼 is integral over 𝑅;

(ii) 𝑅[𝛼] is a finitely generated 𝑅-module;
(iii) 𝛼 is contained in a subring 𝐴 that is finitely generated as an 𝑅-module.
Proof. This lemma is standard; the only extra detail here is to note that in (iii) we do
not need to assume that the subring 𝐴 is commutative: (ii) ⇒ (iii) is immediate taking
𝐴 = 𝑅[𝛼], and for the converse, if 𝐴 ⊆ 𝐵 is a subring that is finitely generated as an
𝑅-module, then 𝑅[𝛼] ⊆ 𝐴 and since 𝑅 is noetherian and 𝐴 is finitely generated as an
𝑅-module, it follows that 𝑅[𝛼] is also finitely generated as an 𝑅-module.
Corollary 10.3.3. If O is an 𝑅-order, then every 𝛼 ∈ O is integral over 𝑅.
10.3.4. We say 𝑅 is integrally closed (in 𝐹) if whenever 𝛼 ∈ 𝐹 is integral over 𝑅,

then in fact 𝛼 ∈ 𝑅. Inside the field 𝐹, the set of elements integral over 𝑅 (the integral
closure of 𝑅 in 𝐹) forms a ring: if 𝛼, 𝛽 are integral over 𝑅 then 𝛼 + 𝛽 and 𝛼𝛽 are
integral since they lie in 𝑅[𝛼, 𝛽] which is a finitely generated submodule of 𝐹. The
integral closure of 𝑅 is itself integrally closed.
Lemma 10.3.5. Suppose that 𝑅 is integrally closed. Then 𝛼 ∈ 𝐵 is integral over 𝑅 if

and only if the minimal polynomial of 𝛼 over 𝐹 has coefficients in 𝑅.
Proof. Let 𝑓 (𝑥) ∈ 𝑅[𝑥] be a monic polynomial that 𝛼 satisfies, and let 𝑔(𝑥) ∈ 𝐹 [𝑥] be
the minimal polynomial of 𝛼. Let 𝐾 be a splitting field for 𝑔(𝑥), and let 𝛼1 , . . . , 𝛼𝑛 be
the roots of 𝑔(𝑥) in 𝐾. Since 𝑔(𝑥) | 𝑓 (𝑥), each such 𝛼𝑖 is integral over 𝑅, and the set
of elements in 𝐾 integral over 𝑅 forms a ring, so each coefficient of 𝑔 is integral over
𝑅 and belongs to 𝐹; but since 𝑅 is integrally closed, these coefficients must belong to
𝑅 and 𝑔(𝑥) ∈ 𝑅[𝑥].
Corollary 10.3.6. If 𝐵 is an 𝐹-algebra with a standard involution, and 𝑅 is integrally

closed, then 𝛼 ∈ 𝐵 is integral over 𝑅 if and only if trd(𝛼), nrd(𝛼) ∈ 𝑅.
We may characterize orders in separable algebras as follows.
Lemma 10.3.7. Let O ⊆ 𝐵 be a subring of a separable 𝐹-algebra 𝐵 such that O𝐹 = 𝐵.

Then O is an 𝑅-order if and only if every 𝛼 ∈ O is integral.
Proof. Let O ⊆ 𝐵 be a subring of an 𝐹-algebra 𝐵 such that O𝐹 = 𝐵. Recall from

Theorem 7.9.4 that a separable 𝐹-algebra is a semisimple 𝐹-algebra such that the
symmetric bilinear pairing (𝛼, 𝛽) ↦→ trd(𝛼𝛽) is nondegenerate.
10.4. MAXIMAL ORDERS 155
We need to show that O is finitely Í generated. Let 𝛼1 , . . . , 𝛼𝑛 be an 𝐹-basis for 𝐵

contained in O. If 𝛽 ∈ÍO then 𝛽 = 𝑖 𝑎 𝑖 𝛼𝑖 with 𝑎 𝑖 ∈ 𝐹. We have 𝛽𝛼𝑖 ∈ O since O is
a ring, so trd(𝛽𝛼𝑖 ) = 𝑗 𝑎 𝑗 trd(𝛼 𝑗 𝛼𝑖 ) with trd(𝛼 𝑗 𝛼𝑖 ) ∈ 𝑅. Now since 𝐵 is separable,
the matrix (trd(𝛼𝑖 𝛼 𝑗 ))𝑖, 𝑗=1,...,𝑛 is invertible, say 𝑟 = det(trd(𝛼𝑖 𝛼 𝑗 )), so we can solve
these equations for 𝑎 𝑗 using Cramer’s rule and we find that 𝑎 𝑗 ∈ 𝑟 −1 𝑅. Consequently
O ⊆ 𝑟 −1 (𝑅𝛼1 ⊕ · · · ⊕ 𝑅𝛼𝑛 ) is a submodule of a finitely generated module so (since 𝑅
is noetherian) O is finitely generated.
10.4 Maximal orders

The integral closure of 𝑅 in 𝐹 is the largest ring containing integral elements. Accord-
ingly, we make the following more general definition.
Definition 10.4.1. An 𝑅-order O ⊆ 𝐵 is maximal if it is not properly contained in
another 𝑅-order.
If 𝐵 is a commutative 𝐹-algebra and 𝑅 is integrally closed in 𝐹, then the integral
closure 𝑆 of 𝑅 in 𝐾 is integrally closed and therefore 𝑆 is a maximal 𝑅-order in
𝐾. However, if 𝐵 is noncommutative, then the set of elements in 𝐵 integral over 𝑅
is no longer necessarily itself a ring, and so the theory of maximal orders is more
complicated. (This may seem counterintuitive at first, but certain aspects of the
noncommutative situation are quite different!) The problem in the noncommutative
setting is that although 𝑅[𝛼] and 𝑅[𝛽] may be finitely generated as 𝑅-modules for
𝛼, 𝛽 ∈ 𝐵, this need not be the case for the 𝑅-algebra generated by 𝛼 and 𝛽.
10.4.2. It follows from Lemma 10.3.7 that a separable 𝐹-algebra 𝐵 has a maximal
𝑅-order, as follows. By Lemma 10.2.7, 𝐵 has an 𝑅-order O (since it has a lattice, taking
the 𝑅-span of an 𝐹-basis), so the collection of 𝑅-orders containing O is nonempty.
Given a chain of 𝑅-orders containing O, by Lemma 10.3.7 the union of these orders is
again an 𝑅-order. Since 𝑅 is noetherian, there exists a maximal element in a chain.
For the rest of this section, we restrict attention and suppose that 𝑅 is a Dedekind
domain. We begin by showing that the property of being a maximal order is a local
property.
Lemma 10.4.3. An 𝑅-order O ⊆ 𝐵 is maximal if and only if O (𝔭) is a maximal
𝑅 (𝔭) -order for all primes 𝔭 of 𝑅.
Proof. If O (𝔭) is maximal for each prime 𝔭 then by Corollary 9.4.7 we see that O is
maximal. Conversely, suppose O is maximal and suppose that O (𝔭) ( O0(𝔭) is a proper
containment of orders for some nonzero prime 𝔭. Then the set O0 = 0
Ñ
𝔮≠𝔭 O𝔮 ∩ O (𝔭)
is an 𝑅-order properly containing O by Lemma 10.2.10 and Theorem 9.4.9.
Lemma 10.4.4. Let O ⊂ 𝐵 be an 𝑅-order. Then for all but finitely many primes 𝔭 of
𝑅, we have that O (𝔭) = O ⊗𝑅 𝑅 (𝔭) is maximal.
Proof. By 10.4.2, there exists a maximal order O0 ⊇ O. By the local-global principle

for lattices (Theorem 9.4.9), we have O𝔭0 = O𝔭 for all but finitely many primes 𝔭.
The structure of (maximal) orders in quaternion algebras over domains of arithmetic

interest is the subject of the second part of this text.
10.5 Orders in a matrix ring

In this section, we study orders in a matrix ring; we restore generality, and let 𝑅 be a
noetherian domain with 𝐹 = Frac 𝑅.
The matrix ring over 𝐹 is just the endomorphism ring of a finite-dimension vector
space over 𝐹, and we seek a similar description for orders as endomorphism rings of
lattices (cf. 10.2.5).
Let 𝑉 be an 𝐹-vector space with dim𝐹 𝑉 = 𝑛 and let 𝐵 = End𝐹 (𝑉). Choosing a
basis of 𝑉 gives an identification 𝐵 = End𝐹 (𝑉) ' M𝑛 (𝐹). Given an 𝑅-lattice 𝑀 ⊆ 𝑉,
we define
End𝑅 (𝑀) := { 𝑓 ∈ End𝐹 (𝑉) : 𝑓 (𝑀) ⊆ 𝑀 } ⊆ 𝐵. (10.5.1)
The left order (10.2.5) is the special case of (10.5.1) where 𝑀 = 𝐼 ⊆ 𝑉 = 𝐵.
Example 10.5.2. If 𝑉 = 𝐹𝑥 1 ⊕ · · · ⊕ 𝐹𝑥 𝑛 and 𝑀 = 𝑅𝑥1 ⊕ · · · ⊕ 𝑅𝑥 𝑛 , then End𝑅 (𝑀) '

M𝑛 (𝑅).
More generally, if 𝑀 is completely decomposable, i.e. 𝑀 = 𝔞1 𝑥1 ⊕ · · · ⊕ 𝔞𝑛 𝑥 𝑛
with each 𝔞𝑖 ⊆ 𝐹 invertible fractional ideals, then we have End𝑅 (𝑀) ⊆ M𝑛 (𝐹) the
subring of matrices whose 𝑖 𝑗-entry belongs to the 𝑅-module
𝔞 𝑗 𝔞−1
𝑖 ' Hom 𝑅 (𝔞𝑖 , 𝔞 𝑗 ) ⊆ Hom𝐹 (𝐹, 𝐹) ' 𝐹
where the isomorphisms come from multiplication. For example, if 𝑛 = 2 then
𝔞2 𝔞−1

𝑅 1
End𝑅 (𝑀) ' ⊆ M2 (𝐹).
𝔞1 𝔞−1
2 𝑅
(Note how the cross terms are aligned

correctly
in the multiplication!) For example, if
𝑅 𝔞−1
𝑀 = 𝑅𝑥1 + 𝔞𝑥2 , then End𝑅 (𝑀) ' .
𝔞 𝑅
Lemma 10.5.3. Let 𝑀 be an 𝑅-lattice of 𝑉. Then End𝑅 (𝑀) is an 𝑅-order in 𝐵 =

End𝐹 (𝑉).
Proof. As in the proof of Lemma 10.2.7, we conclude that End𝑅 (𝑀)𝐹 = 𝐵. Let
𝛼1 , 𝛼2 , . . . , 𝛼𝑛 be an 𝐹-basis for 𝑉 and let 𝑁 = 𝑅𝛼1 ⊕ · · · ⊕ 𝑅𝛼𝑛 . Thus End𝑅 (𝑁) '
M𝑛 (𝑅) is finitely generated as an 𝑅-module.
By Lemma 9.3.5 there exists nonzero 𝑟 ∈ 𝑅 such that 𝑟 𝑁 ⊆ 𝑀 ⊆ 𝑟 −1 𝑁. Therefore,
if 𝜙 ∈ End𝑅 (𝑀), so that 𝜙(𝑀) ⊆ 𝑀, then
(𝑟 2 𝜙) (𝑁) = 𝑟 𝜙(𝑟 𝑁) ⊆ 𝑟 𝜙(𝑀) ⊆ 𝑟 𝑀 ⊆ 𝑁
and thus End𝑅 (𝑀) ⊆ 𝑟 −2 End𝑅 (𝑁); since 𝑅 is noetherian, this implies that End𝑅 (𝑀)
is finitely generated as an 𝑅-module and End𝑅 (𝑀) is an 𝑅-order in 𝐵.
10.5. ORDERS IN A MATRIX RING 157
Lemma 10.5.4. Let O ⊆ 𝐵 = End𝐹 (𝑉) be an 𝑅-order. Then O ⊆ End𝑅 (𝑀) for some
𝑅-lattice 𝑀 ⊆ 𝑉. In particular, if O ⊆ 𝐵 is a maximal 𝑅-order, then O = End𝑅 (𝑀)
for some 𝑅-lattice 𝑀.
Proof. Quite generally, if 𝑁 is any 𝑅-lattice in 𝑉, then 𝑀 = {𝑥 ∈ 𝑁 : O𝑥 ⊆ 𝑁 } is an

𝑅-submodule of 𝑁 with 𝐹 𝑀 = 𝑉 (as in Lemma 10.2.7), thus 𝑀 is an 𝑅-lattice in 𝑉
and O ⊆ End𝑅 (𝑀). If further O is maximal, then the other containment so equality
holds.
Corollary 10.5.5. If 𝑅 is a PID, then every maximal 𝑅-order O ⊆ 𝐵 ' M𝑛 (𝐹) is

conjugate in 𝐵 to M𝑛 (𝑅).
Proof. The Éisomorphism 𝐵 ' M𝑛 (𝐹) arises from a choice of basis 𝑥1 , . . . , 𝑥 𝑛 for 𝑉;
𝑛
letting 𝑁 = 𝑖=1 𝑅𝑥 𝑖 we have End 𝑅 (𝑁) ' M𝑛 (𝑅). The 𝑅-order M𝑛 (𝑅) is maximal
by Exercise 10.6, since a PID is integrally closed.
By Lemma 10.5.4, we have O ⊆ End𝑅 (𝑀) for some 𝑅-lattice 𝑀 ⊆ 𝑉, so if O is
maximal then O = End𝑅 (𝑀). If 𝑅 is a PID then 𝑀 is free as an 𝑅-module, and we
can write 𝑀 = 𝑅𝑦 1 ⊕ · · · ⊕ 𝑅𝑦 𝑛 ; the change of basis matrix from 𝑥 𝑖 to 𝑦 𝑖 then realizes
End𝑅 (𝑀) as a conjugate of End𝑅 (𝑁) ' M𝑛 (𝑅).
Exercises
Let 𝑅 be a noetherian domain with field of fractions 𝐹.
1. Let 𝔠 ⊆ 𝑅 be an ideal. Show that

𝑅 𝑅 𝑎 𝑏
= ∈ M2 (𝑅) : 𝑐 ∈ 𝔠 ⊆ M2 (𝑅)
𝔠 𝑅 𝑐 𝑑
is an 𝑅-order in M2 (𝐹). Note further that if 𝔠 is projective (equivalently, locally

free) as an 𝑅-module, then this 𝑅-order is projective as an 𝑅-module.
2. Let 𝐵 be a finite-dimensional 𝐹-algebra with a standard involution and let O ⊆ 𝐵
be an 𝑅-order. Suppose that 𝑅 is integrally closed in 𝐹. Verify that nrd : O → 𝑅
is a quadratic form over 𝑅.
3. Let O, O0 ⊆ 𝐵 be 𝑅-orders in an 𝐹-algebra 𝐵.
(a) Show that O ∩ O0 is an 𝑅-order.
(b) If O ⊆ O0, show that O0× ∩ O = O× .
4. Let O ⊂ 𝐵 be an 𝑅-order in an 𝐹-algebra 𝐵 and suppose that 𝑅 is integrally
closed. Show that 𝐹 ∩ O = 𝑅.
5. Let 𝐴1 , . . . , 𝐴𝑟 be 𝐹-algebras and let 𝐵 = 𝐴1 × · · · × 𝐴𝑟 . Show that O ⊆ 𝐵 is an
𝑅-order if and only if O is an 𝑅-lattice in 𝐵 and O ∩ 𝐴𝑖 is an 𝑅-order for each 𝑖.
6. Let 𝑅 be integrally closed. Show that M𝑛 (𝑅) is a maximal 𝑅-order in M𝑛 (𝐹).
7. Let 𝐵 = (𝐾, 𝑏 | 𝐹) be a quaternion algebra with 𝑏 ∈ 𝑅 and let 𝑆 be an 𝑅-order
in 𝐾. Let O = 𝑆 + 𝑆 𝑗. Show that O is an 𝑅-order in 𝐵.
8. Suppose that 𝑅 is a PID, and let O ⊆ 𝐵 be an 𝑅-order in the quaternion algebra

𝐵. Let 𝛼 ∈ O be such that 𝑆 = 𝑅[𝛼] is (commutative) domain that is a maximal
𝑅-order in its field of fractions.
(a) Show that 1, 𝛼 extends to an 𝑅-basis for O.
(b) If moreover 𝑆 is a PID, show that there exists 𝛽 ∈ O such that 1, 𝛼, 𝛽, 𝛼𝛽
is an 𝑅-basis for O.
9. Let 𝐵 be an 𝐹-algebra with a standard involution and let 𝛼 ∈ 𝐵. Show that if 𝛼
is integral over 𝑅 then trd(𝛼 𝑛 ) ∈ 𝑅 for all 𝑛 ∈ Z ≥0 . Is the converse true?
10. Generalize Example 10.1.1: Exhibit a division quaternion algebra 𝐵 over Q and
elements 𝛼, 𝛽 ∈ 𝐵 such that 𝛼, 𝛽 are integral over Z but both 𝛼 + 𝛽 and 𝛼𝛽 are
not.
11. Let 𝛼 ∈ M𝑛 (𝐹) have characteristic polynomial with coefficients in 𝑅. Show that
𝛼 is conjugate by an element 𝛽 ∈ GL𝑛 (𝐹) to an element of M𝑛 (𝑅). Explicitly,
how do you find such a matrix 𝛽?
I 12. Let 𝐵 = M𝑛 (𝐹) and let 𝐼 ⊆ 𝐵 be an 𝑅-lattice. Let 𝐼 t = {𝛼t : 𝛼 ∈ 𝐼} be the
transpose lattice. Show that OL (𝐼 t ) = OR (𝐼) t .
I 13. Let 𝐼, 𝐽 ⊆ 𝐵 be 𝑅-lattices. Let 𝐼𝐽 be the 𝑅-submodule of 𝐵 generated by
products 𝛼𝛽 where 𝛼 ∈ 𝐼, 𝛽 ∈ 𝐽; i.e.,
Í 𝑘
𝐼𝐽 := 𝑖=1 𝛼𝑖 𝛽𝑖 : 𝛼𝑖 ∈ 𝐼, 𝛽𝑖 ∈ 𝐽 .
(a) Show that 𝐼𝐽 is an 𝑅-lattice.

(b) Let 𝔭 be a prime of 𝑅. Show that products commute with localization in
the sense that
(𝐼𝐽) ⊗𝑅 𝑅 (𝔭) = (𝐼 ⊗𝑅 𝑅 (𝔭) ) (𝐽 ⊗𝑅 𝑅 (𝔭) ) ⊆ 𝐵 (𝔭) = 𝐵.
I 14. Let O ⊆ 𝐵 be an 𝑅-order in an 𝐹-algebra 𝐵.

(a) Show that OL (O) = OR (O) = O.
(b) Let 𝛼 ∈ 𝐵× , and let 𝛼O = {𝛼𝛽 : 𝛽 ∈ O}. Show that 𝛼O is an 𝑅-lattice
and that OL (𝛼O) = 𝛼O𝛼−1 .
15. Let O ⊆ 𝐵 be an 𝑅-order in an 𝐹-algebra 𝐵. Let 𝛾 ∈ O and let 𝑁 : 𝐵× → 𝐹 ×
be a multiplicative map. Show that 𝛾 ∈ O× if and only if 𝑁 (𝛾) ∈ 𝑅 × , and in
particular, if 𝐵 has a standard involution, then 𝛾 ∈ O× if and only if nrd(𝛾) ∈ 𝑅 × .
Chapter 11
The Hurwitz order
With the preceding chapters on lattices and orders in hand, we are now prepared
to embark on a general treatment of quaternion algebras over number fields and the
arithmetic of their orders. Before we do so, for motivation and pure enjoyment, in this
chapter we consider the special case of the Hurwitz order. Not only is this appropriate
in a historical spirit, it is also instructive for what follows; moreover, the Hurwitz order
has certain exceptional symmetries that make it worthy of specific investigation.
11.1 ⊲ The Hurwitz order

Hurwitz developed the theory of integral quaternions in a treatise [Hur19] in 1919. A
more modern treasure trove of detail about quaternion groups and the Hurwitz order
(as well as many other things) can be found in the book by Conway–Smith [CSm2003];
the review by Baez [Bae2005] also provides an accessible overview.
We consider in this chapter the restriction
of the Hamiltonians from R to Q,
−1, −1
namely, the quaternion algebra 𝐵 = . We further restrict to those elements
Q
with integer coordinates
Zh𝑖, 𝑗i = Z + Z𝑖 + Z 𝑗 + Z𝑘, (11.1.1)
where 𝑘 := 𝑖 𝑗. By Example 10.2.4, this is an order in 𝐵, called the Lipschitz order.
In the rest of this chapter, we will work over Z and so we will simply refer to lattices
and orders.
The Lipschitz order is not a maximal order, and√maximal orders have better
√ prop-
erties. This is analogous to the fact that the ring Z[ −3] is an order in Q( −3) but is
not maximal (not √ integrally closed), properly contained in the better-behaved maximal
order Z[(−1 + −3)/2] of Eisenstein integers. The comparison with the Eisenstein
integers is more than incidental: the element 𝛼 = 𝑖 + 𝑗 + 𝑘 satisfies 𝛼2 + 3 = 0, so it is
natural to consider
−1 + 𝑖 + 𝑗 + 𝑘
𝜔 :=
2
which satisfies 𝜔2 + 𝜔 + 1 = 0. We can enlarge the Lipschitz order to include
𝜔—indeed, this is the only possibility.
159
160 CHAPTER 11. THE HURWITZ ORDER
Lemma 11.1.2. The lattice
O = Z + Z𝑖 + Z 𝑗 + Z𝜔 = Zh𝑖, 𝑗i + Zh𝑖, 𝑗i𝜔 (11.1.3)
in 𝐵 is the unique order that properly contains Zh𝑖, 𝑗i, and O is maximal.
The order O in (11.1.3) is called the Hurwitz order, and it contains Zh𝑖, 𝑗i with
index 2. Note that if 𝛼 ∈ O, then 𝛼 ∈ Zh𝑖, 𝑗i if and only if trd(𝛼) ∈ 2Z.
Proof. By Exercise 11.1, the lattice O is an order.

Suppose that O0 ) Zh𝑖, 𝑗i and let 𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 ∈ O0 with 𝑡, 𝑥, 𝑦, 𝑧 ∈ Q.
Then trd(𝛼) = 2𝑡 ∈ Z, so by Corollary 10.3.6 we have 𝑡 ∈ 12 Z. Similarly, 𝛼𝑖 ∈ O0
therefore trd(𝛼𝑖) = −2𝑥 ∈ Z and 𝑥 ∈ 12 Z, and in the same way 𝑦, 𝑧 ∈ 12 Z. Finally,
nrd(𝛼) = 𝑡 2 + 𝑥 2 + 𝑦 2 + 𝑧 2 ∈ Z, and considerations modulo 4 imply that 𝑡, 𝑥, 𝑦, 𝑧 either
all belong to Z or to 12 + Z; thus 𝛼 ∈ O and so O0 = O.
11.1.4. We can recast this calculation in terms of the local-global dictionary for
lattices (Theorem 9.1.1). Since O[ 12 ] = Zh𝑖, 𝑗i [ 12 ], for every odd prime 𝑝 we have
O ( 𝑝) = Zh𝑖, 𝑗i ( 𝑝) , and O (2) ) Zh𝑖, 𝑗i (2) .
11.2 ⊲ Hurwitz units

We now consider unit groups; in this section, we take 𝑘 := 𝑖 𝑗. An element 𝛾 =
𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 ∈ Zh𝑖, 𝑗i is a unit if and only if nrd(𝛾) = 𝑡 2 + 𝑥 2 + 𝑦 2 + 𝑧2 ∈ Z× ,
i.e. nrd(𝛾) = 1, and since 𝑡, 𝑥, 𝑦, 𝑧 ∈ Z we immediately have
Zh𝑖, 𝑗i × = {±1, ±𝑖, ± 𝑗, ±𝑘 } ' 𝑄 8
is the quaternion group of order 8. In a similar way, taking 𝛾 ∈ O in the Hurwitz

order and allowing 𝑡, 𝑥, 𝑦, 𝑧 ∈ 21 Z so that 2𝑡, 2𝑥, 2𝑦, 2𝑧 all have the same parity, we find
that
O× = 𝑄 8 ∪ (±1 ± 𝑖 ± 𝑗 ± 𝑘)/2
is a group of order 24.
We have O× ; 𝑆4 (the symmetric group on 4 letters) because there is no embedding
𝑄 8 ↩→ 𝑆4 . (The permutation representation 𝑄 8 → 𝑆4 obtained by the action on the
cosets of the unique subgroup h−1i of index 4 factors through the quotient 𝑄 8 →
𝑄 8 /{±1} ' 𝑉4 ↩→ 𝑆4 , where 𝑉4 is the Klein 4-group.) There are 15 groups of order
24 up to isomorphism! We identify the right one as follows.
Lemma 11.2.1. We have O× ' SL2 (F3 ).
Proof. We reduce modulo 3. There is a ring homomorphism

−1, −1
O → O/3O ' F3 h𝑖, 𝑗i ' .
F3
Any quaternion algebra over a finite field is isomorphic to the matrix ring by Wed-
derburn’s little theorem (Exercises 3.16, 6.16, and 7.29). Specifically, the element
11.2. ⊲ HURWITZ UNITS 161
𝜖 = 𝑖 + 𝑗 + 𝑘 has 𝜖 2 = 0 ∈ O/3O. The left ideal 𝐼 generated by 𝜖 is an F3 -vector space,

and we compute that it has basis 𝜖 and 𝑖𝜖 = −1 − 𝑗 + 𝑘. As in (7.6.3) (Proposition
7.6.2) this yields an isomorphism
O/3O → M2 (F3 )

0 −1 1 1
𝑖, 𝑗 ↦→ ,
1 0 1 −1
(a statement that can be explicitly and independently verified in Exercise 11.4). We

obtain a group homomorphism O× → SL2 (F3 ), since the reduced norm corresponds
to the determinant and nrd(O× ) = {1}, and this homomorphism is injective because if
𝛾 ∈ O× has 𝛾 − 1 ∈ 3O then 𝛾 = 1, by inspection. Since #O× = # SL2 (F3 ) = 24, the
map O× ↩→ SL2 (F3 ) is an isomorphism.
11.2.2. The group SL2 (F3 ) acts on the left on the set of nonzero column vectors
F23 up to sign, a set of cardinality (9 − 1)/2 = 4. (More generally, SL2 (F 𝑝 ) acts
on P1 (F 𝑝 ) := (F2𝑝 r {(0, 0)})/F×𝑝 , a set of cardinality 𝑝 + 1.) This action yields a
permutation representation SL2 (F3 ) → 𝑆4 ; the kernel of this map is the subgroup
generated by the scalar matrix −1 and so the representation gives an injective group
homomorphism from PSL2 (F3 ) := SL2 (F3 )/{±1} into 𝑆4 . Since 𝐴4 ≤ 𝑆4 is the
unique subgroup of size 24/2 = 12, we must have PSL2 (F3 ) ' 𝐴4 , giving an exact
sequence
1 → {±1} → O× → 𝐴4 → 1. (11.2.3)
11.2.4. We can also visualize the group O× and the exact sequence (11.2.3), thinking
of the Hamiltonians as acting by rotations (section 2.4). Recall there is an exact
sequence (Corollary 2.4.21)
1 → {±1} → H1 → SO(3) → 1 (11.2.5)
obtained by the left action 𝛼 ↦→ 𝛼𝑣𝛼−1 for 𝛼 ∈ H1 and 𝑣 ∈ H0 ' R3 ; specifically, by

Proposition 2.4.18, a quaternion 𝛼 = cos 𝜃 + 𝐼 (𝛼) sin 𝜃 acts by rotation through the
angle 2𝜃 about the axis 𝐼 (𝛼).
We have been considering

−1, −1 −1, −1
O ↩→ 𝐵 = ↩→ 𝐵 ⊗Q R = = H, (11.2.6)
Q R
and we now consider the corresponding embedding of groups O1 = O× ↩→ H1 . We

are led to think of the group O× /{±1} ' 𝐴4 as the group of symmetries (rigid motions)
of a tetrahedron (or rather, a tetrahedron and its dual), as in Figure 11.2.7.
−i − j + k
i+j+k
j
i −i + j − k y
i−j−k
x
Figure 11.2.7: Symmetries of a tetrahedron, viewed quaternionically
Inside the cube in R3 with vertices (±1, ±1, ±1) = ±𝑖 ± 𝑗 ± 𝑘, we can find four
inscribed tetrahedra, for example, the tetrahedron 𝑇 with vertices
𝑖 + 𝑗 + 𝑘, 𝑖 − 𝑗 − 𝑘, −𝑖 + 𝑗 − 𝑘, −𝑖 − 𝑗 + 𝑘.
Then the elements ±𝑖, ± 𝑗, ±𝑘 act by rotation about the 𝑥, 𝑦, 𝑧 axes by an angle 𝜋 (so
interchanging points with the same 𝑥, 𝑦, 𝑧 coordinate). The element ±𝜔 = ±(−1 + 𝑖 +
𝑗 + 𝑘)/2 rotates by the angle 2𝜋/3 fixing the point (1, 1, 1) and cyclically permuting the
other three points, and by symmetry we understand the action of the other elements of
O× . We therefore call O× the binary tetrahedral group. Following Conway–Smith
[CSm2003, §3.3], we also write 2𝑇 = O× for this group; the notation 𝐴 e4 is also used.
The subgroup 𝑄 8 E 2𝑇 is normal (as it is characteristic, consisting of all elements
of O of order dividing 4), and so we can write 2𝑇 = 𝑄 8 o h𝜔i where h𝜔i ' Z/3Z acts
on 𝑄 8 by conjugation, cyclically rotating the elements 𝑖, 𝑗, 𝑘. Finally, the group 2𝑇
has a presentation (Exercise 11.7)
2𝑇 ' h𝑟, 𝑠, 𝑡 | 𝑟 2 = 𝑠3 = 𝑡 3 = 𝑟 𝑠𝑡 = −1i (11.2.8)
via 𝑟 = 𝑖, 𝑠 = 𝜔 = (−1 + 𝑖 + 𝑗 + 𝑘)/2, and 𝑡 = (−1 + 𝑖 + 𝑗 − 𝑘)/2.
We conclude by noting that the difference between the Lipschitz and Hurwitz
orders is “covered” by the extra units.
Lemma 11.2.9. For every 𝛽 ∈ O, there exists 𝛾 ∈ O× such that 𝛽𝛾 ∈ Zh𝑖, 𝑗i.
11.3. ⊲ EUCLIDEAN ALGORITHM 163
Proof. If 𝛽 ∈ Zh𝑖, 𝑗i already, then we are done. Otherwise, 2𝛽 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 with

all 𝑡, 𝑥, 𝑦, 𝑧 ∈ Z odd. Choosing matching signs, there exists 𝛾 ∈ O× such that 2𝛽 ≡ 2𝛾
(mod 4O). Thus
(2𝛽)𝛾 −1 ≡ 2 (mod 4O)
so 𝛽𝛾 −1 ∈ Z + 2O = Zh𝑖, 𝑗i, so we may take 𝛾 −1 for the statement of the lemma.
11.3 ⊲ Euclidean algorithm

√
The Eisenstein order Z[(−1 + −3)/2] has several nice properties. Perhaps nicest of
all √
is that it is a Euclidean domain, so in particular it is a PID and UFD. (Alas, the ring
Z[ −3] just fails to be Euclidean.)
11.3.1. The Hurwitz order also has a left (or right) Euclidean algorithm generalizing
the commutative case, as follows. There is an embedding 𝐵 ↩→ 𝐵 ⊗Q R ' H, and
inside H ' R4 the Hurwitz order sits as a (Z-)lattice equipped with the Euclidean
inner product, so we can think of the reduced norm by instead thinking of distance.
In the Lipschitz order, we see by rounding coordinates that for all 𝛾 ∈ 𝐵 there exists
𝜇 ∈ Zh𝑖, 𝑗i such that nrd(𝛾 − 𝜇) ≤ 4 · (1/2) 2 = 1—a farthest point occurs at the
center (1/2, 1/2, 1/2, 1/2) of a unit cube. But this is precisely the point where the
Hurwitz quaternions occur, and it follows that for all 𝛾 ∈ 𝐵, there exists 𝜇 ∈ O such
that nrd(𝛾 − 𝜇) < 1. (In fact, we can take nrd(𝛾 − 𝜇) ≤ 1/2; see Exercise 11.8.)
Paragraph 11.3.1 becomes a right Euclidean algorithm as in the commutative case.
Lemma 11.3.2 (Hurwitz order is right norm Euclidean). For all 𝛼, 𝛽 ∈ O with 𝛽 ≠ 0,
there exists 𝜇, 𝜌 ∈ O such that
𝛼 = 𝛽𝜇 + 𝜌 (11.3.3)
and nrd(𝜌) < nrd(𝛽).
Proof. If nrd(𝛼) < nrd(𝛽), we may take 𝜇 = 0 and 𝜌 = 𝛼, so suppose nrd(𝛼) ≥

nrd(𝛽) > 0. Let 𝛾 = 𝛽−1 𝛼 ∈ 𝐵. Then by 11.3.1, there exists 𝜇 ∈ O such that
nrd(𝛾 − 𝜇) < 1. Let 𝜌 = 𝛼 − 𝛽𝜇. Then by multiplicativity of the norm,
nrd(𝜌) = nrd(𝛼 − 𝛽𝜇) < nrd(𝛽).
A similar statement to Lemma 11.3.2 holds for division on the left, i.e., in (11.3.3)
we may take 𝛼 = 𝜇𝛽 + 𝜌 (with possibly different elements 𝜇, 𝜌 ∈ O, of course).
Proposition 11.3.4. Every right ideal 𝐼 ⊆ O is right principal, i.e., there exists 𝛽 ∈ 𝐼
such that 𝐼 = 𝛽O.
Proof. Let 𝐼 ⊆ O be a right ideal. If 𝐼 = {0}, we are done. Otherwise, there exists an
element 0 ≠ 𝛽 ∈ 𝐼 with minimal reduced norm nrd(𝛽) ∈ Z>0 . We claim that 𝐼 = 𝛽O.
For all 𝛼 ∈ 𝐼, by the left Euclidean algorithm in Lemma 11.3.2, there exists 𝜇 ∈ O
such that 𝛼 = 𝜇𝛽 + 𝜌 with nrd(𝜌) < nrd(𝛽); but 𝜌 = 𝛼 − 𝛽𝜇 ∈ 𝐼, so by minimality,
nrd(𝜌) = 0 and 𝜌 = 0, hence 𝛼 = 𝛽𝜇 ∈ 𝛽O as claimed.
Definition 11.3.5. Let 𝛼, 𝛽 ∈ O. We say 𝛽 right divides 𝛼 (or 𝛼 is a right multiple

of 𝛽) and write 𝛽 |R 𝛼 if there exists 𝛾 ∈ O such that 𝛼 = 𝛽𝛾.
A right common divisor of 𝛼, 𝛽 ∈ O is an element 𝛾 ∈ O such that 𝛾 |R 𝛼, 𝛽. A
right greatest common divisor of 𝛼, 𝛽 is a common divisor 𝛾 such that 𝛿 |R 𝛾 for all
common divisors 𝛿 of 𝛼, 𝛽.
It follows from Lemma 11.3.2 in the same way as in the commutative case that if
𝛼, 𝛽 are not both zero, then there exists a greatest common divisor of 𝛼, 𝛽, taking the
last nonzero remainder in the right Euclidean algorithm.
Corollary 11.3.6 (Bézout’s theorem). For all 𝛼, 𝛽 ∈ O not both zero, there exists
𝜇, 𝜈 ∈ O such that 𝛼𝜇 + 𝛽𝜈 = 𝛾 where 𝛾 is a right greatest common divisor of 𝛼, 𝛽.
Proof. By Proposition 11.3.4, we may write 𝛼O + 𝛽O = 𝛾O for some 𝛾 ∈ O, and then
𝛾 ∈ 𝛼O + 𝛽O implies there exists 𝜇, 𝜈 ∈ O such that 𝛼𝜇 + 𝛽𝜈 = 𝛾.
Proposition 11.3.7. Let O0 ⊂ 𝐵 be a maximal order. Then there exists 𝛼 ∈ 𝐵× such
that O0 = 𝛼−1 O𝛼, and in particular O0 ' O as rings.
Proof. By clearing denominators, there exists nonzero 𝑎 ∈ Z such that 𝑎O0 ⊆ O.
Let 𝐼 = 𝑎O0O be the right ideal of O generated by 𝑎O0. Then O0 ⊆ OL (𝐼), and
equality holds since O0 is maximal. By Proposition 11.3.4, we have 𝐼 = 𝛽O for some
𝛽 ∈ 𝐵× . We have OL (𝐼) = 𝛽O𝛽−1 by Exercise 10.14, so O0 = 𝛽O𝛽−1 and we may
take 𝛼 = 𝛽−1 .
Example 11.3.8. The Lipschitz order Zh𝑖, 𝑗i does not enjoy the property that every
right ideal is principal, as the following example shows.
Let 𝐼 = 2O = 2Z + 2𝑖Z + 2 𝑗Z + (1 + 𝑖 + 𝑗 + 𝑘)Z. Then 𝐼 ⊆ Zh𝑖, 𝑗i and 𝐼 has
the structure of a right Zh𝑖, 𝑗i-ideal, in fact 𝐼 = 2Zh𝑖, 𝑗i + (1 + 𝑖 + 𝑗 + 𝑘)Zh𝑖, 𝑗i. We
claim that 𝐼 is not principal as a right Zh𝑖, 𝑗i-ideal. Indeed, suppose 𝐼 = 𝛼Zh𝑖, 𝑗i with
𝛼 ∈ 𝐼. Since 𝛼 ∈ 2O, we have 4 | nrd(𝛼). But 2 ∈ 𝐼 so 2 = 𝛼𝛽 with 𝛽 ∈ Zh𝑖, 𝑗i, so
4 = nrd(2) = nrd(𝛼) nrd(𝛽), whence nrd(𝛼) = 4 and nrd(𝛽) = 1 so 𝛽 ∈ Zh𝑖, 𝑗i × and
so 2O = 𝐼 = 𝛼Zh𝑖, 𝑗i = 2Zh𝑖, 𝑗i. Cancelling the factor 2, we conclude O = Zh𝑖, 𝑗i, a
contradiction.
For more, see Exercise 11.11.
11.4 ⊲ Unique factorization

It does not follow that there is unique factorization in O in the traditional sense, as the
order of multiplication matters. Nevertheless, there is a theory of prime factorization
in O as follows.
Lemma 11.4.1. Let 𝑝 be prime. Then there exists 𝜋 ∈ O such that 𝜋𝜋 = nrd(𝜋) = 𝑝.
Proof. We have nrd(1 + 𝑖) = 12 + 12 = 2, so we may suppose 𝑝 ≥ 3 is odd. Then
O/𝑝O ' (−1, −1 | F 𝑝 ) ' M2 (F 𝑝 ) by Wedderburn’s little theorem, and there exists a
right ideal 𝐼 mod 𝑝 ⊂ O/𝑝O with dimF 𝑝 (𝐼 mod 𝑝) = 2. Let
𝐼 = {𝛼 ∈ O : 𝛼 mod 𝑝 ∈ 𝐼 mod 𝑝}
11.4. ⊲ UNIQUE FACTORIZATION 165
be the preimage of 𝐼 mod 𝑝 in the map O → O/𝑝O. Then 𝑝O ( 𝐼 ( O. Then 𝐼 ⊂ O

is a right ideal, and 𝐼 ≠ O. But 𝐼 = 𝛽O is right principal by Proposition 11.3.4.
We claim that nrd(𝛽) = 𝑝. Since 𝑝 ∈ 𝐼, we have 𝑝 = 𝛽𝜇 for some 𝜇 ∈ O,
whence nrd( 𝑝) = 𝑝 2 = nrd(𝛽) nrd(𝜇) so nrd(𝛽) | 𝑝 2 . We cannot have nrd(𝛽) = 1 or
nrd(𝛽) = 𝑝 2 , as these would imply 𝐼 = O or 𝐼 = 𝑝O, impossible. We conclude that
nrd(𝛽) = 𝑝.
Remark 11.4.2. Once we have developed a suitable theory of norms, the proof that
nrd(𝛽) = 𝑝 above will be immediate: if we define N (𝐼) := #(O/𝐼) then N (𝐼) = 𝑝 2 by
construction, and it turns out that N (𝐼) = nrd(𝛽) 2 .
Theorem 11.4.3 (Lagrange). Every integer 𝑛 ≥ 0 is the sum of four squares, i.e., there
exist 𝑡, 𝑥, 𝑦, 𝑧 ∈ Z such that 𝑛 = 𝑡 2 + 𝑥 2 + 𝑦 2 + 𝑧2 .
Proof. We seek an element 𝛽 ∈ Zh𝑖, 𝑗i such that nrd(𝛽) = 𝑛. By multiplicativity of

the reduced norm, it is sufficient to treat the case where 𝑛 = 𝑝 is prime. We obtain
𝜋 ∈ O such that nrd(𝜋) = 𝑝 by Lemma 11.4.1. But now the result follows from
Lemma 11.2.9, as there exists 𝛾 ∈ O× such that 𝜋𝛾 ∈ Zh𝑖, 𝑗i.
Remark 11.4.4. A counterpart to Lagrange’s theorem (Theorem 11.4.3) is the following

theorem of Legendre and Gauss on sums of three squares: Every integer 𝑛 that is not
of the form 𝑛 = 4𝑎 𝑚 with 𝑚 ≡ 7 (mod 8) can be written as the sum of three squares
𝑛 = 𝑥 2 + 𝑦 2 + 𝑧 2 . We will revisit this classical theorem in Chapter 30 as motivation
for the study of embedding numbers, and the number of such representations will be
given in terms of class numbers, following Gauss. A direct proof of the three square
theorem is given by Mordell [Mor69, §20, Theorem 1], but he notes that “no really
elementary treatment [of this theorem] is known”.
We finish this section with a discussion of ‘unique factorization’ in the Hurwitz
order.
Definition 11.4.5. An element 𝜋 ∈ O is irreducible if whenever 𝜋 = 𝛼𝛽 with 𝛼, 𝛽 ∈ O

then either 𝛼 ∈ O× or 𝛽 ∈ O× .
Lemma 11.4.6. Let 𝜋 ∈ O. Then 𝜋 is irreducible if and only if nrd(𝜋) = 𝑝 ∈ Z is

prime.
Proof. If nrd(𝜋) = 𝑝 is prime and 𝜋 = 𝛼𝛽 then nrd(𝜋) = 𝑝 = nrd(𝛼) nrd(𝛽) so

either nrd(𝛼) = 1 or nrd(𝛽) = 1, thus 𝛼 ∈ O× or 𝛽 ∈ O× . Conversely, suppose 𝜋 is
irreducible and let 𝑝 | nrd(𝜋). Let 𝐼 = 𝜋O + 𝑝O = 𝛼O. Then nrd(𝛼) | nrd( 𝑝) = 𝑝 2 .
We cannot have nrd(𝛼) = 1, as every element of 𝐼 has reduced norm divisible by
𝑝. We similarly cannot have nrd(𝛼) = 𝑝 2 , since this would imply 𝜋 ∈ 𝑝O; but
by Lemma 11.4.1, 𝑝 is reducible, a contradiction. We conclude that nrd(𝛼) = 𝑝.
From 𝜋 ∈ 𝐼 = 𝛼O we obtain 𝜋 = 𝛼𝛽 with 𝛽 ∈ O; by irreducibility, 𝛽 ∈ O× and
nrd(𝜋) = nrd(𝛼) = 𝑝.
Definition 11.4.7. An element 𝛼 ∈ O is primitive if 𝛼 ∉ 𝑛O for all 𝑛 ∈ Z ≥2 .

Theorem 11.4.8 (Conway–Smith). Let 𝛼 ∈ O be primitive and let 𝑎 = nrd(𝛼). Factor

𝑎 = 𝑝 1 𝑝 2 · · · 𝑝 𝑟 into a product of primes. Then there exists 𝜋1 , 𝜋2 , . . . , 𝜋𝑟 ∈ O such
that
𝛼 = 𝜋1 𝜋2 · · · 𝜋𝑟 , and nrd(𝜋𝑖 ) = 𝑝 𝑖 for all 𝑖. (11.4.9)
Moreover, every other such factorization is of the form
𝛼 = (𝜋1 𝛾1 ) (𝛾1−1 𝜋2 𝛾2 ) · · · (𝛾𝑟−1−1 𝜋𝑟 ) (11.4.10)
where 𝛾1 , . . . , 𝛾𝑟 ∈ O× .
Proof. Let 𝐼 = 𝛼O + 𝑝 1 O; as in the proof of Lemma 11.4.6, we find 𝐼 = 𝜋1 O with

nrd(𝜋1 ) = 𝑝 1 , arguing that nrd(𝜋1 ) ≠ 𝑝 21 since 𝛼 ∈ 𝑝 1 O is in contradiction to 𝛼 being
primitive. Then 𝜋1 is unique up to right multiplication by a unit and 𝛼 = 𝜋1 𝛼2 . The
result then follows by induction.
The factorization (11.4.10) is said to be obtained from 𝛼 = 𝜋1 · · · 𝜋𝑟 by unit

migration.
Remark 11.4.11. To look at all possible prime factorizations of 𝛼 as in (11.4.9), it is
necessary to consider the possible factorizations 𝑎 = 𝑝 1 · · · 𝑝 𝑟 . Conway–Smith call
this process metacommutation [CSm2003, Chapter 5]; metacommutation is analyzed
by Cohn–Kumar [CK2015], Forsyth–Gurev–Shrima FGS, and in a very general context
by Chari [Cha2020].
11.5 Finite quaternionic unit groups

We conclude this section by a discussion of quaternion unit groups extending the
discussion 11.2: we classify finite subgroups of H× and realize the possible subgroups
as coming from quaternionic unit groups.
11.5.1. To begin with the classification, suppose that Γ ⊆ H× is a finite subgroup.

Then nrd(Γ) is a finite subgroup of R×>0 , hence identically 1, so Γ ⊆ H1 .
Similarly, if Γ ⊆ H× /R× ' H1 /{±1} is a finite subgroup, then it lifts via the
projection H1 → H1 /{±1} to a finite subgroup of H1 .
So let Γ ⊆ H1 be a finite subgroup. Then
Γ/{±1} ↩→ H1 /{±1} ' SO(3)
the latter isomorphism by Hamilton’s original (!) motivation for quaternion algebras
(Corollary 2.4.21). Therefore Γ/{±1} ⊆ SO(3) is a finite rotation group, and these
groups have been known since antiquity.
Proposition 11.5.2. A finite subgroup of SO(3) is one of the following:
(i) a cyclic group;

(ii) a dihedral group;
(iii) the tetrahedral group 𝐴4 of order 12;
11.5. FINITE QUATERNIONIC UNIT GROUPS 167
(iv) the octahedral group 𝑆4 of order 24; or

(v) the icosahedral group 𝐴5 of order 60.
Cases (iii)–(v) are the symmetry groups of the corresponding Platonic solids and
are called exceptional rotation groups.
Proof. Let 𝐺 ≤ SO(3) be a finite subgroup with #𝐺 = 𝑛 > 1; then 𝐺 must consist of
rotations about a common fixed point (its center of gravity), which we may take to be
the origin. The group 𝐺 then acts on the unit sphere, and every nonidentity element
of 𝐺 acts by rotation about an axis, fixing the poles of its axis on the sphere. Let 𝑉 be
the subset of these poles in the unit sphere; the set 𝑉 will soon be the vertices of our
(possibly degenerate) polyhedron. Let
𝑋 = {(𝑔, 𝑣) : 𝑔 ∈ 𝐺 r {1} and 𝑣 is a pole of 𝑔}.
Since each 𝑔 ∈ 𝐺 r {1} has exactly two poles, we have #𝑋 = 2(𝑛 − 1). On the other
hand, we can also count organizing by orbits. Choose a representative set 𝑣 1 , . . . , 𝑣 𝑟
of poles, one from each orbit of 𝐺 on 𝑉, and let
𝑛𝑖 = # Stab𝐺 (𝑣 𝑖 ) = #{𝑔 ∈ 𝐺 : 𝑔𝑣 𝑖 = 𝑣 𝑖 }
be the order of the stabilizer: this group is a cyclic subgroup about a common axis.
Then
𝑟 𝑟 𝑟
∑︁ ∑︁ 𝑛 ∑︁ 1
2𝑛 − 2 = #𝑋 = #(𝐺𝑣 𝑖 ) (𝑛𝑖 − 1) = (𝑛𝑖 − 1) = 𝑛 1− ,
𝑖=1
𝑛
𝑖=1 𝑖 𝑖=1
𝑛𝑖
by the orbit–stabilizer theorem. Dividing both sides by 𝑛 gives

𝑟
2 ∑︁ 1
2− = 1− . (11.5.3)
𝑛 𝑖=1 𝑛𝑖
Since 𝑛 > 1, we have 1 ≤ 2 − 2/𝑛 < 2; and since each 𝑛𝑖 ≥ 2, we have 1/2 ≤
1 − 1/𝑛𝑖 < 1. Putting these together, we must have 𝑟 = 2, 3.
If 𝑟 = 2, then (11.5.3) becomes 2 = 𝑛/𝑛1 + 𝑛/𝑛2 , with 𝑛/𝑛𝑖 = #(𝐺𝑣 𝑖 ) ≥ 1, so
𝑛1 = 𝑛2 = 𝑛, there is only one axis of rotation, and 𝐺 is cyclic.
If 𝑟 = 3, then the only possibilities for (𝑛1 , 𝑛2 , 𝑛3 ) with 𝑛1 ≤ 𝑛2 ≤ 𝑛3 are
(2, 2, 𝑐), (2, 3, 3), (2, 3, 4), (2, 3, 5); the corresponding groups have sizes 2𝑐, 12, 24, 60,
respectively, and can be identified with 𝐷 2𝑐 , 𝐴4 , 𝑆4 , 𝐴5 by a careful but classical anal-
ysis of orbits. See Armstrong [Arm88, Chapter 19], Grove–Benson [GB2008, §2.4],
or Conway–Smith [CSm2003, §3.3].
In 11.2.4, we gave a quaternionic visualization of the binary tetrahedral group

(lifting the tetrahedral group to H1 ); we repeat this with the two other exceptional
rotation groups, taking again 𝑘 := 𝑖 𝑗.
11.5.4. The octahedral group 𝑆4 pulls back to the binary octahedral group 2𝑂 ⊆ H1
of order 24 · 2 = 48, whose elements act by rigid motions of the octahedron (or dually,
the cube). We make identifications following 11.2.4, shown in Figure 11.5.5.
−i − j + k
−i + j + k
k
−i
i−j+k
−j
j
i −i + j − k y
−k
i−j−k
x i+j−k
Figure 11.5.5: Symmetries of an octahedron and a cube, viewed quaternionically
The binary tetrahedral group 2𝑇 E 2𝑂 of order 24 acts as a subgroup of rigid

motions; the group 2𝑂 is generated by an element which maps to a rotation of order 4
around the 6 faces, i.e., one of the 12 elements
±1 ± 𝑖 ±1 ± 𝑗 ±1 ± 𝑘
√ , √ , √ .
2 2 2
The group 2𝑂 has a Coxeter presentation
2𝑂 ' h𝑟, 𝑠, 𝑡 | 𝑟 2 = 𝑠3 = 𝑡 4 = 𝑟 𝑠𝑡 = −1i
(with −1 central√and (−1) 2 = 1).√One also writes 2𝑂 ' 𝑆e4 .

Let 𝐹
= Q( 2) and 𝑅 = Z[ 2]. If we consider the Hamiltonians restricted to 𝐹
−1, −1
as 𝐵 = , then the group 2𝑂 ⊆ H1 generates an 𝑅-order: letting 𝑖, 𝑗 be the
𝐹 √ √
standard generators and still 𝑘 := 𝑖 𝑗, and letting 𝛼 = (1 + 𝑖)/ 2 and 𝛽 = (1 + 𝑗)/ 2,
then
O2𝑂 = 𝑅 + 𝑅𝛼 + 𝑅𝛽 + 𝑅𝛼𝛽; (11.5.6)
this order contains the scalar extension of the Hurwitz order to 𝑅 and is in fact a
maximal 𝑅-order. (The extension of scalars is necessary: 𝑆4 contains an element
of order 4 which √ lifts to an element of order 8 in 2𝑂; such an element has trace
±(𝜁8 + 𝜁8−1 ) = ± 2.)
11.5.7. Finally, we treat the binary icosahedral group 2𝐼 ⊆ H1 of order 60 · 2 = 120,

acting by rigid motions of the icosahedron (or dually, the dodecahedron). We choose
the regular icosahedron to have vertices at
±𝑖 ± 𝑗 ± 𝑘, ±𝜏𝑖 ± 𝜏 −1 𝑗, ±𝜏 𝑗 ± 𝜏 −1 𝑘, ±𝜏𝑘 ± 𝜏 −1 𝑖
√
where 𝜏 = (1 + 5)/2 is the golden ratio. The elements of order 5 are given by
conjugates and powers of the element 𝜁 = (𝜏 + 𝜏 −1𝑖 + 𝑗)/2, which acts by rotation
about a face. The group 2𝐼 can be presented as
2𝐼 ' h𝑟, 𝑠, 𝑡 | 𝑟 2 = 𝑠3 = 𝑡 5 = 𝑟 𝑠𝑡 = −1i

√
and we have 2𝐼 ' 𝐴 e5 ' SL2 (F5 ). Letting now 𝐹 = Q( 5) and 𝑅 = Z[𝜏], the
𝑅-algebra generated by 2𝐼 is the maximal order
O2𝐼 = 𝑅 + 𝑅𝑖 + 𝑅𝜁 + 𝑅𝑖𝜁 . (11.5.8)
For further references, see Conway–Sloane [CSl88, §8.2], who describe the binary
icosahedral group in detail, calling it the icosian group.
We now consider the related possibilities over Q. (We will return to a general
classification
in section 32.4.) To put ourselves in a situation
like (11.2.6), let 𝐵 =
𝑎, 𝑏 𝑎, 𝑏
be a quaternion algebra over Q such that 𝐵 ⊗Q R = ' H: in this case,
Q R
we say that 𝐵 is definite. By Exercise 2.4, 𝐵 is definite if and only if 𝑎, 𝑏 < 0. Let
O ⊆ 𝐵 be an order in 𝐵; we would like to understand its unit group.
Lemma 11.5.9. The group O× = O1 is finite.
Proof. We may take 𝐵 = (𝑎, 𝑏 | Q) with 𝑎, 𝑏 < 0. Consider the reduced norm
nrd : 𝐵 → Q, given by nrd(𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗) = 𝑡 2 + |𝑎|𝑥 2 + | 𝑏|𝑦 2 + |𝑎𝑏|𝑧2 , so
nrd(𝐵× ) ⊆ Q×>0 . At the same time, nrd(O× ) ⊆ Z× = {±1}, so we conclude O× = O1 .
This group is finite because the restriction nrd | O of the reduced norm to O ' Z4 defines
a (still) positive definite quadratic form, so there are only finitely many elements of O
of any fixed reduced norm. (For a geometric perspective, viewing the elements of O1
as lattice points on an ellipsoid in R4 ), see Proposition 17.5.6.)
In view of Lemma 11.5.9, the classification of finite rotation groups (Proposition

11.5.2) applies. We consider each case in turn.
11.5.10. Among the (nontrivial) cyclic groups, only subgroups of order 2, 4, 6 are
possible over Q. Indeed, a generator satisfies a quadratic equation with integer coeffi-
cients and so belongs to the ring of integers of an imaginary quadratic field; and only
two imaginary
√ quadratic fields have units other than ±1, namely,√the Eisenstein order
Z[(−1 + −3)/2] of discriminant −3 and the Gaussian order Z[ −1] of discriminant
−4 with groups of size 4, 6, respectively. (The more precise question of whether or
not there is a unit of specified order is a question of embedding numbers, the subject
of Chapter 30.)
11.5.11. Next, suppose that O× /{±1} is dihedral, and let 𝑗 ∈ O× r {±1} act by
inversion (equivalently, conjugation) on a cyclic group (of order 2, 3, by 11.5.10),
generated by an element 𝑖. Let 𝐾 = Q(𝑖). Since 𝑗 acts by inversion, we have 𝑗 2 ∈ Q,
and since
𝑗 ∈ O× we have 𝑗 2 = −1. It follows that 𝑗 𝛼 = 𝛼 𝑗 for all 𝛼 ∈ 𝐾. Thus
𝐾, −1
𝐵' , and we have two possibilities:
Q
(i) If 𝑖 has order 4, then 𝐵 ' (−1, −1 | Q) and O contains the order generated
by 𝑖, 𝑗. This is the case treated in section 11.2: O is the Lipschitz order, and
O× ' 𝑄 8 is the quaternion group of order 8.
(ii) Otherwise, 𝑖 = 𝜔 has order 6, and 𝐵 ' (−3, −1 | Q). By Exercise 11.12(a),
we have (−3, −1 | Q) ; (−1, −1 | Q). By an argument similar to Lemma
11.1.2—and boy, there is more of this to come in Chapter 32—we see that
O = Z + Z𝜔 + Z 𝑗 + Z𝜔 𝑗 (11.5.12)
is maximal. The group O× /{±1} ' 𝐷 6 is a dihedral group of order 6, and the
group O× is generated by 𝜔, 𝑗 with relations 𝜔3 = 𝑗 2 = −1 and 𝑗𝜔 = 𝜔−1 𝑗; in
other words, O× ' 𝐶3 o 𝐶4 is the semidirect product of the cyclic group 𝐶3 of
order 3 by the action of the cyclic group 𝐶4 with a generator acting by inversion
on 𝐶3 . Because 𝑖 2 = −1 is central, we also have an exact sequence
1 → 𝐶6 → O× → 𝐶2 → 1
where 𝐶6 ' h𝜔i and 𝐶2 ' h 𝑗i/{±1}. This group is also called the binary
dihedral or dicyclic group of order 12, denoted 2𝐷 6 .
11.5.13. To conclude, suppose that O× /{±1} is exceptional. Each of these groups
contain a dihedral group, so the argument from 11.5.11 applies: the only new group
we see is the (binary) tetrahedral group obtained from the Hurwitz units (section 11.2).
Here is another proof: the group 𝑆4 contains an element of order 4 and 𝐴5 an element
of order 5, and these lift to elements of order 8, 10 in O× , impossible.
We have proven the following theorem.
Theorem 11.5.14. Let 𝐵 = (𝑎, 𝑏 | Q) be a quaternion algebra over Q with 𝑎, 𝑏 < 0,
and let O ⊆ 𝐵 be an order. Then O× is either cyclic of order 2, 4, 6, quaternion of
order 8, binary dihedral of order 12, or binary tetrahedral of order 24.
Moreover, O× is quaternion, binary dihedral, or binary tetrahedral if and only
if O is isomorphic to the Lipschitz order, the order (11.5.12), or the Hurwitz order,
respectively.
Proof. Combine 11.5.10, 11.5.11, and 11.5.13.
Exercises
I 1. Check directly that the Hurwitz order

1+𝑖+ 𝑗 + 𝑘
O = Z + Z𝑖 + Z 𝑗 + Z
2

−1, −1
is indeed an order in 𝐵 = .
Q

−1, −1
2. Let 𝐵 = , and let O ⊆ 𝐵 be the Hurwitz order. For the normalizer
Q
𝑁 𝐵× (O) := {𝛼 ∈ 𝐵× : 𝛼−1 O𝛼 = O}
show the equality 𝑁 𝐵× (Zh𝑖, 𝑗i) = 𝑁 𝐵× (O). [Hint: consider units and their
traces.]
3. (a) Show that the Lipschitz order Zh𝑖, 𝑗i is the unique suborder of the Hurwitz
order O with index 2 (as abelian groups).
(b) Show that
Zh𝑖, 𝑗i = {𝛼 ∈ O : trd(𝛼) is even}.
4. Check that the map
O/3O → M2 (F3 )

0 −1 1 1
𝑖, 𝑗 ↦→ ,
1 0 1 −1
from Lemma 11.2.1 is an F3 -algebra isomorphism.

5. Generalizing the previous exercise, show that for an odd prime 𝑝 that O/𝑝O '
M2 (F 𝑝 ).
6. Draw the subgroup lattice for SL2 (F3 ), indicating normal subgroups (and their
quotients).
I 7. Show explicitly that
2𝑇 ' h𝑟, 𝑠, 𝑡 | 𝑟 2 = 𝑠3 = 𝑡 3 = 𝑟 𝑠𝑡 = −1i
(cf. (11.2.8)).
I 8. Let
Λ = Z4 + Z( 12 , 12 , 21 , 12 ) ⊂ R4
be the image of the Hurwitz order O under the natural embedding O ↩→ H ' R4 .
Show that for every 𝑥 ∈ R4 , there exists 𝜆 ∈ Λ such that k𝜆k 2 ≤ 1/2. [Hint:
without loss of generality we may take 0 ≤ 𝑥𝑖 ≤ 1/2 for all 𝑖; then show we may
take 𝑥 1 + 𝑥2 + 𝑥3 + 𝑥4 ≤ 1; conclude that the maximum value of k𝑥k 2 with these
conditions occurs at the point ( 12 , 12 , 0, 0).]
9. Let 𝐵 be a definite quaternion algebra over Q and let O ⊆ 𝐵 be an order. Show
that O is left Euclidean if and only if O is right Euclidean (with respect to a
norm 𝑁).
10. Let O ⊂ 𝐵 := (−1, −1 | Q) be the Hurwitz order.
(a) Consider the natural ring homomorphism O → O/2O = O ⊗Z F2 giving
the reduction of the algebra O modulo 2. Show that O/2O is an F2 -
algebra, that #(O/2O) = 16, and that (O/2O) × ' 𝐴4 is isomorphic to
the alternating group on 4 elements. Conclude that O/2O ; M2 (F2 ) and

hence that O/2O is not a quaternion algebra over F2 .
(b) Show that the group of ring automorphisms of O/2O is
Aut(O/2O) ' 𝑆4 .
(c) More generally, if 𝐹 is a field of characteristic 2 show that there is an exact
sequence
1 → 𝐹 2 → Aut(O ⊗Z 𝐹) → SL2 (𝐹) → 1
where 𝐹 2 is an additive group. [Hint: let 𝐽 = rad(O ⊗Z 𝐹) be the Jacobson
radical of the algebra, and show that the sequence is induced by 𝐹-linear
automorphisms of 𝐽 and the automorphisms 𝜔 ↦→ 𝜔 + 𝜖 with 𝜖 ∈ 𝐽.]
[This kind of construction, considered instead over the octonions, arises
when constructing the exceptional group 𝐺 2 in characteristic 2 [Wils2009,
§4.4.1].]
11. Although the Lipschitz order just misses being Euclidean with respect to the
norm (see Example 11.3.8), bootstrapping from the Hurwitz order we still obtain
a result on principality by restricting the set of ideals, as follows.
Let 𝐼 ⊆ Zh𝑖, 𝑗i be a right ideal.
(a) Show that 𝐼O = 𝛽O for some 𝛽 ∈ 𝐼O ∩ Zh𝑖, 𝑗i.
(b) Prove that 𝐼 (2) = Z (2) h𝑖, 𝑗i if and only if 𝐼 is generated by elements of odd
reduced norm.
(c) If 𝐼 (2) = Z (2) h𝑖, 𝑗i, show that 𝐼O ∩ Zh𝑖, 𝑗i = 𝐼 and conclude that 𝐼 is right
principal. [Hint: Argue locally.]
12. Let 𝐵 := (−1, −3 | Q), and let
1+ 𝑗 1+ 𝑗
O := Zh𝑖, (1 + 𝑗)/2i = Z + Z𝑖 + Z + Z𝑖 .
2 2
(a) Show that 𝐵 ; (−1, −1 | Q).
(b) Show that O is a maximal order in 𝐵.
(c) Show that O is Euclidean with respect to the reduced norm
nrd(𝑡 + 𝑥𝑖 + 𝑦(1 + 𝑗)/2 + 𝑧𝑖(1 + 𝑗)/2) = 𝑡 2 + 𝑡𝑦 + 𝑥 2 + 𝑥𝑧 + 𝑦 2 + 𝑧2 .
(d) Show that every maximal order in 𝐵 is conjugate to O.
13. Let 𝐺 ≤ O(2) be a finite subgroup such that tr(𝑔), det(𝑔) ∈ Q for all 𝑔 ∈ 𝐺.
Show that 𝐺 is conjugate in O(2) to one of the following: (i) a cyclic group of
order 1, 2, 3, 4, 6 that is a subgroup of SO(2), or (ii) a dihedral group of order
2, 4, 6, 8, 12, not contained in SO(2).
14. Let 𝑝 be an odd prime.
(a) The group GL2 (Z ( 𝑝) ) acts by right multiplication on the set of matrices
𝜋 ∈ M2 (Z ( 𝑝) ) with 𝑝 k det(𝜋) (i.e., 𝑝 exactly divides the numerator of
det(𝜋), written in lowest terms). Show that there are precisely 𝑝 + 1 orbits,
represented by
𝑝 0
𝜋=
0 1
and
1 0
𝜋= , 𝑎 = 0, 1, . . . , 𝑝 − 1.
𝑎 𝑝
[Hint: use column operations.]
(b) Repeat (a) but with SL2 (Z ( 𝑝) ) acting on the set of matrices 𝜋 ∈ M2 (Z ( 𝑝) )
with det(𝜋) = 𝑝, with the same conclusion.
(c) Show that the number of (left or) right ideals of O of reduced norm 𝑝 is
equal to 𝑝 + 1.
(d) Accounting for units, conclude that the number of ways of writing an odd
prime 𝑝 as the sum of four squares is equal to 8( 𝑝 + 1).
15. In the following exercise, we consider a computational problem, suitable for
those with some background in number theory algorithms (see e.g. Cohen
[Coh93]).
(a) Show that one can find 𝑥, 𝑦, 𝑧 ∈ Z such that 𝑥 2 + 𝑦 2 + 𝑧2 = 𝑝𝑚 with 𝑝 - 𝑚
in probabilistic polynomial time in log 𝑝.
(b) Describe the right Euclidean algorithm as applied to 𝛼 = 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 and
𝑝 to obtain 𝜋 ∈ O with nrd(𝜋) = 𝑝. Adjust as in Lemma 11.2.9 to find a
solution to 𝑥 2 + 𝑦 2 + 𝑧2 = 𝑝 with 𝑥, 𝑦, 𝑧 ∈ Z. Estimate the running time of
this algorithm.
Chapter 12
Ternary quadratic forms over local

fields
In this chapter, we classify quaternion algebras over local fields using quadratic forms;
this generalizes the classification of quaternion algebras over R.
12.1 ⊲ The p-adic numbers and local quaternion algebras

Before beginning, we briefly remind the reader about the structure of the 𝑝-adic
numbers. The 𝑝-adics were developed by Hensel, who wanted a uniform way to say
that a Diophantine equation has a (consistent) solution modulo 𝑝 𝑛 for all 𝑛. In the
early 1920s, Hasse used them in the study of quadratic forms and algebras over number
fields. At the time, what is now called the local-global principle then was called the
𝑝-adic transfer from the “small” to the “large”. As references on 𝑝-adic numbers, see
for example Gouvêa [Gou97], Katok [Kat2007], or Koblitz [Kob84].
Just as elements of R can be thought of infinite decimals, an element of Q 𝑝 can be
thought of in its 𝑝-adic expansion
∞
∑︁
𝑎 = (. . . 𝑎 3 𝑎 2 𝑎 1 𝑎 0 .𝑎 −1 𝑎 −2 · · · 𝑎 −𝑘 ) 𝑝 = 𝑎𝑛 𝑝𝑛 (12.1.1)
𝑛=−𝑘
where each 𝑎 𝑖 ∈ {0, . . . , 𝑝 − 1} are the digits of 𝑎. We continue “to the left” because
a decimal expansion is a series in the base 1/10 < 1 and instead we have a base 𝑝 > 1.
Put a bit more precisely, we define the 𝑝-adic absolute value on Q by defined by
|0| 𝑝 := 0 and
|𝑐| 𝑝 := 𝑝 −𝑣𝑝 (𝑐) for 𝑐 ∈ Q× , (12.1.2)
where 𝑣 𝑝 (𝑐) is the power of 𝑝 occurring in 𝑐 in its unique factorization (taken to be
negative if 𝑝 divides the denominator of 𝑐 written in lowest terms). Then the field Q 𝑝
is the completion of Q with respect to | | 𝑝 , that is to say, Q 𝑝 is the set of equivalence
classes of Cauchy sequences of rational numbers, and it obtains a topology induced by
the metric 𝑑 𝑝 (𝑥, 𝑦) = |𝑥 − 𝑦| 𝑝 . We have |𝑎| 𝑝 = 𝑝 𝑘 for 𝑎 as in (12.1.1) with 𝑎 −𝑘 ≠ 0.
175
176 CHAPTER 12. TERNARY QUADRATIC FORMS OVER LOCAL FIELDS
Of course, all of the information in the 𝑝-adic absolute value is encoded in the 𝑝-adic
valuation 𝑣 𝑝 : Q → R ∪ {∞}.
Inside Q 𝑝 is the ring Z 𝑝 of 𝑝-adic integers, the completion of Z with respect to
| | 𝑝 : the ring Z 𝑝 consists of those elements of Q 𝑝 with 𝑎 𝑛 = 0 for 𝑛 < 0. (The ring Z 𝑝
might be thought of intuitively as Z/𝑝 ∞ Z, if this made sense.)
Equipped with their topologies, the ring Z 𝑝 is compact and the field Q 𝑝 is lo-
cally compact. These statements can be understood quite easily by viewing Z 𝑝 in a
slightly different way, as a projective limit with respect to the natural projection maps
Z/𝑝 𝑛+1 Z → Z/𝑝 𝑛 Z:
Z 𝑝 = lim Z/𝑝 𝑛 Z
←−−
𝑛
( ∞
Ö
) (12.1.3)
𝑛 𝑛
= 𝑥 = (𝑥 𝑛 ) 𝑛 ∈ Z/𝑝 Z : 𝑥 𝑛+1 ≡ 𝑥 𝑛 (mod 𝑝 ) for all 𝑛 ≥ 1 .
𝑛=1
In other words, each element of Z 𝑝 is a compatible sequence of elements in Z/𝑝 𝑛 Z for

each 𝑛. The equality (12.1.3) is just a reformulation of the notion of Cauchy sequence
for Z, and so for the purposes of this introduction it can equally well be taken as a
definition.
As for the
Î∞ topology in (12.1.3), each factor Z/𝑝 𝑛 Z is given the discrete topology,
the product 𝑛=0 Z/𝑝 𝑛 Z is given the product topology, and Z 𝑝 is given the subspace
topology. Since each Z/𝑝 𝑛 Z is compact (it is a finite set!), by Tychonoff’s theorem the
product ∞
Î
𝑛=0 Z/𝑝 Z is compact; and Z 𝑝 is closed inside this product (a convergent
𝑛
limit of Cauchy sequences is a Cauchy sequence), so Z 𝑝 is compact and still Hausdorff.

The topology on Z 𝑝 is a bit strange though, as Z 𝑝 is totally disconnected: every
nonempty connected subset is a single point. In fact, Z 𝑝 is homeomorphic to the
Cantor set, which is itself homeomorphic to the product of countably many copies of
{0, 1}. (More generally, every nonempty totally disconnected compact metric space
with no isolated points is homeomorphic to the Cantor set.)
The set Z 𝑝 is a compact neighborhood of 0, as it is the closed ball of radius 1
around 0:
Z 𝑝 = {𝑥 ∈ Q 𝑝 : |𝑥| 𝑝 ≤ 1} = {𝑥 ∈ Q 𝑝 : 𝑣 𝑝 (𝑥) ≥ 0}. (12.1.4)
In a similar way, the disc of radius 1 around 𝑎 ∈ Q 𝑝 is a compact neighborhood of
𝑎 homeomorphic to Z 𝑝 , so Q 𝑝 is locally compact. Being able to make topological
arguments like the one above is the whole point of looking at fields like Q 𝑝 : our
understanding of infinite algebraic objects is informed by topology.
With this review, and topological arguments now at our disposal, we consider
quaternion algebras over Q 𝑝 . The ‘original’ quaternion algebra, of course, was the
division ring H of Hamiltonians over the real numbers (the ‘original’ field with a
topology), and indeed H is the unique division quaternion algebra over R (Corollary
3.5.8). We find a similar result over Q 𝑝 (a special case of Theorem 12.3.2), as follows.
Theorem 12.1.5. There is a unique division quaternion algebra 𝐵 over Q 𝑝 , up to
isomorphism; if 𝑝 ≠ 2, then
𝑒, 𝑝
𝐵'
Q𝑝
12.1. ⊲ THE P-ADIC NUMBERS AND LOCAL QUATERNION ALGEBRAS 177
where 𝑒 ∈ Z is a quadratic nonresidue modulo 𝑝.
For example, if 𝑝 ≡ 3 (mod 4) then −1 is quadratic nonresidue and (−1, 𝑝 | Q 𝑝 )

is the unique division quaternion algebra over Q 𝑝 .
Because we have exactly two such possibilities, we define the Hilbert symbol:
for 𝑎, 𝑏 ∈ Q×𝑝 , we have (𝑎, 𝑏)Q 𝑝 = 1, −1 according as the quaternion algebra (𝑎, 𝑏 |
Q 𝑝 ) ' M2 (Q 𝑝 ) is split or not. According to Theorem 12.1.5, the Hilbert symbol over
Q 𝑝 uniquely identifies the two possible isomorphism classes of quaternion algebras
over Q 𝑝 —just like it does over R.
Our approach to Theorem 12.1.5 uses quadratic forms: we use the classification
of isomorphism classes of quaternion algebras given in terms of similarity classes
of ternary quadratic forms (Theorem 5.1.1). The following proposition then implies
Theorem 12.1.5.
Proposition 12.1.6. There is a unique ternary anisotropic quadratic form 𝑄 over Q 𝑝 ,

up to similarity; if 𝑝 ≠ 2, then 𝑄 ∼ h1, −𝑒, −𝑝i where 𝑒 is a quadratic nonresidue
modulo 𝑝.
Happily, this proposition can be proved using some rather direct manipulations
with quadratic forms and gives a very “hands on” feel; it is also suggests the arguments
we use for a more general result. The main input we need is a quadratic Hensel’s
lemma, or more precisely, the following consequence.
Lemma 12.1.7. For 𝑝 ≠ 2, the classes in Q×𝑝 /Q×2

𝑝 are represented by 1, 𝑒, 𝑝, 𝑒 𝑝 where
𝑒 is a quadratic nonresidue modulo 𝑝.
Proof. Let 𝑎 ∈ Q×𝑝 and let 𝑚 := 𝑣 𝑝 (𝑎). Then 𝑎 = 𝑏 𝑝 𝑚 with 𝑏 := 𝑎/𝑝 𝑚 ∈ Z×𝑝 , and by
squaring 𝑎 ∈ Q×2 ×2 ×2
𝑝 if and only if 𝑏 ∈ Z 𝑝 and 𝑚 is even. We claim that 𝑏 ∈ Z 𝑝 if and
×
only if its reduction 𝑏 modulo 𝑝 is a square in (Z/𝑝Z) . With the forward implication
immediate, suppose 𝑏 ≡ 𝑐2 (mod 𝑝) with 𝑐 ∈ Z×𝑝 , then 𝑏/𝑐2 ∈ 1 + 𝑝Z 𝑝 . But squaring
is a bĳection on 1 + 𝑝Z 𝑝 , by expanding the square root as a convergent series (see
Exercise 12.1) and using that 𝑝 ≠ 2. Thus 𝑏/𝑐2 ∈ Z×2 𝑝 , and the result follows.
We now proceed with the proof when 𝑝 ≠ 2.
Proof of Proposition 12.1.6, 𝑝 ≠ 2. We start by showing that 𝑄(𝑥, 𝑦, 𝑧) = 𝑥 2 − 𝑒𝑦 2 −

𝑝𝑧2 is anisotropic. Suppose 𝑄(𝑥, 𝑦, 𝑧) = 0 with not all 𝑥, 𝑦, 𝑧 ∈ Q 𝑝 zero. Rescaling
by 𝑝, we may assume that 𝑥, 𝑦, 𝑧 ∈ Z 𝑝 and not all 𝑥, 𝑦, 𝑧 ∈ 𝑝Z 𝑝 . We then reduce
modulo 𝑝 to find that 𝑥 2 ≡ 𝑒𝑦 2 (mod 𝑝). If 𝑝 - 𝑦, then (𝑥/𝑦) 2 ≡ 𝑒 (mod 𝑝); but 𝑒 is
a quadratic nonresidue modulo 𝑝, a contradiction. So 𝑝 | 𝑦; thus 𝑝 | (𝑒𝑦 2 + 𝑝𝑧2 ) = 𝑥 2 ,
so 𝑝 | 𝑥; thus 𝑝 2 | (𝑥 2 − 𝑒𝑦 2 ) = 𝑝𝑧2 , so 𝑝 | 𝑧, a contradiction.
To show uniqueness, let 𝑄 be a ternary anisotropic form over Q 𝑝 . Since 𝑝 ≠ 2,
we may diagonalize. In such a diagonal form taken up to similarity, we may also
rescale each coordinate up to squares as well as rescale the entire quadratic form.
Putting this together with Lemma 12.1.7, without loss of generality we may suppose
𝑄(𝑥, 𝑦, 𝑧) = h1, −𝑏, −𝑐i = 𝑥 2 − 𝑏𝑦 2 − 𝑐𝑧2 with 𝑏, 𝑐 ∈ {1, 𝑒, 𝑝, 𝑒 𝑝}, the signs chosen
for convenience. If 𝑏 = 1 or 𝑐 = 1, the form is isotropic by inspection. So we are left
to consider the cases (𝑏, 𝑐) = (𝑒, 𝑒), (𝑒, 𝑝), (𝑒, 𝑒 𝑝), ( 𝑝, 𝑝), ( 𝑝, 𝑒 𝑝), (𝑒 𝑝, 𝑒 𝑝).
• When (𝑏, 𝑐) = (𝑒, 𝑒), we have after rescaling 𝑥 2 + 𝑦 2 − 𝑒𝑧2 . We claim this form is
always isotropic. Indeed, the form reduces to a nondegenerate ternary quadratic
form over F 𝑝 . Such a form is always isotropic by a delightful counting argument
(Exercise 5.5(b), or a second chance in Exercise 12.6!). Lifting, there exist
𝑥, 𝑦, 𝑧 ∈ Z 𝑝 , not all zero modulo 𝑝, such that 𝑥 2 ≡ −𝑦 2 + 𝑒𝑧 2 (mod 𝑝). Since
𝑒 is a nonsquare, we have 𝑝 - 𝑥 (arguing similarly as in the first paragraph).
Let 𝑑 := −𝑦 2 + 𝑒𝑧2 ∈ Q×𝑝 . By the possibilities in Lemma 12.1.7, we must have
𝑑 ∈ Q×2 2
𝑝 ; solving 𝑥 = 𝑑 for 𝑥 ∈ Q 𝑝 then shows that 𝑄 is isotropic.
• The case (𝑒, 𝑝) is our desired form.
• In the third case (eep!), we substitute 𝑥 ← 𝑒𝑥 and divide by 𝑒 to obtain the form
−𝑦 2 + 𝑒𝑥 2 − 𝑝𝑧 2 . We claim that there is an isometry h−1, 𝑒i ' h1, −𝑒i: indeed,
in the first bullet we showed that the quadratic form h−1, 𝑒, −1i is isotropic,
so −𝑥 2 + 𝑒𝑦 2 represents 1; using this representation as the first basis vector,
extending to a basis, and diagonalizing, we conclude that h−1, 𝑒i ' h1, 𝑏i. By
discriminants, we have −𝑒 = 𝑏 up to squares. This brings us back to the first
case.
• In cases ( 𝑝, 𝑝) or (𝑒 𝑝, 𝑒 𝑝), replacing 𝑥 ← 𝑝𝑥 and dividing gives the quadratic
forms 𝑥 2 + 𝑦 2 − 𝑝𝑧2 and 𝑥 2 + 𝑦 2 − 𝑒 𝑝𝑧 2 . If −1 ∈ Z×2
𝑝 , then the form is isotropic;
otherwise, we may take 𝑒 = −1 and we are back to cases (𝑒, 𝑝), (𝑒, 𝑒 𝑝).
• In the final case (keeping pep!), we substitute 𝑥 ← 𝑝𝑥 and divide by −𝑝 to get
𝑦 2 + 𝑒𝑧2 − 𝑝𝑥 2 . If −1 ∈ Z×2 𝑝 , then by substitution we change the middle sign
to return to the first case. Otherwise, we may take 𝑒 = −1, and the form is
isotropic, a contradiction.
This consideration of cases completes the proof.
Although direct, the proof we just gave has the defect that quadratic forms behave
differently in characteristic 2, and so one may ask for a proof that works uniformly in
all characteristics: we give such a proof in the next chapter by extending valuations.
One of the nice applications of this classification is that it gives a necessary
condition for two quaternion algebras to be isomorphic. Let 𝐵 = (𝑎, 𝑏 | Q) be a
quaternion algebra over Q and consider its scalar extension 𝐵 𝑝 = 𝐵 ⊗Q Q 𝑝 ' (𝑎, 𝑏 |
Q 𝑝 ). If 𝐵 0 is another quaternion algebra over Q and 𝐵 ' 𝐵 0, then this implies
𝐵 𝑝 ' 𝐵 0𝑝 for all primes 𝑝, and of course the same is true over R. Perhaps surprisingly,
it turns out that the collection of all of these tests is also sufficient: if 𝐵, 𝐵 0 become
isomorphic over R and over Q 𝑝 for all primes 𝑝, then in fact 𝐵 ' 𝐵 0 are isomorphic
over Q! This profound and powerful principle—detecting global isomorphism from
local isomorphisms, a local-global principle—will be examined in chapter 14.
12.2 Local fields

In this section, we set up notation and basic results from the theory of local fields.
The theory of local fields is described in many places, including Neukirch [Neu99,
Chapters II, V], the classic texts by Cassels [Cas86] and Serre [Ser79]. Weil [Weil74]
approaches number theory from the ground up in the language of local fields, building
up the theory of local division rings.
12.2. LOCAL FIELDS 179
Our motivation for local fields is as follows: we want a topology compatible with
the field operations in which the field is Hausdorff and locally compact (every element
has a compact neighborhood), analogous to what holds over the real and complex
numbers. And to avoid trivialities, we will insist that this topology is not the discrete
topology (where every subset of 𝐹 is open). To carry this out, we begin with some
basic definitions.
Definition 12.2.1. A topological group is a group equipped with a topology such that
the group operation and inversion are continuous. A homomorphism of topological
groups is a group homomorphism that is continuous.
A topological ring is a ring 𝐴 equipped with a topology such that the ring oper-
ations (addition, negation, and multiplication) are continuous. A homomorphism of
topological rings is a ring homomorphism that is continuous. A topological field is a
field that is also a topological ring in such a way that division by a nonzero element is
continuous.
One natural way to equip a ring with a topology is by way of an absolute value. To
get started, we consider such notions first for fields. Throughout this section, let 𝐹 be
a field.
Definition 12.2.2. An absolute value on 𝐹 is a map
| | : 𝐹 → R ≥0
such that:
(i) |𝑥| = 0 if and only if 𝑥 = 0;

(ii) |𝑥𝑦| = |𝑥|| 𝑦| for all 𝑥, 𝑦 ∈ 𝐹; and
(iii) |𝑥 + 𝑦| ≤ |𝑥| + | 𝑦| for all 𝑥, 𝑦 ∈ 𝐹 (triangle inequality).
An absolute value | | on 𝐹 gives 𝐹 the structure of a topological field by the metric

𝑑 (𝑥, 𝑦) = |𝑥 − 𝑦|. Two absolute values | | 1 , | | 2 on 𝐹 are (strictly) equivalent if there
exists 𝑐 > 0 such that |𝑥| 1 = |𝑥| 2𝑐 for all 𝑥 ∈ 𝐹; equivalent absolute values induce the
same topology on 𝐹.
Remark 12.2.3. If | | is an absolute value on 𝐹, then it need not be the case that
𝑥 ↦→ |𝑥| 𝑐 for 𝑐 > 0 is again absolute value, because it need not satisfy the triangle
inequality. In particular, we will find it convenient to consider the square of the usual
absolute value on 𝐹 = C, which suffers from this deficiency. There are various ways
around this problem; perhaps the simplest is just to ignore it.
Definition 12.2.4. An absolute value is nonarchimedean if the ultrametric inequal-

ity
|𝑥 + 𝑦| ≤ sup{|𝑥|, | 𝑦|}
is satisfied for all 𝑥, 𝑦 ∈ 𝐹, and archimedean otherwise.
Example 12.2.5. The fields R and C are topological fields with respect to the usual
archimedean absolute value.
Remark 12.2.6. A field with absolute value is archimedean if and only if it satisfies
the archimedean property: for all 𝑥 ∈ 𝐹 × , there exists 𝑛 ∈ Z such that |𝑛𝑥| > 1. In
particular, a field 𝐹 equipped with an archimedean absolute value has char 𝐹 = 0.
Example 12.2.7. Every field has the trivial (nonarchimedean) absolute value, defined
by |0| = 0 and |𝑥| = 1 for all 𝑥 ∈ 𝐹 × ; the trivial absolute value induces the discrete
topology on 𝐹.
A nonarchimedean absolute value on a field 𝐹 arises naturally by way of a valuation,

as follows.
Definition 12.2.8. A valuation of a field 𝐹 is a map 𝑣 : 𝐹 → R ∪ {∞} such that:
(i) 𝑣(𝑥) = ∞ if and only if 𝑥 = 0;

(ii) 𝑣(𝑥𝑦) = 𝑣(𝑥) + 𝑣(𝑦) for all 𝑥, 𝑦 ∈ 𝐹; and
(iii) 𝑣(𝑥 + 𝑦) ≥ min(𝑣(𝑥), 𝑣(𝑦)) for all 𝑥, 𝑦 ∈ 𝐹.
A valuation is discrete if the value group 𝑣(𝐹 × ) is discrete in R (has no accumulation

points).
Here, we set the convention that 𝑥 + ∞ = ∞ + 𝑥 = ∞ for all 𝑥 ∈ R ∪ {∞}. By

(ii), the value group 𝑣(𝐹 × ) is a subgroup of the additive group R, and so whereas an
absolute value is multiplicative, a valuation is additive.
Example 12.2.9. Each 𝑥 ∈ Q× can be written 𝑥 = 𝑝 𝑟 𝑎/𝑏 with 𝑎, 𝑏 ∈ Z relatively

prime and 𝑝 - 𝑎𝑏; the map 𝑣 𝑝 (𝑥) = 𝑟 defines the 𝑝-adic valuation on Q.
Example 12.2.10. Let 𝑘 be a field and 𝐹 = 𝑘 (𝑡) the field of rational functions over
𝑘. For 𝑓 (𝑡) = 𝑔(𝑡)/ℎ(𝑡) ∈ 𝑘 (𝑡) r {0} with 𝑔(𝑡), ℎ(𝑡) ∈ 𝑘 [𝑡], define 𝑣( 𝑓 (𝑡)) :=
deg ℎ(𝑡) − deg 𝑔(𝑡). Then 𝑣 is a discrete valuation on 𝐹.
Given the parallels between them, it should come as no surprise that a valuation
gives rise to an absolute value on 𝐹 by defining
|𝑥| = 𝑐−𝑣 ( 𝑥) (12.2.11)
for a fixed 𝑐 > 1; the induced topology on 𝐹 is independent of the choice of 𝑐. By

condition (iii), the absolute value associated to a valuation is nonarchimedean.
Example 12.2.12. The trivial valuation is the valuation 𝑣 satisfying 𝑣(0) = ∞ and
𝑣(𝑥) = 0 for all 𝑥 ∈ 𝐹 × . The trivial valuation gives the trivial absolute value on 𝐹.
Two valuations 𝑣, 𝑤 are equivalent if there exists 𝑎 ∈ R>0 such that 𝑣(𝑥) = 𝑎𝑤(𝑥)
for all 𝑥 ∈ 𝐹; equivalent valuations give the same topology on a field. A nontrivial
discrete valuation is equivalent after rescaling (by the minimal positive element in the
value group) to one with value group Z, since a nontrivial discrete subgroup of R is
cyclic; we call such a discrete valuation normalized.
12.2. LOCAL FIELDS 181
12.2.13. Given a field 𝐹 with a nontrivial discrete valuation 𝑣, the valuation ring is
𝑅 := {𝑥 ∈ 𝐹 : 𝑣(𝑥) ≥ 0}. We have 𝑅 × = {𝑥 ∈ 𝐹 : 𝑣(𝑥) = 0} since
𝑣(𝑥) + 𝑣(𝑥 −1 ) = 𝑣(𝑥𝑥 −1 ) = 𝑣(1) = 0
for all 𝑥 ∈ 𝐹 × . The valuation ring is a local domain with unique maximal ideal
𝔭 := {𝑥 ∈ 𝐹 : 𝑣(𝑥) > 0} = 𝑅 r 𝑅 × .
An element 𝜋 ∈ 𝔭 with smallest valuation is called a uniformizer, and comparing

valuations we see that 𝜋𝑅 = (𝜋) = 𝔭. Since 𝔭 ( 𝑅 is maximal, the quotient 𝑘 := 𝑅/𝔭
is a field, called the residue field of 𝑅 (or of 𝐹).
Recall that a topological space is locally compact if each point has a compact
neighborhood (every point is contained in a compact set containing an open set).
Definition 12.2.14. A local field is a Hausdorff, locally compact topological field with
a nondiscrete topology.
In a local field, we can hope to understand its structure by local considerations in a

compact neighborhood, hence the name. Local fields have a very simple classification
as follows.
Theorem 12.2.15. A field 𝐹 with absolute value is a local field if and only if 𝐹 is one
of the following:
(i) 𝐹 is archimedean, and 𝐹 ' R or 𝐹 ' C;

(ii) 𝐹 is nonarchimedean with char 𝐹 = 0, and 𝐹 is a finite extension of Q 𝑝 for some
prime 𝑝; or
(iii) 𝐹 is nonarchimedean with char 𝐹 = 𝑝, and 𝐹 is a finite extension of the Laurent
series field F 𝑝 ((𝑡)) for some prime 𝑝; in this case, there is a (non-canonical)
isomorphism 𝐹 ' F𝑞 ((𝑡)) where 𝑞 is a power of 𝑝.
A field 𝐹 with absolute value | | is a nonarchimedean local field if and only if 𝐹 is

complete with respect to | |, and | | is equivalent to the absolute value associated to a
nontrivial discrete valuation 𝑣 : 𝐹 → R ∪ {∞} with finite residue field.
Proof. See Neukirch [Neu99, Chapter II, §5], Cassels [Cas86, Chapter 4, §1], or Serre
[Ser79, Chapter II, §1].
Although a local field is only locally compact, the valuation ring is itself compact,
as follows.
Lemma 12.2.16. Suppose 𝐹 is nonarchimedean. Then 𝐹 is totally disconnected and

the valuation ring 𝑅 ⊂ 𝐹 is a compact, totally disconnected topological ring.
Proof. To see that 𝐹 is totally disconnected (whence 𝑅 too is totally disconnected),

by translation it suffices to show that the only connected set containing 0 is {0}. Let
𝑥 ∈ 𝐹 × with |𝑥| = 𝛿 > 0. The image | 𝐹 × | ⊆ R>0 is discrete, so there exists 0 < 𝜖 < 𝛿
such that | 𝑦| < 𝛿 implies | 𝑦| ≤ 𝛿 − 𝜖 for all 𝑦 ∈ 𝐹. Thus an open ball is a closed ball
𝐷 (0, 𝛿) = {𝑦 ∈ 𝐹 : | 𝑦| < 𝛿} = {𝑦 ∈ 𝐹 : | 𝑦| ≤ 𝛿 − 𝜖 } = 𝐷 [0, 𝛿 − 𝜖];
since 𝑥 ∈ 𝐹 × and 𝛿 > 0 were arbitrary, the only connected subset containing 0 is {0}.
Next, we show 𝑅 is compact. There is a natural continuous ring homomorphism
∞
Ö
𝜙: 𝑅 → 𝑅/𝔭𝑛
𝑛=1
where each factor 𝑅/𝔭𝑛 is equipped with the discrete topology Ñ and𝑛 the product is
given the product topology. The map 𝜙 is injective, since ∞ 𝑛=1 𝔭 = {0} (every
nonzero element has finite valuation). The image of 𝜙 is obviously closed. Therefore
𝑅 is homeomorphic onto its closed image. But by Tychonoff’s theorem, the product
Î∞
𝑛=1 𝑅/𝔭 of compact sets is compact, and a closed subset of a compact set is compact,
𝑛
thus 𝑅 is compact.
One key property of local fields we will use is Hensel’s lemma.
Lemma 12.2.17 (Hensel’s lemma, univariate). Let 𝐹 be a nonarchimedean local field

with valuation 𝑣 and valuation ring 𝑅, and let 𝑓 (𝑥) ∈ 𝑅[𝑥]. Let 𝑎 ∈ 𝑅 satisfy
𝑚 := 𝑣( 𝑓 (𝑎)) > 2𝑣( 𝑓 0 (𝑎)). Then there exists e
𝑎 ∈ 𝑅 such that 𝑓 (e
𝑎 ) = 0 and e
𝑎≡𝑎
(mod 𝔭𝑚 ).
Proof. The result is straightforward to prove using Taylor expansion or the same
formulas as in Newton’s method.
Perhaps less well-known is the multivariate version.
Lemma 12.2.18 (Hensel’s lemma). Let 𝐹 be a nonarchimedean local field with valu-
ation 𝑣 and valuation ring 𝑅, and let 𝑓 (𝑥 1 , . . . , 𝑥 𝑛 ) ∈ 𝑅[𝑥 1 , . . . , 𝑥 𝑛 ] with 𝑛 ≥ 1.
Let 𝑎 ∈ 𝑅 𝑛 have 𝑚 := 𝑣( 𝑓 (𝑎)) and suppose that

𝜕𝑓
𝑚 > 2 min 𝑣 (𝑎) ≥ 0.
𝑖 𝜕𝑥𝑖
𝑎 ∈ 𝑅 𝑛 such that 𝑓 (e
Then there exists e 𝑎 ) = 0 and
𝑎≡𝑎
e (mod 𝔭𝑚 ).
Proof. One can reduce from several variables to the one variable version of Hensel’s
lemma (Lemma 12.2.17): see Exercise 12.11.
Remark 12.2.19. With essentially the same proof, Hensel’s lemma holds more gener-
ally for 𝑅 a complete DVR (without the condition on the residue field) and becomes
axiomatically the property of Henselian rings.
12.3. CLASSIFICATION VIA QUADRATIC FORMS 183
12.3 Classification via quadratic forms

We now seek to classify quaternion algebras over local fields.
12.3.1. First, suppose 𝐹 is archimedean. When 𝐹 = C, the only quaternion algebra
over C up to isomorphism is 𝐵 ' M2 (C). When 𝐹 = R, by the theorem of Frobenius
(Corollary 3.5.8), there is a unique quaternion division algebra over R.
The classification of quaternion algebras over nonarchimedean local fields is quite
analogous to the classification over R, as follows.
Main Theorem 12.3.2. Let 𝐹 ≠ C be a local field. Then there is a unique division
quaternion algebra 𝐵 over 𝐹 up to 𝐹-algebra isomorphism.
We approach the proof of Main Theorem 12.3.2 from two vantage points. In this
section, we give a proof using quadratic forms; in the next section, we give another
proof by extending the valuation (valid in all characteristics).
To prove this theorem, having dispatched the cases 𝐹 = R, C in 12.3.1 above, from
the previous section we may suppose 𝐹 is a nonarchimedean local field with discrete
valuation 𝑣, valuation ring 𝑅, maximal ideal 𝔭 = 𝜋𝑅 with uniformizer 𝜋, and residue
field 𝑅/𝔭 = 𝑘.
12.3.3. Since 𝑅 is a DVR, all 𝑅-lattices 𝑀 are free (and we only consider those of
finite rank): i.e., 𝑀 ' 𝑅 𝑛 for some 𝑛 ∈ Z ≥0 . Given such an 𝑅-lattice 𝑀, we can reduce
modulo 𝔭 to get 𝑀/𝔭𝑀 ' 𝑀 ⊗𝑅 𝑘 ' 𝑘 𝑛 ; conversely, any lift to 𝑀 of any 𝑘-basis of
𝑀/𝔭𝑀 is an 𝑅-basis for 𝑀, by Nakayama’s lemma.
We recall Main Theorem 5.2.5, Corollary 5.2.6, and Main Theorem 5.4.4: iso-
morphism classes of quaternion algebras over a field 𝐹 are in natural bĳection with
nondegenerate ternary quadratic forms up to similarity, and the matrix algebra corre-
sponds to any isotropic form. So to prove Main Theorem 12.3.2, it is equivalent to
prove the following statement.
Theorem 12.3.4. Let 𝐹 ≠ C be a local field. Then there is a unique anisotropic
ternary quadratic form over 𝐹 up to similarity.
Rescaling shows equivalently that there is a unique anisotropic ternary quadratic
form over a local field 𝐹 ≠ C of discriminant 1 up to isometry. So our task becomes
a hands-on investigation of ternary quadratic forms over 𝐹. The theory of quadratic
forms over 𝐹 is linked to that over its residue field 𝑘, so we first need to examine
isotropy of quadratic forms over a finite field.
Lemma 12.3.5. A quadratic form 𝑄 : 𝑉 → 𝑘 over a finite field 𝑘 with dim 𝑘 𝑉 ≥ 3 is
isotropic.
Proof. The proof is a delightful elementary exercise (Exercise 12.6).
Recall definitions and notation for quadratic forms over 𝑅 provided in section 9.7,
we embark on a proof in the case where char 𝑘 ≠ 2, beginning with the following
lemma.
Lemma 12.3.6. Suppose char 𝑘 ≠ 2. Let 𝑄 : 𝑀 → 𝑅 be a nonsingular quadratic

form over 𝑅. Then the reduction 𝑄 𝑘 : 𝑀 ⊗𝑅 𝑘 → 𝑘 of 𝑄 modulo 𝔭 is nonsingular
(equivalently, nondegenerate) over 𝑘; moreover, 𝑄 is isotropic over 𝑅 if and only if
𝑄 mod 𝔭 is isotropic.
Proof. For the first statement, by definition we have disc 𝑄 ∈ 𝑅 × , so disc 𝑄 𝑘 ∈ 𝑘 × by

reduction.
For the second, we first prove (⇒), let 𝑥 ∈ 𝑀 r {0} have 𝑄(𝑥) = 0. Since 𝑄 is
homogeneous, we may suppose that 𝑥 ∉ 𝔭𝑀 (divide by powers of 𝜋 as necessary),
so its image in 𝑀 ⊗𝑟 𝑘 is nonzero and thereby shows that 𝑄 𝑘 is isotropic. For (⇐),
let 𝑎 ∈ 𝑀 be such that 𝑄 𝑘 (𝑎) = 0 ∈ 𝑘 and 𝑎 has nonzero reduction. Choose a
basis 𝑀 ' 𝑅 𝑛 and write 𝑄(𝑥 1 , . . . , 𝑥 𝑛 ) = 𝑄(𝑥1 𝑒 1 + · · · + 𝑥 𝑛 𝑒 𝑛 ) ∈ 𝑅[𝑥1 , . . . , 𝑥 𝑛 ] in
the standard basis as a homogeneous polynomial of degree 2, let 𝑇 be the associated
symmetric bilinear form and [𝑇] = (𝑇 (𝑒 𝑖 , 𝑒 𝑗 ))𝑖, 𝑗 the Gram matrix. We are almost
ready to apply Hensel’s lemma (Lemma 12.2.18), but need to ensure convergence. We
observe that
𝑛
𝜕𝑄 ∑︁
(𝑥1 , . . . , 𝑥 𝑛 ) = 𝑇 (𝑥𝑖 , 𝑥 𝑗 )𝑥 𝑗 (12.3.7)
𝜕𝑥𝑖 𝑗=1
so the vector of partial derivatives ((𝜕𝑄/𝜕𝑥𝑖 ) (𝑎))𝑖 = [𝑇]𝑎 is just the matrix product
of the Gram matrix with the vector 𝑎 = (𝑎 𝑖 )𝑖 . Working modulo 𝔭, we have disc 𝑄 𝑘 =
2−𝑛 det[𝑇] ∈ 𝑘 × , using that 2 ∈ 𝑘 × , so the kernel of [𝑇] mod 𝔭 is zero. Since
𝑎 has nonzero reduction, we conclude that [𝑇]𝑎 also has nonzero reduction, which
means that min𝑖 𝑣((𝜕𝑄/𝜕𝑥𝑖 ) (𝑎)) = 0. Therefore the hypotheses of Hensel’s lemma
are satisfied with 𝑚 = 1, and we conclude there exists a nonzero e 𝑎 ∈ 𝑀 such that
𝑄(e𝑎 ) = 0 and so 𝑄 is isotropic.
From Lemma 12.3.6, we obtain the following.

Proposition 12.3.8. Suppose char 𝑘 ≠ 2. Let 𝑄 : 𝑀 → 𝑅 be a nonsingular quadratic
form over 𝑅 with 𝑀 free of rank at least 3. Then 𝑄 is isotropic.
Proof. Combine Lemmas 12.3.5 and 12.3.6.
Considering valuations, we also deduce the following from Lemma 12.3.6.

Lemma 12.3.9. Suppose char 𝑘 ≠ 2. Then 𝐹 × /𝐹 ×2 ' (Z/2Z) 2 and is represented
by the classes of 1, 𝑒, 𝜋, 𝑒𝜋 where 𝑒 ∈ 𝑅 × is an element which reduces modulo 𝔭 to a
nonsquare in 𝑘.
We now turn to the proof of our theorem in the case char 𝑘 ≠ 2.
Proof of Theorem 12.3.4 (char 𝑘 ≠ 2). Let 𝑄 ' h𝑎, −𝑏, −𝑐i be an anisotropic ternary
quadratic form over 𝐹. Then 𝑄 is nondegenerate. After rescaling and a change of basis
(Exercise 12.8), we may suppose that 𝑎 = 1 and 0 = 𝑣(𝑏) ≤ 𝑣(𝑐). If 𝑣(𝑏) = 𝑣(𝑐) = 0
then the quadratic form modulo 𝔭 is nonsingular, so by Lemma 12.3.5 it is isotropic
and by Lemma 12.3.6 we conclude 𝑄 is isotropic, a contradiction.
We are left with the case 𝑣(𝑏) = 0 and 𝑣(𝑐) = 1. By Lemma 12.3.9, we may
suppose 𝑏 = 1 or 𝑏 = 𝑒 where 𝑒 is a nonsquare in 𝑘. If 𝑏 = 1, then the form is
12.3. CLASSIFICATION VIA QUADRATIC FORMS 185
obviously isotropic, so 𝑏 = 𝑒. Similarly, 𝑐 = 𝜋 or 𝑐 = 𝑒𝜋. In fact, the latter case is

similar to the former: dividing by 𝑒, we have
h1, −𝑒, −𝑒𝜋i ∼ h𝑒 −1 , −1, −𝜋i ' h−1, 𝑒, −𝜋i
and since h−1, 𝑒i ' h1, −𝑒i (Exercise 12.7), we conclude 𝑄 ∼ h1, −𝑒, −𝜋i.
To finish, we show that the form h1, −𝑒, −𝜋i is anisotropic. Suppose that 𝑥 2 − 𝑒𝑦 2 =
𝜋𝑧 with 𝑥, 𝑦, 𝑧 ∈ 𝐹 3 not all zero. By homogeneity, rescaling by a power of 𝜋 if
2
necessary, we may suppose 𝑥, 𝑦, 𝑧 ∈ 𝑅 and at least one of 𝑥, 𝑦, 𝑧 ∈ 𝑅 × . Reducing

modulo 𝔭 we have 𝑥 2 ≡ 𝑒𝑦 2 (mod 𝔭); since 𝑒 is a nonsquare in 𝑘, we must have
𝑣(𝑥), 𝑣(𝑦) ≥ 1. But this implies that 𝑣(𝑧) = 0 and so 𝑣(𝜋𝑧 2 ) = 1 = 𝑣(𝑥 2 − 𝑒𝑦 2 ) ≥ 2, a
contradiction.
Predictably, the proof when char 𝑘 ≠ 2 involving quadratic forms does not general-
ize in a simple way. However, armed with the above outline and acknowledging these
complications, we now pursue the case char 𝑘 = 2. A key ingredient will understanding
certain binary quadratic forms, as follows.
Lemma 12.3.10. There is a unique anisotropic binary quadratic form over 𝑘, up to
isometry. Moreover, there is a unique (anisotropic) binary quadratic form over 𝑅
whose reduction modulo 𝔭 is anisotropic, up to isometry (over 𝑅).
When char 𝑘 ≠ 2, this unique class of binary forms over 𝑅 is h1, 𝑒i in the notation
above. We will want a similar bit of notation in the case char 𝑘 = 2.
12.3.11. Recall the issues (6.1.4) with inseparability in characteristic 2. Let ℘(𝑘) =
{𝑧 + 𝑧 2 : 𝑧 ∈ 𝑘 } be the Artin–Schreier group of 𝑘. The polynomial 𝑥 2 + 𝑥 + 𝑎 ∈ 𝑘 [𝑥]
is reducible if and only if 𝑎 ∈ ℘(𝑘), and since 𝑘 is finite, 𝑘/℘(𝑘) ' Z/2Z (Exercise
12.9).
Let 𝑡 ∈ 𝑅 be such that its reduction to 𝑘 represents the nontrivial class in 𝑘/℘(𝑘).
Proof of Lemma 12.3.10. We begin with the first part of the statement. Let 𝑄 be an
anisotropic binary quadratic form over 𝑘, with 𝑇 the associated bilinear form. The
quadratic form 𝑄 h−1i is isotropic by Lemma 12.3.5, and 𝑄 is anisotropic, so 𝑄
represents 1, say 𝑄(𝑥) = 1. Extending to a basis, we may rescale the second basis
element 𝑦 so that 𝑇 (𝑥, 𝑦) = 0, 1.
• Suppose 𝑇 (𝑥, 𝑦) = 0, so 𝑄 = h1, −𝑏i for 𝑏 ∈ 𝑘 × . Since 𝑄 is isotropic, 𝑏 ∉ 𝑘 ×2 ,
so char 𝑘 ≠ 2, the class 𝑏 ∈ 𝑘 × /𝑘 ×2 is unique, and indeed 𝑄 is anisotropic.
• Suppose 𝑇 (𝑥, 𝑦) = 1. If char 𝑘 ≠ 2, we may complete the square and reduce to
the previous case, so we suppose char 𝑘 = 2 and 𝑄(𝑥, 𝑦) = 𝑥 2 + 𝑥𝑦 + 𝑐𝑦 2 with
𝑐 ∈ 𝑘. Working now with the Artin–Schreier group, since 𝑄 is anisotropic we
must have 𝑐 + ℘(𝑘) = 𝑡 + ℘(𝑘) ∈ 𝑘, giving uniqueness.
For the second statement, if 𝑄 is a binary quadratic form over 𝑅 whose reduction
modulo 𝔭 is anisotropic, then we can find a change of basis over 𝑘 to put 𝑄 𝑘 in the
unique isometry class found above; lifting this basis, we may suppose that 𝑄 𝑘 is equal
to this fixed form. The statement then follows by using Hensel’s lemma to lift the
identity between any two such lifts, and it is a nice application of Hensel’s lemma in
two (several) variables: see exercise (Exercise 12.12).
We now return to the proof of our theorem.
Proof of Theorem 12.3.4 (char 𝑘 = 2). We first claim that the form [1, 1, 𝑡] h𝜋i is
anisotropic; this follows from a straightforward modification of the argument as in the
proof when char 𝑘 ≠ 2 above.
We now show this form is the unique one up to similarity. Suppose that 𝑄 is a
ternary anisotropic form over 𝑅. Let 𝑥 ∈ 𝑉 be nonzero; since 𝑄 is anisotropic, we
may scale 𝑥 so that 𝑎 := 𝑄(𝑥) ∈ 𝑅. Since dim 𝑉 ≥ 3, there exists nonzero 𝑦 0 ∈ 𝑉
such that 𝑇 (𝑥, 𝑦 0) = 0; rescale 𝑦 0 so that 𝑄(𝑦 0) ∈ 𝑅. Let 𝑦 := 𝑥 + 𝑦 0. Then 𝑇 (𝑥, 𝑦) =
𝑇 (𝑥, 𝑥 + 𝑦 0) = 𝑎 and 𝑏 := 𝑄(𝑦) ∈ 𝑅, so 𝑄 on this basis is 𝑎𝑥 2 + 𝑎𝑥𝑦 + 𝑏 2 ' [𝑎, 𝑎, 𝑏].
We compute that disc [𝑎, 𝑎, 𝑏] = 𝑎(𝑎 − 4𝑏) ≡ 𝑎 2 (mod 𝔭), so disc [𝑎, 𝑎, 𝑏] ∈ 𝑅 ×
and [𝑎, 𝑎, 𝑏] is nonsingular; completing to a basis with a nonzero element in the
orthogonal complement and rescaling, we may suppose without loss of generality that
𝑄 = [𝑎, 𝑎, 𝑏] h𝑐i with 𝑎, 𝑏, 𝑐 ∈ 𝑅 and both 𝑣(𝑎) = 0, 1 and 𝑣(𝑐) = 0, 1, with
disc 𝑄 = 𝑎(𝑎 − 4𝑏)𝑐. This leaves four cases.
• If 𝑣(𝑎) = 𝑣(𝑐) = 0, then reducing modulo 𝔭 we have disc 𝑄 𝑘 = 𝑎 2 𝑐 ≠ 0 so 𝑄 𝑘

is nondegenerate. By Lemma 12.3.5, 𝑄 𝑘 is isotropic. Hensel’s lemma (Lemma
12.2.18) applies to 𝑄, showing that 𝑄 is isotropic and giving a contradiction—we
omit the details.
• Suppose 𝑣(𝑎) = 0 and 𝑣(𝑐) = 1. Rescaling by a unit we may suppose 𝑐 = 𝜋.
The reduction [𝑎, 𝑎, 𝑏] 𝑘 modulo 𝔭 is nondegenerate, so if it is isotropic then
𝑄 is isotropic, a contradiction. Therefore by Lemma 12.3.10, we may suppose
𝑎 = 1 and 𝑏 = 𝑡; this is the desired form.
• Next, consider the case 𝑣(𝑎) = 1 and 𝑣(𝑐) = 0; we may suppose 𝑐 = 1. If
𝑣(𝑏) ≥ 2, then 𝜋 −1 [𝑎, 𝑎, 𝑏] ∼ [1, 1, 𝑏/𝜋] so the argument in the previous case
applies to show that 𝑄 is isotropic. If 𝑣(𝑏) = 1, then we scale 𝑧 by 𝜋 and divide
𝑄 by 𝜋 to reduce to the previous case. Finally, if 𝑣(𝑏) = 0, we scale 𝑦 by 𝜋
and 𝜋 −1 [𝑎, 𝑎, 𝑏𝜋 2 ] ∼ [1, 1, 𝑏(𝜋 2 /𝑎)] is again isotropic, a contradiction as in the
case 𝑣(𝑏) ≥ 2.
• In the case 𝑣(𝑎) = 𝑣(𝑐) = 1, if 𝑣(𝑏) > 0 then dividing through by 𝜋 we reduce to
the first case, so we may suppose 𝑣(𝑏) = 0. Then we multiply by 𝑎 and replace
𝑥 ← 𝑎𝑥, and 𝑦 ← 𝑦/𝑏 and 𝑧 ← 𝑧/𝜋 to work with [1, 𝑎/𝑏, 𝑎/𝑏] h𝑎𝑐/𝜋 2 i,
which after interchanging 𝑥, 𝑦 reduces to the previous case 𝑣(𝑎) = 1, 𝑣(𝑐) = 0.
Having exhausted the cases and the reader, the result now follows.
Corollary 12.3.12. Let 𝐹 be a nonarchimedean local field with valuation ring 𝑅 and
uniformizer 𝜋 ∈ 𝑅. Let 𝐵 be a quaternion algebra over 𝐹.
If char 𝑘 ≠ 2, then 𝐵 is a division algebra if and only if

𝑒, 𝜋
𝐵' , where 𝑒 ∈ 𝑅 × is nontrivial in 𝑘 × /𝑘 ×2
𝐹
and if char 𝐹 = char 𝑘 = 2, then 𝐵 is a division algebra if and only if

𝑡, 𝜋
𝐵' , where 𝑡 ∈ 𝑅 is nontrivial in 𝑘/℘(𝑘).
𝐹
12.4. HILBERT SYMBOL 187
In Theorem 13.3.11, we rephrase this corollary in terms of the unramified quadratic

extension of 𝐹.
Remark 12.3.13. In mixed characteristic where char 𝐹 = 0 and char 𝑘 = 2, in the
extension 𝐾 = 𝐹 [𝑥]/(𝑥 2 + 𝑥 + 𝑡) for 𝑡 nontrivial in 𝑘/℘(𝑘) we can complete the square
√
to obtain 𝐾 = 𝐹 ( 𝑒) with 𝑒 ∈ 𝐹 × r 𝐹 ×2 .
Definition 12.3.14. Let 𝐵 be a quaternion algebra over 𝐹. The Hasse invariant of 𝐵
is defined to be −1 if 𝐵 is a division algebra and +1 if 𝐵 ' M2 (𝐹).
12.4 Hilbert symbol

Let 𝐹 be a local field with char 𝐹 ≠ 2. We record the splitting behavior of quaternion
algebras as follows.
Definition 12.4.1. We define the Hilbert symbol
( , ) 𝐹 : 𝐹 × × 𝐹 × → {±1}

𝑎, 𝑏
by the condition that (𝑎, 𝑏) 𝐹 = 1 if and only if the quaternion algebra ' M2 (𝐹)
𝐹
is split.
The Hilbert symbol is well-defined as a map
𝐹 × /𝐹 ×2 × 𝐹 × /𝐹 ×2 → {±1}
(Exercise 2.4). By Main Theorem 5.4.4(v), we have (𝑎, 𝑏) 𝐹 = 1 if and only if the
Hilbert equation 𝑎𝑥 2 + 𝑏𝑦 2 = 1 has a solution with 𝑥, 𝑦 ∈ 𝐹: this is called Hilbert’s
criterion for the splitting of a quaternion algebra.

𝑎, 𝑏
Remark 12.4.2. The similarity between the symbols and (𝑎, 𝑏) 𝐹 is intentional;
𝐹
but they are not the same, as the former represents an algebra and the latter takes the
value ±1.
In some contexts, the Hilbert
symbol (𝑎, 𝑏) 𝐹 is defined to be the isomorphism class
𝑎, 𝑏
of the quaternion algebra in the Brauer group Br(𝐹), rather than ±1 according
𝐹
to whether or not the algebra is split. Conflating these two symbols is not uncommon
and in certain contexts it can be quite convenient, but we warn that it can lead to
confusion and caution against referring to a quaternion algebra or its isomorphism
class as a Hilbert symbol.
Lemma 12.4.3. Let 𝑎, 𝑏 ∈ 𝐹 × . Then the following statements hold:
(a) (𝑎𝑐2 , 𝑏𝑑 2 ) 𝐹 = (𝑎, 𝑏) 𝐹 for all 𝑐, 𝑑 ∈ 𝐹 × .
(b) (𝑏, 𝑎) 𝐹 = (𝑎, 𝑏) 𝐹 .
(c) (𝑎, 𝑏) 𝐹 = (𝑎, −𝑎𝑏) 𝐹 = (𝑏, −𝑎𝑏) 𝐹 .
(d) (1, 𝑎) 𝐹 = (𝑎, −𝑎) 𝐹 = 1.
(e) If 𝑎 ≠ 1, then (𝑎, 1 − 𝑎) 𝐹 = 1.
(f) If 𝜎 ∈ Aut(𝐹), then (𝑎, 𝑏) 𝐹 = (𝜎(𝑎), 𝜎(𝑏)) 𝐹 .
Proof. Statements (a)–(c) follow from Exercise 2.4. For (d), the Hilbert equation
𝑥 2 + 𝑎𝑦 2 = 1 has the obvious solution (𝑥, 𝑦) = (1, 0). And h𝑎, −𝑎i is isotropic (taking
(𝑥, 𝑦) = (1, 1)) so is a hyperbolic plane and represents 1 as in the proof of Main
Theorem 5.4.4, or we argue
(𝑎, −𝑎) 𝐹 = (𝑎, 𝑎 2 ) 𝐹 = (𝑎, 1) 𝐹 = (1, 𝑎) 𝐹 = 1
by Exercise 2.4. For part (e), by Hilbert’s criterion (𝑎, 1 − 𝑎) 𝐹 = 1 since the quadratic
equation 𝑎𝑥 2 +(1−𝑎)𝑦 2 = 1 has the solution (𝑥, 𝑦) = (1, 1). Finally, part (f): the Hilbert
equation 𝑎𝑥 2 + 𝑏𝑦 2 = 1 has a solution with 𝑥, 𝑦 ∈ 𝐹 if and only if 𝜎(𝑎)𝑥 2 + 𝜎(𝑏)𝑦 2 = 1
has such a solution.
Remark 12.4.4. Staring at the properties in Lemma 12.4.3 and seeking to axiomatize
them, the study of symbols like the Hilbert symbol leads naturally to the definition of
𝐾2 (𝐹). In its various formulations, algebraic 𝐾-theory (𝐾 for the German “Klasse”,
following Grothendieck) seeks to understand certain kinds of functors from rings to
abelian groups in a universal sense, encoded in groups 𝐾𝑛 (𝑅) for 𝑛 ∈ Z ≥0 and 𝑅 a
commutative ring: see e.g. Karoubi [Kar2010]. For a field 𝐹, we have 𝐾0 (𝐹) = Z and
𝐾1 (𝐹) = 𝐹 × . By a theorem of Matsumoto [Mat69] (see also Milnor [Milno71]), the
group 𝐾2 (𝐹) is the universal domain for symbols over 𝐹:
𝐾2 (𝐹) := (𝐹 × ⊗Z 𝐹 × )/h𝑎 ⊗ (1 − 𝑎) : 𝑎 ≠ 0, 1i.
(The tensor product over Z views 𝐹 × as an abelian group and therefore a Z-module.)
The map 𝑎 ⊗ 𝑏 ↦→ (𝑎, 𝑏) 𝐹 extends to a map 𝐾2 (𝐹) → {±1}, a Steinberg symbol, a
homomorphism from 𝐾2 (𝐹) to a multiplicative abelian group. The higher 𝐾-groups
are related to deeper arithmetic of commutative rings. For an introduction, see Weibel
[Weib2013] and Curtis–Reiner [CR87, Chapter 5].
We now turn to be quite explicit about the values of the Hilbert symbol. We begin
with the case where 𝐹 is archimedean. If 𝐹 = C, then the Hilbert symbol is identically
1. If 𝐹 = R, then (
1, if 𝑎 > 0 or 𝑏 > 0;
(𝑎, 𝑏)R = (12.4.5)
−1, if 𝑎 < 0 and 𝑏 < 0.
Lemma 12.4.6. The Hilbert symbol defines a nondegenerate symmetric bimultiplica-

tive pairing
( , ) 𝐹 : 𝐹 × /𝐹 ×2 × 𝐹 × /𝐹 ×2 → {±1}.
By bimultiplicativity, we mean that
(𝑎, 𝑏𝑐) 𝐹 = (𝑎, 𝑏) 𝐹 (𝑎, 𝑐) 𝐹 and (𝑎𝑏, 𝑐) 𝐹 = (𝑎, 𝑐) 𝐹 (𝑏, 𝑐) 𝐹 (12.4.7)
for all 𝑎, 𝑏, 𝑐 ∈ 𝐹 × (equivalent, by symmetry).

Keeping in the vibe of this section, we give a proof under the hypothesis that
char 𝑘 ≠ 2; for a general proof, see Corollary 13.4.6.
Proof (char 𝑘 ≠ 2). This lemma can be read off of the direct computation below
(12.4.9), recording what was computed along the way in the proof of Theorem
12.3.4.
12.4.8. Since the Hilbert symbol is well-defined up to squares, the symbol (𝑎, 𝑏) 𝐹 is
determined by the values with 𝑎, 𝑏 ∈ {1, 𝑒, 𝜋, 𝑒𝜋} where 𝑒 is a nonsquare in 𝑘 × . Let
𝑠 = (−1) (#𝑘−1)/2 , so that 𝑠 = 1, −1 according as −1 is a square in 𝑘. Then:
(𝑎, 𝑏) 𝐹 1 𝑒 𝜋 𝑒𝜋
1 1 1 1 1
𝑒 1 1 −1 −1 (12.4.9)
𝜋 1 −1 𝑠 −𝑠
𝑒𝜋 1 −1 −𝑠 𝑠
The computation of this table is requested in Exercise 12.15. For example, if 𝑎 = 𝑒, we

showed in the proof of Theorem 12.3.4 that (𝑒, 1) 𝐹 = (𝑒, 𝑒) 𝐹 = 1, because 𝑒𝑥 2 + 𝑦 2 − 𝑧2
and 𝑒𝑥 2 + 𝑒𝑦 2 − 𝑧 2 are isotropic.
In general, writing 𝑎 = 𝑎 0 𝜋 𝑣 (𝑎) and 𝑏 = 𝑏 0 𝜋 𝑣 (𝑏) we have
𝑣 (𝑏) 𝑣 (𝑎)
𝑎0 𝑏0
(𝑎, 𝑏) 𝐹 = (−1) 𝑣 (𝑎) 𝑣 (𝑏) (𝑞−1)/2 (12.4.10)
𝔭 𝔭
where 𝑞 = #𝑘 and
𝑐
= 0, ±1 ≡ 𝑐 (𝑞−1)/2 (mod 𝔭) (12.4.11)
𝔭
is the Legendre symbol: see Exercise 12.16. (Multiplicativity can also be read off of
the formula 12.4.10.)
12.4.12. The following easy criteria follow from 12.4.9 (or (12.4.10)):
(a) If 𝑣(𝑎𝑏) = 0, then (𝑎, 𝑏) 𝐹 = 1.

(b) If 𝑣(𝑎) = 0 and 𝑣(𝑏) = 𝑣(𝜋), then
(
𝑎 1 if 𝑎 ∈ 𝑘 ×2 ;
(𝑎, 𝑏) 𝐹 = =
𝔭 −1 if 𝑎 ∈ 𝑘 × r 𝑘 ×2 .
12.4.13. To compute the Hilbert symbol for a local field 𝐹 with char 𝐹 = 0 and
char 𝑘 = 2 is significantly more involved. But we can at least compute the Hilbert
symbol by hand for 𝐹 = Q2 .
To begin, the group Q×2 /Q×2
2 is generated by −1, −3, 2, so representatives are
{±1, ±3, ±2, ±6}. We recall Hilbert’s criterion: (𝑎, 𝑏) 𝐹 = 1 if and only if 𝑎𝑥 2 +𝑏𝑦 2 = 1
has a solution with 𝑥, 𝑦 ∈ 𝐹.
If 𝑎, 𝑏 ∈ Z are odd, then
𝑎𝑥 2 + 𝑏𝑦 2 = 𝑧2 has a nontrivial solution in Q2

⇔ 𝑎 ≡ 1 (mod 4) or 𝑏 ≡ 1 (mod 4);
by homogeneity and Hensel’s lemma, it is enough to check for a solution modulo 4.

This deals with all of the symbols with 𝑎, 𝑏 odd: summarizing, we have in this case
(𝑎, 𝑏)2 = (−1) (𝑎−1) (𝑏−1)/4 . (12.4.14)
By the determination above, we see that (−3, 𝑏) = −1 for 𝑏 = ±2, ±6 and (2, 2)2 =
(−1, 2)2 = 1 the latter by Hilbert’s criterion, as −1 + 2 = 1; knowing multiplicativity
(Lemma 12.4.6), we have uniquely determined all Hilbert symbols, in particular, for
𝑎 ∈ Z odd we have
2
(𝑎, 2)2 = (−1) (𝑎 −1)/8 . (12.4.15)
It is still useful to compute several of these symbols individually, in the same manner
as (12.4.13) (working modulo 8): see Exercise 12.17. We summarize the results here:
(𝑎, 𝑏)2 1 −3 −1 3 2 −6 −2 6
1 1 1 1 1 1 1 1 1
−3 1 1 1 1 −1 −1 −1 −1
−1 1 1 −1 −1 1 1 −1 −1
3 1 1 −1 −1 −1 −1 1 1 (12.4.16)
2 1 −1 1 −1 1 −1 1 −1
−6 1 −1 1 −1 −1 1 −1 1
−2 1 −1 −1 1 1 −1 −1 1
6 1 −1 −1 1 −1 1 1 −1
Remark 12.4.17. Analogously, one can define a symbol [𝑎, 𝑏) 𝐹 for the splitting of
quaternion algebras for 𝐹 a local field with char 𝐹 = 2. This symbol is no longer called
the Hilbert symbol, but many properties
remain: in particular, there is still an analogue
𝑎, 𝑏
of the Hilbert equation, and is split if and only if 𝑏𝑥 2 + 𝑏𝑥𝑦 + 𝑎𝑏𝑦 2 = 1 has a
𝐹
solution with 𝑥, 𝑦 ∈ 𝐹.
Exercises
1. Let 𝑝 be an odd prime. Show
(a) Show the equality
∞
∑︁
(1 − 4𝑥) 1/2 = 1 − 𝐶𝑛 𝑥 𝑛 ∈ Z[[𝑥]]
𝑛=1
of formal series in 𝑥 with coefficients in Z, where

1 2𝑛
𝐶𝑛 := ∈ Z>0
2𝑛 − 1 𝑛
are the Catalan numbers. [Hint: use binomial expansion.]

(b) Let 𝑝 be an odd prime. Show that the squaring map is bĳective on 1 + 𝑝Z 𝑝 .
[Hint: show that the series expansion in (a) converges in Z 𝑝 .]
2. Recall that a topological space is T1 if for every pair of distinct points, each
point has an open neighborhood not containing the other.
(a) Show that a topological space 𝑋 is T1 if and only if {𝑥} is closed for all
𝑥 ∈ 𝑋.
(b) Let 𝐺 be a topological group. Show that 𝐺 is Hausdorff if and only if 𝐺
is T1 .
3. In this exercise we prove some basic facts about topological groups. Let 𝐺 be a
topological group.
(a) Let 𝐻 ≤ 𝐺 be a subgroup. Show that 𝐻 is open if and only if there exists
ℎ ∈ 𝐻 and an open neighborhood of ℎ contained in 𝐻.
(b) Show that if 𝐻 ≤ 𝐺 is an open subgroup, then 𝐻 is closed.
(c) Show that a closed subgroup 𝐻 ≤ 𝐺 of finite index is open.
(d) Suppose that 𝐺 is compact. Show that an open subgroup 𝐻 ≤ 𝐺 is of finite
index, and that every open subgroup contains an open normal subgroup.
I 4. Let 𝐺 be a topological group. Let 𝑈 3 1 be an open neighborhood of 1.
(a) Show that there exists an open neighborhood 𝑉 ⊆ 𝑈 of 1 ∈ 𝑉 such that
𝑉 2 = 𝑉 · 𝑉 ⊆ 𝑈. [Hint: Multiplication is continuous.]
(b) Similarly, show that there exists an open neighborhood 𝑉 ⊆ 𝑈 of 1 ∈ 𝑉
such that 𝑉 −1𝑉 ⊆ 𝑈.
5. Let 𝐺 be a topological group and let 𝐻 ≤ 𝐺 be a closed subgroup. Equip 𝐺/𝐻
with the quotient topology. Show that 𝐺/𝐻 is Hausdorff. [Hint: Use Exercise
12.4(b).]
I 6. Let 𝑘 be a finite field and let 𝑄 : 𝑉 → 𝑘 be a ternary quadratic form. Show that 𝑄
is isotropic. [Hint: Reduce to the case of finding a solution to 𝑦 2 = 𝑓 (𝑥) where
𝑓 is a polynomial of degree 2. If #𝑘 is odd, count squares and the number of
distinct values taken by 𝑓 (𝑥) in 𝑘. Second approach: reduce to the case where
#𝑘 is odd, and show that 𝑥 2 + 𝑦 2 represents a nonsquare, since the squares
cannot be closed under addition!] [This repeats Exercise 5.5!]
I 7. Let 𝑘 be a finite field with char 𝑘 ≠ 2 and let 𝑒 ∈ 𝑘 × . Show directly that there is
an isometry h−1, 𝑒i ' h1, −𝑒i.
I 8. Let 𝑅 be a DVR with field of fractions 𝐹, let 𝑎, 𝑏, 𝑐 ∈ 𝐹 be nonzero and
let 𝑄 = h𝑎, −𝑏, −𝑐i. Show that 𝑄 is similar over 𝐹 to h1, −𝑏 0, −𝑐 0i with
0 = 𝑣(𝑏 0) ≤ 𝑣(𝑐 0). [Hint: first get 𝑣(𝑎), 𝑣(𝑏), 𝑣(𝑐) ∈ {0, 1}.]
I 9. Let 𝑘 be a finite field with even cardinality. Show that #𝑘/℘(𝑘) = 2, where ℘(𝑘)
is the Artin-Schreier group.
10. By Theorem 12.2.15, a complete archimedean local field is isomorphic to R or
C. Extend this classification to division algebras as follows.
The notion of absolute value (Definition 12.2.2) extends to a division algebra
without modification, as does the notion of archimedean and nonarchimedean.
√︁
(a) Show that H has an absolute value |𝛼| = nrd(𝛼) for 𝛼 ∈ H.
(b) Let 𝐷 be a division algebra equipped with an absolute value | |. Show
that if | | is archimedean, then char 𝐷 = 0 and if the restriction of | | to its
center 𝑍 (𝐷) is archimedean.

(c) Show that every division algebra complete with respect to an archimedean
absolute value is isomorphic to R, C, √or H and with the absolute value
equivalent to the absolute value |𝛼| = nrd 𝛼 in each case. [Hint: recall
Theorem 3.5.1.]
I 11. Prove Lemma 12.2.18 using Lemma 12.2.17. [Hint: let 𝑗 be the index that
achieves the minimal valuation among partial derivatives, and consider the
restriction 𝑓 (𝑎 1 , . . . , 𝑎 𝑗−1 , 𝑥, 𝑎 𝑗+1 , . . . , 𝑎 𝑛 ) ∈ 𝑅[𝑥] to one variable.]
I 12. In this exercise, we consider an extension of Hensel’s lemma to several polyno-
mials (in several variables).
Let 𝐹 be a nonarchimedean local field with valuation 𝑣 and valuation ring 𝑅.
(a) Let 𝑓1 , . . . , 𝑓𝑛 ∈ 𝑅[𝑥1 , . . . , 𝑥 𝑛 ] be polynomials with 𝑛 ≥ 1. Let 𝑎 ∈ 𝑅 𝑛
have 𝑚 := min𝑖 𝑣( 𝑓𝑖 (𝑎)). Let

𝜕 𝑓𝑖
D 𝑓 := (𝑎) ∈ 𝑀𝑛 (𝑅).
𝜕𝑥 𝑗 𝑖, 𝑗
Suppose 𝑚 > 2𝑣(det D 𝑓 (𝑎)). Show that there exists e 𝑎 ∈ 𝑅 𝑛 such that
𝑓𝑖 (𝑎) = 0 for all 𝑖 = 1, . . . , 𝑛 and e
𝑎 ≡ 𝑎 (mod 𝔭 ). [Hint: by Taylor
𝑚
expansion, write 𝑓 (𝑎 + 𝜋 𝑡 𝑥) = 𝑓 (𝑎) + D 𝑓 (𝑎) 𝑝 𝑡 𝑥 + 𝑝 2𝑡 𝑟 (𝑥) in vector

notation with 𝑡 ≥ 𝑚, and iteratively solve the system in a manner analogous
to Newton’s method. See Conrad [Con] for a complete development.]
(b) Show that (a) also holds more generally if the number of polynomials 𝑟 has
𝑟 ≤ 𝑛 and there exists an 𝑟 × 𝑟 matrix minor of D 𝑓 (𝑎) whose determinant
has valuation < 𝑚/2. [Hint: see Exercise 12.11.]
(c) Finish the proof of Lemma 12.3.10 as follows. Let 𝑄, 𝑄 0 be two binary
quadratic forms over 𝑅 such that their reductions 𝑄 𝑘 , 𝑄 0𝑘 are anisotropic.
Show that 𝑄 ' 𝑄 0. [Hint: reduce to the case where 𝑄 𝑘 = 𝑄 0𝑘 . If
𝑄 𝑘 ' h1, −𝑒i, rescale the basis using a plain vanilla flavor of Hensel’s
lemma. If 𝑄 𝑘 ' [1, 1, 𝑡], reduce to showing that 𝑄 ' [1, 1, 𝑡]. Consider
a general change of variables in GL2 (𝑅) from [1, 1, 𝑡] that reduces to the
identity modulo 𝔭 and apply the deluxe version of Hensel’s lemma in part
(b) to the resulting system of three equations in four unknowns: there is a
minor with determinant 1 − 4𝑡 ∈ 𝑅 × .]
13. Let 𝐹 ≠ C be a local field and let 𝑄 be a nondegenerate ternary quadratic form
over 𝐹. Let 𝐾 ⊇ 𝐹 be a quadratic field extension. Show that 𝑄 is isotropic over
𝐾.
14. Give another proof of Lemma 12.4.6 that the local Hilbert symbol is bimulti-
plicative using Example 8.2.2 and the Brauer group (section 8.3).
I 15. Show that the table of Hilbert symbols (12.4.9) is correct.
I 16. One can package 12.4.8 together with multiplying by squares to prove the fol-
lowing more general criterion. Let 𝐹 be a nonarchimedean local field with
uniformizer 𝜋, valuation 𝑣 with 𝑣(𝜋) = 1, and residue field 𝑘. Let 𝑞 = #𝑘 and
suppose 𝑞 is odd.
Show that for 𝑎, 𝑏 ∈ 𝐹 × , if we write 𝑎 = 𝑎 0 𝜋 𝑣 (𝑎) and 𝑏 = 𝑏 0 𝜋 𝑣 (𝑏) , then

𝑣 (𝑏) 𝑣 (𝑎)
𝑎0 𝑏0
(𝑎, 𝑏) 𝐹 = (−1) 𝑣 (𝑎) 𝑣 (𝑏) (𝑞−1)/2 .
𝜋 𝜋
I 17. Show that the table of Hilbert symbols (12.4.16) is correct by considering the
equation 𝑎𝑥 2 + 𝑏𝑦 2 ≡ 1 (mod 8).
18. Prove a descent for the Hilbert symbol, as follows. Let 𝐾 be a finite extension
of the local field 𝐹 with char 𝐹 ≠ 2 and let 𝑎, 𝑏 ∈ 𝐹 × . Show that (𝑎, 𝑏) 𝐾 =
(𝑎, Nm𝐾 |𝐹 (𝑏)) 𝐹 = (Nm𝐾 |𝐹 (𝑎), 𝑏) 𝐹 .
Chapter 13
Quaternion algebras over local fields
In this chapter, we approach the classification of quaternion algebras over local fields
in a second way, using valuations.
13.1 Extending the valuation

Recall (section 12.1) the valuation 𝑣 = 𝑣 𝑝 on Q 𝑝 , measuring divisibility by 𝑝. We
have
Z 𝑝 = {𝑥 ∈ Q 𝑝 : 𝑣(𝑥) ≥ 0}, and
(13.1.1)
𝑝Z 𝑝 = {𝑥 ∈ Q 𝑝 : 𝑣(𝑥) > 0};
Indeed, these can profitably be taken as their definition.
For any finite extension 𝐾 ⊇ Q 𝑝 of fields, there is a unique valuation 𝑤 on 𝐾 such
that 𝑤| Q 𝑝 = 𝑣 (so 𝑤 extends 𝑣), defined by
𝑣(Nm𝐾 |Q 𝑝 (𝑥))
𝑤(𝑥) := . (13.1.2)
[𝐾 : Q 𝑝 ]
The integral closure of Z 𝑝 in 𝐾 is the valuation ring {𝑥 ∈ 𝐾 : 𝑤(𝑥) ≥ 0} ⊇ Z 𝑝 , and

its unique maximal ideal is {𝑥 ∈ 𝐾 : 𝑤(𝑥) > 0}, as in (13.1.1).
For example,
√ there is a unique unramified quadratic extension 𝐾 of Q 𝑝 : we have
𝐾 = Q 𝑝 ( 𝑒), where 𝑒 = −3 for 𝑝 = 2 and otherwise 𝑒 ∈ Z is a quadratic nonresidue
modulo 𝑝 for 𝑝 odd. It is common to write 𝐾 = Q 𝑝2 for this field and Z 𝑝2 for its
valuation ring, since the residue field of 𝐾 is F 𝑝2 .
In a completely parallel fashion, let 𝐵 be a division quaternion algebra over Q 𝑝 .
Then there is again a unique valuation 𝑤 extending 𝑣, defined by
𝑤 : 𝐵 → R ∪ {∞}
𝑣(nrd(𝛼)) (13.1.3)
𝛼 ↦→ .
2
The valuation ring
O := {𝛼 ∈ 𝐵 : 𝑤(𝛼) ≥ 0} (13.1.4)
195
196 CHAPTER 13. QUATERNION ALGEBRAS OVER LOCAL FIELDS
is the unique (!) maximal 𝑅-order in 𝐵, consisting of all elements of 𝐵 that are integral
over Z 𝑝 . The set
𝑃 := {𝛼 ∈ 𝐵 : 𝑤(𝛼) > 0} (13.1.5)
is the unique maximal two-sided (bilateral) ideal of O.
Using the unique extension of the valuation, we obtain the following main result
of this chapter (a special case of Theorem 13.3.11).
Theorem 13.1.6. Let 𝑞 := 𝑝 2 . Then the following statements hold.
(a) There is a unique

division
quaternion algebra 𝐵 over Q 𝑝 , up to isomorphism
Q𝑞 , 𝑝
given by 𝐵 ' .
Q𝑝
(b) The valuation ring of 𝐵 is O ' Z𝑞 ⊕ Z𝑞 𝑗.
(c) The maximal ideal 𝑃 = O 𝑗 has 𝑃2 = 𝑝O and O/𝑃 ' Z𝑞 /𝑝Z𝑞 ' F𝑞 .
The method of proof used in this classification can also be used to classify central
division algebras over local fields in much the same manner.
13.2 Valuations
To begin, we briefly review extensions of valuations; for further reading, see the
references given in section 12.2.
Let 𝑅 be a complete DVR with valuation 𝑣 : 𝑅 → Z ≥0 ∪ {∞}, field of fractions
𝐹, maximal ideal 𝔭 generated by a uniformizer 𝜋 (with 𝑣(𝜋) = 1), and residue field
𝑘 := 𝑅/𝔭. Then 𝑅 is an integrally closed PID (every ideal is a power of the maximal
ideal 𝔭), and 𝑅 = {𝑥 ∈ 𝐹 : 𝑣(𝑥) ≥ 0}. Let | | 𝑣 be an absolute value attached to 𝑣, as in
(12.2.11).
Let 𝐾 ⊇ 𝐹 be a finite separable extension of degree 𝑛 := [𝐾 : 𝐹]. Then in fact 𝐾
is also a nonarchimedean local field; more precisely, we have the following lemma.
Lemma 13.2.1. There exists a unique valuation 𝑤 on 𝐾 such that 𝑤| 𝐹 = 𝑣, defined by
𝑣(Nm𝐾 |𝐹 (𝑥))
𝑤(𝑥) := . (13.2.2)
[𝐾 : 𝐹]
The integral closure of 𝑅 in 𝐾 is the valuation ring
𝑆 := {𝑥 ∈ 𝐾 : 𝑤(𝑥) ≥ 0}.
When 𝑤| 𝐹 = 𝑣, we say that 𝑤 extends 𝑣.
Proof. See e.g. Neukrich [Neu99, Chapter II, Theorem (4.8)], Cassels [Cas86, Chapter
7, Theorem 1.1], or Serre [Ser79, Chapter II, §2, Proposition 3].
By the same token using (13.2.2), there exists a unique absolute value | | 𝑤 on 𝐾
which restricts to | | 𝑣 on 𝐹; we pass freely between these two formulations.
13.2. VALUATIONS 197
13.2.3. We say 𝐾 ⊇ 𝐹 is unramified if a uniformizer 𝜋 for 𝐹 is also a uniformizer for

𝐾. We say 𝐾 ⊇ 𝐹 is totally ramified if a uniformizer 𝜋 𝐾 has the property that 𝜋 𝐾
𝑛 is
a uniformizer for 𝐹.
In general, there is a (unique) maximal unramified subextension 𝐾un ⊆ 𝐾, and the
extension 𝐾 ⊇ 𝐾un is totally ramified.
𝐾
𝑒
𝐾un
𝑓
𝐹
We say that 𝑒 = [𝐾 : 𝐾un ] is the ramification degree and 𝑓 = [𝐾un : 𝐹] the inertial
degree, and the fundamental equality
𝑛 = [𝐾 : 𝐹] = 𝑒 𝑓 (13.2.4)
holds.
13.2.5. Suppose that 𝐹 is a local field (equivalently, the residue field 𝑘 is a finite field).
Then for all 𝑓 ∈ Z ≥1 , there is a unique unramified extension of 𝐹 of degree 𝑓 and such
a field corresponds to the unique extension of the residue field 𝑘 of degree 𝑓 . In an
unramified extension 𝐾 ⊇ 𝐹 of degree [𝐾 : 𝐹] = 𝑓 , we have Nm𝐾 |𝐹 (𝐾 × ) = 𝑅 × 𝜋 𝑓 Z ,
so 𝑏 ∈ Nm𝐾 |𝐹 (𝐾 × ) if and only if 𝑓 | 𝑣(𝑏).
If char 𝑘 ≠ 2, then by Hensel’s lemma, the unramified extension of degree 2 is
given by adjoining a square root of an element of 𝑅 which reduces to the unique
nontrivial class in 𝑘 × /𝑘 ×2 ; if char 𝑘 = 2, then the unramified extension of degree 2
is given by adjoining a root of the polynomial 𝑥 2 + 𝑥 + 𝑡 where 𝑡 ∈ 𝑅 reduces to an
element which is nontrivial in the Artin-Schreier group 𝑘/℘(𝑘) (recalling 12.3.11).
Before proceeding further, we describe local fields by their defining polynomials—
we will need this later in the study of norms and strong approximation.
Lemma 13.2.6 (Krasner’s lemma). Let 𝐾 ⊇ 𝐹 be a finite, Galois extension with
absolute value | | 𝑤 . Let 𝛼, 𝛽 ∈ 𝐾, and suppose that for all 𝜎 ∈ Gal(𝐾 | 𝐹) with
𝜎(𝛼) ≠ 𝛼, we have
|𝛼 − 𝛽| 𝑤 < |𝛼 − 𝜎(𝛼)| 𝑤 . (13.2.7)
Then 𝐹 (𝛼) ⊆ 𝐹 (𝛽).
Intuitively, we can think of Krasner’s lemma as telling us when 𝛽 is closer to 𝛼
than any of its conjugates, then 𝐹 (𝛽) contains 𝛼. It is for this reason that we state the
lemma in terms of absolute values (instead of valuations).
Proof. Let 𝜎 ∈ Gal(𝐾 | 𝐹 (𝛽)) have 𝜎(𝛼) ≠ 𝛼. Then by the ultrametric inequality,
|𝜎(𝛼) − 𝛼| 𝑤 = |𝜎(𝛼) − 𝛽 + 𝛽 − 𝛼| 𝑤 ≤ max(|𝜎(𝛼) − 𝛽| 𝑤 , | 𝛽 − 𝛼| 𝑤 )
(13.2.8)
= max(|𝜎(𝛼 − 𝛽)| 𝑤 , | 𝛽 − 𝛼| 𝑤 ) = |𝛼 − 𝛽| 𝑤 ,
the final equality a consequence of (13.2.2) and the fact that Galois conjugates have the
same norm. This contradicts the existence of 𝜎, so 𝜎(𝛼) = 𝛼 for all 𝜎 ∈ Gal(𝐾 | 𝐹 (𝛽)).
By Galois theory, we conclude that 𝐹 (𝛼) ⊆ 𝐹 (𝛽).
Corollary 13.2.9. Let 𝑓 (𝑥) = 𝑥 𝑛 + 𝑎 𝑛−1 𝑥 𝑛−1 + · · · + 𝑎 0 ∈ 𝐹 [𝑥] be a separable, monic

polynomial. Then there exists 𝛿 > 0 such that whenever 𝑔(𝑥) = 𝑥 𝑛 +𝑏 𝑛−1 𝑥 𝑛−1 +· · ·+𝑏 0 ∈
𝐹 [𝑥] has |𝑎 𝑖 − 𝑏 𝑖 | 𝑣 < 𝛿, then
𝐹 [𝑥]/( 𝑓 (𝑥)) ' 𝐹 [𝑥]/(𝑔(𝑥)).
In particular, if 𝑓 (𝑥) is irreducible then 𝑔(𝑥) is irreducible.
Proof. Since 𝑓 (𝑥) is separable, its discriminant disc( 𝑓 ) is nonzero. The discriminant
is a polynomial function in the coefficients, so by continuity (multivariate Taylor
expansion), there exists 𝛿1 > 0 such that if 𝑔(𝑥) = 𝑥 𝑛 + · · · + 𝑏 0 ∈ 𝐹 [𝑥] has |𝑎 𝑖 − 𝑏 𝑖 | 𝑣 <
𝛿1 for all 𝑖, then |disc(𝑔) − disc( 𝑓 )| 𝑣 < |disc( 𝑓 )| 𝑣 ; by the ultrametric inequality,
we conclude that for such 𝑔(𝑥) we have |disc(𝑔)| 𝑣 = |disc( 𝑓 )| 𝑣 so in particular
disc(𝑔) ≠ 0.
Let 𝑔(𝑥) be as in the previous paragraph; then 𝑔(𝑥) is separable. We first consider
the caseÎ where 𝑓 (𝑥) is irreducible. Let 𝐾Î⊇ 𝐹 be a splitting field for the polynomials
𝑛 𝑛
𝑓 (𝑥) = 𝑖=1 (𝑥 − 𝛼𝑖 ) ∈ 𝐾 [𝑥] and 𝑔(𝑥) = 𝑖=1 (𝑥 − 𝛽𝑖 ). Let | | 𝑤 on 𝐾 extend | | 𝑣 . Let
𝜖 := min |𝛼𝑖 − 𝛼 𝑗 | 𝑤 . (13.2.10)

𝑖≠ 𝑗
Finally, let
𝑛
Ö 𝑛
Ö
𝜌(𝑔) = 𝜌(𝑏 0 , . . . , 𝑏 𝑛−1 ) := 𝑔(𝛼𝑖 ) = (𝛼𝑖 − 𝛽 𝑗 ). (13.2.11)
𝑖=1 𝑖, 𝑗=1
The map 𝑔 ↦→ 𝜌(𝑔) is again a polynomial in the coefficients 𝑏 0 , . . . , 𝑏 𝑛−1 (indeed,

it is a polynomial resultant). Therefore there exists 𝛿 > 0, with 𝛿 < 𝛿1 , such that if
2
|𝑎 𝑖 − 𝑏 𝑖 | 𝑣 < 𝛿, then | 𝜌(𝑔)| 𝑤 < 𝜖 𝑛 . Therefore in (13.2.11), there exists 𝑖, 𝑗 such that
|𝛼𝑖 − 𝛽 𝑗 | 𝑤 < 𝜖. Together with (13.2.10), we have
|𝛼𝑖 − 𝛽 𝑗 | 𝑤 < 𝜖 ≤ |𝛼𝑖 − 𝛼 𝑘 | 𝑤
for all 𝑘 ≠ 𝑖. By Krasner’s lemma (Lemma 13.2.6), we conclude that 𝐹 (𝛼𝑖 ) ⊆ 𝐹 (𝛽 𝑗 ).

Since 𝑓 (𝛼𝑖 ) = 𝑔(𝛽 𝑗 ) = 0 and 𝑓 is irreducible, we have
[𝐹 (𝛼𝑖 ) : 𝐹] = 𝑛 ≤ [𝐹 (𝛽 𝑗 ) : 𝐹] ≤ 𝑛
so in fact [𝐹 (𝛽 𝑗 ) : 𝐹] = 𝑛 and 𝐹 (𝛼𝑖 ) = 𝐹 (𝛽 𝑗 ). Finally,
𝐹 [𝑥]/( 𝑓 (𝑥)) ' 𝐹 (𝛼𝑖 ) = 𝐹 (𝛽 𝑗 ) ' 𝐹 [𝑥]/(𝑔(𝑥))
as desired.
The case when 𝑓 (𝑥) = 𝑓1 (𝑥) · · · 𝑓𝑟 (𝑥) is reducible follows by repeating the above
argument on each factor, and finishing using the continuity of multiplication among
the coefficients: the details are requested in Exercise 13.17.
13.3. CLASSIFICATION VIA EXTENSIONS OF VALUATIONS 199
13.3 Classification via extensions of valuations

We now seek to generalize this setup to the noncommutative case; we retain the
notation from the previous section. Let 𝐷 be a central (simple) division algebra over
𝐹 with dim𝐹 𝐷 = [𝐷 : 𝐹] = 𝑛2 . We extend the valuation 𝑣 to a map
𝑤 : 𝐷 → R ∪ {∞}
𝑣(Nm𝐷 |𝐹 (𝛼)) 𝑣(nrd(𝛼)) (13.3.1)
𝛼 ↦→ = ,
[𝐷 : 𝐹] 𝑛
where the equality follows from the fact that Nm𝐷 |𝐹 (𝛼) = nrd(𝛼) 𝑛 (see section 7.8).
Lemma 13.3.2. The map 𝑤 is the unique valuation on 𝐷 extending 𝑣, i.e., the following
hold:
(i) 𝑤(𝛼) = ∞ if and only if 𝛼 = 0.

(ii) 𝑤(𝛼𝛽) = 𝑤(𝛼) + 𝑤(𝛽) = 𝑤(𝛽𝛼) for all 𝛼, 𝛽 ∈ 𝐷.
(iii) 𝑤(𝛼 + 𝛽) ≥ min(𝑤(𝛼), 𝑤(𝛽)) for all 𝛼, 𝛽 ∈ 𝐷.
(iv) 𝑤(𝐷 × ) is discrete in R.
Proof. Since 𝐷 is a division ring, statement (i) is immediate. Statement (ii) follows
from the multiplicativity of nrd and 𝑣. To prove (iii), we may suppose 𝛽 ≠ 0 and so
𝛽 ∈ 𝐷 × . We have
𝑤(𝛼 + 𝛽) = 𝑤((𝛼𝛽−1 + 1) 𝛽) = 𝑤(𝛼𝛽−1 + 1) + 𝑤(𝛽).
But the restriction of 𝑤 to 𝐹 (𝛼𝛽−1 ) is a discrete valuation, thus 𝑤(𝛼𝛽−1 + 1) ≥

min(𝑤(𝛼𝛽−1 ), 𝑤(1)) and by (ii) 𝑤(𝛼 + 𝛽) ≥ min(𝑤(𝛼), 𝑤(𝛽)), as desired. Finally,
(iv) holds since 𝑤(𝐷 × ) ⊆ 𝑣(𝐹 × )/𝑛 and the latter is discrete. The valuation is unique
because it is unique whenever it is restricted to a subfield.
13.3.3. From Lemma 13.3.2, we say that 𝑤 is a discrete valuation on 𝐷 since it

satisfies the same axioms as for a field. It follows from Lemma 13.3.2 that the set
O := {𝛼 ∈ 𝐷 : 𝑤(𝛼) ≥ 0}
is a ring, called the valuation ring of 𝐷.
Proposition 13.3.4. The ring O is the unique maximal 𝑅-order in 𝐷, consisting of all
elements of 𝐷 that are integral over 𝑅.
Proof. First, we prove that
O = {𝛼 ∈ 𝐷 : 𝛼 is integral over 𝑅}. (13.3.5)
We first show the inclusion (⊇) of (13.3.5), and suppose 𝛼 ∈ 𝐷 is integral over
𝑅. Since 𝑅 is integrally closed, by Lemma 10.3.5 the coefficients of the minimal
polynomial 𝑓 (𝑥) ∈ 𝐹 [𝑥] of 𝛼 belong to 𝑅. Since 𝐷 is a division ring, 𝑓 (𝑥) is
irreducible and hence the reduced characteristic polynomial 𝑔(𝑥) is a power of 𝑓 (𝑥)
and thus has coefficients in 𝑅. Up to sign, the constant coefficient of 𝑔(𝑥) is nrd(𝛼),
so 𝑤(𝛼) = 𝑣(nrd(𝛼)) ≥ 0, hence 𝛼 ∈ O.
Next we prove (⊆) in (13.3.5). Suppose 𝛼 ∈ O, so that 𝑤(𝛼) ≥ 0, and let
𝐾 = 𝐹 (𝛼). Let 𝑓 (𝑥) ∈ 𝐹 [𝑥] be the minimal polynomial of 𝛼. We want to conclude
that 𝑓 (𝑥) ∈ 𝑅[𝑥] knowing that 𝑤(𝛼) ≥ 0. But the restriction of 𝑤 to 𝐾 is the unique
extension of 𝑣 to 𝐾, and this is a statement about the extension 𝐾 ⊇ 𝐹 of fields and
therefore follows from the commutative case, Lemma 13.2.1.
We can now prove that O is an 𝑅-order. Scaling an element of 𝐷 × by an appropriate
power of 𝜋 gives it positive valuation, so O𝐹 = 𝐷. To conclude, we must show that O
is finitely generated as an 𝑅-module. Recall that 𝐷 is a central division algebra over 𝐹,
hence a separable 𝐹-algebra, so we may apply Lemma 10.3.7: every 𝛼 ∈ O is integral
over 𝑅 and O is a ring, and the lemma implies that O is an 𝑅-order.
Finally, it follows immediately that O is a maximal 𝑅-order: by Corollary 10.3.3,
every element of an 𝑅-order is integral over 𝑅, and O contains all such elements.
Remark 13.3.6. For a quaternion division algebra 𝐷, we can argue more directly in
the proof of Proposition 13.3.4 using the reduced norm: see Exercise 13.4.
13.3.7. It follows from Proposition 13.3.4 that O is a finitely generated 𝑅-submodule
of 𝐷. But 𝑅 is a PID so in fact O is free of rank [𝐷 : 𝐹] as an 𝑅-module. We have
O× = {𝛼 ∈ 𝐷 : 𝑤(𝛼) = 0} (13.3.8)
since 𝑤(𝛼−1 ) = −𝑤(𝛼), and in particular 𝛼 ∈ O× if and only if nrd(𝛼) ∈ 𝑅 × .
Consequently,
𝑃 := {𝛼 ∈ 𝐷 : 𝑤(𝛼) > 0} = O r O× (13.3.9)
is the unique maximal two-sided (bilateral) ideal of O, as well as the unique left or
right ideal of O. Therefore O is a noncommutative local ring, a noncommutative
ring with a unique maximal left (equivalently, right) ideal.
13.3.10. Let 𝛽 ∈ 𝑃 have minimal (positive) valuation 𝑤(𝛽) > 0. Then for all
0 ≠ 𝛼 ∈ 𝑃 we have 𝑤(𝛼𝛽−1 ) = 𝑤(𝛼) − 𝑤(𝛽) ≥ 0 so 𝛼𝛽−1 ∈ O and 𝛼 ∈ O𝛽. Arguing
on the other side, we have also 𝛼 ∈ 𝛽O. Thus 𝑃 = O𝛽 = 𝛽O = O𝛽O.
Arguing in the same way, we see that every one-sided ideal of O is in fact two-sided,
and every two-sided ideal of O is principally generated by any element with minimal
valuation hence of the form 𝑃𝑟 for some 𝑟 ∈ Z ≥0 .
We are now prepared to give the second proof of the main result in this chapter
(Main Theorem 12.3.2). We now add the hypothesis that 𝐹 is a local field, so that 𝑘 is
a finite field.
Theorem 13.3.11. Let 𝐹 be a nonarchimedean local field. Then the following state-
ments hold.
(a) There is a unique division quaternion algebra 𝐵 over 𝐹, up to 𝐹-algebra iso-
morphism given by
𝐾, 𝜋
𝐵' ,
𝐹
where 𝐾 is the unique quadratic unramified (separable) extension of 𝐹.
13.3. CLASSIFICATION VIA EXTENSIONS OF VALUATIONS 201
(b) Let 𝐵 be as in (a). Then the valuation ring of 𝐵 is O ' 𝑆 ⊕ 𝑆 𝑗, where 𝑆 is the
integral closure of 𝑅 in 𝐾. Moreover, the ideal 𝑃 = O 𝑗 is the unique maximal
ideal; we have 𝑃2 = 𝜋O, and O/𝑃 ⊇ 𝑅/𝔭 is a quadratic extension of finite
fields.
Proof. We begin with existence in part (a), and existence: we prove that 𝐵 = (𝐾, 𝜋 | 𝐹)
is a division algebra. We recall that 𝐵 is a division ring if and only if 𝜋 ∉ Nm𝐾 |𝐹 (𝐾 × )
by Main Theorem 5.4.4 and Theorem 6.4.11. Since 𝐾 ⊇ 𝐹 is unramified, we have
Nm𝐾 |𝐹 (𝐾 × ) = 𝑅 × 𝜋 2Z by 13.2.5. Putting these together gives the result.
Continuing with (a), we now show uniqueness. Let 𝐵 be a division quaternion
algebra over 𝐹. We refer to 13.3.10, and let 𝑃 = O𝛽. Then 𝑤(𝛽) ∈ 12 Z>0 , so
𝑤(𝛽) ≤ 𝑤(𝜋) = 𝑣(𝜋) = 1 ≤ 2𝑤(𝛽) = 𝑤(𝛽2 ); (13.3.12)
we conclude that 𝛽O = 𝑃 ⊇ 𝜋O ⊇ = 𝑃2 𝛽2 O. The map 𝛼 ↦→ 𝛼𝛽 yields an
isomorphism O/𝑃 −→∼ 𝑃/𝑃2 of 𝑘-vector spaces, so
4 = dim 𝑘 (O/𝜋O) ≤ dim 𝑘 (O/𝑃2 ) = dim 𝑘 (O/𝑃) + dim 𝑘 (𝑃/𝑃2 ) = 2 dim 𝑘 (O/𝑃)
(13.3.13)
and thus dim 𝑘 (O/𝑃) ≥ 2, with equality if and only if 𝜋O = 𝑃2 .
As in (13.3.9), we have O r 𝑃 = O× , so the ring O/𝑃 is a division algebra over
𝑘 and hence a finite division ring. By Wedderburn’s little theorem (Exercise 7.29),
we conclude that O/𝑃 is a finite field! So there exists 𝑖 ∈ O such that its reduction
generates O/𝑃 as a finite extension of 𝑘. But 𝑖 satisfies its reduced characteristic
polynomial, a monic polynomial of degree 2 with coefficients in 𝑅, so its reduction
satisfies a polynomial of degree 2 with coefficients in 𝑘. Since 𝑖 is a generator, we
conclude [O/𝑃 : 𝑘] ≤ 2. Together with the conclusion of the previous paragraph,
we conclude that [O/𝑃 : 𝑘] = dim 𝑘 (O/𝑃) = 2, in other words O/𝑃 is a (separable)
quadratic field extension of 𝑘. It then follows from 13.2.5 that 𝐾 := 𝐹 (𝑖) is the unique,
unramified (separable) quadratic extension of 𝐹. Therefore equality holds in (13.3.13)
and 𝑃2 = 𝜋O. Since 𝛽2 O = 𝑃2 = 𝜋O, we have 𝑤(𝛽) = 1/2.
By Exercise 6.2 or 7.26, there exists 𝑏 ∈ 𝐹 × such that 𝐵 ' (𝐾, 𝑏 | 𝐹). Recalling the
first paragraph above, since 𝐵 is a division algebra, we have 𝑏 ∉ Nm𝐾 |𝐹 (𝐾 × ) = 𝑅 × 𝜋 2Z .
Applying Exercise 6.4, we may multiply 𝑏 by a norm from 𝐾 × , so we may suppose
𝑏 = 𝜋, and therefore 𝐵 ' (𝐾, 𝜋 | 𝐹). This concludes the proof of (a).
We turn now to (b), with 𝐵 = (𝐾, 𝜋 | 𝐹) = 𝐾 + 𝐾 𝑗 with 𝑗 2 = 𝜋. Let 𝛼 = 𝑢 + 𝑣 𝑗 ∈ 𝐵
with 𝑢, 𝑣 ∈ 𝐾. Then nrd(𝛼) = Nm𝐾 |𝐹 (𝑢) − 𝜋 Nm𝐾 |𝐹 (𝑣) = 𝑥 − 𝜋𝑦 with 𝑥, 𝑦 ∈ 𝐹 and
𝑣(𝑥) even and 𝑣(𝜋𝑦) odd (as norms from 𝐾). By the ultrametric inequality, we have
𝑤(𝛼) = 𝑣(nrd(𝛼)) ≥ 0 if and only if 𝑣(𝑥), 𝑣(𝑦) ≥ 0 if and only if 𝑢, 𝑣 ∈ 𝑆 (as 𝑆 is the
valuation ring of 𝐾). Therefore O = 𝑆 + 𝑆 𝑗. Since 𝑗 2 = 𝜋, we have 𝑤( 𝑗) = 1/2, so
𝑗O = 𝑃. The remaining statements were proven in the course of proving (a).
Corollary 13.3.14. Let 𝐹 ≠ C be a local field. Let 𝐾 be the unramified quadratic
extension of 𝐹, with h𝜎i = Gal(𝐾 | 𝐹). Then the 𝐹-subalgebra

𝑢 𝜋𝑣
𝐵= : 𝑢, 𝑣 ∈ 𝐾 ⊂ M2 (𝐾)
𝜎(𝑣) 𝜎(𝑢)
is the unique division quaternion algebra over 𝐹 (up to isomorphism).
Proof. Using Theorem 13.3.11(a), we split 𝐵 over 𝐾 as in 2.3.4. (We may also put 𝜋
below the diagonal as in 2.3.12.)
13.4 Consequences
We now observe a few consequences of Theorem 13.3.11.
Corollary 13.4.1. Let 𝐹 be a nonarchimedean local field with valuation 𝑣, let 𝐾 be

a separable, unramified quadratic 𝐹-algebra, and let 𝐵 = (𝐾, 𝑏 | 𝐹) with 𝑏 ∈ 𝐹 × . If
𝑣(𝑏) = 0, then 𝐵 ' M2 (𝐹).
Proof. Either 𝐾 ' 𝐹 × 𝐹 or 𝐾 ⊇ 𝐹 is the unique unramified quadratic field extension.

In the first case, 𝐾 has a zerodivisor so 𝐵 ' M2 (𝐹). In the second case, we conclude as
in the first paragraph of the proof of Theorem 13.3.11, since 𝑏 ∈ 𝑅 × ≤ Nm𝐾 |𝐹 (𝐾 × ).

13.4.2. Let 𝐵 be a division quaternion algebra over 𝐹. In analogy with the case of
field extensions (13.2.4), we define the ramification index of 𝐵 over 𝐹 as 𝑒(𝐵|𝐹) = 2
since 𝑃2 = 𝜋O, and the inertial degree of 𝐵 over 𝐹 as 𝑓 (𝐵|𝐹) = 2 since 𝐵 contains
the unramified quadratic extension 𝐾 of 𝐹, and note the equality
𝑒(𝐵|𝐹) 𝑓 (𝐵|𝐹) = 4 = [𝐵 : 𝐹],
as in the commutative case. (Viewed in this way, 𝐵 is obtained from first an unramified
extension and then a “noncommutative” ramified extension.)
Remark 13.4.3. Theorem 13.3.11, the fundamental result describing division quater-
nion algebras over a local field, is a special case of a more general result as follows.
Let 𝑅 be a complete DVR with maximal ideal 𝔭 = 𝜋𝑅 and 𝐹 := Frac(𝑅).
Let 𝐷 be a (finite-dimensional) division algebra over 𝐹, and let O ⊆ 𝐷 be the
valuation ring and 𝑃 ⊂ O the maximal ideal. Then 𝑃 𝑒 = 𝔭O for some 𝑒 ≥ 1, called the
ramification index; the quotient O/𝑃 is a division algebra over the field 𝑘 = 𝑅/𝔭, and
we let the inertial degree be 𝑓 = dim 𝑘 (O/𝑃). Then 𝑒 𝑓 = dim𝐹 𝐷 = 𝑛2 ; moreover,
if 𝑘 is finite (𝐹 is a local field), then 𝑒 = 𝑓 = 𝑛. For a proof, see Exercise 13.11; or
consult Reiner [Rei2003, Theorems 12.8, 13.3, 14.3]. However, the uniqueness of 𝐷
up to 𝐹-algebra isomorphism no longer holds. If 𝐹 is a local field, then the possibilities
for 𝐷 are classified up to isomorphism by a local invariant inv 𝐷 ∈ ( 𝑛1 Z)/Z ' Z/𝑛Z.
These patch together to give a global result: see Remark 14.6.10.
This classification can be further extended to an arbitrary central simple algebra
𝐵 ' M𝑛 (𝐷) over 𝐹: see Reiner [Rei2003, §17–18].
Splitting of local division quaternion algebras over extension fields is given by the
following simple criterion.
Proposition 13.4.4. Let 𝐵 be a division quaternion algebra over a local field 𝐹, and
let 𝐿 be a separable field extension of 𝐹 of finite degree. Then 𝐿 is a splitting field for
𝐵 if and only if [𝐿 : 𝐹] is even.
13.4. CONSEQUENCES 203
Proof. If 𝐹 is archimedean, then either 𝐹 = C and there is no such 𝐿, or 𝐹 = R and

𝐵 = H and 𝐿 = C, and the result holds. So suppose 𝐹 is nonarchimedean. We have
𝐵 ' (𝐾, 𝜋 | 𝐹) where 𝐾 is the unramified quadratic extension of 𝐹. Let 𝑒, 𝑓 be the
ramification index and inertial degree of 𝐿, respectively. Then [𝐿 : 𝐹] = 𝑛 = 𝑒 𝑓 , and
𝑛 is even if and only if 𝑒 is even or 𝑓 is even. But 𝑓 is even if and only if 𝐿 contains
an unramified quadratic subextension, necessarily isomorphic to 𝐾; but then 𝐾 splits
𝐵 so 𝐿 splits 𝐵.
Having established the claim when 𝑓 is even, suppose that 𝑓 is odd. Then 𝐿 is
linearly disjoint from 𝐾 and 𝐾 ⊗𝐹 𝐿 = 𝐾 𝐿 is the unramified quadratic extension of
𝐿. Therefore 𝐵 ⊗𝐹 𝐿 ' (𝐾 𝐿, 𝜋 | 𝐿). Let 𝑅 𝐿 be the valuation ring of 𝐿 and let 𝜋 𝐿
be a uniformizer for 𝐿. Then Nm𝐾 𝐿/𝐿 (𝐾 𝐿 × ) = 𝑅 ×𝐿 𝜋 2Z
𝐿 . We have 𝜋 = 𝑢𝜋 𝐿 for some
𝑒
×
𝑢 ∈ 𝑅 𝐿 . Putting these together, we see that 𝐵 ⊗𝐹 𝐿 is split if and only if 𝜋 is a norm
from 𝐾 𝐿 if and only if 𝑒 is even.
As a consequence, 𝐵 contains every separable quadratic extension of 𝐹.
Corollary 13.4.5. If 𝐵 is a division quaternion algebra over a local field 𝐹 and 𝐾 ⊇ 𝐹

is a separable quadratic field extension, then 𝐾 ↩→ 𝐵.
Proof. Combine Proposition 13.4.4 with Lemmas 5.4.7 and 6.4.12.
We repeat now Lemma 12.4.6, giving a proof that works without restriction on
characteristic.
Corollary 13.4.6. If char 𝐹 ≠ 2, then the Hilbert symbol defines a symmetric, nonde-
generate bilinear form on 𝐹 × /𝐹 ×2 .
Proof. Let 𝐾 := 𝐹 [𝑥]/(𝑥 2 − 𝑎). The Hilbert symbol gives a well-defined map of sets
𝐹 × /𝐹 ×2 → {±1}
𝑏 ↦→ (𝑎, 𝑏) 𝐹
and we may conclude as in Lemma 12.4.6 if we show that this is a nontrivial group
homomorphism.
First we show it is nontrivial. By Corollary 13.4.5, the field 𝐾 embeds in the
division quaternion algebra 𝐵, so by Exercise 2.5, there exists 𝑏 such that 𝐵 ' (𝑎, 𝑏 |
𝐹), whence (𝑎, 𝑏) 𝐹 = −1.
Next, we show it is a homomorphism. We appeal to Main Theorem 5.4.4. We have
(𝑎, 𝑏) 𝐹 = 1 if and only if 𝑏 ∈ Nm𝐾 |𝐹 (𝐾 × ). So we reduce to showing that if (𝑎, 𝑏) 𝐹 =
(𝑎, 𝑏 0) 𝐹 = −1 for 𝑏, 𝑏 0 ∈ 𝐹 × , then (𝑎, 𝑏𝑏 0) 𝐹 = 1. But by Corollary 7.7.6, since there
is a unique division quaternion algebra, we conclude that 𝑏/𝑏 0 ∈ Nm𝐾 |𝐹 (𝐾 × ); thus
𝑏𝑏 0 = (𝑏 0) 2 (𝑏/𝑏 0) ∈ Nm𝐾 |𝐹 (𝐾 × ) and (𝑎, 𝑏𝑏 0 | 𝐹) ' M2 (𝐹) so (𝑎, 𝑏𝑏 0) 𝐹 = 1, as
claimed.
Remark 13.4.7. The proof of Corollary 13.4.6 (pairing with any 𝐹 × /𝐹 ×2 ) shows that
𝐹 × /Nm𝐾 |𝐹 (𝐾 × ) ' Z/2Z. (13.4.8)

Conversely, if we know (13.4.8) then the properties of the Hilbert symbol are immedi-
ate. Although it was not hard to prove (13.4.8) when char 𝑘 ≠ 2, to establish its truth
when char 𝑘 = 2, one is led to study higher ramification groups (e.g. Serre [Ser79,
Chapter XV]) eventually leading to local class field theory.
The norm groups played an important role in the proof above, so we conclude by
recording the image of the reduced norm nrd(𝐵×𝑣 ) ⊆ 𝐹𝑣× .
Lemma 13.4.9. We have

(
× R×>0 , if 𝐵 ' H;
nrd(𝐵 ) =
𝐹×, otherwise.
Moreover, if 𝐹 is nonarchimedean and O ⊆ 𝐵 is a maximal 𝑅-order, then

nrd(O× ) = 𝑅 × .
Proof. If 𝐵 ' M2 (𝐹) is split, then nrd(𝐵× ) = det(GL2 (𝐹)) = 𝐹 × . So suppose 𝐵 is a

division algebra. If 𝐵 ' H then nrd(𝐵× ) = R×>0 , so we suppose 𝐹 is nonarchimedean.
Then 𝐵 ' (𝐾, 𝜋 | 𝐹) where as above 𝐾 is the unramified quadratic extension of 𝐹 and
𝜋 is a uniformizer. But 𝐹 × = 𝑅 × × h𝜋i, and nrd(𝐾 × ) = Nm𝐾 |𝐹 (𝐾 × ) = 𝑅 × 𝜋 2Z and
nrd( 𝑗) = 𝜋. The result then follows by multiplicativity of the norm.
The second statement follows similarly: if 𝐵 ' M2 (𝐹) then O ' M2 (𝑅) and
nrd(O× ) = det(GL2 (𝑅)) = 𝑅 × ; otherwise O ' (𝑆, 𝜋 | 𝑅) where 𝑆 is the ring of
integers of 𝐾, and nrd(𝑆 × ) = Nm𝐾 |𝐹 (𝑆 × ) ⊇ 𝑅 × and again nrd( 𝑗) = 𝜋.
13.5 Some topology

In this section, we dive into the basic topological adjectives relevant to the objects we
have seen and that will continue to play an important role. Throughout, let 𝐹 be a
local field.
13.5.1. 𝐹 is a locally compact topological field (by definition) but 𝐹 is not itself
compact. The subgroup 𝐹 × = 𝐹 r {0} is given the topology induced from the
embedding
𝐹 × ↩→ 𝐹 × 𝐹
𝑥 ↦→ (𝑥, 𝑥 −1 );
it turns out here that this coincides with the subspace topology 𝐹 × ⊆ 𝐹 (see Exercise
13.15(a)). Visibly, 𝐹 × is open in 𝐹 so 𝐹 × is locally compact.
If 𝐹 is nonarchimedean, with valuation ring 𝑅 and valuation 𝑣, then 𝐹 × is totally
disconnected and further
𝑅 × = {𝑥 ∈ 𝑅 : 𝑣(𝑥) = 0} ⊂ 𝑅
is closed so is a topological abelian group that is compact (and totally disconnected).
Now let 𝐵 be a finite-dimensional 𝐹-algebra.

13.5. SOME TOPOLOGY 205
13.5.2. As an 𝐹-vector space, 𝐵 has a unique topology compatible with the topology on
𝐹 as all norms on a topological vector 𝐹-space extending the norm on 𝐹 are equivalent
(the sup norm is equivalent to the sum of squares norm, etc.): see Exercise 13.13. In
particular, two elements are close in the topology on 𝐵 if and only if their coefficients
are close with respect to a (fixed) basis: for example, two matrices in M𝑛 (𝐹) are close
if and only if all of their coordinate entries are close. (Of course, the precise notion
of “close” depends on the choice of norm.) Consequently, 𝐵 is a complete, locally
compact topological ring, taking a compact neighborhood in each coordinate.
13.5.3. The group 𝐵× is a topological group, with the topology given by the embedding
𝐵× 3 𝛼 ↦→ (𝛼, 𝛼−1 ) ∈ 𝐵 × 𝐵. This topology coincides with the subspace topology
(see Exercise 13.15(b)). From this, we can see that 𝐵× is locally compact: the norm
Nm 𝐵 |𝐹 : 𝐵× → 𝐹 × is a continuous map, so 𝐵× = Nm−1 ×
𝐵 |𝐹 (𝐹 ) is open in 𝐵, and an
open subset of a Hausdorff, locally compact space is locally compact in the subspace
topology (Exercise 13.15(c)).
Example 13.5.4. If 𝐵 = M𝑛 (𝐹), then 𝐵× = GL𝑛 (𝐹) is locally compact: a closed,

bounded neighborhood that avoids the locus of matrices with determinant 0 is a
compact neighborhood. When 𝐹 is archimedean, this is quite visual: a matrix of
nonzero determinant is at some finite distance away from the determinant zero locus!
Note however that GL𝑛 (𝐹) is not itself compact: already 𝐹 × = GL1 (𝐹) is not compact.
Now suppose 𝐹 is nonarchimedean with valuation 𝑣 and valuation ring 𝑅.
13.5.5. We claim that 𝑅 is the maximal compact subring of 𝐹. Indeed, 𝑥 ∈ 𝐹 lies

in a compact subring if and only if 𝑣(𝑥) ≥ 0 if and only if 𝑥 is integral over 𝑅. The
only new implication here is the statement that if 𝑣(𝑥) < 0 then 𝑥 does not lie in a
compact subring, and that is because the sequence 𝑥 𝑛 = 𝑥 𝑛 does not have a convergent
subsequence as |𝑥 𝑛 | → ∞.
Next, let O be an 𝑅-order in 𝐵.
13.5.6. Choosing an 𝑅-basis, we have an isomorphism O ' 𝑅 𝑛 , and this isomorphism

is also a homeomorphism. Therefore, O is compact as the Cartesian power of a
compact set. The group O× is therefore also compact because it is closed: for 𝛾 ∈ O,
we have 𝛾 ∈ O× if and only if Nm 𝐵 |𝐹 (𝛾) ∈ 𝑅 × , the norm map is continuous, and
𝑅 × = {𝑥 ∈ 𝑅 : 𝑣(𝑥) = 0} ⊆ 𝑅 is closed.
Example 13.5.7. For 𝑅 = Z 𝑝 ⊆ 𝐹 = Q 𝑝 and 𝐵 = M𝑛 (Q 𝑝 ), the order O = M𝑛 (Z 𝑝 ) is

compact (neighborhoods of a matrix can be taken as neighbhoods in each coordinate)
and the subgroup O× = GL𝑛 (Z 𝑝 ) is compact: there is no way to “run off to infinity”,
either in a single coordinate or via the determinant.
13.5.8. Suppose 𝐵 = 𝐷 is a division algebra. Then the valuation ring O is the maximal
compact subring of 𝐵, for the same reason as in the commutative case (see 13.5.5,
details requested in Exercise 13.18(a)). There is a filtration
O ⊃ 𝑃 ⊃ 𝑃2 ⊃ . . .
giving rise to a filtration
O× ⊃ 1 + 𝑃 ⊃ 1 + 𝑃 2 ⊃ . . . . (13.5.9)
As in the second proof of Main Theorem 12.3.2, the quotient O/𝑃 is a finite extension
of the finite residue field 𝑘, so (O/𝑃) × is a finite cyclic group. The maximal two-sided
ideal 𝑃 is principal, generated by an element 𝑗 of minimal valuation, and multiplication
by 𝑗 𝑛 gives an isomorphism O/𝑃 −→ ∼ 𝑃 𝑛 /𝑃 𝑛+1 of 𝑘-vector spaces (or abelian groups)
for all 𝑛 ≥ 1.
Furthermore, for each 𝑛 ≥ 1, there is an isomorphism of groups
∼ (1 + 𝑃 𝑛 )/(1 + 𝑃 𝑛+1 )
O/𝑃 ' 𝑃 𝑛 /𝑃 𝑛+1 −→
(13.5.10)
𝛼 ↦→ 1 + 𝛼.
Therefore, O× = lim (O/𝑃 𝑛 ) × is a projective limit of solvable groups, also called a

←−−𝑛
prosolvable group.
Example 13.5.11. If 𝐵 is a division quaternion algebra over Q 𝑝 , with valuation
ring O and maximal ideal 𝑃, then the filtration (13.5.9) has quotients isomorphic to
O/𝑃 ' F 𝑝2 .
13.5.12. We will also want to consider norm 1 groups; for this, we suppose that 𝐵 is
a semisimple algebra. Let
𝐵1 := {𝛼 ∈ 𝐵 : nrd(𝛼) = 1};
some authors also write SL1 (𝐵) := 𝐵1 . Then 𝐵1 is a closed subgroup of 𝐵× , since the
reduced norm is continuous.
If 𝐵 is a division ring and 𝐹 is archimedean, then 𝐵 ' H and 𝐵1 ' H1 ' SU(2) is
compact (it is identified with the 3-sphere in R4 ). In a similar way, if 𝐵 is a divison ring
and 𝐹 is nonarchimedean, then 𝐵1 is compact: for 𝐵 has a valuation 𝑣 and valuation
ring O, and if 𝛼 ∈ 𝐵 has nrd(𝛼) = 1 then 𝑣(𝛼) = 0 and 𝛼 ∈ O, and consequently
𝐵1 ⊆ O× is closed in a compact set so compact.
If 𝐵 is not a division ring, then either 𝐵 is the product of two algebras or 𝐵 is a
matrix ring over a division ring, and in either case 𝐵1 is not compact.
Remark 13.5.13. The locally compact division algebras over a nonarchimedean field
are necessarily totally disconnected. On the other hand, it is a theorem of Pontryagin
[War89, Theorem 27.2] that if 𝐴 is a connected locally compact division ring, then 𝐴
is isomorphic as a topological ring to either R, C, or H.
Exercises

−1, −1
1. Let 𝐵 := .
Q2
(a) Show that 𝐵 is a division ring that is complete with respect to the discrete
valuation 𝑤 defined by 𝑤(𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗) = 𝑣(𝑡 2 + 𝑥 2 + 𝑦 2 + 𝑧2 ) for
𝑡, 𝑥, 𝑦, 𝑧 ∈ Q2 .
13.5. SOME TOPOLOGY 207
(b) Prove that
1 + 𝑖 + 𝑗 + 𝑖𝑗
O := Z2 ⊕ Z2𝑖 ⊕ Z2 𝑗 ⊕ Z2 ⊂𝐵
2
is the valuation ring of 𝐵.
2. Let 𝐵 be a division quaternion algebra over a nonarchimedean local field 𝐹.
Give another proof that the unramified quadratic extension 𝐾 of 𝐹 embeds in 𝐵
as follows.
Suppose it does not: then for all 𝛼 ∈ O, the extension 𝐹 (𝛼) ⊇ 𝐹 is ramified,
so there exists 𝑎 ∈ 𝑅 such that 𝛼 − 𝑎 ∈ 𝑃 ∩ 𝐾 (𝛼); let 𝑃 = 𝑗O and write
𝛼 = 𝛼0 = 𝑎 + 𝑗 𝛼1 , and iterate to conclude that 𝛼 = ∞
Í
𝑛=0 𝑎 𝑛 𝑗 with 𝑎 𝑛 ∈ 𝑅. But
𝑛
𝐹 ( 𝑗) is complete so O ⊆ 𝐹 ( 𝑗), a contradiction.

3. Let 𝐹 be a local field with 𝐹 ; C, let 𝐾 the unramified (separable) quadratic
extension of 𝐹 (take 𝐾 = C if 𝐹 ' R), and let h𝜎i = Gal(𝐾 | 𝐹), so that 𝜎 is the
standard involution on 𝐾. Let 𝐵 be a division quaternion algebra 𝐵 over 𝐹.
Show that
𝑎 𝑏
𝐵' : 𝑎, 𝑏 ∈ 𝐾 ⊆ M2 (𝐾).
𝜋𝜎(𝑏) 𝜎(𝑎)
[Hint: Compute the regular representation 2.3.8.] Identify the maximal order
O its maximal ideal 𝐽 under this identification.
4. Let 𝐵 be a division quaternion algebra over 𝐹. Show that 𝛼 ∈ 𝐵 is integral over
𝑅 if and only if nrd(𝛼), nrd(𝛼 + 1) ∈ 𝑅 if and only if 𝑤(𝛼), 𝑤(𝛼 + 1) ≥ 0, where
𝑤 is the valuation on 𝐵.
5. Extend Theorem 13.3.11 as follows. Let 𝑅 be a complete DVR with field of
fractions
𝐹, and let 𝐵 be a quaternion division algebra over 𝐹. Show that
𝐾, 𝑏
𝐵' where 𝐾 ⊇ 𝐹 is an unramified separable quadratic extension of 𝐹
𝐹
and 𝑏 ∉ Nm𝐾 |𝐹 (𝐾 × ).
6. Let 𝐹 be a nonarchimedean local field with residue field 𝑘 having char 𝑘 ≠ 2,
and let 𝐾 ⊇ 𝐹 be a separable quadratic field extension.
(a) Let 𝑏 ∈ 𝐹 × . Show that if 𝐾 is unramified
√ then 𝑏 ∈ Nm𝐾 |𝐹 (𝐾 × ) if and
only if 𝑣(𝑏) is even; and if 𝐾 = 𝐹 ( 𝑎) is ramified, then 𝑏 ∈ Nm𝐾 |𝐹 (𝐾 × )
if and only if 𝑏 = 𝑐2 or 𝑏 = −𝑎𝑐2 for some 𝑐 ∈ 𝐹 ×2 .
(b) Show that [𝐹 × : Nm𝐾 |𝐹 (𝐾 × )] = 2 and deduce Corollary 13.4.6 in this
case.
7. Let 𝑅 := Q[[𝑡]] be the ring of formal power series over Q; then 𝑅 is a complete
DVR with fraction field 𝐹 = Q((𝑡)), the Laurent series over Q. Let 𝐵0 := (𝑎, 𝑏 |
Q) be a division quaternion algebra over Q, and let 𝐵 := 𝐵0 ⊗Q 𝐹 = (𝑎, 𝑏 | 𝐹).
Show that 𝐵 is a division quaternion algebra over 𝐹, with valuation ring O :=
𝑅 + 𝑅𝑖 + 𝑅 𝑗 + 𝑅𝑖 𝑗.
8. Let 𝐵 be a division quaternion algebra over a nonarchimedean local field 𝐹, and
let O be the valuation ring.
(a) Show that every one-sided (left or right) ideal of O is a power of the
maximal ideal 𝑃 and hence is two-sided.
(b) Let
[O, O] := h𝛼𝛽 − 𝛽𝛼 : 𝛼, 𝛽 ∈ Oi
be the commutator ideal [O, O] of O, the two-sided ideal generated by
commutators of elements of O. Show that 𝑃 = [O, O].
9. Let 𝐹 be a nonarchimedean local field, let 𝐵 = M2 (𝐹) and O = M2 (𝑅). Show
that there are 𝑞 + 1 right O-ideals of norm 𝔭 corresponding to the elements of
P1 (𝑘) or equivalently the lines in 𝑘 2 .
10. Give another proof of Lemma 13.4.9 using quadratic forms.
11. Let 𝐹 be a nonarchimedean local field with valuation ring 𝑅, maximal ideal 𝔭,
and residue field 𝑘. Let 𝐷 be a division algebra over 𝐹 with dim𝐹 𝐷 = 𝑛2 ,
with valuation ring O and maximal two-sided ideal 𝑃. Show that O/𝑃 is finite
extension of 𝑘 of degree 𝑛, and 𝑃 𝑛 = 𝔭O (cf. Remark 13.4.3).
12. Show that (13.5.10) is an isomorphism of (abelian) groups.
13. Let 𝐹 be a field with absolute value | | and 𝑉 a finite-dimensional 𝐹-vector space.
(a) Let 𝑥 1 , . . . , 𝑥 𝑛 be a basis for 𝑉, and define
k𝑎 1 𝑥1 + · · · + 𝑎 𝑛 𝑥 𝑛 k := max(|𝑎 1 |, . . . , |𝑎 𝑛 |)
for 𝑎 𝑖 ∈ 𝐹. Show that 𝑉 is a metric space with distance 𝑑 (𝑥, 𝑦) = k𝑥 − 𝑦k.
(b) Show that the topology on 𝑉 is independent of the choice of basis in (a).
(c) Finally, show that if 𝐹 is complete with respect to | |, then 𝑉 is also
complete.
14. Let 𝐹 be a topological field. Show that the coarsest topology (fewest open sets)
in which multiplication on M𝑛 (𝐹) is continuous is the coordinate topology.
15. Let 𝐹 be a local field.
(a) The group 𝐹 × has the structure of topological group under the embedding
𝑥 ↦→ (𝑥, 𝑥 −1 ) ∈ 𝐹 × 𝐹 (under the subspace topology in 𝐹 × 𝐹). Show that
this topology coincides with the subspace topology 𝐹 × ⊆ 𝐹.
(b) More generally, let 𝐵 be a finite-dimensional 𝐹-algebra. Show that the
the topology on 𝐵× induced by 𝛼 ↦→ (𝛼, 𝛼−1 ) ∈ 𝐵 × 𝐵 coincides with the
subspace topology 𝐵× ⊆ 𝐵.
(c) Show that an open subset of a Hausdorff, locally compact space is locally
compact in the subspace topology.
16. Let 𝐹 be a finite extension of Q2 . Show that (−1, −1) 𝐹 = (−1) [𝐹 :Q2 ] .
17. Finish the proof of Lemma 13.2.9.
18. Let 𝐷 be a division algebra over a nonarchimedean local field 𝐹. We recall (see
13.5.2) that 𝐷 is a complete, locally compact topological ring.
(a) Verify (as in 13.5.5) that O is the maximal compact subring of 𝐵.
(b) Show that 𝐵× /𝐹 × is a compact topological group.
Chapter 14
Quaternion algebras over global fields
In this chapter, we discuss quaternion algebras over global fields and characterize them
up to isomorphism.
14.1 ⊲ Ramification
To motivate the classification of quaternion algebras over Q, we consider by analogy a
classification of quadratic fields. We restrict to the following class of quadratic fields
for the best analogy.
√
Definition 14.1.1. A quadratic field 𝐹 = Q( 𝑑) of discriminant 𝑑 ∈ Z is mildly
ramified if 8 - 𝑑.
√
A quadratic field 𝐹 is mildly ramified if and only if 𝐹 = Q( 𝑚) where 𝑚 ≠ 1 is
√ then 𝑑 = 𝑚 or 𝑑 = 4𝑚 according as 𝑚 = 1, 3 (mod 4).
odd and squarefree;
Let 𝐹 = Q( 𝑑) be a mildly ramified quadratic field of discriminant 𝑑 ∈ Z and
let 𝑅 be its ring of integers. A prime 𝑝 ramifies in 𝐹, i.e. 𝑝𝑅 = 𝔭2 for a prime ideal
𝔭 ⊂ 𝑅, if and only if 𝑝 | 𝑑.
But a discriminant 𝑑 can be either positive or negative; to put this bit of data on
the same footing, we define the set of places of Q to be the primes together with the
symbol ∞, and we make √ the convention that ∞ ramifies in 𝐹 if 𝑑 < 0 and is unramified
if 𝑑 > 0. Let 𝐹 = Q( 𝑑) be a mildly ramified quadratic field, and let Ram(𝐹) be the
set of places that ramify in 𝐹. The set Ram(𝐹) determines 𝐹 up to isomorphism, since
the discriminant of 𝐹 is the product of the odd primes in Ram(𝐹), multiplied by 4 if
2 ∈ Ram(𝐹) and by −1 if ∞ ∈ Ram(𝐹). (For bookkeeping reasons, in this context it
would probably therefore be better to consider 4 and −1 as primes, but we will resist
the inducement here.) However, not every finite set of places Σ occurs: the product 𝑑
corresponding to Σ is a discriminant if and only if 𝑑 ≡ 0, 1 (mod 4). We call this a
parity condition on the set of ramifying places of a mildly ramified quadratic field:
2∈Σ ↔ there are an odd number of places in Σ congruent to −1 (mod 4)
with the convention that ∞ is congruent to −1 (mod 4).
209
210 CHAPTER 14. QUATERNION ALGEBRAS OVER GLOBAL FIELDS
Note that if Σ is a finite subset of places of Q and 2 ∉ Σ, then precisely one of

either Σ or Σ ∪ {∞} satisfies the parity condition; accordingly, if we define 𝑚(Σ) to be
the product of all odd primes in Σ multiplied by −1 if ∞ ∈ Σ, then we can recover Σ
from 𝑚(Σ).
We have proven the following result.
Lemma 14.1.2. The maps 𝐹 ↦→ Ram(𝐹) and Σ ↦→ 𝑚(Σ) furnishes a bĳection

Mildly
√ ramified quadratic fields ↔ Finite subsets of places of Q
Q( 𝑑) up to isomorphism satisfying the parity condition
Squarefree odd integers
n o
↔ .
𝑚≠1
This classification procedure using sets of ramifying primes and discriminants

works as well for quaternion algebras over Q. Let 𝐵 be a quaternion algebra over
Q. When is a prime 𝑝 ramified in 𝐵? In Chapter 12, we saw that the completion
𝐵 𝑝 = 𝐵 ⊗Q Q 𝑝 is either a division ring or the matrix ring M2 (Q 𝑝 ). Further, when 𝐵 𝑝
is a division ring, the valuation ring O 𝑝 ⊂ 𝐵 𝑝 is the unique maximal order, and the
unique maximal ideal 𝑃 𝑝 ⊂ O 𝑝 satisfies 𝑝O 𝑝 = 𝑃2𝑝 . By analogy with the quadratic
case, we say that a place 𝑣 is ramified in 𝐵 if the completion 𝐵 𝑣 is a division ring, and
otherwise 𝑣 is unramified
(or split).
𝑎, 𝑏
Let 𝐵 = . Without loss of generality, we may suppose 𝑎, 𝑏 ∈ Z. There are
Q
only finitely many places where 𝐵 is ramified: by the calculation of the Hilbert symbol
(12.4.12), if 𝑝 is prime and 𝑝 - 2𝑎𝑏, then (𝑎, 𝑏)Q 𝑝 = 1 and 𝑝 is split in 𝐵. Therefore
# Ram 𝐵 < ∞.
We say that 𝐵 is definite if ∞ ∈ Ram 𝐵 and
𝐵 is indefinite otherwise. By definition,
𝑎, 𝑏
𝐵 is definite if and only if 𝐵∞ := 𝐵 ⊗Q R = ' H if and only if 𝑎, 𝑏 < 0 (Exercise
R
2.4).
Let Ram 𝐵 be the set of ramified places of 𝐵. Not every finite subset Σ of places
can occur as Ram 𝐵 for a quaternion algebra 𝐵. It turns out that the parity condition
here is that we must have #Σ even. So again, if Σ is a finite set of primes, then precisely
one of either Σ or Σ ∪ {∞} can occur as Ram 𝐵. We define the discriminant of 𝐵
to be the product disc 𝐵 of primes that ramify in 𝐵, so disc 𝐵 is a squarefree positive
integer. This notion of discriminant is admittedly strange; we relate it to perhaps more
familiar notions in Chapter 15.
The main result of this chapter, specialized to the case 𝐹 = Q, is the following.
Î
Main Theorem 14.1.3. The maps 𝐵 ↦→ Ram 𝐵 and Σ ↦→ 𝑝 ∈Σ 𝑝 furnish bĳections

Quaternion algebras over Q Finite subsets of places of Q
↔
up to isomorphism of even cardinality
↔ { 𝐷 ∈ Z>0 squarefree } .
Î
The composition of these maps is 𝐵 ↦→ 𝑝 ∈Ram 𝐵 𝑝 = disc 𝐵.
14.2. ⊲ HILBERT RECIPROCITY OVER THE RATIONALS 211
As previewed at the end of section 12.1, Main Theorem 14.1.3 is a local-global

principle and provides a convenient way to test when quaternion algebras over Q are
isomorphic: instead of working hard over Q, we can just test for isomorphism over the
local fields Q 𝑝 and R.
We will spend the next two sections giving a self-contained proof of Main Theo-
rem 14.1.3 following Serre [Ser73, Chapters III–IV], assuming two statements from
basic number theory (quadratic reciprocity and the existence of primes in arithmetic
progression), finishing the proof in section 14.3. Although the proofs presented do
not seem to generalize beyond 𝐹 = Q, the argument is simple enough and its structure
is good motivation for the more involved treatment in the Chapter ahead. (It is also
comforting to see a complete proof in the simplest case.)
14.2 ⊲ Hilbert reciprocity over the rationals

To begin, we look into the parity condition: it has a simple reformulation in terms of
the Hilbert symbol (section 12.4). For a place 𝑣 of Q, let Q 𝑣 denote the completion
of Q at the absolute value associated to 𝑣: if 𝑣 = 𝑝 is prime, then Q 𝑣 = Q 𝑝 is the
field of 𝑝-adic numbers; if 𝑣 = ∞, then Q 𝑣 = R. For 𝑎, 𝑏 ∈ Q× , we abbreviate
(𝑎, 𝑏)Q𝑣 = (𝑎, 𝑏) 𝑣 .
Proposition 14.2.1 (Hilbert reciprocity). For all 𝑎, 𝑏 ∈ Q× , we have

Ö
(𝑎, 𝑏) 𝑣 = 1, (14.2.2)
𝑣
where the product taken over all places 𝑣 of Q.
When 𝑝 is odd and divides neither numerator nor denominator of 𝑎 or 𝑏, we have

(𝑎, 𝑏) 𝑝 = 1, so the product (14.2.2) is well-defined. The following corollary is an
equivalent statement.
Corollary 14.2.3. Let 𝐵 be a quaternion algebra over Q. Then the set Ram 𝐵 is finite
of even cardinality.
The law of Hilbert reciprocity, as it turns out, is a core premise in number theory:
it is equivalent to the law of quadratic reciprocity

𝑝 𝑞 𝑝−1 𝑞−1
= (−1) 2 2 (14.2.4)
𝑞 𝑝
for odd primes 𝑝, 𝑞 together with the supplement

−1 𝑝−1 2 𝑝 2 −1
= (−1) 2 and = (−1) 8 (14.2.5)
𝑝 𝑝
for odd primes 𝑝.

We now give a proof of Hilbert reciprocity (Proposition 14.2.1), assuming the law
of quadratic reciprocity and its supplement.
Proof of Proposition 14.2.1. Since each local Hilbert symbol is bilinear, it suffices to
prove the statement when 𝑎, 𝑏 ∈ Z are equal to either −1 or a prime number. The
Hilbert symbol is also symmetric, sowe may interchange
𝑎, 𝑏.
𝑎, 𝑏 −1, −1
If 𝑎 = 𝑏 = −1, then 𝐵 = = is the rational Hamiltonians, and
Q Q
(−1, −1)∞ = (−1, −1)2 = −1 and (−1, −1) 𝑣 = 1 if 𝑣 ≠ 2, ∞, by the computation of
the even Hilbert symbol (12.4.13). Similarly, the cases with 𝑎 = −1, 2 follow from the
supplement (14.2.5), and are requested in Exercise 14.1.
𝑝, 𝑝 −1, 𝑝
So we may suppose 𝑎 = 𝑝 and 𝑏 = 𝑞 are primes. If 𝑝 = 𝑞 then '
Q Q
and we reduce to the previous case, so we may suppose 𝑝 ≠ 𝑞. Since 𝑝, 𝑞 > 0, we
have ( 𝑝, 𝑞)∞ = 1. We have ( 𝑝, 𝑞)ℓ = 1 for all primes ℓ - 2𝑝𝑞, and

𝑞 𝑝
( 𝑝, 𝑞) 𝑝 = (𝑞, 𝑝) 𝑝 = and ( 𝑝, 𝑞)𝑞 =
𝑝 𝑞
by 12.4.12. Finally,
( 𝑝, 𝑞)2 = −1 if and only if 𝑝, 𝑞 ≡ 3 (mod 4)
i.e., ( 𝑝, 𝑞)2 = (−1) ( 𝑝−1) (𝑞−1)/4 , again by the computation of the even Hilbert symbol
(12.4.13). Thus the product becomes

( 𝑝−1) (𝑞−1)/4 𝑝 𝑞
Ö
( 𝑝, 𝑞) 𝑣 = (−1) =1
𝑣
𝑞 𝑝
by quadratic reciprocity.
Hilbert reciprocity has several aesthetic advantages over the law of quadratic reci-
procity. For one, it is simpler to write down! Also, Hilbert believed that his reciprocity
law is a kind of analogue of Cauchy’s integral theorem, expressing an integral as a
sum of residues (Remark 14.6.4). The fact that a normalized product over all places is
trivial also arises quite naturally: if we define for 𝑥 ∈ Q× and a prime 𝑝 the normalized
absolute value
|𝑥| 𝑝 := 𝑝 −𝑣𝑝 ( 𝑥) ,
and |𝑥| ∞ the usual archimedean absolute value, then
Ö
|𝑥| 𝑣 = 1
𝑣
by unique factorization in Z; this is called the product formula for Q, for obvious
reasons.
From the tight relationship between quaternion algebras and ternary quadratic
forms, we obtain the following corollary.
Corollary 14.2.6. Let 𝑄 be a nondegenerate ternary quadratic form over Q. Then the
set of places 𝑣 such that 𝑄 𝑣 is anisotropic is finite and of even cardinality.
14.2. ⊲ HILBERT RECIPROCITY OVER THE RATIONALS 213
In particular, by Corollary 14.2.6, if 𝑄 𝑣 is isotropic for all but one place 𝑣 of Q,

then 𝑄 𝑣 is in fact isotropic for all places 𝑣.
Proof. In the bĳection implied by Main Theorem 5.2.5, the quadratic form 𝑄 cor-
responds to a quaternion algebra 𝐵 = (𝑎, 𝑏 | Q), and by Main Theorem 5.4.4, 𝑄 is
isotropic if and only if 𝐵 is split if and only if (𝑎, 𝑏)Q = 1. By functoriality, the same
is true over each completion Q 𝑣 for 𝑣 a place of Q, and therefore the set of places
𝑣 where 𝑄 𝑣 is isotropic is precisely the set of ramified places in 𝐵. The result then
follows by Hilbert reciprocity.
To conclude this section, we show that every allowable product of Hilbert symbols
is obtained.
Proposition 14.2.7. Let Σ be a finite set of places of Q of even cardinality. Then there
exists a quaternion algebra 𝐵 over Q with Ram 𝐵 = Σ.
Remark 14.2.8. Albert [Alb34, Theorem 2, Theorem 3] already sought to simplify
the presentation of a quaternion algebra by a series of transformations, the content
of which is contained in Proposition 14.2.7; this was further investigated by Latimer
[Lat35].
Just as with Hilbert reciprocity, Proposition 14.2.7 touches on a deep statement in
number theory concerning primes, due to Dirichlet.
Theorem 14.2.9 (Infinitude of primes in arithmetic progression). Given 𝑎, 𝑛 ∈ Z
coprime, there are infinitely many primes 𝑝 ≡ 𝑎 (mod 𝑛).
Proof. See e.g. Serre [Ser73, Chapter VI] or Apostol [Apo76, Chapter 7]. We will
prove this theorem in Exercise 26.11 as a consequence of the analytic class number
formula.
Remark 14.2.10. Theorem 14.2.9 seems to require analysis. (For algebraic proofs in
special cases, see e.g., Neukirch [Neu99, Exercise I.10.1] and Lenstra–Stevenhagen
[LS91].) Ram Murty [Mur88] showed that a “Euclidean proof” of the infinitude of
primes 𝑝 ≡ 𝑎 (mod 𝑛) is possible if and only if 𝑎 2 ≡ 1 (mod 𝑛), and Paul Pollack
[Pol2010] has shown that Schnizel’s Hypothesis H gives a heuristic for this. This
crucial role played by analytic methods motivates part III of this monograph.
We now prove Proposition 14.2.7 assuming Theorem 14.2.9.
Î
Proof. Let 𝐷 := 𝑝 ∈Σ 𝑝 be the product of the primes in Σ, and let 𝑢 := −1 if ∞ ∈ Σ
and 𝑢 := 1 otherwise. Let 𝐷 ♦ := 𝑢𝐷. We consider quaternion algebras of the form
♦ ♦
𝑞 ,𝐷
𝐵=
Q
with 𝑞 ♦ = 𝑢𝑞 (and 𝑞 prime) chosen to satisfy certain congruence conditions ensuring

that Ram 𝐵 = Σ. To this end, we seek a prime 𝑞 such that
♦
𝑞
= −1 for all odd 𝑝 | 𝐷 (14.2.11)
𝑝
and (
1 (mod 8), if 2 - 𝐷;
𝑞♦ ≡ (14.2.12)
5 (mod 8), if 2 | 𝐷.
There exists a prime satisfying the conditions (14.2.11)–(14.2.12) by Theorem 14.2.9,
since the condition to be a quadratic nonresidue is a congruence condition on 𝑞 ♦ and
hence on 𝑞 modulo 𝑝.
We now verify that 𝐵 has Ram 𝐵 = Σ. We have (𝑞 ♦ , 𝐷 ♦ )∞ = 𝑢 by choice of signs
and (𝑞 ♦ , 𝐷 ♦ ) 𝑝 = 1 for all 𝑝 - 2𝑑𝑞. We compute that
♦
𝑞
(𝑞 ♦ , 𝐷 ♦ ) 𝑝 = = −1 for all odd 𝑝 | 𝐷
𝑝
by (14.2.11). For 𝑝 = 2, we find that (𝑞 ♦ , 𝐷 ♦ )2 = −1 or (𝑞 ♦ , 𝐷 ♦ )2 = 1 according as
2 | 𝐷 or not by the computation of the even Hilbert symbol (12.4.13). This shows that
Σ ⊆ Ram 𝐵 ⊆ Σ ∪ {𝑞}.
The final symbol (𝑞 ♦ , 𝐷 ♦ )𝑞 is determined by Hilbert reciprocity (Proposition 14.2.1):
since #Σ is already
♦ ♦even, we must have (𝑞 ♦ , 𝐷 ♦ )𝑞 = 1. Therefore the quaternion
𝑞 ,𝐷
algebra 𝐵 := has Σ = Ram 𝐵.
Q
Example 14.2.13. Let 𝐵 = (𝑎, 𝑏 | Q) be a quaternion algebra of prime discriminant
𝐷 = 𝑝 over Q. Then:
(i) For 𝐷 = 𝑝 = 2, we take 𝑎 = 𝑏 = −1;
(ii) For 𝐷 = 𝑝 ≡ 3 (mod 4), we take 𝑏 = −𝑝 and 𝑎 = −1;
(iii) For 𝐷 = 𝑝 ≡ 1 (mod 4), we take 𝑏 = −𝑝 and 𝑎 = −𝑞 where 𝑞 ≡ 3 (mod 4) is
𝑞
prime and = −1.
𝑝

−𝑝 𝑞
In case (iii), by qudaratic reciprocity =− = 1 so indeed 𝐵 is not ramified at
𝑞 𝑝
𝑝. In the proof of Theorem 14.2.7 above, we would have required the more restrictive
condition 𝑞 ≡ 3 (mod 8), but we can look again at the table of even Hilbert symbols
(12.4.16): since 𝑏 = −𝑝 = −1, 3 (mod 8), we may take 𝑎 = −𝑞 = 1, −3 (mod 8) freely,
so 𝑞 ≡ 3 (mod 4).
Similarly, for discriminant 𝐷 the product of two (distinct) primes:
(i) For 𝐷 = 2𝑝 with 𝑝 ≡ 3 (mod 4), we take 𝑎 = −1 and 𝑏 = 𝑝;
(ii) For 𝐷 = 2𝑝 with 𝑝 ≡ 5 (mod 8), we take 𝑎 = 2 and 𝑏 = 𝑝;
(iii) For 𝐷 = 𝑝𝑞 with 𝑝 ≡ 𝑞 ≡ 3 (mod 4), we take 𝑎 = −1 and
𝑏 = 𝑝𝑞;
𝑞
(iv) For 𝐷 = 𝑝𝑞 with 𝑝 ≡ 1 (mod 4) or 𝑞 ≡ 1 (mod 4) and ≠ 1, we take 𝑎 = 𝑝
𝑝
and 𝑏 = 𝑞.
For other explicit presentations of quaternion algebras over Q with specified dis-
criminant, see Alsina–Bayer [AB2004, §1.1.2]. See Example 15.5.7 for some explicit
maximal orders.
14.3. ⊲ HASSE–MINKOWSKI THEOREM OVER THE RATIONALS 215
14.3 ⊲ Hasse–Minkowski theorem over the rationals

To complete the proof of Main Theorem 14.1.3, we now show that the map 𝐵 ↦→ Ram 𝐵
is injective on isomorphism classes.
Proposition 14.3.1. Let 𝐵, 𝐵 0 be quaternion algebras over Q. Then the following are
equivalent:
(i) 𝐵 ' 𝐵 0;
(ii) Ram 𝐵 = Ram 𝐵 0;
(iii) 𝐵 𝑣 ' 𝐵 0𝑣 for all places 𝑣 ∈ Pl(Q); and
(iv) 𝐵 𝑣 ' 𝐵 0𝑣 for all but one place 𝑣.
The statement of Proposition 14.3.1 is a local-global principle: the global isomor-

phism class is determined by the local isomorphism classes.
Corollary 14.3.2. Let 𝐵 be a quaternion algebra over Q. Then 𝐵 ' M2 (Q) if and
only if 𝐵 𝑝 ' M2 (Q 𝑝 ) for all primes 𝑝.
Proof. Apply Proposition 14.3.1 (i) ⇔ (iv) with 𝐵 0 = M2 (Q) checking all but the
archimedean place.
By the equivalence between quaternion algebras and quadratic forms (see Chapter
5, specifically section 5.2), the statement of Proposition 14.3.1 is equivalent to the
statement that a ternary quadratic form over Q is isotropic if and only if it is isotropic
over all (but one) completions. In fact, the more general statement is true—and again
we come in contact with a deep result in number theory.
Theorem 14.3.3 (Hasse–Minkowski). Let 𝑄 be a quadratic form over Q. Then 𝑄 is

isotropic if and only if 𝑄 𝑣 is isotropic for all places 𝑣 of Q.
We will prove the Hasse–Minkowski theorem by induction on the number of

variables. Of particular interest is the case of (nondegenerate) ternary quadratic forms,
for which we have the following theorem of Legendre.
Theorem 14.3.4 (Legendre). Let 𝑎, 𝑏, 𝑐 ∈ Z be nonzero, squarefree integers that are

relatively prime in pairs. Then the quadratic form
𝑎𝑥 2 + 𝑏𝑦 2 + 𝑐𝑧 2 = 0
has a nontrivial solution 𝑥, 𝑦, 𝑧 ∈ Q if and only if 𝑎, 𝑏, 𝑐 do not all have the same sign
and
−𝑎𝑏, −𝑏𝑐, −𝑎𝑐 are quadratic residues modulo |𝑐|, |𝑎|, | 𝑏|, respectively.
Proof. First, the conditions for solvability are necessary. The condition on signs is
necessary for a solution in R. If 𝑎𝑥 2 + 𝑏𝑦 2 + 𝑐𝑧2 = 0 with 𝑥, 𝑦, 𝑧 ∈ Q not all zero, then
scaling we may suppose 𝑥, 𝑦, 𝑧 ∈ Z satisfy gcd(𝑥, 𝑦, 𝑧) = 1; if 𝑝 | 𝑐 then 𝑝 - 𝑦 (else
𝑝 | 𝑥 and 𝑝 | 𝑧, contradiction), so (𝑥/𝑦) 2 ≡ (−𝑏/𝑎) (mod |𝑐|) and −𝑏𝑎 is a quadratic
residue modulo |𝑐|; the other conditions hold by symmetry.
So suppose the conditions hold. Multiplying through and rescaling by squares

(Exercise 14.8), we may suppose 𝑎, 𝑏 are squarefree (but not necessarily coprime) and
𝑐 = −1, and we seek a nontrivial solution to 𝑎𝑥 2 + 𝑏𝑦 2 = 𝑧 2 . If 𝑎 ∈ Q×2 , then we are
done. Otherwise, we need to solve
√
𝑧 2 − 𝑎𝑥 2

√
𝑧+𝑥 𝑎
= 𝑏 = Nm
𝑦2 Q( 𝑎)/Q
𝑦
√
for 𝑥, 𝑦, 𝑧 ∈ Q and 𝑦 ≠ 0, i.e., we need to show that 𝑏 is a norm from 𝐹 = Q( 𝑎). By
hypothesis, 𝑎, 𝑏 are not both negative and
𝑏 is a square modulo |𝑎| and 𝑎 is a square modulo | 𝑏|. (14.3.5)
We may also suppose |𝑎| ≤ | 𝑏|.

We use complete induction on 𝑚 = |𝑎| + | 𝑏|. If 𝑚 = 2, then we must consider the
equation ±𝑥 2 ± 𝑦 2 = 𝑧2 with the case both negative signs excluded, each of which has
solutions. Now suppose that 𝑚 > 2 so | 𝑏| ≥ 2, and let 𝑝 | 𝑏 be prime divisor. By
hypothesis, there exist integers 𝑡, 𝑏 0 such that 𝑡 2 = 𝑎 + 𝑏𝑏 0; taking a small residue, we
may suppose |𝑡| < | 𝑏|/2. Thus
√
𝑏𝑏 0 = 𝑡 2 − 𝑎 = Nm𝐹 /Q (𝑡 + 𝑎)
so 𝑏𝑏 0 is a norm from 𝐹. Thus 𝑏 is a norm if and only if 𝑏 0 is a norm. But

2
𝑡 − 𝑎 | 𝑏|
| 𝑏 0 | = ≤ + 1 < | 𝑏|
𝑏 4
because | 𝑏| ≥ 2.
Now write 𝑏 0 = 𝑏 00𝑢 2 with 𝑏 00, 𝑢 ∈ Z and 𝑏 00 squarefree. Then | 𝑏 00 | ≤ | 𝑏 0 | < | 𝑏|
and 𝑏 00 is a norm if and only if 𝑏 0 is a norm. With these manipulations, we propagate
the hypothesis that |𝑎| is a square modulo | 𝑏 00 | and | 𝑏 00 | is a square modulo |𝑎|.
Therefore, the induction hypothesis applies to the equation 𝑎𝑥 2 + 𝑏 00 𝑦 2 = 𝑧2 , and the
proof is complete.
Corollary 14.3.6. Let 𝑄 be a nondegenerate ternary quadratic form over Q. Then 𝑄

is isotropic if and only if 𝑄 𝑣 is isotropic for all places 𝑣 of Q (but one).
Proof. If 𝑄 is isotropic, then 𝑄 𝑣 is isotropic for all 𝑣. For the converse, suppose that
𝑄 𝑣 is isotropic for all places 𝑣 of Q. As in the proof of Legendre’s Theorem 14.3.4,
we may suppose 𝑄(𝑥, 𝑦, 𝑧) = 𝑎𝑥 2 + 𝑏𝑦 2 − 𝑧2 . The fact that 𝑄 is isotropic over R
implies that 𝑎, 𝑏 are not both negative. Now let 𝑝 | 𝑎 be odd. The condition that 𝑄 𝑝
is isotropic is equivalent to (𝑎, 𝑏) 𝑝 = (𝑏/𝑝) = 1; putting these together, we conclude
that 𝑏 is a quadratic residue modulo |𝑎|. The same holds for 𝑎, 𝑏 interchanged, so
(14.3.5) holds and the result follows.
We are now in a position to complete the proof of the Hasse–Minkowski theorem.

14.3. ⊲ HASSE–MINKOWSKI THEOREM OVER THE RATIONALS 217
Proof of Theorem 14.3.3. We follow Serre [Ser73, Theorem 8, §IV.3.2]. We may

suppose that 𝑄 is nondegenerate in 𝑛 ≥ 1 variables. If 𝑛 = 1, the statement is vacuous.
If 𝑛 = 2, the after scaling we may suppose 𝑄(𝑥, 𝑦) = 𝑥 2 − 𝑎𝑦 2 with 𝑎 ∈ Q× ; since
𝑄 𝑝 is isotropic for all primes 𝑝, we have 𝑎 ∈ Q×2 𝑝 so in particular 𝑣 𝑝 (𝑎) is even for
all primes 𝑝; since 𝑄 is isotropic at ∞, we have 𝑎 > 0; thus by unique factorization
𝑎 ∈ Q×2 , and the result follows. If 𝑛 = 3, the statement is proven in Corollary 14.3.6.
Now suppose 𝑛 ≥ 4. Write 𝑄 = h𝑎, 𝑏i −𝑄 0 where 𝑄 0 = h𝑐 1 , . . . , 𝑐 𝑛−2 i and
𝑎, 𝑏, 𝑐 𝑖 ∈ Z. Let 𝑑 = 2𝑎𝑏(𝑐 1 · · · 𝑐 𝑛−2 ) ≠ 0. For each prime 𝑝 | 𝑑, since 𝑄 is isotropic,
there exists 𝑡 𝑝 ∈ Q×𝑝 represented by both h𝑎, 𝑏i and 𝑄 0 in Q 𝑝 . (This requires a small
argument, see Exercise 6.14.) Similarly, there exists 𝑡∞ ∈ R× represented by these
forms in R.
By another application of the infinitude of primes in arithmetic progression (Exer-
cise 14.10), there exists 𝑡 ∈ Q× such that:
(i) 𝑡 ∈ 𝑡 𝑝 Q×2
𝑝 for all primes 𝑝 | 𝑑,
(ii) 𝑡 and 𝑡 ∞ have the same sign, and
(iii) 𝑝 - 𝑡 for all primes 𝑝 - 𝑑 except possibly for one prime 𝑞 - 𝑑.
Now the quadratic form h𝑎, 𝑏, −𝑡i is isotropic for all 𝑝 | 𝑑 and at ∞ by construction
and at all primes 𝑝 - 𝑑 except 𝑝 = 𝑞 since 𝑝 - 𝑎𝑏𝑡. Therefore, by case 𝑛 = 3 (using
the “all but one” in Corollary 14.3.6), the form h𝑎, 𝑏, −𝑡i is isotropic.
On the other side, if 𝑛 = 4, then the form h𝑡i 𝑄 0 is isotropic by the same argument.
If 𝑛 ≥ 5, then we apply the induction hypothesis to 𝑄 0: the hypothesis holds, since 𝑄 0
is isotropic at ∞ and all 𝑝 | 𝑑 by construction, and for all 𝑝 - 𝑑 the completion 𝑄 0𝑝 is
a nondegenerate form in ≥ 3 variables over Z 𝑝 so is isotropic by the results of section
12.3, using Hensel’s lemma to lift a solution modulo the odd prime 𝑝.
Putting these two pieces together, we find that 𝑄 is isotropic over Q.
We conclude with the following consequence.
Corollary 14.3.7. Let 𝑄, 𝑄 0 be quadratic forms over Q in the same number of vari-
ables. Then 𝑄 ' 𝑄 0 if and only if 𝑄 𝑣 ' 𝑄 0𝑣 for all places 𝑣.
Proof. The implication (⇒) is immediate. We prove (⇐) by induction on the number
of variables, the case of 𝑛 = 0 variables being clear. By splitting the radical (4.3.9),
we may suppose that 𝑄, 𝑄 0 are nondegenerate. Let 𝑎 ∈ Q× be represented by 𝑄. Since
𝑄 𝑣 ' 𝑄 0𝑣 the quadratic form h−𝑎i 𝑄 0 is isotropic at 𝑣 for all 𝑣, so 𝑄 0 represents 𝑎
(Lemma 5.4.3). In both cases, we can write 𝑄 ' h−𝑎i 𝑄 1 and 𝑄 0 ' h−𝑎i 𝑄 10 for
quadratic forms 𝑄 1 , 𝑄 10 in one fewer number of variables. Finally, by Witt cancellation
(Theorem 4.2.22), from 𝑄 𝑣 ' 𝑄 0𝑣 we have (𝑄 1 ) 𝑣 ' (𝑄 10 ) 𝑣 for all 𝑣, so by induction
𝑄 1 ' 𝑄 10 , and thus 𝑄 ' 𝑄 0.
We now officially complete our proofs.
Proof of Proposition 14.3.1. The implications (i) ⇒ (ii) and (iii) ⇒ (iv) are immediate.
For the implication (ii) ⇒ (iii): either 𝑣 ∈ Ram 𝐵, in which case 𝐵 𝑣 ' 𝐵 0𝑣 is the unique
division algebra over Q 𝑣 (Theorem 12.1.5), or 𝑣 ∉ Ram 𝐵, in which case 𝐵 𝑣 '
M2 (Q 𝑣 ) ' 𝐵 0𝑣 by definition. For the implication (iv) ⇒ (i), recalling Theorem 5.1.1,
by Corollary 14.3.6 applied to the ternary quadratic form associated to 𝐵, we conclude

that this form is isotropic, which by Proposition 5.1.2 implies that 𝐵 ' M2 (Q).
Proof of Main Theorem 14.1.3. The map 𝐵 ↦→ Ram 𝐵 has the desired codomain, by
Hilbert reciprocity (Proposition 14.2.1); it is surjective by Proposition 14.2.7; and it
is injective by Corollaries 14.3.6 and 14.3.7. The second bĳection (with squarefree
integers) is immediate.
To summarize these past few sections, the classification of quaternion algebras

over Q embodies some deep statements in number theory: quadratic reciprocity (and
its reformulation in Hilbert reciprocity), the Hasse–Minkowski theorem (the local-
global principle for quadratic forms), and the proofs use the theorem of the infinitude
of primes in arithmetic progression! It is a small blessing that we can make these
essentially elementary arguments over Q. In the more general case, we must dig more
deeply.
For fun, we conclude this section with a consequence in number theory: Legen-
dre’s three-square theorem (cf. Lagrange’s four-square theorem, Theorem 11.4.3, and
Remark 11.4.4).
Theorem 14.3.8 (Legendre–Gauss). An integer 𝑛 ≥ 0 can be written as the sum of

three squares 𝑛 = 𝑥 2 + 𝑦 2 + 𝑧 2 if and only if 𝑛 is not of the form 𝑛 = 4𝑎 (8𝑏 + 7) with
𝑎, 𝑏 ∈ Z.
Proof. Looking modulo 8, we see that the provided condition is necessary (Exercise
14.3(a)). Conversely, suppose 𝑛 > 0 is not of the form 𝑛 = 4𝑎 (8𝑏 + 7), or equivalently
that −𝑛 ∉ Q×22 (Exercise 14.4). We may suppose 𝑎 = 0, 1.
Let 𝐵 = (−1, −1 | Q) be the rational Hamiltonians. We have Ram 𝐵 = {2, ∞},
which is to say the associated ternary quadratic form 𝑥 2 + 𝑦 2 + 𝑧2 is isotropic over Q 𝑝
for all odd primes 𝑝. Consider the quadratic form 𝑄(𝑥, 𝑦, 𝑧, 𝑤) = 𝑥 2 + 𝑦 2 + 𝑧 2 − 𝑛𝑤 2 .
Then 𝑄 is isotropic over R since 𝑛 > 0, and isotropic over all Q 𝑝 with 𝑝 odd taking
𝑤 = 0. The form is also isotropic over Q2 (Exercise 14.3), lifting a solution modulo
8 via Hensel’s lemma. By the Hasse–Minkowski theorem (Theorem 14.3.3), 𝑄 is
isotropic over Q, so there exist 𝑥, 𝑦, 𝑧, 𝑤 ∈ Q not all zero such that 𝑥 2 + 𝑦 2 + 𝑧2 = 𝑛𝑤 2 .
We must have 𝑤 ≠ 0 by positivity, and dividing through we get 𝑥, 𝑦, 𝑧 ∈ Q not all zero
such that 𝑥 2 + 𝑦 2 + 𝑧 2 = 𝑛. Let 𝛼 = 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 ∈ 𝐵. Then 𝛼2 + 𝑛 = 0 and 𝛼 ∈ 𝐵 is
integral.
Let O0 ⊂ 𝐵 be a maximal order containing 𝛼, and let O be the Hurwitz order. By
Proposition 11.3.7, O0 is conjugate to O; after conjugating, we may suppose 𝛼 ∈ O.
But trd(𝛼) = 0, so necessarily 𝛼 ∈ Zh𝑖, 𝑗i and 𝑥, 𝑦, 𝑧 ∈ Z with nrd(𝛼) = 𝑥 2 + 𝑦 2 + 𝑧2 = 𝑛
as desired.
See also Exercise 14.5 for a variant of the proof of the three-square theorem staying
in the language of quaternions.
14.4. GLOBAL FIELDS 219
14.4 Global fields

In this chapter and in many that remain, we focus on a certain class of fields of
arithmetic interest: a global field is either a finite extension of Q (a number field)
or of F 𝑝 (𝑡) (a function field) for a prime 𝑝. Global fields are strongly governed by
their completions with respect to nontrivial absolute values, which are local fields.
Throughout this text, we will return to this theme that global behavior is governed by
local behavior.
For the rest of this chapter, let 𝐹 be a global field. We quickly introduce in this
section some basic notions from algebraic number theory: for further reference, see
e.g. Neukirch [Neu99, Chapters I–II], Cassels [Cas2010, Chapter II], or Janusz [Jan96,
Chapter II].
Remark 14.4.1. When 𝐹 is a function field, we will often insist that 𝐹 is equipped
with an inclusion 𝐹0 ↩→ 𝐹 where 𝐹0 ' F 𝑝 (𝑡) has pure transcendence degree 1 over
F 𝑝 . (For the geometrically inclined, this corresponds to a morphism 𝑋 → P1 of the
associated curves.) Often this inclusion will not play a role, but it will be important
to treat certain aspects uniformly with the number field case where there is only one
inclusion Q ↩→ 𝐹.
14.4.2. The set of places of 𝐹 is the set Pl 𝐹 of equivalence classes of embeddings
𝜄 𝑣 : 𝐹 → 𝐹𝑣 where 𝐹𝑣 is a local field and 𝜄 𝑣 (𝐹) is dense in 𝐹𝑣 ; two embeddings
𝜄 𝑣 : 𝐹 → 𝐹𝑣 and 𝜄0𝑣 : 𝐹 → 𝐹𝑣0 are said to be equivalent if there is an isomorphism of
topological fields 𝜙 : 𝐹𝑣 → 𝐹𝑣0 such that 𝜄0𝑣 = 𝜙 ◦ 𝜄 𝑣 .
14.4.3. Every valuation 𝑣 : 𝐹 → R ∪ {∞}, up to scaling, defines a place 𝜄 𝑣 : 𝐹 → 𝐹𝑣
where 𝑣 is the completion of 𝐹 with respect to the absolute value induced by 𝑣; we
call such a place nonarchimedean, and using this identification we will write 𝑣 for
both the place of 𝐹 and the corresponding valuation. For a nonarchimedean place 𝑣
corresponding to a local field 𝐹𝑣 , we denote by 𝑅 𝑣 its valuation ring, 𝔭 𝑣 its maximal
ideal, and 𝑘 𝑣 its residue field. If 𝐹 is a function field, then all places of 𝐹 are
nonarchimedean. If 𝐹 is a number field, a place 𝐹 ↩→ R is called a real place and a
place 𝐹 ↩→ C (equivalent to its complex conjugate) is called a complex place. A real
or complex place is archimedean.
14.4.4. Let 𝐾 ⊇ 𝐹 be a finite, separable extension of fields, and let 𝑣 ∈ Pl 𝐹. We say
that a place 𝑤 of 𝐾 is above 𝑣 if 𝑤| 𝐹 = 𝑣, and we write 𝑤 | 𝑣. The set of places 𝑤
above 𝑣 are obtained as follows: since 𝐾 is separable, we have an isomorphism
𝐾 ⊗𝐹 𝐹𝑣 ' 𝐾1 × · · · × 𝐾𝑟 (14.4.5)
where each 𝐾𝑖 ⊇ 𝐹𝑣 is a finite extension of local fields. Indeed, writing 𝐾 =

𝐹 [𝑥]/( 𝑓 (𝑥)) with 𝑓 (𝑥) ∈ 𝐹 [𝑥] the minimal polynomial of a primitive element, we
have
𝐾 ⊗𝐹 𝐹𝑣 ' 𝐹𝑣 [𝑥]/( 𝑓 (𝑥)) ' 𝐹𝑣 [𝑥]/( 𝑓1 (𝑥)) × · · · × 𝐹𝑣 [𝑥]/( 𝑓𝑟 (𝑥))
where 𝑓 (𝑥) = 𝑓1 (𝑥) · · · 𝑓𝑟 (𝑥) ∈ 𝐹𝑣 [𝑥] is the factorization of 𝑓 (𝑥) into irreducibles in
𝐹𝑣 [𝑥], distinct because 𝑓 is separable. Thus each 𝐾𝑖 is a local field by the classification
in Theorem 12.2.15, and the composition
𝐾 ↩→ 𝐾 ⊗𝐹 𝐹𝑣 → 𝐾𝑖
defines a place 𝑤 𝑖 of 𝐾 above 𝑣. Conversely, every place 𝑤 above 𝑣 is equivalent to 𝑤 𝑖

for some 𝑖 [Jan96, Chapter II, Theorem 5.1; Cas78, §9; Neu99, Chapter II, Proposition
(8.3)].
We say that a nonarchimedean place 𝑣 ramifies in 𝐾 if there exists a place 𝑤 | 𝑣
such that 𝐾 𝑤 ⊇ 𝐹𝑣 is ramified (see 13.2.3). Only finitely many places of 𝐹 ramify in
𝐾.
A global field 𝐹 has a set of preferred embeddings 𝜄 𝑣 : 𝐹 ↩→ 𝐹𝑣 corresponding to

each place 𝑣 ∈ Pl 𝐹—equivalently, a preferred choice of absolute values | | 𝑣 for each
place 𝑣 ∈ Pl 𝐹—such that the product formula holds: for all 𝑥 ∈ 𝐹 × ,
Ö
|𝑥| 𝑚
𝑣 = 1.
𝑣
(14.4.6)
𝑣 ∈Pl 𝐹
where 𝑚 𝑣 = 2 if 𝑣 is complex and 𝑚 𝑣 = 1 otherwise. Admittedly, the extra exponents

2 for the complex places are annoying (see Remark 12.2.3)! Often what is done is
to define normalized absolute values k𝑥k 𝑣 := |𝑥| 𝑚𝑣 for 𝑣 ∈ Pl 𝐹, so then (14.4.6)
𝑣
becomes Ö
k𝑥k 𝑣 = 1. (14.4.7)
𝑣 ∈Pl 𝐹
Preferred absolute values are defined as follows.
14.4.8. The set of places Pl(Q) of Q consists of the archimedean real place, induced
by the embedding Q ↩→ R and the usual absolute value |𝑥| ∞ , and the set of nonar-
chimedean places indexed by the primes 𝑝 given by the embeddings Q ↩→ Q 𝑝 , with
the preferred absolute value
|𝑥| 𝑝 = 𝑝 −𝑣𝑝 ( 𝑥) .
The statement of the product formula for 𝑥 ∈ Q is
Ö
|𝑥| ∞ 𝑝 −𝑣𝑝 ( 𝑥) = 1; (14.4.9)
𝑝
𝑝 𝑣𝑝 ( 𝑥) = |𝑥|, and this follows from unique

Î
rearranging, (14.4.9) is equivalent to 𝑝
factorization in Z.
14.4.10. The set of places of F 𝑝 (𝑡) is indexed by monic irreducible polynomials

𝑓 (𝑡) ∈ F 𝑝 [𝑡] with preferred absolute value
|𝑥(𝑡)| 𝑓 = 𝑝 −(deg 𝑓 ) ord 𝑓 ( 𝑥)
and 1/𝑡, the place at infinity, with preferred absolute value
|𝑥(𝑡)| 1/𝑡 = 𝑝 deg 𝑥 ,

14.4. GLOBAL FIELDS 221
where if 𝑥 = 𝑓 /𝑔 is the ratio of relatively prime polynomials 𝑓 , 𝑔 ∈ F 𝑝 [𝑡], then

deg 𝑥 := max(deg 𝑓 , deg 𝑔).
Then the statement of the product formula for 𝑥(𝑡) ∈ F 𝑝 (𝑡) is
Ö
𝑝 deg 𝑥 𝑝 −(deg 𝑓 ) ord 𝑓 ( 𝑥) = 1; (14.4.11)
𝑓
rearranging
Í as over Q, but now also taking the logarithm in base 𝑝, (14.4.9) is equivalent
to 𝑓 (deg 𝑓 ) ord 𝑓 (𝑥) = deg 𝑥 which follows from unique factorization in F 𝑝 [𝑡].
14.4.12. More generally, let 𝐾 ⊇ 𝐹 be a finite, separable extension of global fields.
Let 𝑣 be a place of 𝐹 with a preferred absolute value and let 𝑤 be a place of 𝐾 above
𝑣. Then the preferred absolute value for 𝑤 is the unique one extending 𝑣, namely
𝑤 :𝐹𝑣 ]
|𝑥| 𝑤 = |Nm𝐾𝑤 |𝐹𝑣 (𝑥)| 1/[𝐾
𝑣
for 𝑥 ∈ 𝐾. These absolute values fit together, with

Ö
|𝑥| 𝑤 = |Nm𝐾 |𝐹 (𝑥)| 𝑣 (14.4.13)
𝑤 |𝑣
for all 𝑥 ∈ 𝐾, a consequence of (14.4.5) [Jan96, Chapter II, Theorem 5.2; Cas78, §11,
Theorem, p. 59; Neu99, Chapter II, Corollary (8.4)].
In particular, if 𝐹 satisfies the product formula (14.4.6) with respect to preferred
absolute values, then so does 𝐾, since
Ö Ö Ö Ö
|𝑥| 𝑚
𝑤
𝑤
= |𝑥| 𝑚𝑤
𝑤 = |Nm𝐾 |𝐹 (𝑥)| 𝑚
𝑣 = 1.
𝑣
(14.4.14)
𝑤 𝑣 𝑤 |𝑣 𝑣
Remark 14.4.15. The definitions for the preferred absolute values are pretty dry—
sorry! But we will see later that they are natural from the perspective of Haar
measure: see section 29.3 and ultimately (29.6.3).
We will also make use of the following notation in many places in the text. Let 𝐹
be a global field.
Definition 14.4.16. A set S ⊆ Pl 𝐹 is eligible if S is finite, nonempty, and contains all
archimedean places of 𝐹.
Definition 14.4.17. Let S be an eligible set of places. The ring of S-integers in 𝐹 is
the set
𝑅 (S) := {𝑥 ∈ 𝐹 : 𝑣(𝑥) ≥ 0 for all 𝑣 ∉ S}. (14.4.18)
A global ring is a ring of S-integers in a global field for an associated eligible set S.
The expression (14.4.18) makes sense, since if 𝑣 ∉ S then by hypothesis 𝑣 is
nonarchimedean. When no confusion can result, we will abbreviate 𝑅 = 𝑅 (S) for a
global ring 𝑅.
Example 14.4.19. If 𝐹 is a number field and S consists only of the archimedean places
in 𝐹 then 𝑅 (S) is the ring of integers in 𝐹, the integral closure of Z in 𝐹, also denoted
𝑅 (S) = Z𝐹 . If 𝐹 is a function field, corresponding to a curve 𝑋, then 𝑅 (S) is the ring
of all rational functions with no poles outside S. (So in all cases, it is helpful to think
of the ring 𝑅 (S) as consisting of those elements of 𝐹 with “no poles outside S”.)
14.5 Ramification and discriminant

Let 𝑅 = 𝑅 (S) be a global ring, with S ⊂ Pl 𝐹 eligible. Let 𝐵 be a quaternion algebra
over 𝐹.
Definition 14.5.1. Let 𝑣 ∈ Pl 𝐹. We say that 𝐵 is ramified at 𝑣 if 𝐵 𝑣 = 𝐵 ⊗𝐹 𝐹𝑣 is a
division ring; otherwise we say that 𝐵 is split (or unramified) at 𝑣.
Let Ram 𝐵 denote the set of ramified places of 𝐵.
If 𝑣 ∈ Pl 𝐹 is a nonarchimedean place, corresponding to a prime 𝔭 of 𝑅, we will
also say that 𝐵 is ramified at 𝔭 when 𝐵 is ramified at 𝑣.
Remark 14.5.2. We use the term ramified for the following reason: if 𝐵𝔭 is a division
ring with valuation ring O𝔭 , then 𝔭O𝔭 = 𝑃2 for a two-sided maximal ideal 𝑃: see
Theorem 13.3.11. (Eichler [Eic55-56, §1, Theorem 4] called them characteristic
primes.)
Lemma 14.5.3. The set Ram 𝐵 of ramified places of 𝐵 is finite.
Proof. Write 𝐵 = (𝐾, 𝑏 | 𝐹). Since 𝐹 has only finitely many archimedean places, we
may suppose 𝑣 is nonarchimedean. The extension 𝐾 ⊇ 𝐹 is ramified at only finitely
many places, so we may suppose that 𝐾 ⊇ 𝐹 is unramified at 𝑣. Finally, 𝑣(𝑏) = 0 for
all but finitely many 𝑣, so we may suppose 𝑣(𝑏) = 0. But then under these hypotheses,
𝐵 𝑣 = (𝐾 𝑣 , 𝑏 | 𝐹𝑣 ) is split, by Corollary 13.4.1.
Motivated by the fact that the discriminant of a quadratic field extension is divisible
by ramifying primes, we make the following definition.
Definition 14.5.4. The 𝑅-discriminant of 𝐵 is the 𝑅-ideal
Ö
disc𝑅 (𝐵) = 𝔭⊆𝑅
𝔭∈Ram 𝐵
𝔭∉S
obtained as the product of all primes 𝔭 of 𝑅 = 𝑅 (S) ramified in 𝐵.

Remark 14.5.5. When 𝐹 is a number field and S consists of archimedean places only,
so that 𝑅 = Z𝐹 is the ring of integers of 𝐹, we abbreviate disc𝑅 (𝐵) = disc 𝐵. The
discriminant disc𝑅 (𝐵) discards information about primes in S: only Ram 𝐵 records
information about 𝐵 that is independent of S.
Remark 14.5.6. One could make the same definitions when 𝑅 is more generally a
Dedekind domain. However, unless the residue fields of 𝑅 are finite, this is not as
useful a notion: see Exercise 14.13. (In some sense, this is because the Brauer group
of 𝐹 = Frac 𝑅 is not as simply described as when 𝐹 is a global field, viz. Remark
14.6.10.)
As usual, the archimedean places play a special role for number fields, so we make
the following definition.
Definition 14.5.7. Let 𝐹 be a number field. We say that 𝐵 is totally definite if all
archimedean places of 𝐹 are ramified in 𝐵; otherwise, we say 𝐵 is indefinite.
14.6. QUATERNION ALGEBRAS OVER GLOBAL FIELDS 223
14.5.8. If 𝑣 is a complex place, then 𝑣 is necessarily split, since the only quaternion
algebra over C is M2 (C); therefore, if 𝐵 is a totally definite quaternion algebra over a
number field 𝐹, then 𝐹 is totally real.
14.6 Quaternion algebras over global fields

We now generalize Main Theorem 14.1.3 to the global field 𝐹, deducing results
characterizing isomorphism classes of quaternion algebras. The main result is as
follows.
Main Theorem 14.6.1. Let 𝐹 be a global field. Then the map 𝐵 ↦→ Ram 𝐵 gives a
bĳection

Quaternion algebras over 𝐹 Finite subsets of noncomplex places
↔ .
up to isomorphism of 𝐹 of even cardinality
In other words, if 𝐵 is a quaternion algebra over a global field, then the set of places
of 𝐹 where 𝐵 is ramified is finite and of even cardinality, this set uniquely determines
𝐵 up to isomorphism, and every such set occurs.
Proof. We give a proof in section 26.8, which itself relies on an analytic result (The-
orem 26.8.19) proven in Chapter 29.
Alternatively, this statement can also be viewed a direct consequence of a (hard-
earned) fundamental exact sequence in class field theory: see Remark 14.6.10.
Recall the definition of the Hilbert symbol (as in section 12.4), computed explicitly
for 𝑣 an odd nonarchimedean place (12.4.9): for a place 𝑣 of 𝐹, we abbreviate
(𝑎, 𝑏) 𝐹𝑣 = (𝑎, 𝑏) 𝑣 . We also recall Lemma 14.5.3 that (𝑎, 𝑏) 𝑣 = 1 for all but finitely
many places 𝑣.
Corollary 14.6.2 (Hilbert reciprocity). Let 𝐹 be a global field with char 𝐹 ≠ 2 and let
𝑎, 𝑏 ∈ 𝐹 × . Then Ö
(𝑎, 𝑏) 𝑣 = 1. (14.6.3)
𝑣 ∈Pl 𝐹
Proof. Immediate from Main Theorem 14.6.1: Hilbert reciprocity is equivalent to the
statement that # Ram 𝐵 is even.
Remark 14.6.4. Stating the reciprocity law in the form (14.6.3) is natural from the
point of view of the product formula (14.4.6). And Hilbert reciprocity can be rightly
seen as a law of quadratic reciprocity for number fields (as we saw in section 14.2 for
𝐹 = Q). (For more, see Exercise 14.15.)
Hilbert saw his reciprocity law (Corollary 14.6.2) as an analogue of Cauchy’s inte-
gral theorem [Hil32, p. 367–368]; for more on this analogy, see Vostokov [Vos2009].
Corollary 14.6.5 (Local-global principle for quaternion algebras). Let 𝐵, 𝐵 0 be quater-
nion algebras over 𝐹. Then the following are equivalent:
(i) 𝐵 ' 𝐵 0;
(ii) Ram 𝐵 = Ram(𝐵 0);

(iii) 𝐵 𝑣 ' 𝐵 0𝑣 for all places 𝑣 ∈ Pl 𝐹; and
(iv) 𝐵 𝑣 ' 𝐵 0𝑣 for all but one place 𝑣 ∈ Pl 𝐹.
In particular, 𝐵 ' M2 (𝐹) if and only if Ram 𝐵 = ∅.
Proof. For the equivalence (i) ⇔ (ii) ⇔ (iii), combine Main Theorem 14.6.1 and the
fact that for a noncomplex place 𝑣 there is a unique division algebra over 𝐹𝑣 . The
equivalence (iii) ⇔ (iv) follows from the parity constraint, since if 𝑣 is a place and
Ram 𝐵 r {𝑣} = Σ, then 𝑣 ∈ Ram 𝐵 or not according as #Σ is odd or even.
Remark 14.6.6. Corollary 14.6.5 is a special case of the Albert–Brauer–Hasse–Noether

theorem [AH32, BHN31]: a central simple algebra 𝐴 over 𝐹 such that 𝐴 𝑣 ' M𝑛 (𝐹𝑣 )
for all 𝑣 ∈ Pl 𝐹 has 𝐴 ' M𝑛 (𝐹). See Remark 14.6.10 for further discussion.
The statement of Corollary 14.6.5 is the local-global principle for quaternion alge-
bras: the isomorphism class of a quaternion algebra over a global field is determined
by its isomorphism classes over the collection of local fields obtained as completions
of the global field. In a similar way, we have a local-global principle for quadratic
embeddings as follows.
Proposition 14.6.7 (Local-global principle for splitting/embeddings). Let 𝐾 ⊇ 𝐹 be

a finite separable extension of global fields. Then the following are equivalent:
(i) 𝐾 splits 𝐵, i.e., 𝐵 ⊗𝐹 𝐾 ' M2 (𝐾); and

(ii) For all places 𝑤 ∈ Pl 𝐾, the field 𝐾 𝑤 splits 𝐵.
If dim𝐹 𝐾 = 2, then these are further equivalent to:
(iii) There is an embedding 𝐾 ↩→ 𝐵 of 𝐹-algebras;

(iv) For all places 𝑣 ∈ Pl 𝐹, there is an embedding 𝐾 𝑣 ↩→ 𝐵 𝑣 of 𝐹𝑣 -algebras; and
(v) Every 𝑣 ∈ Ram 𝐵 does not split in 𝐾, i.e., 𝐾 𝑣 is a field for all 𝑣 ∈ Ram 𝐵.
Proof. The equivalence (i) ⇔ (ii) is a consequence of Corollary 14.6.5: they are both
equivalent to Ram 𝐵 𝐾 = ∅, since 𝐾 splits 𝐵 if and only if 𝐵 ⊗𝐹 𝐾 ' M2 (𝐾) if and only
if Ram(𝐵 ⊗𝐹 𝐾) = ∅ if and only if for all places 𝑤 of 𝐾 we have 𝐵 ⊗𝐹 𝐾 𝑤 ' M2 (𝐾 𝑤 ).
The equivalence (i) ⇔ (iii) was given by Lemmas 5.4.7 and 6.4.12.
The implication (iii) ⇒ (iv) is clear. For the implication (iv) ⇒ (v), if 𝑣 ∈ Ram 𝐵,
then 𝐵 𝑣 is a division algebra; so if 𝐾 𝑣 is not a field, then we cannot have 𝐾 𝑣 ↩→ 𝐵 𝑣 .
Finally, for (v) ⇒ (ii), let 𝑤 ∈ Pl 𝐾 with 𝑤 | 𝑣 ∈ Pl 𝐹. If 𝑣 ∉ Ram 𝐵 then already 𝐹𝑣
splits 𝐵; otherwise, 𝑣 ∈ Ram 𝐵 and 𝐾 𝑣 = 𝐾 𝑤 is a field with [𝐾 𝑤 : 𝐹𝑣 ] = 2, so by
Proposition 13.4.4, 𝐾 𝑤 splits 𝐵.
14.6.8. The equivalences (iii) ⇔ (iv) ⇔ (v) in Proposition 14.6.7 hold also for the
separable 𝐹-algebra 𝐾 = 𝐹 × 𝐹: for there is an embedding 𝐹 × 𝐹 ↩→ 𝐵 if and only if
𝐵 ' M2 (𝐹).
We also record the statement of the Hasse–Minkowski theorem over global fields,
generalizing Theorem 14.3.3.
14.7. THEOREMS ON NORMS 225
Theorem 14.6.9 (Hasse–Minkowski). Let 𝐹 be a global field and let 𝑄 be a quadratic

form over 𝐹. Then 𝑄 is isotropic over 𝐹 if and only if 𝑄 𝑣 is isotropic over 𝐹𝑣 for all
places 𝑣 of 𝐹.
Proof. The same comments as in the proof of Main Theorem 14.6.1 apply: we give
a proof in section 26.8. But see also O’Meara [O’Me73, §§65–66] for a standalone
class field theory proof for the case when 𝐹 is a number field.
This local-global principle for isotropy of quadratic forms is also called the Hasse
principle. For a historical overview of the Hasse principle, and more generally Hasse’s
contributions in the arithmetic theory of algebras, see Fenster–Schwärmer [FS2007].
Remark 14.6.10. The fact that quaternion algebras are classified by their ramification
set (Main Theorem 14.6.1) over a global field 𝐹 is a consequence of the following
theorem from class field theory: there is an exact sequence
Ê
0 → Br(𝐹) → Br(𝐹𝑣 ) → Q/Z → 0
𝑣 (14.6.11)
Í
( [ 𝐴 𝑣 ]) 𝑣 ↦→ 𝑣 inv 𝑣 [ 𝐴 𝑣 ]
where the first map is the natural diagonal inclusion [ 𝐴] ↦→ ( [ 𝐴 ⊗ 𝑣 𝐹𝑣 ]) 𝑣 and the
second map is the sum of the local invariant maps inv 𝑣 : Br(𝑘 𝑣 ) → Q/Z from Remark
13.4.3. The class of a quaternion algebra 𝐵 in a Brauer group over a field is 2-torsion
by 8.3.4, and the local invariant inv 𝑣 𝐵 𝑣 is equal to 0, 1/2 according as 𝐵 𝑣 is split or
ramified, and in this way we recover the main classification theorem. (In this sense, the
discriminant of a quaternion algebra captures the Brauer class of a quaternion algebra
at the finite places, and the ramification set captures it fully.) The exact sequence
(14.6.11) is sometimes called the fundamental exact sequence of global class field
theory: see Milne [Milne-CFT, §VIII.4] or Neukirch–Schmidt–Wingberg [NSW2008,
Theorem 8.1.17].
14.7 Theorems on norms

In the previous sections, we have seen how both local-global principles allow a nice,
clean understanding of quaternion algebras—and at the same time, the norm groups
play an important role in this characterization. These themes will continue through
the book, so we develop them here in an important first case by describing the group
nrd(𝐵× ) ≤ 𝐹 × .
We retain our hypotheses that 𝐹 is a global field and 𝐵 a quaternion 𝐹-algebra.
14.7.1. First, we recall the calculation of the local norm groups (Lemma 13.4.9): for
𝑣 ∈ Pl 𝐹, we have
(
× R×>0 , if 𝑣 ∈ Ram 𝐵 is real (i.e., 𝐵 𝑣 ' H);
nrd(𝐵 𝑣 ) =
𝐹𝑣× , otherwise.
Under 𝐵 ↩→ 𝐵 𝑣 , we have nrd(𝐵× ) ≤ nrd(𝐵×𝑣 ) for all places 𝑣 ∈ Pl 𝐹, and this

‘places’ a condition on the reduced norm at precisely the real ramified places.
14.7.2. Let Ω ⊆ Ram 𝐵 be the set of ramified (necessarily real) archimedean places
in 𝐵. (If 𝐹 is a function field, then Ω = ∅.) Let
𝐹>×Ω 0 := {𝑥 ∈ 𝐹 × : 𝑣(𝑥) > 0 for all 𝑣 ∈ Ω} (14.7.3)
be the group of elements that are positive for the embeddings 𝑣 ∈ Ω. For Ω the set of
× and call such elements totally positive.
all real places, we write simply 𝐹>0
By 14.7.1, we have nrd(𝐵× ) ≤ 𝐹>×Ω 0 . In fact, equality holds.
Main Theorem 14.7.4 (Hasse–Schilling). We have nrd(𝐵× ) = 𝐹>×Ω 0 .
To prove this theorem, we will use two lemmas.
Lemma 14.7.5. Let 𝑣 be a noncomplex place of 𝐹. Let 𝑛 𝑣 ∈ 𝐹𝑣× , and if 𝑣 is real

suppose 𝑛 𝑣 > 0. Then there exists 𝑡 𝑣 ∈ 𝐹𝑣 such that 𝑥 2 − 𝑡 𝑣 𝑥 + 𝑛 𝑣 is separable and
irreducible over 𝐹𝑣 . Moreover, if 𝑛 𝑣 ∈ 𝑅 𝑣 then we may take 𝑡 𝑣 ∈ 𝑅 𝑣 .
Proof. We suppose that char 𝐹𝑣 ≠ 2 and leave the other case as an exercise (Exercise
14.22). If 𝑛 𝑣 ∉ 𝐹𝑣×2 , then we can take 𝑡 𝑣 = 0; √this includes the case where 𝑣 is a
real place. So suppose 𝑛 𝑣 ∈ 𝐹𝑣×2 . Let 𝐾 𝑤 = 𝐹𝑣 ( 𝑑 𝑣 ) ⊇ 𝐹𝑣 be a separable quadratic
extension, and in particular 𝑑 𝑣 ∉ 𝐹𝑣×2 . The quadratic form h1, −4𝑛 𝑣 i ' h1, −1i over
𝐹𝑣 is a hyperbolic plane (Definition 5.4.1) so universal; let 𝑥 𝑣 , 𝑦 𝑣 ∈ 𝐹𝑣 be such that
𝑥 2𝑣 − 4𝑛 𝑣 𝑦 2𝑣 = 𝑑 𝑣 . We cannot have 𝑦 𝑣 = 0, else 𝑥 2𝑣 = 𝑑 𝑣 ∈ 𝐹𝑣×2 . Let 𝑡 𝑣 = 𝑥 𝑣 /𝑦 𝑣 . Then
𝑥 2 − 𝑡 𝑣 𝑥 + 𝑛 𝑣 has discriminant 𝑑 𝑣 and so is separable and irreducible.
For the second statement, multiplying by a square we may suppose without loss of
generality that 𝑑 𝑣 ∈ 𝑅, so the equality 𝑡 2𝑣 = 𝑑 𝑣 + 4𝑛 𝑣 implies 𝑣(𝑡 𝑣 ) ≥ 0 and 𝑡 𝑣 ∈ 𝑅.
Next, we want to show that we can approximate a polynomial over a completion

𝐹𝑣 by a polynomial over the global field 𝐹 sufficiently well—the reader is invited to
ignore this on a first reading and accept this intuitively as a consequence of the fact
that 𝐹 is dense in 𝐹𝑣 .
Lemma 14.7.6. Let 𝑣 ∈ Pl 𝐹, let 𝑓 𝑣 (𝑥) = 𝑥 2 − 𝑡 𝑣 𝑥 + 𝑛 𝑣 ∈ 𝐹𝑣 [𝑥] be a separable

polynomial, and let 𝜖 > 0. Then there exists 𝑡, 𝑛 ∈ 𝐹 such that |𝑡 − 𝑡 𝑣 |, |𝑛 − 𝑛 𝑣 | < 𝜖
and such that 𝑓 (𝑥) = 𝑥 2 − 𝑡𝑥 + 𝑛 has
𝐹𝑣 [𝑥]/( 𝑓 (𝑥)) ' 𝐹𝑣 [𝑥]/( 𝑓 𝑣 (𝑥)). (14.7.7)
In particular, 𝑓 (𝑥) is separable, and if 𝑓 𝑣 (𝑥) is irreducible then so is 𝑓 (𝑥).

Further, if already 𝑛 𝑣 ∈ 𝐹 then we may take 𝑛 = 𝑛 𝑣 , and similarly with 𝑡.
Proof. If 𝐹𝑣 is nonarchimedean, the lemma follows from Corollary 13.2.9 and the
fact that 𝐹 is dense in 𝐹𝑣 . The case where 𝐹𝑣 is archimedean is straightforward: see
Exercise 14.23.
The same argument can be applied to several local fields at once, as follows.
Corollary 14.7.8. Let Σ ⊆ Pl 𝐹 be a finite set of noncomplex places. For each 𝑣 ∈ Σ,

let 𝑓 𝑣 (𝑥) = 𝑥 2 − 𝑡 𝑣 𝑥 + 𝑛 𝑣 ∈ 𝐹𝑣 [𝑥] be a separable polynomial, and let 𝜖 > 0. Then
there exists 𝑡, 𝑛 ∈ 𝐹 such that for 𝑓 (𝑥) = 𝑥 2 − 𝑡𝑥 + 𝑛 and for all 𝑣 ∈ Σ we have
|𝑡 − 𝑡 𝑣 |, |𝑛 − 𝑛 𝑣 | < 𝜖 and 𝐹𝑣 [𝑥]/( 𝑓 (𝑥)) ' 𝐹𝑣 [𝑥]/( 𝑓 𝑣 (𝑥)). In particular, 𝑓 (𝑥) is
separable, and if 𝑓 𝑣 (𝑥) is irreducible for some 𝑣 then so is 𝑓 (𝑥).
Further, if all 𝑛 𝑣 = 𝑚 ∈ 𝐹 for 𝑣 ∈ Σ, then we may take 𝑛 = 𝑚, and similarly with 𝑡.
Proof. We repeat
Î the argument of Lemma 14.7.6, using weak approximation (i.e., 𝐹
is dense in 𝑣 𝐹𝑣 ; look ahead to Lemma 28.7.1 and the adjacent discussion) for all
𝑣 ∈ Σ to find 𝑡, 𝑛.
We now conclude with a proof of the theorem on norms.
Proof of Main Theorem 14.7.4. Let 𝑛 ∈ 𝐹>×Ω 0 . We will construct a separable quadratic
extension 𝐾 ⊇ 𝐹 with 𝐾 ↩→ 𝐵 such that 𝑛 ∈ Nm𝐾 |𝐹 (𝐾 × ). To this end, by Proposition
14.6.7, it is enough to find 𝐾 ⊇ 𝐹 such that 𝐾 𝑣 is a field for all 𝑣 ∈ Ram 𝐵.
By Lemma 14.7.5, for all 𝑣 ∈ Ram 𝐵, there exists 𝑡 𝑣 ∈ 𝐹𝑣 such that the polynomial
𝑥 2 − 𝑡 𝑣 𝑥 + 𝑛 ∈ 𝐹𝑣 [𝑥] is separable and irreducible over 𝐹𝑣 ; here if 𝑣 ∈ Ω is real we
use that 𝑣(𝑛) > 0. By Corollary 14.7.8, there exists 𝑡 ∈ 𝐹 such that 𝑥 2 − 𝑡𝑥 + 𝑛
irreducible over each 𝐹𝑣 . Let 𝐾 be the extension of 𝐹 obtained by adjoining a root
of this polynomial. Then 𝐾 𝑣 is a field for each ramified 𝑣, and 𝑛 ∈ Nm𝐾 |𝐹 (𝐾 × ) as
desired.
Exercises
I 1. Complete the proof of Hilbert reciprocity (Proposition 14.2.1) in the remaining
cases (𝑎, 𝑏) = (−1, 2), (2, 2), (−1, 𝑝), (2, 𝑝). In particular, show that

−1, 2 2, 2
' ' M2 (Q)
Q Q
and
𝑎
(𝑎, 𝑝)2 = (𝑎, 𝑝) 𝑝 =
𝑝
for 𝑎 = −1, 2 and all primes 𝑝 (cf. 12.4.13).
2. Derive the law of quadratic reciprocity (14.2.4) and the supplement (14.2.5)
from the statement of Hilbert reciprocity (Proposition 14.2.1).
I 3. Let 𝑛 ∈ Z>0 .
(a) Suppose 𝑛 is of the form 𝑛 = 4𝑎 (8𝑏 + 7) with 𝑎, 𝑏 ∈ Z. Show that there is
no solution to 𝑥 2 + 𝑦 2 + 𝑧2 = 𝑛 with 𝑥, 𝑦, 𝑧 ∈ Z. [Hint: Look modulo 8.]
(b) Suppose 𝑛 is not of the form 𝑛 = 4𝑎 (8𝑏 + 7) with 𝑎, 𝑏 ∈ Z. Show that there
is a solution to 𝑥 2 + 𝑦 2 + 𝑧2 = 𝑛 with 𝑥, 𝑦, 𝑧 ∈ Z2 . [Hint: lift a solution
modulo 8 using Hensel’s lemma.]
I 4. Let 𝑛 ∈ Z be nonzero. Show that 𝑛 is a square in Q2 if and only if 𝑛 is of the
form 𝑛 = 4𝑎 (8𝑏 + 1) with 𝑎, 𝑏 ∈ Z.
√
5. Let 𝑛 > 0 have −𝑛 ∉ Q×2
2 . Let 𝐵 = (−1, −1 | Q) and let 𝐾 = Q( −𝑛). Show that
𝐾 splits 𝐵. [Hint: Use the local-global principle for embeddings (Proposition
14.6.7).] Conclude that there exists 𝛼 ∈ 𝐵 such that 𝛼2 = −𝑛, and conclude as
in Theorem 14.3.8 that 𝑛 is the sum of three squares.
6. Let 𝐹 be a number field. Show that every totally positive element of 𝐹 is a sum
of four squares of elements of 𝐹.
7. Show that the law of Hilbert reciprocity (Proposition 14.2.1) implies the law of
quadratic reciprocity; with the argument given in section 14.1, this completes
the equivalence of these two laws.
8. In the proof of Legendre’s theorem (Theorem 14.3.4), we reduced to the case
𝑎, 𝑏 > 0 and 𝑐 = −1. Show that this reduction is valid.
9. In this exercise, we generalize the proof of Proposition 14.2.7 to give a more
general construction of quaternion quaternion algebras. Let 𝐷 be a squarefree
positive integer and let 𝑢 = −1 if 𝐷 has an odd number of prime divisors,
otherwise 𝑢 := 1.
√
(a) For 𝑏 ∈ Z squarefree, show that 𝐾 := Q( 𝑏) embeds in a quaternion
algebra of discriminant 𝐷 if and only if:
• 𝑏 < 0 if 𝐵 is definite;
• 𝑏 . 1 (mod 8) if 2 | 𝐷; and
𝑏
• for all odd primes 𝑝 | 𝐷, we have ≠ 1.
𝑝
(b) Suppose 𝑏 satisfies the conditions in (a) but with the
further requirement
𝑏
that 𝐷 | 𝑏, i.e., in the third condition we require = 0. Let 𝑞 be an odd
𝑝
prime such that 𝑞 ♦ := 𝑢𝑞 has:
♦
𝑞
• = −1 for all odd 𝑝 | 𝐷;
𝑝♦
𝑞
• = 1 for all odd 𝑝 | (𝑏/𝐷); and
𝑝
• 𝑞 ♦ ≡ 1, 5 (mod 8) according as 2 - 𝐷 or 2 | 𝐷.
Note there exist infinitely many such primes 𝑞 by the infinitude of primes
𝑞♦, 𝑏

in arithmetic progression. Then show that 𝐵 := has disc 𝐵 = 𝐷
Q
and Ram 𝐵 = Σ.
I 10. Let S ⊆ Pl(Q) be eligible. For each 𝑣 ∈ S, let 𝑡 𝑣 ∈ Q×𝑣 be given. Show that there
exists 𝑡 ∈ Q× such that 𝑡 ∈ 𝑡 𝑣 Q×2
𝑣 for all 𝑣 ∈ S and 𝑣 𝑝 (𝑡) = 0 for all 𝑝 ∉ S r {∞}
except (possibly) for one prime 𝑝 = 𝑞.
√
11. Let 𝐹 = Q( 𝑑) be a real quadratic field. Find 𝑎, 𝑏 ∈ Q× (depending on 𝑑) such
that (𝑎, 𝑏 | 𝐹) is a division ring unramified at all finite places.
12. Let 𝐹 be a global field with char 𝐹 ≠ 2 and let 𝐵 be a quaternion algebra over
𝐹. Let 𝐿 ⊇ 𝐹 be a finite extension. An extension 𝐾 ⊇ 𝐹 is linearly disjoint
with 𝐿 over 𝐹 if the multiplication map 𝐾 ⊗𝐹 𝐿 −→ ∼ 𝐾 𝐿 is an isomorphism of
𝐹-algebras.
Show that there exists a splitting field 𝐾 ⊇ 𝐹 for 𝐵 such that 𝐾 is linearly disjoint
with 𝐿 over 𝐹.
13. Show that the notion of discriminant of a quaternion algebra as the product of
ramified primes is not such a great notion when 𝑅 is an arbitrary Dedekind
domain, as follows.
Let 𝑅 = Q[𝑡]; then 𝑅 is a Dedekind domain. Let 𝐹 = Frac 𝑅 = Q(𝑡). Let
𝐵0 = (𝑎, 𝑏 | Q) be a division quaternion algebra over Q and let 𝐵 = 𝐵0 ⊗Q 𝐹 =
(𝑎, 𝑏 | Q(𝑡)). Show that there are infinitely primes at which 𝐵 is “ramified”:
for every prime 𝔭 = (𝑡 − 𝑐)𝑅, show that the algebra 𝐵𝔭 is a division quaternion
algebra over 𝐹𝔭 ' Q((𝑡)). [Hint: See Exercise 13.7.]
14. Using Hilbert reciprocity, one can convert the computation of an even Hilbert
symbol to the computation of several odd Hilbert symbols, as follows.
Let 𝐹 be a number field, let 𝔭 | (2), and let 𝑎, 𝑏 ∈ 𝐹 × . Show that there exist
(computable) 𝑎 0, 𝑏 0 ∈ 𝐹 × such that the following hold:
(i) (𝑎, 𝑏)𝔭 = (𝑎 0, 𝑏 0)𝔭 ; and
(ii) ord𝔮 (𝑎 0) = ord𝔮 (𝑏 0) = 0 for all 𝔮 | (2) with 𝔮 ≠ 𝔭.
Conclude that
Ö
(𝑎, 𝑏)𝔭 = (𝑎 0, 𝑏 0) 𝑣 .
𝑣 ∈Pl 𝐹
𝑣 odd
15. Let 𝐹 be a number field with ring of integers Z𝐹 . We say an ideal 𝔟 ⊆ Z𝐹 is odd
is odd, and 𝑏 ∈ Z𝐹 is odd if (𝑏) is odd. For 𝑎 ∈ Z𝐹 and 𝔟 ⊆ Z𝐹 odd,
if Nm(𝔟)
𝑎
let be the generalized Jacobi symbol, extending the generalized Legendre
𝔟
𝑎 𝑎
symbol by multiplicativity, and write := for 𝑎, 𝑏 ∈ Z𝐹 r {0} with
𝑏 𝑏Z𝐹
𝑏 odd.
(a) Let 𝑎, 𝑏 ∈ Z𝐹 satisfy 𝑎Z𝐹 + 𝑏Z𝐹 = Z𝐹 , with 𝑏 odd, and suppose 𝑎 = 𝑎 0 𝑎 1

with 𝑎 1 odd. Then
Ö
𝑎 𝑏
= (𝑎, 𝑏) 𝑣 .
𝑏 𝑎1
𝑣 |2∞
(b) Suppose that 𝐹 has a computable Euclidean function 𝑁 and let 𝑎, 𝑏 ∈

Z𝐹 r {0} with 𝑏 odd.
Describe an algorithm using (a) to compute the
𝑎
Legendre symbol .
𝑏
16. In this exercise, we give a constructive proof of the surjectivity of the map
𝐵 ↦→ Ram 𝐵 in Main Theorem 14.6.1 in the spirit of the proof of Proposition
14.2.7 (assuming two analytic results).
Let 𝐹 be a number field, and let Σ ⊆ Pl 𝐹 be a finite set of noncomplex places
of 𝐹 of even cardinality. Let 𝑅 = Z𝐹 be the ring of integers of 𝐹.
Î
(a) Let 𝔇 := 𝔭∈Σ 𝔭 be the product of the primes corresponding to nonar-
chimedean places in S. Using weak approximation (see Lemma 28.7.1),
show there exists 𝑎 ∈ 𝔇 such that:
• 𝑣(𝑎) < 0 for all real places 𝑣 ∈ Σ and 𝑣(𝑎) > 0 for all real places
𝑣 ∉ Σ, if there are any; and
• 𝑎𝑅 = 𝔇𝔟 with 𝔇 + 𝔟 = 𝑅 and 2𝑅 + 𝔟 = 𝑅.
In the special case where 𝑅 has narrow class number 1 (that is, every ideal
𝔞 ⊆ 𝑅 is principal 𝔞 = (𝑎) and generated by an element 𝑎 ∈ 𝑅 such that
𝑣(𝑎) > 0 for every real places 𝑣), show that we may take (𝑎) = 𝔇 and
𝔟 = 𝑅.
(b) Show that there exists 𝑡 ∈ 𝑅 coprime to 8𝑎𝑅 such that the following hold:

𝑡
• For all primes 𝔭 | 𝔇 with 𝔭 - 2𝑅, we have = −1;
𝔭 √
• For all primes 𝔭 | 𝔇 with 𝔭 | 2𝑅, the extension 𝐹𝔭 ( 𝑡) is the quadratic
𝑡
unramified extension of 𝐹𝔭 , so = −1 in the sense of the general-
𝔭
ized Kronecker symbol;
𝑡
• For all primes 𝔮 | 𝔟 we have = 1; and
𝔮
• For all prime powers 𝔯 k 8𝑅 with 𝔯 - 𝐷, we have 𝑡 ≡ 1 (mod 𝔯𝑒 ).
𝑒
Show that 𝑡 is well-defined as an element of (𝑅/8𝑎𝑅) × , i.e., if 𝑡 0 ≡ 𝑡

(mod 8𝑎) then 𝑡 0 also satisfies these conditions.
(c) Using the infinitude of primes in arithmetic progression over number fields
(Theorem 26.8.26), show there exists 𝑞 ∈ 𝑅 a prime element (i.e., 𝑞𝑅 is
a prime ideal) such that 𝑞 ≡ 𝑡 (mod 8𝑎) with 𝑡 as in (b) and further
satisfying 𝑣(𝑞) < 0 for all real places 𝑣 ∈ Σ.
𝑎, 𝑞
(d) Show that 𝐵 := has Ram 𝐵 = Σ.
𝐹
17. Let 𝐹 be a global field. Show that two quaternion algebras 𝐵, 𝐵 0 over 𝐹 are
isomorphic if and only if they have the same quadratic subfields (for a quadratic
extension 𝐾 ⊃ 𝐹, we have 𝐾 ↩→ 𝐵 if and only if 𝐾 ↩→ 𝐵 0).
[See work of Garibaldi–Saltman [GS2010] for a discussion of the fields 𝐹 with
char 𝐹 ≠ 2 and the property that two division quaternion algebras over 𝐹 with
the same subfields are necessarily isomorphic. (Roughly speaking, they are the
fields for which nonzero 2-torsion elements of the Brauer group can be detected
using ramification.)]
18. In this exercise, we consider how ramification sets change under base extension.
Let 𝐹 be a global field and let 𝐾 ⊇ 𝐹 be a finite separable extension.
(a) Let 𝐵 be a quaternion algebra over 𝐹 with ramification set Ram 𝐵 and
consider 𝐵 𝐾 = 𝐵 ⊗𝐹 𝐾. Show that
Ram(𝐵 𝐾 ) = {𝑤 ∈ Pl(𝐾) : 𝑤 lies over 𝑣 ∈ Ram 𝐵 and 2 - [𝐾 𝑤 : 𝐹𝑣 ]}.
(b) Suppose 𝐵 is a division algebra and [𝐾 : 𝐹] is odd. Show that 𝐵 𝐾 is a
division algebra.
(c) As a converse to (a), suppose that Σ𝐾 ⊆ Pl(𝐾) is a finite subset of

noncomplex places of 𝐾 of even cardinality with the property that if
𝑤 ∈ Σ𝐾 lies over 𝑣 ∈ Pl 𝐹, then [𝐾 𝑤 : 𝐹𝑣 ] is odd and moreover
{𝑤 ∈ Pl 𝐹 : 𝑤 lies over 𝑣 and [𝐾 𝑤 : 𝐹𝑣 ] is odd} ⊆ Σ𝐾 .
Show that there exists a quaternion algebra 𝐵 over 𝐹 with the property that
Ram(𝐵 𝐾 ) = Σ𝐾 . (We say that the quaternion algebra associated to the set
Σ𝐾 descends to 𝐹.)
(d) As a special case, what do (a) and (c) say when [𝐾 : 𝐹] = 2?
(e) Restate (a) and (b) in terms of the kernel of the map Br(𝐹) [2] → Br(𝐾) [2]
induced by [𝐵] ↦→ [𝐵 𝐾 ] (see Remark 14.6.10).
19. Let 𝑅 be a global ring with 𝐹 = Frac 𝑅, and let 𝐾 ⊇ 𝐹 be a finite Galois
extension with 𝑆 the integral closure of 𝑅 in 𝐾. Let 𝐵 be a quaternion algebra
over 𝐹 and consider 𝐵 𝐾 = 𝐵 ⊗𝐹 𝐾. Then Gal(𝐾 | 𝐹) acts naturally on 𝐵 𝐾 via
𝜎(𝛼 ⊗ 𝑥) = 𝛼 ⊗ 𝜎(𝑥). (This action is not by 𝐾-algebra isomorphism!)
Show that there exists a maximal 𝑆-order O ⊆ 𝐵 𝐾 stable under Gal(𝐾 | 𝐹), i.e.,
𝜎(O) = O for all 𝜎 ∈ Gal(𝐾 | 𝐹).
I 20. Let 𝐹 be a global field, let 𝑣 1 , . . . , 𝑣 𝑟 be places of 𝐹, and for each 𝑣 𝑖 suppose
we are given the condition ramified, split, or inert. Show that there exists a
separable quadratic extension 𝐾 ⊇ 𝐹 that 𝐾 𝑣𝑖 satisfies the given condition for
each 𝑖. [Hint: follow the proof of Main Theorem 14.7.4.]
21. Let 𝐹 be a global field, let 𝐵1 , 𝐵2 , . . . , 𝐵𝑟 be quaternion algebras over 𝐹, and let
𝐵 := 𝐵1 ⊗ 𝐵2 ⊗ · · · ⊗ 𝐵𝑟 . Recalling section 8.2, show (in as many ways as you
can) that 𝐵 ' M2𝑟−1 (𝐵 0) for a quaternion algebra 𝐵 0 over 𝐹. (Recalling 8.3, by
Merkurjev’s theorem this shows the class of every element in the 2-torsion of
the Brauer group Br(𝐹) [2] is represented by a quaternion algebra.)
I 22. Let 𝐹𝑣 be a local field with char 𝐹𝑣 = 2. Let 𝑛 ∈ 𝐹𝑣 . Show that there exists
𝑡 ∈ 𝐹𝑣 such that 𝑥 2 − 𝑡𝑥 + 𝑛 is separable and irreducible.
I 23. Prove Lemma 14.7.6 for 𝑣 an archimedean place.
24. In this advanced exercise following up on Exercise 9.15, we consider features of
quaternion algebras and orders in the case of a global function field, assuming
background in algebraic geometry.
Let 𝑋 be a smooth, projective, geometrically integral curve over a finite field 𝑘;
then 𝑋 is a separated, integral Dedekind scheme. Let 𝒪𝑋 be its structure sheaf.
Let 𝐹 be its function field, and let 𝐵 be a quaternion algebra over 𝐹. Define a
sheaf of 𝒪𝑋 -orders in 𝐵, or simply an 𝒪𝑋 -order in 𝐵, to be an 𝒪𝑋 -lattice ℬ in
𝐵 such that for each open set 𝑈 ⊆ 𝑋, the 𝒪𝑋 (𝑈)-lattice ℬ(𝑈) is a subring of 𝐵.
We recall the local-global dictionary for 𝒪𝑋 -lattices (Exercise 9.15(c)).
In parts (a)–(c) we work out an example: let 𝑋 = P1 with function field 𝐹 = 𝑘 (𝑡)
where div 𝑡 = (0) − (∞). Suppose that char 𝑘 ≠ 2, and let 𝑢 ∈ 𝑘 × r 𝑘 ×2 . Let 𝐵
be the quaternion algebra with Ram(𝐵) = {(0), (∞)}.
(a) Show that 𝐵 = (𝑢, 𝑡 | 𝐹).
(b) Let 𝑈 = Spec 𝑘 [𝑡] = 𝑋 r {∞}. Show that there exists a unique 𝒪𝑋 -order ℬ
in 𝐵 with ℬ(𝑈) = 𝑘 [𝑡] + 𝑘 [𝑡]𝑖 + 𝑘 [𝑡] 𝑗 + 𝑘 [𝑡]𝑖 𝑗 and stalk ℬ(∞) a maximal
𝒪𝑋 , (∞) -order. Describe explicitly ℬ(∞) and ℬ(Spec 𝑘 [1/𝑡]) as orders in
𝐵.
(c) With ℬ from (b), show that ℬ(𝑋) = 𝑘 [𝑖].
Restoring generality, let ℬ be an 𝒪𝑋 -order such that ℬ(𝑈) is a maximal 𝒪𝑋 (𝑈)-
order in 𝐵 for all affine open sets 𝑈.
(d) Show that ℬ(𝑋) has a zero divisor if and only if ℬ(𝑋) ' M2 (𝑘) if and
only if 𝐵 ' M2 (𝐹).
(e) Show that ℬ(𝑋) is a 𝑘-algebra with a nondegenerate standard involution.
(f) Suppose that 𝐵 is a division algebra. Show that either ℬ(𝑋) = 𝑘 or
ℬ(𝑋) ' 𝑘 2 is the quadratic extension of 𝑘.
(g) Still supposing that 𝐵 is a division algebra, show that if ℬ(𝑋) = 𝑘 2 , then
every ramified place of 𝐵 has odd degree. [Hint: show that 𝐵 ' (𝐾, 𝑏 | 𝐹)
where 𝐾 = 𝐹 𝑘 2 is the constant field extension of 𝐹 of degree 2, and
𝑏 ∈ 𝐹 × r 𝑘 × . Compute the Hilbert symbol at 𝑣 ∈ Ram(𝐵) to show 𝑣(𝑏) is
odd.]
Chapter 15
Discriminants
Discriminants measure volume and arithmetic complexity, and they simultaneously

encode ramification. We devote this chapter to their study.
15.1 ⊲ Discriminantal notions

Let 𝑥1 , . . . , 𝑥 𝑛 ∈ R𝑛 , and let 𝐴 be the matrix with columns 𝑥 𝑖 . Then the parallelopiped
with edges from the origin to 𝑥𝑖 has volume |det( 𝐴)|. We can compute this volume in
another way:
det( 𝐴) 2 = det( 𝐴t 𝐴) = det(𝑀) (15.1.1)
where 𝑀 has 𝑖 𝑗th entry equal to the ordinary dot product 𝑥𝑖 · 𝑥 𝑗 .
The absolute discriminant of a number field is a volume and a measure of arithmetic
𝑛 is a Z-basis for Z𝐹 and 𝜄 : 𝐹 ↩→ 𝐹 ⊗Q R ' R
complexity, as follows. If 𝑥1 , . . . , 𝑥√ 𝑛
(normalized with an extra factor of 2 at the complex places), then the volume of Z𝐹
in this embedding is the absolute determinant of the matrix with columns 𝜄(𝑥𝑖 ), and
its square is defined to be the absolute discriminant of 𝐹. Replacing the dot product
in the definition of 𝑀 in (15.1.1) with the trace form (𝑥, 𝑦) ↦→ Tr𝐹 /Q (𝑥𝑦), we see that
the absolute discriminant is a positive integer. A prime 𝑝 is ramified in 𝐹 if and only
if it divides the discriminant, so this volume also records arithmetic properties of 𝐹.
More generally, whenever we have a symmetric bilinear form 𝑇 : 𝑉 × 𝑉 → 𝐹 on
a finite-dimensional 𝐹-vector space 𝑉, there is a volume defined by the determinant
det(𝑇 (𝑥𝑖 , 𝑥 𝑗 ))𝑖, 𝑗 : and when 𝑇 arises from a quadratic form 𝑄, this is volume is the
discriminant of 𝑄 (up to a normalizing factor of 2 in odd degree, see 6.3.1). In
particular, if 𝐵 is a finite-dimensional algebra over 𝐹, there is a bilinear form
𝐵×𝐵→𝐹
(𝛼, 𝛽) ↦→ Tr 𝐵 |𝐹 (𝛼𝛽)
(or, when 𝐵 is semisimple, the bilinear form associated to the reduced trace trd)
and so we obtain a discriminant—a “squared” volume—measuring in some way the
complexity of 𝐵. As in the commutative case, discriminants encode ramification.
In this chapter, we establish basic facts about discriminants, including how they
behave under inclusion (measuring index) and localization. To illustrate, let 𝐵 be a
233
234 CHAPTER 15. DISCRIMINANTS
quaternion algebra over Q and let O ⊂ 𝐵 be an order. We define the discriminant of

O to be
disc(O) := |det(trd(𝛼𝑖 𝛼 𝑗 ))𝑖, 𝑗 | ∈ Z>0 (15.1.2)

𝑎, 𝑏
where 𝛼1 , . . . , 𝛼4 is a Z-basis for O. For example, if 𝐵 = with 𝑎, 𝑏 ∈ Z r {0},
Q
then the standard order O = Z + Z𝑖 + Z 𝑗 + Z𝑘 has
disc(O) = (4𝑎𝑏) 2 ;
indeed, this is the discriminant of the quadratic form h1, −𝑎, −𝑏, 𝑎𝑏i, the reduced norm
restricted to O. If 𝑎, 𝑏 < 0, i.e. 𝐵 is definite, then the reduced√ norm is a Euclidean
norm on 𝐵∞ = 𝐵 ⊗Q R ' H; normalizing with an extra factor 2, the discriminant is
square of the covolume of the lattice O ⊂ 𝐵∞ . For example, the Lipschitz√order Zh𝑖, 𝑗i
(11.1.1) has disc(Zh𝑖, 𝑗i) = 42 , the square of the covolume of the lattice ( 2Z) 4 ⊆ R4 .
If O0 ⊇ O, then disc(O) = [O0 : O] 2 disc(O0); in particular O0 = O if and only if
disc(O0) = disc(O). It follows that the discriminant of an order is always a square, so
we define the reduced discriminant discrd(O) to be the positive integer square root,
and discrd(O) 2 = disc(O). The discriminant of an order measures how far the order
is from being a maximal order. We will show (Theorem 15.5.5) that O is a maximal
order if and only if discrd(O) = disc 𝐵, where disc 𝐵 is the (squarefree) product of
primes ramified in 𝐵.
In an extension of Dedekind domains, the different of the extension is an ideal
whose norm is the discriminant of the extension (see Neukirch [Neu99, §III.2]).
The different is perhaps not as popular as its discriminant cousin, but it has many
nice properties, including easy-to-understand behavior under base extension. Similar
conclusions holds in the noncommutative context (presented in section 15.6).
15.2 Discriminant
For further reference on discriminants, see Reiner [Rei2003, §10, §14].
Let 𝑅 be a noetherian domain and let 𝐹 = Frac 𝑅. Let 𝐵 be a semisimple algebra
over 𝐹 with dim𝐹 𝐵 = 𝑛. For elements 𝛼1 , . . . , 𝛼𝑛 ∈ 𝐵, we define
𝑑 (𝛼1 , . . . , 𝛼𝑛 ) := det(trd(𝛼𝑖 𝛼 𝑗 ))𝑖, 𝑗=1,...,𝑛 . (15.2.1)
Let 𝐼 ⊆ 𝐵 be an 𝑅-lattice.
Definition 15.2.2. The discriminant of 𝐼 is the 𝑅-submodule disc(𝐼) ⊆ 𝐹 generated

by the set
{𝑑 (𝛼1 , . . . , 𝛼𝑛 ) : 𝛼1 , . . . , 𝛼𝑛 ∈ 𝐼}.
15.2.3. If 𝐼 = O, then for 𝛼1 , . . . , 𝛼𝑛 ∈ O we have 𝛼𝑖 𝛼 𝑗 ∈ O and so trd(𝛼𝑖 𝛼 𝑗 ) ∈ 𝑅 for

all 𝑖, 𝑗. Thus 𝑑 (𝛼1 , . . . , 𝛼𝑛 ) ∈ 𝑅 and therefore disc(O) ⊆ 𝑅.
15.2. DISCRIMINANT 235
Remark 15.2.4. When working over Z, it is common to take the discriminant instead to
be the positive generator of the discriminant as an ideal; passing between these should
cause no confusion.
Although Definition 15.2.2 may look unwieldly, it works as well in the commutative
case as in the noncommutative case. Right away, we see that if O ⊆ O0 are 𝑅-orders,
then disc(O0) | disc(O).
The function 𝑑 itself transforms in a nice way under a change of basis, as follows.
Í 15.2.5. Let 𝛼1 , . . . , 𝛼𝑛 ∈ 𝐵 and suppose 𝛽1 , . . . , 𝛽𝑛 ∈ 𝐵 are of the form

Lemma
𝛽𝑖 = 𝑛𝑗=1 𝑚 𝑖 𝑗 𝛼 𝑗 with 𝑚 𝑖 𝑗 ∈ 𝐹. Let 𝑀 = (𝑚 𝑖 𝑗 )𝑖, 𝑗=1,...,𝑛 . Then
𝑑 (𝛽1 , . . . , 𝛽𝑛 ) = det(𝑀) 2 𝑑 (𝛼1 , . . . , 𝛼𝑛 ). (15.2.6)
Proof. By properties of determinants, if 𝛽1 , . . . , 𝛽𝑛 are linearly dependent (over 𝐹)

then 𝑑 (𝛽1 , . . . , 𝛽𝑛 ) = 0 and either 𝛼1 , . . . , 𝛼𝑛 are also linearly dependent or det(𝑀) =
0, and in either case the equality (15.2.6) holds trivially.
So suppose that 𝛽1 , . . . , 𝛽𝑛 are linearly independent, then 𝛼1 , . . . , 𝛼𝑛 are also
linearly independent and the matrix 𝑀, a change of basis matrix, is invertible. By
Gaussian reduction, we can write 𝑀 as a product of elementary matrices (a matrix that
coincides with the identity matrix except for a single off-diagonal entry), permutation
matrices (a matrix interchanging rows suffices), and a diagonal matrix; it is enough
to check that the equality holds when 𝑀 is a matrix of one of these forms. And
for such a matrix, the equality can be checked in a straightforward manner using the
corresponding property of determinants.
Corollary 15.2.7. If 𝐼 is free as an 𝑅-module, and 𝛼1 , . . . , 𝛼𝑛 is an 𝑅-basis for 𝐼, then
disc(𝐼) = 𝑑 (𝛼1 , . . . , 𝛼𝑛 )𝑅.
Proof. The matrix 𝑀 writing any other 𝛽1 , . . . , 𝛽𝑛 ∈ 𝐼 in terms of the basis has
𝑀 ∈ M𝑛 (𝑅) so det(𝑀) ∈ 𝑅, and therefore 𝑑 (𝛽1 , . . . , 𝛽𝑛 ) ∈ 𝑑 (𝛼1 , . . . , 𝛼𝑛 )𝑅 by
Lemma 15.2.5.
15.2.8. More generally, if 𝐼 is completely decomposable with
𝐼 = 𝔞1 𝛼1 ⊕ · · · ⊕ 𝔞𝑛 𝛼𝑛
such as in (9.3.7), then from (15.2.6)
disc(𝐼) = (𝔞1 · · · 𝔞𝑛 ) 2 𝑑 (𝛼1 , . . . , 𝛼𝑛 ).
More generally, the discriminant is well-behaved under automorphisms because

the reduced trace is so.
∼ 𝐵 is an 𝐹-algebra automorphism, then disc(𝜙(𝐼)) =
Corollary 15.2.9. If 𝜙 : 𝐵 −→
disc(𝐼).
Proof. By Proposition 7.8.6, we have trd(𝜙(𝛼𝛽)) = trd(𝛼𝛽) for all 𝛼, 𝛽 ∈ 𝐵. There-
fore, for all 𝛼1 , . . . , 𝛼𝑛 ∈ 𝐵 we have 𝑑 (𝜙(𝛼1 ), . . . , 𝜙(𝛼𝑛 )) = 𝑑 (𝛼1 , . . . , 𝛼𝑛 ); the result
disc(𝜙(𝐼)) = disc(𝐼) follows.
Our primary interest will be in the case 𝐼 = O.
Example 15.2.10. Suppose char 𝐹 ≠ 2. Let 𝐵 := (𝑎, 𝑏 | 𝐹) with 𝑎, 𝑏 ∈ 𝑅. Let

O := 𝑅 ⊕ 𝑅𝑖 ⊕ 𝑅 𝑗 ⊕ 𝑅𝑖 𝑗 be the standard order. Then disc(O) is the principal 𝑅-ideal
generated by
2 0 0 0
0 2𝑎 0 0 ®
© ª
𝑑 (1, 𝑖, 𝑗, 𝑖 𝑗) = det ® = −(4𝑎𝑏) 2 .
0 0 2𝑏 0 ®
«0 0 0 −2𝑎𝑏 ¬
The calculation when char 𝐹 = 2 is requested in Exercise 15.1.
Example 15.2.11. Let 𝐵 := M𝑛 (𝐹) and O := M𝑛 (𝑅). Then disc(O) = 𝑅 (Exercise

15.2).
15.2.12. Let 𝐵 := (𝐾, 𝑏 | 𝐹) be a quaternion algebra over 𝐹 with 𝑏 ∈ 𝑅 and let 𝑆 be

an 𝑅-order in 𝐾. Let O := 𝑆 ⊕ 𝑆 𝑗; then O is an 𝑅-order in 𝐵 by Exercise 10.7. We
have disc(O) = 𝑏 2 disc(𝑆) 2 , by Exercise 15.4.
In particular, let 𝐹 be a nonarchimedean local field, let 𝑅 be its valuation ring and
𝔭 = 𝑅𝜋 its maximal ideal, and let 𝐵 be a division quaternion algebra over 𝐹. Then
by Theorem 13.3.11, we have 𝐵 ' (𝐾, 𝜋 | 𝐹) with 𝐾 ⊇ 𝐹 an unramified separable
quadratic extension of 𝐹. The valuation ring 𝑆 of 𝐾 has disc(𝑆) = 𝑅, so the valuation
ring O = 𝑆 ⊕ 𝑆 𝑗 of 𝐵 has discriminant disc(O) = 𝔭2 .
15.2.13. Equation (15.2.6) and the fact that 𝐼 (𝔭) = 𝐼 ⊗𝑅 𝑅 (𝔭) implies the equality
disc(𝐼 (𝔭) ) = disc(𝐼) (𝔭)
on localizations and for the same reason an equality for the completions disc(𝐼𝔭 ) =
disc(𝐼)𝔭 . In other words, the discriminant respects localization and completion and
can be computed locally. Therefore, by the local-global principle (Lemma 9.4.6),
Ù
disc(𝐼) = disc(𝐼 (𝔭) ).
𝔭
Lemma 15.2.14. If 𝐵 is separable as an 𝐹-algebra and 𝐼 is projective as an 𝑅-module,

then disc(𝐼) is a nonzero projective fractional ideal of 𝑅.
Proof. Since 𝐼 is an 𝑅-lattice, there exist elements 𝛼1 , . . . , 𝛼𝑛 which are linearly

independent over 𝐹. Since 𝐵 is separable, by Theorem 7.9.4, trd is a nondegenerate
bilinear pairing on 𝐵 so disc(𝐼) is a nonzero ideal of 𝑅. It follows from Lemma 15.2.5
that disc(𝐼) is finitely generated as an 𝑅-module, since this is true of 𝐼: we apply 𝑑
to all subsets of a set of generators for 𝐼 as an 𝑅-module. To show that disc(𝐼) is
projective, by 9.2.1 we show that disc(𝐼) is locally principal. Let 𝔭 be a prime ideal of
𝑅. Since 𝐼 is a projective 𝑅-module, its localization 𝐼 (𝔭) is free; thus from Corollary
15.2.7, we conclude that disc(𝐼) (𝔭) = disc(𝐼 (𝔭) ) is principal over 𝑅 (𝔭) and generated
by disc(𝛼1 , . . . , 𝛼𝑛 ) for an 𝑅 (𝔭) -basis 𝛼1 , . . . , 𝛼𝑛 of 𝐼, as desired.
We conclude this section comparing lattices by their index and discriminant as

follows. We recall the definition of index (section 9.6).
Lemma 15.2.15. Let 𝐼, 𝐽 ⊆ 𝐵 be projective 𝑅-lattices. Then
disc(𝐼) = [𝐽 : 𝐼] 2𝑅 disc(𝐽).
Moreover, if 𝐼 ⊆ 𝐽, then disc(𝐼) = disc(𝐽) if and only if 𝐼 = 𝐽.
Proof. For the first statement, we argue locally, and combine (15.2.6) and Lemma 9.6.4.
For the second statement, clearly disc(𝐽) ⊆ disc(𝐼), and if 𝐼 = 𝐽 then equality holds;
and conversely, from disc(𝐼) = [𝐽 : 𝐼] 2𝑅 disc(𝐽) = disc(𝐽) we conclude [𝐽 : 𝐼] 𝑅 = 𝑅,
hence 𝐽 = 𝐼 by Proposition 9.6.8.
Remark 15.2.16. We defined the discriminant for semisimple algebras so that it is given
in terms of the reduced trace. This definition extends to an arbitrary finite-dimensional
𝐹-algebra 𝐵, replacing the reduced trace by the algebra trace Tr 𝐵 |𝐹 . If 𝐵 is a central
simple 𝐹-algebra of dimension 𝑛2 , then 𝑛 trd = Tr 𝐵 |𝐹 so when 𝑛 ∈ 𝐹 × one can recover
the discriminant as we have defined it here from the more general definition; but if
𝑛 = 0 ∈ 𝐹 then the discriminant of 𝐵 computed with the algebra trace will be zero.
15.3 Quadratic forms

Essentially the same definition of discriminant (Definition 15.2.2) applies to quadratic
modules, as follows. We recall 6.3.1, where the discriminant was defined in all
characteristics.
Let 𝑄 : 𝑀 → 𝐿 be a quadratic module over 𝑅 (Definition 9.7.3) with rk 𝑀 = 𝑛
and associated bilinear map 𝑇 : 𝑀 × 𝑀 → 𝐿.
15.3.1. Let 𝑥 1 , . . . , 𝑥 𝑛 ∈ 𝑀 and 𝑓 ∈ 𝐿 ∨ := Hom𝑅 (𝐿, 𝑅). If 𝑛 is even, we define
𝑑 (𝑥1 , . . . , 𝑥 𝑛 ; 𝑓 ) := det( 𝑓 (𝑇 (𝑥𝑖 , 𝑥 𝑗 )))𝑖, 𝑗=1,...,𝑛 . (15.3.2)
If 𝑛 is odd, then by specializing the universal determinant as in 6.3.4, we define
𝑑 (𝑥1 , . . . , 𝑥 𝑛 ; 𝑓 ) := (det/2) ( 𝑓 (𝑇 (𝑥𝑖 , 𝑥 𝑗 )))𝑖, 𝑗=1,...,𝑛 . (15.3.3)
The discriminant of 𝑄 is then the ideal disc(𝑄) ⊆ 𝑅 generated by the set
{𝑑 (𝑥1 , . . . , 𝑥 𝑛 ; 𝑓 ) : 𝑥1 , . . . , 𝑥 𝑛 ∈ 𝑀, 𝑓 ∈ 𝐿 ∨ }. (15.3.4)
15.3.5. If 𝑀, 𝐿 are free with 𝑅-basis 𝑥 1 , . . . , 𝑥 𝑛 and 𝑒, respectively, then letting 𝑓 ∈ 𝐿 ∨

the dual to 𝑒 with 𝑓 (𝑒) = 1 gives
disc(𝑄) = 𝑑 (𝑥1 , . . . , 𝑥 𝑛 ; 𝑓 )𝑅.
In particular, since 𝑀, 𝐿 are projective and therefore locally free over 𝑅, the discrimi-
nant of 𝑄 is locally free and hence a projective 𝑅-ideal.
Lemma 15.3.6. The discriminant of a quadratic module is well-defined up to similarity.
Proof. Let 𝑄 : 𝑀 → 𝐿 and 𝑄 0 : 𝑀 0 → 𝐿 0 be quadratic modules over 𝑅 similar by

𝑔 : 𝑀 −→∼ 𝑀 0 and ℎ : 𝐿 −→ ∼ 𝐿 0. It suffices to check the invariance locally, so to this end
Í𝑛
we may suppose that the modules are free; choose Í𝑛a basis 𝑀 = 𝑖=1 𝑅𝑥𝑖 and 𝐿 = 𝑅𝑒,
0
and let 𝑥𝑖 = 𝑔(𝑥𝑖 ) and 𝑒 = ℎ(𝑒). Then 𝑀 = 𝑖=1 𝑅𝑥𝑖 and 𝐿 0 = 𝑅𝑒 0. Let 𝑓 , 𝑓 0 be
0 0 0
dual to 𝑒, 𝑒 0; then postcomposing 𝑄 and 𝑄 0 by 𝑓 , 𝑓 0 we may suppose 𝐿 = 𝐿 0 = 𝑅 and

ℎ is the identity.
We then have 𝑄 0 (𝑔(𝑥)) = 𝑄(𝑥) for all 𝑥 ∈ 𝑀, so the same is true of the associated
bilinear forms 𝑇, 𝑇 0. But then 𝑑 (𝑥 10 , . . . , 𝑥 𝑛0 ) = 𝑑 (𝑥1 , . . . , 𝑥 𝑛 ), and by 15.3.5 this
implies disc(𝑄) = disc(𝑄 0) as ideals of 𝑅.
15.3.7. Let 𝐵 be a finite-dimensional 𝐹-algebra with a standard involution. Then the

reduced norm is a quadratic form on 𝐵 with associated bilinear form 𝑇 (𝛼, 𝛽) = trd(𝛼𝛽).
Although the bilinear form differs by the presence of this standard involution from the
definition of discriminant in (15.2.1), the resulting discriminants are the same (up to
𝑅 × ): see Exercise 15.13.
Lemma 15.3.8. The quadratic module 𝑄 is nonsingular if and only if disc(𝑄) = 𝐿.

In particular, suppose that 𝑀 ' 𝑅 𝑛 is free with basis 𝑒 𝑖 and 𝐿 = 𝑅, and let
[𝑇] := (𝑇 (𝑒 𝑖 , 𝑒 𝑗 ))𝑖, 𝑗 ∈ M𝑛 (𝑅) be the Gram matrix in this basis. Then 𝑄 is nonsingular
if and only if det( [𝑇]), (det/2) ( [𝑇]) ∈ 𝑅 × according as 𝑛 is even or odd.
Proof. The map 𝑇 : 𝑀 → Hom𝑅 (𝑀, 𝐿) is an isomorphism if and only if it is an

isomorphism in every localization, so we may suppose that 𝑄 is free, with 𝑀 = 𝑅 𝑛
and 𝐿 = 𝑅, which is to say we may prove the second statement in the case where 𝑅 is
local, with maximal ideal 𝔭 and residue field 𝑘 := 𝑅/𝔭. Let 𝑄 mod 𝔭 : 𝑀 ⊗𝑅 𝑘 → 𝑘
be the reduction of 𝑄; its Gram matrix is [𝑇] mod 𝔭 ∈ M𝑛 (𝑘). Over the field 𝑘,
we have that 𝑄 mod 𝔭 is nonsingular if and only if it is nondegenerate if and only if
det[𝑇], (det/2) ( [𝑇]) ≠ 0 according as 𝑛 is even or odd; since 𝑅 is local, these are
equivalent to asking that these values are in 𝑅 × . An application of Nakayama’s lemma
then implies the result.
15.4 Reduced discriminant

In this section, we extract a square root of the discriminant for quaternion orders.
Indeed, in Example 15.2.10, we saw that the discriminant of the standard 𝑅-order
O ⊆ 𝐵 = (𝑎, 𝑏 | 𝐹) is disc(O) = (4𝑎𝑏) 2 𝑅, a square. If O0 is another projective
𝑅-order, then disc(O0) = [O : O0] 2𝑅 disc(O) by Lemma 15.2.15, so in fact the
discriminant of every 𝑅-order is the square of an 𝑅-ideal.
In fact, there is a way to define this square root directly, inspired by vector calculus.
15.4.1. If 𝑢, 𝑣, 𝑤 ∈ R3 then |𝑢 · (𝑣 × 𝑤)|, the absolute value of the so-called mixed

product (or scalar triple product or box product), is the volume of the parallelopiped
defined by 𝑢, 𝑣, 𝑤; identifying R3 ' H0 as in section 2.4, from (2.4.10) we can write
2𝑢 · (𝑣 × 𝑤) = 𝑢 · (𝑣𝑤 − 𝑤𝑣) = − trd(𝑢(𝑣𝑤 − 𝑤𝑣)).

15.4. REDUCED DISCRIMINANT 239
For example, 2 = −2𝑖 · ( 𝑗 × 𝑘) = −𝑖 · ( 𝑗 𝑘 − 𝑘 𝑗) = trd(𝑖 𝑗 𝑘).

More generally (and carefully attending to the factors of 2) we make the following
definition. Let 𝐵 be a quaternion algebra over 𝐹.
15.4.2. For 𝛼1 , 𝛼2 , 𝛼3 ∈ 𝐵, we define
𝑚(𝛼1 , 𝛼2 , 𝛼3 ) := trd((𝛼1 𝛼2 − 𝛼2 𝛼1 )𝛼3 )
= 𝛼1 𝛼2 𝛼3 − 𝛼2 𝛼1 𝛼3 − 𝛼3 𝛼2 𝛼1 + 𝛼3 𝛼1 𝛼2 .
Lemma 15.4.3. The form 𝑚 : 𝐵 × 𝐵 × 𝐵 → 𝐹 is an alternating trilinear form which

is well-defined as a form on 𝐵/𝐹.
Proof. The form is alternating because for all 𝛼1 , 𝛼2 ∈ 𝐵 we have 𝑚(𝛼1 , 𝛼1 , 𝛼2 ) = 0
and
𝑚(𝛼1 , 𝛼2 , 𝛼1 ) = trd((𝛼1 𝛼2 − 𝛼2 𝛼1 )𝛼1 ) = trd(nrd(𝛼1 )𝛼2 ) − trd(𝛼2 nrd(𝛼1 )) = 0
and similarly 𝑚(𝛼1 , 𝛼2 , 𝛼2 ) = 0. The trilinearity follows from the linearity of the
reduced trace. Finally, from these two properties, the descent to 𝐵/𝐹 follows from the
computation 𝑚(1, 𝛼1 , 𝛼2 ) = 0 for all 𝛼1 , 𝛼2 ∈ 𝐵.
(Alternatively, one can check that the pairing descends to 𝐵/𝐹 first, so that the
involution becomes 𝛼 + 𝐹 = −𝛼 + 𝐹, and then the alternating condition is immediate.)

Definition 15.4.4. Let 𝐼 ⊆ 𝐵 be an 𝑅-lattice. The reduced discriminant of 𝐼 is the
𝑅-submodule discrd(𝐼) of 𝐹 generated by
{𝑚(𝛼1 , 𝛼2 , 𝛼3 ) : 𝛼1 , 𝛼2 , 𝛼3 ∈ 𝐼}.
15.4.5. If 𝛼𝑖 , 𝛽𝑖 ∈ 𝐵 with 𝛽𝑖 = 𝑀𝛼𝑖 for some 𝑀 ∈ M3 (𝐹), then
𝑚(𝛽1 , 𝛽2 , 𝛽3 ) = det(𝑀)𝑚(𝛼1 , 𝛼2 , 𝛼3 ) (15.4.6)
by Exercise 15.10. It follows that if 𝐼 ⊆ 𝐽 are projective 𝑅-lattices in 𝐵, then
discrd(𝐼) = [𝐽 : 𝐼] discrd(𝐽).
Lemma 15.4.7. If 𝐼 is a projective 𝑅-lattice in 𝐵, then disc(𝐼) = discrd(𝐼) 2 .
Proof. First, we claim that
𝑚(𝑖, 𝑗, 𝑖 𝑗) 2 = −𝑑 (1, 𝑖, 𝑗, 𝑖 𝑗).
If char 𝐹 ≠ 2, then disc(1, 𝑖, 𝑗, 𝑖 𝑗) = −(4𝑎𝑏) 2 by Example 15.2.10 and
𝑚(𝑖, 𝑗, 𝑖 𝑗) = trd((𝑖 𝑗 − 𝑗𝑖)𝑖 𝑗) = trd(2𝑖 𝑗 (𝑖 𝑗)) = 4𝑎𝑏,
as claimed. See Exercise 15.1 for the case char 𝐹 = 2. This computation verifies the
result for the order O = 𝑅 ⊕ 𝑅𝑖 ⊕ 𝑅 𝑗 ⊕ 𝑅𝑖 𝑗.
The lemma now follows using (15.2.6) and (15.4.6), for it shows that
𝑚(𝛼1 , 𝛼2 , 𝛼3 ) 2 = −𝑑 (1, 𝛼1 , 𝛼2 , 𝛼3 )
for all 𝛼1 , 𝛼2 , 𝛼3 ∈ 𝐵, and the latter generate discrd(𝐼) by Exercise 15.7.
The notions in this section extend more generally to an arbitrary algebra 𝐵 with a
standard involution.
15.5 Maximal orders and discriminants

We now relate discriminants to maximal orders. Throughout this section, we suppose
that 𝑅 is a Dedekind domain. We record the following important principle.
Lemma 15.5.1. Let O ⊆ O0 be 𝑅-orders. Then O = O0 if and only if disc O = disc O0.
Proof. In the nontrivial direction, by Lemma 15.2.15 we have
disc O = [O0 : O] 2𝑅 disc(O0)
so disc O = disc O0 if and only if O = O0.
First, we ensure the existence of maximal orders (cf. 10.4.2) using the discriminant.
Proposition 15.5.2. There exists a maximal 𝑅-order O ⊆ 𝐵, and every order O is

contained in a maximal 𝑅-order O0 ⊆ 𝐵.
Proof. The algebra 𝐵 has at least one 𝑅-order O as the left- or right-order of a lattice
10.2.5. If O is not maximal, then there exists an order O0 ) O with disc(O0) ) disc(O)
by Lemma 15.5.1. If O0 is maximal, we are done; otherwise, we can continue in
this way to obtain orders O = O1 ( O2 ( . . . and an ascending chain of ideals
disc(O1 ) ( disc(O2 ) ( . . . of 𝑅; but since 𝑅 is noetherian, the latter stabilizes after
finitely many steps, and the resulting order is then maximal, by Lemma 15.2.15.
Using the discriminant as a measure of index, we can similarly detect when orders
are maximal. We recall (10.4.3) that the property of being maximal is a local property,
so we begin with the local matrix case.
Lemma 15.5.3. Suppose that 𝑅 is a DVR, and let O ⊆ 𝐵 := M𝑛 (𝐹) be an 𝑅-order.

Then O is maximal if and only if disc O = 𝑅.
Proof. First, suppose O is maximal. Then by Corollary 10.5.5, we conclude O '

M𝑛 (𝑅) (conjugate in 𝐵). By Corollary 15.2.9, we have disc O = disc M𝑛 (𝑅); we
computed in Example 15.2.11 that disc M𝑛 (𝑅) = 𝑅, as claimed. The converse follows
by taking O0 a maximal order containing O (furnished by Proposition 15.5.2) and
applying Lemma 15.5.1.
Example 15.5.4. By 15.2.12, if 𝐹 is a nonarchimedean local field with valuation ring

𝑅 and 𝐵 is a division quaternion algebra over 𝐹, then the valuation ring O ⊂ 𝐵 is the
unique maximal order (Theorem 13.3.11) with disc O = 𝔭2 and discrd O = 𝔭. Arguing
as in Lemma 15.5.3, we find that an 𝑅-order in 𝐵 is maximal if and only if it has
reduced discriminant 𝔭.
Maximality can be detected over global rings in terms of discriminants, as follows.

15.6. DUALITY 241
Theorem 15.5.5. Let 𝑅 be a global ring with field of fractions 𝐹, let 𝐵 be a quaternion
algebra over 𝐹, and let O ⊆ 𝐵 be an 𝑅-order. Then O is maximal if and only if
discrd(O) = disc𝑅 (𝐵). (15.5.6)
Proof. Suppose that O is maximal. Then O𝔭 is maximal for all primes 𝔭 of 𝑅. If
𝐵𝔭 ' M2 (𝐹𝔭 ) is split, then by Lemma 15.5.3, discrd O𝔭 = 𝑅𝔭 ; if 𝐵𝔭 is a division
algebra, then discrd O𝔭 = 𝔭𝑅𝔭 . Since discriminants are defined locally, we conclude
that Ö
discrd(O) = 𝔭 = disc𝑅 (𝐵)
𝔭∈Ram 𝐵rS
if 𝑅 as a global ring is the ring of S-integers.
In the other direction, if (15.5.6) holds, we choose O0 ⊇ O be a maximal 𝑅-
superorder and conclude that disc(O) = disc𝑅 (𝐵) 2 = disc(O0) so O = O0 is maximal
by Lemma 15.5.1.
Example 15.5.7. We recall Example 14.2.13, giving an explicit description of quater-
nion algebras 𝐵 = (𝑎, 𝑏 | Q) of prime discriminant 𝐷 = 𝑝. We now exhibit an explicit
maximal order in each of these algebras.
For 𝑝 = 2, we have 𝐵 = (−1, −1 | Q) and take O ⊆ 𝐵 the Hurwitz order.
For 𝑝 ≡ 3 (mod 4), we took 𝐵 = (−𝑝, −1 | Q). The order O := Zh(1 + 𝑖)/2, 𝑗i =
𝑆 ⊕ 𝑆 𝑗 with 𝑆 := Z[(1 + 𝑖)/2] has discrd O = 𝑝 by 15.2.12, so O is maximal by
Theorem 15.5.5.
For 𝑝 ≡ 1 (mod 4), we had 𝐵 = (−𝑝, −𝑞 | Q) where 𝑞 ≡ 3 (mod 4) is prime and
𝑞
= −1, so that by qudaratic reciprocity −𝑝𝑞 = −𝑞 𝑝 = 1. In this case, let 𝑐 ∈ Z be
𝑝
such that 𝑐2 ≡ −𝑝 (mod 𝑞). Then
1+ 𝑗 𝑖(1 + 𝑗) (𝑐 + 𝑖) 𝑗
O := Z ⊕ Z ⊕Z ⊕Z
2 2 𝑞
is a maximal order: one checks that O is closed under multiplication (in particular, the
basis elements are integral), and then that disc O = 𝑝. The order Zh𝑖, (1 + 𝑗)/2i ⊆ O
has the larger reduced discriminant 𝑝𝑞, hence the need for a denominator 𝑞 in the
fourth element.
For further discussion of explicit maximal orders over Z, see Ibukiyama [Ibu82,
pp. 181–182] or Pizer [Piz80a, Proposition 5.2]. For a more general construction, see
Exercise 15.5.
15.6 Duality
To round out the chapter, we relate the discriminant and trace pairings to the dual and
the different. For a detailed, general investigation of the dual in the context of other
results for orders, see Faddeev [Fad65].
We continue with the hypothesis that 𝑅 is a domain with 𝐹 = Frac 𝑅. Let 𝐵 be an
𝐹-algebra with 𝑛 := dim𝐹 𝐵 < ∞. As the trace pairing will play a significant role in
what follows, we suppose throughout that 𝐵 is separable (in particular, semisimple)
as an 𝐹-algebra with reduced trace trd. Let 𝐼, 𝐽 be 𝑅-lattices in 𝐵.
Definition 15.6.1. The dual of 𝐼 (over 𝑅, with respect to trd) is
𝐼 ♯ := {𝛼 ∈ 𝐵 : trd(𝛼𝐼) ⊆ 𝑅} = {𝛼 ∈ 𝐵 : trd(𝐼𝛼) ⊆ 𝑅}.
Some properties of the dual are evident.
Lemma 15.6.2.
(a) If 𝐼 ⊆ 𝐽 then 𝐼 ♯ ⊇ 𝐽 ♯ .
(b) For all 𝛽 ∈ 𝐵× , we have (𝛽𝐼) ♯ = 𝐼 ♯ 𝛽−1 .
(c) If 𝔭 ⊆ 𝑅 is prime, then (𝐼 (𝔭) ) ♯ = (𝐼 ♯ ) (𝔭) and the same with the completion.
Proof. For parts (a) and (b), see Exercise 15.15. The proof of part (c) is similarly
straightforward.
15.6.3. Suppose that 𝐼 is free over 𝑅 with basis 𝛼1 , . . . , 𝛼𝑛 . Since the trace pairing on
𝐵 is nondegenerate (Theorem 7.9.4), there exists a dual basis 𝛼𝑖♯ ∈ 𝐵 to 𝛼𝑖 under the
reduced trace trd, so that trd(𝛼𝑖♯ 𝛼 𝑗 ) = 0, 1 according as 𝑖 ≠ 𝑗 or 𝑖 = 𝑗.
Then 𝐼 ♯ is free over 𝑅 with basis 𝛼1♯ , . . . , 𝛼𝑛♯ : if 𝛽 = 𝑏 1 𝛼1♯ + · · · + 𝑏 𝑛 𝛼𝑛♯ with
𝑏 1 , . . . , 𝑏 𝑛 ∈ 𝐹, then 𝛽 ∈ 𝐼 ♯ if and only if trd(𝛼𝑖 𝛽) = 𝑏 𝑖 ∈ 𝑅 for all 𝑖.
Lemma 15.6.4. 𝐼 ♯ is an 𝑅-lattice in 𝐵.

Í
Proof. Let 𝛼1 , . . . , 𝛼𝑛 ∈ 𝐼 be an 𝐹-basis for 𝐵, and let 𝐽 = 𝑖 𝑅𝛼𝑖 ⊆ 𝐼. Then there
exists nonzero 𝑟 ∈ 𝑅 such that 𝑟 𝐼 ⊆ 𝐽, so 𝐽 ⊆ 𝐼 ⊆ 𝑟 −1 𝐽. Let 𝛼1♯ , . . . , 𝛼𝑛♯ ∈ 𝐵 be the
dual basis as in 15.6.3. It follows that 𝐽 ♯ = 𝑖 𝑅𝛼𝑖♯ is an 𝑅-lattice, and consequently
Í
by Lemma 15.6.2(a)–(b) we have 𝑟 𝐽 ♯ ⊆ 𝐼 ♯ ⊆ 𝐽 ♯ ; since 𝑅 is noetherian, 𝐼 ♯ is an

𝑅-lattice.
From now on, we suppose that 𝑅 is a Dedekind domain; in particular, 𝐼 is then

projective as an 𝑅-module.
Lemma 15.6.5. The natural inclusion 𝐼 ↩→ (𝐼 ♯ ) ♯ ⊆ 𝐵 is an equality.
Proof. If 𝛼 ∈ 𝐼 and 𝛽 ∈ 𝐼 ♯ then trd(𝛼𝛽) ⊆ 𝑅 and 𝛼 ∈ (𝐼 ♯ ) ♯ . To show that the map is

an equality, we argue locally, so we may suppose that 𝐼 is free over 𝑅 with basis 𝛼𝑖 ;
then by applying 15.6.3 twice, (𝐼 ♯ ) ♯ has basis (𝛼𝑖♯ ) ♯ = 𝛼𝑖 , and equality holds.
Proposition 15.6.6. We have OR (𝐼) = OL (𝐼 ♯ ) and OL (𝐼) = OR (𝐼 ♯ ).
Proof. First the inclusion (⊆). Let 𝛼 ∈ OR (𝐼); then 𝐼𝛼 ⊆ 𝐼, so 𝐼 ♯ 𝐼𝛼 ⊆ 𝐼 ♯ 𝐼 and
trd(𝛼𝐼 ♯ 𝐼) = trd(𝐼 ♯ 𝐼𝛼) ⊆ trd(𝐼 ♯ 𝐼) ⊆ 𝑅
hence 𝛼𝐼 ♯ ⊆ 𝐼 ♯ and 𝛼 ∈ OL (𝐼 ♯ ). Thus OR (𝐼) ⊆ OL (𝐼 ♯ ) ⊆ OR ((𝐼 ♯ ) ♯ ) = OR (𝐼) by

Lemma 15.6.5, so equality holds. A similar argument works on the other side.
The name dual is explained by the following lemma.

15.6. DUALITY 243
Proposition 15.6.7. Let trd(𝐼) = 𝔞 ⊆ 𝐹. Then the map

∼ Hom (𝐼, 𝔞)
𝐼 ♯ −→ 𝑅
(15.6.8)
𝛽 ↦→ (𝛼 ↦→ trd(𝛼𝛽))
is an isomorphism of OR (𝐼), OL (𝐼)-bimodules over 𝑅.
Proof. For 𝛽 ∈ 𝐼, let 𝜙 𝛽 : 𝐼 → 𝔞 be defined by 𝜙 𝛽 (𝛼) = trd(𝛼𝛽) for 𝛼 ∈ 𝐼. The map

𝛽 ↦→ 𝜙 𝛽 ∈ Hom𝑅 (𝐼, 𝔞) from (15.6.8) is an 𝑅-module homomorphism. Moreover, it
a map of OR (𝐼), OL (𝐼)-bimodules: if 𝛾 ∈ OL (𝐼) then 𝛾 ∈ OR (𝐼 ♯ ) by Lemma 15.6.6,
with induced map
𝜙 𝛽𝛾 (𝛼) = trd(𝛼𝛽𝛾) = trd(𝛾𝛼𝛽) = 𝜙 𝛽 (𝛾𝛼) = (𝛾𝜙 𝛽 ) (𝛼) (15.6.9)
and similarly on the other side.

Finally, we prove that the map (15.6.8) is also an isomorphism. Extending scalars
to 𝐹, the trace pairing gives an isomorphism of 𝐹-vector spaces
Hom𝑅 (𝐼, 𝔞) ⊗𝑅 𝐹 ' Hom𝐹 (𝐵, 𝐹) ' 𝐵

𝛽 ↦→ 𝜙 𝛽
because the pairing is nondegenerate (as 𝐵 is separable). So immediately the map is

injective; and it is surjective, because if 𝜙 ∈ Hom𝑅 (𝐼, 𝔞) then 𝜙 = 𝜙 𝛽 for some 𝛽 ∈ 𝐵,
but then 𝜙(𝛼) = trd(𝛼𝛽) ∈ 𝑅 for all 𝛼 ∈ 𝐼, so 𝛽 ∈ 𝐼 ♯ by definition.
Remark 15.6.10. The content of Proposition 15.6.7 is that although one can always
construct the module dual, the trace pairing concretely realizes this module dual as a
lattice. (And we speak of bimodules in the proposition because Hom𝑅 (𝐼, 𝔞) does not
come equipped with the structure of 𝑅-lattice in 𝐵.) This module duality, and the fact
that 𝐼 is projective over 𝑅, can be used to give another proof of Lemma 15.6.5.
The dual asks for elements that pair integrally under the trace. We might also ask
for elements that multiply one lattice into another, as follows.
Definition 15.6.11. Let 𝐼, 𝐽 be 𝑅-lattices. The left colon lattice of 𝐼 with respect to
𝐽 is the set
(𝐼 : 𝐽) L := {𝛼 ∈ 𝐵 : 𝛼𝐽 ⊆ 𝐼}
and similarly the right colon lattice is
(𝐼 : 𝐽) R := {𝛼 ∈ 𝐵 : 𝐽𝛼 ⊆ 𝐼}.
Note that (𝐼 : 𝐼) L = OL (𝐼) is the left order of 𝐼 (and similarly on the right). The
same proof as in Lemma 10.2.7 shows that (𝐼 : 𝐽) L and (𝐼 : 𝐽) R are 𝑅-lattices.
(𝐼𝐽) ♯ = (𝐼 ♯ : 𝐽) R = (𝐽 ♯ : 𝐼) L .
Proof. We have 𝛽 ∈ (𝐼𝐽) ♯ if and only if trd(𝛽𝐼𝐽) ⊆ 𝑅 if and only if 𝛽𝛼 ∈ 𝐽 ♯ for

all 𝛼 ∈ 𝐼 if and only if 𝛽 ∈ (𝐽 ♯ : 𝐼) L . A similar argument works on the other side,
considering trd(𝐼𝐽 𝛽) instead.
Corollary 15.6.13. We have OL (𝐼) = (𝐼 𝐼 ♯ ) ♯ and OR (𝐼) = (𝐼 ♯ 𝐼) ♯ .
Proof. Combining Lemmas 15.6.5 and 15.6.12,
OL (𝐼) = (𝐼 : 𝐼) L = ((𝐼 ♯ ) ♯ : 𝐼) L = (𝐼 𝐼 ♯ ) ♯
and similarly on the right.
Definition 15.6.14. The level of 𝐼 is the fractional ideal lvl(𝐼) = nrd(𝐼 ♯ ) ⊆ 𝐹.
We now relate the above duality to the discriminant.
Definition 15.6.15. The codifferent of O is
codiff(O) := O♯ .
Lemma 15.6.16. OL (codiff (O)) = OR (codiff(O)) = O and O ⊆ codiff (O).
Proof. By Proposition 15.6.6, O = OR (O) = OL (codiff(O)) and similarly on the

right. And O ⊆ codiff(O) since trd(OO) = trd(O) ⊆ 𝑅.
The major role played by the codifferent is its relationship to the discriminant, as
follows.
Lemma 15.6.17. disc(O) = [codiff (O) : O] 𝑅 .
Proof. For a prime 𝔭 ⊆ 𝑅 we have disc(O) (𝔭) = disc(O (𝔭) ) and [O♯(𝔭) : O (𝔭) ] 𝑅 (𝔭) =
( [O♯ : O] 𝑅 ) (𝔭) , and so to establish the equality we may argue locally. Since O (𝔭) is
Í
free over 𝑅 (𝔭) , we reduce to the case where O is free over 𝑅, say O = 𝑖 𝑅𝛼𝑖 . Then
O♯ = 𝑖 𝑅𝛼𝑖♯ with 𝛼1♯ , . . . , 𝛼𝑛♯ ∈ 𝐵 the dual basis, as in 15.6.3.
Í
The ideal disc(O) is principal, generated by 𝑑 (𝛼1 , . . . , 𝛼𝑛 ) = det(trd(𝛼𝑖 𝛼 𝑗 ))𝑖, 𝑗 ; at
the same time, the 𝑅-index [O♯ : O] 𝑅 is generated by det(𝛿) where 𝛿 is the change of
basis from 𝛼𝑖♯ to 𝛼𝑖 . But 𝛿 is precisely the matrix (trd(𝛼𝑖 𝛼 𝑗 ))𝑖, 𝑗 (Exercise 15.14), and
the result follows.
Remark 15.6.18. In certain circumstances, it is preferable to work with an integral

ideal measuring the discriminant, so instead of the codifferent instead a different: we
will want to take a kind of inverse. We study this in the next chapter: see section 16.8.
15.6. DUALITY 245
Exercises
Unless otherwise specified, let 𝑅 be a noetherian domain with field of fractions 𝐹.

𝑎, 𝑏
1. Let char 𝐹 = 2 and let be a quaternion algebra over 𝐹 with 𝑎, 𝑏 ∈ 𝑅 and
𝐹
𝑏 ≠ 0. Show that O = 𝑅 + 𝑅𝑖 + 𝑅 𝑗 + 𝑅𝑖 𝑗 is an 𝑅-order in 𝐵 and compute the
(reduced) discriminant of O.
2. Let 𝐵 = M𝑛 (𝐹) and O = M𝑛 (𝑅) with 𝑛 ≥ 1. Show that disc(O) = 𝑅. [Hint:
Compute directly on a basis {𝑒 𝑖 𝑗 }𝑖, 𝑗 of matrix units, which satisfy 𝑒 𝑖 𝑗 𝑒 𝑖0 𝑗 0 = 𝑒 𝑖 𝑗 0
if 𝑗 = 𝑖 0, otherwise zero.]
3. Suppose 𝑅 is a global ring, so 𝐹 is a global field; let 𝐵 be a quaternion algebra
over 𝐹 and let O ⊆ 𝐵 be an 𝑅-order. Prove that for all primes 𝔭 ⊆ 𝑅, we have
O𝔭 ' M𝑛 (𝑅𝔭 ) if and only if 𝔭 - disc O.
4. Let 𝐵 := (𝐾, 𝑏 | 𝐹) be a quaternion algebra over a field 𝐹 with 𝑏 ∈ 𝐹 × . Let
𝑆 ⊆ 𝐾 be an 𝑅-order with 𝔡 := disc(𝑆); let 𝔟 ⊆ 𝐾 be a fractional 𝑆-ideal (which
can be but need not be invertible), and finally let O := 𝑆 ⊕ 𝔟 𝑗.
(a) Show that O is an 𝑅-order if and only if Nm𝐾 |𝐹 𝔟 ⊆ 𝑏 −1 𝑅.
(b) Compute that discrd O = 𝔡(Nm𝐾 |𝐹 𝔟)𝑏.
5. In this exercise, we consider a construction of maximal orders as crossed
♦ products

𝑞 ,𝑏
in the simplest case over Q, continuing Exercise 14.9. Let 𝐵 := be a
Q
quaternion algebra of discriminant 𝐷, where 𝑏 ∈ Z is squarefree with 𝐷 | 𝑏 and
♦
√︁ with 𝑞 = ±𝑞 ≡ 1 (mod 4), the minus sign if 𝐵 is indefinite.
𝑞 is an odd prime
Let 𝐾 := Q( 𝑞 ♦ ) be the quadratic field of discriminant 𝑞 ♦ . Let 𝑆 ⊆ 𝐾 be the
ring of integers of 𝐾, so disc 𝑆 = 𝑞 ♦ .
♦
𝑞
(a) Show that for all odd primes 𝑝 | (𝑏/𝐷), we have = 1. Conclude
𝑝
there exists an ideal 𝔟 ⊆ 𝑆 such that Nm 𝔟 = 𝑏/𝐷.
(b) Let 𝔮 ⊆ 𝑆 be the unique prime above 𝑞, and let
O := 𝑆 ⊕ (𝔮𝔟) −1 𝑗
Show that O is a maximal order in 𝐵.

(c) Let 𝑐 ∈ Z satisfy 𝑐2 ≡ 𝑞 ♦ (mod 4𝑏/𝐷). Show that the order O in (b) can
be written
1+𝑖 𝐷 (𝑐 + 𝑖) 𝑗
O=Z⊕Z ⊕ Z𝑗 ⊕ Z .
2 2𝑏𝑞
6. Let 𝐵 be a separable 𝐹-algebra with dim𝐹 𝐵 = 𝑛. Show that 𝛼1 , . . . , 𝛼𝑛 ∈ 𝐵 are
linearly independent over 𝐹 if and only if 𝑑 (𝛼1 , . . . , 𝛼𝑛 ) ≠ 0.
7. Let O be an 𝑅-order. Show that disc(O) is generated by
{𝑑 (1, 𝛼1 , . . . , 𝛼𝑛−1 ) : 𝛼1 , . . . , 𝛼𝑛−1 ∈ O}.

8. Let 𝐼 be an 𝑅-lattice in 𝐵 over 𝐹, let 𝐾 be a finite extension field of 𝐹, and let 𝑆

be a domain containing 𝑅 with field of fractions 𝐾. Show that
disc(𝐼 ⊗𝑅 𝑆) = disc(𝐼) ⊗𝑅 𝑆 = disc(𝐼)𝑆.
9. Let char 𝐹 ≠ 2 and let 𝐵 be a quaternion algebra over 𝐹. Let 𝛼, 𝛽 ∈ 𝐵 be such

that 𝐹 (𝛼) ∩ 𝐹 (𝛽) = 𝐹. Recall the discriminant form Δ (Exercise 4.3), and let
trd(𝛼) trd(𝛽)
𝑠 := trd(𝛼𝛽) − .
2
Show that
𝑑 (1, 𝛼, 𝛽, 𝛼𝛽) = −(𝑠2 − 4Δ(𝛼)Δ(𝛽)) 2 = −Δ(𝛼𝛽) 2 .
[Hint: reduce to the case where trd(𝛼) = trd(𝛽) = 0, noting the invariance of
𝑠.]
10. Let 𝐵 be a quaternion algebra over 𝐹. Define 𝑚 : 𝐵 × 𝐵 × 𝐵 → 𝐹 by
𝑚(𝛼1 , 𝛼2 , 𝛼3 ) := trd( [𝛼1 , 𝛼2 ]𝛼3 ) for 𝛼𝑖 ∈ 𝐵. If 𝛽𝑖 = 𝑀𝛼𝑖 for some 𝑀 ∈ M3 (𝐹),
show that
𝑚(𝛽1 , 𝛽2 , 𝛽3 ) = det(𝑀)𝑚(𝛼1 , 𝛼2 , 𝛼3 ).
11. Let 𝐵 be a quaternion algebra over 𝐹. Give another proof that
𝑚(𝛼1 , 𝛼2 , 𝛼3 ) 2 = 𝑑 (1, 𝛼1 , 𝛼2 , 𝛼3 )
(cf. Brzezinski [Brz82, Lemma 1.1(a)]) for all 𝛼𝑖 ∈ 𝐵 as follows:

(a) Suppose 𝐵 = M2 (𝐹). Show that the matrix units

0 1 0 0 1 0
𝑒 12 = , 𝑒 21 = , 𝑒 22 =
0 0 1 0 0 0
span 𝐵/𝐹, and 𝑚(𝑒 12 , 𝑒 21 , 𝑒 22 ) 2 = 𝑑 (1, 𝑒 12 , 𝑒 21 , 𝑒 22 ). Conclude using

Exercise 15.10.
(b) Reduce to (a) in general by taking a splitting field for 𝐵.
12. Suppose 𝑅 = 𝑅 (S) is a global ring with 2 ∈ 𝑅 × . Let 𝐾 ⊃ 𝐹 be a quadratic field
extension and 𝑆 ⊆ 𝐾 an 𝑅-order. Let Ram(𝐾) be the set of places of 𝐹 that are
ramified in 𝐾. Show that 𝑆 is maximal if and only if its discriminant is equal to
Ö
disc𝑅 (𝑆) = 𝔭⊆𝑅
𝔭∈Ram(𝐾 )\S
in analogy with Theorem 15.5.5.

13. Let 𝐵 be a finite-dimensional 𝐹-algebra with a standard involution. Compare
det(trd(𝛼𝑖 𝛼 𝑗 ))𝑖, 𝑗 with det(trd(𝛼𝑖 𝛼 𝑗 ))𝑖, 𝑗
for 𝛼𝑖 ∈ 𝐵, and show that defining the discriminant of an order O ⊆ 𝐵 with

respect to either pairing gives the same result.
15.6. DUALITY 247
I 14. Let 𝐵 be a semisimple 𝐹-algebra with dim𝐹 𝐵 = 𝑛, let 𝐼 be an 𝑅-lattice that is

free over 𝑅 with basis 𝛼1 , . . . , 𝛼𝑛 , and let 𝛼1♯ , . . . , 𝛼𝑛♯ ∈ 𝐵 be the dual basis, so
trd(𝛼𝑖♯ 𝛼 𝑗 ) = 1, 0 according as 𝑖 = 𝑗 or not. Show that the change of basis matrix
from {𝛼𝑖♯ }𝑖 to {𝛼𝑖 }𝑖 is given by (trd(𝛼𝑖 𝛼 𝑗 ))𝑖, 𝑗 .
15. Let 𝐼 ⊆ 𝐵 be an 𝑅-lattice in a separable algebra 𝐵.
(a) If 𝐽 ⊆ 𝐵 is an 𝑅-lattice with 𝐼 ⊆ 𝐽, show that 𝐼 ♯ ⊇ 𝐽 ♯ .
(b) Show that for all 𝛽 ∈ 𝐵× , we have (𝛽𝐼) ♯ = 𝐼 ♯ 𝛽−1

𝑅 𝑅
16. Let 𝑅 be a DVR with maximal ideal 𝔭 = 𝜋𝑅, and let O := for 𝑒 ≥ 0.
𝔭𝑒 𝑅
Compute the codifferent codiff (O): in particular, show that codiff(O) is a
principal two-sided O-ideal, and find a generator. Verify Lemma 15.6.17.
17. Let 𝑅 be a noetherian domain with 𝐹 = Frac 𝑅. Let 𝐵 be a central simple
algebra over 𝐹. Let O ⊆ 𝐵 be an 𝑅-order. We say O is Azumaya if O is
𝑅-simple, which is to say every two-sided ideal 𝐼 ⊆ O is of the form 𝔞O = O𝔞
with 𝔞 = 𝐼 ∩ 𝑅 ⊆ 𝑅.
a) Show that O is Azumaya if and only if every 𝑅-algebra homomorphism
O → 𝐴 is either the zero map or injective.
b) Show that O is Azumaya if and only if O/𝔪O is a central simple algebra
over the field 𝑅/𝔪 for all maximal ideals 𝔪 of 𝑅.
c) Suppose that 𝐵 is a quaternion algebra. Show that the quaternion order O
is Azumaya if and only if disc O = 𝑅. Conclude that the only Azumaya
quaternion algebra over the valuation ring 𝑅 of a local field is M2 (𝑅), and
that the only Azumaya quaternion algebra over Z is M2 (Z).
[See Auslander and Goldman [AG60] or Milne [Milne80, §IV.1].]
18. Let 𝐺 be a finite group of order 𝑛 = #𝐺 and let 𝑅 be a domain with 𝐹 = Frac 𝑅.
Suppose that char 𝐹 - 𝑛. Then 𝐵 := 𝐹 [𝐺] is a separable 𝐹-algebra by Exercise
7.15.
a) Consider the algebra trace Tr 𝐵 |𝐹 and its associated bilinear form. Show
that in the basis of 𝐹 [𝐺] given by the elements of 𝐺 that the trace pairing
is the scalar matrix 𝑛.
b) Now write 𝐵 ' 𝐵1 × · · · × 𝐵𝑟 as a product of simple 𝐹-algebras. Let 𝐾𝑖
be the center of 𝐵𝑖 , and let dim𝐾𝑖 𝐵𝑖 = 𝑛2𝑖 . Show that Tr | 𝐵𝑖 = 𝑛𝑖 trd. Let
O = 𝑅[𝐺], and suppose that O ' O1 × · · · × O𝑟 . Show that
codiff (O) = 𝑛−1 −1

1 O1 × · · · × 𝑛𝑟 O𝑟 .
Chapter 16
Quaternion ideals and invertibility
Much like a space can be understood by studying functions on that space, often the
first task to understand a ring 𝐴 is to understand the ideals of 𝐴 and modules over 𝐴
(in other words, to pursue “linear algebra” over 𝐴). The ideals of a ring that are easiest
to work with are the principal ideals—but not all ideals are principal, and various
algebraic structures are built to understand the difference between these two. In this
chapter, we consider these questions for the case where 𝐴 is a quaternion order.
16.1 ⊲ Quaternion ideals

To get warmed up for the noncommutative situation, we consider ideals of quadratic
rings. An integer 𝑑 ∈ Z is a discriminant if 𝑑 ≡ 0, 1 (mod 4). Let 𝑆 be the quadratic
order of nonsquare discriminant 𝑑 ∈ Z, namely,
√ √
𝑆 = 𝑆(𝑑) := Z ⊕ Z[(𝑑 + 𝑑)/2] ⊂ 𝐾 = Q( 𝑑).
The set of ideals of 𝑆 has a natural multiplicative structure with identity element 𝑆
(giving it the structure of a commutative monoid), but we lack inverses and we would
surely feel more comfortable with a group structure. So we consider nonzero 𝑆-lattices
𝔞 ⊂ 𝐾, and call them fractional ideals of 𝑆; equivalently, they are the 𝑆-submodules
𝑑 −1 𝔞 ⊂ 𝐾 with 𝔞 ⊆ 𝑆 a nonzero ideal and 𝑑 ∈ Z>0 , hence the name fractional ideal
(viz. 9.2.4). To get a group structure, we must restrict our attention to the invertible
fractional ideals 𝔞 ⊂ 𝐾, i.e., those such that there exists a fractional ideal 𝔟 with
𝔞𝔟 = 𝑆. The simplest kind of invertible fractional ideals are the principal ones 𝔞 = 𝑎𝑆
for 𝑎 ∈ 𝐾 × , with inverse 𝔞−1 = 𝑎 −1 𝑆. If a fractional ideal 𝔞 has an inverse then this
inverse is unique, given by
𝔞−1 = {𝑥 ∈ 𝐾 : 𝑥𝔞 ⊆ 𝑆};
and for a fractional ideal 𝔞, we always have 𝔞𝔞−1 ⊆ 𝑆 (but equality may not hold). If
𝑆 = Z𝐾 is the ring of integers (the maximal order) of 𝐾, then all nonzero fractional
ideals of 𝑆 are invertible—in fact, this property characterizes Dedekind domains, in
that a noetherian commutative ring is a Dedekind domain if and only if every nonzero
(prime) ideal is invertible. (See also the summary in section 9.2.)
249
250 CHAPTER 16. QUATERNION IDEALS AND INVERTIBILITY
A fractional ideal 𝔞 of 𝑆 is invertible if and only if 𝔞 is locally principal, i.e.,

𝔞 ⊗Z Z ( 𝑝) = 𝔞 ( 𝑝) = 𝑎 ( 𝑝) Z ( 𝑝) is a principal fractional ideal of the localization 𝑆 ( 𝑝) for
all primes 𝑝. Every locally principal ideal is invertible, and the extent to which the
converse holds is something that arises in an important way more generally in algebraic
geometry. In the language of commutative algebra, a locally principal 𝑆-module is
equivalently a projective 𝑆-module of rank 1.
Suppose that 𝑆 is not maximal; then 𝑆(𝑑) = Z + 𝑓 Z𝐾 for a unique integer 𝑓 ∈ Z>1 ,
the conductor of 𝑆; accordingly, 𝑑 = 𝑑 𝐾 𝑓 2 , where 𝑑 𝐾 ∈ Z is the discriminant of Z𝐾
(a fundamental discriminant). In this case, there is always an ideal of 𝑆 that is not
invertible. Specifically, consider the ideal
√
𝔣 = 𝑓 Z + 𝑑Z ⊆ 𝑆. (16.1.1)
Then 𝔣 is a free Z-module of rank 2 and

√ √
𝔣2 = ( 𝑓 Z + Z 𝑑) 2 = 𝑓 2 Z + 𝑓 𝑑Z = 𝑓 𝔣
so if 𝔣 were invertible, then cancelling we would obtain 𝔣 = 𝑓 𝑆, √ a contradiction. The

source of this example is that 𝔣 = 𝑓 𝑆(𝑑 𝐾 ) since 𝑆(𝑑 𝐾 ) = Z + 𝑑 𝐾 Z, so really this
fractional ideal belongs to the maximal order 𝑆(𝑑 𝐾 ), not to 𝑆. For more on the notion
of invertibility for quadratic orders, see Cox [Cox89, §7], with further connections to
quadratic forms and class numbers.
We now turn to the quaternionic generalization, where noncommutativity presents
some complications. Let 𝐵 be a quaternion algebra over Q and let O ⊂ 𝐵 be an order.
To study ideals of O we must distinguish between left or right ideals, and the product
of two (say) right O-ideals need not be again a right O-ideal! To address this, for
lattices 𝐼, 𝐽 ⊂ 𝐵, we say that 𝐼 is compatible with 𝐽 if the right order of 𝐼 is equal to
the left order of 𝐽, so that what comes between 𝐼 and 𝐽 in the product 𝐼 · 𝐽 “matches
up”.
A lattice 𝐼 ⊂ 𝐵 is right invertible if there exists a lattice 𝐼 0 ⊂ 𝐵 such that
𝐼 𝐼 0 = OL (𝐼)
with a compatible product, and we call 𝐼 0 a right inverse. We similarly define notions
on the left, and we say 𝐼 ⊂ 𝐵 is invertible if there is a two-sided inverse 𝐼 0 ⊂ 𝐵, so
𝐼 𝐼 0 = OL (𝐼) = OR (𝐼 0) and 𝐼 0 𝐼 = OL (𝐼 0) = OR (𝐼)
with both of these products compatible. If a lattice 𝐼 has a two-sided inverse, then this
inverse is uniquely given by
𝐼 −1 := {𝛼 ∈ 𝐵 : 𝐼𝛼𝐼 ⊆ 𝐼}
(defined so as to simultaneously take care of both left and right): we always have that
𝐼 𝐼 −1 ⊆ OL (𝐼), but equality is needed for right invertibility, and the same on the left.
Let O ⊆ 𝐵 be an order. A left fractional O-ideal is a lattice 𝐼 ⊆ 𝐵 such that
O ⊆ OL (𝐼); we similarly define on the right. For a maximal order, all lattices are
invertible (Proposition 16.6.15(b)).
16.2. LOCALLY PRINCIPAL, COMPATIBLE LATTICES 251
Proposition 16.1.2. Let O ⊆ 𝐵 be a maximal order. Then a left or right fractional

O-ideal is invertible.
The simplest kind of invertible lattices are the principal lattices
𝐼 = OL (𝐼)𝛼 = 𝛼OR (𝐼)
with 𝛼 ∈ 𝐵× : its inverse is 𝐼 −1 = 𝛼−1 OL (𝐼) = OR (𝐼)𝛼−1 .

The major task of this chapter will be to interrelate these notions in the quaternionic
context. Let
nrd(𝐼) := gcd({nrd(𝛼) : 𝛼 ∈ 𝐼}),
i.e., nrd(𝐼) is a positive generator of the (finitely generated) subgroup of Q generated
by nrd(𝛼) for 𝛼 ∈ 𝐼. The main result over Q is the following theorem (Main Theorem
16.7.7).
Main Theorem 16.1.3. Let 𝐵 be a quaternion algebra over Q and let 𝐼 ⊂ 𝐵 be an

integral lattice. Then the following are equivalent:
(i) 𝐼 is locally principal, i.e., 𝐼 ( 𝑝) = 𝐼 ⊗Z Z ( 𝑝) is principal for all primes 𝑝;

(ii) 𝐼 is invertible;
(iii) 𝐼 is right invertible;
(iii0) 𝐼 is left invertible;
(iv) nrd(𝐼) 2 = [OR (𝐼) : 𝐼]; and
(iv0) nrd(𝐼) 2 = [OL (𝐼) : 𝐼].
Accordingly, for 𝐼 integral, we may define the right absolute norm of 𝐼 by
N (𝐼) := #(OR (𝐼)/𝐼) = [OR (𝐼) : 𝐼] ∈ Z ≥1
and similarly on the left; by Main Theorem 16.1.3 (iv) ⇔ (iv0), when 𝐼 is locally
principal, the left and right absolute norms coincide (called then just absolute norm)
and are related to the reduced norm by N (𝐼) = nrd(𝐼) 2 .
16.2 Locally principal, compatible lattices

The simplest lattices to understand are those that are principal; but as we saw in
section 9.4, lattices over Dedekind domains are inherently local in nature. We are led
to consider the more general class of locally principal lattices. We work first with
lattices, and later we will keep track of their left and right orders.
Throughout this chapter, let 𝑅 be a Dedekind domain with field of fractions 𝐹, let
𝐵 be a finite-dimensional algebra over 𝐹, and let 𝐼 ⊆ 𝐵 be an 𝑅-lattice.
Definition 16.2.1. 𝐼 is principal if there exists 𝛼 ∈ 𝐵 such that
𝐼 = OL (𝐼)𝛼 = 𝛼OR (𝐼);
we say that 𝐼 is generated by 𝛼.

16.2.2. If 𝐼 is generated by 𝛼 ∈ 𝐵, then since 𝐼 is a lattice (Definition 9.3.1) we have

𝐼 𝐹 = 𝐵𝛼 = 𝐵, so 𝛼 ∈ 𝐵× .
16.2.3. If 𝐼 = OL (𝐼)𝛼, then OR (𝐼) = 𝛼−1 OL (𝐼)𝛼 by Exercise 16.2, so
𝐼 = 𝛼(𝛼−1 OL (𝐼)𝛼) = 𝛼OR (𝐼).
Therefore it is sufficient to check for a one-sided generator (and if we defined the

obvious notions of left principal or right principal, these would be equivalent to the
notion of principal).
The notion of principality naturally extends locally.
Definition 16.2.4. An 𝑅-lattice 𝐼 is locally principal if 𝐼 (𝔭) = 𝐼 ⊗𝑅 𝑅 (𝔭) is a principal

𝑅 (𝔭) -lattice for all primes 𝔭 of 𝑅.
Now let 𝐼, 𝐽 be 𝑅-lattices in 𝐵. We define the product 𝐼𝐽 to be the 𝑅-submodule

of 𝐵 generated by the set
{𝛼𝛽 : 𝛼 ∈ 𝐼, 𝛽 ∈ 𝐽}.
The product 𝐼𝐽 is an 𝑅-lattice: it is finitely generated as this is true of 𝐼, 𝐽 individually,
and there exists a nonzero 𝑟 ∈ 𝐼 (Exercise 9.2) so 𝑟 𝐽 ⊂ 𝐼𝐽 and thus
𝐵 = 𝐹𝐽 = 𝐹 (𝑟 𝐽) ⊆ 𝐹 𝐼𝐽
so equality holds.
When multiplication of two lattices matches up their respective left and right
orders, we give it a name.
Definition 16.2.5. We say that 𝐼 is compatible with 𝐽 if OR (𝐼) = OL (𝐽).
We will also sometimes just say that the product 𝐼𝐽 is compatible to mean that 𝐼 is
compatible with 𝐽. The relation “is compatible with” is in general neither symmetric
nor transitive. (Looking ahead to groupoids in Chapter 17, we might also say that the
two lattices are composable.)
16.2.6. 𝐼 has the structure of a right OR (𝐼)-module and 𝐽 the structure of a left OL (𝐽)-
module. When OR (𝐼) = OL (𝐽) = O, that is, when 𝐼 is compatible with 𝐽, it makes
sense to consider the tensor product 𝐼 ⊗O 𝐽 as an 𝑅-module. The multiplication map
𝐵 ⊗ 𝐵 𝐵 −→∼ 𝐵 defined by 𝛼 ⊗ 𝛽 ↦→ 𝛼𝛽 restricts to give an isomorphism 𝐼 ⊗ 𝐽 −→ ∼ 𝐼𝐽
O
as 𝑅-lattices. In this way, multiplication of compatible lattices can be thought of as a
special case of the tensor product of modules.
We conclude this section with several other basic properties of lattices.
Definition 16.2.7. An 𝑅-lattice 𝐼 is integral if 𝐼 2 ⊆ 𝐼.
In Definition 16.2.7, the product need not be compatible.
Lemma 16.2.8. Let 𝐼 be an 𝑅-lattice. Then the following are equivalent:

16.3. REDUCED NORMS 253
(i) 𝐼 is integral;
(ii) For all 𝛼, 𝛽 ∈ 𝐼, we have 𝛼𝛽 ∈ 𝐼;
(iii) 𝐼 ⊆ OL (𝐼), so 𝐼 is a left ideal of OL (𝐼) in the usual sense;
(iii0) 𝐼 ⊆ OR (𝐼); and
(iv) 𝐼 ⊆ OL (𝐼) ∩ OR (𝐼).
If 𝐼 is integral, then every element of 𝐼 is integral over 𝑅.
Proof. The equivalence (i) ⇔ (ii) follows immediately. For (i) ⇔ (iii), we have 𝐼 𝐼 ⊆ 𝐼
if and only if 𝐼 ⊆ OL (𝐼) by definition of OL (𝐼), and the same argument gives (i)
⇔ (iii0), and this then gives (i) ⇔ (iv). The final statement follows from Lemma
10.3.2.
In light of Lemma 16.2.8, we need not define notions of left integral or right
integral.
For an 𝑅-lattice 𝐼, there exists nonzero 𝑑 ∈ 𝑅 such that 𝑑𝐼 is integral, so every
𝑅-lattice 𝐼 = (𝑑𝐼)/𝑑 is fractional in the sense that it is obtained from an integral lattice
with denominator.
Definition 16.2.9. Let O ⊆ 𝐵 be an 𝑅-order. A left fractional O-ideal is a lattice
𝐼 ⊆ 𝐵 such that O ⊆ OL (𝐼); similarly on the right.
If O, O0 ⊆ 𝐵 are 𝑅-orders, then a fractional O, O0-ideal is a lattice 𝐼 that is a left
fractional O-ideal and a right fractional O0-ideal.
Remark 16.2.10. A left ideal 𝐼 ⊆ O in the usual sense is an integral left O-ideal
in the sense of Definition 16.2.9 if and only if 𝐼 𝐹 = 𝐵, i.e., 𝐼 is a (full) 𝑅-lattice.
(Same for right and two-sided ideals.) If 𝐼 is nonzero and 𝐵 is a division algebra, then
automatically 𝐼 is full and the two notions coincide.
Indeed, suppose 𝐼 ⊆ O is a left ideal of O (in the usual sense). Then O ⊆ OL (𝐼)
so in particular 𝐼 has the structure of an 𝑅-module, and since O is finitely generated as
an 𝑅-module and 𝑅 is noetherian, it follows that 𝐼 is finitely generated. Consequently,
a left ideal 𝐼 ⊆ O is a left fractional O-ideal if and only if 𝐼 𝐹 = 𝐵.
Definition 16.2.11. Let 𝐼 be a left fractional O-ideal. We say that 𝐼 is sated (as a left
fractional O-ideal) if O = OL (𝐼). We make a similar definition on the right and for
two-sided ideals.
Example 16.2.12. By Lemma 15.6.16, codiff(O) is a two-sided sated O-ideal.
Remark 16.2.13. Our notion of sated is sometimes called proper: we do not use this
already overloaded term, as it conflicts with the notion of a proper subset.
16.3 Reduced norms

Next, we extend the reduced norm to lattices; see also Reiner [Rei2003, §24]. To this
end, in this section we suppose that 𝐵 is semisimple.
Definition 16.3.1. The reduced norm nrd(𝐼) of 𝐼 is the 𝑅-submodule of 𝐹 generated
by the set {nrd(𝛼) : 𝛼 ∈ 𝐼}.
Lemma 16.3.2. The reduced norm nrd(𝐼) is a fractional ideal of 𝐹: i.e., it is finitely
generated as an 𝑅-module.
Proof. We first give a proof when 𝐵 has a standard involution, and nrd is a quadratic
form. Since 𝐼 is an 𝑅-lattice we have 𝐼 𝐹 = 𝐵; since nrd(𝐵) ≠ {0}, we have nrd(𝐼) ≠
{0}. And 𝐼 is generated by finitely many 𝛼𝑖 as an 𝑅-module; the 𝑅-module nrd(𝐼) is
then generated by the values 𝑎 𝑖𝑖 = nrd(𝛼𝑖 ) and 𝑎 𝑖 𝑗 = nrd(𝛼𝑖 + 𝛼 𝑗 ) − nrd(𝛼𝑖 ) − nrd(𝛼 𝑗 ),
since then Í
Í Í
nrd 𝑖 𝑐 𝑖 𝛼𝑖 = 𝑖, 𝑗 𝑎 𝑖 𝑗 𝑐 𝑖 𝑐 𝑗 ∈ 𝑖, 𝑗 𝑅𝑎 𝑖 𝑗
for all 𝑐 𝑖 ∈ 𝑅.
Now for the general case. Replacing 𝐼 by 𝑟 𝐼 with 𝑟 ∈ 𝑅 nonzero, we may suppose
that 𝐼 is integral, and hence nrd(𝐼) ⊆ 𝑅. Since 𝐼 is a lattice, there exists 𝑟 ∈ 𝐼 ∩ 𝑅
with 𝑟 ≠ 0. For all 𝔭 such that ord𝔭 (𝑟) = 0, we have 1 ∈ 𝐼 (𝔭) so nrd(𝐼 (𝔭) ) = 𝑅 (𝔭) .
For each of the finitely many primes 𝔭 that remain, we choose an element 𝛼 ∈ 𝐼 such
that ord𝔭 (nrd(𝛼)) is minimal; then nrd(𝛼) generates nrd(𝐼 (𝔭) ), and by the local-global
dictionary, these finitely many elements generate nrd(𝐼).
16.3.3. For a prime 𝔭 of 𝑅 we have nrd(𝐼) (𝔭) = nrd(𝐼 (𝔭) ), so by the local-global
property of lattices (Lemma 9.4.6),
Ù Ù
nrd(𝐼) = nrd(𝐼) (𝔭) = nrd(𝐼 (𝔭) ). (16.3.4)
𝔭 𝔭
16.3.5. If 𝐼 is a principal 𝑅-lattice generated by 𝛼 ∈ 𝐼 then nrd(𝐼) = nrd(𝛼)𝑅; more

generally, if 𝐼 is an 𝑅-lattice and 𝛼 ∈ 𝐵× then nrd(𝛼𝐼) = nrd(𝛼) nrd(𝐼) (Exercise
16.4).
Now suppose that 𝐼, 𝐽 are lattices. Then nrd(𝐼𝐽) ⊇ nrd(𝐼) nrd(𝐽). However, we
need not have equality, as the following example indicates.
Example 16.3.6. It is not always true that nrd(𝐼𝐽) = nrd(𝐼)nrd(𝐽). For

example, if
𝑎𝑅 𝑅 𝑎𝑅 𝑎𝑅
𝑎 ∈ 𝑅 is neither zero nor a unit, then 𝐼 = and 𝐽 = are 𝑅-lattices
𝑎𝑅 𝑅 𝑅 𝑅
in M2 (𝐹) with nrd(𝐼) = nrd(𝐽) = 𝑎𝑅 but 𝐼𝐽 = M2 (𝑅) and so nrd(𝐼𝐽) = 𝑅.
We have OR (𝐽) = M2 (𝑅) = OL (𝐼), so 𝐽 is compatible with 𝐼, and nrd(𝐽 𝐼) =
𝑎 2 𝑅 = nrd(𝐽) nrd(𝐼); but
𝑅 𝑎 −1 𝑅

𝑅 𝑎𝑅
OR (𝐼) = and OL (𝐽) = −1 ,
𝑎𝑅 𝑅 𝑎 𝑅 𝑅
so 𝐼 is not compatible with 𝐽.
The issue present in Example 16.3.6 is that the product is not as well-behaved
for noncommutative rings as for commutative rings; we need the elements coming
between 𝐼 and 𝐽 to match up.
Lemma 16.3.7. Suppose that 𝐼 is compatible with 𝐽 and that either 𝐼 or 𝐽 is locally
principal. Then nrd(𝐼𝐽) = nrd(𝐼) nrd(𝐽).
16.4. ALGEBRA AND ABSOLUTE NORM 255
Proof. By the local-global property for norms (16.3.4) and since localization com-
mutes with multiplication, i.e.,
(𝔞𝔟) (𝔭) = 𝔞 (𝔭) 𝔟 (𝔭) for all (finitely generated) 𝑅-modules 𝔞, 𝔟 ⊆ 𝐹,
we may localize and suppose that either 𝐼 or 𝐽 is principal. Suppose 𝐼 is (right)

principal. Then 𝐼 = 𝛼O for some 𝛼 ∈ 𝐵 where O = OR (𝐼) = OL (𝐽). Then
𝐼𝐽 = (𝛼O)𝐽 = 𝛼(O𝐽) = 𝛼𝐽
and so nrd(𝐼𝐽) = nrd(𝛼) nrd(𝐽) = nrd(𝐼) nrd(𝐽) by 16.3.5. The case where 𝐽 is
principal follows in the same way.
Principal lattices are characterized by reduced norms, as follows.
Lemma 16.3.8. Let 𝐼 be locally principal and let 𝛼 ∈ 𝐼. Then 𝛼 generates 𝐼 if and
only if nrd(𝛼)𝑅 = nrd(𝐼).
Proof. If 𝐼 = 𝛼O then nrd(𝐼) = nrd(𝛼)𝑅 by Lemma 16.3.7.

For the converse, let O = OR (𝐼). We want to show that 𝐼 = 𝛼O, and we know that
𝐼 ⊇ 𝛼O. To prove that equality holds, it suffices to show this locally, so we may suppose
that 𝐼 = 𝛽O. Then 𝛼 = 𝛽𝜇 with 𝜇 ∈ O, and nrd(𝛼) = nrd(𝛽𝜇) = nrd(𝛽) nrd(𝜇). By
hypothesis, nrd(𝜇) ∈ 𝑅 × , and thus 𝜇 ∈ O× , so 𝛽O = 𝛼O.
16.4 Algebra and absolute norm

The reduced norm of an ideal is related to its algebra norm, as follows. We continue
to suppose that 𝐵 is semisimple, so the definitions of left and right norm coincide.
Definition 16.4.1. The (algebra) norm Nm 𝐵 |𝐹 (𝐼) of 𝐼 is the 𝑅-submodule of 𝐹

generated by the set {Nm 𝐵 |𝐹 (𝛼) : 𝛼 ∈ 𝐼}.
Remark 16.4.2. The definition of algebra norm by necessity depends on the choice of
domain 𝑅; indeed, 𝐼 is an 𝑅-lattice.
Proposition 16.4.3. The following are equivalent:
(i) 𝐼 is locally principal;

(ii) Nm 𝐵 |𝐹 (𝐼) = [OL (𝐼) : 𝐼] 𝑅 ; and
(iii) Nm 𝐵 |𝐹 (𝐼) = [OR (𝐼) : 𝐼] 𝑅 .
If 𝐵 is simple with dim𝐹 𝐵 = 𝑛2 , then these are further equivalent to
(iv) nrd(𝐼) 𝑛 = [OL (𝐼) : 𝐼] 𝑅 .

(v) nrd(𝐼) 𝑛 = [OR (𝐼) : 𝐼] 𝑅 .
Proof. Let O = OL (𝐼). Let 𝛼 ∈ 𝐼. Right multiplication by 𝛼 gives an 𝑅-module

isomorphism O −→ ∼ O𝛼 (change of basis between two free 𝑅-modules), so by Lemma
9.6.4 we have [O : O𝛼] 𝑅 = det(𝛼)𝑅 = Nm 𝐵 |𝐹 (𝛼)𝑅, thinking of 𝛼 ∈ End𝐹 (𝐵).
We now prove (i) ⇔ (ii). We may suppose 𝑅 is local, so 𝑅 is a DVR. For all 𝛼 ∈ 𝐼,
the following equality holds:
[O : 𝐼] 𝑅 [𝐼 : O𝛼] 𝑅 = [O : O𝛼] 𝑅 = Nm 𝐵 |𝐹 (𝛼)𝑅. (16.4.4)
To show (i) ⇒ (ii), if 𝐼 = O𝛼 then Nm 𝐵 |𝐹 (𝐼) = Nm 𝐵 |𝐹 (𝛼)𝑅 and so by cancelling

[𝐼 : O𝛼] 𝑅 = 𝑅 in (16.4.4) we obtain (ii). To show (ii) ⇒ (i), suppose that Nm 𝐵 |𝐹 (𝐼) =
[O : 𝐼] 𝑅 . Let 𝛼 ∈ 𝐼 be such that Nm 𝐵 |𝐹 (𝛼) has minimal valuation; then Nm 𝐵 |𝐹 (𝛼)
generates Nm 𝐵 |𝐹 (𝐼). By (16.4.4), cancelling on both sides [𝐼 : O𝛼] 𝑅 = 𝑅, and since
O𝛼 ⊆ 𝐼 we conclude 𝐼 = O𝛼. A similar argument holds on the right, proving (i) ⇔
(iii). Finally, (iii) ⇔ (iv) since Nm 𝐵 |𝐹 (𝛼) = nrd(𝛼) 𝑛 , and the same on the right.
16.4.5. Recalling the proof of Proposition 16.4.3 and the definition of 𝑅-index, we
always have the containment
Nm 𝐵 |𝐹 (𝐼) ⊇ [OL (𝐼) : 𝐼] 𝑅
and the same on the right; by Propostion 16.4.3, equality is equivalent to 𝐼 being locally
principal.
To conclude this section, we suppose for its remainder that 𝐹 is a local with
valuation ring 𝑅 or a global number field. Then the reduced norm is also related to the
absolute norm, an absolute measure of size, as follows.
16.4.6. For a fractional ideal 𝔞 of 𝑅, we define the absolute norm N (𝔞) to be
N (𝔞) := [𝑅 : 𝔞] Z < ∞ (16.4.7)
the index taken as abelian groups, recalling Example 9.6.6. Then
N (𝔞) = |Nm𝐹 /Q (𝔞)|
and if 𝔞 ⊆ 𝑅 then
N (𝔞) = #(𝑅/𝔞),
so this norm is also called the counting norm.
We extend this definition to elements 𝑎 ∈ 𝐹 × by defining N (𝑎) := N (𝑎𝑅).
16.4.8. Similarly, if 𝐼 ⊆ 𝐵 is a locally principal 𝑅-lattice, we define the absolute norm
of 𝐼 to be
N (𝐼) := [OL (𝐼) : 𝐼] Z = [OR (𝐼) : 𝐼] Z , (16.4.9)
the latter equality by taking 𝑅 = Z in Proposition 16.4.3. If 𝐼 is integral then
N (𝐼) = #(OL (𝐼)/𝐼) = #(OR (𝐼)/𝐼).
The absolute norm of 𝐼 is compatible with the absolute norm on 𝑅 via
N (𝐼) = [OR (𝐼) : 𝐼] Z = [𝑅 : Nm 𝐵 |𝐹 (𝐼)] Z = N (Nm 𝐵 |𝐹 (𝐼));

16.5. INVERTIBLE LATTICES 257
and if 𝐵 is simple with dim𝐹 𝐵 = 𝑛2 then
N (𝐼) = N (Nm 𝐵 |𝐹 (𝐼)) = N (nrd(𝐼)) 𝑛 . (16.4.10)
Remark 16.4.11. The absolute norm may also be defined for a global function field,
but there is no canonical ‘ring of integers’ as above.
16.5 Invertible lattices

We are now in a position to investigate the class of invertible lattices. Let 𝐼 ⊆ 𝐵 be an
𝑅-lattice.
Definition 16.5.1. 𝐼 is invertible if there exists an 𝑅-lattice 𝐼 0 ⊆ 𝐵 that is a (two-sided)
inverse to 𝐼, i.e.
𝐼 𝐼 0 = OL (𝐼) = OR (𝐼 0) and 𝐼 0 𝐼 = OL (𝐼 0) = OR (𝐼). (16.5.2)
In particular, both of the products in (16.5.2) are compatible.

16.5.3. If 𝐼, 𝐽 are invertible lattices and 𝐼 is compatible with 𝐽, then 𝐼𝐽 is invertible
(Exercise 16.10).
16.5.4. If 𝐼 is a principal lattice, then 𝐼 is invertible: if 𝐼 = O𝛼 with 𝛼 ∈ 𝐵× and
O = OL (𝐼), then 𝐼 0 = 𝛼−1 O has
𝐼 𝐼 0 = (O𝛼) (𝛼−1 O) = O(𝛼𝛼−1 )O = OO = O
so 𝐼 0 is a right inverse, and
𝐼 0 𝐼 = (𝛼−1 O)(O𝛼) = 𝛼−1 O𝛼 = OR (𝐼)
so 𝐼 0 is also a left inverse.

A candidate for the inverse presents itself quite naturally. If 𝐼 𝐼 0 = OL (𝐼) and
𝐼 0𝐼 = OR (𝐼), then 𝐼 𝐼 0 𝐼 = 𝐼.
Definition 16.5.5. We define the quasi-inverse of 𝐼 as
𝐼 −1 := {𝛼 ∈ 𝐵 : 𝐼𝛼𝐼 ⊆ 𝐼}. (16.5.6)

(a) The quasi-inverse 𝐼 −1 is an 𝑅-lattice and
𝐼 𝐼 −1 𝐼 ⊆ 𝐼.
(b) If O is an 𝑅-order, then O−1 = O.
Proof. Statement (a) follows as in the proof of Lemma 10.2.7, and the inclusion is by
the definition of 𝐼 −1 . For statement (b), if 𝛼 ∈ O, then O𝛼O ⊆ O since O is an order;
conversely, if O𝛼O ⊆ O, then taking 1 ∈ O on left and right we conclude 𝛼 ∈ O.
We now consider the quasi-inverse as an inverse.
(i) 𝐼 −1 is a (two-sided) inverse for 𝐼;

(ii) 𝐼 −1 𝐼 = OR (𝐼) and 𝐼 𝐼 −1 = OL (𝐼);
(iii) 𝐼 is invertible;
(iv) There is a compatible product 𝐼 𝐼 −1 𝐼 = 𝐼 and both 1 ∈ 𝐼 𝐼 −1 and 1 ∈ 𝐼 −1 𝐼.
Proof. The implication (i) ⇒ (ii) is clear. For (ii) ⇒ (i), we need to check the
compatibility of the product: but since 𝐼 −1 𝐼 = OR (𝐼) we have OL (𝐼 −1 ) ⊆ OR (𝐼), and
from the other direction we have the other containment, so these are equal.
The implication (i) ⇒ (iii) is clear. For (iii) ⇒ (i), suppose that 𝐼 0 is an inverse
to 𝐼. Then 𝐼 = 𝐼 𝐼 0 𝐼 so 𝐼 0 ⊆ 𝐼 −1 by definition. Therefore 𝐼 ⊆ 𝐼 𝐼 −1 𝐼 ⊆ 𝐼 and equality
holds throughout. Multiplying by 𝐼 0 on the left and right then gives
𝐼 −1 = (𝐼 0 𝐼)𝐼 −1 (𝐼 𝐼 0) = 𝐼 0 𝐼 𝐼 0 = 𝐼 0 .
Again the implication (i) ⇒ (iv) is immediate. To prove (iv) ⇒ (ii), we need to
show that 𝐼 𝐼 −1 = OL (𝐼) and 𝐼 −1 𝐼 = OR (𝐼); we show the former. By compatibility,
OR (𝐼 −1 ) = OL (𝐼) = O. If 𝐼 𝐼 −1 = 𝐽 then 𝐽 = 𝐼 𝐼 −1 = O(𝐼 𝐼 −1 )O = O𝐽O, so 𝐽 ⊆ O is a
two-sided ideal of O containing 1 hence 𝐽 = O.
Invertibility is a local property, as one might expect.
Lemma 16.5.9. 𝐼 is invertible if and only 𝐼 (𝔭) is invertible for all primes 𝔭.
Proof. We employ Proposition 16.5.8(iv): We have 𝐼 𝐼 −1 𝐼 = 𝐼 if and only if
(𝐼 𝐼 −1 𝐼) (𝔭) = 𝐼 (𝔭) (𝐼 −1 ) (𝔭) 𝐼 (𝔭) = 𝐼 (𝔭)
for all primes 𝔭 and e.g. 1 ∈ 𝐼 𝐼 −1 if and only if 1 ∈ 𝐼 (𝔭) 𝐼 −1

(𝔭)
.
Corollary 16.5.10. If 𝐼 is locally principal, then 𝐼 is invertible.
Proof. Combine 16.5.4 with Lemma 16.5.9.
A compatible product with an invertible lattice respects taking left (and right)
orders, as follows.
Lemma 16.5.11. If 𝐼 is compatible with 𝐽 and 𝐽 is invertible, then OL (𝐼𝐽) = OL (𝐼).
Proof. We always have OL (𝐼) ⊆ OL (𝐼𝐽) (even without 𝐽 invertible). To show the
other containment, suppose that 𝛼 ∈ OL (𝐼𝐽), so that 𝛼𝐼𝐽 ⊆ 𝐼𝐽. Multiplying by 𝐽 −1 ,
we conclude 𝛼𝐼 ⊆ 𝐼 and 𝛼 ∈ OL (𝐼).
Finally, not every lattice is invertible, and it is helpful to have counterexamples at

hand (see also Exercise 16.12).
16.5. INVERTIBLE LATTICES 259

𝑝, 𝑝
Example 16.5.12. Let 𝑝 ∈ Z be prime. Let 𝐵 := and
Q
O := Z ⊕ 𝑝Z𝑖 ⊕ 𝑝Z 𝑗 ⊕ Z𝑖 𝑗
𝐼 := 𝑝 2 Z ⊕ Z𝑖 ⊕ Z 𝑗 ⊕ Z𝑖 𝑗 .
Then O ⊂ 𝐵 is an order and OL (𝐼) = OR (𝐼) = O. We compute that
𝐼 −1 = 𝑝Z ⊕ Z𝑖 ⊕ Z 𝑗 ⊕ Z𝑖 𝑗 (16.5.13)
and
1 1
OL (𝐼 −1 ) = OR (𝐼 −1 ) = Z + Z𝑖 + Z 𝑗 + Z𝑖 𝑗 = Z + O; (16.5.14)
𝑝 𝑝
so in the product
𝐼 𝐼 −1 = 𝐼 −1 𝐼 = 𝑝Z ⊕ 𝑝Z𝑖 ⊕ 𝑝Z 𝑗 ⊕ Z𝑖 𝑗 ( O (16.5.15)
we see 𝐼 is not invertible and the product is not compatible.

Seen a different way, we have 𝐼 = 𝐼 and in the compatible product
𝐼 2 = 𝐼 𝐼 = 𝐼 𝐼 = 𝑝Z ⊕ 𝑝Z𝑖 ⊕ 𝑝Z 𝑗 ⊕ Z𝑖 𝑗 (16.5.16)
we have 𝑖, 𝑗 ∈ OL (𝐼 2 ) = OR (𝐼 2 ) but 𝑖, 𝑗 ∉ O; therefore, 𝐼 is not invertible by Lemma

16.5.11. Indeed,
1 1
OL (𝐼 2 ) = OR (𝐼 2 ) = 𝐼 2 = Z + O.
𝑝 𝑝
Finally, it will convenient to consider invertibility in the context of ideals, labelling
left and right orders as follows.
Definition 16.5.17. Let O, O0 ⊆ 𝐵 be 𝑅-orders and let 𝐼 be a fractional O, O0-ideal.
We say 𝐼 is invertible if 𝐼 is invertible as a lattice and 𝐼 is sated (i.e., O = OL (𝐼) and
O0 = OR (𝐼)).
16.5.18. The condition that 𝐼 is sated in Definition 16.5.17 is important: we must be
careful to work over left and right orders and not some smaller order. Indeed, if 𝐼 is
invertible as an 𝑅-lattice then it is invertible as a fractional OL (𝐼), OR (𝐼)-ideal, but not
for any strictly smaller orders. If 𝐼 0 is an 𝑅-lattice and 𝐼 𝐼 0 = O for some O ⊆ OL (𝐼),
then multiplying on both sides on the left by OL (𝐼) gives
O = 𝐼 𝐼 0 = OL (𝐼)𝐼 𝐼 0 = OL (𝐼)O = OL (𝐼)
and the same on the right. In other words, if we are going to call out an invertible
fractional ideal by labelling actions on left and right, then we require these labels to
be the actual orders that make the inverse work.
Remark 16.5.19. Example 16.1.1 suggested the ‘real issue’ with noninvertible modules
for quadratic orders: as an abelian group,
√
𝔣 = 𝑓 Z + 𝑓 𝑑Z = 𝑓 · 𝑆(𝑑),
so 𝔣 is principal and hence certainly invertible as an ideal of√𝑆(𝑑)—but not as an ideal

of the smaller order 𝑆(𝑑). More generally, if 𝔞 ⊂ 𝐾 = Q( 𝑑) is a lattice in 𝐾 (free
Z-module of rank 2), we define its multiplicator ring as
𝑆(𝔞) := {𝑥 ∈ 𝐾 : 𝑥𝔞 ⊆ 𝔞};
the ring 𝑆(𝔞) is an order√of 𝐾 and so is also called the order of 𝔞. In the example
above, 𝑆(𝔣) = 𝑆( 𝑓 (Z + 𝑑 𝐾 Z)) = 𝑆(𝑑 𝐾 ) ) 𝑆(𝑑). It turns out that every lattice in
𝐾 is invertible as an ideal of its multiplicator ring [Cox89, Proposition 7.4], and this
statement plays an important role in the theory of complex multiplication. (Sometimes,
an ideal 𝔞 ⊆ 𝑆 is called proper or regular if 𝑆 = 𝑆(𝔞); both terms are overloaded in
mathematics, so we will mostly resist this notion.)
Unfortunately, unlike the quadratic case, not every lattice 𝐼 ⊂ 𝐵 is projective as
a left module over its left order (or the same on the right): this is necessary, but not
sufficient. In Chapter 24, we classify orders O with the property that every lattice 𝐼
having OL (𝐼) = O is projective as an O-module: they are the Gorenstein orders.
Remark 16.5.20. Invertible lattices give rise to a Morita equivalence between their
corresponding left and right orders: see Remark 7.2.20.
16.6 Invertibility with a standard involution

In section 16.6, we follow Kaplansky [Kap69], considering invertibility in the presence
of a standard involution. The main result of this chapter is as follows.
Main Theorem 16.6.1. Let 𝑅 be a Dedekind domain with field of fractions 𝐹, and let
𝐵 be a finite-dimensional 𝐹-algebra with a standard involution. Then an 𝑅-lattice 𝐼 is
invertible if and only if 𝐼 is locally principal.
Remark 16.6.2. We can relax the hypothesis that 𝑅 is a Dedekind domain and in-
stead work with a Prüfer domain, a generalization of Dedekind domains to the non-
noetherian context.
We have already seen (Corollary 16.5.10) that the implication (⇒) in Main Theo-
rem 16.6.1 holds without the hypothesis of a standard involution; the reverse implica-
tion is the topic of this section. This implication is not in general true if this hypothesis
is removed (but is true again when 𝐵 is commutative); see Exercise 16.18(a).
Remark 16.6.3. The provenance of the hypothesis that 𝑅 is a Dedekind domain is the
following: if 𝔞 ⊂ 𝑅 is not invertible as an 𝑅-module, and O ⊂ 𝐵 is an 𝑅-order, then
𝔞O is not invertible as an 𝑅-lattice. To make the simplest kind of arguments here, we
would like for all (nonzero) ideals 𝔞 ⊆ 𝑅 to be invertible, and this is equivalent to the
requirement that 𝑅 is a Dedekind domain (see section 9.2).
Throughout this section, let 𝑅 be a Dedekind domain with field of fractions 𝐹, let
𝐵 be a finite-dimensional 𝐹-algebra, and let 𝐼 ⊂ 𝐵 be an 𝑅-lattice. The following
concept will be useful in this section.
Definition 16.6.4. We say 𝐼 is a semi-order if 1 ∈ 𝐼 and nrd(𝐼) ⊆ 𝑅.

16.6. INVERTIBILITY WITH A STANDARD INVOLUTION 261
(For a semi-order 𝐼, we necessarily have nrd(𝐼) = 𝑅 since 1 ∈ 𝐼.)
Lemma 16.6.5. An 𝑅-lattice 𝐼 is a semi-order if and only if 1 ∈ 𝐼 and every 𝛼 ∈ 𝐼 is

integral over 𝑅.
Proof. We have that 𝛼 ∈ 𝐼 is integral over 𝑅 if and only if trd(𝛼) ∈ 𝑅 and nrd(𝛼) ∈ 𝑅
(by Corollary 10.3.6, since 𝑅 is integrally closed) if and only if nrd(𝛼) ∈ 𝑅 and
nrd(𝛼 + 1) = nrd(𝛼) + trd(𝛼) + 1 ∈ 𝑅.
In particular, Lemma 16.6.5 implies that an order is a semi-order (by Corollary

10.3.3); we will see that semi-orders behave enough like orders that we can deduce
local principality from their structure.
16.6.6. Let 𝐼 := {𝛼 : 𝛼 ∈ 𝐼}. Then 𝐼 is an 𝑅-lattice in 𝐵. If 𝐼, 𝐽 are 𝑅-lattices then

𝐼𝐽 = 𝐽 𝐼 (even if this product is not compatible).
If 𝐼 is a semi-order, then 𝐼 = 𝐼 (Exercise 16.15). In particular, if O is an 𝑅-order
then O = O.
Lemma 16.6.7. We have OL (𝐼) = OR (𝐼) and OR (𝐼) = OL (𝐼).
Proof. We have 𝛼 ∈ OL (𝐼) if and only if 𝛼𝐼 ⊆ 𝐼 if and only if 𝛼𝐼 = 𝐼 𝛼 ⊆ 𝐼 if and

only if 𝛼 ∈ OR (𝐼) if and only if 𝛼 ∈ OR (𝐼) = OR (𝐼).
Corollary 16.6.8. If 𝐼 is a semi-order, then OL (𝐼) = OR (𝐼).
Proof. Apply Lemma 16.6.7 with 𝐼 = 𝐼.
By Lemma 16.6.7, the standard involution gives a bĳection between the set of
lattices 𝐼 with OL (𝐼) = O and the set of lattices with OR (𝐼) = O.
16.6.9. Suppose that 𝑅 is a DVR (e.g., a localization of 𝑅 at a prime ideal 𝔭). We will
show how to reduce the proof of Main Theorem 16.6.1 to that of a semi-order.
Since 𝑅 is a DVR, the fractional 𝑅-ideal nrd(𝐼) ⊆ 𝑅 is principal, generated by an
element with minimal valuation: let 𝛼 ∈ 𝐼 achieve this minimum reduced norm. Then
the 𝑅-lattice 𝐽 = 𝛼−1 𝐼 now satisfies 1 ∈ 𝐽 and nrd(𝐽) = 𝑅. Thus 𝐽 is a semi-order,
and 𝐽 is (locally) principal if and only if 𝐼 is (locally) principal.
Proof of Main Theorem 16.6.1. The proof is due to Kaplansky [Kap69, Theorem 2].
The statement is local; localizing, we may suppose 𝑅 is a DVR. By 16.6.9, we reduce
to the case where 𝐼 is a semi-order. In particular, we have 1 ∈ 𝐼. Let 𝛼1 , . . . , 𝛼𝑛 be an
𝑅-basis for 𝐼.
We claim that
𝐼 𝑛+1 = 𝐼 𝑛 (16.6.10)
Since 1 ∈ 𝐼, we have 𝐼 𝑛 ⊆ 𝐼 𝑛+1 . It suffices then to prove that a product of 𝑛 + 1 basis
elements of 𝐼 lies in 𝐼 𝑛 . By the pigeonhole principle, there must be a repeated term 𝛼𝑖
among them. We recall the formula (4.2.16)
𝛼𝛽 + 𝛽𝛼 = trd(𝛽)𝛼 + trd(𝛼) 𝛽 − trd(𝛼𝛽) (16.6.11)

for all 𝛼, 𝛽 ∈ 𝐵. We can use this relation to “push” the second instance of the repeated
element until it meets with its mate, at the expense of terms lying in 𝐼 𝑛 . More precisely,
in the 𝑅-module 𝐼 2 /𝐼, by (16.6.11),
𝛼𝑖 𝛼 𝑗 ≡ −𝛼 𝑗 𝛼𝑖 (mod 𝐼)
for all 𝑖, 𝑗; it follows that in 𝐼 𝑛+1 /𝐼 𝑛 ,
𝜇(𝛼𝑖 𝛼 𝑗 )𝜈 ≡ −𝜇(𝛼 𝑗 𝛼𝑖 )𝜈 (mod 𝐼 𝑛 )
for all 𝜇, 𝜈 appropriate products of basis elements. Therefore we may suppose that
the repetition 𝛼𝑖2 is adjacent; but then 𝛼𝑖 satisfies a quadratic equation and 𝛼𝑖2 =
trd(𝛼𝑖 )𝛼𝑖 − nrd(𝛼𝑖 ) ∈ 𝐼, so in fact the product belongs to 𝐼 𝑛 , and the claim follows.
Now suppose 𝐼 is invertible; we wish to show that 𝐼 is principal. From the
equality 𝐼 𝑛+1 = 𝐼 𝑛 , we multiply both sides of this equation by (𝐼 −1 ) 𝑛 and obtain
𝐼 = O = OL (𝐼) = OR (𝐼). In particular, 𝐼 is principal, generated by 1.
The above proof has the following immediate corollary.
Corollary 16.6.12. An 𝑅-lattice 𝐼 is an 𝑅-order if and only if 1 ∈ 𝐼, every element of

𝐼 is integral, and 𝐼 is invertible. In particular, an invertible semi-order is an order.
We conclude with two consequences.
16.6.13. Let 𝐼, 𝐽 be invertible 𝑅-lattices such that 𝐼 is compatible with 𝐽. Then

nrd(𝐼𝐽) = nrd(𝐼) nrd(𝐽), since it is enough to check this locally, and locally both 𝐼
and 𝐽 are principal and we have proved the statement in this case (Lemma 16.3.7).
16.6.14. In the presence of a standard involution, we can write the inverse in another
way: if 𝐼 is invertible, then
𝐼 𝐼 = nrd(𝐼)OR (𝐼) and 𝐼 𝐼 = nrd(𝐼)OL (𝐼)
by checking these statements locally (where they follow immediately by computing

the norm on a local generator). Since nrd(𝐼) is a fractional 𝑅-ideal and thus invertible
(𝑅 is a Dedekind domain), it follows that if 𝐼 is invertible, then
𝐼 −1 = 𝐼 nrd(𝐼) −1 .
In view of 16.6.14, the following important proposition is natural.
Proposition 16.6.15. Let 𝐵 be a quaternion algebra over 𝐹 and let 𝐼 ⊂ 𝐵 be an

𝑅-lattice. Then the following statements hold.
(a) We have 𝐼 𝐼 = nrd(𝐼)O, where O ⊆ 𝐵 an 𝑅-order satisfying OL (𝐼) ⊆ O, and

similarly 𝐼 𝐼 = nrd(𝐼)O0 with OR (𝐼) ⊆ O0.
(b) If either OL (𝐼) or OR (𝐼) is maximal, then 𝐼 is invertible, and both OL (𝐼) and
OR (𝐼) are maximal.
16.7. ONE-SIDED INVERTIBILITY 263
Proof. We follow Kaplansky [Kap69, Theorems 6–7]. We again may suppose 𝑅 is a

DVR and 𝐼 is a semi-order, so 1 ∈ 𝐼 and nrd(𝐼) = 𝑅; and 𝐼 = 𝐼.
First we prove (a). We need to show that 𝐼 2 is an order. We showed in (16.6.10)
(without extra hypothesis) that 𝐼 3 = 𝐼 4 ; with 𝐵 a quaternion algebra, we will im-
prove this to 𝐼 2 = 𝐼 3 , whence (𝐼 2 ) 2 = 𝐼 4 = 𝐼 2 and consequently 𝐼 2 is closed under
multiplication and hence an 𝑅-order.
Let 𝐽 = 𝐼 3 ; then 𝐽 2 = (𝐼 3 ) 2 = 𝐼 6 = 𝐼 3 = 𝐽, so 𝐽 is an 𝑅-order. Let 𝔭 be the
maximal ideal of 𝑅 and consider the 4-dimensional algebra 𝐽/𝔭𝐽 over 𝑘 = 𝑅/𝔭. Then
𝐼/𝔭𝐼 ⊆ 𝐽/𝔭𝐽 is a 𝑘-subspace containing 1. If (𝐼/𝔭𝐼) 2 = 𝐼/𝔭𝐼, then by dimensions we
contradict (𝐼/𝔭𝐼) 3 = 𝐽/𝔭𝐽; therefore (𝐼/𝔭𝐼) 2 ) 𝐼/𝔭𝐼. If dim 𝑘 (𝐼/𝔭𝐼) ≤ 2, then 𝐼/𝔭𝐼
is a proper 𝑘-subalgebra, impossible. Thus dim 𝑘 (𝐼/𝔭𝐼) ≥ 3 and dim 𝑘 (𝐼/𝔭𝐼) 2 ≥ 4,
and so (𝐼/𝔭𝐼) 2 = 𝐽/𝔭𝐽. By Nakayama’s lemma, it follows that 𝐼 2 = 𝐽 = 𝐼 3 . The
containments follow directly, e.g. OL (𝐼) ⊆ OL (𝐼 𝐼) = OL (O) = O.
For part (b), applying part (a) we have O = OL (𝐼) by maximality and the same on
the right. But since 𝐼 is a semi-order, from Corollary 16.6.8 we have 𝐼 𝐼 = OL (𝐼) =
OR (𝐼) = 𝐼 𝐼 so by definition, 𝐼 has inverse 𝐼.
16.7 One-sided invertibility

In this section, we pause to consider one-sided notions of invertibility. We refresh
our notation, recalling that 𝑅 is a Dedekind domain with 𝐹 = Frac 𝑅 and 𝐵 is a
finite-dimensional algebra over 𝐹 with 𝐼 ⊆ 𝐵 an 𝑅-lattice.
Definition 16.7.1. 𝐼 is right invertible if there exists an 𝑅-lattice 𝐼 0 ⊆ 𝐵, a right

inverse, such that the product 𝐼 𝐼 0 is compatible and 𝐼 𝐼 0 = OL (𝐼).
A right fractional O-ideal 𝐼 is right invertible if 𝐼 is right invertible and sated (viz.
16.5.18).
We similarly define left invertible and left inverse. Applying the same reasoning
as in Lemma 16.5.9, we see that one-sided invertibility is a local property.
Remark 16.7.2. For rings, the (left or) right inverse of an element need not be unique
even though a two-sided inverse is necessarily unique. Similarly, left invertibility does
not imply right invertibility for lattices in general, and so the one-sided notions can be
a bit slippery: see Exercise 16.18(b).
Remark 16.7.3. The compatibility condition in invertibility is important to avoid
trivialities.
Consider
Example 16.3.6: we have 𝐼𝐽 = M2 (𝑅) = OL (𝐼), and if we let
𝑏𝑅 𝑏𝑅
𝐽 = for any nonzero 𝑏 ∈ 𝑅, the equality 𝐼𝐽 = M2 (𝑅) remains true. Not
𝑅 𝑅
every author requires compatibility in the definition of (sided) invertibility.
A natural candidate for the right inverse presents itself: if 𝐼 𝐼 0 = OL (𝐼), then 𝐼 0
maps 𝐼 into OL (𝐼) on the right. We recall the definition of the colon lattices (Definition
15.6.11). Let 𝐼 0 := (OL (𝐼) : 𝐼) R . Then 𝐼 𝐼 0 ⊆ OL (𝐼) by definition; however, in general
equality need not hold and the product need not be compatible. Similarly, since
𝐼 𝐼 −1 𝐼 ⊆ 𝐼 we have 𝐼 𝐼 −1 ⊆ OL (𝐼), but again equality need not hold.
The sided version of Proposition 16.5.8 also holds.

(i) 𝐼 −1 is a right inverse for 𝐼;
(ii) 𝐼 is right invertible;
(iii) There is a compatible product 𝐼 𝐼 −1 𝐼 = 𝐼 and 1 ∈ 𝐼 𝐼 −1 .
Similar equivalences hold on the left.
Proof. This is just a sided restriction of the proof of Proposition 16.5.8. For example,
to show (ii) ⇒ (i), we always have 𝐼 𝐼 −1 𝐼 ⊆ 𝐼 so 𝐼 𝐼 −1 ⊆ OL (𝐼); if 𝐼 0 is a right inverse to
𝐼, then 𝐼 𝐼 0 𝐼 = OL (𝐼)𝐼 = 𝐼 and 𝐼 0 ⊆ 𝐼 −1 , and therefore 𝐼 𝐼 −1 ⊇ 𝐼 𝐼 0 = OL (𝐼). Therefore
𝐼 𝐼 −1 = OL (𝐼) and 𝐼 −1 is a right inverse for 𝐼.
Returning to the setting of the previous section, however, we can show that the
one-sided notions of invertibility are equivalent to the two-sided notion.
Lemma 16.7.5. Suppose 𝐵 has a standard involution. Then an 𝑅-lattice 𝐼 is left
invertible if and only if 𝐼 is right invertible if and only if 𝐼 is invertible.
Proof. We will show that if 𝐼 is right invertible then 𝐼 is left invertible; the other
implications follow similarly. By localizing, we reduce to the case where 𝑅 is a DVR.
By the results of 16.6.9, we may suppose that 𝐼 is a semi-order, so that OL (𝐼) =
OR (𝐼) = O and 𝐼 = 𝐼. Suppose 𝐼 𝐼 0 = O. Then 𝐼 0 𝐼 = O = O, and 𝐼 0 is compatible with
𝐼 since
O = OR (𝐼) = OL (𝐼 0) = OR (𝐼 0)
as desired.
Corollary 16.7.6. Suppose 𝑅 is a Dedekind domain and that 𝐵 has a standard in-
volution. Then an 𝑅-lattice 𝐼 is right invertible with 𝐼 𝐼 0 = OL (𝐼) if and only if
𝐼 0 = (OL (𝐼) : 𝐼) R = 𝐼 −1 .
A similar statement holds for the left inverse; in particular, this shows that a right
inverse is necessarily unique.
Proof. The implication (⇒) is immediate, so we prove (⇐). Let O = OL (𝐼). Then
O = 𝐼 𝐼 0 ⊆ 𝐼 (O : 𝐼) R ⊆ O
so equality must hold, and 𝐼 𝐼 0 = 𝐼 (O : 𝐼) R . By 16.7.5, 𝐼 is invertible, and multiplying

both sides by 𝐼 −1 gives 𝐼 0 = (O : 𝐼) R .
We collect the results of this section in the following theorem.

Main Theorem 16.7.7. Let 𝑅 be a Dedekind domain with 𝐹 = Frac 𝑅, let 𝐵 be a
quaternion algebra over 𝐹, and let 𝐼 ⊆ 𝐵 be an 𝑅-lattice. Then the following are
equivalent:
(i) 𝐼 is locally principal;
(iii) 𝐼 is left invertible;
16.8. INVERTIBILITY AND THE CODIFFERENT 265
(iii0) 𝐼 is right invertible;

(iv) nrd(𝐼) 2 = [OL (𝐼) : 𝐼] 𝑅 ; and
(iv0) nrd(𝐼) 2 = [OR (𝐼) : 𝐼] 𝑅 .
Proof. Main Theorem 16.6.1 proves (i) ⇔ (ii). For the equivalence (ii) ⇔ (iii) ⇔
(iii0), apply Lemma 16.7.5. Finally, the equivalence (i) ⇔ (iv) ⇔ (iv0) is supplied by
Proposition 16.4.3.
16.8 Invertibility and the codifferent

To conclude this chapter, we pick up a remaining thread concerning the (co)different.
Definition 16.8.1. We define the different of O to be the quasi-inverse of the codif-

ferent:
diff(O) := codiff (O) −1 = {𝛼 ∈ 𝐵 : O♯ 𝛼O♯ ⊆ O♯ }.
Lemma 16.8.2. The different diff (O) is an integral two-sided O-ideal.
Proof. By Lemma 15.6.16, we have OO♯ O = O♯ and so if 𝛼 ∈ diff(O) then

O♯ (O𝛼O)O♯ = O♯ 𝛼O♯ = O♯ and diff (O) is a two-sided O-ideal. To prove that
diff (O) ⊆ O, referring to Lemma 15.6.2, starting with O♯ 𝛼O♯ ⊆ O♯ taking 1 ∈ O♯
we have 𝛼O♯ ⊆ O♯ so (𝛼O♯ ) ♯ = (O♯ ) ♯ 𝛼−1 ⊇ (O♯ ) ♯ . By Lemma 15.6.5, we have
(O♯ ) ♯ = O, so O𝛼−1 ⊇ O, so O𝛼 ⊆ O and again taking 1 we get 𝛼 ∈ O.
16.8.3. If codiff(O) is locally principal (section 16.2), then so is diff (O), and by
Proposition 16.4.3 we have
Nm 𝐵 |𝐹 (diff (O)) = [O : diff(O)] 𝑅 = [codiff(O) : O] 𝑅 = disc(O);
so when further 𝐵 is a quaternion algebra, we have
nrd(diff (O)) = discrd(O). (16.8.4)
Invertibility of ideals is detected by the (co)different [Fad65, Proposition 24.1].
Proposition 16.8.5. If codiff(O) is right invertible, then all sated left fractional O-
ideals are right invertible. Similarly, if codiff(O) is left invertible, then all sated right
fractional O-ideals are left invertible.
Proof. To get started, we refresh a few things: by Corollary 15.6.13, we have (𝐼 𝐼 ♯ ) ♯ =

OL (𝐼) = O. The product 𝐼 𝐼 ♯ is compatible by Proposition 15.6.6. By Lemma 15.6.5
we have 𝐼 𝐼 ♯ = O♯ = codiff (O).
Now by hypothesis of invertibility, O♯ (O♯ ) −1 = OL (O♯ ) = O is a compatible
product. Therefore the product 𝐼 ♯ (O♯ ) −1 is compatible, and
𝐼 (𝐼 ♯ (O♯ ) −1 ) = (𝐼 𝐼 ♯ ) (O♯ ) −1 = O♯ (O♯ ) −1 = O. (16.8.6)
A similar argument holds on the right.

We have the following corollary of Proposition 16.8.5, phrased in terms of the

different.
Corollary 16.8.7. Suppose that 𝐵 has a standard involution. Then the following are
equivalent:
(i) codiff(O) is invertible;

(ii) diff(O) is invertible;
(iii) All sated left fractional O-ideals 𝐼 are invertible, with inverse 𝐼 −1 = 𝐼 ♯ diff (O);
and
(iii0) All sated right fractional O-ideals 𝐼 are invertible, with inverse 𝐼 −1 = diff (O)𝐼 ♯ .
Proof. Combine Proposition 16.8.5 and (16.8.6) with Lemma 16.7.5 and Corollary
16.7.6.
We conclude with a criterion to determine invertibility; it is not used in the sequel.
Proposition 16.8.8 (Brandt’s invertibility criterion). Let 𝐼 ⊆ 𝐵 be an 𝑅-lattice. Then

𝐼 is invertible if and only if
nrd(𝐼 ♯ ) discrd(𝐼) ⊆ nrd(𝐼).
Proof. See Kaplansky [Kap69, Theorem 10] or Brzezinski [Brz82, Theorem 3.4].
Exercises
Unless otherwise specified, throughout these exercises let 𝑅 be a Dedekind domain
with field of fractions 𝐹, let 𝐵 be a finite-dimensional 𝐹-algebra, and let 𝐼 ⊆ 𝐵 be an
𝑅-lattice.
√
1. Let 𝑑 ∈ Z be a nonsquare discriminant, and let 𝑆(𝑑) = Z[(𝑑 + 𝑑)/2] be the
quadratic ring of discriminant 𝑑.
√
(a) Suppose that 𝑑 = 𝑑 𝐾 𝑓 2 with 𝑓 > 1. Show that the ideal ( 𝑓 , 𝑑) of 𝑆(𝑑)
is not invertible. √
(b) Consider 𝑑 = −12, and 𝑆 = 𝑆(−12) = Z[ −3]. Show that every invertible
ideal of 𝑆 is principal (so 𝑆 has class number 1), but that 𝑆 is not a PID.
I 2. Show that if 𝐼 = OL (𝐼)𝛼 with 𝛼 ∈ 𝐵× , then OR (𝐼) = 𝛼−1 OL (𝐼)𝛼.
I 3. Show that if 𝐽 is an 𝑅-lattice in 𝐵 and 𝜇 ∈ 𝐵× , then 𝜇𝐽 = 𝐽 if and only if
𝜇 ∈ OL (𝐽) × .
I 4. Show that if 𝛼 ∈ 𝐵 then nrd(𝛼𝐼) = nrd(𝛼) nrd(𝐼). Conclude that if 𝐼 is a
principal 𝑅-lattice, generated by 𝛼 ∈ 𝐼, then nrd(𝐼) = nrd(𝛼)𝑅.
5. Let 𝛼1 , . . . , 𝛼𝑛 generate 𝐼 as an 𝑅-module. Give an explicit example where
nrd(𝐼) is not generated by nrd(𝛼𝑖 ) (cf. Lemma 16.3.2). Moreover, show that for
an 𝑅-lattice 𝐼, there exists a set of 𝑅-module generators 𝛼𝑖 such that nrd(𝐼) is in
fact generated by nrd(𝛼𝑖 ).
16.8. INVERTIBILITY AND THE CODIFFERENT 267
I 6. Suppose that 𝑅 is a Dedekind domain, and let O ⊆ 𝐵 be an 𝑅-order. Let 𝐼 be

a locally principal right fractional O-ideal. Show that 𝐼 can be generated as a
right O-ideal by two elements, and in fact for 𝑎 ∈ nrd(𝐼) nonzero we can write
𝐼 = 𝑎O + 𝛽O with 𝛽 ∈ 𝐵× .
7. Let 𝐹 be a number field, let 𝑅 ⊆ 𝐹 be a (Z-)order, and let 𝔞 ⊆ 𝑅 be a nonzero
ideal. Show that 𝔞 is projective as an 𝑅-module if and only if 𝔞 is invertible if
and only if 𝔞 is locally principal. [These are all automatic when 𝑅 is a Dedekind
domain 9.4.5.]
8. If 𝐼, 𝐽 ⊆ 𝐵 are 𝑅-lattices with 𝐼 ⊆ 𝐽, is it true that 𝐼 −1 ⊇ 𝐽 −1 ?
9. Let 𝐼, 𝐽, 𝐾 ⊆ 𝐵 be 𝑅-lattices. Show that
((𝐼 : 𝐽) L : 𝐾) R = ((𝐼 : 𝐾) R : 𝐽) L .
I 10. Let 𝐼, 𝐽 ⊆ 𝐵 be 𝑅-lattices and suppose that 𝐼 is compatible with 𝐽. Show that 𝐼𝐽
is invertible (with (𝐼𝐽) −1 = 𝐽 −1 𝐼 −1 ) if 𝐼, 𝐽 are invertible, but the converse need
not hold.
11. Let 𝐼, 𝐽 ⊆ 𝐵 be 𝑅-lattices, and suppose that 𝐽 is invertible. Show that (𝐼 : 𝐽) L =
𝐼𝐽 −1 and (𝐼 : 𝐽) R = 𝐽 −1 𝐼.
12. Let 𝑝 be prime, let 𝐵 = ( 𝑝, 𝑝 | Q), and let O := Zh𝑖, 𝑗i = Z ⊕ Z𝑖 ⊕ Z 𝑗 ⊕ Z𝑖 𝑗.
(a) Let 𝐼 = {𝛼 ∈ O : 𝑝 | nrd(𝛼)}. Show that 𝐼 = 𝑝Z ⊕ Z𝑖 ⊕ Z 𝑗 ⊕ Z𝑖 𝑗.
(b) Show 𝐼 = O𝑖 + O 𝑗, that O is a two-sided O-ideal, and that [O : 𝐼] = 𝑝.
(c) Show that 𝐼 ( 𝑝) ≠ O ( 𝑝) 𝛼 for all 𝛼 ∈ 𝐼 ( 𝑝) . [Hint: show that if 𝛼 ∈ 𝐼, then
𝑝 2 | [O : O𝛼].]
(d) Compute that OL (𝐼) = Z + Z𝑖 + Z 𝑗 + Z(𝑖 𝑗/𝑝) ) O, and that 𝐼 = OL (𝐼)𝑖 =
OL (𝐼) 𝑗.
(e) Compute codiff(O) and diff (O) and show they are invertible.
[Compare Lemurell [Lem2011, Remark 6.4].]
I 13. Let 𝐾 be a separable quadratic field extension of 𝐹 and let 𝐼 ⊆ 𝐾 be an 𝑅-lattice.
Let O = OL (𝐼) = OR (𝐼).
(a) Show that 𝐼 𝐼 = 𝐼 𝐼 = nrd(𝐼)O. [Hint: argue as in Proposition 16.6.15.]
(b) Conclude that 𝐼 is invertible as a O-module.
14. Show that if 𝐼, 𝐽 ⊆ 𝐵 are locally principal (hence invertible) 𝑅-lattices, then
[𝐼 : 𝐽] 𝑅 = [𝐽 −1 : 𝐼 −1 ] 𝑅 .
I 15. Let 𝐵 be an 𝐹-algebra with a standard involution . Show that if 𝐼 is a semi-order

then 𝐼 = 𝐼.
16. Let 𝑅 be a Dedekind domain with field of fractions 𝐹, let 𝐾 ⊃ 𝐹 be a separable
quadratic field extension and let 𝑆 be an 𝑅-order in 𝐾. Let 𝑆 𝐾 be the integral
closure of 𝑅 in 𝐾.
(a) Show that there exists a (unique) ideal 𝔣 = 𝔣(𝑆) ⊂ 𝑆 𝐾 (called the conduc-
tor) such that 𝑆 = 𝑅 + 𝔣𝑆 𝐾 .
(b) Now let 𝔟 ⊂ 𝐾 be a fractional 𝑆-ideal. Show that the following are
equivalent:
(i) 𝔟 is a locally principal 𝑆-ideal;
(ii) 𝔟 is invertible as a fractional 𝑆-ideal, i.e., there exists a fractional ideal
𝔟−1 such that 𝔟𝔟−1 = 𝑆 (necessarily 𝔟−1 = (𝑆 : 𝔟));
(iii) There exists 𝑑 ∈ 𝐾 × such that 𝑑𝔟 + 𝔣 ∩ 𝑆 = 𝑆; and
(iv) 𝔟 is proper, i.e., 𝑆 = O(𝔟) = {𝑥 ∈ 𝐾 : 𝑥𝔟 ⊆ 𝔟}.
I 17. Let O ⊆ 𝐵 be an 𝑅-order.
(a) Let 𝛼 ∈ 𝐵× . Show that 𝐼 = O𝛼 is a lattice with OL (𝐼) = OR (𝐼) = O if
and only if 𝛼 ∈ 𝐵× and O𝛼 = 𝛼O. Conclude that the set of invertible
two-sided principal lattices 𝐼 with OL (𝐼) = OR (𝐼) = O forms a group.
(b) Show that the normalizer of O,
𝑁 𝐵× (O) = {𝛼 ∈ 𝐵× : 𝛼O𝛼−1 = O}
is the group generated by 𝛼 ∈ 𝐵× such that O𝛼 is a two-sided O-ideal.

18. The following example is due to Kaplansky [Kap69, pp. 220, 221]. Let 𝑅 be a
DVR with field of fractions 𝐹 and maximal ideal 𝔭 = 𝜋𝑅.
(a) Consider the 𝑅-lattice
𝜋𝑅 𝜋𝑅 𝑅
𝐼 = 𝜋𝑅 𝑅 ® ⊂ 𝐵 = M3 (𝐹)
© ª
𝜋𝑅
«𝑅 𝑅 𝑅¬
Show that 𝐼 is invertible but is not principal.

(b) Consider the 𝑅-lattice
𝜋𝑅 𝜋𝑅 𝑅
𝐼 = 𝜋 2 𝑅 𝜋2 𝑅 𝑅® ⊂ 𝐵 = M3 (𝐹)
© ª
« 𝑅 𝑅 𝑅¬
Show that 𝐼 is left invertible but is not right invertible.

Chapter 17
Classes of quaternion ideals
Having investigated the structure of lattices and ideals in Chapter 16, we now turn to
the study of their isomorphism classes.
17.1 ⊲ Ideal classes

For motivation, let 𝐾 be a quadratic number field and 𝑆 ⊆ 𝐾 an order. We say that
two invertible fractional ideals 𝔞, 𝔟 ⊂ 𝐾 of 𝑆 are in the same class, and write 𝔞 ∼ 𝔟, if
there exists 𝑐 ∈ 𝐾 × such that 𝑐𝔞 = 𝔟; we denote the class of a fractional ideal 𝔞 as [𝔞].
We have 𝔞 ∼ 𝔟 if and only if 𝔞 and 𝔟 are isomorphic as 𝑆-modules. The set Cl 𝑆 of
invertible fractional ideals is a group under multiplication, measuring the failure of 𝑆
to be a PID. The class group Cl 𝑆 is a finite abelian group, by Minkowski’s geometry of
numbers: every class in Cl 𝑆 is represented by an integral ideal 𝔞 ⊆ 𝑆 whose absolute
norm is bounded (depending on 𝑆, but independent of the class), and there are only
finitely many such ideals. For an introduction to orders in quadratic fields and their
class numbers, with further connections to quadratic forms, see Cox [Cox89, §7].
The first treatment of isomorphism classes of quaternion ideals was given by Brandt
[Bra28]. Let 𝐵 be a quaternion algebra over Q. In the consideration of classes of lattices
𝐼 ⊂ 𝐵, we make a choice and consider lattices as right modules—considerations on
the left are analogous, with the map 𝐼 ↦→ 𝐼 allowing passage between left and right.
We say that lattices 𝐼, 𝐽 ⊆ 𝐵 are in the same right class, and write 𝐼 ∼R 𝐽, if there
exists 𝛼 ∈ 𝐵× such that 𝛼𝐼 = 𝐽; equivalently, 𝐼 ∼R 𝐽 if and only if 𝐼 is isomorphic to
𝐽 as right modules over OR (𝐼) = OR (𝐽). The relation ∼R is evidently an equivalence
relation, and the class of a lattice 𝐼 is denoted [𝐼] R .
Let O ⊂ 𝐵 be an order. We define the right class set of O as
ClsR O := {[𝐼] R : 𝐼 ⊂ 𝐵 invertible and OR (𝐼) = O};
equivalently, ClsR O is the set of isomorphism classes of invertible right O-modules

in 𝐵. The standard involution induces a bĳection between ClsR O and the analogously
defined left class set ClsL O; working on the right from now on, we will often abbreviate
Cls O := ClsR O.
269
270 CHAPTER 17. CLASSES OF QUATERNION IDEALS
Unfortunately, the class set Cls O does not have the structure of a group: only
a pointed set, with distinguished element [O] R . One problem is the compatibility
of multiplication discussed in the previous chapter. But even if we allowed products
between incompatible lattices, the product need not be well-defined: the lattices 𝐼𝐽 and
𝐼𝛼𝐽 for 𝛼 ∈ 𝐵× need not be in the same class, because of the failure of commutativity.
(This is the reason we write ‘Cls’ instead of ‘Cl’, as a reminder that it is only a class
set.) In Chapter 19, we will describe the structure that arises naturally instead: a
partially defined product on classes of lattices, a groupoid.
In any case, using the same method of proof (geometry of numbers) as in the
commutative case, we will show that there exists a constant 𝐶 (depending on O)
such that every class in Cls O is represented by an integral ideal 𝐼 ⊆ O with N (𝐼) =
#(O/𝐼) ≤ 𝐶. As a consequence, we have the following fundamental theorem.
Theorem 17.1.1. Let 𝐵 be a quaternion algebra over Q and let O ⊂ 𝐵 be an order.
Then the right class set Cls O is finite.
Accordingly, we call # Cls O ∈ Z ≥1 the (right) class number of O.
Right class sets pass between orders as follows. Let O, O0 ⊂ 𝐵 be orders. If
O ' O0 are isomorphic as rings, then of course this isomorphism induces a bĳection
Cls O −→∼ Cls O0. In fact, O ' O0 if and only if there exists 𝛼 ∈ 𝐵× such that
O = 𝛼−1 O𝛼 by the Skolem–Noether theorem; for historical reasons, we say that
0
O, O0 are of the same type.

Note that 𝐼 = O𝛼 = 𝛼O0 has OL (𝐼) = O and OR (𝐼) = O0 (recalling 10.2.5).
With this in mind, more generally, we say that O0 is connected to O if there exists
an invertible lattice 𝐽 with OL (𝐽) = O and OR (𝐽) = O0, called a connecting ideal.
Because invertible lattices are locally principal, two orders are connected if and only
if they are locally of the same type (i.e., locally isomorphic). If O0 is connected to
O, then right multiplying by a O, O0-connecting ideal 𝐽 yields a bĳection
∼ Cls O0
Cls O −→
(17.1.2)
[𝐼] R ↦→ [𝐼𝐽] R
We define the genus of an order O ⊂ 𝐵 to be the set Gen O of orders in 𝐵 locally

isomorphic to O, and the type set Typ O of O to be the set of 𝑅-isomorphism classes
of orders in the genus of O. The map
Cls O → Typ O
(17.1.3)
[𝐼] R ↦→ class of OL (𝐼)
is a surjective map of sets, so the type set is finite: in other words, up to isomorphism,
there are only finitely many types of orders in the genus of O. All maximal orders in 𝐵
are in the same genus, so in particular there are only finitely many conjugacy classes
of maximal orders in 𝐵. In this way, the right class set of O also organizes the types
of orders arising from O.
The most basic question about the class number is of course its size (as a function
of O). In the case of quadratic fields, the behavior of the class group depends in a
significant way on whether the field is imaginary or real: for negative discriminant
17.2. MATRIX RING 271
√︁
𝑑 < 0, the Brauer–Siegel theorem provides that # Cl 𝑆 is approximately of size | 𝑑|;
in contrast, for positive discriminant 𝑑 > 0, one typically sees a small class group and
a correspondingly large fundamental unit, but this statement is notoriously difficult to
establish unconditionally.
The same dichotomy is at play in the case of quaternion algebras, and to state the
cleanest results we suppose that O is a maximal order. Let 𝐷 := disc 𝐵 = discrd(O)
be the discriminant of 𝐵. If 𝐵 is definite, which is to say ∞ ∈ Ram 𝐵, then 𝐵 is like
an imaginary quadratic field 𝐾: the norm is positive definite. In this case, # Cls O is
approximately of size 𝐷, a consequence of the Eichler mass formula, the subject of
Chapter 25. On the other hand, if 𝐵 is indefinite, akin to a real quadratic field, then
# Cls O = 1, this time a consequence of strong approximation, the subject of Chapter
28. Just as in the commutative case, estimates on the size of the class number use
analytic methods and so must wait until we have developed the required tools.
17.2 Matrix ring

To begin, we first consider classes of ideals for the matrix ring; here, we can use
methods from linear algebra before we turn to more general methods in the rest of the
chapter.
17.2.1. Let 𝑅 be a PID with field of fractions 𝐹, and let 𝐵 = M𝑛 (𝐹). By Corollary
10.5.5, every maximal order of 𝐵 = M𝑛 (𝐹) is conjugate to M𝑛 (𝑅). Moreover, every
two-sided ideal of M𝑛 (𝑅) is principal, generated by an element 𝑎 ∈ 𝐹 × (multiplying
a candidate ideal by matrix units, as in Exercise 7.5(b)), so the group of fractional
two-sided M𝑛 (𝑅)-ideals is canonically identified with the group of fractional 𝑅-ideals,
itself isomorphic to the free abelian group on the (principal) nonzero prime ideals of
𝑅.
Just as in the two-sided case, the right class set for M𝑛 (𝑅) is trivial.
Proposition 17.2.2. Let 𝑅 be a PID with field of fractions 𝐹, and let 𝐵 = M𝑛 (𝐹). Let
𝐼 ⊆ 𝐵 be an 𝑅-lattice with either OL (𝐼) or OR (𝐼) maximal. Then 𝐼 is principal, and
both OL (𝐼) and OR (𝐼) are maximal.
Proof. We may suppose 𝐼 is integral by rescaling by 𝑟 ∈ 𝑅. Replacing 𝐼 by the

transpose 𝐼 t = {𝛼t : 𝛼 ∈ 𝐼} interchanging left and right orders (Exercise 10.12) if
necessary, we may suppose that OL (𝐼) is maximal. Then, by Corollary 10.5.5, we
have OL (𝐼) = 𝛼−1 M𝑛 (𝑅)𝛼 with 𝛼 ∈ 𝐵× , so replacing 𝐼 by 𝛼−1 𝐼 we may suppose
OL (𝐼) = M𝑛 (𝑅).
Now we follow Newman [New72, Theorem II.5]. Let 𝛼1 , . . . , 𝛼𝑚 be 𝑅-module
generators for 𝐼. Consider the 𝑛𝑚 × 𝑛 matrix 𝐴 = (𝛼1 , . . . , 𝛼𝑚 ) t . By row reduction
over 𝑅 (Hermite normal form, proven as part of the structure theorem for finitely
generated modules over a PID), there exists 𝑄 ∈ GL𝑛𝑚 (𝑅) such that 𝑄 𝐴 = (𝛽, 0) t
and 𝛽 ∈ M𝑛 (𝑅). We will show that 𝐼 = M𝑛 (𝑅) 𝛽. Let 𝜈11 , . . . , 𝜈1𝑚 ∈ M𝑛 (𝑅) be
the block matrices in the top 𝑛 rows of 𝑄. Then 𝛽 = 𝜈11 𝛼1 + · · · + 𝜈1𝑚 𝛼𝑚 so 𝛽 ∈ 𝐼
and M𝑛 (𝑅) 𝛽 ⊆ 𝐼. Conversely, let 𝜇11 , . . . , 𝜇 𝑚1 ∈ M𝑛 (𝑅) be the block matrices in
the left 𝑛 columns of 𝑄 −1 ∈ GL𝑛𝑚 (𝑅). Since 𝑄 −1 (𝛽, 0) t = 𝐴, we have 𝜇𝑖1 𝛽 = 𝛼𝑖 so
𝛼𝑖 ∈ M𝑛 (𝑅) 𝛽 for 𝑖 = 1, . . . , 𝑚, thus 𝐼 ⊆ M𝑛 (𝑅) 𝛽. Therefore 𝐼 = M𝑛 (𝑅) 𝛽, and so

OR (𝐼) is maximal (16.2.3).
Returning to the case of quaternion algebras, we have the following corollary of

Proposition 17.2.2.
Corollary 17.2.3. Let 𝑅 be a Dedekind domain and let 𝐵 be a quaternion algebra over
𝐹 = Frac 𝑅. Let 𝐼 ⊆ 𝐵 be an 𝑅-lattice with either OL (𝐼) or OR (𝐼) maximal. Then 𝐼 is
locally principal and both OL (𝐼) and OR (𝐼) are maximal.
Proof. For each prime 𝔭 of 𝑅, we have that 𝑅𝔭 is a DVR and one of two possibilities:
either 𝐵𝔭 ' M2 (𝐹𝔭 ), in which case we can apply Lemma 17.2.2 to conclude 𝐼𝔭 is
principal, or 𝐵𝔭 is a division algebra, and we instead apply 13.3.10 to conclude that 𝐼𝔭
is principal.
17.3 Classes of lattices

For the rest of this chapter, let 𝑅 be a Dedekind domain with field of fractions 𝐹 =
Frac 𝑅, and let 𝐵 be a simple 𝐹-algebra.
Definition 17.3.1. Let 𝐼, 𝐽 ⊆ 𝐵 be 𝑅-lattices. We say 𝐼, 𝐽 are in the same right class,
and we write 𝐼 ∼R 𝐽, if there exists 𝛼 ∈ 𝐵× such that 𝛼𝐼 = 𝐽.
17.3.2. Throughout, we work on the right; analogous definitions can be made on the
left. When 𝐵 has a standard involution, the map 𝐼 ↦→ 𝐼 interchanges left and right.
Lemma 17.3.3. Let 𝐼, 𝐽 ⊆ 𝐵 be 𝑅-lattices. Then the following are equivalent:
(i) 𝐼 ∼R 𝐽;
(ii) 𝐼 is isomorphic to 𝐽 as a right module over OR (𝐼) = OR (𝐽); and
(iii) (𝐽 : 𝐼) L is a principal 𝑅-lattice.
Proof. For (i) ⇒ (ii). If 𝐼 ∼R 𝐽 then 𝐽 = 𝛼𝐼 with 𝛼 ∈ 𝐵× , so OR (𝐽) = OR (𝐼) and the
map left-multiplication by 𝛼 gives a right O-module isomorphism 𝐼 −→∼ 𝐽. Conversely,
∼
for (i) ⇐ (ii), suppose that 𝜙 : 𝐼 −→ 𝐽 is an isomorphism of right O-modules. Then
𝜙 𝐹 : 𝐼 ⊗𝑅 𝐹 = 𝐵 −→ ∼ 𝐽 ⊗ 𝐹 = 𝐵 is an automorphism of 𝐵 as a right 𝐵-module.
𝑅
Then as in Example 7.2.14, such an isomorphism is obtained by left multiplication by
𝛼 ∈ 𝐵× , so by restriction 𝜙 is given by this map as well.
Next, for (i) ⇒ (iii), suppose 𝛼𝐼 = 𝐽 with 𝛼 ∈ 𝐵× . Then
(𝐽 : 𝐼) L = {𝛽 ∈ 𝐵 : 𝛽𝐼 ⊆ 𝐽 = 𝛼𝐼} = 𝛼OL (𝐼)
is principal. The converse follows similarly.
The relation ∼R defines an equivalence relation on the set of 𝑅-lattices in 𝐵, and

the equivalence class of an 𝑅-lattice 𝐼 is denoted [𝐼] R . If 𝐼 is an invertible 𝑅-lattice,
then every lattice in the class [𝐼] R is invertible and we call the class invertible.
In view of Lemma 17.3.3(b), we organize classes of lattices by their right orders.
Let O ⊂ 𝐵 be an 𝑅-order.
17.4. TYPES OF ORDERS 273
Definition 17.3.4. The (right) class set of O is
ClsR O := {[𝐼] R : 𝐼 an invertible right fractional O-ideal}.
In view of 17.3.2, we will soon abbreviate Cls O := ClsR O and drop the subscript
R from the classes, when no confusion can result.
Remark 17.3.5. The notation Cl O is also used for the class set, but it sometimes means
instead the stably free class group or some other variant. We use “Cls” to emphasize
that we are working with a class set.
17.3.6. The set ClsR O has a distinguished element [O] R ∈ ClsR O, so it has the
structure of a pointed set (a set equipped with a distinguished element of the set).
However, in general it does not have the structure of a group under multiplication:
for classes [𝐼] R , [𝐽] R , we have [𝛼𝐽] R = [𝐽] R for 𝛼 ∈ 𝐵× but we need not have
[𝐼𝛼𝐽] R = [𝐼𝐽] R , because of the lack of commutativity.
17.3.7. An argument similar to the one in Proposition 17.2.2, either arguing locally or
with pseudobases (9.3.7), yields the following [CR81, (4.13)].
Let 𝑅 be a Dedekind domain with 𝐹 = Frac 𝑅, and let 𝐼 ⊆ 𝐵 be an 𝑅-lattice with
OL (𝐼) = M𝑛 (𝑅). Then there exists 𝛽 ∈ GL𝑛 (𝐹) and fractional ideals 𝔞1 , · · · , 𝔞𝑛 such
that
𝐼 = M𝑛 (𝑅) diag(𝔞1 , . . . , 𝔞𝑛 ) 𝛽 (17.3.8)
where diag(𝔞1 , . . . , 𝔞𝑛 ) is the 𝑅-module of diagonal matrices with entries in the given
fractional ideal. The representation (17.3.8) is called the Hermite normal form of
the 𝑅-module 𝐼, because it generalizes the Hermite normal form over a PID (allowing
coefficient ideals).
By 9.3.10, the Steinitz class [𝔞1 · · · 𝔞𝑛 ] ∈ Cl 𝑅 is uniquely defined. Switching to
the right, this yields a bĳection
∼ Cls (M (𝑅))
Cl 𝑅 −→ R 𝑛
(17.3.9)
[𝔞] ↦→ [diag(𝔞, 1, . . . , 1) M𝑛 (𝑅)] R
17.4 Types of orders

Next, we consider isomorphism classes of orders. Let O, O0 ⊆ 𝐵 be 𝑅-orders.
Definition 17.4.1. We say O, O0 are of the same type if there exists 𝛼 ∈ 𝐵× such that
O0 = 𝛼−1 O𝛼.
Lemma 17.4.2. The 𝑅-orders O, O0 are of the same type if and only if they are
isomorphic as 𝑅-algebras.
Proof. If O, O0 are of the same type, then they are isomorphic (under conjugation).
∼ O0 is an isomorphism of 𝑅-algebras, then extending scalars
Conversely, if 𝜙 : O −→
to 𝐹 we obtain 𝜙 𝐹 : O𝐹 = 𝐵 −→ ∼ 𝐵 = O0 𝐹 an 𝐹-algebra automorphism of 𝐵. By
the theorem of Skolem–Noether (Corollary 7.7.4), such an automorphism is given by
conjugation by 𝛼 ∈ 𝐵× , so O, O0 are of the same type.
∼ O0 induces a bĳection
17.4.3. If O, O0 are of the same type, then an isomorphism O −→
∼ 0
Cls O −→ Cls O of pointed sets. By Lemma 17.4.2, such an isomorphism is provided
by conjugation O0 = 𝛼−1 O𝛼 for some 𝛼 ∈ 𝐵× . The principal lattice 𝐼 = O𝛼 = 𝛼O0
has OL (𝐼) = O and OR (𝐼) = O0.
Generalizing 17.4.3, the class sets of two orders are in bĳection if they are con-
nected, in the following sense.
Definition 17.4.4. O is connected to O0 if there exists a locally principal fractional
O, O0-ideal 𝐽 ⊆ 𝐵, called a connecting ideal.
The relation of being connected is an equivalence relation on the set of 𝑅-orders.
If two 𝑅-orders O, O0 are of the same type, then they are connected by a principal
connecting ideal (17.4.3).
Definition 17.4.5. We say that O, O0 are locally of the same type or locally iso-
morphic if O𝔭 and O𝔭0 are of the same type (i.e., O𝔭 ' O𝔭0 ) for all primes 𝔭 of
𝑅.
Lemma 17.4.6. The 𝑅-orders O, O0 are connected if and only if O, O0 are locally
isomorphic.
Proof. Let 𝐽 be a connecting ideal, a locally principal fractional O, O0-ideal. Then for
all primes 𝔭 of 𝑅 we have 𝐽𝔭 = O𝔭 𝛼𝔭 with 𝛼𝔭 ∈ 𝐵𝔭 , and consequently O𝔭0 = OR (𝐼𝔭 ) =
𝛼𝔭−1 O𝔭 𝛼𝔭 . Therefore O is locally isomorphic to O0.
Conversely, if O, O0 are locally isomorphic, then for all primes 𝔭 of 𝑅 we have
O𝔭 = 𝛼𝔭−1 O𝔭 𝛼𝔭 with 𝛼𝔭 ∈ 𝐵𝔭 . Since 𝑅 is a Dedekind domain, O𝔭0 = O𝔭 for all but
0
finitely many primes 𝔭, so we may take 𝛼𝔭 ∈ O𝔭 = O𝔭0 for all but finitely many primes
𝔭. Therefore, there exists an 𝑅-lattice 𝐼 with 𝐼𝔭 = O𝔭 𝛼𝔭 by the local-global principle
for lattices, and 𝐼 is a locally principal fractional O, O0-ideal.
Lemma 17.4.7. If O, O0 ⊆ 𝐵 are maximal 𝑅-orders, then OO0 is a O, O0-connecting

ideal.
The product in Lemma 17.4.7 is not necessarily compatible.
Proof. Since O, O0 are 𝑅-lattices, their product 𝐼 := OO0 is an 𝑅-lattice. We visibly

have O ⊆ OL (𝐼) and the same on the right; but O, O0 are maximal, so equality holds
and 𝐼 is a fractional O, O0-ideal. Finally, 𝐼 is invertible by Proposition 16.6.15(b),
hence locally principal by Main Theorem 16.6.1.
In analogy with the class set, we make the following definitions.

Definition 17.4.8. Let O ⊂ 𝐵 be an 𝑅-order. The genus Gen O of O is the set of 𝑅-
orders in 𝐵 locally isomorphic to O. The type set Typ O of O is the set of isomorphism
classes of orders in the genus of O.
17.4.9. The orders in a genus have a common reduced discriminant, since the dis-
criminant can be defined locally and is well-defined on (local) isomorphism classes,
by Corollary 15.2.9.
17.4. TYPES OF ORDERS 275
17.4.10. Recalling section 15.5, there is a unique genus of maximal 𝑅-orders in a

quaternion algebra 𝐵—that is to say, every two maximal orders are locally isomorphic—
and this genus has a well-defined reduced discriminant equal to disc𝑅 𝐵.
The importance of connected orders is attested to by the following result.
Lemma 17.4.11. Let O, O0 be connected 𝑅-orders, and let 𝐽 be a connecting O, O0-

ideal. Then the maps
∼ Cls O0
ClsR O −→ R
[𝐼] R ↦→ [𝐼𝐽] R
[𝐼 𝐽 ] R ← [𝐼 0] R
0 −1
are mutually inverse bĳections. In particular, if O0 ∈ Gen O then # ClsR O =

# ClsR O0.
Proof. By definition, 𝐽 is invertible with OL (𝐽) = O and OR (𝐽) = O0. Therefore

the map 𝐼 ↦→ 𝐼𝐽 induces a bĳection between the set of invertible right O-ideals
and the set of invertible right O0-ideals (Lemma 16.5.11), with inverse given by
𝐼 0 ↦→ 𝐼 0 𝐽 −1 , and each of these products is compatible. This map then induces a
bĳection Cls O −→ ∼ Cls O0, since is compatible with left multiplication in 𝐵, i.e.,
(𝛼𝐼)𝐽 = 𝛼(𝐼𝐽) for all 𝛼 ∈ 𝐵× .
Remark 17.4.12. The equivalence in Lemma 17.4.11 is a form of Morita equivalence:

see Remark 7.2.20.
Lemma 17.4.11 says that the cardinality of the right class set is well-defined on the
genus Gen O; and of course the cardinality of the type set is also well-defined on the
genus (as it is the number of isomorphism classes).
Lemma 17.4.13. The map
ClsR O → Typ O
(17.4.14)
is a surjective map of sets.
Proof. If O0 is connected to O, then there is a connecting O0, O-ideal 𝐼, and [𝐼] R ∈

ClsR O has OL (𝐼) ' O0.
Remark 17.4.15. The fibers of the map (17.4.14) is given by classes of two-sided
ideals: see Proposition 18.5.10.
17.4.16. Let 𝐵 = M2 (𝐹) and O = M2 (𝑅). From

the bĳection (17.3.9), the classes in
𝔞 𝔞
ClR (M2 (𝑅)) are represented by 𝐼𝔞 = for [𝔞] ∈ Cl 𝑅. Consequently
𝑅 𝑅

𝑅 𝔞
OL (𝐼𝔞 ) = .
𝔞−1 𝑅
We will see later (28.5.11) that there is a bĳection

∼ Typ M (𝑅)
Cl 𝑅/(Cl 𝑅) 2 −→ 2
(17.4.17)

𝑅 𝔞
class of [𝔞] up to squares ↦→ class of −1 .
𝔞 𝑅
17.5 ⊲ Finiteness of the class set: over the integers

Over the next two sections, we will show that the set Cls O of invertible right (fractional)
O-ideals is finite using the geometry of numbers. In this section, we carry this out
for the simplest case, when 𝐵 is definite over Q; we consider the general case in the
next section. For further reading on the rich theory of the geometry of numbers, see
Cassels [Cas97], Gruber–Lekkerkerker [GrLe87], and Siegel [Sie89].
Our strategy is as follows: if 𝐽 is an invertible right O-ideal, we will show there
exists 𝛼 ∈ 𝐽 −1 with the property that 𝛼𝐽 = 𝐼 ⊆ O has bounded absolute norm
N (𝐼) = #(O/𝐼) ≤ 𝐶 where 𝐶 ∈ R>0 is independent of 𝐽. The result will then follow
from the fact that there are only finitely many right O-ideals of bounded absolute norm.
We begin with some definitions (generalizing Definition 9.3.1 slightly).
Definition 17.5.1. A Euclidean lattice is a Z-submodule Λ ⊆ R𝑛 with Λ ' Z𝑛 such

that RΛ = R𝑛 . The covolume of a Euclidean lattice Λ is covol(Λ) = vol(R𝑛 /Λ).
É Equivalently, a Euclidean lattice Λ ⊂ R is the Z-span of a basis of R , and if

17.5.2. 𝑛 𝑛
Λ= 𝑖 Z𝑎 𝑖 , then covol(Λ) = |det(𝑎 𝑖 𝑗 )𝑖, 𝑗 |.
Lemma 17.5.3. A subgroup Λ ⊂ R𝑛 is a Euclidean lattice if and only if Λ is discrete

and the quotient R𝑛 /Λ is compact.
Definition 17.5.4. Let 𝑋 ⊆ R𝑛 be a subset.
(a) 𝑋 is convex if 𝑡𝑥 + (1 − 𝑡)𝑦 ∈ 𝑋 for all 𝑥, 𝑦 ∈ 𝑋 and 𝑡 ∈ [0, 1].

(b) 𝑋 is symmetric if −𝑥 ∈ 𝑋 for all 𝑥 ∈ 𝑋.
The main result of Minkowski’s geometry of numbers is the following convex body
theorem.
Theorem 17.5.5 (Minkowski). Let 𝑋 ⊆ R𝑛 be a closed, convex, symmetric subset of

R𝑛 , and let Λ ⊂ R𝑛 be a Euclidean lattice. If vol(𝑋) ≥ 2𝑛 covol(Λ), then there exists
0 ≠ 𝛼 ∈ Λ ∩ 𝑋.
The following proposition can be seen as a generalization of what was done for the
Hurwitz order (11.3.1).
17.6. ⊲ EXAMPLE 277
Proposition 17.5.6. Let 𝐵 be a definite quaternion algebra over Q and let O ⊂ 𝐵

be an order. Then O× = O1 is a finite group, and every right ideal class in Cls O is
represented by an integral right O-ideal with
8
N (𝐼) ≤ discrd(O)
𝜋2
and the right class set Cls O is finite.

𝑎, 𝑏
Proof. Let 𝐵 = , with 𝑎, 𝑏 ∈ Z<0 . Since 𝐵 is definite, there is an embedding
Q
𝐵 ↩→ 𝐵∞ = 𝐵 ⊗Q R ' H. Inside 𝐵∞ ' R4 with Euclidean norm nrd, the order O
sits as a Euclidean lattice. The set O1 is therefore a discrete subset of the compact set
𝐵∞1 ' H1 , so it is finite.
Explicitly, we identify
𝐵∞ −→ ∼ R4
√ √︁ √︁ √︁ (17.5.7)
𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 ↦→ 2 𝑡, 𝑥 |𝑎|, 𝑦 |𝑏|, 𝑧 |𝑎𝑏|
Then 2 nrd(𝛼) = k𝛼k 2 for 𝛼 ∈ 𝐵 in this identification, and we have covol(O) =

discrd(O) (Exercise 17.7).
Let 𝐽 ⊂ 𝐵 be an invertible right fractional O-ideal. To find 𝐼 with [𝐼] = [𝐽] and
𝐼 integral, we look for a small 𝛼 ∈ 𝐽 −1 so that 𝐼 = 𝛼𝐽 ⊆ O will do. As a measure of
(co)volume, counting cosets and applying the definition (16.4.9), we obtain
covol(𝐽 −1 ) = [O : 𝐽 −1 ] Z covol(O) = N (𝐽 −1 ) discrd(O). (17.5.8)
Let 𝑐 > 0 satisfy 𝑐4 = (32/𝜋 2 ) covol(𝐽 −1 ), and let
𝑋 = {𝑥 ∈ R4 : k𝑥k ≤ 𝑐}.
Then 𝑋 is closed, convex, and symmetric, and vol(𝑋) = 𝜋 2 𝑐4 /2 = 16 covol(𝐽 −1 ).

Then by Minkowski’s theorem (Theorem 17.5.5), there exists 0 ≠ 𝛼 ∈ 𝐽 −1 ∩ 𝑋, and
1
N (𝛼𝐽) = Nm 𝐵 |Q (𝛼) N (𝐽) = nrd(𝛼) 2 N (𝐽) = k𝛼k 4 N (𝐽)
4 (17.5.9)
1 8
≤ 𝑐4 N (𝐽) = 2 discrd(O).
4 𝜋
Since 𝛼 is nonzero and 𝐵 is a division algebra, 𝛼 ∈ 𝐵× . Since 𝛼 ∈ 𝐽 −1 , the integral
right fractional O-ideal 𝐼 = 𝛼𝐽 ⊆ O is as desired.
If 𝐼 ⊆ O has N (𝐼) = #(O/𝐼) ≤ 𝐶 for 𝐶 ∈ Z>0 , then 𝐶O ⊆ 𝐼 ⊆ O hence there are
only finitely many possibilities for 𝐼, and the second statement follows.
17.6 ⊲ Example
We pause for an extended example. We steal the following lemma from the future.
Lemma 17.6.1. Let 𝑒 ∈ Z ≥0 . Then every principal right M2 (Z 𝑝 )-ideal 𝐼 with nrd(𝐼) =
𝑝 𝑒 is of the form 𝐼 = 𝛼 M2 (Z 𝑝 ) where
𝑢
𝑝 0 𝑣
𝛼∈ : 𝑢, 𝑣 ∈ Z ≥0 , 𝑢 + 𝑣 = 𝑒, and 𝑐 ∈ Z/𝑝 Z . (17.6.2)
𝑐 𝑝𝑣
Proof. The lemma follows from the theory of invariant factors: a more general state-
ment is proven in Lemma 26.4.1.

−1, −23
Example 17.6.3. Let 𝐵 = , and let
Q
1+ 𝑗 1+ 𝑗
O = Z + Z𝑖 + Z + Z𝑖 .
2 2
We have discrd(O) = disc 𝐵 = 23, so O is a maximal order, and 𝛽 = (1 + 𝑗)/2 satisfies
𝛽2 − 𝛽 + 6 = 0. For convenience, let 𝛼 = 𝑖, so O = Zh𝛼, 𝛽i. Then
𝛼𝛽 + 𝛽𝛼 = 𝛼. (17.6.4)
By Proposition 17.5.6, it is sufficient to compute the (invertible) right O-ideals

𝐼 ⊆ O such that
8
nrd(𝐼) 2 = N (𝐼) ≤ 2 (23) ≤ 18.7
𝜋
so nrd(𝐼) ≤ 4. For nrd(𝐼) = 1, we can only have 𝐼 = O, and the class [𝐼1 ] = [O]. Let
O1 = O.
We move to nrd(𝐼) = 2, and refer to Lemma 17.6.1. Since 𝐵 is split at 2, there is
an embedding
O ↩→ M2 (Z2 )

0 −1 1 0
𝛼, 𝛽 ↦→ , .
1 0 0 𝑏0
where 𝑏 0 = 2 + 8 + 16 + 32 + · · · ∈ Z2 satisfies 𝑏 20 − 𝑏 0 + 6 = 0 and 𝑏 0 ≡ 0 (mod 2).

We have
1 0 0 0 1 0
𝛽, 𝛽 + 1, (𝛼 + 1) 𝛽 ≡ , , (mod 2)
0 0 0 1 1 0
so we obtain the three right ideals
𝐼 (1:0) = 2O + 𝛽O, 𝐼 (0:1) = 2O + (𝛽 − 1)O, 𝐼 (1:1) = 2O + (𝛼 + 1) 𝛽O (17.6.5)
labelled by the corresponding nonzero column. If one of these three ideals is principal,
then it is generated by an element of reduced norm 2. We have
nrd(𝑡 + 𝑥𝛼 + 𝑦𝛽 + 𝑧𝛼𝛽)
= 𝑡 2 + 𝑡𝑦 + 𝑥 2 + 𝑥𝑧 + 6𝑦 2 + 6𝑧2
2 2 (17.6.6)
1 1 23 2 23 2
= 𝑡+ 𝑦 + 𝑥+ 𝑧 + 𝑦 + 𝑧 .
2 2 4 4
17.7. FINITENESS OF THE CLASS SET: OVER NUMBER RINGS 279
So nrd(𝛾) = 2 with 𝛾 ∈ O has 𝑡, 𝑥, 𝑦, 𝑧 ∈ Z and therefore 𝑦 = 𝑧 = 0 and 𝑡 = 𝑥 = 1, i.e.,

𝐼 (1:1) = (𝛼 + 1)O is principal, and the ideals 𝐼 (1:0) , 𝐼 (0:1) are not. But [𝐼 (1:0) ] = [𝐼 (0:1) ]
because 𝛼𝐼 (1:0) = 𝐼 (0:1) because 𝛼 ∈ O× and by (17.6.4)
𝛼(2O + (𝛽 − 1)O) = 2𝛼O + 𝛼(𝛽 − 1)O = 2O − 𝛽𝛼O = 𝐼 (0:1)
(We have 𝛼𝐼 (1:0) ≠ 𝐼 (1:0) precisely because 𝛼 ∉ OL (𝐼).) In this way, we have found
exactly one new right ideal class, [𝐼2 ] = [𝐼 (1:0) ]. We compute its left order to be
𝑖(1 + 3 𝑗)
O2 := OL (𝐼2 ) = Z + 𝛽Z + Z + (2𝑖 𝑗)Z ; O
4
and we also have a new type [O2 ] ≠ [O1 ] ∈ Typ O.
In a similar way, we find 4 right ideals of reduced norm 3, and exactly one new
right ideal class, represented by the right ideal 𝐼3 = 3O + (𝛼 + 1) 𝛽O. For example, we
find that the right ideal 𝐼 0 = 3O + 𝛽O is not principal using (17.6.6): letting
1 0
(𝐼 0 : 𝐼2 ) L = 𝐼 0 𝐼2−1 = 𝐼 𝐼2
2
and we find a shortest vector
(1 − 𝛽)/2 ∈ (𝐼 0 : 𝐼2 ) L ,
so [𝐼 0] = [𝐼2 ].
Repeating this with ideals of reduced norm 4 (Exercise 17.8), we conclude that
Cls O = {[𝐼1 ], [𝐼2 ], [𝐼3 ]}
and letting O3 := OL (𝐼3 ), checking it is not isomorphic to the previous two orders, we
have
Typ O = {[O1 ], [O2 ], [O3 ]}.
17.7 Finiteness of the class set: over number rings

We now turn to the general case.
Main Theorem 17.7.1. Let 𝐹 be a number field, let S ⊆ Pl 𝐹 be eligible and 𝑅 = 𝑅 (S)
be the ring of S-integers in 𝐹. Let 𝐵 be a quaternion algebra over 𝐹, and let O ⊆ 𝐵
be an 𝑅-order in 𝐵. Then the class set Cls O and the type set Typ O is finite.
We call # Cls O the (right) class number of O. (By 17.3.2, the left class number
suitably defined is equal to the right class number.) This result will be drastically
improved upon in Part III of this text from analytic considerations; the proof in this
section, using the geometry of numbers, has the advantage that is easy to visualize, it
works in quite some generality, and it is the launching point for algorithmic aspects.
17.7.2. Before we begin, two quick reductions. The finiteness of the type set follows
from finiteness of the right class set by Lemma 17.4.13. And if 𝑅 = Z𝐹 is the ring of
integers of 𝐹, then the general case follows from the fact that the map
Cls O → Cl(O ⊗𝑅 𝑅 (S) )
(17.7.3)
[𝐼] ↦→ [𝐼 ⊗𝑅 𝑅 (S) ]
is surjective for an eligible set S.
Let 𝐹 be a number field of degree 𝑛 = [𝐹 : Q], let 𝑅 = Z𝐹 be the ring of integers
in 𝐹, and let 𝐵 be a quaternion algebra over 𝐹.
17.7.4. Suppose that 𝐹 has 𝑟 real places and 𝑐 complex places, so that 𝑛 = 𝑟 + 2𝑐.
Then Ö
𝐹 ↩→ 𝐹∞ = 𝐹 ⊗Q R ' 𝐹𝑣 ' R𝑟 × C𝑐 . (17.7.5)
𝑣 |∞
Taking the basis 1, 𝑖 for C, we obtain 𝐹∞ ' R𝑛 , and then in the embedding (17.7.5),
the ring of integers 𝑅 ' Z𝑛 sits discretely inside 𝐹∞ ' R𝑛 as a Euclidean lattice.

𝑎, 𝑏
17.7.6. Suppose 𝐵 = and let 1, 𝑖, 𝑗, 𝑘 be the standard basis for 𝐵 with 𝑘 = 𝑖 𝑗,
𝐹
so 𝐵 = 𝐹 ⊕ 𝐹𝑖 ⊕ 𝐹 𝑗 ⊕ 𝐹 𝑘 ' 𝐹 4 as 𝐹-vector spaces. Then
4
𝐵 ↩→ 𝐵∞ = 𝐵 ⊗Q R ' 𝐵 ⊗𝐹 𝐹∞ ' 𝐹∞ (17.7.7)
in this same basis. Via (17.7.5) in each of the four components, the embedding (17.7.7)
then gives an identification 𝐵∞ ' (R𝑛 ) 4 ' R4𝑛 .
The order 𝑅h𝑖, 𝑗, 𝑘i = 𝑅+𝑅𝑖+𝑅 𝑗 +𝑅𝑘 is discrete in 𝐵∞ exactly because 𝑅 is discrete
in 𝐹. But then implies that an 𝑅-order O is discrete in 𝐵∞ , since [O : 𝑅h𝑖, 𝑗, 𝑘i] Z < ∞.
Therefore O ↩→ R4𝑛 has the structure of a Euclidean lattice.
In the previous section, the real vector space 𝐵∞ was Euclidean under the reduced
norm. In general, that need no longer be the case. Instead, we find a positive definite
quadratic form 𝑄 : 𝐵∞ → R that majorizes the reduced norm in the following sense:
we require that
|Nm𝐹 /Q (nrd(𝛼))| ≤ 𝑄(𝛼) 𝑛 (17.7.8)
for all 𝛼 ∈ 𝐵 ⊆ 𝐵∞ .
Remark 17.7.9. With respect to possible majorants (17.7.8): in general, there are
uncountably many such choices, and parametrizing majorants arises in a geometric
context as part of reduction theory. As it will turn out, the only “interesting” case to
consider here is 17.7.10, by strong approximation (see Theorem 17.8.3).
17.7.10. Let 𝐵 be a totally definite (Definition 14.5.7) quaternion algebra over 𝐹, a
totally real number field. Then the quadratic form
𝑄: 𝐵 → Q
∑︁
𝛼 ↦→ Tr𝐹 /Q (nrd(𝛼)) = 𝑣(nrd(𝛼)) (17.7.11)
𝑣 |∞
17.7. FINITENESS OF THE CLASS SET: OVER NUMBER RINGS 281
is positive definite: 𝐵 𝑣 ' H and so 𝑣(nrd(𝛼)) ≥ 0 with equality if and only if

𝛼 = 0. We call this quadratic form the absolute reduced norm. In this case, by the
arithmetic-geometric mean,
! 1/𝑛
Ö
1/𝑛
Nm𝐹 /Q (nrd(𝛼)) = 𝑣(nrd(𝛼))
𝑣 (17.7.12)
1 ∑︁ 1
≤ 𝑣(nrd(𝛼)) = 𝑄(𝛼)
𝑛 𝑣 𝑛
(with equality if and only if 𝑣(nrd 𝛼) agrees for all 𝑣).

We pause to note the following important consequence of 17.7.10.
Lemma 17.7.13. Let 𝐵 be a totally definite quaternion algebra over a totally real field
𝐹 and let O ⊆ 𝐵 be a Z𝐹 -order. Then the group of units of reduced norm 1
O1 = {𝛾 ∈ O : nrd(𝛾) = 1}
is a finite group.
In Lemma 17.7.13, if 𝐹 = Q then O× = O1 , so we have captured the entire unit
group.
Proof. As in 17.7.10, we equip 𝐵R := 𝐵 ⊗Q R ' H𝑛 ' R4𝑛 with the absolute reduced
norm giving O ↩→ 𝐵R the structure of a Euclidean lattice (17.7.7). We have
O1 = {𝛾 ∈ O : 𝑄(𝛾) = 𝑛} (17.7.14)
by the arithmetic-geometric mean (17.7.12). But the set {𝑥 ∈ 𝐵R : 𝑄(𝑥) = 𝑛} is an
ellipsoid in R4𝑛 so compact, and O is a lattice so discrete. Therefore the intersection
O1 is finite.
17.7.15. We now generalize 17.7.10 to the general case. For 𝑣 an infinite place of 𝐹,
define
𝑄 𝑣 : 𝐵𝑣 → R
𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 ↦→ |𝑣(𝑡)| 2 + |𝑣(𝑎)||𝑣(𝑥)| 2 + |𝑣(𝑏)||𝑣(𝑦)| 2 + |𝑣(𝑎𝑏)||𝑣(𝑧)| 2 ;
then 𝑄 𝑣 is a positive definite quadratic form on 𝐵 𝑣 , and
|𝑣(nrd(𝛼))| = |𝑣(𝑡 2 − 𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧2 )|
≤ |𝑣(𝑡)| 2 + |𝑣(𝑎)||𝑣(𝑥)| 2 + |𝑣(𝑏)||𝑣(𝑦)| 2 + |𝑣(𝑎𝑏)||𝑣(𝑧)| 2 (17.7.16)
= 𝑄 𝑣 (𝛼).
Let 𝑚 𝑣 = 1, 2 depending on if 𝑣 is real or complex, and define
Ö
𝑄 : 𝐵∞ ' 𝐵𝑣 → R
𝑣 |∞
∑︁ (17.7.17)
(𝛼 𝑣 ) 𝑣 ↦→ 𝑚 𝑣 𝑄 𝑣 (𝛼 𝑣 ).
𝑣 |∞
Then 𝑄 is a positive definite quadratic form on 𝐵∞ , again called the absolute reduced

𝑎, 𝑏
norm (relative to 𝑎, 𝑏); it depends on the choice of representation 𝐵 = .
𝐹
Nevertheless, (17.7.16) and the arithmetic-geometric mean yield
1 ∑︁
|Nm𝐹 /Q (nrd(𝛼))| 1/𝑛 ≤ 𝑚 𝑣 |𝑣(nrd(𝛼))|
𝑛
𝑣 |∞
(17.7.18)
1 ∑︁
≤ 𝑚 𝑣 𝑄 𝑣 (𝛼) = 𝑄(𝛼).
𝑛
𝑣 |∞
We are now ready to prove the main result of this section.

Proposition 17.7.19. There exists an explicit constant 𝐶 ∈ R>0 such that for all 𝑅-
orders O, every right ideal class in Cls O is represented by an integral right O-ideal
𝐼 with
N (𝐼) ≤ 𝐶 N (discrd(O)).
Proof. If 𝐵 ' M2 (𝐹), then we appeal to 17.3.7, where such a bound comes from the
finiteness of Cl 𝑅. So we may suppose that 𝐵 is a division ring.
Let
𝑋 = {(𝑥𝑖 )𝑖 ∈ R4𝑛 : 𝑄(𝛼) ≤ 1}. (17.7.20)
Then 𝑋 is closed, convex, and symmetric.
Let O be an 𝑅-order in 𝐵 and let 𝐽 be an invertible right fractional O-ideal. As in
(17.5.8), counting cosets gives
covol(𝐽 −1 ) = N (𝐽) −1 covol(O). (17.7.21)
Let 1/4𝑛
covol(𝐽 −1 )

𝑐 := 2 . (17.7.22)
vol(𝑋)
Then vol(𝑐𝑋) = 𝑐4𝑛 vol(𝑋) = 24𝑛 covol(𝐽 −1 ). By Minkowski’s theorem, there exists
0 ≠ 𝛼 ∈ 𝐽 −1 ∩ 𝑐𝑋, so 𝑄(𝛼) ≤ 𝑐2 . By (17.7.18),
1 𝑛 𝑐2𝑛
|Nm𝐹 /Q (nrd(𝛼))| ≤ 𝑄(𝛼) ≤ .
𝑛𝑛 𝑛𝑛
Consequently
𝑐4𝑛
N (𝛼𝐽) = |Nm𝐹 /Q (nrd(𝛼))| 2 N (𝐽) ≤ N (𝐽)
𝑛2𝑛
24𝑛 N (𝐽) −1 covol(O) 24𝑛 covol(O) (17.7.23)
= N (𝐽) =
𝑛2𝑛 vol(𝑋) 𝑛2𝑛 vol(𝑋)
= 𝐶 N (discrd(O))
with
24𝑛 covol(O)
𝐶 := . (17.7.24)
𝑛2𝑛 vol(𝑋) N (discrd(O))
17.8. EICHLER’S THEOREM 283
The ratio covol(O)/N (discrd(O)) is a constant independent of O: if O0 is another

𝑅-order then
N (discrd(O 0 )) [O : O0] Z N (discrd(O)) N (discrd(O))
= = .
covol(O0) [O : O0] Z covol(O) covol(O)
Since 𝛼 is nonzero and 𝐵 is a division algebra we conclude that 𝛼 ∈ 𝐵× , and since

𝛼 ∈ 𝐽 −1 , the ideal 𝐼 = 𝛼𝐽 is as desired.
Remark 17.7.25. For an explicit version of the Minkowski bound in the totally definite
case, with a careful choice of compact region, see Kirschmer [Kir2005, Theorem
3.3.11].
Lemma 17.7.26. For all 𝐶 > 0, there are only finitely many integral right O-ideals
with N (𝐼) ≤ 𝐶.
Proof. We may suppose 𝐶 ∈ Z. If 𝐼 ⊆ O then N (𝐼) = [O : 𝐼] Z ≤ 𝐶, so 𝐶O ⊆ 𝐼 ⊆

O. But the group O/𝐶O is a finite abelian group and there are only finitely many
possibilities for 𝐼.
We now have the ingredients for our main theorem.
Proof of Main Theorem 17.7.1. Combine Proposition 17.7.19, the reductions in 17.7.2,
and Lemma 17.7.26.
Remark 17.7.27. The finiteness statement (Main Theorem 17.7.1) can be generalized
to the following theorem of Jordan–Zassenhaus. Let 𝑅 be a Dedekind domain with
𝐹 = Frac(𝑅) a global field, let O ⊆ 𝐵 be an 𝑅-order in a finite-dimensional semisimple
algebra 𝐵, and let 𝑉 be a left 𝐵-module. Then there are only finitely many isomorphism
classes 𝐼 ⊆ 𝐵 with O ⊆ OL (𝐼). Specializing to 𝑉 = 𝐵 a quaternion algebra, we recover
the Main Theorem 17.7.1. For a proof, see Reiner [Rei2003, Theorem 26.4]; see also
the discussion by Curtis–Reiner [CR81, §24].
17.8 Eichler’s theorem

In this section, we state a special but conceptually important case of Eichler’s theorem
for number fields: roughly speaking, the class set of an indefinite quaternion order is
in bĳection with a certain class group of the base ring.
Let 𝐹 be a number field with ring of integers 𝑅 = Z𝐹 and let 𝐵 be a quaternion
algebra over 𝐹.
Definition 17.8.1. We say 𝐵 satisfies the Eichler condition if 𝐵 is indefinite.
Definition 17.8.1 introduces a longer (and rather opaque) phrase for something
that we already had a word for, but its use is prevalent in the literature. There are two
options: either 𝐵 is totally definite (𝐹 is a totally real field and all archimedean places
of 𝐹 are ramified in 𝐵) or 𝐵 is indefinite and satisfies the Eichler condition.
17.8.2. Recall 14.7.2 that we define Ω ⊆ Ram 𝐵 to be the set of real ramified places
of 𝐵 and 𝐹>×Ω 0 to be the positive elements for 𝑣 ∈ Ω.
We now define the group ClΩ 𝑅 as
the group of fractional ideals of 𝐹 under multiplication
modulo
the subgroup of nonzero principal fractional ideals
generated by an element in 𝐹>×Ω 0
If Ω is the set of all real places of 𝐹, then ClΩ 𝑅 = Cl+ 𝑅 is the narrow (or strict) class
group. On the other hand, if Ω = ∅, then ClΩ 𝑅 = Cl 𝑅. In general, we have surjective
group homomorphisms Cl+ 𝑅 → ClΩ 𝑅 and ClΩ 𝑅 → Cl 𝑅. In the language of class
field theory, ClΩ 𝑅 is the class group corresponding to the cycle given by the product
of the places in Ω.
Theorem 17.8.3 (Eichler; strong approximation). Let 𝐹 be a number field and let 𝐵
be a quaternion algebra over 𝐹 that satisfies the Eichler condition. Let O ⊆ 𝐵 be a
maximal Z𝐹 -order. Then the reduced norm induces a bĳection
∼ Cl 𝑅
Cls O −→ Ω
(17.8.4)
[𝐼] ↦→ [nrd(𝐼)].
where Ω ⊆ Ram 𝐵 is the set of real ramified places in 𝐵.
Proof. Eichler’s theorem is addressed by Reiner [Rei2003, §34], with a global proof
of the key result [Rei2003, Theorem 34.9] falling over several pages. We will instead
prove a more general version of this theorem as part of strong approximation, when
idelic methods allow for a more efficient argument: see Corollary 28.5.17.
Eichler’s theorem says that when 𝐵 is not totally definite, the only obstruction for
an ideal to be principal in a maximal order is that its reduced norm fails to be (strictly)
principal in the base ring. In particular, we have the following corollary.
Corollary 17.8.5. If # Cl+ 𝑅 = 1, then # Cls O = 1: i.e., every right O-ideal of a
maximal order in an indefinite quaternion algebra is principal.
Proof. Immediate from Eichler’s theorem and the fact that Cl+ 𝑅 surjects onto ClΩ 𝑅,
by 17.8.2.
∼ Cl Z .
Corollary 17.8.6. There is a bĳection Cls M2 (Z𝐹 ) −→ 𝐹
Proof. Immediate from Eichler’s theorem; we proved this more generally for a matrix
ring (17.3.9) using the Hermite normal form.
17.8.7. It is sensible for the class group ClΩ 𝑅 to appear by norm considerations. Let
𝑣 ∈ Ω; then 𝐵 𝑣 ' H, and if 𝛼 ∈ 𝐵× then 𝑣(nrd(𝛼)) > 0, as the reduced norm is
positive.
The class sets of totally definite orders are not captured by Eichler’s theorem, and
for good reason: they can be arbitrarily large, a consequence of the Eichler mass
formula (Chapter 25).
Exercises
Unless otherwise specified, throughout these exercises let 𝑅 be a Dedekind domain
with field of fractions 𝐹, let 𝐵 be a quaternion algebra over 𝐹, and let O ⊆ 𝐵 be an
𝑅-order.
1. Argue for Proposition 17.2.2 directly in a special case as follows. Let 𝐼 ⊆ M2 (𝐹)
be a lattice with OR (𝐼) = M2 (𝑅).
(a) By considering 𝐼 ⊗𝑅 𝐹 show that

𝐹 𝐹 0 0
𝐼⊆ M2 (𝑅) ⊕ M2 (𝑅).
0 0 𝐹 𝐹
(b) Suppose that 𝑅 is a PID. Conclude that 𝐼 is principal.

2. Let O, O0 ⊆ 𝐵 be 𝑅-orders. Show that the map in Lemma 17.4.11 is a bĳection
of pointed sets if and only if O is isomorphic to O0.
I 3. Let O, O0 ⊆ 𝐵 be 𝑅-orders with O ⊆ O0.
(a) If 𝐼 is an invertible right O-ideal, show that 𝐼O0 is an invertible right
O0-ideal. (The product 𝐼O0 is not necessarily compatible.)
(b) Show that the map
Cls O → Cls O0
[𝐼] ↦→ [𝐼O0]
is well-defined, surjective, and has finite fibers. [Hint: let 𝑟 ∈ 𝑅 be

nonzero such that O0 ⊆ 𝑟 −1 O. If 𝐼O0 = 𝐼 0, then 𝐼 0 = 𝐼O0 ⊆ 𝑟 −1 𝐼 ⊆ 𝑟 −1 𝐼 0
so 𝑟 𝐼 0 ⊆ 𝐼 ⊆ 𝐼 0, and conclude there are only finitely many possibilities for
𝐼.]
4. Let O, O0 ⊆ 𝐵 be maximal 𝑅-orders. In this exercise, we prove the following
statement:
There is a unique integral connecting O, O0 ideal 𝐼 of minimal re-
duced norm; moreover, we have nrd(𝐼) = [O : O ∩ O0].
(a) Show that this statement is local, i.e., the statement is true over 𝑅 if and
only if it is true over 𝑅𝔭 for all primes 𝔭 of 𝑅.
(b) Suppose 𝑅 is a DVR. Show that the statement is true if 𝐵 is a division
algebra.
(c) Suppose 𝑅 is a DVR with maximal ideal 𝔭, and that 𝐵 ' M2 (𝐹). Show that
there is a unique 𝛼 ∈ Or𝔭O such that O0 = 𝛼−1 O𝛼 up to left multiplication
by O× , and conclude that 𝐼 = O𝛼 is the unique integral connecting O, O0
ideal of minimal reduced norm. [Hint: 𝑁GL2 (𝐹 ) (M2 (𝑅)) = 𝐹 × GL2 (𝑅).]
(d) Continuing (c), show that nrd(𝛼) = [O : O ∩ O0]. [Hint: the statement
is equivalent under left or right multiplication of 𝛼 by O× ' GL2 (𝑅), so
consider invariant factors.] [For another perspective, see section 23.5.]
I 5. Let O ⊆ 𝐵 be an 𝑅-order and let 𝐼 be an invertible fractional right O-ideal. Let

𝔞 ⊆ 𝑅 be a nonzero ideal. Show that there exists a representative 𝐽 ∈ [𝐼] R (in
the same right ideal class as 𝐼) such that 𝐽 ⊆ O and nrd(𝐽) is coprime to 𝔞.
[Hint: look for 𝛼 ∈ (O : 𝐼) R and then look locally.]
I 6. Prove Lemma 17.5.3: a subgroup Λ ⊂ R𝑛 is a Euclidean lattice if and only if Λ is
discrete (every point of Λ is isolated, i.e., every 𝑥 ∈ Λ has an open neighborhood
𝑈 3 𝑥 such that Λ ∩ 𝑈 = {𝑥}) and the quotient R𝑛 /Λ is compact.
I 7. Let 𝐵 be a definite quaternion algebra over Q and let O ⊂ 𝐵 be an order.
(a) Let 𝐵∞ = 𝐵 ⊗Q R. Show that nrd is a Euclidean norm on 𝐵∞ , and √ O is
discrete in 𝐵∞ with covol(O) = 4 discrd(O). [So it is better to take 2 nrd
instead, to get covol(O) = discrd(O) on the nose.]
(b) Let 𝐾1 , 𝐾2 ⊆ 𝐵 be quadratic fields contained in 𝐵 with 𝐾1 ∩ 𝐾2 = Q. Let
𝑆𝑖 := 𝐾𝑖 ∩ O and 𝑑𝑖 = disc 𝑆𝑖 . Show that
(| 𝑑1 | − 1) (| 𝑑2 | − 1) ≥ 4 discrd(O).
[Hint: write 𝑆𝑖 = Z[𝛼𝑖 ] and consider the order Zh𝛼1 , 𝛼2 i.]
(c) Prove that if 𝛼1 , 𝛼2 ∈ O have
√︁
discrd(O)
nrd(𝛼1 ), nrd(𝛼2 ) <
2
then 𝛼1 𝛼2 = 𝛼2 𝛼1 .
8. Complete Example 17.6.3 by showing explicitly that all right O-ideals of reduced
norm 4 are in the same right ideal class as one of 𝐼1 , 𝐼2 , 𝐼3 .
9. Let 𝑅 be a global ring with # Cl 𝑅 = 1, i.e., every fractional 𝑅-ideal is principal
𝔞 = 𝑎𝑅. Suppose further that # Cls O = 1. Let 𝛼 ∈ O have nrd(𝛼) ≠ 0,
and factor nrd(𝛼) = 𝜋1 𝜋2 · · · 𝜋𝑟 ∈ 𝑅 where 𝜋𝑖 ∈ 𝑅 are pairwise nonassociate
nonzero prime elements (equivalently 𝜋𝑖 𝑅 are pairwise distinct nonzero prime
ideals).
(a) Show that there exist 𝜛1 , 𝜛2 , . . . , 𝜛𝑟 ∈ O such that 𝛼 = 𝜛1 𝜛2 · · · 𝜛𝑟 and
nrd(𝜛𝑖 )𝑅 = 𝜋𝑖 𝑅 for all 𝑖 = 1, . . . , 𝑟.
(b) Show that every other such factorization is of the form
𝛼 = (𝜛1 𝛾1 ) (𝛾1−1 𝜛2 𝛾2 ) · · · (𝛾𝑟−1−1 𝜛𝑟 )
where 𝛾1 , . . . , 𝛾𝑟 ∈ O× .
(c) Suppose that nrd(O× ) = 𝑅 × . Refine part (a) and show that the stronger
conclusion that there exist 𝜛𝑖 such that nrd(𝜛𝑖 ) = 𝜋𝑖 for all 𝑖.
[This generalizes Theorem 11.4.8.]
10. We have seen that maximal orders in (definite) quaternion algebras of discrim-
inant 2 (the Hurwitz order) and discriminant 3 (Exercise 11.12) are Euclidean
with respect to the norm, and in particular they have trivial right class set.
(a) Show that maximal orders O in quaternion algebras of discriminants
5, 7, 13 have # Cls O = 1.
(b) Conclude that the quaternary quadratic forms

𝑡 2 + 𝑡𝑥 + 𝑡𝑦 + 𝑡𝑧 + 𝑥 2 + 𝑥𝑦 + 𝑥𝑧 + 2𝑦 2 − 𝑦𝑧 + 2𝑧2 ,
𝑡 2 + 𝑡𝑧 + 𝑥 2 + 𝑥𝑦 + 2𝑦 2 + 2𝑧2 ,
𝑡 + 𝑡𝑦 + 𝑡𝑧 + 2𝑥 2 + 𝑥𝑦 + 2𝑥𝑧 + 2𝑦 2 + 𝑦𝑧 + 4𝑧 2
2
are multiplicative and universal, i.e., represent all positive integers.

(c) Show that for discriminant 7, 13 the maximal orders are not Euclidean with
respect to the norm.
[We discuss the maximal orders of class number 1 in Theorem 25.4.1. The
maximal order for discriminant 5 is in fact norm Euclidean: see Fitzgerald
[Fit2011].]
11. In this exercise, we show that the group of principal two-sided ideals PIdl(O)
need not be normal in the group of invertible fractional O-ideals Idl(O) of an
order.
Let 𝐵 = (−1, −1 | Q), and let O ⊆ 𝐵 be the Hurwitz order. Let O0 = Z + 5O =
O(5) (cf. Exercise 18.6). Show that
𝐼 0 = 10O0 + (1 − 2𝑖 + 𝑗)O0
is a two-sided invertible O0-ideal, and that
𝐼 0 𝑗 (𝐼 0) −1 = 5O0 + (𝑖 + 3 𝑗 + 𝑘)O0
is not principal.
12. The finiteness of the class group (see Reiner [Rei2003, Lemma 26.3]) can be
proven replacing the geometry of numbers with just the pigeonhole principle, as
follows. Let 𝐵 be a division algebra over a number field 𝐹 with ring of integers
𝑅, and let O ⊆ 𝐵 be an 𝑅-order.
(a) To prove the finiteness of Cls O, show that without loss of generality we
may take 𝐹 = Q.
(b) Show that Nm 𝐵 |Q (𝑥1 𝛼1 + · · · + 𝑥 𝑛 𝛼𝑛 ) ∈ Q[𝑥 1 , . . . , 𝑥 𝑛 ] is a homogeneous
polynomial of degree 𝑛.
(c) Show that there exists 𝐶 ∈ Z>0 such that for all 𝑡 > 0 and all 𝑥 ∈ Z𝑛 with
|𝑥𝑖 | ≤ 𝑡, we have |Nm 𝐵 |𝐹 (𝑥1 𝛼1 + · · · + 𝑥 𝑛 𝛼𝑛 )| ≤ 𝐶𝑡 𝑛 .
(d) Let 𝐼 ⊆ O be a lattice. Let 𝑠 ∈ Z be such that
𝑠 𝑛 ≤ N (𝐼) = #(O/𝐼) ≤ (𝑠 + 1) 𝑛 .
Í
Using the pigeonhole principle, show that there exists 𝛼 = 𝑖 𝑥 𝑖 𝛼𝑖 ∈ 𝐼
with 𝑥 𝑖 ∈ Z and |𝑥𝑖 | ≤ 2(𝑠 + 1) for all 𝑖.
(e) Show that N (𝛼O) ≤ 2𝑛 (𝑠 + 1) 𝑛 𝐶, and conclude that
#(𝐼/𝛼O) ≤ 4𝑛 𝐶.
(f) Let 𝑀 = (4𝑛 𝐶)! and show that 𝑀 𝐼 ⊆ 𝛼O, whence
𝑀O ⊆ 𝐼 0 ⊆ O
where 𝐼 0 = (𝑀𝛼−1 )𝐼. Conclude that the number of possibilities for 𝐼 0 is

finite, hence the number of right classes of lattices 𝐼 ⊆ O is finite, and
hence # Cls O < ∞.
Chapter 18
Two-sided ideals and the Picard group
In this chapter, we treat maximal orders like noncommutative Dedekind domains, and
we consider the structure of two-sided ideals (and their classes), in a manner parallel
to the commutative case.
18.1 ⊲ Noncommutative Dedekind domains

Let 𝑅 be a Dedekind domain with field of fractions 𝐹: then by definition 𝑅 is noetherian,
integrally closed, and all nonzero prime ideals of 𝑅 are maximal. Equivalently, every
ideal of 𝑅 is the product of prime ideals (uniquely up to permutation). To establish this
latter property of unique factorization of ideals, there are two essential ingredients:
first, every proper ideal contains a finite product of prime ideals, and second, every
nonzero prime ideal 𝔭 ⊆ 𝑅 is invertible. The first of these uses that 𝑅 is noetherian
and that nonzero prime ideals of 𝑅 are maximal; the second uses that 𝑅 is integrally
closed.
Here, the theorems are no easier to prove in the case of a quaternion algebra, so
we might as well consider them in more generality. Let 𝐵 be a simple 𝐹-algebra and
let O ⊆ 𝐵 be an 𝑅-order.
To draw the closest analogy with Dedekind domains, we suppose that O ⊂ 𝐵
is maximal: this is the noncommutative replacement for integrally closed. Since O
is finitely generated, if 𝐼 ⊆ O is a two-sided O-ideal, then 𝐼 is a finitely generated
𝑅-submodule, so the noetherian condition on 𝑅 automatically implies that every chain
of ideals of O stabilizes. We say a two-sided ideal 𝑃 ⊆ O is prime if 𝑃 ≠ O and for
all two-sided ideals 𝐼, 𝐽 ⊆ O, we have
𝐼𝐽 ⊆ 𝑃 ⇒ 𝐼 ⊆ 𝑃 or 𝐽 ⊆ 𝑃.
Running parallel to the above, we have the following initial lemma.

Lemma 18.1.1. A nonzero two-sided O-ideal is prime if and only it is maximal, and
every two-sided O-ideal contains a product of prime two-sided O-ideals.
Completing the analogy with the commutative case, we then have the following
theorem.
289
290 CHAPTER 18. PICARD GROUP
Theorem 18.1.2. Let 𝑅 be a Dedekind domain with field of fractions 𝐹 = Frac 𝑅, let
𝐵 be a simple 𝐹-algebra and let O ⊆ 𝐵 be a maximal 𝑅-order. Then the following
statements hold.
(a) If 𝐼 ⊆ 𝐵 is an 𝑅-lattice such that OL (𝐼) = O or OR (𝐼) = O, then 𝐼 is invertible

and both OL (𝐼) and OR (𝐼) are maximal 𝑅-orders.
(b) Multiplication of two-sided O-ideals is commutative, and every nonzero two-
sided O-ideal is the product of finitely many prime two-sided O-ideals, uniquely
up to permutation.
Let Idl(O) be the group of invertible two-sided fractional O-ideals. Put another
way, Theorem 18.1.2 says that if O is maximal, then Idl(O) is isomorphic to the free
abelian group on the set of nonzero prime two-sided O-ideals under multiplication.
We now consider classes of two-sided ideals, in the spirit of section 17.1. Two
candidates present themselves. On the one hand, inside the group Idl(O) of invertible
fractional two-sided O-ideals, the principal fractional two-sided O-ideals (those of
the form O𝛼O = O𝛼 = 𝛼O for certain 𝛼 ∈ 𝐵× ) form a subgroup PIdl(O), and we
could consider the quotient. On the other hand, for a commutative ring 𝑆, the Picard
group Pic(𝑆) is defined to be the group of isomorphism classes of rank one projective
(equivalently, invertible) 𝑆-modules under the tensor product. When 𝑆 is a Dedekind
domain, there is a canonical isomorphism Cl 𝑆 Pic(𝑆).
For simplicity, suppose now that 𝑅 = Z. In this noncommutative setting, we
analogously define Pic O to be the group of isomorphism classes of invertible O-
bimodules (over Z) under tensor product. If 𝐼, 𝐽 ∈ Idl(O), then 𝐼, 𝐽 are isomorphic as
O-bimodules if and only if 𝐽 = 𝑎𝐼 with 𝑎 ∈ Q× , and this yields an isomorphism
Pic O ' Idl(O)/Q× .
Let
𝑁 𝐵× (O) = {𝛼 ∈ 𝐵× : 𝛼O = O𝛼}
be the normalizer of O in 𝐵. By the Skolem–Noether theorem,
𝑁 𝐵× (O)/Q× ' Aut(O)
is the group of Z-algebra (or ring) automorphisms of O.
Theorem 18.1.3. Let 𝐵 be a quaternion algebra over Q of discriminant 𝐷 := disc 𝐵,

and let O ⊂ 𝐵 be a maximal order. Then
Ö
Pic O ' Z/2Z
𝑝 |𝐷
generated by (unique) prime two-sided O-ideals with reduced norm 𝑝 | 𝐷, and there
is an exact sequence
0 → 𝑁 𝐵× (O)/(Q× O× ) → Pic O → Idl(O)/PIdl(O) → 0

(18.1.4)
𝛼(Q× O× ) ↦→ [O𝛼O].
18.2. PRIME IDEALS 291
In particular, Pic O is a finite abelian 2-group. As an application of Theorem

18.1.3, we revisit the map (17.1.3):
Cls O → Typ O
We recall that this map is surjective. The fibers are given by Theorem 18.1.3 (see
Proposition 18.5.10): the fiber above the isomorphism class of O0 is in bĳection with
the set PIdl(O0)\ Idl(O0).
Remark 18.1.5. The structure of Pic O is more complicated when O is not necessarily
a maximal order: in general, the group Pic O is finite but it may be nonabelian (see
Exercise 18.6); worse still, in general the subgroup PIdl(O) may not be a normal
subgroup in Idl(O).
18.2 Prime ideals

Throughout this chapter, let 𝑅 be a Dedekind domain with field of fractions 𝐹 = Frac 𝑅,
let 𝐵 be a simple finite-dimensional 𝐹-algebra, and let O ⊆ 𝐵 be an 𝑅-order.
18.2.1. Let 𝐼 ⊆ O be a nonzero two-sided ideal. In view of Remark 16.2.10, we see

that 𝐼 is automatically an 𝑅-lattice: 𝐼 𝐹 ⊆ 𝐵 is a two-sided ideal of 𝐵, so since 𝐵 is
simple and 𝐼 ≠ {0} we must have 𝐼 𝐹 = 𝐵.
Definition 18.2.2. A two-sided ideal 𝑃 ⊆ O is prime if 𝑃 ≠ O and for all two-sided

ideals 𝐼, 𝐽 ⊆ O we have
𝐼𝐽 ⊆ 𝑃 ⇒ 𝐼 ⊆ 𝑃 or 𝐽 ⊆ 𝑃.
A two-sided O-ideal 𝑀 ⊆ O is maximal if 𝑀 ≠ O and 𝑀 is not properly contained

in another two-sided ideal.
Example 18.2.3. The zero ideal 𝑃 = {0} is prime: see Exercise 18.2.
18.2.4. Let 𝑃 ⊆ O be a two-sided ideal. Then the two-sided O/𝑃-ideals are in

bĳection with the two-sided O-ideals containing 𝑃. If 𝑃 ≠ O, then 𝑃 is prime if and
only if for all two sided O/𝑃-ideals 𝐼/𝑃, 𝐽/𝑃, we have
(𝐼/𝑃) (𝐽/𝑃) = {0} ⇒ 𝐼/𝑃 = {0} or 𝐽/𝑃 = {0}. (18.2.5)
Lemma 18.2.6. If 𝑀 is a maximal two-sided O-ideal, then 𝑀 is prime.
Proof. Suppose 𝐼𝐽 ⊆ 𝑀. Then (𝐼 + 𝑀) (𝐽 + 𝑀) ⊆ 𝑀. But 𝐼 + 𝑀 ⊇ 𝑀 so either

𝐼 + 𝑀 = 𝑀 or 𝐼 + 𝑀 = O by maximality, and the same is true for 𝐽. Since 𝑀 ≠ O we
must have either 𝐼 + 𝑀 = 𝑀 or 𝐽 + 𝑀 = 𝑀, which is to say 𝐼 ⊆ 𝑀 or 𝐽 ⊆ 𝑀.
Proposition 18.2.7.
(a) A nonzero two-sided O-ideal is prime if and only it is maximal.

(b) If 𝑃 ⊆ O is a nonzero prime two-sided O-ideal, then 𝔭 = 𝑃 ∩ 𝑅 is a nonzero

prime ideal of 𝑅, and O/𝑃 is a finite-dimensional simple algebra over the field
𝑅/𝔭.
Proof. We follow Reiner [Rei2003, Theorem 22.3]. The implication (⇒) in (a) follows
from Lemma 18.2.6. Conversely, let 𝑃 be a nonzero prime two-sided O-ideal, and let
𝔭 = 𝑃 ∩ 𝑅. We show 𝔭 is a nonzero prime. By 18.2.1, 𝑃 is an 𝑅-lattice, so 𝔭 ≠ {0};
since 1 ∉ 𝑃, we have 𝔭 ≠ 𝑅 and 𝔭 is nontrivial. If 𝑎, 𝑏 ∈ 𝑅, then 𝑎𝑏 ∈ 𝔭 implies
(𝑎O) (𝑏O) ⊆ 𝑃; since 𝑃 is prime, we have 𝑎O ⊆ 𝑃 or 𝑏O ⊆ 𝑃, so 𝑎 ∈ 𝔭 or 𝑏 ∈ 𝔭.
Now let 𝐽/𝑃 = rad(O/𝑃) be the Jacobson radical of O/𝑃 (see section 7.4). By
Lemma 7.4.8, the ideal 𝐽/𝑃 is nilpotent; by (18.2.5), we conclude 𝐽/𝑃 = {0}. Thus
O/𝑃 is semisimple by Lemma 7.4.2 and thus is a product of simple 𝑅/𝔭-algebras by
the Wedderburn–Artin theorem (Main Theorem 7.3.10). But the simple components
of O/𝑃 are two-sided ideals that annihilate one another; again by (18.2.5), there can
be only one component, and O/𝑃 is simple. Thus O/𝑃 has no nontrivial ideals, and
𝑃 is maximal.
Lemma 18.2.8. Every nonzero two-sided ideal of O contains a (finite) product of

nonzero prime ideals.
Proof. If not, then the set of ideals which do not contain such products is nonempty;
since O is noetherian, there is a maximal element 𝑀. Since 𝑀 cannot itself be prime,
there exist ideals 𝐼, 𝐽, properly containing 𝑀, such that 𝐼𝐽 ⊆ 𝑀. But both 𝐼, 𝐽 contain
products of prime ideals, so the same is true of 𝑀, a contradiction.
We now turn to notions of invertibility.

18.2.9. Let 𝐼 be an invertible two-sided fractional O-ideal (cf. Definition 16.2.9 and
16.5.17). In particular, OL (𝐼) = OR (𝐼) = O. If 𝐽 is another invertible two-sided
fractional O-ideal, then so is 𝐼𝐽, by Lemma 16.5.11: we have OL (𝐼𝐽) = OL (𝐼) = O
and OR (𝐼𝐽) = OR (𝐽) = O. Let Idl(O) be the set of invertible two-sided fractional
O-ideals. Then Idl(O) is a group under multiplication with identity element O.
The structure of Idl(O), and quotients under natural equivalence relations, is the
subject of this chapter.
18.3 Invertibility
We now consider invertibility first in the general context of orders, then for maximal
orders. The general theory of maximal orders over Dedekind domains in simple
algebras was laid out by Auslander–Goldman [AG60]. One of the highlights of this
theory are the classification of such orders: they are endomorphism rings of a finitely
generated projective module over a maximal order in a division algebra. For a quite
general treatment of maximal orders, see the book by Reiner [Rei2003]; in particular,
the ideal theory presented here is also discussed in Reiner [Rei2003, §§22–23].
Lemma 18.3.1. Let 𝐽 be a two-sided O-ideal, not necessarily invertible. If 𝐽 ( O,
then 𝐽 −1 ) O.
18.3. INVERTIBILITY 293
Proof. The 𝑅-lattice 𝐽 −1 has 𝐽 −1 ⊇ O and OL (𝐽 −1 ) ⊆ OR (𝐽) = O and the same result
holds interchanging left and right.
We follow Reiner [Rei2003, Lemma 23.4] (who calls the proof “mystifying”).
Assume for the purposes of contradiction that 𝐽 −1 = O. Since 𝐽 ( O, there exists a
maximal two-sided O-ideal 𝑀 ⊇ 𝐽. Thus 𝑀 −1 ⊆ 𝐽 −1 = O. By Lemma 18.2.6, 𝑀
is prime. Let 𝑎 ∈ 𝑅 ∩ 𝐽 −1 be nonzero. By Lemma 18.2.8, 𝑎O contains a product of
prime two-sided O-ideals, so
𝑀 ⊇ 𝑎O ⊇ 𝑃1 𝑃2 · · · 𝑃𝑟 ,
with each 𝑃𝑖 prime. We may suppose without loss of generality that 𝑟 ∈ Z>0 is minimal
with this property. Since 𝑃1 · · · 𝑃𝑟 ⊆ 𝑀 and 𝑀 is prime, we must have 𝑃𝑖 ⊆ 𝑀, so
𝑃𝑖 = 𝑀 by Proposition 18.2.7. Thus
𝑀 ⊇ 𝑎O ⊇ 𝐽1 𝑀 𝐽2
with 𝐽1 , 𝐽2 two-sided O-ideals. From 𝑎 −1 𝐽1 𝑀 𝐽2 ⊆ O, we have 𝐽1 (𝑎 −1 𝑀 𝐽2 )𝐽1 ⊆ 𝐽1 ,

so by definition 𝑎 −1 𝑀 𝐽2 𝐽1 ⊆ OL (𝐽1 ) = O. Thus 𝑀 (𝑎 −1 𝐽2 𝐽1 )𝑀 ⊆ 𝑀 and 𝑎 −1 𝐽2 𝐽1 ⊆
𝑀 −1 ⊆ O so 𝐽2 𝐽1 ⊆ 𝑎O. This shows that 𝑎O contains the product 𝐽2 𝐽1 of 𝑟 − 1 prime
two-sided O-ideals, contradicting the minimality of 𝑟.
Using this lemma, we arrive at the following proposition for maximal orders.
Proposition 18.3.2. Let 𝐼 ⊆ 𝐵 be an 𝑅-lattice such that OL (𝐼) is a maximal 𝑅-order.

Then 𝐼 is right invertible, i.e., 𝐼 𝐼 −1 = OL (𝐼).
Of course, one can also swap left for right in the statement of Proposition 18.3.2.
Using the standard involution, we proved Proposition 18.3.2 when 𝐵 is a quaternion
algebra (Proposition 16.6.15(b)).
Proof of Proposition 18.3.2. We follow Reiner [Rei2003, Theorem 23.5]. Let O =

OL (𝐼). Let 𝐽 = 𝐼 𝐼 −1 ⊆ O. Then 𝐽 𝐼 = 𝐼 𝐼 −1 𝐼 ⊆ 𝐼, so 𝐽 ⊆ OL (𝐼) = O and 𝐽 is a
two-sided O-ideal. We have
𝐽𝐽 −1 = 𝐼 𝐼 −1 𝐽 −1 ⊆ O,
so 𝐼 −1 𝐽 −1 ⊆ 𝐼 −1 and therefore 𝐽 −1 ⊆ OR (𝐼 −1 ). Additionally,
OR (𝐼 −1 ) ⊇ OL (𝐼) = O; (18.3.3)
but O is maximal, so equality holds in (18.3.3) and therefore 𝐽 −1 ⊆ O. But O ⊆ 𝐽 −1

as well, so 𝐽 −1 = O. If 𝐽 ( O, then we have a contradiction with Lemma 18.3.1; so
𝐽 = O, and the proof is complete.
Putting these ingredients together, we have the following theorem.
Theorem 18.3.4. Let 𝑅 be a Dedekind domain with 𝐹 = Frac 𝑅, let 𝐵 be a simple

𝐹-algebra, and let O ⊆ 𝐵 be a maximal 𝑅-order. Then:
(a) Multiplication of two-sided ideals is commutative: if 𝐼, 𝐽 are two-sided O-ideals,

then 𝐼𝐽 = 𝐽 𝐼.
(b) Every nonzero two-sided O-ideal is invertible and uniquely expressible as a
product of prime two-sided ideals in O.
Proof. For (b), invertibility follows from Proposition 18.3.2. For (b) without unique-
ness, assume for purposes of contradiction that there is a two-sided ideal of O that is
not the product of prime ideals; then there is a maximal counterexample 𝐽. Since 𝐽 is
not prime, there exists a prime 𝑄 with 𝐽 ( 𝑄 ( O, so 𝐽 ⊂ 𝐽𝑄 −1 ( O. If 𝐽 = 𝐽𝑄 −1 ,
so by cancelling 𝑄 = O, a contradiction. Therefore 𝐽𝑄 −1 = 𝑃1 · · · 𝑃𝑟 is the product of
primes by maximality, and 𝐽 = 𝑃1 · · · 𝑃𝑟 𝑄 is the product of primes, a contradiction.
We now prove (a). If 𝑃, 𝑄 ⊆ O are distinct nonzero prime two-sided ideals, and
we let 𝑄 0 = 𝑃−1 𝑄𝑃, then 𝑄 0 ⊆ 𝑃−1 O𝑃 = O is prime and 𝑃𝑄 0 = 𝑄𝑃 ⊆ 𝑄, so 𝑃 ⊆ 𝑄
or 𝑄 0 ⊆ 𝑄; but equality would hold in each case by maximality, and since 𝑃 ≠ 𝑄, we
must have 𝑄 0 = 𝑄, and multiplication is commutative.
Finally, uniqueness of the factorization in (b) follows as in the commutative case.
If 𝑃1 · · · 𝑃𝑟 = 𝑄 1 · · · 𝑄 𝑠 , then 𝑃1 = 𝑄 𝑖 for some 𝑖; multiplying by 𝑃1−1 and repeating
the argument, we find that {𝑃1 , . . . , 𝑃𝑟 } = {𝑄 1 , . . . , 𝑄 𝑠 }, and the result follows.
Corollary 18.3.5. With hypotheses as in Theorem 18.3.4, the group Idl(O) is isomor-
phic to the free abelian group on the set of nonzero prime ideals.
With these arguments in hand, we have the following foundational result for quater-
nion orders.
Theorem 18.3.6. Suppose that 𝑅 is a Dedekind domain. Let 𝐵 be a quaternion algebra
over 𝐹 and let O ⊆ 𝐵 be a maximal 𝑅-order. Then the map
{Prime two-sided O-ideals} ↔ {Prime ideals of 𝑅}

(18.3.7)
𝑃 ↦→ 𝑃 ∩ 𝑅
is a bĳection.
Moreover, if 𝑅 is a global ring, then there is an exact sequence
Ö
0 → Idl(𝑅) → Idl(O) → Z/2Z → 0
𝔭 |𝔇 (18.3.8)
𝔞 ↦→ O𝔞O
where 𝔇 = disc𝑅 (𝐵).
Proof. The map (18.3.7) is defined by Proposition 18.2.7, and it is surjective because
𝔭O ⊆ 𝑃 is contained in a maximal therefore prime ideal.
Next we show that the map is injective. Let 𝑃 be a prime ideal, and work with
completions at a prime 𝔭. Then 𝑃𝔭 = 𝑃 ⊗𝑅 𝑅𝔭 ⊆ O𝔭 is a maximal ideal of O𝔭 . If
𝐵𝔭 ' M2 (𝐹𝔭 ), so O𝔭 ' M2 (𝑅𝔭 ), then the only maximal two-sided ideal is 𝔭O𝔭 ; if
instead 𝐵𝔭 is a division algebra, then there is a unique maximal two-sided ideal 𝑃𝔭
with 𝑃𝔭2 = 𝔭O𝔭 by Theorem 13.3.11. We can also describe this uniformly, by the proof
of Proposition 18.2.7: in all cases, we have 𝑃𝔭 = rad(O𝔭 ).
18.4. PICARD GROUP 295
There is a natural group homomorphism
Idl(𝑅) → Idl(O)
𝔞 ↦→ O𝔞O = 𝔞O
This map is injective, since if 𝔞O = O then 𝔞2 = nrd(𝔞O) = nrd(O) = 𝑅, so 𝔞 = 𝑅.

The cokernel of the map is determined by the previous paragraph.
Remark 18.3.9. Many of the theorems stated in this section (and chapter) hold more
generally for hereditary orders: this notion is pursued in Chapter 21. To see what this
looks like in a more general context, see Curtis–Reiner [CR81, §26B]. A very general
context in which one can make an argument like in section 18.3 was axiomatized by
Asano; for an exposition and several references, see McConnell–Robson [McCR87,
Chapter 5].
18.4 Picard group

We now proceed to consider classes of two-sided ideals. We begin with a natural but
abstract definition, in terms of bimodules. (Recall 20.3.7, that a bimodule is over 𝑅 if
the 𝑅-action on left and right are equal.)
Definition 18.4.1. The Picard group of O over 𝑅 is the group Pic𝑅 (O) of isomorphism
classes of invertible O-bimodules over 𝑅 under tensor product.
Remark 18.4.2. Some authors also write Picent(O) = Pic 𝑍 (O) (O) when considering
the Picard group over the center of O, the most important case. To avoid additional
complication, in this section we suppose that 𝐵 is central over 𝐹, so Pic𝑅 (O) =
Picent(O).
18.4.3. If 𝐼 ⊆ 𝐵 is an 𝑅-lattice that is a fractional two-sided O-ideal, then 𝐼 is a

O-bimodule over 𝑅. Conversely, if 𝐼 is a O-bimodule over 𝑅 then 𝐼 ⊗𝑅 𝐹 ' 𝐵 as
𝐵-bimodules, and choosing such an isomorphism gives an embedding 𝐼 ↩→ 𝐵 as an
𝑅-lattice.
Lemma 18.4.4. Let 𝐼, 𝐽 ⊆ 𝐵 be 𝑅-lattices that are fractional two-sided O-ideals.

Then 𝐼 is isomorphic to 𝐽 as O-bimodules over 𝑅 if and only if there exists 𝑎 ∈ 𝐹 ×
such that 𝐽 = 𝑎𝐼.
Proof. See Exercise 18.9.
18.4.5. By 18.4.3, there is a natural surjective map
Idl(O) → Pic𝑅 (O);
we claim that the kernel of this map is PIdl(𝑅) E Idl(O). By Lemma 19.5.1, every
isomorphism class of invertible O-bimodule is represented by an invertible 𝑅-lattice
𝐼 ⊆ 𝐵, unique up to scaling by 𝐹 × , and if 𝑎 ∈ 𝐹 × then 𝑎O = O if and only if
𝑎 ∈ 𝑅 ∩ O× = 𝑅 × , so the ideal 𝑎𝑅 ∈ PIdl(𝑅) is well-defined. Thus, we obtain a natural

isomorphism
Idl(O)/PIdl(𝑅) −→∼ Pic (O). (18.4.6)
𝑅
Equivalently, the sequence
1 → 𝑅 × → 𝐹 × → Idl(O) → Pic𝑅 (O) → 1
is exact. One might profitably take (18.4.6) as the definition of Pic𝑅 (O).
18.4.7. If O0 is locally isomorphic O (so they are in the same genus), then there is a
O, O0-connecting ideal 𝐽, and the map
Idl(O) → Idl(O0)
𝐼 ↦→ 𝐽 −1 𝐼𝐽
is an isomorphism of groups restricting to the identity on Cl 𝑅, so from (18.4.6) we

obtain an isomorphism
Pic𝑅 (O) ' Pic𝑅 (O0),
analogous to Lemma 17.4.11.
Our remaining task in this section is to examine the structure of Pic𝑅 (O), and to
this end we suppose that 𝐵 is a quaternion algebra over 𝐹.
18.4.8. Suppose that O is a maximal 𝑅-order with 𝐹 = Frac 𝑅 a global field. Then
taking the quotient by PIdl(𝑅) in the first two terms in (18.3.8) yields an exact sequence
Ö
0 → Cl 𝑅 → Pic𝑅 (O) → Z/2Z → 0. (18.4.9)
𝔭 |𝔇
Although this sequence need not split, it does show that the Picard group of the maximal
order O is not far from the class group Cl 𝑅, the difference precisely measured by the
primes that ramify in 𝐵.
In general, for a quaternion 𝑅-order O we have the following result.
Proposition 18.4.10. Pic𝑅 (O) is a finite group.
Proof. If O is maximal, we combine (18.4.9) with the finiteness of Cl 𝑅 and the fact
that there are only finitely many primes 𝔭 dividing the discriminant 𝔇.
Now let O be an 𝑅-order. Then there exists a maximal 𝑅-order O0 ⊇ O. We argue
as in Exercise 17.3. We define a map of sets:
Pic𝑅 (O) → Pic𝑅 (O0)

[𝐼] ↦→ [O0 𝐼O0]
The class up to scaling by 𝐹 × is well-defined, and 𝐼 0 := O0 𝐼O0 ⊇ 𝐼 an 𝑅-lattice with

left and right orders containing O0, but since O0 is maximal these orders equal O0 and
𝐼 0 is invertible.
18.5. CLASSES OF TWO-SIDED IDEALS 297
By the first paragraph, by finiteness of Pic𝑅 (O0), after rescaling we may suppose 𝐼 0
is one of finitely many possibilities. But there exists nonzero 𝑟 ∈ 𝑅 such that 𝑟O0 ⊂ O,
so
𝐼 0 = O0 𝐼O0 ⊆ (𝑟 −1 O)𝐼 (𝑟 −1 O) = 𝑟 −2 𝐼 ⊆ 𝑟 −2 𝐼 0
so 𝑟 2 𝐼 0 ⊆ 𝐼 ⊆ 𝐼 0; since 𝐼 0/𝑟 2 𝐼 0 is a finite group, this leaves only finitely many
possibilities for 𝐼.
Remark 18.4.11. The study of the Picard group is quite general. It was studied in
detail by Fröhlich [Frö73]; see also Curtis–Reiner [CR87, §55].
18.5 Classes of two-sided ideals

In this section, we compare the Picard group to the group of “ideals modulo principal
ideals”.
Let PIdl(O) ≤ Idl(O) be the subgroup of principal two-sided fractional O-ideals
(invertible by 16.5.4). Let
𝑁 𝐵× (O) = {𝛼 ∈ 𝐵× : 𝛼−1 O𝛼 = O}
be the normalizer of O in 𝐵× .
Lemma 18.5.1. There is an exact sequence of groups
1 → O× → 𝑁 𝐵× (O) → PIdl(O) → 1
(18.5.2)
𝛼 ↦→ O𝛼O.
Proof. We have 𝛼 ∈ 𝑁 𝐵× (O) if and only if 𝛼O = O𝛼 if and only if O𝛼O is a
principal two-sided fractional O-ideal, as in Exercise 16.17; this gives a surjective
group homomorphism 𝑁 𝐵× (O) → PIdl(O). The kernel is the set of 𝛼 ∈ 𝐵× such that
𝛼O = O, and this normal subgroup is precisely O× .
Proposition 18.5.3. There is an isomorphism of groups
∼ PIdl(O)/PIdl(𝑅)
𝑁 × (O)/(𝐹 × O× ) −→
𝐵
(18.5.4)
𝛼𝐹 × O× ↦→ class of O𝛼O.
If PIdl(O) E Idl(O) is normal, then the isomorphism (18.5.4) induces a natural exact
sequence
0 → 𝑁 𝐵× (O)/(𝐹 × O× ) → Pic𝑅 (O) → Idl(O)/PIdl(O) → 0
(18.5.5)
𝛼𝐹 × O× ↦→ class of O𝛼O.
Proof. There is an isomorphism 𝑁 𝐵× (O)/O× ' PIdl(O) by (18.5.2). The image of
𝐹 × ≤ 𝑁 𝐵× (O) in PIdl(O) under this map consists of two-sided ideals of the form
O𝑎O with 𝑎 ∈ 𝐹 × ; we have O𝑎O = O if and only if 𝑎 ∈ O× if and only if 𝑎 ∈ 𝑅 × , so
this image is isomorphic to the group PIdl(𝑅) of principal fractional 𝑅-ideals via the
map 𝑎𝑅 ↦→ O𝑎O. The first isomorphism follows. The exact sequence (18.5.5) is then
just rewriting the natural sequence
0 → PIdl(O)/PIdl(𝑅) → Idl(O)/PIdl(𝑅) → Idl(O)/PIdl(O) → 0.
Remark 18.5.6. The moral of Proposition 18.5.3 is that, unlike the commutative
case where the two notions coincide, the two notions of “isomorphism classes of
invertible bimodules” and “ideals modulo principal ideals” are in general different for
a quaternion order. These notions coincide precisely when 𝑁 𝐵× (O)/𝐹 × ' O× /𝑅 × , or
equivalently (by the Skolem–Noether theorem) that every 𝑅-algebra automorphism of
O is inner, which is to say Aut 𝑅 (O) = Inn𝑅 (O) = O× /𝑅 × .
18.5.7. Unfortunately, the subgroup PIdl(O) ≤ Idl(O) need not be normal in general
(Exercise 17.11), so statements like Proposition 18.5.3 depend on the order O having
good structural properties. If O is a maximal order, then Idl(O) is abelian, so the result
holds in this case.
In general, from the proof but using cosets one still obtains the equality
#(Idl(O)/PIdl(O)) · #(𝑁 𝐵× (O)/(𝐹 × O× )) = # Pic𝑅 (O). (18.5.8)
Remark 18.5.9. If O, O0 are connected, then Pic𝑅 (O) ' Pic𝑅 (O0) by 18.4.7 but this
isomorphism need not respect the exact sequence (18.5.5). Each order O “balances” the
contribution of this group between the normalizer 𝑁 𝐵× (O)/(𝐹 × O× ) and the quotient
Idl(O)/PIdl(O)—and these might be of different sizes for O0. We will return to
examine more closely this structure in section 28.9, when strong approximation allows
us to be more precise in measuring the discrepancy.
We conclude with an application to the structure of (right) class sets. We examine
from Lemma 17.4.13 the fibers of the surjective map (17.4.14)
Cls O → Typ O
[𝐼] ↦→ class of OL (𝐼).
Refreshing our notation, let 𝐵 be a central simple 𝐹-algebra and let O ⊂ 𝐵 an 𝑅-order.
Proposition 18.5.10. The map 𝐼 ↦→ [𝐼] induces a bĳection
PIdl(O)\ Idl(O) ↔ {[𝐼] ∈ Cls O : OL (𝐼) ' O}.
Proof. Let O0 be an order of the same type as O. Since (17.4.14) is surjective, there
exists [𝐼] ∈ Cls O such that OL (𝐼) ' O0. We are free to replace O0 by an isomorphic
order, so we may suppose OL (𝐼) = O0. For all [𝐼 0] ∈ Cls O with OL (𝐼 0) ' O0 (running
over the fiber), since OL (𝛼𝐼 0) = 𝛼O0 𝛼−1 for 𝛼 ∈ 𝐵× we may suppose without loss of
generality that the representative 𝐼 0 has OL (𝐼 0) = O0.
We then define a map
PIdl(O0)\ Idl(O0) → {[𝐼 0] ∈ Cls O : OL (𝐼 0) = O0 }

(18.5.11)
𝐽 0 ↦→ [𝐽 0 𝐼]
The map is surjective, because if 𝐽 0 = 𝐼 0 𝐼 −1 then OL (𝐽 0) = OR (𝐽 0) = O, so 𝐽 0 is a two-

sided invertible O0-ideal. It is injective because if [𝐽 0 𝐼] = [𝐾 0 𝐼] for 𝐽 0, 𝐾 0 ∈ Idl(O0)
then 𝐾 0 = 𝛼 0 𝐽 0 with 𝛼 0 ∈ 𝐵× , but further we need OL (𝐾 0) = 𝛼 0O0 𝛼 0−1 = O0, so in fact
[𝐽 0 𝐼] = [𝐾 0 𝐼] if and only if 𝛼 0 ∈ 𝑁 𝐵× (O0), and the result then follows from Lemma
18.5.1.
18.5. CLASSES OF TWO-SIDED IDEALS 299
We have the following corollaries.
Corollary 18.5.12. We have

∑︁ ∑︁ 1
# Cls O = [Idl(O0) : PIdl(O0)] = # Pic𝑅 O
𝑧 O0
[O0 ] ∈Typ O [O0 ] ∈Typ O
where 𝑧 O0 = [𝑁 𝐵× (O0) : 𝐹 × O0× ].
Proof. For the first equality, combine Lemma 17.4.13 and Proposition 18.5.10, com-
puting the size of the fibers. For the second, substitute (18.5.8) and use 18.4.7.
Corollary 18.5.13. Let O𝑖 be representatives of Typ O. For each 𝑖, let 𝐼𝑖 be a

connecting O𝑖 , O-ideal, and let 𝐽𝑖, 𝑗 be representatives of PIdl(O𝑖 )\ Idl(O𝑖 ). Then the
set {𝐽𝑖, 𝑗 𝐼𝑖 }𝑖, 𝑗 is a complete set of representatives for Cls O.
Proof. We choose representatives and take the fibers of the map (17.4.14).
Remark 18.5.14. When PIdl(O) E Idl(O), then in Proposition 18.5.10 we have written
the class set Cls O as a disjoint union of abelian groups. The fact that the bĳection is
noncanonical is due to the fact that we choose a connecting ideal, so without making
choices we obtain only a disjoint union of principal homogeneous spaces (i.e., torsors)
under the groups PIdl(O0)\ Idl(O0).
Exercises
Unless otherwise specified, let 𝑅 be a Dedekind domain with field of fractions 𝐹 =
Frac 𝑅, let 𝐵 be a simple finite-dimensional 𝐹-algebra, and let O ⊆ 𝐵 be an 𝑅-order.
1. Show that the following are equivalent:

(i) O is a maximal 𝑅-order;
(ii) OL (𝐼) = OR (𝐼) = O for all fractional two-sided O-ideals 𝐼; and
(iii) OL (𝐼) = OR (𝐼) = O for all two-sided O-ideals 𝐼 ⊆ O.
2. Show that the zero ideal is a prime ideal of O.
3. Let 𝐽 ⊆ O be a nonzero two-sided ideal of O in the ring-theoretic sense: 𝐽 is an
additive subgroup closed under left and right multiplication by O. Show that 𝐽
is an 𝑅-lattice.

𝑅 𝑅
4. Let 𝑅 be a DVR with maximal ideal 𝔭, and let O = ⊆ 𝐵 = M2 (𝐹).
𝔭 𝑅
Show that the two-sided ideal 𝔭 M2 (𝑅) ⊆ O is not a prime ideal.
√ √ √
5. Let 𝑅 := Z[ −6] and 𝐹 := Q( −6). Let 𝐵 := (2, −6 | 𝐹).
(a) Show that 2𝑅 = 𝔭22 and 3𝑅 = 𝔭23 for primes 𝔭2 , 𝔭3 ⊆ 𝑅.
(b) Show that Ram(𝐵) = {𝔭2 , 𝔭3 }.
(c) Let O be a maximal order in 𝐵. Show that there is a unique two-sided
ideal 𝑃2 such that 𝑃22 = 𝔭2 O.
(d) Prove that [𝑃2 ] ∈ Pic𝑅 (O) has order 4, and conclude that the sequence
(18.4.9) does not split.
(e) Show that we may take
√ √
O = 𝑅 + 𝔭−1 −1
2 ( −6 + 𝑖) + 𝑅 𝑗 + 𝔭2 ( −6 + 𝑖) 𝑗
√
as the maximal order,
√ and then that 𝐼 is generated by 𝑖 and −6𝑖 𝑗/2, and
finally that 𝐼 2 = ( −6 + 𝑖)/2.
6. Let 𝐵 = M𝑛 (𝐹) with 𝑛 ≥ 2, let O = M𝑛 (𝑅), let 𝔭 ⊆ 𝑅 be prime with 𝑘 = 𝑅/𝔭,
and let O(𝔭) = 𝑅 + 𝔭O.
(a) Show that O(𝔭) is an order of reduced discriminant 𝔭3 .
(b) Show that O× ' GL𝑛 (𝑅) normalizes O(𝔭) ⊆ O, so that
O(𝔭) × E O× ' GL𝑛 (𝑅),
and that the map
O× ↩→ Idl(O)
𝛾 ↦→ O𝛾 = 𝛾O
induces an injective group homomorphism PGL𝑛 (𝑘) ↩→ Idl(O). Con-
clude that Idl(O) is not an abelian group.
7. Show that Theorem 18.3.4 holds more generally for 𝐵 a semisimple 𝐹-algebra
(but still O ⊆ 𝐵 maximal). [Hint: Decompose 𝐵 into a product of simple
𝐹-algebras.]
8. Let O be maximal, and let 𝑃1 , . . . , 𝑃𝑟 ⊆ O be distinct prime two-sided ideals.
Let
Ö𝑟 Ö 𝑟
𝑓
𝐼 := 𝑃𝑖𝑒𝑖 and 𝐽 := 𝑃𝑖 𝑖
𝑖=1 𝑖=1
with 𝑒 𝑖 , 𝑓𝑖 ∈ Z.
(a) Prove that 𝐼 ⊆ O if and only if 𝑒 𝑖 ≥ 0 for all 𝑖 = 1, . . . , 𝑛, and in this case
there is a ring isomorphism
𝑟
Ê
O/𝐼 ' O/𝑃𝑖𝑒𝑖 .
𝑖=1
(b) Prove that 𝐼 ⊇ 𝐽 if and only if 𝑒 𝑖 ≤ 𝑓𝑖 for all 𝑖.

Î min(𝑒𝑖 , 𝑓𝑖 ) Î max(𝑒𝑖 , 𝑓𝑖 )
(c) Show 𝐼 + 𝐽 = 𝑟𝑖=1 𝑃𝑖 and 𝐼 ∩ 𝐽 = 𝑟𝑖=1 𝑃𝑖 .
I 9. Prove Lemma 18.4.4: Show that fractional two-sided O-ideals 𝐼, 𝐽 ⊆ 𝐵 are
isomorphic as O-bimodules over 𝑅 if and only if there exists 𝑎 ∈ 𝐹 × such that
𝐽 = 𝑎𝐼. [Hint: Peek at Lemma 19.5.1.]
10. Let 𝐾 ⊇ 𝐹 be a finite, separable extension and let 𝑆 be the integral closure
of 𝑅 in 𝐾. Show that the map 𝐼 ↦→ 𝐼 ⊗𝑅 𝑆 defines a group homomorphism
Pic O → Pic(O ⊗𝑅 𝑆).
Chapter 19
Brandt groupoids
In this chapter, we study the relationship between multiplication and classes of quater-
nion ideals.
19.1 ⊲ Composition laws and ideal multiplication

To guide our investigations, we again appeal to the quadratic case. Let 𝑑 ∈ Z be
a nonsquare discriminant. A subject of classical interest was the set of integral
primitive binary quadratic forms of discriminant 𝑑, namely
Q (𝑑) = {𝑎𝑥 2 + 𝑏𝑥𝑦 + 𝑐𝑦 2 : 𝑎, 𝑏, 𝑐 ∈ Z, 𝑏 2 − 4𝑎𝑐 = 𝑑, and gcd(𝑎, 𝑏, 𝑐) = 1}.
Of particular interest to early number theorists (Fermat, Legendre, Lagrange, and

Gauss) was the set of primes represented by a quadratic form 𝑄 ∈ Q (𝑑); inquiries of
this nature proved to be quite deep, giving rise to the law of quadratic reciprocity and
the beginnings of the theory of complex multiplication and class field theory.
An invertible, oriented change of variables on a quadratic form 𝑄 ∈ Q (𝑑) does
not alter the set of primes represented, so one is naturally led to study the equivalence
classes of quadratic forms under the (right) action of the group SL2 (Z) given by
(𝑄 | 𝑔) (𝑥, 𝑦) = 𝑄((𝑥, 𝑦) · 𝑔) for 𝑔 ∈ SL2 (Z). (19.1.1)
The set Cl(𝑑) of SL2 (Z)-classes of forms in Q (𝑑) is finite, by reduction theory (see
section 35.2): every form in Q (𝑑) is equivalent under the action of SL2 (Z) to a unique
reduced form, of which there are only finitely many. To study this finite set, Gauss
defined a composition law on Cl(𝑑), giving Cl(𝑑) the structure of an abelian group
by an explicit formula. Gauss’s composition law on binary quadratic forms can be
understood using 2 × 2 × 2 Rubik’s cubes, by a sublime result of Bhargava [Bha2004a].
Today, we see this composition law as a consequence of a natural bĳection between
Cl(𝑑) and a set equipped with an obvious group structure. Let 𝑆 = 𝑆(𝑑) be the quadratic
ring of discriminant 𝑑. Define the narrow class group Cl+ (𝑆) as
the group of invertible fractional ideals of 𝑆 under multiplication
301
302 CHAPTER 19. BRANDT GROUPOIDS
modulo
the subgroup of nonzero principal fractional ideals
generated by a totally positive element
(i.e., one that is positive in every embedding into R, so if 𝑑 < 0 then this is no
condition). (Alternatively, Cl+ (𝑆) can be thought of as the group of isomorphism
classes of oriented, invertible 𝑆-modules, under a suitable notion of orientation.) Then
there is a bĳection between Cl(𝑑) and Cl+ (𝑆): explicitly, to the class of the quadratic
form 𝑄 = 𝑎𝑥 2 + 𝑏𝑥𝑦 + 𝑐𝑦 2 ∈ Q (𝑑), we associate the class of the ideal
√
−𝑏 + 𝑑
𝔞 = 𝑎Z + Z ⊂ 𝑆(𝑑). (19.1.2)
2
√
Conversely, the quadratic form is recovered as the norm form on 𝐾 = Q( 𝑑) restricted
to 𝔞: √
−𝑏 + 𝑑
Nm𝐾 /Q 𝑎𝑥 + 𝑦 = 𝑎𝑥 2 + 𝑏𝑥𝑦 + 𝑐𝑦 2 (19.1.3)
2
where 𝑐 = (𝑏 2 − 𝑑)/(4𝑎) ∈ Z.
Much of the same structure can be found in the quaternionic case, with several
interesting twists. It was Brandt who first asked if there was a composition law for
(integral, primitive) quaternary quadratic forms: it would arise naturally from some
kind of multiplication of ideals in a quaternion order, with the analogous bĳection
furnished by the reduced norm form. Brandt started writing on composition laws
for quaternary quadratic forms in 1913 [Bra13], tracing the notion of composition
back to Hermite, who observed a kind of multiplication law (bilinear substitution) for
quaternary forms 𝑥02 + 𝐹 (𝑥1 , 𝑥2 , 𝑥3 ) in formulas of Euler and Lagrange. He continued
on this note during the 1920s [Bra24, Bra25, Bra28, Bra37], when it became clear that
quaternion algebras was the right framework to place his composition laws; in 1943,
he developed this theme significantly [Bra43] and defined his Brandt matrices (that
will figure prominently in Chapter 41.
However, in the set of invertible lattices in 𝐵 under compatible product, one cannot
always multiply! However, this set has the structure of a groupoid: a nonempty set
with an inverse function and a partial product that satisfies the associativity, inverse, and
identity properties whenever they are defined. Groupoids now figure prominently in
category theory (a groupoid is equivalently a small category in which every morphism
is an isomorphism) and many other contexts; see Remark 19.3.11.
Organizing lattices by their left and right orders, which by definition are connected
and hence in the same genus, we define
Brt(O) = {𝐼 : 𝐼 ⊂ 𝐵 invertible 𝑅-lattice and OL (𝐼), OR (𝐼) ∈ Gen O}; (19.1.4)
visibly, Brt(O) depends only on the genus of O. Organizing lattices according to the
genus of orders is sensible: after all, we only apply the composition law to binary
quadratic forms of the same discriminant, and in the compatible product we see
precisely those classes whose left and right orders are connected.
√ In other words,
the set of invertible lattices in the quadratic field 𝐾 = Q( 𝑑) has the structure of a
groupoid if we multiply only those lattices with the same multiplicator ring.
19.1. ⊲ COMPOSITION LAWS AND IDEAL MULTIPLICATION 303

Then the set Brt(O) has the structure of a groupoid under compatible product.
We call Brt(O) the Brandt groupoid of (the genus of) O.
We now consider classes of lattices. A lattice 𝐼 ⊂ 𝐵 has the structure of a
OL (𝐼), OR (𝐼)-bimodule. Two invertible lattices 𝐼, 𝐽 with the same left and right
orders OL (𝐼) = OL (𝐽) and OR (𝐼) = OR (𝐽) are isomorphic as bimodules if and only
if there exists 𝑎 ∈ Q× such that 𝐽 = 𝑎𝐼. Accordingly, we say two lattices 𝐼, 𝐽 ⊂ 𝐵 are
homothetic if there exists 𝑎 ∈ Q× such that 𝐽 = 𝑎𝐼.
For connected orders O, O0 ⊂ 𝐵, we define
Pic(O, O0) := {[𝐼] : 𝐼 ⊂ 𝐵 invertible and OL (𝐼) = O and OR (𝐼) = O0 } (19.1.6)
to be the set of homothety classes of lattices with left order O and right order O0, or
equivalently the set of isomorphism classes of O, O0-bimodules over 𝑅. Restricting to
the subset of lattices with O = O0, and the lattices 𝐼 ⊂ 𝐵 are O-bimodules, we recover
Pic(O, O) = Pic O the Picard group from the previous chapter.
Now let O ⊂ 𝐵 be an order and let O𝑖 be representative orders for the type set
Typ O. Let Ä
BrtCl O := Pic(O𝑖 , O 𝑗 ). (19.1.7)
𝑖, 𝑗

Then the set BrtCl O has the structure of a groupoid that, up to isomorphism, is
independent of the choice of the orders O𝑖 .
In particular, BrtCl O depends only on the genus of O. We call the set BrtCl O the
Brandt class groupoid of (the genus of) O.
Returning to quadratic forms, to each 𝑅-lattice 𝐼 with nrd(𝐼) = 𝑎Z and 𝑎 > 0, we
associate the quadratic form
nrd𝐼 : 𝐼 → Z
nrd𝐼 (𝜇) = nrd(𝜇)/𝑎
Alternatively, up to similarity we can just take the quadratic module nrd | 𝐼 : 𝐼 → nrd(𝐼)
remembering that the quadratic form takes values in nrd(𝐼). The discriminant of an
invertible lattice 𝐼 ⊂ 𝐵 is equal to the common discriminant 𝑁 2 of the genus of its
left or right order. The quadratic forms nrd𝐼 are all locally similar, respecting the
canonical orientation 5.6.7 on 𝐵. Therefore, there is a map
( Quaternary quadratic forms over Z )
BrtCl O → locally similar to nrd | O
up to oriented similarity
[𝐼] ↦→ nrd𝐼
is (well-defined and) surjective. Unfortunately, this map is not injective (a reflection
of the lack of a natural quotient groupoid homomorphism): the Brandt class is a kind
of rigidification of the oriented similarity class. Nevertheless, Theorem 19.1.8 can be
viewed as a generalization of Gauss composition of binary quadratic forms, defining a
partial composition law on (rigidified) classes of quaternary quadratic forms.
19.2 Example

−2, −37
Consider the quaternion algebra 𝐵 := with standard basis 1, 𝑖, 𝑗, 𝑘 = 𝑖 𝑗,
Q
and the maximal order O of reduced discriminant 37 defined by
1+𝑖+ 𝑗 2+𝑖+ 𝑘
O := Z + 𝑖Z + Z+ Z. (19.2.1)
2 4
The type set Typ O of orders connected to O has exactly two isomorphism classes,
represented by O1 = O and
3 − 7𝑖 + 𝑗 2 − 3𝑖 + 𝑘
O2 := Z + 3𝑖Z + Z+ Z.
6 4
These orders are connected by the O2 , O1 -connecting ideal
3−𝑖+ 𝑗 2 + 3𝑖 − 𝑘 3−𝑖+ 𝑗
𝐼 := 3Z + 3𝑖Z + Z+ Z = 3O + O.
2 4 2
There are isomorphisms
Pic(O) ' Pic(O2 ) ' Z/2Z
with the nontrivial class in Pic(O1 ) represented by the principal two-sided ideal 𝐽1 =
𝑗O = O 𝑗 with 𝑗 ∈ 𝑁 𝐵× (O), and the nontrivial class in Pic(O2 ) represented by the
nonprincipal (but invertible) ideal
111 − 259𝑖 + 𝑗
𝐽2 := 𝐼𝐽1 𝐼 −1 = 37O2 + O2 .
6
In particular,
ClsR (O1 ) = {[O1 ], [𝐼], [𝐽2 𝐼]} and ClsR (O2 ) = {[O2 ], [𝐽2 ], [𝐼]},
with [𝐽1 ] = [O1 ].

We can visualize this groupoid as a graph as in Figure 19.2.2, with directed edges
for multiplication:
[IJ1 ] = [J2 I]
[J1 ] = [J1−1 ] [I] [J2 ] = [J2−1 ]

1 2
[O1 ] [I] = [I −1
] [O2 ]
[IJ2 ] = [J1 I]
Figure 19.2.2: BrtCl O, for discrd O = 37
The Brandt class groupoid
BrtCl O = Pic(O1 ) t Cl(O1 , O2 ) t Cl(O2 , O1 ) t Pic(O2 )

19.3. GROUPOID STRUCTURE 305
has 2 + 4 + 4 + 2 = 12 elements; it is generated as a groupoid by the elements

[𝐽1 ], [𝐽2 ], [𝐼], with relations
[𝐽1 ] 2 = [O1 ], [𝐽2 ] 2 = [O2 ], [𝐽2 ] [𝐼] = [𝐼] [𝐽1 ].
Restricting the reduced norm to these lattices, we obtain classes of quaternary

quadratic forms of discriminant 372 :
nrdO1 = 𝑡 2 + 𝑡𝑦 + 𝑡𝑧 + 2𝑥 2 + 𝑥𝑦 + 2𝑥𝑧 + 5𝑦 2 + 𝑦𝑧 + 10𝑧2

nrdO2 = 𝑡 2 + 𝑡𝑥 + 𝑡𝑧 + 4𝑥 2 − 𝑥𝑦 + 4𝑥𝑧 + 5𝑦 2 + 2𝑦𝑧 + 6𝑧 2
nrd𝐼 = 3𝑡 2 − 𝑡𝑥 + 𝑡𝑦 + 𝑡𝑧 + 3𝑥 2 − 3𝑥𝑦 − 𝑥𝑧 + 4𝑦 2 − 𝑦𝑧 + 5𝑧2
nrd𝐼 = 3𝑡 2 + 𝑡𝑥 − 𝑡𝑦 − 𝑡𝑧 + 3𝑥 2 − 3𝑥𝑦 − 𝑥𝑧 + 4𝑦 2 − 𝑦𝑧 + 5𝑧2
nrd 𝐽2 = 2𝑡 2 − 𝑡𝑥 + 𝑡𝑦 + 2𝑥 2 − 2𝑥𝑦 + 𝑥𝑧 + 3𝑦 2 + 2𝑦𝑧 + 10𝑧 2
The quadratic forms nrd𝐼 and nrd𝐼 are isometric but not by an oriented isometry.
19.3 Groupoid structure

We begin with some generalities on groupoids.
Definition 19.3.1. A partial function 𝑓 : 𝑋 → 𝑌 is a function defined on a subset of

the domain 𝑋.
Definition 19.3.2. A groupoid 𝐺 is a set with a unary operation −1 : 𝐺 → 𝐺 and a

partial function ∗ : 𝐺 × 𝐺 → 𝐺 such that ∗ and −1 satisfy the associativity, inverse,
and identity properties (as in a group) whenever they are defined:
(a) [Associativity] For all 𝑎, 𝑏, 𝑐 ∈ 𝐺 such that 𝑎 ∗ 𝑏 is defined and (𝑎 ∗ 𝑏) ∗ 𝑐 is

defined, both 𝑏 ∗ 𝑐 and 𝑎 ∗ (𝑏 ∗ 𝑐) are defined and
(𝑎 ∗ 𝑏) ∗ 𝑐 = 𝑎 ∗ (𝑏 ∗ 𝑐).
(b) [Inverses] For all 𝑎 ∈ 𝐺, there exists 𝑎 −1 ∈ 𝐺 such that 𝑎 ∗ 𝑎 −1 and 𝑎 −1 ∗ 𝑎 are
defined (but not necessarily equal).
(c) [Identity] For all 𝑎, 𝑏 ∈ 𝐺 such that 𝑎 ∗ 𝑏 is defined, we have
(𝑎 ∗ 𝑏) ∗ 𝑏 −1 = 𝑎 and 𝑎 −1 ∗ (𝑎 ∗ 𝑏) = 𝑏. (19.3.3)
A homomorphism 𝜙 : 𝐺 → 𝐺 0 of groupoids is a map satisfying
𝜙(𝑎 ∗ 𝑏) = 𝜙(𝑎) ∗ 𝜙(𝑏)
for all 𝑎, 𝑏 ∈ 𝐺.
19.3.4. Let 𝐺 be a groupoid. Then the products in the identity law (19.3.3) are defined
by the associative and inverse laws, and it follows that 𝑒 = 𝑎 ∗ 𝑎 −1 , the left identity of
𝑎, and 𝑓 = 𝑎 −1 ∗ 𝑎 the corresponding right identity of 𝑎, satisfy 𝑒 ∗ 𝑎 = 𝑎 = 𝑎 ∗ 𝑓
for all 𝑎 ∈ 𝐺. (We may have that 𝑒 ≠ 𝑓 , i.e., the left and right identities for 𝑎 ∈ 𝐺
disagree.) The right identity of 𝑎 ∈ 𝐺 is the left identity of 𝑎 −1 ∈ 𝐺, so we call the set
{𝑒 = 𝑎 ∗ 𝑎 −1 : 𝑎 ∈ 𝐺}
the set of identity elements in 𝐺.
19.3.5. Equivalently, a groupoid is a small category (the class of objects in the category
is a set) such that every morphism is an isomorphism: given a groupoid, we associate
the category whose objects are the elements of the set 𝑆 := {𝑒 = 𝑎 ∗ 𝑎 −1 : 𝑎 ∈ 𝐺} of
identity elements in 𝐺 and the morphisms between 𝑒, 𝑓 ∈ 𝑆 are the elements 𝑎 ∈ 𝐺
such that 𝑒 ∗ 𝑎 and 𝑎 ∗ 𝑓 are defined (see Proposition 19.3.9 below). Conversely, to
a category in which every morphism is an isomorphism, we associate the groupoid
whose underlying set is the union of all morphisms under inverse and composition.
Example 19.3.6. The set of homotopy classes of paths in a topological space 𝑋 forms
a groupoid under composition: the paths 𝛾1 , 𝛾2 : [0, 1] → 𝑋 can be composed to a
path 𝛾2 ◦ 𝛾1 : [0, 1] → 𝑋 if and only if 𝛾2 (0) = 𝛾1 (1).
Example 19.3.7. A disjoint union of groups is a groupoid, with the product defined if
and only if the elements belong to the same group; the set of identities is canonically
in bĳection with the index set of the disjoint union.
19.3.8. Let 𝐺 be a groupoid and let 𝑒, 𝑓 ∈ 𝐺 be identity elements. We say that

𝑒 is connected to 𝑓 if there exists 𝑎 ∈ 𝐺 such that 𝑎 has left identity 𝑒 and right
identity 𝑓 . The relation of being connected defines an equivalence relation on the set
of identity elements in 𝐺, and the resulting equivalence classes are called connected
components of 𝐺. We say 𝐺 is connected if all identity elements 𝑒, 𝑓 ∈ 𝐺 are
connected; connected components of a groupoid are connected.
Viewing the groupoid 𝐺 as a small category as in 19.3.5, we say two objects are
connected if there exists a morphism between them, and the category is connected if
every two objects are connected.
If 𝑒 ∈ 𝐺 is an identity element in a groupoid 𝐺, then the set of elements 𝑎 ∈ 𝐺

with left and right identity equal to 𝑒 has the structure of a group; for the associated
category, this is the automorphism group of the object. More generally, the following
structural result holds.
Proposition 19.3.9. Let 𝐺 be a connected groupoid, and let 𝑒, 𝑓 be identity elements

in 𝐺. Let
𝐺 (𝑒, 𝑓 ) := {𝑎 ∈ 𝐺 : 𝑒 ∗ 𝑎 and 𝑎 ∗ 𝑓 are defined}.
Then the following statements hold.
(a) The set 𝐺 (𝑒, 𝑒) is a group under ∗.

(b) There is a (noncanonical) isomorphism 𝐺 (𝑒, 𝑒) ' 𝐺 ( 𝑓 , 𝑓 ).
(c) The set 𝐺 (𝑒, 𝑓 ) is a principal homogeneous space for 𝐺 (𝑒, 𝑒) ' 𝐺 ( 𝑓 , 𝑓 ).
19.4. BRANDT GROUPOID 307
Proof. The set 𝐺 (𝑒, 𝑒) is nonempty, has the identity element 𝑒 ∈ 𝐺, and if 𝑎 ∈ 𝐺 (𝑒, 𝑒)
then 𝑎∗𝑎 −1 = 𝑎 −1 ∗𝑎 = 𝑒. If 𝑒, 𝑓 are identity elements, since 𝐺 is connected there exists
𝑎 ∈ 𝐺 (𝑒, 𝑓 ), so 𝑎 −1 ∈ 𝐺 ( 𝑓 , 𝑒) and the map 𝐺 (𝑒, 𝑒) → 𝐺 ( 𝑓 , 𝑓 ) by 𝑥 ↦→ 𝑎 ∗ 𝑥 ∗ 𝑎 −1
is an isomophism of groups. Similarly, the set 𝐺 (𝑒, 𝑓 ) has a right, simply transitive
action of 𝐺 (𝑒, 𝑒) under right multiplication by ∗.
19.3.10. The moral of Proposition 19.3.9 is that the only two interesting invariants of
a connected groupoid are the number of identity elements (objects in the category) and
the group of elements with a common left and right identity (the automorphism group
of every one of the objects). A connected groupoid is determined up to isomorphism
of groupoids by these two properties.
Remark 19.3.11. After seeing its relevance in the context of composition of quaternary
forms, Brandt set out general axioms for his notion of a groupoid [Bra27, Bra40].
(Brandt’s original definition of groupoid is now called a connected groupoid.) This
notion has blossomed into an important structure in mathematics that sees quite general
use, especially in homotopy theory and category theory. It is believed that the groupoid
axioms influenced the work of Eilenberg–Mac Lane [EM45] in the first definition of
a category: see e.g., Brown [Bro87] for a survey, Bruck [Bruc71] for context in the
theory of binary structures, as well as the article by Weinstein [Wein96].
Groupoids exhibit many facets of mathematics, arising naturally in functional anal-
ysis (𝐶 ∗ -algebras) and group representations, as Figure 19.3.12 indicates (appearing
in Williams [Will2001, p. 21], and attributed to Arlan Ramsay).
group
actions
groups
sets
equivalence
groupoids relations
Figure 19.3.12: Groupoids, as they relate to other mathematical objects

(In this diagram, for example, a set 𝑋 is a groupoid with only the multiplications
𝑥 ∗ 𝑥 = 𝑥 for 𝑥 ∈ 𝑋. The corner between sets and groups can be explained by a set
with one element which can be made into a group in a unique way.)
19.4 Brandt groupoid

Let 𝑅 be a Dedekind domain with field of fractions 𝐹 and let 𝐵 be a quaternion algebra
over 𝐹.
Proposition 19.4.1. The set of invertible 𝑅-lattices in 𝐵 is a groupoid under inverse

and compatible product; the 𝑅-orders in 𝐵 are the identity elements in this groupoid.
Proof. For the associative law, suppose 𝐼, 𝐽, 𝐾 are invertible 𝑅-lattices with 𝐼𝐽 and
(𝐼𝐽)𝐾 compatible products. Then OR (𝐼) = OL (𝐽) = OL (𝐽𝐾) and OR (𝐼𝐽) = OR (𝐽) =
OL (𝐾) by Lemma 16.5.11, so the products 𝐽𝐾 and 𝐼 (𝐽𝐾) are compatible. Multipli-
cation is associative in 𝐵, and it follows that 𝐼 (𝐽𝐾) = (𝐼𝐽)𝐾. Inverses exist exactly
because we restrict to the invertible lattices.
The law of identity holds as follows: if 𝐼, 𝐽 are invertible 𝑅-lattices such that 𝐼𝐽 is
a compatible product, then (𝐼𝐽)𝐽 −1 is a compatible product since OR (𝐼𝐽) = OR (𝐽) =
OL (𝐽 −1 ), and by associativity
(𝐼𝐽)𝐽 −1 = 𝐼 (𝐽𝐽 −1 ) = 𝐼OL (𝐽) = 𝐼OR (𝐼) = 𝐼,
with a similar argument on the left. If 𝐼 is an invertible 𝑅-lattice, then 𝐼 𝐼 −1 = OL (𝐼)

is an 𝑅-order in 𝐵, and every 𝑅-order O arises by taking 𝐼 = O itself, so the 𝑅-orders
are the identity elements in the groupoid.
Lemma 19.4.2. The connected components of the groupoid of invertible 𝑅-lattices in

𝐵 are identified by the genus of the (left or) right order, and the group defined on such
a component corresponding to an order O is Idl(O), the group of invertible two-sided
O-ideals.
Proof. By Proposition 19.4.1, the identity elements correspond to orders, and two
orders are connected if and only if there is a (invertible, equivalently locally principal)
connecting ideal if and only if they are in the same genus, as in section 17.4. The
second statement follows immediately.
As a consequence of Lemma 19.4.2, the subset of 𝑅-lattices whose (left or) right
order belong to a specified genus of orders is a connected subgroupoid.
Definition 19.4.3. Let O ⊆ 𝐵 be an 𝑅-order. The Brandt groupoid of (the genus of)
O is
Brt(O) = {𝐼 : 𝐼 ⊂ 𝐵 invertible 𝑅-lattice and OL (𝐼), OR (𝐼) ∈ Gen O}.
In the next section, we consider a variant that considers classes of lattices, giving
rise to a finite groupoid.
19.5 Brandt class groupoid

We now organize lattices up to isomorphism as bimodules for their left and right
orders.
Lemma 19.5.1. Let 𝐼, 𝐽 ⊂ 𝐵 be lattices with OL (𝐼) = OL (𝐽) = O and OR (𝐼) =

OR (𝐽) = O0. Then 𝐼 is isomorphic to 𝐽 as O, O0-bimodules if and only if there exists
𝑎 ∈ 𝐹 × such that 𝐽 = 𝑎𝐼.
19.5. BRANDT CLASS GROUPOID 309
Proof. We have 𝐹 = 𝑍 (𝐵). If 𝐽 = 𝑎𝐼 with 𝑎 ∈ 𝐹 × , then multiplication by 𝑎 gives an

𝑅-module isomorphism 𝐼 → 𝐽 that commutes with the left and right actions and so
defines a O, O0-bimodule isomorphism.
∼ 𝐽 is a O, O0-bimodule isomorphism. Then
Conversely, suppose that 𝜙 : 𝐼 −→
𝜙(𝜇𝛼𝜈) = 𝜇𝜙(𝛼)𝜈 for all 𝛼 ∈ 𝐼 and 𝜇, 𝜈 ∈ O. Extending scalars to 𝐵, we obtain a
𝐵-bimodule isomorphism 𝜙 : 𝐼 𝐹 = 𝐵 → 𝐽𝐹 = 𝐵. Let 𝜙(1) = 𝛽. Then for all 𝛼 ∈ 𝐵,
we have 𝜙(𝛼) = 𝜙(1)𝛼 = 𝛽𝛼; but by the same token, 𝜙(𝛼) = 𝛼𝛽 for all 𝛼 ∈ 𝐵, so
𝛽 ∈ 𝑍 (𝐵) = 𝐹.
Definition 19.5.2. Let 𝐼, 𝐽 ⊆ 𝐵 be 𝑅-lattices. We say that 𝐼 is homothetic to 𝐽 if there
exists 𝑎 ∈ 𝐹 × such that 𝐽 = 𝑎𝐼.
Homothety defines an equivalence relation, and we let [𝐼] denote the homothety
class of an 𝑅-lattice 𝐼. The left and right order of a homothety class is well-defined.
19.5.3. The set of homothety classes of invertible 𝑅-lattices 𝐼 ⊆ 𝐵 has the structure
of a groupoid under compatible product, since the compatible product [𝐼𝐽] is well-
defined: if 𝐼 0 = 𝑎𝐼 and 𝐽 0 = 𝑏𝐽 with 𝑎, 𝑏 ∈ 𝐹 × , then [𝐼 0 𝐽 0] = [𝑎𝑏𝐼𝐽] = [𝐼𝐽] since
𝑎, 𝑏 are central.
The map which takes an invertible lattice to its homothety class yields a surjective
homomorphism of groupoids. Taking connected components we obtain a connected
groupoid associated to a (genus of an) 𝑅-order O. Recalling 19.3.10, we note that the
group at an 𝑅-order O is Pic𝑅 (O), but there are still infinitely many orders (objects in
the category).
In order to whittle down to a finite groupoid, we fix representatives of the type set,
and make the following definitions.
19.5.4. For 𝑅-orders O, O0 ⊆ 𝐵, let
Pic𝑅 (O, O0) = {[𝐼] : 𝐼 ⊂ 𝐵 invertible and OL (𝐼) = O, OR (𝐼) = O0 }
be the set of homothety classes of 𝑅-lattices in 𝐵 with left order O and right order
O0; equivalently, by Lemma 19.5.1, Pic(O, O0) is the set of isomorphism classes of
invertible O, O0-bimodules over 𝑅. In particular, Pic𝑅 (O) = Pic𝑅 (O, O).
We have Pic𝑅 (O, O0) ≠ ∅ if and only if O is connected to O0.
Let O ⊂ 𝐵 be an order and let O𝑖 be representative orders for the type set Typ O.
We define Ä
BrtCl O := Pic𝑅 (O𝑖 , O 𝑗 ).
𝑖, 𝑗
Theorem 19.5.5. Let 𝑅 be a Dedekind domain with field of fractions 𝐹, and let 𝐵 be
a quaternion algebra over 𝐹. Let O ⊂ 𝐵 be an order. Then the set BrtCl O has the
structure of a finite groupoid that, up to isomorphism, is independent of the choice of
the orders O𝑖 .
In particular, by Theorem 19.5.5 BrtCl O depends only on the genus of O up to
groupoid isomorphism. We call the set BrtCl O the Brandt class groupoid of (the
genus of) O.
Proof. The groupoid structure is compatible multiplication, with

Pic𝑅 (O𝑖 , O 𝑗 ) Pic𝑅 (O 𝑗 , O 𝑘 ) ⊆ Pic𝑅 (O𝑖 , O 𝑘 )
for all 𝑖, 𝑗, 𝑘; in other words, BrtCl O is a connected subgroupoid of the groupoid of
homothety classes of 𝑅-lattices 19.5.4.
The groupoid is finite, by 19.3.10: the type set Typ O is finite by Main Theorem
17.7.1 and Pic𝑅 (O) is finite by Proposition 18.4.10. Explicitly, if [𝐼𝑖 𝑗 ] ∈ Pic𝑅 (O𝑖 , O 𝑗 )
then the map
Pic𝑅 (O) ' Pic𝑅 (O𝑖 ) → Pic𝑅 (O𝑖 , O 𝑗 )
[𝐼] ↦→ [𝐼 𝐼𝑖 𝑗 ]
is a bĳection of sets, just as in the proof of Proposition 19.3.9. Therefore
# BrtCl O = # Pic𝑅 (O)# Typ O. (19.5.6)
Finally, this subgroupoid is independent of the choices of the orders O𝑖 as follows:
all other choices correspond to O𝑖0 = 𝛼𝑖 O𝑖 𝛼𝑖−1 with 𝛼𝑖 ∈ 𝐵× , and the induced maps
Pic𝑅 (O𝑖 , O 𝑗 ) → Pic(O𝑖0 , O0𝑗 )
[𝐼] ↦→ [𝛼𝑖 𝐼𝛼−1 0
𝑗 ] = [𝐼 ]
together give an isomorphism of groupoids, since

[𝐼 0 𝐽 0] = [𝛼𝑖 𝐼𝛼−1 −1 −1
𝑗 𝛼 𝑗 𝐽𝛼 𝑘 ] = [𝛼𝑖 𝐼𝐽𝛼 𝑘 ]
for all [𝐼] ∈ Pic𝑅 (O𝑖 , O 𝑗 ) and [𝐽] ∈ Pic𝑅 (O 𝑗 , O 𝑘 ).

Remark 19.5.7. Unfortunately, there is not in general a natural equivalence relation on
Brt(O) giving rise to a quotient groupoid homomorphism Brt(O) → BrtCl O. Rather,
we find that BrtCl O is naturally a subgroupoid of Brt(O).
Turning to the invariants 19.3.10, we see that the Brant class groupoid BrtCl O
encodes two things: the group Pic𝑅 (O) and the type set Typ O.
Remark 19.5.8. The modern theory of Brandt composition was investigated by Ka-
plansky [Kap69] and generalized to Azumaya quaternion algebras over commutative
rings by Kneser–Knus–Ojanguren–Parimala–Sridharan [KKOPS86].
19.6 Quadratic forms

We now connect the Brandt class groupoid to quadratic forms. For simplicity, we
suppose char 𝐹 ≠ 2 throughout this section.
19.6.1. We begin by recalling Proposition 4.5.17: for the quaternary quadratic form
nrd : 𝐵 → 𝐹, every oriented similarity of nrd is of the form
𝐵 ↦→ 𝐵
𝑥 ↦→ 𝛼𝑥 𝛽−1
with 𝛼, 𝛽 ∈ 𝐵 (in particular, respecting the canonical orientation 5.6.7 of 𝐵); the
similitude factor of such a map is 𝑢 = nrd(𝛼)/nrd(𝛽).
Let 𝐼 ⊂ 𝐵 be a projective 𝑅-lattice.

19.6.2. Generalizing Exercise 10.2, the reduced norm restricts to give a quadratic form
on 𝐼. We are given that 𝐼 is projective of rank 4 as an 𝑅-module. Therefore the map
nrd𝐼 : 𝐼 → 𝐿 = nrd(𝐼)
is a quaternary quadratic module over 𝑅.

If 𝐽 ⊂ 𝐵 is another projective 𝑅-lattice, and 𝑓 is an oriented similarity from nrd𝐼 to
nrd 𝐽 , then extending scalars by 𝐹 we obtain a oriented self-similarity of nrd : 𝐵 → 𝐵;
by 19.6.1, we conclude that 𝐽 = 𝛼𝐼 𝛽−1 for some 𝛼, 𝛽 ∈ 𝐵× :
𝐼
nrd 𝐼
/𝐿
o 𝛼·𝛽 o 𝑎𝑏 (19.6.3)

𝐽 = 𝛼𝐼 𝛽
nrd 𝐽
/ 𝑎𝑏𝐿
19.6.4. Suppose that nrd(𝐼) = 𝐿 = 𝑎𝑅 is principal. Then there is a similarity
𝐼
nrd 𝐼
/ 𝐿 = 𝑎𝑅
o 𝑎−1 (19.6.5)
𝑎−1
𝐼
nrd 𝐼
/𝑅
In other words, if every value of the quadratic form is divisible by 𝑎, then up to

similarity it is equivalent to consider the quadratic form 𝑎 −1 nrd, taking values in 𝑅.
Lemma 19.6.6. Suppose 𝐼 is invertible. Then the quadratic form nrd𝐼 : 𝐼 → 𝐿 is
locally oriented similar to nrdO : O → 𝑅, where O = OR (𝐼).
Proof. By 19.6.3, if 𝐼 = 𝛼O is principal, then nrd𝐼 is similar to nrdO . If 𝐼 is invertible,
then 𝐼 is locally principal, so for all primes 𝔭 of 𝑅 the quadratic form nrd : 𝐼𝔭 → 𝐿 𝔭 is
similar to nrd : O𝔭 → 𝑅𝔭 where O𝔭 is the right order of 𝐼𝔭 . (A similar statement holds
on the left.)
19.6.7. From Lemma 15.3.6, it follows from Lemma 19.6.6 that
disc(nrd𝐼 ) = disc(O)
and in particular this discriminant is a square.

The quadratic forms nrd𝐼 are all locally similar, respecting the canonical orientation
5.6.7 on 𝐵. Therefore, the map
( Quaternary quadratic forms over Z )
BrtCl O → locally similar to nrdO
up to oriented similarity
[𝐼] ↦→ nrd𝐼
is (well-defined and) surjective.
Remark 19.6.8. The Brandt groupoid is connected as a groupoid. This can also be
viewed in the language of quadratic forms: a connected class of orders is equivalently
a genus of integral ternary quadratic forms, and this is akin to a resolvent for the
quaternary norm forms. We refer to Chapter 23 for further development.
Exercises
1. Verify the computational details in the example of section 19.2.
2. Let 𝐵 = (−1, −11 | Q) with disc 𝐵 = 11 and O = Zh𝑖, (1 + 𝑗)/2i a maximal
order. Compute BrtCl O, in a manner analogous to the example of section 19.2.
3. Let 𝐺 be a groupoid.
(a) Show that if 𝑎, 𝑏, 𝑐 ∈ 𝐺 and both 𝑎 ∗ 𝑏 and 𝑎 ∗ 𝑐 are defined, then
𝑏 ∗ 𝑏 −1 = 𝑐 ∗ 𝑐−1 (and both are defined).
(b) Show that for all 𝑎 ∈ 𝐺 we have (𝑎 −1 ) −1 .
4. Let 𝐺 be a group acting on a nonempty set 𝑋. Let
𝐴(𝐺, 𝑋) = {(𝑔, 𝑥) : 𝑔 ∈ 𝐺, 𝑥 ∈ 𝑋 }.
Show that 𝐴(𝐺, 𝑋) has a natural groupoid structure with (𝑔, 𝑥) ∗ (ℎ, 𝑦) = (𝑔ℎ, 𝑦)
defined if and only if 𝑥 = ℎ𝑦. What are the identity elements?
5. Show that in a homomorphism 𝜙 : 𝐺 → 𝐺 0 of groupoids, the set of identity
elements of 𝐺 maps to the set of identity elements of 𝐺 0.
6. Let C be a small category. Show that there is a unique maximal subcategory
that is a groupoid. [Hint: Discard all nonisomorphisms.]
7. Let 𝑋 be a set and let ∼ be an equivalence relation on 𝑋, thought of as a
subset 𝑆 ⊆ 𝑋 × 𝑋. Equip 𝑆 with the partial binary operation ∗ defined by
(𝑥, 𝑦) ∗ (𝑦, 𝑧) = (𝑥, 𝑧) for (𝑥, 𝑦), (𝑦, 𝑧) ∈ 𝑆 (and (𝑥, 𝑦) ∗ (𝑤, 𝑧) is not defined if
𝑦 ≠ 𝑤). Show that 𝑆 is a groupoid. [This shows that “equivalence relations are
groupoids”, cf. (19.3.12).]
8. Let 𝐹 be a field and let GL(𝐹) = ∞
Ð
𝑛=1 GL𝑛 (𝐹). Show that GL(𝐹) has a natural
structure of groupoid, sometimes called the general linear groupoid over 𝐹.
9. Show that the reduced norm is a homomorphism from the groupoid of invertible
𝑅-lattices in 𝐵 to the group(oid) of fractional 𝑅-ideals in 𝐹.
10. Let 𝑋 be a nonempty topological space, and let 𝑥, 𝑦 ∈ 𝑋. Recall that a path
from 𝑥 to 𝑦 is a continuous map 𝜐0 : [0, 1] → 𝑋 with 𝜐(0) = 𝑥 and 𝜐(1) = 𝑦.
We say that paths 𝜐0 , 𝜐1 : [0, 1] → 𝑋 from 𝑥 to 𝑦 are homotopic if there exists
a continuous map 𝐻 : [0, 1] × [0, 1] → 𝑋 such that 𝐻 (0, 𝑠) = 𝑥 and 𝐻 (1, 𝑠) = 𝑦
for all 𝑠 ∈ [0, 1] and 𝐻 (𝑡, 0) = 𝜐0 and 𝐻 (𝑡, 1) = 𝜐1 (𝑡) for all 𝑡 ∈ [0, 1]. [So
each 𝐻 (𝑡, 𝑠) for fixed 𝑡 ∈ [0, 1] is a path from 𝑥 to 𝑦, and this set of paths varies
continuously.]
(a) Check that being homotopic defines an equivalence relation on the set of
continuous paths from 𝑥 to 𝑦.
(b) Check that paths can be composed (going at twice speed) and that compo-
sition of paths is well-defined on homotopy classes.
(c) Show that composition of homotopy classes of continuous paths is asso-
ciative.
Let Π(𝑋) be the category whose objects are the points of 𝑋 and with mor-
phisms to be the set of homotopy classes of continuous paths from 𝑥 to 𝑦 under
composition.
(d) Show that Π(𝑋) is a category.
(e) Show that Π(𝑋) is a groupoid, called the fundamental groupoid of 𝑋.
(f) Finally, for all 𝑥 ∈ 𝑋, show the set of all morphisms from 𝑥 to 𝑥 in Π(𝑋) is
a group (the more familiar fundamental group 𝜋1 (𝑋, 𝑥) with base point
𝑥).
11. Continuing the previous exercise, show that if 𝑋 is path-connected, then Π(𝑋)
is equivalent as a category to a groupoid with one object. [Hint: choose a point
𝑥 ∈ 𝑋, look at the group(oid) 𝜋1 (𝑋, 𝑥).]
Chapter 20
Integral representation theory
In this chapter, we consider a slightly more general framework on the preceding

chapters: we consider lattices as projective modules, and relate this to invertibility and
representation theory in an integral sense.
20.1 ⊲ Projectivity, invertibility, and representation theory

Let 𝑅 be a Dedekind domain with field of fractions 𝐹 = Frac 𝑅. Finitely generated,
projective 𝑅-modules have played an important role throughout this text, and we now
seek to understand them in the context of orders.
To this end, let 𝐵 be a finite-dimensional 𝐹-algebra and let O ⊆ 𝐵 be an 𝑅-order.
A left O-lattice 𝑀 is an 𝑅-lattice that is a left O-module, i.e., 𝑀 is a finitely generated,
projective (locally free) 𝑅-module that has the structure of a left O-module. We make
a similar definition on the right.
We say that a left (or right) O-lattice 𝑀 is projective if it is a direct summand of
a free left (or right) O-module. Projectivity for lattices in 𝐵 is related to invertibility
as follows (Theorem 20.3.3).
Theorem 20.1.1. Let 𝐼 ⊆ 𝐵 be an 𝑅-lattice. Then 𝐼 is invertible if and only if 𝐼 is

projective as a left OL (𝐼)-module and as a right OR (𝐼)-module.
One can also tease apart left and right invertibility if desired; in the quaternion
context, these are equivalent anyway because of the standard involution (Main Theorem
20.3.9).
Given our efforts to understand invertible lattices, one may think that Theorem
20.1.1 is all there is to say. However, two issues remain. First, there may be finitely
generated (projective) O-modules that are not lattices, and they play a structurally
important role for the order O. Second, and this point is subtle: there may be lattices
𝐼 ⊆ 𝐵 that are projective as a left O-module, but with OL (𝐼) ) O; in other words, such
lattices are invertible over a larger order, even though they still have good properties
as modules over the smaller order.
315
316 CHAPTER 20. INTEGRAL REPRESENTATION THEORY
Example 20.1.2. Let

Z Z
O := ⊆ 𝐵 := M2 (Q)
𝑝Z Z
be the order consisting of integral matrices that are upper triangular modulo a prime
𝑝. We will exhibit both of the issues above. First, we consider O as a left O-module:
it decomposes as

1 0 0 0 Z 0 0 Z
O=O ⊕O = ⊕
0 0 0 1 𝑝Z 0 0 Z
(20.1.3)
Z Z
' ⊕ =: 𝐼1 ⊕ 𝐼2 .
( 𝑝) Z
The two left O-modules 𝐼1 , 𝐼2 are visibly projective, and they are not isomorphic:
intuitively, an isomorphism would have to be multiplication on the left by a 2×2-matrix
that commutes with multiplication O, and so it must be scalar. More precisely, suppose
𝜙 ∈ HomO (𝐼1 , 𝐼2 ) is an isomorphism of left O-modules. Extending scalars, we have

Q
Q𝐼1 = Q𝐼2 = =: 𝑉,
Q
and the extension of 𝜙 gives an element in Aut 𝐵 (𝑉) where 𝐵 = M2 (Q) = EndQ (𝑉),
so commutes with the action of 𝐵 and is therefore central: which is to say 𝜙 is a scalar
matrix, and that is absurd.
The lattice 𝐼 = M2 (Z) is invertible as lattice, since it is an order (!); and it is a
two-sided fractional O-ideal, but it is not sated. We claim that 𝐼 is also a projective
O-module: this follows from the fact that M2 (Z) ' 𝐼2⊕2 as a left O-module, so M2 (Z)
is isomorphic to a direct summand of O ⊕2 .
In this chapter, we establish some basic vocabulary of modules in the language of
the representation theory of an order. In the case of algebras over a field, we defined
a Jacobson radical as a way to measure the failure of the algebra to be semisimple.
Similarly, for every ring 𝐴, we define the Jacobson radical rad 𝐴 to be the intersection
of all maximal left ideals of 𝐴: it again measures the failure of left indecomposable
modules to be simple. There is a left-right symmetry to rad 𝐴, and in fact rad 𝐴 ⊆ 𝐴
is a two-sided 𝐴-ideal.
Locally, the Jacobson radical plays a key role. Suppose 𝑅 is a complete DVR with
unique maximal ideal 𝔭. Then 𝔭 = rad O since it is the maximal ideal. Moreover, we
will see that 𝔭O ⊆ rad O, so O/rad O is a finite-dimensional semisimple 𝑘-algebra.
Much of the structure of O-modules is reflected in the structure of modules over the
quotient O/rad O (see Lemma 20.6.8).
Remark 20.1.4. In representation theory, generally speaking, to study the action of a
group on some kind of object (vector space, simplicial complex, etc.) one introduces
some kind of group ring and studies modules over this ring. The major task becomes
to classify such modules. For example, let 𝑅 be a Dedekind domain with 𝐹 = Frac 𝑅,
and let O be an 𝑅-order in a finite-dimensional 𝐹-algebra 𝐵. A finitely generated
integral representation of O is a finitely generated O-module that is projective as
20.2. PROJECTIVE MODULES 317
an 𝑅-module (in particular, is 𝑅-torsion free). The integral representations of O are

quite complicated! Nevertheless, integral representation theory is a beautiful blend of
number theory, commutative algebra, and linear algebra. In section 20.6, we will see
some of the basic ingredients when 𝑅 is a DVR, and in section 21.4 in the next chapter
we will show that hereditary orders have a tidy integral representation theory. For
more on the subject, see the surveys by Reiner [Rei70, Rei76] as well as the massive
treatises by Curtis–Reiner [CR81, CR87].
20.2 Projective modules

As we will need the notion over several different rings, we start more generally: let 𝐴
be a ring (not necessarily commutative, but with 1). For an introduction to the theory
of projective modules and related subjects, see Lam [Lam99, §2] and Curtis–Reiner
[CR81, §2], and Berrick–Keating [BK2000, §2].
Definition 20.2.1. Let 𝑃 be a finitely generated left 𝐴-module. Then 𝑃 is projective

as a left 𝐴-module if it is a direct summand of a free left 𝐴-module.
A finitely generated free module is projective. The notion of projectivity is quite

fundamental, as the following proposition indicates.
Proposition 20.2.2. Let 𝑃 be a finitely generated left 𝐴-module. Then the following
are equivalent:
(i) 𝑃 is projective;
(ii) There exists a finitely generated left 𝐴-module 𝑄 such that 𝑃 ⊕ 𝑄 is free as a left
𝐴-module.
(iii) Every surjective homomorphism 𝑓 : 𝑀 → 𝑃 (of left 𝐴-modules) has a splitting
𝑔 : 𝑃 → 𝑀 (i.e., 𝑓 ◦ 𝑔 = id 𝑃 );
(iv) Every diagram
𝑃
𝑞
𝑝
~
𝑀
𝑓
/𝑁 /0
of left 𝐴-modules with exact bottom row can be extended as indicated, with
𝑝 = 𝑓 ◦ 𝑞; and
(v) Hom 𝐴 (𝑃, −) is a (right) exact functor.
Proof. See Lam [Lam99, Chapter 2]. In statement (v), given a short exact sequence
0→𝑄→𝑀 →𝑁 →0
then Hom 𝐴 (𝑃, −) is always left exact, so
0 → Hom 𝐴 (𝑃, 𝑄) → Hom 𝐴 (𝑃, 𝑀) → Hom 𝐴 (𝑃, 𝑁) (20.2.3)

is exact; the condition for 𝑃 to be projective is that Hom 𝐴 (𝑃, −) is right exact, so the
full sequence
0 → Hom 𝐴 (𝑃, 𝑄) → Hom 𝐴 (𝑃, 𝑀) → Hom 𝐴 (𝑃, 𝑁) → 0 (20.2.4)
is short exact.
É
20.2.5. A finite direct sum 𝑃 = 𝑖 𝑃𝑖 of finitely generated 𝐴-modules is projective
if and only if each summand
Î 𝑃 𝑖 is projective: indeed, the functor Hom𝑅 (𝑃, −) is
naturally isomorphic to 𝑖 Hom𝑅 (𝑃𝑖 , −), so we apply condition (v) of Proposition
20.2.2.
20.2.6. Localizing Proposition 20.2.2(v), and using the fact that a sequence is exact
if and only if it is exact locally (Exercise 20.1(a)), we see that 𝑃 is projective as a left
O-module if and only if 𝑃 (𝔭) is projective as a left O (𝔭) -module for all primes 𝔭 ⊆ 𝑅
Definition 20.2.7. A left O-lattice is an 𝑅-lattice 𝑀 that is a left O-module.
We make a similar definition on the right.
20.2.8. A left O-lattice 𝑀 is locally free of rank 𝑟 ≥ 1 if 𝑀𝔭 ' O𝔭⊕𝑟 as left O-modules
for all primes 𝔭 ⊆ 𝑅. If follows from 20.2.5 and 20.2.6 that a locally free O-lattice is
projective.
20.3 Projective modules and invertible lattices

Now let 𝑅 be a noetherian domain with 𝐹 := Frac 𝑅, let 𝐵 be a finite-dimensional
𝐹-algebra, and let O ⊆ 𝐵 be an 𝑅-order.
One can extend the base ring of the module while preserving projectivity, as
follows.
Lemma 20.3.1. Let O ⊆ O0 be 𝑅-orders in 𝐵 and let 𝑀 be a left O0-lattice. If 𝑀 is
projective as a left O-module, then 𝑀 is projective as a left O0-module.
Proof. Suppose 𝑀 is projective as a left O-module; then 𝑀 ⊕ 𝑁 ' O𝑟 for some 𝑟 ≥ 0.

Tensor with O0 to get
(O0 ⊗O 𝑀) ⊕ (O0 ⊗O 𝑁) ' (O0) 𝑟 . (20.3.2)

∼ O0 𝑀 = 𝑀,
Since multiplication gives an isomorphism of left O0-modules O0 ⊗O 𝑀 −→
the result follows. (More generally, see Harada [Har63a, Lemma 1.3].)
In the commutative case, an 𝑅-lattice 𝔞 ⊆ 𝐹 is invertible as an 𝑅-module if and

only if 𝔞 is projective as a (left and right) 𝑅-module. Something is true in this more
general context.
Theorem 20.3.3. Let 𝐼 ⊆ 𝐵 be an 𝑅-lattice.
(a) 𝐼 −1 𝐼 = OR (𝐼) if and only if 𝐼 is projective as a left OL (𝐼)-module, and 𝐼 𝐼 −1 =
OL (𝐼) if and only if 𝐼 is projective as a right OR (𝐼)-module.
20.3. PROJECTIVE MODULES AND INVERTIBLE LATTICES 319
(b) 𝐼 is projective as a left OL (𝐼)-module and a right OR (𝐼)-module if and only if 𝐼

is invertible (as an 𝑅-lattice).
The difference between (a) and (b) in Theorem 20.3.3 is the compatibility of the
two products.
−1
∗ −1 Í ∗ (⇒), suppose 𝐼 𝐼 = OR (𝐼); then
Proof. We begin with (a). To prove the implication
there exist 𝛼𝑖 ∈ 𝐼 and 𝛼𝑖 ∈ 𝐼 such that 𝑖 𝛼𝑖 𝛼𝑖 = 1. We may extend the set 𝛼𝑖
to generate 𝐼 as a left OL (𝐼)-module by taking 𝛼𝑖∗ = 0 if necessary. We define the
surjective map Ê
𝑓: 𝑀= OL (𝐼)𝑒 𝑖 → 𝐼
𝑖 (20.3.4)
𝑒 𝑖 ↦→ 𝛼𝑖 .
Consider the map
𝑔:𝐼→𝑀
∑︁
𝛽 ↦→ 𝛽𝛼𝑖∗ 𝑒 𝑖 ;
𝑖
the map 𝑔 is defined because for all 𝛽 ∈ 𝐼, we have 𝛽𝛼𝑖∗ ∈ 𝐼 𝐼 −1 , and as always 𝐼 𝐼 −1 𝐼 ⊆ 𝐼
so 𝐼 𝐼 −1 ⊆ OL (𝐼). The map 𝑔 is a splitting of 𝑓 since
∑︁ ∑︁
( 𝑓 ◦ 𝑔) (𝛽) = 𝛽𝛼𝑖∗ 𝛼𝑖 = 𝛽 𝛼𝑖∗ 𝛼𝑖 = 𝛽.
𝑖 𝑖
Therefore 𝐼 is a direct summand of 𝑀, so 𝐼 is projective as a left OL (𝐼)-module.

Next we prove (⇐). There exists a nonzero 𝑟 ∈ 𝐼 ∩ 𝑅 (Exercise 9.2), so to show
that 𝐼 is left invertible, we may replace 𝐼 with 𝑟 −1 𝐼 and therefore suppose that 1 ∈ 𝐼.
É let 𝛼𝑖 generate 𝐼 as a left OL (𝐼)-module, and consider
Following similar lines as above,
the surjective map 𝑓 : 𝑀 = 𝑖 OL (𝐼)𝑒 𝑖 → 𝐼 by 𝑒 𝑖 ↦→ 𝛼𝑖 . Then since 𝐼 is projective
as a left OL (𝐼)-module, this map splits by a map 𝑔 : 𝐼 → 𝑀; suppose that 𝑔(1) = (𝛼𝑖∗ )𝑖
with 𝛼𝑖∗ ∈ OL (𝐼); then ∑︁
( 𝑓 ◦ 𝑔) (1) = 1 = 𝛼𝑖∗ 𝛼𝑖 . (20.3.5)
𝑖
For all 𝛽 ∈ 𝐼, we have 𝑔(𝛽) = (𝛽𝛼𝑖∗ )𝑖

∈ 𝑀, so 𝛽𝛼𝑖∗ ∈ OL (𝐼) for all 𝑖; therefore for
all 𝛼, 𝛽 ∈ 𝐼 we have 𝛽𝛼𝑖 𝛼 ∈ OL (𝐼)𝐼 ⊆ 𝐼, whence 𝛼𝑖∗ ∈ 𝐼 −1 by definition. Thus from
∗
(20.3.5) we have 1 ∈ 𝐼 −1 𝐼, whence
OR (𝐼) ⊆ 𝐼 −1 𝐼OR (𝐼) = 𝐼 −1 𝐼 ⊆ OR (𝐼)
and thus equality holds.

For part (b), the implication (⇐) follows from (a), and the implication (⇒) for
compatibility follows from Proposition 16.5.8.
Remark 20.3.6. The proof of Theorem 20.3.3 follows what is sometimes called the
dual basis lemma for a projective module: see Lam [Lam99, (2.9)], Curtis–Reiner
[CR81, (3.46)], or Faddeev [Fad65, Proposition 18.2].
20.3.7. Let O, O0 ⊆ 𝐵 be 𝑅-orders. A O, O0-bimodule over 𝑅 is an abelian group 𝑀

with a left O-module and a right O0-module structure with the same action by 𝑅 on
the left and right (i.e., acting centrally, so 𝑟𝑚 = 𝑚𝑟 for all 𝑟 ∈ 𝑅 and 𝑚 ∈ 𝑀). The
𝑅-lattice 𝐼 ⊆ 𝐵 is an OL (𝐼), OR (𝐼)-bimodule over 𝑅.
When the equivalent conditions of Theorem 20.3.3(b) hold, we say that 𝐼 is pro-
jective as a OL (𝐼), OR (𝐼)-bimodule over 𝑅.
Remark 20.3.8. In Theorem 20.3.3, we only considered an 𝑅-lattice 𝐼 as a module

over its left and right orders (i.e., we considered only sated fractional O, O0-ideals),
for the reasons explained in 16.5.18.
Although invertible requires working in this way, it is possible for an 𝑅-lattice 𝐼 to
be projective as a left O-module but still O ( OL (𝐼): for example, if O is a hereditary
order (see Chapter 21) contained properly in a maximal order O ( O0, then O0 is
projective as a left O-module.
Although this may seem a bit complicated, it is refreshing that for quaternion
algebras, all of the sided notions coincide. We recall the equivalences in Main Theorem
16.7.7, building upon them.
Main Theorem 20.3.9. Suppose 𝑅 is a Dedekind domain and 𝐵 is a quaternion

algebra over 𝐹 = Frac 𝑅, and let 𝐼 ⊂ 𝐵 be an 𝑅-lattice. Then the following are
equivalent:
(iii) 𝐼 is left invertible;
(iii0) 𝐼 is right invertible;
(v) 𝐼 is projective as a left OL (𝐼)-module; and
(v0) 𝐼 is projective as a right OR (𝐼)-module.
Proof. The equivalences (ii) ⇔ (ii0) ⇔ (iii) are from Main Theorem 16.7.7 (proven in
Lemma 16.7.5). Theorem 20.3.3(a) gives (v) ⇒ (iii) and (v0) ⇒ (iii0), and Theorem
20.3.3(b) gives (ii) ⇒ (v), (v0).
Example 20.3.10. Consider again Example 16.5.12. The lattice 𝐼 has OL (𝐼) =
OR (𝐼) = O (so has the structure of a sated O, O-bimodule) but 𝐼 is not invertible; from
Main Theorem 20.3.9, it follows that 𝐼 is not projective as a left or right O-module.
20.4 Jacobson radical

Before proceeding further in our analysis of orders, we pause to extend some notions
in sections 7.2 and 7.4 from algebras to rings. We follow Reiner [Rei2003, §6a]; see
also Curtis–Reiner [CR81, §5].
Throughout, let 𝐴 be a ring (not necessarily commutative, but with 1).
Definition 20.4.1. Let 𝑀 be a left 𝐴-module. We say 𝑀 is irreducible or simple

if 𝑀 ≠ {0} and 𝑀 contains no 𝐴-submodules except {0} and 𝑀. We say 𝑀 is
indecomposable if whenever 𝑀 = 𝑀1 ⊕ 𝑀2 with 𝑀1 , 𝑀2 left 𝐴-modules, then either
𝑀1 = {0} or 𝑀2 = {0}.
20.4. JACOBSON RADICAL 321
20.4.2. We generalize Lemma 7.2.7. If 𝐼 is a maximal left ideal of 𝐴, then 𝐴/𝐼 is a

simple 𝐴-module. Conversely, if 𝑀 is a simple 𝐴-module, then for any 𝑥 ∈ 𝑀 nonzero
we have 𝐴𝑥 = 𝑀; therefore 𝑀 ' 𝐴/𝐼 where
𝐼 = ann(𝑥) := {𝛼 ∈ 𝐴 : 𝛼𝑥 = 0}.
Definition 20.4.3. The Jacobson radical rad 𝐴 is the intersection of all maximal left
ideals of 𝐴. The ring 𝐴 is Jacobson semisimple if rad 𝐴 = {0}.
Lemma 20.4.4. The Jacobson radical rad 𝐴 is the intersection of all annihilators of
simple left 𝐴-modules; rad 𝐴 ⊆ 𝐴 is a two-sided 𝐴-ideal.
Proof. The same proof as in Lemma 7.4.5 and Corollary 7.4.6 applies, mutatis mu-
tandis.
Example 20.4.5. If 𝐴 is a commutative local ring, then rad 𝐴 is the unique maximal
ideal of 𝐴.
Example 20.4.6. Let 𝑅 be a complete DVR with maximal ideal 𝔭 = rad 𝑅. Let
𝐹 = Frac 𝑅 and let 𝐷 be a division algebra over 𝐹. Let O ⊆ 𝐷 be the valuation ring,
the unique maximal 𝑅-order (Proposition 13.3.4). Then O has a unique two-sided
ideal 𝑃 by 13.3.10, and so rad O = 𝑃.
Lemma 20.4.7. 𝐴/rad 𝐴 is Jacobson semisimple.
Proof. Let 𝐽 = rad 𝐴. Since 𝐽 𝑀 = {0} for each simple left 𝐴-module 𝑀, we may view
each such 𝑀 as a simple left 𝐴/𝐽-module. Now let 𝛼 ∈ 𝐴 be such that 𝛼+𝐽 ∈ rad( 𝐴/𝐽);
then (𝛼 + 𝐽) 𝑀 = {0}, so 𝛼𝑀 = {0} and 𝛼 ∈ 𝐽; thus rad( 𝐴/𝐽) = {0}, and 𝐴/𝐽 is
Jacobson semisimple.
rad 𝐴 = {𝛽 ∈ 𝐴 : 1 − 𝛼1 𝛽𝛼2 ∈ 𝐴× for all 𝛼1 , 𝛼2 ∈ 𝐴}.
Corollary 20.4.9. rad 𝐴 is the intersection of all maximal right ideals of 𝐴.
Proof. Lemma 20.4.8 gives a left-right symmetric characterization of rad 𝐴.
Corollary 20.4.10. If 𝜙 : 𝐴 → 𝐴 0 is a surjective ring homomorphism, then 𝜙(rad 𝐴) ⊆

rad 𝐴 0 and we have an induced surjective homomorphism 𝐴/rad 𝐴 → 𝐴 0/rad 𝐴 0.
Proof. Let 𝛽 ∈ rad 𝐴, let 𝛼10 , 𝛼20 ∈ 𝐴 0; since 𝜙 is surjective, there exist preimages
𝛼1 , 𝛼2 ∈ 𝐴. By Lemma 20.4.8, 1 − 𝛼1 𝛽𝛼2 ∈ 𝐴× and
𝜙(1 − 𝛼1 𝛽1 𝛼2 ) = 1 − 𝛼10 𝜙(𝛽)𝛼20 ∈ 𝐴 0× ,
so by the same lemma, 𝜙(𝛽) ∈ rad 𝐴 0.
Corollary 20.4.11. Let 𝐼 ⊆ 𝐴 be a two-sided 𝐴-ideal.

(a) If 𝐴/𝐼 is Jacobson semisimple, then rad 𝐴 ⊆ 𝐼.

(b) If 𝐼 ⊆ rad 𝐴, then (rad 𝐴)/𝐼 = rad( 𝐴/𝐼).
Proof. We have a surjection 𝜙 : 𝐴 → 𝐴/𝐼. For (a), we get 𝜙(rad 𝐴) ⊆ rad( 𝐴/𝐼) = {0}
from Corollary 20.4.10, so rad 𝐴 ⊆ 𝐼. For (b), we get rad( 𝐴)/𝐼 ⊆ rad( 𝐴/𝐼) from the
surjection, and applying (a) to ( 𝐴/𝐼)/(rad( 𝐴)/𝐼) we get rad( 𝐴/𝐼) ⊆ rad( 𝐴)/𝐼.
Lemma 20.4.12 (Nakayama’s lemma). Let 𝑀 be a finitely generated left 𝐴-module

such that (rad 𝐴) 𝑀 = 𝑀. Then 𝑀 = {0}.
Proof. If 𝑀 ≠ {0}, let 𝑥1 , . . . , 𝑥 𝑛 be a minimal set of generators for 𝑀 as a left

𝐴-module. Since 𝑥1 ∈ 𝑀 = (rad 𝐴)𝑀, we may write
𝑥 1 = 𝛽1 𝑥 1 + · · · + 𝛽 𝑛 𝑥 𝑛
with 𝛽𝑖 ∈ rad 𝐴. But then 1− 𝛽1 ∈ 𝐴× , so the generator 𝑥1 is redundant, a contradiction.

Corollary 20.4.13. Let 𝑀 be a finitely generated left 𝐴-module, and let 𝑁 ⊆ 𝑀 be a

submodule such that 𝑁 + (rad 𝐴) 𝑀 = 𝑀. Then 𝑁 = 𝑀.
Proof. By hypothesis, 𝑀/𝑁 is finitely generated, and (rad 𝐴) (𝑀/𝑁) = 𝑀/𝑁, so by

Nakayama’s lemma, 𝑀/𝑁 = {0} and 𝑀 = 𝑁.
Lemma 20.4.14. Let 𝐼 be a maximal two-sided ideal of 𝐴. Then 𝐼 contains rad 𝐴.
Proof. If 𝐼 does not contain rad 𝐴, then 𝐼 + rad 𝐴 is a two-sided ideal of 𝐴 containing
rad 𝐴 and properly containing 𝐼. Since 𝐼 is maximal, we have 𝐼 + rad 𝐴 = 𝐴. By (the
corollary to) Nakayama’s lemma, we get 𝐼 = 𝐴, a contradiction.
20.5 Local Jacobson radical

Suppose now that 𝑅 is a complete DVR with fraction field 𝐹 = Frac 𝑅, maximal ideal
𝔭 = rad 𝑅, and residue field 𝑘 = 𝑅/𝔭. Let 𝐵 be a finite-dimensional 𝐹-algebra, and let
O ⊆ 𝐵 be an 𝑅-order.
In this setting, we may identify the Jacobson radical via pullback as follows.
Theorem 20.5.1. Let 𝜙 : O → O/𝔭O be reduction modulo 𝔭. Then
rad O = 𝜙−1 (rad O/𝔭O) ⊇ 𝔭O,
and (rad O) 𝑟 ⊆ 𝔭O for some 𝑟 > 0.
Proof. See Reiner [Rei2003, Theorem 6.15].
Corollary 20.5.2. O/rad O is a (finite-dimensional) semisimple 𝑘-algebra.
Proof. Since rad O ⊇ 𝔭O, we conclude that O/𝔭O is a 𝑘-algebra; it is Jacobson

semisimple by 20.4.7 and hence semisimple by Lemma 7.4.2.
20.6. INTEGRAL REPRESENTATION THEORY 323
Definition 20.5.3. A two-sided ideal 𝐽 ⊆ O is topologically nilpotent if 𝐽 𝑟 ⊆ 𝔭O for

some 𝑟 > 0.
Remark 20.5.4. The order O as a free 𝑅-module has a natural topology induced from
the 𝔭-adic topology on 𝑅; 𝐽 is topologically nilpotent if and only if 𝐽 𝑟 → {0} in this
topology.
Corollary 20.5.5. Let 𝐼 ⊆ O be a two-sided ideal. Then the following are equivalent:
(a) 𝐼 ⊆ rad O;
(b) 𝐼 𝑟 ⊆ rad O for some 𝑟 > 0; and
(c) 𝐼 is topologically nilpotent.
Proof. See Reiner [Rei2003, Exercise 39.1, Exercise 6.3].
20.6 Local integral representation theory

We continue our notation that 𝑅 is a complete DVR. We now turn to some notions in
integral representation theory. In this local case, there is a tight connection between
the representation theory of O (viewed in terms of O-modules) and the representation
theory of the quotient O/𝔭O which is a 𝑘-algebra of finite dimension over 𝑘, since O
is finitely generated as an 𝑅-module.
20.6.1. Recall that a representation of 𝐵 over 𝐹 is the same as a left 𝐵-module. If 𝑀 is

a finitely-generated left O-module, then 𝑉 := 𝑀 ⊗𝑅 𝐹 is a left 𝐵-module, and 𝑀 ⊆ 𝑉
is an 𝑅-lattice. A O-supermodule of 𝑀 is a left O-module 𝑉 ⊇ 𝑀 0 ⊇ 𝑀.
The following result is foundational.
Theorem 20.6.2 (Krull–Schmidt). Every finitely generated left O-module 𝑀 is ex-

pressible as a finite direct sum of indecomposable modules, uniquely determined by
𝑀 up to O-module isomorphism and reordering.
Proof. Since 𝑀 is finitely generated over 𝑅 it is itself noetherian, so the process of

decomposing 𝑀 into direct summands terminates. See Curtis–Reiner [CR81, (6.12)]
or Reiner [Rei2003, §6, Exercise 6] for hints that lead to a proof of the second
(uniqueness) part.
Corollary 20.6.3. Let 𝑀 = 𝑀1 ⊕ · · · ⊕ 𝑀𝑟 be a decomposition into finitely generated

indecomposable left O-modules, and let 𝑁 ⊆ 𝑀 be a direct summand. Then 𝑁 '
𝑀𝑖1 ⊕ · · · ⊕ 𝑀𝑖𝑠 for some subset {𝑖 1 , . . . , 𝑖 𝑠 } ⊆ {1, . . . , 𝑛}.
0 0
É𝑛
Proof. By hypothesis, we can write 𝑖=1 𝑀𝑖 = 𝑁 ⊕ 𝑁 , with 𝑁 a finitely generated
left O-module. By the Krull–Schmidt theorem (Theorem 20.6.2), if we write 𝑁, 𝑁 0 as
the direct sums of indecomposable modules, the conclusion follows.
20.6.4. We saw in 7.2.19 that idempotents govern the decomposition of the 𝐹-algebra 𝐵
into indecomposable left 𝐵-modules. The same argument shows that a decomposition
O = 𝑃 1 ⊕ · · · ⊕ 𝑃𝑟 (20.6.5)
into a direct sum of indecomposable left O-modules corresponds to an idempotent
decomposition 1 = 𝑒 1 + · · · + 𝑒𝑟 , with the 𝑒 𝑖 a complete set of primitive orthogonal
idempotents. Moreover, each 𝑃𝑖 = O𝑒 𝑖 is a projective indecomposable left O-module.
Conversely, if 𝑃 is a projective indecomposable finitely generated left O-module,
then 𝑃 ' 𝑃𝑖 for some 𝑖: taking a set of generators we have a surjective O-module
homomorphism O𝑟 → 𝑃, and since 𝑃 is projective we have 𝑃 ⊆ O𝑟 a direct summand,
so Corollary 20.6.3 applies.
Consequently, if 𝑃 is a projective left O-lattice, then 𝑃 ' 𝑃1⊕𝑛1 ⊕ · · · ⊕ 𝑃𝑟⊕𝑛𝑟 with
𝑛𝑖 ≥ 0 for 𝑖 = 1, . . . , 𝑟.
20.6.6. The decomposition of an order into projective indecomposables is a nice way
to keep track of other orders, as follows. We extend our notation slightly, and define
OL (𝑀) := {𝛼 ∈ 𝐵 : 𝛼𝑀 ⊆ 𝑀 }
for every left O-submodule 𝑀 ⊆ 𝐵.

Take a decomposition of O in (20.6.5); since each 𝑃𝑖 is a left O-module, extending
scalars it is a left 𝐵-module, so
𝑟
Ù
OL (𝑃𝑖 ) = O. (20.6.7)
𝑖=1
Now let 𝐼 ⊆ 𝐵 be an 𝑅-lattice with O ⊂ OL (𝐼) that is projective as an O-module.

By 20.6.4, considering 𝐼 as a left O-module, we have an isomorphism of left O-modules
∼ 𝑃 ⊕𝑛1 ⊕ · · · ⊕ 𝑃 𝑛𝑟
𝜙 : 𝐼 −→ 1 𝑟
with 𝑛𝑖 ≥ 0. We claim that

Ù
OL (𝐼) = OL (𝑃𝑖 ).
𝑖
𝑛𝑖 >0
Indeed, we have 𝛼𝐼 ⊆ 𝐼 if and only if 𝜙(𝛼𝐼) = 𝛼𝜙(𝐼) ⊆ 𝜙(𝐼), since 𝜙 is a O-module

homomorphism so extends to a 𝐵-algebra homomorphism, and finally 𝛼𝜙(𝐼) ⊆ 𝜙(𝐼)
if and only if 𝛼𝑃𝑖 ⊆ 𝑃𝑖 for all 𝑖 with 𝑛𝑖 > 0, as in (20.6.7).
We now relate a decomposition of O into a decomposition of O/𝔭O.
Lemma 20.6.8. Let 𝐽 = rad O. The association 𝐼 ↦→ 𝐼/𝐽 𝐼 gives a bĳection between
isomorphism classes of indecomposable finitely generated projective left O-modules
and isomorphism classes of simple finite-dimensional left O/𝐽-modules.
Proof. The proof requires a bit of fiddling with idempotents, but is otherwise straight-
forward—so it makes a good exercise (Exercise 20.7).
20.7. ∗ STABLE CLASS GROUP AND CANCELLATION 325
Corollary 20.6.9. If 𝐼 is projective indecomposable, then 𝐽 𝐼 ⊆ 𝐼 is the unique maximal

O-submodule of 𝐼.
Proof. By Lemma 20.6.8, since 𝐼 is indecomposable, 𝐼/𝐽 𝐼 is simple so 𝐽 𝐼 is a unique

maximal submodule. If 𝐼 0 ⊆ 𝐼 is another maximal O-submodule, then 𝐽 𝐼 + 𝐼 0 = 𝐼, and
by Nakayama’s lemma 𝐼 0 = 𝐼, a contradiction.
We finish our local study over 𝑅 a complete DVR with composition series for
modules over an order.
Definition 20.6.10. Let 𝑀 be an O-lattice. A composition series for 𝑀 is a strictly

decreasing sequence
𝑀 = 𝑀0 ⊇ 𝑀1 ⊇ 𝑀2 ⊇ . . .
Ñ∞
such that 𝑖=1 𝑀𝑖 = {0} and each composition factor 𝑀𝑖 /𝑀𝑖+1 is simple as a O-
module.
The length of a composition series is the largest integer 𝑟 such that 𝑀𝑟 = {0} if 𝑟
exists (in which case we call the series finite), and otherwise the length is ∞.
20.6.11. If 𝑀 has a finite composition series, then its length ℓ(𝑀) is well-defined,
independent of the series. For example, taking 𝑅 = 𝐹 and O = 𝐵, a finitely generated
𝐵-module is a finite-dimensional 𝐹-vector space, so every composition series is finite
and every 𝐵-module 𝑉 has a well-defined length ℓ(𝑉).
20.6.12. Let 𝑁 ⊆ 𝑀 be a maximal O-submodule. We claim that 𝐽 𝑀 ⊆ 𝑁. Otherwise,

𝑁 + 𝐽 𝑀 = 𝑀 by maximality, so by Nakayama’s lemma (Corollary 20.4.13), 𝑁 = 𝑀,
a contradiction.
20.7 ∗ Stable class group and cancellation

To conclude this chapter, we apply the above results and consider a different way to
form of a group of ideal classes; for further reference on the topics of this section, see
Curtis–Reiner [CR87, §§49–51] or Reiner [Rei2003, §38].
Let 𝑅 be a Dedekind domain with field of fractions 𝐹.
20.7.1. Recall that the group Cl 𝑅 records classes of fractional ideals, or what is
more relevant here, isomorphism classes of projective modules of rank 1. Here is
another way to see the group law on Cl 𝑅: given two such fractional ideals 𝔞, 𝔟 up to
isomorphism, there is an isomorphism of 𝑅-modules
𝔞 ⊕ 𝔟 ' 𝑅 ⊕ 𝔞𝔟,
and the class of 𝔞𝔟 is uniquely determined by this isomorphism by 9.3.10.
We now consider an analogous construction to 20.7.1 in the noncommutative

setting. Let 𝐵 be a simple 𝐹-algebra and O ⊆ 𝐵 an 𝑅-order. We begin with a technical
lemma.
Lemma 20.7.2 (Weak approximation). Let 𝐼 be a locally principal left fractional

O-ideal and let 𝔞 ⊆ 𝑅 be an ideal. Then there exists 𝛽 ∈ 𝐵× such that 𝐼 𝛽 ⊆ O and
(𝐼 𝛽)𝔭 = O𝔭 for all 𝔭 | 𝔞. (20.7.3)
Proof. For each prime 𝔭, we have 𝐼𝔭 = O𝔭 𝛼𝔭 with 𝛼𝔭 ∈ 𝐵𝔭× . Because 𝐹 is dense in

𝐹𝔭 , there exists 𝛽 ∈ (O : 𝐼) R such that 𝛼𝔭 𝛽𝔭 ≡ 1 (mod 𝔭O𝔭 ) for all 𝔭 | 𝔞. By norms,
we have 𝛽 ∈ 𝐵× . Letting 𝜇𝔭 := 𝛼𝔭 𝛽𝔭 , we have 𝜇𝔭 − 1 ∈ 𝔭O𝔭 ⊇ rad O𝔭 by Theorem
20.5.1 so 𝜇𝔭 ∈ O×𝔭 by Lemma 20.4.8. Therefore (𝐼 𝛽)𝔭 = O𝔭 for all 𝔭 | 𝔞.
Proposition 20.7.4. If 𝐼, 𝐼 0 ⊆ 𝐵 are locally principal left fractional O-ideals, then

there exists a locally principal left fractional O-ideal 𝐽 and an isomorphism
𝐼 ⊕ 𝐼0 ' 𝐽 ⊕ O (20.7.5)
of left O-modules.
Proof. We may suppose without loss of generality that 𝐼, 𝐼 0 ⊆ O. Then we have exact
sequences of left O-modules
𝜙
→ O → O/𝐼 → 0
0→𝐼−
𝜙0
0 → 𝐼 0 −−→ O → O/𝐼 0 → 0
The module O/𝐼 is 𝑅-torsion, annihilated by the (nonzero) 𝑅-ideal 𝔞 = [O : 𝐼] 𝑅 , and

similarly with 𝐼 0, annihilated by 𝔞0 = [O : 𝐼 0] 𝑅 . By weak approximation (Lemma
20.7.2), replacing 𝐼 0 with 𝐼 0 𝛽 we may suppose that 𝐼𝔭0 = O𝔭 for all 𝔭 | 𝔞, and hence
𝔞, 𝔞0 are coprime. Then for all primes 𝔭 of 𝑅, we have either (O/𝐼)𝔭 = {0} so 𝜙𝔭 is
surjective, or correspondingly 𝜙𝔭0 is surjective.
Now consider the left O-module homomorphism
𝜙 + 𝜙0 : 𝐼 ⊕ 𝐼 0 → O (20.7.6)
obtained by summing the two natural inclusions. We just showed that (𝜙 + 𝜙 0)𝔭
is surjective for all primes 𝔭, so it follows that 𝜙 + 𝜙 0 is surjective: the cokernel
𝑀 := coker(𝜙 + 𝜙 0) has 𝑀𝔭 = {0} for all 𝔭, and 𝑀 = {0}. Moreover, since O is
projective as a left O-module, the map 𝜙 + 𝜙 0 splits (or note that the map splits locally
for every prime 𝔭, so it splits globally, Exercise 20.1(b)). If we let 𝐽 := ker(𝜙 + 𝜙 0),
we then obtain an isomorphism
𝐼 ⊕ 𝐼 0 ' 𝐽 ⊕ O. (20.7.7)
To conclude, we show that 𝐽 is locally principal. To this end, we localize at a prime

𝔭 and note that 𝐼, 𝐼 0 are locally principal, so
𝐼𝔭 ⊕ 𝐼𝔭0 ' O𝔭⊕2 ' 𝐽𝔭 ⊕ O𝔭 . (20.7.8)
But by the Krull–Schmidt theorem (Theorem 20.6.2) and Exercise 20.8, we can cancel
one copy of O𝔭 from both sides! We conclude that 𝐽𝔭 ' O𝔭 as left O-modules and
therefore by Lemma 17.3.3 that 𝐽𝔭 is (right) principal.
The candidate binary operation in Proposition 20.7.4 has a simple description in

the “coprime” case.
Lemma 20.7.9. Let 𝐼, 𝐼 0 ⊆ O be locally principal integral left O-ideals, and suppose
for every prime 𝔭 ⊆ 𝑅 either 𝐼𝔭 = O𝔭 or 𝐼𝔭0 = O𝔭 . Then
𝐼 ⊕ 𝐼 0 ' O ⊕ 𝐽, where 𝐽 = 𝐼 ∩ 𝐼 0 .
Moreover, writing 𝐼𝔭 = O𝛼𝔭 and 𝐼𝔭0 = O𝛼𝔭0 , we have
𝐽𝔭 = O𝔭 𝛼𝔭 𝛼𝔭0 = O𝔭 𝛼𝔭0 𝛼𝔭 .
By weak approximation (Lemma 20.7.2), the hypothesis of Lemma 20.7.9 can

always be arranged to hold for 𝐼, 𝐼 0, up to isomorphism (as left O-ideals).
Proof. By hypothesis, if 𝜙, 𝜙 : 𝐼, 𝐼 0 ↩→ O are the inclusions, then the map 𝜙 + 𝜙 0 : 𝐼 ⊕

𝐼 0 → O as in (20.7.6) is surjective. We have
ker(𝜙 + 𝜙 0) = {(𝛼, 𝛼 0) ∈ 𝐼 ⊕ 𝐼 0 : 𝛼 + 𝛼 0 = 0} ' 𝐼 ∩ 𝐼 0
by projection onto either coordinate, since 𝛼 = −𝛼 0 ∈ 𝐼 ∩ 𝐼 0. This gives an exact

sequence
0 → 𝐼 ∩ 𝐼0 → 𝐼 ⊕ 𝐼0 → O → 0
and as above 𝐼 ⊕ 𝐼 0 ' 𝐽 ⊕ O with 𝐽 = 𝐼 ∩ 𝐼 0. The final statement follows from the
hypothesis that either 𝐼𝔭 = O𝔭 or 𝐼𝔭0 = O𝔭 , since then 𝛼𝔭 ∈ O×𝔭 or 𝛼𝔭0 ∈ O×𝔭 .
In order to get a well-defined binary operation, we need an equivalence relation:

we will need to identify 𝐽, 𝐽 0 if 𝐽 ⊕ O ' 𝐽 0 ⊕ O. But the copies of O needed for the
axioms start to pile up, so we make the following more general definition.
Definition 20.7.10. Let 𝐽, 𝐽 0 ⊆ 𝐵 be locally principal left O-ideals. We say that 𝐽 is

stably isomorphic to 𝐽 0 if there exists an isomorphism of left O-modules
𝐽 ⊕ O ⊕𝑟 ' 𝐽 0 ⊕ O ⊕𝑟
for some 𝑟 ≥ 0.
Let [𝐽] St denote the stable isomorphism class of a left O-ideal 𝐽 and let StCl O be
the set of stable isomorphism classes of left O-ideals in 𝐵.
Proposition 20.7.11. StCl O is an abelian group under the binary operation (20.7.5),
written [𝐼] St + [𝐼 0] St = [𝐽] St , with identity [O] St .
Accordingly, we call StCl O the stable class group of O; it is also referred to as

the locally free class group of O.
Proof. The operation is well-defined: if [𝐼1 ] St = [𝐼2 ] St via 𝐼1 ⊕ O ⊕𝑟 ' 𝐼2 ⊕ O ⊕𝑟 and

the same with [𝐼10 ] St = [𝐼20 ] St , and we perform the binary operation 𝐼1 ⊕ 𝐼10 ' 𝐽1 ⊕ O
and the same with the subscripts 2 , then
0 0
𝐽1 ⊕ O ⊕ (𝑟 +𝑟 +1) ' (𝐼1 ⊕ O ⊕𝑟 ) ⊕ (𝐼10 ⊕ O ⊕𝑟 )
0
' (𝐼2 ⊕ O ⊕𝑟 ) ⊕ (𝐼20 ⊕ O ⊕𝑟 ) (20.7.12)
⊕ (𝑟 +𝑟 0 +1)
' 𝐽2 ⊕ O
so [𝐽1 ] St = [𝐽2 ] St . It is similarly straightforward to verify that the operation is

associative and commutative and that [O] St is the identity.
To conclude, we show that StCl O has inverses. Let 𝐼 ⊆ O be a locally principal
O-ideal. For each prime 𝔭 ⊆ 𝑅, we have 𝐼𝔭 = O𝔭 𝛼𝔭 with 𝛼𝔭 ∈ 𝐵𝔭× , and 𝛼𝔭 = 1 for all
but finitely many 𝔭. Let 𝐼 0 be the 𝑅-lattice with 𝐼𝔭0 = O𝔭 𝛼𝔭−1 for all 𝔭. Then 𝐼 0 is a
left fractional O-ideal, because this is true locally. By weak approximation (Lemma
20.7.2), there exists 𝛽 ∈ 𝐵× such that (𝐼 0 𝛽)𝔭 = O𝔭 for all 𝔭 such that 𝐼𝔭 ≠ O𝔭 , i.e.,
for all 𝔭 such that 𝛼𝔭 ≠ 1. But now we can perform the group operation as in Lemma
20.7.9: we have [𝐼] St + [𝐼 0] St = [𝐽] St where 𝐽 = 𝐼 ∩ 𝐼 0 𝛽, and for all 𝔭 we have
𝐽𝔭 = O𝔭 𝛼𝔭 𝛼𝔭−1 𝛽𝔭 = O𝔭 𝛽𝔭
so 𝐽 = O𝛽 and 𝐽 ' O, so [𝐼] St + [𝐼 0] St = [O] St and 𝐼 0 is an inverse.
Remark 20.7.13. There is a related group to StCl O, defined as follows. Let 𝐴 be

a ring, and let P ( 𝐴) be the category of finitely generated projective left 𝐴-modules
under isomorphisms. We define the group 𝐾0 ( 𝐴) to be the free abelian group on the
isomorphism classes [𝑃] of objects 𝑃 ∈ P ( 𝐴) modulo the subgroup of relations
[𝑃 ⊕ 𝑃 0] = [𝑃] + [𝑃 0], for 𝑃, 𝑃 0 ∈ P ( 𝐴);
equivalently relations [𝑃] + [𝑃 0] = [𝑄] for each exact sequence
0 → 𝑃 → 𝑄 → 𝑃0 → 0
since such a sequence splits. The group 𝐾0 ( 𝐴) is sometimes called the projective
class group of 𝐴. (The group 𝐾0 ( 𝐴) is the Grothendieck group of the category P ( 𝐴).)
Then for 𝑃, 𝑄 ∈ P (O), we have [𝑃] = [𝑄] ∈ 𝐾0 (O) if and only if 𝑃, 𝑄 are stably
isomorphic [CR87, Proposition 38.22]. Moreover, there is a natural map
𝐾0 (O) → 𝐾0 (𝐵)
[𝑃] ↦→ [𝐹 ⊗𝑅 O],
and we let 𝑆𝐾0 (O) be its kernel, called the reduced projective class group of O. The
abelian group 𝑆𝐾0 (O) is generated by elements [𝑃] − [𝑄] where 𝑃, 𝑄 ∈ P (O) and
𝐹 ⊗𝑅 𝑃 ' 𝐹 ⊗𝑅 𝑄. Finally, we have an isomorphism [CR87, Theorem 49.32]
∼ 𝑆𝐾 (O)
StCl O −→ 0
(20.7.14)
[𝐼] St ↦→ [𝐼] − [O].
In other words, after all of this work—at least for maximal orders—the reduced
projective class group and the stable class group coincide. (For a more general
order, one instead compares to a maximal superorder via the natural extension maps
StCl O → StCl O0.)
The stable class group was first introduced and studied by Swan [Swa60, Swa62]
in this context in the special case where O = Z[𝐺] is the group ring of a finite group
𝐺.
Definition 20.7.15. We say that O has stable cancellation (or the simplification
property) if stable isomorphism implies isomorphism, i.e., if whenever 𝐼, 𝐼 0 are locally
principal left O-ideals with 𝐼 ⊕ O𝑟 ' 𝐼 0 ⊕ O𝑟 for 𝑟 ≥ 0, then in fact 𝐼 ' 𝐼 0.
If we had defined stable isomorphism and cancellation for locally free O-modules,
we would arrive at the same groups and condition, so stable cancellation is also called
the locally free cancellation.
From now on, suppose that 𝑅 is a global ring with 𝐹 = Frac 𝑅, and O ⊂ 𝐵 is a
maximal 𝑅-order in a quaternion algebra 𝐵 over 𝐹. We recall section 17.8, and the
class group ClΩ 𝑅, where Ω ⊆ Ram 𝐵 is the set of real ramified places.
Theorem 20.7.16 (Fröhlich–Swan). Let 𝑅 = 𝑅 (S) be a global ring, let 𝐵 be a quater-

nion algebra over 𝐹, and let O ⊂ 𝐵 be a maximal 𝑅-order. Then the reduced norm
induces an isomorphism
∼ Cl 𝑅
nrd : StCl O −→ (20.7.17)
Ω
of finite abelian groups.
Proof. See Fröhlich [Frö75, Theorem 2, §X], Swan [Swa80, Theorem 9.4], or Curtis–
Reiner [CR87, Theorem 49.32]; we will sketch a proof of a more general version of
this theorem in section 28.10, when we have idelic methods at our disposal.
20.7.18. Since 𝐵 is a quaternion algebra, the notions of invertible and locally principal
coincide. Then there is a surjective map of sets
ClsL O → StCl O
(20.7.19)
[𝐼] L ↦→ [𝐼] St .
Suppose further that 𝐹 = Frac 𝑅 is a number field and 𝑅 is a global ring. Then ClsL O
is a finite set, by Main Theorem 17.7.1; consequently, the stable class group StCl O is
a finite abelian group. However, the map (20.7.19) of sets need not be injective.
The order O has stable cancellation if and only if the map (20.7.19) is injective
(equivalently, bĳective).
20.7.20. Suppose that 𝐵 satisfies the Eichler condition. Then by Eichler’s theorem
∼ Cl 𝑅 compatible
(Theorem 17.8.3), the reduced norm also gives a bĳection Cls O −→ Ω
with the surjective map ClsL O → StCl O (20.7.19) which must therefore also be a
bĳection.
What remains, then, is the case where 𝐵 is definite. We restrict attention to the
case where the base field 𝐹 is a number field, hence a totally real field, and we work
with 𝑅-orders O ⊆ 𝐵, where 𝑅 = Z𝐹 is the ring of integers of 𝐹. Vignéras [Vig76b]
initiated the classification of definite quaternion orders with stable cancellation, and
showed that there are only finitely many such orders. Hallouin–Maire [HM2006] and
Smertnig [Sme2015] extended this classification to certain classes of orders, and the
complete classification was obtained by Smertnig–Voight [SV2019].
Theorem 20.7.21 (Vignéras, Hallouin–Maire, Smertnig, Smertnig–Voight). Up to

isomorphism, there are exactly 316 definite quaternion 𝑅-orders with stable cancella-
tion.
The isomorphisms in Theorem 20.7.21 are as 𝑅-orders; up to ring isomorphism

(identifying Galois conjugates), there are exactly 247.
Example 20.7.22. If O is a definite maximal quaternion Z-order, by Theorem 20.7.16

we have StCl O = # Cl+ Z = 1, so O has stable cancellation if and only if # Cls O = 1.
These orders will be classified in section 25.4: they are the orders of discriminant
𝐷 = 2, 3, 5, 7, 13. (More generally, if 𝑅 = Z𝐹 has # Cl+ Z𝐹 = 1, then a definite,
maximal quaternion 𝑅-order has stable cancellation if and only if # Cls O = 1.)
Remark 20.7.23. Jacobinski [Jaci68] was the first to consider the stable class group for
general orders in the context of his work on genera of lattices; his cancellation theorem
states more generally that if 𝐵 is a central simple algebra over 𝐹 and 𝐵 is not a totally
definite quaternion algebra, then every 𝑅-order O ⊆ 𝐵 has stable cancellation. This
result was reformulated by Fröhlich [Frö75] in terms of ideles and further developed
by Fröhlich–Reiner–Ullom [FRU74]. Swan [Swa80] related cancellation to strong
approximation in the context of 𝐾-groups.
Brzezinski [Brz83b] also defines the spinor class group of an order, a quotient of
its locally free class group; this group measures certain invariants phrased in terms of
quadratic forms.
Remark 20.7.24. More generally, a ring 𝐴 in which every stably free right 𝐴-module
is free is called a (right) Hermite ring by some authors: for further reference and
comparison of terminology, see Lam [Lam2006, Section I.4]. If O has locally free
cancellation, then O is Hermite; however, the converse does not hold in general—
a counterexample is described in detail by Smertnig [Sme2015]. Smertnig–Voight
[SV2019] show that there are exactly 375 definite quaternion 𝑅-orders with the Hermite
property up to isomorphism.
Exercises
Throughout these exercises, let 𝑅 be a noetherian domain with 𝐹 = Frac 𝑅, let 𝐵 be a
finite-dimensional 𝐹-algebra, let O ⊆ 𝐵 be an 𝑅-order, and let 𝐽 = rad O.
I 1. Let 𝑀, 𝑁 be left O-lattices.

(a) Show that a sequence 0 → 𝑀 → 𝑁 → 𝑀 0 → 0 of left O-lattices is exact

if and only if the sequences 0 → 𝑀𝔭 → 𝑁𝔭 → 𝑀𝔭0 → 0 are exact for
all primes 𝔭 ⊆ 𝑅. [Hint: Consider the modules measuring the failure of
exactness and show they are locally zero, hence zero.]
(b) Let 𝜙 : 𝑀 → 𝑁 be a surjective O-module homomorphism. Show that
𝜙 splits (there exists 𝜓 : 𝑁 → 𝑀 such that 𝜙𝜓 = id 𝑁 ) if and only if
𝜙𝔭 : 𝑀𝔭 → 𝑁𝔭 splits for all primes 𝔭 ⊆ 𝑅.
I 2. Suppose 𝑅 is a DVR and 𝐵 is a quaternion algebra. Let 𝐽 = rad O. Show that
𝐽 = 𝐽 and OL (rad O) = OR (rad O).
3. Let 𝑅 be a complete DVR with 𝔭 = rad 𝑅. Show that the 𝔭-adic topology and
the 𝐽-adic topology on O are the same.

𝑅 𝑅
4. Let 𝑅 be a DVR with maximal ideal 𝔭, and let O = ⊆ 𝐵 = M2 (𝐹). Let
𝔭 𝑅
𝐼 ⊆ 𝐵 be a left fractional O-ideal. Show that either 𝐼 is invertible as a O-ideal
or 𝐼 is conjugate to M2 (𝑅) by an element of 𝐵× .
5. Let O be a maximal 𝑅-order, and let 𝑀 be a projective left O-lattice. Show
that 𝑀 is indecomposable if and only if 𝐹 𝑀 is a simple left 𝐵-module. [Hint:
Suppose 𝑊 ⊆ 𝐹 𝑀 is a left 𝐵-submodule of 𝐹 𝑀, and let 𝑁 := 𝑀 ∩ 𝑊. Show
that 𝑀/𝑁 is a projective O-lattice, so the sequence 0 → 𝑁 → 𝑀 → 𝑀/𝑁 → 0
splits.]
I 6. Let 𝐴 be a ring (not necessarily commutative, but with 1. In this exercise, we
prove Lemma 20.4.8, that
rad 𝐴 = {𝛽 ∈ 𝐴 : 1 − 𝛼1 𝛽𝛼2 ∈ 𝐴× for all 𝛼1 , 𝛼2 ∈ 𝐴}.
We first show the inclusion (⊆).

(a) Since rad 𝐴 is a two-sided ideal, it suffices to show that 1 − 𝛽 ∈ 𝐴× . Show
that 𝐴(1 − 𝛽) = 𝐴.
(b) Let 𝛼 ∈ 𝐴 be such that 𝛼(1 − 𝛽) = 1. Repeating the argument, show that
𝐴(1 − (1 − 𝛼)) = 𝐴𝛼 = 𝐴.
(c) Show that 𝛼 is also a right inverse of 1 − 𝛽, so 1 − 𝛽 ∈ 𝐴× .
Next we show the inclusion (⊇).
(d) Let 𝛽 ∈ 𝐴 be such that 1 − 𝛼𝛽𝛾 ∈ 𝐴× for all 𝛼, 𝛾 ∈ 𝐴. Let 𝑀 be a simple
left 𝐴-module. Show that 𝛼𝑀 = {0}. Conclude that 𝛼 ∈ rad 𝐴.
I 7. Suppose 𝑅 is a complete DVR. Prove Lemma 20.6.8: the association 𝐼 ↦→
𝐼/𝐽 𝐼 gives a bĳection between isomorphism classes of indecomposable finitely
generated projective left O-modules and isomorphism classes of simple finite-
dimensional left O/𝐽-modules.
I 8. Let 𝑅 be a complete DVR, and let 𝐼, 𝐼 0, 𝐽 be finitely generated left O-modules
such that
𝐼 ⊕ 𝐽 ' 𝐼0 ⊕ 𝐽
as left O-modules. Prove that 𝐼 ' 𝐼 0 as left O-modules.
9. Let Λ = M𝑛 (O) with 𝑛 ≥ 1. Show that StCl Λ ' StCl O.

10. Show that the Lipschitz order has stable cancellation.
11. Let 𝐵 = (−1, −3 | Q) and let O = Z + Z(3𝑖) + Z(−1 + 𝑗)/2 + Z(3𝑖 + 𝑖 𝑗)/2.
(a) Show that O is an order with discrd O = 9.
(b) Show that O has stable cancellation.
Chapter 21
Hereditary and extremal orders
In this chapter, we consider hereditary orders, those with the simplest kind of module
theory; we characterize these orders in several ways, including showing they have an
extremal property with respect to their Jacobson radical.
21.1 ⊲ Hereditary and extremal orders

Let 𝑅 be a Dedekind domain. Then 𝑅 is hereditary: every submodule of a projective
module is again projective. (Hence the name: projectivity is inherited by a submodule.)
A noetherian domain is hereditary if and only if every ideal of 𝑅 is projective, or
equivalently, that every submodule of a free 𝑅-module is a direct sum of ideals of
𝑅. This property is used in the proof of unique factorization of ideals and makes the
structure theory of modules over a Dedekind quite nice. (Note, however, that every
order in a number field which is not maximal is not hereditary.)
It is important to identify those orders for which projective modules abound. Let
𝐵 be a simple finite-dimensional 𝐹-algebra and let O ⊆ 𝐵 be an 𝑅-order.
Definition 21.1.1. We say O is left hereditary if every left O-ideal 𝐼 ⊆ O is projective

as a left O-module.
We could define also right hereditary, but left hereditary and right hereditary are
equivalent for an 𝑅-order O, and so we simply say hereditary. We have O hereditary
if and only if every O-submodule of a projective finitely generated O-module is
projective—that is to say, projectivity is inherited by submodules. Moreover, being
hereditary is a local property.
Maximal orders are hereditary (Theorem 18.1.2), and one motivation for hereditary
orders is that many of the results from chapter 18 on the structure of two-sided ideals
extend from maximal orders to hereditary orders (Theorem 21.4.9).
Proposition 21.1.2. Suppose O is hereditary. Then the set of two-sided invertible

fractional O-ideals of 𝐵 forms an abelian group under multiplication, generated by
the prime O-ideals.
333
334 CHAPTER 21. HEREDITARY AND EXTREMAL ORDERS
Hereditary orders are an incredibly rich class of objects, and they may be charac-
terized in a number of equivalent ways (Theorem 21.5.1). We restrict to the complete
local case, and suppose now that 𝑅 is a complete DVR with unique maximal ideal 𝔭
and residue field 𝑘 = 𝑅/𝔭.
Just as maximal orders are defined in terms of containment, we say O is extremal
if whenever O0 ⊇ O and rad O0 ⊇ rad O, then O0 = O. If O is not extremal, then
O0 := OL (rad O) ) O (21.1.3)
is a superorder. We then have the following main theorem (Theorem 21.5.1).
Main Theorem 21.1.4. Let 𝑅 be a complete DVR and let O ⊆ 𝐵 be an 𝑅-order in a

simple 𝐹-algebra 𝐵. Let 𝐽 := rad O. Then the following are equivalent:
(i) O is hereditary;
(ii) 𝐽 is projective as a left O-module;
(ii0) 𝐽 is projective as a right O-module;
(iii) OL (𝐽) = O;
(iii0) OR (𝐽) = O;
(iv) 𝐽 is invertible as a (sated) two-sided O-ideal; and
(v) O is extremal.
The fact that hereditary orders are the same as extremal orders is quite remarkable,
and gives tight control over the structure of hereditary orders: extremal orders are
equivalently characterized as endomorphism algebras of flags in a suitable sense, and
so we have the following important corollary for quaternion algebras.
Corollary 21.1.5. Suppose further that 𝐵 is a quaternion algebra. Then an 𝑅-order

O ⊆ 𝐵 is hereditary if and only if either O is maximal or

𝑅 𝑅
O' ⊆ M2 (𝐹) ' 𝐵.
𝔭 𝑅
It is no surprise that we meet again the order from Example 20.1.2! The reader who
is willing to accept Corollary 21.1.5 can profitably move on from this chapter, as the
ring of upper triangular matrices is explicit enough to work with in many cases. That
being said, the methods we encounter here will be useful in framing investigations of
orders beyond the hereditary ones.
21.2 Extremal orders

In this section, we will see how to extend an order to a superorder using the Jacobson
radical, and we will characterize those orders that are extremal with respect to this
process.
We work locally throughout this section; let 𝑅 be a complete DVR with maximal
ideal 𝔭 = rad(𝑅) and residue field 𝑘 = 𝑅/𝔭, and let 𝐹 = Frac 𝑅. Let 𝐵 be a finite-
dimensional separable 𝐹-algebra and let O ⊆ 𝐵 be an 𝑅-order.
21.2. EXTREMAL ORDERS 335
21.2.1. Our motivation comes from the following: we canonically associate a super-
order as follows. Let 𝐽 := rad O and O0 := OL (𝐽). Then O0 ⊇ O. By Corollary
20.5.5, 𝐽 𝑟 ⊆ 𝔭O ⊆ 𝔭O0 for some 𝑟 > 0, and then 𝐽 ⊆ rad O0.
Definition 21.2.2. An 𝑅-order O0 ⊆ 𝐵 radically covers O if O0 ⊇ O and rad O0 ⊇

rad O. We say O is extremal if whenever O0 radically covers O then O0 = O.
We can think of extremal orders as like maximal orders, but under certain inclu-
sions.
Proposition 21.2.3. An 𝑅-order O is extremal if and only if OL (rad O) = O if and

only if OR (rad O) = O.
Proof. The argument is due to Jacobinski [Jaci71, Proposition 1].

We first prove (⇒). Suppose O is extremal, and let 𝐽 = rad O and O0 = OL (𝐽).
By Corollary 20.5.5, 𝐽 is topologically nilpotent as a O-ideal, so the same is true as a
O0 ideal, and 𝐽 ⊆ rad O0 and O0 radically covers O. Since O is extremal, we conclude
O0 = O. The same argument works on the right.
Next we prove (⇐). Let 𝐽 = rad O, suppose O = OL (𝐽); let O0 radically cover O,
and let 𝐽 0 = rad O0. As lattices, we have 𝔭𝑠 O0 ⊆ 𝐽 for some 𝑠 > 0; by Theorem 20.5.1,
(𝐽 0) 𝑟 ⊆ 𝔭O0 for some 𝑟 > 0, so putting these together we have (𝐽 0) 𝑡 ⊆ 𝐽 for some 𝑡 > 0.
Suppose 𝑡 > 1. Since O0 radically covers, we have 𝐽 ⊆ 𝐽 0; thus 𝐽 (𝐽 0) 𝑡−1 ⊆ (𝐽 0) 𝑡 ⊆ 𝐽
and (𝐽 0) 𝑡−1 ⊆ OR (𝐽) = O. But then since ((𝐽 0) 𝑡−1 ) 𝑡 ⊆ (𝐽 0) 𝑡 ⊆ 𝐽, by Corollary 20.5.5,
(𝐽 0) 𝑡−1 ⊆ 𝐽. Continuing in this way, we obtain 𝑡 = 1 and 𝐽 0 ⊆ 𝐽. Therefore 𝐽 = 𝐽 0
and O = OL (𝐽) = OL (𝐽 0) = O0, thus O is extremal.
Lemma 21.2.4. Let O be an 𝑅-order and let O0 ⊆ 𝐵 be an 𝑅-order containing O. Let

𝐽 0 := rad O0. Then O + 𝐽 0 is an 𝑅-order that radically covers O. If further 𝐽 0 ⊆ O,
then 𝐽 0 ⊆ 𝐽.
21.2.5. In view of Lemma 21.2.4, an extremal order is determined by its homomorphic

image in a nice 𝑘-algebra as follows.
Let O be an extremal 𝑅-order and let O0 ⊆ 𝐵 be a maximal 𝑅-order containing O.
Let 𝐽 0 := rad O0. By Lemma 21.2.4, O + 𝐽 0 is an 𝑅-order that radically covers O, so
O + 𝐽 0 = O. Therefore 𝐽 0 ⊆ O. By the second part of Lemma 21.2.4, we immediately
conclude 𝐽 0 ⊆ 𝐽 := rad O. In sum,
𝐽 0 = rad O0 ⊆ 𝐽 = rad O ⊆ O. (21.2.6)
Consider now the reduction map 𝜌 : O0 → O0/𝐽 0. Since 𝐽 0 ⊆ O, if 𝐴 = 𝜌(O)

then O = 𝜌 −1 ( 𝐴). But since 𝔭O0 ⊆ 𝐽 0 and O0 is a maximal 𝑅-order, the codomain is
a nice, finite dimensional 𝑘-algebra, something we will get our hands on in the next
section.
Paragraph 21.2.5 has the following consequence.

Lemma 21.2.7. Suppose that 𝐵 is a division algebra over 𝐹 and let O ⊆ 𝐵 be extremal.
Then O is maximal.
Proof. Recall 13.3.7. The valuation ring O0 ⊇ O has the unique maximal two-sided
ideal 𝐽 0 = rad O0 = O0 \ (O0) × , so O0/𝐽 0 is a field. We have (21.2.6) 𝐽 0 ⊆ 𝐽, but then
𝐽/𝐽 0 = {0} ⊆ O0/𝐽 0 thus 𝐽 = 𝐽 0. Thus O0 radically covers O, and since O is extremal,
O = O0.
Remark 21.2.8. We stop short in our explicit description of local extremal orders in
section 21.2: we gave a construction in 21.3.1 only for 𝐵 ' M𝑛 (𝐹). The results extend
to 𝐵 ' M𝑛 (𝐷) where 𝐷 is a division algebra over 𝐹 by considering lattices in a free
left 𝐷-module: see Reiner [Rei2003, Theorem 39.14].
21.3 ∗ Explicit description of extremal orders

We now turn to an explicit description of extremal orders. In Lemma 10.5.4, we saw
that maximal orders in a matrix algebra 𝐵 = End𝐹 (𝑉) are endomorphism algebras of
lattices. In this section, we extend this to encompass orders that arise from endomor-
phism algebras of a chain of lattices: these orders are “block upper triangular”, and
can be characterized in a number of ways.
21.3.1. Let 𝑉 be a finite-dimensional 𝐹-vector space and let 𝐵 = End𝐹 (𝑉); then 𝑉 is
a simple 𝐵-module. Let 𝑀 ⊆ 𝑉 be an 𝑅-lattice. By Lemma 10.5.4, Λ := End𝑅 (𝑀) is
a maximal 𝑅-order; we have rad Λ = 𝔭Λ.
Choosing a basis for 𝑀, we get Λ ' M𝑛 (𝑅) ⊆ M𝑛 (𝐹) ' 𝐵, and rad Λ = M𝑛 (𝔭).
Now let 𝑍 := 𝑀 ⊗𝑅 𝑘 = 𝑀/𝔭𝑀. Then 𝑍 is a finite-dimensional vector space over
𝑘. Let
E : {0} = 𝑍0 ( 𝑍1 ( · · · ( 𝑍 𝑠−1 ( 𝑍 𝑠 = 𝑍
be a (partial) flag, a strictly increasing sequence of 𝑘-vector spaces. We define
OL (E) := {𝛼 ∈ Λ : 𝛼𝑍𝑖 ⊆ 𝑍𝑖 : 𝑖 = 0, . . . , 𝑠}.
Equivalently, let 𝑀𝑖 be the inverse image of 𝑍𝑖 under the projection 𝑀 → 𝑍; then we

have a chain
𝔭𝑀 = 𝑀0 ( 𝑀1 ( · · · ( 𝑀𝑠−1 ( 𝑀𝑠 = 𝑀 (21.3.2)
and
OL (E) = {𝛼 ∈ Λ : 𝛼𝑀𝑖 ⊆ 𝑀𝑖 : 𝑖 = 0, . . . , 𝑠}.
Lemma 21.3.3. OL (E) ⊆ Λ is an 𝑅-order with
rad OL (E) = {𝛼 ∈ Λ : 𝛼𝑍𝑖 ⊆ 𝑍𝑖−1 } = {𝛼 ∈ Λ : 𝛼𝑀𝑖 ⊆ 𝑀𝑖−1 }.
Proof. That OL (E) is an order follows in the same way as the proof of Lemma 10.2.7.
For the statement on the radical: let 𝐽 = {𝛼 ∈ Λ : 𝛼𝑍𝑖 ⊆ 𝑍𝑖−1 }. Then 𝐽 ⊆ OL (E)
21.3. ∗ EXPLICIT DESCRIPTION OF EXTREMAL ORDERS 337
is a two-sided ideal. We have 𝐽 𝑠 ⊆ 𝔭Λ pushing along the flag, so 𝐽 ⊆ rad OL (E) by

Corollary 20.5.5. Conversely,
𝑠
Ê
OL (E)/𝐽 ' End 𝑘 (𝑍𝑖 /𝑍𝑖−1 );
𝑖=1
each factor is simple, so the sum is (Jacobson) semisimple; therefore 𝐽 ⊆ rad OL (E)
and equality holds.
Example 21.3.4. If we take the trivial flag OL (E) : {0} = 𝑍0 ( 𝑍1 = 𝑍, then
OL (E) = Λ, so this recovers the construction of maximal orders.
Example 21.3.5. Let E be the complete flag of length 𝑠 = 𝑛 + 1 = dim𝐹 𝑉, where
each quotient has dim 𝑘 (𝑍𝑖+1 /𝑍𝑖 ) = 1. Then there exists a basis 𝑧1 , . . . , 𝑧 𝑛 of 𝑍 so
that 𝑍𝑖 has basis 𝑧1 , . . . , 𝑧 𝑛−𝑖 ; We lift this to basis to 𝑥 1 , . . . , 𝑥 𝑛 of 𝑀 (by Nakayama’s
lemma), and in this basis, we have
𝑅 𝑅 𝑅 ... 𝑅
𝔭
© ª
𝑅 𝑅 ... 𝑅®
OL (E) = 𝔭 𝔭
®
𝑅 ... 𝑅®
.. .. .. ..
.
.. ®®
. . . .®
«𝔭 𝔭 𝔭 ... 𝑅¬
consisting of matrices which are upper triangular modulo 𝔭, and
𝔭 𝑅 𝑅 ... 𝑅
0 1 ... 0
𝔭 𝔭
© ª
𝑅 ... 𝑅® ©. .. .. ª®
. ..
rad OL (E) = 𝔭 𝔭 𝔭
® .
... 𝑅 ® = OL (E) . . .® (21.3.6)
.. .. .. .. ®® 0 0 1®
.. ®
...
. . . . .®
«𝜋 0 ... 0¬
«𝔭 𝔭 𝔭 ... 𝔭¬
where the latter is taken to be a block matrix with lower left entry 𝜋 and top right entry
the (𝑛 − 1) × (𝑛 − 1) identity matrix.
Other choices of flag give an order which lie between OL (E) and Λ: we might
think of them as being block upper triangular orders.
Now for the punch line of this section.
Proposition 21.3.7. Let O ⊆ 𝐵 be an 𝑅-order. Then O is extremal if and only if
O = OL (E) for a flag E.
Proof. Let O = OL (E). Let 𝐽 = rad O; we seek to apply Proposition 21.2.3, so we
show that O = OL (𝐽). By Lemma 21.3.3, we have 𝐽 𝑀𝑖 = 𝑀𝑖−1 so OL (𝐽) 𝑀𝑖−1 =
OL (𝐽)𝐽 𝑀𝑖 = 𝐽 𝑀𝑖 = 𝑀𝑖−1 for 𝑖 = 1, . . . , 𝑠. Since 𝑀0 = 𝔭𝑀 ' 𝑀, we conclude
OL (𝐽) = OL (E) = O by definition.
Conversely, suppose O is extremal with 𝐽 = rad O. Let 𝑠 be minimal so that
𝐽 𝑠 = 𝔭O. We may embed O ⊆ Λ for some Λ, and we take the flag
E : {0} = 𝐽 𝑠 𝑍 ( 𝐽 𝑠−1 𝑍 ( · · · ( 𝐽 𝑍 ( 𝑍.
Then O ⊆ O(E) and rad O(E) ⊇ 𝐽 by construction, so since O is extremal, we have

O = O(E).
21.4 Hereditary orders

We now link the orders in the previous two sections to another important type of order.
The theory of extremal and hereditary orders was developed by Brumer [Brum63a,
Brum63b], Drozd–Kirichenko [DK68], Harada [Har63a, Har63b, Har63c], Jacobinski
[Jaci71], and Hĳikata–Nishida [HN94]. An overview of the local and global theory of
hereditary orders is given by Reiner [Rei2003, §§2f, 39–40], and Drozd–Kirichenko–
Roiter [DKR67] and Hĳikata–Nishida [HN94] extend some results from hereditary
orders to Bass orders.
Let 𝑅 be a noetherian domain with 𝐹 = Frac 𝑅, and let 𝐵 be a separable 𝐹-algebra,
and let O ⊆ 𝐵 be an 𝑅-order.
Definition 21.4.1. We say O is left hereditary if every left O-ideal 𝐼 ⊆ O is projective

as a left O-module.
21.4.2. We could similarly define right hereditary, but since an order O is left and
right noetherian, it follows that O is left hereditary if and only if O is right hereditary:
see Exercise 21.8. When 𝐵 is a quaternion algebra, the standard involution inter-
changes and left and right, so the two notions are immediately seen to be equivalent.
Accordingly, we say hereditary for either sided notion.
Example 21.4.3. In the generic case 𝐹 = 𝑅 and O = 𝐵, we note that every semisimple
algebra 𝐵 over a field 𝐹 is hereditary: by Lemma 7.3.5, every 𝐵-module is semisimple
hence the direct sum of simple 𝐵-modules equivalently maximal left ideals, by Lemma
7.2.7.
21.4.4. By 20.2.6, being hereditary is a local property.
The following lemma motivates the name hereditary: projectivity is inherited

by submodules. (Note that since 𝑅 is noetherian, a finitely generated O-module is
noetherian, so every submodule is finitely generated.)
Lemma 21.4.5. Let O be hereditary, and let 𝑃 be a finitely generated projective left
O-module. Then every submodule 𝑀 ⊆ 𝑃 is isomorphic as a left O-module to a finite
direct sum of finitely generated left O-ideals; in particular, 𝑀 is projective.
Proof. We may suppose without loss of generality that 𝑃 ' O𝑟 . We proceed by

induction on 𝑟; the case 𝑟 = 1 holds by definition. Decompose O𝑟 = 𝐸 ⊕ O where
𝐸 ' O𝑟 −1 . From the exact sequence
0 → ker 𝜙 → 𝑀 → 𝑀 ∩ 𝐸 → 0
and projectivity, we find that 𝑀 ' (𝑀 ∩ 𝐸) ⊕ ker 𝜙 where ker 𝜙 ⊆ O is a left ideal of
O. By induction, 𝑀 ∩ 𝐸 is projective, so the same is true of 𝑀.
21.4. HEREDITARY ORDERS 339
Corollary 21.4.6. O is hereditary if and only if every submodule of a projective

O-module is projective.
Proof. The implication (⇒) is Lemma 21.4.5; for the implication (⇐), O is projective
(free!) as a left O-module and every left ideal is a O-submodule 𝐼 ⊆ O, so by hypothesis
𝐼 is projective.
Remark 21.4.7. O is hereditary if and only if every 𝑅-lattice 𝐼 ⊆ 𝐵 with O ⊆ OL (𝐼)

is projective as a left O-module, after rescaling.
However, a bit of a warning is due. If 𝐼 ⊆ 𝐵 is an 𝑅-lattice that is projective as a
left O-module, then we have shown that 𝐼 is projective as a left OL (𝐼)-module (Lemma
20.3.1), whence left invertible (Theorem 20.3.3) as a lattice. But the converse need
not be true; so it is important that in the definition of hereditary we do not require
that every left O-fractional ideal 𝐼 is invertible as a left fractional O-ideal (Definition
16.5.17): the latter carries the extra assumption that 𝐼 is sated. See also Remark 20.3.8.
Lemma 21.4.8. Let O ⊆ 𝐵 be a hereditary 𝑅-order and let O0 ⊇ O be an 𝑅-

superorder. Then O0 is hereditary.
Proof. Let 𝐼 0 ⊆ O0 be a left O0-ideal. Scaling we may take 𝐼 0 ⊆ O, and it is a left

O-ideal. Since O is hereditary, 𝐼 0 is projective as a left O-module; by Lemma 20.3.1,
𝐼 0 is projective as a left O0-module.
One of the desirable aspects of hereditary orders is that many of the results from
chapter 18 on the structure of two-sided ideals extend from maximal orders to heredi-
tary orders. Indeed, section 18.2 made no maximality hypothesis (we held out as long
as we could!).
Theorem 21.4.9. Let 𝑅 be a Dedekind domain and let O be a hereditary 𝑅-order in a

simple 𝐹-algebra 𝐵. Then the set of two-sided invertible fractional O-ideals of 𝐵 forms
an abelian group under multiplication, generated by the invertible prime O-ideals.
Proof. Proven in the same manner as in Theorem 18.3.4; a self-contained proof is

requested in Exercise 21.4.
Remark 21.4.10. Theorem 21.4.9 is proven by Vignéras [Vig80a, Théorème I.4.5],

but there is a glitch in the proof. Let 𝑅 be a Dedekind domain, let 𝐵 be a quaternion
algebra over 𝐹 = Frac 𝑅, and let O ⊆ 𝐵 be an 𝑅-order. Vignéras claims that the
two-sided ideals of O form a group that is freely generated by the prime ideals, and the
proof uses that if 𝐼 is a two-sided ideal then 𝐼 is invertible. This is false for a general
order O (see Example 16.5.12).
If one restricts to the group of invertible two-sided ideals, the logic of the proof
is still flawed. The proof does not use anything about quaternion algebras, and √ works
verbatim for the case where 𝑅 = Z ⊆ 𝐹 = Q and 𝐵 is replaced by 𝐾 = Q( 𝑑 𝐾 ) and
O is replaced by an order of discriminant 𝑑 = 𝑑 𝐾 𝑓 2 √that is not maximal, of conductor
𝑓 ∈ Z>1 , as in section 16.1. Then the ideal 𝔣 = 𝑓 Z + 𝑑Z is not invertible, but 𝔣 ) ( 𝑓 )
and ( 𝑓 ) is invertible but not maximal, so the group of invertible ideals is not generated
by primes.
However, if one supposes that every two-sided ideal is invertible (as a lattice),
then the argument can proceed: this is the class of hereditary orders, and is treated in
Theorem 21.4.9.
Remark 21.4.11. The module theory for hereditary noetherian prime rings, generaliz-
ing hereditary orders, has been worked out by Levy–Robson [LR2011].
21.5 ∗ Classification of local hereditary orders

We now come to the main theorem of this chapter, relating extremal orders, hereditary
orders, their modules and composition series in the local setting.
Theorem 21.5.1. Let 𝑅 be a complete DVR and let 𝐹 := Frac 𝑅. Let 𝐵 be a finite-
dimensional 𝐹-algebra, and let O ⊆ 𝐵 be an 𝑅-order. Let 𝐽 := rad O. Then the
following are equivalent, along with the conditions 0 where ‘left’ is replaced by ‘right’:
(i) O is extremal;
(ii) Every projective indecomposable left O-submodule 𝑃 ⊆ 𝐵 is the minimum
O-supermodule of 𝐽𝑃;
(iii) Every projective indecomposable left O-module 𝑃 has a unique composition
series;
(iv) Every projective indecomposable left O-module 𝑃 has a unique composition
series consisting of projectives;
(v) O is hereditary;
(vi) 𝐽 is projective as a left O-module;
(vii) If 𝑃 is a projective indecomposable left O-module, then 𝐽𝑃 is also projective
indecomposable; and
(viii) 𝐽 is invertible as a (sated) two-sided O-ideal.
Proof. See Hĳikata–Nishida [HN94, §1].
Corollary 21.5.2. A maximal order is hereditary.
Proof. We proved this in Theorem 18.1.2, but here is another proof using Theorem
21.5.1: the property of being maximal is local, and a maximal order is extremal.
To conclude, we classify the lattices of a local hereditary order.

21.5.3. Suppose 𝑅 is a complete DVR and that 𝐵 ' M𝑛 (𝐹). Suppose O ⊆ 𝐵 is a
hereditary 𝑅-order; then by Theorem 21.5.1, O = OL (E) is extremal, arising from a
chain 21.3.2 which by Lemma 21.3.3 is of the form
𝔭𝑀 = 𝑀0 = 𝐽 𝑠 𝑀 ( 𝐽 𝑠−1 𝑀 ( · · · ( 𝑀𝑠−1 = 𝐽 𝑀 ( 𝑀𝑠 = 𝑀,
with each quotient 𝑀𝑖 /𝑀𝑖+1 ' 𝑀/𝐽 𝑀 simple.

We claim that the set 𝑀, 𝐽 𝑀, . . . , 𝐽 𝑠−1 𝑀 form a complete set of isomorphism
classes of indecomposable left O-modules. Indeed, these modules are all mutually
nonisomorphic, because an isomorphism 𝜙 : 𝐽 𝑖 𝑀 −→ ∼ 𝐽 𝑗 𝑀 of left O-modules extends
to an isomorphism 𝜙 ∈ End 𝐵 (𝐵) ' 𝐹 so is given by (right) multiplication by a power
21.5. ∗ CLASSIFICATION OF LOCAL HEREDITARY ORDERS 341
of 𝜋, impossible unless 𝑖 ≡ 𝑗 (mod 𝑠). And if 𝑁 is an indecomposable left O-module,

then 𝐹 𝑁 ' 𝐹 𝑛 is ‘the’ simple 𝐵-module, so 𝑁 is isomorphic to a lattice in 𝑉. Since 𝐽
is invertible, we may replace 𝑁 by 𝐽 𝑟 𝑁 with 𝑟 ∈ Z, to suppose that 𝑀 ⊇ 𝑁 ) 𝐽 𝑀. But
𝑀/𝐽 𝑀 ' O/𝐽 is simple as a left O-module, so 𝑁 = 𝑀. (See also Reiner [Rei2003,
Theorem 39.23].)
Exercises
1. Show that a Dedekind domain is hereditary (cf. Exercise 9.5).
√
2. Let 𝑅 = Z, let 𝐵 = 𝐾 = Q( 𝑑) with 𝑑 the discriminant of 𝐾, and let 𝑆 ⊆ Z𝐾 be
an order. Show that 𝑆 is hereditary if and only if 𝑆 is maximal.
3. Let 𝑅 bea DVR with maximal ideal 𝔭 = 𝜋𝑅 and 𝐹 = Frac 𝑅 with char 𝐹 ≠ 2.
1, 𝜋
Let 𝐵 = and O = 𝑅h𝑖, 𝑗i the standard order. Show directly that rad O =
𝐹
O 𝑗 = 𝑗O, and conclude that O is hereditary (but not a maximal order).
I 4. Give a self-contained proof of Theorem 21.4.9 following Theorem 18.3.4.
(Where does the issue with invertibility arise?)
5. Let 𝑅 be a complete DVR and let O be a hereditary 𝑅-order. Show that O is
hereditary if and only if rad O is an invertible (sated) two-sided O-ideal.
I 6. In this exercise, we prove Lemma 21.2.4 following Reiner [Rei2003, Exercise
39.2]. We adopt the notation from that section, so in particular 𝑅 be a complete
DVR with maximal ideal 𝔭 = rad(𝑅). Let O be an 𝑅-order and let O0 ⊆ 𝐵 be
an 𝑅-order containing O. Let 𝐽 0 = rad O0.
(a) Show that O + 𝐽 0 is an 𝑅-order.
(b) Show that O + 𝐽 0 radically covers O. [Hint: let 𝐽 = rad O, and claim that
𝐽 +𝐽 0 ⊆ rad(O+𝐽 0). For 𝑟 large, show 𝐽 𝑟 ⊆ 𝔭O so (𝐽 +𝐽 0) 𝑟 ⊆ 𝔭O0 +𝐽 0 and
3
(𝐽 0) 𝑟 ⊆ 𝔭O0, and then making 𝑟 even larger show (𝐽 + 𝐽 0) 𝑟 ⊆ 𝔭(O + 𝐽 0).
Conclude using Corollary 20.5.5.]
(c) If further 𝐽 0 ⊆ O, show that 𝐽 0 ⊆ 𝐽.
7. Let 𝑅 be a Dedekind domain with 𝐹 = Frac(𝑅), let 𝐵 be finite-dimensional
𝐹-algebra, and let O ⊆ 𝐵 be a hereditary order. Let 𝑃 be a finitely generated
projective O-module. Show that 𝑃 is indecomposable if and only if 𝑉 := 𝑃 ⊗𝑅 𝐹
is simple as a 𝐵-module.
I 8. Let 𝑅 be a Dedekind domain, and let O ⊆ 𝐵 be an 𝑅-order in a finite-dimensional
𝐹-algebra. Show that O is left hereditary (every left O-ideal is projective) if
and only if it is right hereditary (every right O-ideal is projective). [See Reiner
[Rei2003, Theorem 40.1].]
9. Consider the ring
𝑎 0
𝐴 := : 𝑎 ∈ Z, 𝑏, 𝑐 ∈ Q .
𝑏 𝑐
Show that every submodule of a projective left 𝐴-module is projective, but the
same is not true on the right.
10. Let 𝑅 be a Dedekind domain. Let 𝐵 be a separable 𝐹-algebra, and let 𝐵 '
𝐵1 × · · · × 𝐵𝑟 be its decomposition into simple components, with 𝐵𝑖 = 𝐵𝑒 𝑖 for
central idempotents 𝑒 𝑖 . Let 𝐾𝑖 be the center of 𝐵𝑖 , and let 𝑆𝑖 be the integral
closure of 𝑅 in 𝐾𝑖 .
a) Let O ⊆ 𝐵 be a hereditary 𝑅-order. Show that O ' O1 × · · · × O𝑟 where
O𝑖 = O𝑒 𝑖 , and each O𝑖 is a hereditary 𝑅-order in 𝐵𝑖 .
b) Conversely, if O𝑖 ⊆ 𝐵𝑖 is a hereditary 𝑅-order, then O1 × · · · × O𝑟 is a
hereditary 𝑅-order in 𝐵.
[Hint: use the fact that hereditary orders are extremal.]
11. For the following exercise, we consider integral group rings. Let 𝐺 be a finite
group of order 𝑛 = #𝐺 and let 𝑅 be a Dedekind domain with 𝐹 = Frac 𝑅.
Suppose that char 𝐹 - 𝑛. Then 𝐵 := 𝐹 [𝐺] is a separable 𝐹-algebra by Exercise
7.15. Let O = 𝑅[𝐺].
a) Let O0 ⊇ O be an 𝑅-superorder of O in 𝐵. Show that
O ⊆ O0 ⊆ 𝑛−1 O.
𝑎 𝑔 𝑔 ∈ O0 with 𝑎 𝑔 ∈ 𝐹, show that

Í
[Hint: for ell 𝛼 = 𝑔
Tr 𝐵 |𝐹 (𝛼𝑔) = 𝑛𝑎 𝑔 ∈ 𝑅.
Conclude that O0 ⊆ 𝑛−1 O.]

b) Show that O is maximal if and only if O is hereditary if and only if
𝑛 ∈ 𝑅 × . [Hint: if O is hereditary, then O contains the central idempotent
𝑛−1 𝑔 ∈𝐺 𝑔 by Exercise 21.10.]
Í
c) We define the left conductor of O0 into O to be the colon ideal
(O0 : O) L = {𝛼 ∈ 𝐵 : 𝛼O0 ⊆ O}.
(and similarly on right). Prove that

𝑡
∑︁ 𝑛
(O0 : O) L = codiff(O𝑖0 ).
𝑖=1
𝑛 𝑖
12. Give an explicit description like Example 21.3.5 for OL (E) when dim𝐹 𝑉 = 3, 4.
13. Let 𝑅 be a Dedekind domain, and let O ⊆ 𝐵 be an 𝑅-order in a finite-dimensional
simple 𝐹-algebra. Show that O is maximal if and only if O is hereditary and
rad O ⊆ O is a maximal two-sided ideal.
Chapter 22
Quaternion orders and ternary

quadratic forms
In this chapter, we classify orders over a Dedekind domain in terms of ternary quadratic
forms; this is the integral analogue to what we did over fields in Chapter 5.
22.1 ⊲ Quaternion orders and ternary quadratic forms

We begin our project by returning to the classification over fields: in Chapter 5 and 6
(see Main Theorem 5.2.5 and Theorem 6.4.7), we saw that quaternion algebras over
a field 𝐹 are classified by similarity classes of nondegenerate ternary quadratic forms
over 𝐹. We will soon see that, suitably interpreted, quaternion orders are classified by
similarity classes of integral ternary quadratic forms.
Let 𝑅 be a PID with field of fractions 𝐹 := Frac 𝑅. We recall that the similarity
class of a ternary quadratic form 𝑄 : 𝑅 3 → 𝑅 is determined by the natural change of
variable by GL3 (𝑅) on the domain and by rescaling by 𝑅 × on the codomain, and that
𝑄 is nondegenerate if and only if disc(𝑄) ≠ 0.
Main Theorem 22.1.1. Let 𝑅 be a PID. Then there is a (reduced) discriminant-
preserving bĳection

Nondegenerate ternary quadratic Quaternion orders over 𝑅
↔ .
forms 𝑄 over 𝑅 up to similarity up to isomorphism
One beautiful feature of the bĳection in Main Theorem 22.1.1 is that it can be given
explicitly. Let 𝑄 : 𝑅 3 → 𝑅 be a ternary quadratic form with nonzero discriminant,
and let 𝑒 1 , 𝑒 2 , 𝑒 3 be the standard basis for 𝑅 3 . Then the extension to 𝐹 given by
𝑄 𝐹 : 𝐹 3 → 𝐹 is a ternary quadratic space whose even Clifford algebra (section 5.3) is
a quaternion algebra 𝐵. Moreover, the 𝑅-lattice O with basis
1, 𝑖 := 𝑒 2 𝑒 3 , 𝑗 := 𝑒 3 𝑒 1 , 𝑘 := 𝑒 1 𝑒 2
is closed under multiplication and so defines an 𝑅-order in 𝐵. Explicitly, if the
quadratic form 𝑄 is given by
𝑄(𝑥, 𝑦, 𝑧) = 𝑎𝑥 2 + 𝑏𝑦 2 + 𝑐𝑧2 + 𝑢𝑦𝑧 + 𝑣𝑥𝑧 + 𝑤𝑥𝑦 ∈ 𝑅[𝑥, 𝑦, 𝑧]
343
with (half-)discriminant
𝑁 := 4𝑎𝑏𝑐 + 𝑢𝑣𝑤 − 𝑎𝑢 2 − 𝑏𝑣 2 − 𝑐𝑤 2 ≠ 0,
then we associate the quaternion 𝑅-order O ⊆ 𝐵 with basis 1, 𝑖, 𝑗, 𝑘 and multiplication

laws
𝑖 2 = 𝑢𝑖 − 𝑏𝑐 𝑗 𝑘 = 𝑎𝑖 = 𝑎(𝑢 − 𝑖)
𝑗 2 = 𝑣 𝑗 − 𝑎𝑐 𝑘𝑖 = 𝑏 𝑗 = 𝑏(𝑣 − 𝑗) (22.1.2)
2
𝑘 = 𝑤𝑘 − 𝑎𝑏 𝑖 𝑗 = 𝑐𝑘 = 𝑐(𝑤 − 𝑘).
The other multiplication rules are determined by the skew commutativity relations
(4.2.16) coming from the standard involution; one beautiful consequence is the equality
𝑖 𝑗 𝑘 = 𝑗 𝑘𝑖 = 𝑘𝑖 𝑗 = 𝑎𝑏𝑐.
The 𝑅-order O defined by (22.1.2) is called the even Clifford algebra Clf 0 (𝑄) of
𝑄—its algebra structure is obtained by restriction from the even Clifford algebra of
𝑄 𝐹 —and the reduced discriminant of O is discrd(O) = (𝑁). At least one of the
minors
𝑢 2 − 4𝑏𝑐, 𝑣 2 − 4𝑎𝑐, 𝑤 2 − 4𝑎𝑏
of the Gram matrix of 𝑄 in the standard basis is nonzero since 𝑄 is nondegenerate, so
for example if 𝑤 2 − 4𝑎𝑏 ≠ 0 and char 𝐹 ≠ 2, completing the square we find
𝑤 2 − 4𝑎𝑏, −𝑎𝑁

O⊂𝐵' .
𝐹
It is straightforward to show that the isomorphism class of O is determined by the

similarity class of 𝑄 (using the even Clifford algebra construction). Therefore, the
proof of Main Theorem 22.1.1 amounts to verifying that every quaternion order arises
this way up to isomorphism, and that isomorphic quaternion algebras yield similar
ternary quadratic forms.
To this end, we define an inverse to the even Clifford algebra construction. Let
O ⊂ 𝐵 be a quaternion order over 𝑅 with reduced discriminant discrd(O) generated
by 𝑁 ∈ 𝑅 nonzero. Recalling 15.6, let
(O♯ ) 0 = {𝛼 ∈ O♯ : trd(𝛼) = 0}
be the trace zero elements in the dual of O with respect to the reduced trace pairing.
Then we associate the ternary quadratic form
𝑁 nrd♯ (O) : (O♯ ) 0 → 𝑅

(22.1.3)
𝛼 ↦→ 𝑁 nrd(𝛼);
explicitly, we have
𝑁𝑖 ♯ = 𝑗 𝑘 − 𝑘 𝑗, 𝑁 𝑗 ♯ = 𝑘𝑖 − 𝑖𝑘, 𝑁 𝑘 ♯ = 𝑖 𝑗 − 𝑗𝑖
22.1. ⊲ QUATERNION ORDERS AND TERNARY QUADRATIC FORMS 345
where 1, 𝑖, 𝑗, 𝑘 is an 𝑅-basis of O, so

𝑁 nrd♯ (𝑥, 𝑦, 𝑧) = nrd 𝑥( 𝑗 𝑘 − 𝑘 𝑗) + 𝑦(𝑘𝑖 − 𝑖𝑘) + 𝑧(𝑖 𝑗 − 𝑗𝑖) . (22.1.4)
It is then a bit of beautiful algebra to verify that 𝑁 nrd♯ has discriminant 𝑁 and that
(22.1.3) furnishes an inverse to the even Clifford map.
Just as in the case of fields, the translation from quaternion orders to ternary
quadratic forms makes the classification problem easier: we replace the potentially
complicated notion of finding a lattice closed under multiplication in a quaternion
algebra with the simpler notion of choosing coefficients of a quadratic form.
To conclude this introduction, we state a more general bĳective result stated in
terms of lattices. Let 𝑅 be a Dedekind domain with 𝐹 = Frac 𝑅, let 𝑄 𝐹 : 𝑉 → 𝐹
be a nondegenerate ternary quadratic form. If 𝑀 ⊆ 𝑉 is an 𝑅-lattice, and 𝔩 ⊆ 𝐹 is
a fractional ideal of 𝑅 such that 𝑄(𝑀) ⊆ 𝔩, then we have an induced quadratic form
𝑄 : 𝑀 → 𝔩; we call such a form a quadratic module in 𝑉. Given a fractional ideal
𝔞 ⊆ 𝐹, the twist by 𝔞 of the quadratic module 𝑄 : 𝑀 → 𝔩 in 𝑉 is the quadratic module
𝔞𝑀 → 𝔞2 𝔩. A twisted similarity between quadratic modules 𝑄, 𝑄 0 in 𝑉 is a similarity
between 𝑄 and a twist of 𝑄 0. From these notions in hand, we have the following
theorem (a special case of Main Theorem 22.5.7).
Theorem 22.1.5. Let 𝑅 be a Dedekind domain, and let 𝑄 𝐹 : 𝑉 → 𝐹 be a nondegen-

erate ternary quadratic form. Let 𝐵 := Clf 0 𝑉. Then the even Clifford map yields a
discriminant-preserving bĳection

Quadratic modules in 𝑉 Quaternion orders in 𝐵
↔
up to twisted similarity up to isomorphism
that is functorial with respect to 𝑅.
By functorial with respect to 𝑅, we mean the same thing as in Corollary 5.2.6, but
with respect to any homomorphism 𝑅 → 𝑆 of Dedekind domains. In particular, the bi-
jection in Theorem 22.1.5 is compatible with the bĳections obtained over localizations
of 𝑅, including the bĳection over 𝐹 between quaternion algebras and nondegener-
ate ternary quadratic forms previously obtained. In the language of quadratic forms
(Definition 9.7.13), after some additional work (nailing down the difference between
similarity and isometry), we conclude: if the ternary quadratic module 𝑄 corresponds
to the quaternion order O, then there is a bĳection
Cl 𝑄 ↔ Typ O, (22.1.6)
i.e. the type number of a quaternion order is the same as the class number of the
corresponding ternary quadratic form.
Remark 22.1.7. If we restrict the correspondence to primitive modules 𝑄 : 𝑀 → 𝔩 (i.e.,
𝑄(𝑀) = 𝔩), then we need only remember the underlying lattice 𝑀, and on the right-
hand side we obtain precisely the Gorenstein orders; these orders will be introduced
in 24.1.1 and this correspondence is proven in section 24.2.
22.2 Even Clifford algebras

In this section, we construct the even Clifford algebra associated to a quadratic module:
see Remark 22.2.14 below for further references. The reader who wants to skip over
technicalities at first is encouraged to skip this section and accept 22.3.2 as a definition.
Let 𝑅 be a noetherian domain with 𝐹 = Frac 𝑅. Let 𝑄 : 𝑀 → 𝐿 be a quadratic
module over 𝑅 (see section 9.7), so that 𝑀 is a projective 𝑅-module of finite rank
and 𝐿 is an invertible 𝑅-module (rank 1). Write 𝐿 ∨ := Hom𝑅 (𝐿, 𝑅) and 𝑀 ⊗0 = 𝑅.
(For further reference on tensor algebra, see Matsumura [Mat89, Appendix C] or
Curtis–Reiner [CR81, §12].)
22.2.1. Let
∞
Ê
Ten0 (𝑀; 𝐿) := (𝑀 ⊗ 𝑀 ⊗ 𝐿 ∨ ) ⊗𝑑 .
𝑑=0
0
Now Ten (𝑀; 𝐿) has a natural tensor multiplication law (rearranging tensors), so
Ten0 (𝑀; 𝐿) is a graded 𝑅-algebra. Let 𝐼 0 (𝑄) be the two-sided ideal of Ten0 (𝑀; 𝐿)
defined by
𝐼 0 (𝑄) := h𝑥 ⊗ 𝑥 ⊗ 𝑔 − 𝑔(𝑄(𝑥)) : 𝑥 ∈ 𝑀, 𝑔 ∈ 𝐿 ∨ i ⊆ Ten0 (𝑀; 𝐿); (22.2.2)
note that 𝑄(𝑥) ∈ 𝐿 so 𝑔(𝑄(𝑥)) ∈ 𝑅. We define the even Clifford algebra of 𝑄 to be

the quotient
Clf 0 (𝑄) = Ten0 (𝑀; 𝐿)/𝐼 0 (𝑄). (22.2.3)
Remark 22.2.4. We might try to define

∞
Ê
Ten(𝑀; 𝐿) := 𝑀 ⊗𝑑 ⊗ (𝐿 ∨ ) ⊗ b𝑑/2c = 𝑅 ⊕ 𝑀 ⊕ (𝑀 ⊗ 𝑀 ⊗ 𝐿 ∨ ) ⊕ . . . ;
𝑑=0
unfortunately, Ten(𝑀; 𝐿) does not have a natural tensor multiplication law, because
there is no natural map 𝑀 ⊗ 𝑀 → 𝑀 ⊗ 𝑀 ⊗ 𝐿 ∨ . But see 22.2.16 below for the odd
part.
Example 22.2.5. Under the inclusion 𝑅 ↩→ 𝐹, we have a natural identification
Clf 0 (𝑄) ⊗𝑅 𝐹 Clf 0 (𝑄 𝐹 ). (22.2.6)
We conclude that the 𝑅-lattice in Clf 0 (𝑄 𝐹 ) defined by the image of 𝑅 3 is closed under
multiplication—something that may also be verified directly—and so Clf 0 (𝑄) is an
𝑅-order in Clf 0 (𝑄 𝐹 ).
22.2.7. As in 5.3.7, for all 𝑥, 𝑦 ∈ 𝑀 and 𝑔 ∈ 𝐿 ∨ , the calculation
𝑥 ⊗ 𝑦 ⊗ 𝑔 + 𝑦 ⊗ 𝑥 ⊗ 𝑔 = 𝑔(𝑇 (𝑥, 𝑦)) ∈ 𝑅 (22.2.8)
holds in Clf 0 (𝑄), where 𝑇 (𝑥, 𝑦) = 𝑄(𝑥 + 𝑦) − 𝑄(𝑥) − 𝑄(𝑦) ∈ 𝐿.

22.2. EVEN CLIFFORD ALGEBRAS 347
22.2.9. If 𝑀 ' 𝑅 𝑛 is free with basis 𝑒 1 , . . . , 𝑒 𝑛 and 𝐿 = 𝑅𝑔 is free, then Clf 0 (𝑄) is a
free 𝑅-module with basis
𝑒 𝑖1 ⊗ · · · ⊗ 𝑒 𝑖𝑑 ⊗ 𝑔 −⊗𝑑/2 , 1 ≤ 𝑒 𝑖1 < · · · < 𝑒 𝑖𝑑 ≤ 𝑛, 𝑑 even,
as a consequence of 22.2.8, just as in the case over fields 5.3.9. In particular, by

localizing, if 𝑀 has rank 𝑛 as an 𝑅-module, then Clf 0 (𝑄) is projective of rank 2𝑛−1 as
an 𝑅-module. We write elements of Clf 0 (𝑄) without tensors, for brevity.
22.2.10. The reversal map defined on simple tensors
rev : Clf 0 (𝑄) → Clf 0 (𝑄)

𝑥1 ⊗ · · · ⊗ 𝑥 𝑑 ⊗ (𝑔1 · · · 𝑔 𝑑/2 ) ↦→ 𝑥 𝑑 ⊗ · · · ⊗ 𝑥 1 ⊗ (𝑔1 · · · 𝑔 𝑑/2 )
for 𝑥 𝑖 ∈ 𝑀 and 𝑔𝑖 ∈ 𝐿 ∨ , and extended 𝑅-linearly, is an 𝑅-linear involution.
Theorem 22.2.11. The association 𝑄 ↦→ Clf 0 (𝑄) is a functor from the category of
quadratic 𝑅-modules under similarities to the category of projective 𝑅-algebras with
involution under isomorphism. Moreover, this association is functorial with respect to
𝑅.
We call the association 𝑄 ↦→ Clf 0 (𝑄) in Theorem 22.2.11 the even Clifford
functor.
22.2.12. The statement “functorial with respect to 𝑅” means the following: given
a ring homomorphism 𝑅 → 𝑆, there is a natural transformation between the even
Clifford functors over 𝑅 and 𝑆. Explicitly, given a ring homomorphism 𝑅 → 𝑆 and a
quadratic module 𝑄 : 𝑀 → 𝐿, we have a quadratic module 𝑄 𝑆 : 𝑀 ⊗𝑅 𝑆 → 𝐿 ⊗𝑅 𝑆,
and Clf 0 (𝑄) ⊗𝑅 𝑆 Clf 0 (𝑄 𝑆 ) in a way compatible with morphisms in each category.
In particular, this recovers the identification in Example 22.2.5 arising from 𝑅 ↩→ 𝐹.
Remark 22.2.13. The association 𝑄 ↦→ Clf (𝑄) of the full Clifford algebra is a functor
from the category of quadratic 𝑅-modules under isometries to the category of 𝑅-
algebras with involution under isomorphism that is also functorial with respect to 𝑅.
See Bischel–Knus [BK94].
Proof of Theorem 22.2.11. The construction in 22.2.1 yields an 𝑅-algebra that is pro-
jective as an 𝑅-module; we need to define an association on the level of morphisms.
∼ 𝑀0
Let 𝑄 0 : 𝑀 0 → 𝐿 0 be a quadratic module and ( 𝑓 , ℎ) be a similarity with 𝑓 : 𝑀 −→
∼ 0 0
and ℎ : 𝐿 −→ 𝐿 satisfying 𝑄 ( 𝑓 (𝑥)) = ℎ(𝑄(𝑥)). We mimic the proof of Lemma
5.3.21. We define a map via
Ten0 (𝑀; 𝐿) → Ten0 (𝑀 0; 𝐿 0)

𝑥 ⊗ 𝑦 ⊗ 𝑔 ↦→ 𝑓 (𝑥) ⊗ 𝑓 (𝑦) ⊗ (ℎ−1 ) ∗ (𝑔)
for 𝑥, 𝑦 ∈ 𝑀 and 𝑔 ∈ 𝐿 ∨ and extending multiplicatively, where
(ℎ−1 ) ∗ (𝑔) := 𝑔 ◦ ℎ−1 : 𝐿 0 → 𝑅

is the pullback under ℎ−1 . Then
𝑥 ⊗ 𝑥 ⊗ 𝑔 − 𝑔(𝑄(𝑥)) ↦→ 𝑓 (𝑥) ⊗ 𝑓 (𝑥) ⊗ (ℎ−1 ) ∗ (𝑔) − 𝑔(𝑄(𝑥))
and since
𝑔(𝑄(𝑥)) = 𝑔(ℎ−1 (𝑄 0 ( 𝑓 (𝑥)))) = (ℎ−1 ) ∗ (𝑔) (𝑄 0 ( 𝑓 (𝑥))),
we conclude that 𝐼 0 (𝑄) is mapped to 𝐼 0 (𝑄 0). Repeating with the inverse similarity
( 𝑓 −1 , ℎ−1 ), and composing to get the identity, we conclude that the induced map
Clf 0 (𝑄) → Clf 0 (𝑄 0) is an 𝑅-algebra isomorphism.
Functoriality in the sense of 22.2.12 then follows directly.
Remark 22.2.14. In his thesis, Bichsel [Bic85] constructed an even Clifford algebra of
a line bundle-valued quadratic form on an affine scheme using faithfully flat descent.
A related and more general construction was given by Bischel–Knus [BK94]. Several
other constructions are available: see Auel [Auel2011, §1.8] and the references therein.
The direct tensorial construction given above is given for ternary quadratic modules
by Voight [Voi2011a, (1.10)] and in general by Auel [Auel2011, §1.8] and with further
detail in Auel [Auel2015, §1.2]; for a comparison of this direct construction with
others, see Auel–Bernardara–Bolognesi [ABB2014, §1.5, Appendix A].
Remark 22.2.15. Allowing the quadratic forms to take values in a invertible module
is essential for what follows and for many other purposes: for an overview, see the
introduction to Auel [Auel2011].
22.2.16. Let
∞
Ê
Ten1 (𝑀; 𝐿) := 𝑀 ⊗𝑑 ⊗ (𝐿 ∨ ) ⊗ b𝑑/2c = 𝑀 ⊕ (𝑀 ⊗ 𝑀 ⊗ 𝑀 ⊗ 𝐿 ∨ ) ⊕ . . . .
𝑑=1
𝑑 odd
Then Ten1 (𝑀; 𝐿) is a graded Ten0 (𝑀; 𝐿)-bimodule under the natural tensor multi-
plication. Let 𝐼 1 (𝑄) be the 𝑅-submodule of Ten1 (𝑀; 𝐿) generated by the image of
multiplication of 𝐼 0 (𝑄) by 𝑀 on the left and right: then 𝐼 1 (𝑄) is the Ten0 (𝑀; 𝐿)-
bisubmodule generated by the set of elements of the form
𝑥 ⊗ 𝑥 ⊗ 𝑦 ⊗ 𝑔 − 𝑔(𝑄(𝑥))𝑦 and 𝑦 ⊗ 𝑥 ⊗ 𝑥 ⊗ 𝑔 − 𝑔(𝑄(𝑥))𝑦
with 𝑥, 𝑦 ∈ 𝑀 and 𝑔 ∈ 𝐿 ∨ .
We define the odd Clifford bimodule as
Clf 1 (𝑄) := Ten1 (𝑀; 𝐿)/𝐼 1 (𝑄).
Visibly, Clf 1 (𝑄) is a bimodule for the even Clifford algebra Clf 0 (𝑄).
22.2.17. When 𝐿 = 𝑅, we can combine the construction of the even Clifford algebra
and its odd Clifford bimodule to construct a full Clifford algebra, just as in section
5.3 over a field: see Exercise 22.7. This direct tensorial construction does not extend
in an obvious way when 𝐿 ≠ 𝑅, as we would need to define a multiplication map
𝑀 ⊗ 𝑀 → 𝑀 ⊗ 𝑀 ⊗ 𝐿∨.
22.3. EVEN CLIFFORD ALGEBRA OF A TERNARY QUADRATIC MODULE 349
22.2.18. We will employ exterior calculus in what follows: this is a convenient method
for keeping track of our module maps in a general setting. Let 𝑀 be an 𝑅-module and
let 𝑟 ≥ 1. The 𝑟th exterior power of 𝑀 (over 𝑅) is
𝑀 := 𝑀 ⊗𝑟 /𝐸𝑟
Ó𝑟
where 𝐸𝑟 is the 𝑅-module
𝐸𝑟 := h𝑥 1 ⊗ · · · ⊗ 𝑥𝑟 : 𝑥1 , . . . , 𝑥𝑟 ∈ 𝑀 and 𝑥𝑖 = 𝑥 𝑗 for some 𝑖 ≠ 𝑗i.

⊗𝑟 in
Ó0
𝑀 = 𝑅 (and 1 𝑀 = 𝑀). The image of 𝑥1 ⊗ · · · ⊗ 𝑥Ó
Ó Ó𝑟
We let 𝑟 ∈ 𝑀 𝑀 is
written𝑥1 ∧ · · · ∧ 𝑥𝑟 . If 𝑀 is projective of rank 𝑛 over 𝑅, then 𝑟 𝑀 is projective of
rank 𝑟𝑛 .
22.3 Even Clifford algebra of a ternary quadratic module

Now suppose that 𝑄 : 𝑀 → 𝐿 is a ternary quadratic module, which is to say 𝑀 has
rank 3; in this section, we examine its even Clifford algebra Clf 0 (𝑄). Recall that an
𝑅-order is projective if it is projective as an 𝑅-module. The main result of this section
is as follows.
Theorem 22.3.1. Let 𝑅 be a noetherian domain. Then the association 𝑄 ↦→ Clf 0 (𝑄)
gives a functor from the category of
nondegenerate ternary quadratic modules over 𝑅,
under similarities
to the category of
projective quaternion orders over 𝑅, under isomorphisms.
In the previous section, we defined the even Clifford functor, whose codomain was
the category of projective 𝑅-algebras; in this section, we show that the restriction to
nondegenerate ternary quadratic modules lands in projective quaternion orders.
We begin with some explicit descriptions.
22.3.2. By 22.2.9, the even Clifford algebra Clf 0 (𝑄) is an 𝑅-algebra that is projective
of rank 4 as an 𝑅-module. Explicitly, as an 𝑅-module we have
𝑅 ⊕ 𝑀 ⊗ 𝑀 ⊗ 𝐿∨

0
Clf (𝑄) ' (22.3.3)
𝐼 0 (𝑄)
where 𝐼 0 (𝑄) is the 𝑅-submodule generated by elements of the form
𝑥 ⊗ 𝑥 ⊗ 𝑔 − 1 ⊗ 𝑔(𝑄(𝑥))
for 𝑥 ∈ 𝑀 and 𝑔 ∈ 𝐿 ∨ .
We now explicitly give the even Clifford algebra of a ternary quadratic module in
the free case; this could also be taken as the definition when 𝑅 is a PID and 𝑀 = 𝑅 3 .
22.3.4. Let 𝑀 = 𝑅 3 with standard basis 𝑒 1 , 𝑒 2 , 𝑒 3 be equipped with the quadratic form
𝑄 : 𝑀 → 𝑅 defined by
𝑄(𝑥, 𝑦, 𝑧) = 𝑄(𝑥𝑒 1 + 𝑦𝑒 2 + 𝑧𝑒 3 ) = 𝑎𝑥 2 + 𝑏𝑦 2 + 𝑐𝑧2 + 𝑢𝑦𝑧 + 𝑣𝑥𝑧 + 𝑤𝑥𝑦, (22.3.5)
with 𝑎, 𝑏, 𝑐, 𝑢, 𝑣, 𝑤 ∈ 𝑅. Then
𝑁 := disc(𝑄) = 4𝑎𝑏𝑐 + 𝑢𝑣𝑤 − 𝑎𝑢 2 − 𝑏𝑣 2 − 𝑐𝑤 2 ∈ 𝑅/𝑅 × . (22.3.6)
By 22.2.9, we have
Clf 0 (𝑄) = 𝑅 ⊕ 𝑅𝑖 ⊕ 𝑅 𝑗 ⊕ 𝑅𝑘
where
𝑖 := 𝑒 2 𝑒 3 , 𝑗 := 𝑒 3 𝑒 1 , 𝑘 := 𝑒 1 𝑒 2 .
The reversal involution acts by
𝑖 = 𝑒 3 𝑒 2 = 𝑇 (𝑒 2 , 𝑒 3 ) − 𝑖 = 𝑢 − 𝑖,
and similarly 𝑗 = 𝑣 − 𝑗 and 𝑘 = 𝑤 − 𝑘 by (22.2.8).

We then compute directly the multiplication table:
𝑖 2 = 𝑢𝑖 − 𝑏𝑐 𝑗 𝑘 = 𝑎𝑖
2
𝑗 = 𝑣 𝑗 − 𝑎𝑐 𝑘𝑖 = 𝑏 𝑗 (22.3.7)
2
𝑘 = 𝑤𝑘 − 𝑎𝑏 𝑖 𝑗 = 𝑐𝑘
For example,
𝑖 2 = (𝑒 2 𝑒 3 ) (𝑒 2 𝑒 3 ) = 𝑒 2 (𝑒 3 𝑒 2 )𝑒 3 = 𝑒 2 (𝑢 − 𝑒 2 𝑒 3 )𝑒 3 = 𝑢𝑒 2 𝑒 3 − 𝑒 22 𝑒 23 = 𝑢𝑖 − 𝑏𝑐
and
𝑗 𝑘 = (𝑒 3 𝑒 1 ) (𝑒 1 𝑒 2 ) = 𝑎𝑒 3 𝑒 2 = 𝑎𝑖.
The remaining multiplication laws can be computed in the same way, or by using the
reversal involution and (22.3.7): we compute
𝑎𝑖 = 𝑗 𝑘 = 𝑘 𝑗 = (𝑤 − 𝑘) (𝑣 − 𝑗) = 𝑣𝑤 − 𝑤 𝑗 − 𝑣𝑘 + 𝑘 𝑗
so 𝑘 𝑗 = −𝑣𝑤 + 𝑎𝑖 + 𝑤 𝑗 + 𝑣𝑘. By symmetry, we find:
𝑘 𝑗 = −𝑣𝑤 + 𝑎𝑖 + 𝑤 𝑗 + 𝑣𝑘
𝑖𝑘 = −𝑢𝑤 + 𝑤𝑖 + 𝑏 𝑗 + 𝑢𝑘 (22.3.8)
𝑗𝑖 = −𝑢𝑣 + 𝑣𝑖 + 𝑢 𝑗 + 𝑐𝑘
We note also the formulas
𝑖 𝑗 𝑘 = 𝑗 𝑘𝑖 = 𝑘𝑖 𝑗 = 𝑎𝑏𝑐. (22.3.9)
Example 22.3.10. It is clarifying to work out the diagonal case. Let 𝐵 = (𝑎, 𝑏 | 𝐹)
with 𝑎, 𝑏 ∈ 𝑅, and let
O = 𝑅h𝑖, 𝑗i = 𝑅 + 𝑅𝑖 + 𝑅 𝑗 + 𝑅𝑖 𝑗 ⊂ 𝐵
be the 𝑅-order generated by the standard generators, and let 𝑘 = 𝑖 𝑗. Then:
𝑖2 = 𝑎 𝑗 𝑘 = 𝑏𝑖 = −𝑏𝑖
𝑗2 = 𝑏 𝑘𝑖 = 𝑎 𝑗 = −𝑎 𝑗 (22.3.11)
2
𝑘 = −𝑎𝑏 𝑖 𝑗 = −𝑘 = 𝑘.
Example 22.3.12. Consider
𝑄(𝑥, 𝑦, 𝑧) = 𝑥𝑦 − 𝑧 2
so (𝑎, 𝑏, 𝑐, 𝑢, 𝑣, 𝑤) = (0, 0, −1, 0, 0, 1); then disc(𝑄) = −𝑐𝑤 2 = 1. Then the even
Clifford algebra Clf 0 (𝑄) = 𝑅 + 𝑅𝑖 + 𝑅 𝑗 + 𝑅𝑘 has multiplication table
𝑖2 = 0 𝑗𝑘 = 0
2
𝑗 =0 𝑘𝑖 = 0 (22.3.13)
𝑘2 = 𝑘 𝑖 𝑗 = −𝑘 = 𝑘 − 1.
We find an isomorphism of 𝑅-algebras
∼ 𝑀 (𝑅)
Clf 0 (𝑄) −→ 2
(22.3.14)

0 1 0 0 0 0
𝑖, 𝑗, 𝑘 ↦→ , , .
0 0 −1 0 0 1
22.3.15. Returning to the free quadratic form 22.3.4, the group GL3 (𝑅) acts naturally
on 𝑀 by change of basis, and this induces an action on Clf 0 (𝑄) by 𝑅-algebra auto-
morphism by functoriality. Explicitly, for 𝜌 ∈ GL3 (𝑅), the action on the basis 𝑖, 𝑗, 𝑘
is by the adjugate adj(𝜌) of 𝜌, the 3 × 3 matrix whose entries are the 2 × 2 minors of
𝜌. The verification is requested in Exercise 22.2.
22.3.16. Let 𝐹 = Frac 𝑅. By base extension, we have a quadratic form 𝑄 𝐹 : 𝑉 → 𝐹
where 𝑉 = 𝑀 ⊗𝑅 𝐹, and by functoriality 22.2.12 with respect to the inclusion 𝑅 ↩→ 𝐹,
we have an inclusion Clf 0 (𝑄) ↩→ Clf 0 (𝑄 𝐹 ) realizing Clf 0 (𝑄) as an 𝑅-order in the
𝐹-algebra Clf 0 (𝑄 𝐹 ).
Lemma 22.3.17. The reversal involution is a standard involution on Clf 0 (𝑄 𝐹 ).
Proof. To check that the involution is standard, we could appeal to Exercise 3.19, but
we find it more illustrative to exhibit the involution on a universal element, yielding a
rather beautiful formula. We choose a basis for 𝑉 and work with the presentation for
Clf 0 (𝑄 𝐹 ) as in 22.3.4.
Let 𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 with 𝑡, 𝑥, 𝑦, 𝑧 ∈ 𝐹. Then 𝛼 = 2𝑡 + 𝑢𝑥 + 𝑣𝑦 + 𝑤𝑧 − 𝛼, and
we find that
𝛼2 − (𝛼 + 𝛼)𝛼 + 𝛼𝛼 = 𝛼2 − 𝜏(𝛼)𝛼 + 𝜈(𝛼) = 0
where
𝜏(𝛼) = 2𝑡 + 𝑢𝑥 + 𝑣𝑦 + 𝑤𝑧
𝜈(𝛼) = 𝑡 2 + 𝑢𝑡𝑥 + 𝑣𝑡𝑦 + 𝑤𝑡𝑧
(22.3.18)
+ 𝑏𝑐𝑥 2 + (𝑢𝑣 − 𝑐𝑤)𝑥𝑦 + (𝑢𝑤 − 𝑏𝑣)𝑥𝑧
+ 𝑎𝑐𝑦 2 + (𝑣𝑤 − 𝑎𝑢)𝑦𝑧 + 𝑎𝑏𝑧 2
so that the reversal map 𝛼 ↦→ 𝜏(𝛼) − 𝛼 defines a standard involution.
discrd(Clf 0 (𝑄)) = disc(𝑄)𝑅.
Proof. The construction of the even Clifford algebra is functorial with respect to
localization, and the statement itself is local, so we may suppose that 𝑀 = 𝑅 3 , 𝐿 = 𝑅
are free with the presentation for O = Clf 0 (𝑄) as in 22.3.4.
We refer to section 15.4 and Lemma 15.4.7: we compute
𝑚(𝑖, 𝑗, 𝑘) = trd((𝑖 𝑗 − 𝑗𝑖)𝑘)

= trd(−2𝑎𝑏𝑐 + 𝑎𝑢 2 + 𝑐𝑤 2 − 𝑎𝑢𝑖 + (𝑏𝑣 − 𝑢𝑤) 𝑗 − 𝑐𝑤𝑘) (22.3.20)
2 2 2
= −4𝑎𝑏𝑐 + 𝑎𝑢 + 𝑐𝑤 − 𝑢𝑣𝑤 + 𝑏𝑣 = − disc(𝑄)
and discrd(O) = 𝑚(𝑖, 𝑗, 𝑘)𝑅 as claimed.

Alternatively, we compute directly that
2 𝑢 𝑣 𝑤
𝑢 𝑢 2 − 2𝑏𝑐
© ª
𝑐𝑤 𝑏𝑣 ®
𝑑 (1, 𝑖, 𝑗, 𝑘) = 2
𝑣 − 2𝑎𝑐
®
𝑣 𝑐𝑤 𝑎𝑢 ® (22.3.21)
2
𝑤 − 2𝑎𝑏 ¬
«𝑤 𝑏𝑣 𝑎𝑢
= −(4𝑎𝑏𝑐 + 𝑢𝑣𝑤 − 𝑎𝑢 − 𝑏𝑣 − 𝑐𝑤 2 ) 2 = − disc(𝑄) 2
2 2
so disc(O) = disc(𝑄) 2 𝑅, and the result follows by taking square roots (as ideals).
Corollary 22.3.22. If 𝑄 is nondegenerate, then Clf 0 (𝑄) is an 𝑅-order in the quaternion

algebra 𝐵 = Clf 0 (𝑄 𝐹 ).
Proof. The standard involution has discriminant disc(nrd) = disc(O) 2 = disc(𝑄 𝐹 ) ≠

0; the result then follows from the characterization of algebras with nondegenerate
standard involution (Main Theorem 4.4.1 and Theorem 6.4.1).
Remark 22.3.23. Corollary 22.3.22 gives a characteristic independent proof of the

fact that the even Clifford algebra of a nondegenerate ternary quadratic form over 𝐹
is a quaternion algebra over 𝐹: we proved this in 5.3.23 and Exercise 6.10 (when
char 𝐹 = 2).
Intermediate between the general abstract definition and the explicit description in
the free case is the situation where the modules are completely decomposable, and we
can work with a pseudobasis.
Example 22.3.24. Let 𝑅 be a Dedekind domain. Then we can write
𝑀 = 𝔞𝑒 1 ⊕ 𝔟𝑒 2 ⊕ 𝔠𝑒 3 and 𝐿=𝔩
for fractional ideals 𝔞, 𝔟, 𝔠, 𝔩. Let 𝑉 = 𝑀 ⊗𝑅 𝐹 ' 𝐹 3 with basis 𝑒 1 , 𝑒 2 , 𝑒 3 , so then

𝑀 ↩→ 𝑉 is a ternary 𝑅-lattice. Then we may take 𝑄 𝐹 : 𝑉 → 𝐹 to have the form
22.3.5, and Clf 0 (𝑄 𝐹 ) = 𝐵 is a quaternion algebra with O := Clf 0 (𝑄) ⊆ 𝐵 an 𝑅-order.
Extending the description in 22.3.4, we find that
O = 𝑅 ⊕ 𝔟𝔠𝔩−1 𝑖 ⊕ 𝔞𝔠𝔩−1 𝑗 ⊕ 𝔞𝔟𝔩−1 𝑘 (22.3.25)
where 𝑖, 𝑗, 𝑘 satisfy the multiplication table (22.3.7). We can verify directly that O is
closed under multiplication: for example, if 𝛼 ∈ 𝔟𝔠𝔩−1 so 𝛼𝑖 ∈ O, then
(𝛼𝑖) 2 = 𝑢𝛼𝑖 − 𝛼2 𝑏𝑐 ∈ O
since 𝑄(𝔟𝑒 2 ) = 𝔟2 𝑄(𝑒 2 ) ⊆ 𝔩 so 𝑏 = 𝑄(𝑒 2 ) ∈ 𝔩𝔟−2 and therefore
𝛼2 𝑏𝑐 ∈ (𝔟𝔠𝔩−1 ) 2 (𝔩𝔟−2 ) (𝔩𝔠−2 ) = 𝑅.

√ √
Example 22.3.26.√ Let 𝐹 = Q( 10) and 𝑅 = Z𝐹 = Q[ 10] be the ring of integers.
Then 𝔭 = (3, 4 + 10) is a prime ideal over 3 that is not principal.
Let 𝑄 : 𝑀 = 𝑅 3 → 𝔭 be the quadratic module
√
𝑄(𝑥, 𝑦, 𝑧) = 3𝑥 2 + 3𝑦 2 + (4 + 10)𝑧 2 .
We have 𝔭 = 𝑄(𝑅 3 ). The even Clifford algebra is then
O = Clf 0 (𝑄) = 𝑅 + 𝔭−1𝑖 + 𝔭−1 𝑗 ⊕ 𝔭−1 𝑘
with the multiplication law

√
𝑖 2 = −3(4 + 10) 𝑗 𝑘 = 3𝑖
√
𝑗 2 = −3(4 + 10) 𝑘𝑖 = 3 𝑗 (22.3.27)
2
√
𝑘 = −9 𝑖 𝑗 = (4 + 10)𝑘.
We have √ √ √
discrd(O) = 4(9) (4 + 10)𝔭−3 = (2, 10) 5 (3, 2 + 10) 2
and in particular 𝔭 - discrd(O), and
√ √
−3(4 + 10), −3(4 + 10)
O⊂𝐵=
𝐹
√
with disc 𝐵 = (2 + 𝑤)𝑅, so Ram 𝐵 = {(2, 10), 𝔭, ∞1 , ∞2 } where ∞1 , ∞2 are the two
real places of 𝐹.
22.4 Over a PID

In the previous two sections, we observed that the construction of the even Clifford
algebra gives a functorial association from nondegenerate ternary quadratic modules
to quaternion orders. In this section, we show that this functor gives a bĳection on
classes over a PID, following Gross–Lucianovic [GrLu2009, §4].
Main Theorem 22.4.1. Suppose that 𝑅 is a PID. Then the association 𝑄 ↦→ Clf 0 (𝑄)
induces a discriminant-preserving bĳection

Nondegenerate ternary quadratic Quaternion orders over 𝑅
↔ (22.4.2)
forms over 𝑅 up to similarity up to isomorphism
that is functorial with respect to 𝑅.
Remark 22.4.3. The bĳection can also be rephrased in terms of the orbits of a group
(following Gross–Lucianovic [GrLu2009]). The group GL3 (𝑅) has a natural twisted
action on quadratic forms by (𝑔𝑄) (𝑥, 𝑦, 𝑧) = (det 𝑔) (𝑄(𝑔 −1 (𝑥, 𝑦, 𝑧) t )), i.e., the usual
action with an extra scalingÓfactor of det 𝑔 ∈ 𝑅 × . This is the natural action on the
𝑅-module Ó Sym2 ((𝑅 3 ) ∨ ) ⊗ 3 𝑅 3 , or equivalently on the set of quadratic modules
3
𝑄 : 𝑅 → 3 𝑅 3 . Main Theorem 22.1.1 states that the nondegenerate orbits of this
action are in functorial bĳection with the set of isomorphism classes of quaternion
orders over 𝑅.
We prove this theorem in a few steps. Throughout this section, let 𝑅 be a PID.
First, we prove that the map (22.4.2) is surjective, or equivalently that the even
Clifford functor is essentially surjective from the category of nondegenerate ternary
quadratic forms to the category of quaternion orders.
Proposition 22.4.4. Every quaternion 𝑅-order is isomorphic to the even Clifford

algebra of a nondegenerate ternary quadratic form.
Proof. We work explicitly with the multiplication table, hoping to make it look like
(22.3.7).
Let O be a quaternion 𝑅-order. Since 𝑅 is a PID, O is free as an 𝑅-module. We
need a slight upgrade from this, a technical result supplied by Exercise 22.1: in fact,
O has an 𝑅-basis containing 1.
So let 1, 𝑖, 𝑗, 𝑘 be an 𝑅-basis for O. Since every element of O is integral over 𝑅,
satisfying its reduced characteristic polynomial of degree 2 over 𝑅, we have
𝑖 2 = 𝑢𝑖 + 𝑙
𝑗2 = 𝑣 𝑗 + 𝑚
𝑘 2 = 𝑤𝑘 + 𝑛
for some 𝑙, 𝑚, 𝑛, 𝑢, 𝑣, 𝑤 ∈ 𝑅. The product 𝑗 𝑘 = 𝑟 − 𝑎𝑖 + 𝑞 𝑗 + 𝛼𝑘 can be written as an

𝑅-linear combination of 1, 𝑖, 𝑗, 𝑘, with 𝑞, 𝑟, 𝑎, 𝛼 ∈ 𝑅. Letting 𝑘 0 := 𝑘 − 𝑞, we have
𝑗 𝑘 0 = 𝑗 (𝑘 − 𝑞) = 𝑟 − 𝑎𝑖 + 𝛼𝑘 = (𝑟 + 𝛼𝑞) − 𝑎𝑖 + 𝛼𝑘 0 .
22.4. OVER A PID 355
So changing the basis, we may suppose 𝑗 𝑘 is an 𝑅-linear combination of 1, 𝑖, 𝑘 (no 𝑗

term). By symmetry, in the product 𝑘𝑖 we may suppose that the coefficient of 𝑘 is zero
and in 𝑖 𝑗 the coefficient of 𝑖 is zero. Therefore:
𝑗 𝑘 = 𝑟 − 𝑎𝑖 + 𝛼𝑘
𝑘𝑖 = 𝑠 − 𝑏 𝑗 + 𝛽𝑖
𝑖 𝑗 = 𝑡 − 𝑐𝑘 + 𝛾 𝑗
As before, the other products can be calculated using the standard involution: for
example, we have
𝑖𝑘 + 𝑘𝑖 = − trd(𝑘𝑖) + trd(𝑘)𝑖 + trd(𝑖)𝑘

= − trd(𝑘 (𝑢 − 𝑖)) + 𝑤𝑖 + 𝑢𝑘
= (−𝑢𝑤 + 2𝑠 − 𝑏𝑣 + 𝛽𝑢) + 𝑤𝑖 + 𝑢𝑘
so
𝑖𝑘 = (𝑠 + 𝛽𝑢 − 𝑏𝑣 − 𝑢𝑤) + (𝑤 − 𝛽)𝑖 + 𝑏 𝑗 + 𝑢𝑘. (22.4.5)
But now from these multiplication laws, we compute that the trace of left multiplication
𝑖 is Tr(𝑖) = 0 + 𝑢 + 𝛾 + 𝑢 = 2𝑢 + 𝛾. But in a quaternion algebra, we have Tr(𝑖) =
2 trd(𝑖) = 2𝑢, so we must have 𝛾 = 0. By symmetry, we find that 𝛼 = 𝛽 = 0. Finally,
associativity implies relations on the structure constants in the multiplication table:
we have
𝑗 (𝑘 𝑘) = ( 𝑗 𝑘)𝑘
−𝑛 𝑗 = (𝑟 − 𝑎𝑖) (𝑤 − 𝑘) = 𝑟𝑤 − 𝑎𝑤𝑖 − 𝑟 𝑘 + 𝑎𝑖𝑘 (22.4.6)
−𝑛 𝑗 = (𝑟𝑤 + 𝑎𝑠 − 𝑎𝑏𝑣 − 𝑎𝑢𝑤) + 𝑎𝑏 𝑗 + (𝑎𝑢 − 𝑟)𝑘
using (22.4.5) with 𝛽 = 0; so equality of coefficients of 𝑗, 𝑘 implies 𝑟 = 𝑎𝑢 and
𝑛 = −𝑎𝑏. By symmetry, we find 𝑠 = 𝑏𝑣, 𝑡 = 𝑐𝑤 and 𝑚 = −𝑎𝑐, 𝑛 = −𝑎𝑏, so we have
the following multiplication table:
𝑖 2 = 𝑢𝑖 − 𝑏𝑐 𝑗 𝑘 = 𝑎𝑖
2
𝑗 = 𝑣 𝑗 − 𝑎𝑐 𝑘𝑖 = 𝑏 𝑗
𝑘 2 = 𝑤𝑘 − 𝑎𝑏 𝑖 𝑗 = 𝑐𝑘
This matches precisely the multiplication table (22.3.7) for the even Clifford algebra
of the quadratic form 𝑄(𝑥, 𝑦, 𝑧) = 𝑎𝑥 2 + 𝑏𝑦 2 + 𝑐𝑧2 + 𝑢𝑦𝑧 + 𝑣𝑥𝑧 + 𝑤𝑥𝑦.
22.4.7. More generally, if 𝑅 is a domain and O is a quaternion 𝑅-order such that O
is free as an 𝑅-module with basis 1, 𝑖, 𝑗, 𝑘, then the proof of Proposition 22.4.4 shows
O has a basis 1, 𝑖, 𝑗, 𝑘 satisfying the multiplication laws (22.3.7) of an even Clifford
algebra; we call such a basis a good basis for O. Moreover, we have seen that given
a basis 1, 𝑖, 𝑗, 𝑘, there exist unique 𝜂(𝑖), 𝜂( 𝑗), 𝜂(𝑘) ∈ 𝑅 (in fact, certain coefficients of
the multiplication table) such that
1, 𝑖 − 𝜂(𝑖), 𝑗 − 𝜂( 𝑗), 𝑘 − 𝜂(𝑘)
is a good basis.
To conclude, we need to show that if two quaternion 𝑅-orders are isomorphic, then
they correspond to similar ternary quadratic forms. To this end, we define an inverse.
22.4.8. Let O ⊆ 𝐵 be a quaternion 𝑅-order with 𝑅-basis 1, 𝑖, 𝑗, 𝑘. Let 𝑁 ∈ 𝑅 be such
that (𝑁) = discrd(O); then 𝑁 ≠ 0 and is well-defined up to multiplication by 𝑅 × . Let
1♯ , 𝑖 ♯ , 𝑗 ♯ , 𝑘 ♯ be the dual basis (see 15.6.3); then trd(𝑖 ♯ ) = trd(1 · 𝑖 ♯ ) = 0 and similarly
for 𝑗 ♯ , 𝑘 ♯ , so
(O♯ ) 0 = {𝛼 ∈ O♯ : trd(𝛼) = 0} = 𝑅𝑖 ♯ + 𝑅 𝑗 ♯ + 𝑅𝑘 ♯ .
We define a candidate quadratic form
𝑁 nrd♯ (O) (𝑥, 𝑦, 𝑧) = 𝑁 nrd(𝑥𝑖 ♯ + 𝑦 𝑗 ♯ + 𝑧𝑘 ♯ ), (22.4.9)
well-defined up to similarity (along the way, we chose a basis and a generator for
discrd(O)).
Example 22.4.10. We return to Example 22.3.10. The 𝑅-order O has reduced dis-
criminant 𝑁 = 4𝑎𝑏. The (rescaled) dual basis is
𝑁𝑖 ♯ = 2𝑏𝑖, 𝑁 𝑗 ♯ = 2𝑎 𝑗, 𝑁 𝑘 ♯ = −2𝑘
and 𝑖 ♯ , 𝑗 ♯ , 𝑘 ♯ is a basis for (O♯ ) 0 ; thus

1
𝑁 nrd(𝑥𝑖 ♯ + 𝑦 𝑗 ♯ + 𝑧𝑘 ♯ ) = (−4𝑎𝑏 2 𝑥 2 − 4𝑎 2 𝑏𝑦 2 + 4𝑎𝑏𝑧2 ) = −𝑏𝑥 2 − 𝑎𝑦 2 + 𝑧 2 .
𝑁
Example 22.4.11. We return to Example 22.3.12. We have

♯ ♯ ♯ 0 0 0 −1 −1 0
𝑖 , 𝑗 ,𝑘 = , ,
1 0 0 0 0 1
and 𝑁 = 1 so

−𝑧 −𝑦
𝑁 nrd(𝑥𝑖 ♯ + 𝑦 𝑗 ♯ + 𝑧𝑘 ♯ ) = det = 𝑥𝑦 − 𝑧2 .
𝑥 𝑧
Proposition 22.4.12. If 𝑄 : 𝑅 3 → 𝑅 is a nondegenerate ternary quadratic form with

disc 𝑄 = 𝑁, then 𝑁 nrd♯ (Clf 0 (𝑄)) is similar to 𝑄. If O is a quaternion 𝑅-order with
discrd(O) = (𝑁), then 𝑁 nrd♯ (O) : 𝑅 3 → 𝑅 has Clf 0 (𝑁 nrd♯ (O)) ' O.
Proof. Proposition 22.4.4 shows that the even Clifford functor induces a surjective
map from similarity classes of nondegenerate ternary quadratic forms over 𝑅 to iso-
morphism classes of quaternion 𝑅-orders. If we prove the first statement, then the
second follows from set theory (and can be verified in a similar way).
We start with the quadratic form (22.3.5) with O satisfying the multiplication laws
(22.3.7). Let 𝑁 := disc(𝑄). We claim that
𝑁𝑖 ♯ = 𝑗 𝑘 − 𝑘 𝑗 = (𝑎𝑢 + 𝑣𝑤) − 2𝑎𝑖 − 𝑤 𝑗 − 𝑣𝑘

𝑁 𝑗 ♯ = 𝑘𝑖 − 𝑖𝑘 = (𝑏𝑣 + 𝑢𝑤) − 𝑤𝑖 − 2𝑏 𝑗 − 𝑢𝑘 (22.4.13)
𝑁 𝑘 ♯ = 𝑖 𝑗 − 𝑗𝑖 = (𝑐𝑤 + 𝑢𝑣) − 𝑣𝑖 − 𝑢 𝑗 − 2𝑐𝑘.
22.4. OVER A PID 357
We see that trd(𝑁𝑖 ♯ ) = 0 and the same with 𝑗 ♯ , 𝑘 ♯ . We recall the alternating trilinear
form 𝑚 (defined in 15.4.2). By (22.3.20) we have
𝑚(𝑖, 𝑗, 𝑘) = −𝑁 = trd(𝑖( 𝑗 𝑘 − 𝑘 𝑗)) = − trd(𝑖( 𝑗 𝑘 − 𝑘 𝑗)) = trd(𝑖(𝑁𝑖 ♯ ))
and
𝑚( 𝑗, 𝑗, 𝑘) = 0 = trd( 𝑗 ( 𝑗 𝑘 − 𝑘 𝑗)) = − trd( 𝑗 (𝑁𝑖 ♯ ))
and similarly trd(𝑘 (𝑁𝑖 ♯ )) = 0. The other equalities follow similarly, and this verifies
the dual basis (22.4.13). In particular, we have trd(𝑁𝑖 ♯ ) = trd(𝑁 𝑗 ♯ ) = trd(𝑁 𝑘 ♯ ) = 0.
We then compute the quadratic form on this basis and claim that
nrd(𝑁 (𝑥𝑖 ♯ + 𝑦 𝑗 ♯ + 𝑧𝑘 ♯ )) = 𝑁 (𝑎𝑥 2 + 𝑏𝑦 2 + 𝑐𝑧 2 + 𝑢𝑦𝑧 + 𝑣𝑥𝑧 + 𝑤𝑥𝑦)

(22.4.14)
= 𝑁𝑄(𝑥, 𝑦, 𝑧).
Indeed,
2 nrd(𝑁𝑖 ♯ ) = trd((𝑁𝑖 ♯ ) (𝑁𝑖 ♯ )) = − trd((𝑁𝑖 ♯ ) 2 ) = −(−2𝑎)𝑁 = 2𝑎𝑁

since only the term trd(𝑁𝑖 ♯ 𝑖) = 𝑁 is nonzero; and
trd(𝑁𝑖 ♯ 𝑁 𝑗 ♯ ) = − trd(𝑁𝑖 ♯ 𝑁 𝑗 ♯ ) = −𝑤𝑁.

The other equalities follow by symmetry. Then the claim (22.4.14) implies that
𝑁 nrd(𝑥𝑖 ♯ + 𝑦 𝑗 ♯ + 𝑧𝑘 ♯ ) = 𝑄(𝑥, 𝑦, 𝑧), as desired.
Corollary 22.4.15. Let O be a quaternion 𝑅-order. Then
O = 𝑅 + discrd(O) (O♯ ) 0 (O♯ ) 0 = 𝑅 + discrd(O)O♯ O♯ .
Proof. If we take the identifications in the proof of Proposition 22.4.12 working within
𝐵 ⊇ O, we see that Clf 0 (𝑁 nrd♯ (O)) is spanned over 𝑅 by the elements
1, 𝑁𝑖 ♯ 𝑗 ♯ , 𝑁 𝑗 ♯ 𝑘 ♯ , 𝑁 𝑘 ♯ 𝑖 ♯
where discrd(O) = (𝑁). In order to see that the other factors belong to this ring, we
compute
(𝑁𝑖 ♯ ) 2 = −𝑎𝑁
(𝑁 𝑗 ♯ ) 2 = −𝑏𝑁 (22.4.16)
♯ 2
(𝑁 𝑘 ) = −𝑐𝑁.
and
(𝑁 𝑗 ♯ ) (𝑁𝑖 ♯ ) = −𝑁 𝑘
(𝑁 𝑘 ♯ ) (𝑁 𝑗 ♯ ) = −𝑁𝑖 (22.4.17)
♯ ♯
(𝑁𝑖 ) (𝑁 𝑘 ) = −𝑁 𝑗 .
If we want to throw in the factors with 1♯ as well, then we check:
𝑁1♯ = 2𝑁 − 𝑖𝑖 ♯ − 𝑗 𝑗 ♯ − 𝑘 𝑘 ♯
= 𝑁 − 2(𝑎𝑏𝑐 + 𝑢𝑣𝑤) + (𝑎𝑢 + 𝑣𝑤)𝑖 + (𝑏𝑣 + 𝑢𝑤) 𝑗 + (𝑐𝑤 + 𝑢𝑣)𝑘.
satisfies
(𝑁1♯ ) 2 − 𝑁 (𝑁1♯ ) + 𝑁 (𝑎𝑏𝑐 + 𝑢𝑣𝑤) = 0 (22.4.18)
and
(𝑁1♯ ) (𝑁𝑖 ♯ ) = −𝑁 (𝑎𝑢 + 𝑣𝑤 − 𝑎𝑖 − 𝑣𝑘) = (𝑁𝑖 ♯ ) (𝑁1♯ ) + 𝑁 (𝑤 𝑗 − 𝑣𝑘)

(𝑁1♯ ) (𝑁 𝑗 ♯ ) = −𝑁 (𝑏𝑣 + 𝑤𝑢 − 𝑤𝑖 − 𝑏 𝑗) = (𝑁 𝑗 ♯ ) (𝑁1♯ ) + 𝑁 (−𝑤𝑖 + 𝑢𝑘)
(𝑁1♯ ) (𝑁 𝑘 ♯ ) = −𝑁 (𝑐𝑤 + 𝑢𝑣 − 𝑢 𝑗 − 𝑐𝑘) = (𝑁 𝑘 ♯ ) (𝑁1♯ ) + 𝑁 (𝑣𝑖 − 𝑢 𝑗).
The result follows.
Finally, we officially combine our work to prove the main theorem of this section.
Proof of Main Theorem 22.4.1. Combine Propositions 22.4.4 and 22.4.12.
Remark 22.4.19. Just as in section 5.5, we may ask about embeddings of a quadratic
ring in an order. However, moving from the rational to the integral is a bit tricky, and
the issue of embeddings is a theme that will return with gusto in Chapter 30. In that
context, it will be more natural to look at a different ternary quadratic form to measure
embeddings; just as in the case of trace zero, it is related to but not the same as the one
obtained in the above bĳection.
22.5 Twisting and final bĳection

In this final section, we conclude with the final bĳection. We must keep track of the
extra data of an ideal class, and along the way allow coefficient ideals. Throughout,
let 𝑅 be a Dedekind domain.
We first need the following slightly revised notion of similarity (one that ‘glues
together’ local similarities) allowing scaling by fractional ideals.
Definition 22.5.1. Let 𝑄 : 𝑀 → 𝔩 be a quadratic module with 𝔩 a fractional 𝑅-ideal.

The twist of 𝑄 by a fractional 𝑅-ideal 𝔲 is the quadratic form 𝔲 ⊗ 𝑄 : 𝔲 ⊗ 𝑀 → 𝔲2 𝔩
defined by (𝔲 ⊗ 𝑄) (𝑢 ⊗ 𝑥) = 𝑢 2 𝑄(𝑥).
A twisted similarity between quadratic modules 𝑄 and 𝑄 0 is a similarity between
𝑄 and a twist 𝔲 ⊗ 𝑄 0 for some fractional 𝑅-ideal 𝔲.
Example 22.5.2. If 𝔲 = 𝑢𝑅 is a principal fractional ideal, then twisted similarities

between 𝑄 and 𝔞𝑄 0 = 𝑢𝑄 0 are precisely those obtained from a similarity between 𝑄
and 𝑄 0, multiplied by 𝑢. In particular, if 𝑅 is a PID, then the notions of similarity and
twisted similarity coincide.
Example 22.5.3. Two quadratic modules 𝑄, 𝑄 0 : 𝑀, 𝑀 0 → 𝔩 with the same codomain

are twisted similar if and only if they are similar. Indeed, if 𝔲2 𝔩 = 𝔩, then 𝔲 = 𝑅.
22.5.4. Second, we extend the definition of the inverse in 22.4.8 using the reduced
norm to the noetherian domain 𝑅 as follows. Let O ⊆ 𝐵 be a quaternion 𝑅-order.
22.5. TWISTING AND FINAL BIJECTION 359
Since 𝑅 is a Dedekind domain, the reduced discriminant discrd(O) ⊆ 𝑅 of O is an

invertible 𝑅-ideal. Define the map
nrd♯ (O) : (O♯ ) 0 → 𝐹

(22.5.5)
𝛼 ↦→ nrd(𝛼).
Lemma 22.5.6. For a quaternion order O, the map nrd♯ (O) defines a ternary
quadratic module with values in discrd(O) −1 .
Proof. The reduced norm defines a quadratic map, so we only need to verify that the
codomain is valid. To this end, we may check locally since reduced norm and reduced
discriminant commute with localization. Reducing to the local case, suppose 𝑅 is now
a local Dedekind domain hence a PID. Choosing a basis, we verified in (22.4.14) that
𝑁 nrd♯ (O) ⊆ 𝑅, where discrd(O) = (𝑁); the result follows.
Main Theorem 22.5.7. Let 𝑅 be a Dedekind domain. Then the associations

( Nondegenerate ternary quadratic ) ( Quaternion orders )
modules over 𝑅 ↔ over 𝑅 up to
up to twisted similarity isomorphism
(22.5.8)
𝑄 ↦→ Clf 0 (𝑄)
nrd♯ (O) ← O
are mutually inverse, discriminant-preserving bĳections that are also functorial with
respect to 𝑅.
Proof. We proved a version of this statement when 𝑅 is a PID in Theorem 22.4.1.

More generally, we work now with a pseudobasis instead of a basis, explaining the
presence of the twisted similarity.
The surjectivity of the even Clifford map follows by generalizing the argument in
Proposition 22.4.4 and 22.4.7 to show that O has a good pseudobasis: see Exercise
22.5.
Let 𝑄 : 𝑀 → 𝔩 be a quadratic module with O := Clf 0 (𝑄). Returning to Example
22.3.24, we may write
𝑀 = 𝔞𝑒 1 ⊕ 𝔟𝑒 2 ⊕ 𝔠𝑒 3
for fractional ideals 𝔞, 𝔟, 𝔠, and with 𝑄 𝐹 (𝑥, 𝑦, 𝑧) := 𝑄(𝑥𝑒 1 + 𝑦𝑒 2 + 𝑧𝑒 3 ) in the usual
form provided by (22.3.5). Let 𝑁 := disc(𝑄 𝐹 ). Then
O = 𝑅 ⊕ 𝔟𝔠𝔩−1𝑖 ⊕ 𝔞𝔠𝔩−1 𝑗 ⊕ 𝔞𝔟𝔩−1 𝑘.
Consider now nrd♯ (O) : (O♯ ) 0 → discrd(O) −1 : as in 22.4.8 we have
(O♯ ) 0 = 𝔩(𝔟𝔠) −1𝑖 ♯ ⊕ 𝔩(𝔞𝔠) −1 𝑗 ♯ ⊕ 𝔩(𝔞𝔟) −1 𝑘 ♯ . (22.5.9)
To prepare our twisted similarity, let
𝔡 := 𝔞𝔟𝔠𝔩−1 ' 𝑀 ⊗ 𝐿∨ .
Ó3
(22.5.10)
Then
𝔡(O♯ ) 0 = 𝔞𝑖 ♯ ⊕ 𝔟 𝑗 ♯ ⊕ 𝔠𝑘 ♯ . (22.5.11)
We claim that the reduced norm on 𝔡(O♯ ) 0 has values in 𝑁 −1 𝔩 and is similar to 𝑄. The
claim follows from the same calculation in the proof of Proposition 22.4.12, namely
that nrd(𝑥𝑖 ♯ + 𝑦 𝑗 ♯ + 𝑧𝑘 ♯ ) = 𝑁 −1 𝑄 𝐹 (𝑥, 𝑦, 𝑧)! We conclude that nrd♯ (O) is twisted
similar to 𝑄.
We conclude with the following application to quadratic forms.
Corollary 22.5.12. Let 𝑄 : 𝑀 → 𝔩 be a ternary quadratic module and O := Clf 0 (𝑄).

Then the even Clifford map induces a bĳection Cl 𝑄 ↔ Typ O.
Proof. We first claim that the even Clifford map induces an injection Gen 𝑄 → Gen O,
giving an injection Cl 𝑄 → Typ O. Indeed, let 𝑄 0 ∈ Gen 𝑄, so 𝑄 0 : 𝑀 0 → 𝔩0 is locally
isometric to 𝑄. Let O0 := Clf 0 (𝑄 0). Since 𝑄 0, 𝑄 are locally isometric, they are locally
similar, so O0, O are locally isomorphic by Main Theorem 22.4.1, so O0 ∈ Gen O.
And if 𝑄 0 ' 𝑄 are isometric, again they are (twisted) similar, so by Main Theorem
22.5.7 we have O0 ' O.
To finish, we need to show that the even Clifford map is surjective. We pass from
similarity classes to isometry classes in the same way as in the proof of Corollary
5.2.6. To this end, let O0 ∈ Gen(O). Let 𝑄 0 := 𝔡 nrd♯ (O0) as in (22.5.11). By the
same rescaling argument given in Corollary 5.2.6, applying a similarity to 𝑄 0 we may
further suppose that disc 𝑄 𝐹0 = disc 𝑄 𝐹 ∈ 𝐹 × /𝐹 ×2 . By surjectivity in Main Theorem
22.5.7, for every prime 𝔭 of 𝑅, there exists a twisted similarity from 𝑄 0(𝔭) to 𝑄 (𝔭) over
𝑅 (𝔭) —and since each 𝑅 (𝔭) is a PID, by Example 22.5.2, these are in fact similarities.
Taking such a similarity and considering it as a similarity over 𝐹, again repeating the
same argument as at the end of Corollary 5.2.6, we conclude that 𝑄 0 and 𝑄 are locally
isometric, so 𝑄 0 ∈ Gen 𝑄. Finally, if O0 ' O, repeating these arguments one more
time over 𝑅 (first to go from twisted similar to similar, then to note the similarity gives
rise to an isometry) we conclude that 𝑄 0 ' 𝑄.
Remark 22.5.13. The correspondence between ternary quadratic forms and quaternion
orders has a particularly rich history. Perhaps the earliest prototype is due to Hermite
[Herm1854], who examined the product of automorphs of ternary quadratic forms.
Early versions of the correspondence were given by Latimer [Lat37, Theorem 3], Pall
[Pall46, Theorems 4–5], and Brandt [Bra43, §3ff] over Z by use of explicit formulas.
Various attempts were made to generalize the correspondence to Dedekind do-
mains, with the thorny issue being how to deal with a nontrivial class group. Eichler
[Eic53, §14, p. 96] gave such an extension. Peters [Pet69, §4] noted that Eichler’s cor-
respondence was not onto due to class group issues, and he gave a rescaled version that
gives a bĳection for Gorenstein orders. Eichler’s correspondence was further tweaked
by Nipp [Nip74, §3], who opted for a different scaling factor that is not restricted to a
class of orders, but his correspondence fails to be onto [Nip74, p.536].
These correspondences were developed further by Brzezinski [Brz80, §3], [Brz85,
§3], where he connected the structure of orders to relatively minimal models of the cor-
responding integral conic; see also Remark 24.3.11. He revisited the correspondence
again in the context of Gorenstein orders [Brz82, §3] and Bass orders [Brz83b, §2].
Lemurell [Lem2011, Theorem 4.3] gives a concise account of the correspondence of
Brzezinski over a PID (the guts of which are contained in [Brz82, (3.2)]).
More recently, Gross–Lucianovic [GrLu2009, §4] revisited the correspondence
over a PID or local ring, and they extended it to include quadratic forms of nonzero
discriminant and without restricting to Gorenstein orders; this extension is impor-
tant for automorphic reasons, connected to Fourier coefficients of modular forms
on PGSp(6), as developed by Lucianovic in his thesis [Luc2003]. Balaji [Bal2007,
Theorem 3.1] studied degenerations of ternary quadratic modules in the context of
orthogonal groups and Witt invariants and showed that the even Clifford functor is
bĳective over a general scheme. Finally, Voight [Voi2011a, Theorem B] gave a gen-
eral and functorial correspondence without any of the above restrictive hypotheses,
including the functorial inverse to the even Clifford functor provided above.
Remark 22.5.14. In the most general formulation of the correspondence, allowing arbi-
trary ternary quadratic modules discriminant over all sorts of rings, Voight [Voi2011a,
Theorem A] characterizes the image of the even Clifford functor, as follows. Let 𝐵
be an 𝑅-algebra that is (faithfully) projective of rank 4 as an 𝑅-module. Then 𝐵 is
a quaternion ring if 𝐵 ' Clf 0 (𝑄) for a ternary quadratic module 𝑄. Then 𝐵 is a
quaternion ring if and only if 𝐵 has a standard involution and for all 𝑥 ∈ 𝐵, the trace
of left (or right) multiplication by 𝑥 on 𝐵 is equal to 2 trd(𝑥).
For example, if we take the quadratic form 𝑄 : 𝑅 3 → 𝑅 defined by 𝑄(𝑥, 𝑦, 𝑧) = 0
identically, the multiplication table on Clf 0 (𝑄) gives the commutative ring
Clf 0 (𝑄) ' 𝑅[𝑖, 𝑗, 𝑘]/(𝑖, 𝑗, 𝑘) 2 .
One can see this as a kind of deformation of a quaternion algebra (in an algebro-
geometric sense), letting 𝑎, 𝑏 → 0.
Exercises
I 1. Let 𝑅 be a PID or local noetherian domain. Let 𝐴 be an 𝑅-algebra that is free
of finite rank as an 𝑅-module. Show that 𝐴 has an 𝑅-basis including 1. [Hint:
show that the quotient 𝐴/𝑅 is torsion-free, hence free; since free modules are
projective, the sequence 0 → 𝑅 → 𝐴 → 𝐴/𝑅 → 0 splits, giving 𝐴 ' 𝑅 ⊕ 𝐴/𝑅.]
I 2. For a free quadratic ternary form (as in 22.3.4), show that a change of basis
𝜌 ∈ GL3 (𝑅) acts on 𝑖, 𝑗, 𝑘 ∈ Clf 0 (𝑄) by the adjugate matrix adj(𝜌) ∈ GL3 (𝑅)
(where the entries of adj(𝜌) are the 2 × 2-cofactors of 𝜌 and 𝜌 adj(𝜌) = det(𝜌)).
3. Let 𝑅 be a domain and let Pic 𝑅 be the group of isomorphism classes of invertible
𝑅-modules (equivalently, classes of fractional 𝑅-ideals in 𝐹). Show that up to
twisted similarity, the target of a quadratic module only depends on its class in
Pic 𝑅/2 Pic 𝑅. [See Example 9.7.5.]
4. Finish the direct verification in Example 22.3.24 that O is closed under multi-
plication.
I 5. Let 𝑅 be a Dedekind domain, and let O be a quaternion 𝑅-order. Show that
there exist 𝑖, 𝑗, 𝑘 ∈ O and 𝔞, 𝔟, 𝔠 ⊂ 𝐹 fractional 𝑅-ideals such that O = 𝑅 + 𝔞𝑖 +
𝔟 𝑗 + 𝔠𝑘 and such that 1, 𝑖, 𝑗, 𝑘 satisfy the multiplication rules (22.3.7) for some
𝑎, 𝑏, 𝑐, 𝑢, 𝑣, 𝑤 ∈ 𝐹—called a good pseudobasis for O. [Hint: revisit what goes
into 22.4.7; a simple observation will suffice!]
6. Let 𝑄 : 𝑅 3 → 𝑅 be a ternary quadratic form and let O := Clf 0 (𝑄). Show that
disc nrd | O = (disc 𝑄) 2 .
[Hint: see (22.3.18).]

7. Let 𝑄 : 𝑀 → 𝑅 be a quadratic form over 𝑅. Construct a Clifford algebra with a
universal property analogous to Proposition 5.3.1, and recover the even Clifford
algebra and odd Clifford bimodule.
8. Let 𝑄 : 𝑀 → 𝑅 be a quadratic form such that there exists 𝑥 ∈ 𝑀 such that
𝑄(𝑥) ∈ 𝑅 × .
(a) Show that the odd Clifford bimodule Clf 1 (𝑄) is free of rank 1 as a Clf 0 (𝑄).
(b) Generalize this result to case where 𝑄 : 𝑀 → 𝐿 is a quadratic module.
9. Let 𝑄 : 𝑀 → 𝑅 be a quadratic form over 𝑅 with 𝑀 of odd rank as an 𝑅-module
and let 𝐹 = Frac 𝑅. Let 𝑆 := 𝑍 (Clf 𝑄) ↩→ 𝐾 := 𝑍 (Clf 𝑄 𝐹 ) be the center of the
Clifford algebra of 𝑄. Show that 𝑆 is an 𝑅-order in 𝐾.
I 10. Show that nrd(O♯ ) = nrd((O♯ ) 0 ). [Hint: use (22.4.18).]
11. Let 𝑅 be a Dedekind domain with 𝐹 = Frac 𝑅, let 𝐵 be a quaternion algebra
over 𝐹, and let O ⊆ 𝐵 be an 𝑅-order. Let 𝑆 ⊆ O be an 𝑅-order.
(a) Suppose 𝑆 ⊆ O is integrally closed. Prove that O is projective of rank 2
as a left 𝑆-module.
(b) If 𝑆 is not integrally closed, then show that (a) need not hold by the
following example. Let 𝑅 = Z and 𝐹 = Q, let 𝐵 = (−1, −1 | Q), let
O = Z + Z𝑝𝑖 + Z 𝑗 + Z𝑖 𝑗
for an odd prime 𝑝 (that is 𝑝𝑖, not 𝜋!). Let 𝑆 = Z[ 𝑝𝑖] ⊆ O. Show that O
is not projective as a left 𝑆-module.
(c) Show that the property that O is projective as an 𝑆-module is a local
property (over primes of 𝑅).
(d) In light of (c), suppose that 𝑅 is a PID, and write O in a good basis (22.3.7).
Suppose that 𝑆 = 𝑅[𝑖] with 𝑖 2 = 𝑢𝑖 − 𝑏𝑐. Show that O is projective as an
𝑆-module if and only if the quadratic form 𝑏𝑥 2 + 𝑢𝑥𝑦 + 𝑐𝑦 2 represents a
unit.
(e) Using (d), conclude in general that if 𝑆 has conductor coprime to discrd O,
show that O is projective as an 𝑆-module.
12. Let 𝑅 be a global ring with 𝐹 := Frac 𝑅, let 𝐵, 𝐵 0 be quaternion algebras over
𝐹, and let O ⊆ 𝐵 and O0 ⊆ 𝐵 0 be 𝑅-orders. Consider the quaternary quadratic
forms 𝑄 := nrd | O : O → 𝑅 and similarly 𝑄 0 on O0.
(a) Show that 𝑄 is isometric to 𝑄 0, then there is an isomorphism of 𝐹-algebras
∼ 𝐵 0.
𝐵 −→
(b) In light of (a), suppose O, O0 ⊆ 𝐵 = 𝐵 0. Show that O, O0 are isomorphic

as 𝑅-orders if and only if 𝑄, 𝑄 0 are isometric as (quaternary) quadratic
modules. [Hint: if 𝑄 is isometric to 𝑄 0, then show that there is a similarity
on the trace zero elements of the duals, thereby giving an isomorphism
O −→∼ O0.]
13. Let 𝑄 : 𝑀 → 𝐿 be a quadratic module. Show that the even Clifford algebra
Clf 0 (𝑄) with its map 𝜄 : 𝑀 ⊗ 𝑀 ⊗ 𝐿 ∨ → Clf 0 (𝑄) has the following universal
property: if 𝐴 is an 𝑅-algebra and 𝜄 𝐴 : 𝑀 ⊗ 𝑀 ⊗ 𝐿 → 𝐴 is an 𝑅-module
homomorphism such that
(i) 𝜄 𝐴 (𝑥 ⊗ 𝑥 ⊗ 𝑓 ) = 𝑓 (𝑄(𝑥)) for all 𝑥 ∈ 𝑀 and 𝑓 ∈ 𝐿 ∨ , and
(ii) 𝜄 𝐴 (𝑥 ⊗ 𝑦 ⊗ 𝑓 )𝜄 𝐴 (𝑦 ⊗ 𝑧 ⊗ 𝑔) = 𝑓 (𝑄(𝑦))𝜄 𝐴 (𝑥 ⊗ 𝑧 ⊗ 𝑔) for all 𝑥, 𝑦, 𝑧 ∈ 𝑀 and
𝑓 , 𝑔 ∈ 𝐿∨,
then there exists a unique 𝑅-algebra homomorphism 𝜙 : Clf 0 (𝑄) → 𝐴 such that
the diagram
𝑀 ⊗ 𝑀 ⊗ 𝐿∨
𝜄 / Clf (𝑄)
𝜙
&
𝜄𝐴
𝐴
commutes. Conclude that the pair (Clf 0 (𝑄), 𝜄) is unique up to unique isomor-
phism.
Chapter 23
Quaternion orders
In the previous chapter, we gave a rather general classification of quaternion orders in

terms of ternary quadratic modules. In this chapter, we take a guided tour of the most
important animals in the zoo of quaternion orders, identifying those with good local
properties. We continue in the next chapter with a second visit to the zoo.
23.1 ⊲ Highlights of quaternion orders

We begin in this section by providing some highlights of this tour. Let 𝐵 be a quaternion
algebra over Q of discriminant 𝐷 := disc 𝐵 and let O ⊂ 𝐵 be an order with reduced
discriminant 𝑁 := discrd(O). Then 𝑁 = 𝐷 𝑀 with 𝑀 ∈ Z ≥1 , and O is maximal if and
only if 𝑁 = 𝐷.
23.1.1 (Maximal orders). The nicest orders are undoubtedly the maximal orders, those
not properly contained in another order. An order is maximal if and only if it is locally
maximal (Lemma 10.4.3), i.e. 𝑝-maximal for all primes 𝑝; globally, an order O is
maximal if and only if 𝑁 = 𝐷 (i.e., 𝑀 = 1).
We have either 𝐵 ' M2 (Q 𝑝 ) or 𝐵 is a division algebra over Q 𝑝 (unique up
to isomorphism). If 𝐵 is split, then a maximal order is isomorphic (conjugate) to
M2 (Z 𝑝 ), and the corresponding ternary quadratic form is the determinant 𝑥𝑦 − 𝑧 2
(see Example 22.3.12). If instead 𝐵 is division, then the unique maximal order is the
valuation ring, with corresponding anisotropic form 𝑥 2 − 𝑒𝑦 2 + 𝑝𝑧 2 for 𝑝 ≠ 2, where
𝑒 ∈ Z is a quadratic nonresidue modulo 𝑝 (and for 𝑝 = 2, the associated form is
𝑥 2 + 𝑥𝑦 + 𝑦 2 + 2𝑧2 ).
Maximal orders have modules with good structural properties: all lattices 𝐼 ⊂ 𝐵
with left or right order equal to a maximal order O are invertible (Theorem 18.1.2).
There is a combinatorial structure, called the Bruhat–Tits tree, that classifies
maximal orders in M2 (Q 𝑝 ) (as endomorphism rings of lattices, up to scaling): the
Bruhat–Tits tree is a 𝑝 + 1-regular tree (see section 23.5).
Examining orders beyond maximal orders is important for the development of

the theory: already the Lipschitz order—an order which arises when considering if a
365
366 CHAPTER 23. QUATERNION ORDERS
positive integer is the sum of four squares—is properly contained inside the Hurwitz
order (Chapter 11).
23.1.2 (Hereditary orders). More generally, we say that O is hereditary if every left
or right fractional O-ideal (i.e., lattice 𝐼 ⊆ 𝐵 with left or right order containing O) is
invertible. Maximal orders are hereditary, and being hereditary is a local property. A
hereditary Z 𝑝 -order O 𝑝 ⊆ 𝐵 𝑝 is either maximal or

Z𝑝 Z𝑝 𝑎 𝑏
O𝑝 ' = : 𝑎, 𝑏, 𝑐, 𝑑 ∈ Z 𝑝 ⊆ M2 (Q 𝑝 ) ' 𝐵 𝑝
𝑝Z 𝑝 Z 𝑝 𝑝𝑐 𝑑
with associated ternary quadratic form 𝑥𝑦 − 𝑝𝑧2 . Thus O ⊂ 𝐵 is hereditary if and only
if discrd(O) = 𝐷 𝑀 is squarefree, so in particular gcd(𝐷, 𝑀) = 1.
Hereditary orders share the nice structural property of maximal orders: all lattices
𝐼 ⊂ 𝐵 with hereditary left or right order are invertible. The different ideal diff O 𝑝 is
generated by any element 𝜇 ∈ O 𝑝 such that 𝜇2 ∈ 𝑝Z 𝑝 .
23.1.3 (Eichler orders). More generally, we can consider orders that are “upper trian-
gular modulo 𝑀” with gcd(𝐷, 𝑀) = 1 (i.e., avoiding primes that ramify in 𝐵). The
order
Z𝑝 Z𝑝
⊆ M2 (Z 𝑝 )
𝑝𝑒 Z 𝑝 Z 𝑝
is called the standard Eichler order of level 𝑝 𝑒 in M2 (Q 𝑝 ). A Z 𝑝 -order O 𝑝 ⊆
M2 (Q 𝑝 ) is an Eichler order if O 𝑝 is isomorphic to a standard Eichler order. The
ternary quadratic form associated to an Eichler order of level 𝑝 𝑒 is 𝑥𝑦 − 𝑝 𝑒 𝑧 2 .
Globally, we say O ⊂ 𝐵 is a Eichler order of level 𝑀 if discrd(O) = 𝑁 = 𝐷 𝑀
with gcd(𝐷, 𝑀) = 1 and O 𝑝 is an Eichler order of level 𝑝 𝑒 for all 𝑝 𝑒 k 𝑀. In
particular, O 𝑝 is maximal at all primes 𝑝 | 𝐷. Every hereditary order is Eichler, and
an Eichler order is hereditary if and only if its level 𝑀 (or reduced discriminant 𝑁) is
squarefree. A maximal Z 𝑝 -order O 𝑝 ⊆ M2 (Z 𝑝 ) is an Eichler order of level 1 = 𝑝 0 .
Eichler orders play a crucial role in the context of modular forms, as we will see in the
final part of this monograph.
This local description of Eichler orders also admits a global description. The
standard Eichler order O 𝑝 can be written
𝑝 −𝑒 Z 𝑝

Z Z𝑝 Z
O𝑝 = 𝑒 𝑝 = M2 (Z 𝑝 ) ∩ 𝑒 𝑝 = M2 (Z 𝑝 ) ∩ 𝜛 −1 M2 (Z 𝑝 )𝜛
𝑝 Z𝑝 Z𝑝 𝑝 Z𝑝 Z𝑝
as the intersection of a (unique) pair of maximal orders, with

0 1
𝜛 := 𝑒 ∈ 𝑁GL2 (Q 𝑝 ) (O 𝑝 ) (23.1.4)
𝑝 0
a generator of the group 𝑁GL2 (Q 𝑝 ) (O 𝑝 )/Q×𝑝 O×𝑝 ' Z/2Z, and 𝜛 2 = 𝑝 𝑒 . The different
diff O 𝑝 is the two-sided ideal generated by 𝜛.
From the local-global dictionary, it follows that O ⊂ 𝐵 is Eichler if and only if O
is the intersection of two (not necessarily distinct) maximal orders.
23.2. MAXIMAL ORDERS 367
23.2 Maximal orders

Throughout this chapter, we impose the following notation: let 𝑅 be a Dedekind
domain with field of fractions 𝐹 = Frac 𝑅, let 𝐵 be a quaternion algebra over 𝐹, and
let O ⊆ 𝐵 an 𝑅-order.
23.2.1. We make the following convention. When we say “𝑅 is local”, we mean that
𝑅 is a complete DVR, and in this setting we let 𝔭 = 𝜋𝑅 be its maximal ideal, and
𝑘 = 𝑅/𝔭 the residue field. When we want to return to the general context, we will say
“𝑅 is Dedekind”.
Recall that an 𝑅-order is maximal if it is not properly contained in another order. We

begin in this section by summarizing the properties of maximal orders for convenience.
23.2.2. Being maximal is a local property (Lemma 10.4.3), so the following are
equivalent:
(i) O is a maximal 𝑅-order;

(ii) O (𝔭) is a maximal 𝑅 (𝔭) -order for all 𝔭 ⊆ 𝑅; and
(iii) O𝔭 is a maximal 𝑅𝔭 -order for all 𝔭 ⊆ 𝑅.
We recall (Theorem 15.5.5) that an order over a global ring 𝑅 is maximal if and
only if discrd(O) = disc𝑅 (𝐵). Local maximal 𝑅-orders have the nice local description.
23.2.3. Suppose 𝑅 is local and that 𝐵 ' M2 (𝐹) is split. Then by Corollary 10.5.5,
every maximal 𝑅-order in M2 (𝐹) is conjugate to M2 (𝑅) by an element of GL2 (𝑅),
i.e. O ' M2 (𝑅). We have discrd(O) = 𝑅. All two-sided ideals of O are powers of
rad(O) = 𝔭O, and
O/rad(O) ' M2 (𝑘).
The associated ternary quadratic form is similar to 𝑄(𝑥, 𝑦, 𝑧) = 𝑥𝑦 − 𝑧2 , by Example
22.3.12 and the classification theorem (Main Theorem 22.1.1). Finally,
𝑁 𝐵× (O) = 𝑁GL2 (𝐹 ) (M2 (𝑅)) = 𝐹 × O× . (23.2.4)
23.2.5. Suppose 𝑅 is local but now that 𝐵 is a division algebra. Then the valuation
ring O ⊂ 𝐵 is the unique maximal 𝑅-order by Proposition 13.3.4.
Suppose further that the residue field 𝑘 is finite (equivalently, that 𝐹 is a local
field). Then Theorem 13.3.11 applies, and we have

𝐾, 𝜋
O ' 𝑆 ⊕ 𝑆𝑗 ⊆
𝐹
where 𝐾 ⊇ 𝐹 is the unique quadratic unramified extension of 𝐹 and 𝑆 its valuation

ring. We computed in 15.2.12 that discrd(O) = 𝔭. All two-sided ideals of O are
powers of the unique maximal ideal rad(O) = 𝑃 = O 𝑗O, and ℓ := O/rad(O) is a
quadratic field extension of 𝑘. By Exercise 13.8, we have 𝑃 = [O, O] equal to the
commutator. We also have 𝑃 = diff O; this can be computed directly, or it follows
from the condition that nrd(diff O) = discrd O = 𝔭.
Write 𝑆 = 𝑅[𝑖] with 𝑖 2 = 𝑢𝑖 − 𝑏, and 𝑢, 𝑏 ∈ 𝑅. Then 1, 𝑖, 𝑗, 𝑘 where 𝑘 = −𝑖 𝑗 is an

𝑅-basis for O. We have
𝑘 2 = 𝑖( 𝑗𝑖) 𝑗 = 𝑖(𝑖 𝑗) 𝑗 = nrd(𝑖)𝜋,
so trd(𝑘) = 0, and 𝑘 = −𝑘 = 𝑖 𝑗. This gives multiplication table
𝑖 2 = 𝑢𝑖 − 𝑏 𝑗 𝑘 = −𝜋𝑖
2
𝑗 =𝜋 𝑘𝑖 = 𝑏 𝑗 (23.2.6)
2
𝑘 = 𝑏𝜋 𝑖𝑗 = 𝑘
realizing the basis as a good basis in the sense of 22.4.7; the associated ternary
quadratic form is
nrd♯ (O) (𝑥, 𝑦, 𝑧) = −𝜋𝑥 2 + 𝑏𝑦 2 + 𝑢𝑦𝑧 + 𝑧 2 = −𝜋𝑥 2 + Nm𝐾 |𝐹 (𝑧 + 𝑦𝑖). (23.2.7)
Finally, since the valuation ring is the unique maximal order and conjugation
respects integrality, we have
𝑁 𝐵× (O) = 𝐵× . (23.2.8)
Finally, lattices over maximal orders are necessarily invertibility, as follows.
23.2.9. All lattices 𝐼 ⊂ 𝐵 with left or right order equal to a maximal order O are
invertible, by Theorem 18.1.2, proven in Proposition 18.3.2. (We also gave a different
proof of this fact in Proposition 16.6.15(b).)
The classification of two-sided ideals and their classes follows from that of hered-
itary orders: see 23.3.19.
23.3 Hereditary orders

Hereditary orders were investigated in section 21.4 in general; here, we provide a
quick development specific to quaternion algebras. As mentioned before, a good
general reference for (maximal and) hereditary orders is Reiner [Rei2003, Chapters
3–6, 9].
We recall that O is hereditary if every left (or right) ideal 𝐼 ⊆ O is projective
as a left (or right) O-module. Being hereditary is a local property 21.4.4 because
projectivity is.
23.3.1. Suppose 𝑅 is local. By Main Theorem 21.1.4 and Corollary 21.1.5, the
following are equivalent:
(i) O is hereditary;
(ii) rad O is projective as a left (or right) O-module;
(iii) OL (rad O) = OR (rad O) = O;
(iv) rad O is invertible as a (sated) two-sided O-ideal; and
(v) either O is maximal or

𝑅 𝑅
O' ⊆ M2 (𝐹) ' 𝐵.
𝔭 𝑅
We now spend some time investigating ‘the’ local hereditary order that is not
maximal. So until further notice, suppose 𝑅 is local and let

𝑅 𝑅
O0 (𝔭) := ⊆ M2 (𝑅).
𝔭 𝑅
To avoid clutter, we will just write O = O0 (𝔭). We have
discrd(O) = [M2 (𝑅) : O] 𝑅 disc(M2 (𝑅)) = 𝔭. (23.3.2)
23.3.3. A multiplication table for O is obtained from the one for M2 (𝑅) in Example
22.3.12, scaling 𝑗 by 𝜋 in (22.3.14), which gives the same multiplication laws as
(22.3.13) except now 𝑖 𝑗 = −𝜋𝑘 and 𝑐 is scaled by 𝜋. Therefore, the similarity class of
ternary quadratic forms associated to O is represented by
𝑄(𝑥, 𝑦, 𝑧) = 𝑥𝑦 − 𝜋𝑧2 . (23.3.4)
23.3.5. Let 𝐽 := rad(O). Then
𝔭 𝑅
𝐽= (23.3.6)
𝔭 𝔭
by (21.3.6), and we find
𝑘 0
O/𝐽 ' '𝑘×𝑘 (23.3.7)
0 𝑘
as 𝑘-algebras. Now let
0 1
𝜛= . (23.3.8)
𝜋 0
Then a direct calculation yields
𝐽 = O𝜛 = 𝜛O (23.3.9)
in agreement with 23.3.1(ii)–23.3.1(iii), and 𝐽 is an invertible O-ideal. Since 𝜛 2 = 𝜋,
we have
𝐽 2 = 𝔭O. (23.3.10)
In particular, 𝐽 −1 = 𝜋 −1 𝐽, and the powers of 𝐽 give a filtration
O ) 𝐽 ) 𝔭O ) 𝐽 3 ) . . . . (23.3.11)
23.3.12. We compute that
𝔭−1 𝑅

𝑅
𝜛 M2 (𝑅)𝜛 −1 =
𝔭 𝑅
and hence
O = M2 (𝑅) ∩ 𝜛 M2 (𝑅)𝜛 −1
is the intersection of two maximal orders.
Lemma 23.3.13. The group Idl(O) = PIdl(O) is generated by 𝐽 = rad O, with

𝐽 2 = 𝔭O.
Proof. We have Idl(O) = PIdl(O) since 𝑅 is local: invertible is equivalent to principal

(Main Theorem 16.6.1).
Let 𝐼 ⊆ O be an invertible two-sided O-ideal. Then by (23.3.11), we can replace
𝐼 by a power of 𝐽 −1 and suppose that O ( 𝐼 ⊆ 𝐽. Invertible means locally principal,
so 𝐼 = O𝛼 = 𝛼O and
[O : 𝐼] 𝑅 = nrd(𝐼) 2 = nrd(𝛼) 2 𝑅 | [O : 𝐽] 𝑅 = 𝔭2 .
Thus [O : 𝐼] 𝑅 = 𝔭2 , so [𝐼 : 𝐽] 𝑅 = 𝑅 and 𝐼 = 𝐽. (This also follows directly from

Proposition 16.4.3.)
Here is a second computational proof. The image 𝐼/𝐽 ⊆ O/𝐽 ' 𝑘 × 𝑘 (23.3.7) is
a two-sided ideal, therefore

𝑅 𝔭 𝔭 𝔭
𝐼= or 𝐼 = .
𝔭 𝔭 𝔭 𝑅
But
𝑅 𝔭 0 1 0 𝑅 𝑅 𝔭
= * ;
𝔭 𝔭 0 0 0 𝔭 𝔭 𝔭
we get a contradiction with the other possibility by multiplying instead on the left. A
third “pure matrix multiplication” proof is also requested in Exercise 23.1.
The second statement was already proven in (23.3.10).
Corollary 23.3.14. We have 𝑁 𝐵× (O)/(𝐹 × O× ) ' Z/2Z generated by 𝜛, and

(
𝐹 ×2 𝑅𝔭× , if 𝑒 is even;
nrd(𝑁 𝐵× (O)) =
𝐹×, if 𝑒 is odd.
Proof. By (18.5.4), we have an isomorphism
𝑁 𝐵× (O)/(𝐹 × O× ) ' PIdl(O)/PIdl(𝑅);
by Lemma 23.3.13, the latter is generated by 𝐽 = 𝜛O with 𝐽 2 = 𝜋O. The computation

of reduced norms is immediate.
23.3.15. Lemma 23.3.13 also implies the description
𝐽 = [O, O] (23.3.16)
as the commutator. Since O/𝐽 is commutative, we know [O, O] ⊆ 𝐽; but [O, O] (

𝐽 2 = 𝔭O since O/𝔭O is noncommutative. We also have 𝐽 = diff O for the same
reason, since nrd(diff O) = discrd O = 𝔭. (A matrix proof of these facts are requested
in Exercise 23.8.)
23.3.17. We now classify the left O-lattices, up to isomorphism. Each such O-lattice
is projective, since O is hereditary.
By the Krull-Schmidt theorem (Theorem 20.6.2), every O-lattice can be written as
the direct sum of indecomposables, so it is enough to classify
the indecomposables;
𝐹
and we did just this in 21.5.3. Explicitly, we have 𝑉 = ' 𝐹 2 the simple 𝐵 =
𝐹

𝑅 𝑅
M2 (𝐹)-module, and we take the O-lattice 𝑀 = ⊂ 𝑉. We have 𝐽 𝑀 = and
𝑅 𝔭

𝔭
𝐽2 𝑀 = = 𝜋𝑀, and 𝑀, 𝐽 𝑀 give a complete set of indecomposable left O-modules.
𝔭
As expected, O = 𝐽 𝑀 ⊕ 𝑀 is a decomposition of O into projective indecomposable
left O-modules.
The preceding local results combine to determine global structure. Now let 𝑅 be
a global ring.
Lemma 23.3.18. O is hereditary if and only if discrd(O) is squarefree.
Proof. We argue locally; and then we use the characterization (iv), the computation
of the reduced discriminant (23.3.2), and the same argument as in Theorem 15.5.5 to
finish.
23.3.19. Let O be a hereditary (possibly maximal) 𝑅-order. By Theorem 21.4.9, we
know that the group Idl(O) is an abelian group generated by the prime (equivalently,
maximal) invertible two-sided ideals. We claim that the map
{Prime two-sided invertible O-ideals} ↔ {Prime ideals of 𝑅}
(23.3.20)
𝑃 ↦→ 𝑃 ∩ 𝑅
is a bĳection, generalizing Theorem 18.3.6. If 𝔭 - 𝔑 then we have 𝑃 = 𝔭O; and
if 𝔭 | 𝔇 then we have a prime two-sided ideal 𝑃 = O ∩ rad(O𝔭 ) with 𝑃2 = 𝔭O.
Otherwise, 𝔭 | 𝔑 but 𝔭 - 𝔇, so O𝔭 is hereditary but not maximal; from the local
description in Lemma 23.3.13, we get a prime ideal 𝑃 = O ∩ rad(O𝔭 ) with 𝑃2 = 𝔭O
as in the ramified case. This proves (23.3.20), and that the sequence
Ö
0 → Idl(𝑅) → Idl(O) → Z/2Z → 0 (23.3.21)
𝔭 |𝔑
is exact.
Taking the quotient by PIdl(𝑅), we obtain the exact sequence
Ö
0 → Cl 𝑅 → Pic𝑅 (O) → Z/2Z → 0. (23.3.22)
𝔭 |𝔑
In particular, if 𝔇 = (1), then Pic𝑅 (O) ' Cl 𝑅. Finally, the group of two-sided ideals
modulo principal two-sided ideals is related to the Picard group by the exact sequence
(18.5.5):
0 → 𝑁 𝐵× (O)/(𝐹 × O× ) → Pic𝑅 (O) → Idl(O)/PIdl(O) → 0
𝛼𝐹 × O× ↦→ [𝛼O] = [O𝛼]
(This exact sequence is sensitive to O even within its genus: see Remark 18.5.9.)
23.4 Eichler orders

We now consider a more general class of orders inspired by the hereditary orders.
Definition 23.4.1. An Eichler order O ⊆ 𝐵 is the intersection of two (not necessarily
distinct) maximal orders.
23.4.2. By the local-global dictionary for lattices (and orders), the property of being
an Eichler order is local. Moreover, from 23.3.12, it follows that a hereditary order is
Eichler.
Proposition 23.4.3. Suppose 𝑅 is local and O ⊆ 𝐵 = M2 (𝐹). Then the following are
equivalent:
(i) O is Eichler;

𝑅 𝑅
(ii) O ' 𝑒 ;
𝔭 𝑅

𝑅 0
(iii) O contains an 𝑅-subalgebra that is 𝐵× -conjugate
to ; and
0 𝑅
(iv) O is the intersection of a uniquely determined pair of maximal orders (not
necessarily distinct).
Proof. We follow Hĳikata [Hĳ74, 2.2(i)]. Apologies in advance for all of the explicit
matrix multiplication!
We prove (i) ⇒ (ii) ⇔ (iii) and then (ii) ⇒ (iv) ⇒ (i). The implications (ii) ⇒
(iii) and (iv) ⇒ (i) are immediate.
So first (i) ⇒ (ii). Suppose O = O1 ∩O2 . All maximal orders in 𝐵 are 𝐵× -conjugate
to M2 (𝑅), so there exist 𝛼1 , 𝛼2 ∈ 𝐵× such that O𝑖 = 𝛼𝑖−1 M2 (𝑅)𝛼𝑖 for 𝑖 = 1, 2.
Conjugating by 𝛼1 , we may suppose 𝛼1 = 1 and we write 𝛼 = 𝛼2−1 for convenience, so
O ' M2 (𝑅) ∩ 𝛼 M2 (𝑅)𝛼−1 . Scaling by 𝜋, we may suppose 𝛼 ∈ M2 (𝑅) r 𝜋 M2 (𝑅).
By row and column operations (Smith normal form, proven as part of the structure
theorem for finitely generated modules over a PID), there exist 𝛽, 𝛾 ∈ GL2 (𝑅) such
that
1 0
𝛽𝛼𝛾 =
0 𝜋𝑒
is in standard invariant form with 𝑒 ≥ 0. Then
O ' 𝛽O𝛽−1 = M2 (𝑅) ∩ 𝛽𝛼 M2 (𝑅)𝛼−1 𝛽−1 (23.4.4)
since 𝛽 ∈ GL2 (𝑅), and
𝔭−𝑒

1 0 𝑅 𝑅 1 0 𝑅
𝛽𝛼 M2 (𝑅)𝛼−1 𝛽−1 = = 𝑒 (23.4.5)
0 𝜋𝑒 𝑅 𝑅 0 𝜋 −𝑒 𝔭 𝑅
so
𝔭−𝑒

𝑅 𝑅 𝑅 𝑅 𝑅
O' ∩ 𝑒 = 𝑒 . (23.4.6)
𝑅 𝑅 𝔭 𝑅 𝔭 𝑅
23.4. EICHLER ORDERS 373

𝑅 0
To show (iii) ⇒ (ii), we may suppose O ⊇ = 𝑅𝑒 11 + 𝑅𝑒 22 with 𝑒 11 =
0 𝑅

1 0 0 0 𝐹 0
, 𝑒 22 = . Then O𝑒 11 ⊆ . Let 𝜋𝑖 𝑗 be the projection onto the
0 0 0 1 𝐹 0
𝑖 𝑗-coordinate. Then

𝑅 0
𝜋11 (O𝑒 11 ) = trd(O𝑒 11 )𝑒 11 ⊆ 𝑅 =
0 0
and so equality holds. Therefore

0 0
O ⊇ 𝜋21 (O) = 𝑎
𝔭 0
for some 𝑎 ∈ Z. Arguing again with the other matrix unit 𝑒 22 we conclude that

𝑅 𝔭𝑏
O= 𝑎 (23.4.7)
𝔭 𝑅
with 𝑎, 𝑏 ∈ Z. Multiplying

𝑅 𝔭𝑏 𝑅 𝔭𝑏 𝑅 + 𝔭𝑎+𝑏 𝔭𝑏
=
𝔭𝑎 𝑅 𝔭𝑎 𝑅 𝔭𝑎 𝑅 + 𝔭𝑎+𝑏
we conclude that 𝑒 =𝑎 + 𝑏 ≥ 0. Such

an order is maximal if and only if 𝑎 + 𝑏 = 0: if
𝑅 𝔭𝑏 𝑅 𝔭−𝑎 0 1
𝑎 ≥ 0, then 𝑎 ⊆ 𝑎 and similarly if 𝑎 ≤ 0. The element 𝛼 = 𝑎
𝔭 𝑅 𝔭 𝑅 𝜋 0
has
0 1 𝑅 𝔭𝑏 0 𝜋 −𝑎

𝑅 𝑅
𝛼−1 O𝛼 = 𝑎 = 𝑒 (23.4.8)
𝜋 0 𝔭𝑎 𝑅 1 0 𝔭 𝑅
(and normalizes the given subalgebra) so the result is proven.

𝑅 𝑅
To conclude, we show (ii) ⇒ (iv). Let O0 ⊇ O = 𝑒 be a maximal 𝑅-
𝔭 𝑅

𝑅 0
order. Since ⊆ O0, the argument of the previous paragraph applies, and
0 𝑅
𝑅 𝔭−𝑐

0
O = 𝑐 with 𝑐 ∈ Z satisfying 0 ≤ 𝑐 ≤ 𝑒. The intersection of another such
𝔭 𝑅

𝑅 𝔭𝑎
maximal orders with the parameter 𝑑 is the order 𝑏 where 𝑎 = max(𝑐, 𝑑)
𝔭 𝑅

𝑅 𝑅
and 𝑏 = − min(𝑐, 𝑑) so is equal to 𝑒 if and only if 𝑒 = 𝑎 = max(𝑐, 𝑑) and
𝔭 𝑅
0 = 𝑏 = min(𝑐, 𝑑), which uniquely determine 𝑐, 𝑑 up to swapping.
Remark 23.4.9. There is a further important equivalent characterization of Eichler

orders as being maximal or residually split: see Lemma 24.3.6.
Corollary 23.4.10. Every superorder of an Eichler order is Eichler.

Proof. The corollary is local, so we may apply Proposition 23.4.3(iii) to every super-
order.
Definition 23.4.11. Suppose 𝑅 is local. The standard Eichler order of level 𝔭𝑒 in

M2 (𝐹) is the order
𝑒 𝑅 𝑅
O0 (𝔭 ) := 𝑒 .
𝔭 𝑅
By Proposition 23.4.3, if 𝑅 is local then an order O ⊆ M2 (𝐹) is Eichler if and only

if O is conjugate to a standard Eichler order.
Suppose until further notice that 𝑅 is local, and let O = O0 (𝔭𝑒 ) be the standard
Eichler order of level 𝔭𝑒 with 𝑒 ≥ 0.
23.4.12. First two basic facts about the Eichler order O of level 𝔭𝑒 : We have
discrd(O) = [M2 (𝑅) : O] 𝑅 = 𝔭𝑒
and its associated ternary quadratic form 𝑄(𝑥, 𝑦, 𝑧) = 𝑥𝑦 − 𝜋 𝑒 𝑧2 as in 23.3.3.
23.4.13. Let
0 1
𝜛= ∈ O.
𝜋𝑒 0
Then
O = M2 (𝑅) ∩ 𝜛 −1 M2 (𝑅)𝜛
as in (23.4.6); by Proposition 23.4.3 these two orders are the uniquely determined pair
of maximal orders containing O. We have 𝜛 2 = 𝜋 𝑒 , and so 𝜛 ∈ 𝑁 𝐵× (O). It follows
(and can be checked directly) that 𝐼 = O𝜛 = 𝜛O is a two-sided O-ideal. If 𝑒 = 0, we
have 𝐼 = O.
Proposition 23.4.14. Suppose that 𝑒 ≥ 1. Then we have 𝑁 𝐵× (O)/(𝐹 × O× ) = h𝜛i '

Z/2Z. Moreover, the group Idl(O) = PIdl(O) is abelian, generated by 𝐼 and 𝔭O with
the single relation 𝐼 2 = 𝔭𝑒 O.
Proof. Let 𝛼 ∈ 𝑁 𝐵× (O). Then by uniqueness of the intersection in 23.4.13, conjuga-

tion by 𝛼 permutes these two orders, so we have a homomorphism 𝑁 𝐵× (O) to a cyclic
group of order 2. This homomorphism is surjective, since 𝜛 transposes the orders. If
𝛼 is in the kernel, then 𝛼 ∈ 𝑁 𝐵× (M2 (𝑅)) = 𝐹 × GL2 (𝑅) and unconjugating the second
factor we similarly get 𝜛𝛼𝜛 −1 ∈ 𝐹 × GL2 (𝑅), so
𝛼 ∈ 𝐹 × (GL2 (𝑅) ∩ 𝜛 −1 GL2 (𝑅)𝜛) = 𝐹 × O× .
Again since 𝑅 is local, we have Idl(O) = PIdl(O), and by (18.5.4), we have an

isomorphism
𝑁 𝐵× (O)/(𝐹 × O× ) ' PIdl(O)/PIdl(𝑅)
so Idl(O) is generated by 𝐼 and the generator 𝔭 for PIdl(𝑅).
23.4. EICHLER ORDERS 375
23.4.15. It is helpful to consider the Jacobson radical of an Eichler order, to compare

to the hereditary case.
𝔭 𝑅
𝐽= 𝑒 .
𝔭 𝔭
We claim that 𝐽 = rad O. We verify directly that 𝐽 ⊆ O is a two-sided ideal and
𝑓
𝔭 𝔭
𝐽 2 = 𝑒+1 ⊆ 𝔭O, (23.4.16)
𝔭 𝔭𝑓
where 𝑓 = min(𝑒, 2), so by Corollary 20.5.5, 𝐽 ⊆ rad O; on the other hand, the
quotient
O/𝐽 ' 𝑘 × 𝑘 (23.4.17)
is semisimple, so rad O ⊆ 𝐽 by Corollary 20.4.11(a).
However, the radical is not an invertible (sated) two-sided O-ideal unless O is
hereditary (𝑒 = 1), by 23.3.1. Indeed, we verify that
𝔭−1

𝑅
OL (𝐽) = 𝑒−1 = OR (𝐽) (23.4.18)
𝔭 𝑅
(Exercise 23.6); this recovers O if and only if 𝑒 = 1, and if 𝑒 ≥ 2 then it is an Eichler

order of level 𝔭𝑒−2 (conjugating as in (23.4.8)). By (23.4.16), if 𝑒 ≥ 2 then 𝐽 2 = 𝔭𝐽,
and so we certainly could not have 𝐽 invertible!
We now repackage these local efforts into a global characterization.
23.4.19. Suppose that 𝑅 is a global ring. Let disc𝑅 𝐵 = 𝔇 and let O be an Eichler
order with discrd O = 𝔑. If 𝔭 | 𝔇, then 𝐵𝔭 has a unique maximal order, so (as an
‘intersection’) O𝔭 is necessarily the maximal order. If 𝔭 - 𝔇, and ord𝔭 𝔑 = 𝑒 ≥ 0,
then O𝔭 is isomorphic to the standard Eichler order of level 𝔭𝑒 .
We have 𝔑 = 𝔇𝔐 with 𝔐 ⊆ 𝑅 and we just showed that 𝔐 is coprime to 𝔇. We
call 𝔐 the level of the Eichler order O. The pair 𝔇, 𝔐 (or 𝔇, 𝔑) determines a unique
genus of Eichler 𝑅-orders, i.e., this data uniquely determines the isomorphism class
of O𝔭 for each 𝔭.
Putting together Proposition 23.4.14 together with 23.3.19 for the remaining primes
where the order is maximal, we have an exact sequence
Ö
0 → Idl(𝑅) → Idl(O) → Z/2Z → 0 (23.4.20)
𝔭 |𝔑
and we may take the quotient by PIdl 𝑅 to get

Ö
0 → Cl 𝑅 → Pic𝑅 (O) → Z/2Z → 0 (23.4.21)
𝔭 |𝔑
Remark 23.4.22. Eichler [Eic56a] developed his orders in detail for prime level, and
employed them in the study of modular correspondences [Eic56c] and the trace formula
[Eic73] in the case of squarefree level. (As mentioned in Remark 24.1.5, Eichler’s
earlier work [Eic36, §6] over Q included some more general investigations, but his later
work seemed mostly confined to the hereditary orders.) Hĳikata [Hĳ74, §2.2] later
studied these orders in the attempt to generalize Eichler’s result beyond the squarefree
case; calling the orders split (like our name, residually split). Pizer [Piz73, p. 77] may
be the first who explicitly called them Eichler orders.
Remark 23.4.23. An 𝑅-order O ⊆ M𝑛 (𝐹) that contains a GL𝑛 (𝐹)-conjugate of the
diagonal matrices diag(𝑅, . . . , 𝑅) (equivalently, containing 𝑛 orthogonal idempotents)
is said to be tiled. By 23.4.3, an order O ⊆ M2 (𝐹) is tiled if and only if it is
Eichler. Tiled orders also go by other names (including graduated orders) and they
arise naturally in many contexts, including representation theory [Pl83] and modular
forms.
23.5 Bruhat–Tits tree

In the previous section, we examined Eichler orders as the intersection of two maximal
orders. There is a beautiful and useful combinatorial construction—a tree—which
keeps track of the containments among maximal orders in aggregate, as follows. For
further reading, see Serre [Ser2003, §II.1].
Let 𝐹 be a nonarchimedean local field with valuation ring 𝑅, maximal ideal 𝔭 = 𝜋𝑅,
and residue field 𝑘 := 𝑅/𝔭, and let 𝑞 := #𝑘. Let 𝐵 := M2 (𝐹), and let 𝑉 := 𝐹 2 as
column vectors, so that 𝐵 := End𝐹 (𝑉) acts on the left.
Recalling again section 10.5, every maximal order O ⊂ 𝐵 has O = End𝑅 (𝑀)
where 𝑀 ⊂ 𝑉 is an 𝑅-lattice. So to understand maximal orders, it is equivalent to
understand lattices (and their containments) and the positioning of one lattice inside
another.
Lemma 23.5.1. Let 𝐿, 𝑀 ⊂ 𝑉 be 𝑅-lattices. Then there exists an 𝑅-basis 𝑥1 , 𝑥2 of 𝐿

such that 𝜋 𝑓1 𝑥1 , 𝜋 𝑓2 𝑥2 is an 𝑅-basis of 𝑀 with 𝑓1 , 𝑓2 ∈ Z and 𝑓1 ≤ 𝑓2 .
Lemma 23.5.2. We have End𝑅 (𝐿) = End𝑅 (𝑀) if and only if there exists 𝑎 ∈ 𝐹 × such
that 𝑀 = 𝑎𝐿.
Proof. If 𝑀 = 𝑎𝐿 for 𝑎 ∈ 𝐹 × , then End𝑅 (𝐿) = End𝑅 (𝑀). Conversely, suppose

that End𝑅 (𝐿) = End𝑅 (𝑀). Replacing 𝑀 by 𝑎𝑀 with 𝑎 ∈ 𝐹 × , we may suppose
without loss of generality that 𝐿 ⊆ 𝑀. By Lemma 23.5.1, we may identify 𝐿 = 𝑅 2
with the standard basis and 𝑀 = 𝜋 𝑓1 𝑒 1 ⊕ 𝜋 𝑓2 𝑒 2 with 𝑓1 , 𝑓2 ∈ Z ≥0 ; rescaling again,
and interchanging the basis elements if necessary, we may suppose 𝑒 1 = 0. Then
End𝑅 (𝑀) = End𝑅 (𝐿) ' M2 (𝑅) implies 𝑓2 = 0 and 𝐿 = 𝑀.
With this lemma in mind, we make the following definition (recalling this definition
made earlier in the context of algebras).
Definition 23.5.3. Two 𝑅-lattices 𝐿, 𝐿 0 ⊂ 𝑉 are homothetic if there exists 𝑎 ∈ 𝐹 ×

such that 𝐿 0 = 𝑎𝐿.
23.5. BRUHAT–TITS TREE 377
The relation of homothety is an equivalence relation on the set of 𝑅-lattices in 𝑉,

and we write [𝐿] for the homothety class of 𝐿.
23.5.4. Let 𝐿 ⊂ 𝑉 be an 𝑅-lattice. In a homothety class of lattices, there is a unique
lattice 𝐿 0 ⊆ 𝐿 in this homothety class satisfying any of the following equivalent
conditions:
(i) 𝐿 0 ⊆ 𝐿 is maximal;
(ii) 𝐿 0 * 𝜋𝐿; and
(iii) 𝐿/𝐿 0 is cyclic as an 𝑅-module (has one generator).
These equivalences follow from Lemma 23.5.1: they are equivalent to 𝑓1 = 0, and
correspond to a maximal scaling of 𝐿 0 by a power of 𝜋 within 𝐿. For such an 𝐿 0, we
have 𝐿/𝐿 0 ' 𝑅/𝜋 𝑓 𝑅 for a unique 𝑓 ≥ 0.
Definition 23.5.5. Let T be the graph whose vertices are homothety classes of 𝑅-
lattices in 𝑉 and where an undirected edge joins two vertices (exactly) when there exist
representative lattices 𝐿, 𝐿 0 for these vertices such that
𝜋𝐿 ( 𝐿 0 ( 𝐿. (23.5.6)
Equivalently, by Lemma 23.5.2, the vertices of T are in bĳection with maximal

orders in 𝐵 = M2 (𝐹) by [𝐿] ↦→ End𝑅 (𝐿) for every choice of 𝐿 ∈ [𝐿].
23.5.7. The adjacency relation (23.5.6) implies 𝐿 0 ( 𝜋𝐿 ( 𝜋𝐿 0, so it is sensible to
have undirected edges.
A class [𝐿 0] has an edge to 𝐿 if and only if the representative 𝐿 0 in 23.5.4 has
𝑓 = 1.
Proposition 23.5.8. The graph T is a connected tree such that each vertex has degree
𝑞 + 1.
Proof. We have 𝐿/𝜋𝐿 ' 𝑘 2 , and so the lattices 𝐿 0 satisfying (23.5.6) are in bĳection
with 𝑘-subspaces of dimension 1 in 𝐿/𝜋𝐿; such a subspace is given by a choice of
generator up to scaling, so there are exactly (𝑞 2 − 1)/(𝑞 − 1) = 𝑞 + 1 such, and each
vertex has 𝑞 + 1 adjacent vertices. The graph is connected: given two vertices, we may
choose representative lattices 𝐿, 𝐿 0 such that 𝐿 0 ⊆ 𝐿 as in 23.5.4. The quotient 𝐿/𝐿 0
is cyclic, so by induction the lattices 𝐿 𝑖 = 𝜋 𝑖 𝐿 + 𝐿 0 for 𝑖 = 0, . . . , 𝑓 have 𝐿 𝑖 adjacent
to 𝐿 𝑖+1 , and 𝐿 0 = 𝐿 and 𝐿 𝑓 = 𝐿 0, giving a path from [𝐿] to [𝐿 0].
The following argument comes from Dasgupta–Teitelbaum [DT2008, Proposition
1.3.2]. Suppose there is a nontrivial cycle in T
𝜋 𝑣 𝐿 = 𝐿 𝑠 ( 𝐿 𝑠−1 ( · · · ( 𝐿 1 ( 𝐿 0 = 𝐿 (23.5.9)
so that 𝑣 ≥ 1. We may suppose this cycle is minimal, meaning that no intermediate

lattices are equivalent. The quotient 𝐿/𝐿 𝑠 = 𝐿/𝜋 𝑣 𝐿 ' (𝑅/𝔭 𝑣 ) 2 is not cyclic; let 𝑖 be
the largest index such that 𝐿/𝐿 𝑖 is cyclic but 𝐿/𝐿 𝑖+1 is not. Thus 𝐿/𝐿 𝑖+1 ' 𝑅/𝔭𝑖 ⊕ 𝑅/𝔭,
and so 𝜋 𝑖 annihilates 𝐿/𝐿 𝑖+1 and 𝜋 𝑖 𝐿 ⊆ 𝐿 𝑖+1 . Since 𝐿/𝐿 𝑖 is cyclic, just as in the
previous paragraph, we conclude 𝐿 𝑖−1 = 𝜋 𝑖−1 𝐿 + 𝐿 𝑖 . Putting these together, we find
that 𝜋𝐿 𝑖−1 = 𝜋 𝑖 𝐿 + 𝜋𝐿 𝑖 ⊆ 𝐿 𝑖+1 , the latter by definition of the adjacency between 𝐿 𝑖

and 𝐿 𝑖+1 . By adjacency, 𝐿 𝑖 ( 𝜋𝐿 𝑖1 ⊆ 𝐿 𝑖+1 . And again by adjacency, 𝐿 𝑖+1 is maximal
inside 𝐿 𝑖 , so 𝜋𝐿 𝑖−1 = 𝐿 𝑖+1 . This contradicts the minimality of the cycle; we conclude
that T has no cycles.
We call T the Bruhat–Tits tree for GL2 (𝐹). The Bruhat–Tits tree for 𝐹 = Q2 is
sketched in Figure 23.5.10. We write Ver(T ) and Edg(T ) for the set of vertices and
edges of T .
Figure 23.5.10: Bruhat–Tits tree for GL2 (Q2 )

23.5.11. We define a transitive action of GL2 (𝐹) on T as follows.
Let 𝐿 ⊆ 𝑉 be a lattice. Choose a 𝑅-basis for 𝐿 and put the vectors in the columns
of a matrix 𝛽 ∈ M2 (𝐹). Since these columns span 𝑉 over 𝐹, we have 𝛽 ∈ GL2 (𝐹),
and the matrix 𝛽 is well-defined up to a change of basis over 𝑅; therefore the coset
𝛽 GL2 (𝑅) ∈ GL2 (𝐹)/GL2 (𝑅) is well-defined. (Check that the action of change of
basis on columns is given by matrix multiplication the right.) Therefore a homothety
class [𝐿] gives a well-defined element of GL2 (𝐹)/(𝐹 × GL2 (𝑅)). Conversely, given
such a class we can consider the 𝑅-lattice spanned by its columns, and its homothety
class is well-defined. We have shown there is a bĳection
Ver(T ) ↔ GL2 (𝐹)/(𝐹 × GL2 (𝑅)). (23.5.12)
The group GL2 (𝐹) acts transitively on the left on the cosets GL2 (𝐹)/(𝐹 × GL2 (𝑅))
and we transport via the bĳection (23.5.12) to an action on Ver(T ).
We claim this action preserves the adjacency relation on T : if 𝐿 ⊇ 𝐿 0 are adjacent,
then by invariant factors we can choose a basis 𝑥1 , 𝑥2 for 𝐿 such that 𝑥1 , 𝜋𝑥2 is a basis
for 𝐿 0, i.e.,
1 0
𝛽0 = 𝛽 . (23.5.13)
0 𝜋
23.5. BRUHAT–TITS TREE 379
If 𝛼 ∈ GL2 (𝐹), then multiplying (23.5.13) on the left by 𝛼 shows that 𝛼𝐿, 𝛼𝐿 0 are
adjacent.
23.5.14. The tree T has a natural notion of distance 𝑑 between two vertices, given
by the length of the shortest path between them, giving each edge of T length 1.
Consequently, we have a notion of distance 𝑑 (O, O0) between every two maximal
orders O, O0 ⊆ 𝐵.
Lemma 23.5.15. Let 𝐿, 𝐿 0 be lattices with bases 𝑥 1 , 𝑥2 and 𝜋 𝑓 𝑥1 , 𝜋 𝑒+ 𝑓 𝑥2 , respectively.

0 0
Then
basis 𝑥 1 , 𝑥2 , we have O = End𝑅 (𝐿) ' M2 (𝑅) and O = End𝑅 (𝐿 ) '
in the
−𝑒
𝑅 𝔭
, and 𝑑 (O, O0) = 𝑒.
𝔭𝑒 𝑅
Proof. The statement on endomorphism rings comes from Example 10.5.2; we may
suppose up to homothety that 𝐿 0 has basis 𝑥1 , 𝜋 𝑒 𝑥2 ; the maximal lattices as in 23.5.4 are
given by 𝐿 𝑖 = 𝑅𝑥 1 + 𝔭𝑖 𝑅𝑥 2 with 𝑖 = 0, . . . , 𝑒, so the distance is 𝑑 ( [𝐿], [𝐿 0]) = 𝑒.
23.5.16. Importantly, now, we turn to Eichler orders: they are the intersection of two
unique maximal orders, and so correspond to a pair of vertices in T , or equivalently a
path. By Lemma 23.5.15, the standard Eichler order of level 𝔭𝑒 corresponds to a path
of length 𝑒, and by transitivity the same is true of every Eichler order. The normalizer
𝜛 of an Eichler order 23.4.13 acts by swapping the two vertices. Each vertex of the
path corresponds to the 𝑒 + 1 possible maximal superorders.
In this way, the Bruhat–Tits tree provides a visual way to keep track of many
calculations with Eichler orders.
Remark 23.5.17. The theory of Bruhat–Tits trees beautifully generalizes to become

the theory of buildings, pioneered by Tits; see the survey by Tits [Tit79], as well as
introductions by Abramenko–Brown [AB2008].
Exercises
Unless otherwise specified, let 𝑅 be a Dedekind domain with 𝐹 = Frac 𝑅 and let
O ⊆ 𝐵 be an 𝑅-order in a quaternion algebra 𝐵.
1. Let 𝑅 be a DVR with maximal ideal 𝔭 = (𝜋), and let

𝑅 𝑅
O= .
𝔭 𝑅

𝑥 𝑦
(a) Suppose that 𝛼 = ∈ 𝑁 𝐵× (O). After scaling, we may suppose
𝜋𝑧 𝑤
that 𝑥, 𝑦, 𝑧, 𝑤 ∈ 𝑅. By determinants, show that if 𝑥, 𝑤 ∈ 𝑅 × , then 𝛼 ∈ O× .
(b) Compute
𝑥 2 − 𝜋𝑦 2

𝜋(𝑤𝑦 − 𝑥𝑧)
𝛼𝜛𝛼 = .
𝜋(𝑤 2 − 𝜋𝑧2 ) −𝜋(𝑤𝑦 − 𝑥𝑧)
Show that 𝜋 | 𝑥, and then 𝜋 | 𝑤, whence 𝛼 ∈ 𝜛O.

0 1
(c) Conclude that 𝑁 𝐵× (O)/(𝐹 × O× ) is generated by 𝜛 = .
𝜋 0
[See also the matrix proof by Eichler [Eic56a, Satz 5], which uses a normal form
for one-sided ideals simplifying the above computation.]
I 2. Let 𝐾 ⊇ 𝐹 be a field extension, let 𝑆 be the integral closure of 𝑅 in 𝐾, and let
𝔞, 𝔟 ⊂ 𝐹 be fractional ideals of 𝑅. Show that 𝔞 = 𝔟 if and only if 𝔞𝑆 = 𝔟𝑆.
I 3. Extend Lemma 13.4.9, and show that if 𝑅 is a DVR and O is an Eichler 𝑅-order
then nrd(O× ) = 𝑅 × .
I 4. Suppose 𝑅 is local and O is a hereditary order. Show that if 𝜇 ∈ O has 𝜇2 = 𝜋,
then 𝜇 generates 𝑁 𝐵× (O)/(𝐹 × O× ).
5. Suppose 𝑅 is local and O ⊆ 𝐵 = M2 (𝐹) is the intersection of two maximal
orders. Give another (independent) proof that O is isomorphic to a standard
Eichler order which replaces matrix calculations in Proposition 23.4.3 with some
representation theory as follows.
(a) Write O ' M2 (𝑅) ∩ O0. Let 𝑒 11 be the top-left matrix unit and let
𝐼 = M2 (𝑅)𝑒 11 . Show 𝐼 0 = O0 𝑒 11 is an 𝑅-lattice in 𝑉 = M2 (𝐹)𝑒 11 ' 𝐹 2 .
(b) Use elementary divisors to show that there exists an 𝑅-basis 𝑥1 , 𝑥2 of 𝐼
such that 𝑥1 , 𝜋 𝑒 𝑥2 is an 𝑅-basis for 𝐼 0.
1 0
(c) Show that the corresponding change of basis matrix 𝛼 = has
0 𝜋𝑒
0
𝐼 = 𝛼𝐼, and use this to identify O with the standard Eichler order of level
𝔭𝑒 .
[See Brzezinski [Brz83a, Proposition 2.1].]
I 6. Let 𝑅 be local and let O be the standard Eichler order of level 𝔭𝑒 for 𝑒 ≥ 1. Let
𝐽 = rad O. Show that
𝔭−1

𝑅
OL (𝐽) = 𝑒−1 = OR (𝐽).
𝔭 𝑅
I 7. Prove Lemma 23.5.1. [Hint: use direct matrix methods or the theory of invariant
factors.]
8. Let 𝑅 be local and let O be a hereditary quaternion 𝑅-order. Show that rad O =
[O, O] is the commutator (cf. Exercise 13.8) and that diff O = rad O.
9. Let 𝑅 be local. Let O, O0 ⊆ 𝐵 be maximal 𝑅-orders. Recall that O, O0 are
vertices in the Bruhat–Tits tree. Define the distance dist(O, O0) to be the
distance in the Bruhat–Tits tree between the respective vertices. Show that
[O : O ∩ O0] = dist(O, O0) = [O0 : O ∩ O0].
10. Let 𝑅 be local, and let O be an Eichler order of level 𝔭𝑒 . Consider the graph
whose vertices are 𝑅-superorders O0 ⊇ O in 𝐵 and with a directed edge whenever
the containment O0 ) O is proper and minimal. What does this graph look like?
[Hint: use the Bruhat–Tits tree; it helps to draw the Eichler orders of the same
level at the same height.]
Chapter 24
Quaternion orders: second meeting
In this chapter, we continue our tour of quaternion orders with some more advanced
species.
24.1 ⊲ Advanced quaternion orders

Let 𝐵 be a quaternion algebra over Q of discriminant 𝐷 = disc 𝐵 and let O ⊂ 𝐵 be an
order with reduced discriminant 𝑁 = discrd(O). Then 𝑁 = 𝐷 𝑀 with 𝑀 ∈ Z ≥1 .
24.1.1 (Gorenstein and Bass orders). Although Eichler orders may lose the property
that all of its ideals are invertible, we may still insist on the invertibility of its dual.
Recall (Definition 15.6.15) that the codifferent of an order is the lattice codiff(O) = O♯
obtained as the dual of the trace pairing over 𝑅. We say O is Gorenstein if codiff(O)
is invertible, or equivalently (Corollary 16.8.7) every sated left or right fractional O-
ideal (lattice 𝐼 ⊆ 𝐵 with left or right order equal to O) is invertible. Hereditary orders
are Gorenstein, since for a hereditary order every left or right fractional O-ideal (not
necessarily sated) is invertible.
Being Gorenstein is a local property because invertibility is so. An Eichler order is
Gorenstein, but there are Gorenstein orders that are not Eichler. An order is Gorenstein
if and only if its associated ternary quadratic form is primitive, i.e. the greatest common
divisor of its coefficients is 1, or equivalently its values generate Z.
We say O is Bass if every superorder O0 ⊇ O (including O0 = O) is Gorenstein. A
Bass order is Gorenstein, but not always conversely. The fact that every superorder is
Gorenstein reflects into good structural properties of a Bass order. Most importantly,
a Z 𝑝 -order O is Bass if and only if it contains either Z 𝑝 × Z 𝑝 or the ring of integers
in a quadratic extension 𝐾 ⊇ Q 𝑝 (these order are sometimes called primitive; we call
them basic). This embedded subalgebra makes it possible to calculate explicitly with
the order, with important applications to the arithmetic of modular forms, a topic we
pursue in the final part of this book.
In summary, there is a chain of proper implications
maximal ⇒ hereditary ⇒ Eichler ⇒ Bass ⇒ Gorenstein (24.1.2)
381
382 CHAPTER 24. QUATERNION ORDERS: SECOND MEETING
for orders O ⊆ 𝐵, and each of these implications is strict (: in general).

Given an order O, we construct its radical idealizer as
O♮ = OL (rad O) = OR (rad O).
We have O ⊆ O♮ , and equality holds if and only if O is hereditary. Iterating, we obtain
a canonically attached sequence of superorders:
O = O0 ( O1 = O ♮ ( · · · ( O 𝑠 (24.1.3)
where O𝑠 is hereditary. A more refined classification of orders involves dissecting the
chain (24.1.3) explicitly.
There is one final way of classifying orders that extends nicely to the noncom-
mutative context, due to Brzezinski. By way of analogy, we √ recall that orders in a
quadratic field are characterized by conductor. Let 𝐾 = Q( 𝑑 𝐾 ), let Z𝐾 be the ring
of integers of 𝐾, and let 𝑑 𝐾 ∈ Z be the discriminant of Z𝐾 . An order 𝑆 in 𝐾 is of the
form 𝑆 = Z + 𝑓 Z𝐾 , where 𝑓 ∈ Z ≥1 is the conductor of 𝑆 (as in section 16.1), and thus
the discriminant of 𝑆 is 𝑑 = 𝑓 2 𝑑 𝐾 . Even in classical considerations, these orders arise
naturally when considering binary quadratic forms of nonfundamental discriminant.
Proposition 24.1.4. Let 𝐵 be a quaternion algebra over Q and let O ⊆ 𝐵 be an order.
Then there exists a unique integer 𝑓 (O) ≥ 1 and Gorenstein order Gor(O) such that
O = Z + 𝑓 (O) Gor(O).
Two orders O, O0 are isomorphic if and only if 𝑓 (O) = 𝑓 (O0) and Gor(O) ' Gor(O0).
The order Gor(O) is called the Gorenstein saturation of O, and we call 𝑓 (O) the
Gorenstein conductor of the order (also sometimes called the Brandt invariant): an
order is Gorenstein if and only if 𝑓 (O) = 1, so this gives a ready supply of orders that
are not Gorenstein. In particular, to classify all orders, via the operation of Gorenstein
saturation, it is enough to classify the Gorenstein orders.
Remark 24.1.5. The first attempt to tame the zoo of quaternion orders was Eichler
[Eic36, Satz 12], who classified what he called primitive (our basic) orders. Later
this was generalized by Brzezinski [Brz83a, §5], who also clarified certain aspects
[Brz90, §1]. A nice summary of facts about quaternion orders is given by Lemurell
[Lem2011].
24.2 Gorenstein orders

In this section we define the well-behaved Gorenstein orders. See Remark 24.2.24 for
more context on the class of Gorenstein rings.
Recall the definition of the codifferent (Definition 15.6.15):
codiff(O) = O♯ := {𝛼 ∈ 𝐵 : trd(𝛼O) ⊆ 𝑅} ⊆ 𝐵.
We have codiff (O) a two-sided sated O-ideal with O ⊆ codiff(O) (Lemma 15.6.16),
and disc(O) = [codiff(O) : O] 𝑅 . We already saw in section 16.8 the importance of
the following class of orders.
24.2. GORENSTEIN ORDERS 383
Definition 24.2.1. O is Gorenstein if codiff(O) is invertible.

24.2.2. The property of being a Gorenstein order is local (O is Gorenstein if and only
if O𝔭 is Gorenstein for all primes 𝔭), since invertibility is a local property.
(i) O is Gorenstein;
(ii) codiff (O) is projective as a O-bimodule;
(iii) O∨ = Hom𝑅 (O, 𝑅) is projective as an O-bimodule; and
(iv) All sated left or right fractional O-ideals are invertible as O-ideals.
Proof. For the equivalence (i) ⇔ (ii), because codiff (O) is sated it follows from
Theorem 20.3.3 that O is Gorenstein if and only if codiff (O) is projective as an O-
bimodule. For (ii) ⇔ (iii), by Proposition 15.6.7, we have an isomorphism codiff (O) '
Hom𝑅 (O, 𝑅) of O-bimodules over 𝑅. Finally, (i) ⇔ (iv) follows from Corollary
16.8.7.
24.2.4. We call in for relief as well Main Theorem 20.3.9: the equivalent sided notions
(on the left and right) in Proposition 24.2.3 are also all equivalent. In particular, a
suitably defined notion of left Gorenstein or right Gorenstein would also be equivalent.
The Gorenstein condition can be detected on the level of norms as follows.
nrd(codiff (O)) discrd(O) ⊆ 𝑅,
and O is Gorenstein if and only if equality holds.
Proof. We refer to Proposition 16.4.3, and 16.4.5: we have
[O : O♯ ] 𝑅 ⊇ Nm 𝐵 |𝐹 (O♯ ) = nrd(O♯ ) 2 ,
with equality if and only if O♯ is locally principal. But by Lemma 15.6.17, we have
[O♯ : O] 𝑅 = disc(O) = discrd(O) 2 , and combining these gives the result.
Our next main result connects the Gorenstein condition to a property of the corre-
sponding ternary quadratic module. Let 𝑄 : 𝑀 → 𝐿 be a ternary quadratic module.
We follow Gross–Lucianovic [GrLu2009, Propositions 6.1–6.2], and consider the odd
Clifford bimodule.
Proposition 24.2.6. Left multiplication gives a pairing
Clf 0 (𝑄) × Clf 1 (𝑄) → Clf 1 (𝑄)/𝑀 ' 𝑀 ⊗ 𝐿∨

Ó3
that induces an isomorphism
Hom𝑅 (Clf 0 (𝑄),

Ó3 ∼ Clf 1 (𝑄)
𝑀 ⊗ 𝐿 ∨ ) −→ (24.2.7)
of left Clf 0 (𝑄)-modules.

We have a similar argument on the right.

Remark 24.2.8. If 𝑅 is a Dedekind domain, then choosing a pseudobasis for 𝑀 we may
write 𝑀 = 𝔞𝑒 1 + 𝔟𝑒 2 + 𝔠𝑒 3 for fractional Ó
ideals 𝔞, 𝔟, 𝔠, and without loss of generality
may take 𝐿 = 𝔩 a fractional ideal. Then 3 𝑀 = 𝔞𝔟𝔠(𝑒 ∧ 𝑒 ∧ 𝑒 ) ' 𝔞𝔟𝔠 is just the
1 2 3
𝑀 ⊗ 𝐿 ' 𝔞𝔟𝔠𝔩−1 as in (22.5.10). Restricting to this case
∨
Ó3
Steinitz class of 𝑀, and
is still quite illustrative.
Proof. By construction of the Clifford algebra, we have as 𝑅-modules that
Clf 1 (𝑄) ' 𝑀 ⊕ 3 𝑀 ⊗ 𝐿 ∨ ,

Ó
so Clf 1 (𝑄)/𝑀 ' 3 𝑀 ⊗ 𝐿 ∨ . Multiplication in Clf(𝑄) induces a pairing that induces

Ó
a homomorphism of Clf 0 (𝑄)-bimodules by associativity of multiplication in Clf(𝑄).
To conclude that the pairing induces an isomorphism, we can argue locally and suppose
that 𝑀 ' 𝑅 3 with basis 𝑒 1 , 𝑒 2 , 𝑒 3 and 𝐿 ' 𝑅 is the quadratic form
𝑄(𝑥, 𝑦, 𝑧) = 𝑎𝑥 2 + 𝑏𝑦 2 + 𝑐𝑧 2 + 𝑢𝑦𝑧 + 𝑣𝑥𝑧 + 𝑤𝑥𝑦.
Then 1, 𝑒 2 𝑒 3 , 𝑒 3 𝑒 1 , 𝑒 1 𝑒 2 is an 𝑅-basis for Clf 0 (𝑄) and
Clf 1 (𝑄) = (𝑅𝑒 1 + 𝑅𝑒 2 + 𝑅𝑒 3 ) + 𝑅𝑒 1 𝑒 2 𝑒 3
with Clf 1 (𝑄)/𝑀 ' 𝑅𝑒 1 𝑒 2 𝑒 3 .

We then compute the dual basis of Clf 1 (𝑄) to that of Clf 0 (𝑄) as
𝑒 1 𝑒 2 𝑒 3 − 𝑢𝑒 1 − 𝑣𝑒 2 − 𝑤𝑒 3 , 𝑒 1 , 𝑒 2 , 𝑒 3 ; (24.2.9)
for example,
𝑒 2 𝑒 3 (𝑒 1 𝑒 2 𝑒 3 ) = 𝑒 2 (−𝑒 1 𝑒 3 + 𝑣)𝑒 2 𝑒 3 ≡ −𝑒 2 𝑒 1 𝑒 3 (𝑒 2 𝑒 3 ) (mod 𝑀)
so
𝑒 2 𝑒 3 (𝑒 1 𝑒 2 𝑒 3 − 𝑢𝑒 1 − 𝑣𝑒 2 − 𝑤𝑒 3 ) ≡ 𝑒 1 (𝑒 2 𝑒 3 ) 2 − 𝑢𝑒 1 𝑒 2 𝑒 3
≡ 𝑒 1 (𝑢𝑒 2 𝑒 3 ) − 𝑢𝑒 1 𝑒 2 𝑒 3 = 0 (mod 𝑀).
The other products can be computed in a similarly direct fashion.
Recall (Definition 9.7.14) that 𝑄 : 𝑀 → 𝐿 is primitive if 𝑄(𝑀) generates 𝐿 as an

𝑅-module.
Theorem 24.2.10. O is Gorenstein if and only if its associated ternary quadratic
module 𝑄 = 𝜓O is primitive.
Proof. The statement is local, so we may Ó

suppose that 𝑅 is a local domain with maximal
ideal 𝔭, and that 𝑀 ' 𝑅 3 and 𝐿 ' 𝑅, so 3 𝑀 ⊗ 𝐿 ∨ ' 𝑅. Let 𝐽 = Clf 1 (𝑄) be the odd
Clifford bimodule, thought of as a left O-module. By Proposition 24.2.6 (specifically,
(24.2.7)), we have Clf 1 (𝑄) ' Hom𝑅 (O, 𝑅) as left O-modules. By Proposition 24.2.3
and 24.2.4 (or repeating the argument on the right), we want to show that 𝐽 is principal.
Suppose that 𝑄 is primitive. Let 𝛼 = 𝑥𝑒 1 + 𝑦𝑒 2 + 𝑧𝑒 3 . We then compute that

1 𝑥 0 𝑐𝑧 + 𝑣𝑥 −(𝑏𝑦 + 𝑢𝑧) 𝑒
ª© 1 ª
−𝑐𝑧 (𝑎𝑥 + 𝑣𝑧 + 𝑤𝑦) ® 𝑒 2 ®
© ª ©
𝑒 2 𝑒 3 ® 𝑦 𝑣𝑦
𝛼 ®=
(𝑏𝑦 + 𝑢𝑧) −(𝑎𝑥 + 𝑤𝑦) 0
® ®
𝑒 3 𝑒 1 ® 𝑧 ® 𝑒3 ®
« 𝑒 1 𝑒 2 ¬ «0 𝑥 𝑦 𝑧 ¬ «𝑒 1 𝑒 2 𝑒 3 ¬
and the determinant of the matrix in the middle is precisely 𝑄(𝑥, 𝑦, 𝑧) 2 . So 𝑄(𝛼) ∈ 𝑅 ×
if and only if 𝛼O = 𝐽.
Conversely, suppose that 𝑄 is not primitive. Then 𝑄 ≡ 0 (mod 𝔭). If 𝛼 =
𝑥𝑒 1 + 𝑦𝑒 2 + 𝑧𝑒 3 + 𝑡𝑒 1 𝑒 2 𝑒 3 with 𝑥, 𝑦, 𝑧, 𝑡 ∈ 𝑅, then
𝛼𝑒 2 𝑒 3 ≡ 𝑥𝑒 1 𝑒 2 𝑒 3 (mod 𝔭)
and symmetrically with the other products, so 𝛼O has rank ≤ 2 over 𝑅/𝔭, and it follows
that 𝛼O ≠ 𝐽 for all 𝛼.
Corollary 24.2.11. Let 𝑅 be a Dedekind domain. Then the associations
( Nondegenerate primitive ternary ) ( Gorenstein quaternion )
quadratic modules over 𝑅 ↔ orders over 𝑅
up to twisted similarity up to isomorphism
(24.2.12)
𝑄 ↦→ Clf 0 (𝑄)
nrd♯ (O) ← O
are mutually inverse discriminant-preserving bĳections that are also functorial with
respect to 𝑅.
Proof. We restrict the bĳection in Main Theorem 22.5.7 and apply Theorem 24.2.10.

Remark 24.2.13. In view of Corollary 24.2.11, the issues in the correspondence with
ternary quadratic forms for non-Gorenstein orders amounted to the failure to account
for the codomain of the quadratic module: non-Gorenstein orders are obtained from
quadratic modules 𝑄 : 𝑀 → 𝐿 where 𝑄(𝑀) ( 𝐿.
24.2.14. From 23.4.12 and Theorem 24.2.10, we conclude that every Eichler order is
Gorenstein; a direct proof is given in Exercise 24.1.
Therefore, non-Gorenstein orders abound: indeed, any order corresponding to
an imprimitive form will do. More generally, we construct a canonically associated
Gorenstein order containing a given order as follows.
Proposition 24.2.15. There exists a unique ideal 𝔣(O) ⊆ 𝑅 and unique Gorenstein
order Gor(O) ⊆ 𝐵 such that
O = 𝑅 + 𝔣(O) Gor(O). (24.2.16)
In fact, we have
𝔣(O) = discrd(O) nrd(O♯ )
Gor(O) = 𝑅 + nrd(O♯ ) −1 O♯ O♯ .
Before we begin the proof, we make a definition and then consider the key ingredient
of the proof: how rescaling the module affects the even Clifford algebra.
Definition 24.2.17. In (24.2.16), we call 𝔣(O) ⊆ 𝑅 the conductor and Gor(O) the
Gorenstein saturation of O.
Remark 24.2.18. Brzezinski [Brz83a] writes 𝔟(O) instead of 𝔣(O) and calls it the
Brandt invariant. Other authors call Gor(O) the Gorenstein closure of O, but this
terminology may be confusing: Gor(O) is not necessarily the smallest Gorenstein
order containing O (see Exercise 24.11).
24.2.19. Suppose that the ternary qudratic module 𝑄 : 𝑀 → 𝐿 corresponds to a

quaternion 𝑅-order O. For a nonzero ideal 𝔞 ⊆ 𝑅, we define 𝑄(𝔞) : 𝑀 → 𝔞−1 𝐿 to
be just 𝑄 but with values taken in 𝔞−1 ⊇ 𝐿. We claim that for all 𝑅-ideals 𝔞, we have
under the correspondence
𝑄(𝔞) : 𝑀 → 𝔞−1 𝐿 O(𝔞) := 𝑅 + 𝔞O. (24.2.20)
The fact that Clf 0 (𝑄 [𝔞]) = 𝑅 + 𝔞O is visible from the construction of the even Clifford
algebra (22.2.3); it is also visible from the description (22.3.25) in Example 22.3.24.
In the other direction, we have
nrd♯ (O(𝔞)) : (O(𝔞) ♯ ) 0 = 𝔞−1 (O♯ ) 0 → discrd(O(𝔞)) −1 = 𝔞−3 discrd(O) −1
and now we twist by 𝔞 to get
𝔞 ⊗ nrd(O(𝔞) ♯ ) : (O♯ ) 0 → 𝔞−1 discrd(O) −1 . (24.2.21)
Proof of Proposition 24.2.15. We argue using ternary quadratic modules: our proof
amounts to replacing a potentially imprimitive form with a primitive form, following
Theorem 24.2.10.
Let 𝑄 = 𝜓O : 𝑀 → 𝐿 be the ternary quadratic module associated to O, well-
defined up to twisted similarity. We may take 𝐿 = 𝔩 ⊆ 𝐿 ⊗𝑅 𝐹 ' 𝐹 and we do so for
concreteness; accordingly, suppose 𝔩 is a fractional ideal of 𝑅. Then 𝑄(𝑀) = 𝔫 ⊆ 𝔩 is a
finitely generated nonzero 𝑅-submodule; since 𝑅 is a Dedekind domain, 𝔫 is invertible.
Let
𝔣 = 𝔣(O) := 𝔫𝔩−1 ⊆ 𝑅.
Let Gor(𝑄) = 𝑄(𝔣) : 𝑀 → 𝔫 be the primitive ternary quadratic module obtained
by restricting the codomain. Then Clf 0 (Gor(𝑄)) is a Gorenstein order by Theorem
24.2.10, and
Clf 0 (𝑄) = 𝑅 + 𝔣(O) Clf 0 (Gor(𝑄));
by (24.2.20), so we let Gor(O) := Clf 0 (Gor(𝑄)). Uniqueness follows directly from
(24.2.20): if O = 𝑅 + 𝔞O0 and O0 is Gorenstein, then 𝑄(𝔞−1 ) : 𝑀 → 𝔞𝔩 is primitive
and 𝑄(𝑀) = 𝔫 = 𝔞𝔩, thus 𝔞 = 𝔣(O) and O0 = Clf 0 (𝑄(𝔞)) = Gor(O).
To prove the remaining statements, we recall Corollary 22.4.15 to get
O = 𝑅 + discrd(O)O♯ O♯
in all cases, and the primitivity of
nrd(O(𝔣−1 ) ♯ ) ⊗ 𝔣−1 : (O♯ ) 0 → 𝔣 discrd(O) −1
as in (24.2.21) is equivalent to
nrd((O♯ ) 0 ) = 𝔣 discrd(O) −1
i.e. 𝔣 = discrd(O) nrd((O♯ ) 0 ). Finally, nrd(O♯ ) = nrd((O♯ ) 0 ) is proven in Exercise

22.10.
Lemma 24.2.22. Let O0 ⊆ 𝐵 be an 𝑅-order. Then O ' O0 as 𝑅-orders if and only if

Gor(O) ' Gor(O0) and 𝔣(O) = 𝔣(O0).
Proof. Immediate from the uniqueness claim in Proposition 24.2.15.
The translation of the Gorenstein property in terms of primitivity of the ternary

quadratic form makes it quite accessible. For example, we have the following result
that shows that the Gorenstein condition is stable under base change.
Proposition 24.2.23. Let 𝐾 ⊇ 𝐹 be a finite extension and let 𝑆 be the integral closure
of 𝑅 in 𝐾. Then O is a Gorenstein 𝑅-order if and only if O ⊗𝑅 𝑆 is a Gorenstein
𝑆-order.
Proof. Let 𝑄 : 𝑀 → 𝐿 be the ternary quadratic module corresponding to O; denoting

extension of scalars by subscripts, we have 𝑄 𝑆 : 𝑀𝑆 → 𝐿 𝑆 corresponding to O𝑆 .
We want to show that 𝑄 is primitive if and only if 𝑄 𝑆 is primitive, which is to say
𝑄(𝑀) = 𝐿 if and only if 𝑄 𝑆 (𝑀) = 𝑄(𝑀) 𝑆 = 𝐿 𝑆 , and this statement is true as it holds
for fractional 𝑅-ideals (Exercise 23.2).
Remark 24.2.24. Gorenstein rings were introduced by Gorenstein [Gor52] in the

context of plane curves (the results of his Ph.D. thesis); Bass [Bas62, Footnote 2]
writes: “After writing this paper I discovered from Professor Serre that these rings
have been encountered by Grothendieck, the latter having christened them ‘Gorenstein
rings.’ They are described in his setting by the fact that a certain module of differentials
is locally free of rank one.” Bass [Bas63] gives a survey of (commutative) Gorenstein
rings, noting their ubiquity; see also the later survey by Huneke [Hun99]. Gorenstein
rings are truly abundant: they include coordinate rings of affine plane curves and
curves with only double points as singularities, complete intersections, and integral
group rings of finite groups.
The above-mentioned paper of Bass [Bas63] also gave rise to the class of epony-
mous orders in which every superorder is Gorenstein; over a complete DVR, these
orders were completely classified (and related to hereditary and Gorenstein orders) by
Drozd–Kirichenko–Roiter [DKR67]. Gorenstein orders in the context of quaternion
algebras, were studied by Brzezinski [Brz82]. (Brzezinski [Brz87] also considers more
general orders in which every lattice is locally principal: but for quaternion algebras,
these are again the Gorenstein orders.)
24.3 Eichler symbol

Just as local quadratic extensions are classified as being either ramified, inert, or split,
it is helpful to have a similar classification for quaternion orders.
We first work locally, and we suppose until further notice that 𝑅 is local (as in
23.2.1).
24.3.1. The 𝑘-algebra O/rad O is semisimple and has a standard involution (since
rad O is preserved by it), so by Example 7.4.9 this standard involution is nondegenerate;
by classification (Main Theorem 4.4.1 and Theorem 6.4.1) we have one of three
posibilities: O/rad O is either 𝑘, a separable quadratic 𝑘-algebra, or a quaternion
algebra over 𝑘.
We give symbols to each of the possibilities in 24.3.1 as follows.
Definition 24.3.2. Let 𝐽 := rad O. We define the Eichler symbol



 ∗, if O/𝐽 is a quaternion algebra;


O



 1, if O/𝐽 ' 𝑘 × 𝑘, and we say O is residually split;
:= 0, if O/𝐽 ' 𝑘, and we say O is residually ramified; and
𝔭 

 if O/𝐽 is a (separable) quadratic field extension of 𝑘, and
 −1,



 we say O is residually inert.

O
For formatting reasons, we will also write = (O | 𝔭). The similarity of
𝔭
the Eichler symbol to other quadratic-like symbols is intentional: because of the
arguments to the symbol, it should not be confused with the others. If the reader
finds this overloading of symbols unpleasant, they may wish to use the symbol 𝜀𝔭 (O)
instead.
24.3.3. Recall the definition of the discriminant quadratic form
Δ:𝐵→𝐹
(24.3.4)
Δ(𝛼) = trd(𝛼) 2 − 4 nrd(𝛼)
that computes the discriminant of 𝐹 [𝛼] = 𝐹 [𝑥]/(𝑥 2 − trd(𝑥)𝑥 + nrd(𝑥)) in the basis
1, 𝛼. The form factors through a map Δ : 𝐵/𝐹 → 𝐹.
𝑎
Suppose that #𝑘 < ∞. For 𝑎 ∈ 𝑅, let denote the generalized Kronecker sym-
𝔭
bol, defined to be 0, 1, −1 according as if 𝐹 [𝑥]/(𝑥 2 − 𝑎) is ramified, split (isomorphic
𝑎
to 𝐹 × 𝐹), or inert. If char 𝑘 ≠ 2, then is the Legendre symbol. We then have the
𝔭
following characterization (Exercise 24.4):

O Δ(𝛼)
(a) = ∗ if and only if takes on all of the values −1, 0, 1 for 𝛼 ∈ O.
𝔭 𝔭
O Δ(𝛼)
(b) = 𝜖 if and only if takes the values {0, 𝜖 } for 𝛼 ∈ O.
𝔭 𝔭
24.3. EICHLER SYMBOL 389
We now consider each of possible values of the Eichler symbol in turn.

O
24.3.5. We have = ∗ if and only if rad O = 𝔭O, by dimension considerations. If
𝔭
O
further #𝑘 < ∞, then = ∗ if and only if O ' M2 (𝑅), since then the only quaternion
𝔭
algebra over 𝑘 is M2 (𝑘), and we can lift matrix units using Hensel’s lemma.
Lemma 24.3.6. The order O is Eichler if and only if O is maximal or residually split.
Proof. If O is Eichler, then either O is maximal or O/rad O ' 𝑘 × 𝑘 by (23.4.17).

Conversely, a maximal order is an Eichler order by definition, so suppose O is residually
split, i.e., O/rad O ' 𝑘 × 𝑘. Then (as in Lemma 20.6.8), the nontrivial orthogonal
idempotents of 𝑘 × 𝑘 lift to orthogonal idempotents 𝑒 1 , 𝑒 2 ∈ O. Since 𝑒 1 𝑒 2 = 0,
immediately 𝐵 ' M2 (𝐹). Conjugating in 𝐵, we may suppose 𝑒 1 = 𝑒 11 and therefore
𝑒 2 = 𝑒 22 , and then the result follows from Proposition 23.4.3.
For fun, here is a second proof. We have a decomposition O = 𝐼1 ⊕ 𝐼2 = O𝑒 1 ⊕ O𝑒 2 .
Thus O = OL (𝐼1 ) ∩ OL (𝐼2 ). Tensoring up to 𝐹, we get 𝐵 = (𝐼1 ) 𝐹 ⊕ (𝐼2 ) 𝐹 , so
by dimensions we have (𝐼1 ) 𝐹 ' (𝐼2 ) 𝐹 ' 𝐹 2 the simple 𝐵-module, and each 𝐼𝑖 is
isomorphic to an 𝑅-lattice 𝑀𝑖 ⊆ 𝐹 2 . Thus OL (𝐼𝑖 ) ' End𝐹 (𝑀𝑖 ) ' M2 (𝑅), and each
OL (𝐼𝑖 ) is maximal, so O is the intersection of two maximal orders.
24.3.7. Residually inert orders in a division quaternion algebra 𝐵 over Q 𝑝 were studied
by Pizer [Piz76b], and he described them as follows. Let 𝐾 = Q 𝑝2 be the unramified
extension of Q 𝑝 . Then 𝐾 ↩→ 𝐵. Consider the left regular representation 𝐵 → M2 (𝐾);
it has image
𝑧 𝑤
, with 𝑧, 𝑤 ∈ 𝐾 (24.3.8)
𝑝𝑤 𝑧
The valuation ring of 𝐵 consists of those with 𝑧, 𝑤 ∈ Z 𝑝2 where Z 𝑝2 is the valuation
ring of 𝐾. Pizer then considers those orders with 𝑧 ∈ Z 𝑝2 and 𝑝 𝑟 | 𝑤. He [Piz80a,
Remark 1.5, Proposition 1.6] connected the residually inert and residually split orders
by noting the striking resemblance between (24.3.8) and the standard Eichler order,
remarking:
Thus O0𝑝 and O 𝑝 [the Eichler order and the Pizer order] are both essentially

𝑅 𝑅
subrings of 2𝑟 +1 fixed by certain (different!) Galois actions induced
𝑝 𝑅
by the Galois group of 𝐿/Q 𝑝 and thus they can be viewed as twisted
versions of each other. Hence O and O0 are locally isomorphic at all
primes 𝑞 ≠ 𝑝 while at 𝑝 they are almost isomorphic. Thus it should not
be too surprising that there are close connections between [them].
Pizer works explicitly and algorithmically [Piz80a] with residually inert orders, with
applications to computing modular forms of certain nonsquarefree level.
24.3.9. We can also interpret the Eichler symbol in terms of the reduction of the
associated ternary quadratic form 𝑄.
We have (O | 𝔭) = ∗ if and only if 𝑄 mod 𝔭 is nondegenerate, defining a smooth

conic over 𝑘.
If (O | 𝔭) = 1, then by Lemma 24.3.6, O is Eichler, and 𝑄 mod 𝔭 ∼ 𝑄(𝑥, 𝑦) = 𝑥𝑦
is degenerate of rank 2 by 23.4.12, cutting out two intersecting lines over 𝑘; and
conversely.
Suppose (O | 𝔭) = −1. Let 𝑖 ∈ O generate the quadratic field ℓ = O/rad O over
𝑘, let 𝐾 = 𝐹 (𝑖) and let 𝑆 be the integral closure of 𝑅 in 𝐾. Then 𝐾 is unramified
over 𝐹, and 𝔭𝑆 is the maximal ideal of 𝑆. Now the 𝑆-order O𝑆 := O ⊗𝑅 𝑆 has
O𝑆 /rad O𝑆 ' ℓ × ℓ, and (O𝑆 | 𝔭𝑆) = 1. Therefore 𝑄 mod 𝔭 over ℓ is degenerate
of rank 2, so the same is true over 𝑘, and since we are not in the previous case, it is
defined by an irreducible quadratic polynomial (the norm form from ℓ to 𝑘). Therefore
𝑄 mod 𝔭 cuts out two lines defined over ℓ and conjugate under Gal(ℓ | 𝑘). In particular,
a residually inert order is Gorenstein.
The only possibilities that remain are that 𝑄 mod 𝔭 is identically zero or has rank
1 (defined by the square of a linear factor), and correspondingly cuts out the whole
projective plane or a double line. These are the cases (O | 𝔭) = 0.
24.3.10. It follows from 24.3.9 that the ternary quadratic form associated to a residually
inert order is similar to
𝑄(𝑥, 𝑦, 𝑧) = 𝜋 𝑒 𝑥 2 + Nm𝐾 |𝐹 (𝑧 + 𝑦𝑖)
just as in (23.2.7); we have 𝑒 odd if and only if 𝐵 is a division algebra and 𝑒 even if
and only if 𝐵 is split.
The residually ramified orders do not admit such a simple classification; we will
pursue them further in the coming sections.
Remark 24.3.11. Let 𝑅 be a DVR, let O be a Gorenstein 𝑅-order, and let 𝑄 be a ternary
quadratic form over 𝑅 representing the similarity class associated to O. Then 𝑄 is
primitive and so defines an integral model C of a conic 𝐶 ⊆ P2 over 𝐹 = Frac 𝑅. By
discriminants, this conic has good reduction if and only if O is maximal. Moreover, we
saw by direct calculation that this conic has the simplest kind of bad reduction—regular,
with just one node over 𝑘—if and only if O is hereditary. This is no coincidence: in
fact, C is normal if and only if C is Bass. See Brzezinski [Brz80] for more on the
relationship between integral models of conics and quaternion orders and the follow-up
work [Brz85] where the increasing sequence of Bass orders ending in an hereditary
order corresponds to a sequence of elementary blowup transformations.
Another motivation to study the Eichler symbol is that it controls the structure of
unit groups, as follows.
Lemma 24.3.12. Let #𝑘 = 𝑞. Then 1 + 𝔭O ⊆ O× , and


 𝑞(𝑞 − 1) 2 (𝑞 + 1), if (O | 𝔭) = ∗;
 𝑞 2 (𝑞 − 1) 2 ,

if (O | 𝔭) = 1;


[O× : 1 + 𝔭O] = 2 2
 𝑞 (𝑞 − 1),
 if (O | 𝔭) = −1;

 𝑞 3 (𝑞 − 1),

if (O | 𝔭) = 0.

24.4. CHAINS OF ORDERS 391
Proof. Since 𝐽 ⊆ 𝔭O, if 𝜇 ∈ 1 + 𝔭O then 𝜇 − 1 is topologically nilpotent, hence

𝜇 ∈ O× .
Now the indices. If (O | 𝔭) = ∗, then we are computing the cardinality # GL2 (𝑘) =
(𝑞 2 − 1) (𝑞 2 − 𝑞). If (O | 𝔭) = 1, then O/𝐽 ' 𝑘 × 𝑘 and dim 𝑘 𝐽/𝔭O = 2, so the index
is (𝑞 − 1) 2 𝑞 2 . Similarly if (O | 𝔭) = −1, we get (𝑞 2 − 1)𝑞 2 . If (O | 𝔭) = 0, then
O/𝐽 ' 𝑘 and so the index is (𝑞 − 1)𝑞 3 .
We conclude with a global definition, and so we restore 𝑅 to a Dedekind domain.
Definition 24.3.13. For a nonzero prime 𝔭 ⊆ 𝑅, we define the Eichler symbol at 𝔭

to be
O O𝔭
:= ,
𝔭 𝔭𝑅𝔭
i.e., the Eichler symbol of the completion at 𝔭.
O
We say O is locally residually inert if ∈ {∗, −1} for all primes 𝔭.
𝔭
The analogously defined locally residually split orders already have a name: they
are the Eichler orders that are not maximal.
24.4 Chains of orders

We have a few more classes of orders to consider, but before we continue our tour we
pause to consider an aspect of the more general classification: we seek to put every
order in a chain of superorders, ending in a maximal order.
Suppose throughout this section that 𝑅 is local.
Definition 24.4.1. The radical idealizer of O is
O♮ := OL (rad O) ∩ OR (rad O).
24.4.2. By Exercise 20.2, we have OL (rad O) = OR (rad O) = O♮ , so the symmetric

definition can be replaced by either order in the intersection.
24.4.3. We recall our motivation to study extremal orders 21.2.1: O♮ radically covers
O, and by Proposition 21.2.3 we have O♮ = O if and only if O is extremal. By Theorem
21.5.1, O is extremal if and only if O is hereditary. Iterating, we have a canonically
associated chain of orders
O = O0 ( O1 = O♮ ( · · · ( O𝑟 (24.4.4)
terminating in an order O𝑟 that is hereditary. (By 23.3.1 and Proposition 23.4.3(iv),

either O𝑟 is maximal or O𝑟 is contained in exactly two possible maximal orders.) We
call O𝑟 the hereditary closure of O.
24.4.5. Suppose (O | 𝔭) = 1, i.e., O is an Eichler order (Lemma 24.3.6). Suppose

O has level 𝔭𝑒 . Then O♮ is an Eichler order of level 𝔭𝑒−2 from (23.4.18)—it had
to be Eichler of some level by Corollary 23.4.10. So the chain (24.4.4) is of length
b𝑒/2c with quotients dim 𝑘 (O𝑖 /O𝑖+1 ) = 2. On the Bruhat–Tits tree, the Eichler order
corresponds to a path of length 𝑒 by 23.5.16, and O♮ is the path of length 𝑒 − 2 obtained
by plucking away the vertices on the ends. (If desired, one can refine this chain by
squeezing in an extra Eichler order in between each step.)
24.4.6. If O = 𝑅 + 𝔭O0 for an order O0, then O♮ = O0, by Exercise 24.5.
In general, write O = 𝑅 +𝔭 𝑓 Gor(O) where 𝔭 𝑓 = 𝔣(O) is the Gorenstein conductor
of O and Gor(O) is the Gorenstein saturation as in Proposition 24.2.15. Then the chain
of radical idealizers begins
O ( O1 = 𝑅 + 𝔭 𝑓 −1 Gor(O) ( · · · ( O 𝑓 = Gor(O).
For each 𝑖, we have dim 𝑘 (O𝑖 /O𝑖+1 ) = 3.
We next consider the chain of superorders over a (local) residually inert order.
Proposition 24.4.7. Let O be a residually inert 𝑅-order. Then the following statements
hold.
(a) rad O = rad O♮ ∩ O = 𝔭O♮ .
(b) Suppose O is not maximal. Then O♮ is the unique minimal superorder of O.
Moreover, O♮ is residually inert and we have [O♮ : O] 𝑅 = 𝔭2 .
Proof. We begin with the first part of (a), and we show rad O♮ ∩ O = rad O. As in
21.2.1, O♮ is a radical cover so rad O ⊆ rad O♮ ∩ O. But arguing as in the proof
of Lemma 21.2.4, we know that rad O♮ is topologically nilpotent as a O0-ideal and
𝔭𝑟 O0 ⊆ 𝔭O for large 𝑟, so rad O♮ ∩ O is topologically nilpotent as a O-ideal, and
rad O♮ ∩ O ⊆ rad O.
We therefore have a map
O/rad O ↩→ O♮ /rad O♮ ; (24.4.8)
since O/rad O is a quadratic field, we must have (O♮| 𝔭) = ∗, −1, i.e., O♮
is either
maximal or residually inert—and in the latter case, (24.4.8) is an isomorphism.
Let 𝑖 ∈ O generate ℓ = O/rad O as a quadratic extension of 𝑘. Let 𝐾 = 𝐹 (𝑖), and
let 𝑆 be the integral closure of 𝑅 in 𝐾. Then 𝐾 is (separable and) unramified over 𝐹.
We claim that O𝑆 = O ⊗𝑅 𝑆 is residually split. Indeed, we have an isomorphism of
𝑘-algebras
∼ O /𝔭O
(O/𝔭O) ⊗ 𝑘 ℓ −→ 𝑆 𝑆
and since ℓ is separable over 𝑘, an identification (Exercise 7.18)

rad((O/𝔭O) ⊗ 𝑘 ℓ) = rad(O/𝔭) ⊗ ℓ
giving (rad O) 𝑆 = rad(O𝑆 ) and
O/rad O ⊗ 𝑘 ℓ ' O𝑆 /rad O𝑆 .
But O/rad O = ℓ ⊗ 𝑘 ℓ ' ℓ × ℓ. This shows (O𝑆 | 𝔭𝑆) = 1.
To conclude, the statements that rad O♮ ∩O = 𝔭O♮ in (a) and that [O♮ : O] 𝑅 = 𝔭2 in
(b) hold for O𝑆 by (23.4.18), so they hold for O. Minimality follows from Proposition
24.4.12.
24.4. CHAINS OF ORDERS 393
24.4.9. As a consequence of Proposition 24.4.7, if O is a residually inert 𝑅-order with

discrd(O) = 𝔭𝑒 , then the radical idealizer chain
O ( O1 = O♮ ( · · · ( O𝑟
has length 𝑟 = b𝑒/2c, with O𝑟 maximal, and dim 𝑘 (O𝑖 /O𝑖+1 ) = 2 for all 𝑖; accordingly,
the case 𝑒 even occurs exactly when 𝐵 ' M2 (𝐹) and 𝑒 odd occurs exactly when 𝐵 is a
division algebra.
We conclude the section showing that under certain hypotheses, the radical idealizer
is a minimal (proper) superorder. The results are due to Drozd–Kirichenko–Roiter
[DKR67, Propositions 1.3, 10.3]; we follow Curtis–Reiner [CR81, Exercises 37.5,
37.7].
Definition 24.4.10. Let 𝐼 0 ⊆ 𝐼 be left fractional O-ideals in 𝐵. We say 𝐼 0 is (left)
hypercharacteristic in 𝐼 if for every left O-module homomorphism 𝜙 : 𝐼 0 → 𝐼 we
have 𝜙(𝐼 0) ⊆ 𝐼 0.
Lemma 24.4.11. The map O0 ↦→ 𝐼 0 = (O0) ♯ gives an inclusion-reversing bĳection
from the set of 𝑅-superorders O0 ⊇ O to the set of right hypercharacteristic 𝑅-
sublattices 𝐼 0 ⊆ O♯ .
Proof. Inclusing-reversing follows from Lemma 15.6.2(a). By the proof of Lemma

17.3.3, we have a natural identification HomO (𝐼 0, 𝐼) ' (𝐼 : 𝐼 0) R given by right mutli-
plication.
We first show that if O0 ⊇ O, then (O0) ♯ ⊆ O is right hypercharacteristic; so we
verify (O♯ : (O0) ♯ ) R ⊆ (O0) ♯ . If (O0) ♯ 𝛼 ⊆ O♯ , then
((O0) ♯ 𝛼) ♯ = 𝛼−1 O0 ⊇ (O♯ ) ♯ = O
so 𝛼O ⊆ O0, thus 𝛼 = 𝛼 · 1 ∈ O0 = OR ((O0) ♯ ). Conversely, given 𝐼 0 ⊆ O♯

hypercharacteristic, we have 𝛼 ∈ (𝐼 0) ♯ if and only if trd(𝐼 0 𝛼) ⊆ 𝑅 if and only if
𝛼 ∈ (𝐼 0 : O) R = (𝐼 0 : 𝐼 0) = OR (𝐼 0), so (𝐼 0) ♯ = OR (𝐼 0) is an 𝑅-order.
Proposition 24.4.12. Let 𝑅 be local and let O be a Gorenstein 𝑅-order that is not
maximal and such that O is indecomposable as a left O-module. Then there is a unique
minimal 𝑅-superorder O0 ) O and O0 = O♮ .
Proof. Since O is projective indecomposable as a O-module, and O is Gorenstein

so O♯ is projective, we must have O♯ = O𝛼 for some 𝛼 ∈ 𝐵× . Let 𝐽 = rad O. By
Corollary 20.6.9, 𝐽O♯ is the unique O-submodule of O♯ . If 𝐽O♯ 𝛽 = O♯ for some
𝛽 ∈ 𝐵× , then 𝐽 = O𝛼𝛽𝛼−1 , so OL (𝐽) = O and O is extremal; but an extremal order
that is indecomposable is already maximal, a contradiction. Therefore, if 𝐽O♯ 𝛽 ⊆ O♯
then the inclusion is strict, and by maximality we have 𝐽O♯ 𝛽 ⊆ 𝐽O♯ ; that is to say,
𝐽O♯ is hypercharacteristic in O♯ . By Lemma 24.4.11, and inclusion-reversing, we see
that
O0 = (𝐽O♯ ) ♯ = (𝐽O𝛼) ♯ = 𝛼−1 𝐽 ♯
is the unique minimal 𝑅-superorder, and O0 = OR (𝐽 ♯ ) = OL (𝐽) by Proposition
15.6.6.
24.4.13. Let 𝑅 be local and O be a Gorenstein 𝑅-order. By Lemma 20.6.8, the

condition that O is indecomposable
is equivalent to the condition that O/rad O is
O
simple as a 𝑘-algebra, i.e., ≠ 1 or equivalently, O is not Eichler. We considered
𝔭
the Eichler case in 24.4.5, so suppose further that O is not Eichler (and in particular,
not maximal). Then O♮ ⊇ O is the minimal 𝑅-superorder over O by Proposition
24.4.12. If O♮ is not itself Gorenstein, then we may associate its Gorenstein saturation
Gor(O♮ ). In this way, we obtain a canonical chain of Gorenstein superorders of O.
24.5 Bass and basic orders

Given the importance of the Gorenstein condition, we will want to give a name to the
condition that every superorder is Gorenstein.
In this section, we restore our hypothesis that 𝑅 is a Dedekind domain.
Definition 24.5.1. An order O is Bass if every order O0 ⊇ O is Gorenstein.
24.5.2. Since the Gorenstein condition is local by 24.2.2, the Bass condition is also
local. Moreover, an Eichler (i.e., residually split) order is Bass, because Eichler
orders are Gorenstein by 24.2.14, and every superorder of an Eichler order is Eichler
(Corollary 23.4.10).
For the rest of this section, we investigate the local structure of Bass orders, and
we suppose that 𝑅 is local. We do not use the following proposition, but we state it for
context.
Proposition 24.5.3. Suppose 𝑅 is local. Then the following are equivalent:
(i) Every O-ideal is generated by two elements;

(ii) O is Bass; and
(iii) Every O-lattice is isomorphic to a direct sum of O-ideals.
Proof. See Drozd–Kirichenko–Roiter [DKR67, Propositions 12.1, 12.5] or Curtis–

Reiner [CR81, §37]; the implications (i) ⇒ (ii) ⇒ (iii) always hold, and the implication
(iii) ⇒ (i) holds because 𝐵 is a quaternion algebra.
The residually inert orders, those with Eichler symbol (O | 𝔭) = −1, give a source
of Bass orders, following Brzezinski [Brz83a, §3].
Proposition 24.5.4. Let O be a residually inert 𝑅-order. Then O is Bass.
Proof. We begin by arguing as in the proof of Proposition 24.4.7: letting 𝑖 ∈ O

generate the field extension O/𝑃 ⊇ 𝑘, we then make a base extension to 𝐾 = 𝐹 (𝑖)
where 𝑖 ∈ O generates O/rad O: then O𝑆 is residually split, i.e., O𝑆 is Eichler.
We then appeal to Proposition 24.2.23: the Gorenstein condition is stable under
base change. So for every superorder O0 ⊇ O, we have a superorder O𝑆0 ⊇ O𝑆 and so
O𝑆0 is Gorenstein by 24.5.2, so O0 is Gorenstein. Therefore O is Bass.
24.5. BASS AND BASIC ORDERS 395
Combining
24.5.5. 24.5.2 with Proposition 24.5.4, we see that if O is not a Bass order,
O
then = 0.
𝔭
Another rich source of Bass orders are the basic orders.
Definition 24.5.6. We say O is basic if O contains a maximal 𝑅-order in a maximal

commutative 𝐹-subalgebra 𝐾 ⊆ 𝐵.
Remark 24.5.7. Local basic orders in a division quaternion algebra were studied by
Hĳikata–Pizer–Shemanske [HPS89b]; they gave the global orders O such that O (𝔭) is
basic if 𝔭 ∈ Ram(𝐵) and Eichler if 𝔭 ∉ Ram(𝐵) the name special. The remaining types
of local basic orders (residually ramified and residually inert for the matrix ring) were
studied by Brzezinski [Brz90], and further worked on by Jun [Jun97]. The role of the
embedded maximal quadratic 𝑅-algebra 𝑆 is that one can compute embedding numbers
for them (see for example the epic work of Hĳikata–Pizer–Shemanske [HPS89a]), and
therefore compute explicitly with the trace formula.
Other authors use the term primitive instead of basic, but this quickly gets confusing
as the word primitive is used for the ternary quadratic forms and the two notions do not
coincide. The following propositions show that the sound of the word basic conveys
the right meaning.
Proposition 24.5.8. Suppose 𝑅 is local. Then O is basic if and only if O is Bass.
Proof. See Chari–Smertnig–Voight [CSV2019, Theorem 1.1]. To illustrate the proof,

we explain here only the case where 2 ∈ 𝑅 × , following Brzezinski [Brz90, Proposition
1.11].
Let 𝑄 : 𝑀 → 𝑅 be the ternary quadratic form associated to O, a representative up
to twisted similarity chosen so that 𝑄 is integral. We argue explicitly with a quadratic
form (22.3.5) the multiplication table of the order as a Clifford algebra in a good basis
(22.3.7).
First, suppose O is Bass. Then O is Gorenstein, so after rescaling we may suppose
𝑄 ∼ h−1, 𝑏, 𝑐i, and the multiplication table reads:
𝑖 2 = −𝑏𝑐 𝑗 𝑘 = −𝑖
2
𝑗 =𝑐 𝑘𝑖 = 𝑏 𝑗 (24.5.9)
2
𝑘 =𝑏 𝑖 𝑗 = 𝑐𝑘
If ord𝔭 (𝑏) ≤ 1 or ord𝔭 (𝑐) ≤ 1, then correspondingly 𝑅[ 𝑗] or 𝑅[𝑘] are maximal

(again by valuation of discriminant), so suppose ord𝔭 (𝑏), ord𝔭 (𝑐) ≥ 2. Consider the
𝑅-submodule O0 generated by 1, 𝑖 0, 𝑗, 𝑘 with 𝑖 0 = 𝑖/𝜋. Then the integrality of the
multiplication table remains intact so O0 is an order, with new coefficients 𝑎 0 = 𝜋,
𝑏 0 = 𝑏/𝜋, 𝑐 0 = 𝑐/𝜋. The corresponding ternary quadratic form 𝑄 0, by our hypotheses
on valuations, is now imprimitive and O0 is not Gorenstein, and this contradicts the
fact that O is Bass.
Conversely, suppose O is basic. If the maximal 𝑅-order 𝑆 ⊆ O is in an unramified
subalgebra 𝐾 = 𝐹𝑆 ⊆ 𝐵, then it generates an unramified extension over the residue
field and (O | 𝔭) ≠ 0, and then by 24.5.5 we know O is Bass. Now at least one of the
products 𝑏𝑐, 𝑎𝑐, 𝑎𝑏 has valuation 1 because O contains a maximal 𝑅-order; without
loss of generality we may take, after rescaling, 𝑎 = −1 and ord𝔭 (𝑏) = 1 ≤ ord𝔭 (𝑐)
with 𝑆 = 𝑅[𝑘]. Therefore 𝑄 is primitive and O is Gorenstein.
We now compute that rad O = h𝜋, 𝑖, 𝑗, 𝑘i and so
O♮ = OL (rad O) = 𝑅 + 𝔭−1𝑖 + 𝑅 𝑗 + 𝑅𝑘
is the minimal superorder, and it is still basic. By minimality, to show that O is

Bass it suffices to show that O0 = O♮ is Bass. In our new parameters we have
𝑎 0, 𝑏 0, 𝑐 0 = −𝜋, 𝑏/𝜋, 𝑐/𝜋, so ord𝔭 (𝑏 0) = 0 and ord𝔭 (𝑐 0) = ord𝔭 (𝑐) − 1. Swapping 𝑖 0, 𝑗 0
interchanges 𝑎 0, 𝑏 0, and we are back in the original situation but with ord𝔭 (𝑐) reduced.
By induction, we can continue in this way until ord𝔭 (𝑐) = 0, when then 𝑅[ 𝑗] is a
maximal order in an unramified extension, and we are done.
Proposition 24.5.10. Suppose 𝐹 = Frac(𝑅) is a number field. Then for the 𝑅-order
O, the following are equivalent:
(i) O is basic;
(ii) O (𝔭) is basic for all primes 𝔭 of 𝑅; and
(iii) O is Bass.
In other words, for orders in a quaternion algebra over a number field, being basic
is a local property and it is equivalent to being Bass.
Proof. See Chari–Smertnig–Voight [CSV2019, Theorem 1.2], building on work of

Eichler [Eic36, Satz 8] for the case 𝐹 = Q.
Remark 24.5.11. One can similarly define basic orders for a general Dedekind domain
𝑅, but the preceding results are not known in this level of generality.
24.5.12. We established several other important features along the way in Proposition
24.5.8 that we now record. Suppose 𝑅 is local with 2 ∈ 𝑅 × and suppose that O is a
residually ramified Bass (i.e., basic) order. Then the quadratic form associated to O
is similar to h−1, 𝑏, 𝑐i with ord𝔭 (𝑏) = 1 ≤ ord𝔭 (𝑐), and so the multiplication table
(24.5.9) holds. The unique superorder O♮ has [O♮ : O] 𝑅 = 𝔭 and associated ternary
quadratic form h−1, 𝑏, −𝑐/𝑏i (see Exercise 24.13).
♮
O O
Corollary 24.5.13. If O is not hereditary, then
♮ = .
𝔭 𝔭
Since O is not hereditary, we cannot have O maximal, so (O | 𝔭) = 1, 0, −1.

Proof. ♮
O O
If = 1, then O is Eichler, and so too are its superorders. If = −1 and O♮ is
𝔭 𝔭
♮
O O
not maximal, then = −1 by Proposition 24.4.7(b). For the case = 0, we
𝔭 𝔭
appeal to 24.5.12.
24.6. TREE OF ODD BASS ORDERS 397
We can repackage what we have done for basic orders to give another description
in terms of its hereditary closure.
Proposition 24.5.14. Let O be a basic, nonhereditary 𝑅-order with discrd(O) = 𝔭𝑛 .

Suppose 2 ∈ 𝑅 × , and let 𝑆 ⊆ O be a maximal 𝑅-order in the 𝐹-algebra 𝐾. Let 𝐽 be the
Jacobson radical of the hereditary closure of O. Then the following statements hold.
(a) Suppose O is residually inert. Then O = 𝑆 + 𝐽 𝑚 where 𝑚 = 𝑛/2, 𝑛 − 1 according

as 𝐵 is ramified or split.
(b) Suppose O is residually ramified. Then O = 𝑆 + 𝐽 𝑚 where 𝑚 = 𝑛 − 1.
Proof. The statement follows by induction using the explicit descriptions of these
orders in 24.3.7, 24.5.12, and 24.5.12. See Brzezinski [Brz90, Proposition 1.12].
24.6 Tree of odd Bass orders

To conclude this chapter, we draw a picture of the containments of Bass orders.
Suppose 𝑅 is local with finite residue field 𝑘 and 2 ∈ 𝑅 × . We put together
the radical idealizer chains in the residually split 24.4.5, residually inert 24.4.9, and
residually ramified 24.5.12 cases. The resulting tree of Bass orders is shown in Figure
24.6.1.
discrd (O)
B ' M2 (F ) split B ramified
=
R
M2 (R)
det p
p2
p3
p4

O
= −1 0 1 ∗ −1 0
p
res. inert res. split res. inert
hπ 2e i ⊥ NmKun |F (Eichler) hπ 2e+1 i ⊥ NmKun |F
hπ e i ⊥ H
Figure 24.6.1: Tree of local Bass orders, odd characteristic residue field
Each vertex of this graph represents an isomorphism class of Bass order; there is an
edge between two vertices if and only if there is a minimal containment between them.
In each case, such a containment is given by the radical idealizer except when the order
is residually split (in which case it hops by two, skipping the minimal superorder).
For the trees when 2 is a uniformizer in 𝑅, and many other explicit calculations,
see Lemurell [Lem2011, §5], as well as Pacetti–Sirolli [PS2014, §5].
Exercises
Unless otherwise specified, let 𝑅 be a Dedekind domain with 𝐹 = Frac 𝑅 and let
O ⊆ 𝐵 be an 𝑅-order in a quaternion algebra 𝐵.
1. Show that every Eichler order O is Gorenstein by showing O♯ is locally principal

by direct computation.
2. Show that codiff (O) is right invertible if and only if
codiff(O) (codiff(O) 2 ) ♯ = O
and codiff (O) is left invertible if and only if
(codiff(O) 2 ) ♯ codiff (O) = O.
[Hint: show that (codiff (O) −1 ) ♯ = codiff (O) 2 .]
𝑅 is local. Show that if O is an 𝑅-order and 𝐵 is a division algebra,

I 3. Suppose
O
then = 1, 0.
𝔭
I 4. Suppose 𝑅 is local with finite residue field. Recall the discriminant quadratic
form and the generalized Kronecker symbol 24.3.3.

O Δ(𝛼)
(a) Show that = ∗ if and only if takes on all of the values −1, 0, 1
𝔭 𝔭
for 𝛼 ∈ O.
O Δ(𝛼)
(b) For 𝜖 = −1, 0, 1, show that = 𝜖 if and only if takes the values
𝔭 𝔭
{0, 𝜖 } for 𝛼 ∈ O.
I 5. Let 𝑅 be local and suppose O = 𝑅 + 𝔭O0 for an order O0. Show that rad O = 𝔭O0
and O♮ = O0. [Hint: Argue as in 23.4.15.]
6. Let 𝑄 : 𝑀 → 𝐿 be a ternary quadratic module, and let 𝔞 be a fractional ideal.
Write 𝔞𝑄 = 𝑄 ⊗ 𝔞 : 𝑀 ⊗ 𝔞 → 𝐿 ⊗ 𝔞2 for the twist. Show that there is a bĳection
Cl 𝑄 ↔ Cl 𝔞𝑄, and conclude that there is a bĳection Typ O ↔ Typ Gor(O),
where Gor(O) is the Gorenstein saturation.
7. (a) Show that if discrd(O) is cubefree (i.e., there is no prime 𝔭 of 𝑅 such that
𝔭3 | discrd(O)) then O is a Bass order.
O
(b) Show that there is a (local) Gorenstein order O with = 0 that is not
𝔭
Bass with discrd(O) = 𝔭4 .
24.6. TREE OF ODD BASS ORDERS 399

O
8. Suppose that O is Gorenstein with ≠ 1. Let O0 be the hereditary closure
𝔭
of O. Show that 𝑁 𝐵× (O) ≤ 𝑁 𝐵× (O0), and further that equality holds when O is
residually inert.
9. Let 𝑅 be local and let O be a residually ramified quaternion 𝑅-order that is
Gorenstein but not a Bass order. Let O♮ be the unique minimal order containing
O. Show that 𝔣(O♮ ) = 𝔭. [See Brzezinski [Brz83a, Lemma 4.4].]
10. Let 𝑅 be local with 2 ∈ 𝑅 × , and let 𝑄 = h−1, 𝑏, 𝑐i : 𝑅 3 → 𝑅 with 0 = ord𝔭 (𝑏) <
ord𝔭 (𝑐). Let O = Clf 0 (𝑄) be its even Clifford algebra. Show that

O 𝑏
= .
𝔭 𝔭
11. Let 𝑅 be local, let O = M2 (𝑅), let O1 be the standard Eichler order of level
𝔭𝑒 and let O2 = 𝑅 + 𝔭𝑒 O. Show that O2 ( O1 ( O, that Gor(O2 ) = O, but
Gor(O1 ) = O1 ; conclude that the Gorenstein saturation is not (necessarily) the
smallest Gorenstein superorder.
12. Let 𝑅 be local and let O be the standard Eichler order of level 𝔭𝑒 with 𝑒 ≥ 2.
Show that O contains no integrally closed quadratic 𝑅-order that is a domain
(even though O contains 𝑅 × 𝑅).
I 13. Let 𝑅 be local with 2 ∈ 𝑅 × .
the proof of Proposition 24.5.8) that O is a local Bass order
(a) Show(using
O
with = 0 if and only if its corresponding ternary quadratic form is
𝔭
similar to h−1, 𝑏, 𝑐i with ord𝔭 (𝑏) = 1 ≤ ord𝔭 (𝑐).
(b) In case (a), show that the minimal overorder O0 corresponds to the (simi-
larity class of) ternary quadratic form h−1, 𝑏, −𝑐/𝑏i.
Part III
Analysis
401
Chapter 25
The Eichler mass formula
In Part II of this text, we investigated arithmetic and algebraic properties of quaternion

orders and ideals. Now in Part III, continuing this investigation we turn to the use of
analytic methods. In this first introductory chapter, we restrict to a special case and
consider zeta functions of quaternion orders over the rationals; as an application, we
obtain a formula for the mass of the class set of a definite quaternion order.
25.1 ⊲ Weighted class number formula

Gauss conjectured (in the language of binary quadratic forms) that there are finitely
many imaginary quadratic orders of class number 1 [Gau86, Article 303]. Approaches
to this problem involve beautiful and deep mathematics. Given that we want to prove
some kind of lower bound for the class number (in terms of the discriminant of the
order), it is natural to seek an analytic expression for it. The analytic class number
formula of Dirichlet provides such an expression, turning the class number problem
of Gauss into a (still very hard, but tractable) problem of estimation.
We recall section 14.1, which gives a classification of quaternion algebras over Q,
and sections 16.1 and 17.1, providing background on ideal classes in quaternion orders.
With this motivation in hand, we are led to ask: what are the definite quaternion orders
of class number 1? The method to prove Dirichlet’s formula generalizes to definite
quaternion orders as well, as pursued by Eichler in his mass formula. This chapter
gives an overview of the Eichler mass formula in the simplest case for a maximal order
in a definite quaternion algebra over Q. (The reader who is already motivated and
ready for action may consider skipping to the next chapter.)
Theorem 25.1.1 (Eichler mass formula over Q, maximal orders). Let 𝐵 be a definite
quaternion algebra over Q of discriminant 𝐷 and let O ⊂ 𝐵 be a maximal order. Then
∑︁ 1 𝜑(𝐷)
=
𝑤𝐽 12
[𝐽 ] ∈Cls O
where 𝑤 𝐽 := #OL (𝐽) × /{±1} and 𝜑(𝐷) := #(Z/𝐷Z) × =

Î
𝑝 |𝐷 ( 𝑝 − 1) is the Euler
totient function.
403
404 CHAPTER 25. THE EICHLER MASS FORMULA
The Eichler mass formula does not quite give us a formula for the class number—
rather, it gives us a formula for a “weighted” class number. That being said, we remark
that 𝑤 𝐽 ≤ 24 (see Theorem 11.5.14), and very often 𝑤 𝐽 = 1 (i.e., OL (𝐽) = {±1}). In
order to convert the Eichler mass formula into a formula for the class number itself,
one needs to understand the unit groups of left orders: this can be understood either
as a problem in representation numbers of ternary quadratic forms or of embedding
numbers of quadratic orders into quaternion orders, and we will take this subject up in
earnest in Chapter 30.
Over Q, the Eichler mass formula was first proven by Hey [Hey29, II, (80)], a
Ph.D. student of Artin, along the same lines as the proof sketched below. This formula
was also stated by Brandt [Bra28, §67]. We gradually warm up to this theorem
by considering a broader analytic context. We see the analytic class number for an
imaginary quadratic field as coming from the residue of its zeta function, and we then
pursue a quaternionic generalization.
25.2 ⊲ Imaginary quadratic class number formula

√
To introduce the circle of ideas, let 𝐾 := Q( 𝑑) be a quadratic field of discriminant
𝑑 ∈ Z and let Z𝐾 be its ring of integers. We encode information about the field 𝐾 by
its zeta function.
25.2.1. Over Q, we define the Riemann zeta function

∞
∑︁ 1
𝜁 (𝑠) := 𝑠
(25.2.2)
𝑛=1
𝑛
as the prototypical such function; this series converges for Re 𝑠 > 1, by the comparison
test. By unique factorization, there is an Euler product
Ö −1
1
𝜁 (𝑠) = 1− (25.2.3)
𝑝
𝑝𝑠
where the product is over all primes 𝑝. The function 𝜁 (𝑠) can be meromorphically
continued to the right half-plane Re 𝑠 > 0 using the fact that the sum
∞
∑︁ (−1) 𝑛
𝜁2 (𝑠) =
𝑛=1
𝑛𝑠
converges for Re 𝑠 > 0 and
𝜁 (𝑠) + 𝜁2 (𝑠) = 21−𝑠 𝜁 (𝑠)
so that
1
𝜁 (𝑠) = 𝜁2 (𝑠)
21−𝑠 −1
25.2. ⊲ IMAGINARY QUADRATIC CLASS NUMBER FORMULA 405
and the right-hand side makes sense for Re 𝑠 > 0 except for possible poles where
21−𝑠 = 1. For real values of 𝑠 > 1, we have
∫ ∞ ∫ ∞
1 d𝑥 d𝑥 𝑠
= ≤ 𝜁 (𝑠) ≤ 1 + =
𝑠−1 1 𝑥𝑠 1 𝑥𝑠 𝑠−1
so
1 ≤ (𝑠 − 1)𝜁 (𝑠) ≤ 𝑠;
therefore, as 𝑠 approaches 1 from above, we have lim𝑠&1 (𝑠 − 1)𝜁 (𝑠) = 1, so 𝜁 (𝑠) has
a simple pole at 𝑠 = 1 with residue
res𝑠=1 𝜁 (𝑠) = 1. (25.2.4)
25.2.5. For the quadratic field 𝐾, modeled after (25.2.2) we define the Dedekind zeta
function by
∑︁ 1
𝜁 𝐾 (𝑠) := (25.2.6)
𝔞⊆Z
N(𝔞) 𝑠
𝐾
where N(𝔞) := #(Z𝐾 /𝔞) is the absolute norm, the sum is over all nonzero ideals of
Z𝐾 , and the series is defined for 𝑠 ∈ C with Re 𝑠 > 1. (We recall that N(𝔞) is the
positive generator of Nm𝐾 |Q (𝔞), so we could equivalently work with the algebra norm,
if desired.)
We can also write the Dedekind zeta function as a Dirichlet series
∞
∑︁ 𝑎𝑛
𝜁 𝐾 (𝑠) = (25.2.7)
𝑛=1
𝑛𝑠
where 𝑎 𝑛 := #{𝔞 ⊆ Z𝐾 : N(𝔞) = 𝑛} is the number of ideals in Z𝐾 of norm 𝑛 ≥ 1. By

unique factorization of ideals, we again have an Euler product expansion
Ö −1
1
𝜁 𝐾 (𝑠) = 1− , (25.2.8)
𝔭
N(𝔭) 𝑠
the product over all nonzero prime ideals 𝔭 ⊂ Z𝐾 .

In order to introduce a formula that involves the class number, we group the ideals
in (25.2.6) by their ideal class: for [𝔟] ∈ Cl(𝐾), we define the partial zeta function
∑︁ 1
𝜁 𝐾 , [𝔟] (𝑠) :=
𝔞⊆Z𝐾
N(𝔞) 𝑠
[𝔞]=[𝔟]
so that ∑︁
𝜁 𝐾 (𝑠) = 𝜁 𝐾 , [𝔟] (𝑠). (25.2.9)
[𝔟] ∈Cl(𝐾 )
In general, for [𝔟] ∈ Cl(𝐾), we have [𝔞] = [𝔟] if and only if there exists 𝑎 ∈ 𝐾 × such
that 𝔞 = 𝑎𝔟, but since 𝔞 ⊆ Z𝐾 , in fact
𝑎 ∈ 𝔟−1 = {𝑎 ∈ Z𝐾 : 𝑎𝔟 ⊆ Z𝐾 };
this gives a bĳection

{𝔞 ⊆ Z𝐾 : [𝔞] = [𝔟]} ↔ 𝔟−1 /Z×𝐾 ,
(since the generator of an ideal is unique up to units). Thus
1 ∑︁ 1
𝜁 𝐾 , [𝔟] (𝑠) = . (25.2.10)
N(𝔟) 𝑠
−1 ×
N(𝑎) 𝑠
0≠𝑎 ∈𝔟 /Z𝐾
for each class [𝔟] ∈ Cl(𝐾).

Everything we have done so far works equally as well for real as for imaginary
quadratic fields. But to make sense of 𝔟−1 /Z×𝐾 in the simplest case, we want Z×𝐾 to
be a finite group, which by Dirichlet’s unit theorem means exactly that 𝐾 is Q or√an
imaginary quadratic field. So from now on in this section, we suppose 𝐾 = Q( 𝑑)
with 𝑑 < 0. Then 𝑤 := #Z×𝐾 = 2, except when 𝑑 = −3, −4 where 𝑤 = 6, 4, respectively.
Under this hypothesis, the sum (25.2.10) can be transformed into a sum over
lattice points with the fixed factor 𝑤 of overcounting. Before estimating the sum over
reciprocal norms, we first estimate the count. Let Λ ⊂ C be a lattice. We can estimate
the number of lattice points 𝜆 ∈ Λ with |𝜆| ≤ 𝑥 by the ratio 𝜋𝑥 2 /𝐴, where 𝐴 is the
area of a fundamental parallelogram 𝑃 for Λ: roughly speaking, this says that we can
tile a circle of radius 𝑥 with approximately 𝜋𝑥 2 /𝐴 parallelograms 𝑃.
More precisely, the following lemma holds.
Lemma 25.2.11. Let Λ ⊂ C be a lattice with area(C/Λ) = 𝐴. Then there is a constant
𝐶 such that for all 𝑥 > 1,
2

#{𝜆 ∈ Λ : |𝜆| ≤ 𝑥} − 𝜋𝑥 ≤ 𝐶𝑥.

𝐴
We leave this lemma as an exercise (Exercise 25.3) in tiling a circle with radius 𝑥
with fundamental parallelograms for the lattice Λ. With a bit of manipulation (Exercise
25.4), this lemma can be used to prove the analytic class number formula.
√ 25.2.12 (Analytic class number formula, imaginary quadratic field). Let

Theorem
𝐾 = Q( 𝑑) be an imaginary quadratic field with discriminant 𝑑 < 0. Then
2𝜋ℎ
res𝑠=1 𝜁 𝐾 (𝑠) = √︁
𝑤 | 𝑑|
where ℎ is the class number of 𝐾 and 𝑤 is the number of roots of unity in 𝐾.
This formula simplifies slightly if we cancel the pole at 𝑠 = 1 with 𝜁 (𝑠), as follows.
Like in the Dirichlet series, we can combine terms in (25.2.8) to get
ÖÖ −1 Ö
1
𝜁 𝐾 (𝑠) = 1− = 𝐿 𝑝 ( 𝑝 −𝑠 ) −1
N(𝔭) 𝑠
𝑝 𝔭| 𝑝 𝑝
where


 (1 − 𝑇) 2 , if 𝑝 splits in 𝐾;


𝐿 𝑝 (𝑇) := 1 − 𝑇, if 𝑝 ramifies in 𝐾; and (25.2.13)

1 − 𝑇 2,

if 𝑝 is inert in 𝐾.

25.2. ⊲ IMAGINARY QUADRATIC CLASS NUMBER FORMULA 407
The condition of being split, ramified, or inert in 𝐾 is recorded in a function:



 1, if 𝑝 splits in 𝐾;


𝜒( 𝑝) := 𝜒𝑑 ( 𝑝) = 0, if 𝑝 ramifies in 𝐾; and (25.2.14)

 −1,

if 𝑝 is inert in 𝐾

for prime 𝑝 and extended to all positive integers by multiplicativity. If 𝑝 - 𝑑 is an odd
prime, then
𝑑
𝜒( 𝑝) =
𝑝
is the usual Legendre symbol, equal to 1 or −1 according as if 𝑑 is a quadratic residue
or not modulo 𝑝. Then in all cases, we have
𝐿 𝑝 (𝑇) = (1 − 𝑇) (1 − 𝜒( 𝑝)𝑇).
Expanding the Euler product term-by-term and taking a limit, we conclude
𝜁 𝐾 (𝑠) = 𝜁 (𝑠)𝐿 (𝑠, 𝜒) (25.2.15)
where −1
Ö 𝜒( 𝑝) ∑︁ 𝜒(𝑛)
𝐿 (𝑠, 𝜒) := 1− 𝑠 = . (25.2.16)
𝑝
𝑝 𝑛
𝑛𝑠
The function 𝐿(𝑠, 𝜒) is in fact holomorphic for all Re 𝑠 > 0; this follows from the
Í
fact that the partial sums 𝑛 ≤𝑥 𝜒(𝑛) are bounded and the mean value theorem. So in
particular the series
𝜒(2) 𝜒(3) 𝜒(4)

𝐿(1, 𝜒) = 1 + + + +...
2 3 4
converges (slowly). Combining (25.2.15) with the analytic class number formula
yields:
2𝜋ℎ
𝐿 (1, 𝜒) = √︁ ≠ 0. (25.2.17)
𝑤 | 𝑑|
For example, taking 𝑑 = −4, so 𝜒(2) = 0 and 𝜒( 𝑝) = (−1/𝑝) = (−1) ( 𝑝−1)/2 ,
1 1 1 1 1
𝐿(1, 𝜒) = 1 − + − + − +...
3 5 7 9 11
−1
Ö (−1) ( 𝑝−1)/2 𝜋
= 1− = = 0.7853 . . . .
𝑝 ≥3
𝑝 4
Remark 25.2.18. The fact that 𝐿(1, 𝜒) ≠ 0, and its generalization to complex characters
𝜒, is the key ingredient to prove Dirichlet’s theorem on primes in arithmetic progression
(Theorem 14.2.9), used in the classification of quaternion algebras over Q. The
arguments to complete the proof are requested in Exercise 26.11.
Remark 25.2.19. To approach the class number problem of Gauss, we would then seek
lower bounds on 𝐿 (1, 𝜒) in terms of the absolute discriminant | 𝑑|. Indeed, the history
of class number problems is both long and beautiful. The problem of determining all
positive definite binary quadratic forms with small class number was first posed by
Gauss [Gau86, Article 303]. This problem was later seen to be equivalent to finding all
imaginary quadratic fields of small class number (as in section 19.1). It would take al-
most 150 years of work, with important work of Heegner [Heeg52] and culminating in
the results of Stark [Sta67] and Baker [Bak71], to determine those fields with class num-
ber 1: there are exactly nine, namely 𝑑 = −3, −4, −7, −8, −11, −19, −43, −67, −163.
See Goldfeld [Gol85] or Stark [Sta2007] for a history of this problem. For more
specifically on the analytic class number formula for imaginary quadratic fields, see
the survey by Weston [Wes] as well as the book by Serre [Ser73, Chapter VI].
25.3 ⊲ Eichler mass formula: over the rationals

We are now prepared to consider the analogue of the analytic class number formula
(Theorem 25.2.12) for quaternion orders: the Eichler mass formula, which is a weighted
class number formula. We follow Eichler [Eic55-56, Eic56a], and in this section we
give an overview with proofs omitted—a full development will be given in the next
chapter, in more generality.
Let 𝐵 be a quaternion algebra over Q of discriminant 𝐷 and let O ⊂ 𝐵 be an order.
We define the zeta function of O to be
∑︁ 1
𝜁O (𝑠) := , (25.3.1)
N(𝐼) 𝑠
𝐼 ⊆O
where the sum over all invertible (nonzero, integral) right O-ideals and
N(𝐼) := [O : 𝐼] = #(O/𝐼) ∈ Z>0 .
(By Main Theorem 16.1.3, we have N(𝐼) the totally positive generator of nrd(𝐼) 2 , so
we could equivalently work with the reduced norm.)
Let 𝑎 𝑛 be the number of invertible right O-ideals of reduced norm 𝑛 > 0 (with
positive generator chosen, as usual). Then N(𝐼) = Nm(𝐼) = nrd(𝐼) 2 by 16.4.10, so
∞
∑︁ 𝑎𝑛
𝜁O (𝑠) = 2𝑠
. (25.3.2)
𝑛=1
𝑛
To establish an Euler product for 𝜁O (𝑠), in due course we will give a kind of factoriza-
tion formula for right ideals of O—but by necessity, writing an ideal as a compatible
product will involve the entire set of orders connected to O! A direct consequence of
the local-global dictionary for lattices (Theorem 9.4.9) is that
𝑎 𝑚𝑛 = 𝑎 𝑚 𝑎 𝑛 (25.3.3)
whenever 𝑚, 𝑛 are coprime. Next, we will count the ideals of a given reduced norm
𝑞 = 𝑝 𝑒 a power of a prime: the answer will depend on the local structure of the order
25.3. ⊲ EICHLER MASS FORMULA: OVER THE RATIONALS 409
O 𝑝 . Indeed, 𝜁O (𝑠) has an Euler product

Ö
𝜁O (𝑠) = 𝜁O, 𝑝 ( 𝑝 −𝑠 ) −1 (25.3.4)
𝑝
with 𝜁O, 𝑝 (𝑇) ∈ 1 + 𝑇Z[𝑇]. In particular, 𝜁O (𝑠) only depends on the genus (local
isomorphism classes) of O.
For simplicity, we first consider the case where O is a maximal order. Since there
is a unique genus of maximal orders, the zeta function is independent of the choice of
O and so we will write 𝜁 𝐵 (𝑠) := 𝜁O (𝑠) for O maximal. Then by a local count, we will
show that (
2 1, if 𝑝 | 𝐷;
𝜁 𝐵, 𝑝 (𝑇) = (1 − 𝑇 ) · (25.3.5)
1 − 𝑝𝑇 2 , if 𝑝 - 𝐷.
From (25.3.5),
Ö 1

𝜁 𝐵 (𝑠) = 𝜁 (2𝑠)𝜁 (2𝑠 − 1) 1− . (25.3.6)
𝑝 |𝐷
𝑝 2𝑠−1
In particular, since 𝜁 (𝑠) has a simple pole at 𝑠 = 1 with residue 1 and 𝜁 (2) = 𝜋 2 /6
(Exercise 25.1),
𝜋2 Ö

1
res𝑠=1 𝜁 𝐵 (𝑠) = lim (𝑠 − 1)𝜁 𝐵 (𝑠) = 1− (25.3.7)
𝑠&1 12 𝑝
𝑝 |𝐷
(We could also look to cancel the poles of 𝜁 𝐵 (𝑠) in a similar way to define an 𝐿-function
for 𝐵, holomorphic for Re 𝑠 > 0.)
Now we break up the sum (25.3.1) according to right ideal class:
∑︁
𝜁 𝐵 (𝑠) = 𝜁 𝐵, [𝐽 ] (𝑠)
[𝐽 ] ∈Cls O
where ∑︁ 1
𝜁 𝐵, [𝐽 ] (𝑠) := . (25.3.8)
N(𝐼) 𝑠
𝐼 ⊆O
[𝐼 ]=[𝐽 ]
Since [𝐼] = [𝐽] if and only if 𝐼 = 𝛼𝐽 for some invertible 𝛼 ∈ 𝐽 −1 , and 𝜇𝐽 = 𝐽 if and
only if 𝜇 ∈ OL (𝐽) × , we conclude that
1 ∑︁ 1
𝜁 𝐵, [𝐽 ] (𝑠) = (25.3.9)
N(𝐽) 𝑠 −1 ×
N(𝛼) 𝑠
0≠𝛼∈𝐽 /OL ( 𝐽 )
where the sum is taken over the nonzero elements 𝛼 ∈ 𝐽 −1 up to right multiplication
by units OL (𝐽) × in the left order.
In order to proceed, we now suppose that 𝐵 is definite (ramified at ∞) and hence
that #OL (𝐽) × < ∞ (see Lemma 17.7.13); this is the analogue with the case of an
imaginary quadratic field, and each 𝐽 has the structure of a lattice in the Euclidean
space R4 via the embedding
𝐽 ↩→ 𝐵 ↩→ 𝐵 ⊗Q R ' H ' R4 . (25.3.10)
Let 𝑤 𝐽 = #OL (𝐽) × /{±1}. We again argue by counting lattice points to prove the
following proposition.
Proposition 25.3.11. The function 𝜁 𝐵, [𝐽 ] (𝑠) has a simple pole at 𝑠 = 1 with residue
𝜋2
res𝑠=1 𝜁 𝐵, [𝐽 ] (𝑠) = .
𝑤𝐽 𝐷
Proof sketch. From a more general result (Theorem 26.2.12, proven in the next section
and used to prove the analytic class number formula itself), we will show that
1 vol((R4 ) ≤1 )
res𝑠=1 𝜁 𝐵, [𝐽 ] (𝑠) = (25.3.12)
2𝑤 𝐽 N(𝐽) covol(𝐽)
where under (25.3.10) we have
𝜋2
vol((R4 ) ≤1 ) = vol({𝑥 ∈ R4 : |𝑥| ≤ 1}) =
2
and
covol(O) 𝐷/4
covol(𝐽) = = .
N(𝐽) N(𝐽)
Putting all of these facts together,
𝜋2 4N(𝐽) 𝜋2
res𝑠=1 𝜁 𝐵, [𝐽 ] (𝑠) = = . (25.3.13)
4𝑤 𝐽 N(𝐽) 𝐷 𝑤𝐽 𝐷

In particular the pole of each zeta function 𝜁 𝐵, [𝐽 ] (𝑠) is almost independent of the
class [𝐽], with the only relevant term being 𝑤 𝐽 the number of units.
Combining Proposition 25.3.13 with Proposition 25.3.11,
𝜋 2 ∑︁ 𝜋2 Ö

1 1
res𝑠=1 𝜁 𝐵 (𝑠) = = 1− (25.3.14)
𝐷 𝑤𝐽 12 𝑝
[𝐽 ] ∈Cls O 𝑝 |𝐷
and we conclude the following theorem.

Theorem 25.3.15 (Eichler mass formula, maximal orders). Let 𝐵 be a definite quater-
nion algebra over Q of discriminant 𝐷 and let O ⊂ 𝐵 be a maximal order. Then
∑︁ 1 𝜑(𝐷)
= . (25.3.16)
𝑤𝐽 12
[𝐽 ] ∈Cls O
Remark 25.3.17. The Eichler mass formula is also very similar to the mass formula
for the number of isomorphism classes of supersingular elliptic curves: this is no
coincidence, and its origins will be explored in section 42.2.
To extend the Eichler mass formula to a more general class of orders, one only
needs to replace the local calculation in 25.3.5 by a count of invertible ideals in the
order. First we treat the important case of Eichler orders (see 23.1.3).
25.4. CLASS NUMBER ONE AND TYPE NUMBER ONE 411
Theorem 25.3.18 (Eichler mass formula, Eichler orders over Q). Let O ⊂ 𝐵 be an
Eichler order of level 𝑀 in a definite quaternion algebra 𝐵 of discriminant 𝐷. Then
∑︁ 1 𝜑(𝐷)𝜓(𝑀)
=
𝑤𝐽 12
[𝐽 ] ∈Cls O
where
Ö Ö 1

𝑒 𝑒−1
𝜓(𝑀) = (𝑝 + 𝑝 )=𝑀 1+ .
𝑝
𝑝𝑒 k 𝑀 𝑝 |𝑀
The most general formula is written in terms of the Eichler

symbol 24.3. Just to
O
recall two important cases: if 𝑝 | 𝑁 = discrd(O), then = −1 if O 𝑝 is the maximal
𝑝
O
order in the division algebra 𝐵 𝑝 and = 1 if O 𝑝 is an Eichler order.
𝑝
Main Theorem 25.3.19 (Eichler mass formula, general case over Q). Let 𝐵 be a
definite quaternion algebra over Q and O ⊂ 𝐵 be an order with discrd(O) = 𝑁. Then
∑︁ 1 𝑁 Ö
= 𝜆(O, 𝑝)
𝑤𝐽 12
[𝐽 ] ∈Cls O 𝑝|𝑁
where
 1 + 1/𝑝, if (O | 𝑝) = 1;
1 − 𝑝 −2




𝜆(O, 𝑝) = = 1 − 1/𝑝, if (O | 𝑝) = −1; and (25.3.20)
O −1 
1− 𝑝  1 − 1/𝑝 2 ,

if (O | 𝑝) = 0.
𝑝 
Main Theorem 25.3.19 was proven by Brzezinski [Brz90, (4.6)] and more generally
over number rings by Körner [Kör87, Theorem 1].
25.4 Class number one and type number one

The Eichler mass formula can be used to solve the class number 1 problem for definite
quaternion orders over Z, and in fact it is much easier than for imaginary quadratic
fields! We begin with the case of maximal orders.
Theorem 25.4.1. Let O be a maximal order in a definite quaternion algebra over Q

of discriminant 𝐷. Then # Cls O = 1 if and only if 𝐷 = 2, 3, 5, 7, 13.
Proof. A calculation by hand; see Exercise 25.5.
Remark 25.4.2. The primes 𝑝 = 2, 3, 5, 7, 13 in Theorem 25.4.1 are also the primes
𝑝 such that the modular curve 𝑋0 ( 𝑝) has genus 0. This is not a coincidence, and
reflects a deep correspondence between classical and quaternionic modular forms (the
Eichler–Shimizu–Jacquet–Langlands correspondence): see Remark 41.5.13.
The list of all definite quaternion orders (over Z) of class number 1 was determined
by Brzezinski [Brz95]. (Brzezinski mistakenly lists an order of class number 2, and
so he counts 25, not 24; he corrects this later in a footnote [Brz98, Footnote 1].)
Theorem 25.4.3 (Brzezinski). There are exactly 24 isomorphism classes of definite

quaternion orders over Z with # Cls O = 1.
The list of orders with class number 1 is given in Table 25.4.4. We provide
𝑁 := discrd(O), 𝐷 := disc(𝐵), we list the Eichler symbols (O | 𝑝) for the relevant
primes 𝑝 ≤ 13, we say if the order is maximal, hereditary (but not maximal), Eich-
ler/residually split or residually inert (but not hereditary), Bass (but not Eichler), or
non-Gorenstein. This list includes the three norm Euclidean maximal orders (Exercise
25.6) of discriminant 𝐷 = 2, 3, 5.
(O | 𝑝)
𝑁 𝐷 2 3 5 7 11 13 Class 𝑄
2 2 −1 ∗ ∗ ∗ ∗ ∗ maximal 𝑥 2 − 𝑥𝑦 − 𝑥𝑧 + 𝑦 2 + 𝑦𝑧 + 𝑧2
3 3 ∗ −1 ∗ ∗ ∗ ∗ maximal 𝑥 2 − 𝑥𝑦 + 𝑦 2 + 𝑧2
4 2 0 ∗ ∗ ∗ ∗ ∗ Bass 𝑥 2 + 𝑦2 + 𝑧2
5 5 ∗ ∗ −1 ∗ ∗ ∗ maximal 𝑥 − 𝑥𝑦 − 𝑥𝑧 + 𝑦 2 + 𝑦𝑧 + 2𝑧2
2
6 2 −1 1 ∗ ∗ ∗ ∗ hereditary 𝑥 2 − 𝑥𝑦 + 𝑦 2 + 2𝑧 2
6 3 1 −1 ∗ ∗ ∗ ∗ hereditary 𝑥 + 𝑥𝑧 + 𝑦 2 − 𝑦𝑧 + 2𝑧2
2
7 7 ∗ ∗ ∗ −1 ∗ ∗ maximal 𝑥 2 − 𝑥𝑧 + 𝑦 2 + 2𝑧2
8 2 −1 ∗ ∗ ∗ ∗ ∗ Bass 𝑥 + 𝑥𝑦 − 𝑥𝑧 + 𝑦 2 − 𝑦𝑧 + 3𝑧2
2
8 2 0 ∗ ∗ ∗ ∗ ∗ Bass 𝑥 2 + 𝑦 2 + 2𝑧2
10 2 −1 ∗ 1 ∗ ∗ ∗ hereditary 𝑥 − 𝑥𝑧 + 𝑦 2 − 𝑦𝑧 + 3𝑧2
2
10 5 1 ∗ −1 ∗ ∗ ∗ hereditary 𝑥 + 𝑥𝑦 + 𝑥𝑧 + 2𝑦 2 + 2𝑦𝑧 + 2𝑧 2
2
12 2 0 1 ∗ ∗ ∗ ∗ Bass 𝑥 2 + 2𝑦 2 − 2𝑦𝑧 + 2𝑧2

12 3 −1 −1 ∗ ∗ ∗ ∗ residually inert 𝑥 2 + 𝑥𝑦 + 𝑦 2 + 4𝑧2
12 3 1 −1 ∗ ∗ ∗ ∗ Eichler 𝑥 − 𝑥𝑦 + 𝑥𝑧 + 2𝑦 2 − 𝑦𝑧 + 2𝑧 2
2
12 3 0 −1 ∗ ∗ ∗ ∗ Bass 𝑥 2 + 𝑦 2 + 3𝑧2
13 13 ∗ ∗ ∗ ∗ ∗ −1 maximal 𝑥 − 𝑥𝑦 + 2𝑦 2 + 𝑦𝑧 + 2𝑧 2
2
16 2 0 ∗ ∗ ∗ ∗ ∗ non-Gorenstein 2(𝑥 2 − 𝑥𝑦 − 𝑥𝑧 + 𝑦 2 + 𝑦𝑧 + 𝑧2 )
16 2 0 ∗ ∗ ∗ ∗ ∗ Bass 𝑥 2 + 2𝑦 2 + 2𝑧 2
18 2 −1 −1 ∗ ∗ ∗ ∗ Bass 𝑥 + 𝑥𝑧 + 𝑦 2 − 𝑦𝑧 + 5𝑧2
2
18 2 −1 1 ∗ ∗ ∗ ∗ Eichler 𝑥 2 + 𝑥𝑧 + 2𝑦 2 + 2𝑦𝑧 + 3𝑧2

20 5 0 ∗ −1 ∗ ∗ ∗ Bass 𝑥 2 + 2𝑦 2 − 2𝑦𝑧 + 3𝑧2
22 2 −1 ∗ ∗ ∗ 1 ∗ hereditary 𝑥 2 + 𝑥𝑧 + 2𝑦 2 + 3𝑧2
24 3 0 −1 ∗ ∗ ∗ ∗ non-Gorenstein 2(𝑥 2 − 𝑥𝑦 + 𝑦 2 + 𝑧2 )
28 7 −1 ∗ ∗ −1 ∗ ∗ Bass 𝑥 + 𝑥𝑦 − 𝑥𝑧 + 3𝑦 2 − 2𝑦𝑧 + 3𝑧2
2
Table 25.4.4: Definite quaternion orders over Z with class number 1
Proof of Theorem 25.4.3. Suppose # Cls O = 1. We apply the mass formula (Main
Theorem 25.3.19). We note that 𝜆(O, 𝑝) ≥ 1 − 1/𝑝, in all three cases, so

1 𝑁 Ö 𝑁 Ö 1 𝜑(𝑁)
1≥ = 𝜆(O, 𝑝) ≥ 1− = (25.4.5)
𝑤 12 12 𝑝 12
𝑝|𝑁 𝑝|𝑁
and therefore 𝜑(𝑁) ≤ 12. By elementary number theory, this implies that
2 ≤ 𝑁 ≤ 16 or 𝑁 = 18, 20, 21, 22, 24, 26, 28, 30, 36, 42.
This immediately gives a finite list of possibilities for the discriminant 𝐷 = disc 𝐵 ∈
{2, 3, 5, 7, 11, 13, 30, 42}, as 𝐷 must be a squarefree product of an odd number of
primes.
By Exercise 17.3, if O ⊆ O0 then there is a natural surjection Cls O → Cls O0,
which is to say an order has at least as large a class number as any superorder. So
we must have # Cls O0 = 1 for a maximal order O0 ⊆ 𝐵, and by Theorem 25.4.1, this
then reduces us to 𝐷 = 2, 3, 5, 7, 13. Because # Cls O0 = 1, the maximal order O0 is
unique up to conjugation, and so fixing a choice of such a maximal order O0 up to
isomorphism we may suppose O ⊆ O0. But now the index [O0 : O] = 𝑀 = 𝐷/𝑁 is
explicitly given, and there are only finitely many suborders of bounded index; for each,
we may compute representatives of the class set in a manner similar to the example in
section 17.6. (We are aided further by deciding what assignment of Eichler symbols
and unit orders would be necessary in each case.)
There is similarly an interest in definite quaternion orders O of type number 1:
these are the orders with the property that the “local-to-global principle applies for
isomorphisms”, i.e., if O𝔭0 ' O𝔭 for all primes 𝔭 then O0 ' O. If an order has class
number 1 then it has type number 1, by Lemma 17.4.13, but one may have # Typ O = 1
but # Cls O > 1. Since an order has the same type number as its Gorenstein saturation,
i.e. Typ O = Typ Gor(O), it suffices to classify the Gorenstein orders with this property.
By the bĳection between ternary forms and quaternion orders, this is equivalently
the problem of enumerating one-class genera of primitive ternary quadratic forms.
The list was drawn up by Jagy–Kaplansky–Schiemann [JKS97] (with early work due
to Watson [Wats75]), and has been independently confirmed by Lorch–Kirschmer
[LK2013].
Theorem 25.4.6 (Watson, Jagy–Kaplansky–Schiemann, Lorch–Kirschmer). There
are exactly 794 primitive ternary quadratic forms of class number 1, corresponding
to 794 Gorenstein quaternion orders of type number 1. The largest prime dividing
a discriminant is 23, and the largest (reduced) discriminant is 28 33 72 = 338688.
There are exactly 9 corresponding to maximal quaternion orders: they have reduced
discriminants
𝐷 = 2, 3, 5, 7, 13, 30, 42, 70, 78.
Remark 25.4.7. The generalization of the class number 1 problem to quadratic forms
of more variables was pursued by Watson, who showed that one-class genera do not
exist in more than ten variables [Wats62]. Watson also tried to compile complete lists
in low dimensions, followed by work of Hanke, and recently the complete list has been
drawn up in at least 3 variables over Q by Lorch–Kirschmer [LK2013] and over totally
real fields for maximal lattices by Kirschmer [Kir2014].
Exercises
1. A short and fun proof of the equality 𝜁 (2) = 𝜋 2 /6 is due to Calabi [BCK93].
(a) Expand (1 − 𝑥 2 𝑦 2 ) −1 in a geometric series and integrate termwise over
𝑆 = [0, 1] × [0, 1] to obtain
∫∫
2 2 −1 1 1 1
(1 − 𝑥 𝑦 ) d𝑥 d𝑦 = 1 + 2 + 2 + . . . = 1 − 𝜁 (2).
𝑆 3 5 4
(b) Show that the substitution

sin 𝑢 sin 𝑣
(𝑥, 𝑦) := ,
cos 𝑣 cos 𝑢
has Jacobian 1 − 𝑥 2 𝑦 2 and maps the open triangle
𝑇 := {(𝑢, 𝑣) : 𝑢, 𝑣 > 0 and 𝑢 + 𝑣 < 𝜋/2}
bĳectively to the interior of 𝑆.

(c) Conclude that
𝜋2
∫∫ ∫∫
(1 − 𝑥 2 𝑦 2 ) −1 d𝑥 d𝑦 = d𝑢 d𝑣 =
𝑆 𝑇 8
and thus 𝜁 (2) = 𝜋 2 /6.

2. In this exercise, we give a very jazzy proof that 𝜁 (𝑘) ∈ Q𝜋 𝑘 for all 𝑘 ∈ 2Z ≥1 ,
due to Zagier [Zag94, p. 498].
(a) We start with 𝑘 = 4. Define
1 1 1
𝑓 (𝑚, 𝑛) := 3
+ 2 2
+ 3
𝑚𝑛 2𝑚 𝑛 𝑚 𝑛
for 𝑚, 𝑛 ∈ Z>0 , and prove that
1
𝑓 (𝑚, 𝑛) − 𝑓 (𝑚 + 𝑛, 𝑛) − 𝑓 (𝑚, 𝑚 + 𝑛) = .
𝑚 2 𝑛2
(b) Prove
∞
∑︁ 5
𝜁 (2) 2 = 𝑓 (𝑛, 𝑛) = 𝜁 (4).
𝑛=1
2
Conclude that 𝜁 (4) = 𝜋 4 /90 using Exercise 25.1.

(c) In general, for 𝑘 ∈ 2Z ≥2 , let
𝑘−2
1 1 ∑︁ 1 1
𝑓 (𝑚, 𝑛) = + +
𝑚𝑛 𝑘−1 2 𝑟 =2 𝑚 𝑟 𝑛 𝑘−𝑟 𝑚 𝑘−1 𝑛
and check that

∑︁ 1
𝑓 (𝑚, 𝑛) − 𝑓 (𝑚 + 𝑛, 𝑛) − 𝑓 (𝑚, 𝑚 + 𝑛) = .
0< 𝑗<𝑘
𝑚 𝑗 𝑛 𝑘− 𝑗
𝑗 even
Conclude in a similar way as in (b) that

∑︁ 𝑘 +1
𝜁 ( 𝑗)𝜁 (𝑘 − 𝑗) = 𝜁 (𝑘)
0< 𝑗<𝑘
2
𝑗 even
so by induction 𝜁 (𝑘) ∈ Q𝜋 𝑘 .
I 3. Prove Lemma 25.2.11 as follows.
(a) Let 𝑃 be a fundamental parallelogram for Λ, and for 𝜆 ∈ Λ let 𝑃𝜆 := 𝑃 + 𝜆.
For 𝑥 > 1, let
𝐷 (𝑥) := {𝑧 ∈ C : | 𝑧| ≤ 𝑥}
and
𝑁 (𝑥) := #{𝜆 ∈ Λ : 𝜆 ∈ 𝐷 (𝑥)}

𝑁 𝑃 (𝑥) := #{𝜆 ∈ Λ : 𝑃𝜆 ⊆ 𝐷 (𝑥)}
𝑁 𝑃+ (𝑥) := #{𝜆 ∈ Λ : 𝑃𝜆 ∩ 𝐷 (𝑥) ≠ ∅}.
Show that
𝑁 𝑃 (𝑥) ≤ 𝑁 (𝑥) ≤ 𝑁 𝑃+ (𝑥).
(b) Show that 𝑁 𝑃 (𝑥) ≤ 𝜋𝑥 2 /𝐴 ≤ 𝑁 𝑃+ (𝑥).
(c) Let 𝑙 be the length of a long diagonal in 𝑃. Show that for all 𝜆 ∈ Λ ∩ 𝐷 (𝑥),
we have 𝑃𝜆 ⊆ 𝐷 (𝑥 + 𝑙), so
𝜋(𝑥 + 𝑙) 2
𝑁 (𝑥) ≤ 𝑁 𝑃 (𝑥 + 𝑙) ≤ .
𝐴
Similarly, show that if 𝑃𝜆 ∩ 𝐷 (𝑥 − 𝑙) ≠ ∅ then 𝑃𝜆 ⊆ 𝐷 (𝑥) and 𝜆 ∈ 𝐷 (𝑥),
so
𝜋(𝑥 − 𝑙) 2
≤ 𝑁 𝑃+ (𝑥 − 𝑙) ≤ 𝑁 (𝑥).
𝐴
(d) Conclude that Lemma 25.2.11 holds with 𝐶 := 𝜋(2𝑙 + 𝑙 2 )/𝐴.
I 4. Using the previous exercise, we now prove the analytic class number formula
(Theorem 25.2.12).
(a) Let 𝔟 ⊂ C be a fractional ideal, and let
𝑏 𝑛 := #{𝑎 ∈ 𝔟−1 : Nm(𝑎) = 𝑛}.
Show that
∑︁ 𝜋𝑥 √
𝑏 − ≤𝐶 𝑥

𝑛 ≤𝑥 𝑛

𝐴
for a constant 𝐶 that does not depend on 𝑥 and

√︁
−1 | 𝑑|
𝐴 = Nm(𝔟 ) .
2
[Hint: Apply Lemma 25.2.11 to 𝔟−1 .]
(b) Consider the Dirichlet series
∞
1 ∑︁ 𝜋 1
𝑓 (𝑠) := 𝑏𝑛 − .
𝑤𝑁 (𝔟) 𝑛=1
𝑠 𝐴 𝑛𝑠
Show by the comparison test that 𝑓 (𝑠) converges for all Re 𝑠 > 1/2.
(c) Show that
2𝜋
res𝑠=1 𝜁 𝐾 (𝑠) = lim (𝑠 − 1)𝜁 𝐾 , [𝔟] (𝑠) = √︁ .
𝑠&1 𝑤 | 𝑑|
(d) Sum the residues over [𝔟] ∈ Cl(𝐾) to derive the theorem.
I 5. In this exercise, we prove Theorem 25.4.1: if O is a maximal order in a definite
quaternion algebra over Q of discriminant 𝐷, then # Cls O = 1 if and only if
𝐷 = 2, 3, 5, 7, 13. By the Eichler mass formula (Theorem 25.3.15), we have
# Cls O = 1 if and only if
1 𝜑(𝐷)
=
𝑤 12
where 𝑤 = #O/{±1}.
(a) Show (cf. 11.5.13) that if 𝐷 > 3 then 𝑤 ≤ 3.
(b) Show that # Cls O = 1 for 𝐷 = 2, 3. (The case 𝐷 = 2 is the Hurwitz
order and 𝐷 = 3 is considered in Exercise 11.12. In fact, these orders are
Euclidean with respect to the reduced norm: see the next exercise.)
(c) Show that if 𝐷 is a squarefree positive integer with an odd number of prime
factors and 𝜑(𝐷)/12 ∈ {1, 1/2, 1/3}, then 𝐷 ∈ {5, 7, 13, 42}.
(d) Prove that # Cls O = 1 for 𝐷 = 5, 7, 13 (cf. Exercise 17.10).
(e) Show that # Cls O = 2 for 𝐷 = 42.
6. Let O be a definite quaternion order over Z. If O is Euclidean, then # Cls O = 1,
and we saw in 11.3.1 that the Hurwitz order O of discriminant 𝐷 = 2 is Euclidean
with respect to the reduced norm.
(a) Show that if O is norm Euclidean, then O is maximal.
(b) Show that O is Euclidean with respect to the reduced norm if and only if
for all 𝛾 ∈ 𝐵, there exists 𝜇 ∈ O such that nrd(𝛾 − 𝜇) < 1.
(c) Show that if O is maximal, then O is norm Euclidean if and only if
𝐷 = 2, 3, 5.
7. Generalizing the previous exercise, we may ask for the Euclidean ideal classes
in maximal orders. We will show that there are no nonprincipal Euclidean two-
sided ideal classes in maximal definite quaternion orders over Z. [This exercise
was suggested by Pete L. Clark.]
Let O be a maximal definite quaternion order over Z of discriminant 𝐷, and let

𝐼 be a two-sided O-ideal. We say that 𝐼 is (norm) Euclidean if for all 𝛾 ∈ 𝐵,
there exists 𝜇 ∈ 𝐼 such that nrd(𝛾 − 𝜇) < nrd(𝐼).
(a) Show that if 𝐼 is principal, then 𝐼 is Euclidean if and only if O is Euclidean.
In general, show that 𝐼 is Euclidean if and only if 𝛼𝐼 is Euclidean for all
𝛼 ∈ 𝐵× , so we may ask if [𝐼] is Euclidean.
(b) Show that if [𝐼] is nontrivial and Euclidean, then # Cls O = 2 and PicZ (O)
is cyclic, generated by [𝐼]. [Hint: argue by induction on nrd(𝐽). Then use
the fact that Pic(O) is a group of exponent 2.]
(c) Show that # Cls O = 2 if and only if 𝐷 = 11, 17, 19, 30, 42, 70, 78.
(d) Show that PicZ (O) = 1 for 𝐷 = 11, 17, 19, and for the remaining dis-
criminants 𝐷 = 30, 42, 70, 78 that the nontrivial class [𝐼] is not norm
Euclidean.
Chapter 26
Classical zeta functions
In this chapter, we prove the Eichler mass number for a definite quaternion order over
a totally real field using classical analytic methods.
26.1 ⊲ Eichler mass formula

In the previous chapter, we saw a sketch of how analytic methods with quaternionic
zeta functions provide a weighted class number formula for a quaternion order in a
definite quaternion algebra over Q, analogous to the analytic class number formula of
Dirichlet for a quadratic field. The main result of this section is then the generalization
of the Eichler mass formula to a definite quaternion order over a totally real number
field. In this section, we give the statement of this result.
26.1.1. Let 𝐹 be a totally real number field of degree 𝑛 = [𝐹 : Q], absolute dis-
criminant 𝑑 𝐹 , and ring of integers 𝑅 := Z𝐹 . Let ℎ 𝐹 be the class number of 𝐹.
Let ∑︁ 1
𝜁 𝐹 (𝑠) :=
𝔞⊆𝑅
N(𝔞) 𝑠
be the Dedekind zeta function of 𝐹, where N(𝔞) = [𝑅 : 𝔞] ∈ Z>0 . Let 𝐵 be a totally
definite quaternion algebra over 𝐹 of discriminant 𝔇. Let O ⊂ 𝐵 be an 𝑅-order with
reduced discriminant discrd(O) = 𝔑.
O
For a prime 𝔭 | 𝔑 with N(𝔭) = 𝑞, let ∈ {−1, 0, 1} be the Eichler symbol
𝔭
(Definition 24.3.2), and let
 1 + 1/𝑞, if (O | 𝑝) = 1;
1 − N(𝔭) −2




𝜆(O, 𝔭) := = 1 − 1/𝑞, if (O | 𝑝) = −1; (26.1.2)
O
N(𝔭) −1 

1−  1 − 1/𝑞 2 , if (O | 𝑝) = 0.
𝔭 
26.1.3. We saw in Lemma 17.7.13 that for each definite order O, the group O1 of
units of reduced norm 1 is a finite group; we will see in Lemma 26.5.1 that further the
group O× /𝑅 × is finite. For a right O-ideal 𝐽, the automorphism group of 𝐽 (as a right
419
420 CHAPTER 26. CLASSICAL ZETA FUNCTIONS
O-module) consists of right multiplication maps by elements 𝜇 ∈ 𝐵× with 𝜇𝐽 = 𝐽,

i.e., 𝜇 ∈ OL (𝐽) × .
It was already evident in the Eichler mass formula (and remains a general principle
in mathematics) that one often gets a better count of objects when they are weighted
by the inverse size of the automorphism group, so we weight a right ideal class [𝐽] by
[OL (𝐽) × : 𝑅 × ] −1 and make the following definition of a weighted class number.
Definition 26.1.4. Define the mass of Cls O to be
∑︁
mass(Cls O) = [OL (𝐽) × : 𝑅 × ] −1 .
[𝐽 ] ∈Cls O
Main Theorem 26.1.5 (Eichler mass formula). With notation as in 26.1.1, we have
2𝜁 𝐹 (2) 3/2 Ö
mass(Cls O) = 𝑑 ℎ 𝐹 N(𝔑) 𝜆(O, 𝔭). (26.1.6)
(2𝜋) 2𝑛 𝐹 𝔭 |𝔑
26.1.7. The functional equation for the Dedekind zeta function relates 𝑠 to 1− 𝑠, giving
an alternative way of writing (26.1.6) as
2𝜁 𝐹 (2) 3/2 | 𝜁 𝐹 (−1)|
𝑑 = . (26.1.8)
(2𝜋) 2𝑛 𝐹 2𝑛−1
We notice that the Eichler mass formula then implies that 𝜁 𝐹 (−1) ∈ Q.
Remark 26.1.9. More generally, the rationality of the values 𝜁 𝐹 (−𝑛) with 𝑛 ∈ Z>0 is
a theorem of Siegel [Sie69] and Deligne–Ribet [DR80].
Remark 26.1.10. The weighting in the mass is what makes Main Theorem 26.1.5 so
simple. In the (unlikely) situation where 𝑤 𝐽 = 𝑤 O is independent of 𝐽, we would have
a formula for the class number, but more generally we will need to take account of unit
groups by computing embedding numbers of cyclotomic quadratic orders: we will do
this in Chapter 30.
We now make the formula (26.1.6) a bit more explicit for the case of Eichler orders.
26.1.11. Let O be an Eichler order of level 𝔐, so that 𝔑 = 𝔇𝔐 with 𝔇, 𝔐 coprime.
Then
  −1, if 𝔭 | 𝔇;
O



= 1, if 𝔭 | 𝔐;
𝔭 
 ∗,

if 𝔭 - 𝔑.

Accordingly, we define the generalized Euler 𝜑-function and Dedekind 𝜓-function by
Ö Ö 1

𝜑(𝔇) := (N(𝔭) − 1) = N(𝔇) 1−
N(𝔭)
𝔭 |𝔇 𝔭 |𝔇
Ö Ö 1

𝜓(𝔐) := N(𝔭) 𝑒−1 (N(𝔭) + 1) = N(𝔐) 1+
N(𝔭)
𝑒 𝔭 k𝔐 𝔭 |𝔐
(recalling 𝔇 is squarefree, with the natural extension 𝜑(𝔇) = #(𝑅/𝔇) × for all 𝔇).
The 𝜓-function computes a unit index: see Lemma 26.6.7.
26.2. ANALYTIC CLASS NUMBER FORMULA 421
The Eichler mass formula (Main Theorem 26.1.5) for Eichler orders then reads as
follows.
Theorem 26.1.12 (Eichler mass formula, Eichler orders). With notation as in 26.1.11,
we have
2𝜁 𝐹 (2) 3/2
mass(Cls O) = 𝑑 ℎ 𝐹 𝜑(𝔇)𝜓(𝔐). (26.1.13)
(2𝜋) 2𝑛 𝐹
Remark 26.1.14. The Eichler mass formula in the form (26.1.13) for maximal orders
was proven by Eichler (working over a general totally real field) using the techniques in
this chapter [Eic38b, Satz 1], and was extended to squarefree level 𝔑 (i.e., hereditary
orders) again by Eichler [Eic56a, §4]. This was extended by Brzezinski [Brz90, (4.6)]
to a general formula over Q and by Körner [Kör87, Theorem 1], using idelic methods.
The classical method to prove Main Theorem 26.1.5 is similar to the one we
sketched over Q in chapter 25, with some added technicalities of working over a
number field. We follow this approach, first proving the formula when O is a maximal
order, and then deducing the general case. We will return in chapter 29 and reconsider
the Eichler mass formula from an idelic point of view, thinking of it as a special case
of a volume formula (for a finite set of “quotient points”). It is hoped that this chapter
will serve to show both the power and limits of classical methods before we build upon
them using idelic methods.
26.2 Analytic class number formula

In this section, in preparation for the quaternionic case we briefly review what we
need from the analytic class number formula for a number field 𝐹. References for this
material include Borevich–Shafarevich [BS66, Chapter 5], Lang [Lang94, Chapter
VI], and Neukirch [Neu99, Chapter VII].
We begin by setting some notation that will be used throughout the rest of this
chapter. Let 𝐹 be a number field of degree 𝑛 := [𝐹 : Q] with 𝑟 real places and 𝑐
complex places, so that 𝑛 = 𝑟 + 2𝑐. Let 𝑅 := Z𝐹 be the ring of integers in 𝐹, and let 𝑑 𝐹
be the discriminant of 𝐹. Let 𝑤 𝐹 be the number of roots of unity in 𝐹, let ℎ 𝐹 := # Cl 𝑅
be the class number of 𝐹, and let Reg𝐹 be the regulator of 𝐹 (the covolume of 𝑅 ×
under the Minkowski embedding).
Define the Dedekind zeta function for 𝑠 ∈ C with Re(𝑠) > 1 by
∑︁ 1
𝜁 𝐹 (𝑠) :=
𝔞⊆𝑅
N(𝔞) 𝑠
where the sum is over all nonzero ideals of 𝑅 and N(𝔞) = #(𝑅/𝔞) = [𝑅 : 𝔞] is the
absolute norm; we have N(𝔞) = Nm(𝔞) with norm taken from 𝐹 to Q and positive
generator chosen.
26.2.1. The Dedekind zeta function converges for Re 𝑠 > 1 and has an Euler product
Ö −1
1
𝜁 𝐹 (𝑠) = 1− (26.2.2)
𝔭
N(𝔭) 𝑠
where the product is over all nonzero primes of 𝑅—this follows formally from unique
factorization of ideals after one shows that the pruned product converges.
The Dedekind zeta function has properties analogous to the Riemann zeta function,
which is the special case 𝐹 = Q. In particular, we can extend 𝜁 𝐹 (𝑠) in a manner
analogous to 25.2.1 to Re 𝑠 > 0. For 𝑎 ∈ C we write 𝜁 𝐹∗ (𝑎) for the leading coefficient
in the Laurent series expansion for 𝜁 𝐹 at 𝑠 = 𝑎.
Theorem 26.2.3 (Analytic class number formula). 𝜁 𝐹 (𝑠) has analytic continuation to
Re 𝑠 > 0, with a simple pole at 𝑠 = 1 having residue
2𝑟 (2𝜋) 𝑐
𝜁 𝐹∗ (1) = lim (𝑠 − 1)𝜁 𝐹 (𝑠) = √︁ ℎ 𝐹 Reg𝐹 . (26.2.4)
𝑠→1 𝑤 𝐹 | 𝑑𝐹 |
Remark 26.2.5. The formula (26.2.4) is known as Dirichlet’s analytic class number
formula (even though the original form of Dirichlet’s theorem concerned quadratic
forms rather than classes of ideals, so is closer to Theorem 25.2.12).
Example 26.2.6. When 𝐹 is an imaginary quadratic field (𝑟 = 0 and 𝑐 = 1) we have

Reg𝐹 = 1 and Theorem 26.2.3 is Theorem 25.2.12.
Before we finish this section, we review a few ingredients from the proof of the
analytic class number formula (26.2.4) to set up the Eichler mass formula.
26.2.7. We first write the Dedekind zeta function as a sum over ideals in a given ideal
class [𝔟] ∈ Cl(𝑅): we define the partial zeta function
∑︁ 1
𝜁 𝐹 , [𝔟] (𝑠) := (26.2.8)
𝔞⊆𝑅
N(𝔞) 𝑠
[𝔞]=[𝔟]
convergent for Re 𝑠 > 1 by comparison to the harmonic series, so that

∑︁
𝜁 𝐹 (𝑠) = 𝜁 𝐹 , [𝔟] (𝑠). (26.2.9)
[𝔟] ∈Cl 𝑅
Now note that [𝔞] = [𝔟] if and only if 𝔞 = 𝑎𝔟 for some nonzero
𝑎 ∈ 𝔟−1 = {𝑥 ∈ 𝐹 : 𝑥𝔟 ⊆ 𝑅},
so there is a bĳection between nonzero ideals 𝔞 ⊆ 𝑅 such that [𝔞] = [𝔟] and the set of
nonzero elements in 𝔟−1 /𝑅 × . So
1 ∑︁ 1
𝜁 𝐹 , [𝔟] (𝑠) = . (26.2.10)
N(𝔟) 𝑠 Nm(𝑎) 𝑠
0≠𝑎 ∈𝔟−1 /𝑅 ×
One now reduces to a problem concerning lattice points in a fundamental domain for
the action of 𝑅 × , and examining the residue of the pole at 𝑠 = 1 fits into a more general
framework (invoked again below).
26.2. ANALYTIC CLASS NUMBER FORMULA 423
Definition 26.2.11. A cone 𝑋 ⊆ R𝑛 is a subset closed under multiplication by positive

scalars, so 𝑡 𝑋 = 𝑋 for all 𝑡 ∈ R>0 .
Theorem 26.2.12. Let 𝑋 ⊆ R𝑛 be a cone. Let 𝑁 : 𝑋 → R>0 be a function satisfying
𝑁 (𝑡𝑥) = 𝑡 𝑛 𝑁 (𝑥) for all 𝑥 ∈ 𝑋, 𝑡 ∈ R>0 .
Suppose that
𝑋 ≤1 := {𝑥 ∈ 𝑋 : 𝑁 (𝑥) ≤ 1} ⊆ R𝑛 (26.2.13)
is a bounded subset with volume vol(𝑋 ≤1 ). Let Λ ⊆ R𝑛
be a (full) Z-lattice in R𝑛 , and
let
∑︁ 1
𝜁Λ,𝑋 (𝑠) := .
𝜆∈𝑋 ∩Λ
𝑁 (𝜆) 𝑠
Then 𝜁Λ,𝑋 (𝑠) converges for Re 𝑠 > 1 and has a simple pole at 𝑠 = 1 with residue
∗ vol(𝑋 ≤1 )
𝜁Λ,𝑋 (1) = lim (𝑠 − 1)𝜁Λ,𝑋 (𝑠) = .
𝑠&1 covol(Λ)
Proof. See Borevich–Shafarevich [BS66, Chapter 5, Section 1.1, Theorem 1].
26.2.14. To apply Theorem 26.2.12 for 𝜁 𝐹 , [𝔟] (𝑠), we embed 𝐹 ↩→ 𝐹R ' R𝑟 × C𝑐 and
we equip 𝐹R with the inner product
𝑟
∑︁ 𝑐
∑︁
h𝑥, 𝑦i = 𝑥𝑖 𝑦 𝑖 + 2 Re(𝑥𝑟 + 𝑗 𝑦 𝑟 + 𝑗 ) (26.2.15)
𝑖=1 𝑗=1
for 𝑥 = (𝑥𝑖 )𝑖 , 𝑦 = (𝑦 𝑖 )𝑖 ∈ 𝐹R . This inner product modifies the usual one by rescaling
complex coordinates, and the volume form vol induced by h , i is 2𝑐 times the standard
Lebesgue volume on R𝑟 × C𝑐 . With this convention, we have h𝑥, 1i = Tr𝐹 |Q (𝑥) and
√︁
covol(𝑅) = | 𝑑 𝐹 |.
We then take Λ to be the image of 𝔟−1 , and take 𝑋 to be a cone fundamental
domain for the action of the unit group 𝑅 × . The absolute norm N(𝑥) = |Nm𝐹 |Q (𝑥)|
then satisfies the required homogeneity property, and 𝑋 ≤1 is bounded, so by Theorem
26.2.12,
1 vol(𝑋 ≤1 )
𝜁 𝐹∗ , [𝔟] (1) = . (26.2.16)
N(𝔟) covol(Λ)
We have √︁
covol(𝑅) | 𝑑𝐹 |
covol(Λ) = = . (26.2.17)
N(𝔟) N(𝔟)
It requires a bit more work to compute vol(𝑋 ≤1 ).
Proposition 26.2.18. We have
2𝑟 (2𝜋) 𝑐 Reg𝐹
vol(𝑋 ≤1 ) = (26.2.19)
𝑤𝐹
Proof. See Exercise 26.3: the proof is well-summarized as a “change of variables”,

but the reader may prefer the idelic point of view (Chapter 29) instead, where the
integrals are ‘easier’. A detailed proof can be found in Borevich–Shafarevich [BS66,
§5.1.3], Lang [Lang94, §VI.3, Theorem 3], and Neukirch [Neu99, §VII.5].
Plugging (26.2.19) and (26.2.17) into (26.2.16),

2𝑟 (2𝜋) 𝑐
𝜁 𝐹∗ , [𝔟] (1) = √︁ Reg𝐹 ; (26.2.20)
𝑤 𝐹 | 𝑑𝐹 |
note in particular that this does not depend on the class [𝔟]! The analytic class number
formula (Theorem 26.2.3) then follows as
∑︁ 2𝑟 (2𝜋) 𝑐
𝜁 𝐹∗ (1) = 𝜁 𝐹∗ , [𝔟] (1) = √︁ Reg𝐹 ℎ 𝐹 . (26.2.21)
[𝔟] 𝑤 𝐹 | 𝑑𝐹 |
26.3 Classical zeta functions of quaternion algebras

We now embark on a proof in our quaternionic setting, mimicking the above. We
retain our notation on the number field 𝐹. We further let throughout 𝐵 be a quaternion
algebra over 𝐹 of discriminant 𝔇 and let O ⊆ 𝐵 be an 𝑅-order. (Our emphasis will be
on the case O a maximal order, but many definitions carry through.)
To begin, in this section we define the classical zeta function and show it has an
Euler product.
26.3.1. Let 𝐼 be an invertible, integral right O-ideal, so that 𝐼 ⊆ O, and by definition
𝐼 is sated so OR (𝐼) = O. Recall we have defined N(𝐼) = #(O/𝐼); we have N(𝐼) =
N(nrd(𝐼)) 2 (Paragraph 16.4.10).
For example, if 𝔞 ⊆ 𝑅 is a nonzero ideal then N(𝔞O) = N(𝔞) 4 .
We then define the (classical) zeta function of O to be
∑︁ 1 ∑︁ 𝑎 𝔫 (O)
𝜁O (𝑠) := = (26.3.2)
𝐼 ⊆O
N(𝐼) 𝑠 𝔫
N(𝔫) 2𝑠
where the first sum is over all (nonzero) integral, invertible right O-ideals 𝐼 and in the
second sum we define
𝑎 𝔫 (O) := #{𝐼 ⊆ O : nrd(𝐼) = 𝔫} (26.3.3)
(and 𝑎 𝔫 (O) is finite by Lemma 17.7.26).

Lemma 26.3.4. If O, O0 are locally isomorphic, then 𝑎 𝔫 (O) = 𝑎 𝔫 (O0) for all 𝔫.
Proof. We use the local-global dictionary for lattices (Theorem 9.4.9). To ease paren-
theses in the notation, we work in the completion, but one can also work just in the
localization. For all 𝔭, we have O𝔭0 = 𝜈𝔭−1 O𝔭 𝜈𝔭 for some 𝜈𝔭 ∈ 𝐵𝔭× , and we may take
𝜈𝔭 = 1 for all but finitely many 𝔭; the element 𝜈𝔭 is well-defined up to left multiplication
by O×𝔭 and right multiplication by O𝔭0× .
26.3. CLASSICAL ZETA FUNCTIONS OF QUATERNION ALGEBRAS 425
Then to an integral, invertible right O-ideal 𝐼, we associate the unique lattice 𝐼 0 such
that 𝐼𝔭0 = 𝜈𝔭−1 𝐼𝔭 𝜈𝔭 ; such a lattice is well-defined independent of the choice of 𝜈𝔭 . By
construction, OR (𝐼𝔭0 ) = O𝔭0 so OR (𝐼 0) = O0. And since 𝐼 is integral, 𝐼𝔭 ⊆ O𝔭 whence
𝐼𝔭0 ⊆ 𝜈𝔭−1 𝐼𝔭 𝜈𝔭 ⊆ O𝔭0 and 𝐼 is locally integral hence integral. Since 𝐼 is invertible,
𝐼 is locally principal, so 𝐼 0 is also locally principal, hence invertible. Finally, again
checking locally, we have nrd(𝐼 0) = nrd(𝐼).
Repeating this argument going from 𝐼 0 to 𝐼, we see that the corresponding sets of
ideals are in bĳection, as claimed.
26.3.5. From Lemma 26.3.4, we see that 𝜁O (𝑠) only depends on the genus of O. Since
there is a unique genus of maximal orders in 𝐵, following the number field case we
will write 𝜁 𝐵 (𝑠) = 𝜁O (𝑠) where O is any maximal order.
Our next order of business is to establish an Euler product for 𝜁O (𝑠). We prove a
more general result on the factorization of invertible lattices.
Lemma 26.3.6. Let 𝐼 be an invertible, integral lattice and suppose that nrd(𝐼) = 𝔪𝔫
with 𝔪, 𝔫 ⊆ 𝑅 coprime ideals. Then there exists a unique invertible, integral lattice 𝐽
such that 𝐼 is compatible with 𝐽 −1 with 𝐼𝐽 −1 integral and nrd(𝐽) = 𝔪.
Proof. We use the local-global dictionary for lattices, and we define 𝐽 ⊆ 𝐵 to be the
unique lattice such that
 𝐼 = O (𝔭) , if 𝔭 - 𝔪𝔫;
 (𝔭)



𝐽 (𝔭) := 𝐼 (𝔭) , if 𝔭 | 𝔪; (26.3.7)

 O (𝔭) ,

if 𝔭 | 𝔫.

We have OR (𝐽) = O and nrd(𝐽) = 𝔪, since these statements hold locally. Integrality
and invertibility are local; since these are true for 𝐼 they are true for 𝐽. Finally, we
compute that (𝐼𝐽 −1 ) (𝔭) = O (𝔭) for all 𝔭 - 𝔫 and (𝐼𝐽 −1 ) (𝔭) = 𝐼 (𝔭) for 𝔭 | 𝔫, so 𝐼𝐽 −1
is locally integral and hence integral. The uniqueness of 𝐽 can be verified directly
(Exercise 26.4).
26.3.8. Consider the situation of Lemma 26.3.6. Let 𝐼 0 = 𝐼𝐽 −1 . Then 𝐼 = 𝐼 0 𝐽, and 𝐼 0

is integral, invertible (Paragraph 16.5.3) and compatible with 𝐽. Since nrd(𝐼 0) = 𝔫,
we have “factored” 𝐼.
We have OR (𝐼 0) = OL (𝐽) by compatibility, but this common order is only locally
isomorphic to O, since 𝐼 0, 𝐽 are locally principal but not necessarily principal. So in
a sense, this factorization occurs not over O but over the genus of O—but this is a
harmless extension.
Proposition 26.3.9. If 𝔪, 𝔫 are coprime, then 𝑎 𝔪𝔫 (O) = 𝑎 𝔪 (O)𝑎 𝔫 (O).
Proof. Write 𝐴𝔫 (O) for the set of integral, invertible right O-ideals 𝐼 with nrd(𝐼) = 𝔫.
Then #𝐴𝔫 (O) = 𝑎 𝔫 (O). According to Lemma 26.3.6, there is a map
𝐴𝔪𝔫 (O) → 𝐴𝔫 (O)

(26.3.10)
𝐼 ↦→ 𝐽
We claim that this map is surjective and that each fiber has cardinality 𝑎 𝔪 (O). Indeed,
these statements follow at the same time from the following observation: if 𝐽 ∈ 𝐴𝔫 (O)
with O0 = OL (𝐽), then for each 𝐼 0 ∈ 𝐴𝔪 (O0), we have 𝐼 0 compatible with 𝐽 and
𝐼 = 𝐼 0 𝐽 ∈ 𝐴𝔪 , and conversely; so the fiber of (26.3.10) is identified with 𝐴𝔪 (O0), of
cardinality 𝑎 𝔪 (O0) = 𝑎 𝔪 (O) by Lemma 26.3.4.
26.3.11. From Proposition 26.3.9 and unique factorization of ideals in 𝑅, we find that
𝜁O has an Euler product Ö
𝜁O (𝑠) = 𝜁O𝔭 (𝑠) (26.3.12)
𝔭
where
∞
∑︁ 1 ∑︁ 𝑎 𝔭𝑒 (O)
𝜁O𝔭 (𝑠) := = . (26.3.13)
𝐼𝔭 ⊆O𝔭
N(𝐼𝔭 ) 𝑠
𝑒=0
N(𝔭) 2𝑠
Remark 26.3.14. Zeta functions of semisimple algebras over a number field can be
defined in the same way as in (26.3.2), following Solomon [Sol77]: see the survey on
analytic methods in noncommutative number theory by Bushnell–Reiner [BR85].
Remark 26.3.15. The world of 𝐿-functions is rich and very deep: for a beautiful
survey of the analytic theory of automorphic 𝐿-functions in historical perspective,
see Gelbart–Miller [GM2004]. In particular, we have not given a general definition
of zeta functions (or 𝐿-functions) in this section, but it is generally agreed that the
Selberg class incorporates the minimal essential featuers: definition as a Dirichlet
series, meromorphic continuation to the complex plane, Euler product, and functional
equation. See e.g. Conrey–Ghosh [CG93] and the references therein.
26.4 Counting ideals in a maximal order

We now count ideals of prime power norm. By the local-global dictionary, there is a
bĳection
∼ {𝐼 ⊆ O : nrd(𝐼 ) = 𝔭𝑒 }.
{𝐼 ⊆ O : nrd(𝐼) = 𝔭𝑒 } −→ 𝔭 𝔭 𝔭
so it suffices to count the number of ideals in the local case. In this section, we carry
out this count for maximal orders.
So let 𝔭 ⊂ 𝑅 be a (nonzero) prime and let 𝑞 := Nm(𝔭). Let O𝔭 ⊂ 𝐵𝔭 be a maximal
order. Let
𝑎 𝔭𝑒 (O𝔭 ) = #{𝐼𝔭 = 𝛼𝔭 O𝔭 ⊆ O𝔭 : nrd(𝐼𝔭 ) = 𝔭𝑒 }
count the number of right integral O𝔭 -ideals of norm 𝔭𝑒 . Since O𝔭 is maximal, every
nonzero ideal is invertible; and because 𝑅𝔭 is a DVR, all such invertible ideals are
principal.
Lemma 26.4.1. Let O𝔭 ⊂ 𝐵𝔭 be a maximal order and let 𝑒 ∈ Z ≥0 .
(a) If 𝐵𝔭 is a division ring, then every right integral O𝔭 -ideal is a power of the
maximal ideal and 𝑎 𝔭𝑒 (O𝔭 ) = 1.
26.4. COUNTING IDEALS IN A MAXIMAL ORDER 427
(b) If 𝐵𝔭 ' M2 (𝐹𝔭 ), so that O𝔭 ' M2 (𝑅𝔭 ), then the set of right integral O𝔭 -ideals
of reduced norm 𝔭𝑒 is in bĳection with the set
𝑢
𝜋 0 𝑣
: 𝑢, 𝑣 ∈ Z ≥0 , 𝑢 + 𝑣 = 𝑒 and 𝑐 ∈ 𝑅/𝔭
𝑐 𝜋𝑣
and
𝑎 𝔭𝑒 (O𝔭 ) = 1 + 𝑞 + · · · + 𝑞 𝑒 . (26.4.2)
Proof. For (a), if 𝔭 is ramified then by the work of section 13.3, there is a unique
maximal order O𝔭 with a unique (two-sided) maximal ideal 𝐽𝔭 having nrd(𝐽𝔭 ) = 𝔭,
and all ideals of O𝔭 are powers of 𝐽𝔭 .
To prove (b), we appeal to the theory of elementary divisors (applying column
operations, acting on the right). Suppose O𝔭 = M2 (𝑅𝔭 ). Let 𝐼𝔭 = 𝛼𝔭 O𝔭 be a right
integral O𝔭 -ideal of norm 𝔭𝑒 and let 𝜋 be a uniformizer for 𝔭. Then by the theory of
elementary divisors, we can write
𝑢
𝜋 0
𝛼𝔭 =
𝑐 𝜋𝑣
for unique 𝑢, 𝑣 ∈ Z ≥0 with 𝑢 + 𝑣 = 𝑒 and 𝑐 ∈ 𝑅 is uniquely defined as element of

Í
𝑅/𝔭 𝑣 (Exercise 26.6). It follows that the number of such ideals is equal to 𝑒𝑣=0 𝑞 𝑣 =
1+𝑞 +···+𝑞 . 𝑒
26.4.3. There is an alternate bĳection that is quite useful. We say an integral right
O-ideal 𝐼 is 𝔭-primitive if it does not contain 𝔭O (so we cannot write 𝐼 = 𝔭𝐼 0 with 𝐼 0
integral).
For a commutative ring 𝐴, we define the projective line over 𝐴 to be the set
P1 ( 𝐴) := {(𝑥, 𝑦) ∈ 𝐴2 : 𝑥 𝐴 + 𝑦 𝐴 = 𝐴}/𝐴×
and write equivalence classes (𝑥 : 𝑦) ∈ P1 ( 𝐴).

Then for O𝔭 = M2 (𝑅𝔭 ), there is a bĳection
P1 (𝑅/𝔭𝑒 ) → {𝐼𝔭 ⊆ O𝔭 : 𝐼𝔭 primitive and nrd(𝐼𝔭 ) = 𝔭𝑒 }

(26.4.4)

𝑎 0
(𝑎 : 𝑐) ↦→ O + 𝔭 𝑒 O𝔭
𝑐 0 𝔭
Any ideal of the form in the right-hand side of (26.4.4) is a primitive right integral
O𝔭 -ideal with reduced norm 𝔭𝑒 . Conversely, suppose that 𝐼𝔭 = 𝛼𝔭 O𝔭 is primitive. We
have nrd(𝛼𝔭 ) ≡ 0 (mod 𝔭𝑒 ). We find a “standard form” for 𝐼𝔭 by looking at the left
kernel of 𝛼𝔭 . Let
𝐿 := {𝑥 ∈ (𝑅/𝔭𝑒 ) 2 : 𝑥𝛼𝔭 ≡ 0 (mod 𝔭𝑒 )}.
We claim that 𝐿 is a free 𝑅/𝔭𝑒 -module of rank 1. Indeed, 𝐿 is one-dimensional over
𝑅/𝔭 since 𝐼𝔭 is primitive and so 𝛼𝔭 . 0 (mod 𝔭); by Hensel’s lemma, it follows that
𝐿 is also one-dimensional. Therefore, there is a unique generator (𝑎 : 𝑐) ∈ P1 (𝑅/𝔭𝑒 )
for 𝐿. We therefore define an map 𝐼𝔭 ↦→ (−𝑐 : 𝑎) and verify that this furnishes an
inverse to (26.4.4).
Since #P1 (𝑅/𝔭𝑒 ) = 𝑞 𝑒 + 𝑞 𝑒−1 for 𝑒 ≥ 1, we recover the count (26.4.2) as

b𝑒/2c
∑︁
𝑎 𝔭𝑒 (O𝔭 ) = #P1 (𝑅/𝔭𝑒−2𝑖 ) = 𝑞 𝑒 + 𝑞 𝑒−1 + · · · + 𝑞 + 1.
𝑖=0
26.4.5. Lemma 26.4.1 implies a factorization of 𝜁 𝐵 𝑝 (𝑠) = 𝜁O𝔭 (𝑠). Write

∞ −1
∑︁ 1 1
𝜁 𝐹𝑝 (𝑠) = = 1 − (26.4.6)
𝑒=0
𝑞 𝑒𝑠 𝑞𝑠
Î
so that 𝜁 𝐹 (𝑠) = 𝔭 𝜁 𝐹𝑝 (𝑠).

−1 (
1 1, if 𝔭 is ramified;
𝜁 𝐵𝔭 (𝑠) = 1 − 2𝑠 · −1
𝑞 1 − 1/𝑞 2𝑠−1 , if 𝔭 is split.
Equivalently,
(
𝜁 𝐹𝔭 (2𝑠), if 𝔭 is ramified;
𝜁 𝐵𝔭 (𝑠) =
𝜁 𝐹𝔭 (2𝑠)𝜁 𝐹𝔭 (2𝑠 − 1), if 𝔭 is split.
Proof. We use Lemma 26.4.1. If 𝐵𝔭 is a division ring, then Lemma 26.4.1(a) applies,
and the result is immediate. For the second case, we compute
∞ ∞
∑︁ 1 + 𝑞 + · · · + 𝑞𝑒∑︁ 1 − 𝑞 𝑒+1
𝜁 𝐵𝔭 (𝑠) = =
𝑒=0
𝑞 2𝑒𝑠 𝑒=0
(1 − 𝑞)𝑞 2𝑒𝑠
∞ ∞
!
1 ∑︁ 1 ∑︁ 1
= −𝑞
1 − 𝑞 𝑒=0 𝑞 2𝑒𝑠 𝑞 (2𝑠−1)𝑒

𝑒=0
(26.4.8)
1 1 𝑞
= −
1 − 𝑞 1 − 1/𝑞 2𝑠 1 − 1/𝑞 2𝑠−1
−1 −1
1 1
= 1 − 2𝑠 1 − 2𝑠−1
𝑞 𝑞
as claimed.
We have proven the following result.

Theorem 26.4.9 (Factorization of 𝜁 𝐵 (𝑠), maximal order). Let 𝐵 be a quaternion
algebra of discriminant 𝔇 = disc 𝐵. Then
Ö Ö
𝜁 𝐵 (𝑠) = 𝜁 𝐵𝔭 (𝑠) = 𝜁 𝐹 (2𝑠)𝜁 𝐹 (2𝑠 − 1) (1 − N(𝔭) 1−2𝑠 ). (26.4.10)
𝔭 𝔭 |𝔇
Proof. Combine the Euler product 26.3.11 with Corollary 26.4.7.

26.5. EICHLER MASS FORMULA: MAXIMAL ORDERS 429
Corollary 26.4.11. 𝜁 𝐵 (𝑠) has a simple pole at 𝑠 = 1 with residue
𝜁 𝐹∗ (1) Ö
𝜁 𝐵∗ (1) = lim (𝑠 − 1)𝜁 𝐵 (𝑠) = 𝜁 𝐹 (2) (1 − N(𝔭) −1 ). (26.4.12)
𝑠→1 2
𝔭 |𝔇
Proof. Since 𝜁 𝐹 (𝑠) has only a simple pole at 𝑠 = 1, with residue computed in Theorem
26.2.3, there is a single simple pole of 𝜁 𝐵 (𝑠) at 𝑠 = 1.
26.5 Eichler mass formula: maximal orders

We now finish the proof of the Eichler mass formula (Main Theorem 26.1.5) for
maximal orders (26.1.13). In the next section, we will deduce the general formula
from it: for a nonmaximal order, there are extra factors at each prime dividing the
discriminant, and it is simpler to account for those in a separate step.
In this section, we now suppose that 𝐵 is definite, so 𝐹 is a totally real field (and
𝑑 𝐹 > 0). In particular, 𝐵 is a division algebra. We saw in 26.1.3 that it was natural
to weight ideal classes inversely by the size of their automorphism group (modulo
scalars). To this end, and noting 𝑅 × E O× is central so normal, we prove the following
lemma.
Lemma 26.5.1. The group O× /𝑅 × is finite.
Proof. In Lemma 17.7.13, we proved that
O1 := {𝛾 ∈ O× : nrd(𝛾) = 1}
is a finite group by embedding O ↩→ 𝐵R ' R4𝑛 as a Euclidean lattice with respect to

the absolute reduced norm (see 17.7.10). Since O1 ∩ 𝑅 × = {±1}, the reduced norm
gives an exact sequence
O1 O× nrd 𝑅 ×
1→ → × −−→ ×2 . (26.5.2)
{±1} 𝑅 𝑅
By Dirichlet’s unit theorem, the group 𝑅 × is finitely generated (of rank 𝑟 + 𝑐 − 1), so
the group 𝑅 × /𝑅 ×2 is a finite abelian 2-group. The result follows.
We will examine unit groups in detail in Chapter 32. With this finiteness statement
in hand, we make the following definition.
Definition 26.5.3. Define the mass of O to be

∑︁ 1
mass(Cls O) :=
𝑤𝐽
[𝐽 ] ∈Cls O
where 𝑤 𝐽 = [OL (𝐽) × : 𝑅 × ] ∈ Z ≥1 .

Theorem 26.5.4 (Eichler’s mass formula). Let O be a maximal order in a totally

definite quaternion algebra 𝐵 of discriminant 𝔇. Then
2
mass(Cls O) = ℎ 𝐹 𝑑 𝐹3/2 𝜑(𝔇)
(2𝜋) 2𝑛
Î
where 𝜑(𝔇) = 𝔭 |𝔇 (N(𝔭) − 1).
Following the strategy in the classical case (to prove the analytic class number
formula), to prove Theorem 26.5.4 we will write 𝜁O (𝑠) as a sum over right ideal
classes and analyze its residue at 𝑠 = 1 by a volume computation.
26.5.5. For an integral invertible right O-ideal 𝐽, let

∑︁ 1
𝜁O, [𝐽 ] (𝑠) := . (26.5.6)
N(𝐼) 𝑠
𝐼 ⊆O
[𝐼 ]=[𝐽 ]
Then ∑︁
𝜁O (𝑠) = 𝜁O, [𝐽 ] (𝑠).
[𝐽 ] ∈Cls O
We have [𝐼] = [𝐽] if and only if 𝐼 ' 𝐽 if and only if 𝐼 = 𝛼𝐽 for nonzero 𝛼 ∈ 𝐽 −1 .
Since 𝜇𝐽 = 𝐽 if and only if 𝜇 ∈ OL (𝐽) × (Exercise 16.3), it follows that
1 ∑︁ 1
𝜁O, [𝐽 ] (𝑠) = . (26.5.7)
N(𝐽) 𝑠 N(𝛼) 𝑠
0≠𝛼∈𝐽 −1 /OL ( 𝐽 ) ×
By Lemma 26.5.1, we have
𝑤 𝐽 := [OL (𝐽) × : 𝑅 × ] ∈ Z>0 . (26.5.8)
Then (26.5.7) becomes

1 ∑︁ 1
𝜁O, [𝐽 ] (𝑠) = . (26.5.9)
𝑤 𝐽 N(𝐽) 𝑠 N(𝛼) 𝑠
0≠𝛼∈𝐽 −1 /𝑅 ×
Proposition 26.5.10. Let 𝔑 := discrd(O). Then 𝜁O, [𝐽 ] (𝑠) has a simple pole at 𝑠 = 1
with residue
∗ 2𝑛 (2𝜋) 2𝑛 Reg𝐹
𝜁O, [𝐽 ] (1) = .
8𝑤 𝐽 𝑑 𝐹2 N(𝔑)
Proof. We relate residue to volumes using Theorem 26.2.12. We recall (again) 17.7.10:
this gives
𝐽 −1 ↩→ 𝐵 ↩→ 𝐵R := 𝐵 ⊗Q R ' H𝑛 ' R4𝑛
the structure of a Euclidean lattice Λ ⊆ R4𝑛 with respect to the absolute reduced norm.
We take the function 𝑁 in Theorem 26.2.12 to be the absolute norm N (recalling
16.4.8).
26.5. EICHLER MASS FORMULA: MAXIMAL ORDERS 431
We claim that
𝑑 𝐹2 N(𝔑)
covol(O) = . (26.5.11)
2𝑛
By compatible real scaling, it is enough to prove that this relation holds for a single
order O, and we choose the 𝑅-order
O = 𝑅 ⊕ 𝑅𝑖 ⊕ 𝑅 𝑗 ⊕ 𝑅𝑘. (26.5.12)
√ √ 4
The lattice 𝑅 ⊆ 𝐹R has covolume 𝑑 𝐹 , so 𝑅 4 has covolume 𝑑 𝐹 = 𝑑 𝐹2 ; the Z-order
Z ⊕ Z𝑖 ⊕ Z 𝑗 ⊕ Z𝑘 has reduced discriminant 4 and covolume 1; and putting these
together, the formula (26.5.11) is verified.
Then (26.5.11) and N(𝐽) = [O : 𝐽] Z = [𝐽 −1 : O] imply that
covol(O) 𝑑 𝐹2 N(𝔑)
covol(Λ) = = 𝑛 . (26.5.13)
N(𝐽) 2 N(𝐽)
Next, the group OL (𝐽) × acts on 𝐽 −1 (and on 𝐵R ); and this group contains 𝑅 × with
finite index 𝑤 𝐽 = [OL (𝐽) : 𝑅 × ], so
1
vol(OL (𝐽) × \𝐵R ) = vol(𝑅 × \𝐵R ). (26.5.14)
𝑤𝐽
Multiplication provides an identification
𝐵R, ≤1 ' 𝐹R, ≤1 × (H1 ) 𝑛 ,
so
𝑋 ≤1 = 𝑅 × \𝐵R, ≤1 ' (𝐸\𝐹R, ≤1 ) × ({±1}\(H1 ) 𝑛 ) (26.5.15)
where 𝐸 ≤ 𝑅 × is acting by squares. Thus
2𝑛−1 1
vol(𝑅 × \𝐹R, ≤1 ) = Reg𝐹 = Reg𝐹 . (26.5.16)
2(2𝑛 ) 4
Therefore
(2𝜋 2 ) 𝑛 Reg𝐹
vol(𝑋 ≤1 ) = . (26.5.17)
8𝑤 𝐽
From Theorem 26.2.12 together with (26.5.13) and (26.5.17),
∗ 4𝑛 (2𝜋 2 ) 𝑛 Reg𝐹 2𝑛 (2𝜋) 2𝑛 Reg𝐹

𝜁O, [𝐽 ] (1) = = . (26.5.18)
8𝑤 𝐽 𝑑 𝐹2 N(𝔑) 8𝑤 𝐽 𝑑 𝐹2 N(𝔑)
We now conclude the proof.
Proof of Theorem 26.5.4. We now suppose that O ⊂ 𝐵 is a maximal order, and write
𝜁 𝐵 (𝑠) and 𝜁 𝐵, [𝐽 ] (𝑠). We compare the evaluation of residues given by Corollary
26.4.11 and Proposition 26.5.10. Since 𝜁 𝐵 (𝑠) and each 𝜁 𝐵, [𝐽 ] (𝑠) have simple poles at
𝑠 = 1, we get ∑︁
𝜁 𝐵∗ (1) = ∗
𝜁 𝐵, [𝐽 ] (1).
[𝐽 ] ∈Cls O
From (26.4.12),
𝜁 𝐹∗ (1) Ö 𝜁 ∗ (1) 𝜑(𝔇)

1
𝜁 𝐵∗ (1) = 𝜁 𝐹 (2) 1− = 𝜁 𝐹 (2) 𝐹 . (26.5.19)
2 N(𝔭) 2 N(𝔇)
𝔭 |𝔇
From the analytic class number formula (Theorem 26.2.3),
2𝑛
𝜁 𝐹∗ (1) = √ ℎ 𝐹 Reg𝐹
2 𝑑𝐹
since 𝑤 𝐹 = 2 (as 𝐹 is totally real).
Adding the residues from Lemma 26.5.10, we find that
2𝑛 𝜁 𝐹 (2) 𝜑(𝔇) 2𝑛 (2𝜋) 2𝑛 Reg𝐹 ∑︁ 1

√ ℎ 𝐹 Reg𝐹 = . (26.5.20)
4 𝑑𝐹 N(𝔇) 8𝑑 𝐹2 N(𝔇) [𝐽 ] ∈Cls O
𝑤𝐽
Cancelling, we find
∑︁ 1 2
mass(Cls O) = = 𝜁 (2)𝑑 𝐹3/2 ℎ 𝐹 𝜑(𝔇)
2𝑛 𝐹
(26.5.21)
[𝐽 ] ∈Cls O
𝑤𝐽 (2𝜋)
and this concludes the proof.
Remark 26.5.22. For an alternative direct approach in this setting using Epstein zeta
functions, see Sands [San2017].
26.6 Eichler mass formula: general case

We now consider the general case of the Eichler mass formula, involving two steps.
First, we relate the class set of a suborder to the class set of a (maximal) superorder;
second, we compute the fibers of this map via a group action of the units.
For these steps, we refresh our notation and allow 𝐵 to be a definite or indefinite
26.6.1. Let O0 ⊇ O be an 𝑅-superorder, and suppose that there is a prime 𝔭 such that
O𝔮0 = O𝔮 for all primes 𝔮 ≠ 𝔭. We refine the map from Exercise 17.3(b) as follows.
For 𝐼 ⊆ O a right O-ideal, we define the right O0-ideal 𝜌(𝐼) = 𝐼O0 ⊆ O0 obtained by
extension. Then 𝜌 induces a map
Cls O → Cls O0
(26.6.2)
[𝐼] ↦→ [𝐼O0]
that is well-defined and surjective (Exercise 26.5(a)). Let [𝐼 0] ∈ Cls O0 and consider
the set
𝜌 −1 (𝐼 0) = {𝐼 ⊆ O : 𝐼O = 𝐼 0 },
the fiber of the extension map over 𝐼 0.
26.6. EICHLER MASS FORMULA: GENERAL CASE 433
We define an action of the group O𝔭0× on 𝜌 −1 (𝐼 0) as follows. Write 𝐼𝔭0 = 𝛽𝔭 O𝔭0 .

Then to 𝜇𝔭 ∈ O𝔭0× , we associate the unique lattice 𝐼 h𝜇𝔭 i (the notation to suggest “the
lattice generated by 𝜇𝔭 ”) such that
𝐼 h𝜇𝔭 i𝔭 = 𝛽𝔭 𝜇𝔭 O𝔭
and 𝐼 h𝜇𝔭 i𝔮 = 𝐼𝔮 = 𝐼𝔮0 for all 𝔮 ≠ 𝔭, using the local-global dictionary (Theorem 9.4.9).
This defines a right action of O𝔭0× ; it acts simply transitively on 𝜌 −1 (𝐼 0), and the kernel
of this action is visibly the subgroup O×𝔭 . Therefore
#𝜌 −1 (𝐼 0) = [O𝔭0× : O×𝔭 ].
We now look at the classes in the fiber. If 𝜇𝔭 , 𝜈𝔭 ∈ O𝔭0× have [𝐼 h𝜇𝔭 i] = [𝐼 h𝜈𝔭 i] ∈
Cls O, then there exists 𝛼 ∈ 𝐵× such that
𝛼𝐼 h𝜇𝔭 i = 𝐼 h𝜈𝔭 i
and by extension 𝛼𝐼 0 = 𝐼 0, so 𝛼 ∈ OL (𝐼 0), and conversely. Therefore, we have a

bĳection Ä
Cls O ↔ OL (𝐼 0) × \𝜌 −1 (𝐼 0). (26.6.3)
[𝐼 0 ] ∈Cls O0
(See also Pacetti–Sirolli [PS2014, §3].)
Proposition 26.6.4. Let O0 ⊇ O be an 𝑅-superorder, and suppose that there is a prime

𝔭 such that O𝔮0 = O𝔮 for all primes 𝔮 ≠ 𝔭. Then
mass(Cls O) = [O𝔭0× : O×𝔭 ] mass(Cls O0).
Proof. By (26.6.3), we conclude that

−1
∑︁ 1 ∑︁ ∑︁ 1 𝑤𝐼0
mass(Cls O) = =
𝑤𝐼 𝑤𝐼 𝑤𝐼
[𝐼 ] ∈Cls O [𝐼 0 ] ∈Cls O0 𝐼 O0 =𝐼 0
∑︁ 1 (26.6.5)
= [O𝔭0× : O×𝔭 ]
𝑤𝐼0
[𝐼 0 ] ∈Cls O0
= [O𝔭0× : O×𝔭 ] mass(Cls O0)
as claimed.
In order to apply Proposition 26.6.4, we need to compute the index of unit groups,
a quantity that depends on the (locally defined) Eichler symbol. For a prime 𝔭, we
define
 1 + 1/𝑞, if (O | 𝑝) = 1;
1 − Nm(𝔭) −2




𝜆(O, 𝔭) := = 1 − 1/𝑞, if (O | 𝑝) = −1; (26.6.6)
O
Nm(𝔭) −1 

1−  1 − 1/𝑞 2 , if (O | 𝑝) = 0.
𝔭 
Lemma 26.6.7. Let O0 ⊇ O be a containment of 𝑅-orders with O0 maximal. Then

(
1, if 𝔭 is split in 𝐵;
[O𝔭0× : O×𝔭 ] = [O𝔭0 : O𝔭 ]𝜆(O, 𝔭) ·
(1 − 1/𝑞) −1 , if 𝔭 is ramified in 𝐵.
Proof. We follow Körner [Kör85, §3]. To prove the lemma, we may localize at 𝔭 and
so we drop the subscripts. Let 𝑛 ∈ Z ≥1 be such that 𝔭𝑛 O0 ⊆ 𝔭O. Then
[O0× : 1 + 𝔭O0] [1 + 𝔭O0 : 1 + 𝔭𝑛 O0]

[O0× : O× ] = .
[O× : 1 + 𝔭O] [1 + 𝔭O : 1 + 𝔭𝑛 O0]
For 𝛾, 𝛿 ∈ 1 + 𝔭O, we have 𝛾𝛿−1 ∈ 1 + 𝔭𝑛 O0 if and only if 𝛾 − 𝛿 ∈ 𝔭𝑛 O0. Therefore
[1 + 𝔭O : 1 + 𝔭𝑛 O0] = [𝔭O : 𝔭𝑛 O0] = [O : 𝔭𝑛−1 O0]
and similarly with O0, all indices taken as abelian groups. Therefore
[1 + 𝔭O0 : 1 + 𝔭𝑛 O0]
= [O0 : O].
[1 + 𝔭O : 1 + 𝔭𝑛 O0]
For the other terms, we recall Lemma 24.3.12. We divide up into the cases, noting
that if(O0 | 𝔭) = −1 then we must have 𝜀 = −1, 0 by classification (Exercise 24.3);
this leaves 6 cases to compute. For example, if (O0 | 𝔭) = ∗ and (O0 | 𝔭) = 1, then
[O0× : 1 + 𝔭O0] 𝑞(𝑞 − 1) 2 (𝑞 + 1) 1

×
= 2 2
=1+ .
[O : 1 + 𝔭O] 𝑞 (𝑞 − 1) 𝑞
The other cases follow similarly (Exercise 26.8).
We can now finish the job.
Proof of Main Theorem 26.1.5. We first invoke Theorem 26.5.4 for a maximal order
O0 ⊇ O to get
2 Ö 1

0 3/2
mass(Cls O ) = ℎ 𝐹 𝑑 𝐹 Nm(𝔇) 1− .
(2𝜋) 2𝑛 Nm(𝔭)
𝔭 |𝔇
By Proposition 26.6.4 and Lemma 26.6.7, we have

Ö
mass(Cls O) = mass(Cls O0) [O𝔭0× : O×𝔭 ]
𝔭 |𝔑
Ö −1 Ö
1
= mass(Cls O0) [O0 : O] Z 1− 𝜆(O, 𝔭) (26.6.8)
Nm(𝔭)
𝔭 |𝔇 𝔭 |𝔑
2 Ö
= 2𝑛
ℎ 𝐹 𝑑 𝐹3/2 Nm(𝔑) 𝜆(O, 𝔭)
(2𝜋) 𝔭 |𝔑
using Nm(𝔑) = Nm(𝔇) [O0 : O] Z .

26.7. CLASS NUMBER ONE 435
26.7 Class number one

It is helpful to get a sense of the overall size of the mass, as follows.
26.7.1. Let 𝑚(𝔇, 𝔐) be the mass of an(y) Eichler order of level 𝔐. Then in analogy
with the Brauer–Siegel theorem,
3
log 𝑚(𝔇, 𝔐) ∼ log 𝑑 𝐹 + log ℎ 𝐹 + log Nm(𝔇𝔐) (26.7.2)
2
as 𝑑 𝐹 Nm(𝔇𝔐) → ∞ with the degree 𝑛 fixed: see Exercise 26.9. In particular, for
𝐹 = Q,
log 𝑚(𝐷, 𝑀) ∼ log(𝐷 𝑀).
Since 𝐹 is totally real, one typically expects ℎ 𝐹 to be small in comparison to
𝑑 𝐹 —but there is a family of real quadratic fields with small regulator first studied by
Chowla with log ℎ 𝐹 ∼ 12 log 𝑑 𝐹 , a result due to Montgomery–Weinberger [MW77].
To conclude this section, as in section 25.4 (over Q), the Eichler mass formula can
now be used to solve class number one problems for quaternion orders for definite
quaternion orders (over totally real fields). This effort was undertaken recently by
Kirschmer–Lorch [KL2016]: a complete list of definite orders of type number one is
given, and again because # Typ O ≤ # Cls O, the following theorem can be proven.
Theorem 26.7.3 (Kirschmer–Lorch). There are 4194 one-class genera of primitive,
positive definite ternary quadratic forms (equivalently, definite quaternion orders O
with # Typ O = 1, up to isomorphism): they occur over 30 possible base fields of
degrees up to 5.
There are exactly 154 isomorphism classes of definite quaternion orders O with
# Cls O = 1; of these, 144 are Gorenstein and 10 are non-Gorenstein.
Remark 26.7.4. Kirschmer–Lorch [KL2016] also enumerate two-class genera; a com-
plete list is available online [KLwww]. Cerri–Chaubert–Lezowski [CCL2013] also
consider totally definite Euclidean orders over totally real fields, giving the complete
list over Q and over quadratic fields: all of them are Euclidean under the reduced norm.
26.8 Functional equation and classification

To conclude this chapter, we discuss the functional equation and important applications
to the classification of quaternion algebras over number fields. This section serves as
preview of the material in chapter 29 (motivation for it!) but we frame results in the
same vein as the results of this chapter.
Following Riemann, we complete 𝜁 (𝑠) := ∞ −𝑠 to the function
Í
𝑛=1 𝑛
𝜉 (𝑠) := 𝜋 −𝑠/2 𝜁 (𝑠)Γ(𝑠/2),
where Γ(𝑠) is the complex Γ-function (Exercise 26.2). Riemann proved that 𝜉 (𝑠)
extends to a meromorphic function on C and satisfies the functional equation
𝜉 (1 − 𝑠) = 𝜉 (𝑠). (26.8.1)
(It is also common to multiply 𝜉 (𝑠) by 𝑠(1 − 𝑠) to cancel the poles at 𝑠 = 0, 1.) We
will prove this in section 29.1—the impatient reader is encouraged to flip ahead!
26.8.2. This result extends to Dedekind zeta functions (retaining the notation from
section 26.2). Define
ΓR (𝑠) := 𝜋 −𝑠/2 Γ(𝑠/2), ΓC (𝑠) := 2(2𝜋) −𝑠 Γ(𝑠). (26.8.3)
We then define the completed Dedekind zeta function to be
𝜉 𝐹 (𝑠) := | 𝑑 𝐹 | 𝑠/2 ΓR (𝑠) 𝑟 ΓC (𝑠) 𝑐 𝜁 𝐹 (𝑠). (26.8.4)
Then 𝜉 𝐹 (𝑠) satisfies the functional equation
𝜉 𝐹 (1 − 𝑠) = 𝜉 𝐹 (𝑠) (26.8.5)
for all 𝑠 ∈ C. We prove (26.8.5) as Corollary 29.10.3(a) using idelic methods; this
proof will also motivate the completion defined above. For now, we borrow from the
future. The functional equation gives 𝜁 𝐹 (𝑠) meromorphic continuation to C via
𝑠−1/2
| 𝑑𝐹 | Γ(𝑠/2) 𝑟 Γ(𝑠) 𝑐
𝜁 𝐹 (1 − 𝑠) = 𝜁 𝐹 (𝑠) . (26.8.6)
4𝑐 𝜋 𝑛 Γ((1 − 𝑠)/2) 𝑟 Γ(1 − 𝑠) 𝑐
26.8.7. Using the functional equation (26.8.6), we can rewrite (26.2.4) to obtain the
tidier expression
ℎ 𝐹 Reg𝐹
𝜁 𝐹∗ (0) = lim 𝑠−(𝑟 +𝑐−1) 𝜁 𝐹 (0) = ; (26.8.8)
𝑠→0 𝑤𝐹
in particular, 𝜁 𝐹 has a zero at 𝑠 = 0 of order 𝑟 + 𝑐 − 1, the rank of the unit group of 𝑅
by Dirichlet’s unit theorem.
In terms of the completed Dedekind zeta function, we find 𝜉 𝐹 (𝑠) has analytic
continuation to C r {0, 1} with simple poles at 𝑠 = 0, 1 and residues
2𝑟 +𝑐 ℎ 𝐹 Reg𝐹
𝜉 𝐹∗ (0) = 𝜉 𝐹∗ (1) = . (26.8.9)
𝑤𝐹
Example 26.8.10. When 𝐹 is an imaginary quadratic field (𝑟 = 0 and 𝑐 = 1) we have

Reg𝐹 = 1 and 𝜁 𝐹∗ (0) = 𝜁 𝐹 (0), so ℎ 𝐹 /𝑤 𝐹 = 𝜁 𝐹 (0), and in particular if | 𝑑 𝐹 | > 4 then
ℎ 𝐹 = 2𝜁 𝐹 (0).
We now turn to a quaternionic generalization.
26.8.11. The factorization of 𝜁 𝐵 (𝑠) in Theorem 26.4.9 implies a functional equation

for 𝜁 𝐵 (𝑠) via the functional equation for 𝜁 𝐹 (𝑠). This functional equation is simplest
to state for a completed zeta function. We recall that
Ö
𝜁 𝐵 (𝑠) = 𝜁 𝐹 (2𝑠)𝜁 𝐹 (2𝑠 − 1) 1 − Nm(𝔭) 1−2𝑠 ;
𝔭 |𝔇
26.8. FUNCTIONAL EQUATION AND CLASSIFICATION 437
just as the function 𝜁 𝐹 completes to 𝜉 𝐹 with a simple functional equation, we analo-

gously complete 𝜁 𝐵 to
Ö Ö
𝜉 𝐵 (𝑠) := 𝜉 𝐹 (2𝑠)𝜉 𝐹 (2𝑠 − 1) Nm(𝔭) 𝑠 1 − Nm(𝔭) 1−2𝑠 (2𝑠 − 1) (26.8.12)
𝔭 |𝔇 𝑣 ∈Ω
where Ω ⊆ Ram 𝐵 be the set of real, ramified places in 𝐵. The definition (26.8.12) is
motivated by the simplicity of the functional equation; see also Remark 26.8.15 below.
Written in a different way,
𝜉 𝐵 (𝑠) = (2𝜋) 𝑡 (| 𝑑 𝐹 | 4 Nm(𝔇) 2 ) 𝑠/2 Γ 𝐵 (𝑠)𝜁 𝐵 (𝑠) (26.8.13)
where 𝑡 is the number of split real places, so that # Pl(𝐹) = #Ω + 𝑡,
Γ 𝐵 (𝑠) := ΓR (2𝑠) 𝑟 ΓR (2𝑠 + 1) 𝑟 −𝑡 ΓR (2𝑠 − 1) 𝑡 ΓC (2𝑠) 𝑐 ΓC (2𝑠 − 1) 𝑐 , (26.8.14)
and we have used the formula
(2𝑠 − 1)ΓR (2𝑠 − 1) = 2(𝑠 − 1/2)𝜋 −(2𝑠−1)/2 Γ(𝑠 − 1/2)

= (2𝜋)𝜋 −(2𝑠+1)/2 Γ(𝑠 + 1/2)
= 2𝜋ΓR (2𝑠 + 1).
Remark 26.8.15. The completion factors (26.8.12) are not arbitrarily chosen; they have
a natural interpretation from an idelic perspective. Perhaps this serves as a motivation
for working idelically: namely, that it helps to nail down these kinds of quantities! For
more, see section 29.8.
Some properties can be read off easily from (26.8.12).
Proposition 26.8.16 (Analytic continuation, functional equation). Let 𝑚 = # Ram 𝐵.
(a) 𝜉 𝐵 (𝑠) has meromorphic continuation to C and is holomorphic in C r {0, 1/2, 1}
with simple poles at 𝑠 = 0, 1.
(b) 𝜉 𝐵 (𝑠) satisfies the functional equation
𝜉 𝐵 (1 − 𝑠) = (−1) 𝑚 𝜉 𝐵 (𝑠). (26.8.17)
(c) 𝜉 𝐵 (𝑠) has a pole of order 2 − 𝑚 at 𝑠 = 1/2; in particular, if 𝑚 ≥ 2 then 𝜉 𝐵 (𝑠)

is holomorphic at 𝑠 = 1/2.
Proof. Statement (a) follows from (26.8.12), recalling that 𝜉 𝐹 (𝑠) is holomorphic in
C r {0, 1} by 26.8.7 with simple poles at 𝑠 = 0, 1. Part (c) follows similarly from (b),
since the other factors in (26.8.12) have a simple zero at 𝑠 = 1/2.
To prove (b), we consider each term in the definition of (26.8.12). The functional
equation (26.8.5) for 𝜉 𝐹 (𝑠) with 𝑠 ← 1 − 𝑠 implies
𝜉 𝐹 (2(1 − 𝑠))𝜉 𝐹 (2(1 − 𝑠) − 1) = 𝜉 𝐹 (2 − 2𝑠)𝜉 𝐹 (1 − 2𝑠)

= 𝜉 𝐹 (1 − (2 − 2𝑠))𝜉 𝐹 (1 − (1 − 2𝑠)) (26.8.18)
= 𝜉 𝐹 (2𝑠 − 1)𝜉 𝐹 (2𝑠).
For
ℓ(𝑠) = 𝑞 𝑠 (1 − 𝑞 1−2𝑠 ) = 𝑞 𝑠 − 𝑞 1−𝑠
and 𝑞 > 0 we have ℓ(1 − 𝑠) = −ℓ(𝑠), so with 𝑞 = Nm(𝔭) the factors 𝔭 | 𝔇 are taken
into account. Finally, 2(1 − 𝑠) − 1 = −(2𝑠 − 1) takes care of 𝑣 ∈ Ω, and (b) follows.
Proposition 26.8.16 shows how algebraic properties of 𝐵 correspond to analytic

properties of 𝜉 𝐵 . A deeper investigation ultimately will reveal the following funda-
mental result.
Theorem 26.8.19 (Sign of functional equation, holomorphicity). 𝜉 𝐵 (𝑠) satisfies the

functional equation
𝜉 𝐵 (1 − 𝑠) = 𝜉 𝐵 (𝑠). (26.8.20)
Moreover, if 𝐵 is a division algebra, then 𝜉 𝐵 (𝑠) is holomorphic at 𝑠 = 1/2.
Proof. This theorem was proven by Hey [Hey29, §3] (more generally, for division
algebras over Q) following the same general script as in the proof of the functional
equation for the Dedekind zeta function (26.8.5), as proven first by Hecke: the key
ingredient is Poisson summation. The argument is also given by Eichler [Eic38a, Part
V]. We instead prove this theorem in the language of ideles (Main Theorem 29.2.6),
as it simplifies the calculations—and so for continuity of ideas in the exposition, we
borrow from the future.
Assuming Theorem 26.8.19, we can now deduce the main classification theorem
(Main Theorem 14.6.1) for quaternion algebras over number fields. First, we have
Hilbert reciprocity as an immediate consequence.
Corollary 26.8.21 (Hilbert reciprocity, cf. Corollary 14.6.2). # Ram 𝐵 is even.
Proof. Immediate from (26.8.17) and (26.8.20).
Next we conclude the all-important local-global principle.
Corollary 26.8.22. We have 𝐵 ' M2 (𝐹) if and only if 𝐵 𝑣 ' M2 (𝐹𝑣 ) for all (but one)
places 𝑣 ∈ Pl 𝐹.
Proof. The implication (⇒) is immediate. For the converse (⇐), by Proposition
26.8.16(c), 𝜉 𝐵 (𝑠) has a pole of order 2 − 𝑚 at 𝑠 = 1/2, so if 𝑚 ≤ 1 then 𝜉 𝐵 (𝑠) is not
holomorphic at 𝑠 = 1/2; but then by Theorem 26.8.19, 𝐵 is not a division algebra, so
𝐵 ' M2 (𝐹) (and the order of pole is necessarily 2, and 𝐵 𝑣 ' M2 (𝐹𝑣 ) for all 𝑣).
From this corollary, we are able to deduce the Hasse norm theorem for quadratic
extensions.
Theorem 26.8.23 (Hasse norm theorem). Let 𝐾 ⊇ 𝐹 be a separable quadratic field

extension and let 𝑏 ∈ 𝐹 × . Then 𝑏 ∈ Nm𝐾 /𝐹 (𝐾 × ) if and only if 𝑏 ∈ Nm𝐾𝑣 /𝐹𝑣 (𝐾 𝑣× ) for
all (but one) places 𝑣 ∈ Pl 𝐹.
Proof. Consider the quaternion algebra 𝐵 = (𝐾, 𝑏 | 𝐹). Then by Main Theorem 5.4.4,
we have 𝑏 ∈ Nm𝐾 |𝐹 (𝐾 × ) if and only if 𝐵 ' M2 (𝐹). By Corollary 26.8.22, this holds
if and only if 𝐵 𝑣 ' M2 (𝐹𝑣 ) for all (but one) places 𝑣. Repeating the application of
Main Theorem 5.4.4, this holds if and only if 𝐵 𝑣 ' M2 (𝐹𝑣 ) for all (but one) 𝑣.
We may similarly conclude the all important local-global principle for quadratic
forms.
Theorem 26.8.24 (Hasse–Minkowski theorem). Let 𝑄 be a quadratic form over 𝐹.

Then 𝑄 is isotropic over 𝐹 if and only if 𝑄 𝑣 is isotropic over 𝐹𝑣 for all places 𝑣 of 𝐹.
Proof. The implication (⇒) is immediate, so we prove (⇐). We may suppose without
loss of generality that 𝑄 is nondegenerate. If 𝑛 = dim𝐹 𝑉 = 1, the theorem is vacuous.
Suppose 𝑛 = 2. Then after scaling we may suppose 𝑄 = h1, −𝑎i, and 𝑄 is isotropic√
if and only if 𝑎 is a square. Suppose for purposes of contradiction that 𝐾 = 𝐹 ( 𝑎)
is a field. Since 𝑄 𝑣 is isotropic for all 𝑣, we have 𝐾 𝑣 ' 𝐹𝑣 × 𝐹𝑣 for all 𝑣, and thus
𝜁 𝐾 (𝑠) = 𝜁 𝐹 (𝑠) 2 . But as Dedekind zeta functions, both 𝜁 𝐹 (𝑠) and 𝜁 𝐾 (𝑠) have poles
of order 1 at 𝑠 = 1 (we evaluated the residue in the analytic class number formula,
Theorem 26.2.3), a contradiction.
Suppose 𝑛 = 3. Again after rescaling we√may suppose 𝑄 = h1, −𝑎, −𝑏i, and 𝑄 is
isotropic if and only if 𝑏 is a norm from 𝐹 [ 𝑎]: then the equivalence follows from
Theorem 26.8.23. √
Next, suppose 𝑛 = 4, and 𝑄 = h1, −𝑎, −𝑏, 𝑐i. Let 𝐾 = 𝐹 ( 𝑎𝑏𝑐). By extension,
𝑄 is isotropic over 𝐾 and all of its completions. But now 𝑄 ' h1, −𝑎, −𝑏, 𝑎𝑏i over
𝐾. Let 𝐵 = (𝑎, 𝑏 | 𝐾). Then by Main Theorem 5.4.4, we have 𝐵 𝑤 ' M2 (𝐾 𝑤 ) for all
𝑤; thus by Corollary 26.8.22 we have 𝐵 ' M2 (𝐾), so 𝐾 splits 𝐵. By 5.4.7, we have
𝐾 ↩→ 𝐵, so there exist 𝑥, 𝑦, 𝑧 ∈ 𝐹 such that
nrd(𝛼) = nrd(𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗) = −𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧 2 = −𝑎𝑏𝑐;
dividing by 𝑎𝑏 we have 𝑧 2 − 𝑎(𝑦/𝑎) 2 − 𝑏(𝑥/𝑏) 2 + 𝑐 = 0 so 𝑄(𝑧, 𝑦/𝑎, 𝑥/𝑏, 1) = 0.

Finally, when 𝑛 ≥ 5, we make an argument like at the end of proof of Theorem
14.3.3: we follow Lam [Lam2005, Theorem VI.3.8] and Milne [Milne-CFT, Theorem
VIII.3.5(b)], but we are brief. Write 𝑄 = 𝑄 1 ⊥ 𝑄 2 where 𝑄 1 = h𝑎, 𝑏i and dim𝐹 𝑉2 ≥
3. Choosing a ternary subform and looking at its quaternion algebra, we find a finite
set 𝑇 ⊆ Pl 𝐹 such that 𝑄 2 is isotropic for all 𝑣 ∉ 𝑇. For each 𝑣 ∈ 𝑇, let 𝑄(𝑧 𝑣 ) = 0
and let 𝑐 𝑣 = 𝑄 1 (𝑧 𝑣 ) = −𝑄 2 (𝑧 𝑣 ). Choose 𝑥, 𝑦 ∈ 𝐹 × close enough so that 𝑧 = 𝑄 1 (𝑥)
has 𝑧𝑧 𝑣 ∈ 𝐹𝑣×2 . The form 𝑄 0 = h𝑐i ⊥ 𝑄 2 in 𝑛 − 1 variables is isotropic for all 𝑣: for
𝑣 ∈ 𝑇 this was arranged, and for 𝑣 ∉ 𝑇 already 𝑄 2 was isotropic at 𝑣. By induction on
𝑛, we conclude that 𝑄 0 is isotropic; diagonalizing, we may write 𝑄 = h𝑑i ⊥ 𝑄 0, and it
follows that 𝑄 is isotropic.
Corollary 26.8.25. Let 𝑄, 𝑄 0 be quadratic forms over 𝐹 in the same number of

variables. Then 𝑄 ' 𝑄 0 if and only if 𝑄 𝑣 ' 𝑄 0𝑣 for all places 𝑣 ∈ Pl 𝐹.
Proof. Apply the same method of proof as in Corollary 14.3.7: see Exercise 26.10.
We may now conclude the classification with one further input.

Theorem 26.8.26 (Infinitude of primes in arithmetic progression over number fields).

Let 𝔫 ⊆ Z𝐹 be a nonzero ideal, let 𝑎 ∈ (Z𝐹 /𝔫) × , and for each 𝑣 | ∞ real let 𝜖 𝑣 ∈ {±1}.
Then there are infinitely many prime elements 𝑝 ∈ Z𝐹 such that sgn(𝑣( 𝑝)) = 𝜖 𝑣 and
𝑝 ≡ 𝑎 (mod 𝔫).
Proof. The theorem generalizes Dirichlet’s theorem on the infinitude of primes in

arithmetic progression (Theorem 14.2.9): see Lang [Lang94, Theorem VIII.4.10] or
Neukirch [Neu99, Theorem VII.13.4].
Proof of Main Theorem 14.6.1, 𝐹 a number field. First, the map 𝐵 ↦→ Ram 𝐵 has the
correct codomain by Hilbert reciprocity (Corollary 26.8.21). Surjectivity follows
by Exercise 14.16 (using Theorem 26.8.26. To conclude, we show injectivity. We
refer to Corollary 5.2.6, giving a bĳection between quaternion algebras over 𝐹 up to
isomorphism and ternary quadratic forms of discriminant 1 up to isometry; and we
recall Theorem 12.3.4, that (rescaling) there is a unique anisotropic ternary quadratic
form of discriminant 1 up to isometry. Therefore Corollary 26.8.25 implies that the
map 𝐵 ↦→ Ram 𝐵 is injective, since the set Ram 𝐵 records those places 𝑣 where the
ternary quadratic form attached to 𝐵 is anisotropic.
We will give another proof of Main Theorem 14.6.1 over global fields using
the characterization of idelic norms in Proposition 27.5.15 (avoiding fiddling with
quadratic forms and the use of primes in arithmetic progression).
Remark 26.8.27. For the readers who accept the fundamental exact sequence of class
field theory as in Remark 14.6.10, the arguments above can be run in reverse, and the
analytic statement in Theorem 26.8.19 can be deduced as a consequence.
Exercises
1. Prove Proposition 26.2.18 that
vol(𝑋 ≤1 ) =
𝑤𝐹
in the special case of a real quadratic field.

I 2. Let ∫ ∞
d𝑥
Γ(𝑠) := 𝑥 𝑠 𝑒 −𝑥
0 𝑥
be the complex Γ-function, defined for Re 𝑠 > 0. Verify the following basic
properties of Γ(𝑠).
√
(a) Γ(1) = 1 and Γ(1/2) = 𝜋.
(b) Γ(𝑠 + 1) = 𝑠Γ(𝑠) for all Re 𝑠 > 0, and Γ(𝑛) = (𝑛 − 1)! for 𝑛 ≥ 1.
(c) Γ(𝑠) has meromorphic continuation to C, holomorphic away from simple
poles at Z ≤0 .
(d) Γ(𝑠) has no zeros in C.
I 3. In this exercise, we prove Proposition 26.2.18. Let 𝐹 be a number field with

ring of integers 𝑅, let 𝑋 ⊆ 𝐹R be a cone fundamental domain for 𝑅 × . Let Reg𝐹
be the regulator of 𝐹 and 𝑤 𝐹 the number of roots of unity in 𝐹.
(a) Let 𝑉 = 𝐹R be the ambient space. Let 𝜇(𝑅 × ) ≤ 𝑅 × be the group generated
by a fundamental system of units, so 𝑅 × /𝜇(𝑅 × ) ' 𝑅tors
× . Show that
∫
2𝑐
vol((𝑉/𝑅 × ) ≤1 ) = d𝑥 d𝑧
𝑤 𝐹 𝑉≤1 /𝜇 (𝑅× )
with 𝑥 𝑖 , 𝑧 𝑗 standard coordinates on R𝑟 × C𝑐 in multi-index notation.
(b) Let 𝜌 𝑗 , 𝜃 𝑗 be polar coordinates on C𝑐 , and restrict the domain 𝑉 to the
domain 𝑉 + with 𝑥𝑖 > 0 for all 𝑖. Let 𝑊 + be the projection of 𝑉 + onto the
𝑥, 𝜌-coordinate plane and let 𝑥𝑟 + 𝑗 = 𝜌 2𝑗 . Show that
∫ ∫
𝑟 𝑐
d𝑥 d𝑧 = 2 𝜋 d𝑥.
𝑉≤1 /𝜇 (𝑅 × ) 𝑊 +,≤1 /𝜇 (𝑅 × )
(c) Apply the change of variables 𝑢 𝑖 = log 𝑥𝑖 to obtain

∫ ∫
d𝑥 = d𝑢
+ /𝜇 (𝑅 × )
𝑊≤1 𝑃
where 𝑃 is the fundamental parallelogram for the additive (logarithmic)

action of 𝑅 × . Conclude that
vol(𝑋 ≤1 ) = .
𝑤𝐹
I 4. Show that the ideal 𝐽 in Lemma 26.3.6 is unique: more specifically, show that if
𝐼 is an invertible, integral lattice and suppose that nrd(𝐼) = 𝔪𝔫 with 𝔪, 𝔫 ⊆ 𝑅
coprime ideals, then an invertible, integral lattice 𝐽 such that 𝐼 is compatible
with 𝐽 −1 with 𝐼𝐽 −1 integral and nrd(𝐽) = 𝔪 is unique.
I 5. Let 𝐹 be a number field with ring of integers 𝑅, let 𝐵 be a quaternion algebra
over 𝐹, and let O ⊆ O0 ⊆ 𝐵 be 𝑅-orders. For 𝐼 ⊆ O a right O-ideal, we define
the right O0-ideal 𝜌(𝐼) = 𝐼O0 ⊆ O0 obtained by extension.
(a) Show that 𝜌 induces a (well-defined) surjective map
Cls O → Cls O0
[𝐼] ↦→ [𝐼O0]
of pointed sets with finite fibers.
(b) For the case where O is the Lipschitz order and O0 the Hurwitz order, show
that the map in (a) is a bĳection (cf. Lemma 11.2.9).
I 6. Let 𝑅 be a DVR with uniformizer 𝜋 and let 𝐼 be a (invertible) integral right
M2 (𝑅)-ideal. Show that 𝐼 is generated by
𝑢
𝜋 0
𝑥=
𝑐 𝜋𝑣
where 𝑢, 𝑣 ∈ Z ≥0 and 𝑐 ∈ 𝑅/𝜋 𝑣 are unique.
7. Generalize Exercise 11.14 as follows. For 𝑛 ∈ Z, let
𝑟 4 (𝑛) := #{(𝑡, 𝑥, 𝑦, 𝑧) ∈ Z4 : 𝑡 2 + 𝑥 2 + 𝑦 2 + 𝑧2 = 𝑛}
and let 𝑟 40 (𝑛) := 𝑟 4 (𝑛)/8.

(a) Show that 𝑟 40 (2𝑒 ) = 1 for all 𝑒 ≥ 1 and 𝑟 40 ( 𝑝 𝑒 ) = 1 + 𝑝 + · · · + 𝑝 𝑒 for all
𝑒 ≥ 1 and 𝑝 odd. [Hint: relate the count to the number of right ideals and
inspect the coefficients of the zeta function.]
(b) Show that 𝑟 40 is multiplicative: 𝑟 40 (𝑚𝑛) = 𝑟 40 (𝑚)𝑟 40 (𝑛) if gcd(𝑚, 𝑛) = 1.
(c) Conclude that
( Í
8 𝑑 |𝑛 𝑑, 𝑛 odd;
𝑟 4 (𝑛) = Í
24 𝑑 |𝑚 𝑑, 𝑛 = 2𝑒 𝑚 even with 𝑚 odd.
I 8. Finish the proof of Lemma 26.6.7 by checking the remaining cases.

9. Prove (26.7.2). Specifically, for a number field 𝐹 and coprime ideals 𝔇, 𝔐 with
𝔇 squarefree and coprime to 𝔐, define the mass
2𝜁 𝐹 (2) 3/2
𝑚(𝐹, 𝔇, 𝔐) := 𝑑 ℎ 𝐹 𝜑(𝔇)𝜓(𝔐).
(2𝜋) 2𝑛 𝐹
Let 𝔑 := 𝔇𝔐. Show for fixed 𝑛 that
3
log 𝑚(𝐹, 𝔇, 𝔐) ∼ log 𝑑 𝐹 + log ℎ 𝐹 + log Nm(𝔑) (26.8.28)
2
as 𝑑 𝐹 Nm(𝔑) → ∞, as follows.
(a) Show that
Ö −𝑛 Ö −1
1 1
𝜁Q (2) 𝑛 = 1− ≤ 𝜁 𝐹 (2) ≤ 1− = 𝜁Q (2𝑛)
𝑝
𝑝2 𝑝
𝑝 2𝑛
so 𝜁 𝐹 (2) 1.
(b) Show that
Nm(𝔑)
𝜑(𝔇)𝜓(𝔐) Nm(𝔑) (log log Nm(𝔐)) .
log log Nm(𝔇)
[Hint: you may need some elementary estimates from analytic number
theory, adapted for this purpose; you may wish to start with the case
𝐹 = Q.]
(c) Conclude (26.8.28).
10. Prove Corollary 26.8.25: if 𝑄, 𝑄 0 are quadratic forms over 𝐹 in the same number
of variables, then 𝑄 ' 𝑄 0 if and only if 𝑄 𝑣 ' 𝑄 0𝑣 for all places 𝑣 ∈ Pl 𝐹. [Hint:
see Corollary 14.3.7.]
I 11. Use Dirichlet’s analytic class number formula to prove the theorem on arithmetic
progressions (Theorem 14.2.9) as follows.
(a) Let 𝐹 = Q(𝜁 𝑚 ). Show that

Ö
𝜁 𝐹 (𝑠) = 𝜁 (𝑠) 𝐿 (𝑠, 𝜒)
𝜒≠1
where 𝜒 runs over all nontrivial Dirichlet characters 𝜒 : (Z/𝑚Z) × → C× ,

and
∞
∑︁ 𝜒(𝑛)
𝐿 (𝑠, 𝜒) .
𝑛=1
𝑛𝑠
[Hint: Factor according to the decomposition 𝑚 = 𝑟𝑒 𝑓 .]

(b) Use partial summation and the fact that the partial sums are bounded to
show that each 𝐿(𝑠, 𝜒) for 𝜒 ≠ 1 is holomorphic at 𝑠 = 1.
(c) Conclude from the analytic class number formula that 𝐿 (1, 𝜒) ≠ 0 for
𝜒 ≠ 1.
(d) For gcd(𝑎, 𝑚) = 1, using (c) show that as 𝑠 & 1 that
∑︁ log 𝜁 (𝑠)
𝑝 −𝑠 = + 𝑂 (1)
𝜑(𝑚)
𝑝≡𝑎 (mod 𝑚)
and conclude that the set of primes 𝑝 with 𝑝 ≡ 𝑎 (mod 𝑚) is infinite.

I 12. Let 𝐹 be a nonarchimedean local field with valuation ring 𝑅 having maximal
ideal 𝔭 and residue field 𝑘 of size 𝑞 := #𝑘. For 𝑛 ≥ 1, let 𝐵𝑛 := M𝑛 (𝐹) and
O𝑛 := M𝑛 (𝑅). In this exercise we generalize Lemma 26.4.1(b).
(a) Show that the set of right integral O𝑛 -ideals is in bĳection with the set


 𝜋 𝑢1 0 0 ··· 0 


𝜋 𝑢2 0 ··· 0
 © ª 


 𝑐 21 ® 


𝜋 𝑢3 ··· 0

 ® 

𝑐 31 𝑐 32 ® : 𝑢 1 , . . . , 𝑢 𝑛 ∈ Z ≥0 , 𝑐 𝑖 𝑗 ∈ 𝑅/𝔭𝑖 .
 . .. .. .. ® 
.
0 ®
 . . ® 


 . . 


··· 𝑢
 
 𝑐 𝑛1
« 𝑐 𝑛2 𝑐 𝑛3 𝜋 𝑛¬ 

[Hint: appeal to the theory of elementary divisors, applying column oper-
ations acting on the right.]
(b) Let 𝑎 𝔭𝑒 (O𝑛 ) be the number of right ideals of reduced norm 𝔭𝑒 in O. Show
that
𝑒
∑︁
𝑎 𝔭𝑒 (O𝑛 ) = 𝑎 𝔭 𝑓 (O𝑛−1 )𝑞 (𝑛−1) (𝑒− 𝑓 ) .
𝑓 =0
(c) Let
∑︁ 1
𝜁O𝑛 (𝑠) := ,
N(𝐼)
𝐼 ⊆O𝑛
the sum over nonzero right ideals of O𝑛 . Show that 𝜁O1 (𝑠) = 𝜁 𝐹 (𝑠) =
(1 − 𝑞 −𝑠 ) −1 .
(d) Show for 𝑛 ≥ 2 that

∞
∑︁ 𝑎 𝔭𝑒 (O𝑛 )
𝜁O𝑛 (𝑠) = = 𝜁O𝑛−1 (𝑛𝑠/(𝑛 − 1))𝜁O1 (𝑛𝑠 − (𝑛 − 1)).
𝑒=0
𝑞 𝑛𝑒𝑠
(e) Conclude that

𝑛−1
Ö
𝜁O𝑛 (𝑠) = 𝜁 𝐹 (𝑛𝑠)𝜁 𝐹 (𝑛𝑠 − 1) · · · 𝜁 𝐹 (𝑛𝑠 − (𝑛 − 1)) = 𝜁 𝐹 (𝑛𝑠 − 𝑖)
𝑖=0
(cf. Corollary 26.4.7).

Chapter 27
Adelic framework
We have already seen that the local-global dictionary is a powerful tool in understanding
the arithmetic of quaternion algebras. In this section, we formalize this connection by
consideration of adeles and ideles.
The basic idea: we want to consider all of the completions of a global field at once.
There are at least two benefits to this approach:
• We will gain notational efficiency, resulting in brief and well-behaved proofs

that would be difficult or impossible to state clearly in classical language.
• Each completion is a locally compact field and so amenable to harmonic analysis,
and by extension to the adele ring and its group of units, we can do harmonic
analysis on global objects.
The adelic framework, and its use in class field theory, is a vast topic whose complete
development deserves its own book. We do our best in this chapter to develop this
notation and state what is needed for the case of quaternion algebras. For further
background reading, see Childress [Chi2009] and the references given at the start of
section 27.4
27.1 ⊲ The rational adele ring

In this first section, we work purely over Q to give a concrete flavor to the abstract
definitions to come.
27.1.1. Recall in section 12.1 that for a prime 𝑝 we defined Z 𝑝 = lim Z/𝑝 𝑟 Z as a
←−−𝑟
projective limit, and each Z 𝑝 is compact. We can package these together to make the
direct product ring
Ö
Z :=
b Z𝑝 (27.1.2)
𝑝
equipped with the product topology: as a profinite group, it is Hausdorff, compact,

and totally disconnected.
445
446 CHAPTER 27. ADELIC FRAMEWORK
We can see bZ itself as projective limit as follows. By the Sun Zi theorem (CRT),
we have an isomorphism
∼ lim Z/𝑛Z
Ö
Z=
b lim Z/𝑝 𝑟 Z −→
←−− ←−−
𝑝 𝑟 𝑛
of topological rings, with the projective limit indexed by positive integers partially
ordered under divisibility; so under this isomorphism, we may identify
Z = lim Z/𝑛Z
b
← −−
𝑛 (27.1.3)
∞
∈ ∞
Î
= (𝑎 𝑛 ) 𝑛=1 𝑛=1 Z/𝑛Z : 𝑎 𝑚 ≡ 𝑎 𝑛 (mod 𝑛) for all 𝑛 | 𝑚
The natural ring homomorphism Z → b Z which takes every element to its reduction
modulo 𝑛 is injective; the image of Z is discrete and dense in b
Z again by the CRT. One
warning is due: b
Z is not a domain.
27.1.4. We now make the ring b Z a bit bigger so that it contains Q as a subring. If
Î
we were to take the ring 𝑝 Q 𝑝 , a product of locally compact rings, unfortunately
we wouldÎno longer have something that is locally compact (see Exercise 27.1): the
product 𝑝 Q 𝑝 is much too big, allowing denominators in every component, whereas
the image of Q will only have denominators in finitely many positions. We should also
keep track of archimedean information at the same time.
With these in mind we define, for each finite set S of primes, the ring
Ö Ö
𝑈S := R × Q𝑝 × Z𝑝 (27.1.5)
𝑝 ∈S 𝑝∉S
equipped with the product topology, so that 𝑈S is locally compact. For example,
𝑈∅ = R × b
Z. (27.1.6)
To assemble these rings together, allowing more denominators and arbitrarily large
sets 𝑆, we take the injective limit of 𝑈S under the natural directed system 𝑈S ↩→ 𝑈S0
for S ⊆ S 0. The resulting object is the restricted direct product of Q 𝑝 relative to Z 𝑝
and is called the adele ring Q of Q:
Ö0 Ö0
Q := R × Q𝑝 = Q𝑣
𝑝 𝑣 ∈Pl Q
Î (27.1.7)
:= R × 𝑥 = (𝑥 𝑝 ) 𝑝 ∈ 𝑝 Q 𝑝 : 𝑥 𝑝 ∈ Z 𝑝 for all but finitely many 𝑝
Î
= 𝑥 = (𝑥 𝑣 ) 𝑣 ∈ 𝑣 Q 𝑣 : |𝑥 𝑣 | 𝑣 ≤ 1 for all but finitely many 𝑣
We declare the sets 𝑈S ⊆ Q with the product topology to be open in Q; and with
this basis of open neighborhoods of 0 (open in 𝑈S for some S), we have given Q the
structureÎof a topological ring. The sets 𝑈S ⊆ Q are also closed. Note that the topology
on Q ⊂ 𝑣 Q 𝑣 is not the subspace topology.
27.1. ⊲ THE RATIONAL ADELE RING 447
27.1.8. For a finite set S, we write

Ö0
Q6S := Q𝑣 (27.1.9)
𝑣∉S
for the projection of Q onto the factors away from S. We also write
Ö
QS := Q𝑣 . (27.1.10)
𝑣 ∈𝑆
We embed each of these into Q extending by zero and identify them with their images,
so that Q = Q6S × QS .
Remark 27.1.11. Our notation Q for the adele ring returns to the notation of Weil
[Weil82] but is not standard; more typically, the adele ring is denoted A (which we
find markedly problematic).
Remark 27.1.12. Although Q6S and QS are rings and (via projection) are naturally
quotient rings of Q, they are not subrings because they do not contain 1. This subtlety
should cause no confusion in what follows (especially because we will be focused on
the multiplicative case and working with groups, where there is no issue extending by
the multiplicative identity 1).
We have a natural continuous embedding Q ↩→ Q 𝑣 for all 𝑣 ∈ Pl Q, and this
extends to a diagonal embedding Q ↩→ Q.
Lemma 27.1.13. The diagonal embedding Q ↩→ Q is a continuous injective ring
homomorphism and the image is closed and discrete as a subring of Q.
Proof. The embedding is continuous because it is so in each component. Because Q

is a topological group under addition, to prove the remaining part it is enough to find
a neighborhood 0 ∈ 𝑈 ⊆ Q such that 𝑈 ∩ Q = {0}. We take

𝑈 := (−1, 1) × b
Z = (𝑥 𝑣 ) 𝑣 : |𝑥∞ | ∞ < 1 and |𝑥 𝑝 | 𝑝 ≤ 1 for all primes 𝑝 .
By definition, 𝑈 is open in Q as it is open in 𝑈 ∅ (for a reminder, see (27.1.6)). And if

𝑎 ∈ 𝑈 ∩ Q, then 𝑎 ∈ Z 𝑝 for all 𝑝, so 𝑎 ∈ Z, and |𝑎| ∞ < 1, and thus 𝑎 = 0.
Lemma 27.1.14. The image of Q ↩→ Q is cocompact, i.e., Q/Q is compact.
Proof. Let 𝑊 := [0, 1] × b Z. Then 𝑊 is compact. By strong approximation—for a

snapshot review, flip ahead to Theorem 28.1.9 and its corollary—we have Q = Q + 𝑊.
Therefore the continuous quotient map Q → Q/Q restricted to 𝑊 is surjective, so Q/Q
equal to the image of the compact set 𝑊 is compact. (𝑊 is a fundamental set for the
action of Q on Q: see Exercise 27.2.)
27.1.15. The proof of Lemma 27.1.14 shows that the natural map R × b Z → Q/Q is
surjective; its kernel is Z diagonally embedded, so we have an isomorphism
∼ (R × b
Q/Q −→ Z)/Z
of topological groups. The resulting topological group Sol := (R × b Z)/Z is called a

solenoid: it is compact, Hausdorff, connected, but not path-connected (Exercise 27.5),
which can be visualized as in Figure 27.1.16.
···
···
···
···
···
Figure 27.1.16: The solenoid
Very often, we will want to tease apart the nonarchimedean and archimedean parts
of the adele ring Q, and will write
Ö0
b := Q {∞} =
Q Q𝑝 ' b
Z ⊗Z Q (27.1.17)

𝑝
b × R.
so that extending by zero we have Q = Q
27.2 ⊲ The rational idele group

Having dealt with the additive version in adeles, now we talk about the multiplicative
version, ideles.
27.2.1. We define the rational idele group to be
Ö0
Q× = R× × Q×𝑝 = R× × Qb×
𝑝 (27.2.2)
×
Î
= 𝑥 = (𝑥 𝑣 ) 𝑣 ∈ 𝑣 Q𝑣 : |𝑥 𝑣 | = 1 for all but finitely many 𝑣
That is to say, Q× is the restricted direct product of the spaces Q×𝑝 with respect to Z×𝑝 .
The topology is such that for S a finite set of primes, the set
Ö Ö
𝑉𝑆 = R× × Q×𝑝 × Z×𝑝
𝑝 ∈S 𝑝∉S
is open and closed as a subgroup of Q× .
Remark 27.2.3. Chevalley first used the name élément idéal for elements of Q× , but at
Hasse’s suggestion he abbreviated it to idèle; the name adèle was then shorthand for
an “additive idele”. Anglifying, we drop the accents on these words.
27.2. ⊲ THE RATIONAL IDELE GROUP 449
27.2.4. The topology on Q× is not the subspace topology Q× ⊂ Q, because inversion

need not be continuous. Instead, we think of Q× as a subset of Q × Q via the map
𝑥 ↦→ (𝑥, 𝑥 −1 ), and then Q× inherits the structure of a topological group.
Lemma 27.2.5. The diagonal map Q× ↩→ Q× is an injective continuous group homo-
morphism, and the image of Q× is closed and discrete.
Proof. Since Q is closed and discrete in Q and Q× ⊆ Q × Q has the subspace topology,
so too is Q× closed and discrete.
27.2.6. We now give an explicit description of the quotient Q× /Q× : we will see it is
not compact.
There is a canonical isomorphism of topological groups
Q×𝑝 ' h𝑝i × Z×𝑝
by 𝑝-adic valuation. Since h𝑝i = 𝑝 Z ' Z, we have a topological group isomorphism

Ö0 Ö Ê
Q× = R× × Q×𝑝 ' {±1} × R>0 × Z×𝑝 × Z. (27.2.7)
𝑝 𝑝 𝑝
A direct sum appears because an element of the restricted direct product is a 𝑝-adic
unit for all but finitely many 𝑝. We project Q× onto the product of the first and last
factor, getting a continuous surjective map
Ê
Q× → {±1} × Z. (27.2.8)
𝑝
Looking at 𝑟 ∈ Q× ⊆ Q× , if we write 𝑟 = 𝜖 𝑝 𝑝 𝑛( 𝑝) , where 𝜖 ∈ {±1} and

Î
𝑛( 𝑝) = ord 𝑝 (𝑟), then 𝑟 ↦→ ×

É(𝜖, (𝑛( 𝑝))× 𝑝 ) in the projection. Therefore Q is canonically
identified with {±1} × 𝑝 Z in Q . So the projection map (27.2.8) restricts to an
isomorphism on the diagonally embedded Q× . Therefore
Ö
Q× ' Q× × R>0 × Z×𝑝 . (27.2.9)
𝑝
By the logarithm map, there is an isomorphism R>0 ' R, so

Q× ' Q× × R × b
Z× (27.2.10)
and we have an isomorphism of topological groups
Q× /Q× ' R × b
Z× . (27.2.11)
(This is not a solenoid!)
b×/Q× ' b
In a similar way, we see that Q Z× , where Q×>0 = {𝑥 ∈ Q : 𝑥 > 0}, and so
>0
b×/Q× is compact.
Q
Remark 27.2.12. In 27.2.6 we used that Z is a UFD and Z× = {±1}; for a general
number field, we face problems associated with units and the class group of the field,
and the relevant exact sequences will not split!
27.3 ⊲ Rational quaternionic adeles and ideles

In the remainder of this chapter, we generalize the above construction to the adele ring
and idele group of a global field and then a quaternion algebra over a global field. For
the reader on a brisk read, in this section we briefly consider the constructions for a
quaternion algebra over Q.
Let 𝐵 be a quaternion algebra over Q, and let O ⊂ 𝐵 be an order.
27.3.1. The adele ring of 𝐵 is
Ö0 Î
𝐵 := 𝐵 𝑣 = 𝛼 = (𝛼 𝑣 ) 𝑣 ∈ 𝑣 ∈Pl Q 𝐵 𝑣 : 𝛼 𝑣 ∈ O 𝑣 for all but finitely many 𝑣 ,
𝑣
(27.3.2)
the restricted direct product of the topological rings 𝐵 𝑣 with respect to O 𝑣 for places 𝑣
of Q; this definition is independent of the choice of order O, because any two orders are
equal at all but finitely many places by the local-global dictionary for lattices (Theorem
9.4.9). We embed 𝐵 ↩→ 𝐵 diagonally: the image is discrete, closed, and cocompact
since the same is true for Q ↩→ Q (Lemmas 27.1.13–27.1.14). We write
Ö0
b :=
𝐵 𝐵𝑝, (27.3.3)
𝑝
b × 𝐵∞ . We also define
so extending by zero we may identify 𝐵 = 𝐵
Ö
O
b := O 𝑝 ⊆ 𝐵.
b (27.3.4)
𝑝
The idele group of 𝐵 is 𝐵× := 0𝑣 𝐵×𝑣 , the restricted direct product of the topo-
Î
logical groups 𝐵×𝑣 with respect to O×𝑣 ; we similarly define
Ö Ö0
b× :=
O O×𝑝 ≤ 𝐵b× := 𝐵×𝑝 . (27.3.5)
𝑝 𝑝
Working adelically is notationally quite convenient, as the following lemma illus-

trates (Lemma 27.6.8 for 𝐹 = Q).
Lemma 27.3.6. The set of invertible right fractional O-ideals is in bĳection with
b× /O
𝐵 b× via the map 𝐼 ↦→ b
𝛼Ob× , where 𝐼 𝑝 = 𝛼 𝑝 O 𝑝 and b
𝛼 = (𝛼 𝑝 ) 𝑝 ; this map induces a
bĳection
ClsR O ↔ 𝐵× \ 𝐵
b× /O
b×
[𝐼] R ↦→ 𝐵× b
𝛼Ob× .
The most fundamental result in this chapter is the following (see Main Theorem
27.6.14, taking 𝐹 = Q).
Theorem 27.3.7. Let 𝐵 be a division quaternion algebra over Q. Then 𝐵× ≤ 𝐵× is
cocompact and the set 𝐵× \ 𝐵
b× /O
b× is finite.
In particular, combining Lemma 27.3.6 and Theorem 27.3.7, we conclude that the
class set of O is finite, something we proved earlier using the geometry of numbers for
𝐵 definite in section 17.5.
27.4. ADELES AND IDELES 451
27.4 Adeles and ideles

In pursuit of more generality, we now repeat the constructions of adeles and ideles over
a global field. For further reference, see e.g. Cassels [Cas2010], Goldstein [Gol71],
Knapp [Kna2016, Chapter VI], Lang [Lang94, Chapter VII], Neukirch [Neu99, Chap-
ter VI], or Ramakrishnan–Valenza [RM99, Chapter 5].
Throughout the rest of this chapter, let 𝐹 be a global field.
27.4.1. We recall notation from section 14.4 for convenience. The set of places of 𝐹
is denoted Pl 𝐹. For a place 𝑣 of 𝐹, we denote by 𝐹𝑣 the completion of 𝐹 at the place
𝑣, with preferred (normalized) absolute value | | 𝑣 so that the product formula holds in
𝐹: see 14.4.12. If 𝑣 is nonarchimedean, we let
𝑅 𝑣 := {𝑥 ∈ 𝐹𝑣 : |𝑥| 𝑣 ≤ 1} = {𝑥 ∈ 𝐹𝑣 : 𝑣(𝑥) ≥ 0} (27.4.2)
be the valuation ring of 𝐹𝑣 , where we write 𝑣 also for the discrete valuation associated
to the place 𝑣. If 𝐹 is a number field, we will sometimes denote an archimedean place
by writing 𝑣 | ∞, and for an archimedean place we just take 𝑅 𝑣 = 𝐹𝑣 . A set S ⊆ Pl 𝐹
of places is eligible if it is finite, nonempty, and contains all archimedean places.
27.4.3. The adele ring of 𝐹 is the restricted direct product of 𝐹𝑣 with respect to 𝑅 𝑣 :
Ö0
𝐹 := 𝐹𝑣
𝑣
Î (27.4.4)
:= {(𝑥 𝑣 ) 𝑣 ∈ 𝑣 𝐹𝑣 : 𝑥 𝑣 ∈ 𝑅 𝑣 for all but finitely many 𝑣}
Î
= {(𝑥 𝑣 ) 𝑣 ∈ 𝑣 𝐹𝑣 : |𝑥 𝑣 | 𝑣 ≤ 1 for all but finitely many 𝑣}
with the restricted direct product topology. The Îtopology is uniquely characterized
(as a topological ring) by the condition that 𝑅 := 𝑣 𝑅 𝑣 (with the product topology)
is open. Accordingly,
Î a subset 𝑈 ⊆ 𝐹 is open if and only if for all 𝑎 ∈ 𝐹, the set
(𝑎 + 𝑈) ∩ 𝑣 𝑅 𝑣 is open in the product topology.
Giving 𝐹 the discrete topology, we have an isomorphism of topological rings

∼ 𝐹: see Exercise 27.14.
Q ⊗Q 𝐹 −→
27.4.5. We embed 𝐹 ⊆ 𝐹 under the product of the embeddings 𝐹 ↩→ 𝐹𝑣 , i.e., by

𝑥 ↦→ (𝑥) 𝑣 ; this map is well-defined because |𝑥| 𝑣 ≤ 1 for all but finitely many places 𝑣 of
𝐹. The image of 𝐹 in 𝐹 has the discrete topology and is closed in 𝐹; the quotient 𝐹/𝐹
is compact (i.e., 𝐹 is cocompact in 𝐹): for 𝑥 ∈ 𝐹, find 𝑎 ∈ 𝐹 such that 𝑥 𝑣 − 𝑎 𝑣 ∈ 𝑅 𝑣
for all 𝑣 by the Sun Zi theorem (CRT).
27.4.6. Let S ⊂ Pl 𝐹 be an eligible set. We will write

Ö0 Ö
𝐹6S := 𝐹𝔭 , 𝐹S := 𝐹𝑣 (27.4.7)
𝔭∉S 𝑣 ∈S
and extending by zero we identify these sets with their images in 𝐹, so that 𝐹 = 𝐹6S ×𝐹S ;
we call 𝐹6S the S-finite adele ring of 𝐹.
We pass now to the multiplicative situation.
27.4.8. The idele group of 𝐹 is the restricted direct product of 𝐹𝑣× with respect to 𝑅 ×𝑣 :
Ö0
𝐹 × := 𝐹𝑣×
𝑣 (27.4.9)
:= (𝑥 𝑣 ) 𝑣 ∈ 𝑣 𝐹𝑣× : |𝑥 𝑣 | 𝑣 = 1 for all but finitely many 𝑣 .
Î
27.4.10. The topology ×

Î × on 𝐹 (as a topological ring) is uniquely characterized by the
condition that 𝑣 𝑅 𝑣 (with the productÎtopology) is open. Thus, 𝑈 ⊆ 𝐹 × is open if
and only if for all 𝑎 ∈ 𝐹 × , the set 𝑎𝑈 ∩ 𝑣 𝑅 ×𝑣 is open in the product topology.
Note that 𝐹 × does not have the topology induced from being a subspace of 𝐹,
since inversion is not a continuous operation. In general, if 𝐴 is a topological ring, 𝐴×
becomes a topological group when 𝐴× is given the relative topology from
𝐴× ↩→ 𝐴 × 𝐴
𝑥 ↦→ (𝑥, 𝑥 −1 ).
(See Exercise 27.13.)
Just as 𝐹 ⊆ 𝐹 is discrete, 𝐹 × ⊆ 𝐹 × is also discrete.
Definition 27.4.11. The group 𝐶𝐹 := 𝐹 × /𝐹 × is the idele class group of 𝐹.
The justification for calling this the idele class group is given in section 27.5.
27.4.12. As above, if S ⊂ Pl 𝐹 is an eligible set, we define the S-finite ideles

Ö0
𝐹6S× := 𝐹𝑣×
𝑣∉S
which is missing the product at the places in S (a finite product), namely

Ö
𝐹S× := 𝐹𝑣× ;
𝑣 ∈S
extending by 1, we identify these with their images in 𝐹 × , so that 𝐹 × = (𝐹6S ) × × 𝐹S× .
27.4.13. With respect to the normalized absolute values 14.4.12, we have a natural
map
𝐹 × → R>0
(27.4.14)
Ö
(𝑥 𝑣 ) 𝑣 ↦→ k𝑥 𝑣 k 𝑣 .
𝑣
When 𝐹 is a number field, the map (27.4.14) is surjective; when 𝐹 is a function field
with constant field F𝑞 , the image is 𝑞 Z , the cyclic subgroup of R>0 generated by 𝑞. Let
𝐹 (1) := {𝑥 = (𝑥 𝑣 ) 𝑣 :
Î
𝑣 k𝑥 𝑣 k 𝑣 = 1} (27.4.15)
so that 𝐹 (1) is the kernel of (27.4.14). Then 𝐹 × ≤ 𝐹 (1) by the product formula (14.4.6).
27.5. CLASS FIELD THEORY 453
The following theorem is fundamental.
Theorem 27.4.16. The quotient 𝐹 (1) /𝐹 × is compact, i.e., 𝐹 × is cocompact in 𝐹 (1) .
Proof. We give a proof in a more general context in Main Theorem 27.6.14 below. Or
see e.g. Cassels [Cas2010, §16, p. 69] for a direct proof.
Theorem 27.4.16 is equivalent (!) to the Dirichlet unit theorem and the finiteness
of the class group in the number field case, and finite generation of the unit group of a
coordinate ring of a curve and the finiteness of the group of rational divisors of degree
zero in the function field case [Cas2010, §§17–18].
Via the projection map 𝐹 (1) → 𝐹6S× , we have 𝐹 × cocompact also in 𝐹6S× .
27.5 Class field theory

Let 𝐹 sep be a separable closure of 𝐹. In this chapter, we summarize the idelic approach
to class field theory; unfortunately, we must omit most proofs, as a full treatment would
require a lengthy development—but the reader who is willing to accept the statements
should be able to digest what follows and will hopefully be motivated to dig deeper!
For further reading, see Tate [Tate2010], Lang [Lang94, Chapters XI-XI], Neukirch
[Neu99, Chapters IV–VI], or Janusz [Jan96, Chapter V].
27.5.1. Let 𝑅 = 𝑅 (S) be a global ring (the ring of S-integers) for the eligible set
S ⊆ Pl 𝐹. Then 𝑅 is a Dedekind domain with field of fractions 𝐹. The class group of
𝑅 admits an idelic description, embodying the definitions above, as follows.
To simplify notation, throughout we abbreviate 𝐹6𝑆 = 𝐹,b as we take the set S to be
fixed. To an invertible fractional ideal 𝔞 ⊆ 𝐹 of 𝑅, we have 𝔞𝔭 = 𝑅𝔭 for all but finitely
many primes 𝔭, so we can consider its idelic image (𝔞𝔭 )𝔭 ⊆ 𝐹 b under the product of
completions. Since 𝔞 is locally principal, we can write each 𝔞𝔭 = 𝑎 𝔭 𝑅𝔭 with 𝑎 𝔭 ∈ 𝐹𝔭× ,
well-defined up to an element of 𝑅𝔭× ; putting these together we obtain an element
b×
𝑎 = (𝑎 𝔭 )𝔭 ∈ 𝐹
b
and
𝔞 = 𝔞𝑅
b b= b
𝑎𝑅b ⊆ 𝐹.
b
We recover 𝔞 = b 𝔞 ∩ 𝐹 from Lemmas 9.4.6 and 9.5.3. Therefore the group of invertible
fractional ideals of 𝑅 is canonically isomorphic to the quotient
b× / 𝑅
Idl 𝑅 ' 𝐹 b× . (27.5.2)
The principal (invertible) fractional ideals correspond to the image of 𝐹 × in 𝐹

b× .
Therefore there is a canonical isomorphism
∼ 𝐹
Cl 𝑅 −→ b× / 𝑅
b× 𝐹 ×
(27.5.3)
[𝔞] ↦→ b
𝑎𝑅b× 𝐹 × .
27.5.4. Suppose 𝐹 is a number field. If S consists of the set of archimedean places,

then 𝑅 = Z𝐹 is the ring of integers, and Cl 𝑅 is the usual class group. For larger sets S,
we have a natural quotient map Cl Z𝐹 → Cl 𝑅 obtained by the quotient by the classes
of primes 𝔭 corresponding to nonarchimedean places in S.
More generally, one may restrict (27.5.3) to the subgroup of principal fractional
ideals which have a totally positive generator; we then obtain the narrow (or strict)
S-class group
∼ 𝐹
Cl+ 𝑅 −→ b×/ 𝑅
b× 𝐹 × . (27.5.5)
>0
In idelic class field theory, it is often convenient to move between quotients of the
finite ideles and quotients of the full idele group as follows.
Î
Lemma 27.5.6. Let 𝐹∞ := 𝑣 |∞ 𝐹𝑣 and let
×
𝐹∞,>0 := {(𝑎 𝑣 ) 𝑣 ∈ 𝐹∞ : 𝑎 𝑣 > 0 for all 𝑣 real}.
Then the map

b× → 𝐹 × ' 𝐹
𝐹 b× × 𝐹∞
×
(27.5.7)
𝑎 ↦→ (b
b 𝑎 , 1)
induces isomorphisms of topological groups
∼ 𝐹 × /𝐹 × 𝐹 ×
b× /𝐹 × −→
𝐹 ∞
× × ∼ × × ×
(27.5.8)
𝐹 /𝐹 −→ 𝐹 /𝐹 𝐹
b
>0 ∞,>0 .
Proof. Composing with the projection to 𝐹 × /𝐹 × 𝐹∞,>0

× ×
, we see that the kernel is 𝐹>0
× ×
and that the induced map is surjective; similarly for the quotient by 𝐹 𝐹∞ . The few
details are requested in Exercise 27.9.
Using Lemma 27.5.6, we have

∼ 𝐹 × /( 𝑅
Cl 𝑅 −→ b× 𝐹∞
× ×
𝐹 )
+ ∼ × b× ×
Cl 𝑅 −→ 𝐹 /( 𝑅 𝐹 𝐹 × ).
∞,>0
27.5.9. Class field theory relates class groups to abelian extensions. For example, let
Ö Ö
𝐻=𝑅 b× × 𝐹S× = 𝑅 ×𝑣 × 𝐹𝑣× ≤ 𝐶𝐹 := 𝐹 × /𝐹 × .
𝑣∉S 𝑣 ∈S
Then 𝐻 ≤ 𝐶𝐹 is an open subgroup of finite index, and the projection map

∼ 𝐹
𝐶𝐹 /𝐻 −→ b×/ 𝑅
b×
is an isomorphism, which together with (27.5.3) gives an isomorphism to Cl 𝑅. So we

are led to consider the finite-index open subgroups of 𝐶𝐹 .
The main theorem of idelic class field theory for finite extensions is as follows.
27.5. CLASS FIELD THEORY 455
Theorem 27.5.10. There is a bĳection

{𝐾 ⊆ 𝐹 sep : 𝐾 ⊇ 𝐹 finite abelian} ↔ {𝐻 ≤ 𝐶𝐹 : 𝐻 finite-index open}
(27.5.11)
𝐾 ↦→ 𝐹 × Nm𝐾 /𝐹 𝐶𝐾
∼ Gal(𝐾 | 𝐹).
together with functorial isomorphisms 𝐶𝐹 /𝐻 −→
∼ Gal(𝐾 | 𝐹) is called the Artin isomorphism for 𝐻, 𝐾.
The map 𝐶𝐹 /𝐻 −→
Proof. See e.g. Tate [Tate2010, §5].
27.5.12. Rewriting the main theorem (Theorem 27.5.10) slightly, we see that if 𝐻 ≤ 𝐹 b×
×
is an open finite-index subgroup containing 𝐹>0 , then there is a finite abelian extension
𝐾 ⊇ 𝐹 with the Artin isomorphism
∼ Gal(𝐾 | 𝐹).
b× /𝐻 −→
𝐹
27.5.13. Combining the surjections 𝐶𝐹 → Gal(𝐾 | 𝐹), we obtain a continuous homo-

morphism
𝜃 : 𝐶𝐹 → lim Gal(𝐾 | 𝐹) = Gal(𝐹 ab | 𝐹)
←−−
𝐾
called the global Artin homomorphism, where 𝐹 ab ⊆ 𝐹 sep is the maximal abelian
extension of 𝐹 in 𝐹 sep .
If 𝐹 is a number field, then 𝜃 is surjective; let 𝐷 𝐹 be the connected component of
1 in 𝐶𝐹 . Then 𝐷 𝐹 is a closed subgroup with
𝐷 𝐹 ' R × (R/Z) 𝑐 × Sol𝑟 +𝑐−1 (27.5.14)
(see Exercise 27.10). We therefore have an isomorphism 𝐶𝐹 /𝐷 𝐹 ' Gal(𝐹 ab | 𝐹).

If 𝐹 is a function field with finite constant field 𝑘, then 𝜃 is injective, and 𝜃 (𝐶𝐹 ) is the
dense subgroup of automorphisms 𝜎 ∈ Gal(𝐹 ab | 𝐹) whose restriction to Gal(𝑘 al | 𝑘) '
Z lies in Z, i.e., acts by an integer power of the Frobenius. See Tate [Tate2010, §5.4–
b
5.7].
We conclude with a nice application to the classification of quaternion algebras.
Proposition 27.5.15. Let Σ ⊆ Pl 𝐹 be a finite subset of noncomplex places of 𝐹 of
even cardinality. Then there exists a quaternion algebra 𝐵 over 𝐹 with Ram 𝐵 = Σ.
Proof. Let 𝐾 ⊇ 𝐹 be a separable quadratic extension that is inert (an unramified field
extension) at every 𝑣 ∈ Σ: such an extension exists by Exercise 14.20. By the main
theorem of class field theory, we have [𝐶𝐹 : 𝐹 × Nm𝐾 /𝐹 𝐶𝐾 ] = [𝐾 : 𝐹] = 2, where
𝐶𝐹 = 𝐹 × /𝐹 × and similarly 𝐶𝐾 are idele class groups. Therefore
[𝐶𝐹 : 𝐹 × Nm𝐾 /𝐹 𝐶𝐾 ] = [𝐹 × : 𝐹 × Nm𝐾 /𝐹 (𝐾 × )] = 2 (27.5.16)
as well.
For each 𝑣 ∈ Σ, let 𝜋 𝑣 be a uniformizer for 𝑅 𝑣 and if 𝑣 is real let 𝜋 𝑣 = −1. Since
𝐾 𝑣 ⊇ 𝐹𝑣 is an unramified field extension, we have 𝜋 𝑣 ∉ Nm𝐾𝑣 /𝐹𝑣 (𝐾 𝑣× ). For 𝑣 ∈ Σ,
let 𝜋 𝑣 = (1, . . . , 1, 𝜋 𝑣 , . . . ) ∈ 𝐹 × . Then 𝜋 𝑣 ∉ Nm𝐾 /𝐹 (𝐾 × ).
We claim that 𝜋 𝑣 ∉ 𝐹 × Nm𝐾 /𝐹 (𝐾 × ). Otherwise, there would exist 𝑎 ∈ 𝐹 × such

that 𝑎𝜋 𝑣 ∈ Nm𝐾 /𝐹 (𝐾 × ), so 𝑎 ∈ Nm𝐾𝑤 0 /𝐹𝑣 0 (𝐾 𝑤 × ) for all 𝑤 0 | 𝑣 0 with 𝑣 0 ≠ 𝑣, and
0
×
𝑎 ∉ Nm𝐾𝑣 /𝐹𝑣 (𝐾 𝑣 ); but then 𝑎 is a local norm at all but one real place, so by the
Hasse norm theorem (Theorem 26.8.23), 𝑎 ∈ Nm𝐾 /𝐹 (𝐾 × ) is a global norm, but this
contradicts that 𝑎 is not a local norm at 𝑣.
Î
Now let 𝑝 = 𝑣 ∈Σ 𝜋 𝑣 . Since #Σ is even, by the previous paragraph and (27.5.16)
we get
𝑝 = 𝑏𝑢 ∈ 𝐹 × Nm𝐾 /𝐹 (𝐾 × ).

𝐾, 𝑏
Consider the quaternion algebra . For all places 𝑣 ∈ Σ, we have 𝐾 𝑣 a field
𝐹
and 𝑏 = 𝜋 𝑣 𝑢 −1 ×
𝑣 ∉ Nm𝐾 𝑣 /𝐹𝑣 (𝐾 𝑣 ), so 𝑣 ∈ Ram 𝐵. At every other place 𝑣 ∉ Σ, we
0
−1 ×
have either that 𝐾 𝑣 is not a field or 𝑏 = 𝑢 𝑣 ∈ Nm𝐾𝑣 /𝐹𝑣 (𝐾 𝑣 ), and in either case
𝑣 0 ∉ Ram 𝐵.
27.6 Noncommutative adeles

We retain notation from the previous section, in particular the abbreviations 𝑅 = 𝑅 (S)
and 𝐹6𝑆 = 𝐹.
b Let 𝐵 be a finite-dimensional simple algebra over the global field 𝐹. In
this section, we extend idelic notions to 𝐵; the main case of interest is where 𝐵 is a
quaternion algebra over 𝐹. We recall the topology on 𝐵 𝑣 for places 𝑣, discussed in
section 13.5. Let O ⊆ 𝐵 be an 𝑅-order.
27.6.1. The adele ring of 𝐵 is the restricted direct product of the topological rings 𝐵 𝑣
with respect to O 𝑣 :
Ö0 Î
𝐵 := 𝐵 𝑣 = {(𝛼 𝑣 ) 𝑣 ∈ 𝑣 𝐵 𝑣 : 𝛼 𝑣 ∈ O 𝑣 for all but finitely many 𝑣 ∉ S}
𝑣 ∈Pl(𝐹 )
The topology onÎ𝐵 (as a topological ring) is uniquely characterized by the property
that the subring 𝑣 O 𝑣 is open with the product topology.
By the local-global dictionary for lattices (Theorem 9.4.9), the definition of 𝐵 is
independent of the choice of order O and eligible set S (and base ring 𝑅 = 𝑅 (S) ).
27.6.2. Just as in 27.4.5, we embed 𝐵 ↩→ 𝐵 diagonally. A basis for 𝐵 as an 𝐹-vector

space shows that 𝐵 is a free 𝐹-module of finite rank. Then, since the image 𝐹 ↩→ 𝐹
is discrete, closed, and cocompact, arguing in each coordinate (with respect to the
chosen basis), we conclude that 𝐵 ↩→ 𝐵 is discrete, closed, and cocompact. (Details
are requested in Exercise 27.11.)
We now turn to the multiplicative structure, the main object of our concern.
27.6.3. The idele group of 𝐵 is the restricted direct product of the topological groups
𝐵×𝑣 with respect to O×𝑣 :
Ö0
𝐵× := 𝐵×𝑣 = (𝛼 𝑣 ) 𝑣 ∈ 𝑣 𝐵×𝑣 : 𝛼 𝑣 ∈ O×𝑣 for all but finitely many 𝑣 ;
Î
𝑣
27.6. NONCOMMUTATIVE ADELES 457
equivalently, 𝐵× is the unit group of 𝐵 with the topology as in 27.4.10. The topology
Î on
𝐵× as a topological group is characterized by the condition that the subgroup 𝑣 O×𝑣
is open with the product topology. Again, 𝐵× is independent of the choice of O and
eligible set S because any two such constructions differ at only finitely many places.
27.6.4. The S-finite adele ring is

Ö0
𝐵6S := 𝐵𝑣 ; (27.6.5)
𝑣∉S
extending by zero, we may identify 𝐵6S with its image in 𝐵. The S-finite adele ring has
a compact open subring Ö
O6S := O 𝑣 ⊆ 𝐵6S . (27.6.6)
𝑣∉S
We similarly define the S-finite idele group with its compact open subgroup
Ö0 Ö
𝐵6S× := 𝐵×𝑣 ⊃ O×𝑣 =: O×6S . (27.6.7)
𝑣∉S 𝑣∉S
When no confusion can result (S is clear from context), we will drop the superscript
b = 𝐵6S and O× = O
and replace with hats, writing simply 𝐵 b× , etc.
6S
Just as in 27.5.1, the ideles provide a convenient way of encoding fractional ideals,
as follows.
Lemma 27.6.8. The set of locally principal, right fractional O-ideals is in bĳection
b×/O
with 𝐵 b× via the map 𝐼 ↦→ b
𝛼Ob× , where 𝐼𝔭 = 𝛼𝔭 O𝔭 and b
𝛼 = (𝛼𝔭 )𝔭 ; this map induces
a bĳection
ClsR O ↔ 𝐵× \ 𝐵 b×/O b×
(27.6.9)
[𝐼] R ↦→ 𝐵× b
𝛼Ob× .
Proof. Let 𝐼 be a locally principal right fractional O-ideal, so 𝐼𝔭 = 𝛼𝔭 O𝔭 for all primes
𝔭 of 𝑅, with 𝛼𝔭 well-defined up to right multiplication by an element of O×𝔭 , so to 𝐼
we associate (𝛼𝔭 O×𝔭 )𝔭 = b𝛼Ob× ∈ 𝐵 b×/Ob× . Conversely, given b 𝛼 ∈ 𝐵b×/Ob× we recover
𝐼 =b𝛼Ob ∩ 𝐵 from Lemmas 9.4.6 and 9.5.3.
The equivalence relation defining the (right) class set is given by left multiplication
by 𝐵× , so the second statement follows.
Remark 27.6.10. We recall by Main Theorem 16.6.1 that for 𝐵 a quaternion algebra,
an 𝑅-lattice 𝐼 ⊂ 𝐵 is locally principal if and only if it is invertible.
27.6.11. In analogy with 27.4.13, we have a natural multiplicative map
k k : 𝐵× → R>0
(27.6.12)
Ö
𝛼 = (𝛼 𝑣 ) 𝑣 ↦→ |nrd(𝛼 𝑣 )| 𝑣
𝑣
and we define
𝐵 (1) = kerk k = {𝛼 = (𝛼 𝑣 ) 𝑣 :
Î
𝑣 |nrd(𝛼 𝑣 )| 𝑣 = 1} (27.6.13)
By the product formula (14.4.6), we have 𝐵× ≤ 𝐵 (1) .

By comparison, we have also the groups
𝐵1 = {𝛼 ∈ 𝐵 : nrd(𝛼) = 1}
𝐵1 = {𝛼 ∈ 𝐵 : nrd(𝛼) = 1}
satisfying 𝐵1 ≤ 𝐵1 ≤ 𝐵 (1) .
The following theorem is fundamental (see Fujisaki [Fuj58, Theorem 8.3], Weil
[Weil82, Lemma 3.1.1]).
Main Theorem 27.6.14 (Fujisaki’s lemma). Let 𝐵 be a finite-dimensional division
algebra over a global field 𝐹. Then the following statements hold.
(a) 𝐵× ≤ 𝐵 (1) is cocompact.
b× is cocompact and the set
(b) For any eligible set S, the subgroup 𝐵× ≤ 𝐵6S× = 𝐵
𝐵× \ 𝐵
b×/O
b× is finite.
Proof. The natural place to prove this result is after some more serious analysis has
been done—but it is too important to wait for this. The small amount of input needed,
which can be seen as an (ineffective) idelic version of the Minkowski convex body
theorem (Theorem 17.5.5), is as follows. There exists a compact subset 𝐸 such that
for all 𝛽 ∈ 𝐵 (1) ,
the map 𝐵 → 𝐵\𝐵 is not injective when restricted to 𝛽𝐸 . (27.6.15)
For the proof of (27.6.15), see Exercise 29.11: in a nutshell, there is a measure 𝜇 on 𝐵 in
which 𝜇(𝐵\𝐵) < ∞, and a compact 𝐸 with 𝜇(𝐸) satisfies 27.6.15, as 𝜇(𝛽𝐸) = 𝜇(𝐸).
We first quickly prove part (b), assuming part (a). We have O b× open in 𝐵 b× , so the
open cover {𝐵× b 𝛼Ob× } b× can be reduced to a finite cover, whence the double coset
b∈ 𝐵
𝛼
space is finite.
We now turn to prove (a), which we do in steps.
Step 1: Setup. From the set 𝐸 granted above, we define
𝑋 := 𝐸 − 𝐸 = {𝜂 − 𝜂 0 : 𝜂, 𝜂 0 ∈ 𝐸 }. (27.6.16)
Since 𝐸 is compact, we conclude that 𝑋 ⊆ 𝐵 is compact (subtraction is continuous).

We will also use the set of products 𝑋 𝑋, which is again compact in 𝐵.
Step 2: Measuring differences. We claim that for all 𝛽 ∈ 𝐵 (1) , we have 𝛽𝑋 ∩ 𝐵× ≠ ∅.
Indeed, let 𝛽 ∈ 𝐵 (1) . By (27.6.15), there exist distinct 𝜂, 𝜂 0 ∈ 𝐸 (depending on 𝛽) such
that
𝛽(𝜂 − 𝜂 0) ∈ 𝐵. (27.6.17)
27.6. NONCOMMUTATIVE ADELES 459
Because 𝜂 ≠ 𝜂 0 and 𝐵 is a division algebra (here is where we use this hypothesis), we

have 𝛽(𝜂 − 𝜂 0) ∈ 𝐵× , as claimed.
A similar statement holds on the right, multiplying by 𝛽−1 .
Step 3: Cocompactness. We claim that there exists a compact set 𝐾 ⊆ 𝐵× × 𝐵× such

that for all 𝛽 ∈ 𝐵 (1) , there exist 𝛽 ∈ 𝐵× and 𝜈 ∈ 𝐵 (1) such that 𝛽 = 𝛽𝜈 and (𝜈, 𝜈 −1 ) ∈ 𝐾.
To prove this claim, first define 𝑇 := 𝐵× ∩ 𝑋 𝑋. Since 𝑇 is the intersection of a
discrete set and a compact set, we conclude that 𝑇 is finite. Let 𝑇 −1 := {𝛾 −1 : 𝛾 ∈ 𝑇 }
and let
𝐾 := 𝑇 −1 𝑋 × 𝑋. (27.6.18)
To check the claim, let 𝛽 ∈ 𝐵 (1) . By Step 2, 𝛽𝑋∩𝐵× ≠ ∅ and (similarly) 𝑋 𝛽−1 ∩𝐵× ≠ ∅.
Therefore there exist 𝜈, 𝜈 0 ∈ 𝑋 and 𝛽, 𝛽 0 ∈ 𝐵× such that
𝛽𝜈 = 𝛽 and 𝜈 0 𝛽−1 = 𝛽 0. (27.6.19)
Therefore
𝛽 0 𝛽 = (𝜈 0 𝛽−1 ) (𝛽𝜈) = 𝜈 0 𝜈 ∈ 𝐵× ∩ 𝑋 𝑋.
Then 𝜈 −1 ∈ 𝑇 −1 𝑋, and 𝑇 −1 𝑋 is compact. We have shown that 𝛽 = 𝛽𝜈 −1 with 𝛽 ∈ 𝐵×

and (𝜈 −1 , 𝜈) ∈ 𝐾 = 𝑇 −1 𝑋 × 𝑋, and this proves the claim.
Step 4: Conclusion. By the definition of the topology on 𝐵× , the set
{𝜈 ∈ 𝐵 (1) : (𝜈, 𝜈 −1 ) ∈ 𝐾 }
is compact; then, by the claim in Step 3, this set surjects onto 𝐵× \𝐵 (1) . We conclude
that 𝐵× \𝐵 (1) is compact, completing the proof of part (a).
Corollary 27.6.20. If 𝐵 is a division algebra or 𝐵 ' M𝑛 (𝐹), and O ⊆ 𝐵 is an 𝑅-order,

then the class set ClsR O is finite.
Proof. If 𝐵 is a division algebra, we combine Lemma 27.6.8 and Main Theorem

27.6.14. Otherwise, 𝐵 ' M𝑛 (𝐹). If O is maximal then O ' M𝑛 (𝑅) and the result
follows from 17.3.7; and then the result for a general order O follows from Exercise
17.3(b).
Remark 27.6.21. Corollary 27.6.20 covers all quaternion algebras 𝐵. This finiteness
statement generalizes to the theorem of Jordan–Zassenhaus: see Remark 17.7.27.
27.6.22. The idelic point of view (Lemma 27.6.8) also makes it clear why the class
number is independent of the order within its genus (Lemma 17.4.11): the idelic
description only depends on the local orders, up to isomorphism.
We have two other objects that admit a nice idelic description.

27.6.23. The genus of an order and its type set (see section 17.4) can be similarly
described. Let O be an 𝑅-order, and let O0 ∈ Gen O be an order in the genus of O.
Then O0 is locally isomorphic to O, so there exists b𝜈 ∈ 𝐵 b× such that b b𝜈 −1 = O
𝜈 Ob b0 ,
well defined up to right multiplication by an element of the normalizer 𝑁 𝐵b× ( O).
b We
0
recover O = O 0
b ∩ 𝐵, so this gives a bĳection
b×/𝑁 b× ( O).
Gen O ↔ 𝐵 b
𝐵
Two such orders are isomorphic if and only if there exists 𝛽 ∈ 𝐵× such that 𝛽O𝛽−1 = O0,
so we have a bĳection
Typ O ↔ 𝐵× \ 𝐵 b×/𝑁 b× ( O).
𝐵
b (27.6.24)
Corollary 27.6.25. The type set Typ O is finite.
Proof. In view of (27.6.24), the double coset 𝐵× \ 𝐵

b×/𝑁 b× ( O)
𝐵
b is a further quotient of
× × ×
the set 𝐵 \ 𝐵 /O which is finite by Main Theorem 27.6.14.
b b
27.6.26. Referring to section 18.5 locally, we see that the group of locally principal
two-sided O-ideals Idl(O) is in bĳection with
b× \𝑁 b× ( O)/
O b O b× = 𝑁 b× ( O)/
b O b× = O
b× \𝑁 b× ( O)
b (27.6.27)
𝐵 𝐵 𝐵
where
𝑁 𝐵b× ( O)
b = {b b× : b
𝛼∈𝐵 𝛼Ob = Ob
b𝛼 }
is the normalizer of O b× . Furthermore, the group of isomorphism classes of locally

b in 𝐵
principal two-sided O-ideals is therefore in bĳection with
b× = 𝑁 b× ( O)/ b× = 𝑁 𝐵× (O) O
b× \𝑁 b× ( O).

𝑁 𝐵× (O)\𝑁 𝐵b× ( O)/
b O
𝐵
b 𝑁 𝐵× (O) O
𝐵
b
27.7 Reduced norms

To conclude, we consider reduced norms; we specialize and suppose that 𝐵 is a
quaternion algebra.
27.7.1. Since S contains all archimedean places, by the local norm calculation (Lemma
13.4.9), we have nrd( 𝐵b× ) = 𝐹 b× . By the Hasse–Schilling theorem on norms (Main
Theorem 14.7.4), we have nrd(𝐵× ) = 𝐹>×Ω 0 , where Ω ⊆ Ram 𝐵 is the set of real
ramified places and 𝐹>×Ω 0 is the set of elements positive at all 𝑣 ∈ Ω (recalling 14.7.2).
Therefore, the reduced norm (in each component) yields a surjective map
nrd : 𝐵× \ 𝐵
b×/O
b× → 𝐹 × \𝐹
>Ω 0
b×/nrd( O
b× ). (27.7.2)
We will now see that the group 𝐹>×Ω 0 \𝐹

b×/nrd( O
b× ) in (27.7.2) is a certain class
group of 𝑅.
Lemma 27.7.3. The subgroup nrd( O b× )𝐹 × ≤ 𝐹 b× is a finite-index open subgroup

>Ω 0
containing 𝐹>0 , the group of totally positive elements of 𝐹 × . If moreover O is maximal,
×
b× ) = 𝑅
then nrd( O b× .
Proof. In Lemma 13.4.9, we saw that if O𝔭 is maximal, then nrd(O×𝔭 ) = 𝑅𝔭× ; for
the finitely many remaining 𝔭 ⊆ 𝑅, the 𝑅𝔭 -order O𝔭 is of finite index in a maximal
superorder, so nrd(O×𝔭 ) is a finite index open subgroup of 𝑅𝔭× . Putting these together,
we conclude nrd( O b× ) is a finite index open subgroup of 𝑅
b× .
But 𝐹b×/ 𝑅
b× 𝐹 × ' ClΩ 𝑅 is a finite group and therefore
>Ω 0
b× : nrd( O
[𝐹 b× )𝐹 × ] = [ 𝐹
b× : 𝑅
b× 𝐹 × ] [ 𝑅
b× 𝐹 × : nrd( O
b× )𝐹 × ] < ∞.
>Ω 0 >Ω 0 >Ω 0 >Ω 0
Finally, we have 𝐹>×Ω 0 ≥ 𝐹>0

× , as the latter possibly requires further positivity.
27.7.4. Let
𝐺 (O) := 𝐹>×Ω 0 nrd( O
b× ) ≤ 𝐹
b× . (27.7.5)
From Lemma 27.7.3, 𝐺 (O) is a finite-index open subgroup containing 𝐹>0 × . By class
field theory 27.5.12, there exists a class field 𝐾 for 𝐺 (O), i.e., there exists a finite
abelian extension 𝐾 ⊇ 𝐹 and an Artin isomorphism
∼ Gal(𝐾 | 𝐹).
Cl𝐺 (O) 𝑅 = 𝐹 × /𝐺 (O) −→ (27.7.6)
The group Cl𝐺 (O) 𝑅 only depends on the genus of O: if O0 ∈ Gen O then O0 is locally
isomorphic to O, so there exists b b× such that O
𝜈∈𝐵 b0 = b
𝜈 −1 Ob b0× ) = nrd( O
b𝜈 so nrd( O b× ).
Example 27.7.7. Suppose 𝐹 is a number field and S is the set of archimedean places
of 𝐹, so that 𝑅 = Z𝐹 is the ring of integers in 𝐹. Suppose further that O is maximal.
Then 𝐺 (O) = 𝐹>×Ω 0 𝑅
b× . Recalling 17.8.2, let Ω be the set of ramified (necessarily real)
archimedean places of 𝐵, and let ClΩ Z𝐹 := 𝐹>×Ω 0 \𝐹 b× / 𝑅
b× , equivalently, ClΩ Z𝐹 is the
group of fractional ideals of Z𝐹 modulo the subgroup of principal fractional ideals
generated by elements in 𝐹>×Ω 0 . Then Cl𝐺 (O) Z𝐹 = ClΩ Z𝐹 by definition, so we have
a surjective map
nrd : Cls O → Cl𝐺 (O) Z𝐹 = ClΩ Z𝐹 .
The two extreme cases: if 𝐵 is unramified at all real places, then Ω = ∅, and
ClΩ Z𝐹 = Cl Z𝐹 is the class group; if 𝐵 is ramified at all real places, then ClΩ Z𝐹 =
Cl+ Z𝐹 is the narrow class group.
Remark 27.7.8. It is a fundamental result of Eichler (Theorem 17.8.3) that whenever

there exists 𝑣 ∈ S such that 𝐵 is unramified at 𝑣, then the reduced norm map (27.7.2)
is injective, and hence bĳective, giving a bĳection between the class set of O and a
certain class group of 𝑅. This topic is the main result in the next chapter.
Exercises
Unless otherwise specified, let 𝐹 be a global field and let 𝐵 be a quaternion algebra
over 𝐹.
1. If we take the direct product instead of the restricted direct product in the
definition of the adele ring, we lose local compactness. More precisely, let
Î
{𝑋𝑖 }𝑖 ∈𝐼 be a collection of nonempty topological spaces. Show that 𝑋 = 𝑖 ∈𝐼 𝑋𝑖
is locally compact if and only if each 𝑋𝑖 is locally compact and all but finitely
many 𝑋𝑖 are compact.
2. Review the language of group actions and fundamental sets (section 34.1).
a) Equip Q with the discrete topology. We have a group action Q Q (by
Z × [0, 1] ⊆ Q
addition). Show that b b × R is a fundamental set for this action.
[Hint: Review the arguments in Lemmas 27.1.13–27.1.14.]
b) Similarly, show that Q× Q× and that bZ× × R>0 ⊆ Qb × R is a fundamental
set for this action.
3. For a prime 𝑝, let 𝑝b = ( 𝑝, 1, . . . , 1, 𝑝, 1, . . . ) ∈ Q be the adele which is equal to
𝑝 in the 𝑝th and ∞th component and 1 elsewhere.
a) Show that the sequence 𝑝b, ranging over primes 𝑝, does not converge in
Q× ; conclude that Q× is not compact.
b) However, show that the sequence 𝑝b has a subsequence converging to the
identity in the quotient Q× /Q× .
Î
Z = lim Z/𝑛Z ' 𝑝 Z 𝑝 .
4. Recall that b
←−−𝑛
𝛼= ∞
Í
(a) Prove that each b 𝛼∈b Z has a unique representation as b 𝑛=1 𝑐 𝑛 𝑛! where
𝑐 𝑛 ∈ Z and 0 ≤ 𝑐 𝑛 ≤ Î𝑛.
(b) Prove that b Z× ' bZ× ∞ 𝑛=1 Z/𝑛Z as profinite groups. [Hint: Consider the
product of the 𝑝-adic logarithm maps and use the fact that for every prime
power 𝑝 𝑒 there are infinitely many primes 𝑞 such that 𝑝 𝑒 k (𝑞 − 1).]
(c) Prove for every 𝑛 ∈ Z>0 that the natural map Z/𝑛Z → b Z/𝑛bZ is an isomor-
phism.
(d) Prove that there is a bĳection from Z ≥0 to the set of open subgroups of b Z
mapping 𝑛 ↦→ 𝑛b Z.
5. We recall from 27.1.15 the solenoid Sol = (R × b Z)/Z (with Z embedded diago-
nally, and given the quotient topology).
(a) Prove that Sol is a compact, Hausdorff, and connected topological group.
(b) Prove that Sol ' Q/Q as topological groups.
(c) Prove that Sol ' lim R/ 𝑛1 Z with respect to the directed system 𝑛 ∈ Z ≥1
←−−𝑛
under divisibility.
(d) Show that the group of path components of Sol is isomorphic to b Z/Z, and
conclude that Sol is not path connected. [Hint: the neutral path component
is the image of {0} × R ⊆ Q b × R.]
6. Recall that by definition, a set 𝑈 ⊆ 𝐹 is open if and only if 𝑈 ∩ 𝐹6S is open in 𝐹6S
S ⊆ Pl 𝐹. Show that
for all eligible Î Î 𝑈 ⊆ 𝐹 is open if and only if for all 𝑎 ∈ 𝐹
that (𝑎 + 𝑈) ∩ 𝑣 𝑅 𝑣 is open in 𝑣 𝑅 𝑣 .
7. Show that the topology on 𝐹 × agrees with the subspace topology induced on
𝐹 × ↩→ 𝐹 × 𝐹 by 𝑥 ↦→ (𝑥, 𝑥 −1 ).
8. Show that if S is eligible and O ⊆ 𝐵 is an 𝑅 (S) -order, then
Î
𝐵 = {(𝑥 𝑣 ) 𝑣 ∈ 𝑣 𝐵𝑣 : 𝑥 𝑣 ∈ O 𝑣 for all but finitely many 𝑣}
and
𝐵× = (𝑥 𝑣 ) 𝑣 ∈ 𝑣 𝐵×𝑣 : 𝑥 𝑣 ∈ O×𝑣 for all but finitely many 𝑣
Î
and therefore that this definition is independent of the choice of O (and S).
I 9. Prove Lemma 27.5.6.
10. Returning to 27.5.13, let 𝐷 𝐹 ≤ 𝐶𝐹 be the connected component of 1 in the idele
class group of 𝐹. Show that 𝐷 𝐹 is a closed subgroup with
𝐷 𝐹 ' R × (R/Z) 𝑐 × Sol𝑟 +𝑐−1
where 𝑟 is the number of real places of 𝐹 and 𝑐 the number of complex places.
Interpret this isomorphism explicitly for 𝐹 a quadratic field for both 𝐹 real and
imaginary: what ‘explains’ the factors that appear?
I 11. Let 𝐵 be a finite-dimensional 𝐹-algebra.
(a) Show that 𝐵 is discrete and closed in 𝐵. [Hint: 𝐹 is discrete in 𝐹 by the
product formula.]
(b) Show that 𝐵 is cocompact in 𝐵 (under the diagonal embedding), i.e., that
𝐵/𝐵 is compact.
(c)
12. Give a fundamental system of neighborhoods of 0 in 𝐵 b× .
b and of 1 in 𝐵
I 13. Let 𝐴 be a topological ring.
(a) Suppose that 𝐴× ⊆ 𝐴 has the induced topology. Give an example to show
that the map 𝑥 ↦→ 𝑥 −1 on 𝐴× is not necessarily continuous.
(b) Now embed
𝐴× ↩→ 𝐴 × 𝐴
𝑥 ↦→ (𝑥, 𝑥 −1 )
and give 𝐴× the subspace topology. Show that 𝐴× in this topology is a

topological group.
I 14. Let S ⊂ Pl 𝐹 be an eligible set.
Î
(a) Show that 𝑅 (S) is discrete in 𝐹S = 𝑣 ∈S 𝐹𝑣 . [Hint: it is enough to show
this for a neighborhood of 0, and then use the fact that the norm must be
an integer.]
Î
(b) Prove that if O is an 𝑅 (S) -order in 𝐵, then O is discrete in 𝐵S = 𝑣 ∈S 𝐵𝑣 .
15. Let 𝐹 be a global field and let 𝐾 be a finite separable extension of 𝐹.
Î
(a) Show that 𝐾 ' 𝐹 ⊗𝐹 𝐾. [Hint: Use the fact that 𝐹𝑣 ⊗𝐹 𝐾 ' 𝑤 𝐾 𝑤
where 𝑤 runs over the places above 𝑣.]
(b) Show that
Î
𝐾 = (𝑥 𝑤 ) 𝑤 ∈ 𝑤 𝐾 𝑤 : |Nm𝐾𝑤 /𝐹𝑣 𝑥 𝑣 | 𝑣 ≤ 1 for almost all 𝑣
but that the inclusion

Î
𝐵 = 𝐵 ⊗𝐹 𝐹 ⊂ {(𝑥 𝑣 ) 𝑣 ∈ 𝑣 𝐵𝑣 : |nrd(𝑥 𝑣 )| 𝑣 ≤ 1 for almost all 𝑣}
is strict, so the corresponding statement is false for 𝐵.

16. Let 𝑅 = 𝑅 (S) be a global ring and O be an 𝑅-order in 𝐵. Show that the set
b× /𝑁 b× ( O),
of 𝑅-orders which are connected to O is in bĳection with 𝐵 b where
𝐵
𝑁 𝐵b× ( O)
b is the normalizer of O
b in 𝐵 ×
b .
b× ) = 𝑅
17. Extend Lemma 27.7.3 as follows: if O is an Eichler order, then nrd( O b× .
Chapter 28
Strong approximation
28.1 ⊲ Beginnings
We have already seen in several places in this book how theorems about quaternion
algebras over global fields are often first investigated locally, and then a global result is
recovered using some form of approximation. Approximation provides a way to trans-
fer analytic properties (encoded in congruences or bounds) into global elements. In
this chapter, we develop robust approximation theorems and investigate their arithmetic
applications.
We begin by reviewing weak and strong approximation over Q, taking a breath in
preparation for the idelic efforts to come.
28.1.1. The starting point is the Sun Zi theorem (CRT): given a finite, nonempty set
S of primes, and for each 𝑝 ∈ S an exponent 𝑛 𝑝 ∈ Z ≥1 and an element 𝑥 𝑝 ∈ Z/𝑝 𝑛 𝑝 Z,
there exists 𝑥 ∈ Z such that 𝑥 ≡ 𝑥 𝑝 (mod 𝑝 𝑛 𝑝 ) for all 𝑝 ∈ S. These congruences can
be equivalently formulated in the 𝑝-adic metric by lifting Î to 𝑥 𝑝 ∈ Z 𝑝 and asking for
|𝑥 −𝑥 𝑝 | < 𝑝 −𝑛 𝑝 for 𝑝 ∈ S; or equivalently, the map Z → 𝑝 ∈S Z 𝑝 has dense image for
any finite, nonempty set of primes S (giving the target the product topology). We may
therefore think of the CRT as an approximation theorem, in the sense that it allows us
to find an integer that simultaneously approximates a finite number of 𝑝-adic integers
arbitrarily well.
We may generalize 28.1.1 (recalling notation from section 27.1, in particular

27.1.8) as follows.
Theorem 28.1.2 (Weak approximation). Let S ⊆ Pl Q be a finite, nonempty set of

Î
places of Q. Then the image of Q ↩→ QS := 𝑣 ∈S Q 𝑣 is dense.
Proof. Let 𝑥 𝑣 ∈ Q 𝑣 for each 𝑣 ∈ S, and let 𝜖 > 0. We want to show
there exists 𝑥 ∈ Q such that |𝑥 − 𝑥 𝑣 | 𝑣 < 𝜖 for all 𝑣 ∈ S. (28.1.3)
We proceed by considering increasingly more general cases; our intent is to com-

municate (concrete, effective) meaning behind the symbols and to set us up to prove
465
466 CHAPTER 28. STRONG APPROXIMATION
strong approximation below; for a short, uniform proof (which reproves the CRT), see
Exercise 28.1.
Case 1. Suppose ∞ ∉ S and 𝑥 𝑝 ∈ Z 𝑝 for all 𝑝 ∈ S. Then (28.1.3) holds by the CRT,
as in 28.1.1; in fact, we have infinitely many 𝑥 ∈ Z for this purpose.
Case 2. Suppose ∞ ∉ S, but 𝑥 𝑝 ∈ Q 𝑝 for 𝑝 ∈ S. We employ continuity of
multiplication to reduce to the previous case, as follows. We consider the least
common denominator:
Ö
𝑑 := 𝑝 max(0,−𝑣𝑝 ( 𝑥 𝑝 )) ∈ Z>0 . (28.1.4)
𝑝 ∈S
Then 𝑑𝑥 𝑝 ∈ Z 𝑝 for all 𝑝 ∈ S. By the case just established by the CRT, there exists
𝑥 0 ∈ Z such that |𝑥 0 − 𝑑𝑥 𝑝 | 𝑝 < 𝜖 | 𝑑| 𝑝 for all 𝑝 ∈ S, so taking 𝑥 := 𝑥 0/𝑑 and dividing
through we conclude that (28.1.3) holds.
Case 3. To conclude, suppose ∞ ∈ S. We employ an additive translation: we find
a rational number close to 𝑥∞ and add to it a small solution to the previous case, as
follows. Since Q ⊆ R is dense, there exists 𝑦 ∈ Q such that | 𝑦 − 𝑥∞ | ∞ < 𝜖/2. Let
𝑦 𝑝 := 𝑥 𝑝 − 𝑦. From case 2, we find 𝑦 0 ∈ Q such that | 𝑦 0 − 𝑦 𝑝 | 𝑝 < 𝜖 for all 𝑝 ∈ S. By
case 1, there exist infinitely many 𝑚 ∈ Z such that |1 − 𝑚| 𝑝 < min(1, | 𝑦 0 − 𝑦 𝑝 | 𝑝 ) for
all 𝑝 ∈ S; for such 𝑚, we have 𝑚 ≡ 1 (mod 𝑝) so |𝑚| 𝑝 = 1 and
| 𝑦 0 − 𝑚𝑦 𝑝 | 𝑝
| (𝑦 0/𝑚) − 𝑦 𝑝 | 𝑝 = = | 𝑦 0 − 𝑦 𝑝 + (1 − 𝑚)𝑦 𝑝 | 𝑝 = | 𝑦 0 − 𝑦 𝑝 | 𝑝 < 𝜖 (28.1.5)
|𝑚| 𝑝
by the ultrametric inequality. By choosing 𝑚 large enough, we may ensure that

| 𝑦 0/𝑚| ∞ < 𝜖/2. Let 𝑥 := 𝑦 0/𝑚 + 𝑦. Then by (28.1.5)
|𝑥 − 𝑥 𝑝 | 𝑝 = | (𝑦 0/𝑚) + 𝑦 − (𝑦 𝑝 + 𝑦)| 𝑝 = | (𝑦 0/𝑚) − 𝑦 𝑝 | 𝑝 < 𝜖 (28.1.6)
and
|𝑥 − 𝑥∞ | ∞ ≤ |𝑥 − 𝑦| ∞ + | 𝑦 − 𝑥∞ | ∞ = | 𝑦 0/𝑚| ∞ + 𝜖/2 < 𝜖 (28.1.7)
proving (28.1.3).
Theorem 28.1.8 (Strong approximation). Let S ⊆ Pl Q be a nonempty set of places.

Then the image of Q ↩→ Q6S := 0𝑣∉S Q 𝑣 is dense.
Î
28.1.9. Written out in the standard basis of open sets, strong approximation is equiv-
alent to: given a finite set T ⊆ Pl Q disjoint from S, elements 𝑥 𝑣 ∈ Q 𝑣 for 𝑣 ∈ T, and
𝜖 > 0, there exists 𝑥 ∈ Q such that |𝑥 − 𝑥 𝑣 | 𝑣 < 𝜖 for all 𝑣 ∈ T and 𝑥 ∈ Z 𝑝 for all
𝑝 ∉ S t T with 𝑝 ≠ ∞.
Weak approximation follows from strong approximation by forgetting S and weakly
approximating 𝑥 𝑣 for 𝑣 ∈ T. Indeed, the difference between the ‘weak’ and the
‘strong’ is meaningful here. In weak approximation, we satisfy only a finite number
of conditions, with no control over the rational number at places 𝑣 ∉ S. By contrast, in
strong approximation, the role of the set S is switched, and have specified conditions
28.2. ⊲ STRONG APPROXIMATION FOR SL2 (Q) 467
at all primes 𝑣 ∉ S, either by approximation at finitely many places in T as in 28.1.9

or by the assertion of integrality at the rest. (This results in some asymmetry in the
conclusion for the real place; we could restore this by defining Z∞ := (−1, 1), but this
would not add content as we could just include this interval in T.)
Proof of Theorem 28.1.8. We prove the statement in its formulation 28.1.9. Naturally,
we return to the proof of weak approximation. Without loss of generality (proving a
stronger statement), we may assume #S = 1.
If S = {∞}, we apply step 2 of weak approximation over the set T: the result
𝑥 = 𝑥 0/𝑑 already has 𝑥 ∈ Z𝑞 for 𝑞 ∉ T, since then 𝑞 - 𝑑.
So suppose S = {ℓ} with ℓ prime. We return to case 3. To define 𝑦, we note instead
that Z[1/𝑝] ⊆ R is dense, so we may take 𝑦 ∈ Z[1/ℓ], so in particular 𝑦 ∈ Z𝑞 for
𝑞 ≠ ℓ. We just showed that we may take 𝑦 0 ∈ Z𝑞 for 𝑞 ∉ T. And for the integers 𝑚, we
claim we may take 𝑚 = ℓ 𝑘 for 𝑘 ∈ Z ≥0 : indeed, we are applying case 1 (CRT) and, as
in 28.1.1, this asks for 𝑚 ≡ 1 (mod 𝑝 𝑛 𝑝 ) for 𝑝 ∈ T (with 𝑛 𝑝 ∈ Z ≥1 , and ℓ ≠ 𝑝), so
we just need to take 𝑘 to be a common multiple of the orders of ℓ ∈ (Z/𝑝 𝑛 𝑝 Z) × . With
this strengthening, we have 𝑥 = 𝑦 0/𝑚 + 𝑦 ∈ Z𝑞 for all 𝑞 ∉ S t T.
28.1.10. As is hopefully evident from the proof of strong approximation when S =

{∞}, aside from continuity of multiplication, the key statement needed was that the
map Z → Z/𝑚Z is surjective for all 𝑚 ∈ Z, as provided by the CRT. Or more zippily,
what is needed is that the image of Z ↩→ b
Z is dense.
28.1.11. In weak approximation, we can replace the additive group Q with with the
multiplicative group Q× : the image of Q× ↩→ 𝑣 ∈S Q×𝑣 is dense a fortiori.
Î
However, the embedding Q× ↩→ Q×6S is not dense: that is to say, we do not have
strong approximation for Q× . Indeed, taking S = {∞} we have Q× = Q b× ; and since
6S
Z× ∩ Q× = Z× = {±1}, the open set b
b Z× r {±1} is disjoint from Q× . In view of 28.1.10,
the problem is also indicated by the fact that Z× = {±1} does not surject onto (Z/𝑚Z) ×
for 𝑚 ≥ 7.
28.2 ⊲ Strong approximation for SL2 (Q)

We now consider approximation in the noncommutative context. For motivation in
this second phase of the introduction, we consider the simplest case where 𝐵 = M2 (Q)
and take S = {∞}, and Q6𝑆 = Q;b by analogy, this is like considering a noncommutative
generalization of the CRT.
28.2.1. Recall that 𝐵 𝑣 = M2 (Q 𝑣 ) ' Q4𝑣 has the coordinate topology (see section 13.5);
therefore weak and strong approximation for 𝐵 = M2 (Q) follow from these statements
for Q, and weak approximation for GL2 (Q) follows as the determinant is continuous.
28.2.2. We should not expect the embedding GL2 (Q) ↩→ GL2 ( Q) b to be dense any
more than it was for Q× = GL1 (Q), as in 28.1.11. In fact, we rediscover the same issue
by taking determinants: the map GL2 (Z) → GL2 (Z/𝑚Z) is not surjective, because
det(GL2 (Z)) = ±1 whereas det(GL2 (Z/𝑚Z)) = (Z/𝑚Z) × .
Once we restrict to the subgroup of determinant 1, we find a dense subgroup once

again.
Theorem 28.2.3. The image of SL2 (Q) ↩→ SL2 ( Q)

b is dense.
Theorem 28.2.3 is known as strong approximation for the group SL2 (Q). We
give a quick proof of Theorem 28.2.3 in two steps. In preparation, we recall from
27.2.6 that
b × = Q× b
Q Z×
(“denominators can be handled globally”) and prove an analogous decomposition.
GL2 ( Q)
b = GL2 (Q) GL2 (b
Z)
(28.2.5)
SL2 ( Q)
b = SL2 (Q) SL2 (b
Z).
Proof. We begin with the first statement. The inclusion GL2 (Q) GL2 (b Z) ⊆ GL2 ( Q)b
holds; we prove the other containment. Let b 𝛼 ∈ GL2 ( Q). Consider the collection of
b
lattices (𝐿 𝑝 ) 𝑝 with 𝐿 𝑝 = 𝛼 𝑝 Z2𝑝 ⊆ Q2𝑝 . Since 𝛼 𝑝 ∈ GL2 (Z 𝑝 ) for all but finitely many
𝑝, we have 𝐿 𝑝 = Z2𝑝 for all but finitely many 𝑝. By the local-global dictionary for
lattices (Theorem 9.4.9), there exists a unique lattice 𝐿 ⊆ Q2 whose completions are
𝐿 𝑝 . We now rephrase this adelically (and succinctly): letting b 𝐿=b 𝛼bZ2 ⊆ Q b2 , we take
𝐿 = 𝐿 ∩ Q . Choose a basis for 𝐿 and put the columns in a matrix 𝛼, so 𝐿 = 𝛼Z2 .
b 2
Then b 𝐿 = 𝛼b Z2 = b𝛼bZ2 , and there exists 𝛾 ∈ GL2 (bZ) such that b

𝛼 = 𝛼b 𝛾 . This completes
the inclusion.
To get down to SL2 , we take determinants. Let b 𝛼 ∈ SL2 ( Q)
b and write it as b 𝛼 = 𝛼b𝛾
with 𝛼 ∈ GL2 (Q) and b 𝛾 ∈ GL2 ( Z). Then
b
𝛼) = det(𝛼) det(b
1 = det(b 𝛾 ) ∈ Q× bZ× = Q b×

× × −1 0
but Q ∩Z = {±1}; multiplying both 𝛼, b
b 𝛾 by on the right and left respectively,
0 1
𝛾 ) = 1, i.e., 𝛼 ∈ SL2 (Q) and b
if necessary, we may take det(𝛼) = det(b 𝛾 ∈ SL2 (b Z).
Now for the slightly magical second step.
Theorem 28.2.6. The map
SL2 (Z) → SL2 (Z/𝑚Z)
is surjective for all 𝑚 ∈ Z; equivalently, the image of SL2 (Z) ↩→ SL2 (b

Z) is dense.
The statement is nontrivial: a matrix modulo 𝑚 can certainly be lifted to a matrix

in Z whose determinant will be congruent to 1 modulo 𝑚, but the hard part is to ensure
that the lifted matrix has determinant equal to 1.
28.2. ⊲ STRONG APPROXIMATION FOR SL2 (Q) 469
Proof of Theorem 28.2.6. Let 𝛼 ∈ M2 (Z) be such that 𝛼 maps to the desired matrix
in SL2 (Z/𝑚Z), so in particular det(𝛼) ≡ 1 (mod 𝑚). By the theory of elementary
divisors (Smith normal form), there exist matrices 𝜇, 𝜈 ∈ SL2 (Z) such
that 𝜇𝛼𝜈 is
𝑎 0
diagonal; so without loss of generality, we may suppose that 𝛼 = with 𝑎𝑏 ≡ 1
0 𝑏
(mod 𝑚). Let

0 𝑎 −(1 − 𝑎𝑏)
𝛼 = . (28.2.7)
1 − 𝑎𝑏 𝑏(2 − 𝑎𝑏)
Then 𝛼 0 ≡ 𝛼 (mod 𝑚) and
det(𝛼 0) = 𝑎𝑏(2 − 𝑎𝑏) + (1 − 𝑎𝑏) 2 = 1 (28.2.8)
so 𝛼 0 ∈ SL2 (Z), as claimed. (Compare Shimura [Shi71, Lemma 1.38].)
Remark 28.2.9. The proof of Theorem 28.2.6 extends in two ways. First, we can
replace Z with a PID and the same proof works. Second, arguing by induction, one
can show that the map SL𝑛 (Z) → SL𝑛 (Z/𝑚Z) is surjective for all 𝑛 ≥ 2 and 𝑚 ∈ Z.
We are now ready to prove strong approximation for SL2 (Q).
Proof of Theorem 28.2.3. Consider the closure of SL2 (Q) in SL2 ( Q)

b in the idelic
topology; we obtain a closed subgroup. Since SL2 (Q) ≥ SL2 (Z), by Theorem
28.2.6 the closure contains SL2 (b
Z), but then by Lemma 28.2.4, it contains all of
SL2 (Q) SL2 (b
Z) = SL2 ( Q)!
b Therefore SL2 (Q) ≤ SL2 ( Q)
b is dense.
With the preceding context, we are now ready to state a more general formulation
of strong approximation for indefinite quaternion algebras over Q. The following
theorem is a special case of Main Theorem 28.5.3.
Theorem 28.2.10 (Strong approximation). Let 𝐵 be an indefinite quaternion algebra

over Q. Then 𝐵1 is dense in 𝐵
b1 .
If 𝐵 is definite, then O1 is a finite group () and so cannot be dense in the infinite

idelic group 𝐵b1 . This theorem has the following important applications.
Theorem 28.2.11. Let O be an Eichler order in an indefinite quaternion algebra over

Q. Then the following statements hold.
(a) Every locally principal right O-ideal is in fact principal, i.e., # Cls O = 1.
(b) Every order O0 locally isomorphic to O is in fact isomorphic to O, i.e., # Typ O =
1.
(c) For any integer 𝑚, the reduction map O1 → (O/𝑚O) 1 is surjective.
Proof. Specialize Main Theorem 28.5.3, Corollary 28.5.6, and Corollary 28.5.14,
respectively, using 28.5.16.
28.3 Elementary matrices

Before embarking on our more general idelic quest, we pause to give a second proof
of strong approximation for SL2 using elementary matrices.
28.3.1. Let 𝑅 be a domain. An elementary matrix (or transvection) in SL𝑛 (𝑅) is a

matrix which differs from the identity in one off-diagonal entry; such a matrix acts by
an elementary row operation (add a multiple of a row to a different row) on the left and
by an elementary column
operation
on the right. For 𝑛 = 2, the elementary matrices
1 𝑏 1 0
are those of the form or with 𝑏, 𝑐 ∈ 𝐹.
0 1 𝑐 1
28.3.2. If 𝐹 is a field, then SL𝑛 (𝐹) is generated by elementary matrices by the theory
of echelon forms (Exercise 28.3).
Lemma 28.3.3. Let 𝑅 be a Euclidean domain. Then SL2 (𝑅) is generated by elemen-
tary matrices.
Proof. The calculation

1 1 1 0 1 1 0 1
= =: 𝜂
0 1 −1 1 0 1 −1 0
shows that 𝜂 is in the subgroup of elementary matrices.

We now follow the usual proof of the elementary divisor theorem. Let 𝛼 =
𝑎 𝑏
∈ SL2 (𝑅). First, suppose 𝑏 = 0. Then det(𝛼) = 𝑎𝑑 = 1; adding 𝑎 times the
𝑐 𝑑
second row to the first, we may suppose 𝑏 = 1; then multiplying by 𝜂 on the right we
may suppose 𝑎 = 1; elementary row and column operations then give 𝑏 = 𝑐 = 0, and
then 𝑑 = 1. Similarly, if 𝑎 = 0, multiplying by 𝜂 gives 𝑏 = 0 and we repeat.
So we may suppose 𝑎, 𝑏 ≠ 0. By the Euclidean algorithm under the norm 𝑁, there
exists 𝑞, 𝑟 ∈ 𝑅 such that 𝑎 = 𝑏𝑞 + 𝑟 and 𝑁 (𝑟) < 𝑁 (𝑏). Applying the elementary
matrix which adds −𝑞 times the second column to the first, we may suppose 𝑎 = 0 or
𝑁 (𝑎) < 𝑁 (𝑏). If 𝑎 = 0, we are done by the previous paragraph; otherwise, we multiply
on the right by 𝜂 which swaps columns, and repeat. Because 𝑁 takes nonnegative
integer values, this procedure terminates after finitely many steps.
Remark 28.3.4. Lemma 28.3.3 holds for general 𝑛 ≥ 2, and it follows from the above
by induction: see Exercise 28.4.
This theory of elementary matrices has the following striking consequence.
Proposition 28.3.5. Let 𝑅 be a Dedekind domain. Then for all ideals 𝔪 ⊆ 𝑅, the map
SL2 (𝑅) → SL2 (𝑅/𝔪)
is surjective.
28.4. STRONG APPROXIMATION AND THE IDEAL CLASS SET 471
Proof. We may suppose 𝔪 is nonzero. Then by the CRT, 𝑅/𝔪 is a finite product of local
Artinian principal ideal rings. Therefore 𝑅/𝔪 is Euclidean and by a generalization
of Lemma 28.3.3, the group SL2 (𝑅/𝔪) is generated by elementary matrices: see
Exercise 28.6. Every elementary matrix in SL2 (𝑅/𝔪) lifts to an elementary matrix in
SL2 (𝑅), and the statement follows.
Corollary 28.3.6. Let 𝐹 be a global field, and let 𝑅 ⊆ 𝐹 be a global ring with eligible
set S. Then the image of the map
Ö0
SL2 (𝐹) ↩→ SL2 (𝐹6𝑆 ) = SL2 (𝐹𝑣 )
𝑣∉S
is dense.
Proof. For brevity, we write 𝐹b = 𝐹6𝑆 and 𝑅b = Î 𝑣∉S 𝑅 𝑣 . We first show that
SL2 (𝑅) ↩→ SL2 ( 𝑅)
b is dense. If 𝑈 ⊆ SL2 ( 𝑅)
b is open, then 𝑈 contains a standard
open neighborhood of the form
{ 𝛽b ∈ SL2 ( 𝑅)
b : 𝛽b ≡ 𝛼𝔪 (mod 𝔪)}
for some 𝛼𝔪 ∈ SL2 (𝑅/𝔪) and 𝔪 ⊆ 𝑅. The surjectivity in Proposition 28.3.5 then
implies that 𝑈 ∩ SL2 (𝑅) ≠ ∅.
For the statement itself, we again argue with elementary matrices. Let b 𝛼 = (𝛼 𝑣 ) 𝑣 ∈
SL2 ( 𝐹);
b then 𝛼 𝑣 ∈ SL2 (𝑅 𝑣 ) for all but finitely many 𝑣. For these finitely many 𝑣, we
know SL2 (𝐹𝑣 ) is generated by elementary matrices by Lemma 28.3.2, so by strong
approximation in 𝐹 (Lemma 28.7.2) we can approximate 𝛼 𝑣 by an element of SL2 (𝐹)
that belongs to any open neighborhood of 𝛼 𝑣 ; for the remaining places we apply the
previous paragraph, and we finish using the continuity of multiplication.
28.4 Strong approximation and the ideal class set

In this section, we provide one more motivation for strong approximation, relating it
to the ideal class set as previewed in Eichler’s theorem (see section 17.8).
We adopt the following notation for the rest of this chapter. Let 𝑅 be a global ring
with eligible set S and 𝐹 = Frac 𝑅 its global field. Let 𝐵 be a quaternion algebra over
𝐹 and let O ⊆ 𝐵 be an 𝑅-order.
28.4.1. By 27.7.1, the reduced norm map
nrd : Cls O = 𝐵× \ 𝐵
b×/O
b× → 𝐹 × \𝐹
>Ω 0
b×/nrd( O
b× ) (28.4.2)
is surjective. Then by class field theory 27.7.4, the codomain 𝐹>×Ω 0 \𝐹

b×/nrd( O
b× ) =
Cl𝐺 (O) 𝑅 admits a description as a class group.
The important point that we will soon see: the reduced norm map is injective,
therefore bĳective, when strong approximation holds for the group 𝐵1 . But before we
get there, we have some explaining to do.
We now investigate the injectivity of the reduced norm map (28.4.2). This map is
only a map of (pointed) sets, so first we show that it suffices to look at an appropriate
kernel.
b× , the map b
28.4.3. For all 𝛽b ∈ 𝐵 𝛼O ↦→ b
𝛼O 𝛽b gives a bĳection
Cls O = 𝐵× \ 𝐵
b× /O
b× ↔ 𝐵× \ 𝐵
b× /O
b0× = Cls O0
where O0 := 𝐵 ∩ 𝛽b−1 O b 𝛽b is connected (locally isomorphic) to O. So it is sensible

to consider the maps (28.4.2) for all orders O0 connected to O, i.e., the entire genus
Gen O. We recall that the type set is finite (Corollary 27.6.25, or Main Theorem 17.7.1
in the number field case using the geometry of numbers).
Our investigations will involve the kernels of the reduced norm maps:
𝐵1 := {𝛼 ∈ 𝐵× : nrd(𝛼) = 1} ≤ 𝐵
b1 := {b b× : nrd(b
𝛼∈𝐵 𝛼) = 1} (28.4.4)
Example 28.4.5. If 𝐵 = (𝑎, 𝑏 | 𝐹), then 𝐵1 admits the Diophantine description
𝐵1 ' {(𝑥, 𝑦, 𝑧, 𝑤) ∈ 𝐹 4 : 𝑥 2 − 𝑎𝑦 2 − 𝑏𝑧 2 + 𝑎𝑏𝑤 2 = 1},
and the group 𝐵b1 consists of local solutions at all primes that belong to 𝑅 4𝑣 for almost
all places 𝑣 ∈ Pl 𝐹.
Lemma 28.4.6. Let O ⊆ 𝐵 be an 𝑅-order. Then the reduced norm map (28.4.2) is
injective for all orders O0 ∈ Gen O if and only if 𝐵
b1 ⊆ 𝐵× O
b0× for all O0 ∈ Gen O.
Proof. If (28.4.2) is injective, then given b 𝛼 ∈ 𝐵 b1 we have nrd(b b× ) = nrd( O

𝛼O b× ) so
× ×
𝛼O = 𝑧O for some 𝑧 ∈ 𝐵 and b
b b b × ×
𝛼 ∈ 𝑧O ⊆ 𝐵 O .
b × b×
For the converse, since nrd : 𝐵× → 𝐹>×Ω 0 and nrd : O b× → nrd( O b× ) are both
surjective, to show nrd is injective for O we may show that if nrd(b 𝛼) = nrd( 𝛽)b ∈𝐹 b×
then b × × ×
𝛼O = 𝑧 𝛽bO for some 𝑧 ∈ 𝐵 . We consider (b
b b −1 −1
𝛼 𝛽b ) ( 𝛽bO 𝛽b ) = (b
b −1 0
𝛼 𝛽b ) O where
b
as above O0 = 𝐵 ∩ 𝛽bO b 𝛽b−1 ∈ Gen O. Since b 𝛼 𝛽b−1 ∈ 𝐵b1 , by hypothesis b𝛼 𝛽b−1 = 𝑧b𝜇0 =
b𝜇 𝛽b−1 ) where 𝑧 ∈ 𝐵× and 𝜇
𝑧( 𝛽b b∈ O b× , and consequently b 𝛼O b𝜇O
b = 𝑧 𝛽b b = 𝑧 𝛽bO,b and
hence the map is injective.
28.4.7. We have 𝐵× O
b× ∩ 𝐵 b1 if and only if nrd(O× ) = 𝐹 × ∩ nrd( O
b1 = 𝐵1 O
>Ω 0
b× )
(Exercise 28.10).
28.4.8. Suppose that 𝐵1 is dense in 𝐵b1 . Then we claim that 𝐵

b1 ⊆ 𝐵1 Ob1 ⊆ 𝐵× O b× for
all orders O. 𝛼 = (𝛼𝔭 )𝔭 ∈ 𝐵
b Indeed, if b 1
b then b 𝛼O 1
b ≤𝐵 b is open, so there exists 𝛼 ∈ 𝐵1
1
such that 𝛼 = b𝛾 with b

𝛼b 𝛾∈O b1 , and b
𝛼 = 𝛼b𝛾 −1 ∈ 𝐵1 O
b1 .
We should not expect hypothesis of 28.4.8 to hold for all quaternion algebras: see
Exercise 28.7.
28.5. STATEMENT AND FIRST APPLICATIONS 473
28.5 Statement and first applications

In this section, we set up and state the strong approximation theorem, and then derive
some applications. Throughout, we abbreviate 𝐵6S = 𝐵. b
Definition 28.5.1. We say 𝐵 is S-indefinite (or 𝐵 satisfies the S-Eichler condition)

if S contains a place which is unramified in 𝐵.
28.5.2. If 𝐹 is a number field, then this definition agrees with Definition 17.8.1; and
since a complex place is necessarily split and S contains the archimedean places, if 𝐵
is S-definite over a number field 𝐹 then 𝐹 is a totally real number field.
Main Theorem 28.5.3 (Strong approximation). Let 𝐵 be a quaternion algebra over a

global field and suppose 𝐵 is S-indefinite. Then 𝐵1 is dense in 𝐵
b1 .
28.5.4. One can think of strong approximation from the following informal perspective:
if 𝐵1𝑣 is not compact, then there is enough room for 𝐵1 to “spread out” in 𝐵1𝑣 so that
correspondingly 𝐵1 is dense in the S-finite part 𝐵b1 .
1 Î 1
The hypothesis that 𝐵S = 𝑣 ∈S 𝐵 𝑣 is noncompact is certainly necessary for the
conclusion that 𝐵1 is dense in 𝐵 b1 . Indeed, if 𝐵S1 = Î 𝑣 ∈S 𝐵1𝑣 is compact, then since
𝐵1 is discrete in 𝐵1 , the subgroup 𝐵1 𝐵S1 ≤ 𝐵1 is closed in 𝐵1 , and 𝐵1 𝐵S1 ≠ 𝐵1 . On
the other hand, if 𝐵1 is dense in 𝐵b1 , then adding the components for 𝑣 ∈ S we have
𝐵1 𝐵S1 ≤ 𝐵1 dense. This is a contradiction.
We give two proofs of strong approximation over the next two sections. For the
moment, we consider some applications.
Our main motivation for strong approximation is the following proposition. We
recall the class group 27.7.4 associated to O.
Theorem 28.5.5. If 𝐵 is S-indefinite, then the reduced norm map (28.4.2)
nrd : Cls O = 𝐵× \ 𝐵
b× /O
b× → Cl𝐺 (O) 𝑅 = 𝐹 × \𝐹
>Ω 0
b× /nrd(O× )
is a bĳection for all 𝑅-orders O ⊆ 𝐵: in particular, if 𝐼 is a locally principal right

O-ideal, then 𝐼 is principal if and only if nrd(𝐼) is principal in the class group Cl𝐺 (O) 𝑅.
Proof. Combine Lemma 28.4.6 and 28.4.8.
Corollary 28.5.6. If 𝐵 is S-indefinite and Cl𝐺 (O) 𝑅 is trivial, then Typ O is trivial,
i.e., every order O0 locally isomorphic to O is in fact isomorphic to O.
Proof. The class set Cls O maps surjectively onto Typ O by Lemma 17.4.13, and the
latter is trivial by Theorem 28.5.5.
28.5.7. More generally, we can grapple with the type set of O, measured by a different
(generalized) class group. Recall (27.6.24) that
Typ O ↔ 𝐵× \ 𝐵
b× /𝑁 b× ( O).
𝐵
b
Let
𝐺 𝑁 (O) := 𝐹>×Ω 0 nrd(𝑁 𝐵b× ( O))
b ≤𝐹b× . (28.5.8)
b× ≤ 𝑁 b× ( O),
Since O b we have 𝐺 𝑁 (O) ≥ 𝐺 (O). Define accordingly the class group
𝐵
b× /𝐺 𝑁 (O).
Cl𝐺 𝑁 (O) 𝑅 = 𝐹 (28.5.9)
Then there is a surjective map of abelian groups
Cl𝐺 (O) → Cl𝐺 𝑁 (O) 𝑅.
Corollary 28.5.10. If 𝐵 is S-indefinite, then the reduced norm map induces a bĳection
∼ Cl
nrd : Typ O −→ 𝐺 𝑁 (O) 𝑅.
Proof. We take the further quotient by the normalizer in the bĳection in Theorem
28.5.5.
28.5.11. Returning to 17.4.16, for 𝐵 = M2 (𝐹) and O = M2 (𝑅) we compute that

nrd(𝑁 𝐵b× ( O))
b =𝐹 b2 𝑅b× , so Cl𝐺 𝑁 (O) 𝑅 = Cl 𝑅/(Cl 𝑅) 2 . Thus by Corollary 28.5.10, the

𝑅 𝔞
types of maximal orders in O are given by −1 for [𝔞] in a set of representatives
𝔞 𝑅
of (Cl 𝑅)/(Cl 𝑅) 2 .
Two other immediate applications of strong approximation that served as motiva-
tion are now apparent.
Corollary 28.5.12. Suppose 𝐵 is S-indefinite. Then
b1 = 𝐵1 O
𝐵 b1 and 𝐵 1 = 𝐵 1 O1 . (28.5.13)
Proof. The inclusion 𝐵1 Ob1 ⊆ 𝐵b1 holds, and the converse holds when 𝐵1 is dense in
b1 by 28.4.8. For the second statement, we have 𝐵 = 𝐵
𝐵 b × 𝐵S and O = O
b × 𝐵S , so we
1
take norm 1 units and multiply both sides of (28.5.13) by 𝐵S = 𝑣 ∈S 𝐵1𝑣 .
Î

Corollary 28.5.14. Suppose 𝐵 is S-indefinite. Let 𝔪 ⊆ 𝑅 be an ideal. Then the

reduction map
O1 → (O/𝔪O) 1
is surjective. Moreover, O1 is dense in O
b1 .
Proof. For all 𝛼𝔪 ∈ (O/𝔪O) 1 , by strong approximation the open set

b1 : 𝛽b ≡ 𝛼𝔪 (mod 𝔪O)} ⊆ O
{ 𝛽b ∈ O b1
contains an element 𝛼 ∈ 𝐵1 ∩ O b1 = O1 mapping to 𝛼𝔪 in the reduction map. The

second statement follows as above from the definition of the topology.
We now give a name to a large classes of orders where the group 𝐺 (O) governing
principality is explicitly given.
28.5. STATEMENT AND FIRST APPLICATIONS 475
Definition 28.5.15. We say that an 𝑅-order O ⊆ 𝐵 is locally norm-maximal if

b× ) = 𝑅
nrd( O b× .
Equivalently, O is locally norm-maximal if and only if the reduced norm maps

nrd : O×𝔭 → 𝑅𝔭× are surjective for all nonzero primes 𝔭 of 𝑅.
Example 28.5.16. If O is maximal, then O is locally norm-maximal (Lemma 13.4.9);

more generally if O is Eichler then O is locally norm-maximal (Exercise 23.3).
Certain special cases of Theorem 28.5.5 are important in applications. Recall that
Ω ⊆ Ram 𝐵 is the set of real ramified places, and ClΩ 𝑅 as defined in 17.8.2 is class
group associated to Ω, a quotient of the narrow class group.
Corollary 28.5.17. Suppose 𝐹 is a number field and let S be the set of archimedean
places of 𝐹. Suppose 𝐵 is S-indefinite and O ⊆ 𝐵 is locally norm-maximal 𝑅-order.
Then nrd : Cls O → ClΩ 𝑅 is a bĳection.
Proof. This is just a restatement of Theorem 28.5.5 once we note that Cl𝐺 (O) 𝑅 =
ClΩ 𝑅 by Example 27.7.7.
Proposition 28.5.18. Let T ⊇ S be a set of primes of 𝑅 that generate Cl𝐺 (O) 𝑅 and
suppose 𝐵 is T-indefinite. Then every class in Cls O contains an integral (invertible
right) O-ideal whose reduced norm is supported in the set T.
Proof. Let 𝑅 (T) denote the (further) localization of 𝑅 at the primes in T. We apply
Theorem 28.5.5 to the order O (T) := O ⊗𝑅 𝑅 (T) : we conclude that there is a bĳection
Cls O (T) −→∼ Cl
𝐺 (O (T ) ) 𝑅 (T) . But Cl𝐺 (O (T ) ) 𝑅 (T) is the quotient of Cl𝐺 (O) 𝑅 by the
primes in T, and so by hypothesis is trivial. Therefore if 𝐼 is a right O-ideal, then
𝐼 (T) := 𝐼 ⊗𝑅 𝑅 (T) has 𝐼 (T) = 𝛼O (T) for some 𝛼 ∈ 𝐵× . Let 𝐽 = 𝛼−1 𝐼. Then [𝐽] R = [𝐼] R
and 𝐽𝔭 = O𝔭 for all primes 𝔭 ∉ T and so 𝐽 has reduced norm supported in 𝑇. Replacing
𝐽 by 𝑎𝐽 with 𝑎 ∈ 𝑅 nonzero and supported in T, we may suppose further that 𝐽 ⊆ O
is integral, and the result follows.
Example 28.5.19. Let 𝐵 be a definite quaternion algebra over a totally real (number)
field 𝐹 and let S be the set of archimedean places, so 𝑅 = Z𝐹 . Let O be a locally
norm-maximal 𝑅-order in 𝐵. Suppose that ClΩ 𝑅 = {1} and let 𝔭 ⊆ 𝑅 be a prime of
𝑅 unramified in 𝐵. Then by Proposition 28.5.18, every ideal class in Cls O contains
an integral O-ideal whose reduced norm is a power of 𝔭.
As a special case, we may take 𝐹 = Q: then ClΩ Z = Cl+ Z = {1}. Therefore,
if 𝐵 is a definite quaternion algebra of discriminant 𝐷 over Q, and O ⊆ 𝐵 a locally
norm-maximal order (e.g., an Eichler order), then for a prime 𝑝 - 𝐷, every invertible
right O-ideal class is represented by an integral ideal whose reduced norm is a power
of 𝑝.
28.6 Further applications

We continue with further applications of strong approximation.
Our next consequence of strong approximation is a refinement the Hasse–Schilling
theorem on norms (Main Theorem 14.7.4) as follows.
Theorem 28.6.1 (Eichler’s theorem on norms). Suppose 𝐵 is S-indefinite, and let

𝑛 ∈ 𝑅 ∩ 𝐹>×Ω 0 . Then there exists 𝛼 ∈ 𝐵× integral over 𝑅 such that nrd(𝛼) = 𝑛.
Proof. By Main Theorem 14.7.4, there exists 𝛼 ∈ 𝐵× such that nrd(𝛼) = 𝑛. For each
prime 𝔭, the set
𝑈𝔭 = {𝛽𝔭 ∈ 𝐵𝔭1 : trd(𝛽𝔭 𝛼) ∈ 𝑅𝔭 } (28.6.2)
is (closed and) open since trd is continuous, and further 𝑈𝔭 is nonempty: if 𝔭 ∈ Ram 𝐵
then already 𝛼𝔭 is integral
over 𝑅𝔭 and 1 ∈ 𝑈𝔭 , and otherwise 𝐵𝔭 ' M2 (𝐹𝔭 ) and we
0 −𝑛
may suppose 𝛼𝔭 = is in rational canonical form whereby
1 𝑡

0 1 0 −𝑛 1 𝑡
trd = trd = 𝑛 + 1 ∈ 𝑅𝔭 (28.6.3)
−1 0 1 𝑡 0 𝑛

0 1
shows 𝛽𝔭 = ∈ 𝑈𝔭 .
−1 0
b1
Î
b :=
Let 𝑈 𝔭 𝑈𝔭 ∩ 𝐵 ; then 𝑈 is open and nonempty. By strong approximation,
b
there exists 𝛽 ∈ 𝑈 1
b ∩ 𝐵 . Thus trd(𝛽𝛼) ∈ Ñ𝔭 𝑅𝔭 = 𝑅 and nrd(𝛽𝛼) = nrd(𝛼) = 𝑛.
Therefore 𝛽𝛼 is as desired.
Let 𝑅>× Ω 0 := 𝑅 × ∩ 𝐹>×Ω 0 be the subgroup of units that are positive at the places
𝑣 ∈ Ω (the set of real, ramified places in 𝐵).
Corollary 28.6.4. Suppose 𝐵 is S-indefinite and that ClΩ 𝑅 is trivial. Let O ⊆ 𝐵 be

an Eichler 𝑅-order. Then
nrd(O× ) = 𝑅>× Ω 0 .
Proof. Let 𝑢 ∈ 𝑅>× Ω 0 . We repeat the argument of Theorem 28.6.1, with a slight
modification. Let 𝔐 be the level of O.
Let 𝔭 | 𝔐 be a prime that divides the level 𝔐. We choose an isomorphism
𝐵𝔭 ' M2 (𝐹𝔭 ) such that 𝛼𝔭 is in rational canonical form, and let O𝔭0 be the standard
Eichler order in M2 (𝐹𝔭 ) of the same level as O. Define
𝑈𝔭 = {𝛽𝔭 ∈ 𝐵𝔭1 : 𝛽𝔭 𝛼 ∈ O𝔭0 }.
This is again an open condition because multiplication is continuous, and the calcula-
tion −1
𝑡𝑢 1 0 −𝑢 1 0
=
−1 0 1 𝑡 0 𝑢
shows also that 𝑈𝔭 ≠ ∅.
28.6. FURTHER APPLICATIONS 477
For all other primes 𝔭 - 𝔐, we define 𝑈𝔭 as in (28.6.2). As in the proof of Theorem

28.6.1, we find 𝛽 ∈ 𝐵1 such that nrd(𝛽𝛼) = 𝑢 and 𝛾 0 = 𝛽𝛼 has 𝛾𝔭0 ∈ O𝔭0 for all 𝔭 | 𝔐.
Since 𝛾 0 is integral, 𝛾 0 belongs to an 𝑅-order O0 that is equal to O𝔭0 at all 𝔭 | 𝔐 and is
maximal at all 𝔭 - 𝔐. The order O0 is therefore an Eichler order of level 𝔐.
Finally, since Cl𝐺 (O) 𝑅 = ClΩ 𝑅 is trivial, the type set Typ(O) is also trivial
(Corollary 28.5.6), so we conclude O = 𝜈O0 𝜈 −1 for some 𝜈 ∈ 𝐵× , and 𝛾 = 𝜈𝛾 0 𝜈 −1 ∈ O
has nrd(𝛾) = 𝑢 as desired.
Example 28.6.5. Suppose O is an Eichler order in an indefinite quaternion algebra
over 𝐹 = Q. Then by Corollary 28.6.4, nrd(O× ) = {±1}, in particular, there exists
𝛾 ∈ O× such that nrd(𝛾) = −1.
Remark 28.6.6. We will prove a stronger version of Corollary 28.6.4 after we have
developed the theory of selectivity: see Corollary 31.1.11.
To conclude this section, we consider a variant of principalization of right fractional
ideals: we ask further that the generator has totally positive reduced norm.
28.6.7. Suppose 𝐹 is a number field. Let
𝐵×>0 = {𝛼 ∈ 𝐵× : 𝑣(nrd(𝛼)) > 0 for all real places 𝑣}.
The reduced norm gives a map 𝐵× /𝐵×>0 → 𝐹>×Ω 0 /𝐹>0 × and the quotient is a finite
abelian 2-group, so in particular 𝐵×>0 ≤ 𝐵× has finite index.

Let 𝐼, 𝐽 ⊆ 𝐵 be 𝑅-lattices. We say 𝐼, 𝐽 are in the same narrow right class if there
exists 𝛼 ∈ 𝐵×>0 such that 𝛼𝐼 = 𝐽; accordingly, we let Cls+R O be the narrow (right)
class set of O. As in Lemma 27.6.8, there is a bĳection
b×/O
Cls+R O ↔ 𝐵×>0 \ 𝐵 b×
choosing a local generator. The projection map Cls+R O → ClsR O has finite fibers as
𝐵×>0 ≤ 𝐵× has finite index, so since ClsR O is a finite set, so too is Cls+R O.
Corollary 28.6.8. Let 𝐹 be a number field and suppose 𝐵 is S-definite. Then the map
nrd : Cls+R O ↔ 𝐵×>0 \ 𝐵

b×/O
b× → 𝐹 × \𝐹
>0
b×/nrd( O
b× ) =: Cl+
𝐺 (O) 𝑅 (28.6.9)
induced by the reduced norm is a bĳection.
Proof. Repeating the argument in the proof of Lemma 28.4.6, we see that the map
(28.6.9) is injective for all orders O0 ∈ Gen O if and only if 𝐵
b1 ⊆ 𝐵× O
>0
b0× for all
0
O ∈ Gen O. And the latter holds by strong approximation (Corollary 28.5.12):
b01 ⊆ 𝐵× O
b1 ⊆ 𝐵1 O
𝐵 b0×
>0
for all 𝑅-orders O0.

Proposition 28.6.10. Let 𝐹 be a number field. Suppose that 𝐵 is S-indefinite and that
Cl+𝐺 (O) 𝑅 = Cl𝐺 (O) 𝑅. For each real place 𝑣 not in Ω, let 𝜖 𝑣 ∈ {±1}. Then there exists
𝛾 ∈ O× such that sgn(𝑣(nrd(𝛾))) = 𝜖 𝑣 for all 𝑣 real not in Ω.
In particular, if 𝐹 = Q and O is a locally norm-maximal Z-order in an indefinite
quaternion algebra 𝐵, then there exists 𝛾 ∈ O× with nrd(𝛾) = −1.
Proof. Let 𝑎 ∈ 𝐹>×Ω 0 be such that 𝑣(𝑎) = 𝜖 𝑣 for all 𝑣 real not in Ω and the class of
𝑎𝑅 is trivial in Cl𝐺 (O) 𝑅: these constraints together impose congruence conditions on
elements in a real cone. By the Hasse–Schilling norm theorem, there exists 𝛼 ∈ 𝐵×
such that nrd(𝛼) = 𝑎. Thus the class of nrd(𝛼O) in Cl+𝐺 (O) 𝑅 = Cl𝐺 (O) 𝑅 is trivial.
But then by (28.6.9) (a consequence of strong approximation), there exists 𝛽 ∈ 𝐵×>0
such that 𝛽O = 𝛼O, and therefore 𝛽 = 𝛼𝛾 with 𝛾 ∈ O× . Since 𝛽 is totally positive, for
all real places 𝑣 ∉ Ω we have sgn(𝑣(nrd(𝛾))) = sgn(𝑣(nrd(𝛼))) = 𝜖 𝑣 , completing the
proof.
For the second statement, we only need to note that Cl+ Z = Cl Z = {1} and recall
that nrd(O× ) ≤ {±1}.
Remark 28.6.11. More generally, let 𝐵 be a central simple algebra over the global field
𝐹. We say 𝐵 satisfies the S-Eichler condition if there exists a place 𝑣 ∈ S such that
𝐵 𝑣 is not a division algebra. (In this text, for quaternion algebras we prefer to use the
term S-indefinite because it readily conveys the notion, but both terms are common.)
If 𝐹 is a number field and S is the set of archimedean places of 𝐹, then 𝐵 satisfies
the S-Eichler condition if and only if 𝐵 is not a totally definite quaternion algebra
(Exercise 28.5). So the condition is a mild condition, and the quaternion algebra case
requires special effort.
When 𝐵 satisfies the S-Eichler condition, then 𝐵1 ↩→ 𝐵 b1 is dense, and for O ⊆ 𝐵
a maximal 𝑅-order, a locally principal right fractional O-ideal 𝐼 ⊆ 𝐵 is principal if
and only if its reduced norm nrd(𝐼) ⊆ 𝑅 is trivial in ClΩ 𝑅, where Ω is the set of real
places 𝑣 ∈ Pl 𝐹 such that 𝐵 𝑣 ' M𝑛 (H), generalizing the quaternion case.
Eichler proved Theorem 28.5.5 and the more general statement of the previous
paragraph [Eic37, Satz 2], also providing several reformulations and applications
[Eic38a, Eic38c]. For a full exposition, see Reiner [Rei2003, §34].
28.7 First proof

Now we proceed with the proof of strong approximation in Theorem 28.5.3; we follow
roughly the same lines as in the proof of Eichler’s theorem on norms, but here instead
we will be concerned with traces. In other words, we replace strong approximation of
elements by strong approximation of traces, and then we just have to chase conjugacy
classes.
We start with a statement of weak and strong approximation for the global field 𝐹.
Lemma 28.7.1 (Weak approximation for 𝐹). Let S ⊆ Pl 𝐹 be a finite nonempty set of
places. Then the images of the maps
Ö Ö
𝐹 ↩→ 𝐹S := 𝐹𝑣 and 𝐹 × ↩→ 𝐹S× := 𝐹𝑣×
𝑣 ∈S 𝑣 ∈S
are dense.
Proof. See e.g. Neukirch [Neu99, Theorem II.3.4] or O’Meara [O’Me73, §11E].
Lemma 28.7.2 (Strong approximation forÎ𝐹). Let S ⊆ Pl 𝐹 be a finite nonempty set
of places. Then the image of 𝐹 ↩→ 𝐹6S := 𝑣∉S 𝐹𝑣 is dense.
28.7. FIRST PROOF 479
Proof. See e.g. Neukirch [Neu99, Exercise III.1.1] or O’Meara [O’Me73, §33G].
Î
We recall that 𝐵 is a quaternion algebra over 𝐹 and we write 𝐵S := 𝑣 ∈S 𝐵𝑣 .
Proposition 28.7.3 (Weak approximation for 𝐵). Let S ⊆ Pl 𝐹 be a finite nonempty

set of places. Then the images
𝐵 ↩→ 𝐵S and 𝐵× ↩→ 𝐵S× and 𝐵1 ↩→ 𝐵S1
are dense.
Î
Proof. By weak approximation for 𝐹 (Lemma 28.7.1), we have 𝐹 dense in 𝑣 ∈S 𝐹𝑣 .
Choosing an 𝐹-basis for 𝐵, we have 𝐵 ' 𝐹 4 as topological 𝐹-vector Î spaces, and so
by approximating in each coordinate, we conclude that 𝐵 is dense in 𝑣 ∈S 𝐵 𝑣 . The
multiplicative case follows a fortiori, restricting open neighborhoods.
Finally we treat 𝐵1 . By Exercise 7.30, we know that 𝐵1 = [𝐵× , 𝐵× ] is the
commutator. Let (𝛾 𝑣 ) 𝑣 ∈
Î 1
𝑣 ∈S 𝐵 𝑣 . Then for each 𝑣 ∈ S, we can write 𝛾 𝑣 =
𝛼 𝑣 𝛽 𝑣 𝛼 𝑣 𝛽 𝑣 with 𝛼 𝑣 , 𝛽 𝑣 ∈ 𝐵 . Then by weak approximation for 𝐵× , we can find a
−1 −1 ×
sequence 𝛼𝑛 ∈ 𝐵× such that 𝛼𝑛 → (𝛼 𝑣 ) 𝑣 ∈ 𝐵S× , and similarly with 𝛽𝑛 → (𝛽 𝑣 ) 𝑣 . Then

since multiplication is continuous, we conclude that 𝛾𝑛 = 𝛼𝑛 𝛽𝑛 𝛼𝑛−1 𝛽−1𝑛 → (𝛾 𝑣 ) 𝑣 ∈
𝐵1 .
Next, we need to approximate polynomials: this kind of lemma was first performed
in section 14.7 to prove the Hasse–Schilling theorem of norms, and here we need
another variant.
Lemma 28.7.4. Let S ⊆ Pl 𝐹 be a finite nonempty set of places and Σ ⊆ Pl 𝐹 a finite

set of noncomplex places disjoint from S. Let b 𝑡∈𝐵b be such that 𝑥 2 − 𝑡 𝑣 𝑥 + 1 ∈ 𝐹𝑣 [𝑥]
is irreducible for 𝑣 ∈ Σ.
Let 𝜖 > 0. Then there exists 𝑡 ∈ 𝐹 such that:
• 𝑓 (𝑥) = 𝑥 2 − 𝑡𝑥 + 1 is irreducible and separable over 𝐹;

• |𝑡 − 𝑡 𝑣 | 𝑣 < 𝜖 for all 𝑣 ∈ Σ;
• 𝑓 (𝑥) is irreducible over 𝐹𝑣 for all 𝑣 ∈ Σ; and
• |𝑡 − 𝑡 𝑣 | 𝑣 ≤ 1 for all 𝑣 ∉ S ∪ Σ.
Proof. We argue as in Lemma 14.7.6 (and Corollary 14.7.8), but instead of weak
approximation we now use strong approximation (Lemma 28.7.2). Our job is a bit
easier because we are only asking for irreducibility, not separability.
Since S is nonempty, by strong approximation for 𝐹 we can find 𝑡 arbitrarily close
𝑡, thus ensuring that the desired inequalities hold and that 𝑓 (𝑥) is irreducible over
to b
𝐹𝑣 for 𝑣 ∈ Σ; to ensure that 𝑓 (𝑥) is separable, we need only avoid the locus 𝑡 2 = 4,
and similarly we may ensure 𝑓 (𝑥) is irreducible.
We now embark on the proof of strong approximation.

Proof of Main Theorem 28.5.3. We follow Vignéras [Vig80a, Théorème III.4.3]; see
also Miyake [Miy2006, Theorems 5.2.9–5.2.10] for the case 𝐹 = Q. We show that the
closure of 𝐵1 is equal to 𝐵
b1 . Let b b1 ; we will find a sequence of elements
𝛾 = (𝛾 𝑣 ) 𝑣 ∈ 𝐵
1
of 𝐵 converging to b 𝛾.
Step 1: Setup. We claim it is enough to consider the case where 𝛾 𝑣 = 1 for all but
finitely many 𝑣, by a Cantor-style diagonalization argument. Indeed, for a finite set
T ⊆ Pl 𝐹 disjoint from S, we let b 𝛾 [T ] be the idele obtained from b 𝛾 replacing 𝛾 𝑣 = 1
for 𝑣 ∉ T. Then for a sequence of subsets T eventually containing each place 𝑣, we
have b 𝛾 . Thus, if we can find sequences in 𝐵1 converging to each b
𝛾 [T ] → b 𝛾 [T ] we can
diagonalize to find a sequence converging to b 𝛾 , since Pl 𝐹 is countable.
So we may suppose without loss of generality that 𝛾 𝑣 = 1 for all but finitely many
𝑣. To find a sequence in 𝐵1 converging to b 𝛾 , our strategy in the proof is as follows:
in shrinking open neighborhoods of b 𝛾 we first find an element in 𝐵1 whose reduced
characteristic polynomial is close to an element in the open neighborhood, and then
we conjugate by 𝐵× to get the limits themselves to match.
To this end, let O ⊂ 𝐵 be a reference 𝑅-order, let T ⊆ Pl 𝐹 be a finite set of places
disjoint from S containing the primes 𝑣 where 𝛾 𝑣 ≠ 1 and the ramified places of 𝐵 not
in S. We consider open neighborhoods of the form
Ö Ö
𝑈= 𝛾𝑣 𝑈𝑣 × O1𝑣
𝑣 ∈T 𝑣∉S∪T
where 𝑈 𝑣 ⊆ 𝐵1𝑣 is an open neighborhood of 1.

Step 2: Polynomial approximation. Now comes the polynomial approximation step:
we will show that there exists b𝛼∈ 𝐵b× such that 𝐵1 ∩ b𝛼−1𝑈b𝛼 ≠ ∅. (This is about as
good as could be expected at this stage: if we argue by approximating polynomials,
we should only be able to expect to get something up to conjugation.)
We define the idele b
𝑡 in each component, as follows.
• If 𝑣 ∉ S ∪ T, we take 𝑡 𝑣 = trd(𝛾 𝑣 ) = 2.
• If 𝑣 ∈ T is unramified in 𝐵, we take 𝑡 𝑣 = trd(𝛾 𝑣 ).
• If 𝑣 ∈ T is ramified in 𝐵, we choose 𝜇 𝑣 ∈ 𝛾 𝑣 𝑈 𝑣 such that 𝜇 𝑣 ∉ 𝐹𝑣 is separable
over 𝐹𝑣 , and take 𝑡 𝑣 = trd(𝜇 𝑣 ). Since 𝐵 𝑣 is a division algebra, we have 𝐹𝑣 [𝜇 𝑣 ]
is a field, and so its reduced characteristic polynomial is irreducible.
• If 𝑣 ∈ S is ramified in 𝐵, we choose 𝑡 𝑣 ∈ 𝐹𝑣 such that 𝑥 2 − 𝑡 𝑣 𝑥 + 1 is irreducible;
such an element 𝑡 𝑣 exists by Lemma 14.7.5).
The only places that remain are those 𝑣 ∈ S such that 𝑣 is unramified in 𝐵. By
hypothesis that 𝐵 is S-indefinite, we know that there is at least one such split place
𝑣 spl ∈ S remaining (and in particular, 𝑣 spl ∉ Ram 𝐵 ∪ T).
We now apply our polynomial approximation Lemma 28.7.4, we conclude that
there exists 𝑡 ∈ 𝐹 such that 𝑓 (𝑥) = 𝑥 2 − 𝑡𝑥 + 1 is separable and irreducible over 𝐹,
and irreducible over 𝐹𝑣 for all 𝑣 ∈ Ram 𝐵, and such that 𝑡 is arbitrarily close to b
𝑡. Let
𝐾 = 𝐹 [𝑥]/( 𝑓 (𝑥)). Then either Ram 𝐵 = ∅ and 𝐾 ↩→ 𝐵 automatically, or Ram 𝐵 ≠ ∅
so 𝑓 (𝑥) is irreducible and then 𝐾 ↩→ 𝐵 by the local-global principle for embeddings
(Proposition 14.6.7). Let 𝛽 ∈ 𝐵1 have minimal polynomial 𝑓 (𝑥). Since b 𝑡 ∈ trd(𝑈)
28.7. FIRST PROOF 481
and the reduced trace is an open (linear) map, with a closer approximation we may
𝛾 0 ∈ 𝑈 such that trd(𝛽) = trd(b
suppose trd(𝛽) ∈ trd(𝑈), and therefore there exists b 𝛾 0)
so that 𝛽, b0
𝛾 have the same irreducible minimal polynomial. By the Skolem–Noether
𝛼∈𝐵
theorem (Corollary 7.7.3), there exists b b× such that
𝛼−1 b
𝛽=b 𝛾 0b
𝛼. (28.7.5)
Step 3: Finding a sequence. We then Ñ

repeat the above argument with a sequence of
open neighborhoods 𝑈𝑛 3 b𝛼 such that 𝑛 𝑈𝑛 = {b 𝛾 }; we obtain a sequence
𝛼𝑛−1 b
𝐵1 3 𝛽𝑛 = b 𝛾𝑛0 b 𝛼𝑛−1𝑈𝑛 b
𝛼𝑛 ∈ b 𝛼𝑛 . (28.7.6)
𝛾𝑛0 ∈ 𝑈𝑛 , we have b
Since b 𝛾𝑛0 → b
𝛾 , and in particular for 𝑣 ∈ Pl 𝐹 r S ∪ T, we have
0
𝛾𝑛,𝑣 → 𝛾 𝑣 = 1.
Step 4: Harmonizing the sequence. By ‘harmonizing’ the conjugating elements b 𝛼𝑛 ,
we will realize a sequence in 𝐵1 tending to b 𝛾 as desired, in two (subset)steps. First, by
Main Theorem 27.6.14, 𝐵 b×/𝐵× is compact. So restricting to a subsequence, we have
b b ×
𝛼𝑛 = 𝛿 𝑛 𝜇 𝑛 with 𝜇 𝑛 ∈ 𝐵 and b 𝛿𝑛 → b b× . Second, by weak approximation
𝛿 = (𝛿 𝑣 ) 𝑣 ∈ 𝐵
for 𝐵 (Proposition 28.7.3), 𝐵 is dense in 𝑣 ∈T 𝐵×𝑣 , so there is a sequence 𝜈𝑛 from 𝐵×
× Î
such that 𝜈𝑛 → (𝛿 𝑣 ) 𝑣 ∈T .
Step 5: Conclusion. To conclude, we consider the sequence
(𝜈𝑛 𝜇 𝑛 ) 𝛽𝑛 (𝜈𝑛 𝜇 𝑛 ) −1 = (𝜈𝑛 b

𝛿−1 𝛾𝑛0 ( b
𝑛 )b 𝛿 𝑛 𝜈𝑛−1 ). (28.7.7)
𝛿−1
𝛾 . For 𝑣 ∈ T, we have 𝜈𝑛,𝑣 b
We claim that this sequence tends to b −1
𝑛,𝑣 → 𝛿 𝑣 𝛿 𝑣 = 1 so
(𝜈𝑛,𝑣 𝛿−1 0 −1
𝑛,𝑣 )𝛾 𝑛,𝑣 (𝛿 𝑛,𝑣 𝜈 𝑛,𝑣 ) → 𝛾 𝑣 . (28.7.8)
0
On the other hand, for 𝑣 ∈ Pl 𝐹 r (S ∪ T), we have 𝛾𝑛,𝑣 → 1, so
(𝜈𝑛,𝑣 𝛿−1 0 −1
𝑛,𝑣 )𝛾 𝑛,𝑣 (𝛿 𝑛,𝑣 𝜈 𝑛,𝑣 ) → 1 = 𝛾 𝑣 . (28.7.9)
Putting together these cases, we conclude that (𝜈𝑛 𝜇 𝑛 ) 𝛽𝑛 (𝜈𝑛 𝜇 𝑛 ) −1 → b

𝛾 , and therefore
𝛾 is in the closure of 𝐵1 .
b
Remark 28.7.10. Strong approximation has a more general formulation, as follows.

Let 𝐺 be a semisimple, simply-connected algebraic group over the global field 𝐹. Let
S be an eligible set containing a place 𝑣 such that 𝐺 (𝐹𝑣 ) is not compact. Then 𝐺 (𝐹)
is dense in 𝐺 (𝐹6S ), and we say 𝐺 satisfies strong approximation (relative to S). Over
number fields, strong approximation was established by Kneser [Kne66a, Kne66b] and
Platonov [Pla69, Pla69-70], and over function fields by Margulis and Prasad [Pra77];
see also Platonov–Rapinchuk [PR94, Theorem 7.12]. For a survey with discussion
and bibliography, see Rapinchuk [Rap2014] and the description of Kneser’s work by
Scharlau [Scha2009, §2.1].
28.8 Second proof

Because of its importance, we now give a second proof of strong approximation; this
has the same essential elements, but uses some facts from group theory to simplify
away the final steps of the previous proof. We follow Swan [Swa80, §14], who
references Kneser [Kne66a, Kne66b], Platonov [Pla69-70], and Prasad [Pra77]; see
also the exposition by Kleinert [Klt2000, §4.2].
Let 𝑍 := cl(𝐵1 ) ≤ 𝐵 b1 be the closure of 𝐵1 in 𝐵
b1 . We embed 𝐵1𝑣 ↩→ 𝐵
b1 by
(. . . , 1, 𝛼 𝑣 , 1, . . . ) in the 𝑣th component, for 𝑣 ∉ S.
Lemma 28.8.1. If 𝐵1𝑣 ⊆ 𝑍 for almost all 𝑣 ∉ S, then 𝑍 = 𝐵
b1 .
Proof. Suppose that 𝐵1𝑣 ⊆ 𝑍 for all 𝑣 ∉ T where T is a finite set. Let b 𝛾 ∈ 𝐵 b1 . If
𝛾 𝑣 = 1 for all 𝑣 ∈ T, then b 𝛾 is a limit of elements in 𝑍 (approximating at a finite
𝛾 ∈ 𝑍. Otherwise, by weak approximation for 𝐵 (Proposition 28.7.3), there
level), so b
exists 𝛾 ∈ 𝐵1 such that 𝛾 is near 𝛾 𝑣 for all 𝑣 ∈ T. Let 𝛽b have 𝛽 𝑣 = 1 for 𝑣 ∈ T and
𝛽 𝑣 = 𝛾 −1 𝛾 𝑣 for 𝑣 ∉ T; then 𝛽b ∈ 𝑍, and b
𝛾 is the limit of the 𝛾 𝛽.
b
Now we consider
𝑍1 := {b
𝛾 ∈ 𝑍 : 𝛾 𝑣 = 1 for all but finitely many 𝑣}. (28.8.2)
b1 is a normal subgroup.
Lemma 28.8.3. 𝑍1 E 𝐵
Proof. Let b𝛾∈𝐵 b1 and let b𝛼 ∈ 𝑍1 with 𝛼 𝑣 = 1 for 𝑣 ∉ T with T a finite set. By weak
approximation for 𝐵 (Proposition 28.7.3), there exists 𝛾 ∈ 𝐵1 with 𝛾 close to 𝛾 𝑣 for
all 𝑣 ∈ T. Therefore 𝛾 −1 b 𝛾 −1 b
𝛼 𝛾 is near b 𝛾 for 𝑣 ∈ T and 𝛾 −1 𝛼 𝑣 𝛾 = 𝛾 −1
𝛼b 𝑣 𝛼 𝑣 𝛾 𝑣 = 1 for
𝑣 ∉ T, so is the limit of such in 𝑍. Therefore b 𝛾 −1 b𝛾 ∈ 𝑍1 , thus b
𝛼b 𝛾 −1 𝑍1 b
𝛾 ⊆ 𝑍1 and
𝑍1 E 𝐵 b1 .
Lemma 28.8.4. Let 𝐹 be an infinite field. Then PSL2 (𝐹) is a simple group.
Proof. See e.g. Grove [Grov2002, Theorem 1.13]. Briefly, the result can be proven
using Iwasawa’s criterion, since SL2 (𝐹) acts doubly transitively on the linear subspaces
of 𝐹 2 : the kernel of the action is the center {±1}, and the stabilizer subgroup of a
standard basis element is the subgroup of upper triangular matrices, whose conjugates
generate SL2 (𝐹).
Proof of Main Theorem 28.5.3 (Strong approximation). We show that 𝑍 = 𝐵 b1 . By

1
Lemma 28.8.1, it is enough to show that 𝐵 𝑣 ⊆ 𝑍 for almost all 𝑣. By Lemma 28.8.3,
each 𝑍1 ∩ 𝐵1𝑣 E 𝐵1𝑣 is a normal subgroup; by Lemma 28.8.4, either this normal
subgroup is either scalar, or we have 𝑍1 ∩ 𝐵1𝑣 = 𝐵1𝑣 ≤ 𝑍1 ≤ 𝑍 and we are done. So
it suffices to show that for almost all 𝑣, we have 𝑍1 ∩ 𝐵1𝑣 nonscalar, which is to say,
𝑍1 ∩ 𝐵1𝑣 ≠ {±1}.
So let 𝑤 ∈ Pl 𝐹 be unramified. We perform polynomial approximation, as in Step 2
of the first proof. For 𝑣 ∈ Ram 𝐵, let 𝑥 2 − 𝑡 𝑣 𝑥 + 1 be a separable irreducible polynomial
(which exists by Lemma 14.7.5) with 𝑡 𝑣 ∈ 𝑅 𝑣 , and do the same for 𝑤; let b 𝑡 be the
corresponding idele, with 𝑡 𝑣 = 1 for the remaining places 𝑣. By Lemma 28.7.4, there
exists 𝑡 ∈ 𝐹 such that 𝑓 (𝑥) = 𝑥 2 − 𝑡𝑥 + 1 is:
28.9. ∗ NORMALIZER GROUPS 483
• irreducible and separable over 𝐹,

• irreducible over 𝐹𝑣 for all 𝑣 ∈ Ram 𝐵, and
• such that 𝑡 arbitrarily well-approximates b
𝑡 ∈ 𝑅,
b so we may suppose 𝑡 ∈ 𝑅.
By the local-global principle for embeddings (Proposition 14.6.7), there exists 𝛽 ∈ 𝐵1

with 𝑓 (𝛽) = 0; but moreover, 𝛽 is integral and so belongs to a maximal order. In this
way, we manufacture a sequence 𝛽𝑛 with trd(𝛽𝑛 ) → b 𝑡. Repeating this with 𝑡 𝑣 → 1 for
𝑣 ∈ Ram 𝐵 and diagonalizing, we may suppose 𝑡 𝑣 = 1 for all 𝑣 ≠ 𝑤.
Let O be a maximal 𝑅-order. The type set Typ O is finite by Corollary 27.6.25;
choose representatives O𝑖 for Typ O. After conjugating the elements 𝛽𝑛 , we may
suppose without loss of generality that each 𝛽𝑛 belongs to one of the orders O𝑖 . By
the pigeonhole principle, there is an order containing infinitely many, so restricting to
a subsequence we may suppose 𝛽𝑛 ∈ O1 for all 𝑛.
But now the kicker: O b1 is compact, so we may restrict to a convergent subse-
b b 1
quence 𝛽𝑛 → 𝛽 ∈ O . By construction, each 𝛽𝑛 has separable reduced characteristic
polynomial, and trd(𝛽𝑛,𝑣 ) → 1 for all 𝑣 ≠ 𝑤, so 𝛽𝑛,𝑣 → 𝛽 𝑣 = 1 for all 𝑣 ≠ 𝑤. But
trd(𝛽𝑛,𝑤 ) → 𝑡 𝑤 , and 𝑥 2 − 𝑡 𝑤 𝑥 + 1 is irreducible, so 𝛽 𝑤 ∉ 𝐹𝑤× , as desired.
28.9 ∗ Normalizer groups

In this section, we apply strong approximation to the normalizer of an order, and
we compare the normalizers for locally isomorphic orders. We recall the notation
from section 18.5. Let Idl(O) be the group of invertible two-sided fractional O-
ideals, let PIdl(O) ≤ Idl(O) the subgroup of principal fractional O-ideals, and let
PIdl(𝑅) ≤ PIdl(O) be the image of the group of principal fractional 𝑅-ideals.
We suppose throughout this section that PIdl(O) E Idl(O) is a normal subgroup.
This is true whenever Idl O is abelian, which holds when O is Eichler order (us-
ing Lemma 23.3.13 for the primes where 𝔭 is maximal and Proposition 23.4.14 the
remaining primes).
28.9.1. Recall there is a natural exact sequence
1 → 𝑁 𝐵× (O)/(𝐹 × O× ) → Pic𝑅 (O) → Idl(O)/PIdl(O) → 1 (18.5.5)
obtained by considering the class of a bimodule as a two-sided ideal modulo principal

ideals. The idelic dictionary (27.6.26) gives another proof of exactness: there is a
canonical bĳection
Idl(O) −→∼ 𝑁 ( O)/
b O b× ,
b× 𝐵
to obtain Idl(O)/PIdl(O) we take the quotient by 𝑁 (O) and to obtain Idl(O)/PIdl(𝑅)

we take the quotient by 𝐹 × ; therefore the exact sequence (18.5.5) can be rewritten
1 → 𝐹 × \𝑁 𝐵× (O)/O× → 𝐹 × \𝑁 𝐵b× ( O)/

b O b× → 𝑁 𝐵× (O)\𝑁 b× ( O)/
𝐵
b O b× → 1 (28.9.2)
and its exactness is now visible.

28.9.3. We have a map of pointed sets
Idl(O) → Cls O
𝐼 ↦→ [𝐼]
and PIdl(O) is the kernel of this map, the preimage of the trivial class in Cls O. The
composition of this map with the reduced norm gives a group homomorphism:
𝑐 : Idl(O) → Cl𝐺 (O) (𝑅) = 𝐹>×Ω 0 \𝐹

b×/nrd( O)
b
𝐼 ↦→ [nrd(𝐼)]
Lemma 28.9.4. Suppose that 𝐵 is S-indefinite. Then PIdl(O) = ker 𝑐, i.e.,
PIdl(O) = {𝐼 ∈ Idl(O) : [nrd(𝐼)] is trivial in Cl𝐺 (O) (𝑅)}.
Proof. By strong approximation (Theorem 28.5.5), the reduced norm gives a bĳection
nrd : Cls O → Cl𝐺 (O) (𝑅); thus 𝐼 ∈ Idl(O) is principal, belonging to PIdl(O), if and
only if [nrd(𝐼)] is trivial in Cl𝐺 (O) (𝑅).
Now let O, O0 be locally isomorphic orders (in the same genus) with connecting
O, O0-ideal 𝐽.
28.9.5. By 18.4.7, there is an isomorphism of groups

∼ Idl(O0)
Idl(O) −→
(28.9.6)
𝐼 ↦→ 𝐽 −1 𝐼𝐽.
which induces an isomorphism Pic𝑅 (O) ' Pic𝑅 (O0).
We now come to the first major result of this section.
Proposition 28.9.7. Suppose that 𝐵 is S-indefinite. Then the map (28.9.6) induces a
commutative diagram
1 / 𝑁 𝐵× (O)/(𝐹 × O× ) / Pic𝑅 (O) / Idl(O)/PIdl(O) /1
o o o

1 / 𝑁 𝐵× (O0)/(𝐹 × O0× ) / Pic𝑅 (O0) / Idl(O0)/PIdl(O0) /1
with vertical maps isomorphisms.
Proof. We verify that 𝐽 −1 PIdl(O)𝐽 = PIdl(O0), from which both statements follow;
and this verification comes from Lemma 28.9.4, as
[nrd(𝐽 −1 𝐼𝐽)] = [nrd(𝐼)] ∈ Cl𝐺 (O) 𝑅 = Cl𝐺 (O0 ) 𝑅
(recall 27.7.4), so 𝐼 ∈ PIdl(O) if and only if 𝐽 −1 𝐼𝐽 ∈ PIdl(O0).

28.9. ∗ NORMALIZER GROUPS 485
Proposition 28.9.7 says that when 𝐵 is S-indefinite, then the structure of the
normalizer group, the Picard group, and group of ideals modulo principal ideals are
all isomorphic for all orders in a genus. The same is not true when 𝐵 is S-definite;
we always have an isomorphism in the middle, but for locally isomorphic orders,
the Picard group may be distributed differently between the normalizer and the ideal
group.
By chasing a few diagrams, we can be more specific about the structure of Idl(O)
by seeking out primitive ideals.
Definition 28.9.8. The Atkin–Lehner group of O is
AL(O) := 𝐽 ∈ Idl(O) : [nrd(𝐽)] ∈ (Cl𝐺 (O) 𝑅) 2 /Idl(𝑅).

(28.9.9)
The definition (28.9.9) makes sense because Idl(𝑅) is indeed a subgroup: since if
𝔞 ∈ Idl(𝑅) then [nrd(𝔞O)] = [𝔞] 2 ∈ (Cl𝐺 (O) 𝑅) 2 .
Example 28.9.10. Suppose that O is an Eichler order with discrd O = 𝔑. Then

(23.4.21) gives an isomorphism
Ö
Idl(O)/Idl(𝑅) ' Z/2Z.
𝔭 |𝔑
Î
The group AL(O) is therefore an abelian 2-group, isomorphic to 𝔭 |𝔑 Z/2Z when
(Cl𝐺 (O) 𝑅) 2 is trivial. For example, for O = M2 (𝑅) ⊂ 𝐵 = M2 (𝐹), we have AL(O)
the trivial group.
28.9.11. There is a group homomorphism
Idl(𝑅) → Idl(O)
(28.9.12)
𝔞 ↦→ 𝔞O;
this map is injective, since 𝔞O = O implies 𝔞 = 𝑅. We obtain an exact sequence
1 → Cl 𝑅 → Pic𝑅 (O) → Idl(O)/Idl(𝑅) → 1 (28.9.13)
(compare to (18.5.5)). From Lemma 28.9.4 and the fact that PIdl(𝑅) ⊆ ker(𝑐), we
obtain an exact sequence
𝑐
1 → 𝑁 𝐵× (O)/(𝐹 × O× ) → Pic𝑅 (O) →
− Cl𝐺 (O) (𝑅). (28.9.14)
From (28.9.14), we see that 𝑐(Idl(𝑅)/PIdl(𝑅)) = (Cl𝐺 (O) 𝑅) 2 ; further, we have

𝑐(Cl 𝑅) = (Cl𝐺 (O) 𝑅) 2 with

ker 𝑐| Cl 𝑅 = (Cl 𝑅) [2] ↑O := img Cl𝐺 (O) (𝑅) [2] → Cl(𝑅) ≤ (Cl 𝑅) [2]. (28.9.15)
We write (Cl 𝑅) [2] ↑O because this is the subgroup of ideal classes that lift to the group
Cl𝐺 (O) (𝑅) [2] (having order dividing 2). Therefore the following diagram commutes,
with exact rows and columns:

1 1 1

1 / (Cl 𝑅) [2] ↑O / 𝑁 𝐵× (O)/(𝐹 × O× ) / AL(O)

1 / Cl 𝑅 / Pic𝑅 (O) / Idl(O)/Idl(𝑅) /1
𝑐

1 / (Cl𝐺 (O) 𝑅) 2 / Cl𝐺 (O) (𝑅) / Cl𝐺 (O) (𝑅)/(Cl𝐺 (O) 𝑅) 2 /1

1
(28.9.16)
The second main result of this section is the following.
Proposition 28.9.17. Suppose 𝐵 is S-indefinite. Then there is a (non-canonically)
split exact sequence
1 → (Cl 𝑅) [2] ↑O → 𝑁 𝐵× (O)/(𝐹 × O× ) → AL(O) → 1. (28.9.18)
Proof. The snake lemma implies that the top row of (28.9.16) is exact; the sequence
is split by choosing for each class in AL(O) a generator of a representative ideal.
Here is second, self-contained proof which captures the above discussion. Say that
a two-sided (integral) O-ideal 𝐼 is 𝑅-primitive if 𝐼 is not divisible by any integral ideal
of the form 𝔞O with 𝔞 ( 𝑅. (If 𝐼 is integral but not 𝑅-primitive, with 𝐼 ⊆ 𝔞O and 𝔞 as
small as possible, then 𝔞−1 𝐼 ⊆ O is integral and now 𝑅-primitive.) Let 𝛼 ∈ 𝑁 𝐵× (O).
Then O𝛼O = 𝐼 ∈ Idl(O), so we can factor 𝐼 = 𝔠𝐽 uniquely with 𝔠 a fractional ideal of
𝑅 and 𝐽 an 𝑅-primitive ideal. We have
𝑎𝑅 = nrd(𝛼)𝑅 = nrd(𝐼) = 𝔠2 nrd(𝐽) = 𝔠2 𝔞 (28.9.19)
and so
1 = [(𝑎)] = [𝔠] 2 [𝔞] ∈ Cl𝐺 (O) (𝑅)
and in particular [nrd(𝐽)] ∈ (Cl𝐺 (O) 𝑅) 2 . Therefore there is a map
𝑁 𝐵× (O) → AL(O). (28.9.20)
This map is surjective by strong approximation (see Lemma 28.9.4): if 𝐽 ∈ Idl(O)
has nrd(𝐽) = 𝔞 and [𝔞] = [𝔠−1 ] 2 ∈ (Cl𝐺 (O) 𝑅) 2 , then [nrd(𝔠𝐽)] = 1 ∈ Cl𝐺 (O) (𝑅) so
by Theorem 28.5.5, there exists 𝛼 ∈ 𝐵× such that O𝛼O = 𝔠𝐽 and since 𝔠𝐽 ∈ Idl(O)
we have 𝛼 ∈ 𝑁 𝐵× (O). The map is split by this construction, with a choice of 𝐽
up to Idl 𝑅. The kernel of the map in (28.9.20) consists of 𝛼 ∈ 𝑁 𝐵× (O) such that
O𝛼O = 𝔠O with 𝔠 ∈ Idl(𝑅), and from the preceding paragraph [𝔠2 ] = 1 ∈ Cl𝐺 (O) (𝑅)
so [𝔠] ∈ Cl𝐺 (O) (𝑅) [2]; however, the kernel also contains 𝐹 × O× , whence the class of
𝔠 is well-defined only in Cl(𝑅). Therefore the kernel is canonically identified with
(Cl 𝑅) [2] ↑O .
28.10. ∗ STABLE CLASS GROUP 487
Corollary 28.9.21. Suppose that O is an Eichler order with discrd O = 𝔑 and that 𝐵
is S-indefinite. Then
𝑁 𝐵× (O)/(𝐹 × O× ) ' AL(O) × (Cl 𝑅) [2] ↑O
is an abelian 2-group with rank at most 𝜔(𝔑) + ℎ2 (𝑅), where 𝜔(𝔑) is the number of
prime divisors of 𝔑 and ℎ2 (𝑅) = dimF2 (Cl 𝑅) [2].
Proof. Combine Proposition 28.9.17 and Example 28.9.10.
𝑁GL2 (𝐹 ) (M2 (𝑅))/(𝐹 × GL2 (𝑅)) ' (Cl 𝑅) [2].
Proof. Apply Corollary 28.9.21 with Cl𝐺 (O) (𝑅) = Cl(𝑅) so (Cl 𝑅) [2] ↑O = (Cl 𝑅) [2]
and AL(O) the trivial group by Example 28.9.10.
Remark 28.9.23. Corollary 28.9.21 corrects Vignéras [Vig80a, Exercise III.5.4] to

account for possible class group factors.
28.10 ∗ Stable class group

We conclude this epic chapter with a final result on the stable class group, restoring
some generality; we announced a special case of this theorem as Theorem 20.7.16.
Theorem 28.10.1 (Fröhlich–Swan). Let 𝑅 be a global ring with 𝐹 = Frac 𝑅, let 𝐵 be

a central simple 𝐹-algebra, and let O ⊂ 𝐵 be an 𝑅-order. Let Ω ⊆ Pl 𝐹 be the set of
real places 𝑣 of 𝐹 such that 𝐵 𝑣 ' M𝑚 (H) for some 𝑚 ≥ 1. Then the reduced norm
induces an isomorphism
∼ 𝐹 × \𝐹
nrd : StCl O −→ b×/nrd( O
b× ) (28.10.2)
>Ω 0
of finite abelian groups. In particular, if 𝐵 is a quaternion algebra and O is locally

norm-maximal order, then
StCl O ' ClΩ 𝑅
where Ω ⊆ Ram 𝐵 is the set of real ramified places in 𝐵.
Proof. We give only a sketch of the proof. For further details, see the references given
for the proof of Theorem 20.7.16.
We first show that the map (28.10.2) is well-defined. Suppose that 𝐼 ⊕ O𝑟 ' 𝐼 0 ⊕ O𝑟
with 𝑟 ≥ 0. If 𝑟 = 0, we are done; so suppose 𝑟 ≥ 1. Extending scalars, we
find an isomorphism 𝜙 : 𝐵𝑟 +1 → 𝐵𝑟 +1 of left 𝐵-modules, represented by an element
𝛾 ∈ GL𝑟 +1 (𝐵) acting on the left. In a similar way, associated to 𝐼 ⊕ O𝑟 we obtain a
class
GL𝑟 +1 (𝐵)b𝛼 GL𝑟 +1 ( O)
b ∈ GL𝑟 +1 (𝐵)\ GL𝑟 +1 ( 𝐵)/GL
b 𝑟 +1 ( O)
b
by choosing in each completion an isomorphism with O𝑟𝔭+1 represented by a matrix,

well-defined up to a change of basis on the right (and on the left by a global isomor-
phism). Now by strong approximation in the advanced version announced in Remark
28.7.10, the reduced norm induces a bĳection
× b×/nrd( O
b× )
nrd : GL𝑟 +1 (𝐵)\ GL𝑟 +1 ( 𝐵)/GL𝑟 +1 ( O) → 𝐹>Ω 0 \ 𝐹
b b
after we check that the codomain is indeed the image of the reduced norm. Then
𝛼 0) = nrd(𝛾b
nrd(b 𝛼) = nrd(𝛾) nrd(b 𝛼) ∈ 𝐹>×Ω 0 \𝐹
𝛼) = nrd(b b×/nrd( O
b× )
so the map is well-defined.

Similarly, the map (28.10.2) is a group homomorphism: an isomorphism 𝐼 ⊕ 𝐼 0 '
𝐽 ⊕ O gives nrd(𝛾 𝛽) = nrd(𝛽) = nrd(b 𝛼b𝛼 0). The map is surjective. If [𝐼] St is in the
𝛼1 corresponds to 𝐼 ⊕ O; by
𝛼1 ) where b
kernel and nrd(𝐼) is trivial, then so too is nrd(b
strong approximation, this means that there exists 𝛽 ∈ GL2 (𝐵) and 𝜇 b ∈ GL2 ( O) b such
that b 𝜇 and so via 𝛽 we have 𝐼 ⊕ O ' O ⊕2 , which means [𝐼] St = [O].
𝛼1 = 𝛽b
Exercises
Unless otherwise specified, let 𝑅 be a global ring with eligible set S and 𝐹 = Frac 𝑅,
and let 𝐵 be a quaternion algebra over 𝐹, and let O ⊂ 𝐵 be an 𝑅-order.
1. Give another proof weak approximation for Q, as follows.

(a) Let S ⊆ Pl Q be a finite, nonempty set of places. Show that, for each
𝑣 ∈ S, there exists 𝑦 𝑣 ∈ Q× such Î that | 𝑦 𝑣 | 𝑣 < 1 and | 𝑦 𝑣 | 𝑣 > 1 for all
0
𝑣 0 ∈ S r {𝑣}. [Hint: let 𝑦 ∞ := 𝑝 ∈S 1/𝑝 and to get 𝑦 𝑝 , multiply 𝑦 ∞ by a

power of 𝑝.]
(b) For the elements 𝑦 𝑣 constructed in (a), show that for all 𝑣 ∈ S we have
1/(1 + 𝑦 𝑛𝑣 ) → 1 ∈ Q 𝑣 and 1/(1 + 𝑦 𝑛𝑣 ) → 0 ∈ Q 𝑣 0 for 𝑣 0 ≠ 𝑣. [How does
this relate to the proof of the CRT?]
(c) Prove weak approximation. [Hint: given 𝑥 𝑣 ∈ Q 𝑣 for each 𝑣 ∈ S, show
that ∑︁ 𝑥 𝑣
𝑧 𝑛 := → 𝑥 𝑣 ∈ Q𝑣
𝑣 ∈S
1 + 𝑦 𝑛𝑣
as 𝑛 → ∞.]
2. Show that for all 𝑁 ≥ 1, the group SL2 (Z/𝑁Z) is generated by two elements of
order 𝑁.
3. Let 𝐹 be a field and let 𝑛 ∈ Z ≥2 . Show that the elementary matrices (which differ
from the identity matrix in exactly one off-diagonal place) generate SL𝑛 (𝐹) as
a group. [Hint: Argue by induction, and reduce a matrix to the identity by
elementary row and column operations.]
4. Let 𝑛 ≥ 2.
(a) Let 𝑅 be a Euclidean domain. Show that the elementary matrices generate
SL𝑛 (𝑅). [Hint: In view of Lemma 28.3.3, argue by induction.]
28.10. ∗ STABLE CLASS GROUP 489
(b) Using (a), show that Proposition 28.3.5 and Corollary 28.3.6 hold for SL𝑛 .
5. Let 𝐵 be a central simple algebra over the global field 𝐹. We say 𝐵 satisfies the
S-Eichler condition if there exists a place 𝑣 ∈ S such that 𝐵 𝑣 is not a division
algebra. Show that if 𝐹 is a number field and S is the set of archimedean places
of 𝐹, then 𝐵 satisfies the S-Eichler condition if and only if 𝐵 is not a totally
definite quaternion algebra.
6. In this exercise, we provide details in the proof of Proposition 28.3.5, considering
elementary matrices of rings that are not necessarily domains.
Let 𝐴 be a commutative ring (with 1), not necessarily a domain. We say 𝐴 is
Euclidean if there exists a function 𝑁 : 𝐴 → Z ≥0 such that for all 𝑎, 𝑏 ∈ 𝐴 with
𝑏 ≠ 0, there exists 𝑞, 𝑟 ∈ 𝐴 such that 𝑎 = 𝑞𝑏 + 𝑟 and 𝑁 (𝑟) < 𝑁 (𝑏).
In the first parts, we suppose 𝐴 is Euclidean with respect to 𝑁 and show that the
situation is quite analogous to the case when 𝐴 is a domain.
(a) Show that for all 𝑏 ∈ 𝐴 we have 𝑁 (𝑏) ≥ 𝑁 (0) with equality if and only if
𝑏 = 0.
(b) Let 𝑚 = min{𝑁 (𝑏) : 𝑏 ≠ 0}. Show that if 𝑁 (𝑏) = 𝑚 then 𝑏 ∈ 𝐴× .
(c) Show that every ideal of 𝐴 is principal. [We call 𝐴 a principal ideal ring.]
We now consider examples.
(d) Let 𝐴 be an Artinian local principal ideal ring with a unique maximal ideal
𝔪 = 𝜋 𝐴. (For example, we may take 𝐴 = 𝑅/𝔭𝑒 where 𝑅 is a Dedekind
domain, 𝔭 is a nonzero prime ideal, and 𝑒 ∈ Z ≥0 .) Show that for all 𝑥 ∈ 𝐴
with 𝑥 ≠ 0, there exists 𝑢 ∈ 𝐴× and a unique 𝑛 ∈ Z ≥0 such that 𝑥 = 𝑢𝜋 𝑛 .
ConcludeÎ that 𝐴 is Euclidean with 𝑁 (𝑥) = 𝑛.
(e) If 𝐴 ' 𝑟𝑖=1 𝐴𝑖 with each 𝐴𝑖 an Artinian
Í local principal ideal ring, show
that 𝐴 is Euclidean under 𝑁 (𝑥) = 𝑟𝑖=1 𝑁𝑖 (𝜋𝑖 (𝑥)) where 𝑁𝑖 is as given in
(a) for 𝐴𝑖 and 𝜋𝑖 : 𝐴 → 𝐴𝑖 is the projection.
We conclude with the application.
(f) Let 𝐴 be a Euclidean ring. Show that SL2 ( 𝐴) is generated by elementary
matrices. [Hint: Show that the proof in Lemma 28.3.3 carries over.]
7. Consider the quaternion algebra 𝐵 := (−11, −17 | Q) and let O = Z ⊕ Z𝑖 ⊕ Z 𝑗 ⊕
Z𝑖 𝑗. Then 𝐵 is definite, so strong approximation does not apply. Indeed, show
that 𝐵b1 * 𝐵× O
b× as follows.
(a) Find 𝑎, 𝑏, 𝑐, 𝑑, 𝑚 ∈ Z with 3 - 𝑚 such that
𝑎 2 + 11𝑏 2 = 𝑐2 + 17𝑑 2 = 3𝑚.

b× be such that
𝛼 = (𝛼 𝑝 ) 𝑝 ∈ 𝐵
(b) Now let b
(𝑎 + 𝑏𝑖) (𝑐 − 𝑑 𝑗)
𝛼3 = (𝑎 + 𝑏𝑖) (𝑐 + 𝑑 𝑗) −1 =
3𝑚
and 𝛼 𝑝 = 1 if 𝑝 ≠ 3. Show that b 𝛼∈𝐵 𝛼 ∉ 𝐵× O
b1 and b b× . [Hint: Observe
that nrd | O only represents 9 by ±3.]
(c) Prove that the right O-ideals 𝐼1 := 3O + (𝑎 + 𝑏𝑖)O and 𝐼2 := 3O + (𝑐 + 𝑑𝑖)O

are not principal. How does this relate to (b)?
8. Let 𝐹 be a number field with ring of integers 𝑅. Show that there is a finite set S
of (rational) primes such that every totally positive element of 𝑅 can be written
as a sum of four squares of elements of 𝐹 whose denominator is a product of
primes in S. (We may not be able to write every such element as sum of four
squares from 𝑅, but we only need denominators in S.)
9. Suppose that 𝐵 is S-indefinite. Suppose O is locally norm-maximal. Give a
direct proof using strong approximation that if O ⊆ 𝐵 is an 𝑅-order and 𝐼 is an
invertible right fractional O-ideal, then 𝐼 is principal if and only if [nrd(𝐼)] is
trivial in ClΩ 𝑅. [Hint: If 𝛼 ∈ 𝐵× satisfies nrd(𝛼)𝑅 = nrd(𝐼), consider 𝛼−1 𝐼.]
I 10. Show that 𝐵× O
b× ∩ 𝐵 b1 if and only if nrd(O× ) = 𝐹 × ∩ nrd( O
b1 = 𝐵1 O
>Ω 0
b× ) (see
28.4.7).
11. Suppose that 𝐵 is S-indefinite. Show that # Typ O is a power of 2.
12. Suppose disc 𝐵 = 𝔇 and O is an Eichler order of level 𝔐 and reduced discrim-
inant 𝔑 = 𝔇𝔐. Show that Cl𝐺 𝑁 (O) 𝑅 is the quotient of (Cl 𝑅)/(Cl 𝑅) 2 by the
subgroup generated by the classes of ideals 𝔭 | 𝔇 together with 𝔮 | 𝔑 such that
ord𝔮 (𝔑) is odd.
𝑅 𝑅
Conclude that for 𝐵 = M2 (𝐹) and O = , the type set Typ O is represented
𝔑 𝑅

𝑅 𝔟
by the isomorphism classes of the orders for [𝔟] ∈ Cl𝐺 𝑁 (O) 𝑅.
𝔑𝔟−1 𝑅
Chapter 29
Idelic zeta functions
In this chapter, we present an idelic formulation of zeta functions associated to central

simple algebras over global fields and we prove that they have analytic continuation
and functional equation.
To read beyond the first two introductory sections, the additional prerequisite of
real analysis at the level of measure theory is recommended; happily, the statements
of the essential conclusions (for example, the functional equation of the zeta function
of a quaternion algebra) can be understood without this background.
29.1 ⊲ Poisson summation and the Riemann zeta function

We begin Í with−𝑠some essential motivation. Recall that the Riemann zeta function
𝜁 (𝑠) := ∞ 𝑛=1 𝑛 defined in (25.2.2). Following Riemann, we complete to the function
𝜉 (𝑠) := 𝜁 (𝑠)ΓR (𝑠), where
ΓR (𝑠) := 𝜋 −𝑠/2 Γ(𝑠/2) (29.1.1)
and Γ(𝑠) is the complex Γ-function (Exercise 26.2). Then 𝜉 (𝑠) extends to a meromor-
phic function on C and satisfies the functional equation
𝜉 (1 − 𝑠) = 𝜉 (𝑠). (29.1.2)
(More generally, we saw in 26.8.2 that the Dedekind zeta function of a number field
can be completed in an analogous manner, again with functional equation.)
In this introductory section, we sketch a proof of the functional equation and see
it as a consequence of Poisson summation (arising naturally in Fourier analysis, still
following Riemann). Then, following Tate we reinterpret this extra factor in a manner
that realizes the zeta function as a zeta integral on an adelic space, giving a uniform
description and making the whole setup more suitable for analysis.
To begin, we the function 𝜉 (𝑠) itself as an integral. Looking at one term in 𝜉 (𝑠),
we have
∫ ∞ ∫ ∞
d𝑥 2 d𝑢
𝜋 −𝑠/2 Γ(𝑠/2)𝑛−𝑠 = 𝜋 −𝑠/2 𝑛−𝑠 𝑒 −𝑥 𝑥 𝑠/2 = 𝑒 − 𝜋𝑛 𝑢 𝑢 𝑠/2 . (29.1.3)
0 𝑥 0 𝑢
491
492 CHAPTER 29. IDELIC ZETA FUNCTIONS
Accordingly, we define the theta function

∞
∑︁ ∞
∑︁
− 𝜋𝑛2 𝑢 2𝑢
Θ(𝑢) := 𝑒 =1+2 𝑒 − 𝜋𝑛 (29.1.4)
𝑛=−∞ 𝑛=1
studied by Jacobi, convergent absolutely to a holomorphic function on the right half-

plane Re 𝑢 > 0. Summing over 𝑛 ≥ 1 gives
1 ∞
∫
d𝑢
𝜉 (𝑠) = (Θ(𝑢) − 1)𝑢 𝑠/2 (29.1.5)
2 0 𝑢
valid for Re 𝑠 > 1, where the integral converges (so we may justify interchanging
summation and integration). We will soon rewrite this integral so as to extend its
definition to all 𝑠 ∈ C apart from 𝑠 = 0, 1.
Let 𝑓 : R → C be a Schwartz function, so 𝑓 is infinitely differentiable and every
derivative decays rapidly (for all 𝑚, 𝑛 ≥ 0 we have | 𝑓 (𝑚) (𝑥)| = 𝑂 (𝑥 −𝑛 ) as |𝑥| → ∞).
We define the Fourier transform 𝑓 ∨ : R → C of 𝑓 by
∫ ∞
𝑓 ∨ (𝑦) = 𝑒 −2 𝜋𝑖𝑥 𝑦 𝑓 (𝑥) d𝑥. (29.1.6)
−∞
Theorem 29.1.7 (Poisson summation). We have

∞
∑︁ ∞
∑︁
𝑓 (𝑚) = 𝑓 ∨ (𝑛), (29.1.8)
𝑚=−∞ 𝑛=−∞
with both sums converging absolutely.
Proof. The condition that 𝑓 is Schwartz ensures good analytic behavior allowing
the interchange of sum and integral, the details of which we elide. Let 𝑔(𝑥) :=
Í∞
𝑚=−∞ Í 𝑓 (𝑥 + 𝑚). Then 𝑔 is periodic with period 1, so by Fourier expansion we have
𝑔(𝑥) = ∞ 𝑛=−∞ 𝑎 𝑛 𝑒
2 𝜋𝑖𝑛𝑥 where
∫ 1 ∫ 1 ∞
∑︁
𝑎 𝑛 := 𝑔(𝑥)𝑒 −2 𝜋𝑖𝑛𝑥 d𝑥 = 𝑓 (𝑥 + 𝑚)𝑒 −2 𝜋𝑖𝑛𝑥 d𝑥
0 0 𝑚=−∞
∞ ∫
∑︁ 1 ∫ ∞
−2 𝜋𝑖𝑛( 𝑥+𝑚)
= 𝑓 (𝑥 + 𝑚)𝑒 d𝑥 = 𝑓 (𝑥)𝑒 −2 𝜋𝑖𝑛𝑥 d𝑥 = 𝑓 ∨ (𝑛).
𝑚=−∞ 0 −∞
Thus
∞
∑︁ ∞
∑︁
𝑓 (𝑚) = 𝑔(0) = 𝑓 ∨ (𝑛).
𝑚=−∞ 𝑛=−∞
2
− 𝜋 𝑥 . Then 𝑓 is Schwartz. By contour integration and us-
∫ ∞ take2 𝑓 (𝑥) = 𝑒
Now
ing −∞ 𝑒 − 𝜋 𝑥 d𝑥 = 1, we conclude that 𝑓 ∨ (𝑦) = 𝑓 (𝑦) for all 𝑦. For 𝑢 > 0, let
29.2. ⊲ IDELIC ZETA FUNCTIONS, AFTER TATE 493
𝑓𝑢 (𝑥) := 𝑓 (𝑢𝑥); then 𝑓𝑢∨ (𝑦) = 𝑢 −1 𝑓1/𝑢 (𝑦) by change of variable. Applying Poisson
summation to 𝑓√𝑢 (𝑥) then gives
1
Θ(𝑢) = √ Θ(1/𝑢). (29.1.9)
𝑢
The equation (29.1.9) implies the functional equation (29.1.2) for 𝜉 (𝑠) as follows:
we split up the integral (29.1.5) as
∫ 1 ∫ ∞
1 1 d𝑢 1 d𝑢
𝜉 (𝑠) = − + Θ(𝑢)𝑢 𝑠/2 + (Θ(𝑢) − 1)𝑢 𝑠/2
𝑠 2 0 𝑢 2 1 𝑢
and apply the change of variable 𝑢 ← 1/𝑢 to obtain

∫ 1 ∫ ∞ ∫ ∞
𝑠/2 d𝑢 (1−𝑠)/2 d𝑢 2 d𝑢
Θ(𝑢)𝑢 = Θ(𝑢)𝑢 = + (Θ(𝑢) − 1)𝑢 (1−𝑠)/2 .
0 𝑢 1 𝑢 𝑠−1 1 𝑢
Putting these together, we have
1 ∞
∫
d𝑢 1 1
𝜉 (𝑠) = (Θ(𝑢) − 1) (𝑢 𝑠/2 + 𝑢 (1−𝑠)/2 ) − − (29.1.10)
2 1 𝑢 𝑠 1−𝑠
which is sensible as a meromorphic function for all 𝑠 ∈ C, holomorphic except for

𝑠 = 0, 1, with the right-hand side visibly invariant under 𝑠 ← 1 − 𝑠. This establishes
(29.1.2).
This method extends to prove the functional equation for the 𝐿-series 𝐿 (𝑠, 𝜒)
where 𝜒 is a Dirichlet character (now involving a Gauss sum); Hecke extended this
method (generalizing the appropriate theta functions) to prove the functional equation
for a wider class of functions, including the Dedekind zeta functions.
29.2 ⊲ Idelic zeta functions, after Tate

Now convinced of the utility of integral representations, we seek to put the finite places
on an equal footing. In this way, the inclusion
Î Z ⊆ R in the Fourier analysis above is
replaced by Q ⊆ Q. Recall that 𝜁 (𝑠) = 𝑝 𝜁 𝑝 (𝑠) where
∞
∑︁
𝜁 𝑝 (𝑠) := 𝑝 −𝑒𝑠 = (1 − 𝑝 −𝑠 ) −1 ;
𝑒=0
we recover these factors from an integral. First we need a measure to integrate against.
We have Z 𝑝 = lim Z/𝑝 𝑛 Z as a projective limit with compatible projection maps
←−−𝑛
𝜋 𝑛 : Z 𝑝 → Z/𝑝 𝑛 Z. We define the measure on Z 𝑝 as the projective limit of the
counting measures on each Z/𝑝 𝑛 Z with total measure 1, i.e., for a set 𝐸 ⊆ Z 𝑝 we
define
#𝜋 𝑛 (𝐸)
𝜇 𝑝 (𝐸) := lim
𝑛→∞ 𝑝𝑛
when this limit exists. The measure extends additively to Q 𝑝 = Z 𝑝 [1/𝑝] and is
invariant under additive translation
𝜇 𝑝 (𝑎 + 𝐸) = 𝜇 𝑝 (𝐸) for all 𝑎 ∈ Z 𝑝 ;
accordingly, 𝜇 𝑝 is the standard Haar measure on Q 𝑝 (see 29.3). We have normalized

the measure so that ∫
𝜇 𝑝 (Z 𝑝 ) = d𝜇 𝑝 (𝑥) = 1.
Z𝑝
Since Z×𝑝 = Z 𝑝 r 𝑝Z 𝑝 and Z 𝑝 =

Ã 𝑝−1
𝑎=0 (𝑎 + 𝑝Z 𝑝 ) we have 𝜇( 𝑝Z 𝑝 ) = 1/𝑝 and
𝜇 𝑝 (Z×𝑝 ) = 1 − 1/𝑝. (29.2.1)
Similarly, we have a standard Haar measure 𝜇×𝑝 on Q×𝑝 by

−1
1 d𝜇 𝑝 (𝑥)
d𝜇×𝑝 (𝑥) := 1 − ;
𝑝 |𝑥| 𝑝
the measure is invariant under 𝑥 ← 𝑎𝑥 for 𝑎 ∈ Q×𝑝 as well as under the substitution
𝑥 ← 𝑥 −1 , and with our normalization we have
−1
1
𝜇×𝑝 (Z×𝑝 ) = 1− 𝜇 𝑝 (Z×𝑝 ) = 1.
𝑝
For a complex number 𝑠 ∈ C, on our way to recover 𝜁 𝑝 (𝑠) we consider the

(Lebesgue) integral ∫
|𝑥| 𝑠𝑝 d𝜇×𝑝 (𝑥). (29.2.2)
Z 𝑝 r{0}
This looks a bit weird at first, because it is not over a subgroup of Q×𝑝 or anything.
Nevertheless, it works! For every nonzero 𝑥 ∈ Z 𝑝 r {0}, we may write 𝑥 = 𝑝 𝑒 𝑥0 with
𝑥0 ∈ Z×𝑝 and 𝑒 ≥ 0, therefore the integral can be written as a sum over the level sets
𝑝 𝑒 Z×𝑝 :
∫ ∞
∑︁
|𝑥| 𝑠𝑝 d𝜇×𝑝 (𝑥) = 𝑝 −𝑒𝑠 𝜇× (Z×𝑝 ) = (1 − 𝑝 −𝑠 ) −1 = 𝜁 𝑝 (𝑠). (29.2.3)
Z 𝑝 r{0} 𝑒=0
It is more common to rewrite this as an integral over Q×𝑝 by letting Ψ 𝑝 be the charac-
teristic function of Z 𝑝 r {0}, so that
∫ ∫
|𝑥| 𝑠𝑝 d𝜇×𝑝 (𝑥) = |𝑥| 𝑠𝑝 Ψ 𝑝 (𝑥) d𝜇×𝑝 (𝑥). (29.2.4)
Z 𝑝 r{0} Q×𝑝
In a similar fashion, we define the measure 𝜇∞ on R by d𝑥/|𝑥|, and to match

2
(29.1.1) we define Ψ∞ (𝑥) = 𝑒 − 𝜋 𝑥 . Putting these together, on the idele group Q×
we define the product measure 𝜇× = 𝑣 𝜇×𝑣 , the function Ψ(𝑥) = 𝑣 Ψ 𝑣 (𝑥 𝑣 ), and
Î Î
29.2. ⊲ IDELIC ZETA FUNCTIONS, AFTER TATE 495
absolute value k𝑥k = 𝑣 |𝑥 𝑣 | 𝑣 (trivial on Q× by the product formula). We have then

Î
repackaged the zeta function as an adelic integral
∫
𝜉 (𝑠) = |𝑥| 𝑠 Ψ(𝑥) d𝜇× (𝑥). (29.2.5)
Q×
This is really nice!

Rewritten in this idelic way, Tate [Tate67] in his Ph.D. thesis elegantly proved the
functional equation for a wide class of zeta functions (and 𝐿-functions):
[T]he role of Hecke’s complicated theta-formulas for theta functions

formed over a lattice in the 𝑛-dimensional space of classical number the-
ory can be played by a simple Poisson formula [...], the number theoretic
analogue of the Riemann–Roch theorem” [Tate67, p. 305–306].
In this chapter, we use this method of Poisson summation and idelic integrals to
prove the basic properties of the zeta function of a central simple algebra over a global
field (Main Theorem 29.10.1). Translated back into classical language, we prove as
a consequence the key result (the crux of which is Theorem 26.8.19) announced in
section 26.8.
Main Theorem 29.2.6. Let 𝐹 be a number field with ring of integers 𝑅 = Z𝐹 and let
𝐵 be a quaternion algebra over 𝐹 with maximal order O. Let
∑︁
𝜁 𝐵 (𝑠) := N (𝐼) −𝑠
𝐼 ⊆O
be the sum over nonzero right O-ideals, where N is the absolute (counting) norm
(16.4.7) and let 𝜉 𝐵 (𝑠) be its completion (26.8.13). Then 𝜉 𝐵 (𝑠) has meromorphic
continuation to C, holomorphic away from {0, 1/2, 1} with simple poles at 𝑠 = 0, 1,
and it satisfies the functional equation
𝜉 𝐵 (1 − 𝑠) = 𝜉 𝐵 (𝑠). (29.2.7)
Moreover, if 𝐵 is a division algebra, then 𝜉 𝐵 (𝑠) is holomorphic at 𝑠 = 1/2.
Main Theorem 29.2.6 is an analytic result with key arithmetic consequences,

including the classification of quaternion algebras over global fields, as we saw in
section 26.8. More than that, the evaluation of the residue will further give rise to a
volume formula (Main Theorem 39.1.8) that generalizes the Eichler mass formula (as
in the proof of Proposition 26.5.10).
The developments in this chapter have a rich history, and they generalize vastly
beyond this text: see Remark 29.10.24. We restrict ourselves to the case of zeta
functions both because this suffices for our main applications and because it is a good
stepping stone to the more general theory. Although this chapter is quite technical, the
reader’s forbearance will ultimately be rewarded!
29.3 Measures
In this section, we define the local measures we will use. As references for this
section and the next, see Bekka–de la Harpe–Valette [BHV2008, Appendices A and
B], Deitmar [Dei2005], Deitmar–Echterhoff [DE2009, Chapters 1 and 3, Appendix
B], Loomis [Loo53], and Ramakrishnan–Valenza [RM99, Chapters 1–4] as references
on harmonic analysis, and Vignéras [Vig80a, §II.4] and Weil [Weil74, Chapter XI] for
the present context.
Let 𝐺 be a Hausdorff, locally compact, second countable topological group. For
example, we may take 𝐺 = SL2 (R) or 𝐺 = SL𝑛 (R) (or more generally a semisimple
real Lie group). A Borel measure on 𝐺 is a countably additive function, with values
in [0, ∞], defined on the 𝜎-algebra generated by open sets in 𝐺 under complement
and finite or countable unions.
Definition 29.3.1. A Radon measure on 𝐺 is a Borel measure that is finite on

compact sets. A left Haar measure on 𝐺 is a nonzero Radon measure 𝜇 that is left
translation-invariant, so 𝜇(𝑔𝐸) = 𝜇(𝐸) for all Borel subsets 𝐸 ⊆ 𝐺 and all 𝑔 ∈ 𝐺.
The notion of right Haar measure is defined similarly.

Remark 29.3.2. Since 𝐺 is second countable and locally compact, every open subset
of 𝐺 is 𝜎-compact (the union of a countable collection of compact subspaces), so a
Radon measure is necessarily regular, i.e., the measure of a Borel set is the infimum
of the measures of its open supersets and the measure of an open set is the supremum
of the measures of its compact subsets: see Rudin [Rud87, Theorem 2.18].
Proposition 29.3.3. 𝐺 admits a left Haar measure that is unique up to scaling by an

element of R>0 .
Proof. See e.g. Deitmar–Echterhoff [DE2009, Theorem 1.3.4] or Diestel–Spalsbury

[DS2014, §7.2].
We will construct Haar measures explicitly as we need them, so we need not appeal
to the general result of Proposition 29.3.3 beyond the uniqueness statement which is
itself straightforward to establish: see Exercise 29.4.
Example 29.3.4. On 𝐺 = R𝑛 under addition, a left (and right) Haar measure is given
by the usual Lebesgue measure.
Example 29.3.5. Suppose 𝐺 has the discrete topology (all sets are open). Then the
counting measure 𝜇(𝐸) = #𝐸 ∈ Z ≥0 ∪ {∞} is a left (and right) Haar measure.
From now on, let 𝜇 be a left Haar measure on 𝐺.
29.3.6. In general, a left Haar measure need not also be right translation-invariant. For
𝑔 ∈ 𝐺, the measure defined by 𝜇𝑔 (𝐸) := 𝜇(𝐸𝑔) for a Borel set 𝐸 is again a left Haar
measure, so by Proposition 29.3.3, we have 𝜇𝑔 = Δ𝐺 (𝑔)𝜇 for some Δ𝐺 (𝑔) ∈ R>0 .
Since 𝜇 is unique up to scaling, the function Δ𝐺 (𝑔) does not depend on 𝜇.
29.3. MEASURES 497
Definition 29.3.7. The map Δ𝐺 : 𝐺 → R>0 is the modular function of 𝐺. We say 𝐺

is unimodular if Δ𝐺 is identically 1.
By 29.3.6, 𝐺 is unimodular if and only if every left Haar measure is a right Haar
measure.
Lemma 29.3.8. The modular function Δ𝐺 is a homomorphism.
Example 29.3.9. If 𝐺 is abelian or discrete (see Example 29.3.5), then 𝐺 is unimodu-

lar: every left Haar measure is a right Haar measure. (There are many other important
classes of groups that are unimodular, including discrete (e.g., finite) groups, compact
groups, and the others we will in what follows.)
29.3.10. In this paragraph, we briefly review the theory of integration we need; for
further details, see Deitmar–Echterhoff [DE2009, Appendix B.1]. Let 𝑓 : 𝐺 → C
be a complex-valued function on 𝐺. We say 𝑓 is measurable if for all Borel sets
𝐸 ⊆ C, the subset 𝑓 −1 (𝐸) is measurable; we say 𝑓 is real(-valued) if 𝑓 (𝐺) ⊆ R and
is nonnegative if 𝑓 (𝐺) ⊆ R ≥0 .
If 𝐸 is a measurable set then the characteristic function 1𝐸 of 𝐸 (equal to 1 on 𝐸
and 0 outside 𝐸) is defined to have integral
∫ ∫
1𝐸 (𝑥) d𝜇(𝑥) := 𝜇(𝐸) = d𝜇(𝑥). (29.3.11)
𝐺 𝐸
A step function is a finite C-linear combination of characteristic functions of measur-

able sets, and we define the integral of a step function by linearity using (29.3.11).
If 𝑓 is measurable and nonnegative, we define
∫ ∫
𝑔 a nonnegative step function such
𝑓 (𝑥) d𝜇(𝑥) := sup 𝑔(𝑥) d𝜇(𝑥) :
𝐺 𝐺 that 𝑔(𝑥) ≤ 𝑓 (𝑥) for all 𝑥 ∈ 𝐺
∫
and say 𝑓 is integrable if 𝐺 𝑓 (𝑥) d𝜇(𝑥) < ∞.
If 𝑓 is real we say 𝑓 is integrable if both 𝑓 + := max( 𝑓 , 0) and 𝑓 − := − min( 𝑓 , 0)
are integrable. If 𝑓 is real and integral, we then define
∫ ∫ ∫
𝑓 (𝑥) d𝜇(𝑥) := 𝑓 + (𝑥) d𝜇(𝑥) − 𝑓 − (𝑥) d𝜇(𝑥). (29.3.12)
𝐺 𝐺 𝐺
Finally, a general 𝑓 is integrable if Re 𝑓 and Im 𝑓 are integrable, in which case

∫ ∫ ∫
𝑓 (𝑥) d𝜇(𝑥) := Re 𝑓 (𝑥) d𝜇(𝑥) + 𝑖 Im 𝑓 (𝑥) d𝜇(𝑥).
𝐺 𝐺 𝐺
29.3.13. Let 𝐻 be another Hausdorff, locally compact, second countable topological

group, and let 𝜙 : 𝐺 → 𝐻 be continuous surjection with kernel 𝑁 := ker 𝜙. Then we
have an exact sequence
𝜙
1→𝑁 →𝐺−
→ 𝐻 → 1.
We say that left Haar measures on 𝐺, 𝑁, 𝐻 are compatible if for all integrable functions
𝑓 : 𝐺 → C,
∫ ∫ ∫
𝑓 (𝑥) d𝜇(𝑥) = 𝑓 (𝑧𝑦) d𝜇(𝑧) d𝜇(𝜙(𝑦)). (29.3.14)
𝐺 𝐻 𝑁
Given measures on two terms, there exists a unique compatible measure on the third—
but note, this measure depends on the exact sequence (Exercise 29.7).
29.4 Modulus and Fourier inversion

We continue our background review with notation from the previous section; we now
treat the modulus of an automorphism and present Fourier inversion.
29.4.1. Let 𝜙 ∈ Aut(𝐺) be a continuous automorphism. Then the measure defined by

𝜇 𝜙 (𝐸) := 𝜇(𝜙(𝐸)) is again a left Haar measure. By Proposition 29.3.3, there exists a
unique k𝜙k ∈ R>0 such that
𝜇(𝜙𝐸) = k𝜙k𝜇(𝐸) (29.4.2)
for all measurable sets 𝐸. In particular, for all integrable functions 𝑓 on 𝐺 we have
∫ ∫ ∫
𝑓 (𝑥) d𝜇(𝑥) = 𝑓 (𝜙(𝑥)) d𝜇(𝜙(𝑥)) = k𝜙k 𝑓 (𝜙(𝑥)) d𝜇(𝑥). (29.4.3)
𝐺 𝐺 𝐺
Definition 29.4.4. The modulus of 𝜙 ∈ Aut(𝐺) is k𝜙k ∈ R>0 as defined in (29.4.2).
The definition of the modulus is independent of the choice of Haar measure 𝜇,

since Haar measure is unique up to scaling and such a scalar cancels in (29.4.2).
Example 29.4.5. Let 𝑔 ∈ 𝐺; then the conjugation automorphism 𝜙𝑔 (𝑥) = 𝑔 −1 𝑥𝑔 has

k𝜙𝑔 k = Δ𝐺 (𝑔), because 𝜇 is left-invariant. In particular, 𝐺 is unimodular if and only
if all conjugation maps have trivial modulus.
Now let 𝐴 be a Hausdorff, locally compact, second countable topological ring and
let 𝜇 be a left Haar measure on the additive group of 𝐴. Since 𝐴 is abelian, 𝐴 is
unimodular and 𝜇 is also a right Haar measure.
29.4.6. Let 𝑎 ∈ 𝐴× . Then the left multiplication map 𝜆 𝑎 : 𝐴 → 𝐴 by 𝑥 ↦→ 𝑎𝑥 is

a continuous automorphism of 𝐴 as an additive abelian group, so we may define its
modulus k𝑎k := k𝜆 𝑎 k by (29.4.2): symbolically, we write
d𝜇(𝑎𝑥)
k𝑎k = . (29.4.7)
d𝜇(𝑥)
We have k𝑎𝑏k = k𝑎k k𝑏k for all 𝑎, 𝑏 ∈ 𝐴× (Exercise 29.6).

29.4. MODULUS AND FOURIER INVERSION 499
29.4.8. The measure on 𝐴× defined by
d𝜇(𝑥)
d𝜇× (𝑥) := (29.4.9)
k𝑥k
is a (multiplicative) Haar measure by 29.4.6.
We conclude this introductory section with the Fourier inversion formula.
29.4.10. Let C1 := {𝑧 ∈ C : | 𝑧| = 1} be the circle group. A unitary character

of 𝐴 is a continuous homomorphism 𝜒 : 𝐴 → C1 , considering 𝐴 as an additive
group. Let 𝐴∨ := Hom( 𝐴, C1 ) be the unitary character group of 𝐴 under pointwise
multiplication, and equip 𝐴∨ with the compact-open topology (as a closed subset of
the set of all continuous maps 𝐴 → C1 ).
Remark 29.4.11. We reserve the term character for continuous group homomorphisms
𝐴 → 𝐾 × , where 𝐾 is a field (of values for the character); this notion makes sense
for any field 𝐾. Some authors call unitary characters just characters, then calling our
characters instead quasi-characters.
Now suppose that there exists 𝜓 ∈ 𝐴∨ such that the map
𝐴 → 𝐴∨
𝑥 ↦→ (𝑦 ↦→ 𝜓(𝑥𝑦))
is an isomorphism of topological groups. (This is a hypothesis on 𝐴, and may depend

on a choice; in the cases we consider below, we will identify a standard such 𝜓.)
Definition 29.4.12. For 𝑓 : 𝐴 → C continuous and integrable, the Fourier transform

of 𝑓 (relative to 𝜓, 𝜇) is the function
𝑓∨: 𝐴 → C
∫
𝑓 ∨ (𝑥) = 𝑓 (𝑦)𝜓(𝑥𝑦) d𝜇(𝑦).
𝐴
Theorem 29.4.13 (Fourier inversion). There exists a unique Haar measure 𝜏 on 𝐴

(depending on 𝜓) such that for all 𝑓 : 𝐴 → C continuous and integrable with 𝑓 ∨
(defined relative to 𝜏) continuous and integrable, we have
∫
𝑓 (𝑥) = 𝑓 ∨ (𝑦)𝜓(𝑥𝑦) d𝜏(𝑦). (29.4.14)
𝐴
Proof. The proof of this theorem is beyond the scope of this textbook, and we use it as
a black box; see e.g. Deitmar–Echterhoff [DE2009, Theorem 3.5.8] or Folland [Fol95,
Theorems 4.32–4.33].
The normalized measure 𝜏 in Theorem 29.4.13 is called the self-dual measure on

𝐴 (with respect to 𝜓).
29.5 Local measures and zeta functions: archimedean case

Let 𝐵 be a finite-dimensional simple algebra over the local field 𝐹 = R. A good general
reference for the next three sections is Weil [Weil82, Chapter II].
29.5.1. The (additive) Haar measure 𝜇 on 𝐵 is the usual (Lebesgue) measure, normal-
ized as follows:Íletting 𝑛 = dimR 𝐵, we choose an R-basis 𝑒 1 , . . . , 𝑒 𝑛 for 𝐵, so that we
may write 𝑥 = 𝑖 𝑥𝑖 𝑒 𝑖 ∈ 𝐵 with 𝑥𝑖 are coordinates on 𝐵, and we define
d𝑥 := | 𝑑 (𝑒 1 , . . . , 𝑒 𝑗 )| 1/2 d𝑥1 · · · d𝑥 𝑛
where 𝑑 is the discriminant defined by (15.2.1) and the reduced trace is taken on 𝐵 as
an R-algebra. By Lemma 15.2.5, we see that this measure is independent of the choice
of basis 𝑒 𝑖 .
Another application of Lemma 15.2.5 then gives the modulus
k𝛼k = |Nm 𝐵 |R (𝛼)| (29.5.2)
for all 𝛼 ∈ 𝐵× .
Example 29.5.3. We compute:
d𝑥 = d𝑥, for 𝑥 ∈ R;
√
d𝑧 = 2 d𝑥 d𝑦, for 𝑧 = 𝑥 + 𝑦 −1 ∈ C;
d𝛼 = 4 d𝑡 d𝑥 d𝑦 d𝑧, for 𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 ∈ H;
Ö
d𝛼 = d𝑥𝑖 𝑗 , for 𝛼 = (𝑥 𝑖 𝑗 )𝑖, 𝑗 ∈ M𝑛 (R);
𝑖, 𝑗
2
Ö √
d𝛼 = 2𝑛 d𝑥 𝑖 𝑗 d𝑦 𝑖 𝑗 , for 𝛼 = (𝑥𝑖 𝑗 + 𝑦 𝑖 𝑗 −1)𝑖, 𝑗 ∈ M𝑛 (C).
𝑖, 𝑗
And:
k𝑥k = |𝑥|, for 𝑥 ∈ R× ;

k𝑧k = | 𝑧| 2 , for 𝑧 ∈ C× ;
k𝛼k = nrd(𝛼) 2 , for 𝛼 ∈ H× ;
k𝛼k = |det(𝛼)| 𝑛 , for 𝛼 ∈ GL𝑛 (R);
2𝑛
k𝛼k = |det(𝛼)| , for 𝛼 ∈ GL𝑛 (C).
The modulus for R is the usual absolute value, whereas the modulus for C is the square
of the absolute value (cf. Remark 12.2.3), explaining conventions on normalized
absolute values in the product formula 14.4.6.
29.5.4. As in (29.4.9), the (multiplicative) Haar measure 𝜇× on 𝐵× is defined by
d𝜇(𝛼)
d𝜇× (𝛼) := . (29.5.5)
k𝛼k
29.5. ARCHIMEDEAN MEASURES AND ZETA FUNCTIONS 501
29.5.6. We define the standard unitary character
𝜓 : 𝐵 → C1
𝛼 ↦→ exp(−2𝜋𝑖 trd(𝛼))
with reduced trace taken on 𝐵 as an R-algebra. We note that 𝜓(𝛼𝛽) = 𝜓(𝛽𝛼) for all
𝛼, 𝛽 ∈ 𝐵.
A fundamental result in standard Fourier analysis (generalizing the case 𝐵 = R, C
and following from it in the same way) is the following proposition.
Proposition 29.5.7. The standard unitary character induces an isomorphism 𝐵 −→ ∼ 𝐵∨
of topological groups, and the measure 𝜇 defined in 29.5.1 is self-dual (with respect
to 𝜓).
In light of Proposition 29.5.7, we will also write 𝜏 = 𝜇 for the measure defined
above.
29.5.8. The exact sequence
nrd
1 → H1 → H× −−→ R>0 → 1
defines a measure 𝜏 1 on H1 by compatibility, taking the normalized measures 𝜏 × on

H× and on R>0 ≤ R× .
Lemma 29.5.9. We have 𝜏 1 (H1 ) = 4𝜋 2 .
Proof. Let 𝜌 > 0 and let
𝐸 = {𝛼 ∈ H× : nrd(𝛼) ≤ 𝜌 2 }
be the punctured ball of radius 𝜌. Let 𝑓 be the function k𝛼k on 𝐸, zero elsewhere. On
the one hand, ∫ ∫
× d𝜏(𝛼)
𝑓 (𝛼) d𝜏 (𝛼) = k𝛼k = 𝜏(𝐸).
H × 𝐸 k𝛼k
Recalling that 𝜏 is 4 times Lebesgue measure, and that the Lebesgue measure of a
sphere of radius 𝜌 has volume 𝜋 2 𝜌 4 /2, we get 𝜏(𝐸) = 2𝜋 2 𝜌 4 . On the other hand, by
compatibility, this integral is equal to
∫ ∫ ∫ 𝜌∫
1 × 2 2𝑟 d𝑟
𝑓 (𝛼1 𝑟) d𝜏 (𝛼1 ) d (𝑟 ) = k𝛼1 𝑟 kd𝜏 1 (𝛼1 ) 2
R>0 H 1 0 H 1 𝑟
𝜌4
∫ 𝜌
= 2𝜏 1 (H1 ) 𝑟 3 d𝑟 = 𝜏 1 (H1 ) .
0 2
We conclude that 𝜏 1 (H1 ) = 4𝜋 2 .
We will need further functions to integrate, so we make the following definition.
When we consider functions on 𝐵, we Í may think of choosing an R-basis 𝑒 1 , . . . , 𝑒 𝑛 of
𝐵, identifying 𝐵 ' R𝑛 and writing 𝑥 = 𝑖 𝑥𝑖 𝑒 𝑖 ∈ 𝐵; our definitions will be independent
of this choice of basis.
Definition 29.5.10. A function 𝑓 : 𝐵 → C decays rapidly if for all 𝑛 ≥ 0 we have

| 𝑓 (𝑥)| = 𝑂 ((max𝑖 |𝑥𝑖 |) −𝑛 ) as max𝑖 |𝑥𝑖 | → ∞.
A function 𝑓 : 𝐵 → C is Schwartz(–Bruhat) if 𝑓 is infinitely differentiable and
every partial derivative of 𝑓 decays rapidly.
29.5.11. Let ∗ : 𝐵 → 𝐵 be the conjugate transpose involution (see 8.4.3) on 𝐵. Let
𝑄(𝛼) := trd(𝛼𝛼∗ ) for 𝛼 ∈ 𝐵. We compute that
𝑄(𝑥) = 𝑥 2 , for 𝑥 ∈ R;
2 2 2
√
𝑄(𝑧) = 2| 𝑧| = 2(𝑥 + 𝑦 ), for 𝑧 = 𝑥 + 𝑦 −1 ∈ C;
𝑄(𝛼) = 2 nrd(𝛼), for 𝛼 ∈ H;
∑︁ (29.5.12)
𝑄(𝛼) = 𝑥𝑖2𝑗 , for 𝛼 = (𝑥 𝑖 𝑗 )𝑖, 𝑗 ∈ M𝑛 (R); and
𝑖, 𝑗
∑︁
𝑄(𝛼) = 2 | 𝑧𝑖 𝑗 | 2 , for 𝛼 = (𝑧 𝑖 𝑗 )𝑖, 𝑗 ∈ M𝑛 (C).
𝑖, 𝑗
Definition 29.5.13. The standard function on 𝐵 is

Ψ: 𝐵 → C
(29.5.14)
Ψ(𝛼) = exp(−𝜋𝑄(𝛼)).
It is straightforward to verify that the standard function is Schwartz.

Definition 29.5.15. For a Schwartz function Φ, we define the (local) zeta function
∫
𝑍 𝐵Φ (𝑠) := Φ(𝛼) k𝛼k 𝑠 d𝜏 × (𝛼) (29.5.16)
𝐵×
wherever this integral converges. We abbreviate 𝑍 𝐵Ψ (𝑠) = 𝑍 𝐵 (𝑠) for Φ = Ψ the

standard function.
As in (26.8.3), we define
ΓR (𝑠) := 𝜋 −𝑠/2 Γ(𝑠/2)

(29.5.17)
ΓC (𝑠) := 2(2𝜋) −𝑠 Γ(𝑠).
𝑍R (𝑠) = ΓR (𝑠),
𝑍C (𝑠) = 𝜋ΓC (𝑠),
𝑍H (𝑠) = 𝜋(2𝑠 − 1)ΓR (2𝑠)ΓR (2𝑠 − 1) = 2𝜋 2 ΓR (2𝑠)ΓR (2𝑠 + 1),
𝑍M2 (R) (𝑠) = 𝜋ΓR (2𝑠)ΓR (2𝑠 − 1),
𝑍M2 (C) (𝑠) = 2𝜋 3 ΓC (2𝑠)ΓC (2𝑠 − 1).
Proof. We have ∫ ∞
2 d𝑥
𝑍R (𝑠) = 2 𝑥 𝑠 𝑒− 𝜋 𝑥 ;
0 𝑥
29.6. LOCAL MEASURES: COMMUTATIVE NONARCHIMEDEAN CASE 503
making the substitution 𝑥 ← 𝜋𝑥 2 gives the result. A similar argument with polar
coordinates gives 𝑍C (𝑠). The remaining integrals are pretty fun, so they are left as
Exercise 29.8.
Lemma 29.5.18 explains the provenance of the definitions of ΓR (𝑠), ΓC (𝑠) from
(26.8.3), and ultimately their appearance in (26.8.14). (A quick check on the con-
stant in front is provided by 𝑍 𝐵 (1) = 1; these particular choices of constants follow
convention.) By comparison, Lemma 29.5.18 then shows that in general 𝑍 𝐵Φ (𝑠) has
meromorphic continuation to C.
29.6 Local measures: commutative nonarchimedean case

Now let 𝐹 be a nonarchimedean local field, with valuation 𝑣, valuation ring 𝑅, and
maximal ideal 𝔭 = 𝜋𝑅 ⊆ 𝑅. Let 𝑞 := #𝑅/𝔭.
We begin by defining a Haar measure normalized so that 𝑅 has measure 1; then
we extend this to find the normalization in which the measure is self-dual with respect
to a standard unitary character.
29.6.1. We have 𝑅 lim 𝑅/𝔭𝑛 with projection maps 𝜋 𝑛 : 𝑅 → 𝑅/𝔭𝑛 . We define the
←−−𝑛
(additive) measure 𝜇 on 𝑅 by
#𝜋 𝑛 (𝐸)
𝜇(𝐸) := lim
𝑛→∞ 𝑞𝑛
for a subset 𝐸 ⊂ 𝑅, when this limit exists. It is straightforward to check that 𝐸 defines a
Haar measure on 𝑅, and the measure extends to 𝐹 by additivity. In this normalization,
we have 𝜇(𝑅) = 1, and more generally 𝜇(𝑎 + 𝔭𝑛 ) = 1/𝑞 𝑛 for 𝑎 ∈ 𝑅 and 𝑛 ≥ 0.
Let | | be the preferred absolute value (see 14.4.12), with | 𝜋| = 1/𝑞.
Lemma 29.6.2. For all 𝑥 ∈ 𝐹 × , we have
k𝑥k = |𝑥|; (29.6.3)
in particular, if 𝑥 ∈ 𝑅, then |𝑥| = N(𝑅𝑥) −1 where N is the counting norm (16.4.7).
Proof. By multiplicativity, we may suppose 𝑥 ∈ 𝑅. Then [𝑅 : 𝑥𝑅] = 𝑞 𝑣 ( 𝑥) = N(𝑅𝑥).

Adding up cosets of 𝑥𝑅 in 𝑅, we obtain
𝜇(𝑅) = [𝑅 : 𝑥𝑅] 𝜇(𝑥𝑅) = 𝑞 𝑣 ( 𝑥) 𝜇(𝑥𝑅);
so k𝑥k = 𝜇(𝑥𝑅)/𝜇(𝑅) = 𝑞 −𝑣 ( 𝑥) = |𝑥|, as claimed.
Example 29.6.4. Since 𝜇(𝑅 × ) = 𝜇(𝑅) − 𝜇(𝔭) and
𝜇(𝔭) = k𝜋k𝜇(𝑅) = 1/𝑞,
we have
1
𝜇(𝑅 × ) = 1 − . (29.6.5)
𝑞
29.6.6. We normalize the (multiplicative) Haar measure 𝜇× on 𝐹 × by defining

1 d𝜇(𝑥) d𝜇(𝑥)
d𝜇× (𝑥) := = (1 − 1/𝑞) −1
𝜇(𝑅 × ) k𝑥k k𝑥k
so that 𝜇× (𝑅 × ) = 1.
Next, we consider the Fourier transform in this context.
29.6.7. We first define an additive homomorphism
h i𝐹 : 𝐹 → R/Z
as follows.
(a) If 𝐹 = Q 𝑝 , we define h𝑥iQ 𝑝 ∈ Q to be such that 0 ≤ h𝑥iQ 𝑝 < 1 and 𝑥 − h𝑥iQ 𝑝 ∈
Z𝑝. Í
(b) If 𝐹 = F𝑞 ((𝑡)) and 𝑥 = 𝑖 𝑎 𝑖 𝑡 𝑖 then we take h𝑥i𝐹 := TrF𝑞 /F 𝑝 (𝑎 −1 )/𝑝.
(c) In general, if 𝐹 ⊇ 𝐹0 is a finite separable extension of fields, then we define
h𝑥i𝐹 := hTr𝐹 /𝐹0 𝑥i𝐹0 .
Definition 29.6.8. The standard unitary character of 𝐹 is
𝜓 𝐹 : 𝐹 → C1
𝜓 𝐹 (𝑥) = exp(2𝜋𝑖h𝑥i𝐹 ).
Proposition 29.6.9. The standard unitary character 𝜓 𝐹 defines an isomorphism

∼ 𝐹∨
𝐹 −→
𝑥 ↦→ (𝑦 ↦→ 𝜓(𝑥𝑦))
of topological groups.
Proposition 29.6.10. The following statements hold.

(a) If 𝐹 is a local number field (so a finite extension of Q 𝑝 for a prime 𝑝), then the
measure
𝜏 := |disc(𝑅)| 1/2 𝜇 = N (disc 𝑅) −1/2 𝜇
is self-dual (with respect to 𝜓).
(b) If 𝐹 is a local function field, then 𝜇 is self-dual with respect to 𝜓.
Proof. First consider (a), and suppose 𝐹 is a finite extension of Q 𝑝 . We seek to satisfy
(29.4.14); the equation holds up to a constant, so we may choose appropriate 𝑓 and 𝑥.
We choose 𝑓 as the characteristic function of 𝑅 and 𝑥 = 0, so that 𝑓 (0) = 1 and
∫ ∫ ∫ ∫
𝑓 (𝑧)𝜓(𝑦𝑧) d𝜇(𝑧) d𝜇(𝑦) = 𝑓 (𝑧) 𝜓(𝑦𝑧) d𝜇(𝑦) d𝜇(𝑧). (29.6.11)
𝐹 𝐹 𝐹 𝐹
29.7. LOCAL ZETA FUNCTIONS: NONARCHIMEDEAN CASE 505
By character theory,
∫
𝜓(𝑦𝑧) d𝜇(𝑦) = 𝜇({𝑦 ∈ 𝐹 : 𝜓(𝑦𝑧) = 1}) = 𝜇({𝑦 ∈ 𝐹 : Tr𝐹 /Q 𝑝 (𝑦𝑧) ∈ Z 𝑝 })
𝐹
so (29.6.11) is equal to 𝜇(𝑅 ♯ ) where
𝑅 ♯ = codiff(𝑅) = {𝑥 ∈ 𝐹 : Tr𝐹 /Q 𝑝 (𝑥𝑅) ∈ Z 𝑝 }.
Let 𝑥𝑖 be a Z 𝑝 -basis for 𝑅 with 𝑥𝑖♯ the dual basis, giving a Z 𝑝 -basis for 𝑅 ♯ . By Lemma
15.6.17 we have
disc(𝑅) = [𝑅 ♯ : 𝑅] Z 𝑝
so since 𝜇(𝑅) = 1 by additivity we have
𝜇(𝑅 ♯ ) = |disc(𝑅)| −1 .
It follows then that

𝜏 = 𝜇(𝑅 ♯ ) −1/2 𝜇 = |disc(𝑅)| 1/2 𝜇
is self-dual.
29.7 Local zeta functions: nonarchimedean case

Continuing with 𝐹 a nonarchimedean local field, let 𝐵 be a finite-dimensional simple
(not necessarily central) algebra over 𝐹, with maximal order O, and 𝑛2 = dim 𝑍 (𝐵) 𝐵.
29.7.1. The (additive) Haar measure 𝜇 on 𝐵 is defined as in 29.6.1 as a projective

limit, normalized so that 𝜇(O) = 1.
We compute that k𝛼k 𝐵 = kNm 𝐵 |𝐹 (𝛼) k 𝐹 , so that if 𝐵 is central over 𝐹 with
dim𝐹 𝐵 = 𝑛2 , then k𝛼k 𝐵 = knrd(𝛼) 𝑛 k 𝐹 . In the remainder of this section, we work
over 𝐵 and drop the subscript 𝐵 .
29.7.2. For the (multiplicative) Haar measure 𝜇× on 𝐵× , we use the same normalization
factor (29.6.5) as for 𝐹 × , defining the normalized measure
d𝜇(𝛼)
d𝜇× (𝛼) := (1 − 1/𝑞) −1 . (29.7.3)
k𝛼k
29.7.4. We extend the standard unitary character on 𝐹, defined in 29.6.7, to a standard

unitary character on 𝐵 by
𝜓 𝐵 (𝛼) := 𝜓 𝐹 (trd(𝛼))
for 𝛼 ∈ 𝐵. Again, 𝜓(𝛼𝛽) = 𝜓(𝛽𝛼) for all 𝛼, 𝛽 ∈ 𝐵.
If 𝐹 is a local number field containing Q 𝑝 , let 𝑅0 = Z 𝑝 ; if 𝐹 is a local function

field, let 𝑅0 = 𝑅.
Definition 29.7.5. The absolute discriminant of 𝐵 is
D (𝐵) := N (disc 𝑅0 (O)) = |disc 𝑅0 (O)| −1 ∈ Z>0 .
The absolute discriminant is well-defined, independent of the choice of maximal

order.
Example 29.7.6. If 𝐵 is a quaternion algebra over 𝐹, then
D (𝐵) = N (disc 𝑅0 (O)) = N (disc 𝑅0 (𝑅)) 4 N (discrd 𝑅 (O)) 2 .

∼
Proposition 29.7.7. The standard unitary character 𝜓 defines an isomorphism 𝐵 −→
∨
𝐵 of topological groups, and the measure
𝜏 := D (𝐵) −1/2 𝜇 (29.7.8)
is self-dual with respect to 𝜓.
Proof. The same arguments as in Propositions 29.6.9 and 29.6.10 apply, with appro-
priate modifications.
29.7.9. Having normalized the multiplicative measure 𝜇× and seeing the relevant
modification in (29.7.8), just as in the case of 𝐵 = 𝐹 we define
𝜏 × := D (𝐵) −1/2 𝜇× . (29.7.10)
We may now define the local zeta function in the nonarchimedean context.
Definition 29.7.11. A function 𝑓 : 𝐵 → C is Schwartz–Bruhat if 𝑓 is locally constant
(for every 𝛼 ∈ 𝐵, there exists an open neighborhood 𝑈 3 𝑥 such that 𝑓 |𝑈 is constant)
with compact support. A standard function Ψ on 𝐵 is the characteristic function of
a maximal order.
Taking an open cover of the support, we see that every Schwartz–Bruhat function
can be expressed as a (finite) C-linear combination of characteristic functions of
compact open subsets of 𝐵.
Definition 29.7.12. For a Schwartz–Bruhat function Φ, we define the (local) zeta
function ∫
𝑍 𝐵Φ (𝑠) := k𝛼k 𝑠 Φ(𝛼) d𝜏 × (𝛼). (29.7.13)
𝐵×
We write ∫
𝑍 𝐵 (𝑠) := 𝑍 𝐵Ψ (𝑠) = k𝛼k 𝑠 d𝜏 × (𝛼) (29.7.14)
O∩𝐵 ×
for the zeta function with respect to a standard function Ψ; this is well-defined (inde-
pendent of Ψ) as any two maximal orders are conjugate, and k𝛼k is well-defined on
conjugacy classes by 29.7.1.
We begin with a basic convergence estimate.
Lemma 29.7.15. The function 𝑍 𝐵Φ (𝑠) converges for Re 𝑠 > 1.
29.7. LOCAL ZETA FUNCTIONS: NONARCHIMEDEAN CASE 507
Proof. We reduce to the case where Φ is the characteristic function of a compact open
set 𝑈, since Φ is a (finite) C-linear combination of such. Since 𝑈 is compact, k𝛼k is
bounded on 𝑈, so too is k𝛼k 𝑠−1 for fixed Re 𝑠 > 1; thus
∫ ∫
k𝛼k 𝑠 d𝜏 × (𝛼) = k𝛼k 𝑠−1 d𝜏(𝛼)
𝑈 ∩𝐵 × O∩𝐵 ×
is bounded by a constant multiple of 𝜏(𝑈) < ∞, and therefore 𝑍 𝐵Φ (𝑠) is bounded by

comparison.
This integral representation recovers the classical zeta function we studied earlier
in the number field case (section 26.4).
29.7.16. Since there is a unique ideal in 𝑅 of absolute norm 𝑞 𝑒 for each 𝑒 ≥ 0, we
have
∞
∑︁ 1
𝜁 𝐹 (𝑠) = 𝑒𝑠
= (1 − 𝑞 −𝑠 ) −1 .
𝑒=0
𝑞
In like fashion, for 𝑒 ≥ 0, let 𝑎 𝔭𝑒 = 𝑎 𝔭𝑒 (O) be the number of (necessarily principal)
right ideals of O of reduced norm 𝔭𝑒 and thereby absolute norm 𝑞 𝑛𝑒 . We then define
∞
∑︁ 1 ∑︁ 𝑎 𝔭𝑒
𝜁 𝐵 (𝑠) := = ,
N(𝐼) 𝑠 𝑞 𝑛𝑒𝑠
𝐼 ⊆O 𝑒=0
the sum as usual over all nonzero right ideals of O.

(a) We have
𝑍 𝐵 (𝑠) = 𝜏 × (O× )𝜁 𝐵 (𝑠);
in particular, the domain of convergence of 𝑍 𝐵 (𝑠) is the same as 𝜁 𝐵 (𝑠).
(b) If 𝐵 is a quaternion algebra, then 𝜁 𝐵 (𝑠) converges for Re 𝑠 > 1/2, and
(
𝜁 𝐹 (2𝑠), if 𝐵 is a division algebra;
𝜁 𝐵 (𝑠) =
𝜁 𝐹 (2𝑠)𝜁 𝐹 (2𝑠 − 1), if 𝐵 ' M2 (𝐹).
(c) If 𝐵 ' M𝑛 (𝐹), then 𝜁 𝐵 (𝑠) converges for Re 𝑠 > 1 − 1/𝑛, and
𝑛−1
Ö
𝜁 𝐵 (𝑠) = 𝜁 𝐹 (𝑛𝑠 − 𝑖).
𝑖=0
Proof. First part (a). For 𝑒 ≥ 0, choose representatives for the 𝑎 𝔭𝑒 = 𝑎 𝔭𝑒 (O) classes
in O/O× with reduced norm 𝔭𝑒 . Every element 𝛼 ∈ O ∩ 𝐵× can be written as the
product of one of the representatives and an element of O× ; since k𝛼k = |nrd(𝛼) 𝑛 | we
have
∫ ∞
× × ×
∑︁ 𝑎 𝔭𝑒
𝑍 𝐵 (𝑠) = 𝑠
k𝛼k d𝜏 (𝛼) = 𝜏 (O ) 𝑛𝑒𝑠
= 𝜏 × (O× )𝜁 𝐵 (𝑠) (29.7.18)
O∩𝐵 ×
𝑒=0
𝑞
as claimed. For (b) and (c), we gave a formula for 𝑎 𝔭𝑒 for 𝐵 = M𝑛 (𝐹) in Exercise
26.12 and for 𝐵 a quaternion algebra in Corollary 26.4.7 (which extends to the function
field case without change). Of course, these identities only hold within their respective
domains of convergence, which for 𝜁 𝐹 (𝑠) = (1 − 𝑞 −𝑠 ) −1 is Re 𝑠 > 0, thereby giving
the rest.
In the proof of the functional equation, we will need the following proposition,
recalling duality (section 15.6).
Proposition 29.7.19. For Ψ the characteristic function of O, we have
(
∨ 𝜏(O) = D (𝐵) −1/2 , if 𝛼 ∈ O♯ ;
Ψ (𝛼) =
0, otherwise.
Moreover,
∨
𝑍 𝐵Ψ (𝑠) = D (𝐵) 𝑠−1/2 𝑍 𝐵Ψ (𝑠). (29.7.20)
Proof. For the first statement, by definition we have

∫
Ψ∨ (𝛼) = 𝜓(𝛼𝛽) d𝜏(𝛽).
O
If 𝛼 ∈ O♯ then 𝜓(𝛼𝛽) = 1 for all 𝛽 ∈ O, and we obtain
Ψ∨ (𝛼) = 𝜏(O) = D (𝐵) −1/2 𝜇(O) = D (𝐵) −1/2 .
Otherwise, 𝛼 ∉ O♯ , and by character theory Ψ∨ (𝛼) = 0.

Since O is maximal, we have O♯ = O𝛿 for some 𝛿 ∈ 𝐵× with k𝛿k = D (𝐵).
Therefore
∫ ∫
∨
𝑍 𝐵Ψ (𝑠) = D (𝐵) −1/2 k𝛼k 𝑠 d𝜏 × (𝛼) = D (𝐵) −1/2 k𝛼𝛿k 𝑠 d𝜏 × (𝛼)
× ×
∫O ∩𝐵 O∩𝐵
♯
𝑠−1/2 𝑠 × 𝑠−1/2 Ψ
= D (𝐵) k𝛼k d𝜏 (𝛼) = D (𝐵) 𝑍 𝐵 (𝑠),
O∩𝐵 ×
(29.7.21)
proving the second statement.
We conclude this section with some hopefully illustrative computations of measure.

Lemma 29.7.22. Let 𝐵 be a quaternion algebra over 𝐹. Then
(
1 + 1/𝑞, if 𝐵 is a division ring;
𝜇× (O× ) = (29.7.23)
1 − 1/𝑞 2 , if 𝐵 ' M2 (𝐹)
Proof. If 𝐵 is a division ring, then O is the valuation ring; let 𝐽 ⊆ O be the maximal
ideal, so O/𝐽 ' F𝑞2 hence 𝜇(𝐽) = 1/𝑞 2 , and then
1 − 1/𝑞 2
𝜇× (O× ) = (1 − 1/𝑞) −1 (𝜇(O) − 𝜇(𝐽)) = = 1 + 1/𝑞
1 − 1/𝑞
29.8. IDELIC ZETA FUNCTIONS 509
Similarly, if 𝐵 ' M2 (𝐹) then O ' M2 (𝑅), and from the exact sequence
1 → 1 + 𝔭 M2 (𝑅) → GL2 (𝑅) → GL2 (𝑘) → 1
where 𝑘 = 𝑅/𝔭, we compute 𝜇(1 + 𝔭 M2 (𝑅)) = 𝜇(𝔭 M2 (𝑅)) = 1/𝑞 4 and
𝜇× (GL2 (𝑅)) = (1 − 1/𝑞) −1 𝜇(1 + 𝔭O)# GL2 (𝑘)

(𝑞 2 − 1) (𝑞 2 − 𝑞)
= = 1 − 1/𝑞 2 .
𝑞 4 (1 − 1/𝑞)
29.7.24. Let 𝐵 be a quaternion algebra. Then the reduced norm yields an exact
sequence
nrd
1 → O1 → O× −−→ 𝑅 × → 1
so we have an induced compatible measure 𝜏 1 on O1 . Thus
𝜏 × (O× ) D (𝐵) −1/2 𝜇× (O× )

𝜏 1 (O1 ) = =
𝜏 × (𝑅 × ) D (𝑅) −1/2 𝜇× (𝑅 × )
(𝑑 𝐹 discrd(O) 2 ) −1/2 𝜇× (O× )
4
= (29.7.25)
𝑑 𝐹−1/2
(
(𝑞 − 1) −1 , if 𝐵 is a division algebra;
= 𝑑 𝐹−3/2 (1 2
− 1/𝑞 ) ·
1, if 𝐵 ' M2 (𝐹).
29.8 Idelic zeta functions

In this section, we now define (global) zeta functions in an idelic context. Let 𝐹 be a
global field and let 𝐵 be a central simple algebra over 𝐹.
We first define the classical zeta function before generalizing to idelic zeta func-
tions. We define Ö
𝜁 𝐵 (𝑠) := 𝜁 𝐵𝑣 (𝑠), (29.8.1)
𝑣-∞
the product over all nonarchimedean places of the local zeta functions 𝜁 𝐵𝑣 (𝑠) defined
in 29.7.16. We studied these (global) zeta functions over number fields in Chapters 25
and 26.
Example 29.8.2. By Lemma 29.7.17(c) we have
𝑛−1
Ö
𝜁M𝑛 (𝐹 ) (𝑠) = 𝜁 𝐹 (𝑛𝑠 − 𝑖).
𝑖=0
With this in mind, we construct idelic zeta functions as products of local zeta
functions, according to the product measures and characters, as follows.
29.8.3. For 𝑣 ∈ Pl 𝐹, we have defined measures 𝜇 𝑣 on 𝐵 𝑣 in 29.5.1 for 𝑣 archimedean

and 29.7.1 for 𝑣 nonarchimedean. We define the measure
Ö
𝜇 := 𝜇𝑣
𝑣 ∈Pl 𝐹
on 𝐵 as the product measure (on the restricted direct product). In particular, for every
eligible set S ⊂ Pl 𝐹 of places and every Î O in 𝐵, the measure 𝜇 restricts to
Î 𝑅 (S) -order
the (convergent) product measure on 𝑣 ∈S 𝐵 𝑣 × 𝑣∉S O 𝑣 —and this uniquely defines
the 𝜇 as a Haar measure on 𝐵.
Definition 29.8.4. The absolute discriminant D (𝐵) of 𝐵 is defined as the product of
the local absolute discriminants over all nonarchimedean places:
Ö
D (𝐵) := D (𝐵 𝑣 ).
𝑣-∞
The absolute discriminant is well-defined because D (𝐵 𝑣 ) = 1 for all but finitely

many 𝑣.
Example 29.8.5. If 𝐹 is number field, let 𝑑 𝐹 ∈ Z be the discriminant of 𝐹. Multiplying
the factors 29.7.6 together, if 𝐵 is a quaternion algebra we find that
(
Ö
2 | 𝑑 𝐹 | 4 , if 𝐹 is a number field;
D (𝐵) = N (𝔭) ·
𝔭∈Ram(𝐵)
1, if 𝐹 is a function field;
the product taken over all nonarchimedean ramified places.

29.8.6. In light of Proposition 29.7.7, we define the Tamagawa measure on 𝐵 by
Ö
𝜏 := D (𝐵) −1/2 𝜇 = 𝜏𝑣 , (29.8.7)
𝑣
with 𝜏𝑣 = 𝜇 𝑣 for 𝑣 archimedean and 𝜏𝑣 = D (𝐵 𝑣 ) −1/2 𝜇 𝑣 defined in (29.7.8) for 𝑣

nonarchimedean.
Remark 29.8.8. The Tamagawa measure has an intrinsic definition, independent of
local normalizations [Weil82, Chapter II]. This intrinsic definition is, among other
things, important for generalizations (which are quite broad, see Remark 29.11.9). It
is then a nontrivial effort to unravel definitions to verify that the general notion of
Tamagawa measure reduces to the one defined above. We content ourselves with the
characterization 29.8.12 below.
29.8.9. If 𝐹 is a number field, we define the product character
Ö
𝜓 := 𝜓𝑣
𝑣
on 𝐵. If 𝐹 is a function field, we align our local characters as follows: let 𝜔 ∈ Ω𝐹 /F𝑞

be a nonzero meromorphic 1-form, and for a place 𝑣 ∈ Pl 𝐹 let
𝜓 𝑣 (𝛼) = exp(2𝜋𝑖htrd(𝛼)𝜔/d𝑡 𝑣 i𝐹𝑣 )
Î
where 𝑡 𝑣 is a uniformizer at 𝑡, and we again define 𝜓 := 𝑣 𝜓 𝑣 . In both cases,
𝜓(𝛼𝛽) = 𝜓(𝛽𝛼) for all 𝛼, 𝛽 ∈ 𝐵, as this holds for 𝜓 𝑣 for all 𝑣.
In all cases, we define the function
Ö
Φ := Φ𝑣
𝑣
where Φ 𝑣 is the characteristic function of a maximal order O 𝑣 for 𝑣 nonarchimedean

and Φ 𝑣 is the standard function (Definition 29.5.13) if 𝑣 is archimedean.
Proposition 29.8.10. The product character 𝜓 has the following properties.
(a) The map

∼ 𝐵∨
𝐵 −→
(29.8.11)
𝛼 ↦→ (𝛽 ↦→ 𝜓(𝛼𝛽)
is an isomorphism of topological groups.
(b) The image of 𝐵 under (29.8.11) is the group of unitary characters that are trivial
on 𝐵.
(c) The product measure 𝜏 is self-dual with respect to 𝜓.
Proof sketch. In the number field case, the statement follows from the local versions
and the product formula; in the function field case, it follows similarly and reduces to
the fact that the sum of the residues is zero.
29.8.12. The Tamagawa measure 𝜏 is then also characterized as the unique Haar
measure on 𝐵 that is self-dual with respect to the Fourier transform associated to the
standard character.
29.8.13. Moving now from additive to multiplicative, we define the normalized Haar
measure Ö
𝜇× = 𝜇×𝑣 (29.8.14)
𝑣
on 𝐵× ,where 𝜇×𝑣
is defined in 29.5.5 for 𝑣 archimedean and (29.7.3) for 𝑣 nonar-
chimedean. Finally, we define the Tamagawa measure on 𝐵× by
Ö
𝜏 × := D (𝐵) −1/2 𝜇× = 𝜏𝑣× , (29.8.15)
𝑣
where 𝜏𝑣× = 𝜇×𝑣 for 𝑣 archimedean and 𝜏𝑣× = D (𝐵 𝑣 ) −1/2 𝜇×𝑣 as in (29.7.10) for 𝑣
nonarchimedean.
With measures finally (!) in hand, we can now idelically integrate functions against
them.
Definition 29.8.16. A Schwartz–Bruhat Î function on 𝐵 is a finite linear combination

of functions 𝑓 : 𝐵 → C with 𝑓 = 𝑣 𝑓 𝑣 , where each 𝑓 𝑣 is Schwartz–Bruhat and
𝑓 𝑣 = Ψ 𝑣 is a standard function for all but finitely many 𝑣.
Theorem 29.8.17 (Poisson summation). For a Schwartz–Bruhat function Φ, we have

∑︁ ∑︁
Φ(𝛽) = Φ∨ (𝛽).
𝛽 ∈𝐵 𝛽 ∈𝐵
Proof. Symmetrize to obtain

∑︁
(ΣΦ) (𝛼) := Φ(𝛼 + 𝛽). (29.8.18)
𝛽 ∈𝐵
Using 𝜓 to identify 𝐵 with the dual of 𝐵/𝐵 via Proposition 29.8.10 (and attending
carefully to relevant analytic concerns), the symmetrized function is equal to its Fourier
series ∑︁
(ΣΦ) (𝛼) = 𝑎 𝛽 𝜓(−𝛽𝛼) (29.8.19)
𝛽 ∈𝐵
where ∫
𝑎𝛽 = (ΣΦ) (𝛼)𝜓(𝛽𝛼) d𝜏(𝛼)
𝐵/𝐵
∫ ∑︁
= Φ(𝛼 + 𝛾)𝜓(𝛽(𝛼 + 𝛾)) d𝜏(𝛼) (29.8.20)
𝐵/𝐵 𝛾 ∈𝐵
∫
= Φ(𝛼)𝜓(𝛽𝛼) d𝜏(𝛼) = Φ∨ (𝛽)
𝐵
since 𝜓(𝛾) = 1 for all 𝛾 ∈ 𝐵. Therefore

∑︁ ∑︁
Φ(𝛽) = (ΣΦ) (0) = Φ∨ (𝛽).
Remark 29.8.21. When 𝐹 is a function field, the alignment in 𝜓 implies the Riemann–
Roch theorem for the curve with function field 𝐹: see Exercise 29.13.
Definition 29.8.22. For a Schwartz–Bruhat function Φ, we define the (idelic) zeta

function ∫
Φ
𝑍 𝐵 (𝑠) := Φ(𝛼) k𝛼k 𝑠 d𝜏 × (𝛼);
𝐵×
when Φ = Ψ is a standard function, we write simply 𝑍 𝐵 (𝑠).
We have
Φ
Ö
𝑍 𝐵 (𝑠) = 𝑍 𝐵Φ𝑣𝑣 (𝑠) (29.8.23)
𝑣 ∈Pl 𝐹
wherever the product is absolutely convergent; the local zeta functions 𝑍 𝐵𝑣 (𝑠) are
defined in (29.5.16) for archimedean 𝑣 and (29.7.13) for nonarchimedean 𝑣.
We can make the comparison to the classical zeta function (29.8.1) explicit up to
an idelic volume, as follows.

(a) If 𝐹 is a number field, then

Ö
𝜏× (O
𝑍 𝐵 (𝑠) = b b× )𝜁 𝐵 (𝑠) 𝑍 𝐵𝑣 (𝑠);
𝑣 |∞
if 𝐹 is a function field, then
𝑍 𝐵 (𝑠) = 𝜏 × (O× )𝜁 𝐵 (𝑠).
(b) Suppose 𝐵 is a quaternion algebra. If 𝐹 is a number field with 𝔇 := disc 𝐵, then
𝜏× (O
b b× ) −1 = | 𝑑 𝐹 | 2 𝜁 𝐹 (2)𝜑(𝔇) (29.8.25)
Î
where 𝜑(𝔇) = 𝔭 |𝔇 (Nm(𝔭) − 1). If 𝐹 is a function field, then
Ö
𝜏 × (O× ) −1 = 𝜁 𝐹 (2) (1 − 1/𝑞 𝑣 ).
𝑣 ∈Ram(𝐵)
Proof. For part (a), we use Lemma 29.7.17 for the relationship between zeta functions.
For part (b), we combine Lemmas 29.5.18 and 29.7.17(b) and we use Lemma 29.7.22
for the local computation of measure and the equality
Ö √︁
(1 − N(𝔭) −1 ) D (𝐵) = | 𝑑 𝐹 | 2 𝜑(𝔇)
𝔭 |𝔇
in the number field case.
29.8.26. For the modulus k k, we find that

Ö
k𝛼k = k𝛼 𝑣 k 𝑣 .
𝑣
In particular, recalling 27.6.11 we have the group
𝐵 (1) = 𝛼 ∈ 𝐵× : k𝛼k = 1 ,

and by the product formula we have 𝐵× ≤ 𝐵 (1) .

We now restrict the measure on 𝐵 (1) to 𝐵× . Let 𝑛2 = dim𝐹 𝐵. Then
k𝛼k 𝐵 = knrd(𝛼) 𝑛 k 𝐹 .
We have an exact sequence
1 → 𝐵 (1) → 𝐵× →
− k𝐵× k → 1; (29.8.27)
we have k𝐵× k = R>0 if 𝐹 is a number field and k𝐵× k ≤ 𝑞 𝑛Z if 𝐹 is a function field

with constant field F𝑞 . Noting this, we take the measure on k𝐵× k defined by 𝑛−1 d𝑡/𝑡
in both cases. By compatibility (see 29.3.13), we obtain a measure 𝜏 (1) on 𝐵 (1) .
29.9 Convergence and residue

In this section, we establish convergence properties of idelic zeta functions, and com-
pute their residue at 𝑠 = 1. Throughout, 𝐹 is a global field and 𝐵 is a central simple
algebra over 𝐹. Let 𝑛2 = dim𝐹 𝐵; we allow 𝑛 = 1, so 𝐵 = 𝐹.
Proposition 29.9.1. Let Φ be a Schwartz–Bruhat function on 𝐵. Then the following
statements hold.
Φ
(a) 𝑍 𝐵 (𝑠) is absolutely convergent for Re(𝑠) > 1.
Φ
(b) 𝑍 𝐵 (𝑠) has a simple pole at 𝑠 = 1 with residue
Φ Φ∨ (0)𝜁 𝐹∗ (1)
res𝑠=1 𝑍 𝐵 (𝑠) =
𝑛
where 𝜁 𝐹∗ (1) = res𝑠=1 𝜁 𝐹 (1).
Proof. In a nutshell, we prove this proposition by comparison to the classical zeta
function 𝜁M𝑛 (𝐹 ) (𝑠) of the matrix ring, for which we can compare explicitly with a
Dirichlet series. Î
We may suppose without loss of generality that Φ = 𝑣 Φ 𝑣 , since by definition Φ
is a linear combination of such. Let 𝑆 ⊆ Pl(𝐹) be the (finite) set of places 𝑣 of 𝐹 such
that one of the following holds:
• 𝑣 is archimedean (if 𝐹 is a number field);
• 𝐵 𝑣 ; M𝑛 (𝐹𝑣 ); or
• Φ 𝑣 ≠ Ψ 𝑣 , i.e., Φ 𝑣 is not the standard function.
Then
Φ Φ
Ö Ö
𝑍 𝐵 (𝑠) = 𝑍 𝐵𝑣 (𝑠) 𝑍M𝑛 (𝐹𝑣 ) (𝑠). (29.9.2)
𝑣 ∈𝑆 𝑣∉𝑆
where the first product is finite. By Lemma 29.7.15 for 𝑣 ∈ 𝑆 archimedean and com-
parison to Lemma 29.5.18 for 𝑣 archimedean, the first product is absolutely convergent
for Re 𝑠 > 1. For the second (infinite) product, by Lemma 29.7.17(c), we have
𝑛−1
Ö 𝑛−1
Ö
𝜁M𝑛 (𝐹𝑣 ) (𝑠) = 𝜁 𝐹𝑣 (𝑛𝑠 − 𝑖) = (1 − 𝑞 𝑖−𝑛𝑠 ) −1 (29.9.3)
𝑖=0 𝑖=0
where 𝑞 𝑣 is the size of the residue field of 𝐹𝑣 ; note 𝜁M𝑛 (𝐹𝑣 ) (𝑠) −1 = 𝑖 (1 − 𝑞 𝑖−𝑛𝑠 ) is
Î
holomorphic. Putting these together gives
Ö Ö
𝜏× (O
𝑍M𝑛 (𝐹𝑣 ) (𝑠) = b b× )𝜁M (𝐹 ) (𝑠)
𝑛 𝑍M𝑛 (𝐹𝑣 ) (𝑠) −1 .
𝑣∉𝑆 𝑣 ∈𝑆
and so we reduce to showing that 𝜁M𝑛 (𝐹 ) (𝑠) is absolutely convergent for Re(𝑠) > 1.
But multiplying (29.9.3) gives
𝑛−1
Ö
𝜁M𝑛 (𝐹 ) (𝑠) = 𝜁 𝐹 (𝑛𝑠 − 𝑖)
𝑖=0
29.9. CONVERGENCE AND RESIDUE 515
which is absolutely convergent (by comparison to the harmonic series) whenever

𝑛(Re 𝑠) − (𝑛 − 1) = 𝑛(Re 𝑠) − 𝑛 + 1 > 1, i.e., when Re 𝑠 > 1. This proves part (a).
To prove part (b), we dig deeper. Staring at (29.9.3), we identify the terms
𝜁 𝐹𝑣 (𝑛𝑠 − 𝑛 + 1) that contribute to a pole at 𝑠 = 1; by absolute convergence, for Re 𝑠 > 1
we have
Φ Φ𝑣 !
𝑍 𝐵 (𝑠) Ö
Φ𝑣
Ö 𝑍 𝐵𝑣 (𝑠) Ö 𝑛−2
Ö
× ×
= 𝑍 𝐵𝑣 (𝑠) 𝜏 (O ) 𝜁 𝐹 (𝑛𝑠 − 𝑖) .
𝜁 𝐹 (𝑛𝑠 − 𝑛 + 1) 𝜁 (𝑛𝑠 − 𝑛 + 1) 𝑣∉𝑆 𝑣 𝑣 𝑖=0 𝑣
𝑣 ∈𝑆 𝐹𝑣
𝑣 |∞
𝑣-∞
(29.9.4)
The first two products are finite and defined at 𝑠 = 1, and the third (infinite) product
is now absolutely convergent for Re 𝑠 > 1 − 1/𝑛, so in a neighorhood of 𝑠 = 1.
Accordingly, to compute the desired residue, we may compute a limit.
Let 𝑣 be a nonarchimedean place of 𝐹 and let 𝑞 𝑣 be the size of its residue field.
For Re 𝑠 > 1, we have
∫
Φ𝑣
𝑍 𝐵𝑣 (𝑠) = Φ 𝑣 (𝛼 𝑣 ) k𝛼 𝑣 k 𝑠 d× 𝛼 𝑣
𝐵 ×𝑣
∫ (29.9.5)
= Φ 𝑣 (𝛼 𝑣 ) k𝛼k d𝛼 𝑣 (1 − 𝑞 −1
𝑠−1
𝑣 ) −1
𝐵 ×𝑣
coming from the normalization factors between additive and multiplicative Haar mea-
sure at the nonarchimedean places, where 𝑞 𝑣 is the size of the residue field of 𝐹𝑣 .
Therefore
Φ𝑣
𝑍 𝐵𝑣 (𝑠) 1 − 𝑞 −𝑛𝑠+𝑛−1
𝑣
∫
lim = lim Φ 𝑣 (𝛼 𝑣 ) k𝛼k 𝑠−1 d𝛼 𝑣
𝑠&1 𝜁 𝐹𝑣 (𝑛𝑠 − 𝑛 + 1) 𝑠&1 1 − 𝑞 −1
𝑣 𝐵 ×
𝑣 (29.9.6)
∫
∨
= Φ 𝑣 (𝛼 𝑣 ) d𝛼 𝑣 = Φ 𝑣 (0).
𝐵𝑣
In a similar way, if 𝑣 is archimedean we have (without annoying normalization factors)

∫
Φ𝑣
lim 𝑍 𝐵𝑣 (𝑠) = Φ 𝑣 (𝛼 𝑣 ) d𝛼 𝑣 = Φ∨𝑣 (0).
𝑠&1 𝐵𝑣
Applying (29.9.6), we conclude

Φ
𝑍 𝐵 (𝑠) Ö Φ𝑣
Ö 𝑍 𝐵Φ𝑣𝑣 (𝑠)
lim = lim 𝑍 (𝑠) lim
𝑠&1 𝜁 𝐹 (𝑛𝑠 − 𝑛 + 1) 𝑠&1 𝐵 𝑣 𝑠&1 𝜁 𝐹𝑣 (𝑛𝑠 − 𝑛 + 1)
𝑣 |∞
Ö
𝑣-∞ (29.9.7)
= Φ∨𝑣 (0) = Φ∨ (0);
𝑣
the interchange of the product and the limit is justified by absolute convergence of the
product in a neighborhood of 𝑠 = 1. Therefore
Φ Φ∨ (0)𝜁 𝐹∗ (1)
res𝑠=1 𝑍 𝐵 (𝑠) = Φ∨ (0) res𝑠=1 𝜁 𝐹 (𝑛𝑠 − 𝑛 + 1) =
𝑛
as claimed, finishing the proof of (b).
29.10 Main theorem

We now establish the main analytic properties of the idelic zeta function, including
meromorphic continuation and evaluation of residues. Our basic reference is Weil
[Weil82, Section III.1]; see Remark 29.10.24 for historical comments and further
references. The proof follows the same strategy as in section 29.2, with a key role
played by Poisson summation and conceptual clarity brought by idelic methods.
Main Theorem 29.10.1. Let 𝐹 be a global field, let 𝐵 be a central division algebra
over 𝐹 with 𝑛2 = dim𝐹 𝐵. Let Φ be a Schwartz–Bruhat function on 𝐵. Then the
following statements hold.
Φ
(a) The function 𝑍 𝐵 (𝑠) has meromorphic continuation to C. Moreover:
Φ
(i) If 𝐹 is a number field, then 𝑍 𝐵 (𝑠) is holomorphic in C r {0, 1} with simple
poles at 𝑠 = 0, 1 and residues
Φ Φ∨ (0)𝜁 𝐹∗ (1) Φ Φ(0)𝜁 𝐹∗ (1)

res𝑠=1 𝑍 𝐵 (𝑠) = , res𝑠=0 𝑍 𝐵 (𝑠) = − .
𝑛 𝑛
Φ
(ii) If 𝐹 is a function field with field of constants F𝑞 , then 𝑍 𝐵 (𝑠) is holomorphic
in C r {𝑠 : 𝑞 𝑠 = 𝑞 0 , 𝑞 1 } with simple poles when 𝑞 𝑠 = 𝑞 0 , 𝑞 1 and residues
Φ Φ∨ (0)𝜁 𝐹∗ (1) Φ Φ(0)𝜁 𝐹∗ (1)

res𝑞 𝑠 =𝑞1 𝑍 𝐵 (𝑠) = , res𝑞 𝑠 =𝑞0 𝑍 𝐵 (𝑠) = − .
𝑛 𝑛
Φ
(b) 𝑍 𝐵 (𝑠) satisfies the functional equation
Φ Φ∨
𝑍 𝐵 (1 − 𝑠) = 𝑍 𝐵 (𝑠). (29.10.2)
(c) We have
𝜏 (1) (𝐵× \𝐵 (1) ) = 𝜁 𝐹∗ (1).
The value of the residue 𝜁 𝐹∗ (1) is given by the analytic class number formula
(Theorem 26.2.3). Taking 𝑛 = 1 gives the following important special case (without
requiring this as input).
Corollary 29.10.3. Let 𝐹 be a global field and let 𝜁 𝐹 (𝑠) := 𝑣-∞ (1 − 𝑞 −𝑠 −1

Î
𝑣 ) , the
product over all nonarchimedean places of 𝐹 and 𝑞 𝑣 the cardinality of the residue
field at 𝑣. Then the following statements hold.
(a) If 𝐹 is a number field with absolute discriminant 𝑑 𝐹 , let 𝜉 𝐹 (𝑠) be the completed
Dedekind zeta function as defined in (26.8.4). Then 𝜉 𝐹 (𝑠) is holomorphic in
C r {0, 1}, with simple poles at 𝑠 = 0, 1, and satisfies the functional equation
𝜉 𝐹 (1 − 𝑠) = 𝜉 𝐹 (𝑠).
(b) If 𝐹 is a function field of genus 𝑔 with field of constants F𝑞 , let 𝜉 𝐹 (𝑠) :=
𝑞 (𝑔−1)𝑠 𝜁 𝐹 (𝑠). Then 𝜁 𝐹 (𝑠) is holomorphic in C r {𝑠 : 𝑞 𝑠 = 𝑞 0 , 𝑞 1 }, with simple
poles when 𝑞 𝑠 = 𝑞 0 , 𝑞 1 , and satisfies the functional equation 𝜉 𝐹 (1 − 𝑠) = 𝜉 𝐹 (𝑠).
29.10. MAIN THEOREM 517
Î
Proof. Taking 𝑛 = 1 gives 𝐵 = 𝐹; we take Φ = Ψ = 𝑣 Ψ 𝑣 to be the standard
function.
In the number field case, we have 𝜉 (𝑠) = | 𝑑 𝐹 | 𝑠/2 𝑍 𝐹 (𝑠). By Lemma 29.8.24 and
Lemma 29.5.18 for the archimedean contribution, we see that 𝜉 𝐹 (𝑠) = 𝑐 𝐹 | 𝑑 𝐹 | 𝑠/2 𝑍 𝐹 (𝑠)
where 𝑐 𝐹 is a constant depending on 𝐹. We accordingly conclude holomorphicity
from Main Theorem 29.10.1(a). For the functional equation, by (29.7.20), we have
𝑍 𝐹𝑣 (𝑠) = D (𝐹𝑣 ) 𝑠−1/2 𝑍 𝐹𝑣 (𝑠). Taking the product and reading off Main Theorem
29.10.1(b) gives
𝜉 (1 − 𝑠) = 𝑐 𝐹 | 𝑑 𝐹 | (1−𝑠)/2 𝑍 𝐹 (1 − 𝑠) = | 𝑑 𝐹 | (1−𝑠)/2+𝑠−1/2 𝑍 𝐹 (𝑠) = 𝑐 𝐹 | 𝑑 𝐹 | 𝑠/2 𝑍 𝐹 = 𝜉 (𝑠).
The function field case in concluded in a similar but more direct manner.
Main Theorem 29.10.1 extends to the case of a matrix algebra (over a division
algebra), with some additional complications because of the existence of zerodivisors:
see Theorem 29.10.20 for the case 𝐵 = M2 (𝐹). As we will see, the proof is close
to uniform in the number field and function field cases; we have separated these two
cases in the statement for clarity, so in particular the poles in the function field case
occur at 𝑠 ∈ 2𝜋𝑖(log 𝑞)Z and 𝑠 ∈ 1 + 2𝜋𝑖(log 𝑞)Z.
Remark 29.10.4. Weil [Weil82] chooses a different normalization, giving a functional
equation relating 𝑠 to 𝑛 − 𝑠 and with residues Φ∨ (0)𝜁 𝐹∗ (1) at 𝑠 = 0 and −Φ(0)𝜁 𝐹∗ (1)
at 𝑠 = 𝑛.
Proof of Main Theorem 29.10.1. We work with Tamagawa measure throughout, so to

ease notation we abbreviate d× := d𝜏 × . To break up the integral, we define


 1, 0 < 𝑡 < 1;


𝜆(𝑡) := 1/2, 𝑡 = 1; (29.10.5)

 0,

𝑡 > 1.

We break up 𝐵× by 𝜆 according to the norm: for 𝛼 ∈ 𝐵× , we define
𝑓+ (𝛼) := 𝜆(k𝛼k −1 ),
(29.10.6)
𝑓− (𝛼) := 𝜆(k𝛼k).
Then 𝑓+ (𝛼−1 ) = 𝑓− (𝛼) and 𝑓+ (𝛼) + 𝑓− (𝛼) = 1 for all 𝛼 ∈ 𝐵× , and so defining
∫
Φ,±
𝑍 𝐵 (𝑠) = 𝑓± (𝛼) k𝛼k 𝑠 Φ(𝛼) d× 𝛼, (29.10.7)
𝐵×
we obtain
Φ Φ,+ Φ,−
𝑍 𝐵 (𝑠) = 𝑍 𝐵 (𝑠) + 𝑍 𝐵 (𝑠). (29.10.8)
Φ,+
We claim that the function 𝑍 𝐵 (𝑠) is holomorphic. Indeed, by Proposition 29.9.1,
Φ Φ,+
𝑍 𝐵 (𝑠) converges absolutely for Re 𝑠 > 1; thus the same is true for 𝑍 𝐵 (𝑠). But if
Φ,+
𝑍 𝐵 (𝑠) converges absolutely at 𝑠 = 𝑠0 , then it does so for all Re(𝑠) ≤ Re(𝑠0 ), so
Φ,+
𝑍 𝐵 (𝑠) is holomorphic in C.
Φ,−
For the remaining piece 𝑍 𝐵 (𝑠), we have
∫ !
Φ,−
∑︁
𝑍 𝐵 (𝑠) = 𝑓− (𝛼) k𝛼k 𝑠
Φ(𝛽𝛼) d× 𝛼. (29.10.9)
𝐵 × \𝐵 × 𝛽 ∈𝐵 ×
Poisson summation (Theorem 29.8.17) gives

∑︁ ∑︁
Φ(𝛽) = Φ∨ (𝛽).
For 𝛼 ∈ 𝐵× , let Φ 𝛼 (𝛽) = Φ(𝛽𝛼) be the right translate. Then

∫
Φ∨𝛼 (𝛽) = Φ(𝛾𝛼)𝜓(𝛽𝛾) d𝜏(𝛾)
𝐵
∫ (29.10.10)
1 1 ∨ −1
= Φ(𝛾𝛼)𝜓(𝛽𝛾𝛼−1 ) d𝜏(𝛾) = Φ (𝛼 𝛽)
k𝛼k 𝐵 k𝛼k
using that 𝜓 is well-defined on conjugacy classes in the last equality. Plugging this
into Theorem 29.8.17 gives
∑︁ 1 ∑︁ ∨ −1
Φ(𝛽𝛼) = Φ (𝛼 𝛽). (29.10.11)
𝛽 ∈𝐵
k𝛼k 𝛽 ∈𝐵
At this point, we use the hypothesis that 𝐵 is a division algebra over 𝐹. With this
assumption in hand, ∑︁ ∑︁
Φ(𝛽𝛼) = Φ(𝛽𝛼) − Φ(0).
𝛽 ∈𝐵 × 𝛽 ∈𝐵
From (29.10.11) applied to the sum in (29.10.9), we obtain

∫ !
Φ,−
∑︁
𝑍 𝐵 (𝑠) = 𝑓− (𝛼) k𝛼k 𝑠−1 Φ∨ (𝛼−1 𝛽) + Φ∨ (0) − Φ(0) k𝛼k d× 𝛼
𝐵 × \𝐵 × 𝛽 ∈𝐵 ×
(29.10.12)
valid for Re(𝑠) > 1.
Next, we make the substitution
𝑠 ← 1 − 𝑠, 𝛼 ← 𝛼−1
Φ∨ ,+
in the definition of 𝑍 𝐵 (𝑠) (29.10.7). The Tamagawa measure d× 𝛼 is invariant under
inversion 𝛼 ← 𝛼−1 , so
∫
Φ∨ ,+
𝑍 𝐵 (1 − 𝑠) = 𝑓+ (𝛼−1 ) k𝛼k −(1−𝑠) Φ∨ (𝛼−1 ) d× 𝛼
𝐵×
∫ (29.10.13)
= 𝑓− (𝛼) k𝛼k 𝑠−1 Φ∨ (𝛼−1 ) d× 𝛼.
𝐵×
We break up this integral according to its dependence on 𝐵× as follows. We write

∫ !
Φ∨ ,+
∑︁
𝑍 𝐵 (1 − 𝑠) = 𝑓− (𝛼) k𝛼k 𝑠−1
Φ (𝛼 𝛽) d× 𝛼
∨ −1
(29.10.14)
𝐵 × \𝐵 × 𝛽 ∈𝐵 ×
(replacing 𝛽−1 ← 𝛽 in the sum): in writing the integral this way, we integrate
over any measurable set 𝐵× that maps injectively under the continuous quotient map
𝐵× → 𝐵× \𝐵× . Combining (29.10.12) and (29.10.14), we obtain
∫
Φ,− Φ∨ ,+
𝑍 𝐵 (𝑠) − 𝑍 𝐵 (1 − 𝑠) = 𝜈(k𝛼k) d× 𝛼 (29.10.15)
𝐵 × \𝐵 ×
where
𝜈(𝑡) := Φ∨ (0)𝑡 𝑠−1 − Φ(0)𝑡 𝑠 𝜆(𝑡).
The function 𝜈 only depends on lengths in 𝐵× . Recalling 29.8.26, in particular the
exact sequence (29.8.27), we obtain
∫ ∫ !
× (1) × (1) ×
𝜈(k𝛼k) d 𝛼 = 𝜏 (𝐵 \𝐵 ) 𝜈(𝑡) d 𝑡 (29.10.16)
𝐵 × \𝐵 × k𝐵 × k
where as in the previous paragraph,

∫
𝜏 (1) (𝐵× \𝐵 (1) ) = d𝜏 (1) 𝛼
𝐵 × \𝐵 (1)
is the volume of any measurable set in 𝐵 (1) that maps injectively under the quotient
map.
When 𝐹 is a number field,
Φ∨ (0)
∫ ∫ 1 ∫ 1
d𝑡 Φ(0) d𝑡
𝜈(𝑡) d× 𝑡 = 𝑡 𝑠−1 − 𝑡𝑠
k 𝐵× k 𝑛 0 𝑡 𝑛 0 𝑡
∨ (29.10.17)
1 Φ (0) Φ(0)
=− + .
𝑛 1−𝑠 𝑠
When 𝐹 is a function field with constant field F𝑞 ,
∞
! ∞
!
Φ∨ (0) 1 ∑︁
∫
−𝑑 (𝑠−1) Φ(0) 1 ∑︁ −𝑑𝑠
𝜈(𝑡) d𝑡 = + 𝑞 − + 𝑞
k𝐵 × k 𝑛 2 𝑑=1 𝑛 2 𝑑=1
(29.10.18)
1 + 𝑞 −𝑠

1 ∨ 1 + 𝑞 𝑠−1
=− Φ (0) + Φ(0) .
2𝑛 1 − 𝑞 𝑠−1 1 − 𝑞 −𝑠
We now finish the proof, beginning with the case that 𝐹 is a number field. Com-
bining (29.10.8), (29.10.15), and (29.10.17), we obtain
Φ Φ,+ Φ∨ ,+
𝑍 𝐵 (𝑠) = 𝑍 𝐵 (𝑠) + 𝑍 𝐵 (1 − 𝑠)
1
∨
Φ (0) Φ(0)
(29.10.19)
− 𝜏 (1) (𝐵× \𝐵 (1) ) + .
𝑛 1−𝑠 𝑠
The substitution 𝑠 ← 1 − 𝑠 interchanges Φ and Φ∨ , since (Φ∨ ) ∨ (𝑠) = Φ(𝑠); from this
Φ
we conclude the functional equation (b) and that 𝑍 𝐵 (𝑠) has meromorphic continuation.
Φ,+ Φ∨ ,+ Φ
Since 𝑍 𝐵 (𝑠) and 𝑍 𝐵 (𝑠) are entire, we conclude that 𝑍 𝐵 (𝑠) is holomorphic in C
away from 𝑠 = 0, 1. Proposition 29.9.1(b) gives
Φ Φ∨ (0)𝜁 𝐹∗ (1)
𝑛
which proves (a); on the other hand, (29.10.19) tells us
Φ 𝜏 (1) (𝐵× \𝐵 (1) )Φ∨ (0)

𝑛
We conclude that 𝜏 (1) (𝐵× \𝐵 (1) ) = 𝜁 𝐹∗ (1), which proves (c).
The case where 𝐹 is a function field follows in a similar manner.
The above extends to the case of a matrix algebra over a division algebra, with
additional complications. To keep our eye on the prize, we treat just the case 𝐵 =
M2 (𝐹).
Theorem 29.10.20. Let 𝐹 be a global field and let 𝐵 = M2 (𝐹). Let Φ be a Schwartz–
Bruhat function on 𝐵. Then the conclusions of Main Theorem 29.10.1 hold with the
Φ
exception that 𝑍 𝐵 (𝑠) may also fail to be holomorphic at 𝑠 = 1/2 (if 𝐹 is a number
field) or 𝑞 𝑠 = 𝑞 1/2 (if 𝐹 is a function field).
Proof. Let 𝐵1 := (𝑎, 𝑏 1 | 𝐹) and 𝐵2 = (𝑎, 𝑏 2 | 𝐹) be nonisomorphic division
quaternion algebras with Ram(𝐵1 ) ∩Ram(𝐵2 ) = ∅, neither ramified at an archimedean
place, and let 𝐵3 = (𝑎, 𝑏 1 𝑏 2 | 𝐹). Then by multiplicativity of the Hilbert symbol, we
have Ram(𝐵3 ) = Ram(𝐵1 ) t Ram(𝐵2 ). For 𝑖 = 1, 2, 3 and for each 𝑣, let Φ𝑖,𝑣 = Φ 𝑣
if 𝑣 ∉ Ram(𝐵𝑖 ) and Φ𝑖,𝑣 = Ψ 𝑣 the standard function if 𝑣 ∈ Ram(𝐵𝑖 ), and let
Î
Φ𝑖 = 𝑣 Φ𝑖,𝑣 . Then
Φ1 Φ2
Φ 𝑍 𝐵1 (𝑠)𝑍 𝐵2 (𝑠)
𝑍 𝐵 (𝑠) = Φ3
. (29.10.21)
𝑍 𝐵3 (𝑠)
Φ
The meromorphic continuation and functional equation for 𝑍 𝐵 then follow from the
Φ𝑖 Φ
corresponding properties of 𝑍 𝐵𝑖 (𝑠) and (29.10.21); we conclude also that 𝑍 𝐵 (𝑠) has
the claimed simple poles, and the values at the residues hold peeling off the factor
𝜁 𝐹∗ (1)/2 because they hold at each 𝑣.
Φ
To conclude parts (a) and (b), we show that 𝑍 𝐵 (𝑠) is holomorphic away from 𝑠 =
0, 1/2, 1 (if 𝐹 is a number field) and 𝑞 𝑠 = 𝑞 1/2 (if 𝐹 is a function field). Decomposing
the sum, we may suppose without loss of generality that Φ 𝑣 = Ψ 𝑣 is a standard function
for all but finitely many places 𝑣. Let 𝐵 0 be a division quaternion algebra over 𝐹 that
is unramified at archimedean places and all places where Φ 𝑣 ≠ Ψ 𝑣 . Then
Φ Φ0
Ö 𝑍 𝐵Ψ𝑣𝑣 (𝑠)
𝑍 𝐵 (𝑠) = 𝑍 𝐵0 (𝑠) Ψ0
. (29.10.22)
𝑣 ∈Ram(𝐵0 ) 𝑍 𝐵0𝑣 (𝑠)
𝑣
𝑍 𝐵Ψ𝑣𝑣 (𝑠)
Ψ0
= 𝜁 𝐹𝑣 (2𝑠 − 1) = (1 − 𝑞 1−2𝑠
𝑣 ) −1
𝑍 𝐵0𝑣 (𝑠)
𝑣
which is holomorphic away from 𝑞 𝑠𝑣 = 𝑞 1/2

𝑣 . When further 𝐹 is a number field, we
vary 𝐵 0 to conclude holomorphicity away from 𝑠 = 1/2 (Exercise 29.10).
To keep things tidy, we conclude the proof of part (c) in Theorem 29.11.3.
As an important special case, we have the following theorem, recalling the calcu-
lation of D (𝐵) (Example 29.8.5).
Theorem 29.10.23. Let 𝐵 be a quaternion algebra over 𝐹 and let D (𝐵) ∈ Z>0 be the
absolute discriminant of 𝐵. Then the zeta function 𝑍 𝐵 (𝑠) has the following properties.
(a) 𝑍 𝐵 (𝑠) has meromorphic continuation to C. More precisely:

(i) If 𝐹 is a number field, then 𝑍 𝐵 (𝑠) is holomorphic away from 𝑠 = 0, 1/2, 1
with simple poles at 𝑠 = 0, 1 and residues
𝜁 ∗ (1) 𝜁 𝐹∗ (1)
res𝑠=1 𝑍 𝐵 (𝑠) = √︁𝐹 , res𝑠=0 𝑍 𝐵 (𝑠) = − ;
2 D (𝐵) 2
moreover, 𝑍 𝐵 (𝑠) is holomorphic at 𝑠 = 1/2 if and only if 𝐵 is a division

algebra.
(ii) If 𝐹 is a function field of a curve of genus 𝑔 over F𝑞 , then 𝑍 𝐵 (𝑠) is
holomorphic away from 𝑞 𝑠 = 𝑞 0 , 𝑞 1/2 , 𝑞 1 , with simple poles at 𝑞 𝑠 = 𝑞 0 , 𝑞 1
and residues
𝜁 𝐹∗ (1) 𝜁 𝐹∗ (1)
res𝑞 𝑠 =𝑞1 𝑍 𝐵 (𝑠) = √︁ , res𝑞 𝑠 =𝑞0 𝑍 𝐵 (𝑠) = − ;
2𝑞 2(2−2𝑔) D (𝐵) 2
moreover, 𝑍 𝐵 (𝑠) is holomorphic where 𝑞 𝑠 = 𝑞 1/2 if and only if 𝐵 is a

division algebra.
(b) 𝑍 𝐵 (𝑠) satisfies a functional equation. More precisely:
(i) If 𝐹 is a number field, then
𝑍 𝐵 (1 − 𝑠) = D (𝐵) 1/2−𝑠 𝑍 𝐵 (𝑠).
(ii) If 𝐹 is a function field, then

1/2−𝑠
𝑍 𝐵 (1 − 𝑠) = 𝑞 4(2−2𝑔) D (𝐵) 𝑍 𝐵 (𝑠).
Proof. We apply Main Theorem 29.10.1 and Theorem 29.10.20 with Φ as in 29.8.9
with Φ 𝑣 the characteristic function of a maximal order or the standard function ac-
cording as 𝑣 is nonarchimedean or archimedean, so Φ(0) = 1.
In the number field case, we have Φ∨𝑣 = Φ 𝑣 self-dual when 𝑣 is archimedean, and
by (29.7.20) we have
Φ∨
𝑍 𝐵𝑣𝑣 (𝑠) = D (𝐵 𝑣 ) 1/2−𝑠 𝑍 𝐵Φ𝑣 (𝑠)
and Φ∨𝑣 (0) = D (𝐵 𝑣 ) −1/2 when 𝑣 is nonarchimedean. Multiplying these together and
applying Theorem 29.10.1(c) gives
Φ Φ∨ Φ
𝑍 𝐵 (1 − 𝑠) = 𝑍 𝐵 (𝑠) = D (𝐵) 1/2−𝑠 𝑍 𝐵 (𝑠).
In the function field case, a similar argument holds but with the character modified
by a global differential as in 29.8.9: for the relevant additional factor, see Exercise
29.13.
To conclude, we note that when 𝐵 ' M2 (𝐹) then by Lemma 29.8.24 the nonar-
chimedean part of 𝑍 𝐵 (𝑠) is given by 𝜁 𝐹 (2𝑠)𝜁 𝐹 (2𝑠 − 1), and this has a double pole at
𝑠 = 1/2 (or accordingly 𝑞 𝑠−1/2 = 1).
Remark 29.10.24. In her 1929 Ph.D. thesis, Hey [Hey29] defined the zeta function of a
division algebra over Q, proving that it has an Euler product and functional equation—
this was a tour de force in algebraic and analytic number theory, especially at the time!
For more on Hey’s contribution and its role in the development of class field theory,
see the perspective by Roquette [Roq2006, §9], including Zorn’s observation that the
functional equation yields an analytic proof of the classification of division algebras.
Hey’s thesis was never published (though the content appears in Deuring [Deu68, VII,
§8]), and classical treatments of the zeta function for the most part gave way to the
development of Chevalley’s adeles and ideles.
Tate’s thesis [Tate67] (from 1950, published in 1967) is usually given as the
standard reference for the adelic recasting of zeta and 𝐿-functions of global fields as
above: it gives the general definition of a zeta function associated to a local field, an
integrable function, and a quasi-character [Tate67, §2]; see also the Bourbaki seminar
by Weil [Weil66]. But already in 1946, Matchett (also a student of Emil Artin) wrote
a Ph.D. thesis at Indiana University [Mat46] beginning the redevelopment of Hecke’s
theory of zeta and 𝐿-functions in terms of adeles and ideles. At the time, Iwasawa also
contributed to the development of the theory; see his more recently published letter to
Dieudonné [Iwa92].
These results were generalized to central simple algebras over local fields by
Godement [God58a, God58b], Fujisaki [Fuj58], and Weil [Weil82, Chapter III] (and
again Weil [Weil74, Chapter XI]) in the same style as Iwasawa and Tate; and then they
were further generalized (allowing representations) by Godement–Jacquet [GJ72],
providing motivation for the Langlands program.
As an example of the generalizations indicated by Remark 29.10.24, we conclude
with a slightly more general statement.
Theorem 29.10.25. Let 𝐹 be a global field, let 𝐵 be a central division algebra over 𝐹,
let Φ be a Schwartz–Bruhat function on 𝐵, and let 𝜒 : 𝐵× → C be a unitary character
such that 𝜒 restricted to 𝐹 × is trivial. Define
∫
Φ
𝐿 𝐵 (𝑠, 𝜒) := Φ(𝛼) 𝜒(𝛼) k𝛼k 𝑠 d𝜏 × (𝛼).
𝐵×
29.11. TAMAGAWA NUMBERS 523
Φ
Then the function 𝐿 𝐵 (𝑠, 𝜒) is absolutely convergent for Re 𝑠 > 1. If 𝜒 is nontrivial,
Φ
then 𝐿 𝐵 (𝑠, 𝜒) has holomorphic continuation to C and satisfies the functional equation
Φ Φ∨
𝐿 𝐵 (𝑠, 𝜒) = 𝐿 𝐵 (𝑠, 𝜒−1 ).
Proof. Proven in the same way as Main Theorem 29.10.1, just keeping track of the
character 𝜒. The (possible) residues at 𝑠 = 0, 1 (or more generally in the function field
case 𝑞 𝑠 = 𝑞 0 , 𝑞 1 ) are multiplied by
∫
𝜒−1 (𝛼) d𝜏 (1) (𝛼)
𝐵 (1) /𝐹 ×
which is zero when 𝜒 is nontrivial by character theory and therefore the 𝐿-function is
fact holomorphic at these points.
29.11 Tamagawa numbers

Building on Main Theorem 29.10.1, in this section we compute the measure of certain
quotients with respect to the normalized idelic measure above. Let 𝐵 be a quaternion
algebra over the global field 𝐹.
Lemma 29.11.1. Let Ω be the set of real ramified places in 𝐵, and let
𝐹Ω(1) := {𝑥 ∈ 𝐹 (1) : 𝑥 𝑣 > 0 for all 𝑣 ∈ Ω}.
Then the sequence

nrd
1 → 𝐵1 → 𝐵 (1) −−→ 𝐹Ω(1) → 1
is exact, giving a compatible measure on 𝐵1 denoted 𝜏 1 . Under this measure, we have
𝜏 (1) (𝐵× \𝐵 (1) ) = 𝜏 1 (𝐵1 \𝐵1 )𝜏 (1) (𝐹 × \𝐹 (1) ). (29.11.2)
Proof. The surjectivity of the reduced norm is locally established, and its kernel is
𝐵1 by definition. The image of 𝐵× under the reduced norm is 𝐹>×Ω 0 by the Hasse–
Schilling theorem of norms (Main Theorem 14.7.4), with kernel 𝐵1 . Moreover, the
natural inclusion
𝐹>×Ω 0 \𝐹>(1)
Ω0
↩→ 𝐹 × \𝐹 (1)
is also surjective by weak approximation. Putting these together, we obtain (29.11.2),
which we might be tempted to summarize in the exact sequence of pointed sets
nrd
1 → 𝐵1 \𝐵1 → 𝐵× \𝐵 (1) −−→ 𝐹 × \𝐹 (1) → 1
but we will resist the temptation.
We now prove the final result in this chapter.

Theorem 29.11.3. Let 𝐵 be a quaternion algebra over a global field 𝐹. Then
𝜏 (1) (𝐵× \𝐵 (1) ) = 𝜏 (1) (𝐹 × \𝐹 (1) ) = 𝜁 𝐹∗ (1)
and
𝜏 1 (𝐵1 \𝐵1 ) = 1.
Proof. The equality 𝜏 (1) (𝐹 × \𝐹 (1) ) = 𝜁 𝐹∗ (1) is the statement of Main Theorem
29.10.1(c) applied to 𝐵 = 𝐹. Then 𝐵 is a division algebra, the equality 𝜏 (1) (𝐵× \𝐵 (1) ) =
𝜁 𝐹∗ (1) is again Main Theorem 29.10.1(c), and the equality 𝜏 1 (𝐵1 \𝐵1 ) = 1 then follows
from (29.11.2).
To conclude, we make the appropriate modifications in the remaining case, and
suppose that 𝐵 = M2 (𝐹). Then 𝐵1 = SL2 (𝐹) and similarly 𝐵1 = SL2 (𝐹). By the exact
sequence (29.11.2), we may show 𝜏 1 (SL2 (𝐹)\ SL2 (𝐹)) = 𝜏 1 (SL2 (𝐹)/SL2 (𝐹)) = 1.
We will do Fourier analysis on 𝐹 2 , extended from 𝐹, with self-dual measure 𝜏 and
character 𝜓. Let Φ be a Schwartz function on 𝐹 2 . The Fourier transform is
∫
∨
Φ (𝑦) = Φ(𝑥)𝜓(𝑦 t · 𝑥) d𝑥
𝐹2
and Poisson summation reads

∑︁ ∑︁
Φ(𝑥) = Φ∨ (𝑦). (29.11.4)
𝑥 ∈𝐹 2 𝑦 ∈𝐹 2
The group SL2 (𝐹) acts on column vectors 𝐹 2 r {0} by left multiplication, with the
stabilizer of 𝐹 2 r {0} given by SL2 (𝐹). Thus
∫ ∫ !
∑︁
∨
Φ (0) = Φ(𝑥) d𝜏(𝑥) = Φ(𝛼𝑥) − Φ(0) d𝜏 1 (𝛼). (29.11.5)
𝐹2 SL2 (𝐹 )/SL2 (𝐹 )
𝑥 ∈𝐹 2
From (29.11.4), we derive

∑︁ 1 ∑︁ ∨
Φ(𝛼𝑥) = Φ ((𝛼t ) −1 𝑦) (29.11.6)
k𝛼k
𝑥 ∈𝐹 2 2 𝑦 ∈𝐹
and k𝛼k SL2 (𝐹 ) = 1. Plugging in (29.11.6) into (29.11.5), we have
∫ ∫
© ∑︁ ∨
Φ(𝑥) d𝜏(𝑥) = Φ ((𝛼t ) −1 𝑦) − Φ(0) ® d𝜏 1 (𝛼). (29.11.7)
ª

𝐹2 SL2 (𝐹 )/SL2 (𝐹 )
« 𝑦 ∈𝐹 2 ¬
Replacing Φ ← Φ∨ and 𝛼 ← (𝛼t ) −1 (preserving the measure) in (29.11.7) gives
∫ ∫
∨ © ∑︁
Φ(0) = Φ (𝑥) d𝜏(𝑥) = Φ(𝛼𝑦) − Φ∨ (0) ® d𝜏 1 (𝛼).
ª

𝐹2 SL2 (𝐹 )/SL2 (𝐹 )
« 𝑦 ∈𝐹 2 ¬
(29.11.8)
Subtracting (29.11.8) from (29.11.5) gives

∫
∨
Φ∨ (0) − Φ(0) d𝜏 1 (𝛼)

Φ (0) − Φ(0) =
SL2 (𝐹 )/SL2 (𝐹 )
= 𝜏 (SL2 (𝐹)/SL2 (𝐹)) (Φ∨ (0) − Φ(0));

1
choosing Φ such that Φ(0) ≠ Φ∨ (0), we obtain 𝜏 1 (SL2 (𝐹)/SL2 (𝐹)) = 1.
Remark 29.11.9. We return to Remark 29.8.8. In general, there is a natural, intrinsic

measure on the adelic points of a semisimple algebraic group G over a number field 𝐹,
called the Tamagawa measure. With respect to the Tamagawa measure, the volume
vol(G(𝐹)/G(𝐹)) is finite, and the Tamagawa number of G (over 𝐹) is defined as
𝜏(G) := vol(G(𝐹)/G(𝐹)).
For example, in the above we computed the volume for the group G associated to the
group 𝐵1 of quaternions of reduced norm 1.
In the late 1950s, Tamagawa defined the Tamagawa measure [Tam66]. Weil
[Weil82] (based on notes from lectures at Princeton 1959–1960), computed that 𝜏(𝐺) =
1 for 𝐺 a classical semisimple and simply connected group; the conjecture that this
holds in general was known as Weil’s conjecture on Tamagawa numbers. This difficult
conjecture was proven by the efforts of many people: see Scharlau [Scha2009, §2] for
a history. In particular, the calculation of the Tamagawa number of the orthogonal
group of a quadratic form recovers the Smith–Minkowski–Siegel mass formula that
computes the mass of a genus of lattice.
Exercises
1. Let 𝜇 𝑝 be the standard Haar measure on Q 𝑝 .
(a) Let 𝐷 (𝑎, 𝛿) := {𝑥 ∈ Q 𝑝 : |𝑥 − 𝑎| < 𝛿} be the open ball of radius 𝛿 ∈ R>0
around 𝑎 ∈ Q 𝑝 . Compute the measure 𝜇 𝑝 (𝐷 (𝑎, 𝛿)), and repeat with the
closed ball of radius 𝛿 ∈ R ≥0 .
(b) Show that ∫
log 𝑝
log(|𝑥| 𝑝 ) d𝜇 𝑝 (𝑥) = − .
Z𝑝 𝑝−1
2. Let 𝐹 := F𝑞 (𝑇); then 𝐹 is the function field of P1 , a curve of genus 𝑔 = 0 and
1
𝜁 𝐹 (𝑠) = .
(1 − 𝑞 −𝑠 ) (1 − 𝑞 1−𝑠 )
Let 𝐵 := M2 (𝐹). Verify directly that 𝜁 𝐵 (𝑠) as defined in (29.8.1) satisfies

𝜁 𝐵 (1 − 𝑠) = 𝑞 4−8𝑠 𝜁 𝐵 (𝑠), and then compare this with Theorem 29.10.23.
3. Let 𝐺 be a locally compact, second countable topological group, and let 𝐻 ≤ 𝐺
be a subgroup.
(a) Show that 𝐺/𝐻 is locally compact. [Hint: Use Exercise 12.4.]
(b) If 𝐻 is closed, show that 𝐺/𝐻 is Hausdorff (repeating Exercise 12.5).

[We do not need 𝐺 to be Hausdorff in this exercise.]
4. Let 𝐺 be a Hausdorff, locally compact, second countable topological group.
(a) Let 𝜇, 𝜇 0 be two Haar measures on 𝐺. Show that there exists 𝜅 ∈ R>0 such
that 𝜇 0 = 𝜅𝜇.
(b) Let 𝜇 be a left Haar measure on 𝐺. Show that 𝜇(𝐺) < ∞ if and only if 𝐺
is compact. [Hint: Use Exercise 12.4.]
I 5. Let 𝐺 be a Hausdorff, locally compact, second countable topological group. Let
𝜇 be a left Haar measure on 𝐺. Show that the modular function Δ𝐺 : 𝐺 → R>0
is a homomorphism of groups.
I 6. Let 𝐴 be a Hausdorff, locally compact, second countable topological ring, and
let 𝑎, 𝑏 ∈ 𝐴× . Show that
k𝑎𝑏k = k𝑎k k𝑏k.
7. Consider the exact sequences
𝜙
1 → {±1} → R× −
→ R×>0 → 1
where the map 𝜙 is given by either the quotient by ±1 (equivalently, the absolute
value) or by the map 𝑥 ↦→ 𝑥 2 . Equip R× and R×>0 with the standard Haar measure
d𝑥/|𝑥|. Compute the unique compatible measures on {±1} for the two choices
of 𝜙 and show they differ by a factor 2.
8. Finish the proof of Lemma 29.5.18. [Hint: It may help to use the Iwasawa
decomposition (Proposition 33.4.2 and Lemma 36.2.7).]
I 9. Prove Proposition 29.6.9, as follows. Let 𝐹 be a nonarchimedean local field
and let 𝜓 : 𝐹 → C1 be a nontrivial unitary character of 𝐹, for example the
standard unitary character. For each 𝑥 ∈ 𝐹, let 𝜓 𝑥 : 𝐹 → C1 be defined by
𝜓 𝑥 (𝑦) = 𝜓(𝑥𝑦).
(a) Show that 𝜓 𝑥 is again a unitary character of 𝐹 and that 𝜓 𝑥 is trivial if and
only if 𝑥 = 0.
(b) Show that
Ψ : 𝐹 → 𝐹∨
𝑥 ↦→ 𝜓 𝑥
defines a continuous, injective group homomorphism. [Hint: recall that
𝐹 ∨ inherits a compact-open topology, so a basis of neighborhoods of the
identity is given by { 𝑓 ∈ 𝐹 ∨ : 𝑓 (𝐾) ⊆ 𝑉 } for 𝐾 ⊆ 𝐹 compact and 𝑉 ⊆ 𝐹 ∨
an open neighborhood of 1. Given such 𝐾, 𝑉, show that there exists an
open neighborhood 𝑈 ⊆ 𝐹 of 0 such that 𝑥𝐾 ⊆ 𝜓 −1𝑉 for all 𝑥 ∈ 𝑈.]
(c) Show that Ψ(𝐹) is dense in 𝐹 ∨ . [Hint: if 𝜓 𝑥 (𝑦) = 0 for all 𝑥 ∈ 𝐹, then
𝑦 = 0.]
(d) Prove that Ψ−1 is continuous (on its image). [Hint: work in the other
direction in (b).]
(e) Show that Ψ(𝐹) is complete, hence closed, subgroup of 𝐹 ∨ . Conclude

that Ψ is an isomorphism of topological groups.
I 10. Complete the proof of Theorem 29.10.20 for 𝐹 a number field, as follows. By
choice of 𝐵 0, we showed that 𝑍 𝐵Φ (𝑠) is holomorphic except for 𝑠 = 0, 1 and
possibly when 𝑞 𝑠𝑣 = 𝑞 1/2 0
𝑣 for 𝑣 ∈ Ram(𝐵 ). Show (by varying the choice of 𝐵 )
0
Φ
that 𝑍 𝐵 (𝑠) is holomorphic away from 𝑠 = 0, 1/2, 1.
I 11. Let 𝐵 be a quaternion algebra over a global field 𝐹. Show that there exists a
compact 𝐸 ⊆ 𝐵 such that the map 𝐵 → 𝐵\𝐵 is not injective on 𝐸. [Hint:
accept the results of section 29.3 and take 𝐸 with measure 𝜇(𝐸) > 𝜇(𝐵\𝐵) and
integrate.]
12. Following the proof of Theorem 29.11.3, show that if 𝐵 = M𝑛 (𝐹) that
𝜏 (1) (𝐵× \𝐵 (1) ) = 𝜏 (1) (𝐹 × \𝐹 (1) ) = 𝜁 𝐹∗ (1)
and
𝜏 1 (𝐵1 \𝐵1 ) = 1.
13. This exercise assumes some background in algebraic curves, see e.g. Silverman
[Sil2009, Chapter II]—for a wider survey, see Rosen [Ros2002]. Let 𝐹 be
the function field of a curve 𝑋 over F𝑞 . The divisor group Div 𝑋 is the free
abelian group on the set of places 𝑣; it has a degree map deg : Div 𝑋 → Z with
deg 𝑣 = [𝑘 𝑣 : F𝑞 ], where 𝑘 𝑣 is the residue field at 𝑣 An element 𝑓 ∈ 𝐹 × has a
Í
divisor div 𝑓 = 𝑣 ordÍ 𝑣 ( 𝑓 )𝑣. A nonzero meromorphic differential 𝜔 on 𝑋 has
also a divisor 𝐾 := 𝑣 𝑎 𝑣 𝑣 where 𝑎 𝑣 = ord 𝑣 (𝜔/d𝑡Í 𝑣 ) where 𝑡 𝑣 is a uniformizer
at 𝑣; we call 𝐾 a canonical divisor. Given 𝐷 = 𝑣 𝑑 𝑣 𝑣 ∈ Div 𝑋, let
Ö
𝐿(𝐷) := 𝔭−𝑑
𝑣
𝑣
⊆ 𝐹,
𝑣
and 𝐿(𝐷) := 𝐿 (𝐷) ∩ 𝐹 = { 𝑓 ∈ 𝐾 : ord 𝑣 ( 𝑓 ) ≥ −𝑑 𝑣 } and finally ℓ(𝐷) :=

dimF𝑞 𝐿(𝐷). Define the genus of 𝑋 or of 𝐹 by 𝑔 := ℓ(𝐾) ∈ Z ≥0 .
(a) Show that the characteristic function Φ of 𝐿 (𝐷) is 𝑞 deg 𝐷−deg 𝐾 /2 times the
characteristic function of 𝐿(𝐾 − 𝐷).
(b) Apply Poisson summation to Φ to obtain
∑︁ ∑︁
1 = 𝑞 deg 𝐷−deg 𝐾 /2 1
𝑓 ∈𝐿 (𝐷) 𝑓 ∈𝐿 (𝐾 −𝐷)
1
ℓ(𝐷) = deg 𝐷 − deg 𝐾 + ℓ(𝐾 − 𝐷).
2
(c) Plug in 𝐷 = 0 to (b) to get 𝐿 (0) = F𝑞 the constant functions and conclude
deg 𝐾 = 2𝑔 − 2.
(d) Conclude the Riemann–Roch theorem
ℓ(𝐷) − ℓ(𝐾 − 𝐷) = deg 𝐷 + 1 − 𝑔. (29.11.10)
14. Prove Theorem 29.10.25.
Chapter 30
Optimal embeddings
To conclude our analytic part, we apply idelic methods to understand embeddings of

quadratic orders into quaternion orders.
30.1 ⊲ Representation numbers

A subject of classical (and continuing) interest is the number of representations of an
integer by an integral quadratic form, in particular as a sum of squares. Because of
the subject of this text, we consider quadratic forms in three and four variables where
quaternions provide insight.
Lagrange famously proved that every positive integer is the sum of four squares.
We proved Lagrange’s theorem (Theorem 11.4.3) by viewing the sum of four squares
as the reduced norm on the Lipschitz order, and concluded the argument by comparison
to the Hurwitz order, which is Euclidean. In Theorem 14.3.8, we proved Legendre’s
three-square theorem: every integer 𝑛 ≥ 0 not of the form 4𝑎 (8𝑏 + 7) is the sum of
three squares. The proof is harder for three squares than for four (see Remark 11.4.4):
our proof used the Hasse–Minkowski theorem (the local-global principle for quadratic
forms over Q) and again the fact that the Hurwitz order has class number 1. (We gave
a variant in Exercise 14.5, where we used the local-global principle for embeddings.)
In each of these cases, we may also ask for a count of the number of such repre-
sentations. For 𝑘 ∈ Z ≥1 and 𝑛 ∈ Z ≥0 , let
𝑟 𝑘 (𝑛) := #{(𝑥1 , 𝑥2 , . . . , 𝑥 𝑘 ) ∈ Z 𝑘 : 𝑥 12 + 𝑥22 + · · · + 𝑥 2𝑘 = 𝑛} (30.1.1)
be the number of ways of writing 𝑛 as the sum of √𝑘 squares; equivalently, this is

the number of lattice points on the sphere of radius 𝑛 in R 𝑘 . The number 𝑟 4 (𝑛) is
computed in terms of the factorization of 𝑛 in the Hurwitz (or Lipschitz) orders, and
has a simple answer: we saw in Exercise 11.14 that for an odd prime 𝑝,
𝑟 4 ( 𝑝) = 8( 𝑝 + 1)
and we upgraded this in Exercise 26.7 to a general formula for 𝑟 4 (𝑛) in terms of the
sum of (odd) divisors of 𝑛.
529
530 CHAPTER 30. OPTIMAL EMBEDDINGS
We may similarly ask for a formula for 𝑟 3 , but it is more difficult both to state and
to prove. Define
prim
𝑟3 (𝑛) = #{(𝑥, 𝑦, 𝑧) ∈ Z3 : 𝑥 2 + 𝑦 2 + 𝑧2 = 𝑛 and gcd(𝑥, 𝑦, 𝑧) = 1} (30.1.2)
as the number of primitive representations of 𝑛 as the sum of three squares. Then

prim
𝑟 3 (𝑛) = 𝑟 3 (𝑛) if 𝑛 is squarefree, and more generally
∑︁
prim
𝑟 3 (𝑛) = 𝑟 3 (𝑛/𝑑 2 ).
𝑑 2 |𝑛
Let ℎ(𝑑) = # Pic 𝑆 𝑑 be the class number of the quadratic order of discriminant 𝑑 < 0;
equivalently, ℎ(𝑑) is the number of reduced primitive integral positive definite binary
quadratic forms of discriminant 𝑑.
prim
Gauss [Gau86, Section 291] showed that 𝑟 3 (𝑛) is a constant multiple of ℎ(−4𝑛)
as follows.
Theorem 30.1.3 (Gauss). We have 𝑟 3 (1) = 6, 𝑟 3 (3) = 8, and for 𝑛 ∈ Z ≥0 :


 0, if 𝑛 ≡ 0, 4, 7 (mod 8);
prim


𝑟 3 (𝑛) = 12ℎ(−4𝑛), if 𝑛 ≡ 1, 2, 5, 6 (mod 8) and 𝑛 ≠ 1;

 8ℎ(−4𝑛) = 24ℎ(−𝑛), if 𝑛 ≡ 3 (mod 8) and 𝑛 ≠ 3.


(One can√ uniformly include the cases 𝑛 = 1, 3 by accounting for the extra roots of
unity in Q( −𝑛).)
Theorem 30.1.3 is a special case of Theorem 30.4.7—see Exercise 30.4—but for
historical and motivational reasons, we also give in the next section an essentially
self-contained proof for the case 𝑛 ≡ 1, 2 (mod 4), following Venkov [Ven22, Ven29].
The main obstacle to generalizing Gauss’s theorem (Theorem 30.1.3) is that quater-
nion orders need not have class number 1: a generalization with this hypothesis
“following the general plan described by Venkov” is given by Shemanske [Shem86].
Another annoyance is the growing technicality of the local computations giving the
explicit constants involved. Both of these issues are in some sense resolved by em-
ploying idelic methods (hence the placement of this chapter in this text) and even the
proof of Gauss’s theorem itself is simplified by these methods (in the next section).
The result is Theorem 30.4.7: representations are spread across the genus of an order,
with the constants given by local factors (computed in this chapter for maximal orders
and then Eichler orders).
30.1.4. For indefinite quaternion orders, strong approximation applies, and we are
almost always able to prove that the contribution to each order in the genus is equal,
with one quite subtle issue known as selectivity: in certain rare circumstances, a
quadratic order embeds in precisely half of the orders in a genus. We pursue selectivity
in the next chapter (Chapter 31): technical and rather extraordinary, it is a subject that
demands care.
Happily, a locally norm-maximal order (such as an Eichler order) over Z is not
selective!
30.2. SUMS OF THREE SQUARES 531
An important application of this theory is a refinement of the mass formula to a

class number formula. Recall the Eichler mass formula (Theorem 25.3.18): if 𝐵 is a
definite quaternion algebra over Q of discriminant 𝐷 and O ⊂ 𝐵 is an Eichler order of
level 𝑀, then
∑︁ 1 𝜑(𝐷)𝜓(𝑀)
=
𝑤𝐽 12
[𝐽 ] ∈Cls O
where 𝑤 𝐽 = #OL (𝐽)/{±1}. We can account for the necessary correction:

∑︁ 𝜑(𝐷)𝜓(𝑀) ∑︁ 1

# Cls O = 1= + 1− .
12 𝑤𝐽
[𝐽 ] ∈Cls O [𝐽 ] ∈Cls O
𝑤𝐽 >1
The latter “error term” is accounted for by (finite) cyclic subgroups spread across
representative left orders OL (𝐽)—and this is precisely the contribution computed
idelically above!
Theorem 30.1.5 (Eichler class number formula). Let 𝐵 be a definite quaternion

algebra over Q of discriminant 𝐷, and let O ⊂ 𝐵 be an Eichler order of level 𝑀.
Let 𝑁 = 𝐷 𝑀 = discrd O. Then
𝜑(𝐷)𝜓(𝑀) 𝜖 2 𝜖3
# Cls O = + +
12 4 3
where
 Ö Ö
−4

−4
1− 1+ if 4 - 𝑁;



 ,
𝜖 2 = 𝑝 |𝐷 𝑝 𝑝
𝑝 |𝑀

 0, if 4 | 𝑁;


and
 Ö Ö
−3

−3
1− 1+ if 9 - 𝑁;



 ,
𝜖 3 = 𝑝 |𝐷 𝑝 𝑝
𝑝 |𝑀

 0, if 9 | 𝑁.


30.2 Sums of three squares

In this section, we prove the theorem of Gauss showing that the number of represen-
tations of an integer as a sum of three squares is a class number. For further reading,
see also Grosswald [Gro85, Chapter 4] and the references therein.
A few parts are easy to establish. The count 𝑟 3 (1) = 6 is immediate. If 4 | 𝑛, then
prim
𝑟 3 (𝑛) = 0 since 𝑥 2 + 𝑦 2 + 𝑧 2 ≡ 0 (mod 4) implies 𝑥 ≡ 𝑦 ≡ 𝑧 ≡ 0 (mod 2). Similarly,
prim
if 𝑛 ≡ 7 (mod 8) then 𝑟 3 (𝑛) = 𝑟 3 (𝑛) = 0 by the three-square theorem.
The two remaining cases lie deeper. A proof using quaternions is due to Venkov
[Ven22, Ven29]; alternate accounts are given by Hanlon [Hanlon81, Chapter 2] and
Rehm [Reh76]. To accomplish the task of giving an argument that is as self-contained
as possible and still previews the ideas and structure contained in this chapter, we give
a proof in the case 𝑛 ≡ 1, 2 (mod 4).
:=
√ of Theorem 30.1.3
Proof √ for 𝑛 ≡ 1, 2 (mod 4). Suppose 𝑛 ≡ 1, 2 (mod 4). Let 𝑆
Z[ −𝑛] ⊂ 𝐾 := Q( −𝑛). Then 𝑆 is maximal
√ and ramified at 2, i.e., 𝑆 ⊗ Z2 is the ring
of integers of the field 𝐾 ⊗ Q2 = Q2 ( −𝑛).
Let 𝐵 = (−1, −1 | Q) be the rational Hamiltonians and O ⊂ 𝐵 the Hurwitz order.
We consider the set
𝑊 = {𝛼 = 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 ∈ O : 𝛼 is primitive and nrd(𝛼) = 𝑛}; (30.2.1)

prim
then #𝑊 = 𝑟 3 (𝑛). By the three-square theorem, 𝑊 ≠ ∅, so let 𝛼 ∈ 𝑊. We embed
𝐾 ↩→ 𝐵
√
−𝑛 ↦→ 𝛼
and for convenience identify 𝐾 with its image. By the Skolem–Noether theorem
(Corollary 7.1.5), every other element 𝛼 0 ∈ 𝑊 is the form 𝛼 0 = 𝛽−1 𝛼𝛽 with 𝛽 ∈ 𝐵×
(but not necessarily conversely!). Let
𝐸 := {𝛽 ∈ 𝐵× : 𝛽−1 𝛼𝛽 ∈ 𝑊 }. (30.2.2)
The set 𝐸 has a right action of the normalizer 𝑁 𝐵× (O) (checking that primitivity is
preserved).
We relate 𝐸 to the group of fractional ideals Idl 𝑆 as follows. Let 𝔟 ⊆ 𝐾 be a
fractional 𝑆-ideal. Since O is (right) Euclidean, 𝔟O = 𝛽O for some 𝛽 ∈ 𝐵× that is
well-defined up to right multiplication by O× . The heart of the proof is the following
claim: the map
Idl 𝑆 → 𝐸/𝑁 𝐵× (O)
(30.2.3)
𝔟 ↦→ 𝛽𝑁 𝐵× (O)
is a well-defined, surjective map of sets. The most efficient (and clear) proof of this
claim is idelic, and we prove it in two steps.
First, the map (30.2.3) is well-defined: that is to say, 𝛽 ∈ 𝐸. Write 𝔟𝑆b = b b so
𝑏 𝑆,
that 𝛽O
b=b 𝑏Ob and 𝛽 = b
𝑏b𝜇 with 𝜇b∈ O ×
b . Then b𝑏 commutes with 𝛼 (in 𝐾b ⊆ 𝐵),
b so
𝛼 0 = 𝛽−1 𝛼𝛽 = 𝜇
b−1b
𝑏 −1 𝛼b b−1 𝛼b
𝜇=𝜇
𝑏b b−1 Ob
𝜇∈𝜇 b−1 Ob
𝜇⊆𝜇 b𝜇 ∩ 𝐵 = O
b ∩ 𝐵 = O.
To show that the map is surjective, we need to establish one other important point
𝜈 ∈ 𝑁 𝐵b× ( O)
comparing the global and the idelic: we claim that there exists b b such that
𝜈 −1 𝛼b
b 𝜈 = 𝛽−1 𝛼𝛽. (30.2.4)
The existence of b 𝜈 may be established locally. We prove this in Proposition 30.5.3:

for 𝑝 ≠ 2, it amounts to showing that two elements of M2 (Z 𝑝 ) with square −𝑛 are
conjugate under GL2 (Z 𝑝 ), and for 𝑝 = 2 it follows from the description of the valuation
ring: it is here where we use the fact that 𝑆 is maximal at 2, but this is only a technical
issue. (The reader may accept (30.2.4) and proceed, or pause here and work this out,
in Exercise 30.3.)
Given (30.2.4), we see that 𝛽b𝜈 −1 centralizes 𝛼 in 𝐵× and so belongs to 𝐾 b× , whence
b× 𝑁 b× ( O).
𝛽∈𝐾 b
𝐵
30.3. OPTIMAL EMBEDDINGS 533
Finally, since 2 is ramified in 𝐾—using 𝑛 = 1, 2 (mod 4) again—and the nontrivial

class in 𝑁 𝐵b× ( O)/
b O b× ' Z/2Z is represented by an element of reduced norm 2, we have
b× 𝑁 b× ( O)
𝐾 b× O
b =𝐾 b× ,
𝐵
and therefore 𝛽 = b 𝜇 with b

𝔟b 𝔟∈𝐾 b× and 𝜇 b∈ O b× . Therefore b
𝔟Ob = 𝛽O
b and so 𝛽O = 𝔟O
where 𝔟 = 𝔟𝑆.
b b
From the claim, we conclude the theorem. We have 𝛽−1 𝛼𝛽 = 𝛼 if and only if 𝛽
centralizes 𝛼 if and only if 𝛽 ∈ 𝐾 × , so the desired elements 𝛽−1 𝛼𝛽 up to 𝑁 𝐵× (O) are
uniquely determined by the class [𝔟] ∈ Pic 𝑆 = Idl 𝑆/PIdl 𝑆. Finally, when 𝑛 > 1 we
have 𝑆 × = {±1} so #O× /𝑆 × = 12, and the result follows.
Importantly, Venkov’s proof of Gauss’s theorem given above is explicit and con-
structive, given at least one representation as a sum of three squares.
The early observation made in the proof above is that the sum of three squares is
the restriction of the reduced norm to the trace zero elements of the Hurwitz order.
One then seeks a similar statement for quadratic forms 𝑄 = nrd | O0 obtained more
generally. (This is almost the same thing as the ternary quadratic form associated to
O itself via the Clifford algebra construction in Chapter 22; the difference is that for
the latter we take the dual of the order and scale, as in (22.1.3).)
Just as in the proof above, for a representation 𝑄(𝑥,√ 𝑦, 𝑧) = 𝑛 corresponding to
𝛼 ∈ O with 𝛼2 +𝑛 = 0, we obtain an embedding 𝑆 = Z[ −𝑛] ↩→ O of a quadratic order
into the quaternion order; conversely, to such an embedding we find a representation. It
is more convenient to work with embeddings, as they possess more structure. Viewed
in this way, we may equivalently restrict the reduced norm from O to the order 𝑆
itself, and then we are asking for the representation of a binary quadratic form by a
quaternary quadratic form.
30.3 Optimal embeddings

We now begin in earnest. We start by considering quadratic embeddings into quater-
nions, both rationally and integrally. Let 𝑅 be a Dedekind domain with 𝐹 = Frac 𝑅,
and let 𝐵 be a quaternion algebra over 𝐹.
Let 𝐾 be a separable quadratic 𝐹-algebra: then either 𝐾 ⊇ 𝐹 is a separable
quadratic field extension or 𝐾 ' 𝐹 × 𝐹. Suppose that 𝐾 ↩→ 𝐵.
30.3.1. The set of embeddings of 𝐾 in 𝐵 is identified with the set 𝐾 × \𝐵× , by 7.7.12:
if 𝜙 : 𝐾 ↩→ 𝐵 is another embedding, then by the Skolem–Noether theorem there
exists 𝛽 ∈ 𝐵× such that 𝜙(𝛼) = 𝛽−1 𝛼𝛽 for all 𝛼 ∈ 𝐾 with 𝛽 well-defined up to left
multiplication by 𝐾 × , the centralizer of 𝐾 under conjugation by 𝐵× .
We now turn to the integral theory. Let O ⊆ 𝐵 be a quaternion 𝑅-order, and let
𝑆 ⊆ 𝐾 be a quadratic 𝑅-order; we will be interested in embeddings 𝜙 : 𝑆 ↩→ O. Such
an embedding gives an embedding 𝜙 : 𝐾 ↩→ 𝐵 by extending scalars. We keep the
embeddings for various suborders organized as follows.
Definition 30.3.2. An 𝑅-algebra embedding 𝜙 : 𝑆 ↩→ O is optimal if
𝜙(𝐾) ∩ O = 𝜙(𝑆).
Let
Emb𝑅 (𝑆, O) := {Optimal embeddings 𝑆 ↩→ O}. (30.3.3)
When no confusion can result, we drop the subscript 𝑅 and write simply Emb(𝑆, O).
30.3.4. If an embedding 𝜙 : 𝑆 ↩→ O is not optimal, then it is optimal for a larger order

𝑆 0 ⊇ 𝑆. Accordingly, there is a natural decomposition
Ä
{Embeddings 𝑆 ↩→ O} = Emb(𝑆 0, O). (30.3.5)
𝑆0 ⊇𝑆
Lemma 30.3.6. An 𝑅-algebra embedding 𝜙 : 𝑆 ↩→ O is optimal if and only if the

induced 𝑅𝔭 -algebra embeddings 𝑆𝔭 ↩→ O𝔭 are optimal for all primes 𝔭 ⊆ 𝑅.
Proof. Immediate from the local-global dictionary for lattices (Theorem 9.4.9).
Lemma 30.3.6 says that for an embedding 𝜙, the property of being optimal is a
local property.
We define
𝐸 = 𝐸 𝑆,O := {𝛽 ∈ 𝐵× : 𝛽−1 𝐾 𝛽 ∩ O = 𝛽−1 𝑆𝛽}
(30.3.7)
= {𝛽 ∈ 𝐵× : 𝐾 ∩ 𝛽O𝛽−1 = 𝑆}.
(The set 𝐸 also depends on the fixed embedding 𝐾 ↩→ 𝐵, but this ‘reference’ embed-
ding will remain fixed throughout.) In equation 30.3.7, we see two different ways to
think about embeddings: we either move 𝑆 and see how it fits into O, or we fix 𝐾 and
move O.
Lemma 30.3.8. The map

∼ Emb(𝑆, O)
𝐾 × \𝐸 −→
(30.3.9)
𝛽 ↦→ 𝜙 𝛽
where 𝜙 𝛽 (𝛼) := 𝛽−1 𝛼𝛽 for 𝛼 ∈ 𝑆 is a bĳection.
Proof. Immediate from 30.3.1.
We further organize our optimal embeddings up to conjugation as follows.
30.3.10. Let O1 ≤ Γ ≤ 𝑁 𝐵× (O). Then the image of Γ in 𝑁 𝐵× (O)/𝐹 × has finite index.
For example, we may take Γ = O× . (The scalars do not play a role in the theory of
embeddings: they act by the identity under conjugation.)
For 𝛾 ∈ Γ, and an optimal embedding 𝜙 ∈ Emb(𝑆, O), we obtain a new embedding
via 𝛼 ↦→ 𝛾 −1 𝜙(𝛼)𝛾, i.e., Γ acts on the right on Emb(𝑆, O) by conjugation, and
correspondingly on the right on 𝐸 by right multiplication.
Let
Emb(𝑆, O; Γ) := {Γ-conjugacy classes of optimal 𝑆 ↩→ O} (30.3.11)

30.3. OPTIMAL EMBEDDINGS 535
and
𝑚(𝑆, O; Γ) := # Emb(𝑆, O; Γ). (30.3.12)
The quantity 𝑚(𝑆, O; Γ) only depends on the type (isomorphism class) of O (trans-
porting Γ under the isomorphism of orders, of course).
By Lemma 30.3.8, there is a bĳection
∼ 𝐾 × \𝐸/Γ.
Emb(𝑆, O; Γ) −→ (30.3.13)
We conclude this section by a comparison: for groups Γ sitting between unit groups
and norm 1 unit groups, we can compare embedding numbers as follows.
Lemma 30.3.14. If O1 ≤ Γ ≤ O× , then
𝑚(𝑆, O; Γ) = 𝑚(𝑆, O; O× ) [nrd(O× ) : nrd(Γ) nrd(𝑆 × )]. (30.3.15)
Proof. We have a surjective map
Emb(𝑆, O; Γ) → Emb(𝑆, O; O× )
and the lemma amounts to looking at the fibers. From (30.3.13), we turn instead to
𝐾 × \𝐸/Γ → 𝐾 × \𝐸/O× .
For 𝛽 ∈ 𝐸, the fiber over 𝐾 × 𝛽O× is
𝐾 × \𝐾 × 𝛽O×/Γ ↔ 𝐾 𝛽× \𝐾 𝛽× O×/Γ ↔ 𝐾 𝛽× ∩ O× )\O×/Γ (30.3.16)
where 𝐾 𝛽 = 𝛽−1 𝐾 𝛽. But by hypothesis on 𝛽, we have 𝐾 𝛽× ∩ O× = 𝑆 𝛽× , so the fiber

is in bĳection with
𝐾 × \𝐾 × 𝛽O×/Γ ↔ 𝑆 𝛽× \O×/Γ.
Finally, the reduced norm gives a homomorphism O× → 𝑅 × with kernel O1 , so since

O1 ≤ Γ ≤ O× , we have
#𝑆 𝛽× \O×/Γ = # nrd(O× )/nrd(𝑆 𝛽× Γ) = [nrd(O× ) : nrd(Γ) nrd(𝑆 × )]
independent of 𝛽, giving the result.
Remark 30.3.17. The term optimal goes back at least to Schilling [Schi35], but the
notion was studied in the context of maximal orders as well by Chevalley [Chev34],
Hasse [Hass34], and Noether [Noe34]. The theory of optimal embeddings was devel-
oped thereupon by Eichler [Eic56a, §3]; many key ideas can be seen transparently in
Eichler [Eic73, Chapter II, §§3–5]. For further history up to the present, see Remark
30.6.18.
30.4 Counting embeddings, idelically: the trace formula

In this section, we give a formula the number of conjugacy classes of embeddings,
using local-global (idelic) methods. We retain notation from the previous section, but
now specialize to the case where 𝑅 is a global ring with 𝐹 = Frac 𝑅. We use adelic
(mostly idelic) notation as in 27.6.4.
We began in the previous section with a first embedding 𝐾 ↩→ 𝐵. As a reminder,
the existence of such an embedding is determined by a local-global principle as follows.
30.4.1. By the local-global principle for embeddings (Proposition 14.6.7, and 14.6.8
for the case 𝐾 ' 𝐹 × 𝐹), there exists an 𝐹-algebra embedding 𝐾 ↩→ 𝐵 if and only if
there exist 𝐹𝑣 -algebra embeddings 𝐾 𝑣 ↩→ 𝐵 𝑣 for all 𝑣 ∈ Pl 𝐹 if and only if 𝐾 𝑣 is a
field for all 𝑣 ∈ Ram 𝐵.
30.4.2. The definitions in the previous section extend to each completion. Let 𝐾
b=
𝐾 ⊗𝐹 𝐹 and similarly 𝑆 = 𝑆 ⊗𝑅 𝑅. Let
b b b
Γ = (Γ 𝑣 ) 𝑣 ≤ 𝑁 𝐵b× O
b b
be a subgroup whose image in 𝑁 𝐵b× ( O)/

b 𝐹 b× has finite index. For example, we can take
Γ the congruence closure of Γ.
b
We then analogously define
b O;
Emb𝑅b ( 𝑆, bb Γ-conjugacy classes of optimal 𝑆b ↩→ O}
Γ) := {b b (30.4.3)
and
b O;
𝑚( 𝑆, bb b O;
Γ) := # Emb𝑅b ( 𝑆, bbΓ). (30.4.4)
30.4.5. As in Lemma 30.3.8, we define
b× : 𝛽b−1 𝐾 b = 𝛽b−1 𝑆b𝛽b

𝐸b = 𝐸 b b := 𝛽b ∈ 𝐵 b𝛽b ∩ O
𝑆, O
and obtain a bĳection

b O;
Emb( 𝑆, bb ∼ 𝐾
Γ) −→ b× \𝐸/
bb Γ. (30.4.6)
under conjugation.
We now show that the number of global embeddings is counted by a class number
times the number of local embeddings. In view of Lemma 30.3.14, we may focus on
Γ=O
the case b b× . As usual, we write Cls O for the right class set of O.
Theorem 30.4.7 (Trace formula). Let ℎ(𝑆) := # Pic 𝑆. Then

∑︁
𝑚(𝑆, OL (𝐼); OL (𝐼) × ) = ℎ(𝑆)𝑚( 𝑆,
b O; b× ).
b O (30.4.8)
[𝐼 ] ∈Cls O
See Remark 30.4.24 for an explanation of the name trace formula.

30.4. COUNTING EMBEDDINGS, IDELICALLY: THE TRACE FORMULA 537
Proof. We decompose the set 𝐾 × \𝐸/

bO b× in two different ways.
First, there is a natural map
𝐾 × \𝐸/
bO b× → 𝐾
b× \𝐸/
bO b× (30.4.9)
which is a surjective map of pointed sets. The fiber of (30.4.9) over the identity
element is
𝐾 × \𝐾
b× /𝑆b× ' Pic 𝑆. (30.4.10)
We claim that the fibers of (30.4.9) may be similarly identified. Indeed, the fiber over
b× 𝛽bO
𝐾 b× consists of the double cosets 𝐾 ×b b× with 𝐾 ×b
𝜈 𝛽bO 𝜈 ∈ 𝐾 × \𝐾
b× , and
𝐾 ×b b× = 𝐾 × 𝛽bO
𝜈 𝛽bO b×
if and only if
𝜈 𝛽b = 𝜌 𝛽b
b b𝜇
with 𝜌 ∈ 𝐾 × and 𝜇 b× , if and only if

b∈ O
𝜌 −1b b𝜇 𝛽b−1 ∈ 𝐾
𝜈 = 𝛽b b× ∩ 𝛽bO
b× 𝛽b−1 = 𝑆b× (30.4.11)
b if and only if 𝐾 ×b
(since 𝛽b ∈ 𝐸), 𝜈 ⊆ 𝐾 × 𝑆b× , as claimed. From the claim and (30.4.6),
we conclude that
#(𝐾 × \𝐸/
bO b× ) = ℎ(𝑆)𝑚( 𝑆, b O;
b Ob× ). (30.4.12)
b× ∈ 𝐸/
On the other hand, each 𝛽bO bO b× defines a right O-ideal;
b intersecting with 𝐵
and organizing these right ideals by their classes, we will now show that they give rise
to optimal embeddings of the corresponding left order. For brevity, write O𝐼 = OL (𝐼).
There is a map of pointed sets
𝐾 × \𝐸/
bO b× → 𝐵× \ 𝐵
b× /O ∼ Cls O.
b× −→ (30.4.13)
Choose representatives
Ä
𝐵× \ 𝐵
b× /O
b× = 𝐵× b b×
𝛼𝐼 O (30.4.14)
[𝐼 ] ∈Cls O
𝛼𝐼 O
so that 𝐼 = b b ∩ 𝐵. Then
O𝐼 = O 𝐿 (𝐼) = b b𝛼−1 ∩ 𝐵.
𝛼𝐼 Ob𝐼
Let
𝐸 𝐼 = {𝛽 ∈ 𝐵× : 𝐾 ∩ 𝛽O𝐼 𝛽−1 = 𝑆}. (30.4.15)
Now if
b× ∈ 𝐸/
𝛽bO bO b× ,
then there exists a unique 𝐼 such that
𝐵× 𝛽bO
b× ⊆ 𝐵× b b× ,
𝛼𝐼 O
and therefore
b× = (𝛽b
𝛽bO b×
𝛼𝐼 ) O
for some 𝛽 ∈ 𝐵× . If 𝛽 0 ∈ 𝐵× is another, then
𝛽−1 𝛽 0 ∈ b b× b
𝛼𝐼 O 𝛼−1 ×
𝐼 ∩ 𝐵 = O𝐼
so the class 𝛽O×𝐼 is well-defined.

𝛼−1
We claim that 𝛽 ∈ 𝐸 𝐼 , and conversely if 𝛽 ∈ 𝐸 𝐼 then 𝛽b = 𝛽b𝐼 ∈ 𝐸. Indeed,
b
𝐾 ∩ 𝛽O𝐼 𝛽−1 = 𝐾 ∩ ( 𝛽b
b𝛼−1 b𝛼−1 ) (b
𝛼𝐼 Ob
𝐼 ) (b 𝐼 𝛼𝐼 𝛽b−1 )
b 𝛽b−1
= 𝐾 ∩ 𝛽bO
so 𝐾 ∩ 𝛽O𝐼 𝛽−1 = 𝑆 if and only if 𝐾

b ∩ 𝛽bOb 𝛽b−1 = 𝑆.
b Therefore there is a bĳection
∼ ∼
Ä Ä
𝐾 × \𝐸/
bO b× −→ 𝐾 × \𝐸 𝐼 /O×𝐼 −→ Emb(𝑆, O𝐼 ; O×𝐼 )
𝐼 𝐼 (30.4.16)
× bb× × ×
𝐾 𝛽O ↦→ 𝐾 𝛽O𝐼 .
Putting (30.4.12) and (30.4.16) together and counting, the theorem follows.
When O has class number 1, we hit the embedding number on the nose.
Corollary 30.4.17. If # Cls O = 1, then
𝑚(𝑆, O; O× ) = ℎ(𝑆)𝑚( 𝑆,
b O; b× ).
b O
Proof. Immediate from Theorem 30.4.7.
We recall that the genus of O (Definition 17.4.8) is the set Gen O of 𝑅-orders
O0 ⊆ 𝐵 locally isomorphic to O.
Corollary 30.4.18. If Emb(𝑆𝔭 , O𝔭 ) ≠ ∅ for all primes 𝔭, then there exists an order
O0 ∈ Gen O such that Emb(𝑆, O0) ≠ ∅.
Proof. By the trace formula, we have 𝔭 𝑚(𝑆𝔭 , O𝔭 ; O×𝔭 ) = 𝑚( 𝑆, b× ) > 0, so since

Î b O;
b O
all terms in the formula are nonnegative integers, we have 𝑚(𝑆, OL (𝐼); OL (𝐼) × ) > 0
for some [𝐼] ∈ Cls O, and OL (𝐼) ∈ Gen O.
Remark 30.4.19. The possible failure of local optimal embeddings to ‘glue’ to a global
optimal embedding into O is measured by the phenomenon of selectivity, studied in
Chapter 31.
30.4.20. More generally, an isomorphism 𝜙 : O −→ ∼ O0 of orders induces a bĳection

0 × 0×
Emb𝑅 (𝑆, O) ↔ Emb𝑅 (𝑆, O ), and 𝜙(O ) = O , so we may group together the terms
in (30.4.8) according to the type set Typ O. Let
ℎ(O) := [Idl O : PIdl O]

30.5. LOCAL EMBEDDING NUMBERS: MAXIMAL ORDERS 539
be index of the subgroup principal two-sided fractional O-ideals inside the invertible
ones, studied in section 18.5. By Proposition 18.5.10, the equation (30.4.8) then
becomes
∑︁
ℎ(O0)𝑚(𝑆, O0; O0× ) = ℎ(𝑆)𝑚( 𝑆,b O; b× ).
b O (30.4.21)
[O0 ] ∈Typ O
Remark 30.4.22. The foundational formula (30.4.8) is proven by counting a set two
different ways. As such, it admits a purely combinatorial proof: Brzezinski [Brz89]
shows that it follows from looking at “transitive actions of groups on pairs of sets and
on relations invariant with respect to these actions” [Brz89, p. 199].
We also record the following corollary (in the vein of Remark 30.4.22).
Corollary 30.4.23. Left multiplication defines a group action
𝐾 × \𝐾
b× /𝑆b× 𝐾 × \𝐸/
bO b×
that defines a free action of the group Pic 𝑆 on the set

Ä
Emb(𝑆, OL (𝐼); OL (𝐼) × ).
[𝐼 ] ∈Cls O
Proof. In the proof of the trace formula, we identified 𝐾 × \𝐾 b× /𝑆b× ' Pic 𝑆 and estab-
lished an idelic to global bĳection in (30.4.16); in (30.4.11), each fiber of the map
(30.4.9) is identified with Pic 𝑆, describing the orbits of the action. (See also Exercise
30.12, where a direct proof is requested.)
Remark 30.4.24. Theorem 30.4.7 is called a trace formula as it can be applied to

compute the trace of certain matrices (called Brandt matrices) that encode the action
of Hecke operators on a space of modular forms. We return to this important point of
view in detail in Chapter 41, and apply the above formula to compute traces in section
41.5.
30.5 Local embedding numbers: maximal orders

In view of Theorem 30.4.7, we see that up to a class number of the base ring, optimal
embeddings are counted in purely local terms. In this section and the next, we compute
the relevant local embedding numbers; after that, we will return to the global setting
to put the results together.
To this end, in this section and the next, we suppose 𝑅 is local (as in 23.2.1): so 𝑅
is a complete DVR with maximal ideal 𝔭 = 𝜋𝑅 and finite residue field 𝑘 := 𝑅/𝔭.
30.5.1. Since we are now local, the 𝑅-order 𝑆 is free and so 𝑆 = 𝑅[𝛾] for some (not
unique) 𝛾 ∈ 𝑆. Let 𝑓 𝛾 (𝑥) := 𝑥 2 − 𝑡 𝛾 𝑥 + 𝑛 𝛾 be the minimal polynomial of 𝛾, and let
𝑑 𝛾 := 𝑡 𝛾2 − 4𝑛 𝛾 be the discriminant of 𝑓 𝛾 , equal to the discriminant of 𝑆 in 𝑅/𝑅 ×2 . An
𝑅-algebra embedding from 𝑆 is then determined uniquely by the image of 𝛾.
We now compute (local) embedding numbers in the case of a maximal order; to

do so, we introduce some notation.
30.5.2. Recalling that 𝐾 ⊇ 𝐹 is a separable quadratic 𝐹-algebra and 𝔭 is the maximal

ideal of 𝑅, we define a symbol recording the splitting of the prime 𝔭 in 𝐾 (mirroring
the Kronecker symbol):


 1, if 𝐾 ' 𝐹 × 𝐹 is split;
𝐾 

:= 0, if 𝐾 ⊇ 𝐹 is a ramified field extension;
𝔭 
 −1

if 𝐾 ⊇ 𝐹 is an unramified field extension.

If 𝑞 = #𝑘 is odd, and 𝐾 ' 𝐹 [𝑥]/(𝑥 2 − 𝑑), then

𝐾 𝑑
=
𝔭 𝔭
is the usual Legendre symbol.
(a) We have 𝑚(𝑆, M2 (𝑅); GL2 (𝑅)) = 1.

(b) Suppose 𝐵 is a division algebra and O ⊆ 𝐵 its valuation ring. If 𝐾 is a field and
𝑆 = 𝑅𝐾 is integrally closed, then
𝑚(𝑆, O; 𝑁 𝐵× (O)) = 1

× 𝐾
𝑚(𝑆, O; O ) = 1 − .
𝔭
If 𝐾 is not a field or if 𝑆 is not integrally closed, then Emb(𝑆, O) = ∅.
We recall that 𝑁GL2 (𝐹 ) (M2 (𝑅)) = 𝐹 × GL2 (𝑅), so (a) also includes the case of the
normalizer.
Proof. First, part (a). We have at least one embedding 𝜙 : 𝑆 ↩→ End𝑅 (𝑆) ' M2 (𝑅)
given by the regular representation of 𝑆 on itself (in a basis): in the basis 1, 𝛾, we have

0 −𝑛 𝛾
𝛾 ↦→ (30.5.4)
1 𝑡𝛾
a matrix in rational canonical form. This embedding is optimal, because if 𝑥, 𝑦 ∈ 𝐹
satisfy
𝑥 −𝑛 𝛾 𝑦
𝜙(𝑥 + 𝑦𝛾) = ∈ M2 (𝑅)
𝑦 𝑥 + 𝑡𝛾 𝑦
then 𝑥, 𝑦 ∈ 𝑅 already, so 𝜙(𝐾) ∩ M2 (𝑅) = 𝜙(𝑆).
To finish (a), we need to show that the embedding (30.5.4) is the unique one, up to
conjugation by GL2 (𝑅). So let 𝜓 : 𝑆 ↩→ M2 (𝑅) be another optimal embedding. Then
via 𝜓, the 𝑅-module 𝑀 = 𝑅 2 (column vectors) has the structure of a left 𝑆-module; the
condition that 𝜓 is optimal translates into the condition that the (left) multiplicator ring
30.5. LOCAL EMBEDDING NUMBERS: MAXIMAL ORDERS 541
of 𝑅 2 in 𝐾 is precisely 𝑆, and therefore 𝑀 is invertible as a left 𝑆-module by Exercise

16.13, and therefore principal, generated by 𝑥 ∈ 𝑀, so that 𝑀 = 𝑆𝑥 = 𝑅𝑥 + 𝑅𝛾𝑥. In
the 𝑅-basis 𝑥, 𝛾𝑥, the left regular representation has the form (30.5.4), completing the
proof.
Here
is a second quick matrix proof (finding explicitly the cyclic basis). Let
𝑎 𝑏
𝛾 ↦→ ∈ M2 (𝑅) be an optimal embedding; then at least one of 𝑏, 𝑐, 𝑑 − 𝑎 ∈ 𝑅 × .
𝑐 𝑑
Therefore there exist 𝑥1 , 𝑥2 ∈ 𝑅 such that 𝑞(𝑥1 , 𝑥2 ) = 𝑐𝑥12 + (𝑑 − 𝑎)𝑥1 𝑥2 − 𝑏𝑥 22 ∈ 𝑅 × .
Let 𝑥 = (𝑥1 , 𝑥2 ) t ; then 𝛾𝑥 = (𝑎𝑥1 + 𝑏𝑥2 , 𝑐𝑥 1 + 𝑑𝑥 2 ) t . Let 𝛼 ∈ M2 (𝑅) be the matrix
with columns 𝑥, 𝛾𝑥. Then det 𝛼 = 𝑞(𝑥 1 , 𝑥2 ) ∈ 𝑅 × , so in fact 𝛼 ∈ GL2 (𝑅). We then
compute
−1 𝑎 𝑏 0 −(𝑎𝑑 − 𝑏𝑐)
𝛼 𝛼=
𝑐 𝑑 1 𝑎+𝑑
as claimed. (This matrix proof will be generalized in the next section, when we work
more generally with Eichler orders.)
Next, part (b). By Corollary 13.4.5, there exists an embedding 𝐾 ↩→ 𝐵 if and
only if 𝐾 is a field, so suppose 𝐾 is a field. By Proposition 13.3.4, the valuation
ring O is the unique maximal 𝑅-order in 𝐵, consisting of all integral elements, so
the embedding restricts to an embedding 𝑆 ↩→ O by uniqueness. Suppose 𝑆 = 𝑅𝐾 ;
then an embedding 𝑆 ↩→ O extends to 𝐾 ↩→ 𝐵 so is conjugate 30.3.1 to every other
under the action of 𝐵× . But the valuation ring is unique, thus 𝑁 𝐵× (O) = 𝐵× . Thus
𝑚(𝑆, O; 𝑁 𝐵× (O)) = 1. Finally, we have 𝑁 𝐵× (O)/(𝐹 × O× ) ' Z/2Z, generated by any
𝑗 ∈ O with nrd( 𝑗) = 𝜋. If the extension 𝐾 ⊇ 𝐹 is ramified, then 𝐾 = 𝐹 ( 𝑗) so 𝑗
centralizes 𝐾, and 𝑚(𝑆, O; 𝑁 𝐵× (O)) = 𝑚(𝑆, O; O× ) = 1. If instead 𝐾 ⊇ 𝐹 is inert,
then 𝐵 ' (𝐾, 𝑗 | 𝐹), and conjugation by 𝑗 normalizes but does not centralize 𝐾, and
𝑚(𝑆, O; O× ) = 2.
In all cases, the local embedding numbers are finite.
Corollary 30.5.5. 𝑚(𝑆, O; Γ) < ∞.
Proof. Recalling that the image of Γ in 𝑁 𝐵× (O)/𝐹 × is a subgroup of finite index, we

know that the natural surjective map
Emb(𝑆, O; Γ) → Emb(𝑆, O; 𝑁 𝐵× (O))
is finite-to-one; applying this argument twice we reduce to the case that Γ = O× .

Let O0 ⊇ O be a maximal 𝑅-order. Then O× ≤ O0× . Moreover, each 𝜙 ∈
Emb(𝑆, O) gives (by composing with O ↩→ O0) an embedding 𝜙 : 𝑆 ↩→ O0 that is
optimal for some superorder 𝑆 0 ⊇ 𝑆. Thus
∑︁
𝑚(𝑆, O; O× ) ≤ [O0× : O× ]𝑚(𝑆 0, O0; O0× );
𝑆0 ⊇𝑆
There are only finitely many superorders 𝑆 0 in the sum, since the integral closure 𝑅𝐾
contains all 𝑆 0 so [𝑅𝐾 : 𝑆 0] 𝑅 | [𝑅𝐾 : 𝑆] 𝑅 —or equally well, compare discriminants.
Applying Proposition 30.5.3, we conclude 𝑚(𝑆, O; O× ) < ∞.
30.6 ∗ Local embedding numbers: Eichler orders

At this point, the presentation of local embedding numbers begins to run off the rails:
for more general classes of orders, formulas for local embedding numbers are rarely
as simple as in Proposition 30.5.3. To avoid tumbling too far, we present formulas for
the case where O is a local Eichler (residually split) order following Hĳikata [Hĳ74,
Theorem 2.3]. See Remark 30.6.18 for further reference.
In this section, we retain the assumption that 𝑅 is local and the notation in 30.5.1.
Further, we suppose O is an Eichler order of level 𝔭𝑒 with 𝑒 ≥ 0.
30.6.1. Let
0 1
𝜛= ∈ O;
𝜋𝑒 0
then
𝑁 𝐵× (O)/𝐹 × O× = h𝜛i. (30.6.2)
If 𝑒 = 0, then 𝜛 ∈ O×
and 𝑁 𝐵× (O) = 𝐹 × O× ;
if instead 𝑒 ≥ 1, then h𝜛i ' Z/2Z.
For 𝜙 ∈ Emb(𝑆, O), let 𝜙 𝜛 be defined by
𝜙 𝜛 : 𝑆 ↩→ O
𝜙 𝜛 (𝛼) = 𝜛 −1 𝜙(𝛼)𝜛.
By (30.6.2), 𝜙, 𝜙 0 ∈ Emb(𝑆, O) are equivalent in Emb(𝑆, O; 𝑁 𝐵× (O)) if and only if

𝜙 0 is equivalent to either 𝜙 or 𝜙 𝜛 in Emb(𝑆, O; O× ).
Lemma 30.6.3. Let 𝜙 ∈ Emb(𝑆, O). Then there exist 𝑥 ∈ 𝑅 and 𝜈 ∈ 𝑁 𝐵× (O) such
that
𝑓 𝛾 (𝑥) = 𝑥 2 − 𝑡 𝛾 𝑥 + 𝑛 𝛾 ≡ 0 (mod 𝔭𝑒 )
and
𝑥 1
𝜈 −1 𝜙(𝛾)𝜈 = . (30.6.4)
− 𝑓 𝛾 (𝑥) 𝑡𝛾 − 𝑥
Proof. We may suppose that O is the standard Eichler order. Let 𝜙 : 𝑆 ↩→ O be an
embedding, with
𝑎 𝑏
𝜙(𝛾) = ∈O
𝑐𝜋 𝑒 𝑑
so that 𝑎, 𝑏, 𝑐, 𝑑 ∈ 𝑅. We have 𝑡 𝛾 = trd(𝛾) = 𝑎 + 𝑑 and 𝑛 𝛾 = nrd(𝛾) = 𝑎𝑑 − 𝑏𝑐𝜋 𝑒 . We
observe that 𝜙 is optimal if and only if at least one of 𝑏, 𝑐, 𝑎 − 𝑑 ∈ 𝑅 × : indeed, there
exists 𝑧 ∈ 𝑅 such that (𝛾 − 𝑧)/𝜋 ∈ O if and only if all belong to 𝔭.
Suppose now that 𝜙 is optimal.
We have three cases.
× 1 0
If 𝑏 ∈ 𝑅 , then take 𝜈 = ∈ O× ; we compute
0 𝑏 −1
𝑓 𝛾 (𝑎) = 𝑎 2 − (𝑎 + 𝑑)𝑎 + (𝑎𝑑 − 𝑏𝑐𝜋 𝑒 ) = −𝑏𝑐𝜋 𝑒
and
𝑎 1 𝑎 1
𝜈 −1 𝜙(𝛾)𝜈 = = (30.6.5)
𝑏𝑐𝜋 𝑒 𝑑 − 𝑓 𝛾 (𝑎) 𝑡𝛾 − 𝑎
30.6. ∗ LOCAL EMBEDDING NUMBERS: EICHLER ORDERS 543
as desired.
0 1
If 𝑐 ∈ 𝑅× , then we take 𝜈 = 𝜛 = 𝑒 ; now
𝜋 0

𝑑 𝑐
𝜈 −1 𝜙(𝛾)𝜈 = (30.6.6)
𝑏𝜋 𝑒 𝑎
and we apply the previous case.
Finally, if 𝑎 − 𝑑 ∈ 𝑅 × , we may suppose 𝑏 ∈ 𝔭 as well as 𝑐 ∈ 𝔭 if 𝑒 = 0, and then
1 1
we take 𝜈 = to find
0 1

𝑎 − 𝑐𝜋 𝑒 𝑎 − 𝑑 + 𝑏 − 𝑐𝜋 𝑒
𝜈 −1 𝜙(𝛾)𝜈 = (30.6.7)
𝑐𝜋 𝑒 𝑐𝜋 𝑒 + 𝑑
to reduce again to the first case.
Definition 30.6.8. An optimal embedding 𝜙 ∈ Emb(𝑆, O) is normalized and associ-

ated to 𝑥 ∈ 𝑅 if
𝑥 1
𝜙(𝛾) =
− 𝑓 𝛾 (𝑥) 𝑡 𝛾 − 𝑥
(as in (30.6.4)).
The statement of Lemma 30.6.3 is that for all 𝜙 ∈ Emb(𝑆, O), either the class
of 𝜙 or of 𝜙 𝜛 in Emb(𝑆, O; O× ) is represented by a normalized embedding. To
conclude our efforts, we need to check which of these are equivalent. We recall that
𝑑 𝛾 = 𝑡 𝛾2 − 4𝑛 𝛾 .
Lemma 30.6.9. Let 𝜙, 𝜙 0 be normalized embeddings associated to 𝑥, 𝑥 0 ∈ 𝑅. Then
the following statements hold.
(a) 𝜙, 𝜙 0 are conjugate by O× if and only if 𝑥 ≡ 𝑥 0 (mod 𝔭𝑒 ).
(b) If 𝑑 𝛾 ∈ 𝑅 × or 𝑒 = 0, then 𝜙 𝜛 is equivalent to 𝜙 0 in Emb(𝑆, O; O× ) if and only
if 𝑥 0 ≡ 𝑡 𝛾 − 𝑥 (mod 𝔭𝑒 ).
(c) If 𝑑 𝛾 ∉ 𝑅 × and 𝑒 ≥ 1, then 𝜙 𝜛 is equivalent to 𝜙 0 in Emb(𝑆, O; O× ) if and only
if 𝑥 0 ≡ 𝑡 𝛾 − 𝑥 (mod 𝔭𝑒 ) and 𝑓 𝛾 (𝑥) . 0 (mod 𝔭𝑒+1 ).
Proof. First (a). If 𝜙, 𝜙 0 are conjugate by some 𝜇 ∈ O× , we reduce modulo 𝔭𝑒 M2 (𝑅)

to obtain the ring of upper triangular matrices, and see that the diagonal entries of
𝜙(𝛾), 𝜙 0 (𝛾) are congruent modulo 0
𝔭 and in particular 𝑥 ≡ 𝑥 (mod 𝔭 ). Conversely,
𝑒 𝑒
1 0
if 𝑥 ≡ 𝑥 0 (mod 𝔭𝑒 ), then let 𝜇 = 0 ; we confirm that
𝑥 −𝑥 1
𝑥0

1
𝜇−1 𝜙(𝛾)𝜇 = = 𝜙 0 (𝛾).
− 𝑓 𝛾 (𝑥 0) 𝑡 𝛾 − 𝑥 0
In preparation for (b) and (c), we note that

𝑡 − 𝑥 − 𝑓 𝛾 (𝑥)/𝜋 𝑒
𝜙 𝜛 (𝛾) = 𝛾 𝑒 . (30.6.10)
𝜋 𝑥
We now prove (b). If 𝑒 = 0, the statement is true: 𝜛 ∈ O× and all embeddings are
conjugate. Suppose 𝑒 ≥ 1. Since 𝑓 𝛾 (𝑥) ≡ 0 (mod 𝔭𝑒 ) we have
𝑑 𝛾 ≡ (𝑥 − 𝑥) 2 = (𝑥 − (𝑡 𝛾 − 𝑥)) 2 = (𝑡 𝛾 − 2𝑥) 2 (mod 𝔭𝑒 )

1 𝑢
so 𝑡 𝛾 − 2𝑥 ∈ 𝑅 × . For 𝑢 ∈ 𝑅, let 𝜇 = and consider
0 1
𝑡 − 𝑥 − 𝑢𝜋 𝑒 𝑢(𝑡 𝛾 − 2𝑥) − 𝑢 2 𝜋 𝑒 − 𝜋 −𝑒 𝑓 𝛾 (𝑥)

𝜇−1 𝜙 𝜛 (𝛾)𝜇 = 𝛾
𝜋𝑒 𝑥 + 𝑢𝜋 𝑒
(30.6.11)
𝑡 − 𝑥 𝑢(𝑡 𝛾 − 2𝑥) − 𝑓 𝛾 (𝑥)/𝜋 𝑒
≡ 𝛾 𝑒 (mod 𝔭𝑒 ).
𝜋 𝑥
Thus we may choose 𝑢 so that the top-right entry of (30.6.11) is equal to 1. The result
then follows from (a).
We conclude with (c), and we are given 𝑒 ≥ 1 and 𝑑 𝛾 ∉ 𝑅 × . If 𝜋 −𝑒 𝑓 𝛾 (𝑥) ∈ 𝑅 ×
then from (30.6.10) and looking back at (30.6.5), already 𝜙 𝜛 is conjugate under
O× to the normalized embedding associated to 𝑡 𝛾 − 𝑥, so by (a), we then have 𝜙 𝜛
equivalent to 𝜙 0 if (and only if) 𝑥 ≡ 𝑡 − 𝑥 (mod 𝔭𝑒 ). To finish, suppose 𝑓 𝛾 (𝑥) ∈ 𝔭𝑒+1
and 𝑑 ∈ 𝔭. Reducing modulo 𝔭𝑒 M2 (𝑅), the ring O/(𝔭𝑒 M2 (𝑅)) consists of upper-
triangular matrices and its unit group is generated by diagonal matrices and units
𝜇 as in the previous paragraph. But then by (30.6.11), the top-right entry of every
O× -conjugate of 𝜙 𝜛 (𝛾) belongs to 𝔭, and so it cannot be equal to 𝜙 0 (𝛾).
The statements above give a way to compute the local embedding number in terms
of arithmetic of 𝑅.
Proposition 30.6.12. For 𝑠 ∈ Z ≥1 , let
𝑀 (𝑠) := {𝑥 ∈ 𝑅/𝔭𝑠 : 𝑓 𝛾 (𝑥) ≡ 0 (mod 𝔭𝑠 )}. (30.6.13)
Then for 𝑒 ≥ 1,
(
× #𝑀 (𝑒), if 𝑑 𝛾 ∈ 𝑅 × ;
𝑚(𝑆, O; O ) =
#𝑀 (𝑒) + # img (𝑀 (𝑒 + 1) → 𝑅/𝔭𝑒 ) , otherwise.
Proof. By Lemma 30.6.3, the set Emb(𝑆, O; O× ) is represented by the set of normal-
ized embeddings and their conjugates under 𝜛. By Lemma 30.6.9(a), the normalized
embeddings according to 𝑥 ∈ 𝑀 (𝑠) are distinct; by (b)–(c), the remaining conjugate
embeddings are new when they lift to 𝑀 (𝑒 + 1).
√ √
Example 30.6.14. Let 𝑅 = Z2 and 𝑆 = Z2 [ −1], so 𝛾 = −1 and 𝑓 𝛾 (𝑥) = 𝑥 2 + 1. We
have (
{1 mod 2}, if 𝑠 = 1;
𝑀 (𝑠) =
∅, if 𝑠 ≥ 2.
Therefore by Proposition 30.6.12, if O is an Eichler order of level 2𝑒 over Z2 , then
(
√ 1, if 𝑒 ≤ 1;
𝑚(Z2 [ −1], O; O× ) = (30.6.15)
0, if 𝑒 ≥ 2.
30.6. ∗ LOCAL EMBEDDING NUMBERS: EICHLER ORDERS 545
We record an important special case.

Lemma 30.6.16. If 𝑒 = 1 and 𝑆 is a maximal 𝑅-order in 𝐾 (equivalently, 𝑆 is integrally
closed in 𝐾), then
𝐾
𝑚(𝑆, O; O× ) = 1 + .
𝔭
Proof. Referring to Proposition 30.6.12, we first compute #𝑀 (1), the number of
solutions to 𝑓 𝛾 (𝑥) ≡ 0 (mod 𝔭), so over the finite field 𝑘 = 𝑅/𝔭: this number is
𝐾
2, 1, 0 according as = 1, 0, −1. If 𝑑 ∈ 𝑅 × , we are done; so suppose 𝑑 ∉ 𝑅 × ,
𝔭
𝐾
hence = 0. Let 𝑥 ∈ 𝑀 (2). We claim that (𝛾 − 𝑥)/𝜋 is integral over 𝑅,
𝔭
contradicting that 𝑆 is maximal. Indeed, 𝜋 2 | 𝑓 𝛾 (𝑥) = 𝑥 2 − 𝑡 𝛾 𝑥 + 𝑛 𝛾 = nrd(𝛾 − 𝑥); and
𝔭 | 𝑑 𝛾 = 𝑡 𝛾2 − 4𝑛 𝛾 = (𝑡 𝛾 − 2𝑥) 2 , so 𝔭 | (𝑡 𝛾 − 2𝑥) = trd(𝛾 − 𝑥). Thus 𝑀 (2) = ∅ and so
the count remains the same.
To conclude, when 𝑞 = #𝑘 is odd, we can make Proposition 30.6.12 completely

explicit.
Lemma 30.6.17. Suppose #𝑘 = 𝑞 is odd, that 𝑒 ≥ 1, and let 𝑓 := ord𝔭 (𝑑 𝛾 ).
(a) If 𝑓 = 0, then
× 𝐾
𝑚(𝑆, O; O ) = 1 + .
𝔭
(b) If 𝑒 < 𝑓 , then
(
× 2𝑞 (𝑒−1)/2 , if 𝑒 is odd;
𝑚(𝑆, O; O ) =
𝑞 𝑒/2−1 (𝑞 + 1), if 𝑒 is even.
(c) If 𝑒 = 𝑓 , then
( 𝑓 −1)/2 , if 𝑓 is odd;
𝑞


×

𝑚(𝑆, O; O ) = 𝑓 /2 + 𝑞 𝑓 /2−1 𝐾
𝑞
 1+ , if 𝑓 is even.
 𝔭
(d) If 𝑒 > 𝑓 > 0, then
 0, if 𝑓 is odd;


×

𝑚(𝑆, O; O ) = 𝑓 /2−1 (𝑞 + 1) 1 + 𝐾
𝑞
 , if 𝑓 is even.
 𝔭
Proof. Since the residue field 𝑘 has odd characteristic, we can complete the square
and without loss of generality we may suppose that trd(𝛾) = 0, and
𝑀 (𝑠) = 𝑥 ∈ 𝑅/𝔭𝑠 : 𝑥 2 ≡ 𝑑 (mod 𝔭𝑠 ) .

We will abbreviate 𝑚 = 𝑚(𝑆, O; O× ) and repeatedly refer to Proposition 30.6.12.

First suppose 𝑓 = 0. Then 𝑑 𝛾 ∈ 𝑅 × , so by Proposition 30.6.12, we have 𝑚 =

#𝑀 (𝑒). But by Hensel’s lemma, #𝑀 (𝑒) = 0 or 2 according as 𝑑 is a square or not in
𝑅 for all 𝑒 ≥ 1.
Next suppose that 𝑒 < 𝑓 . The solutions to the equation 𝑥 2 ≡ 0 (mod 𝔭𝑠 ) are
those with 𝑥 ≡ 0 (mod 𝔭 d𝑠/2e ). Thus #𝑀 (𝑒) = 𝑞 𝑒− d𝑒/2e = 𝑞 b𝑒/2c and we see that
# img (𝑀 (𝑒 + 1) → 𝑅/𝔭𝑒 ) = 𝑞 𝑒− d(𝑒+1)/2e , so 𝑚 = 2𝑞 (𝑒−1)/2 if 𝑒 is odd and 𝑚 =
𝑞 𝑒/2 + 𝑞 𝑒/2−1 = 𝑞 𝑒/2−1 (𝑞 + 1) if 𝑒 is even.
If 𝑒 = 𝑓 , then again #𝑀 (𝑒) = 𝑞 b𝑒/2c . To count the second contributing set, we
must solve 𝑥 2 ≡ 𝑑 𝛾 (mod 𝔭𝑒+1 ). If 𝑒 = 𝑓 is odd then this congruence has no solution.
If instead 𝑒 is even then we must solve 𝑦 2 = (𝑥/𝜋 𝑓 /2 ) 2 ≡ 𝑑 𝛾 /𝜋 𝑓 (mod 𝔭) where 𝜋
is a uniformizer at 𝔭. This latter congruence has zero or two solutions according as 𝑑
is a square, and given such a solution 𝑦 we have the solutions 𝑥 ≡ 𝑦 (mod 𝜋 𝑓 /2+1 ) to
the original congruence, and hence there are 0 or 2𝑞 𝑓 −( 𝑓 /2+1) = 2𝑞 𝑓 /2−1 solutions, as
claimed.
Finally, suppose 𝑒 > 𝑓 > 0. If 𝑓 is odd, there are no solutions to 𝑥 2 ≡ 𝑑 𝛾
(mod 𝔭𝑒 ). If 𝑓 is even, there are no solutions if 𝑑 𝛾 is not a square and otherwise
the solutions are 𝑥 ≡ 𝑦 (mod 𝔭𝑒− 𝑓 /2 ) as above so they total 2𝑞 𝑓 /2 + 2𝑞 𝑓 /2−1 =
2𝑞 𝑓 /2−1 (𝑞 + 1).
On the other hand, when the residue field 𝑘 has even characteristic, the computa-
tions become even more involved!
Remark 30.6.18. Eichler studied optimal embeddings [Eic38b, §2] very early on,
computing the contribution of units (coming from embeddings of Z[𝜔] and Z[𝑖] in a
maximal order O) in the mass formula. He then [Eic56a, §3] studied more generally
optimal embeddings of quadratic orders into his Eichler orders of squarefree level.
Hĳikata [Hĳ74, §2] studied optimal embeddings in the context of computing traces
of Hecke operators on Γ0 (𝑁) (general 𝑁), with embedding numbers given for certain
orders. See also Eichler’s treatment [Eic73, §3] in the context of the basis problem for
modular forms, as well as Pizer’s presentation [Piz76a, §3]. (See Remark 41.5.13 for
further detail.) Brzezinski [Brz91, Corollary 1.16] (a typo has it appear as Corollary
1.6) gives a general formula for Eichler orders (which is to say, a generalization of
Lemma 30.6.17 to include 𝑞 even)—the proof method is different than the method of
Hĳikata above, and the answer is organized a bit differently than Lemma 30.6.17.
But these papers are just the beginning, and there is a cornucopia of further results.
Many of these are obtained in pursuit of progressively more general forms of the trace
formula (see for example the summary of results by Hashimoto [Hash77]) for Eichler
orders. Shimizu [Shz63, §§26–27] considered embedding numbers over totally real
fields in computing the contribution of elliptic elements to formulas for the dimension
of spaces of cusp forms and later for the trace formula [Shz65, §3]. The contributions
of elliptic elements over totally real fields was also pursued by Prestel [Pre68, §5] and
more generally for embeddings by Schneider [Schn75] (and quite explicitly for real
quadratic fields [Schn77]) and Vignéras [Vig76a, §4].
Pizer [Piz76b, §§3–5] considered optimal embeddings for residually split orders
(see 24.3.7) over Q: these were then applied to further cases of the basis problem for
modular forms [Piz76c, Piz80b]. Then Hĳikata–Pizer–Shemanske [HPS89b, §§1–5]
developed in a uniform manner the optimal embedding theory for basic orders (they
30.7. GLOBAL EMBEDDING NUMBERS 547
called them special, cf. Remark 24.5.7): the application to the trace formula is then
contained in their monograph [HPS89a]. Brzezinski [Brz90, §3] also obtains recursive
formulas for optimal embedding numbers of a local Bass (equivalently, basic) order
(in characteristic not 2), using an effective description of the automorphism group of
the order.
30.7 Global embedding numbers

We now combine the ingredients from the previous three sections to arrive at a formula
for global embedding numbers.
We return to the setting of section 30.4, with 𝑅 a global ring.
30.7.1. Our global ring 𝑅 = 𝑅 (T) comes from an eligible set T ⊆ Pl 𝐹. (This set is
usually denoted S, but we do not want any confusion with the quadratic 𝑅-algebra
𝑆 ⊆ 𝐾.)
For all but finitely many places 𝑣 ∉ T, we have O 𝑣 ' M2 (𝑅 𝑣 ) maximal and Γ 𝑣 =
O×𝑣 . By Proposition 30.5.3(a), for such places 𝑣, we have # Emb𝑅𝑣 (𝑆 𝑣 , O 𝑣 ; O×𝑣 ) = 1.
Therefore the number 𝑚( 𝑆, b O;
bb b O;
Γ) = # Emb𝑅b ( 𝑆, bbΓ) of adelic embeddings is given
by the (well-defined, finite) product
Ö
b O;
𝑚( 𝑆, bbΓ) = 𝑚(𝑆 𝑣 , O 𝑣 ; Γ 𝑣 ) (30.7.2)
𝑣∉T
(well-defined and finite).
We arrive at the following first global result.
Theorem 30.7.3. Let 𝔑 = discrd(O). Then

∑︁ Ö
𝑚(𝑆, OL (𝐼); OL (𝐼) × ) = ℎ(𝑆) 𝑚(𝑆𝔭 , O𝔭 ; O×𝔭 ).
[𝐼 ] ∈Cls O 𝔭 |𝔑
Proof. For all 𝔭 - 𝔑, we have O𝔭 ' M2 (𝑅𝔭 ), so the result follows by combining
Theorem 30.4.7 with 30.7.1.
Working with the normalizer instead, we find the following result.
Proposition 30.7.4. Suppose that O is a maximal order with # Cls O = 1. Then
𝑚(𝑆, O; 𝑁 𝐵× (O)) = ℎ(𝑆).
Proof. As in the proof of Theorem 30.7.3, but with a further appeal to Proposition
30.5.3 to show that all local embeddings numbers are equal to 1.
An important illustrative special case is the following.

Example 30.7.5. Let 𝔇 = disc𝑅 𝐵 and suppose that O ⊆ 𝐵 is an Eichler 𝑅-order of

squarefree level 𝔐, so discrd O = 𝔇𝔐. Suppose further that 𝑆 is a maximal 𝑅-order
in 𝐾. Then Theorem 30.7.3 reads
∑︁ Ö Ö
𝐾

𝐾
𝑚(𝑆, OL (𝐼); OL (𝐼) × ) = ℎ(𝑆) 1− 1+ ,
𝔭 𝔭
[𝐼 ] ∈Cls O 𝔭 |𝔇 𝔭 |𝔐
with the local embedding numbers computed in Proposition 30.5.3(b) for 𝔭 | 𝔇 and
Lemma 30.6.16 for 𝔭 | 𝔐.
Suppose further that 𝐵 is T-indefinite and that # ClΩ 𝑅 = 1; then # Cls O = # ClΩ 𝑅
by Corollary 28.5.17 (an application of strong approximation), so
Ö Ö
𝐾

𝐾
𝑚(𝑆, O; O× ) = ℎ(𝑆) 1− 1+ . (30.7.6)
𝔭 𝔭
𝔭 |𝔇 𝔭 |𝔐
Embeddings of cyclotomic orders are of particular interest.

Example 30.7.7. Consider the case 𝐹 = Q and 𝑅 = Z. Let 𝐷 = disc 𝐵 and suppose
that O is an Eichler order of level 𝑀, so 𝑁 = 𝐷 𝑀 = discrd O. We recall 𝐷 is
squarefree and gcd(𝐷, 𝑀) = 1. Suppose 𝐵 is indefinite. Then by Theorem 28.2.11
(an application of strong approximation), # Cls O = 1. √
If 𝐾 ⊇ Q is√a cyclotomic quadratic extension, then either 𝐾 = Q(𝜔) = Q( −3) or
𝐾 = Q(𝑖) = Q( −4) with corresponding rings of integers Z𝐾 = Z[𝜔] and Z𝐾 = Z[𝑖],
each with ℎ(Z𝐾 ) = 1. With local embedding numbers computed in Lemma 30.6.17,
Theorem 30.7.3 then gives
 Ö Ö
−3

−3
1− 1+ , if 9 - 𝑀;


×


𝑚(Z[𝜔], O; O ) = 𝑝 |𝐷 𝑝
𝑝 |𝑀
𝑝

 0, if 9 | 𝑀.


Similarly, using (30.6.15),
 Ö Ö
−4

−4
1− 1+ if 4 - 𝑀;



 ,
×
𝑚(Z[𝑖], O; O ) = 𝑝 |𝐷 𝑝
𝑝 |𝑀
𝑝

 0, if 4 | 𝑀.


Absent further hypothesis, it is difficult to tease apart the term 𝑚(𝑆, O; O× ) from
the sum over left orders in Theorem 30.4.7. In the next chapter, we will show that the
hypothesis that 𝐵 is T-indefinite is almost enough.
30.8 Class number formula

In this section, we explain how the theory of optimal embeddings can be used to
convert the mass formula into a class number formula, following Eichler.
Suppose throughout this section that 𝐵 is T-definite. By Lemma 26.5.1, the group
O× /𝑅 × is finite; let 𝑤 O = [O× : 𝑅 × ].
30.8. CLASS NUMBER FORMULA 549
30.8.1. To a nontrivial cyclic subgroup of O× /𝑅 × , we associate the quadratic field 𝐾

it generates over 𝐹; such a field is necessarily cyclotomic, with 𝐾 ' 𝐹 (𝜁2𝑞 ) for 𝑞 the
order of the cyclic subgroup. (Only certain 𝐹 (𝜁2𝑞 ) arise as quadratic extensions of 𝐹,
and different values of 𝑞 can give rise to the same quadratic field up to isomorphism.)
Conversely, to a quadratic field 𝐾 ⊇ 𝐹, we obtain a (possibly trivial) cyclic
subgroup (𝐾 × ∩ O× )/𝑅 × .
Lemma 30.8.2. Every nontrivial 𝛼𝑅 × ∈ O× /𝑅 × belongs to a unique maximal cyclic

subgroup.
Proof. Since O× /𝑅 × is finite, 𝛼 belongs to at least one maximal cyclic subgroup; if it

belonged to two, then the corresponding quadratic fields would both contain the field
corresponding to 𝛼, hence by degrees would be equal, so by maximality the cyclic
subgroups would have to be equal.
30.8.3. Recall that 𝑚(𝑆, O; O× ) counts optimal embeddings 𝜙 : 𝑆 ↩→ O up to conjuga-

tion by O× . Since O× /𝑅 × is finite, the set Emb(𝑆, O) is itself finite. Precisely two em-
beddings give rise to the same image 𝜙(𝑆), differing by the (necessarily nontrivial) stan-
dard involution. The stabilizer of O× on 𝜙 ∈ Emb(𝑆, O) is O× ∩ 𝜙(𝑆) = 𝜙(𝑆) × ' 𝑆 × .
Let 𝑤 𝑆 = [𝑆 × : 𝑅 × ]. We have shown that
2𝑤 𝑆
𝑚(𝑆, O; O× ) = #{𝜙(𝑆) ⊆ O : 𝜙 ∈ Emb(𝑆; O)} . (30.8.4)
𝑤O

1 1 ∑︁ 1 ∑︁
1− = 1− 𝑚(𝑆, O; O× ).
𝑤 O 2 𝑞 ≥2 𝑞
[𝑆 × :𝑅 × ]=𝑞
Proof. We count off the group O× /𝑅 × by maximal cyclic subgroups, keeping track
of the trivial class. By Lemma 30.8.2, every nontrivial 𝛼𝑅 × ∈ O× /𝑅 × belongs to
a unique maximal cyclic subgroup of some order 𝑞 ≥ 2: such a subgroup is of the
form 𝜙(𝑆) × /𝑅 × with 𝜙(𝑆) ⊆ O an optimally embedded order, and has 𝑞 − 1 nontrivial
elements. Therefore
∑︁ ∑︁
𝑤O − 1 = (𝑞 − 1)#{𝜙(𝑆) ⊆ O : 𝜙 ∈ Emb(𝑆; O)}.
𝑞 ≥2 𝑆 ⊆𝐾𝑞
[𝑆 × :𝑅 × ]=𝑞
Plugging in (30.8.4), we obtain

∑︁ ∑︁ 𝑤O
𝑤O − 1 = (𝑞 − 1)𝑚(𝑆, O; O× ) ;
𝑞 ≥2 𝑆 ⊆𝐾𝑞
2𝑞
[𝑆 × :𝑅 × ]=𝑞
dividing through by 𝑤 O gives the result.

We now recall the Eichler mass formula (Main Theorem 26.1.5), giving an explicit
formula for the weighted class number
∑︁ 1
mass(Cls O) := ,
𝑤𝐼
[𝐼 ] ∈Cls O
where 𝑤 𝐼 = [OL (𝐼) × : 𝑅 × ].

Main Theorem 30.8.6 (Eichler class number formula). Let 𝑅 be a global ring with
eligible set T, let 𝐵 be an T-definite quaternion algebra over 𝐹 = Frac 𝑅, and let O ⊂ 𝐵
be an 𝑅-order. Then
∑︁
1 ∑︁ 1 b× )
# Cls O = mass(Cls O) + 1− b O;
ℎ(𝑆)𝑚( 𝑆, b O
2 𝑞 ≥2 𝑞 × ×
[𝑆 :𝑅 ]=𝑞
where the inner sum is over all quadratic 𝑅-algebras 𝑆 ⊇ 𝑅 such that [𝑆 × : 𝑅 × ] = 𝑞 ∈
Z ≥2 , and ℎ(𝑆) = # Pic 𝑆.
Proof. We apply Proposition 30.8.5 to each order OL (𝐼) for [𝐼] ∈ Cls O and sum. We
obtain
∑︁ 1

1− = # Cls O − mass(Cls O)
𝑤𝐼
[𝐼 ] ∈Cls O
∑︁
1 ∑︁ 1 ∑︁
= 1− 𝑚(𝑆, OL (𝐼); OL (𝐼) × )
2 𝑞 ≥2 𝑞 × ×
[𝑆 :𝑅 ]=𝑞 [𝐼 ] ∈Cls O
∑︁
1 ∑︁ 1 b O;
b Ob× ),
= 1− ℎ(𝑆)𝑚( 𝑆,
2 𝑞 ≥2 𝑞 × ×
[𝑆 :𝑅 ]=𝑞
the last equality by Theorem 30.4.7.

The expressions in the Eichler class number formula (Main Theorem 30.8.6) are
arithmetically involved but can be computed effectively. In special cases, they give
reasonably simple formulas, as previously advertised.
Example 30.8.7. When 𝐹 = Q, Main Theorem 30.8.6 yields the formula given in
Theorem 30.1.5, with the computation of local class numbers given in Example 30.7.7.
Remark 30.8.8. The formula for the class number for Eichler orders of squarefree
level (i.e., hereditary orders) was given by Eichler [Eic56a, Satz 10–11]; for Eichler
orders over Q of arbitrary level (as in Example 30.8.7), it was given by Pizer [Piz76a,
Theorem 16]. Main Theorem 30.8.6 is proven by Vignéras [Vig80a, Corollaire V.2.5]
and Körner [Kör87, Theorem 2]. For further reference and discussion (in the context
of computing local embedding numbers), see Remark 30.6.18.
30.8.9. Let 𝐾 ⊃ 𝐹 be a quadratic field extension, let 𝑅𝐾 be the integral closure of 𝑅
in 𝐾, and let 𝑆 ⊆ 𝐾 be an 𝑅-order of conductor 𝔣. Then there is an exact sequence
(𝑅𝐾 /𝔣𝑅𝐾 ) ×
1 → 𝑆× → 𝑅𝐾
×
→ → Pic 𝑆 → Pic 𝑅𝐾 → 1 (30.8.10)
(𝑅/𝔣) ×
30.9. TYPE NUMBER FORMULA 551
giving rise to the formula of Dedekind
ℎ(𝑅𝐾 ) Ö
𝐾 1

ℎ(𝑆) = × : 𝑆 × ] N (𝔣)
[𝑅𝐾
1−
𝔭 N (𝔭)
(30.8.11)
𝔭 |𝔣

𝐾
where N is the absolute norm and is given (globally) as in 30.5.2:
𝔭


 −1, if 𝔭 is inert in 𝐾;
𝐾 

= 0, if 𝔭 is ramified in 𝐾;
𝔭 
 1,

if 𝔭 splits in 𝐾.

30.9 Type number formula

We continue with the hypotheses of the previous section. A further application of the
strategy to compute the class number is to also compute the type number. The methods
are indeed quite similar: rearranging Corollary 18.5.12, we have
# Cls O ∑︁ 1

# Typ O = + 1− (30.9.1)
# Pic𝑅 O 0
𝑧 O0
[O ] ∈Typ O
where 𝑧O0 = [𝑁 𝐵× (O0) : 𝐹 × O0× ]. But now the structure of the normalizer groups
come into play, and one can give a type number formula similar to the class number
formula 30.8.6 in terms of certain embedding numbers at least for Eichler orders.
Unfortunately, even over Q, these formulas quickly get very complicated! To give a
sense of what can be proven, in this section we provide a type number formula in a
special but interesting case due originally to Deuring [Deu51], and we refer to Remark
30.9.12 for further reference.
Proposition 30.9.2 (Deuring). Let 𝐵 be a definite quaternion algebra over Q with
disc 𝐵 = 𝑝 prime and let O ⊂ 𝐵 be a maximal order. Then # Typ O = 1 for 𝑝 = 2, 3,
and for 𝑝 ≥ 5,
1 1
# Typ O = # Cls O + [ℎ(−𝑝)] + ℎ(−4𝑝)
2 4
where [ℎ(−𝑝)] = ℎ(−𝑝) when 𝑝 ≡ 3 (mod 4) and is 0 otherwise.
Proof. In light of (30.9.1), we begin by considering the Picard group Pic O (with
𝑅 = Z): by 18.4.8, we have an isomorphism Pic(O) ' Z/2Z generated by the unique
right ideal 𝐽 ⊆ O with nrd(𝐽) = 𝑝. The ideal 𝐽 is automatically two-sided and contains
all elements of reduced norm divisible by 𝑝 (see 13.3.7); and thus 𝐽 is principal if and
only if there exists an element 𝛼 ∈ O with nrd(𝛼) = 𝑝 if and only if 𝑧 O = 2.
Therefore, (30.9.1) reads
1 1 ∑︁
# Typ O = # Cls O + #{𝐽 0 ⊆ O0 principal right ideal : nrd(𝐽 0) = 𝑝}.
2 2
[O0 ] ∈Typ O
(30.9.3)
We now compute this sum in terms of embedding numbers. First, the map 𝛼 ↦→ 𝛼O
gives
#{𝐽 ⊆ O principal right ideal : nrd(𝐽) = 𝑝}
1 (30.9.4)
= #{𝛼 ∈ O : nrd(𝛼) = 𝑝}
2𝑤 O
where 𝑤 O = [O× : Z× ].
Next, we claim that if 𝛼 ∈ O has nrd(𝛼) = 𝑝, then trd(𝛼) = 0, i.e., 𝛼2 + 𝑝 = 0.
Indeed, if 𝑡 = trd(𝛼) then the field 𝐾 = Q(𝛼) has discriminant 𝑡 2 − 4𝑝 < 0 (since
√
𝐵 is definite) so |𝑡| < 2 𝑝. If 𝑡 ≠ 0, then the polynomial 𝑥 2 − 𝑡𝑥 + 𝑝 splits modulo
𝑝, so 𝑝 splits in 𝐾; but 𝐾 𝑝 ↩→ 𝐵 𝑝 and 𝐵 𝑝 is a division algebra, so 𝐾 𝑝 is a field, a
contradiction. Thus 𝑡 = 0.
With these results in hand, we can bring in the theory of embedding numbers. We
have
#{𝛼 ∈ O : nrd(𝛼) = 𝑝} = #{𝛼 ∈ O : 𝛼2 + 𝑝 = 0}
∑︁
= # Emb(𝑆, O). (30.9.5)
√
𝑆 ⊇Z[ − 𝑝]
The group O× /{±1} acts by conjugation on Emb(𝑆, O) without fixed points as in

30.8.3: since 𝑝 ≠ 2, 3, we have 𝑆 × = {±1}. Thus
# Emb(𝑆, O) = 𝑤 O 𝑚(𝑆, O; O× ). (30.9.6)
Combining (30.9.4), (30.9.5), and (30.9.6), and plugging into (30.9.3), we have
1 1 ∑︁ ∑︁
# Typ O = # Cls O + 𝑚(𝑆, O0; O0× ). (30.9.7)
2 4 √
𝑆 ⊇Z[ − 𝑝] [O0 ] ∈Typ O
√
By (30.4.21) (rewriting Theorem 30.4.7), for 𝑆 ⊇ Z[ −𝑝] we have
∑︁
ℎ(O0)𝑚(𝑆, O0; O0× ) = ℎ(𝑆)𝑚( 𝑆,
b O;
b Ob× )
[O0 ] ∈Typ O
where ℎ(O0) = [Idl O0 : PIdl O0]; but ℎ(O0) = 1 whenever 𝑚(𝑆, O0; O0× ) ≠ 0 by the
first paragraph, so we may substitute into (30.9.7) to get
1 1 ∑︁
b× ).
# Typ O = # Cls O + b O;
ℎ(𝑆)𝑚( 𝑆, b O (30.9.8)
2 4 √
𝑆 ⊇Z[ − 𝑝]
The order O is maximal, so the adelic embedding number is the product of local em-
bedding numbers computed in Proposition 30.5.3: there is only a possible contribution
at 𝑝, since 𝑝 ≠ 2 the order 𝑆 is maximal, and 𝐾 is ramified so 𝑚(𝑆 𝑝 , O 𝑝 ; O×𝑝 ) = 1,
b O;
thus 𝑚( 𝑆, b O b× ) = 1.
√
Finally, the order Z[ −𝑝] of discriminant −4𝑝 is maximal whenever 𝑝 ≡ 1 (mod 4)
and the sum becomes simply ℎ(−4𝑝); when 𝑝 ≡ 3 (mod 4), this order is contained in
the maximal order of discriminant −𝑝, so the sum is ℎ(−𝑝) + ℎ(−4𝑝). The result is
proven.
Remark 30.9.9. For an alternate direct proof of Proposition 30.9.2 working with elliptic
curves, see Cox [Cox89, Theorem 14.18].
30.9.10. The sum of class numbers in Proposition 30.9.2 can be rewritten uniformly
in terms of the ring of integers as follows:

 1, if 𝑝 ≡ 1 (mod 4);
√



[ℎ(−𝑝)] + ℎ(−4𝑝) = # Cl Q( −𝑝) · 4, if 𝑝 ≡ 3 (mod 8);

 2, if 𝑝 ≡ 7 (mod 8).


When 𝑝 ≡ 1 (mod 4), there is nothing to do. For 𝑝 ≡ 3 (mod 4), we have
(
3ℎ(−𝑝), if 𝑝 ≡ 3 (mod 8);
ℎ(−4𝑝) =
ℎ(−𝑝), if 𝑝 ≡ 7 (mod 8);
√
according as 2 is inert or split in 𝐾 = Q( −𝑝).
Remark 30.9.11. In section 42.1–42.2, we relate quaternion algebras to supersingular
elliptic curves; in this language, Proposition 30.9.2 gives rise to a formula for the
number of supersingular elliptic curves defined over F 𝑝 up to isomorphism.
Remark 30.9.12. Eichler [Eic56a, Satz 11] gave a type number formula for definite
hereditary orders over a totally real field; this formula has an error which was corrected
by Peters [Pet69, Satz 14, Satz 15] over fields of class number one and by Pizer [Piz73,
Theorem A] in general. Pizer [Piz76a, Theorem 26] gives a formula for the type
number for (general) Eichler orders over Q. Finally, Vignéras [Vig80a, Corollaire
V.2.6] gives a “structural” type number formula (without explicit evaluation of the
sum) for Eichler orders, and Körner [Kör87, Theorem 3] gives a general type number
formula. For a generalization to totally definite orders in central simple algebras of
prime index over global fields, see Brzezinski [Brz97].
Exercises
1. As in section 30.1, for 𝑘 ≥ 1 let
𝑟 𝑘 (𝑛) := #{(𝑥1 , 𝑥2 , . . . , 𝑥 𝑘 ) ∈ Z 𝑘 : 𝑥12 + 𝑥 22 + · · · + 𝑥 2𝑘 = 𝑛}.
We gave formulas for 𝑟 3 (𝑛), 𝑟 4 (𝑛). For completeness, observe that 𝑟 1 (𝑛) = 2, 0
according as 𝑛 is a square or not, and give a formula for 𝑟 2 (𝑛) in terms of the
factorization of 𝑛 in the ring Z[𝑖].
I 2. Let 𝐵 be a quaternion algebra over Q, let O ⊂ 𝐵 be an order, let 𝐾 be a quadratic
field with an embedding 𝐾 ↩→ 𝐵 and suppose 𝑆 = 𝐾 ∩ O is the ring of integers
of 𝐾.
(a) Let 𝔟 ⊂ 𝐾 be an invertible fractional 𝑆-ideal. Show that 𝔟O∩𝐾 = 𝔟. [Hint:
since 1 ∈ O, we have 𝔟O ∩ 𝐾 ⊇ 𝔟. For the other inclusion, consider
(𝔟O ∩ 𝐾) · 𝔟−1 𝔟 ⊆ (𝔟𝔟−1 O ∩ 𝐾) · 𝔟 = 𝔟.]

(b) Rewrite the proof in (a) idelically.

I 3. Let 𝑛 ≡ 1, 2 (mod 4).
(a) Let 𝑝 be an odd prime, and let 𝛼, 𝛼 0 ∈ M2 (Z 𝑝 ) satisfy 𝛼2 + 𝑛 = 0. Show
that there exists 𝜇 ∈ GL2 (Z 𝑝 ) such that 𝛼 0 = 𝜇−1 𝛼𝜇.
−1, −1
(b) Let 𝐵2 = and O2 its valuation ring. Show that if 𝛼, 𝛼 0 satisfy the
Q2
same reduced characteristic polynomial, then there exists 𝜈 ∈ 𝑁 𝐵2× (O2 )
such that 𝛼 0 = 𝜈 −1 𝛼𝜈. [Hint: 𝑁 𝐵2× (O2 ) = 𝐵2× .]
(c) Put together (a) and (b) to conclude (30.2.4).
4. Deduce the theorem of Gauss (Theorem 30.1.3) from Theorem 30.4.7.
√
5. Let 𝑚, 𝑛 ∈ Z>0 be odd, coprime, with 𝑚 < 2 𝑛. Show that there are 12ℎ(𝑚 2 −
4𝑛) solutions to the equations
𝑎2 + 𝑏2 + 𝑐2 + 𝑑 2 = 𝑛
𝑎+𝑏+𝑐+𝑑 =𝑚
with 𝑎, 𝑏, 𝑐, 𝑑 ∈ Z. [Hint: consider the unit (1 − 𝑖 − 𝑗 − 𝑘)/2 of the Hurwitz

order.]
6. Specialize Theorem 30.1.5 to the case 𝐷 = 𝑝 and 𝑁 = 1, as follows. Let 𝐵 be a
definite quaternion algebra over Q of discriminant 𝐷 = 𝑝 prime and let O ⊂ 𝐵
be a maximal order (Eichler order of level 𝑀 = 1). Show that



 1, if 𝑝 = 2, 3;



 ( 𝑝 − 1)/12, if 𝑝 ≡ 1 (mod 12);


# Cls O = ( 𝑝 + 7)/12, if 𝑝 ≡ 5 (mod 12);

( 𝑝 + 5)/12, if ≡ 7 (mod 12); and



 𝑝

 ( 𝑝 + 13)/12,
 if 𝑝 ≡ 11 (mod 12).

7. Let 𝑅 be local and O be an Eichler 𝑅-order of level 𝔭 (so O is hereditary, but
not maximal). Let 𝐾 be a quadratic 𝐹-algebra and𝑆 ⊆ 𝐾 an 𝑅-order. Show that
𝐾
𝑚(𝑆, O; O× ) = ∅ if and only if 𝑆 is maximal and = −1.
𝔭
8. Let 𝐵 be an indefinite quaternion algebra over Q of discriminant 𝐷 and let O ⊆ 𝐵
be an Eichler order. Let 𝑆 ⊆ 𝐾 be a quadratic order of discriminant 𝑑 𝑆 = 𝑑𝑓 2 ,
where 𝑑 is a fundamental discriminant.
(a) Suppose O is maximal. Show that there exists an optimal embedding
𝑆 ↩→ O if and only if for all 𝑝 | 𝐷, we have (𝐾 | 𝑝) ≠ 1 and 𝑝 - 𝑓 .
(b) Suppose O is Eichler of level 𝑀 and (for simplicity) that 𝑁 := 𝐷 𝑀 is odd.
Show that there exists an optimal embedding 𝑆 ↩→ O if and only if the
following conditions hold:
• For all 𝑝 | 𝐷, we have (𝐾 | 𝑝) ≠ 1.
• For all 𝑝 𝑒 k 𝑀, either (𝐾 | 𝑝) = 1 or 𝑝 𝑒 | 𝑑 𝑆 .
[Hint: recall that # Typ O = 1.]
[It is possible to extend this to all 𝑁 by analyzing local embeddings at 𝑝 = 2

working over Q2 .]
√
9. Let 𝐹 = Q( 𝑑) be a real quadratic field of discriminant 𝑑 > 0. Show that there
exists 𝑞 ≠ 2, 3 such that [𝐹 (𝜁2𝑞 ) : 𝐹] = 2 if and only if 𝑑 = 5, 8, 12.
10. Let 𝑅 be a global ring with field of fractions 𝐹, let 𝐵 be a quaternion algebra
over 𝐹, and let O, O0 ⊆ 𝐵 be maximal 𝑅-orders. Let 𝐾 ⊆ 𝐵 be a quadratic
𝐹-algebra in 𝐵, and suppose that 𝐾 ∩ O = 𝐾 ∩ O0; let 𝑆 be this common,
optimally embedded quadratic 𝑅-order. Show that there exists an invertible
fractional 𝑆-ideal 𝔞 ⊆ 𝐾 such that O𝔞 = 𝔞O0. [Hint: the statement is local, so
put the Eichler order O ∩ O0 and generator 𝛾 ∈ 𝑆 in standard form; show that
there exists 𝜇𝔭 ∈ O× ' GL2 (𝑅) and 𝛼𝔭 ∈ 𝑆𝔭 such that 𝜇𝔭 𝛼𝔭 = 𝜛𝔭 generates
the normalizer.]
11. In this exercise, we prove a local-global principle for optimal embeddings in a
self-contained manner. [For a much more precise result, read on to Chapter 31!]
Let 𝐹 be a number field with ring of integers 𝑅. Let 𝐵 be a quaternion algebra
over 𝐹 and let O ⊆ 𝐵 be an 𝑅-order. Let 𝐾 be a quadratic 𝐹-algebra and let
𝑆 ⊆ 𝐾 be an 𝑅-order. Suppose that 𝐾 ↩→ 𝐵; equivalently, by the local-global
principle for embeddings (Proposition 14.6.7), every place 𝑣 ∈ Pl 𝐹 does not
split in 𝐾.
(a) Suppose that for all primes 𝔭 of 𝑅, there exists an optimal embedding
𝜙𝔭 : 𝑆𝔭 ↩→ O𝔭 . Show that there exists an order O0 ∈ Gen(O) (so O0 is
locally isomorphic to O) and an optimal embedding 𝜙 : 𝑆 ↩→ O0. [Hint:
choose a maximal order containing 𝑆 and apply finitely many local cor-
rections.]
(b) Now suppose that # Cl𝐺 (O) 𝑅 = 1. Show that there exists an optimal
embedding 𝜙 : 𝑆 ↩→ O.
(c) Show that parts (a) and (b) follow from the trace formula (Theorem 30.4.7).
12. Give a direct proof of Corollary 30.4.23.
Chapter 31
Selectivity
In the previous chapter, we saw that (conjugacy classes of) embeddings of a quadratic
order into a quaternion algebra are naturally distributed over the genus of a quaternion
order; in applications, we want to compare the number of embeddings over orders in
a genus. Such a comparison can be thought of as a strong integral refinement of the
local-global principle for embeddings of quadratic fields (Proposition 14.6.7), which
belongs to the more general framework of the Albert–Brauer–Hasse–Noether theorem.
This chapter is quite technical, and it may be skipped on a first reading. To reward
the reader who persists, we conclude this chapter with the construction of isospectral,
nonisometric hyperbolic Riemannian manifolds, following Vignéras.
31.1 Selective orders

To get a preview of what selectivity is all about, right off the bat we give an example
of the failure for a quadratic order to embed equitably in the genus of an order.
√ √
Example 31.1.1. Let 𝐹 := Q( −5) and 𝑅 := Z𝐹 = Z[ −5]. √ Then Cl 𝑅 ' Z/2Z, and
the nontrivial class is represented by the ideal 𝔭 = h2, 1 + −5i ⊆ Z𝐹 with 𝔭2 = 2Z𝐹 .
By class field theory, the Hilbert class√field 𝐾 ⊇ 𝐹√ is a quadratic extension, and the
genus theory of Gauss gives 𝐾 = 𝐹 ( −1) = 𝐹 ( 5). The maximal order of 𝐾 is
Z𝐾 = Z𝐹 [𝑤] where
√ ! √ √
√ 1+ 5 −1 + −5
𝑤 = −1 =
2 2
√
satisfies 𝑤 2 − −5𝑤 − 1 = 0, a polynomial of discriminant −5 + 4 = −1.
We take 𝐵 = M2 (𝐹) and the maximal order O = M2 (Z𝐹 ). By 17.3.7, there is a
bĳection Cl 𝑅 −→ ∼ Cls O, with the nontrivial right ideal class represented by
R

𝔭 0 𝔭 𝔭
𝐼= M2 (𝑅) = ;
0 𝑅 𝑅 𝑅
its left order is
𝑅 𝔭
O0 = OL (𝐼) = .
𝔭−1 𝑅
557
558 CHAPTER 31. SELECTIVITY
These two orders are not isomorphic and up to isomorphism represent the two types
of maximal 𝑅-orders in M2 (𝐹).
We claim that there is an embedding Z𝐾 ↩→ O but no embedding Z𝐾 ↩→ O0.
The first part of the claim is easy: taking the rational canonical form, we take the
embedding

0 √1
𝑤 ↦→ 𝛼 = . (31.1.2)
1 −5
The proof that Z𝐾 6↩→ O0 is more difficult. (The embedding (31.1.2) does not
extend to O0 because of the off-diagonal coefficients; and we cannot conjugate this
embedding in an obvious way because the ideal 𝔭 is not principal.) Such an embedding
would be specified by a matrix

𝑎 𝑏√ 𝑅 𝔭
𝛼0 = ∈ −1
𝑐 −𝑎 + −5 𝔭 𝑅
with √
− det(𝛼 0) = 𝑎 2 − −5𝑎 + 𝑏𝑐 = 1; (31.1.3)
so the content in the second claim is that there is no solution to the quadratic equation
(31.1.3). √
Indeed, suppose there is a solution. Let 𝑓 (𝑥) = 𝑥 2 − −5𝑥 − 1 ∈ Z𝐹 [𝑥], so that
𝑓 (𝑎) + 𝑏𝑐 = 0. We may factor 𝑏Z𝐹 = 𝔭𝔟 with 𝔟 ⊆ Z𝐹 and [𝔟] ∈ Cl Z𝐹 nontrivial;
by parity, there exists a prime 𝔮 | 𝔟 with [𝔮] nontrivial. Factoring 𝑐Z𝐹 = 𝔭−1 𝔠 with
𝔠 ⊆ Z𝐹 , we have 𝑏𝑐Z𝐹 = 𝔟𝔠 ⊆ 𝔮, so 𝑓 (𝑎) = −𝑏𝑐 ≡ 0 (mod 𝔮). But 𝑓 (𝑥) has trivial
discriminant, and modulo a prime 𝔮 ⊆ Z𝐹 it either splits (into distinct linear factors) or
remains irreducible. And by the Artin map, 𝑓 (𝑥) splits modulo 𝔮 if and only if 𝔮 splits
in 𝐾 if and only if the class [𝔮] ∈ Cl Z𝐹 is trivial. Putting these two pieces together,
we have 𝑓 (𝑎) ≡ 0 (mod 𝔮) and 𝑓 (𝑥) is irreducible modulo 𝔮. This is a contradiction,
and there can be no solution.
With this cautionary but illustrative example in hand, we state our main theorem.
We return to the idelic notation of section 30.4. We will consider embeddings in the
context of strong approximation (see Chapter 28).
The following notation will be in use throughout this chapter.
31.1.4. Let 𝑅 = 𝑅 (T) be a global ring with eligible set T and let 𝐹 = Frac 𝑅 be its field
of fractions. Let 𝐵 be a quaternion algebra over 𝐹 and suppose that 𝐵 is T-indefinite.
Let O ⊆ 𝐵 be an 𝑅-order.
Let 𝐾 ⊇ 𝐹 be a separable quadratic 𝐹-algebra and let 𝑆 ⊆ 𝐾 be an 𝑅-order.
Suppose that Emb( 𝑆;b O)
b ≠ ∅, which is to say, for all primes 𝔭 ⊆ 𝑅, the 𝑅𝔭 -algebra 𝑆𝔭
embeds optimally into O𝔭 .
Our struggle will be to understand when the local optimal embeddings glue together
to give a global optimal embedding. As a start, we know by Corollary 30.4.18 that
there exists some order O0 ∈ Gen O in the genus of O (i.e., locally isomorphic to O)
such that Emb(𝑆, O0) ≠ ∅.
31.1. SELECTIVE ORDERS 559
Definition 31.1.5. We say that Gen O is genial for 𝑆 if Emb(𝑆, O0) ≠ ∅ for all
O0 ∈ Gen O. If Gen O is not genial, i.e., there exists O0 ∈ Gen O such that
Emb(𝑆, O0) = ∅,
then we say that Gen O is optimally selective for 𝑆.

By definition, Gen O is genial for 𝑆 if and only if 𝑆 embeds optimally in every
order O0 that is locally isomorphic to O.
31.1.6. We define the following condition, called the optimal selectivity condition:
(OS) 𝐾 is a subfield of the class field 𝐻𝐺 𝑁 (O) of 𝐹 obtained from Cl𝐺 𝑁 (O) 𝑅.
In particular, if 𝐾 is not a field, then (OS) does not hold. We now state our main
theorem, with notation and hypotheses in 31.1.4.
Main Theorem 31.1.7 (Optimal selectivity). Suppose that O is an Eichler order.
(a) Gen O is optimally selective for 𝑆 if and only if the optimal selectivity condition
(OS) holds.
(b) If Gen O is optimally selective for 𝑆, then Emb(𝑆, O0) ≠ ∅ for precisely half of
the types [O0] ∈ Typ O.
(c) In all cases,
𝑚(𝑆, O0; O0× ) = 𝑚(𝑆, O; O× )
for all O0 ∈ Gen O whenever both sides are nonzero.
Since the optimal selectivity condition (OS) only depends on 𝐾, if Gen O is
optimally selective for 𝑆 then it is optimally selective for all 𝑅-orders in 𝐾.
Remark 31.1.8. It was first noted by Chevalley [Chev36] in the more general situation
of matrix algebras that it was possible for a commutative order to embed into some, but
not all, maximal orders. An approach to selectivity is sketched by Vignéras [Vig80a,
Théorème III.5.15], but there are some glitches [CF99, Remark 3.4]. Maclachlan
[Macl2008, Theorem 1.4] gives a proof of Main Theorem 31.1.7(a)–(b) for hereditary
orders (Eichler orders of squarefree level). For a more detailed literature survey and
further comments, see 31.7.7.
31.1.9. When Gen O is optimally selective for 𝑆, then we can refine Main Theorem
31.1.7(b) detecting the half of types of orders for which there is an optimal embedding
of 𝑆.
From our hypothesis Emb( 𝑆, b O)
b ≠ ∅, we know that 𝑆 embeds into some order
in the genus of O; we might as well take this to be O itself, so we suppose that
Emb(𝑆, O) ≠ ∅. Let O0 ∈ Gen O. Then O0 is connected to O, so O0 = OL (𝐼) for an
invertible right O-ideal 𝐼 ⊆ O. Let
𝔞 = [O : O ∩ O0] = [O0 : O ∩ O0].
Then Emb(𝑆, O0) ≠ ∅ if and only if Frob𝔞 is trivial in Gal(𝐾 | 𝐹).

For example, if 𝔞 = 𝔭 is prime, then Frob𝔭 is trivial in Gal(𝐾 | 𝐹) if and only if 𝔭 is

not inert in 𝐾. For maximal orders, we can equivalently formulate the index in terms
of distance on the Bruhat–Tits tree (see section 23.5 and Exercise 23.9).
The core application of the optimal selectivity theorem is the following corollary.
Corollary 31.1.10. Suppose that Gen O is genial for 𝑆. Then
ℎ(𝑆)
𝑚(𝑆, O; O× ) = b O;
𝑚( 𝑆, b× ).
b O
# Cls O
Proof. We combine Main Theorem 31.1.7 and Theorem 30.4.7.
We conclude this introduction with a second application, a generalization of Corol-

lary 28.6.4. Let Ω be the set of real, ramified places of 𝐵, and let 𝑅>× Ω 0 := 𝑅 × ∩ 𝐹>×Ω 0
be the subgroup of 𝑅 × of elements that are positive at all 𝑣 ∈ Ω.
Corollary 31.1.11. Let O ⊆ 𝐵 be an Eichler 𝑅-order. Then
nrd(O× ) = 𝑅>× Ω 0 .
Proof. Let 𝑢 ∈ 𝑅>× Ω 0 . We repeat the argument of Corollary 28.6.4: we find 𝛾 0 ∈ O0

with nrd(𝛾 0) = 𝑢 and O0 ∈ Gen O. We may suppose further that Gen O is not selective
for 𝑅[𝛾 0] by shrinking the open set to ensure that 𝐾 = Frac 𝑅[𝛾 0] * 𝐻𝐺 𝑁 (O) . Let
𝑆 = 𝐾 ∩ O0. Then 𝑆 ⊆ O0 is optimally embedded; and by Main Theorem 31.1.7(c),
there exists an optimal embedding 𝜙 : 𝑆 ↩→ O, hence 𝜙(𝛾 0) = 𝛾 ∈ O has nrd(𝛾) = 𝑢
as desired.
31.2 Selectivity conditions

In this brief section, we make the somewhat opaque optimal selectivity condition (OS)
explicit for Eichler orders.
Proposition 31.2.1. Let O be an Eichler order of level 𝔐. Then Condition (OS) holds
if and only if all of the following four conditions hold:
(a) The extension 𝐾 ⊇ 𝐹 and the quaternion algebra 𝐵 are ramified at the same
(possibly empty) set of archimedean places of 𝐹;
(b) 𝐾 and 𝐵 are unramified at all nonarchimedean places 𝑣 ∈ Pl 𝐹;
(c) Every nonarchimedean place 𝑣 ∈ T splits in 𝐾; and
(d) If 𝔭 ⊂ 𝑅 is a nonzero prime and ord𝔭 (𝔐) is odd, then 𝔭 splits in 𝐾.
Proof. We determine the class field 𝐻𝐺 𝑁 (O) obtained from the group 𝐺 𝑁 (O) =
𝐹>×Ω 0 nrd(𝑁 𝐵b× ( O)).
b
Recall we have 𝐺 (O) = 𝐹>×Ω 0 nrd( O
b× ) = 𝐹 × 𝑅
>Ω 0
b× , since O is an Eichler order
and therefore locally norm-maximal, so 𝐻𝐺 (O) is the maximal abelian extension of 𝐹
unramified away from the real places in Ram(𝐵) and such that the remaining places
𝑣 ∈ T split completely.
31.3. ∗ SELECTIVITY SETUP 561
The normalizer nrd(𝑁 𝐵b× ( O))

b is the restricted direct product of local normalizers,
computed in (23.2.8) for 𝔭 | 𝔇 and Corollary 23.3.14 for 𝔭 - 𝔇: for the latter,
(
𝐹𝔭×2 𝑅𝔭× , if ord𝔭 (𝔐) is even;
nrd(𝑁 𝐵𝔭× (O𝔭 )) =
𝐹𝔭× , if ord𝔭 (𝔐) is odd.
Therefore, the quotient Cl𝐺 (O) 𝑅 → Cl𝐺 𝑁 (O) 𝑅 factors through the quotient by squares
Cl𝐺 (O) 𝑅/(Cl𝐺 (O) 𝑅) 2 and then the further quotient by the primes 𝔭 | 𝔇 = disc 𝐵 and
𝔭 | 𝔐 with ord𝔭 (𝔐) odd.
We now conclude the proof. If 𝐾 is not a field then all of the desired conclusions
hold; so suppose 𝐾 is a field. Since 𝐾 ↩→ 𝐵, if 𝑣 ∈ Pl 𝐹 ramifies in 𝐵 then 𝑣 also
ramifies in 𝐵. A containment 𝐾 ⊆ 𝐻𝐺 𝑁 (O) is permitted at archimedean places if and
only if the archimedean ramification in 𝐾 ⊇ 𝐹 is no bigger than this. In a similar
way, the conditions in the previous paragraph establish (c)–(d), and 𝐾 is unramified at
all nonarchimedean places 𝑣. To conclude (b), if 𝔭 | 𝔇 then 𝔭 splits in 𝐻𝐺 𝑁 (O) and
therefore in 𝐾; but we are assuming that 𝐾 ↩→ 𝐵 so 𝐾𝔭 ↩→ 𝐵𝔭 , a contradiction since
𝐵𝔭 is a division algebra, and there can be no such 𝔭.
31.3 ∗ Selectivity setup

We now embark on a proof of the selectivity theorem (Main Theorem 31.1.7); this goal
will occupy us for the remainder of this chapter. In this section, we begin to isolate
the problem: there is a group that is at worst Z/2Z and is usually trivial, and we pin it
down using strong approximation, the reduced norm, and class field theory. Our basic
reference is Vignéras [Vig80a, Théorème III.5.15], and the surrounding text.
Our notation is as in 31.1.4.
31.3.1. To establish the main theorem in the case where 𝐾 ' 𝐹 × 𝐹 is straightforward.
We leave this case as an exercise (Exercise 31.1).
We suppose throughout the rest of this chapter that 𝐾 is a field.
Let O1 ≤ Γ ≤ 𝑁 𝐵× (O) (as in 30.3.10). Recall that there is a bĳection (30.3.13)
Emb(𝑆, O; Γ) ↔ 𝐾 × \𝐸/Γ
where
𝐸 = {𝛽 ∈ 𝐵× : 𝐾 𝛽 ∩ O = 𝑆 𝛽 }
and we abbreviate conjugation 𝐾 𝛽 = 𝛽−1 𝐾 𝛽 for conciseness. Conjugating if necessary,
we may suppose that 1 ∈ 𝐸, i.e., we start with an order and an optimal embedding
𝐾 ∩ O = 𝑆.
We employ idelic notation as in section 30.4. The inclusion 𝐵× ↩→ 𝐵 b× gives an
inclusion
𝐸/Γ ↩→ 𝐸/ bb Γ (31.3.2)
with
b× : 𝐾

b := 𝛽b ∈ 𝐵
𝐸 b𝛽b ∩ O
b = 𝑆b𝛽b .
The hypothesis of strong approximation allows us to identify precisely the image
of the map (31.3.2) via the reduced norm in the following way.
31.3.3. As in Theorem 28.5.5 (a motivating application of strong approximation), the

reduced norm induces a bĳection
𝐵× \ 𝐵
b× /b ∼ 𝐹 × \𝐹
Γ −→ b× /nrd(b
Γ) = Cl𝐺 (Γ) 𝑅 (31.3.4)
>Ω 0
where 𝐺 (Γ) := 𝐹>×Ω 0 nrd(b

Γ).
𝐸/Γ = { 𝛽bb
Γ ∈ 𝐸/
bb b ∈ 𝐺 (Γ)} ⊆ 𝐸/
Γ : nrd( 𝛽) bb Γ.
That is to say, if 𝛽b ∈ 𝐸,
b then there exists 𝛽 ∈ 𝐸 such that 𝛽bb
Γ = 𝛽b
Γ if and only if
b ∈ 𝐺 (Γ).
nrd( 𝛽)
Proof. We find a 𝛽 ∈ 𝐵× (without the condition that 𝛽 ∈ 𝐸) immediately from the

b𝛾 ∈ 𝛽bb
bĳection (31.3.4). But 𝛽 = 𝛽b Γ ≤ 𝑁 𝐵b× ( O)
Γ and b b gives
b𝛽 ∩ O
𝐾 b𝛾b = 𝐾
b𝛽 ∩ O
b = 𝑆b𝛽
and intersecting with 𝐵 we find 𝛽 ∈ 𝐸.
Lemma 31.3.5 points the way more generally, at least to detect if there is an
embedding in the first place in an order. First, we need to give representatives of the
type set.
31.3.6. Recalling 28.5.7, there is a bĳection
Typ O ↔ 𝐵× \ 𝐵
b× /𝑁 b× ( O);
𝐵
b
explicitly, every isomorphism class of order in Typ O is of the form
O0 = b bb𝜈 −1 ∩ 𝐵
b𝜈 −1 ∩ 𝐵 = O
𝜈 Ob
b× in 𝐵× \ 𝐵
𝜈∈𝐵
(yes, the choice of inverse is deliberate), with the class of b b× /𝑁 b× ( O)
b
𝐵
uniquely defined.
In the presence of strong approximation (Corollary 28.5.10), we have a further
bĳection
Typ O ↔ Cl𝐺 𝑁 (O) 𝑅
where
𝐺 𝑁 (O) = 𝐺 (𝑁 𝐵× (O)) = 𝐹>×Ω 0 nrd(𝑁 𝐵b× ( O)).
b
Now we look back at embeddings and reduced norms.

31.3.7. Let
b = {nrd( 𝛽)
nrd( 𝐸) b : 𝛽b ∈ 𝐸 b× .
b} ⊆ 𝐹 (31.3.8)
The set 𝐸 b = {nrd( 𝛽)
b does not obviously have a group structure, so nrd( 𝐸) b : 𝛽b ∈ 𝐸
b}
is a subset. But 𝐸b is stable under 𝐾b× , so nrd( 𝐸) b× ).
b is a union of cosets of nrd( 𝐾
The set nrd( 𝐸)
b is quite large, because it contains reduced norms from 𝐾 b× ⊆ 𝐸 b and
𝑁 b× ( O)
𝐵
b ⊆ 𝐸.b
31.3. ∗ SELECTIVITY SETUP 563
31.3.9. By the main theorem of class field theory (Theorem 27.5.10), the Artin map
gives a bĳection
∼ Gal(𝐾 | 𝐹) ' Z/2Z.
𝐹 × /𝐹 × Nm𝐾 |𝐹 (𝐾 × ) −→ (31.3.10)
Let S ⊆ T be the set of archimedean places of 𝐹. Recall the isomorphism (27.5.8)
𝐹 × /𝐹 × 𝐹∞,>0
× ∼ 𝐹 × /𝐹 × .
−→ 6𝑆 >0
We have
× ×
𝐹∞,>0 ≤ Nm𝐾 |𝐹 (𝐾∞ ) ≤ Nm𝐾 |𝐹 (𝐾 × )
(the latter embedded at the infinite place with the other components 1), and by the same
argument as in Lemma 27.5.6, the image of 𝐹 × Nm𝐾 |𝐹 (𝐾 × ) under the isomorphism
(27.5.8) is 𝐹>×Σ 0 Nm𝐾 |𝐹 (𝐾6𝑆× ) where Σ ⊆ Pl(𝐹) is the set of places ramified in 𝐾 (going
from real to complex in the extension 𝐹 ⊆ 𝐾) and
𝐹>×Σ 0 := {𝑎 ∈ 𝐹 × : 𝑣(𝑎) > 0 for all real 𝑣 ∈ Σ}.
Therefore we have an isomorphism

∼ 𝐹 × /𝐹 × Nm
𝐹 × /𝐹 × Nm𝐾 |𝐹 (𝐾 × ) −→ ×
6𝑆 >Σ 0 𝐾 |𝐹 (𝐾6𝑆 ). (31.3.11)
b× = 𝐹 × to obtain the map

We then further project from the target of (31.3.11) to 𝐹 6T
𝐹 × /𝐹 × Nm𝐾 |𝐹 (𝐾 × ) → 𝐹
b× /𝐹 × Nm𝐾 |𝐹 ( 𝐾
>Σ 0
b× ). (31.3.12)
𝐹>×Σ 0 Nm𝐾 |𝐹 ( 𝐾
b× ) ≤ 𝐹
b×
with total index at most 2, and index equal to 2 if and only if every nonarchimedean
place 𝑣 ∈ T is split in 𝐾.
Proof. In the projection (31.3.12), we start with a group of order 2; in order to keep
it this size, the projection away from the nonarchimedean places in T must be an
isomorphism, which holds if and only if for all nonarchimedean places 𝑣 ∈ T we must
have 𝑣 split in 𝐾.
We conclude this setup section with an overview.
31.3.14. Selectivity arises from an examination of layers in the following selectivity

sandwich:
(OS) 𝑚 𝑠
𝐹>×Σ 0 nrd( 𝐾
b× ) ≤ 𝐹 × nrd( 𝐾
>Ω 0
b× ) nrd(𝑁 b× ( O))
𝐵
b ≤ 𝐹 × nrd( 𝐸)
>Ω 0
b×
b ≤𝐹 (31.3.15)
In terms of the sandwich bread, the left-most group 𝐹>×Σ 0 nrd( 𝐾

b× ) has index at most
b× by Lemma 31.3.13. The set 𝐹 × nrd( 𝐸)
2 in the right-most group 𝐹 b is stable under
>Ω 0
multiplication by 𝐹>×Σ 0 nrd( 𝐾b× ) so is a union of its cosets in 𝐹

b× ; therefore, it is actually
a subgroup.
Much ado about a (possible) group of order two!
Again by Lemma 31.3.13, the sandwich collapses if there is a nonarchimedean
place 𝑣 ∈ T that is inert in 𝐾 ⊇ 𝐹, so there is only work to do when every 𝑣 ∈ T is
split. Under this assumption, we will show in Lemma 31.4.1 that the first inequality
labelled (OS) is an equality if and only if the optimal selectivity condition (OS) holds.
In Propositions 31.5.1 and 31.5.7, we will show that the middle inequality labelled 𝑚
is always an equality and that such an equality implies equality of embedding numbers
(when they are nonzero). Last but not least, in Proposition 31.4.4 we will show that the
final inequality labelled 𝑠 is an equality if and only if there is no selectivity obstruction,
i.e., Emb(𝑆, O0) ≠ ∅ for all O0 ∈ Gen O.
31.4 ∗ Outer selectivity inequalities

In this section, we consider the outer ends of the selectivity sandwich 31.3.14.
The left-most inequality is interpreted in the language of class field theory as
follows.
b× ) ≤ 𝐹 × nrd( 𝐾
>Ω 0
b× ) nrd(𝑁 b× ( O))
𝐵
b (31.4.2)
with index at most 2, and equality holds if and only if either the optimal selectivity
condition (OS) holds or there exists a nonarchimedean place 𝑣 ∈ T inert or ramified
in 𝐾.
Proof. Recall that 𝐺 𝑁 (O) = 𝐹>×Ω 0 nrd(𝑁 𝐵b× ( O)).

b In the side sandwich
b× ) ≤ 𝐹 × nrd( 𝐾
>Ω 0
b× ) nrd(𝑁 b× ( O))
𝐵
b ≤𝐹b× (31.4.3)
we again have total index at most 2. By class field theory and the Galois correspondence
relative to the corresponding tower of class groups, we have
b× ) = 𝐹 × nrd( 𝐾
>Ω 0
b× ) nrd(𝑁 b× ( O))
𝐵
b <𝐹b×
(so first equality, then strict inequality in (31.4.3)) if and only if 𝐾 ⊆ 𝐻𝐺 𝑁 (O) if and
only if (OS) holds. The result then follows by Lemma 31.3.13.
We next consider the right-most inequality, and we show that it contains the
obstruction to selectivity.
Proposition 31.4.4. Let [O0] ∈ Typ O be represented by the class 𝐵×b

𝜈 𝑁 𝐵b× ( O).
b Then
0
Emb(𝑆, O ) ≠ ∅ if and only if nrd(b ×
𝜈 ) ∈ 𝐹>Ω 0 nrd( 𝐸).
b
Proof. Define
0 0
𝐸 0 = {𝛽 0 ∈ 𝐵× : 𝐾 𝛽 ∩ O0 = 𝑆 𝛽 }
31.5. ∗ MIDDLE SELECTIVITY EQUALITY 565
∼ 𝐾 × \𝐸 0, and similarly 𝐸
so that Emb𝑅 (𝑆, O0) −→ b0.
Suppose Emb(𝑆, O ) ≠ ∅, represented by 𝛽 ∈ 𝐸 0. Then
0 0
b𝛽0 ∩ O
𝐾 b𝛽0 ∩ O
b0 = 𝐾 bb𝜈 −1 = 𝑆b𝛽0
so 𝛽 0b
𝜈 = 𝛽b ∈ 𝐸.
b Therefore
𝜈 ) = nrd(𝛽 0−1 ) nrd( 𝛽)

nrd(b b ∈ 𝐹 × nrd( 𝐸).
>Ω 0
b
Conversely, suppose that nrd(b 𝜈 ) ∈ 𝐹>×Ω 0 nrd( 𝐸);

b then there exists 𝑎 ∈ 𝐹 × and
>Ω 0
𝛽b ∈ 𝐸 𝜈 ) = 𝑎 nrd( 𝛽).
b such that nrd(b b Since 𝛽b ∈ 𝐸, b we have
b𝛽b ∩ O
𝐾 b = 𝑆b𝛽b
thus if 𝛽b0 = 𝛽b
b𝜈 −1 we get
b𝛽b0 ∩ O
𝐾 bb𝜈 −1 = 𝐾
b𝛽b0 ∩ O
b0 = 𝑆b𝛽b0 (31.4.5)
and 𝛽b0 ∈ 𝐸
b0. We have
nrd( 𝛽b0) = nrd( 𝛽b

b𝜈 −1 ) = 𝑎 −1 ∈ 𝐹 ×
>Ω 0
So by Lemma 31.3.5, there exists 𝛽 0 ∈ 𝐸 0 mapping to 𝛽b0, and Emb(𝑆, O0) ≠ ∅ as

claimed.
The following corollary indicates the significance of the preceding proposition.
Corollary 31.4.6. If 𝐹>×Ω 0 nrd( 𝐸)
b = 𝐹 b× , then Emb(𝑆, O0) ≠ ∅ for all orders O0 ∈
×
Gen O in the genus of O. Otherwise, 𝐹>Ω 0 nrd( 𝐸) b <𝐹b× has index 2 and Emb(𝑆, O0) ≠
∅ for precisely half of the types of orders in Typ O: we have Emb(𝑆, O0) ≠ ∅ for
O0 = b 𝜈 −1 ∩ 𝐵
𝜈 Ob
𝜈 ) ∈ 𝐹>×Ω 0 nrd( 𝐾
if and only if nrd(b b× ).
In particular, in the latter case we have # Typ O even.

Proof. We apply Proposition 31.4.4, with indexing of the type set as in 31.3.6.
31.5 ∗ Middle selectivity equality

In this section, we pursue the middle inequality in the selectivity sandwich 31.3.14.
First, we show that equality in this middle equality implies equality of embedding
numbers, whenever they are nonzero.
𝐹>×Ω 0 nrd( 𝐾
b× ) nrd(𝑁 b× ( O))
𝐵
b ≤ 𝐹 × nrd( 𝐸)
>Ω 0
b (31.5.2)
with index at most 2. If equality holds in (31.5.2), then whenever O0 ∈ Gen O and
Emb(𝑆, O0) is nonempty, we have
𝑚(𝑆, O; O× ) = 𝑚(𝑆, O0; O0× ).
Proof. The statement about index follows from the layering of the sandwich (31.3.15).
For the second statement, suppose that Emb(𝑆, O0) is nonempty; then by Proposition
31.4.4, we have O0 = O bb𝜈 −1 ∩ 𝐵 with nrd(b𝜈 ) ∈ 𝐹>×Ω 0 nrd( 𝐸).
b If equality holds in
×
(31.5.2), then there exists 𝑎 ∈ 𝐹 , b 𝛼∈𝐾 ×
b , and 𝜂b ∈ 𝑁 b× ( O)
b such that
>Ω 0 𝐵
𝜈 ) = 𝑎 nrd(b
nrd(b 𝛼) nrd(b
𝜂). (31.5.3)
We restore notation from Proposition 31.4.4, and modify the argument in the
converse. We define the map
b0
𝐸→𝐸
(31.5.4)
𝛽 ↦→ 𝛽b0 = b
𝛼 𝛽b𝜈 −1
𝜂b
b𝛽 ∩ O
We argue as in (31.4.5). From 𝛽 ∈ 𝐸, we have 𝐾 b× , so
𝛼∈𝐾
b = 𝑆b𝛽 . We have b
𝛼
b = 𝐾.
𝐾 b b And 𝜂b ∈ 𝑁 b× ( O), so O = O. Therefore
b b 𝜂
b b
𝐵
b𝛽b0 ∩ O
𝐾 bb𝜈 −1 = 𝐾
b𝛽b0 ∩ O
b0 = 𝑆b𝛽b0 (31.5.5)
and indeed 𝛽b0 ∈ 𝐸

b0. Finally, by (31.5.3)
nrd( 𝛽b0) = nrd(b

𝛼 𝛽b𝜈 −1 ) = 𝑎 −1 nrd(𝛽) ∈ 𝐹>×Ω 0 .
𝜂b
By Lemma 31.3.5, there exists 𝛽 0 ∈ 𝐸 0 such that 𝛽 0O b0× , well-defined up to

b0× = 𝛽b0O
× 0 0×
O . Therefore, (31.5.4) descends to a map 𝐸 → 𝐸 /O , and it further descends to a
map
𝐸/O× → 𝐸 0/O0×
(31.5.6)
𝛽O× ↦→ 𝛽 0O0×
because
𝛼 𝛽𝜇b
b 𝜈 −1 O
𝜂b b0 = b
𝛼 𝛽𝜇bb𝜈 −1 = b
𝜂Ob 𝛼 𝛽𝜇Ob𝜈 −1 = b
b𝜂b 𝛼 𝛽Ob𝜈 −1 = b
b𝜂b 𝛼 𝛽b𝜈 −1 O
𝜂b b0 .
This map works as well interchanging the roles of O and O0, and after a little chase,
we verify that the map (31.5.6) is bĳective. Taking orbits under 𝐾 × on the left, we
conclude the proof.
In fact, equality holds in the middle for Eichler orders.

Proposition 31.5.7. If O is an Eichler order, then the inequality (31.5.2) is an equality.
Proof. We will prove that
b× ) nrd(𝑁 b× ( O))
nrd( 𝐾 b = nrd( 𝐸);
b (31.5.8)
𝐵
the inclusion (≤) was direct, so we prove (≥). The desired equality is now idelic, so
we reduce to checking in the completion at a prime 𝔭. If 𝐵𝔭 is a division algebra, then
O𝔭 is maximal, and 𝑚(𝑆𝔭 , O𝔭 ; 𝑁 𝐵𝔭× (O𝔭 )) = 1 by Proposition 30.5.3(b) we have the
stronger equality 𝐸 𝔭 = 𝑁 𝐵𝔭× (O𝔭 ) = 𝐵𝔭× .
31.6. ∗ OPTIMAL SELECTIVITY CONCLUSION 567
Otherwise, 𝐵𝔭 ' M2 (𝐹𝔭 ) is split. Without loss of generality, we may suppose

that O𝔭 is a standard Eichler order. Further, by Lemma 30.6.3, after conjugating
by a normalizing element if necessary, we may suppose that the reference optimal
embedding 𝑆𝔭 ↩→ O𝔭 is normalized. But then this embedding extends to an optimal
embedding 𝑆𝔭 ↩→ M2 (𝑅𝔭 ): the upper-right entry is 1.
𝛽
Now let 𝛽𝔭 ∈ 𝐸 𝔭 be arbitrary, with associated embedding 𝜙𝔭 𝔭 : 𝑆𝔭 ↩→ O𝔭 . We
repeat the argument in the previous paragraph: by Lemma 30.6.3, replacing 𝛽𝔭 by 𝛽𝔭 𝜈𝔭
𝛽
if necessary with 𝜈𝔭 ∈ 𝑁 𝐵𝔭× (O𝔭 ), we may suppose that 𝜙𝔭 𝔭 is normalized, and therefore
extends to an optimal embedding into M2 (𝑅𝔭 ). But by Proposition 30.5.3(a), we have
𝑚(𝑆𝔭 , M2 (𝑅𝔭 ); GL2 (𝑅𝔭 )) = 1—all optimal embeddings into M2 (𝑅𝔭 ) are conjugate
𝛽 𝜇
under GL2 (𝑅𝔭 )—so there exists 𝜇𝔭 ∈ GL2 (𝑅𝔭 ) such that 𝜙𝔭 𝔭 = 𝜙𝔭 𝔭 . Therefore
𝛽𝔭 ∈ 𝜇𝔭 𝐾𝔭× so
nrd(𝛽𝔭 ) ∈ nrd(𝐾𝔭× )𝑅𝔭× ≤ nrd(𝐾𝔭× ) nrd(O×𝔭 ) ≤ nrd(𝐾𝔭× ) nrd(𝑁 𝐵𝔭× (O𝔭 )) (31.5.9)
as claimed.
31.6 ∗ Optimal selectivity conclusion

We now officially complete the proof of the selectivity theorem for Eichler orders.
Proof of Main Theorem 31.1.7. By 31.3.1, we may suppose 𝐾 is a field. We refer to

the selectivity sandwich (31.3.15), using Proposition 31.5.7 to simplify the middle
equality:
(OS) 𝑠
𝐹>×Σ 0 nrd( 𝐾 b× ) nrd(𝑁 b× ( O))
b× ) ≤ 𝐹 × nrd( 𝐾 𝑚 ×
b = 𝐹>Ω 0 nrd( 𝐸) b× ,
b ≤𝐹
>Ω 0 𝐵
with total index at most 2.

By Proposition 31.4.4, we have that Gen O is optimally selective for 𝑆 if and only
if the right-most inequality (labelled 𝑠) is strict. Such an inequality is strict if and only
if the total index is 2 and the left-most inequality is an equality. By Lemma 31.4.1,
this happens if and only if the condition (OS) holds. This proves (a).
Statement (b) is a restatement of Corollary 31.4.6, and statement (c) follows from
the second statement in Proposition 31.5.1.
It has been a long and pretty technical road, so as refreshment we work through an
example (cf. Maclachlan [Macl2008, §4, Example 1]).
Example 31.6.1. Let 𝐹 be the totally real cubic field Q(𝑏) where 𝑏 3 − 4𝑏 − 1 = 0;
then 𝐹 has discriminant 229 and ring of integers 𝑅 = Z𝐹 = Z[𝑏]. The usual class
group Cl 𝑅 is trivial, but the narrow class group is Cl+ 𝑅 ' Z/2Z, represented by the
ideal 𝔭 = (𝑏 + 1)Z𝐹 of norm 2—the ideal 𝔭 is principal, but there is no generator√that
+ +
is totally positive.
narrow class field 𝐾 = 𝐻 ⊇ 𝐹 is quadratic, with 𝐻 = 𝐹 ( 𝑏).
The
−1, 𝑏
Let 𝐵 = . Then 𝑏 is positive at precisely one real place and negative at
𝐹
the other two, and 𝑏 ∈ Z×𝐹 . Computing the Hilbert symbol at the even primes, we
conclude that Ram(𝐵) is equal to two real places. In particular, 𝐵 is indefinite. The
class group Cl𝐺 (O) 𝑅 with modulus equal to these two real places is equal to Cl+ 𝑅, as
we see by the real signs of 𝑏.
Next, we compute representatives of the type set of maximal orders for 𝐵. By strong
approximation (Corollary 28.5.10), we have Typ O in bĳection with Cl𝐺 𝑁 (O) 𝑅, so
we need to compute the idelic normalizer: but 𝐵 is unramified at all nonarchimedean
places, and
𝑁 𝐵b× ( O)
b =𝑁 b× b×
b) (M2 ( 𝑅)) = 𝐹 O .
b
GL2 ( 𝐹
Thus nrd(𝑁 𝐵b× ( O))

b =𝐹b×2 𝑅
b× , and
𝐺 𝑁 (O) = 𝐹>×Ω 0 𝐹
b×2 𝑅
b× = 𝐹 × 𝑅
>Ω 0
b× = 𝐺 (O).
In other words, the quotient map Cl𝐺 (O) 𝑅 → Cl𝐺 𝑁 (O) 𝑅 is an isomorphism, still a
group of order 2. We conclude that # Typ O = 2.
We compute a maximal order
𝑏2𝑖 + 𝑗 𝑏2 + 𝑖 𝑗
O = O1 = Z 𝐹 ⊕ Z 𝐹 𝑖 ⊕ Z 𝐹 ⊕ Z𝐹 .
2 2
We conjugate this order by an ideal of reduced norm 𝔭 to get the second representative
(𝑏 2 + 𝑏 + 1) + (𝑏 + 1)𝑖 + 𝑗 (𝑏 + 1) + (𝑏 2 + 𝑏 + 1)𝑖 + 𝑖 𝑗
O2 = Z 𝐹 ⊕ Z 𝐹 𝑖 ⊕ Z 𝐹 ⊕ Z𝐹 .
2 2
Therefore these orders represent the two types of maximal orders, and Typ O =
{[O1 ], [O2 ]}.
With all of these elements in place, we can observe selectivity (Main Theorem
31.1.7). We saw that both 𝐾 and 𝐵 are ramified at no nonarchimedean places and
exactly the same set of real places. In particular, the field 𝐾 ↩→ 𝐵 embeds by the
local-global principle. Let 𝑆 = Z𝐾 = Z𝐹 [𝑤] be the maximal order in 𝐾. Then
𝑤 2 − 𝑏𝑤 + 1 = 0. Then Emb( 𝑆;b O)
b ≠ ∅ (Proposition 30.5.3(a)).
The optimal selective condition (OS) holds because we took it so, 𝐾 = 𝐻 + . It
follows that 𝑆 embeds in exactly one of O1 or O2 . We find that
𝑏2 + 𝑖 𝑗
𝛼= ∈ O1
2𝑏
satisfies 𝛼2 − 𝑏𝛼 + 1 = 0 as desired; so 𝑆 embeds in O1 (and not O2 ).
Remark 31.6.2. Without the hypothesis of strong approximation, it is very difficult to
tease apart the contributions from different orders in the genus: indeed, the generating
series for representation numbers for a definite quaternion order give coefficients of
modular forms, discussed in Chapter 41.
31.7 ∗ Selectivity, without optimality

To conclude this chapter, we compare the above with a weaker condition than optimal
selectivity, and close with connections to the literature. We continue notation and
hypotheses from 31.1.4.
31.7. ∗ SELECTIVITY, WITHOUT OPTIMALITY 569
Definition 31.7.1. We say that Gen O is selective for 𝑆 if there exists O0 ∈ Gen O
such that there is no embedding 𝑆 ↩→ O0 of 𝑅-algebras.
The difference between Definition 31.1.5 and Definition 31.7.1 is that in the latter,
we do not insist that the embedding is optimal. It may happen that Gen O is selective
for 𝑆, but Gen O is not optimally selective for 𝑆: such a situation arises exactly when
there is an order O0 ∈ Gen O such that 𝑆 embeds in O0 but does not optimally embed
in O0.
Example 31.7.2. We return to Example 31.1.1. We saw that Gen O is optimally

√ order Z𝐾 . By Main Theorem 31.1.7, Gen O is also optimally
selective for the maximal
selective for 𝑆 = Z𝐹 [ −1] ⊆ Z𝐾 .
We claim that Gen O is not selective for 𝑆. For O = M2 (Z𝐾 ), we take the
normalized embedding

0 1
𝑤 ↦→ 𝛼 = .
−1 0

𝑅 𝔭
The surprise is that we can also find an embedding 𝑆 ↩→ O0 = −1 , just not an
𝔭 𝑅
optimal one: we take
√
0 − −5 √2
𝑤 ↦→ 𝛼 = . (31.7.3)
2 −5
(The argument given in Example 31.1.1 no longer applies, because the polynomial
𝑥 2 − 1 has nontrivial discriminant, giving just enough room for the prime 𝔭 to sneak
in.)
For sanity (to show there is no contradiction with the main theorem of optimal
selectivity), we confirm that the embedding (31.7.3) does not define an optimal em-
bedding into O0. We have 2Z𝐹 = 𝔭2 , so
√
0 1 − −5 2√
𝛼 +1= ∈ 𝔭O0
2 1 + −5
so the order 𝑅 + 𝔭−1 (𝛼 0 + 1) ⊇ 𝑅 + 𝑅𝛼 0 ' 𝑆 embeds in O0.
31.7.4. If Gen O is optimally selective for 𝑆 but not selective for 𝑆, then 𝑆 ⊆ O is
optimal but there is an order O0 such that 𝜙 0 : 𝑆 ↩→ O0 is an embedding but not an
optimal embedding. Let 𝑆 0 = 𝜙 0 (𝐾) ∩ O0 ) 𝑆. So there exists a prime 𝔭 | [𝑆 0 : 𝑆] 𝑅 ,
and in particular, 𝑆 is not maximal at 𝔭. In particular, if 𝑆 is integrally closed, then
Gen O is selective for 𝑆 if and only if Gen O is optimally selective for 𝑆.
Theorem 31.7.5 (Chinburg–Friedman, Chan–Xu, Guo–Qin). Let O be an Eichler

order of level 𝔐 and suppose that Gen O is optimally selective for 𝑆. Then Gen O is
selective for 𝑆 if and only if the following condition holds:
(S) If 𝔭 | disc𝑅 𝑆 and ord𝔭 (𝔐) ≠ ord𝔭 (disc𝑅 𝑆), then 𝔭 splits in 𝐾 ⊇ 𝐹.
Proof. Our very setup (section 31.3) is designed to count optimal embeddings, so
to avoid lengthening this chapter, we refer the reader to Chinburg–Friedman [CF99,
Theorem 3.3] for the case of maximal orders, and Chan–Xu [CX2004, Theorem 4.7]
and Guo–Qin [GQ2004, Theorem 2.5] (independently) for Eichler orders.
Remark 31.7.6. The condition Proposition 31.2.1(d) (one part of (OS) for Eichler
orders) is not visible in the selectivity theorem for Eichler orders, but is implied by it:
by (b), the extension 𝐾 ⊇ 𝐹 is unramified so 𝔭 is unramified in 𝐾, thus ord𝔭 (disc𝑅 (𝑆))
is even and necessarily not equal to ord𝔭 (𝔐) if the latter is odd.
31.7.7. Chinburg–Friedman [CF99, Theorem 3.3] prove Theorem 31.7.5 for maximal
orders, and they applied this theorem to embeddings in maximal arithmetic groups
[CF99, Theorem 4.4]. Chinburg–Friedman proved their results in the language of the
Bruhat–Tits tree of maximal orders. This selectivity theorem was then generalized to
Eichler orders by Chan–Xu [CX2004, Theorem 4.7] and Guo–Qin [GQ2004, Theorem
2.5] (independently). Interestingly, while Guo–Qin follow Chinburg–Friedman in their
proof, Chan–Xu instead use results on exceptional spinor genera and their results are
phrased and proven in the language of indefinite integral quadratic forms. (These
results are given for number fields, but the proofs adapt to global fields as pursued
here.)
Some selectivity theorems beyond those for Eichler orders are also known. Arenas-
Carmona [A-C2013, Theorem 1.2] considers more general intersections of maximal
orders. Linowitz [Lin2012, Theorems 1.3–1.4] gives a selectivity theorem for (op-
timal) embeddings into arbitrary orders, subject to some additional technical (copri-
mality) hypotheses. More generally, selectivity theorem have been pursued in the
more context of central simple algebras: see e.g. Linowitz–Shemanske [LS2012] and
Arenas-Carmona [A-C2012].
However, these selectivity results do not prove Main Theorem 31.1.7 on the nose,
either because they deal with selectivity instead of optimal selectivity or do not prove
the more powerful statement that the embedding numbers are in fact equal. On the
latter point, a general setup to establish equality of embedding numbers can be found
in work of Linowitz–Voight [LV2015, §2].
31.8 ∗ Isospectral, nonisometric manifolds

We conclude with an application to geometry. We need to borrow from the future, so
the reader is invited to read this section lightly, using it as present and future motivation;
and then to return to this section after a more careful reading of Chapter 38.
In 1966, Kac [Kac66] famously asked: “Can one hear the shape of a drum?” Put
another way, if you know the frequencies at which a drum vibrates, can you determine
its shape? This beautiful question has led to an almost countless number of articles:
see Giraud–Thas [GT2010] for a survey.
To restate the question in a mathematical framework, let 𝑀 be a connected, compact
Riemannian manifold. Associated to 𝑀 is the Laplace operator, defined by Δ( 𝑓 ) :=
− div(grad( 𝑓 )) for 𝑓 ∈ 𝐿 2 (𝑀) a square-integrable function on 𝑀. The eigenvalues of
Δ on the space 𝐿 2 (𝑀) form an infinite, discrete sequence of nonnegative real numbers
31.8. ∗ ISOSPECTRAL, NONISOMETRIC MANIFOLDS 571
0 = 𝜆0 < 𝜆1 ≤ 𝜆2 ≤ . . . , called the spectrum of 𝑀. By the study of the wave

equation, the spectrum of 𝑀 can essentially be thought of as the frequencies produced
by a drum “in the shape of 𝑀”. Two Riemannian manifolds are said to be Laplace
isospectral if they have the same spectra. We are led to ask: if two Riemannian
manifolds are Laplace isospectral, are they isometric?
A general algebraic method for constructing Laplace isospectral manifolds is due
to Sunada [Sun85] (see the surveys by Gordon [Gor2000, Gor2009]), arising from
almost conjugate subgroups of a finite group of isometries acting on a manifold—
providing a negative answer to Kac’s question by Gordon–Webb–Wolpert [GWW92].
But preceding Sunada, in 1980 Vignéras [Vig80b] constructed such examples—indeed,
one of the applications in her book on quaternion algebras [Vig80a] was to explain
this construction!
Theorem 31.8.1 (Vignéras). For every 𝑚 ≥ 2, there exist Laplace isospectral and
nonisometric manifolds of dimension 𝑚.
We sketch a proof of this theorem in this section, with attention to the particular
detail of selectivity. In 1994, Maclachlan–Rosenberger [MacRos94] claimed to have
produced a pair of Laplace isospectral, nonisometric hyperbolic 2-orbifolds of genus
0, but then Buser–Flach–Semmler [BFS2008] later showed that these examples were
too good to be true! The subtle issue they found: the phenomenon of selectivity.
Our construction is quaternionic, of course. We consider the situation of sections
38.2–38.3, specifically the setup in 38.2.1 and 38.3.1. Let 𝐹 be a number field with 𝑟
real places and 𝑐 complex places, so that [𝐹 : Q] = 𝑟 + 2𝑐 = 𝑛. Let 𝐵 be a quaternion
algebra over 𝐹 and suppose that 𝐵 is split at 𝑡 real places. We have an embedding
(38.2.2) 𝜄 : 𝐵 ↩→ M2 (R) 𝑡 × M2 (C) 𝑐 . Letting H := (H2 ) 𝑡 × (H3 ) 𝑐 as in (38.2.9) and
P𝐵×>0 := 𝐵×>0 /𝐹 × ,
we see that 𝜄(P𝐵×>0 ) ≤ Isom+ (H ) acts on H by linear fractional transformations as

orientation-preserving isometries.
Let 𝑅 = Z𝐹 be the ring of integers of 𝐹 and let O ⊂ 𝐵 be an 𝑅-order. Then
the group Γ1 (O) := 𝜄(O1 /{±1}) ≤ P𝐵×>0 is a discrete subgroup acting properly on
H by isometries. Suppose now that 𝐵 is a division algebra, and further that the only
elements of O1 of finite order are ±1. (We will soon see that this assumption can be
satisfied; in some sense, it is quite typical.) Let 𝑋 1 (O) := Γ1 (O)\H be the quotient,
a Riemannian manifold of dimension 𝑚 := 2𝑡 + 3𝑐.
The condition of Laplace isospectrality is implied by the following arithmetic
condition in terms of conjugacy classes of embeddings.
Theorem 31.8.2. Let O0 ∈ Gen O. Suppose that for every quadratic field 𝐾 ⊇ 𝐹 and
every quadratic 𝑅-order 𝑆 ⊆ 𝐾, we have the equality of embedding numbers
𝑚(𝑆, O; O1 ) = 𝑚(𝑆, O0; O01 ). (31.8.3)
Then 𝑋 1 (O) and 𝑋 1 (O0) are Laplace isospectral.

Proof. See Vignéras [Vig80b, Théorème 6]. The statement there is in terms of all
embeddings, not just optimal embeddings—but the total count of conjugacy classes
of embeddings of an quadratic order is a sum of the corresponding count of optimal
embeddings of superorders (as in 31.7), so it is sufficient to have a genial order.
The key ingredient in the proof is the Selberg trace formula, which allows us
to show that the spectra of the Laplace operators agree by the stronger condition of
representation equivalence: for a more general point of view on this deduction, see
Deturck–Gordon [DG89].
The rub is in the equality (31.8.3). The restriction of the equivalence classes to
units of reduced norm 1 is harmless: by Lemma 30.3.14, we have 𝑚(𝑆, O; O1 ) =
𝑚(𝑆, O; O× ) [nrd(O× ) : nrd(𝑆 × )], and as a consequence of strong approximation we
have [nrd(O× ) : nrd(𝑆 × )] = [nrd( Ob× : nrd( 𝑆b× )]. Thus if O0 ∈ Typ O, then (31.8.3)
holds if and only if 𝑚(𝑆, O; O ) = 𝑚(𝑆, O0; O0× ).
×
Finally, selectivity enters! We suppose that O is an Eichler order. By Main

Theorem 31.1.7(c), we have the desired equality when Gen O is genial (i.e., not
optimally selective)—and by part (b), this equality may fail for some 𝑆.
Corollary 31.8.4. Suppose that Gen O is genial and O0 ∈ Gen O has O0 ; O. Suppose
further that 𝜎(Ram(𝐵)) ≠ Ram(𝐵) for all 𝜎 ∈ Aut(𝐹). Then 𝑋 1 (O) and 𝑋 1 (O0) are
Laplace isospectral, nonisometric Riemannian manifolds of dimension 𝑚.
Proof. Isospectrality follows from Theorem 31.8.2 with Main Theorem 31.1.7. To
show that 𝑋, 𝑋 0 are not isometric, since O0 ; O we know that PO0 is not conjugate
to PO in 𝐵× , but we need this for the groups Γ1 (O0), Γ1 (O) in Isom+ (H ), which is
slightly larger (see Remark 38.2.11). We leave the details to Exercise 31.3 (or see
Linowitz–Voight [LV2015, Proposition 2.24]).
The remainder of the proof of Theorem 31.8.1 involves finding suitable data
𝐹, 𝐵, O. We exhibit a pair for 𝑚 = 2 following Linowitz–Voight [LV2015, Example
5.2], giving a pair of compact hyperbolic 2-manifolds of genus 6 which are isospectral
but not isometric.
Example 31.8.5. Let 𝐹 = Q(𝑤) where 𝑤 is a root of the polynomial 𝑥 4 −5𝑥 2 −2𝑥+1 = 0.
Then 𝐹 is a totally real quartic field with discriminant 𝑑 𝐹 = 5744 = 24 359, Galois
group 𝑆4 , class number # Cl 𝑅 = 1 and narrow class number # Cl+ 𝑅 = 2. Let 𝐵 be
the quaternion algebra over 𝐹 which is ramified at the prime ideal 𝔭13 of norm 13
generated by 𝑏 := 𝑤 3 − 𝑤 2 − 4𝑤 and three of the four real places, with split place
𝑎, 𝑏
𝑤 ↦→ −0.751024 . . .: then 𝐵 = where 𝑎 = 𝑤 3 − 𝑤 2 − 3𝑤 − 1 is a root of
𝐹
𝑥 4 + 8𝑥 3 + 12𝑥 2 − 1.
A maximal order O ⊂ 𝐵 is given by
(𝑤 3 + 1) + 𝑤 2𝑖 + 𝑗 (𝑤 + 1) + (𝑤 3 + 1)𝑖 + 𝑖 𝑗
O = 𝑅 ⊕ 𝑅𝑖 ⊕ 𝑅⊕ 𝑅;
2 2
O has type number 2, so there exists two isomorphism classes of maximal orders
O1 = O and O2 .
31.8. ∗ ISOSPECTRAL, NONISOMETRIC MANIFOLDS 573
We claim that O1 and O2 have no elements of finite order other than ±1. Indeed,
if we had such an element of order 𝑞 then 𝐹 (𝜁2𝑞 ) is a cyclotomic quadratic extension
of 𝐹, whence Q(𝜁2𝑞 ) + ⊆ 𝐹; but 𝐹 is primitive, so the only cyclotomic quadratic
√ √
extensions
√ of 𝐹 √are 𝐾 = 𝐹 ( −1) and 𝐾 = 𝐹 ( −3). But as 𝔭13 splits completely in
𝐹 ( −1) and 𝐹 ( −3), neither field embeds into 𝐵. We conclude that the groups Γ𝑖
are torsion free.
Since 𝐵 is ramified at a finite place, the genus of O is genial by Theorem 31.2.1,
and since Aut(𝐹) is trivial, the hypothesis of Corollary 31.8.4 are satisfied: 𝑋11 , 𝑋21
are Laplace isospectral, but not isometric.
Finally, by Theorem 39.1.13, we have area(𝑋𝑖 ) = 20𝜋, so 𝑔(𝑋𝑖 ) = 6 for 𝑖 = 1, 2.
Fundamental domains for these are given in Figure 31.8.6.
Figure 31.8.6: Fundamental domain for the genus 6 manifolds 𝑋 (Γ11 ),𝑋 (Γ12 )
We obtain a second example by choosing the split real place 𝑤 ↦→ −1.9202 . . ., and
since 𝐹 is not Galois, as in the case of the 2-orbifold pairs 2 and 3, these are pairwise
nonisometric.
For an example with 𝑚 = 3, see Exercise 31.4.
1 1
Exercises
I 1. Prove Main Theorem 31.1.7 in the case 𝐾 ' 𝐹 × 𝐹: to be precise, show that
an 𝑅-order 𝑆 ⊆ 𝐹 × 𝐹 embeds equally in all Eichler 𝑅-orders. [Hint: we must
have 𝐵 ' M2 (𝐹), so reduce to the case where 𝑆 is embedded in the diagonal
and then conjugate.]
2. The following exercise gives insight into the proof of Theorem 31.7.5 on selec-
tivity. Let 𝑅 be local, and let O be an Eichler order of level 𝔭𝑒 .
Let 𝜙 : 𝑆 ↩→ O be an optimal embedding that is normalized and associated to
𝑥 ∈ 𝑅, so represented by

𝑥 1
𝛼=
− 𝑓 𝛾 (𝑥) 𝑡 − 𝑥
as in Definition 30.6.8.
(a) Compute 𝜈 −1 𝛼𝜈 for the matrix

𝑎 𝑏
𝜈= .
𝑐 𝑑
(b) Show that the off-diagonal entries are equal to
(𝑎𝑑 − 𝑏𝑐) −1 Nm𝐾 |𝐹 (𝑏𝛾 − (𝑏𝑥 + 𝑑)),

(𝑎𝑑 − 𝑏𝑐) −1 Nm𝐾 |𝐹 (𝑎𝛾 − (𝑎𝑥 + 𝑐))
so belong to (det 𝜈) −1 Nm𝐾 |𝐹 (𝐾 × ).

(c) Suppose 𝛼 0 = 𝜈 −1 𝛼𝜈 gives a normalized, optimal embedding. Show that
if 𝔭 is inert in 𝐾 ⊇ 𝐹, then det 𝜈 has even valuation.
I 3. With notation as in section 31.8, let Γ, Γ0 ≤ 𝐵× /𝐹 × and 𝜄 : 𝐵×>0 /𝐹 × ↩→
Isom+ (H ). Then 𝑋 (Γ), 𝑋 (Γ0) are isometric if and only if there exists a permuta-
tion 𝜎 of the factors of 𝐺 := Isom+ (H ) and 𝜈 ∈ 𝜄(𝐵× ) such that 𝜎(Γ) = 𝜈Γ0 𝜈 −1
and a Q-algebra automorphism 𝜏 : 𝐵 → 𝐵 such that the diagram
𝐵×
𝜏 / 𝐵×
𝜄 𝜄

𝐺
𝜎 /𝐺 𝜈 /𝐺
commutes, where 𝜈 acts on 𝐺 by conjugation. [Hint: use the Skolem–Noether

theorem.]
4. In this exercise, we exhibit Laplace isospectral, nonisometric hyperbolic 3-
manifolds, following Linowitz–Voight [LV2015, Example 6.3]. Let 𝐹 = Q(𝑤)
be the quintic field withdiscriminant −43535 and defining polynomial 𝑥 5 − 𝑥 4 +
3

3𝑤 − 2, −13
3𝑥 3 − 3𝑥 + 1. Let 𝐵 := .
𝐹
(a) Show that 𝐵 is ramified at the three real places of 𝐹 and the prime ideal
𝔭 = (𝑤 4 − 𝑤 3 + 3𝑤 2 − 𝑤 − 2) of norm 13.
(b) Let O be a maximal order in 𝐵. Show that the type number of O is equal
to 2.
(c) Let O = O1 and O2 be representatives of Typ O. Show that the only
elements of finite order in O1𝑖 are ±1.
(d) Show that Gen O is genial.
(e) Let 𝑋𝑖1 := Γ1 (O𝑖 )\H3 , with notation as in section 31.8. Show that 𝑋11 and
𝑋21 are Laplace isospectral, nonisometric 3-manifolds.
(f) Show that vol(𝑋11 ) = vol(𝑋21 ) = 51.024566 . . .
Part IV
Geometry and topology
575
Chapter 32
Unit groups
Having moved from algebra and arithmetic to analysis, and in particular the study of
class numbers, in this part we consider geometric aspects of quaternion algebras, and
the unit group of a quaternion order acting by isometries on a homogeneous space.
32.1 ⊲ Quaternion unit groups

By way of analogy, we consider what happens for quadratic orders. In this case, just as
with class groups, the behavior of unit groups is quite different depending on whether
the asociated quadratic field 𝐾 is real or imaginary.
In the imaginary case, the unit group is finite, as the norm equation Nm𝐾 |Q (𝛾) = 1
has only finitely many solutions for integral 𝛾: these are elements of a 2-dimensional
lattice in C that lie on the unit circle. Such an element is a root of unity that satisfies
a quadratic equation over Q, and so only two imaginary quadratic orders having units
other than ±1 are the Gaussian
√ order Z[𝑖] of discriminant −4 and the Eisenstein order
Z[𝜌] with 𝜌 := (−1 + −3)/2 of discriminant −3: see Figure 32.1.1.
i ρ −ρ2
−1 1 −1 1 −1 1
−i ρ2 −ρ
Figure 32.1.1: Units in imaginary quadratic orders
Orders O in a definite quaternion algebra 𝐵 over Q behave like orders in an

imaginary quadratic field. The unit group of such an order is finite, as the solutions
to nrd(𝛼) = 1 with 𝛼 ∈ O are elements of a 4-dimensional lattice in R4 again with
bounded size. In section 11.5, after a close investigation of the case of Hurwitz units,
577
578 CHAPTER 32. UNIT GROUPS
we classified the possibilities, embedding O× /{±1} ↩→ H1 /{±1} ' SO(3) as a finite

rotation group: the groups O× that arise over Q are either cyclic of order 2, 4, 6,
quaternion 𝑄 8 of order 8, binary dihedral 2𝐷 6 of order 12, or the binary tetrahedral
group 2𝑇 of order 24. In this chapter, we take up this task in the context of a general
definite quaternion order, and realize all finite rotation groups using quaternions.
Now we turn to real quadratic fields √and correspondingly indefinite quaternion
algebras. For the real quadratic√order Z[ 𝑑] with 𝑑 > 0, the units are solutions to
the Pell equation Nm𝐾 |Q (𝑥 − 𝑦 𝑑) = 𝑥 2 − 𝑑𝑦 2 = ±1 with 𝑥, 𝑦 ∈ Z. All solutions
up to sign are given by powers of a fundamental solution √ × which can be computed
explicitly using continued
√ fractions; consequently, Z[ 𝑑] = h−1, 𝑢i ' Z/2Z × Z
where 𝑢 = 𝑥 + 𝑦 𝑑 is the fundamental unit. The fundamental unit often (but not
always) has large height (in the sense that 𝑥, 𝑦 ∈ Z are large in absolute value), being
of exponential size in the discriminant,
√ by theorems of Schur and Siegel. The unit
group of the ring of integers of Q( 𝑑) for 𝑑 ≡ 1 (mod 4) is treated in a similar way,
by considering the norm equation 𝑥 2 − 𝑥𝑦 + 𝑐𝑦 2 = ±1 where 𝑐 = (1 − 𝑑)/4.
For quaternions, we are led to consider units in the standard order

𝑎, 𝑏
O := Z ⊕ Z𝑖 ⊕ Z 𝑗 ⊕ Z𝑖 𝑗 ⊆ 𝐵 :=
Q
in an indefinite quaternion algebra, with 𝑎, 𝑏 ∈ Z and say 𝑎 > 0. The norm condition
nrd(𝛾) = ±1 for 𝛾 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 then reads
𝑡 2 − 𝑎𝑥 2 − 𝑏𝑦 2 + 𝑎𝑏𝑧2 = ±1 (32.1.2)
with 𝑡, 𝑥, 𝑦, 𝑧 ∈√Z. Amusingly, this “quaternion Pell equation” includes the Pell
equation for Z[ 𝑎] by setting 𝑦 = 𝑧 = 0, and in fact by considering embeddings of
quadratic orders (the subject of Chapter 30), we see that this equation combines all Pell
equations satisfying certain congruence conditions. Combining these Pell equations,
we see that the group of solutions is an infinite, noncommutative group. (The case of
an order different from the standard one will give a different norm equation, but the
same conclusions.) See Jahangiri [Jah2010] for a Diophantine interpretation of the
structure of the unit group of a quaternion order as a quaternionic Pell equation.
We will seek to understand the group O× by its action on a suitable space, and in
this way we are led to consider groups acting discretely on symmetric spaces; we will
discover that the group O× is finitely presented and in particular finitely generated,
so we still can think of a set of fundamental solutions (given by generators) whose
products generate all solutions to (32.1.2). For example, we may take O = M2 (Z) ⊆
1 1 0 1
M2 (Q), where O× = GL2 (Z), generated by the elementary matrices , .
0 1 1 0
Accordingly, our investigation is detailed but fruitful, involving the theory of Fuchsian
and Kleinian groups.
In this chapter, we begin by discussing the general structure of these groups.
32.2 Structure of units

Throughout this chapter, we use the following notation, recalling our notation for
global fields from section 14.4. Let 𝐹 be a global field, let S ⊆ Pl(𝐹) be an eligible set
32.2. STRUCTURE OF UNITS 579
of places of 𝐹, and let 𝑅 = 𝑅S be the global ring associated to S, the ring of S-integers
of 𝐹. (As always, the reader may keep the case 𝐹 = Q, S = {∞}, and 𝑅 = Z in mind.)
Further, let 𝐵 be a quaternion algebra over 𝐹, and let O ⊂ 𝐵 be an 𝑅-order.
We are interested in the structure of the group O× . Since 𝑍 (𝐵× ) = 𝐹 × , we have
𝑅 ≤ 𝑍 (O× ) central. We understand the structure of 𝑅 × by Dirichlet’s unit theorem,
×
as follows.
32.2.1. From Dirichlet’s unit theorem (and its extension to S-units and the function
field case), the group 𝑅 × of units is a finitely generated abelian group of rank #S − 1,
so that
𝑅 × ' Z/𝑤Z ⊕ Z#S−1 (32.2.2)
where 𝑤 is the number of roots of unity in 𝐹. (The proof is briefly recalled in 32.3.1.)
The group O× is (in general) noncommutative, so we should not expect a description
like 32.2.1. But to get started, we consider the quotient O× /𝑅 × , and the reduced norm
map which gets us back into 𝑅 × .
32.2.3. We recall the theorem on norms (see section 14.7): as before, let
Ω := {𝑣 ∈ Ram 𝐵 : 𝑣 real} ⊆ Pl 𝐹 (32.2.4)
be the set of real ramified places in 𝐵 (recalling that complex places cannot be ramified),
and
𝐹>×Ω 0 := {𝑥 ∈ 𝐹 × : 𝑣(𝑥) > 0 for all 𝑣 ∈ Ω} (32.2.5)
the set of elements that are positive at the places 𝑣 ∈ Ω. If 𝐹 is a function field, then
Ω = ∅, and 𝑅>× Ω 0 = 𝑅 × . The Hasse–Schilling norm theorem (Main Theorem 14.7.4)
says that nrd(𝐵× ) = 𝐹>×Ω 0 . Letting
𝑅>× Ω 0 := 𝑅 × ∩ 𝐹>×Ω 0 , (32.2.6)

we conclude that nrd(O× ) ≤ 𝑅>× Ω 0 . (Strictly speaking, we only needed the containment
nrd(𝐵× ) ≤ 𝐹>×Ω 0 which follows directly from local considerations; but we pursue finer
questions below.)
32.2.7. In light of 32.2.3, the reduced norm gives an exact sequence
nrd
1 → O1 → O× −−→ 𝑅>× Ω 0 (32.2.8)
where
O1 := {𝛾 ∈ O× : nrd(𝛾) = 1} (32.2.9)
is the subgroup of units of reduced norm 1.
Since nrd(𝑅 × ) = 𝑅 ×2 by the squaring map, we have O1 ∩ 𝑅 × = {±1}, so (32.2.8)
yields
×
O1 O× nrd 𝑅>Ω 0
1→ → × −−→ ×2 (32.2.10)
{±1} 𝑅 𝑅
Since the group 𝑅 × is finitely generated, the group 𝑅>× Ω 0 /𝑅 ×2 is a finite, elementary
abelian 2-group. (In general, the reduced norm (the last) map in (32.2.10) need not be
surjective.)
32.2.11. In general, the exact sequence (32.2.10) does not split, so the group O× /𝑅 ×
will be a nontrivial extension of O1 /{±1} by an elementary abelian 2-group.
Example 32.2.12. If 𝐵 = M2 (𝐹) and O = M2 (𝑅), then O× = GL2 (𝑅) and O× /𝑅 × =

GL2 (𝑅)/𝑅 × =: PGL2 (𝑅). The reduced norm is the determinant, which is surjective,
and so the exact sequence (32.2.10) can be extended to
det
1 → PSL2 (𝑅) → PGL2 (𝑅) −−→ 𝑅 × /𝑅 ×2 → 1.
32.2.13. In light of Example 32.2.12, it is natural to write PO× := O× /𝑅 × and

PO1 := O1 /{±1}.
Remark 32.2.14. Some authors write GL1 (O) = O× and SL1 (O) = O1 , and this
notation suggests generalizations. In such situations, it is natural to write PGL1 (O) =
O× /𝑅 × and PSL1 (O) = O1 /{±1}.
32.2.15. Suppose 𝐹 is a number field. Then the group 𝑅>× Ω 0 /𝑅 ×2 is canonically

isomorphic to a quotient of class groups, as follows. Let Z𝐹 be the ring of integers
of 𝐹, and let ClΩ Z𝐹 denotes the class group of 𝐹 given by the quotient of the group
of fractional ideals by the subgroup of principal ideals with a generator positive at all
places in Ω. Then 𝑅>× Ω 0 /𝑅 ×2 is isomorphic to the quotient of ClΩ Z𝐹 by Cl Z𝐹 and the
group generated by the finite primes in S.
32.3 Units in definite quaternion orders

In this section, we show that for a definite quaternion order O, the scalar units 𝑅 × are
of finite index in O× —i.e., the group PO× = O× /𝑅 × is a finite group.
32.3.1. To build intuition, suppose 𝐹 is a number field with 𝑟 real places and 𝑐 complex
places. Recall the proof of Dirichlet’s unit theorem: we define a map
𝑅 × → RS
(32.3.2)
𝑥 ↦→ (𝑚 𝑣 log|𝑥| 𝑣 ) 𝑣 .
The kernel of this map is the group of roots of unity (the torsion subgroup of 𝑅 × ).
Í
The image lies inside the trace zero hyperplane 𝑣 ∈S 𝑥 𝑣 = 0 by the product formula
(14.4.6), and it is discrete and cocompact inside this hyperplane, so it is isomorphic
to ZS−1 . In particular, 𝑅 × is finite if and only if #S = 1; since S always contains
the set of archimedean places of size 𝑟 + 𝑐, we see that 𝑅 × is finite if and only if
(𝑟, 𝑐) = (1, 0), (0, 1), so 𝐹 = Q or 𝐹 is an imaginary quadratic field.
Remark 32.3.3. Informally, one might say that 𝑅 × is finite only when the completions
at the places in S provide “no room” for the unit group to become infinite. This is
analogous to the informal case for strong approximation in 28.5.4: if there is a place
𝑣 ∈ S where 𝐵1𝑣 is not compact, then there is enough room for 𝐵1 to “spread out” and
become dense.
32.3. UNITS IN DEFINITE QUATERNION ORDERS 581
32.3.4. Recall (Definition 28.5.1) that 𝐵 is S-definite if S ⊆ Ram(𝐵), i.e., every place
in S is ramified in 𝐵. In particular, if 𝐹 is a number field, then since a complex place
is split and S contains the archimedean places, if 𝐵 is S-definite then 𝐹 is totally real;
in this case, when S is exactly the set of archimedean places, we simply say that 𝐵 is
definite.
32.3.5. Consider the setup in analogy with Dirichlet’s unit theorem 32.3.1. We
consider the embedding of 𝐵 into the completions at all places in S:
Ö
𝐵 ↩→ 𝐵S := 𝐵𝑣 .
𝑣 ∈S
Î
By Exercise 27.14, 𝑅 is discrete in 𝐹S = 𝑣 ∈S 𝐹𝑣 and O is discrete in 𝐵6S (the point
being that in the number field case, S contains all archimedean places). Consequently,
the injections Ö
O× /𝑅 × ↩→ (𝐵S ) × /(𝐹S ) × := 𝐵×𝑣 /𝐹𝑣×
𝑣 ∈𝑆
Ö (32.3.6)
1 1
O ↩→ (𝐵S ) = 𝐵1𝑣
𝑣 ∈𝑆
have discrete image.
Depending on whether the place 𝑣 is nonarchimedean (split or ramified) or
archimedean (split real, ramified real, or complex), we have a different target compo-
nent 𝐵×𝑣 /𝐹𝑣× or 𝐵1𝑣 . The major task of Part IV is to describe these possibilities in detail
and look at the associated symmetric spaces.
We begin with the simplest case, where the unit groups involved are finite.
Proposition 32.3.7. The group O× /𝑅 × is finite if and only if O1 is finite if and only if
𝐵 is S-definite.
Proof. By the exact sequence (32.2.10), the group O× /𝑅 × is finite if and only if the
group O1 is finite.
First, suppose that 𝐵 is S-definite. Then by definition, for each 𝑣 ∈ S, the
completion 𝐵 𝑣 is a division algebra over 𝐹𝑣 . But each 𝐵1𝑣 is compact, from the
topological discussion in section 13.5. Therefore in (32.3.6), the group O1 is a closed,
discrete subgroup of a compact group—hence finite.
Now suppose 𝐵 is not S-definite. Then there is a place 𝑣 0 ∈ S that is unramified; we
will correspondingly find an element of infinite order (like solutions to the quaternion
Pell equation coming from the original Pell’s equation (32.1.2)). We have 𝐵 𝑣0 '
M2 (𝐹𝑣0 ), so there exists 𝛼 ∈ 𝐵 be such that the reduced characteristic polynomial
splits in 𝐹𝑣0 ; we may suppose without loss of generality that 𝐾 = 𝐹 [𝛼] is a field. Let
𝑆 be the integral closure of 𝑅 in 𝐾 (not to be confused with the set S). Then by the
Dirichlet S-unit theorem (32.2.1), the rank of 𝑆 × /𝑅 × is at least 1: the set of places
𝑤 ∈ Pl(𝐾) such that 𝑤 lies above 𝑣 ∈ S contains at least one element from each 𝑣
and two above 𝑣 0 , because it is split. So there is an element 𝛾 ∈ 𝑆 × /𝑅 × of infinite
order. As 𝑅-lattices, the order 𝑆 ∩ O has finite 𝑅-index and hence finite index in 𝑆,
so 𝑆 × /(𝑆 ∩ O) × is a finite group, and therefore a sufficiently high power of 𝛾 lies in
(𝑆 ∩ O) × ⊆ O× , and O× is infinite.
Example 32.3.8. Let 𝐵 = (−1, −1 | Q) and let O be the Z-order generated by 𝑖, 𝑗, so

that S = {∞}. Then 𝐵 is S-definite, and O× = h𝑖, 𝑗i ' 𝑄 8 is the quaternion group of
order 8.
Now consider S = {2, ∞}; then 𝐵 is still S-definite. We find
O[1/2] × = h2, 𝑖, 𝑗, 1 + 𝑖i
(32.3.9)
O[1/2] × /Z[1/2] × = h2, 𝑖, 𝑗, 1 + 𝑖i/h−1, 2i ' 𝑄 8 o Z/2Z
(Exercise 32.2).
Finally, if we take S = {5, ∞}, then 𝐵 is no longer S-definite; and O[1/5] × contains
the element 2 + 𝑖 of norm 5 ∈ Z[1/5] × and infinite order.
32.3.10. Suppose 𝐵 is S-definite. Then by Proposition 32.3.7, the group O× /𝑅 × is
finite. Since we have an embedding
O× /𝑅 × ↩→ 𝐵× /𝐹 ×
it follows that a O× /𝑅 × is a finite subgroup of P𝐵× , so a classification of finite

subgroups of P𝐵× gives a list of possible definite unit groups; we make this our task
in the remainder of this chapter.
32.4 Finite subgroups of quaternion unit groups

We now embark on a classification of finite subgroups of P𝐵× = 𝐵× /𝐹 × and P𝐵1 =
𝐵1 /{±1}; this is akin to first getting acquainted with the roots of unity in a number
field. Suppose throughout the rest of this chapter that 𝐹 is a number field; we allow 𝐵
to be definite or indefinite.
We begin in this section with the classification of the possible groups up to iso-
morphism that goes back at least to Klein [Kle56, Chapter II]: the original book dates
back to 1884 and is undoubtedly one of the most influential books of 19th century
mathematics. See also the descriptions by Coxeter [Coxtr40] and Lamotke [Lamo86,
Chapters I–II] for a presentation of the regular solids, finite rotation groups, as well as
finite subgroups of SL2 (C).
Proposition 32.4.1. Let Γ < P𝐵× be a finite group. Then Γ is cyclic, dihedral, or an
exceptional group 𝐴4 , 𝑆4 , 𝐴5 .
We met these groups already in Proposition 11.5.2, and the proof is an extension
of this result.
Proof. Let 𝑣 be an archimedean place of 𝐹. Then the natural map 𝐵× → 𝐵×𝑣 /𝐹𝑣× has
kernel 𝐹𝑣× ∩ 𝐵× = 𝐹 × , so the group homomorphism 𝐵× /𝐹 × ↩→ 𝐵×𝑣 /𝐹𝑣× is injective.
First suppose that 𝑣 is a ramified (real) place, so 𝐵 𝑣 ' H and
𝐵×𝑣 /𝐹𝑣× ' H× /R× ' H1 /{±1}.
By Corollary 2.4.21, we have H1 /{±1} ' SO(3) so Γ is a finite rotation group: these
are classified in Proposition 11.5.2.
32.5. CYCLIC SUBGROUPS 583
In general, we seek to conjugate the group Γ in order to reduce to the case above.
We may prove the lemma after making a base extension of 𝐹, so we may suppose
that 𝑣 is complex, with 𝐵 𝑣 ' M2 (C). Then 𝐵×𝑣 /𝐹𝑣× ' PGL2 (C), and via the injection
𝐵× /𝐹 × ↩→ PGL2 (C) we obtain a finite subgroup Γ ⊆ PGL2 (C). The natural map
SL2 (C) → PGL2 (C) is surjective, as we may rescale every invertible matrix by a
square root of its determinant to have determinant 1, and its kernel is {±1}, giving an
isomorphism PSL2 (C) ' PGL2 (C). We then lift Γ under the projection SL2 (C) →
PSL2 (C) to a finite group (containing −1). We have
H1 ' SU(2) = { 𝐴 ∈ SL2 (C) : 𝐴∗ 𝐴 = 1} ↩→ SL2 (C) (32.4.2)
as in 2.4.2. If h , i denotes the canonical (Hermitian) inner product on C2 defined by

h𝑧, 𝑤i = 𝑧∗ 𝑤 (as column vectors), then SU(2) is precisely the group of matrices of
determinant 1 preserving h , i, i.e.,
SU(2) = { 𝐴 ∈ SL2 (C) : h𝐴𝑧, 𝐴𝑤i = h𝑧, 𝑤i for all 𝑧, 𝑤 ∈ C2 }
since h𝐴𝑧, 𝐴𝑤i = 𝑧∗ ( 𝐴∗ 𝐴)𝑤 = 𝑧 ∗ 𝑤 if and only if 𝐴∗ 𝐴 = 1. We now define a

Γ-invariant Hermitian inner product on C2 by averaging: for 𝑧, 𝑤 ∈ C2 , we define
1 ∑︁
h𝑧, 𝑤iΓ := h𝛾𝑧, 𝛾𝑤i.
#Γ 𝛾 ∈Γ
Choose an orthonormal basis for h , iΓ and let 𝑇 ∈ SL2 (C) be the change of basis matrix
relative to the standard basis. Then h𝑧, 𝑤iΓ = h𝑇 𝑧, 𝑇 𝑤i and therefore 𝑇Γ𝑇 −1 ⊂ SU(2).
The result now follows from (32.4.2) and the previous case.
32.5 Cyclic subgroups

In the next few sections, we discuss each of the possibilities in Proposition 32.4.1 in
turn, following Chinburg–Friedman [CF2000]. We begin with cyclic subgroups.
There are always many subgroups of P𝐵× of order 2: every nonscalar element
𝛼 ∈ 𝐵× with trace zero has 𝛼2 ∈ 𝐹 × .
Proposition 32.5.1. Let 𝑚 > 2 and let 𝜁 𝑚 ∈ 𝐹 al be a primitive 𝑚th root of unity. Then
P𝐵× contains a cyclic subgroup of order 𝑚 if and only if 𝜁 𝑚 + 𝜁 𝑚 −1 ∈ 𝐹 and 𝐹 (𝜁 )
𝑚
splits 𝐵. Such a cyclic subgroup is unique up to conjugation in P𝐵× .
Proof. First we prove (⇐). Suppose 𝜁 𝑚 + 𝜁 𝑚 −1 ∈ 𝐹 and 𝐹 (𝜁 ) splits 𝐵. If in fact

𝑚
1 0
𝜁 𝑚 ∈ 𝐹, then 𝐹 splits 𝐵, i.e. 𝐵 ' M2 (𝐹); then 𝛾 := has order 𝑚 in
0 𝜁𝑚
P𝐵× ' PGL2 (𝐹). Otherwise, since 𝜁 𝑚 + 𝜁 𝑚 −1 ∈ 𝐹, we have [𝐹 (𝜁 ) : 𝐹] = 2, with 𝜁
𝑚 𝑚
a root of the polynomial 𝑇 2 − (𝜁 𝑚 + 𝜁 𝑚 −1 )𝑇 + 1. By Lemma 5.4.7, this implies there is
an embedding 𝐹 (𝜁 𝑚 ) ↩→ 𝐵; let 𝜁 be the image of 𝜁 𝑚 under this embedding. If 𝜁 has

order 𝑑 in P𝐵× , then Q(𝜁 𝑚 + 𝜁 𝑚
−1 , 𝜁 𝑑 ) ⊆ 𝐹; since 𝜁 ∉ 𝐹, we must have 𝑑 = 𝑚 if 𝑚 is
𝑚 𝑚
odd or 𝑑 = 𝑚/2 if 𝑚 is even. Let 𝛾 = 1 + 𝜁. Then 𝛾 2 𝜁 −1 = 2 + 𝜁 + 𝜁 −1 ∈ 𝐹 × , so 𝛾 has
order 𝑚 in P𝐵× .
Now we prove (⇒). Suppose that 𝛾 ∈ 𝐵× has image in P𝐵× of order 𝑚 > 2, so
that 𝛾 𝑚 = 𝑎 ∈ 𝐹 × . We do calculations in the commutative 𝐹-algebra 𝐾 := 𝐹 [𝛾]. Let
𝜍 := 𝛾𝛾 −1 ∈ 𝐾 × . Then
𝜍 𝑚 = 𝛾 𝑚 (𝛾 𝑚 ) −1 = 𝑎𝑎 −1 = 1 (32.5.2)
so 𝜍 𝑚 = 1. If 𝜍 𝑑 = 1 for 𝑑 | 𝑚 then 𝛾 𝑑 = 𝜍 𝑑 𝛾 𝑑 = 𝛾 𝑑 so 𝛾 𝑑 ∈ 𝐹 × and thus 𝑑 = 𝑚;
thus 𝜍 has order 𝑚 in 𝐵× . Applying the standard involution again gives
𝛾 = 𝛾𝜍 = 𝜍𝛾 = 𝜍 𝜍𝛾; (32.5.3)
thus 𝜍 = 𝜍 −1 , so 𝜍 ∉ 𝐹 and trd(𝜍) = 𝜍 + 𝜍 −1 ∈ 𝐹. Taking an appropriate power to

match up the root of unity, we conclude 𝜁 𝑚 + 𝜁 𝑚 −1 ∈ 𝐹. Finally, either 𝐾 is a quadratic
field in 𝐵, in which case 𝐾 splits 𝐵 by Lemma 5.4.7, or 𝐾 is not a field and 𝐵 ' M2 (𝐹),
in which case 𝐹 already splits 𝐵.
We conclude with uniqueness. Continuing from the previous paragraph, we have
shown that 𝛾 + 𝛾 = (1 + 𝜍)𝛾 ∈ 𝐹 × , so 𝛾 and 1 + 𝜍 generate the same cyclic subgroup
of P𝐵× , where 𝜍 𝑚 = 1. If 𝐾 = 𝐹 (𝜍) is a field, then all embeddings 𝐹 (𝜁 𝑚 ) ↩→ 𝐵 are
conjugate in 𝐵× by the Skolem–Noether theorem (Corollary 7.1.5), and consequently
every two cyclic subgroups of order 𝑚 are conjugate. Otherwise, the reduced charac-
teristic polynomial of 𝜍 factors, so 𝐵 ' M2 (𝐹), and its roots (the eigenvalues of 𝜍)
belong to 𝐹. If the eigenvalues
are repeated, then up to conjugation in GL2 (𝐹), 𝜍 is
1 𝑏
a scalar multiple of with 𝑏 ∈ 𝐹, and therefore has infinite order, impossible.
0 1

1 0
Thus the roots are distinct, and 𝜍 is conjugate to a multiple of and so 𝜆 is a
0 𝜆
primitive 𝑚th root of unity and the cyclic subgroup is unique up to conjugation.
32.5.4. The splitting condition in Proposition 32.5.1 can alternatively be phrased in

local-to-global terms (Proposition 14.6.7): 𝐾 = 𝐹 (𝜁 𝑚 ) splits 𝐵 if and only if every
place 𝑣 ∈ Ram 𝐵 is not split in 𝐾. Since the field 𝐹 (𝜁 𝑚 ) is totally complex, every
archimedean place splits, and so when 𝐾 ≠ 𝐹 we have 𝐾 ↩→ 𝐵 if and only if no prime
𝔭 ∈ Ram 𝐵 splits in 𝐾.
32.5.5. The proof of Proposition 32.5.1 describes the cyclic subgroup explicitly, up to
conjugation (still with 𝑚 > 2):
(i) If 𝜁 𝑚 ∈ 𝐹, then 𝐵 ' M2 (𝐹) and every cyclic subgroup

PGL2 (𝐹) of order 𝑚
of
1 0
is conjugate to the subgroup generated by 𝛾𝑚 = ;
0 𝜁𝑚
(ii) Otherwise, 𝐾 = 𝐹 (𝜁 𝑚 ) is a quadratic extension of 𝐹 with 𝐾 ↩→ 𝐵, and every
subgroup of P𝐵× of order 𝑚 is conjugate to the subgroup generated by the image
of 𝛾𝑚 = 1 + 𝜁 𝑚 .
The 𝐹-algebra 𝐾𝑚 = 𝐹 [𝛾𝑚 ] is separable and uniquely determined up to isomorphism.
In contrast to Proposition 32.5.1, there are a great many cyclic subgroups of order
𝑚 = 2 in P𝐵× , described as follows.
32.6. ∗ DIHEDRAL SUBGROUPS 585
32.5.6. If 𝛾 ∈ P𝐵× has order 𝑚 = 2, then 2 × ×

√ 𝛾 = 𝑎 ∈ 𝐹 and 𝛾 ∉ 𝐹 . Therefore, either
×2
𝑎 ∉ 𝐹 , equivalently 𝐾 = 𝐹 [𝛾] ' 𝐹 ( 𝑎) is a field, and the embedding 𝐾 ↩→ 𝐵 is
unique up to conjugation in 𝐵× by the Skolem–Noether theorem;or 𝑎 ∈ 𝐹 ×2 , in which
1 0
case after rescaling 𝛾 2 = 1 so 𝐵 ' M2 (𝐹) and 𝛾 is conjugate to .
0 −1
The following corollary shows that we can often reduce to the case of an even order
subgroup.
Corollary 32.5.7. Let 𝑚 ≥ 1 be odd. Then P𝐵× contains a cyclic subgroup of order
𝑚 if and only if P𝐵× contains a cyclic subgroup of order 2𝑚.
Proof. If 𝑚 = 1, then we are all set by 32.5.6. If 𝑚 ≥ 3, then Proposition 32.5.1

applies, and we see that the hypotheses hold for 𝑚 if and only if they hold for 2𝑚, since
𝜁2𝑚 = −𝜁 𝑚 .
Corollary 32.5.8. P𝐵× contains a cyclic subgroup of order 2𝑚 if and only if P𝐵1
contains a cyclic subgroup of order 𝑚.
Proof. The corollary follows from 32.5.5: the subgroup of P𝐵× of order 2𝑚 generated
by 𝛾2𝑚 yields the subgroup of P𝐵1 of order 𝑚 generated by 𝜁2𝑚 , and vice versa.
32.6 ∗ Dihedral subgroups

We now turn to the dihedral case, where we show that every cyclic subgroup extends
(in general, in many ways) to a dihedral subgroup, continuing to follow Chinburg–
Friedman [CF2000, Lemma 2.3].
Lemma 32.6.1. Let 𝑚 ≥ 2. Then the following statements hold.
(a) Every cyclic subgroup of P𝐵× of order 𝑚 is contained in a dihedral subgroup of

order 2𝑚; in particular, P𝐵× contains a dihedral subgroup of order 2𝑚 if and
only if it contains a cyclic subgroup of order 𝑚.
(b) Let 𝛾 ∈ 𝐵× have order 𝑚 in P𝐵× , and let 𝐾 = 𝐹 [𝛾]. For 𝑗 ∈ 𝐵× , we have
𝐾, 𝑏
h𝛾, 𝑗i ⊆ P𝐵× dihedral if and only if 𝑗 2 = 𝑏 ∈ 𝐹 and 𝐵 = 𝐾 + 𝐾 𝑗 ' .
𝐹
Proof. First (a). The implication (⇒) is immediate, so we prove (⇐). Let 𝑚 ≥ 2
and suppose that P𝐵× contains a cyclic subgroup of order 𝑚, generated by the image
of 𝛾 ∈ 𝐵× , and let 𝐾 = 𝐹 [𝛾]. Let 𝑗 ∈ 𝐵 be orthogonal to 𝐾 under nrd. Then
𝑗 2 = 𝑏 ∈ 𝐹 × , and 𝑗 𝛼 = 𝛼 𝑗 = 𝛼−1 𝑗 ∈ P𝐵× for all 𝛼 ∈ 𝐾, so the subgroup h𝛾, 𝑗i is
𝐾, 𝑏
dihedral of order 2𝑚 in P𝐵× , and 𝐵 ' , as in Exercise 6.2.
𝐹
Now (b). We just showed (⇒) in the previous part, so we show (⇐). Let Γ ⊆ P𝐵×
be a dihedral subgroup of order 2𝑚, where Γ = h𝛾, 𝑗i has 𝛾 ∈ 𝐵× generating a cyclic
subgroup of order 𝑚 in P𝐵× and 𝑗 ∈ 𝐵× satisfies
𝑗 −1 𝛾 𝑗 = 𝛾 −1 ∈ 𝐵× /𝐹 × .

𝐾, 𝑏
Let 𝐾 = 𝐹 [𝛾]. We claim that 𝐵 = 𝐾 + 𝐾 𝑗 ' .
𝐹
First we show 𝑗 −1 𝛾 𝑗 = 𝛾. This follows from a direct argument using reduced
norm and trace (see Exercise 32.4), but we have also the following argument. Since
𝐾 = 𝐹 [𝛾] is semisimple (see 32.5.5) and conjugation by 𝑗 acts as an 𝐹-algebra
automorphism of 𝐾 = 𝐹 [𝛾], it is either the identity or the standard involution, and
thus 𝑗 −1 𝛾 𝑗 = 𝛾, 𝛾. But we cannot have 𝑗 −1 𝛾 𝑗 = 𝛾, because then 𝐾 [ 𝑗] ⊆ 𝐵 would be
a commutative subalgebra of dimension ≥ 3, a contradiction.
Now by (4.2.16), expanding the trace gives
trd( 𝑗 𝛾) = 𝑗 𝛾 + 𝛾 𝑗 = 𝑗 𝛾 + 𝛾(trd( 𝑗) − 𝑗)
(32.6.2)
= 𝑗 𝛾 − 𝛾 𝑗 + trd( 𝑗)𝛾 = trd( 𝑗)𝛾.
Since 1, 𝛾 are linearly independent we conclude trd( 𝑗) = trd( 𝑗 𝛾) = 0, i.e., 𝑗 is

𝐾, 𝑏
orthogonal to 𝐾 under nrd, so 𝑗 2 = 𝑏 ∈ 𝐹 × and 𝐵 = .
𝐹
The dihedral subgroups of order 2𝑚 for 𝑚 > 2 are classified as follows.
Lemma 32.6.3. Let 𝑚 > 2. Then the set of dihedral subgroups of order 2𝑚 up to
conjugation in P𝐵× are in bĳection with the group
×)
Nm𝐾𝑚 /𝐹 (𝐾𝑚
(32.6.4)
h𝛿i𝐹 ×2
−1 .
where 𝐾𝑚 is as in 32.5.5 and 𝛿 = 2 + 𝜁 𝑚 + 𝜁 𝑚
Proof. Let Γ = h𝛾, 𝑗i and Γ0 = h𝛾, 𝑗 0i be two dihedral subgroups as in Lemma

32.6.1(b) with 𝑗 2 = 𝑏 and ( 𝑗 0) 2 = 𝑏 0. Then 𝑗 0 ∈ 𝐾 ⊥ = 𝐾 𝑗 so 𝑗 0 = 𝛽 𝑗 with 𝛽 ∈ 𝐾 × ,
and nrd( 𝑗 0) = 𝑏 0 = nrd(𝛽)𝑏, so 𝑏𝑏 0 ∈ Nm𝐾 |𝐹 (𝐾 × ) (as in Exercise 6.4). Then we
claim that Γ, Γ0 are conjugate in P𝐵× if and only if
𝑏𝑏 0 ∈ 𝐹 ×2 h𝛿i.
If Γ0 = 𝛼−1 Γ𝛼 with 𝛼 ∈ 𝐵× , then conjugation by 𝛼 normalizes the unique cyclic

subgroup on both sides, so 𝛾 𝑟 𝑗 0 = 𝛼−1 𝑗 𝛼 for some 𝑟, and therefore
−𝛿𝑟 𝑏 0 = nrd(𝛾 𝑟 𝑗 0) = nrd( 𝑗) = −𝑏
as desired. Conversely, if 𝑏𝑏 0 ∈ 𝐹 ×2 h𝛿i then so too for nrd(𝛽) = 𝑏 0 𝑏 −1 ; rescaling 𝛽

by 𝐹 × and replacing 𝛽 by 𝛾 𝛽 if necessary, we may suppose nrd(𝛽) = Nm𝐾 |𝐹 (𝛽) = 1
(without changing Γ0); by Hilbert’s theorem 90, there exists 𝛼 ∈ 𝐾 × such that 𝛽 = 𝛼𝛼−1 ,
and conjugation by 𝛼 again normalizes the cyclic subgroup and satisfies
𝛼−1 𝑗 𝛼 = 𝛼𝛼−1 𝑗 = 𝛽 𝑗
as desired.
32.7. ∗ EXCEPTIONAL SUBGROUPS 587
Remark 32.6.5. One can rephrase Lemma 32.6.3 in terms of a global equivalence
relation, further encompassing the case 𝑚 = 2: see Chinburg–Friedman [CF2000,
Lemma 2.4].
1
Corollary 32.6.6.
group P𝐵 contains a dihedral group of order 2𝑚 > 4 if and
The
𝐾2𝑚 , −1
only if 𝐵 ' .
𝐹
Proof. We first prove (⇒). If P𝐵1 contains a dihedral group Γ = h𝛾, 𝑗i of order 2𝑚
then it contains a cyclic subgroup of order 𝑚 so by Corollary 32.5.8 the group P𝐵×
contains a cyclic subgroup of order 2𝑚, which we may take to be generated by 𝛾2𝑚
as in 32.5.5 with 𝐾2𝑚 = 𝐹 [𝛾2𝑚 ] = 𝐹 [𝛾]; by hypothesis we have 𝑗 2 = − nrd( 𝑗) = −1,
𝐾2𝑚 , −1
and so 𝐵 ' as in the classification in Lemma 32.6.1.
𝐹
Next we prove the converse implication (⇐). We refer to 32.5.5. In case (i) where
𝜁2𝑚 ∈ 𝐹, we have nrd(𝜁2𝑚 −1 𝛾 ) = 𝜁 2 𝜁 = 1 so we may take Γ = h𝜁 −1 𝛾 , 𝑗i; in case
𝑚 2𝑚 𝑚 2𝑚 𝑚
(ii), where 𝜁2𝑚 ∉ 𝐹, we take Γ = h𝜁2𝑚 , 𝑗i.
32.7 ∗ Exceptional subgroups

Finally we treat exceptional groups (cf. Gehring–Maclachlan–Martin–Reid [GMMR97,
p. 3635]). We found quaternionic
√ realizations of the exceptional groups in section 11.5
when 𝐵 ' (−1, −1 | 𝐹) (and 5 ∈ 𝐹 for 𝐴5 ).
(a) P𝐵× contains a subgroup isomorphic to 𝐴4 if and only if P𝐵1 contains a subgroup
isomorphic to 𝐴4 if and only if 𝐵 ' (−1, −1 | 𝐹).
(b) P𝐵× contains a subgroup isomorphic to 𝑆4 if and only if it contains a subgroup
isomorphic to 𝐴4 ; and√P𝐵1 contains a subgroup isomorphic to 𝑆4 if and only if
𝐵 ' (−1, −1 | 𝐹) and 2 ∈ 𝐹.
(c) P𝐵× contains a subgroup isomorphic to 𝐴5 if and only if√P𝐵1 contains a subgroup
isomorphic to 𝐴5 if and only if 𝐵 ' (−1, −1 | 𝐹) and 5 ∈ 𝐹.
Any two such exceptional subgroups of P𝐵× (or P𝐵1 ) are conjugate by an element of
𝐵× if and only if they are isomorphic as groups.
Proof. First we prove (a); let Γ ⊆ P𝐵× be a subgroup with Γ ' 𝐴4 . The reduced
norm gives a homomorphism Γ → nrd(Γ) ⊆ 𝐹 × /𝐹 ×2 , but 𝐴4 has no nontrivial
homomorphic image of exponent 2, so nrd(Γ) ⊆ 𝐹 ×2 . Therefore, there is a unique
lift of Γ to 𝐵1 /{±1}, and the map 𝐵1 /{±1} → 𝐵× /𝐹 × is an isomorphism from this
lift to 𝐻. This shows the first implication; its converse follows from the injection
P𝐵1 ↩→ P𝐵× . For the second implication, let 𝑖, 𝑗 ∈ 𝐵1 generate the 𝑉4 -subgroup (the
normal subgroup of index 3 isomorphic to the Klein 4 group) of 𝐴4 in P𝐵1 . Then
𝑖, 𝑗 ∉ 𝐹 × , and 𝑖 2 = − nrd(𝑖) = −1 = 𝑗 2 ; and similarly (𝑖 𝑗) 2 = −1 implies 𝑗𝑖 = −𝑖 𝑗. By
Lemma 2.2.5, we conclude 𝐵 ' (−1, −1 | 𝐹). The converse follows from the Hurwitz
unit group 11.2.4.
For part (b), the implication (⇒) is immediate; the implication (⇐) follows by
taking the Hurwitz units and adjoining the element 1 + 𝑖, as in 11.5.4 but working
modulo scalars. For the second statement, an element of order√4 in 𝐵1 /{±1} lifts to
an element of order 8 in 𝐵1 and therefore has reduced trace ± 2 ∈ 𝐹; the converse
follows again from the explicit construction in 11.5.4.
For part (c), we argue similarly. Since 𝐴5 is generated by its subgroups
√ isomorphic
to 𝐴4 , we may apply (a) to get a lift, and by the reduced trace we get 5 ∈ 𝐹.
The uniqueness statement is requested in Exercise 32.5.
Exercises
Unless otherwise indicated, let 𝐹 be a number field with ring of integers 𝑅, let 𝐵 be a
quaternion algebra over 𝐹 and let O ⊆ 𝐵 be an 𝑅-order.
1. Let 𝐹 be a totally√real field, let 𝐾 ⊇ 𝐹 be a totally imaginary quadratic extension

of 𝐹, so 𝐾 = 𝐹 ( 𝑑) with 𝑑 totally negative. Let 𝑆 be the ring of integers of 𝐾.
Consider the group homomorphism
𝜙 : 𝑆× → 𝑆×
𝑢 ↦→ 𝑢/𝑢
where is the nontrivial 𝐹-involution of 𝐾.

(a) Show that if 𝑢 ∈ 𝜇(𝑆)𝑅 × then 𝑢 ∈ ker 𝜙.
(b) Show that 𝜙(𝑢) is a root of unity for all 𝑢 ∈ 𝑆 × . [Hint: It is an algebraic
integer of absolute value 1 under all complex embeddings.]
(c) Let 𝜇(𝑆) be the subgroup of roots of unity of 𝑆 × , and let 𝜓 : 𝑆 × →
𝜇(𝑆)/𝜇(𝑆) 2 be the map induced by 𝜙. Show that if 𝑢 ∈ ker 𝜓, so 𝜙(𝑢) = 𝜁 2
with 𝜁 ∈ 𝜇(𝑆), then 𝜁 −1 𝑢 ∈ 𝑅 × . Conclude that ker 𝜓 = 𝜇(𝑆)𝑅 × .
(d) Show that [𝑆 × : 𝜇(𝑆)𝑅 × ] ≤ [𝑆 × : 𝑆 1 𝑅 × ] ≤ 2.
[The index [𝑆 × : 𝜇(𝑆)𝑅 × ] is known as the Hasse unit index.]

−1, −1
2. Let 𝐵 = and let O be the Z-order generated by 𝑖, 𝑗. Prove that
Q
O[1/2] × ' h2, 𝑖, 𝑗, 1 + 𝑖i and describe O[1/2] × /Z[1/2] × as an extension of 𝑄 8
by Z/2Z.
3. Show that
× : 𝑅 ×2 ]
[𝑅Ω
[O× : O1 𝑅 × ] = [nrd(O× ) : 𝑅 ×2 ] = × : nrd(O× )]
[𝑅Ω
where 𝑅>× Ω 0 is the subgroup of units positive at ramified infinite places of 𝐵,

defined in (32.2.6).
4. Let 𝐵 be a quaternion algebra over a field 𝐹 with char 𝐹 ≠ 2, and let 𝛾, 𝑗 ∈ 𝐵×
be such that 𝑗 −1 𝛾 𝑗 = 𝛾 −1 . Show by looking at the reduced norm and trace that
𝑗 −1 𝛾 𝑗 = 𝛾 (cf. Lemma 32.6.1).
32.7. ∗ EXCEPTIONAL SUBGROUPS 589
I 5. Prove the uniqueness statement in Proposition 32.7.1: Show that every two
isomorphic exceptional subgroups of P𝐵× are conjugate by an element of 𝐵× ,
and the same for 𝐵1 .
6. Let Γ ≤ O1 be a maximal finite subgroup. Combining results from sections
11.2 and 11.5 and this chapter, prove the following.
(a) Γ is isomorphic to one of the following groups:
• cyclic of order 2𝑚 with 𝑚 ≥ 1,
• binary dihedral (dicyclic) 2𝐷 2𝑚 of order 4𝑚 with 𝑚 ≥ 1,
• binary tetrahedral 2𝑇 of order 24,
• binary octahedral 2𝑂 of order 48, or
• binary icosahedral 2𝐼 of order 120.
In the latter three cases,√we call Γ exceptional. √
(b) If Γ ' 2𝑂 then 𝐹 ⊇ Q( 2) and if Γ ' 2𝐼 then 𝐹 ⊇ Q( 5).
−1, −1
(c) If Γ is exceptional, then 𝐵 ' .
𝐹
7. Continuing with the previous exercise, suppose that 𝐵 is totally definite, so 𝐹 is
totally real. Prove the following statements.
(a) If O1 does not contain an element of order 4, then O1 is √
cyclic.
(b) If O1 is quaternion 𝑄 8 ' 2𝐷 4 or exceptional, then 𝑅[ −1] ↩→ O and
discrd(O) is only divisible by primes
√ dividing 2.
(c) If O1 ' 2𝐷 2𝑚 with 𝑚 ≥ 3, then 𝑅[ −1], 𝑅[𝜁2𝑚 ] ↩→ O and discrd(O) is
−1 .
only divisible by primes dividing 𝜆22𝑚 − 4, where 𝜆2𝑚 := 𝜁2𝑚 + 𝜁2𝑚
8. Continuing further with the previous exercise, we compare O× and O1 .
(a) Show that [O× : O1 ] = 1, 2, 4. [Hint: Use Exercise 32.1.]
(b) Show that if O1 ' 2𝑂, 2𝐼, then O× = O1 .
(c) Show that if O1 ' 2𝑇, then [O× : O1 ] ≤ 2, and equality holds if and only
if there exists 𝛾 ∈ (1 + 𝑖)𝐹 × ∩ O× such that nrd(𝛾) ∉ 𝐹 ×2 .
[For a complete account covering all cases, see Vignéras-Guého [VG74].]
Chapter 33
Hyperbolic plane
In this chapter, we give background on the geometry of the hyperbolic plane.
33.1 ⊲ The beginnings of hyperbolic geometry

We have seen that the group of unit Hamiltonians H1 acts by rotations of Euclidean
space and therefore by isometries of the unit sphere, and that in spherical geometry the
discrete subgroups are beautifully realized as classical finite groups: cyclic, dihedral,
and the symmetry groups of the Platonic solids.
Replacing H with M2 (R), the group SL2 (R) of determinant 1 matrices possesses a
much richer supply of discrete subgroups. In fact, PSL2 (R) can be naturally identified
with a circle bundle over the hyperbolic plane, and so the structure of quaternionic
unit groups is naturally phrased in the language of hyperbolic geometry. Indeed, it
was work on automorphic functions and differential equations invariant under discrete
subgroups of PSL2 (R) that provided additional early original motivation to study
hyperbolic space: their study provides an incredibly rich interplay between number
theory, algebra, geometry, and topology, with quaternionic applications. This interplay
is the subject of the final parts of the text.
In this chapter, we provide a rapid introduction to the hyperbolic plane. Hyperbolic
geometry has its roots preceding the quaternions, in efforts during the early 1800s to
understand Euclid’s axioms for geometry. Since the time of Euclid, there had been
attempts to prove the quite puzzling parallel postulate (given a line and a point not
on the line, there is a unique line through the point parallel to the given line) from
the other four simpler, self-evident axioms for geometry. In hyperbolic geometry, the
parallel postulate fails to hold—there are always infinitely many distinct lines through
a point that do not intersect a given line—and so it is a non-Euclidean geometry.
The underpinnings of what became hyperbolic geometry can be found in work by
Euler and Gauss in their study of curved surfaces (the differential geometry of surfaces).
It was then Lobachevsky and Bolyai who suggested that curved surfaces of constant
negative curvature could be used in non-Euclidean geometry, and finally Riemann who
generalized this to what are now called Riemannian manifolds. Klein coined the term
“hyperbolic” for this geometry because its formulae can be obtained from spherical
591
592 CHAPTER 33. HYPERBOLIC PLANE
geometry by replacing trigonometric functions by their hyperbolic counterparts. See

[Sco83, §1] for a nice overview of the 2-dimensional geometries.
Hyperbolic geometry, and in particular the hyperbolic plane, remains an important
prototype for understanding negatively-curved spaces in general. Milnor [Milno82]
gives a comprehensive early history of hyperbolic geometry; see also the survey by
Cannon–Floyd–Kenyon–Parry [CFKP97], which includes an exposition of five models
for hyperbolic geometry. (It is also possible to work out hyperbolic geometry in a
manner akin to what Euclid did for his geometry without a particular model, following
Lobachevsky [LP2010].)
For further references on hyperbolic plane geometry, see Jones–Singerman [JS87,
Chapter 5], Anderson [And2005], Ford [For72], Katok [Kat92, Chapter 1], Iversen
[Ive92, Chapter III], and Beardon [Bea95, Chapter 7]. There are a wealth of geometric
results and formulas from Euclidean geometry that one can try to reformulate in the
world of hyperbolic plane geometry, and the interested reader is encouraged to pursue
these further.
33.2 Geodesic spaces

In geometry, we need notions of length, distance, and the straightness of a path. These
notions are defined for a certain kind of metric space, as follows.
Let 𝑋 be a metric space with distance 𝜌. An isometry 𝑔 : 𝑋 −→ ∼ 𝑋 is a bĳective
map that preserves distance, i.e., 𝜌(𝑥, 𝑦) = 𝜌(𝑔(𝑥), 𝑔(𝑦)) for all 𝑥, 𝑦 ∈ 𝑋. (Any
distance-preserving map is automatically injective and so becomes an isometry onto
its image.) The set of isometries Isom(𝑋) forms a group under composition.
33.2.1. A path from 𝑥 to 𝑦, denoted 𝜐 : 𝑥 → 𝑦, is a continuous map 𝜐 : [0, 1] → 𝑋

where 𝜐(0) = 𝑥 and 𝜐(1) = 𝑦. (More generally, we can take the domain to be any
compact real interval.) The length ℓ(𝜐) of a path 𝜐 is the supremum of sums of
distances between successive points over all finite subdivisions of the path (the path
is rectifiable if this supremum is finite). Conversely, if 𝑋 is a set with a notion of
(nonnegative) length of path, then one recovers a candidate (intrinsic) metric as
𝜌(𝑥, 𝑦) = inf ℓ(𝜐), (33.2.2)

𝜐:𝑥→𝑦
a metric when this infimum exists (i.e., there exists a path of finite length 𝑥 → 𝑦) for
all 𝑥, 𝑦 ∈ 𝑋. If the distance on 𝑋 is of the form (33.2.2), we call 𝑋 a length metric
space or a path metric space, and by construction ℓ(𝑔𝜐) = ℓ(𝜐) for all paths 𝜐 and
𝑔 ∈ Isom(𝑋).
Example 33.2.3. The space 𝑋 = R𝑛 with the ordinary Euclidean metric is a path
metric space; it is sometimes denoted E𝑛 as Euclidean space (to emphasize the role of
the metric).
33.2.4. If 𝑋 is a path metric space and 𝜐 achieves the infimum in (33.2.2), then we say
𝜐 is a geodesic segment in 𝑋. A geodesic is a continuous map (−∞, ∞) → 𝑋 such
that the restriction to every compact interval defines a geodesic segment. If 𝑋 is a path
33.2. GEODESIC SPACES 593
metric space such that every two points in 𝑋 are joined by a geodesic segment, we say
𝑋 is a geodesic space, and if this geodesic is unique we call 𝑋 a uniquely geodesic
space.
33.2.5. If 𝑋 is a geodesic space, then an isometry of 𝑋 maps geodesic segments

to geodesic segments, and hence geodesics to geodesics: i.e., if 𝑔 ∈ Isom(𝑋) and
𝜐 : 𝑥 → 𝑦 is a geodesic segment, then 𝑔𝜐 : 𝑔𝑥 → 𝑔𝑦 is a geodesic segment. After all,
𝑔 maps the set of paths 𝑥 → 𝑦 bĳectively to the set of paths 𝑔𝑥 → 𝑔𝑦, preserving
distance.
33.2.6. In the context of differential geometry (our primary concern), these notions
can be made concrete with coordinates. Suppose 𝑈 ⊆ R𝑛 is an open subset. Then
a convenient way to specify the length of a path in 𝑈 is with a length element in
real-valued coordinates. To illustrate, the ordinary metric on R𝑛 is given by the length
element √︃
d𝑠 := d𝑥 12 + · · · + d𝑥 𝑛2 ,
so if 𝜐 : [0, 1] → 𝑈 is a piecewise continuously differentiable function written as

𝜐(𝑡) = (𝑥1 (𝑡), . . . , 𝑥 𝑛 (𝑡)), then
√︄
∫ 1 2 2
d𝑥 1 d𝑥 𝑛
ℓ(𝜐) = +···+ d𝑡 (33.2.7)
0 d𝑡 d𝑡
as usual.
More generally, if 𝜆 : 𝑈 → R>0 is a positive continuous function, then the length
element 𝜆(𝑥) d𝑠 defines a metric (33.2.2) on 𝑈, as follows. The associated length
(33.2.7) is symmetric, nonnegative, and satisfies the triangle inequality. To show that
𝜌(𝑥, 𝑦) > 0 when 𝑥 ≠ 𝑦, by continuity 𝜆 is bounded below by some 𝜂 > 0 on a suitably
small 𝜖 ball neighborhood of 𝑥 not containing 𝑦, so every path 𝜐 : 𝑥 → 𝑦 has ℓ(𝜐) ≥ 𝜖𝜂
and 𝜌(𝑥, 𝑦) > 0.
In this context, we also have a notion of orientation, and we may restrict to
isometries that preserve this orientation. We return to this in section 33.8, rephrasing
this in terms of Riemannian geometry.
Remark 33.2.8. The more general study of geometry based on the notion of length in a
topological space (the very beginnings of which are presented here) is the area of metric
geometry. Metric geometry has seen significant recent applications in group theory and
dynamical systems, as well as many other areas of mathematics. For further reading,
see the texts by Burago–Burago–Ivanov [BBI2001] and Papadopoulous [Pap2014].
In particular, geodesic spaces are quite common in mathematics, including com-
plete Riemannian manifolds; Busemann devotes an entire book to the geometry of
geodesics [Bus55]. Uniquely geodesic spaces are less common; examples include
simply connected Riemannian manifolds without conjugate points, CAT(0) spaces,
and Busemann convex spaces.
The following theorem nearly characterizes geodesic spaces.
Theorem 33.2.9 (Hopf–Rinow). Let 𝑋 be a complete and locally compact length

metric space. Then 𝑋 is a geodesic space and every bounded closed set in 𝑋 is
compact.
Proof. See e.g. Bridson–Haefliger [BH99, Proposition 3.7]).
33.3 Upper half-plane

We now present the first model of two-dimensional hyperbolic space (see Figure
33.3.2).
Definition 33.3.1. The upper half-plane is the set
H2 := {𝑧 = 𝑥 + 𝑖𝑦 ∈ C : Im(𝑧) = 𝑦 > 0}.
H = H2
Figure 33.3.2: Upper half-plane H2
Definition 33.3.3. The hyperbolic length element on H2 is defined by

√︁
d𝑥 2 + d𝑦 2 |d𝑧|
d𝑠 := = ; (33.3.4)
𝑦 Im 𝑧
As described in 33.2.6, the hyperbolic length element induces a metric on H2 , and

this provides it with the structure of a path metric space.
Definition 33.3.5. The set H2 equipped with the hyperbolic metric is (a model for)
the hyperbolic plane.
Remark 33.3.6. The space H2 can be intrinsically characterized as the unique two-
dimensional (connected and) simply connected Riemannian manifold with constant
sectional curvature −1.
The hyperbolic metric and the Euclidean metric on H2 are equivalent, inducing the
same topology (Exercise 33.1). However, lengths and geodesics are different under
these two metrics, as we will soon see.
33.3.7. The group GL2 (R) acts on C via linear fractional transformations:

𝑎𝑧 + 𝑏 𝑎 𝑏
𝑔𝑧 = , for 𝑔 = ∈ SL2 (R) and 𝑧 ∈ C;
𝑐𝑧 + 𝑑 𝑐 𝑑
33.3. UPPER HALF-PLANE 595
since
(𝑎𝑧 + 𝑏) (𝑐𝑧 + 𝑑) 𝑎𝑐|𝑧| 2 + 𝑎𝑑𝑧 + 𝑏𝑐𝑧 + 𝑏𝑑
𝑔𝑧 = = (33.3.8)
|𝑐𝑧 + 𝑑| 2 |𝑐𝑧 + 𝑑| 2
we have
det 𝑔
Im 𝑔𝑧 = Im 𝑧. (33.3.9)
|𝑐𝑧 + 𝑑| 2
and so if Im 𝑧 > 0, then Im 𝑔𝑧 > 0 if and only if det 𝑔 > 0. Therefore, the subgroup
GL+2 (R) = {𝑔 ∈ GL2 (R) : det(𝑔) > 0}
preserves the upper half-plane H2 . Moreover, because the action of GL+2 (R) is holo-
morphic, it is orientation-preserving.
The kernel of this action, those matrices acting by the identity as linear fractional
transformations, are the scalar matrices, since (𝑎𝑧 + 𝑏)/(𝑐𝑧 + 𝑑) = 𝑧 identically if
and only if 𝑐 = 𝑏 = 0 and 𝑎 = 𝑑. Taking the quotient we get a faithful action of
PGL+2 (R) = GL+2 (R)/R× on H2 . There is a canonical isomorphism
∼ PSL (R)
PGL+ (R) −→
2 2
1
𝑔 ↦→ √︁ 𝑔.
det(𝑔)
with the same action on the upper half-plane.
33.3.10. The determinant det : PGL2 (R) → R× /R×2 ' {±1} has the inverse image
of +1 equal to PGL+2 (R) both open and closed in PGL2 (R); therefore, any 𝑔 with
det(𝑔) < 0 together with PGL+2 (R) generates PGL2 (R): for example, we may take

−1 0
𝑔= (33.3.11)
0 1
In view of (33.3.9), we extend the action of PGL2 (R) on H2 by defining for 𝑔 ∈
PGL2 (R) and 𝑧 ∈ H2
(
𝑔𝑧, if det 𝑔 > 0;
𝑔·𝑧= (33.3.12)
𝑔𝑧 = 𝑔𝑧, if det 𝑔 < 0.
The elements 𝑔 ∈ PGL2 (R) with det 𝑔 < 0 act anti-holomorphically and so are
orientation-reversing. The matrix 𝑔 in (33.3.11) then acts by 𝑔(𝑧) = −𝑧.
This action also arises naturally from another point of view. Let
H2− = {𝑧 ∈ C : Im 𝑧 < 0}
be the lower half-plane, let H2+ = H2 , and let
H2± = H2+ ∪ H2− = {𝑧 ∈ C : Im 𝑧 ≠ 0} = C − R.
Then PGL2 (R) acts on H2± (it preserves R hence also its complement in C). Complex
conjugation 𝑧 ↦→ 𝑧 interchanges H2+ and H2− , and there is a canonical identification
∼ H2
H2± /h i −→
from which we obtain the action (33.3.12) of PGL2 (R) on H2 .
Remark 33.3.13. The fact that PSL2 (R) has elements 𝑔 ∈ PSL2 (R) that are matrices
up to sign means that whenever we do a computation with a choice of matrix, implicitly
we are also checking that the computation goes through with the other choice of sign.
Most of the time, this is harmless—but in certain situations this sign plays an important
role!
Let Isom+ (H2 ) ≤ Isom(H2 ) be the subgroup of isometries of H2 that preserve
orientation.
Theorem 33.3.14. The group PSL2 (R) acts on H2 via orientation-preserving isome-
tries, i.e., PSL2 (R) ↩→ Isom+ (H2 ).
Proof. Because the metric is defined by a length element d𝑠, we want to show that
d(𝑔𝑠) = d𝑠 for all 𝑔 ∈ PSL2 (R), i.e.,
|d(𝑔𝑧)| |d𝑧|
=
Im(𝑔𝑧) Im 𝑧
for all 𝑔 ∈ PSL2 (R). Since |d(𝑔𝑧)| = |d𝑔(𝑧)/d𝑧||d𝑧|, it is equivalent to show that
|d𝑔(𝑧)/d𝑧| 1
= (33.3.15)
Im 𝑔𝑧 Im 𝑧
for all 𝑔 ∈ PSL2 (R).

Let 𝑔 ∈ PSL2 (R) act by
𝑎𝑧 + 𝑏
𝑔(𝑧) =
𝑐𝑧 + 𝑑
with 𝑎𝑑 − 𝑏𝑐 = 1. Then

d𝑔 (𝑐𝑧 + 𝑑)𝑎 − (𝑎𝑧 + 𝑏)𝑐 1
(𝑧) = = ; (33.3.16)
d𝑧 (𝑐𝑧 + 𝑑) 2 |𝑐𝑧 + 𝑑| 2
by (33.3.9),
Im 𝑧
Im 𝑔𝑧 = ,
|𝑐𝑧 + 𝑑| 2
so taking the ratio, the two factors |𝑐𝑧 + 𝑑| 2 exactly cancel, establishing (33.3.15).
Finally, the action is holomorphic so (by the Cauchy–Riemann equations) it lands
in the orientation-preserving subgroup.
33.3.17. The action of PSL2 (R) extends to the boundary as follows. We define the
circle at infinity to be the boundary
bd H2 := R ∪ {∞} ⊆ C ∪ {∞}.
(The name comes from viewing H2 in stereographic projection as a half-sphere with

circular boundary.) The group PSL2 (R) still acts on bd H2 by linear fractional trans-
formations. We define the completed upper half-plane to be
H2∗ := H2 ∪ bd H2 .
33.4. CLASSIFICATION OF ISOMETRIES 597
The topology on H2∗ is the same as the Euclidean topology on H2 , and we take a
fundamental system of neighborhoods of the point ∞ to be sets of the form
{𝑧 ∈ H2 : Im 𝑧 > 𝑀 } ∪ {∞}
for 𝑀 > 0 and a system of neighborhoods of the point 𝑧0 to be
{𝑧0 } ∪ {| 𝑧 − (Re 𝑧0 + 𝑚𝑖)| < 𝑚}
i.e. open disks tangent to the real axis at 𝑧0 , together with 𝑧 0 .

Remark 33.3.18. Although the hyperbolic plane cannot be embedded in R3 , it can be
locally embedded. One way to visualize plane hyperbolic geometry (locally) is by
the pseudosphere, the surface of revolution generated by a tractrix: it is a surface
with constant negative curvature and so locally models the hyperbolic plane (Figure
33.3.19).
Figure 33.3.19: Pseudosphere
Remark 33.3.20. We will compactify quotients of H2 in other ways below. In that

context, we will add only a subset of the boundary of H2∗ ; this overloading should
cause no confusion.
33.4 Classification of isometries

On our way to classifying isometries, we pause to identify three natural subgroups of
SL2 (R):
cos 𝑡 sin 𝑡
𝐾 = SO(2) = : 𝑡 ∈ R ' R/2𝜋Z
− sin 𝑡 cos 𝑡

𝑎 0
𝐴= : 𝑎 ∈ R×>0 ' R×>0 ' R (33.4.1)
0 1/𝑎

1 𝑏
𝑁= : 𝑏 ∈ R ' R.
0 1
𝑎𝑖 + 𝑏
We have 𝐾 = StabSL2 (R) (𝑖) since = 𝑖 if and only if 𝑑 = 𝑎 and 𝑐 = −𝑏, and
𝑐𝑖 + 𝑑
𝑎 0
then the determinant condition implies 𝑎 2 + 𝑏 2 = 1. An element acts by
0 1/𝑎
𝑧 ↦→ 𝑎2 𝑧, fixing the origin and stretching along lines through the origin. An element
1 𝑛
acts by the translation 𝑧 ↦→ 𝑧 + 𝑏.
0 1
Proposition 33.4.2 (Iwasawa decomposition). The multiplication map gives a home-

omorphism
∼ SL (R).
𝑁 × 𝐴 × 𝐾 −→ 2
In particular, for all 𝑔 ∈ SL2 (R), we can write uniquely 𝑔 = 𝑛𝑔 𝑎 𝑔 𝑘 𝑔 with 𝑛𝑔 ∈ 𝑁,

𝑎 𝑔 ∈ 𝐴, and 𝑘 𝑔 ∈ 𝐾 in a way continuously varying in 𝑔.
Proof. The multiplication map 𝑁 × 𝐴 × 𝐾 → SL2 (R) is continuous and open, so we

need to show it is bĳective. It is injective, because checking directly we see that
𝑁 𝐴 ∩ 𝐾 = {1} = 𝑁 ∩ 𝐴.
This map is surjective as follows. Let 𝑔 ∈ SL2 (R), and let 𝑧 = 𝑔(𝑖). Let 𝑛𝑔 =
√
1 − Re 𝑧 1/ 𝑦 0
∈ 𝑁, so that (𝑛𝑔 𝑔) (𝑖) = 𝑦𝑖. Let 𝑎 𝑔 = √ ∈ 𝐴; then
0 1 0 𝑦
(𝑎 𝑔 𝑛𝑔 𝑔) (𝑖) = 𝑖, so 𝑎 𝑔 𝑛𝑔 𝑔 ∈ StabSL2 (R) (𝑖) = 𝐾, and peeling back we get 𝑔 ∈ 𝑁 𝐴𝐾,
proving surjectivity.
Remark 33.4.3. We have 𝐴𝑁 = 𝑁 𝐴, and we showed in the proof of Proposition 33.4.2

that 𝑁 𝐴 acts transitively on H2 (by 𝑧 ↦→ 𝑎 2 𝑧 + 𝑏). In section 34.6, we reinterpret this
as providing a direct link between H2 and SL2 (R).
Lemma 33.4.4. The group SL2 (R) is generated by the subgroups 𝐴, 𝑁, and the
0 −1
element , which acts on H2 by 𝑧 ↦→ −1/𝑧.
1 0

𝑎 𝑏
Proof. Let ∈ SL2 (R). The lemma follows by performing row reduction on
𝑐 𝑑
the matrix using the given generators. We find that if 𝑐 ≠ 0, then

𝑎 𝑏 1 𝑎/𝑐 0 −1 𝑐 0 1 𝑑/𝑐
=
𝑐 𝑑 0 1 1 0 0 1/𝑐 0 1
and if 𝑐 = 0 then
𝑎 𝑏 𝑎 0 1 𝑏/𝑎
= .
0 1/𝑎 0 1/𝑎 0 1
The subgroups 𝑁, 𝐴, 𝐾 can be characterized by their traces; with a view to working

on PSL2 (R), we consider the absolute traces:
|Tr(𝐾)| = [0, 2], |Tr( 𝐴)| = [2, ∞), and |Tr(𝑁)| = {2}.
33.4. CLASSIFICATION OF ISOMETRIES 599
Definition 33.4.5. An element 𝑔 ∈ PSL2 (R) with 𝑔 ≠ ±1 is called elliptic, hyperbolic,

or parabolic according to whether |Tr(𝑔)| < 2, |Tr(𝑔)| > 2, or |Tr(𝑔)| = 2.
Every nonidentity element 𝑔 ∈ PSL2 (R) belongs to one of these three types (even
though 𝑔 need not belong to one of the subgroups 𝑁, 𝐴, 𝐾 individually).
Lemma 33.4.6. An element 𝑔 ∈ PSL2 (R) is


 elliptic 




 a unique fixed point in H2 ,

 
 

hyperbolic if and only if 𝑔 has two fixed points on bd H2 ,
 a unique fixed point on bd H2 .
  
 parabolic 
  
  

𝑎 𝑏
Proof. Let 𝑔 = have det(𝑔) = 𝑎𝑑 − 𝑏𝑐 = 1. We look to solve the equation
𝑐 𝑑
𝑎𝑧 + 𝑏
=𝑧
𝑐𝑧 + 𝑑
or equivalently 𝑐𝑧 2 + (𝑑 −𝑎)𝑧−𝑏 = 0. The discriminant is (𝑑 −𝑎) 2 +4𝑏𝑐 = (𝑎+𝑑) 2 −4 =

Tr(𝑔) 2 − 4. Therefore 𝑔 is elliptic if and only if this discriminant is negative if and
only if there is a unique root in H2 ; 𝑔 is parabolic if and only if this discriminant is
zero if and only if there is a unique root in bd H2 ; and 𝑔 is hyperbolic if and only if
this discriminant is positive if and only if there are two roots in bd H2 .
33.4.7. Let 𝑔 ∈ PSL2 (R). If 𝑔 is elliptic, then 𝑔 acts by hyperbolic rotation in

a neighborhood around its fixed point, as in Figure 33.4.8; every such element is
conjugate to an element of 𝐾, fixing 𝑖. (Indeed, in the unit disc model with its fixed
point as the center, an elliptic element acts literally by rotation in the disc; see section
33.7.)
elliptic
Figure 33.4.8: Action on H2 by an elliptic element
A hyperbolic element can be thought of as a translation along the geodesic between

the two fixed points on bd H2 , as in Figure 33.4.9.

hyperbolic
Figure 33.4.9: Action on H2 by a hyperbolic element

Moving these fixed points to 0, ∞, every such element is conjugate to an element of 𝐴,
acting by 𝑧 ↦→ 𝑎 2 𝑧 with 𝑎 ≠ 1, as in Figure 33.4.10.
Figure 33.4.10: Action on H2 by a hyperbolic element with fixed points 0, 1

Finally, a parabolic element should be thought of as a limit of the other two types,
where correspondingly the fixed point tends to the boundary or the two fixed points
move together; every such element is conjugate to an element of 𝑁, acting by translation
𝑧 ↦→ 𝑧 + 𝑛 for some 𝑛 ∈ R.
Lemma 33.4.11. For all 𝑧, 𝑧 0 ∈ H2 , there exists 𝑔 ∈ PSL2 (R) such that 𝑔𝑧, 𝑔𝑧 0 ∈ R>0𝑖
are pure imaginary.
Proof. The proof follows directly by using translations and scaling (33.4.1), requested
in Exercise 33.5.
33.5 Geodesics
In this section, we prove two important theorems: we describe geodesics, giving a
formula for the distance, and we classify isometries.
Theorem 33.5.1. The hyperbolic plane H2 is a uniquely geodesic space. The unique
geodesic passing through two distinct points 𝑧, 𝑧 0 ∈ H2 is a semicircle orthogonal to
R or a vertical line, and
| 𝑧 − 𝑧0| + | 𝑧 − 𝑧0|
𝜌(𝑧, 𝑧 0) = log (33.5.2)
| 𝑧 − 𝑧0| − | 𝑧 − 𝑧0|
| 𝑧 − 𝑧 0 |2
cosh 𝜌(𝑧, 𝑧 0) = 1 + . (33.5.3)
2 Im(𝑧) Im(𝑧 0)
33.5. GEODESICS 601
Proof. We first prove the imaginary axis is a geodesic with 𝑧, 𝑧 0 ∈ R>0𝑖. Let 𝜐(𝑡) =
𝑥(𝑡) + 𝑖𝑦(𝑡) : 𝑧 → 𝑧 0 be a path; then
∫ 1 √︁ ∫ 1
(d𝑥/d𝑡) 2 + (d𝑦/d𝑡) 2 d𝑦/d𝑡
ℓ(𝜐) = d𝑡 ≥ d𝑡
0 𝑦(𝑡) 0 𝑦(𝑡)
0 (33.5.4)
𝑧
= log 𝑦(1) − log 𝑦(0) = log

𝑧
with equality if and only if 𝑥(𝑡) = 0 identically and d𝑦/d𝑡 ≥ 0 for all 𝑡 ∈ [0, 1]. This
is achieved for the path
𝜐(𝑡) = (|𝑧|(1 − 𝑡) + |𝑧 0 |𝑡)𝑖;
so 𝜌(𝑧, 𝑧 0) = log | 𝑧 0/𝑧|, and the imaginary axis is the unique geodesic.
For arbitrary points 𝑧, 𝑧 0 ∈ H2 , we apply Lemma 33.4.11. The statement on
geodesics follows from the fact that the image of R>0𝑖 under an element of PSL2 (R)
is either a semicircle orthogonal to R or a vertical line (Exercise 33.6).
The formula (33.5.2) for the case 𝑧, 𝑧 0 ∈ R>0𝑖 follows from (33.5.4) and plugging
in along the imaginary axis; the general case then follows from the invariance of both
𝜌(𝑧, 𝑧 0) and
| 𝑧 − 𝑧0| + | 𝑧 − 𝑧0|
log
| 𝑧 − 𝑧0| − | 𝑧 − 𝑧0|
under 𝑔 ∈ PSL2 (R), checked on the generators in Lemma 33.4.4 (Exercise 33.10).
Finally, the formula (33.5.3) follows directly from formulas for hyperbolic cosine,
requested in Exercise 33.11.
Theorem 33.5.5. We have
Isom(H2 ) ' PGL2 (R)
and
Isom+ (H2 ) ' PGL+2 (R) ' PSL2 (R).
Proof. Let 𝑍 = {𝑡𝑖 : 𝑡 > 0} be the positive part of the imaginary axis. By Theorem
33.5.1, 𝑍 is the unique geodesic through any two points of 𝑍.
Let 𝜙 ∈ Isom(H2 ). Then 𝜙(𝑍) is a geodesic (33.2.5), so by Exercise 33.7, there
exists 𝑔 ∈ PSL2 (R) such that 𝑔𝜙 fixes 𝑍 pointwise. Replacing 𝜙 by 𝑔𝜙, we may
suppose 𝜙 fixes 𝑍 pointwise.
Let 𝑧 = 𝑥 + 𝑖𝑦 ∈ H2 and 𝑧 0 = 𝑥 0 + 𝑖𝑦 0 = 𝜙(𝑧). For all 𝑡 > 0,
𝜌(𝑧, 𝑖𝑡) = 𝜌(𝜙𝑧, 𝜙(𝑖𝑡)) = 𝜌(𝑧 0, 𝑖𝑡).
Plugging this into the formula (33.5.3) for the distance, we obtain
(𝑥 2 + (𝑦 − 𝑡) 2 )𝑦 0 = (𝑥 02 + (𝑦 0 − 𝑡) 2 )𝑦.
Dividing both sides by 𝑡 2 and taking the limit as 𝑡 → ∞, we find that 𝑦 = 𝑦 0, and
consequently that 𝑥 2 = 𝑥 02 and 𝑥 = ±𝑥 0. The choice of sign ± varies continuously over
the connected set H2 and so must be constant. Therefore 𝜙(𝑧) = 𝑧 or 𝜙(𝑧) = −𝑧 for all
𝑧 ∈ H2 . The latter generates PGL2 (R) over PSL2 (R) (33.3.10), and both statements
in the theorem follow.
33.6 Hyperbolic area and the Gauss–Bonnet formula

In this section, we consider hyperbolic area. We measure hyperbolic area by consider-
ing a small Euclidean rectangle whose sides are parallel to the axes at the point (𝑥, 𝑦);
the hyperbolic length of the sides are d𝑥/𝑦 and d𝑦/𝑦, and we obtain the hyperbolic
area form from the product.
Definition 33.6.1. We define the hyperbolic area form by
d𝑥 d𝑦
d𝐴 = .
𝑦2
For a subset 𝑆 ⊆ H2 , we define the hyperbolic area of 𝑆 by

∫∫
𝜇(𝑆) = d𝐴
𝑆
when this integral is defined.
Proposition 33.6.2. The hyperbolic area is invariant under Isom(H2 ).
Proof. We verify that the hyperbolic area form is invariant. We first check this for the
orientation-reversing isometry
𝑔(𝑧) = 𝑔(𝑥 + 𝑖𝑦) = −𝑥 + 𝑖𝑦 = −𝑧;
visibly d(𝑔 𝐴) = d𝐴 in this case.

It suffices then to consider 𝑔 ∈ PSL2 (R). Let 𝑧 = 𝑥 + 𝑖𝑦 and let
𝑎𝑧 + 𝑏
𝑤 = 𝑔(𝑧) = = 𝑢 + 𝑖𝑣,
𝑐𝑧 + 𝑑
𝑦
with 𝑎𝑑 − 𝑏𝑐 = 1. By (33.3.9), 𝑣 = . We compute that
|𝑐𝑧 + 𝑑| 2
d𝑔 1
(𝑧) = . (33.6.3)
d𝑧 (𝑐𝑧 + 𝑑) 2
Now 𝑔 is holomorphic; so by the Cauchy–Riemann equations, its Jacobian is given by
𝜕 (𝑢, 𝑣) d𝑔 2

= (𝑧) = 1
;
𝜕 (𝑥, 𝑦) d𝑧 |𝑐𝑧 + 𝑑| 4
therefore
d𝑢 d𝑣 𝜕 (𝑢, 𝑣) d𝑥 d𝑦 1 |𝑐𝑧 + 𝑑| 4
d(𝑔 𝐴) = = = d𝑥 d𝑦 = d𝐴.
𝑣2 𝜕 (𝑥, 𝑦) 𝑣 2 |𝑐𝑧 + 𝑑| 4 𝑦2
A major role will be played in what follows by hyperbolic polygons, defined

formally as follows.
33.6. HYPERBOLIC AREA AND THE GAUSS–BONNET FORMULA 603
33.6.4. Let 𝑧, 𝑧 0 ∈ H2∗ be distinct. Then there is a unique geodesic in H2 whose

closure in H2∗ passes through 𝑧, 𝑧 0, and the segment from 𝑧 to 𝑧 0 is denoted [𝑧, 𝑧 0].
Definition 33.6.5. A hyperbolic polygon is a connected, closed subset of H2∗ whose
boundary consists of finitely many geodesic sides of the form [𝑧, 𝑧 0] with 𝑧, 𝑧 0 vertices.
A hyperbolic triangle is a hyperbolic polygon with three sides.
Definition 33.6.6. A subset 𝐴 ⊆ H2 is convex if the geodesic segment between any
two points in 𝐴 lies inside 𝐴.
An example of a (convex) hyperbolic polygon is given in Figure 33.6.7.
Figure 33.6.7: A hyperbolic polygon
A hyperbolic triangle is visibly convex; for more on convexity, see Exercises 33.8–
33.9. The following key formula expresses the hyperbolic area of a triangle in terms
of its angles.
Theorem 33.6.8 (Gauss–Bonnet formula). Let 𝑇 be a hyperbolic triangle with angles
𝛼, 𝛽, 𝛾. Then
𝜇(𝑇) = 𝜋 − (𝛼 + 𝛽 + 𝛾).
Proof. See e.g. Katok [Kat92, Theorem 1.4.2].
33.6.9. Let 𝑃 be a convex hyperbolic polygon with 𝑛 sides. By convexity, each

side meets at each vertex a unique side, so 𝑃 has 𝑛 vertices with angles 𝜃 1 , . . . , 𝜃 𝑛 .
Triangulating 𝑃 and applying the Gauss–Bonnet formula, we conclude that
𝜇(𝑃) = (𝑛 − 2)𝜋 − (𝜃 1 + · · · + 𝜃 𝑛 ).
Remark 33.6.10. Theorem 33.6.8 is called the Gauss–Bonnet formula because it is

closely related to the more general formula relating curvature to Euler characteristic.
The simplest formula of this kind is
∫
𝐾 d𝐴 = 2𝜋 𝜒(𝑋) (33.6.11)
𝑋
for a compact Riemann surface 𝑋. The expression (33.6.11) is quite remarkable: it

says that the total curvature of 𝑋 is determined by its topology; if you flatten out a
surface in one place, the curvature is forced to rise somewhere else. If instead one has
a surface 𝑋 with geodesic boundary, then the formula (33.6.11) becomes
∫ ∑︁
𝐾 d𝐴 + (𝜋 − 𝜃 𝑖 ) = 2𝜋 𝜒(𝑋)
𝑋 𝑖
∫ vertices. For a triangle 𝑋 with constant curvature −1

where 𝜃 𝑖 are the angles at the
and angles 𝛼, 𝛽, 𝛾, we have 𝑋 𝐾 d𝐴 = −𝜇(𝑋) and 𝜒(𝑋) = 𝑉 − 𝐸 + 𝐹 = 1 (as for any
polygon), so we find
−𝜇(𝑋) + 3𝜋 − (𝛼 + 𝛽 + 𝛾) = 2𝜋
and we recover Theorem 33.6.8.
33.7 Unit disc and Lorentz models

In this section, we consider two other models for the hyperbolic plane.
First, we consider the unit disc model.
Definition 33.7.1. The hyperbolic unit disc is the (open) unit disc
D2 = {𝑤 ∈ C : |𝑤| < 1}
equipped with the hyperbolic metric
2|d𝑤|
d𝑠 = .
1 − |𝑤| 2
The hyperbolic unit disc D2 is also called the Poincaré model of planar hyperbolic
geometry. The circle at infinity is the boundary
bd D2 = {𝑤 ∈ C : |𝑤| = 1}.
33.7.2. For all 𝑧 0 ∈ H2 , the maps

∼ D2
𝜙 𝑧0 : H2 −→ 𝜙−1 2 ∼
→ H2
𝑧0 : D −
𝑧 − 𝑧0 𝑧0 𝑤 − 𝑧0 (33.7.3)
𝑧 →
↦ 𝑤= 𝑤 ↦→ 𝑧 =
𝑧 − 𝑧0 𝑤−1
define a conformal equivalence between H2 and D2 with 𝑧0 ↦→ 𝜙(𝑧0 ) = 0. A

particularly nice choice is 𝑧0 = 𝑖, giving
𝑧−𝑖 𝑤+1
𝜙(𝑧) = , 𝜙−1 (𝑤) = −𝑖 . (33.7.4)
𝑧+𝑖 𝑤−1
The hyperbolic metric on D2 is the pushforward of (induced from) the hyperbolic

metric on H2 via the identification (33.7.4) (Exercise 33.12). Ordinarily, one would
decorate the pushforward metric, but because we will frequently move between the
33.7. UNIT DISC AND LORENTZ MODELS 605
upper half-plane and unit disc as each has its advantage, we find it notationally simpler
to avoid this extra decoration. The distance on D2 is
|1 − 𝑤𝑤 0 | + |𝑤 − 𝑤 0 |
𝜌(𝑤, 𝑤 0) = log
|1 − 𝑤𝑤 0 | − |𝑤 − 𝑤 0 |
(33.7.5)
|𝑤 − 𝑤 0 | 2
cosh 𝜌(𝑤, 𝑤 0) = 1 + 2
(1 − |𝑤| 2 ) (1 − |𝑤 0 | 2 )
so that
1 + |𝑤|
𝜌(𝑤, 0) = log = 2 tanh−1 |𝑤|. (33.7.6)
1 − |𝑤|
The map 𝜙 (33.7.3) maps the geodesics in H2 to geodesics in D2 , and as a Möbius
transformation, maps circles and lines to circles and lines, preserves angles, and maps
the real axis to the unit circle; therefore the geodesics in D2 are diameters through the
origin and semicircles orthogonal to the unit circle, as in Figure 33.7.7.
Figure 33.7.7: Hyperbolic geodesics

Accordingly, triangles from the upper half-plane map to triangles in the unit disc.
33.7.8. Via the map 𝜙, the group PSL2 (R) acts on D2 as

𝑎 𝑏
𝜙 PSL2 (R)𝜙−1 = PSU(1, 1) = ∈ M2 (C) : |𝑎| 2 − | 𝑏| 2 = 1 /{±1}.
𝑏 𝑎
Explicitly, an isometry of D2 is a map of the form

𝑤−𝑎
𝑤 ↦→ 𝑒 𝑖 𝜃
1 − 𝑎𝑤
for 𝑎 ∈ C with |𝑎| < 1 and 𝜃 ∈ R. (A direct substitution can be used to give an alternate
verification that these transformations are isometries of D2 with the hyperbolic metric.)
The orientation reversing isometry 𝑔(𝑧) = −𝑧 on H2 acts by 𝑔(𝑤) = 𝑤 on D2 with
the choice 𝑝 = 𝑖 (Exercise 33.13).
The induced area on D2 is given by
4 d𝑥 d𝑦
d𝐴 =
(1 − 𝑥 2 − 𝑦 2 ) 2
for 𝑤 = 𝑥 + 𝑦𝑖.
Second, we present the Lorentz model.
Definition 33.7.9. The Lorentz metric on R3 is the indefinite metric

d𝑠2 = −d𝑡 2 + d𝑥 2 + d𝑦 2 .
33.7.10. The indefinite Lorentz metric is associated to the quadratic form
𝑞(𝑡, 𝑥, 𝑦) = −𝑡 2 + 𝑥 2 + 𝑦 2
in the natural way. Lengths in this metric can be positive or nonpositive. However, on
the hyperboloid
𝑡 2 − 𝑥 2 − 𝑦 2 = 1,
the metric becomes positive definite: a nonzero tangent vector to the hyperboloid has
positive length (Exercise 33.16). The hyperboloid can be thought of as the sphere
of radius 𝑖 about the origin with respect to 𝑞; taking an imaginary radius shows that
hyperbolic geometry is dual in some sense to spherical geometry, where S2 ⊆ R3 has
real radius 1.
Definition 33.7.11. The Lorentz hyperboloid is the set

L2 = {(𝑡, 𝑥, 𝑦) ∈ R3 : 𝑞(𝑡, 𝑥, 𝑦) = −1, 𝑡 > 0}
equipped with the Lorentz metric.
The Lorentz hyperboloid is the upper sheet of the (two-sheeted) hyperboloid; it
is also called the hyperboloid model or the Lorentz model of planar hyperbolic
geometry, and it can be visualized as in Figure 33.7.13. (The choice of signs has to do
with the physics of spacetime.)
The map
L2 → D2
(33.7.12)
(𝑡, 𝑥, 𝑦) ↦→ (𝑥 + 𝑖𝑦)/(𝑡 + 1)
is bĳective and identifies the metrics on L2 and D2 (Exercise 33.15). Moreover, the
map (33.7.12) maps geodesics in D2 to intersections of the hyperboloid with planes
through the origin.
t2 − x2 − y 2 = 1
L2
cone
t2 = x2 + y 2
Figure 33.7.13: The hyperboloid model L2

33.8. RIEMANNIAN GEOMETRY 607
By pullback, Isom+ (L2 ) ' PSL2 (R). However, other isometries are also apparent:
a linear change of variables that preserves the quadratic form 𝑞 also preserves the
Lorentz metric. Let
−1 0 0
O(2, 1) = {𝑔 ∈ GL3 (R) : 𝑔 𝑡 𝑚𝑔 = 𝑚}, where 𝑚 = 0 1 0®.
© ª
«0 0 1¬
Then
d𝑠2 = 𝑣 𝑡 𝑚𝑣, where 𝑣 = (d𝑡, d𝑥, d𝑦) 𝑡 ,
so if 𝑔 ∈ O(2, 1) then immediately d𝑔𝑠2 = d𝑠2 .
Next, if 𝑔 ∈ O(2, 1), then 𝑔 𝑡 𝑚𝑔 = 𝑚 implies det(𝑔) = ±1. The elements of O(2, 1)
that map the hyperboloid to itself comprise the subgroup
SO(2, 1) = {𝑔 ∈ O(2, 1) : det(𝑔) = 1};
let SO+ (2, 1) ≤ SO(2, 1) be the further subgroup that maps the upper sheet of the
hyperboloid to itself, the connected component of the identity.
Remark 33.7.14. We have proven that there is an isomorphism of Lie groups
SO+ (2, 1) ' PSL2 (R); (33.7.15)
it corresponds to the isomorphism of Lie algebras 𝔰𝔬2,1 ' 𝔰𝔩2 .
33.8 Riemannian geometry

The hyperbolic metric (33.3.4) is induced from a Riemannian metric as follows.
33.8.1. A Riemannian metric d𝑠2 on an open set 𝑈 ⊆ R𝑛 is a function that assigns

to each point 𝑥 ∈ 𝑈 a (symmetric, positive definite) inner product on the tangent space
T 𝑥 (𝑈) at 𝑥 ∈ 𝑈, varying differentiably. Such an inner product defines the length of
√ vector k k, the angle between two tangent vectors, and the length element
a tangent
d𝑠 = d𝑠2 . In coordinates, we write
∑︁
d𝑠2 = 𝜂𝑖 𝑗 d𝑥 𝑖 d𝑥 𝑗
𝑖, 𝑗
for standard coordinates 𝑥𝑖 on R𝑛 , and the matrix (𝜂𝑖 𝑗 ) is symmetric, positive definite,
and differentiable. The metric determines a volume formula as
√︁
d𝑉 = det 𝜂 d𝑥1 · · · d𝑥 𝑛 .
A Riemannian metric gives 𝑈 ⊆ R𝑛 the structure of a path metric space, as

explained in 33.2.6: if 𝜐 : [0, 1] → 𝑈 is continuously differentiable, then we define its
length to be
∫ ∫ 1
ℓ(𝜐) = d𝑠 = k𝜐 0 (𝑡) k 𝑑𝑡.
𝜐 0
If 𝜙 : R 𝑘 → R𝑛 is continuously differentiable, the pullback metric 𝜙∗ (d𝑠2 ) is

defined by
𝜙∗ (d𝑠2 ) (𝑣, 𝑤) = d𝑠2 (D 𝑓 (𝑣), D 𝑓 (𝑤))
where 𝑣, 𝑤 ∈ T𝑧 (𝑈) and D is the derivative map.
The language of 33.8.1 gives another way to interpret the hyperbolic metric on H2 .
This point of view extends to provide a description of the full isometry group PSL2 (R)
as the unit tangent bundle of H2 , as follows.
33.8.2. The tangent space to H2 at a point 𝑧 ∈ H2 is T𝑧 H2 ' C and the tangent bundle
T(H2 ) := {(𝑧, 𝑣) : 𝑧 ∈ H2 , 𝑣 ∈ T𝑧 H2 }
is trivial (parallelizable), with T(H2 ) ' H2 × C. The Riemannian metric on H2 is

then defined by the metric on T𝑧 H2 given by
𝑣·𝑤
h𝑣, 𝑤i =
(Im 𝑧) 2
for 𝑣, 𝑤 ∈ T𝑧 (H2 ), a rescaling of the usual inner product on C over R. In particular,

k𝑣k = |𝑣|/(Im 𝑧) for 𝑣 ∈ T𝑧 (H2 ). The angle between two geodesics at an intersection
point 𝑧 ∈ T𝑧 H2 is then defined to be the angle between their tangent vectors in T𝑧 H2 ;
this notion of an angle coincides with the Euclidean angle measure.
The action of PSL2 (R) on H2 extends to an action on T(H2 ) in the expected way:

d𝑔(𝑧) 𝑎𝑧 + 𝑏 1
𝑔(𝑧, 𝑣) = 𝑔𝑧, 𝑣 = , 𝑣 .
d𝑧 𝑐𝑧 + 𝑑 (𝑐𝑧 + 𝑑) 2
Since isometries of H2 are differentiable, they act on the tangent bundle by differentials
preserving the norm and angle, and therefore Isom(H2 ) acts conformally or anti-
conformally on H2 .
If we restrict to the unit tangent bundle
UT(H2 ) := {(𝑧, 𝑣) ∈ T(H2 ) : k𝑣k 2𝑧 = 1}
then we obtain a bĳection

∼ UT(H2 )
PSL2 (R) −→

d𝑔
𝑔 ↦→ 𝑔𝑖, (𝑖)𝑖
d𝑧
(Exercise 33.17).
Remark 33.8.3. The natural generalization of Euclid’s geometry is performed on a
Riemannian manifold 𝑋 that is homogeneous, i.e., the isometry group Isom(𝑋) acts
transitively on 𝑋, as well as isotropic, i.e., Isom(𝑋) acts transitively on frames (a basis
of orthonormal tangent vectors) at a point. In this way, homogeneous says that every
point “looks the same”, and isotropic says that the geometry “looks the same in every
direction” at a point. Taken together, these natural conditions are quite strong, and
there are only three essentially distinct simply connected homogeneous and isotropic
geometries in any dimension, corresponding to constant sectional curvatures zero,
positive, or negative: these are Euclidean, spherical, and hyperbolic geometry, respec-
tively. Put this way, the hyperbolic plane is the unique complete, simply connected
Riemann surface with constant sectional curvature −1. For more on geometries in this
sense, we encourage the reader to consult Thurston [Thu97].
To conclude this section, we briefly review a few facts from the theory of Rieman-
nian manifolds.
33.8.4. A (topological) 𝑛-manifold is a (second-countable) Hausdorff topological
space 𝑋 locally homeomorphic to R𝑛 , i.e., for every 𝑥 ∈ 𝑋, there exists an open
neighborhood 𝑈 3 𝑥 and a continuous map 𝜙 : 𝑈 ↩→ R𝑛 that is a homeomorphism
onto an open subset; the map 𝜙 : 𝑈 → 𝜙(𝑈) ⊆ R𝑛 is called a chart (at 𝑥 ∈ 𝑋), and an
open cover of charts is called an atlas.
Two charts 𝜙1 : 𝑈1 → R𝑛 and 𝜙2 : 𝑈2 → R𝑛 are (𝐶 ∞ -)smoothly compatible if the
transition map
𝜙12 = 𝜙2 𝜙−1
1 : 𝜙1 (𝑈1 ∩ 𝑈2 ) → 𝜙2 (𝑈1 ∩ 𝑈2 )
(see Figure 33.8.5) is 𝐶 ∞ -smooth (i.e., has continuous partial derivatives of all orders).
An atlas on a manifold is smooth on a manifold if all charts are smoothly compatible.
A smooth manifold is a manifold equipped with a smooth atlas.
X
U2
U1
φ1 φ2
φ12
φ2 (U2 )
φ1 (U1 )
Rn Rn
Figure 33.8.5: A manifold, by its atlas
A morphism of smooth manifolds is a continuous map 𝑓 : 𝑌 → 𝑋 such that for
the atlases {(𝜙𝑖 , 𝑈𝑖 )}𝑖 of 𝑋 and {(𝜓 𝑗 , 𝑉 𝑗 )} 𝑗 of 𝑌 , each map
𝜙𝑖 𝑓 𝜓 −1
𝑗 : 𝜓 𝑗 (𝑉 𝑗 ∩ 𝑓
−1
(𝑈𝑖 )) → 𝜙𝑖 ( 𝑓 (𝑉 𝑗 ) ∩ 𝑈𝑖 )
is smooth. An isomorphism (diffeomorphism) of smooth manifolds is a bĳective

∼ 𝑋 such that 𝑓 and 𝑓 −1 are (𝐶 ∞ -)smooth.
morphism 𝑓 : 𝑌 −→
By the same definition as 33.8.1, we define a Riemannian metric on a smooth
𝑛-manifold.
One could similarly define 𝐶 𝑘 -smooth manifolds for any 1 ≤ 𝑘 ≤ ∞.
33.8.6. We similarly define a complex 𝑛-manifold, and morphisms between them, by
replacing R by C and smooth by holomorphic in the definition of a smooth manifold. A
Riemann surface is a complex 1-manifold. For further reference, see e.g. Donaldson
[Don2011] or Miranda [Mir95].
A complex 1-manifold (Riemann surface) defines a smooth, orientable Rieman-
nian 2-manifold by choosing the standard Euclidean metric on the complex plane;
conversely, a conformal structure on a smooth, oriented Riemannian 2-manifold de-
termines a complex 1-manifold. In other words, the category of Riemann surfaces is
equivalent to the category of smooth, orientable Riemannian 2-manifolds with confor-
mal transition maps and with conformal morphisms.
Example 33.8.7. The field C of complex numbers is the “original” Riemann surface,
and every open subset of C is a Riemann surface.
The simplest nonplanar example of a Riemann surface is the Riemann sphere
P1 (C) = C ∪ {∞}. The atlas on P1 (C) is given by the open sets
𝑈1 = P1 (C) r {∞} = C and 𝑈2 = P1 (C) r {0}
and charts 𝜙1 : 𝑈1 → C by 𝜙1 (𝑧) = 𝑧 and 𝜙2 : 𝑈2 → C by 𝜙2 (𝑧) = 1/𝑧; the map
𝜙2 𝜙−1
1 (𝑧) = 1/𝑧 is analytic on 𝜙1 (𝑈1 ∩ 𝑈2 ) = C r {0}. Topologically, the Riemann
sphere is the one-point compactification of C, and becomes a sphere by stereographic
projection.
Example 33.8.8. The inverse function theorem implies that if 𝑋 is a smooth projective
algebraic variety over C, then 𝑋 (C) has the canonical structure of a compact, complex
manifold.
Exercises
1. Show that the hyperbolic metric has the same topology as the Euclidean metric
in two ways.
(a) Show directly that open balls nest: for all 𝑧 ∈ H2 and all 𝜖 > 0, there exist
𝜂1 , 𝜂2 > 0 such that
𝜌(𝑧, 𝑤) < 𝜂1 ⇒ |𝑧 − 𝑤| < 𝜖 ⇒ 𝜌(𝑧, 𝑤) < 𝜂2
for all 𝑤 ∈ H2 .
(b) Show that the collection of Euclidean balls coincides with the collection
of hyperbolic balls. [Hint: applying an isometry, reduce to the case of
balls around 𝑖 and check this directly; it is perhaps even clearer moving to
the unit disc model.]
2. Check that in R𝑛 , the metric specified in (33.2.7)

∫ √︃
ℓ(𝜐) = 𝑥10 (𝑡) 2 + · · · + 𝑥 𝑛0 (𝑡) 2 d𝑡
𝜐
has lines as geodesics.

3. From differential geometry, the curvature of a Riemann surface with metric
√︃
𝑑𝑠 = 𝑓 (𝑥, 𝑦) d𝑥 2 + 𝑔(𝑥, 𝑦) d𝑦 2
is given by the formula
√ ! √︁ ! !
1 𝜕 1 𝜕 𝑔 𝜕 1 𝜕 𝑓
− √︁ √︁ + √
𝑓 𝑔 𝜕𝑥 𝑓 𝜕𝑥 𝜕𝑦 𝑔 𝜕𝑦
for suitably nice functions 𝑓 , 𝑔. Using this formula, verify that the curvature of
H2 and D2 is −1.
4. Consider C with the standard metric. Let
Isomℎ (C) = {𝑔 ∈ Isom(C) : 𝑔 is holomorphic} ≤ Isom(C).
Exhibit an isomorphism of groups

ℎ 𝑎 𝑏
Isom (C) ' ∈ GL2 (C) : |𝑎| = 1
0 1
and an isometry of C that is not holomorphic. [Hint: An invertible holomorphic
map C → C is of the form 𝑧 ↦→ 𝑎𝑧 + 𝑏.]
I 5. Show that for every 𝑧, 𝑧 0 ∈ H2 , there exists 𝑔 ∈ PSL2 (R) such that Re 𝑔𝑧 =
Re 𝑔𝑧 0 = 0. [Hint: Work in D2 .]
I 6. Show that the image of R>0𝑖 under an element of PSL2 (R) is either a semicircle
orthogonal to R or a vertical line. [Hint: Look at the endpoints.]
I 7. We consider the action of PSL2 (R) on geodesics in H2 .
(a) Show that PSL2 (R) acts transitively on the set of geodesics in H2 .
(b) Show that given 𝑧, 𝑧 0 ∈ H2 , there exists 𝑔 ∈ PSL2 (R) such that 𝑔𝑧 = 𝑧 0
and such that 𝑔 that maps the geodesic through 𝑧 and 𝑧 0 to itself.
(c) Show that every isometry of H2 that maps a geodesic to itself and fixes
two points on this geodesic is the identity.
I 8. Let 𝑧 1 , 𝑧2 ∈ H2 be distinct. Let
𝐻 (𝑧1 , 𝑧2 ) = {𝑧 ∈ H2 : 𝜌(𝑧, 𝑧1 ) ≤ 𝜌(𝑧, 𝑧2 )}
be the locus of points as close to 𝑧1 as to 𝑧 2 , and let 𝐿(𝑧 1 , 𝑧2 ) = bd 𝐻 (𝑧 1 , 𝑧2 ).
Show that 𝐻 (𝑧 1 , 𝑧2 ) is a convex (Definition 33.6.6) half-plane, and that
𝐿 (𝑧1 , 𝑧2 ) = {𝑧 ∈ H2 : 𝜌(𝑧, 𝑧1 ) = 𝜌(𝑧, 𝑧2 )}
is geodesic and equal to the perpendicular bisector of the geodesic segment from
𝑧1 to 𝑧2 .
9. Show that a hyperbolic polygon is convex if and only if it is the intersection of

finitely many half-planes 𝐻 (𝑧1 , 𝑧2 ) as in Exercise 33.8.
I 10. Show that the expression
| 𝑧 − 𝑧0| + | 𝑧 − 𝑧0|
| 𝑧 − 𝑧0| − | 𝑧 − 𝑧0|
with 𝑧, 𝑧 0 ∈ H2 is invariant under 𝑔 ∈ PSL2 (R). [Hint: check this on a
convenient set of generators.]
I 11. Show that
| 𝑧 − 𝑧0| + | 𝑧 − 𝑧0| | 𝑧 − 𝑧 0 |2
cosh log =1+
| 𝑧 − 𝑧0| − | 𝑧 − 𝑧0| 2 Im(𝑧) Im(𝑧 0)
for all 𝑧, 𝑧 0 ∈ H2 .
I 12. Verify that the hyperbolic metric on D2 is induced from the hyperbolic metric
on H2 from the identification (33.7.4), as follows.
(a) Show that
2| 𝜙 0 (𝑧)| 1
2
= .
1 − | 𝜙(𝑧)| Im 𝑧
(b) Let 𝑤 = 𝜙(𝑧), and using part (a) conclude that
2 |d𝑤| |d𝑧|
2
= .
1 − |𝑤| Im 𝑧
13. Show that the orientation-reversing isometry 𝑔(𝑧) = −𝑧 induces the map
(𝜙𝑔𝜙−1 ) (𝑤) = 𝑤
on D2 via the conformal transformation 𝜙 : H2 → D2 in (33.7.4).

14. Show that the Iwasawa decomposition (Proposition 33.4.2) can be given explic-
itly as
1 (𝑎𝑐 + 𝑏𝑑)/𝑟 2 1/𝑟 0 𝑠 −𝑡

𝑎 𝑏
= ∈ 𝑁 𝐴𝐾 = SL2 (R)
𝑐 𝑑 0 1 0 𝑟 𝑡 𝑠
√
where 𝑟 = 𝑐2 + 𝑑 2 , 𝑠 = 𝑑/𝑟, 𝑡 = 𝑐/𝑟.
15. Show that the map
L2 → D2
𝑥 + 𝑖𝑦
(𝑡, 𝑥, 𝑦) ↦→
𝑡+1
identifies the metrics on L2 and D2 , via pullback.
I 16. Show that the Lorentz metric restricted to the hyperboloid is an honest (Rie-
mannian) metric. [Hint: Show that a tangent vector 𝑣 at a point 𝑝 satisfies
𝑏( 𝑝, 𝑣) = 0, where 𝑏 is the bilinear form associated to 𝑞; then show that the
orthogonal complement to 𝑝 has signature +2.]
∼ UT(H2 ) defined by the action of 𝑔

I 17. Show that there is a bĳection PSL2 (R) −→
2
on a fixed base point in UT(H ). [Hint: Observe that elliptic elements rotate
the tangent vector.]
Chapter 34
Discrete group actions
Our ongoing goal in this part of the text is to understand quotient spaces that locally
look like (products of) hyperbolic spaces. In order to get off the ground, here we put
the previous two chapters in a more general context, seeking to understand nice group
actions on topological spaces and indicating how these fit in with more general notions
in topology. Pathologies exist! Our goal in this chapter is to provide basic context (for
further references see 34.5.4) before turning to the central case of interest: a discrete
group acting properly on a locally compact, Hausdorff topological space.
34.1 Topological group actions

Group actions will figure prominently in what follows, so we set a bit of notation.
There are many references for topological groups, including Arhangel’skii–Tkachenko
[AT2008, Chapter 1] and McCarty [McC2011, Chapter V].
Let 𝐺 be a group and let 𝑋 be a set. Recall that a left action of 𝐺 on 𝑋 is a map
𝐺×𝑋→𝑋
(34.1.1)
(𝑔, 𝑥) ↦→ 𝑔𝑥
satisfying 1𝑥 = 𝑥 and (𝑔𝑔 0)𝑥 = 𝑔(𝑔 0𝑥) for all 𝑥 ∈ 𝑋 and all 𝑔, 𝑔 0 ∈ 𝐺. A right action
is instead a map 𝑋 × 𝐺 → 𝑋 by (𝑥, 𝑔) ↦→ 𝑥𝑔 satisfying 𝑥1 = 𝑥 and 𝑥(𝑔 0 𝑔) = (𝑥𝑔 0)𝑔
for all 𝑥 ∈ 𝑋 and 𝑔, 𝑔 0 ∈ 𝑔. We will need to consider actions both on the left and the
right; if not specified, a left action is assumed.
We will also sometimes write 𝐺 𝑋 for an action of 𝐺 on 𝑋.
Example 34.1.2. A group 𝐺 acts on itself by left multiplication, the (left) regular
group action of 𝐺. If 𝐻 ≤ 𝐺 is a subgroup, then 𝐻 also acts on 𝐺 by left multiplication.
For example, if 𝑉 is an R-vector space with dimR 𝑉 = 𝑛, and Λ ⊆ 𝑉 is a (full) Z-lattice
in 𝑉, then Λ ' Z𝑛 is a group and Λ acts on 𝑉 by translation.
Another important and related example is the left action of 𝐺 on the set of right
cosets 𝑋 = 𝐺/𝐻 again by multiplication, namely
𝑔(𝑥𝐻) = 𝑔𝑥𝐻 for 𝑔 ∈ 𝐺 and 𝑥𝐻 ∈ 𝐺/𝐻.
615
616 CHAPTER 34. DISCRETE GROUP ACTIONS
Let 𝐺 act on 𝑋. The 𝐺-orbit of 𝑥 ∈ 𝑋 is 𝐺𝑥 = {𝑔𝑥 : 𝑔 ∈ 𝐺}. The set of 𝐺-orbits

forms the quotient set 𝐺\𝑋 = {𝐺𝑥 : 𝑥 ∈ 𝑋 }, with a natural surjective quotient map
𝜋 : 𝑋 → 𝐺\𝑋.
Remark 34.1.3. We write 𝐺\𝑋 for the quotient, as 𝐺 acts on the left; for a right action,
we write 𝑋/𝐺, etc.
Example 34.1.4. A group 𝐺 acts transitively on a nonempty set 𝑋 if and only if 𝐺\𝑋
is a single point; in this case, we call 𝑋 homogeneous under 𝐺. In particular, if
𝐻 ≤ 𝐺, then the action of 𝐺 on 𝐺/𝐻 is transitive.
If 𝐻 ≤ 𝐺, then the quotient set 𝐻\𝐺 is the set of left cosets of 𝐻 in 𝐺. For
example, if Λ = Z𝑛 ≤ R𝑛 = 𝑉, then Λ\𝑉 ' (R/Z) 𝑛 .
For 𝑥 ∈ 𝑋, we define the stabilizer of 𝑥 by Stab𝐺 (𝑥) = {𝑔 ∈ 𝐺 : 𝑔𝑥 = 𝑥}.
Definition 34.1.5. The action of 𝐺 on 𝑋 is free (and we say 𝐺 acts freely on 𝑋) if
Stab𝐺 (𝑥) = {1} for all 𝑥 ∈ 𝑋, i.e., 𝑔𝑥 = 𝑥 implies 𝑔 = 1 for all 𝑥 ∈ 𝑋.
Definition 34.1.6. Let 𝑋 0, 𝑋 be sets with an action of 𝐺. A map 𝑓 : 𝑋 0 → 𝑋 is
𝐺-equivariant if 𝑓 (𝑔𝑥 0) = 𝑔( 𝑓 (𝑥 0)) for all 𝑥 0 ∈ 𝑋 0 and 𝑔 ∈ 𝐺, i.e., the following
diagram commutes:
𝑋0
𝑓
/𝑋
𝑔 𝑔

𝑋0
𝑓
/𝑋
If 𝑓 : 𝑋 0 → 𝑋 is 𝐺-equivariant, then 𝑓 induces a map
𝐺\𝑋 0 → 𝐺\𝑋
𝐺𝑥 0 ↦→ 𝐺 𝑓 (𝑥 0),
well-defined by the 𝐺-equivariance of 𝑓 , and the following diagram commutes:
𝑋0
𝑓
/𝑋
𝜋0 𝜋 (34.1.7)

𝐺\𝑋 0 / 𝐺\𝑋
Now topology enters. Let 𝐺 be a topological group (Definition 12.2.1), a group

with a topology in which the multiplication and inversion maps are continuous. Let 𝑋
be a topological space, and let 𝐺 act on 𝑋. We want to consider only those actions in
which the topology on 𝐺 and on 𝑋 are compatible.
Definition 34.1.8. The action of 𝐺 on 𝑋 is continuous if the map 𝐺 × 𝑋 → 𝑋 is
continuous.
Example 34.1.9. The left regular action of a group on itself is continuous—indeed,
combined with continuity of inversion (and existence of the identity), this is the very
definition of a topological group.
34.1. TOPOLOGICAL GROUP ACTIONS 617
Lemma 34.1.10. Suppose 𝐺 has the discrete topology. Then an action of 𝐺 on 𝑋 is

continuous if and only if for all 𝑔 ∈ 𝐺 the left-multiplication map
𝜆𝑔 : 𝑋 → 𝑋
𝑥 ↦→ 𝑔𝑥
is continuous; and when this holds, each 𝜆 𝑔 is a homeomorphism.
From now on, suppose 𝐺 acts continuously on 𝑋; more generally, whenever 𝐺 is a

topological group acting on a topological space 𝑋, we will implicitly suppose that the
action is continuous.
34.1.11. The quotient 𝐺\𝑋 is equipped with the quotient topology, so that the quotient
map 𝜋 : 𝑋 → 𝐺\𝑋 is continuous: a subset 𝑉 ⊆ 𝐺\𝑋 is open if and only if 𝜋 −1 (𝑉) ⊆ 𝑋
is open.
The projection 𝜋 is an open map, which is to say ifÐ𝑈 ⊆ 𝑋 is open then 𝜋(𝑈) =
𝐺𝑈 ⊆ 𝐺\𝑋 is open: if 𝑈 is open then 𝜋 −1 (𝜋(𝑈)) = 𝑔 ∈𝐺 𝑔𝑈 is open, so 𝜋(𝑈) is
open by definition of the topology.
34.1.12. If 𝐺 acts continuously on 𝑋, then the topologies on 𝐺 and 𝑋 are related by

this action. In particular, for all 𝑥 ∈ 𝑋, the natural map
𝐺 → 𝐺𝑥 ⊆ 𝑋
𝑔 ↦→ 𝑔𝑥
is continuous (it is the restriction of the action map to 𝐺 × {𝑥}). Let 𝐾 = Stab𝐺 (𝑥).
Then this map factors naturally as
∼ 𝐺𝑥
𝐺/𝐾 −→ (34.1.13)
where we give 𝐺/𝐾 the quotient topology; then (34.1.13) then a bĳective continuous
map, a topological upgrade of the orbit–stabilizer theorem. The map (34.1.13) need
not always be a homeomorphism (Exercise 34.5), but we will see below that it becomes
a homeomorphism under further nice hypotheses (Exercise 34.6, Proposition 34.4.11).
In order to work concretely with the quotient 𝐺\𝑋, it is convenient to choose

representatives of each orbit as follows. We write cl for topological closure and int for
topological interior.
Definition 34.1.14. A fundamental set for 𝐺 𝑋 is a subset ◊ ⊆ 𝑋 such that:
(i) cl(int(◊)) = ◊;
(ii) 𝐺◊ = 𝑋; and
(iii) int(◊) ∩ int(𝑔◊) = ∅ for all 1 ≠ 𝑔 ∈ 𝐺.
The condition (i) ensures our basic intuition about tilings (and avoids fundamental
sets that contain an extraneous number of isolated points); condition (ii) says that ◊
tiles 𝑋; and condition (iii) shows that the tiles only overlap along the boundary, and
there is no redundancy in the interior. If there is a fundamental set for 𝐺 𝑋, then
the action is faithful.
34.1.15. Let ◊ ⊆ 𝑋 be a fundamental set for 𝐺 𝑋. Then 𝐺 induces an equivalence

∼ 𝐺\𝑋 is a bĳection.
relation on ◊, and 𝐺\◊ −→
Remark 34.1.16. In chapter 37, we place further restrictions on a fundamental set

to ensure that they retain good properties, calling such a set a fundamental domain
(Definition 37.1.11).
34.2 ⊲ Summary of results

We pause to provide a quick summary of the results in this chapter for the special case
of discrete subgroups of PSL2 (R) (proven in Theorem 34.5.1 and Proposition 34.7.2).
The reader who is willing to accept the theorem below, and will stick to this case, can
profitably skip the rest of this chapter.
The group PSL2 (R) has a natural topology from the metric on SL2 (R) ⊆ M2 (R)
(see 34.6.1): intuitively, two matrices in PSL2 (R) are close if after a choice of sign all
of their entries are close.
Theorem 34.2.1. Let Γ ≤ PSL2 (R) be a subgroup and equip Γ with the subspace
topology. Then the following are equivalent:
(i) Γ is discrete;
(ii) For all 𝑧 ∈ H2 , we have # StabΓ (𝑧) < ∞ and there exists an open neigborhood
𝑈 3 𝑧 such that 𝛾𝑈 ∩ 𝑈 ≠ ∅ implies 𝛾 ∈ StabΓ (𝑧);
(iii) For all compact subsets 𝐾 ⊆ H2 , we have 𝐾 ∩ 𝛾𝐾 ≠ ∅ for only finitely many
𝛾 ∈ Γ; and
(iv) For all 𝑧 ∈ H2 , the orbit Γ𝑧 ⊆ H2 is discrete and # StabΓ (𝑧) < ∞.
Moreover, if these equivalent conditions hold, then the quotient Γ\H2 is Hausdorff,
and the quotient map 𝜋 : H2 → Γ\H2 is a local isometry at all points 𝑧 ∈ H2 with
StabΓ (𝑧) = {1}.
A discrete subgroup Γ ≤ PSL2 (R) is called a Fuchsian group.
34.3 Covering space and wandering actions

Throughout the remainder of this chapter, let 𝑋 be a Hausdorff topological space with
a (continuous) action of a Hausdorff topological group 𝐺.
Definition 34.3.1. We say the action of 𝐺 on 𝑋 is a covering space action if for all
𝑥 ∈ 𝑋, there exists an open neighborhood 𝑈 3 𝑥 such that 𝑔𝑈 ∩ 𝑈 = ∅ for all 𝑔 ∈ 𝐺
with 𝑔 ≠ 1.
34.3. COVERING SPACE AND WANDERING ACTIONS 619
34.3.2. If 𝐺 𝑋 is a covering space action, then the quotient map 𝜋 : 𝑋 → 𝐺\𝑋 is a

local homeomorphism, i.e., for every 𝑥 ∈ 𝑋, there exists an open neighborhood 𝑈 3 𝑥
such that 𝜋|𝑈 : 𝑈 → 𝜋(𝑈) ⊆ 𝐺\𝑋 is a homeomorphism. A local homeomorphism
need not conversely be a covering space map.
If 𝐺 acts by a covering space action, then 𝐺 acts freely on 𝑋. This is too strong a
hypothesis on the group actions we will consider in the rest of this book, so we need
to look for something weaker. So we consider the following.
Definition 34.3.3. We say that the action of 𝐺 is wandering if for all 𝑥 ∈ 𝑋, there
exists an open neighborhood 𝑈 3 𝑥 such that 𝑔𝑈 ∩𝑈 ≠ ∅ for only finitely many 𝑔 ∈ 𝐺.
Example 34.3.4. If 𝐺 is a finite group, then every action of 𝐺 is wandering.
34.3.5. If the action of 𝐺 is wandering, then for all 𝑥 ∈ 𝑋, the orbit 𝐺𝑥 ⊆ 𝑋 is closed
and discrete.
Wandering actions generalize covering space actions, and can be equivalently
characterized, as follows.
Lemma 34.3.6. The following are equivalent:
(i) The action of 𝐺 is wandering; and
(ii) For all 𝑥 ∈ 𝑋 we have # Stab𝐺 (𝑥) < ∞, and there exists an open neighborhood
𝑈 3 𝑥 such that 𝑔𝑈 ∩ 𝑈 ≠ ∅ implies 𝑔 ∈ Stab𝐺 (𝑥).
If 𝐺 acts freely, then these are further equivalent to:
(iii) The action of 𝐺 is by a covering space action.
Proof. The implication (ii) ⇒ (i) is immediate; we prove the converse. Let 𝑈 be a
neighborhood of 𝑥 ∈ 𝑋 such that 𝑔𝑈 ∩ 𝑈 ≠ ∅ for only finitely many 𝑔 ∈ 𝐺. We have
# Stab𝐺 (𝑥) < ∞ since 𝑔 ∈ Stab𝐺 (𝑥) implies 𝑥 ∈ 𝑔𝑈 ∩ 𝑈. Let
{𝑔 ∈ 𝐺 : 𝑔𝑈 ∩ 𝑈 ≠ ∅ and 𝑔𝑥 ≠ 𝑥} = {𝑔1 , . . . , 𝑔𝑛 }. (34.3.7)
Since 𝑋 is Hausdorff, for all 𝑖 there exist open neighborhoods 𝑉𝑖 , 𝑊𝑖 ⊆ 𝑋 of 𝑥, 𝑔𝑖 𝑥,

respectively, such that 𝑉𝑖 ∩ 𝑊𝑖 = ∅. Replacing 𝑉𝑖 with 𝑉𝑖 ∩ 𝑈 (still with 𝑥 ∈ 𝑉𝑖 and
𝑉𝑖 ∩ 𝑊𝑖 = ∅), we may suppose that 𝑉𝑖 ⊆ 𝑈. Since 𝐺 acts continuously, there exists
an open neighborhood 𝑊𝑖0 ⊆ 𝑋 of 𝑥 such that 𝑔𝑖 𝑊𝑖0 ⊆ 𝑊𝑖 . Let 𝑈𝑖 := 𝑉𝑖 ∩ 𝑊𝑖0. Since
𝑉𝑖 ⊆ 𝑈 we have 𝑈𝑖 ⊆ 𝑈 for all 𝑖. Further, 𝑥 ∈ 𝑈𝑖 and
𝑈𝑖 ∩ 𝑔𝑖 𝑈𝑖 ⊆ 𝑈𝑖 ∩ 𝑔𝑖 𝑊𝑖0 ⊆ 𝑉𝑖 ∩ 𝑊𝑖 = ∅. (34.3.8)
We claim that 𝑈 0 = 𝑖 𝑈𝑖 has the desired property in (ii). Suppose that 𝑔𝑈 0 ∩𝑈 0 ≠ ∅

Ñ
for some 𝑔 ∈ 𝐺. Since 𝑈𝑖 ⊆ 𝑈 for all 𝑖 we have 𝑈 0 ⊆ 𝑈, so 𝑔𝑈 ∩ 𝑈 ≠ ∅; then by
(34.3.7), either 𝑔𝑥 = 𝑥 or 𝑔 = 𝑔𝑖 for some 𝑖. We cannot have 𝑔 = 𝑔𝑖 , since this would
imply 𝑔𝑖 𝑈 0 ∩𝑈 0 ⊆ 𝑔𝑖 𝑈𝑖 ∩𝑈𝑖 = ∅, contradicting (34.3.8). So 𝑔𝑥 = 𝑥, and 𝑔 ∈ Stab𝐺 (𝑥)
as claimed.
Finally, if 𝐺 acts freely, then Stab𝐺 (𝑥) is trivial for all 𝑥, whence the equivalence
(ii) ⇔ (iii).
34.3.9. Suppose the action of 𝐺 is wandering. We recall Lemma 34.3.6(ii). At a point

𝑥 with open neighborhood 𝑈 3 𝑥 and finite stabilizer Stab𝐺 (𝑥), we can replace 𝑈 by
Ñ
𝑔 ∈Stab𝐺 ( 𝑥) 𝑔𝑈 so that 𝑈 3 𝑥 is an open neighborhood on which Stab𝐺 (𝑥) acts (see
Exercise 34.12). Then the projection map factors as
𝜋|𝑈 : 𝑈 → Stab𝐺 (𝑥)\𝑈 → 𝐺\𝑋
and the latter map Stab𝐺 (𝑥)\𝑈 → 𝐺\𝑋 is a homeomorphism onto its image; we say
𝜋 is a local homeomorphism modulo stabilizers. If the action of 𝐺 is free, then we
recover 34.3.2.
Remark 34.3.10. If 𝐺 has the discrete topology and the condition in Lemma 34.3.6(ii)
holds, then some authors call the action of 𝐺 properly discontinuous. This is probably
because 𝐺 is then as broken (“discontinuous”) as possible: 𝐺 has the discrete topology,
and we should be able to find neighborhoods that pull apart the action of 𝐺. (Klein
[Kle79, p. 321] uses the term discontinuous because “points that are ‘equivalent’ with
respect to [the group] are separated”.) This nomenclature is strange because we still
want the action to be continuous, just by a discrete group. Adding to the potential
confusion is the issue that different authors give different definitions of “properly
discontinuous” depending on their purposes; most of these can be seen to be equivalent
under the right hypotheses on the space, but not all. We avoid this term.
It turns out that a wandering action is too weak a property in this level of generality
for us to work with. However, it is close, and we will shortly see that it suffices with
additional hypotheses on the space 𝑋.
Remark 34.3.11. Let 𝑋 be a topological space, and let 𝐺 be a set of continuous maps
𝑋 → 𝑋. Then there is a natural map 𝐺 ↩→ 𝑋 𝑋 defined by 𝑔 ↦→ (𝑔𝑥) 𝑥 . We give 𝑋 𝑋
the compact-open topology and 𝐺 the subspace topology, so a subbasis of the topology
on 𝐺 is given by
𝑉 (𝐾, 𝑈) = { 𝑓 ∈ 𝐺 : 𝑓 (𝐾) ⊆ 𝑈}
for 𝐾 ⊆ 𝑋 compact and 𝑈 ⊆ 𝑋 open.
If 𝑋 is Hausdorff and locally compact, then the compact-open topology on 𝐺 is
the weakest topology (smallest, fewest open sets) for which the map 𝐺 × 𝑋 → 𝑋
is continuous (also called an admissible topology on 𝐺) [McC2011, §VII, pp. 171–
172]. Under the hypotheses of Exercise 34.4, this implies that the topology of pointwise
convergence and the compact-open topology coincide.
34.4 Hausdorff quotients and proper group actions

In this section we define proper group actions; to motivate this definition, we first
ask for conditions that imply that a quotient space is Hausdorff. Throughout this
section and the next, let 𝑋 be a Hausdorff topological space and let 𝐺 be a Hausdorff
topological group acting continuously on 𝑋.
Lemma 34.4.1. The following are equivalent:
(i) The quotient 𝐺\𝑋 is Hausdorff;
34.4. HAUSDORFF QUOTIENTS AND PROPER GROUP ACTIONS 621
(ii) If 𝐺𝑥 ≠ 𝐺 𝑦 ∈ 𝐺\𝑋, then there exist open neighborhoods 𝑈 3 𝑥 and 𝑉 3 𝑦 such

that 𝑔𝑈 ∩ 𝑉 = ∅ for all 𝑔 ∈ 𝐺; and
(iii) The image of the action map
𝐺×𝑋→𝑋×𝑋
(34.4.2)
(𝑔, 𝑥) ↦→ (𝑥, 𝑔𝑥)
is closed.
Proof. The implication (i) ⇔ (ii) follows directly from properties of the quotient map:
the preimage of open neighborhoods separating 𝐺𝑥 and 𝐺 𝑦 under the continuous
projection map have the desired properties, and conversely the pushforward of the
given neighborhoods under the open projection map separate 𝐺𝑥 and 𝐺 𝑦.
To conclude, we prove (i) ⇔ (iii). We use the criterion that a topological space is
Hausdorff if and only if the diagonal map has closed image. The continuous surjective
map 𝜋 : 𝑋 → 𝐺\𝑋 is open, so the same is true for
𝜋 × 𝜋 : 𝑋 × 𝑋 → (𝐺\𝑋) × (𝐺\𝑋).
Therefore the diagonal 𝐺\𝑋 ↩→ (𝐺\𝑋) × (𝐺\𝑋) is closed if and only if its preimage
is closed in 𝑋 × 𝑋. But this preimage consists exactly of the orbit relation
{(𝑥, 𝑥 0) ∈ 𝑋 × 𝑋 : 𝑥 0 = 𝑔𝑥 for some 𝑔 ∈ 𝐺},
and this is precisely the image of the action map (34.4.2).
The conditions Lemma 34.4.1(i)–(ii) can sometimes be hard to verify, so it is

convenient to have a condition that implies Lemma 34.4.1(iii); this definition will seek
to generalize the situation when 𝐺 is compact. First, we make a definition.
Definition 34.4.3. Let 𝑓 : 𝑋 → 𝑌 be a continuous map.
(a) We say 𝑓 : 𝑋 → 𝑌 is quasi-proper if 𝑓 −1 (𝐾) is compact for all compact 𝐾 ⊆ 𝑌 .

(b) We say 𝑓 is proper if 𝑓 is quasi-proper and closed (the image of every closed
subset is closed).
Example 34.4.4. If 𝑋 is compact, then every continuous map 𝑓 : 𝑋 → 𝑌 is proper

because 𝑓 is closed and if 𝐾 ⊆ 𝑌 is compact, then 𝐾 is closed, so 𝑓 −1 (𝐾) ⊂ 𝑋 is
closed hence compact, since 𝑋 is compact.
Lemma 34.4.5. Suppose that 𝑌 is locally compact and Hausdorff, and let 𝑓 : 𝑋 → 𝑌
be continuous and quasi-proper. Then 𝑋 is locally compact, and 𝑓 is proper.
Proof. For the first statement, cover 𝑌 with open relatively compact sets 𝑈𝑖 ⊆ 𝐾𝑖 ; then
𝑉𝑖 = 𝑓 −1 (𝑈𝑖 ) is an open cover of 𝑋 by relatively compact sets.
Next, we claim that 𝑓 is in fact already proper; that is to say, we show that 𝑓
is closed. Let 𝑊 ⊆ 𝑋 be a closed set and consider a sequence {𝑦 𝑛 } 𝑛 from 𝑓 (𝑊)
with 𝑦 𝑛 → 𝑦. Let 𝐾 be a compact neighborhood of 𝑦 containing {𝑦 𝑛 }; taking a
subsequence, we may suppose all 𝑦 𝑛 ∈ 𝐾. Let 𝑥 𝑛 ∈ 𝑓 −1 (𝑦 𝑛 ) ∩ 𝑊 be primages. Since
𝑓 is quasi-proper, we have 𝑓 −1 (𝐾) is compact. Suppose for a moment that 𝑓 −1 (𝐾)

is sequentially compact (for example, if 𝑋 is second countable or metrizable). Then
again taking a subsequence, we may suppose that 𝑥 𝑛 → 𝑥 with 𝑥 ∈ 𝑊 since 𝑊 is
closed. By continuity, 𝑓 (𝑥 𝑛 ) → 𝑓 (𝑥) = 𝑦, so 𝑓 (𝑊) is closed. To avoid the extra
hypothesis that 𝑓 −1 (𝐾) is sequentially compact, replace the sequence {𝑦 𝑛 } with a net;
the argument proceeds identically.
Remark 34.4.6. There is an alternate characterization of proper maps as follows: a

continuous map 𝑓 : 𝑋 → 𝑌 is proper if and only if the map 𝑓 × id : 𝑋 × 𝑍 → 𝑌 × 𝑍 is
closed for every topological space 𝑍. See 34.5.4 for more discussion.
Partly motivated by Lemma 34.4.1(iv), we make the following definition.
Definition 34.4.7. The action of 𝐺 on 𝑋 is proper (𝐺 acts properly on 𝑋) if the
action map
𝜆:𝐺×𝑋→𝑋×𝑋
(34.4.8)
(𝑔, 𝑥) ↦→ (𝑥, 𝑔𝑥)
is proper.
Proposition 34.4.9. If 𝐺 is compact, then every (continuous) action of 𝐺 on (a
Hausdorff space) 𝑋 is proper.
Proof. Let 𝐾 ⊆ 𝑋 × 𝑋 be compact; then 𝐾 is closed (because 𝑋 is Hausdorff). Let

𝐾1 be the projection of 𝐾 onto the first factor. Then 𝐾1 is compact, and 𝜆−1 (𝐾) is a
closed subset of the compact set 𝐺 × 𝐾1 , so it is compact. This shows that the action
map is quasi-proper. Finally, the action map is closed. We factor the map as
𝐺×𝑋→𝐺×𝑋×𝑋 →𝑋×𝑋
(𝑔, 𝑥) ↦→ (𝑔, 𝑥, 𝑔𝑥) ↦→ (𝑥, 𝑔𝑥);
the first map is the graph of a continuous map to a Hausdorff space and is closed
(Exercise 34.9); the second (projection) map is closed, as 𝐺 is compact (by a standard
application of the tube lemma). Therefore the composition of these maps is closed.
Example 34.4.10. If 𝐺 is a finite discrete group, then 𝐺 acts properly by Proposition

34.4.9.
Proper actions have many of the properties we need.
Proposition 34.4.11. Let 𝐺 act properly on 𝑋. Then the following are true.
(a) 𝐺\𝑋 is Hausdorff.
(b) The orbit 𝐺𝑥 ⊆ 𝑋 is closed for all 𝑥 ∈ 𝑋.
(c) The natural map
𝜄 𝑥 : 𝐺/Stab𝐺 (𝑥) → 𝐺𝑥
𝑔 ↦→ 𝑔𝑥 ∈ 𝑋
is a homeomorphism.
34.5. PROPER ACTIONS ON A LOCALLY COMPACT SPACE 623
(d) The group Stab𝐺 (𝑥) is compact for all 𝑥 ∈ 𝑋.
Proof. For part (a), by Lemma 34.4.1, it is enough to note that by definition the image
of the action map 𝜆 in (34.4.8) is closed. Part (b) follows in the same way, as
𝐺𝑥 ' {𝑥} × 𝐺𝑥 = 𝜆(𝐺 × {𝑥}).
This also implies part (c) (cf. 34.1.12): the map 𝜄 𝑥 is bĳective and continuous, and
it is also closed (whence a homeomorphism) since 𝜄 𝑥 is a factor of the closed map
𝐺 → 𝐺𝑥.
Finally, for part (d), let 𝜆 : 𝐺 × 𝑋 → 𝑋 × 𝑋 be the action map and let 𝑥 ∈ 𝑋. Then
by definition that 𝜆−1 (𝑥, 𝑥) = Stab𝐺 (𝑥) × {𝑥} ' Stab𝐺 (𝑥), so by definition Stab𝐺 (𝑥)
is compact.
34.5 Proper actions on a locally compact space

When 𝑋 is locally compact, our central case of interest, there are several equivalent
characterizations of a proper discrete action 𝐺 𝑋. For more about proper group
actions and covering spaces, see Lee [Lee2011, Chapter 12].
Recall our running assumption that 𝑋 and 𝐺 are Hausdorff.
Theorem 34.5.1. Suppose that 𝑋 is locally compact and let 𝐺 act (continuously) on
𝑋. Then the following are equivalent:
(i) 𝐺 is discrete and acts properly on 𝑋;
(ii) For all compact subsets 𝐾 ⊆ 𝑋, we have 𝐾 ∩ 𝑔𝐾 ≠ ∅ for only finitely many
𝑔 ∈ 𝐺;
(iii) For all compact subsets 𝐾, 𝐿 ⊆ 𝑋, we have 𝐾 ∩ 𝑔𝐿 ≠ ∅ for only finitely many
𝑔 ∈ 𝐺; and
(iv) For all 𝑥, 𝑦 ∈ 𝑋, there exist open neighborhoods 𝑈 3 𝑥 and 𝑉 3 𝑦 such that
𝑈 ∩ 𝑔𝑉 ≠ ∅ for only finitely many 𝑔 ∈ 𝐺.
Moreover, if 𝑋 is a locally compact metric space and 𝐺 acts by isometries, then these
are further equivalent to:
(v) The action of 𝐺 on 𝑋 is wandering; and
(vi) For all 𝑥 ∈ 𝑋, the orbit 𝐺𝑥 ⊆ 𝑋 is discrete and # Stab𝐺 (𝑥) < ∞.
Proof. First, we show (i) ⇒ (ii). Let 𝜆 : 𝐺 × 𝑋 → 𝑋 × 𝑋 be the action map. Let
𝐾 ⊆ 𝑋 be compact. Then
𝜆−1 (𝐾 × 𝐾) = {(𝑔, 𝑥) ∈ 𝐺 × 𝑋 : 𝑥 ∈ 𝐾, 𝑔𝑥 ∈ 𝐾 }
is compact by definition. The projection of 𝜆−1 (𝐾 × 𝐾) onto 𝐺 is compact, and since

𝐺 is discrete, this projection is finite and includes all 𝑔 ∈ 𝐺 such that 𝐾 ∩ 𝑔𝐾 ≠ ∅.
Next we show (ii) ⇔ (iii): The implication (ii) ⇐ (iii) is immediate, and conversely
we apply (ii) to the compact set 𝐾 ∪ 𝐿 to conclude
𝐾 ∩ 𝑔𝐿 ⊆ (𝐾 ∪ 𝐿) ∩ 𝑔(𝐾 ∪ 𝐿) ≠ ∅
for only finitely many 𝑔 ∈ 𝐺.

Next we show (ii) ⇒ (iv). For all 𝑥 ∈ 𝑋, since 𝑋 is locally compact there is a
compact neighborhood 𝐾 ⊇ 𝑈 3 𝑥, with 𝑈 open and 𝐾 compact. If 𝑈 ∩ 𝑔𝑈 ≠ ∅ then
𝐾 ∩ 𝑔𝐾 ≠ ∅ and this happens for only finitely many 𝑔 ∈ 𝐺.
Finally, we show (iv) ⇒ (i). We first show that the action map is quasi-proper, and
conclude that it is proper by Lemma 34.4.5. Let 𝐾 ⊆ 𝑋 × 𝑋 be compact. By (iv), for
all (𝑥, 𝑦) ∈ 𝐾, there exist open neighborhoods 𝑈 3 𝑥 and 𝑉 3 𝑦 such that the set
𝑊 = {𝑔 ∈ 𝐺 : 𝑔𝑈 ∩ 𝑉 ≠ ∅}
is finite. The set 𝑈 × 𝑉 3 (𝑥, 𝑦) is an open neighborhood of (𝑥, 𝑦) ∈ 𝐾, and so the

collection of these neighborhoods ranging over (𝑥, 𝑦) ∈ 𝐾 is an open cover of 𝐾, so
Ð many 𝑈𝑖 × 𝑉𝑖 3 (𝑥𝑖 , 𝑦 𝑖 ) suffice, and with corresponding sets #𝑊𝑖 < ∞. Let
finitely
𝑊 = 𝑖 𝑊𝑖 ⊆ 𝐺. Let 𝐾1 ⊆ 𝑋 be the projection of 𝐾 onto the first coordinate. We
claim that 𝜆−1 (𝐾) ⊆ 𝑊 × 𝐾1 . Indeed, if 𝜆(𝑔, 𝑥) = (𝑥, 𝑔𝑥) ∈ 𝐾 then 𝑥 ∈ 𝐾1 and
(𝑥, 𝑔𝑥) ∈ 𝑈𝑖 × 𝑉𝑖 for some 𝑖, so 𝑔𝑥 ∈ 𝑔𝑈𝑖 ∩ 𝑉𝑖 and 𝑔 ∈ 𝑊𝑖 , and thus (𝑔, 𝑥) ∈ 𝑊 × 𝐾1 .
Since #𝑊 < ∞ and 𝐾1 is compact, 𝑊 × 𝐾1 is compact; and then since 𝐾 is compact,
𝐾 is closed so 𝜆−1 (𝐾) ⊆ 𝑊 × 𝐾1 is also closed, hence compact.
To conclude that 𝐺 is discrete, we argue as follows. For all 𝑥 ∈ 𝑋, the orbit 𝐺𝑥 ⊆ 𝑋
is discrete: taking 𝑈 = 𝑉 and a neighborhood 𝑈 3 𝑥 with 𝑈 ∩ 𝑔𝑈 ≠ ∅ for only finitely
many 𝑔 ∈ 𝐺, we see that 𝑈 ∩ 𝐺𝑥 is finite so 𝐺𝑥 is discrete (as 𝑋 is Hausdorff). By
Proposition 34.4.11(d), the map
𝐺/Stab𝐺 (𝑥) → 𝐺𝑥
is a homeomorphism for all 𝑥 ∈ 𝑋. Therefore, Stab𝐺 (𝑥) (the preimage of 𝑥) is an open,

finite (Hausdorff) neighborhood of 1; but then Stab𝐺 (𝑥) is discrete, and transporting
we conclude that the topological group 𝐺 has an open cover by discrete sets, and thus
𝐺 is discrete. This completes the equivalence (i)–(iv).
The implication (iv) ⇒ (v) holds in all cases: taking 𝑥 = 𝑦, the neighborhood
𝑈 ∩ 𝑉 is as required in the definition of a wandering action. The implication (v) ⇒
(vi) also holds in all cases from 34.3.5 and Lemma 34.3.6.
To conclude, we show (vi) ⇒ (ii) under the extra hypothesis that 𝑋 is a metric
space with 𝐺 acting by isometries. Assume for purposes of contradiction that there
exist infinitely many 𝑔𝑛 ∈ 𝐺 such that 𝐾 ∩ 𝑔𝑛 𝐾 ≠ ∅, and accordingly let 𝑥 𝑛 ∈ 𝐾 with
𝑔𝑛 𝑥 𝑛 ∈ 𝐾. The points 𝑥 𝑛 accumulate in 𝐾, so we may suppose 𝑥 𝑛 → 𝑥 ∈ 𝐾; by taking
a further subsequence, we may suppose also that 𝑔𝑛 𝑥 𝑛 → 𝑦 ∈ 𝐾. We then claim that
the set {𝑔𝑛 𝑥} 𝑛 accumulates near 𝑦. Since # Stab𝐺 (𝑥) < ∞, we may suppose without
loss of generality that the points 𝑔𝑛 𝑥 are all distinct. Then, given 𝜖 > 0,
𝜖 𝜖
𝜌(𝑔𝑛 𝑥, 𝑦) ≤ 𝜌(𝑔𝑛 𝑥, 𝑔𝑛 𝑥 𝑛 ) + 𝜌(𝑔𝑛 𝑥 𝑛 , 𝑦) = 𝜌(𝑥, 𝑥 𝑛 ) + 𝜌(𝑔𝑛 𝑥 𝑛 , 𝑦) < + = 𝜖,
2 2
−1 𝑔 ∈ 𝐺. By the Cauchy criterion,
for 𝑛 sufficiently large, so 𝑔𝑛 𝑥 → 𝑦. Let ℎ 𝑛 = 𝑔𝑛+1 𝑛
𝜌(ℎ 𝑛 𝑥, 𝑥) = 𝑑 (𝑔𝑛 𝑥, 𝑔𝑛+1 𝑥) < 𝜖
for 𝑛 sufficiently large. Since ℎ 𝑛 𝑥 ≠ 𝑥 for all 𝑛, this contradicts that the orbit 𝐺𝑥 is
discrete, having no limit points.
34.6. SYMMETRIC SPACE MODEL 625
Remark 34.5.2. The hypothesis “𝑋 is a metric space with 𝐺 acting by isometries”

providing the equivalent condition Theorem 34.5.1(v) is necessary: see Exercise
34.11.
34.5.3. From Lemma 34.3.6 and the implication Theorem 34.5.1(v) ⇒ (i), we see that
proper actions generalize covering space actions when 𝑋 is locally compact metric
space and 𝐺 acts by isometries. In fact, a more general statement is true: if 𝐺 is a
discrete group with a covering space action on 𝑋 such that 𝐺\𝑋 is Hausdorff, then 𝐺
acts properly on 𝑋. The (slightly involved) proof in general is requested in Exercise
34.16.
Remark 34.5.4. Bourbaki discusses proper maps [Bou60, Chapter I, §10] and more
generally groups acting properly on topological spaces [Bou60, Chapter III, §§1,4];
the definition of proper is equivalent to ours as follows. Let 𝑓 : 𝑋 → 𝑌 be continuous,
and say 𝑓 is Bourbaki proper to mean that 𝑓 × id : 𝑋 × 𝑍 → 𝑌 × 𝑍 is closed for every
topological space 𝑍. If 𝑓 is Bourbaki proper, then 𝑓 is proper [Bou60, Chapter I, §10,
Proposition 6]. In the other direction, if 𝑓 is proper then 𝑓 is closed and 𝑓 −1 (𝑦) is
compact for all 𝑦 ∈ 𝑌 , and this implies that 𝑓 is Bourbaki proper [Bou60, Chapter I,
§10, Theorem 1].
34.6 Symmetric space model

In this section, before proceeding with our treatment of discrete group actions in our
case of interest, we pause to give a very important way to think about hyperbolic space
in terms of symmetric spaces. The magical formulas in hyperbolic geometry beg for a
more conceptual explanation: what is their provenance? Although it is important for
geometric intuition to begin with a concrete model of hyperbolic space and ask about
its isometries directly, from this point of view it is more natural to instead start with
the desired group and have it act on itself in a natural way.
34.6.1. Let 𝐺 = SL2 (R). As a matrix group, 𝐺 comes with a natural metric. The
space M2 (R) ' R4 has the usual structure of a metric space, with

2 2 2 2 2 𝑎 𝑏
k𝑔k = 𝑎 + 𝑏 + 𝑐 + 𝑑 , if 𝑔 = ∈ M2 (R).
𝑐 𝑑
We give SL2 (R) ⊂ M2 (R) the subspace metric and PSL2 (R) the quotient metric.
Intuitively, in this metric 𝑔, ℎ ∈ PSL2 (R) are close if there exist matrices representing
𝑔, ℎ (corresponding to a choice of sign) with all four entries of the matrix close in R.
34.6.2. Recall from 34.1.12 that if 𝐺 acts (continuously and) transitively on 𝑋, then
for all 𝑥 ∈ 𝑋, the natural map 𝑔 ↦→ 𝑔𝑥 gives a continuous bĳection
∼ 𝐺𝑥 = 𝑋.
𝐺/Stab𝐺 (𝑥) −→
Let 𝑋 = H2 be the hyperbolic plane and let 𝐺 = SL2 (R). Then 𝐺 acts transitively on
𝑋. The stabilizer of 𝑥 = 𝑖 is the subgroup 𝐾 = Stab𝐺 (𝑥) = SO(2) ≤ SL2 (R), so there
is a continuous bĳection
∼ H2 = 𝑋
𝐺/𝐾 = SL2 (R)/SO(2) −→
(34.6.3)
𝑔𝐾 ↦→ 𝑔𝑖.
From the Iwasawa decomposition (Proposition 33.4.2), it follows that
SL2 (R)/SO(2) ' 𝑁 𝐴. (34.6.4)
In fact, the map (34.6.3) is a homeomorphism. To prove this, we observe the

following beautiful equation: for 𝑔 ∈ SL2 (R),
k𝑔k 2 = 2 cosh 𝜌(𝑖, 𝑔𝑖). (34.6.5)
This formula follows directly from the formula (33.5.3) for distance; the calculation is
requested in Exercise 34.17. It follows that the map 𝐺 → 𝑋 is open, and thus (34.6.3)
is a homeomorphism. In fact, by (34.6.5), if we reparametrize the metric on either
SL2 (R)/SO(2) or H2 by the appropriate factor involving the hyperbolic cosine, the
map (34.6.3) becomes an isometry.
We conclude this section with a view to a more general setting where the above
situation applies.
34.6.6. Let 𝐺 be a connected, Hausdorff, locally compact topological group. We recall

(section 29.3) that 𝐺 has a Haar measure, a Borel measure 𝜇 that is left-translation
invariant (so 𝜇(𝑔 𝐴) = 𝜇( 𝐴) for all Borel sets 𝐴 ⊆ 𝐺 and 𝑔 ∈ 𝐺). The Haar measure is
unique up to scaling, with the Haar measure on 𝐺 = R𝑛 the usual Lebesgue measure.
𝐺 has a maximal compact subgroup 𝐾 ≤ 𝐺, unique up to conjugation in 𝐺,
and the quotient 𝑋 = 𝐺/𝐾 is homeomorphic to Euclidean space—in particular, 𝑋 is
contractible.
A lattice Γ ≤ 𝐺 is a discrete subgroup such that 𝜇(Γ\𝐺) < ∞. A lattice Γ acts
properly on 𝑋 by left multiplication.
Remark 34.6.7. More generally, a (globally) symmetric space is a space of the form
𝐺/𝐾 where 𝐺 is a Lie group and 𝐾 ≤ 𝐺 a maximal compact subgroup. Alternatively,
it can be defined as a space where every point has a neighborhood where there is an
isometry of order 2 fixing the point. For more reading on the theory of symmetric
spaces, and the connection to differential geometry and Lie groups, see the book by
Helgason [Hel2001].
34.7 Fuchsian groups

We now specialize to our case of interest and consider the group PSL2 (R) acting by
isometries on the geodesic space H2 . A gentle introduction to the geometry of discrete
groups is provided by Beardon [Bea95], with a particular emphasis on Fuchsian groups
and their fundamental domains—in the notes at the end of each chapter are further
bibliographic pointers. See also Jones–Singerman [JS87].
34.7. FUCHSIAN GROUPS 627
Lemma 34.7.1. Let Γ ≤ SL2 (R). Then the following are equivalent:
(i) Γ is discrete;
(ii) If 𝛾𝑛 ∈ Γ and 𝛾𝑛 → 1, then 𝛾𝑛 = 1 for almost all 𝑛; and
(iii) For all 𝑀 ∈ R>0 , the set {𝛾 ∈ Γ : k𝛾k ≤ 𝑀 } is finite.
Proof. The equivalence (i) ⇔ (ii) is requested in Exercise 34.15. The implication (i)
⇔ (iii) follows from the fact that the ball of radius 𝑀 in SL2 (R) is a compact subset of
M2 (R), and a subset of a compact set is finite if and only if it is discrete. Slightly more
elaborately, a sequence of matrices with bounded norm has a subsequence where the
entries all converge; since the determinant is continuous, the limit exists in SL2 (R) so
Γ is not discrete.
In particular, we find from Lemma 34.7.1 that a discrete subgroup of SL2 (R) is
countable.
Proposition 34.7.2. Let Γ ≤ PSL2 (R) be a subgroup (with the subspace topology).
Then Γ has a wandering action on H2 if and only Γ is discrete.
Proof. The implication ⇒ is a consequence of Theorem 34.5.1(v) ⇒ (i). Conversely,

suppose that Γ is discrete; we show that Theorem 34.5.1(vi) holds: that for all 𝑥 ∈ 𝑋,
the orbit 𝐺𝑥 ⊆ 𝑋 is discrete and # Stab𝐺 (𝑥) < ∞.
First we show that the stabilizer of a point is finite. We may work in the unit
disc D2 and take the point to be 𝑤 = 0 ∈ D2 , as in 33.7.8. The stabilizer of 𝑤 = 0
in SU(1, 1) is SO(2) ' R/(2𝜋)Z, so its stabilizer in Γ is a discrete subgroup of the
compact group SO(2) and is necessarily finite (indeed, cyclic).
Next we show that orbits of Γ on H2 are discrete. We apply the identity (34.6.5).
This identity with Lemma 34.7.1 shows that the orbit Γ𝑖 is discrete. But for all
𝑧 ∈ H2 , there exists 𝜙 ∈ PSL2 (R) such that 𝜙(𝑖) = 𝑧, and conjugation by 𝜙 induces an
isomorphism Γ −→ ∼ 𝜙−1 Γ𝜙 of topological groups. Since
𝜌(𝑧, 𝑔𝑧) = 𝜌(𝜙(𝑖), 𝑔𝜙(𝑖)) = 𝜌(𝑖, (𝜙−1 𝑔𝜙)𝑖)
applying the above argument to 𝜙−1 Γ𝜙 shows that the orbit Γ𝑧 is discrete. This
concludes the proof.
Alternatively, here is a self-contained proof that avoids the slightly more involved
topological machinery. We again work in the unit disc D2 . First we prove (⇐).
Since Γ is discrete, there is an 𝜖-neighborhood 𝑈 3 1 with 𝑈 ⊆ PSU(1, 1) such that
𝑈 ∩ Γ = {1}; therefore, if

𝑎 𝑏
𝛾= ∈ Γ \ {1}
𝑏 𝑎
then | 𝑏| > 𝜖 or (without loss of generality) |𝑎 − 1| > 𝜖. We claim that in either case

𝑏
|𝛾(0)| = > 𝜖,
𝑎
and thus the orbit is discrete. Indeed, if | 𝑏| > 𝜖, then since |𝑎| < 1 anyway immediately
| 𝑏/𝑎| > 𝜖; if |𝑎 − 1| > 𝜖 then |𝑎| < 1 − 𝜖 so |𝑎| 2 < 1 − 𝜖 2 and 1/|𝑎| 2 > 1 + 𝜖 2 , giving
2 2
= 1 − |𝑎| > (1 + 𝜖 2 ) − 1 = 𝜖 2 .
𝑏
𝑎 |𝑎| 2
For (⇒), suppose that Γ is not discrete; then there is a sequence 𝛾𝑛 ∈ Γ \ {1} of
elements such that 𝛾𝑛 → 1. Therefore, for all 𝑧 ∈ H2 , we have 𝛾𝑛 𝑧 → 𝑧 and 𝛾𝑛 𝑧 = 𝑧
for only finitely many 𝑛, so every neighborhood of 𝑧 contains infinitely many distinct
points 𝛾𝑛 𝑧.
With this characterization, we make the following important definition.
Definition 34.7.3. A Fuchsian group is a discrete subgroup of PSL2 (R).
A Fuchsian group Γ acts by orientation-preserving isometries on H2 ; this action is

proper and wandering by Theorem 34.5.1.
34.7.4. A Fuchsian group Γ ≤ PSL2 (R) is elementary if there is a nonempty Γ-

invariant set in H2 ∪bd H2 that contains at most two points. Equivalently, an elementary
group is a cyclic subgroup or a (possibly) dihedral group—in particular, an elementary
group is virtually abelian (has a finite index, abelian subgroup). The elementary
groups are easy to analyze, but their inclusion into theorems about more general
Fuchsian groups can cause problems; and so in general we are only interested in
non-elementary groups.
Non-elementary Fuchsian groups Γ are categorized by the set of limit points
𝐿 (Γ) ⊆ bd H2 of Γ𝑧 with 𝑧 ∈ H2 . If 𝐿 (Γ) = bd H2 , then Γ is said to be a Fuchsian
group of the first kind; otherwise Γ is of the second kind, and 𝐿 (Γ) is a nowhere-
dense perfect subset of bd H2 , topologically a Cantor set. We will see later that if Γ
has quotient with finite hyperbolic area, then Γ is finitely generated of the first kind.
34.8 Riemann uniformization and orbifolds

Our understanding of group actions has an important consequence for the classification
of Riemann surfaces, and we pause (again) to provide this application.
First, we have the important structural result.
Theorem 34.8.1 (Riemann uniformization theorem). Every (connected and) simply

connected Riemann surface 𝐻 is isomorphic to either the Riemann sphere P1 (C), the
complex plane C, or the hyperbolic plane H2 .
A consequence of Riemann uniformization is as follows.
34.8.2. The universal cover 𝑋e of a compact Riemann surface 𝑋 is simply connected, so

by the theory of covering spaces, 𝑋 is a quotient 𝑋 ' Γ\ 𝑋
e where Γ is the fundamental
group of 𝑋, a subgroup of isometries of 𝑋 e acting by a covering space action.
When 𝑋 e = P1 (C), the only possible group Γ (acting freely) is trivial. When 𝑋
e = C,
by classification one sees that the only Riemann surfaces of the form 𝑋 = C/Γ are the
34.8. RIEMANN UNIFORMIZATION AND ORBIFOLDS 629
plane 𝑋 = C, the punctured plane C× ' C/h𝑢i with 𝑢 ∈ C× , and complex tori C/Λ
where Λ ⊂ C is a lattice with Λ ' Z2 . We will embark on a classification of these tori
up to isomorphism by their 𝑗-invariants in section 40.1.
All other Riemann surfaces are hyperbolic with 𝑋 e = H2 , and so are of the form
2
𝑋 = Γ\H with Γ ≤ PSL2 (R) a torsion-free Fuchsian group.
Remark 34.8.3. Klein and Poincaré conjectured the uniformization theorem for alge-
braic curves over C, with rigorous proofs were given by Poincaré: see Gray [Gray94].
Finally, before departing our topological treatment, we consider quotients of man-
ifolds by the (continuous) action of a group. As we have seen, it is quite restrictive
to suppose that this group action is free: we will still want to take quotients by such
groups. Accordingly, we need to model not spaces that are locally modelled by R𝑛 but
those that are locally modelled by the quotient of R𝑛 by a finite group.
Definition 34.8.4. An 𝑛-orbifold 𝑋 is a (second-countable) Hausdorff topological
space that is locally homeomorphic to a quotient 𝐺\R𝑛 with 𝐺 a finite group acting
(continuously). An atlas for an orbifold 𝑋 is the data
(i) An open cover {𝑈𝑖 }𝑖 ∈𝐼 of charts 𝑈𝑖 closed under finite intersection; and
(ii) For each 𝑖 ∈ 𝐼, an open subset 𝑉𝑖 ⊆ R𝑛 equipped with the (continuous) action
of a finite group 𝐺 𝑖 𝑉𝑖 , and a homeomorphism
∼ 𝐺 \𝑉
𝜙𝑖 : 𝑈𝑖 −→ 𝑖 𝑖
satisfying the atlas axiom: for all 𝑈𝑖 ⊆ 𝑈 𝑗 , there exists an injective group homomor-
phism 𝑓𝑖 𝑗 : 𝐺 𝑖 ↩→ 𝐺 𝑗 and a 𝐺 𝑖 -equivariant map 𝜓𝑖 𝑗 : 𝑉𝑖 ↩→ 𝑉 𝑗 satisfying 𝜙 𝑗 ◦𝜓𝑖 𝑗 = 𝜙𝑖
(see Figure 34.8.5).
Ui
Uj
φj
∼
φi
∼
Vi /Gi
Vj /Gj
ψij
Vi Vj
Gi Gj
Figure 34.8.5: An orbifold, by its atlas

Orbifolds were introduced by Thurston [Thu97, Chapter 13], who adds a wealth
of motivation and examples; see also the surveys by Scott [Sco83, §2] and Gordon
[Gor2012] as well as the chapter by Ratcliffe [Rat2006, Chapter 13].
34.8.6. We can further ask that the transition maps 𝑓𝑖 𝑗 in an atlas be smooth to get
a smooth orbifold, preserve a 𝐺 𝑖 -Riemannian metric to get a Riemann orbifold,
etc.; replacing R𝑛 by C𝑛 and smooth by holomorphic, we similarly define a complex
𝑛-orbifold, locally modelled on the quotient 𝐺\C𝑛 with 𝐺 a finite group acting
holomorphically.
Definition 34.8.7. Let 𝑋 be an 𝑛-orbifold.
(a) A point 𝑧 ∈ 𝑋 such that there exists a chart 𝑈𝑖 3 𝑧 with group 𝐺 𝑖 ≠ {1} fixing 𝑧
is called an orbifold point of 𝑋, with stabilizer group (or isotropy group) 𝐺 𝑖 ;
the set of orbifold points of 𝑋 is called the orbifold set of 𝑋.
(b) If 𝑧 ∈ 𝑈𝑖 is an isolated orbifold point and its stabilizer group is cyclic, we call 𝑧
a cone point.
34.8.8. The prototypical example of an orbifold is the quotient of C by a finite group of

rotations. Such a group is necessarily cyclic (as a finite subgroup of C× ) of some order
𝑚 ≥ 2; the quotient is a cone, a fundamental set for the action being a segment with
angle 2𝜋/𝑚, and the fixed point is a cone point of order 𝑚. The cone is homeomorphic
to R2 but it is not isometric: away from the cone point, this space is locally isometric
to R2 , but at the cone point the angle is less than 2𝜋, so shortest paths that do not start
or end at the cone point never go through the cone point.
34.8.9. Let 𝑋 be a manifold and let 𝐺 be a finite group acting (continuously) on 𝑋
such that action of 𝐺 is wandering (Definition 34.3.3). We define an orbifold [𝑋/𝐺]
as follows: by Lemma 34.3.6 and 34.3.9, we can refine an atlas of 𝑋 to one consisting
of open neighborhoods 𝑈𝑖 on which 𝐺 𝑈𝑖 acts, and we make this into an orbifold
atlas by taking 𝐺 𝑖 = 𝐺 for each 𝑖; the atlas axiom is tautologically satisfied.
When 𝑋 is smooth, complex, Riemann, etc., we ask that 𝐺 act diffeomorphically,
holomorphically, etc., to obtain an orbifold with the same properties.
A full, suitable definition of the category of orbifolds—in particular, morphisms
between them—is more subtle than it may seem. In this text, we will be primarily
interested in an accessible and well-behaved class of orbifolds obtained as the quotient
of a manifold.
Definition 34.8.10. An orbifold is good if is of the form [𝑋/𝐺], i.e., it arises as the
quotient of a manifold by a finite group.
34.8.11. The quotient [𝑋/𝐺] of a Riemannian manifold 𝑋 by a discrete group 𝐺 of
isometries acting properly is a good Riemann orbifold.
Example 34.8.12. A complex 1-orbifold is good if and only if it has a branched cover
by a Riemann surface. By Exercise 34.18, the only complex 1-orbifolds that are not
good are the teardrop, a sphere with one cone point, and the football, a sphere with
two cone points of different orders.
34.8.13. Good (topological) compact, oriented 2-orbifolds admit a classification (ex-

tending the usual classification of surfaces by genus) up to homeomorphism by their
signature (𝑔; 𝑒 1 , . . . , 𝑒 𝑘 ), where 𝑔 is the genus of the underlying topological surface
and the 𝑒 1 , . . . , 𝑒 𝑘 are the orders of the (necessarily cyclic) nontrivial stabilizer groups.
34.8.14. Putting these two pieces together, now let Γ ≤ PSL2 (R) be a Fuchsian group.
Then the discrete group Γ H2 acts properly. Then Γ\H2 has the structure of a good
complex 1-orbifold, by the main theorem (Theorem 34.2.1(ii)).
Remark 34.8.15. From certain topological points of view, especially with an eye
towards generalizations, an orbifold is best understood as a topological groupoid (a
point of view first noticed by Haefliger [Hae84, §4]): the objects of the category are
elements of the disjoint union of the charts 𝑈𝑖 , and a morphism from 𝑢 𝑖 ∈ 𝑈𝑖 to 𝑢 𝑗 ∈ 𝑈 𝑗
is the germ of a local homeomorphism that commutes with the projections. For more
on the categorical perspective of orbifolds as groupoids, see Moerdĳk [Moe2002] and
Moerdĳk–Pronk [MP1997].
Exercises
Unless otherwise specified, let 𝐺 𝑋 be an action of a group 𝐺 on a set 𝑋.
1. Show that the quotient map 𝜋 : 𝑋 → 𝐺\𝑋 is defined by a universal property:

if 𝑓 : 𝑋 → 𝑍 is a 𝐺-equivariant map where 𝑍 has a trivial 𝐺-action, then 𝑓
factors uniquely through 𝜋, i.e., there exists a unique map ℎ : 𝐺\𝑋 → 𝑍 making
the diagram
𝑓
/𝑍
𝑋 =
𝜋
ℎ
𝐺\𝑋
commute.
I 2. Prove Lemma 34.1.10, in the following form. Let 𝐺 be a topological group
acting on a topological space 𝑋.
(a) Show that if the action is continuous, then for all 𝑔 ∈ 𝐺 the map 𝑋 → 𝑋
by 𝑥 ↦→ 𝑔𝑥 is continuous (therefore, a homeomorphism).
(b) Show the converse of (a) if 𝐺 is discrete.
3. Let 𝐺 be a topological group acting continuously on a topological space 𝑋.
Show that the orbits of 𝐺 are closed (𝐺𝑥 ⊆ 𝑋 is closed for all 𝑥 ∈ 𝑋) if and only
if 𝐺\𝑋 is T1 (see Exercise 12.2).
4. Let 𝑋 be a metric space. Then Isom(𝑋) has naturally the topology of pointwise
convergence, as follows. There is an embedding
Ö
Isom(𝑋) ↩→ 𝑋 𝑋 = 𝑋
𝑥 ∈𝑋
𝑔 ↦→ (𝑔(𝑥)) 𝑥 ∈𝑋 .
The product 𝑋 𝑋 has the product topology, and so Isom(𝑋) (and every space of
maps from 𝑋 to 𝑋) has an induced subspace topology. A basis of open sets for
Isom(𝑋) in this topology consists of finite intersections of open balls
𝑉 (𝑔; 𝑥, 𝜖) = {ℎ ∈ Isom(𝑋) : 𝜌(𝑔(𝑥), ℎ(𝑥)) < 𝜖 }.
Equip the group 𝐺 = Isom(𝑋) with the topology of pointwise convergence.
(a) Show that 𝐺 is a topological group.
(b) Show that 𝐺 acts continuously on 𝑋.
5. Let 𝐺 = Z be given the discrete topology, and let 𝐺 𝑋 = R/Z act by
𝑥 ↦→ 𝑥 + 𝑛𝑎 ∈ R/Z for 𝑛 ∈ Z for 𝑎 ∈ R − Q. Show that this action is free and
continuous, and show that for all 𝑥 ∈ 𝑋 the map (34.1.13)
𝐺/Stab𝐺 (𝑥) = 𝐺 → 𝐺𝑥
𝑔 ↦→ 𝑔𝑥
is (continuous and bĳective but) not a homeomorphism, giving 𝐺𝑥 ⊆ 𝑋 the
subspace topology.
I 6. Let 𝐺 act (continuously and) transitively on 𝑋. Suppose that 𝐺, 𝑋 are (Hausdorff
and) locally compact, and suppose further that 𝐺 has a countable base of open
sets. Let 𝑥 ∈ 𝑋 and let 𝐾 = Stab𝐺 (𝑥). Show that the natural map 𝐺/𝐾 → 𝑋 is
a homeomorphism.
7. Let 𝐺 𝑋 be a free and wandering action, and let 𝑈 be an open set such
that 𝑔𝑈 ∩ 𝑈 = ∅ for all 𝑔 ≠ 1. Show that the map 𝐺 × 𝑈 → 𝜋 −1 (𝜋(𝑈)) is a
homeomorphism and the restriction 𝜋 : 𝐺 × 𝑈 → 𝜋(𝑈) ' 𝑈 is a (split) covering
map.
8. Let 𝑋 be (Hausdorff and) locally compact, let 𝑥 ∈ 𝑋, and let 𝑈 3 𝑥 be an open
neighborhood. Show that there exists an open neighborhood 𝑉 3 𝑥 such that
𝐾 = cl(𝑉) ⊆ 𝑈 is compact.
I 9. Let 𝑋, 𝑌 be (Hausdorff) topological spaces, let 𝑓 : 𝑋 → 𝑌 be a continuous map,
and let
gr( 𝑓 ) : 𝑋 → 𝑋 × 𝑌
𝑥 ↦→ (𝑥, 𝑓 (𝑥))
be the graph of 𝑓 . Show that 𝑓 is a closed map.
10. One way to weaken the running hypothesis that 𝑋 is Hausdorff in this chapter
is to instead assume only that 𝑋 is locally Hausdorff: every 𝑥 ∈ 𝑋 has an open
neighborhood 𝑈 3 𝑥 such that 𝑈 is Hausdorff.
Show that a weakened version of Lemma 34.3.6(i) ⇒ (ii) is not true with only
the hypothesis that 𝑋 is locally Hausdorff: that is, exhibit a locally Hausdorff
topological space 𝑋 with a (continuous) wandering action of a group 𝐺 such
that 𝜋 : 𝑋 → 𝐺\𝑋 is not a local homeomorphism, and so Lemma 34.3.6(ii) does
not hold. [Hint: Let 𝑋 be the bug-eyed line and 𝐺 ' Z/2Z acting by 𝑥 ↦→ −𝑥
on R× and swapping points in the doubled origin.]
11. Let 𝐺 = Z and let 𝐺 ↩→ 𝑋 = R2 r {(0, 0)} act by 𝑛 · (𝑥, 𝑦) = (2𝑛 𝑥, 𝑦/2𝑛 ).
In other words, 𝐺 is the group of continuous maps 𝑋 → 𝑋 generated by
(𝑥, 𝑦) ↦→ (2𝑥, 𝑦/2).
(a) Show that the action of 𝐺 on 𝑋 is free and wandering.
(b) Show that the quotient 𝐺\𝑋 is not Hausdorff.
(c) Let 𝐾 = {(𝑡, 1−𝑡) : 𝑡 ∈ [0, 1]}. Then 𝐾 is compact. Show that 𝐾 ∩ 𝑔𝐾 ≠ ∅
for infinitely many 𝑔 ∈ 𝐺. [So Theorem 34.5.1(v) holds but (ii) does not,
and in particular that the action of 𝐺 is not proper. Can you see this directly
from the definition of proper?]
12. Let 𝑋 be a Hausdorff topological space with a continuous action of a Hausdorff
topological group 𝐺. Suppose that the action of 𝐺 is wandering. Show that
for all 𝑥 ∈ 𝑋, there is an open neighborhood 𝑈 3 𝑥 such that the finite group
Stab𝐺 (𝑥) acts on 𝑈 (i.e., 𝑔𝑈 ⊆ 𝑈 for all 𝑔 ∈ Stab𝐺 (𝑥)).
13. Show that a subgroup Γ ≤ R𝑛 is discrete if and only if Γ = Z𝑣 1 + · · · + Z𝑣 𝑚
with 𝑣 1 , . . . , 𝑣 𝑚 ∈ Γ linearly independent over R. As a consequence, show that
Γ ≤ R𝑛 is a lattice if and only if Γ is discrete with 𝑚 = 𝑛.
14. Exhibit an injective group homomorphism SO(𝑛) ↩→ SO(𝑛 + 1) and a homeo-
morphism
S𝑛 ' SO(𝑛 + 1)/SO(𝑛),
where S𝑛 = {𝑥 ∈ R𝑛+1 : k𝑥k 2 = 1} is the 𝑛-dimensional sphere, analogous to
(34.6.3).
I 15. Let 𝐺 be a topological group with a countable system of fundamental open
neighborhoods of 1 ∈ 𝐺 (for example, this holds if 𝐺 is metrizable). Show that
𝐺 is discrete if and only if whenever {𝑔𝑛 } 𝑛 is a sequence from 𝐺 with 𝑔𝑛 → 1,
then 𝑔𝑛 = 1 for all but finitely many 𝑛.
16. Let 𝐺 be a discrete group with a (continuous) covering space action on a
Hausdorff space 𝑋 such that 𝐺\𝑋 is Hausdorff. Show that 𝐺 acts quasi-properly
on 𝑋.
I 17. Show that for 𝑔 ∈ SL2 (R),
k𝑔k 2 = 2 cosh 𝜌(𝑖, 𝑔𝑖)
(cf. 34.6.1). [Hint: Use the formula (33.5.2).]

18. (a) Show that a compact, complex 1-orbifold is good if and only if it has a
branched cover by a compact Riemann surface.
(b) Use the Riemann–Hurwitz theorem to show that the only compact, complex
1-orbifolds that are not good are the teardrop (a sphere with one cone point)
and the football (a sphere with two cone points of different orders).
(c) Show that every finitely generated discrete group of isometries of a simply
connected Riemann surface with compact quotient has a torsion free sub-
group of finite index. [Hint: find a torsion free subgroup of finite index by
avoiding the finitely many conjugacy classes of torsion in Γ.] Use this to
give another proof of (b).
19. Show that the stabilizer group of an orbifold point is well-defined up to group
isomorphism, independent of the chart.
20. The group SL2 (R) acts on P1 (R) = R∪ {∞} by linear fractional transformations.
Show that 𝐺 = SL2 (Z) ≤ SL2 (R) is discrete, but 𝐺 does not act properly on
P1 (R). [So discrete groups can act on locally compact spaces without necessarily
acting properly.]
Chapter 35
Classical modular group
In this chapter, we introduce the classical modular group PSL2 (Z) ≤ PSL2 (R), a
discrete group acting on the upper half-plane that has received extensive study because
of the role it plays throughout mathematics. We examine the group in detail via a
fundamental domain and conclude with some applications to number theory. This
chapter will serve as motivation and example for the generalizations sought later in
this part of the text.
There are very many references for the classical modular group, including Apostol
[Apo90, Chapter 2], Diamond–Shurman [DS2005, Chapter 2], and Serre [Ser73,
Chapter VII].
35.1 ⊲ The fundamental set

Definition 35.1.1. The classical modular group is the subgroup of PSL2 (R) defined
by

𝑎 𝑏
PSL2 (Z) = 𝛾 = : 𝑎, 𝑏, 𝑐, 𝑑 ∈ Z, 𝑎𝑑 − 𝑏𝑐 = 1 /{±1}.
𝑐 𝑑
The group PSL2 (Z) acts faithfully on the upper half-plane H2 by linear fractional
transformations; equipping H2 with the hyperbolic metric, this action is by orientation-
preserving isometries.
Since Z ⊆ R is discrete, so too is SL2 (Z) ⊆ M2 (Z) ⊆ M2 (R) discrete and therefore
PSL2 (Z) ≤ PSL2 (R) is a Fuchsian group (Definition 34.7.3).
35.1.2. Our first order of business is to try to understand the structure of the classical
modular group in terms of this action. Let

0 −1 1 1
𝑆 := , 𝑇 := ∈ PSL2 (Z).
1 0 0 1
Then 𝑆𝑧 = −1/𝑧 for 𝑧 ∈ H2 , so 𝑆 maps the unit circle {𝑧 ∈ C : | 𝑧| = 1} to itself, fixing

the point 𝑧 = 𝑖; and 𝑇 𝑧 = 𝑧 + 1 for 𝑧 ∈ H2 acts by translation. We compute that 𝑆 2 = 1
635
636 CHAPTER 35. CLASSICAL MODULAR GROUP
(in PSL2 (Z)) and

0 −1
𝑆𝑇 =
1 1
so (𝑆𝑇) 3 = 1.
35.1.3. In a moment, we will see that PSL2 (Z) is generated by 𝑆 and 𝑇, with a minimal
set of relations given by 𝑆 2 = (𝑆𝑇) 3 = 1. To do so, we examine a fundamental set (cf.
Definition 34.1.14) for the action of PSL2 (Z) on H2 , as follows. Let
◊ := {𝑧 ∈ H2 : |Re 𝑧| ≤ 1/2 and | 𝑧| ≥ 1}.
√ 2
√ ◊ is a hyperbolic triangle with vertices at 𝜔 = (−1 + −3)/2 and −𝜔 =
The set
(1 + −3)/2 and ∞. The translates of ◊ by words in 𝑆, 𝑇 tessellate the plane as in
Figure 35.1.4.
T −1 1 T
T −1 S S TS
ST S ST ST −1 ST −1 S
Figure 35.1.4: Tessellation of H2 into fundamental domains for SL2 (Z)

By the Gauss–Bonnet formula 33.6.8 (or Exercise 35.1),
𝜋 𝜋
area(◊) = 𝜋 − 2 = . (35.1.5)
3 3
The elements 𝑆, 𝑇 act on the edges of this triangle as in Figure 35.1.6.
i
ω −ω 2 = −ω
S
Figure 35.1.6: Action of 𝑆, 𝑇 on ◊

35.1. ⊲ THE FUNDAMENTAL SET 637
In the unit disc, the triangle ◊ is as in Figure 35.1.7.
z7→w= z−i
z+i
−−−−−−−→
Figure 35.1.7: ◊ in D2
The following three lemmas describe the relationship of the set ◊ to Γ.
Lemma 35.1.8. For all 𝑧 ∈ H2 , there exists a word 𝛾 ∈ h𝑆, 𝑇i such that 𝛾𝑧 ∈ ◊.
Proof. In fact, we can determine such a word algorithmically. First, we translate 𝑧 so

that |Re 𝑧| ≤ 1/2. If | 𝑧| ≥ 1, we are done; otherwise, if | 𝑧| < 1, then

−1 Im 𝑧
Im = > Im 𝑧. (35.1.9)
𝑧 | 𝑧| 2
We then repeat this process, obtaining a sequence of elements 𝑧 = 𝑧 1 , 𝑧2 , . . . with
Im 𝑧1 < Im 𝑧2 < . . . . We claim that this process terminates after finitely many steps.
Indeed, by (33.3.9)

Im 𝑧 𝑎 𝑏
Im(𝑔𝑧) = , for 𝑔 = ∈ PSL2 (R),
|𝑐𝑧 + 𝑑| 2 𝑐 𝑑
and the number of 𝑐, 𝑑 ∈ Z such that |𝑐𝑧 + 𝑑| < 1 is finite: the set Z + Z𝑧 ⊆ C is a
lattice, so there are only finitely many elements of bounded norm. (Alternatively, the
orbit Γ𝑧 is discrete by Theorem 34.5.1—or the direct argument given in Proposition
34.7.2—therefore, its intersection with the compact set
𝐾 = {𝑧 0 ∈ H2 : | Re(𝑧 0)| ≤ 1/2 and Im 𝑧 ≤ Im 𝑧 0 ≤ 1}
is finite.) Upon termination, we have found a word 𝛾 in 𝑆, 𝑇 such that 𝛾𝑧 ∈ ◊.
The procedure exhibited in the proof of Lemma 35.1.8 is called a reduction

algorithm.
Lemma 35.1.10. Let 𝑧, 𝑧 0 ∈ ◊, and suppose 𝑧 ∈ int(◊) lies in the interior of ◊. If
𝑧 0 = 𝛾𝑧 with 𝛾 ∈ Γ, then 𝛾 = 1 and 𝑧 = 𝑧 0.

0 𝑎 𝑏
Proof. Let 𝑧 = 𝛾𝑧 with 𝛾 = ∈ Γ. We have Im 𝑧 0 = (Im 𝑧)/|𝑐𝑧 + 𝑑| 2 . First
𝑐 𝑑
suppose that Im 𝑧 0 ≥ Im 𝑧; then
|𝑐𝑧 + 𝑑| 2 = (𝑐 Re 𝑧 + 𝑑) 2 + 𝑐2 (Im 𝑧) 2 ≤ 1. (35.1.11)

√
Since Im 𝑧 > Im 𝜔 = 3/2, from (35.1.11) we conclude that 𝑐2 ≤ 4/3 so |𝑐| ≤ 1.
If 𝑐 = 0 then 𝑎𝑑 − 𝑏𝑐 = 𝑎𝑑 = 1 so 𝑎 = 𝑑 = ±1, and then 𝑧 0 = 𝛾𝑧 = 𝑧 ± 𝑏, which
immediately implies 𝑏 = 0 so 𝛾 = 1 as claimed. If instead |𝑐| = 1, then the conditions
(𝑐 Re 𝑧 + 𝑑) 2 ≤ 1 − (Im 𝑧) 2 ≤ 1 − 3/4 = 1/4 and |Re 𝑧| < 1/2
together imply 𝑑 = 0; but then |𝑐𝑧 + 𝑑| = | 𝑧| ≤ 1, and since 𝑧 ∈ int(◊) we have | 𝑧| > 1,
a contradiction.
If instead Im 𝑧 0 < Im 𝑧, we interchange the roles of 𝑧, 𝑧 0 and have strict inequality
in (35.1.11); by the same argument and the weaker inequality |Re 𝑧| ≤ 1/2, we then
obtain | 𝑧| < 1, a contradiction.
Lemma 35.1.12. The elements 𝑆, 𝑇 generate Γ = PSL2 (Z).
Proof. Let 𝑧 = 2𝑖 ∈ int(◊). Let 𝛾 ∈ Γ, and let 𝑧 0 = 𝛾𝑧. By Lemma 35.1.8, there exists
𝛾 0 a word in 𝑆, 𝑇 such that 𝛾 0 𝑧 0 ∈ ◊. By Lemma 35.1.10, we have 𝛾 0 𝑧 0 = (𝛾 0 𝛾)𝑧 = 𝑧,
so 𝛾 0 𝛾 = 1 and 𝛾 = 𝛾 0 ∈ h𝑆, 𝑇i.
Although we have worked in PSL2 (Z) throughout, it follows from Lemma 35.1.12
that the matrices 𝑆, 𝑇 also generate SL2 (Z), since 𝑆 2 = −1. See Exercise 35.3 for
another proof of Lemma 35.1.12.
Corollary 35.1.13. The set ◊ is a fundamental set for PSL2 (Z) H2 .
Proof. The statement follows from Lemmas 35.1.8 and 35.1.10 (recalling the definition
of fundamental set, Definition 34.1.14).
35.1.14. If 𝑧 ∈ ◊ has StabΓ (𝑧) ≠ {1}, then we claim that one of the following holds:
(i) 𝑧 = 𝑖, and StabΓ (𝑖) = h𝑆i ' Z/2Z;

(ii) 𝑧 = 𝜔, and StabΓ (𝜔) = h𝑆𝑇i ' Z/3Z; or
(iii) 𝑧 = −𝜔2 , and StabΓ (−𝜔2 ) = h𝑇 𝑆i = 𝑇 StabΓ (𝜔)𝑇 −1 .

𝑎 𝑏
Indeed, let 𝛾𝑧 = 𝑧 with 𝛾 = and 𝛾 ≠ 1. Then 𝑐𝑧2 + (𝑑 − 𝑎)𝑧 − 𝑏 = 0, so 𝑐 ≠ 0
𝑐 𝑑
and √
(𝑎 − 𝑑) + 𝐷
𝑧= where 𝐷 = Tr(𝛾) 2 − 4 ∈ Z<0 .
2𝑐
√
Thus 𝐷 = −4 or 𝐷 = −3. In either case, since 𝑧 ∈ ◊ we have Im 𝑧 ≥ 3/2, we
must have 𝑐 = ±1, and replacing 𝛾 ← −𝛾 we may take 𝑐 = 1. If 𝐷 = −4, then
Tr(𝛾) = 𝑎 + 𝑑 = 0 so 𝑧 = 𝑎 + 𝑖, and 𝑎 = 0 = 𝑑 and 𝑐 = 1 = −𝑏, i.e., 𝑧 = 𝑖 and we √ are in
case (i). If the discriminant is −3, then a similar argument gives 𝑧 = ((𝑎 ± 1) + −3)/2
so 𝑎 = 0, and we are in cases (ii) or (iii).
Therefore, if 𝛾 ∈ PSL2 (Z) has finite order, then 𝛾 fixes a point, thus a conjugate
of 𝛾 would fix a point in ◊, and therefore by the above 𝛾 is conjugate in PSL2 (Z) to
either 𝑆 or 𝑆𝑇.
35.1. ⊲ THE FUNDAMENTAL SET 639
Let 𝑌 = Γ\H2 . Gluing together the fundamental set, we obtain a homeomorphism
𝑌 = Γ\H2 ' P1 (C) r {∞} ' C.
The orbit of the limit point ∞ under Γ is P1 (Q) ⊆ bd H2 ; letting H2∗ = H2 ∪ P1 (Q),
there is a homeomorphism
𝑋 = Γ\H2∗ ' P1 (C)
as in Figure 35.1.15.
∞
∞
' ω
i
ω
Figure 35.1.15: 𝑌 as orbifold and as Riemann surface
Away from the orbits Γ𝑖, Γ𝜔 with nontrivial stabilizer, the complex structure on H2
descends and gives the quotient 𝑌 r {Γ𝑖, Γ𝜔} the structure of a Riemann surface. By
studying the moduli of lattices, later we will give an explicit holomorphic identification
𝑗 : 𝑌 → C.
35.1.16. By 34.8.11, the quotient 𝑌 has the structure of a good complex 1-orbifold,
when we keep track of the two nontrivial stabilizers.
Alternatively, we can also give 𝑋 the structure of a compact Riemann surface as
follows. Let 𝑧0 ∈ ◊ ∪ {∞}. If 𝑧0 = ∞, we take the chart 𝑧 ↦→ 𝑒 2 𝜋𝑖𝑧 . Otherwise, let
𝑒 = # StabΓ (𝑧0 ) < ∞, let 𝑤 = (𝑧 − 𝑧0 )/(𝑧 − 𝑧0 ) be the local coordinate as in (33.7.3),
and take the chart 𝑧 ↦→ 𝑤 𝑒 at 𝑧0 .
Lemma 35.1.17. Every relation among 𝑆, 𝑇 is obtained from 𝑆 2 = (𝑆𝑇) 3 = 1 after

conjugation by Γ, so that Γ has the presentation
Γ ' h𝑆, 𝑇 | 𝑆 2 = (𝑆𝑇) 3 = 1i.
Thus Γ = PSL2 (Z) is the free product of Z/2Z and Z/3Z.
Proof. Consider a relation 𝛿1 · · · 𝛿𝑟 = 1 where 𝛿𝑖 ∈ {𝑆, 𝑇, 𝑇 −1 } (recall 𝑆 = 𝑆 −1 ).

We may suppose that 𝛿𝑖 ≠ 𝛿𝑖+1 −1 for 𝑖 = 1, . . . , 𝑟 − 1 or else we can cancel adjacent
terms. Define 𝛾𝑖 := 𝛿1 · · · 𝛿𝑖 for 𝑖 = 1, . . . , 𝑟, so that 𝛾𝑟 = 1 and 𝛾𝑖+1 = 𝛾𝑖 𝛿𝑖+1 for all

𝑖 = 1, . . . , 𝑟 − 1. Let 𝑧0 = 2𝑖 and let
𝑧 𝑖 := 𝛾𝑖 𝑧0 = (𝛿1 · · · 𝛿𝑖 )𝑧 0 .
We claim that 𝑧𝑖+1 and 𝑧𝑖 are in adjacent Γ-translates of ◊: indeed,

𝛾𝑖+1 ◊ ∩ 𝛾𝑖 ◊ ∈ 𝛾𝑖 (𝛿𝑖+1 ◊ ∩ ◊) ≠ ∅ (35.1.18)
is a side of both. Draw the geodesic (shortest path, across the corresponding side)
between 𝑧𝑖+1 and 𝑧𝑖 for each 𝑖. Because 𝛿𝑖 ≠ 𝛿𝑖+1 −1 , there is no backtracking; taken
together, they define a loop in the upper half-plane, since 𝑧𝑟 = 𝛾𝑟 𝑧 0 = 𝑧0 .

We first conjugate the relation by 𝑇 or 𝑇 −1 so 𝛿1 = 𝑆. If 𝛿𝑟 = 𝑆 as well, then
we conjugate by 𝑆 and begin again. So without loss of generality 𝛿𝑟 = 𝑇, 𝑇 −1 ; we
explain the case 𝛿𝑟 = 𝑇, the case 𝛿𝑟 = 𝑇 −1 is similar. Then our relation looks like
𝑆𝛿2 · · · 𝛿𝑟 −1𝑇 = 1, and so 𝛾𝑟 = 1 = 𝛾𝑟 −1 𝛿𝑟 = 𝛾𝑟 −1𝑇 implies 𝑧𝑟 −1 = 𝛾𝑟 −1 𝑧0 = 𝑇 −1 𝑧0 =
2𝑖 − 1.
If 𝑧𝑖 = 𝛾𝑖 𝑧 0 = 𝑧0 for some 0 < 𝑖 < 𝑟, then 𝛾𝑖 = 𝛿1 · · · 𝛿𝑖 = 1 and similarly
𝛿𝑖+1 · · · 𝛿𝑟 = 1, so we may argue separately with each such relation and so may
suppose that 𝑧𝑖 ≠ 𝑧0 for all 𝑖 = 1, . . . , 𝑟 − 1. It follows that the loop intersects ◊ only
in the path from 𝑧0 to 𝑧1 and from 𝑧𝑟 −1 to 𝑧 0 , because any other intersection would
necessarily have source or target 𝑧 0 .
We observe that 𝜔 is in the interior of the loop, since by continuity any path from 𝑖
to 2𝑖 − 1/2 in H2 that does not√intersect ◊ must go from right to left through a highest
value −1/2 + 𝑖𝑡 with 0 < 𝑡 ≤ 3/2.
The proof proceeds by induction on the number of points in the intersection of
the interior of the loop with the set of vertices Γ𝜔. We have shown in the previous
paragraph that if there are no such points, then the relation is trivial. In the general
case with relation 𝑆𝛿2 · · · 𝛿𝑟 −1𝑇 = 1, expanding (𝑆𝑇) 3 = 1 we get 𝑇 = (𝑆𝑇 −1 ) 2 𝑆 and
substituting we obtain another relation 𝑆𝛿2 · · · 𝛿𝑟 −1 (𝑆𝑇 −1 ) 2 𝑆 = 1, as in Figure 35.1.19.
zr = 2i − 1 z0 = 2i
δr = T
z1 = 12 i
Figure 35.1.19: Simplifying a relation

Reading this relation as above, we see that the loop encloses one fewer point in Γ𝜔
(it starts and ends with a backtracking step), and so the same is true for the conjugate
relation 𝛿2 · · · 𝛿𝑟 −1 (𝑆𝑇 −1 ) 2 = 1. Cancelling any new adjacent terms and conjugating
the relation does not change the number of enclosed interior vertices, so the result
holds by induction.
35.2. BINARY QUADRATIC FORMS 641
(Alternatively, for a proof in the style of Lemma 35.1.10, see Exercise 35.5.)
Remark 35.1.20. Alperin [Alp93] uses the action on the irrational numbers to show
directly that PSL2 (Z) is the free product of Z/2Z and Z/3Z (but note the typo 𝛽(𝑧) =
1 − 1/𝑧 on the first page).
35.2 Binary quadratic forms

We pause to give an application to quadratic forms and class groups, after Gauss. An
integral binary quadratic form, abbreviated in this section to simply form, is an
expression
𝑄(𝑥, 𝑦) = 𝑎𝑥 2 + 𝑏𝑥𝑦 + 𝑐𝑦 2 ∈ Z[𝑥, 𝑦].
We define the discriminant of a form 𝑄 to be
disc(𝑄) := 𝑏 2 − 4𝑎𝑐.
A form 𝑄 is primitive if gcd(𝑎, 𝑏, 𝑐) = 1, and 𝑄 is positive definite if 𝑄(𝑥, 𝑦) > 0

for all nonzero (𝑥, 𝑦) ∈ R2 ; after completing the square, we see that a form is positive
definite if and only if 𝑎 > 0 and disc(𝑄) < 0.
For 𝑑 < 0, let
Q 𝑑 := {𝑄(𝑥, 𝑦) = 𝑎𝑥 2 + 𝑏𝑥𝑦 + 𝑐𝑦 2 : 𝑎 > 0, disc(𝑄) = 𝑑}
be the set of primitive, positive definite forms of discriminant 𝑑. The group Γ acts on
Q 𝑑 the right by change
of variable: for 𝛾 ∈ Γ, we define (𝑄 𝛾 ) (𝑥, 𝑦) = 𝑄(𝛾(𝑥, 𝑦) t ), so
𝑟 𝑠
that if 𝛾 = then
𝑡 𝑢
(𝑄 𝛾 ) (𝑥, 𝑦) = 𝑄(𝑟𝑥 + 𝑠𝑦, 𝑡𝑥 + 𝑢𝑦).
We verify that disc(𝑄 𝛾 ) = disc(𝑄) = 𝑑 for 𝛾 ∈ Γ. We say that 𝑄, 𝑄 0 are (Γ-)equivalent

if 𝑄 0 = 𝑄 𝛾 for some 𝛾 ∈ Γ = SL2 (Z).
We claim that the number of equivalence classes ℎ(𝑑) := #Q 𝑑 /Γ is finite. Indeed,
to every 𝑄 ∈ Q 𝑑 , we associate the unique root
√︁
−𝑏 + |𝑑|𝑖
𝑧𝑄 = ∈ H2
2𝑎
of 𝑄(𝑧, 1) = 0. Then 𝑧 𝑄 𝛾 = 𝛾 −1 (𝑧) for 𝛾 ∈ Γ. Therefore, by the reduction theory of

the previous section, we can replace 𝑄 up to equivalence by a form such that 𝑧 𝑄 ∈ ◊.
If we further insist that Re 𝑧 < 1/2 and Re 𝑧 < 0 if | 𝑧| = 1, then this representative is
unique, as in Figure 35.2.1.
Figure 35.2.1: A choice of unique representative in ◊
Thus
1 𝑏 1
− ≤ Re 𝑧 𝑄 = − <
2 2𝑎 2
so −𝑎 < 𝑏 ≤ 𝑎, or equivalently,
| 𝑏| ≤ 𝑎 and (𝑏 ≥ 0 if | 𝑏| = 𝑎);
and
𝑏2 − 𝑑 𝑐
| 𝑧𝑄 | = = ≥1
4𝑎 2 𝑎
so 𝑎 ≤ 𝑐 and 𝑏 ≥ 0 if equality holds. In sum, every positive definite form 𝑄 is
equivalent to a (SL2 (Z)-)reduced form satisfying
| 𝑏| ≤ 𝑎 ≤ 𝑐 with 𝑏 ≥ 0 if | 𝑏| = 𝑎 or 𝑎 = 𝑐.
We now show that there are only finitely many reduced forms with given discrim-
inant 𝑑 < 0, i.e., that ℎ(𝑑) < ∞. The inequalities | 𝑏| ≤ 𝑎 ≤ 𝑐 imply that
| 𝑑| = 4𝑎𝑐 − 𝑏 2 ≥ 3𝑎 2 ,
√︁
so 𝑎 ≤ |𝑑|/3 and |𝑏| ≤ 𝑎, so there are only finitely many possibilities for 𝑎, 𝑏; and
then 𝑐 = (𝑏 2 − 𝑑)/(4𝑎) is determined. This gives an efficient method to compute the
set Q 𝑑 /Γ efficiently. √ √
Let 𝑆 = Z ⊕ Z[(𝑑 + 𝑑)/2] ⊂ 𝐾 = Q( 𝑑) be the quadratic ring of discriminant
𝑑 < 0. Let Pic(𝑆) be the group of invertible fractional ideals of 𝑆 modulo principal
ideals. Then there is a bĳection
Q 𝑑 /Γ ↔ Pic(𝑆)
" √ !#
2 2 −𝑏 + 𝑑
[𝑎𝑥 + 𝑏𝑥𝑦 + 𝑐𝑦 ] ↦→ [𝔞] = 𝑎,
2
(Exercise 35.9). In the same stroke, we have proven the finiteness of the class number
# Pic(𝑆) < ∞.
35.3. MODULI OF LATTICES 643
35.3 Moduli of lattices

In this section, we realize PSL2 (Z)\H2 as a moduli space of complex lattices.
35.3.1. A (complex) lattice Λ ⊂ C is a subgroup Λ = Z𝑧1 + Z𝑧2 with 𝑧 1 , 𝑧2 linearly

independent over R; the elements 𝑧1 , 𝑧2 are a basis for Λ.
Two lattices Λ, Λ0 are homothetic if there exists 𝑢 ∈ C× such that Λ0 = 𝑢Λ, and
we write Λ ∼ Λ0. Let Λ = Z𝑧 1 + Z𝑧2 be a lattice. Then without loss of generality
(interchanging 𝑧1 , 𝑧2 ), we may suppose Im(𝑧2 /𝑧 1 ) > 0, and then we call 𝑧1 , 𝑧2 an
oriented basis. Then there is a homothety
1
Λ∼ Λ = Z + Z𝜏
𝑧1
where 𝜏 = 𝑧 2 /𝑧1 ∈ H2 .
Lemma 35.3.2. Let Λ = Z + Z𝜏 and Λ = Z + Z𝜏 0 be lattices with 𝜏, 𝜏 0 ∈ H2 . Then

Λ ∼ Λ0 if and only if Γ𝜏 = Γ𝜏 0.
Proof. Since 𝜏, 𝜏 0 ∈ H2 , the bases 1, 𝜏 and 1, 𝜏 0 are oriented. We have Λ = Z +

Z𝜏 ∼ Z + Z𝜏 0 = Λ0 if and only if there exists an invertible change of basis matrix
𝑎 𝑏
𝑔= ∈ GL2 (Z) and 𝑢 ∈ C× such that
𝑐 𝑑
0
𝑎 𝑏 𝜏 𝜏
𝑢 =
𝑐 𝑑 1 1
so 𝑢(𝑎𝜏 + 𝑏) = 𝜏 0 and 𝑢(𝑐𝜏 + 𝑑) = 1. Eliminating 𝑢 gives equivalently
𝑎𝜏 + 𝑏
𝜏0 = .
𝑐𝜏 + 𝑑
Therefore 𝑔 ∈ SL2 (Z), and since 𝑔 is well-defined as an element of Γ = PSL2 (Z), the
result follows.
35.3.3. By Lemma 35.3.2, there is a bĳection
𝑌 = Γ\H2 → {Λ ⊂ C lattice}/∼
(35.3.4)
Γ𝜏 ↦→ [Z + Z𝜏];
that is to say, 𝑌 = Γ\H2 parametrizes complex lattices up to homothety.

To a lattice Λ, we associate the complex torus C/Λ (of rank 1); two such tori C/Λ
and C/Λ0 are isomorphic as Riemann surfaces if and only if Λ ∼ Λ0. Therefore, the
space 𝑌 also parametrizes complex tori.
We return to this interpretation in section 40.1.

35.4 Congruence subgroups

The finite-index subgroups of PSL2 (Z) play a central role, and of particular importance
are those subgroups defined by congruence conditions on the entries.
Definition 35.4.1. Let 𝑁 ∈ Z ≥1 . The full congruence subgroup Γ(𝑁) E PSL2 (Z)
of level 𝑁 is
Γ(𝑁) := ker(PSL2 (Z) → PSL2 (Z/𝑁Z))

𝑎 𝑏 1 0
= 𝛾 ∈ PSL2 (Z) : 𝛾 = ≡± (mod 𝑁) .
𝑐 𝑑 0 1
To avoid confusion, from now on we will now write Γ(1) = PSL2 (Z).
35.4.2. By strong approximation for SL2 (Z) (Theorem 28.2.6), the map SL2 (Z) →
SL2 (Z/𝑁Z) is surjective for all 𝑁 ≥ 1, so there is an exact sequence
1 → Γ(𝑁) → Γ(1) → PSL2 (Z/𝑁Z) → 1.
Definition 35.4.3. A subgroup Γ ≤ Γ(1) is a congruence subgroup if Γ ≥ Γ(𝑁) for

some 𝑁 ≥ 1; if so, the minimal such 𝑁 is called the level of Γ.
Remark 35.4.4. Noncongruence subgroups (finite-index subgroups not containing
Γ(𝑁) for any 𝑁 ≥ 1) also play a role in the structure of the group SL2 (Z): see the
recent survey by Li–Long [LL2012] and the references therein.
35.4.5. In addition to the congruence groups Γ(𝑁) themselves, we will make use of
two other important congruence subgroups for 𝑁 ≥ 1:

𝑎 𝑏
Γ0 (𝑁) := 𝛾 = ∈ PSL2 (Z) : 𝑐 ≡ 0 (mod 𝑁)
𝑐 𝑑

∗ ∗
= 𝛾 ∈ PSL2 (Z) : 𝛾 ≡ ± (mod 𝑁) (35.4.6)
0 ∗

1 ∗
Γ1 (𝑁) := 𝛾 ∈ PSL2 (Z) : 𝛾 ≡ ± (mod 𝑁)
0 1
Visibly, Γ(𝑁) ≤ Γ1 (𝑁) ≤ Γ0 (𝑁). We accordingly write
𝑌0 (𝑁) := Γ0 (𝑁)\H2
(35.4.7)
𝑋0 (𝑁) := Γ0 (𝑁)\H2∗
where H2∗ := H2 ∪ P1 (Q), and similarly 𝑌1 (𝑁) and 𝑌 (𝑁).

In the remainder of this section, we consider as an extended example the case
𝑁 = 2. We can equally well write

𝑎 𝑏
Γ(2) = ± ∈ PSL2 (Z) : 𝑏 ≡ 𝑐 ≡ 0 (mod 2) (35.4.8)
𝑐 𝑑
35.4. CONGRUENCE SUBGROUPS 645
From 35.4.2,
Γ(1)/Γ(2) ' PSL2 (Z/2Z) = GL2 (F2 ) ' 𝑆3 (35.4.9)
the nonabelian group of order 6, so in particular [Γ(1) : Γ(2)] = 6.
We can uncover the structure of the group Γ(2) in a manner similar to what we did
for Γ(1) in section 35.1—the details are requested in Exercise 35.10. The group Γ(2)
is generated by
1 2 1 0
and
0 1 2 1
which act on H2 by 𝑧 ↦→ 𝑧 + 2 and 𝑧 ↦→ 𝑧/(2𝑧 + 1), respectively, and a fundamental set

◊ is given in Figure 35.4.10.

1 2
0 1

1 0
2 1
1 1
2 (−1 + i) 2 (1 + i)
−1 0 1
Figure 35.4.10: A fundamental set for Γ(2) H2
(In fact, later we will see from more general structural results that Γ(2) is freely gen-
erated by these two elements, so it is isomorphic to the free group on two generators.)
The action Γ(2) H2 is free: by 35.1.14, if 𝛾𝑧 = 𝑧 with 𝑧 ∈ H2 and 𝛾 ∈ Γ(2) ≤
Γ(1), then 𝛾 is conjugate in Γ(1) to either 𝑆, 𝑆𝑇; but Γ(2)
E Γ(1) is normal, so
0 −1 0 −1
without loss of generality either 𝑆 = or 𝑆𝑇 = belongs to Γ(2), a
1 0 1 1
contradiction.
Let 𝑌 (2) := Γ(2)\H2 . Then gluing together the fundamental set, there is a
homeomorphism
𝑌 (2) ' P1 (C) r {0, 1, ∞}.
The limit points of ◊ in bd H2 are the points −1, 0, 1, ∞ and the points −1, 1 are
identified in the quotient (by translation). The orbit of these points under Γ(2) is
P1 (Q) ⊆ bd H2 , so letting H2∗ = H2 ∪ P1 (Q), there is a homeomorphism
𝑋 (2) := Γ(2)\H ∗ ' P1 (C). (35.4.11)

We have a natural holomorphic projection map
𝑋 (2) = Γ(2)\H2∗ → 𝑋 (1) = Γ(1)\H2∗ ; (35.4.12)
via (35.4.9), the group GL2 (F2 ) acts on 𝑋 (2) by automorphisms:
GL2 (F2 ) 𝑋 (2) → 𝑋 (2)

𝛾(Γ(2)𝑧) = Γ(2)𝛾𝑧
where 𝛾 ∈ Γ(1) is a lift, so the map (35.4.12) is obtained as the quotient by GL2 (F2 ).
Finally, the congruence conditions (35.4.6) imply that Γ0 (2) = Γ1 (2) has index 2 in
Γ(2), with the quotient generated by 𝑇, and we obtain a fundamental set by identifying
the two ideal triangles in ◊ above.
Exercises
1. Prove that 𝑌 (1) = SL2 (Z)\H2 has area(𝑌 (1)) = 𝜋/3 by direct integration
(verifying the Gauss–Bonnet formula).

1 0
2. Show that PSL2 (Z) is generated by 𝑇 and 𝑈 = . [So PSL2 (Z) is generated
1 1
by two parabolic elements (of infinite order), just as it is generated by elements
of order two and three.]
3. Prove Lemma 35.1.12 using Lemma 28.3.3 (elementary matrices).
4. In this exercise, we link the fact that PSL2 (Z) is generated by 𝑆, 𝑇 to a kind of
continued fraction via the Euclidean algorithm. [So the reduction algorithm is
a way to visualize the Euclidean algorithm.] Let 𝑎, 𝑏 ∈ Z ≥1 with 𝑎 ≥ 𝑏.
(a) Show that there exist unique 𝑞, 𝑟 ∈ Z such that 𝑎 = 𝑞𝑏 − 𝑟 and 𝑞 ≥ 2 and
0 ≤ 𝑟 < 𝑏.
From (a), define inductively 𝑟 0 = 𝑎, 𝑟 1 = 𝑏, and 𝑟 𝑖−1 = 𝑞 𝑖 𝑟 𝑖 − 𝑟 𝑖+1 with
0 ≤ 𝑟 𝑖+1 < 𝑟 𝑖 ; we then have 𝑟 1 > 𝑟 2 > · · · > 𝑟 𝑡 > 𝑟 𝑡+1 = 0 for some 𝑡 > 0.
b) Show that gcd(𝑎, 𝑏) = 𝑟 𝑡 , and if gcd(𝑎, 𝑏) = 1 then
𝑎 1
= 𝑞1 − .
𝑏 1
𝑞2 −
1
···−
𝑞𝑡
Such a continued fraction is called a negative-regular or Hirzebruch–

Jung continued fraction. [The Hirzebruch–Jung continued fraction plays
a role in the resolution of singularities [Jun08, Hir53].]
c) Show (by induction) that

0 1 0 1 𝑎 𝑟
··· = 𝑡 .
−1 𝑞 𝑡 −1 𝑞 1 𝑏 0
35.4. CONGRUENCE SUBGROUPS 647
For all 𝑞 ∈ Z, write

0 1
∈ h𝑆, 𝑇i ⊆ PSL2 (Z)
−1 𝑞
as a word in 𝑆, 𝑇, and interpret the action of this matrix in terms of the

reductionalgorithm
to the fundamental set ◊ for PSL2 (Z).
𝑎 𝑏
d) Let 𝐴 = ∈ SL2 (Z). Show that gcd(𝑎, 𝑐) = 1 and conclude from
𝑐 𝑑
(c) that there exists 𝑊 ∈ h𝑆, 𝑇i such that
1 𝑏0

𝑊𝐴 =
0 1
with 𝑏 0 ∈ Z. Conclude that h𝑆, 𝑇i = PSL2 (Z). (So how, in the end, does
this procedure to write 𝐴 in terms of 𝑆 and 𝑇 relate to the one given by the
reduction algorithm in Lemma 35.1.8?)
5. In this exercise, we give a “matrix proof” that a complete set of relations satisfied
by 𝑆, 𝑇 in PSL2 (Z) are 𝑆 2 = (𝑆𝑇) 3 = 1.
(a) Show that it suffices to show that no word 𝑆(𝑆𝑇) 𝑒1 𝑆(𝑆𝑇) 𝑒2 . . . 𝑆(𝑆𝑇) 𝑒𝑛
with 𝑒 𝑖 = 1, 2 is equal to 1.
(b) Observe that 𝑆(𝑆𝑇) = 𝑇 and 𝑆(𝑆𝑇) 2 have at least one off-diagonal entry
nonzero and can be represented with a matrix whose entries all have the
same sign.
𝑎 𝑏
(c) Show that if 𝐴 = has at least one off-diagonal entry nonzero and all
𝑐 𝑑
entries of the same sign, then these properties hold also for both 𝑆(𝑆𝑇) 𝐴
and 𝑆(𝑆𝑇) 2 𝐴. Conclude that (a) holds.
[This argument is given by Fine [Fin89, Theorem 3.2.1].]
6. Show that the commutator subgroup Γ0 E Γ = PSL2 (Z) (the subgroup generated
by commutators 𝛾𝛿𝛾 −1 𝛿−1 for 𝛾, 𝛿 ∈ Γ) has index 6 and Γ/Γ0 ' Z/6Z.
7. Compute the class number ℎ(𝑑) and the set of reduced (positive definite) binary
quadratic forms of discriminant 𝑑 = −71.
8. Let Q 𝑑 be the set of primitive,
Ð positive definite binary quadratic forms of
discriminant 𝑑 < 0 and let Q = 𝑑 Q 𝑑 .
(a) Show that the group GL2 (Z) acts naturally on Q by change of variables,
with PGL2 (Z) acting faithfully.
(b) Consider the action of PGL2 (Z) on H2 . Show that every 𝑄 ∈ Q 𝑑 is
equivalent to a GL2 (Z)-reduced form 𝑎𝑥 2 + 𝑏𝑥𝑦 + 𝑐𝑦 2 satisfying
0 ≤ 𝑏 ≤ 𝑎 ≤ 𝑐.
[Hint: Find a nice fundamental set ◊ for PGL2 (Z).]

(c) By transport, 35.1.14 computes the stabilizer of PSL2 (Z) on 𝑄 ∈ Q 𝑑 .
Compute StabPGL2 (Z) (𝑄) for 𝑄 ∈ Q 𝑑 .
I 9. Let √ √
𝑆 = Z ⊕ Z[(𝑑 + 𝑑)/2] ⊂ 𝐾 = Q( 𝑑)
be the quadratic ring of discriminant 𝑑 < 0. Let Pic(𝑆) be the group of invertible
fractional ideals of 𝑆 modulo principal ideals. Show that the map
Q 𝑑 /Γ → Pic(𝑆)
" √ !#
2 2 −𝑏 + 𝑑
[𝑎𝑥 + 𝑏𝑥𝑦 + 𝑐𝑦 ] ↦→ [𝔞] = 𝑎,
2
is a bĳection, where Q 𝑑 /Γ is the set of (SL2 (Z)-)equivalence classes of (primi-

tive, positive definite) binary quadratic forms of discriminant 𝑑.
I 10. Show that the elements
1 2 1 0
,
0 1 2 1
generate Γ(2) using the fundamental set
◊ = {𝑧 ∈ H2 : |Re 𝑧| ≤ 1, |2𝑧 ± 1| ≥ 1}.
[Hint: adapt the method used in section 35.1.]

Chapter 36
Hyperbolic space
In this chapter, we extend the notions introduced for the hyperbolic plane to hyperbolic
space in three dimensions; we follow essentially the same outline, and so our exposition
is similarly brief.
36.1 Hyperbolic space

A general, encyclopedic reference for hyperbolic geometry is the book by Ratcliffe
[Rat2006]. For further reference, see also Elstrodt–Grunewald–Mennicke [EGM98,
Chapter 1], Iversen [Ive92, Chapter VIII], and Marden [Mard2007].
Definition 36.1.1. The upper half-space is the set
H3 := C × R>0 = {(𝑥, 𝑦) = (𝑥 1 + 𝑥2𝑖, 𝑦) ∈ C × R : 𝑦 > 0}.
Hyperbolic space is the set H3 equipped with the metric induced by the hyperbolic
length element
|d𝑥| 2 + d𝑦 2 d𝑥12 + d𝑥22 + d𝑦 2
d𝑠2 := = .
𝑦2 𝑦2
36.1.2. The space H3 is the unique three-dimensional (connected and) simply con-
nected Riemannian manifold with constant sectional curvature −1. The volume ele-
ment corresponding to the hyperbolic length element is accordingly
d𝑥 1 d𝑥2 d𝑦
d𝑉 := .
𝑦3
36.1.3. A vertical half-plane in hyperbolic space is a set of points with 𝑦 arbitrary
and the coordinate 𝑥 confined to a line in C. The hyperbolic length element restricted
to every vertical half-plane is (equivalent to) the hyperbolic length element on the
hyperbolic plane. Therefore, H3 contains many isometrically embedded copies of H2 .
36.1.4. The sphere at infinity is the set
bd H3 = P1 (C) = C ∪ {∞}
649
650 CHAPTER 36. HYPERBOLIC SPACE
(analogous to the circle at infinity for H2 ), with the image of C corresponding to the
locus of points with 𝑡 = 0. We then define the completed upper half-space to be
H3∗ := H3 ∪ bd H3 .
The topology on H3∗ is defined by taking a fundamental system of neighborhoods of

the point at ∞ to be sets of the form
{(𝑥, 𝑦) ∈ H3 : 𝑦 > 𝑀 } ∪ {∞}
for 𝑀 > 0, and the open balls tangent to 𝑧 ∈ C together with 𝑧.
36.1.5. The metric space H3 is complete, and the topology on H3 is the same as the
topology induced by the Euclidean metric.
The geodesics in H3 are the Euclidean hemicircles orthogonal to C and vertical half-
lines: every two points lie in a vertical hyperbolic plane (see 36.1.3), so this statement
can be deduced from the case of the hyperbolic plane. (Alternatively, by applying an
element of PSL2 (C) it is enough to show that the vertical axis 𝑍 = {(0, 𝑦) : 𝑦 > 0}
is a geodesic, and arguing as in (33.5.4) we obtain the result.) Accordingly, H3 is a
uniquely geodesic space.
36.1.6. Just as in distinct points determine a geodesic, so do three distinct points

determine a geodesic plane, the union of all geodesics through the third point and
a point on the geodesic between the other two (the choice taken arbitrarily). In a
geodesic plane, the geodesic between two points in the plane is contained in the plane.
By the preceding paragraph, the geodesic planes in H3 are the Euclidean hemispheres
orthogonal to C and the vertical half-planes, as in Figure 36.1.7.
Figure 36.1.7: Geodesic lines and planes in H3

36.2. ISOMETRIES 651
36.2 Isometries
Analogous to the case of H2 , with orientation-preserving isometries given by PSL2 (R)
acting by linear fractional transformations, in this section we identify the isometries
of hyperbolic space H3 as coming similarly from PSL2 (C).
36.2.1. The group PSL2 (C) acts on the sphere at infinity P1 (C) by linear fractional
transformations. We extend this action to H3 (with almost the same definition!) as
follows. We identify
H3 ↩→ H = C + C 𝑗
(𝑥, 𝑦) ↦→ 𝑧 = 𝑥 + 𝑦 𝑗
where we recall that 𝑗𝑥 = 𝑥 𝑗 for 𝑥 ∈ C = R + R𝑖 ⊆ H. We then define the action map
SL2 (C) × H3 → H3
(36.2.2)
(𝑔, 𝑧) ↦→ 𝑔𝑧 = (𝑎𝑧 + 𝑏) (𝑐𝑧 + 𝑑) −1

𝑎 𝑏
for 𝑔 = ∈ SL2 (C). If 𝑧 = 𝑥 + 𝑦 𝑗, then in coordinates (Exercise 36.3)
𝑐 𝑑
(𝑎𝑥 + 𝑏) (𝑐𝑥 + 𝑑) + 𝑎𝑐𝑦 2 + 𝑦 𝑗

𝑔(𝑧) = (36.2.3)
k𝑐𝑧 + 𝑑 k 2
where
k𝑐𝑧 + 𝑑 k 2 = nrd(𝑐𝑧 + 𝑑) = |𝑐𝑥 + 𝑑| 2 + |𝑐| 2 𝑦 2 .
Therefore the image of this map lies in H3 . (Compare this formula with the action of
SL2 (R) in (33.3.8).)
Lemma 36.2.4. The map (36.2.2) defines a group action of SL2 (C) on H3 .
Proof. We define the quaternionic projective line to be the set
P1 (H) := {(𝛼, 𝛽) : 𝛼, 𝛽 ≠ (0, 0) ∈ H× }/∼
under the equivalence relation (𝛼, 𝛽) ∼ (𝛼𝛾, 𝛽𝛾) for 𝛾 ∈ H× , and we denote by
(𝛼 : 𝛽) ∈ P1 (H) the equivalence class of (𝛼, 𝛽). We verify that the group SL2 (C) acts
on P1 (H) by
𝑎 𝑏
· (𝛼 : 𝛽) = (𝑎𝛼 + 𝑏𝛽 : 𝑐𝛼 + 𝑑𝛽);
𝑐 𝑑
the left action of SL2 (C) commutes with the right action of H× . The restriction of this
action to H3 ↩→ P1 (H) by 𝑧 ↦→ (𝑧 : 1) is
(𝑧 : 1) ↦→ (𝑎𝑧 + 𝑏 : 𝑐𝑧 + 𝑑) = ((𝑎𝑧 + 𝑏) (𝑐𝑧 + 𝑑) −1 : 1)
as above.
We now show that PSL2 (C) acts on H3 by isometries. This can be verified directly
by the formula, with some effort; we prefer to verify this on a convenient set of
generators, and so we are first led already to the following decomposition of SL2 (C)
(cf. Proposition 33.4.2).
36.2.5. Let

𝑎 𝑏
𝐾 := SU(2) = ∈ M2 (C) : |𝑎| 2 + | 𝑏| 2 = 1 ' H1
−𝑏 𝑎

𝑎 0
𝐴 := : 𝑎 ∈ R×>0 ' R
0 1/𝑎

1 𝑏
𝑁 := : 𝑏 ∈ C ' C.
0 1
We have 𝐾 = StabSL2 (C) ( 𝑗): from (36.2.3), we see that 𝑔 𝑗 = (𝑎 𝑗 + 𝑏) (𝑐 𝑗 + 𝑑) −1 if

and only if |𝑐| 2 + | 𝑑| 2 = 1 and 𝑎𝑐 + 𝑏𝑑 = 0; plugging the first equation into the second,
and using 𝑎𝑑 − 𝑏𝑐 = 1 gives 𝑎 = 𝑑 and then 𝑏 = −𝑐.
Letting 𝑧 = 𝑥 + 𝑦 𝑗, the other elements act as:

𝑎 0
(𝑧) = 𝑎 2 (𝑥 + 𝑦 𝑗),
0 1/𝑎
(36.2.6)
1 𝑏
(𝑧) = (𝑥 + 𝑏) + 𝑦 𝑗 .
0 1
Lemma 36.2.7 (Iwasawa decomposition). The multiplication map gives a homeomor-

phism
𝑁 × 𝐴 × 𝐾 −→∼ SL (C).
2
Proof. We apply the same method as in the proof of Proposition 33.4.2. For surjec-
1 −𝑥
tivity, we let 𝑧 = 𝑔( 𝑗) = 𝑥 + 𝑦 𝑗, let 𝑛𝑔 = ∈ 𝑁 so that (𝑛𝑔 𝑔) ( 𝑗) = 𝑦 𝑗; then let
0 1
√
1/ 𝑦 0
𝑎𝑔 = √ ∈ 𝐴, so (𝑎 𝑔 𝑛𝑔 𝑔) ( 𝑗) = 𝑗 and 𝑎 𝑔 𝑛𝑔 𝑔 ∈ StabSL2 (C) ( 𝑗) = 𝐾.
0 𝑦
Lemma 36.2.8. The group SL2 (C) is generated by the subgroups 𝐴, 𝑁, and the
0 −1
element , which acts on H3 by
1 0

0 −1 1
(𝑧) = −𝑧 −1 = (−𝑥 + 𝑦 𝑗) (36.2.9)
1 0 k𝑧k 2
where k𝑧k 2 = nrd(𝑧) = |𝑥| 2 + 𝑦 2 for 𝑧 = 𝑥 + 𝑦 𝑗.
Proof. The proof is identical to the one in Lemma 33.4.4.

𝑎 0
Remark 36.2.10. In fact, the generators are redundant, but we will not use
0 1/𝑎
this fact here.
36.2. ISOMETRIES 653
We are now ready to investigate the consequences of this decomposition for the
geometry of hyperbolic space.
Theorem 36.2.11. The map (36.2.2) defines a faithful, transitive action of PSL2 (C)
on H3 by isometries.
Proof. We use the generators in Lemma 36.2.8. The fact that the action is faithful
3 3
follows directly. For transitivity,we show that H is the orbit of 𝑗. If 𝑧 = 𝑥 + 𝑦 𝑗 ∈ H
1 −𝑥
then we first apply a translation to reduce to the case 𝑧 = 𝑦 𝑗 and then reduce
0 1
to the case of the hyperbolic plane.
Next, we show that PSL2 (C) ↩→ Isom+ (H3 ). Verification that d𝑔(𝑠) = d𝑠 for 𝑔 a
generator in one of the first two cases of Lemma 36.2.8 is immediate, from the definition
of the metric; the third case can be checked directly (Exercise 36.4). Orientation is
preserved in each case.
36.2.12. The group PSL2 (C) acts transitively on geodesics and consequently on pairs
of points at a fixed distance: by the transitive action of PSL2 (C) on H3 , every point
can be mapped to 𝑗; and applying an element of StabSL2 (C) 𝑗 = SU(2), every other
point 𝑢 can be brought to 𝑡 𝑗 with 𝑡 ≥ 1, with log 𝑡 = 𝜌( 𝑗, 𝑢) by the distance in the
hyperbolic plane. It follows that
|𝑧 − 𝑧 0 | 2 |𝑥 − 𝑥 0 | 2 + (𝑦 − 𝑦 0) 2
cosh 𝜌(𝑧, 𝑧 0) = 1 + = 1 + (36.2.13)
2𝑦𝑦 0 2𝑦𝑦 0
by verifying (36.2.13) in the special case where 𝑧 = 𝑗 and 𝑧 0 = 𝑦 𝑗 with 𝑦 > 0, and then
using the preceding transitive action and the fact that the right-hand side of (36.2.13)
is invariant under the action of SL2 (C), verified again using the generators in Lemma
36.2.8.
Isom+ (H3 ) ' PSL2 (C) (36.2.15)
and
Isom(H3 ) ' PSL2 (C) o Z/2Z (36.2.16)
where the nontrivial element of Z/2Z acts by complex conjugation on PSL2 (C) and
(𝑧, 𝑡) ↦→ (𝑧, 𝑡) on H3 .
Proof. We argue as in Theorem 33.5.5. Let 𝜙 ∈ Isom(H3 ), and let
𝑍 = {𝑦 𝑗 : 𝑦 > 0} ⊆ H3 .
Then 𝑍 is a geodesic (see 36.1.5), so 𝜙(𝑍) is also a geodesic. By transitivity, there

exists an isometry 𝑔 ∈ PSL2 (C) that maps 𝜙( 𝑗) back to 𝑗, and we may suppose without
loss of generality that 𝜙( 𝑗) = 𝑗, and arguing as in the case of H2 we may suppose
in fact that 𝜙 fixes each point of 𝑍. Let H = R + R 𝑗 ⊆ H3 . Then H is a geodesic
half-plane containing 𝑍, so 𝜙(H ) is as well and must be a vertical half-plane. The

𝑒𝑖 𝜃 0
isometric rotation fixes 𝑍, and applying such a rotation we may suppose
0 𝑒 −𝑖 𝜃
further that 𝜙 fixes H .
Now let 𝑧 = 𝑥 + 𝑦 𝑗 and 𝜙(𝑧) = 𝑧 0 = 𝑥 0 + 𝑦 0 𝑗. Let 𝑟 + 𝑠 𝑗 ∈ H . Then
𝜌(𝑧, 𝑟 + 𝑠 𝑗) = 𝜌(𝜙(𝑧), 𝜙(𝑟 + 𝑠 𝑗)) = 𝜌(𝑧 0, 𝑟 + 𝑠 𝑗)
so from (36.2.13)
|𝑥 − 𝑟 | 2 + (𝑦 − 𝑠) 2 |𝑥 0 − 𝑟 | 2 + (𝑦 0 − 𝑠) 2
= ;
2𝑠𝑦 2𝑠𝑦 0
letting 𝑠 → ∞ we find that 𝑦 = 𝑦 0 and |𝑥 − 𝑟 | = |𝑥 0 − 𝑟 | for all 𝑟 ∈ R, thus Re 𝑥 = Re 𝑥 0
and Im 𝑥 = ± Im 𝑥 0. By continuity, the sign is determined uniquely by 𝑔, and we
conclude that either 𝑔(𝑧) = 𝑧 or 𝑔(𝑧) = 𝑥 + 𝑦 𝑗, as claimed.
36.2.17. The isometry group PSL2 (C) also admits a ‘purely geometric’ definition via
the Poincaré extension, as follows.
An element 𝑔 ∈ SL2 (C) as a Möbius transformation, induces a biholomorphic
map of the Riemann sphere P1 (C) = C ∪ {∞}. This map can be represented as a
composition of an even number (at most four) inversions in circles in P1 (C), or circles
and lines in C (Exercise 36.2). We have identified P1 (C) = bd H3 as the boundary,
and for each circle in P1 (C) there is a unique hemisphere in H3 which intersects bd H3
in this circle; if this circle is a line, then we take a vertical half-plane. We then lift
the action of 𝑔 ∈ PSL2 (C) one inversion at a time with respect to the corresponding
hemisphere or half-plane. It turns out that the action of this product does not depend
on the choice of the circles.
To verify that PSL2 (C) acts by isometries, we need to know that inversion in a
hemisphere or vertical half-plane is an isometry of H3 ; after observing that the first
two types of generators in Lemma 36.2.8 (stretching and translating) are isometries,
one reduces to the case of checking that inversion in the unit hemisphere, defined by
𝑧
𝑧 ↦→ ,
k𝑧k 2
is an isometry; and this boils down to the same calculation as requested in Exercise
36.4.
36.2.18. We have a similar classification of isometries of H3 as in the case of H2 as
follows. Let 𝑔 ∈ PSL2 (C).
(i) If ± Tr(𝑔) ∈ (−2, 2), then 𝑔 is elliptic: it has two distinct fixed points in bd H3
and fixes every point in the geodesic between them, called its axis, acting by
(hyperbolic) rotation around its axis.
(ii) If ± Tr(𝑔) ∈ R r [−2, 2], then 𝑔 is hyperbolic; if ± Tr(𝑔) ∈ C \ R, then 𝑔 is
loxodromic. (Some authors combine these two cases.) In these cases, 𝑔 has two
fixed points in bd H3 and the line through these two points is stabilized, and 𝑔
has no fixed point in H3 .
(iii) Finally and otherwise, if ± Tr(𝑔) = ±2, then 𝑔 is parabolic: it has a unique
fixed point in bd H3 and no fixed point in H3 .
36.3. UNIT BALL, LORENTZ, AND SYMMETRIC SPACE MODELS 655
36.3 Unit ball, Lorentz, and symmetric space models

Definition 36.3.1. The hyperbolic unit ball is the (open) unit disc
D3 := {𝑤 = (𝑤 1 , 𝑤 2 , 𝑤 3 ) ∈ R3 : k𝑤k 2 < 1} (36.3.2)
equipped with the hyperbolic metric

2kd𝑤k
d𝑠 := (36.3.3)
1 − k𝑤k 2
and volume
d𝑤 1 d𝑤 2 d𝑤 3
d𝑉 := 8 . (36.3.4)
(1 − k𝑤k) 3
The sphere at infinity is the boundary
bd D3 = {𝑤 ∈ R3 : k𝑤k = 1}.
36.3.5. The maps

∼ D3
𝜙 : H3 −→ ∼ H3
𝜙−1 : D3 −→
𝑧 ↦→ 𝑤 = (𝑧 − 𝑗) (1 − 𝑗 𝑧) −1 𝑤 ↦→ 𝑧 = (𝑤 + 𝑗) (1 + 𝑗 𝑤) −1
define a conformal equivalence between H3 and D3 with 𝑗 ↦→ 𝜙( 𝑗) = 0. The

hyperbolic metric on D2 is the pushforward of (induced from) the hyperbolic metric
on H3 via the identification (36.3.5). We find that
k𝑤 − 𝑤 0 k 2
cosh 𝜌(𝑤, 𝑤 0) = 1 + 2 . (36.3.6)
(1 − k𝑤k 2 ) (1 − k𝑤 0 k 2 )
In the unit ball model, the geodesics are intersections of D3 of Euclidean circles
and straight lines orthogonal to the sphere at infinity, and similarly geodesic planes
are intersections of D3 with Euclidean spheres and Euclidean planes orthogonal to the
sphere at infinity.
36.3.7. The isometries of D3 are obtained by pushforward from H3 . Explicitly, we
first identify
D3 ↩→ H
(36.3.8)
𝑤 ↦→ 𝑤 1 + 𝑤 2𝑖 + 𝑤 3 𝑗 .
We then define the involution
∗
:H→H
(36.3.9)
𝛼 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 ↦→ 𝑘𝛼𝑘 −1 = 𝑡 + 𝑥𝑖 + 𝑦 𝑗 − 𝑧𝑘
and the group

𝛼 𝛽
SU2 (H, ∗ ) = : 𝛼, 𝛽 ∈ H, nrd(𝛼) − nrd(𝛽) = 1 . (36.3.10)
𝛽∗ 𝛼∗
We find that
SU2 (H, ∗ ) ' 𝜙 SL2 (C)𝜙−1
with 𝜙 as in 36.3.5. The group SU2 (H, ∗ ) acts on D3 by
𝑔𝑤 = (𝛼𝑤 + 𝛽) (𝛽∗ 𝑤 + 𝛼∗ ) −1 .
36.3.11. Finally, there is the Lorentz model
L3 := {(𝑡, 𝑥) ∈ R4 : −𝑡 2 + 𝑥 12 + 𝑥22 + 𝑥32 = −1, 𝑡 > 0} (36.3.12)
with
d𝑠2 := −d𝑡 2 + d𝑥12 + d𝑥22 + d𝑥32
and orientation-preserving isometries given by the subgroup SO+ (3, 1) ≤ SO(3, 1)
of elements mapping L3 to itself. The relationship between the Lorentz model and
the upper half-space model relies on the exceptional isomorphism of Lie algebras
𝔰𝔬3,1 ' 𝔰𝔩2,C and the double cover SL2 (C) → SO+ (3, 1).
To conclude, we find the symmetric space model of H3 , analogous to section 34.6.
36.3.13. The group 𝐺 := SL2 (C) has the structure of a metric space induced from the
usual structure on M2 (C) ' C4 . Since 𝐺 acts transitively on H3 , and the stabilizer of
𝑗 is 𝐾 = SU(2),
𝐺/𝐾 = SL2 (C)/SU(2) −→ ∼ H3
(36.3.14)
𝑔𝐾 ↦→ 𝑔 𝑗;
from the Iwasawa decomposition (Lemma 36.2.7), there is a homeomorphism
SL2 (C)/SU(2) ' 𝑁 𝐴.
From the identity

k𝑔k 2 = 2 cosh 𝜌( 𝑗, 𝑔 𝑗) (36.3.15)
for 𝑔 ∈ SL2 (C), proven in the same way as (34.6.5), the map (36.3.14) is a homeo-
morphism, and even an isometry under the explicit reparametrization (36.3.15) of the
metric.
Remark 36.3.16. Similar statements about the unit tangent bundle hold for PSL2 (C)
in place of PSL2 (R), as in 33.8.2.
Remark 36.3.17. More generally, one defines hyperbolic upper half-space
|d𝑥| 2 + d𝑦 2
H𝑛 := {(𝑥, 𝑦) ∈ R𝑛 × R : 𝑦 > 0} with d𝑠2 = .
𝑦2
The space H𝑛 is a uniquely geodesic space and a model for hyperbolic 𝑛-space. The
geodesics in H𝑛 are orthocircles, and via a conformal map. The upper half-space maps
isometrically to the (open) unit ball model
d𝑥12 + · · · + d𝑥 𝑛2
D𝑛 := {𝑥 ∈ R𝑛 : |𝑥| < 1} with d𝑠2 = 4
(1 − 𝑥12 − · · · − 𝑥 𝑛2 ) 2
36.4. BIANCHI GROUPS AND KLEINIAN GROUPS 657
and the hyperboloid model

L𝑛 := {(𝑡, 𝑥) ∈ R𝑛+1 : −𝑡 2 + 𝑥12 + · · · + 𝑥 𝑛2 = −1, 𝑡 > 0}
with
d𝑠2 = −d𝑡 2 + d𝑥12 + · · · + d𝑥 𝑛2 .
These models (and more) are introduced and compared in Cannon–Floyd–Kenyon–
Parry [CFKP97], and treated in detail in the works by Benedetti–Petronio [BP92] and
Ratcliffe [Rat2006].
Hyperbolic 𝑛-space H𝑛 also admits a symmetric space description, as follows.
The group of isometries of H𝑛 is SO(𝑛, 1), and the subgroup of orientation-preserving
isometries is SO+ (𝑛, 1), the component of SO(𝑛, 1) containing the identity matrix.
The stabilizer of every point in H𝑛 is conjugate to SO(𝑛) (rotation around the origin
in the unit ball model, with the fixed point at the origin), and it follows that
H𝑛 ' SO+ (𝑛, 1)/SO(𝑛).
36.4 Bianchi groups and Kleinian groups

Theorem 36.4.1. Let 𝐺 := PSL2 (C) and let Γ ≤ 𝐺 be a subgroup. Then the following
are equivalent:
(i) Γ is discrete (with the subspace topology);
(ii) For all 𝑧 ∈ H3 , we have # StabΓ (𝑧) < ∞ and there exists an open neigborhood
(iii) For all compact subsets 𝐾 ⊆ H3 , we have 𝐾 ∩ 𝛾𝐾 ≠ ∅ for only finitely many
𝛾 ∈ Γ; and
(iv) For all 𝑧 ∈ H3 , the orbit Γ𝑧 ⊆ H3 is discrete and # StabΓ (𝑧) < ∞.
Moreover, if these equivalent conditions hold, then the quotient Γ\H3 is Hausdorff,
and the quotient map 𝜋 : H3 → Γ\H3 is a local isometry at all points 𝑧 ∈ H3 with
StabΓ (𝑧) = {1}.
Proof. Combine Theorem 34.5.1 and the appropriately modified proof of Proposition
34.7.2. The stabilizer of a point is finite because the stabilizer of 𝑤 = 0 in SU2 (H, ∗ )
is SU(2), so its stabilizer in Γ is a discrete subgroup of the compact group SU(2) thus
is necessarily finite (not necessarily cyclic). In particular, a subgroup Γ ≤ PSL2 (C) is
discrete if and only if the action of Γ on H3 is wandering, hence proper.
Definition 36.4.2. A Kleinian group is a discrete subgroup of PSL2 (C).
Let 𝐹 ⊆ C be an imaginary quadratic field with ring of integers 𝑅 = Z𝐹 . Since
𝑅 ⊆ C is discrete, and PSL2 (𝑅) ⊆ PSL2 (C) is discrete.
Definition 36.4.3. The Bianchi group over 𝐹 is the Kleinian group PSL2 (𝑅) ⊆
PSL2 (C).
Remark 36.4.4. The Bianchi groups are so named after work of Bianchi [Bia1892];
he studied them as discrete groups acting on hyperbolic space and found generators in
certain cases. For more, see the book by Fine [Fin89].
36.5 Hyperbolic volume

In this section, we consider volumes of hyperbolic polyhedra, following Milnor’s
chapter in Thurston [Thu97, Chapter 7], published also in Milnor [Milno82, Appendix];
see also the full treatment by Ratcliffe [Rat2006, §10.4].
Definition 36.5.1. The Lobachevsky function is defined to be
L: R → R
∫ 𝜃 (36.5.2)
L (𝜃) = − log |2 sin 𝑡| d𝑡.
0
The Lobachevsky function is also called Clausen’s integral or more conventionally

the log sine integral.
36.5.3. The first derivative of the Lobachevsky function is L 0 (𝜃) = − log |2 sin 𝜃| by
the fundamental theorem of calculus, so L attains its maximum value at L (𝜋/6) =
0.50747 . . . and minimum at L (5𝜋/6) = −L (𝜋/6). The second derivative is L 00 (𝜃) =
− cot 𝜃. A graph of this function is sketched in Figure 36.5.4.
L (θ)
1/2
θ
π/2 π
−1/2
Figure 36.5.4: The Lobachevsky function L
Lemma 36.5.5. L (𝜃) is odd, periodic with period 𝜋, and satisfies the identity
𝑛−1
∑︁
L (𝑛𝜃) = 𝑛 L (𝜃 + 𝑗 𝜋/𝑛) (36.5.6)
𝑗=0
for all 𝑛 ∈ Z.
Proof. Since L 0 (𝜃) = − log |2 sin 𝜃| is an even function and L (0) = 0, we conclude
L (𝜃) is an odd function, i.e., L (−𝜃) = −L (𝜃) for all 𝜃 ∈ R.
Let 𝑛 ∈ Z. From
𝑛−1
Ö
𝑧𝑛 − 1 = (𝑧 − 𝑒 2 𝜋𝑖 𝑗/𝑛 )
𝑗=0
36.5. HYPERBOLIC VOLUME 659
substituting 𝑧 = 𝑒 −2𝑖𝑡 for 𝑡 ∈ R and using |𝑒 2𝑖 𝜃 − 1| = |1 − 𝑒 2𝑖 𝜃 | = |2 sin 𝜃| for all 𝜃 ∈ R

gives
𝑛−1
Ö 𝑛−1
Ö
|2 sin(𝑛𝑡)| = |1 − 𝑒 2𝑖𝑛𝑡 | = |𝑒 −2𝑖𝑡 − 𝑒 2 𝜋𝑖 𝑗/𝑛 | = |2 sin(𝑡 + 𝑗 𝜋/𝑛)|
𝑗=0 𝑗=0
for all 𝑡 ∈ R. Integrating and changing variables 𝑥 = 𝑛𝑡 gives

∫ 𝜃 𝑛−1 ∫ 𝜃+ 𝑗 𝜋/𝑛
1 ∑︁
log |2 sin 𝑥|d𝑥 = log |2 sin 𝑥|d𝑥 (36.5.7)
𝑛 0 𝑗=0 𝑗 𝜋/𝑛
which yields
𝑛−1 𝑛−1
1 ∑︁ ∑︁
L (𝑛𝜃) = L (𝜃 + 𝑗 𝜋/𝑛) − L ( 𝑗 𝜋/𝑛) (36.5.8)
𝑛 𝑗=0 𝑗=0
for all 𝜃 ∈ R. Plugging in 𝜃 = 𝜋/𝑛 into (36.5.8) yields by telescoping
1
L (𝜋) = L (𝜋) − L (0) = L (𝜋)
𝑛
so L (𝜋) = 0.
Now since L 0 (𝜃) is periodic with period 𝜋 and L (0) = L (𝜋) = 0, we conclude
that L (𝜃 + 𝜋) = L (𝜃) is also periodic with period 𝜋. Finally,
𝑛−1
∑︁ 𝑛−1
∑︁ 𝑛−1
∑︁ 𝑛−1
∑︁
L ( 𝑗 𝜋/𝑛) = − L (− 𝑗 𝜋/𝑛) = − L ((𝑛 − 𝑗)𝜋/𝑛) = L ( 𝑗 𝜋/𝑛)
𝑗=0 𝑗=0 𝑗=0 𝑗=0
Í𝑛−1
so 𝑗=0 L ( 𝑗 𝜋/𝑛) = 0, and the result follows from (36.5.8).
L (2𝜃) = 2L (𝜃) + 2L (𝜃 + 𝜋/2) = 2L (𝜃) − 2L (𝜋/2 − 𝜃).
Corollary 36.5.9 is called the duplication formula for L.
Proof. Take 𝑛 = 2 in Lemma 36.5.5.
With the relevant function having been defined, we now return to our geometric
application.
Definition 36.5.10. An ideal tetrahedron is a tetrahedron whose vertices lie on the

sphere at infinity and whose edges are (infinite) geodesics.
36.5.11. An ideal tetrahedron is determined by the three dihedral angles 𝛼, 𝛽, 𝛾 along

the edges meeting at any vertex; the sum of these angles is 2𝜋, as the shadow triangle
made in C has angles that sum to 2𝜋. For an illustration, see Figure 36.5.12.
γ β
γ β
α
α α
β γ
β γ
γ β
α
α
Figure 36.5.12: Ideal tetrahedra and its shadow in C
Proposition 36.5.13. The volume of an ideal tetrahedron with dihedral angles 𝛼, 𝛽, 𝛾

is L (𝛼) + L (𝛽) + L (𝛾).
Proof. We follow Milnor [Milno82, Appendix, Lemma 2]; see also Thurston(–Milnor)
[Thu97, Theorem 7.2.1] and Ratcliffe [Rat2006, Theorem 10.4.10]. We may suppose
without loss of generality that one vertex is at ∞ and the finite face lies on the unit
sphere. Projecting onto the unit disc in the 𝑥-plane, we obtain a triangle inscribed
in the unit circle with angles 𝛼, 𝛽, 𝛾 with 𝛼 + 𝛽 + 𝛾 = 2𝜋. We make the simplifying
assumption that all three angles are acute (the argument for the case of an obtuse angle
is similar). We take the barycentric subdivision of the triangle and add up 6 volumes.
We integrate the volume element d𝑥1 d𝑥 2 d𝑦/𝑦 3 over the region 𝑇 (𝛼) defined by the
inequalities
√︁
𝑦 ≥ 1 − |𝑥| 2 , 0 ≤ 𝑥2 ≤ 𝑥1 tan 𝛼, 0 ≤ 𝑥 1 ≤ cos 𝛼. (36.5.14)
Integrating with respect to 𝑦 we have
1 d𝑥1 d𝑥2 ∞
∫∫∫ ∫∫
d𝑥 1 d𝑥 2 d𝑦
= −
𝑇 ( 𝛼) 𝑦3 2 𝑦 2 𝑦=√1−| 𝑥 |2
∫∫
1 d𝑥1 d𝑥2
=−
2 0≤𝑥1 ≤cos
0≤𝑥 ≤𝑥 tan 𝛼
𝛼 1 − 𝑥 12 − 𝑥 22
2 1
36.5. HYPERBOLIC VOLUME 661
We substitute 𝑥 1 = cos 𝜃, so d𝑥1 = − sin 𝜃d𝜃 and 𝜋/2 ≥ 𝜃 ≥ 𝛼; by partial fractions, we

have ∫ 𝑎 +𝑢
𝑎d𝑢 1
= log

.
𝑎2 − 𝑢2 2 𝑎−𝑢

√︃
So with 𝑎 = 1 − 𝑥12 = sin 𝜃, integrating with respect to 𝑥 2 gives
∫ 𝜋/2 ∫ cos 𝜃 tan 𝛼
1 sin 𝜃d𝑥2
d𝜃
2 𝛼 0 sin2 𝜃 − 𝑥22
sin 𝜃 + 𝑥2 cos 𝜃 tan 𝛼
∫ 𝜋/2
1
= d𝜃 log
4 𝛼 sin 𝜃 − 𝑥2 𝑥2 =0
∫ 𝜋/2
1 sin 𝜃 cos 𝛼 + cos 𝜃 sin 𝛼 (36.5.15)
= log d𝜃
4 𝛼 sin 𝜃 cos 𝛼 − cos 𝜃 sin 𝛼
∫ 𝜋/2
1 2 sin(𝜃 + 𝛼)
= log d𝜃
4 𝛼 2 sin(𝜃 − 𝛼)
1
= − (L (𝜋/2 + 𝛼) − L (2𝛼) − L (𝜋/2 − 𝛼)) .
4
Finally, we use the duplication formula (Corollary 36.5.9), which reads
L (2𝛼) = 2L (𝛼) + L (𝛼 + 𝜋/2) − L (𝜋/2 − 𝛼);
substituting gives the volume L (𝛼)/2, and summing over the other 5 triangles gives
the result.
We define now a standard tetrahedron for use in computing volumes. Let 𝑇𝛼,𝛾
be the tetrahedron with one vertex at ∞ and the other vertices 𝐴, 𝐵, 𝐶 on the unit
hemisphere projecting to 𝐴 0, 𝐵 0, 𝐶 0 in C with 𝐴 0 = 0 to make a Euclidean triangle
with angle 𝜋/2 at 𝐵 0 and 𝛼 at 𝐴 0. The dihedral angle along the ray from 𝐴 to ∞ is
𝛼. Suppose that the dihedral angle along 𝐵𝐶 is 𝛾. The acute angles determine the
isometry class of 𝑇𝛼,𝛾 , and we call 𝑇𝛼,𝛾 the standard tetrahedron with angles 𝛼, 𝛾.
For an illustration, see Figure 36.5.18.
1
vol(𝑇𝛼,𝛾 ) = L (𝛼 + 𝛾) + L (𝛼 − 𝛾) + 2L (𝜋/2 − 𝛼) .
4
Proof. One proof realizes the standard tetrahedron as a signed combination of ideal
tetrahedra, and uses Proposition 36.5.13. A second proof just repeats the integral
(36.5.15) to get
∫ 𝜋/2
1 2 sin(𝜃 + 𝛼)
vol(𝑇𝛼,𝛾 = log
d𝜃
4 𝛾 2 sin(𝜃 − 𝛼)
(36.5.17)
1
= − L (𝜋/2 + 𝛼) − L (𝛼 + 𝛾) − L (𝜋/2 − 𝛼) + L (𝛾 − 𝛼)
4
which rearranges to give the result.
B γ
C
A0
α
C0 Im x
B0
Re x
Figure 36.5.18: Standard tetrahedron
36.5.19. By Exercise 36.11, we have the Fourier expansion

∞
1 ∑︁ sin(2𝑛𝜃)
L (𝜃) = . (36.5.20)
2 𝑛=1 𝑛2
The series (36.5.20) converges rather slowly, but twice integrating the Laurent series
expansion for cot 𝜃 as in Exercise 36.12 gives
∞
!
∑︁ | 𝐵2𝑛 | (2𝜃) 2𝑛+1
L (𝜃) = 𝜃 1 − log|2𝜃| +
𝑛=1
4𝑛 (2𝑛 + 1)!
where
∞
𝑥 ∑︁ 𝑥 𝑘 𝑥 1 𝑥2 1 𝑥4
= 𝐵 𝑘 = 1 − + − +...
𝑒 𝑥 − 1 𝑘=0 𝑘! 2 6 2! 30 4!
so | 𝐵2 | = 1/6, | 𝐵4 | = 1/30, etc. are the Bernoulli numbers.
36.6 Picard modular group

In this section, analogous to the case of the classical modular group PSL2 (Z) we
consider the special case of a full Bianchi group with 𝐾 = Q(𝑖).
36.6. PICARD MODULAR GROUP 663
Definition 36.6.1. The group PSL2 (Z[𝑖]) is called the (full) Picard modular group.
Throughout this section, we write Γ = PSL2 (Z[𝑖]). In order to understand the
structure of the group Γ, we follow the same script as in section 35.1, and first we seek
a fundamental set.
Proposition 36.6.2. Let
◊ = {𝑧 = 𝑥 + 𝑦 𝑗 ∈ H3 : |𝑧| 2 ≥ 1, |Re 𝑥| ≤ 1/2, 0 ≤ Im 𝑥 ≤ 1/2}. (36.6.3)
Then ◊ is a fundamental set for Γ H3 , and PSL2 (Z[𝑖]) is generated by the elements

1 1 1 𝑖 𝑖 0 0 −1
, , , . (36.6.4)
0 1 0 1 0 −𝑖 1 0
The set ◊ in Proposition 36.6.2 is displayed in Figure 36.6.5.
Re x
Im x
Figure 36.6.5: A fundamental set ◊ for the Picard group PSL2 (Z[𝑖]) H3
Proof. First, we show that for all 𝑧 ∈ H3 , there exists a word 𝛾 in the matrices
3
(36.6.4) such that 𝛾𝑧 ∈ ◊ via an explicit reduction
algorithm. Let 𝑧 = 𝑥 + 𝑦 𝑗 ∈ H .
1 𝑏
Recalling the action (36.2.6), the element for 𝑏 ∈ Z[𝑖] act by translation
0 1
𝑧 ↦→ 𝑧 + 𝑏, so repeatedly applying matrices from the first two among
(36.6.4),
we may
𝑖 0
suppose that |Re 𝑥|, |Im 𝑥| ≤ 1/2. Then applying the element , which acts by
0 −𝑖
𝑧 ↦→ (𝑖𝑧) (−𝑖) −1 = 𝑖 2 (𝑥 − 𝑦 𝑗) = −𝑥 + 𝑦 𝑗, we may suppose Im 𝑥 ≥ 0. Now if 𝑧 ∈ ◊,
2 2
which k𝑧k ≥ 1, we are done. Otherwise, k𝑧k < 1, and we apply the matrix
is to say
0 −1
𝛾= , which by (36.2.9) acts by
1 0
1
𝛾𝑧 = 𝑧 0 = (−𝑥 + 𝑦 𝑗) = 𝑥 0 + 𝑦 0 𝑗
k𝑧k 2
and 𝑦 0 = 𝑦/k𝑧k 2 > 𝑦, and we repeat. Since Γ𝑧 is discrete, this terminates after finitely
many steps: the set
Γ𝑧 ∩ 𝑧 0 = 𝑥 0 + 𝑦 0 𝑗 ∈ H3 : |Re 𝑥 0 |, |Im 𝑥 0 | ≤ 1/2, 𝑦 ≤ 𝑦 0 ≤ 1

is discrete and compact, hence finite.

Next, if 𝑧, 𝑧 0 ∈ ◊ and 𝑧 ∈ int(◊) with 𝑧 0 = 𝛾𝑧 for 𝛾 ∈ Γ, then 𝛾 = 1 and 𝑧 = 𝑧 0;
this can be proven directly as in Lemma 35.1.10 (the details are requested in Exercise
36.7). It follows that the matrices (36.6.4) generate PSL2 (Z[𝑖]) as in Lemma 35.1.12,
taking instead 𝑧0 = 2 𝑗 ∈ int(◊).
A slightly more convenient set of generators, together with the gluing relations
they provide on the fundamental set, is given in Figure 36.6.7.
Remark 36.6.6. By a deeper investigation into the structure of the fundamental set ◊,
in chapter 37 we will find a presentation for Γ as
Γ ' h𝛾1 , 𝛾2 , 𝛾3 , 𝛾4 | 𝛾12 = 𝛾22 = 𝛾42 = 1,

(𝛾3 𝛾1 ) 3 = (𝛾3 𝛾2 ) 2 = (𝛾3 𝛾4 ) 2 = (𝛾2 𝛾1 ) 2 = (𝛾4 𝛾1 ) 3 = 1i.
y
γ3
i 0

γ2 = 0 −i
i 1

γ4 = 0 −i
√
2
− 21 + 21 i + 2 j
j
0 −1
γ1 = 1 0
√
1 3
2 + 2 j
√
γ2 1
+ 12 i + 2
2 2 j
γ1 γ4
Im x
γ3
Re x
Figure 36.6.7: Generators for Picard group PSL2 (Z[𝑖]) H3

36.6.8. We compute the volume of this fundamental domain using formulas from the
previous section. First, we use symmetry to triangulate (tetrahedralize) ◊, as in Figure
36.6.9.
T
◊
x2 = Im x
x1 = Re x
Figure 36.6.9: Triangulation of ◊

Let
𝑇 := {𝑧 = 𝑥1 + 𝑥 2𝑖 + 𝑦 𝑗 ∈ H3 : 0 ≤ 𝑥2 ≤ 𝑥1 ≤ 1/2, 𝑥12 + 𝑥 22 + 𝑦 2 ≥ 1}. (36.6.10)
Applying the symmetries of Γ, we see that vol(◊) = 4 vol(𝑇). We have 𝑇 = 𝑇𝛼,𝛾 a
standard tetrahedron, with 𝛼 = 𝜋/4 and dihedral angle 𝛾 = 𝜋/3.
Now by the hard-earned volume formula (Corollary 36.5.16) we have
1
vol(𝑇) = L (𝜋/4 + 𝜋/3) + L (𝜋/4 − 𝜋/3) + 2L (𝜋/4) . (36.6.11)
4
By Lemma 36.5.5 with 𝑛 = 3, we have
1
L (3𝜋/4) = L (𝜋/4) + L (𝜋/4 + 𝜋/3) + L (𝜋/4 + 2𝜋/3); (36.6.12)
3
since
L (3𝜋/4) = L (𝜋 − 𝜋/4) = −L (𝜋/4)
L (𝜋/4 + 2𝜋/3) = L (𝜋/4 + 2𝜋/3 − 𝜋) = L (𝜋/4 − 𝜋/3),
substituting (36.6.12) into (36.6.11) gives

1 1 1
vol(𝑇) = 2−1− L (𝜋/4) = L (𝜋/4) = 0.07633 . . . (36.6.13)
4 3 6
and
2
vol(◊) = 4 vol(𝑇) = L (𝜋/4) = 0.30532 . . .
3
36.6.14. We conclude with a beautiful consequence of this volume calculation, giving

a preview of the volume formula we will prove later. By the Fourier expansion
(36.5.20), we have
∞ ∞
1 ∑︁ sin(𝑛𝜋/2) 1 ∑︁ 𝜒(𝑛)
L (𝜋/4) = = (36.6.15)
2 𝑛=1 𝑛2 2 𝑛=1 𝑛2
where


 0, if 2 | 𝑛;


𝜒(𝑛) = 1, if 𝑛 ≡ 1 (mod 4);

 −1,

if 𝑛 ≡ −1 (mod 4).

is the nontrivial Dirichlet character modulo 4. We can analytically continue the sum
(36.6.15) to C via the 𝐿-series
∞
∑︁ 𝜒(𝑛)
𝐿(𝑠, 𝜒) =
𝑛=1
𝑛𝑠
for 𝑠 ∈ C with Re 𝑠 > 1 whose general study was the heart of Part III of this text. Here,
we can just observe that 𝐿 (2, 𝜒) = 2L (𝜋/4) = 0.915965 . . ., so
1
vol(◊) = vol(Γ\H3 ) = 𝐿(2, 𝜒) = 0.30532 . . .
3
More generally, the volume of the quotient by a Bianchi group is connected to an

𝐿-value attached to the associated imaginary field; we will pursue this topic further in
chapter 39.
Exercises

𝑎 𝑏
1. For 𝑧 ∈ H3 and 𝑔 = ∈ PSL2 (C), show that
𝑐 𝑑
(𝑎𝑧 + 𝑏) (𝑐𝑧 + 𝑑) −1 = (𝑧𝑐 + 𝑑) −1 (𝑧𝑎 + 𝑏).
I 2. Inversion in the circle of radius 𝑟 in C centered at the origin is defined by the

map
𝑟2

𝑧
𝑧 ↦→ 𝑟 2 2
= ,
| 𝑧| 𝑧
z0
Figure 36.6.16: Inversion in the circle of radius 𝑟
Sending 0 ↦→ ∞ and ∞ ↦→ 0 under this map, we obtain an anti-holomorphic

map P1 (C) → P1 (C). Inversion in a line in C is reflection in the line.
Verify that every element of PSL2 (C) can be written as a composition of at
most four inversions in circles and lines in C (or equivalently, by stereographic
projection, circles in P1 (C)).
I 3. Verify (36.2.3) for the action of SL2 (C) on H3 ⊆ H.

0 −1
I 4. Let 𝑔 = ∈ SL2 (C). Show that 𝑔 acts on H3 by
1 0
1
𝑔𝑧 = (−𝑥 + 𝑦 𝑗)
k𝑧k 2
where k𝑧k 2 = |𝑥| 2 + 𝑦 2 , and that 𝑔 is a hyperbolic isometry.

5. Let Γ ≤ PSL2 (C) be a subgroup (with the subspace topology). Show that Γ has
a wandering action on H3 if and only if Γ is discrete (cf. Proposition 34.7.2).
6. Show that the reduction algorithm in Proposition 36.6.2 recovers the Euclidean
algorithm for Z[𝑖] in a manner analogous to Exercise 35.4 for Z.
I 7. Consider the fundamental set ◊ for Γ = PSL2 (Z[𝑖]) (36.6.3). Show that if
𝑧, 𝑧 0 ∈ ◊ and 𝑧 ∈ int(◊) with 𝑧 0 = 𝛾𝑧 for 𝛾 ∈ Γ, then 𝛾 = 1 and 𝑧 = 𝑧 0 (cf.
Lemma 35.1.10).
√
8. Let 𝜔 = 𝑒 2 𝜋𝑖/3 ∈ C. The field Q(𝜔) = Q( −3) is Euclidean under the norm,
just like Q(𝑖). Give a description of a fundamental domain √ for the group
PSL2 (Z[𝜔]) analogous to Proposition 36.6.2. [The fields Q( 𝑑) with 𝑑 < 0 are
Euclidean if and only if 𝑑 = −3, −4, −7, −8, −11, so similar—but increasingly
difficult—arguments can be given in each of these cases. See Fine [Fin89, §4.3]
for presentations.]
I 9. We consider hyperplane bisectors in H3 (cf. Exercise 33.8). Let 𝑧1 , 𝑧2 ∈ H3 be

distinct. Let
𝐻 (𝑧 1 , 𝑧2 ) = {𝑧 ∈ H3 : 𝜌(𝑧, 𝑧1 ) ≤ 𝜌(𝑧, 𝑧2 )}
be the locus of points as close to 𝑧1 as to 𝑧2 , and let
𝐿(𝑧1 , 𝑧2 ) = bd 𝐻 (𝑧 1 , 𝑧2 ).
Show that 𝐻 (𝑧1 , 𝑧2 ) is a convex half-space (for every two points in the half-space,
the geodesic between them is contained in the half-space), and that
𝐿(𝑧 1 , 𝑧2 ) = {𝑧 ∈ H3 : 𝜌(𝑧, 𝑧1 ) = 𝜌(𝑧, 𝑧2 )}
is geodesic and equal to the perpendicular bisector of the geodesic segment from
𝑧1 to 𝑧2 .
10. Prove the duplication formula for the Lobachevsky function L (𝜃) using the
double angle formula, given that L (𝜋/2) = 0.
I 11. In this exercise, we prove the Fourier expansion
∞
1 ∑︁ sin(2𝑛𝜃)
L (𝜃) = . (36.6.17)
2 𝑛=1 𝑛2
(a) Define the dilogarithm function by

∞
∑︁ 𝑧𝑛
Li2 (𝑧) = ;
𝑛=1
𝑛2
show that this series converges for | 𝑧| < 1 and that

∫ 𝑧
log(1 − 𝑤)
Li2 (𝑧) = − d𝑤.
0 𝑤
(b) Prove that
𝜋2
+ 𝜋𝜃 − 𝜃 2 .
2𝑖L (𝜃) = Li2 (𝑒 2𝑖 𝜃 ) − (36.6.18)
6
[Hint: Differentiate both sides for 0 < 𝜃 < 𝜋, using the limiting value as
𝜃 → 0 to compute the limiting value Li2 (1) = 𝜋 2 /6.]
(c) Take imaginary parts of (36.6.18) to prove (36.6.17).
I 12. As in 36.5.19, we define the 𝐵 𝑘 ∈ Q for 𝑘 ≥ 0 by
∞
𝑥 ∑︁ 𝑥 𝑘 𝑥 1 𝑥2 1 𝑥4
= 𝐵 𝑘 = 1 − + − +.... (36.6.19)
𝑒 𝑥 − 1 𝑘=0 𝑘! 2 6 2! 30 4!
(a) Plug in 𝑥 = 2𝑖𝑧 into (36.6.19) to get

∞
∑︁ (2𝑖𝑧) 𝑘
𝑧 cot 𝑧 = 1 + 𝐵𝑘 .
𝑘=2
𝑘!
(b) Integrate twice in (a) to prove

∞
!
∑︁ | 𝐵2𝑛 | (2𝜃) 2𝑛+1
L (𝜃) = 𝜃 1 − log|2𝜃| + .
𝑛=1
4𝑛 (2𝑛 + 1)!
[See also Exercise 40.3.]

Chapter 37
Fundamental domains
We have seen in sections 35.1 and 36.6 that understanding a nice fundamental set for
the action of a discrete group Γ is not only useful to visualize the action of the group
by selecting representatives of the orbits, but it is also instrumental for many other
purposes, including understanding the structure of the group itself. In this chapter, we
pursue a general construction of nice fundamental domains for the action of a discrete
group of isometries.
37.1 ⊲ Dirichlet domains for Fuchsian groups

In this introductory section, we preview the results in this chapter specialized to the
case of Fuchsian groups. Let Γ ⊂ PSL2 (R) be a Fuchsian group; then Γ is discrete,
acting properly by isometries on the hyperbolic plane H2 , with metric 𝜌(·, ·) and
hyperbolic area 𝜇.
A natural way to produce fundamental sets that provide appealing tessellations of
H2 is to select in each orbit the points closest to a fixed point 𝑧0 ∈ H2 , as follows.
Definition 37.1.1. The Dirichlet domain for Γ centered at 𝑧0 ∈ H2 is
◊(Γ; 𝑧 0 ) = {𝑧 ∈ H2 : 𝜌(𝑧, 𝑧0 ) ≤ 𝜌(𝛾𝑧, 𝑧0 ) for all 𝛾 ∈ Γ}.
As the group Γ will not vary, we suppress the dependence on Γ and often write
simply ◊(𝑧 0 ) = ◊(Γ; 𝑧0 ).
37.1.2. The set ◊(𝑧0 ) is an intersection
Ù
◊(𝑧0 ) = 𝐻 (𝛾; 𝑧0 ) (37.1.3)
𝛾 ∈Γ
where
𝐻 (𝛾; 𝑧0 ) := {𝑧 ∈ H2 : 𝜌(𝑧, 𝑧0 ) ≤ 𝜌(𝛾𝑧, 𝑧0 ) = 𝜌(𝑧, 𝛾 −1 𝑧 0 )}. (37.1.4)
In particular, since each 𝐻 (𝛾; 𝑧0 ) is closed, we conclude from (37.1.3) that ◊(𝑧 0 ) is
closed.
671
672 CHAPTER 37. FUNDAMENTAL DOMAINS
The sets 𝐻 (𝛾; 𝑧0 ) can be further described as follows. If 𝑧0 = 𝛾 −1 𝑧0 , then

𝐻 (𝛾; 𝑧0 ) = H2 . So suppose 𝑧0 ≠ 𝛾 −1 𝑧 0 . Then by Exercise 33.8, 𝐻 (𝛾; 𝑧 0 ) is a
(half!) half-plane consisting of the set of points as close to 𝑧 0 as 𝛾 −1 𝑧 0 , and 𝐻 (𝛾; 𝑧0 ) is
convex: if two points lie in the half-plane then so does the geodesic segment between
them. The boundary
bd 𝐻 (𝛾; 𝑧0 ) = 𝐿 (𝛾; 𝑧 0 ) := {𝑧 ∈ H2 : 𝜌(𝑧, 𝑧0 ) = 𝜌(𝑧, 𝛾 −1 𝑧0 )}
is the geodesic perpendicular bisector of the geodesic segment from 𝑧0 to 𝛾 −1 𝑧 0 , as in
Figure 37.1.5.
H(γ; z0 )
L(γ; z0 )
z0
γ −1 z0
Figure 37.1.5: The half-plane 𝐻 (𝛾; 𝑧 0 ) and its boundary 𝐿(𝛾; 𝑧0 )

From the description in 37.1.2, the sketch of a Dirichlet domain looks like Figure
37.1.6.
◊(z0 )
Figure 37.1.6: Sketch of a Dirichlet domain

The fundamental sets we have seen are in fact examples of Dirichlet domains.
Example 37.1.7. We claim that the Dirichlet domain for Γ = PSL2 (Z) centered at
𝑧0 = 2𝑖 is in fact the fundamental set for Γ introduced in section 35.1, i.e.,
◊(Γ; 𝑧0 ) = ◊(2𝑖) = {𝑧 ∈ H2 : |Re 𝑧| ≤ 1/2, |𝑧| ≥ 1}. (37.1.8)
Recall the generators 𝑆, 𝑇 ∈ Γ with 𝑆𝑧 = −1/𝑧 and 𝑇 𝑧 = 𝑧 + 1. By (33.5.3)
|𝑧 − 2𝑖| 2
cosh 𝜌(𝑧, 2𝑖) = 1 + .
4 Im 𝑧
37.1. ⊲ DIRICHLET DOMAINS FOR FUCHSIAN GROUPS 673
Let 𝑧 ∈ H2 . Visibly
𝜌(𝑧, 2𝑖) ≤ 𝜌(𝑇 𝑧, 2𝑖) ⇔ Re 𝑧 ≥ −1/2 (37.1.9)
or put another way

𝐻 (𝑇; 2𝑖) = {𝑧 ∈ H2 : Re 𝑧 ≥ −1/2}.
Similarly, 𝐻 (𝑇 −1 ; 2𝑖) = {𝑧 ∈ H2 : Re 𝑧 ≤ 1/2}. Equivalently, the geodesic perpendic-
ular bisector between 2𝑖 and 2𝑖 ± 1 are the lines Re 𝑧 = ±1/2.
In the same manner, we find that
|𝑧 − 2𝑖| 2 |(−1/𝑧) − 2𝑖| 2 4|𝑧| 2 |𝑧 − 𝑖/2| 2

𝜌(𝑧, 2𝑖) ≤ 𝜌(𝑆𝑧, 2𝑖) ⇔ ≤ =
Im 𝑧 Im(−1/𝑧) |𝑧| 2 Im 𝑧
⇔ |𝑧 − 2𝑖| 2 ≤ 4|𝑧 − 𝑖/2| 2
so
𝜌(𝑧, 2𝑖) ≤ 𝜌(𝑆𝑧, 2𝑖) ⇔ |𝑧| ≥ 1 (37.1.10)
and 𝐻 (𝑆; 2𝑖) = {𝑧 ∈ H2 : |𝑧| ≥ 1}. To see this geometrically, the geodesic between
2𝑖 and 𝑆(2𝑖) = (1/2)𝑖 is along the imaginary axis with midpoint at 𝑖, and so the
perpendicular bisector 𝐿(𝑆; 2𝑖) is the unit semicircle.
The containment (⊆) in (37.1.8) then follows directly from (37.1.9)–(37.1.10).
Conversely, we show the containment (⊇) for the interior—since ◊(2𝑖) is closed, this
implies the full containment. Let 𝑧 ∈ H2 have |Re 𝑧| < 1/2 and |𝑧| > 1, and suppose
that 𝑧 ∉ ◊(2𝑖); then there exists 𝛾 ∈ PSL2 (Z) such that 𝑧 0 = 𝛾𝑧 has 𝜌(𝑧 0, 2𝑖) < 𝜌(𝑧, 2𝑖),
without loss of generality (replacing 𝑧 0 by 𝑆𝑧 0 or 𝑇 𝑧 0) we may suppose |Re 𝑧 0 | ≤ 1/2
and |𝑧 0 | ≥ 1; but then by Lemma 35.1.10, we conclude that 𝑧 0 = 𝑧, a contradiction.
(Note that the same argument works with 𝑧0 = 𝑡𝑖 for all 𝑡 ∈ R>1 .)
With this example in hand, we see that Dirichlet domains have quite nice structure.
To make this more precise, we upgrade our notion of fundamental set (Definition
34.1.14) as follows.
Definition 37.1.11. A fundamental set ◊ for Γ is locally finite if for each compact set
𝐾 ⊂ H2 , we have 𝛾𝐾 ∩ ◊ ≠ ∅ for only finitely many 𝛾 ∈ Γ.
A fundamental domain for Γ H2 is a fundamental set ◊ ⊆ H2 such that
𝜇(bd ◊) = 0.
The first main result of this section is as follows (Theorem 37.5.3).
Theorem 37.1.12. Let 𝑧 0 ∈ H2 satisfy StabΓ (𝑧0 ) = {1}. Then the Dirichlet domain
◊(Γ; 𝑧0 ) is a connected, convex, locally finite fundamental domain for Γ with geodesic
boundary.
By geodesic boundary we mean that the boundary bd ◊(𝑧0 ) is a finite or countable

union of geodesic segments. As for the hypothesis: for a compact set 𝐾 ⊂ H2 , there
are only finitely many points 𝑧 ∈ 𝐾 such that StabΓ (𝑧) ≠ {1}, indeed there are only
finitely many 𝛾 ∈ Γ such that 𝛾𝐾 ∩ 𝐾 ≠ ∅ (as Γ is discrete), and every such 𝛾 ≠ 1 has
at most one fixed point in H2 (Lemma 33.4.6).
37.2 ⊲ Ford domains

In this section, we reinterpret Dirichlet domains in the unit disc D2 , as it is more
convenient to compute and visualize distances this model. Let 𝑧 0 ∈ H2 . We apply the
map (33.7.3)
𝜙 : H2 → D2
𝑧 − 𝑧0
𝑤=
𝑧 − 𝑧0
with 𝑧 0 ↦→ 𝜙(𝑧0 ) = 𝑤 0 = 0. Then by (33.7.6),
1 + |𝑤|
𝜌(𝑤, 0) = log = 2 tanh−1 |𝑤| (37.2.1)
1 − |𝑤|
is an increasing function of |𝑤|.
Example 37.2.2. The Dirichlet domain from Example 37.1.7 looks like Figure 37.2.3,
drawn in D2 (with 𝑧0 = 2𝑖).
φ(ω)
φ(i) 0 = φ(2i) φ(∞)
φ(−ω 2 )
Figure 37.2.3: Dirichlet domain in D2

Let Γ ≤ PSL2 (R) be a Fuchsian group, and (recalling 33.7.8) to ease notation, we
identify Γ with Γ 𝜙 . We analogously define a Dirichlet domain ◊(𝑤 0 ) for a Fuchsian
group Γ centered at 𝑤 0 ∈ D2 and
𝜙(◊(𝑧0 )) = ◊(𝑤 0 ) ⊂ D2
if 𝜙(𝑧0 ) = 𝑤 0 . In particular, the statement of Theorem 37.1.12 applies equally well to
◊(𝑤 0 ) ⊆ D2 .
For simplicity (and without loss of generality), we only consider the case where
𝑤 0 = 0, and then from (37.2.1)
◊(Γ; 0) = {𝑤 ∈ D2 : |𝑤| ≤ |𝛾𝑤| for all 𝛾 ∈ Γ}. (37.2.4)
Let
𝑑 𝑐
𝑔= ∈ PSU(1, 1) D2
𝑐 𝑑
with 𝑐, 𝑑 ∈ C satisfying |𝑑| 2 − |𝑐| 2 = 1.
37.2. ⊲ FORD DOMAINS 675
37.2.5. We now pursue a tidy description of the set (37.2.4). From (37.2.1), we have
𝜌(𝑤, 0) ≤ 𝜌(𝑔𝑤, 0) if and only if

𝑑𝑤 + 𝑐
|𝑤| ≤ ; (37.2.6)

𝑐𝑤 + 𝑑
expanding out (37.2.6) and with a bit of patience (Exercise 37.5), we see that this is
equivalent to simply
|𝑐𝑤 + 𝑑| ≥ 1.
But we can derive this more conceptually, as follows. The hyperbolic metric (Definition
33.7.1) on D2 is invariant, so
|d𝑤| |d(𝑔𝑤)|
d𝑠 = = = d(𝑔𝑠);
(1 − |𝑤|) 2 (1 − |𝑔𝑤|) 2
so, by the chain rule,
2
d𝑔 1 − |𝑔𝑤|
d𝑤 (𝑤) = 1 − |𝑤| .

Therefore

d𝑔 1
|𝑤| ≤ |𝑔𝑤| ⇔ d𝑤 (𝑤) = |𝑐𝑤 + 𝑑| 2 ≤ 1
⇔ |𝑐𝑤 + 𝑑| ≥ 1. (37.2.7)

d𝑔
The equivalence (37.2.7) also gives 𝜌(𝑤, 0) = 𝜌(𝑔𝑤, 0) if and only if (𝑤) = 1,
d𝑤
i.e., 𝑔 acts as a Euclidean isometry at the point 𝑤 (preserving the length of tangent
vectors in the Euclidean metric). So we are led to make the following definition.
Definition 37.2.8. The isometric circle of 𝑔 is

d𝑔
𝐼 (𝑔) = 𝑤 ∈ C : (𝑤) = 1 = {𝑤 ∈ C : |𝑐𝑤 + 𝑑| = 1}.
d𝑤
We have 𝑐 = 0 if and only if 𝑔(0) = 0 if and only if 𝑔 ∈ StabPSU(1,1) (0), and in this
case, 𝑔𝑤 = (𝑑/𝑑)𝑤 with |𝑑/𝑑| = 1 is rotation about the origin. Otherwise, 𝑐 ≠ 0, and
then 𝐼 (𝑔) is a circle with radius 1/|𝑐| and center −𝑑/𝑐 ∈ C.
37.2.9. Summarizing, for all 𝑔 ∈ PSU(1, 1), we have


 <
 
 ext(𝐼 (𝑔)),
 
 
 


𝜌(𝑤, 0) = 𝜌(𝑔𝑤, 0) according as 𝑤 ∈ 𝐼 (𝑔),
  

>  int(𝐼 (𝑔)).

  
In particular, Ù
◊(Γ; 0) = cl ext 𝐼 (𝛾).
𝛾 ∈Γ−StabΓ (0)
This characterization is due to Ford [For72, Theorem 7, §20].

This description of a Dirichlet domain as the intersection of the exteriors of iso-

metric circles is due to Ford, and so we call a Dirichlet domain in D2 centered at 0 a
Ford domain, as in Figure 37.2.10.
Figure 37.2.10: A typical Ford domain
Remark 37.2.11. In the identification H2 → D2 , the preimage of isometric circles

corresponds to the corresponding perpendicular bisector; this is the simplification
provided by working in D2 (the map 𝜙 is a hyperbolic isometry, whereas isometric
circles are defined by a Euclidean condition).
37.3 ⊲ Generators and relations

Continuing with our third and final survey section focused on Fuchsian groups, we
consider applications to the structure of a Fuchsian group Γ. For more, see Beardon
[Bea95, §9.3] and Katok [Kat92, §3.5].
Let ◊ = ◊(Γ; 𝑧 0 ) be a Dirichlet domain centered at 𝑧0 ∈ H2 . A consequence of
the local finiteness of a Dirichlet domain is the following theorem (Theorem 37.4.2).
Theorem 37.3.1. Γ is generated by the set
{𝛾 ∈ Γ : ◊ ∩ 𝛾◊ ≠ ∅}.
So by Theorem 37.3.1, to find generators, we must look for “overlaps” in the

tessellation provided by ◊. If 𝑧 ∈ ◊ ∩ 𝛾◊ with 𝛾 ∈ Γ r {1}, then there exists 𝑧, 𝑧 0 ∈ ◊
such that 𝑧 = 𝛾𝑧 0, and hence
𝜌(𝑧, 𝑧0 ) ≤ 𝜌(𝛾 −1 𝑧, 𝑧0 ) = 𝜌(𝑧 0, 𝑧0 ) ≤ 𝜌(𝛾𝑧 0, 𝑧0 ) = 𝜌(𝑧, 𝑧0 ) (37.3.2)
(twice applying the defining property of ◊), so equality holds and (viz. 37.1.2) 𝑧 ∈
bd ◊. Since the boundary of ◊ is geodesic, to understand generators we should
organize the matching provided along the geodesic boundary of ◊.
We will continue to pass between H2 and D2 , as convenient.
37.3. ⊲ GENERATORS AND RELATIONS 677
37.3.3. A side is a geodesic segment of positive length of the form ◊ ∩ 𝛾◊ with

𝛾 ∈ Γ r {1}. If two sides intersect in D2 , the point of intersection is called a vertex of
◊; equivalently, a vertex is a single point of the form ◊ ∩ 𝛾◊ ∩ 𝛾 0◊ with 𝛾, 𝛾 0 ∈ Γ.
However, we make the following convention (to simplify the arguments below): if
𝐿 = ◊ ∩ 𝛾◊ is a maximal geodesic subset of ◊ and 𝛾 2 = 1, or equivalently if 𝛾𝐿 = 𝐿,
then 𝛾 fixes the midpoint of 𝐿, and we consider 𝐿 to be the union of two sides that
meet at the vertex equal to the midpoint. The representation of a side as ◊ ∩ 𝛾◊ is
unique when 𝛾 2 ≠ 1.
Because ◊ is locally finite, there are only finitely many vertices in a compact
neighborhood (Exercise 37.7).
An ideal vertex is a point of the closure of ◊ in D2∗ that is the intersection of the
closure of two sides in D2∗ , as in Figure 37.3.4.
ideal vertex
vertex
Figure 37.3.4: A vertex and an ideal vertex
37.3.5. Let 𝑆 denote the set of sides of ◊. We define a labeled equivalence relation
on 𝑆 by
𝑃 = {(𝛾, 𝐿, 𝐿 ∗ ) : 𝐿 ∗ = 𝛾(𝐿)} ⊂ Γ × (𝑆 × 𝑆). (37.3.6)
We say that 𝑃 is a side pairing if 𝑃 induces a partition of 𝑆 into pairs, and we denote
by 𝐺 (𝑃) the projection of 𝑃 to Γ. Since (𝛾, 𝐿, 𝐿 ∗ ) ∈ 𝑃 implies (𝛾 −1 , 𝐿 ∗ , 𝐿) ∈ 𝑃, we
conclude that 𝐺 (𝑃) is closed under inverses.
Lemma 37.3.7. A Dirichlet domain ◊ has a side pairing 𝑃.
Proof. Let 𝐿 be a side with 𝐿 ⊆ ◊ ∩ 𝛾◊ for a unique 𝛾. Recalling the convention in

37.3.3, if 𝛾 2 ≠ 1, then equality holds, and
𝛾 −1 𝐿 = ◊ ∩ 𝛾 −1 ◊ = 𝐿 ∗ ≠ 𝐿
so by uniqueness, the equivalence class of 𝐿 contains only 𝐿, 𝐿 ∗ . If 𝛾 2 = 1, then 𝐿

meets 𝛾𝐿 = 𝐿 ∗ at the fixed point of 𝛾, and again the equivalence class of 𝐿 contains
on 𝐿, 𝐿 ∗ . In either case, we conclude that 𝑃 (37.3.6) is a side pairing.
37.3.8. We now provide a standard picture of ◊ in a neighborhood of a point

𝑤 ∈ bd ◊.
Because ◊ is locally finite, there is an an open neighborhood
Ð𝑛 of 𝑤 and finitely
many distinct 𝛿0 , 𝛿1 , . . . , 𝛿 𝑛 ∈ Γ with 𝛿0 = 1 such that 𝑈 ⊆ 𝑖=0 𝛿𝑖 ◊ and 𝑤 ∈ 𝛿𝑖 ◊ for
all 𝑖. Shrinking 𝑈 if necessary, we may suppose that 𝑈 contains no vertices of ◊ except
possibly for 𝑤 and intersects no sides of 𝛿𝑖 ◊ except those that contain 𝑤. Therefore,
we have a picture as in Figure 37.3.9.
δ2 ◊
δ1 ◊
···
δ1 ◊
w ◊
◊ = δ0 ◊
δn ◊
Figure 37.3.9: Standard picture
When 𝑛 = 1, then either 𝑤 can be either a vertex (fixed point of 𝛿1 ) or not.
37.3.10. Let 𝑣 = 𝑣 1 be a vertex of ◊. The standard picture in a neighborhood of 𝑣 can

be reinterpreted as in Figure 37.3.11.
L∗1 = ◊ ∩ γ1 ◊ ν = γ ν
2 1 1
L1 = ◊ ∩ γ1−1 ◊ γ2 L2
γ1
ν = ν1 ..
◊ .
γm γm−1
L∗m
L∗m−1
Lm νm




y
γ1−1 γ2−1 ◊ = δ2
δm ◊ = γ1−1 · · · γm
−1
◊
···
δm γ1−1 ◊ = δ1
◊
···
Figure 37.3.11: The standard picture in a neighborhood of 𝑣

37.3. ⊲ GENERATORS AND RELATIONS 679
Let 𝐿 1 be the side containing 𝑣 1 traveling clockwise from the interior. Then by the
side pairing (Lemma 37.3.7), there is a paired side 𝐿 1∗ with 𝐿 1∗ = 𝛾1 𝐿 1 and 𝛾1 ∈ 𝐺 (𝑃).
(In fact, then 𝐿 1 = ◊ ∩ 𝛾1−1 ◊ and 𝐿 1∗ = ◊ ∩ 𝛾1 ◊.) Let 𝑣 2 = 𝛾1 𝑣 1 . Then 𝑣 2 is a vertex
of ◊, and so is contained in a second side 𝐿 2 . Continuing in this way, with 𝑣 𝑖+1 = 𝛾𝑖 𝑣 𝑖 ,
by local finiteness we find after finitely many steps a final side 𝐿 ∗𝑚 with next vertex
𝑣 𝑚 = 𝑣1.
In terms of the standard picture (Figure 37.3.9), we see that 𝛿1 = 𝛾1−1 and by
induction 𝛿𝑖 = (𝛾𝑖 · · · 𝛾1 ) −1 , since (𝛾𝑖 · · · 𝛾1 ) (𝑣 1 ) = 𝑣 𝑖+1 . Thus 𝛾𝑖+1 = 𝛿𝑖+1
−1 𝛿 for
𝑖
𝑖 = 0, . . . , 𝑚 − 1. Let 𝛿 = 𝛿 𝑚 . Then 𝛿(𝑣) = 𝑣, and 𝛿 acts by counterclockwise
hyperbolic rotation with fixed point 𝑣—and 𝑚 is the smallest nonzero index with
this property. It follows that for all 0 ≤ 𝑗 ≤ 𝑛, writing 𝑗 = 𝑞𝑚 + 𝑖 with 𝑞 ≥ 0
𝑞
and 0 ≤ 𝑖 < 𝑚 we have 𝛿 𝑗 = 𝛿 𝑚 𝛿𝑖 , and in particular that 𝑚 | (𝑛 + 1). Similarly,
◊ ∩ Γ𝑣 1 = {𝑣 1 , . . . , 𝑣 𝑚−1 }.
Let 𝑒 = (𝑛 + 1)/𝑚. Then 𝛿 𝑒 = 1, and we call this relation the vertex cycle
relation for 𝑣. If 𝑣 0 = 𝛾𝑣, then the vertex cycle relation for 𝑣 0 is the conjugate relation
(𝛾 −1 𝛿𝛾) 𝑒 = 𝛾 −1 𝛿 𝑒 𝛾 = 1. Let 𝑅(𝑃) be the set of (conjugacy classes of) cycle relations
arising from Γ-orbits of vertices in ◊.
Example 37.3.12. We compute the set 𝑅(𝑃) of cycle relations for Γ = PSL2 (Z). The
two Γ-orbits of vertices for ◊ are represented by 𝑖 and 𝜌. The vertex 𝑖 is a fixed point
of 𝛿1 = 𝑆, and we obtain the vertex cycle relation 𝑆 2 = 1, and 𝑆 = 𝑆 −1 .
At the vertex 𝜌, we have a picture as in Figure 37.3.13.
δ1 ◊ = T −1 ◊ δ0 ◊ = ◊
T = γ1
L1 L∗1 δ
δ2 ◊ = T −1 S◊
L∗2 i L2
ρ = ν1 S = γ2 ν2 = −ρ2 ST −1 S◊
Figure 37.3.13: The cycle relation at 𝜌
We find that 𝛿1 = 𝑇 −1 and 𝛿 = 𝛿2 = 𝑇 −1 𝑆, and 𝑒 = 6/2 = 3 so 𝛿3 = (𝑇 −1 𝑆) 3 = 1.

Taking inverses (so 𝛿−1 acting instead clockwise), we find the relation (𝑆𝑇) 3 = 1.
Proposition 37.3.14. The set 𝐺 (𝑃) generates Γ with 𝑅(𝑃) a set of defining relations.
In other words, the free group on 𝐺 (𝑃) modulo the normal subgroup generated by the
relations 𝑅(𝑃) is isomorphic to Γ via the natural evaluation map.
Proof. Let Γ∗ ≤ Γ be the subgroup generated by 𝐺 (𝑃). By Theorem 37.3.1, we need

to show that if ◊ ∩ 𝛾◊ ≠ ∅ then 𝛾 ∈ Γ∗ . So let 𝑤 ∈ ◊ ∩ 𝛾◊ with 𝛾 ∈ Γ − {1}. We refer
to the standard picture (see 37.3.8); we have 𝛾 = 𝛿 𝑗 for some 𝑗. For all 𝑖 = 0, 1, . . . , 𝑛,
the intersection 𝛿𝑖 ◊ ∩ 𝛿𝑖+1 ◊ is a side of 𝛿𝑖+1 ◊, so ◊ ∩ 𝛿𝑖+1 −1 𝛿 ◊ is a side of ◊, and
𝑖
−1 𝛿 ∈ 𝐺 (𝑃) is a side pairing element. Since 𝛿 = 1, by induction we find that
thus 𝛿𝑖+1 𝑖 0
𝛿𝑖 ∈ Γ∗ for all 𝑖, so 𝛾 = 𝛿 𝑗 ∈ Γ∗ as claimed.
We now turn to relations, and we give an algorithmic method for rewriting a relation
in terms of the cycle relations. Let 𝛾1 𝛾2 · · · 𝛾 𝑘 = 1 be a relation with each 𝛾𝑖 ∈ 𝐺 (𝑃),
and let 𝑧𝑖 = (𝛾1 · · · 𝛾𝑖 )𝑧 0 for 𝑖 = 1, . . . , 𝑘. Exactly because 𝛾1 ∈ 𝐺 (𝑃), we have that ◊

and 𝛾1 ◊ share a side, and since ◊ is connected, we can draw a path 𝑧 0 → 𝑧 1 through
the corresponding side. Continuing in this way, we end up with a path 𝑧0 → 𝑧 𝑘 = 𝑧0 ,
hence a closed loop, as in Figure 37.3.15.
z3
.. z2
.
z1
zk−2 z0 = zk
zk−1
Figure 37.3.15: A closed loop
Let 𝑉 be the intersection of the Γ orbit of the vertices of ◊ with the interior of the
loop; by local finiteness, this intersection is a finite set, and we proceed by induction
on its cardinality. The proof boils down to the fact that this loop retracts onto the loops
around vertices obtained from cycle relations, as H2 is simply connected.
If the path from 𝑧0 → 𝑧 1 crosses the same side as the path 𝑧 𝑘−1 → 𝑧 𝑘 = 𝑧0 , then
𝑧1 = 𝑧 𝑘−1 and so 𝛾 𝑘−1 = 𝛾1 , since StabΓ (𝑧0 ) = {1} (see Figure 37.3.16).
z1 = zk−1
z0 = zk
Figure 37.3.16: Simplifying a relation: setup
Conjugating the relation by 𝛾1 and repeating if necessary, we may suppose that 𝛾 𝑘−1 ≠
𝛾1 , so 𝑧 𝑘−1 ≠ 𝑧1 ; the set 𝑉 is conjugated so it remains the same size. In particular, if 𝑉
is empty, then this shows that the original relation is conjugate to the trivial relation.
Otherwise, the path 𝑧0 → 𝑧 1 crosses a side and there is a unique vertex 𝑣 on this
side that is interior to the loop (working counterclockwise): see Figure 37.3.17.
v z1 z10
zk−1 zk−1
z0 = zk z0 = zk
Figure 37.3.17: Simplifying a relation: using the cycle relation
The cycle relation 𝛿 𝑒 = 1 at 𝑣 traces a loop around 𝑣, and without loss of generality we
may suppose 𝛿 𝑒 = 𝛼𝛾1 with 𝛼 a word in 𝐺 (𝑃). Therefore, substituting this relation
into the starting relation, we obtain a relation 𝛾2 · · · 𝛾 𝑘 (𝛿 𝑒 𝛼) with one fewer interior
vertex; the result then follows by induction.
37.4. DIRICHLET DOMAINS 681
In section 37.5, we consider a partial converse to Proposition 37.3.14, due to

Poincaré: given a convex hyperbolic polygon with a side pairing that satisfies certain
conditions, there exists a Fuchsian group Γ with the given polygon as a fundamental
domain.
37.4 Dirichlet domains

We now consider the construction of Dirichlet fundamental domains in a general
context. Let (𝑋, 𝜌) be a complete, locally compact geodesic space. In particular, 𝑋 is
connected, and by the theorem of Hopf–Rinow (see Theorem 33.2.9), closed balls in
𝑋 are compact.
Let Γ be a discrete group of isometries acting properly on 𝑋. Right from the get
go, we prove our first important result: we exhibit generators for a group based on a
fundamental set with a basic finiteness property.
Definition 37.4.1. Let 𝐴 ⊆ 𝑋. We say 𝐴 is locally finite for Γ if for each compact set
𝐾 ⊂ 𝑋, we have 𝛾𝐾 ∩ 𝐴 ≠ ∅ for only finitely many 𝛾 ∈ Γ.
The value of a locally finite fundamental set is explained by the following theorem.
Theorem 37.4.2. Let ◊ be a locally finite fundamental set for Γ. Then Γ is generated
by the set
{𝛾 ∈ Γ : ◊ ∩ 𝛾◊ ≠ ∅}. (37.4.3)
Proof. Let Γ∗ ≤ Γ be the subgroup of Γ generated by the elements (37.4.3). We want

to show Γ∗ = Γ.
For all 𝑥 ∈ 𝑋, by Theorem 37.4.18, there exists 𝛾 ∈ Γ such that 𝛾𝑥 ∈ ◊. If there is
another 𝛾 0 ∈ Γ with 𝛾 0𝑥 ∈ ◊, then
𝛾 0𝑥 ∈ ◊ ∩ 𝛾 0 𝛾 −1 ◊
so 𝛾 0 𝛾 −1 ∈ Γ∗ and in particular Γ∗ 𝛾 = Γ∗ 𝛾 0. In this way, we define a map
𝑓 : 𝑋 → Γ∗ \Γ
(37.4.4)
𝑥 ↦→ Γ∗ 𝛾
for all 𝛾 ∈ Γ such that 𝛾𝑥 ∈ ◊.

We now show that 𝑓 is locally constant. Let 𝑥 ∈ 𝑋. Ð Since ◊ is locally finite, for
every compact neighborhood 𝐾 3 𝑥 we can write 𝐾 ⊆ 𝑖 𝛾𝑖 ◊ with a finite union, and
by shrinking 𝐾 we may suppose that 𝑥 ∈ 𝛾𝑖 ◊ for all 𝑖. In particular, 𝑓 (𝑥) = Γ∗ 𝛾𝑖−1 for
any 𝑖. But then if 𝑦 ∈ 𝐾, then 𝑦 ∈ 𝛾𝑖 ◊ for some 𝑖, so 𝑓 (𝑦) = Γ∗ 𝛾𝑖−1 = 𝑓 (𝑥). Thus 𝑓 is
locally constant.
But 𝑋 is connected so every locally constant function is in fact constant and
therefore 𝑓 takes only the value Γ∗ . Let 𝛾 ∈ Γ and let 𝑥 ∈ ◊. Then
Γ∗ = 𝑓 (𝑥) = 𝑓 (𝛾 −1 𝑥) = Γ∗ 𝛾
so 𝛾 ∈ Γ∗ , and the proof is complete.

We now seek a locally finite fundamental set with other nice properties: we will
choose in each Γ-orbit the closest points to a fixed point 𝑥 0 ∈ 𝑋. So we first must
understand the basic local properties of intersections of these half-spaces (as in 37.1.2).
37.4.5. For 𝑥1 , 𝑥2 ∈ 𝑋, define the closed Leibniz half-space
𝐻 (𝑥 1 , 𝑥2 ) = {𝑥 ∈ 𝑋 : 𝜌(𝑥, 𝑥 1 ) ≤ 𝜌(𝑥, 𝑥 2 )}. (37.4.6)
If 𝑥1 = 𝑥2 , then 𝐻 (𝑥1 , 𝑥2 ) = 𝑋. If 𝑥1 ≠ 𝑥2 , then 𝐻 (𝑥1 , 𝑥2 ) consists of the set of points

as close to 𝑥1 as 𝑥2 , so
int 𝐻 (𝑥1 , 𝑥2 ) = {𝑥 ∈ 𝑋 : 𝜌(𝑥, 𝑥 1 ) < 𝜌(𝑥, 𝑥 2 )}. (37.4.7)
and
bd 𝐻 (𝑥1 , 𝑥2 ) = 𝐿 (𝑥1 , 𝑥2 ) = {𝑥 ∈ 𝑋 : 𝜌(𝑥, 𝑥 1 ) = 𝜌(𝑥, 𝑥 2 )}
is called the hyperplane bisector (or equidistant hyperplane or separator) between
𝑥1 and 𝑥 2 .
Remark 37.4.8. In this generality, unfortunately hyperplane bisectors are not neces-
sarily geodesic (Exercise 37.10).
Definition 37.4.9. A set 𝐴 ⊆ 𝑋 is star-shaped with respect to 𝑥 0 ∈ 𝐴 if for all 𝑥 ∈ 𝐴,
the geodesic between 𝑥 and 𝑥0 belongs to 𝐴.
A set 𝐴 ⊆ 𝑋 that is star-shaped is path connected, so connected.
Lemma 37.4.10. A Leibniz half-plane 𝐻 (𝑥1 , 𝑥2 ) is star-shaped with respect to 𝑥1 .
Proof. Let 𝑥 ∈ 𝐻 (𝑥 1 , 𝑥2 ) and let 𝑦 be a point along the unique geodesic from 𝑥 to 𝑥1 .
Then
𝜌(𝑥 1 , 𝑦) + 𝜌(𝑦, 𝑥) = 𝜌(𝑥 1 , 𝑥).
If 𝑦 ∉ 𝐻 (𝑥1 , 𝑥2 ), then 𝜌(𝑥 2 , 𝑦) < 𝜌(𝑥1 , 𝑦), and so
𝜌(𝑥2 , 𝑥) ≤ 𝜌(𝑥2 , 𝑦) + 𝜌(𝑦, 𝑥) < 𝜌(𝑥1 , 𝑦) + 𝜌(𝑦, 𝑥) = 𝜌(𝑥 1 , 𝑥)
contradicting that 𝑥 ∈ 𝐻 (𝑥1 , 𝑥2 ). So 𝑦 ∈ 𝐻 (𝑥1 , 𝑥2 ) as desired.

Now let 𝑥 0 ∈ 𝑋.
Definition 37.4.11. The Dirichlet domain for Γ centered at 𝑥0 ∈ 𝑋 is
◊(Γ; 𝑥0 ) = {𝑥 ∈ 𝑋 : 𝜌(𝑥, 𝑥 0 ) ≤ 𝜌(𝛾𝑥, 𝑥 0 ) for all 𝛾 ∈ Γ}.
We often abbreviate ◊(𝑥0 ) = ◊(Γ; 𝑥 0 ).

37.4.12. Since 𝜌(𝛾𝑥, 𝑥 0 ) = 𝜌(𝑥, 𝛾 −1 𝑥0 ),
Ù
◊(𝑥0 ) = 𝐻 (𝑥0 , 𝛾 −1 𝑥0 );
𝛾 ∈Γ
each half-space is closed and star-shaped with respect to 𝑥0 , so the same is true of
◊(𝑥0 ). In particular, ◊(𝑥0 ) is connected.
37.4. DIRICHLET DOMAINS 683
A Dirichlet domain satisfies a basic finiteness property, as follows.
Lemma 37.4.13. If 𝐴 ⊂ 𝑋 is a bounded set, then 𝐴 ⊆ 𝐻 (𝛾; 𝑥0 ) for all but finitely
many 𝛾 ∈ Γ.
In particular, for all 𝑥 ∈ 𝑋 we have 𝑥 ∈ 𝐻 (𝛾; 𝑥 0 ) for all but finitely many 𝛾 ∈ Γ.
Proof. Since 𝐴 is bounded,
sup({𝜌(𝑥, 𝑥 0 ) : 𝑥 ∈ 𝐴}) = 𝑟 < ∞.
By Theorem 34.5.1, the orbit Γ𝑥0 is discrete and # StabΓ (𝑥0 ) < ∞; since closed balls
are compact by assumption, there are only finitely many 𝛾 ∈ Γ such that
𝜌(𝛾𝑥0 , 𝑥0 ) = 𝜌(𝑥 0 , 𝛾 −1 𝑥0 ) ≤ 2𝑟
and for all remaining 𝛾 ∈ Γ and all 𝑥 ∈ 𝐴,
𝜌(𝑥, 𝛾 −1 𝑥0 ) ≥ 𝜌(𝑥0 , 𝛾 −1 𝑥 0 ) − 𝜌(𝑥, 𝑥 0 ) > 2𝑟 − 𝑟 = 𝑟 ≥ 𝜌(𝑥, 𝑥 0 )
so 𝑥 ∈ 𝐻 (𝛾; 𝑥0 ), as in Figure 37.4.14.
γ4−1 x0
γ1−1 x0
x
A γ −1 x0
ρ (x, x0 )

r ρ x, γ −1 x0
x0
γ3−1 x0
γ2−1 x0
Figure 37.4.14: The bounded set 𝐴 and the orbit of 𝑥 0
This concludes the proof.
37.4.15. Arguing in a similar way as in Lemma 37.4.13, one can show (Exercise 37.8):
if 𝐾 is a compact set, then 𝐾 ∩ 𝐿 (𝛾; 𝑥0 ) ≠ ∅ for only finitely many 𝛾 ∈ Γ.
int ◊(𝑥0 ) = {𝑥 ∈ ◊ : 𝜌(𝑥, 𝑥 0 ) < 𝜌(𝛾𝑥, 𝑥 0 ) for all 𝛾 ∈ Γ r StabΓ (𝑥0 )}.
It follows from Lemma 37.4.16 that 𝑥 ∈ bd ◊(𝑥 0 ) if and only if there exists
𝛾 ∈ Γ r StabΓ (𝑥0 ) such that 𝜌(𝑥, 𝑥 0 ) = 𝜌(𝛾𝑥, 𝑥 0 ).
Proof. Let 𝑥 ∈ ◊, and let 𝑈 3 𝑥 be a bounded open neighborhood of 𝑥. By Lemma

37.4.13, we have 𝑈 ⊆ 𝐻 (𝛾; 𝑥 0 ) for all but finitely many 𝛾 ∈ Γ, so
Ù
𝑈∩◊=𝑈∩ 𝐻 (𝑥0 , 𝛾𝑖−1 𝑥 0 )
𝑖
the intersection over finitely many 𝛾𝑖 ∈ Γ with 𝛾𝑖 ∉ StabΓ (𝑥0 ) (see Figure 37.4.17).
x
x
U
Figure 37.4.17: The neighborhood 𝑈
Thus Ù
𝑈 ∩ int(◊) = 𝑈 ∩ int 𝐻 (𝑥0 , 𝛾𝑖−1 𝑥0 ).
𝑖
The lemma then follows from (37.4.7).
The first main result of this chapter is the following theorem.

Theorem 37.4.18. Let 𝑥0 ∈ 𝑋, and suppose StabΓ (𝑥0 ) = {1}. Then ◊(Γ; 𝑥0 ) is a
locally finite fundamental set for Γ that is star-shaped with respect to 𝑥0 and whose
boundary consists of hyperplane bisectors.
Specifically, in a bounded set 𝐴, by Lemma 37.4.13
Ø
𝐴 ∩ bd ◊(Γ; 𝑥 0 ) ⊆ 𝐿 (𝑥0 , 𝛾𝑖 𝑥0 )
𝑖
for finitely many 𝛾𝑖 ∈ Γ r {1}.
Proof. Abbreviate ◊ = ◊(𝑥 0 ). We saw that ◊ is (closed and) star-shaped with respect
to 𝑥0 in 37.4.12. Ð
Now we show that 𝑋 = 𝛾 𝛾◊. Let 𝑥 ∈ 𝑋. The orbit Γ𝑥 is discrete, so the distance
𝜌(Γ𝑥, 𝑥 0 ) = inf ({𝜌(𝛾𝑥, 𝑥 0 ) : 𝛾 ∈ Γ}) (37.4.19)

is minimized at some point 𝛾𝑥 ∈ ◊ with 𝛾 ∈ Γ. Thus ◊(𝑥0 ) contains at least one point
from every Γ-orbit, and consequently .
We now refer to Lemma 37.4.16. Since 𝑋 is complete, this lemma implies that
cl(int(◊)) = ◊. And int(◊)∩int(𝛾◊) = ∅ for all 𝛾 ∈ Γr{1}, because if 𝑥, 𝛾𝑥 ∈ int(◊)
with 𝛾 ≠ 1 then
𝜌(𝑥, 𝑥 0 ) < 𝜌(𝛾𝑥, 𝑥 0 ) < 𝜌(𝛾 −1 (𝛾𝑥), 𝑥0 ) = 𝜌(𝑥, 𝑥 0 ), (37.4.20)
a contradiction.
37.5. HYPERBOLIC DIRICHLET DOMAINS 685
Finally, we show that 𝑋 is locally finite. It suffices to check this for a closed disc
𝐾 ⊆ 𝑋 with center 𝑥0 and radius 𝑟 ∈ R ≥0 . Suppose that 𝛾𝐾 meets ◊ with 𝛾 ∈ Γ; then
by definition there is 𝑥 ∈ ◊ such that 𝜌(𝑥 0 , 𝛾 −1 𝑥) ≤ 𝑟. Then
𝜌(𝑥0 , 𝛾 −1 𝑥0 ) ≤ 𝜌(𝑥0 , 𝛾 −1 𝑥) + 𝜌(𝛾 −1 𝑥, 𝛾 −1 𝑥 0 ) ≤ 𝑟 + 𝜌(𝑥, 𝑥 0 ).
Since 𝑥 ∈ ◊, we have 𝜌(𝑥, 𝑥 0 ) ≤ 𝜌(𝛾 −1 𝑥, 𝑥 0 ) ≤ 𝑟, so
𝜌(𝑥 0 , 𝛾 −1 𝑥 0 ) ≤ 𝑟 + 𝑟 = 2𝑟.
This setup can be seen in Figure 37.4.21.
γK
x
x ∈ γK ∪ ◊
x0
γ −1 x γ −1 x ∈ K
γ −1 x0
K
Figure 37.4.21: ◊ is locally finite and star-shaped
For the same reason as in Lemma 37.4.13, this can only happen for finitely many
𝛾 ∈ Γ.
Remark 37.4.22. Dirichlet domains are sometimes also called Voronoi domains.
37.5 Hyperbolic Dirichlet domains

We now specialize to the case 𝑋 = H where H = H2 or H = H3 with volume 𝜇; then
Γ is a Fuchsian or Kleinian group, respectively.
Definition 37.5.1. A fundamental domain for Γ 𝑋 is a connected fundamental

set ◊ ⊆ 𝑋 such that 𝜇(bd ◊) = 0.
(A domain in topology is sometimes taken to be an open connected set; one also

sees closed domains, and our fundamental domains are taken to be of this kind.)
We now turn to Dirichlet domains in this context.
37.5.2. In Theorem 37.4.18, the hypothesis that StabΓ (𝑧0 ) = {1} is a very mild
hypothesis. If 𝐾 is a compact set, then since ◊ is locally finite, the set of points 𝑧 ∈ 𝐾
with StabΓ (𝑧) ≠ {1} is a finite set of points when H = H2 and a finite set of points
together with finitely many geodesic axes when H = H3 .
In spite of 37.5.2, we prove a slightly stronger and more useful version of Theorem
37.4.18, as follows. If Γ0 = StabΓ (𝑧 0 ) is nontrivial, we modify the Dirichlet domain
by intersecting with a fundamental set for Γ0 ; the simplest way to do this is just to
choose another point which is not fixed by an element of Γ0 and intersect.
Theorem 37.5.3. Let 𝑧0 ∈ H , let Γ0 = StabΓ (𝑧 0 ), and let 𝑢 0 ∈ H be such that
StabΓ0 (𝑢 0 ) = {1}. Then
◊(Γ; 𝑧 0 ) ∩ ◊(Γ0 ; 𝑢 0 )
is a connected, convex, locally finite fundamental domain for Γ with geodesic boundary
in H .
Proof. Abbreviate
◊ = ◊(Γ; 𝑧0 ) ∩ ◊(Γ0 ; 𝑢 0 ).
First, we show that 𝑧 0 ∈ ◊: we have 𝑧 0 ∈ ◊(Γ; 𝑧 0 ), and by Theorem 37.4.18,
◊(Γ0 ; 𝑢 0 ) is a fundamental set for Γ0 so there exists 𝛾0 ∈ Γ0 such that 𝛾0 𝑢 0 = 𝑢 0 ∈
◊(Γ0 ; 𝑢 0 ).
Ð
Now we show that ◊ is a fundamental set for Γ. First we show H = 𝛾 𝛾◊. Let
𝑧 ∈ H , and let 𝛾 ∈ Γ be such that 𝜌(𝛾𝑧, 𝑧0 ) is minimal as in (37.4.19). Let 𝛾0 ∈ Γ0
such that 𝛾0 (𝛾𝑧) ∈ ◊(Γ0 ; 𝑢 0 ). Then
𝜌(𝛾0 (𝛾𝑧), 𝑧0 ) = 𝜌(𝛾𝑧, 𝑧0 )
so 𝛾0 𝛾𝑧 ∈ ◊. And int(◊) ∩ int(𝛾◊) = ∅ for all 𝛾 ∈ Γ r {1}, because if 𝑧, 𝛾𝑧 ∈ int(◊)
with 𝛾 ≠ 1, then either 𝛾 ∉ Γ0 in which case we obtain a contradiction as in (37.4.20),
or 𝛾 ∈ Γ0 − {1} and then we arrive at a contradiction from the fact that ◊(Γ0 ; 𝑢 0 ) is a
fundamental set.
We conclude by proving the remaining topological properties of ◊. We know that
◊ is locally finite, since it is the intersection of two locally finite sets. We saw in 37.1.2
that the Leibniz half-spaces in H2 are convex with geodesic boundary, and the same is
true in H3 by Exercise 36.9. It follows that ◊ is convex, as the intersection of convex
sets. Thus Ø
bd ◊ ⊆ 𝐿 (𝑧0 , 𝛾 −1 𝑧0 )
𝛾 ∈Γr{1}
is geodesic and measure zero, since 𝐿 (𝑧0 , 𝛾 −1 𝑧0 ) intersects a compact set for only
finitely many 𝛾 by 37.4.15.
37.6 Poincaré’s polyhedron theorem

Continuing with the notation from the previous section, we now turn to a partial
converse for Theorem 37.5.3 for H ; we need one additional condition. Let ◊ be a
convex (finite-sided) hyperbolic polyhedron equipped with a side pairing 𝑃.
37.6.1. First suppose ◊ ⊆ H2 . For a vertex 𝑣 of ◊, let 𝜗(◊, 𝑣) be the interior angle
of ◊ at 𝑣. We say that ◊ satisfies the cycle condition if for all vertices 𝑣 of ◊ there
exists 𝑒 ∈ Z>0 such that
∑︁ 2𝜋
𝜗(◊, 𝑣 𝑖 ) = .
𝑣 ∈Γ𝑣∩◊
𝑒
𝑖
37.6. POINCARE’S POLYHEDRON THEOREM 687
Put another way, ◊ satisfies the cycle condition if the sum of the interior angles for a
Γ-orbit of vertices as in the standard picture is an integer submultiple of 2𝜋.
Now suppose ◊ ⊆ H3 . Now we work with edges: for an edge ℓ of ◊, let 𝜗(◊, ℓ)
be the interior angle of ◊ at ℓ. We say that ◊ satisfies the cycle condition if for all
edges ℓ of ◊ there exists 𝑒 ∈ Z>0 such that
∑︁ 2𝜋
𝜗(◊, ℓ𝑖 ) = .
ℓ𝑖 ∈Γ𝑒∩◊
𝑒
Lemma 37.6.2. Let ◊ be a Dirichlet domain. Then ◊ satisfies the cycle condition.
Proof. We explain the case where ◊ ⊆ H2 ; the case of H3 is similar. Let 𝑣 be a vertex
of ◊. Referring to the standard picture 37.3.8,
𝑛
∑︁ 𝑛
∑︁
2𝜋 = 𝜗(𝛿 𝑗 ◊, 𝑣) = 𝜗(◊, 𝛿−1
𝑗 𝑣).
𝑗=0 𝑗=0
In 37.3.10, we proved that 𝛿 𝑚 acts by hyperbolic rotation around 𝑣 and has order
𝑒 = (𝑛 + 1)/𝑚, and
𝑛
∑︁ 𝑚−1
∑︁ 𝑚−1
∑︁
𝜗(◊, 𝛿−1
𝑗 𝑣) = 𝑒 𝜗(◊, 𝛿𝑖−1 𝑣) = 𝑒 𝜗(◊, 𝑣 𝑖 )
𝑗=0 𝑖=0 𝑖=0
with ◊ ∩ Γ𝑣 = {𝑣 1 , . . . , 𝑣 𝑚−1 }. Combining these two equations, we see that the cycle
condition is satisfied.
37.6.3. There is another condition at certain points at infinity that a fundamental

domain must satisfy. We say a point 𝑧 ∈ bd H is a infinite vertex if 𝑧 lies in the
intersection of two faces and is tangent to both. We define a sequence of tangency
vertices analogous to the cycle transformations to get an infinite vertex sequence and
a infinite vertex transformation. We say that the side/face pairing is complete if the
tangency vertex transformation is parabolic.
Theorem 37.6.4 (Poincaré polygon theorem). Let ◊ be a convex finite-sided hyperbolic
polygon/polyhedron with a side/face-pairing 𝑃. Suppose that ◊ satisfies the cycle
condition and 𝑃 is complete.
Then the group Γ := h𝐺 (𝑃)i generated by side/face-pairing elements is a Fuch-
sian/Kleinian group, ◊ is a fundamental domain for Γ, and 𝑅(𝑃) forms a complete set
of relations for 𝐺 (𝑃).
Proof. Unfortunately, it is beyond the scope of this book to give a complete proof of
Theorem 37.6.4. See Epstein–Petronio [EP94, Theorem 4.14] or Ratcliffe [Rat2006,
§11.2, §13.5]; our statement is a special case of the theorem by Maskit [Mas71], but
see Remark 37.6.5.
Remark 37.6.5. The proof of Poincaré’s theorem [Poi1882, Poi1883] has a bit of a
notorious history. From the very beginning, to quote Epstein–Petronio [EP94, §9,
p. 164]:
It is clear that Poincaré understood very well what was going on. How-
ever, the papers are not easy to read. In particular, the reader of the
three-dimensional case is referred to the treatment of the two-dimensional
case for proofs; this is fully acceptable for a trail-blazing paper, but not
satisfactory in the long term.
There are a number of reasonable published versions of Poincaré’s The-
orem in dimension two. Of these, we would single out the version by
de Rham [dR71] as being particularly careful and easy to read. Most
published versions of Poincaré’s Theorem applying to all dimensions are
unsatisfactory for one reason or another.
This sentiment is echoed by Maskit [Mas71], who presents a proof for polygons with
an extension to polyhedron, with the opening remark:
There are several published proofs of [Poincaré’s classical] theorem, but

there is some question as to their validity; Siegel [Sie65, p. 115] has com-
mented on this and given an apparently valid proof under fairly restrictive
conditions. None of the published proofs are as general as they might be,
and they all have a convexity condition that is never really used.
This note is an attempt to clarify the situation. The problem and the solu-
tion presented below arose during the course of several informal conver-
sations. Present at one or more of these conversations were L. V. Ahlfors,
L. Bers, W. Magnus, J. E. McMillan, and B. Maskit.
Epstein and Petronio [EP94, §9, p. 165] then have this to say:
Maskit’s paper contains a nice discussion of completeness, though again it

is not a constructive approach. He limits his discussion to hyperbolic space
in dimensions two and three. We are not confident that the arguments in
the paper are complete. For example, there seems to be an assumption
that the quotient of a metric space, such that the inverse image of a point
is finite, is metric. This is false, as is shown by identifying 𝑥 with −𝑥 in
[−1, 1], provided 0 ≤ 𝑥 < 1.
Maskit published a book [Mas88] containing an expanded version of the proof for
polyhedra, to which Epstein and Petronio [EP94, §9, p. 164] review:
The treatment in [Mas88] is difficult to understand. For example in H.9 on

page 75, it is claimed that a metric is defined in a certain way, and this fact
is said to be “easy to see”, but it seems to us an essential and non-trivial
point, which is not so easy to see, particularly when the group generated
by the face-pairings is not discrete. . . . The Proposition in IV.1.6 on
page 79 of this book is incorrect—a counter-example is given in Example
9.1—because there are no infinite cycles or infinite edges according to the
definitions in the book.
37.7. SIGNATURE OF A FUCHSIAN GROUP 689
37.7 Signature of a Fuchsian group

As an application of Theorem 37.5.3, we relate area and signature for good orbifolds
obtained from Fuchsian groups.
37.7.1. We first recall 34.8.13: a good compact, oriented 2-orbifold 𝑋 is classified
up to homeomorphism by its signature (𝑔; 𝑒 1 , . . . , 𝑒 𝑘 ), where 𝑔 is the genus of the
underlying topological surface and the 𝑒 1 , . . . , 𝑒 𝑘 are the orders of the (necessarily
cyclic) nontrivial stabilizer groups. This extends to good orbifolds 𝑌 with finitely
many points removed: we define the signature (𝑔; 𝑒 1 , . . . , 𝑒 𝑘 ; 𝛿) where 𝛿 ≥ 0 is the
number of punctures.
Now let Γ ≤ PSL2 (R) be a Fuchsian group such that the quotient 𝑌 (Γ) = Γ\H2
has finite hyperbolic area 𝜇(𝑌 (Γ)).
Theorem 37.7.2 (Siegel). Suppose that 𝜇(Γ\H2 ) has finite hyperbolic area. Then
every Dirichlet domain ◊ has finitely many sides.
Proof. The proof estimates the contribution to the volume from infinite vertices: for
Fuchsian groups, see Siegel [Sie45, p. 716–718], or the expositions of this argument
by Gel’fand–Graev–Pyatetskii-Shapiro [GG90, Chapter 1, §1.5] and Katok [Kat92,
Theorem 4.1.1].
37.7.3. By Theorem 37.5.3 and Siegel’s theorem (Theorem 37.7.2), a Dirichlet domain
◊ for Γ is connected, convex, hyperbolic polygon. In particular, there are 𝑚 ∈
Z ≥0 vertex cycles, which by 37.3.10 correspond to cyclic stabilizer groups of orders
𝑒 1 , . . . , 𝑒 𝑚 ∈ Z ≥1 listed so that 𝑒 1 , . . . , 𝑒 𝑘 ≥ 2, and finitely many 𝛿 ∈ Z ≥0 infinite
vertex cycles, corresponding to 𝛿 stabilizer groups that are infinite cyclic.
𝑘 !
2
∑︁ 1
𝜇(Γ\H ) = 2𝜋 (2𝑔 − 2) + 1− +𝛿 .
𝑖=1
𝑒𝑖
Proof. Let ◊ be a Dirichlet domain for Γ with 2𝑛 sides and 𝑛 finite or infinite vertices.
The hyperbolic area of ◊ is given by the Gauss–Bonnet formula 33.6.9: summing
vertex cycles using the cycle condition (Lemma 37.6.2), we have
𝑘
∑︁ 2𝜋
𝜇(◊) = (2𝑛 − 2)𝜋 − .
𝑖=1
𝑒𝑖
The quotient Γ\H2 is a (punctured) oriented topological surface of genus 𝑔, with 𝑘 + 𝛿

vertices, 𝑛 edges, and 1 face. By Euler’s formula, we have
2 − 2𝑔 = (𝑘 + 𝛿) − 𝑛 + 1
so
𝑛 − 1 = 2𝑔 − 2 + (𝑘 + 𝛿).
Therefore
𝑘
! 𝑘 !
∑︁ 1 ∑︁ 1
𝜇(◊) = 2𝜋 (2𝑔 − 2) + (𝑘 + 𝛿) − = 2𝜋 (2𝑔 − 2) + 1− +𝛿
𝑒
𝑖=1 𝑖 𝑖=1
𝑒𝑖
as claimed.
37.8 The (6, 4, 2)-triangle group

We pause to refresh the quaternionic thread and consider a particularly nice example,
showing how quaternion algebras arise naturally in the context of Fuchsian groups.
Consider a hyperbolic triangle 𝑇 with angles 𝜋/2, 𝜋/6, 𝜋/6 and vertices placed as
in Figure 37.8.1.
z−i
i z 7→ w = z+i
w
π/6 π/6
z
Figure 37.8.1: The hyperbolic triangle 𝑇 with angles 𝜋/2, 𝜋/6, 𝜋/6
By symmetry, we may define the side-pairing 𝑃 as shown in Figure 37.8.1. This
polygon satisfies the cycle condition, so by the Poincaré polygon theorem (Theorem
37.6.4), there exists a Fuchsian group Δ ⊂ PSL2 (R) generated by the two side pairing
elements in 𝑃 and with fundamental domain 𝑇. In this section, we construct this group
explicitly and observe some interesting arithmetic consequences.
37.8.2. We seek the position of the vertex 𝑧 ∈ H2 corresponding to 𝑤 ∈ D2 . Zooming
in, we obtain the picture as in Figure 37.8.3.
w
1 π/6
c
π/4
− dc π/12 0
2
|cw + d| = 1
p
d2 −c2 = 1 ⇒ d = 1 + c2
Figure 37.8.3: Finding the position of 𝑧
37.8. THE (6, 4, 2)-TRIANGLE GROUP 691
The edge containing 𝑤 and its complex conjugate is defined by an isometric circle
|𝑐𝑤 + 𝑑| = 1 with 𝑑 2 − 𝑐2 = 1 and by symmetry 𝑐, 𝑑 ∈ R>0 . With the angles as labeled,
we find that 𝑤 = (−𝑑 + 𝑒 𝜋𝑖/12 )/𝑐, and since arg(𝑤) = 3𝜋/4 we compute that
√︁
1 + 𝑐2 − cos(𝜋/12) = sin(𝜋/12) (37.8.4)
√ √︁
so 𝑐2 = 2 cos(𝜋/12) sin(𝜋/12) = sin(𝜋/6) √ = 1/2 thus 𝑐 = 1/ 2 and 𝑑 = √3/2, so
this isometric circle is defined by |𝑤 + 3| 2 = √ 2, and 𝑤 = (−1 + 𝑖)/(1 + 3). By
coincidence, we find that 𝑧 = 𝑧2 = (−1 + 𝑖)/(1 + 3)√= 𝑤 as well.√ The intersection of
this circle with the imaginary axis is the point 𝑧3 = ( 3 − 1)𝑖/ 2.
The unique element mapping the sides meeting at 𝑖 is obtained by pulling back
rotation by −𝜋/4 in the unit disc model; it is given by the matrix
−1
1 −𝑖 𝑒 − 𝜋𝑖/4

0 1 −𝑖 1 1 −1
𝛿4 = = √ (37.8.5)
1 𝑖 0 𝑒 𝜋𝑖/4 1 𝑖 2 1 1
with
𝑧−1
𝛿4 (𝑧) = , for 𝑧 ∈ H2 ,
𝑧+1
and 𝛿44 = 1 ∈ PSL2 (R). From a similar computation, we find that the other side pairing
element acting by hyperbolic rotation around 𝑧 3 is
√
1 0√ −1 + 3
𝛿2 = √ (37.8.6)
2 −1 − 3 0
and 𝛿22 = 1 ∈ PSL2 (R). We have (𝛿2 𝛿4 ) (𝑧4 ) = 𝛿2 (𝑧 2 ) = 𝑧 4 , so the element

√ √
−1 1 1+ 3 1− 3
𝛿6 = (𝛿2 𝛿4 ) = √ √ (37.8.7)
2 1 + 3 −1 + 3
√
fixes the vertex 𝑧4 , with 𝛿6 (𝑧 1 ) = (1 + −3)/2 and 𝛿66 = 1 ∈ PSL2 (R). These actions
are summarized in Figure 37.8.8.
z1 = i

1 −1
δ4 = · √12
1 1
√
0√ −1 + 3 √1
δ2 = ·
−1 − 3 0 2
√
π/6 3−1
√
2
i = z3
−1+i z4 √ √
√
1+ 3
= z2 −1 1 1 + √3 1 − √3
δ6 = (δ2 δ4 ) = 2 1+ 3 −1 + 3
Figure 37.8.8: Side-pairing elements for the (6, 4, 2)-triangle group

So recording these cycle relations,

Δ = h𝛿2 , 𝛿4 | 𝛿22 = 𝛿44 = (𝛿2 𝛿4 ) 6 = 1i
= h𝛿2 , 𝛿4 , 𝛿6 | 𝛿22 = 𝛿44 = 𝛿66 = 𝛿2 𝛿4 𝛿6 = 1i.
37.8.9. From this setup, we can identify a quaternion algebra obtained from (appro-
priately scaled) generators of Δ. We have the characteristic polynomials
√ √
𝛿22 + 1 = 𝛿42 − 2𝛿4 + 1 = 𝛿62 − 3𝛿6 + 1 = 0.
To obtain rational traces, and to simplify the presentation a bit further we consider the
Z-subalgebra 𝐵 ⊆ M2 (R) generated by
√ √ √
𝑖 := 2𝛿4 , 𝑗 := 3𝛿6 , 𝑘 := − 6𝛿2 ;
we have
𝑖 2 = 2𝑖 − 2 𝑗 𝑘 = −3𝑖
2
𝑗 = 3𝑗 − 3 𝑘𝑖 = −2 𝑗 (37.8.10)
𝑘 2 = −6 𝑖 𝑗 = −𝑘
so as in (22.1.2) with (𝑎, 𝑏, 𝑐, 𝑢, 𝑣, 𝑤) = (−3, −2, −1, 2, 3, 0), we obtain an order
O = Z + Z𝑖 + Z 𝑗 + Z𝑘 of reduced discriminant −24 + 12 + 18 = 6 and with associated
ternary quadratic form
−3𝑥 2 − 2𝑦 2 − 𝑧2 + 2𝑦𝑧 + 3𝑥𝑧.
37.8.11. We now try to simplify the presentation (37.8.10) as much as possible. This
kind of activity is more aesthetics than mathematics, but for this example there is a
preferred way of writing the algebra and order (regrettably, not in a good basis) as
follows. We write 𝑖 0 = 𝑖 − 1 and 𝑗 0 = 𝑖( 𝑗 − 3/2𝑖) = −3𝑖 + 𝑘 + 3 so that now (𝑖 0) 2 = −1
and ( 𝑗 0) 2 = 3 and 𝑗 0𝑖 0 = −𝑖 0 𝑗 0; the remaining basis element of the order in terms of
these generators can be taken to be 𝑘 0 = (1 + 𝑖 0 + 𝑗 0 + 𝑖 0 𝑗 0)/2 = 3 − 𝑖 − 𝑗 + 𝑘.
Throwing
37.8.12. away primes from the previous paragraph, we have the algebra
−1, 3
𝐵 := of discriminant disc 𝐵 = 6 and order
Q
1 + 𝑖 + 𝑗 + 𝑖𝑗
O := Z ⊕ Z𝑖 ⊕ Z 𝑗 ⊕ Z𝑘, 𝑘= (37.8.13)
2
with 𝑘 2 − 𝑘 − 1 = 0. (By 23.1.1, since discrd(O) = disc 𝐵 = 6, we conclude that O is
a maximal order in 𝐵.)
This algebra came with the embedding
𝜄∞ : 𝐵 ↩→ M2 (R)
√
0 −1 3 0
√
𝑖, 𝑗 ↦→ ,
1 0 0 − 3
√ √
𝑡 + 𝑦 √3 −𝑥 + 𝑧√ 3
𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑖 𝑗 ↦→ = 𝑔.
𝑥+𝑧 3 𝑡−𝑦 3
37.9. UNIT GROUP FOR DISCRIMINANT 6 693
Following the transformations back, we recover the triangle group as a subgroup of

𝐵× /𝐹 × after rescaling to be
𝛿2 := 3𝑖 + 𝑖 𝑗 = −1 + 2𝑖 − 𝑗 + 2𝑘,
𝛿4 := 1 + 𝑖, (37.8.14)
𝛿6 := 3 + 3𝑖 + 𝑗 + 𝑖 𝑗 = 2 + 2𝑖 + 2𝑘;
and we confirm
𝛿22 = −6, 𝛿44 = −4, 𝛿66 = −1728, 𝛿2 𝛿4 𝛿6 = −12
so the images of these elements in 𝐵× /𝐹 × generate the group Δ.

In fact, the elements 𝛿2 , 𝛿4 , 𝛿6 ∈ 𝑁 𝐵× (O) normalize the order O:
𝛿2−1 𝑖𝛿2 = −1 + 𝑖 − 𝑗 + 2𝑘 𝛿2−1 𝑗 𝛿2 = − 𝑗 𝛿2−1 𝑘𝛿2 = 1 − 𝑘

𝛿4−1 𝑖𝛿4 = 𝑖 𝛿4−1 𝑗 𝛿4 = −𝑖 𝑗 𝛿4−1 𝑘𝛿4 = 1 + 𝑖 + 𝑗 − 𝑘
𝛿6−1 𝑖𝛿6 = −1 + 𝑖 − 𝑗 + 2𝑘 𝛿6−1 𝑗 𝛿6 = 2 − 𝑖 + 2 𝑗 − 4𝑘 𝛿6−1 𝑘𝛿6 = 1 + 𝑗 − 𝑘
37.9 Unit group for discriminant 6

In this section, we continue the example from the previous section, specifically the
order O in 37.8.12. Let
Γ := 𝜄∞ (O1 /{±1}) ≤ PSL2 (R).
By organized enumeration, we will see that Γ is a Fuchsian group and compute a

fundamental domain for the action of Γ. We return to this example in section 38.1
as a basic example of the general theory of arithmetic groups arising from quaternion
algebras.
37.9.1. Moving to the unit disc via the map 𝑧 →
↦ 𝜙(𝑧) = 𝑤 = (𝑧 − 𝑖)/(𝑧 + 𝑖),
√
𝑡 − 𝑖𝑥 3(𝑦 − 𝑖𝑧)
𝑔𝜙 = √ .
3(𝑦 + 𝑖𝑧) 𝑡 + 𝑖𝑥
To avoid cumbersome notation, we identify 𝑔 with 𝑔 𝜙 .

The isometric circle of such an element 𝑔 is a
1 −(𝑡 + 𝑖𝑥)
circle of radius √︁ centered at √
3(𝑦 2 + 𝑧 2 ) 3(𝑦 + 𝑖𝑧)
when 𝑦 2 + 𝑧2 ≠ 0; when 𝑦 2 + 𝑧 2 = 0, the center 𝑤 = 0 ∈ D2 is stabilized, with stabilizer

𝑖 0
StabΓ (0) = h𝑆i where 𝑆 = acts by rotation by 𝜋. We have
0 −𝑖
Ù
◊(Γ; 0) = cl ext 𝐼 (𝛾).
𝛾 ∈Γ−h𝑆 i
To make a fundamental domain, we need to intersect ◊(0) with a fundamental domain

for h𝑆i, and we take the left half-plane. Then
◊ = ◊(Γ; 0) ∩ {𝑤 ∈ D2 : Re 𝑤 ≤ 0}
is a fundamental domain for Γ.
37.9.2. We list elements 𝑔 = 𝑡 + 𝑥𝑖 + (𝑦 + 𝑧𝑖) 𝑗 with det 𝑔 = 𝑡 2 + 𝑥 2 − 3(𝑦 2 + 𝑧2 ) = 1

and 𝑡, 𝑥, 𝑦, 𝑧 ∈ 12 Z whose doubles are all of the same parity. We enumerate them in
increasing order of the (square) inverse radius 𝑦 2 + 𝑧2 (ignoring the factor 3). The
case 𝑦 2 + 𝑧2 = 0 gives us the stabilizer. By parity, we cannot have 𝑦 2 + 𝑧 2 = 1/4.
If 𝑦 2 + 𝑧2 = 1/2 then we find 𝑦, 𝑧 = ±1/2 and 𝑡 2 + 𝑥 2 = 5/2. Sifting out all of the
possibilities, we find
3 1 1 1 1 3 1 1
𝑔 = ± ± 𝑖 ± 𝑗 ± 𝑖𝑗 or 𝑔 = ± ± 𝑖 ± 𝑗 ± 𝑖 𝑗.
2 2 2 2 2 2 2 2
√︁
All of these elements
√ have the radius of 𝐼√(𝑔) equal to 2/3 = 0.82 . . ., and the centers
are ±1.15 ± 0.57 −1 and ±0.57 ± 1.15 −1. The corresponding external domain is
given in Figure 37.9.3.
Figure 37.9.3: A starting external domain
37.9.4. If we continue in this way, listing elements according to decreasing radius, we

find that all remaining elements have too small a radius to cut away anything extra from
the external domain. The corresponding external domain looks like Figure 37.9.5.
Figure 37.9.5: The external domain
It follows that the ◊ is cut out by the left half with side pairing as in Figure 37.9.6.
γ2
γ1
γ3
Figure 37.9.6: ◊ ⊂ D2
Pulling back to the upper half-plane, we obtain the domain in Figure 37.9.7.
√
ν1 = −1
γ1 = i
√ √
− 3+ −1
2 = ν2 γ2 = −1 + i + k γ3 = −1 + i − j + k
√
−1+√−1
ν6
= ν3
1+ 3 ν5
√ √
ν4 = (2 − 3) −1
Figure 37.9.7: ◊ ⊂ H2
The corresponding tessellation of the upper half-plane looks like the one in Figure
37.9.8.
0
Figure 37.9.8: Tessellation of H2 by Γ1
In addition to this side pairing, we check the cycle relations on the four orbits of
vertices: the fixed points 𝑣 1 , 𝑣 3 , 𝑣 5 and the orbit 𝑣 2 , 𝑣 4 , 𝑣 6 . In conclusion,
Γ = h𝛾1 , 𝛾2 , 𝛾3 | 𝛾12 = 𝛾23 = 𝛾33 = (𝛾3 𝛾2 𝛾1 ) 2 = 1i (37.9.9)
and letting 𝛾4 = 𝛾3 𝛾2 𝛾1 = −2𝑖 + 𝑗, we can rewrite this more symmetrically as
Γ = h𝛾1 , 𝛾2 , 𝛾3 , 𝛾4 | 𝛾12 = 𝛾23 = 𝛾33 = 𝛾42 = 𝛾4 𝛾3 𝛾2 𝛾1 = 1i.
The order of the stabilizers tell us the angles at each vertex, and so by the Gauss–Bonnet
formula (Theorem 33.6.8) we compute that the area is
𝜇(◊) = 3𝜋 − (2/3 + 2/3 + 1)𝜋 = 2𝜋/3.

37.9.10. Finally, the quotient 𝑋 (Γ) = Γ\H2 has the structure of a good complex 1-
orbifold (see 34.8.14), a Riemann surface but with finitely many orbifold points. Since
the fundamental domain ◊ is compact, via the continuous surjective projection map
◊ → 𝑋 (Γ) we see that 𝑋 (Γ) is compact, and 𝜇(𝑋 (Γ)) = 𝜇(◊) = 2𝜋/3. This orbifold
‘folds’ up to a surface with topological genus 0, so the signature 34.8.13 of 𝑋 (Γ) is
(0; 2, 2, 3, 3): see Figure 37.9.11.
ν2 , ν4 , ν6
orbifold signature
ν3 ν5
(0; 2, 2, 3, 3)
ν1
Figure 37.9.11: 𝑋 (Γ) as an orbifold and Riemann surface
We have seen that the norm 1 group contains the (2, 4, 6)-triangle group Δ, so we
have a map 𝑋 (Γ) → 𝑋 (Δ) = Δ\H2 ; by Gauss–Bonnet, we have
𝜇(𝑋 (Δ)) = 2(𝜋 − (1/2 − 1/4 − 1/6)𝜋) = 𝜋/6,
we see
[Δ : Γ] = 𝜇(𝑋 (Γ))/𝜇(𝑋 (Δ)) = (2𝜋/3)/(𝜋/6) = 4,
4
as is visible from Figure 37.9.12 for Γ ≤ Δ.
Figure 37.9.12: Triangulation of ◊
Remark 37.9.13. The discriminant 6 quaternion algebra has been a favorite to study,
and the fundamental domain described above is also given by e.g. Alsina–Bayer
[AB2004, §5.5.2] and Kohel–Verrill [KV2003, §5.1]. We return to this example

in section 43.2 in the context of abelian surfaces with quaternionic multiplication.
Exercises
1. Let Γ = PSL2 (Z). Describe the Dirichlet domain ◊(𝑧) for an arbitrary 𝑧 ∈ H2
with Im 𝑧 > 1.
2. Let Γ = PSL2 (Z[𝑖]) (cf. section 36.6). Show that
◊(Γ; 2 𝑗) = {𝑧 = 𝑥 + 𝑦 𝑗 ∈ H3 : |Re 𝑥|, |Im 𝑥| ≤ 1/2 and k𝑧k ≥ 1}
and that
𝑖 0
StabΓ (2 𝑗) = ' Z/2Z
0 −𝑖
so ◊(Γ; 2 𝑗) is a union of two copies of a fundamental set for Γ.
3. LetΓ be the cyclic
Fuchsian group generated by the isometry 𝑧 ↦→ 4𝑧, represented
2 0
by ∈ PSL2 (R). Give an explicit description of the Dirichlet domain
0 1/2
◊(Γ; 𝑖) ⊂ H2 and its image ◊(Γ; 0) ⊂ D2 with 𝑖 ↦→ 0.

𝑎 𝑏
4. Let 𝑔 = ∈ PSL2 (R) satisfy 𝑔𝑖 ≠ 𝑖. Let 𝐻 be the perpendicular bisector
𝑐 𝑑
between 𝑖 and 𝑔𝑖.
(a) Show that k𝑔k 2 > 2.
(b) Under the map H2 → D2 taking 𝑖 ↦→ 0, show that the perpendicular
bisector between 𝑖 and 𝑔𝑖 is the isometric circle 𝐼 (𝑔 −1 ) inside D2 .
k𝑔k 2 − 2
(c) Show that 𝐻 is the half-circle of square radius 2 centered at
(𝑐 + 𝑑 2 − 1) 2
𝑎𝑐 + 𝑏𝑑
∈ R, where k𝑔k 2 = 𝑎 2 + 𝑏 2 + 𝑐2 + 𝑑 2 . [Hint: as a check along
𝑐 + 𝑑2 − 1
2
the way, 𝐻 is described by the equation | (𝑑 − 𝑖𝑐)𝑧 + 𝑖(𝑎 + 𝑖𝑏)| = | 𝑧 + 𝑖|.]
5. Let
𝑑 𝑐
𝑔= ∈ PSU(1, 1) D2
𝑐 𝑑
with 𝑐, 𝑑 ∈ C satisfying |𝑑| 2 − |𝑐| 2 = 1. Show directly that |𝑔𝑤| = |𝑤| for
𝑤 ∈ D2 if and only if
|𝑐𝑤 + 𝑑| = 1
by expanding out and simplifying.
6. Show that for all 𝑔 ∈ PSU(1, 1), we have 𝑔𝐼 (𝑔) = 𝐼 (𝑔 −1 ), where 𝐼 (𝑔) is the
isometric circle of 𝑔. (Equivalently, show that if 𝑔 ∈ PSL2 (R) that 𝑔𝐿 (𝑔; 𝑧0 ) =
𝐿(𝑔 −1 ; 𝑧 0 ) for all 𝑧0 ∈ H2 .)
I 7. Let ◊ be a Dirichlet domain for a Fuchsian group Γ. Show that in every compact
set, there are only finitely many sides and finitely many vertices of ◊.
8. Let Γ be a discrete group of isometries acting properly on a locally compact,

complete metric space 𝑋, and let 𝑥 0 ∈ 𝑋. Recall the definition of 𝐻 (𝛾; 𝑥 0 )
(37.4.6) for 𝛾 ∈ Γ and 𝐿 (𝛾; 𝑧0 ) = bd 𝐻 (𝛾; 𝑧0 ). Show that if 𝐾 is a compact set,
then 𝐾 ∩ 𝐿(𝛾; 𝑥0 ) ≠ ∅ for only finitely many 𝛾 ∈ Γ.
9. Extend Theorem 37.4.18 as follows. Let (𝑋, 𝜌) be a complete, locally compact
geodesic space, let Γ be a discrete group of isometries acting properly on 𝑋. Let
𝑥 0 ∈ 𝑋, let Γ0 = StabΓ (𝑥0 ), and let 𝑢 0 ∈ 𝑋 be such that StabΓ0 (𝑢 0 ) = {1}. Show
that Γ0 is a discrete group of isometries acting properly on 𝑋, and
◊(Γ; 𝑧 0 ) ∩ ◊(Γ0 ; 𝑢 0 )
is a locally finite fundamental set for Γ that is star-shaped with respect to 𝑥0 and
whose boundary consists of hyperplane bisectors.
10. Consider the egg of revolution, a surface of revolution obtained from convex
curves with positive curvature as in Figure 37.9.14.
top
geodesic
bisector
bottom
Figure 37.9.14: Egg of revolution
An egg of revolution has the structure of a geodesic space with the induced
metric from R3 . Show that the separator between the top and bottom, a circle of
revolution, is not geodesic.
[In fact, Clairaut’s relation shows that the geodesic joining two points in the
same circle of revolution above crest in the 𝑥-axis never lies in this circle of
revolution.]
11. In this exercise, we consider a Fuchsian group constructed from a regular quadri-
lateral.
(a) Show that for every 𝑒 ≥ 2, there exists a regular (equilateral and equian-
gular) quadrilateral ◊ ⊂ D2 , unique up to isometry, with interior angle
𝜋/(2𝑒).
Conclude from Poincaré’s theorem that there is a Fuchsian group, unique
up to conjugation in PSL2 (R), with fundamental domain ◊ and side pairing
(b) Give a presentation for this group for all 𝑒 ≥ 2, and find explicit matrix
generators for the special case when 𝑒 = 2.
Figure 37.9.15: A special quadrilateral Fuchsian group

Chapter 38
Quaternionic arithmetic groups
In this chapter, we now apply our topological and geometric interpretation of discrete
groups in our case of interest: quaternion unit groups.
38.1 ⊲ Rational quaternion groups

The classical modular group PSL2 (Z) (chapter 35) was obtained as follows: we
• started with the matrix algebra M2 (Q),

• looked inside for integral elements to find the order M2 (Z),
• took its unit group GL2 (Z), and finally
• restricted attention to the Fuchsian group PSL2 (Z) to get a faithful action on the
upper half-plane by orientation-preserving isometries.
The wonderful thing that gives life to this part of our monograph is this: the same
thing works if we replace M2 (Q) with a quaternion algebra 𝐵 over a global field! In
this section, we derive key aspects of this program for quaternion algebras over Q in a
self-contained way, before embarking on a more general study in the remainder of the
chapter.
In section 32.1, we already dealt with the case when the quaternion
algebra was
𝑎, 𝑏
definite, finding a finite unit group; so here we take 𝐵 = indefinite. Without
Q
loss of generality (for convenience of presentation), we may
further
suppose 𝑎, 𝑏 > 0
𝑎, 𝑏
and both 𝑎, 𝑏 ∈ Z. To be indefinite means that 𝐵 ⊗Q R = ' M2 (R); we obtain
R
such an embedding via a conjugate of the left-regular representation (2.3.2)
𝜄∞ : 𝐵 ↩→ M2 (R)
√ √ √
(38.1.1)

𝑡 + 𝑥 𝑎√ (𝑦 + 𝑧 𝑎) 𝑏
𝑡 + 𝑥𝑖 + 𝑦 𝑗 + 𝑧𝑘 ↦→ √ √
(𝑦 − 𝑧 𝑎) 𝑏 𝑡 − 𝑥 𝑎.
This embedding is not unique, but another embedding would correspond to post-
composition by an R-algebra automorphism of M2 (R), which by the Skolem–Noether
701
702 CHAPTER 38. QUATERNIONIC ARITHMETIC GROUPS
theorem (Corollary 7.1.4) is given by conjugation by an element of GL2 (R), and so

we can live with a choice and this ambiguity.
Let
O := Zh𝑖, 𝑗i = Z + Z𝑖 + Z 𝑗 + Z𝑘
then O ⊂ 𝐵 is an order, and 𝜄∞ (O) ⊆ M2 (Z) consists of the subset of matrices in
(38.1.1) with 𝑡, 𝑥, 𝑦, 𝑧 ∈ Z. Following the script above, we define
O1 := {𝛾 ∈ O× : nrd(𝛾) = 1}
and let
Γ1 (O) := 𝜄(O1 )/{±1} ⊆ PSL2 (R).
Lemma 38.1.2. The group Γ1 (O) ≤ PSL2 (R) is a Fuchsian group.
Recall that a Fuchsian group (Definition 34.7.3) is a discrete subgroup of PSL2 (R);
by Theorem 34.2.1, a Fuchsian group acts properly on the upper half-plane H2 by
orientation-preserving isometries.
Proof. Because Γ1 (O) is a group, it suffices to find an open neighborhood of 1
containing no other element of Γ1 (O). We take

𝑔11 𝑔12
𝑈= ± ∈ PSL2 (R) : |𝑔11 − 1|, |𝑔12 |, |𝑔21 |, |𝑔22 − 1| < 1/2 .
𝑔21 𝑔22
If 𝛾 = (𝑔𝑖 𝑗 )𝑖, 𝑗 ∈ 𝑈 ∩ Γ1 (O), then

|2(𝑡 − 1)| = |𝑔11 + 𝑔22 − 2| ≤ |𝑔11 − 1| + |𝑔22 − 1| < 1
√
|2𝑦 𝑏| = |𝑔12 + 𝑔21 | ≤ |𝑔12 | + |𝑔21 | < 1
and since 𝑡, 𝑦 ∈ Z we have 𝑡 = 1 and 𝑦 = 0. Then
√
|𝑥 𝑎| = |𝑔11 | < 1/2
√
| 𝑧 𝑎𝑏| = |𝑔12 | < 1/2
and since 𝑎, 𝑏, 𝑥, 𝑧 ∈ Z with 𝑎, 𝑏 ≠ 0, we conclude 𝑥 = 𝑧 = 0, and 𝛾 = ±1.
If 𝐵 ' M2 (Q), then Γ1 (O) = PSL2 (Z) and we investigated this case already in
detail. So suppose from now on that 𝐵 ; M2 (Q), or equivalently that 𝐵 is a division
algebra over Q.
Proposition 38.1.3. The quotient Γ1 (O)\H2 is compact.
This proposition is analogous to the finiteness of the class set (as in section 17.5),
and the proof is again inspired by the geometry of numbers. We give a proof in Main
Theorem 38.4.3.
The compactness result above implies nice properties for Γ1 (O), read off from a
fundamental domain (as presented in section 37.1). Let 𝑧0 ∈ H2 have trivial stabilizer
StabΓ1 (O) 𝑧0 = {1}. Then the Dirichlet domain
◊ = ◊(Γ1 (O); 𝑧0 ) = {𝑧 ∈ H2 : 𝜌(𝑧, 𝑧0 ) ≤ 𝜌(𝛾𝑧, 𝑧0 ) for all 𝛾 ∈ Γ1 (O)}

38.2. ISOMETRIES FROM QUATERNIONIC GROUPS 703
is a closed, locally finite fundamental domain for Γ1 (O) with geodesic sides by Theo-
rem 37.1.12.
Corollary 38.1.4. ◊ is a compact, finite-sided hyperbolic polygon, and the group

Γ1 (O) is finitely presented.
Proof. We write Γ = Γ1 (O). Since Γ\H2 is compact by Proposition 38.1.3, the

distance 𝜌(Γ𝑧 0 , Γ𝑧) for Γ𝑧 ∈ 𝑋 (Γ) is bounded. Thus by construction, the Dirichlet
domain is contained in a bounded set and is therefore compact. Since ◊ is locally
finite, we conclude that 𝛾◊ ∩ ◊ ≠ ∅ for only finitely many 𝛾 ∈ Γ. But the set of such
elements are the side pairing elements and they generate Γ (Theorem 37.3.1), so ◊
has finitely many sides and Γ is finitely generated. Thus ◊ has finitely many vertices,
the set of vertex cycle relations is finite, and these generate the relations (Proposition
37.3.14). Thus Γ is finitely presented.
Two subgroups 𝐻1 , 𝐻2 ≤ 𝐺 are commensurable if 𝐻1 ∩ 𝐻2 has finite index in

both 𝐻1 , 𝐻2 .
38.1.5. If Γ ≤ PSL2 (R) is commensurable with Γ1 (O), we say that Γ is an arithmetic

Fuchsian group with defining quaternion algebra 𝐵. This definition is independent
of the choice of order O: another suborder or superorder has finite index, so the
corresponding unit groups will also have finite index.
For every Fuchsian group Γ ≤ PSL2 (R) commensurable with Γ1 (O), the conclu-
sions of Proposition 38.1.3 and Corollary 38.1.4 remain true: for a containment Γ0 ≤ Γ
of finite index, the corresponding map Γ0\H2 → Γ\H2 is finite-to-one, and the desired
properties pass from one quotient to the other.
In this way, we have completed our task: starting with an indefinite quaternion
algebra 𝐵 over Q, we constructed a Fuchsian group Γ1 (O) generalizing PSL2 (Z) acting
on the upper half-plane. We pursue a more general construction in this chapter (con-
sidering an indefinite quaternion over a number field) and its geometry and arithmetic
properties in the remainder of this book.
38.2 Isometries from quaternionic groups

In the remainder of this chapter, we investigate discrete groups obtained from unit
groups of quaternion algebra over number fields. Throughout, let 𝐹 be a number field
with 𝑟 real places and 𝑐 complex places, so that [𝐹 : Q] = 𝑟 + 2𝑐 = 𝑛. Let 𝐵 be a
38.2.1. Suppose that 𝐵 is split at 𝑡 real places and ramified at the remaining 𝑟 − 𝑡 real
places, so that
∼ M (R) 𝑡 × H𝑟 −𝑡 × M (C) 𝑐 .
𝐵 ↩→ 𝐵R := 𝐵 ⊗Q R −→ (38.2.2)
2 2
Let
𝜄 = (𝜄1 , . . . , 𝜄𝑡+𝑐 ) : 𝐵 → M2 (R) 𝑡 × M2 (C) 𝑐 (38.2.3)
denote the map (38.2.2) composed with the projection onto the matrix ring factors.
Then as long as 𝑡 + 𝑐 > 0, we have
𝜄(𝐵× ) ⊂ GL2 (R) 𝑡 × GL2 (C) 𝑐 . (38.2.4)
We have 𝑡 = 𝑐 = 0 if and only if 𝐹 is totally real and 𝐵 is totally definite; in this
case, the geometry disappears, and we are reduced to considering finite unit groups
(see section 32.3). There is still more to say for this case, and we will return to it in
chapter 41 in the context of modular forms. But in this part of the text we are following
geometric threads, so
suppose from now on that 𝑡 + 𝑐 > 0; (38.2.5)
this is another way of saying that 𝐵 is indefinite or equivalently satisfies the Eichler
condition (Definition 28.5.1).
38.2.6. We now restrict to a subgroup acting faithfully via orientation-preserving
isometries. Recall that
×
𝐹>0 = {𝑥 ∈ 𝐹 × : 𝑣(𝑥) > 0 for all real places 𝑣}
is the group of totally positive elements of 𝐹. Let
𝐵×>0 := {𝛼 ∈ 𝐵× : nrd(𝛼) ∈ 𝐹>0
×
} (38.2.7)
be the group of units of 𝐵 of totally positive reduced norm. Then 𝐹 × ⊂ 𝐵×>0 because
× for all 𝑎 ∈ 𝐹 × . Let
nrd(𝑎) = 𝑎 2 ∈ 𝐹>0
P𝐵×>0 := 𝐵×>0 /𝑍 (𝐵×>0 ) = 𝐵×>0 /𝐹 × (38.2.8)
be the quotient by the center; then 𝜄 induces an inclusion
P𝜄(𝐵×>0 ) ⊂ PSL2 (R) 𝑡 × PSL2 (C) 𝑐
(we have PSL2 (C) ' PGL2 (C) and PGL+2 (R) ' PSL2 (R), rescaling by the determi-
nant). Therefore, the group P𝜄(𝐵×>0 ) acts on
H := (H2 ) 𝑡 × (H3 ) 𝑐 (38.2.9)

on the left faithfully by linear fractional transformations as orientation-preserving
isometries (Theorems 33.3.14 and 36.2.14), that is to say
P𝜄(𝐵×>0 ) ⊂ Isom+ (H ). (38.2.10)
Remark 38.2.11. One can equally well consider P𝜄(𝐵× ) ⊆ Isom(H ). Indeed,
PSL2 (R) 𝑡 × PSL2 (C) 𝑐 ' Isom+ (H2 ) 𝑡 × Isom+ (H3 ) 𝑐 ≤ Isom+ (H )
is the subgroup of isometries that preserve each factor and preserving orientation. The
full group Isom(H ) includes more: permuting factors of the same kind is an isometry,
and with these we have an isomorphism of groups
Isom(H ) ' (Isom(H2 ) 𝑡 × Isom(H3 ) 𝑐 ) o (𝑆𝑡 × 𝑆 𝑐 ) (38.2.12)
38.3. DISCRETENESS 705
where 𝑆 𝑚 denotes the symmetric group. Since PSL2 (R) ' Isom+ (H2 ) ≤ Isom(H2 )
has index 2 and the same for H3 , we conclude that
[Isom(H ) : PSL2 (R) 𝑡 × PSL2 (C) 𝑐 ] = 2𝑡+𝑐 𝑡!𝑐!.
The orientation-preserving subgroup Isom+ (H ) ≤ Isom(H ) has index 2, contain-
ing for example elements which reverse orientation in two components and preserve
orientation in the others; if the orientation is reversed in two components H2 , then the
resulting isometry is orientation-preserving but not holomorphic in these components.
38.3 Discreteness
We now seek discrete subgroups, as follows.
38.3.1. Let 𝑅 = Z𝐹 be the ring of integers of 𝐹 and let O ⊂ 𝐵 be an 𝑅-order. Let
O1 := {𝛾 ∈ O× : nrd(𝛾) = 1} ≤ O× (38.3.2)
be the subgroup of units of reduced norm 1, and let
PO1 := O1 /𝑍 (O1 ) = O1 /{±1}
(38.3.3)
Γ1 (O) := P𝜄(O1 )
Then by (38.2.10), we have Γ1 (O) < Isom+ (H ).
Definition 38.3.4. A subgroup Γ ≤ Isom+ (H ) is arithmetic if Γ is commensurable
with Γ1 (O) for a quaternion algebra 𝐵 and an order O ⊆ 𝐵 (with respect to some
embedding 𝜄).
If Γ is commensurable with Γ1 (O) for an order O then it is commensurable with
Γ1 (O0) for every other order O0, since every two orders have finite 𝑅-index, thus finite
index—so we could equally well compare to one fixed (e.g. maximal) order. The class
of arithmetic groups contains in particular the quaternionic unit groups with which
we started, but contains other groups of interest (including subgroups and discrete
supergroups with finite index).
Remark 38.3.5. There is a more general definition of arithmetic group which reduces
to this one; see section 38.5.
Right away, we show that arithmetic groups are discrete. To do so, we will need
two short lemmas.
Lemma 38.3.6. Let 𝐾, 𝑋 be Hausdorff topological spaces, and suppose 𝐾 is compact.
Let 𝜋 : 𝐾 × 𝑋 → 𝑋 be the projection, and let 𝑌 ⊆ 𝐾 × 𝑋 be discrete and closed. Then
𝜋(𝑌 ) ⊆ 𝑋 is discrete.
Proof. First, 𝑌 has no limit points: a limit point of 𝑌 would belong to 𝑌 , but then 𝑌 is
discrete so every point is isolated point. For the same reason, every subset of 𝑌 is also
(discrete and) closed: a limit point of the subset would be a limit point of 𝑌 .
Now let 𝑥 ∈ 𝜋(𝑌 ) and let 𝑌𝑥 = 𝑌 r 𝜋 −1 (𝑥). Then 𝑌𝑥 ⊆ 𝐾 × 𝑋 is closed. The set
(𝐾 × 𝑋) r 𝑌𝑥 is open and contains 𝐾 × {𝑥}, so by the tube lemma, it contains an open
set 𝐾 × 𝑈. Then 𝑈 3 𝑥 is the desired neighborhood.
Lemma 38.3.7. Let 𝐺 be a Hausdorff topological group and let 𝐻 ≤ 𝐺 be a discrete

subgroup. Then 𝐻 is closed.
Proof. Since 𝐻 is discrete, there is a neighborhood 𝑈 3 1 such that 𝑈 ∩ 𝐻 = {1}.
Further, there exists a neighborhood 𝑉 ⊆ 𝑈 such that 𝑉 −1𝑉 ⊆ 𝑈 (multiplication and
inversion are continuous, see Exercise 12.4).
We show 𝐺 r 𝐻 is open. For 𝑥 ∈ 𝐺, we have 𝑥𝑉 3 𝑥 an open neighborhood, and
if ℎ, ℎ 0 ∈ 𝑥𝑉 ∩ 𝐻 then 𝑥 −1 ℎ, 𝑥 −1 ℎ 0 ∈ 𝑉 and so (𝑥 −1 ℎ) −1 (𝑥 −1 ℎ 0) = ℎ−1 ℎ 0 ∈ 𝑉 −1𝑉 ⊆ 𝑈.
Therefore ℎ−1 ℎ 0 = 1 by the hypothesis on 𝑈, so ℎ = ℎ 0. Thus 𝑥𝑉 contains at most one
element of 𝐻. Since 𝐺 is Hausdorff, when 𝑥 ∉ 𝐻 we can shrink 𝑉 if necessary to get
𝑥𝑉 ∩ 𝐻 = ∅, as desired.
Proposition 38.3.8. Let Γ ≤ Isom+ (H ) be an arithmetic subgroup. Then Γ is discrete.
Proof. It is enough to prove the proposition for Γ = Γ1 (O), as discreteness is preserved
between commensurable groups (having finite index in their intersection).
The image 𝑂 ↩→ 𝐵R as in (38.2.2) is discrete by 17.7.6: we argued using coordi-
nates and noted that 𝑅 = Z𝐹 ↩→ 𝐹R is discrete. Therefore the image
O1 ↩→ 𝐵R1 ' (H1 ) 𝑟 −𝑡 × SL2 (R) 𝑡 × SL2 (C) 𝑐 (38.3.9)
is discrete (by restriction). Further, since H1
is compact, by Lemmas 38.3.6 and 38.3.7,
the image of O1 ↩→ SL2 (R) 𝑡 × SL2 (C) 𝑐 under the projection is discrete. (Any further
projection turns out not to be discrete; see Exercise 38.2.)
38.3.10. The group Γ1 (O) is a Fuchsian group if and only if 𝑡 = 1 and 𝑐 = 0, i.e. 𝐹
is totally real and 𝐵 is ramified at all but one real place; Γ1 (O) is a Kleinian group if
and only if 𝑡 = 0 and 𝑐 = 1, i.e. 𝐹 has exactly one complex place and 𝐵 is ramified at
all real places.
Just as for Fuchsian and Kleinian groups, discrete groups admit several equivalent
characterizations as follows. (For the notion of a good orbifold, see Definition 34.8.10.)
Proposition 38.3.11. Let Γ ≤ Isom+ (H ) be a subgroup. Then the following are
equivalent:
(i) Γ is discrete (with the subspace topology);
(ii) For all 𝑧 ∈ H , we have # StabΓ (𝑧) < ∞ and there exists an open neigborhood
(iii) For all compact subsets 𝐾 ⊆ H , we have 𝐾 ∩ 𝛾𝐾 ≠ ∅ for only finitely many
𝛾 ∈ Γ; and
(iv) For all 𝑧 ∈ H , the orbit Γ𝑧 ⊆ H is discrete and # StabΓ (𝑧) < ∞.
Moreover, if these equivalent conditions hold, then the quotient Γ\H has the structure
of a good Riemann orbifold of dimension 𝑚 = 2𝑡 + 3𝑐, and the quotient map
𝜋 : H → Γ\H
is a local isometry at all points 𝑧 ∈ H with StabΓ (𝑧) = {1}.
Proof. We proved this statement in Propositions 34.7.2 and 36.4.1 when 𝑡 = 0 or 𝑐 = 0;
the general case follows similarly.
38.4. COMPACTNESS AND FINITE GENERATION 707
38.4 Compactness and finite generation

We now consider further properties of arithmetic groups. Refreshing our notation, let
Γ ≤ Isom+ (H ) be an arithmetic group arising from a quaternion algebra 𝐵.
38.4.1. The arithmetic discrete groups arising from the case 𝐵 = M2 (𝐹) of the matrix
ring are of particular interest: they include the case 𝐹 = Q giving rise to the classical
modular group Γ = PSL2 (Z) studied in chapter 35 as well as the case 𝐹 = Q(𝑖) giving
rise to the Picard group Γ = PSL2 (Z[𝑖]) examined in section 36.6 and more generally
the Bianchi groups. Let 𝐵 = M2 (𝐹). Then necessarily 𝑟 = 𝑡 (the matrix ring is
already split!), so H = (H2 ) 𝑟 × (H3 ) 𝑐 , and the embedding 𝜄 in (38.2.3) has the simple
description
𝜄 : M2 (𝐹) ↩→ M2 (𝐹)R ' M2 (R) 𝑟 × M2 (C) 𝑐

𝑎 𝑏 𝑎𝑣 𝑏𝑣 (38.4.2)
𝛼= ↦→
𝑐 𝑑 𝑐𝑣 𝑑𝑣 𝑣
where we embed matrices componentwise, indexed by the archimedean places of 𝐹.
Much is written about the compactification of 𝑌 (Γ), and unfortunately it would
take us too far afield to fully treat this important topic: the groups so obtained behave
differently in several respects than the case where 𝐵 is a division algebra.
We suppose throughout this section that 𝐵 is a division algebra.

Let 𝑋 (Γ) := Γ\H be the quotient, a good Riemann orbifold.
Main Theorem 38.4.3 (Hey). The orbifold 𝑋 (Γ) is compact.
Proof. We follow Zassenhaus [Zas72, §1]; see also Kleinert [Klt2000, Theorem 1.1].
First, the group Γ is commensurable with 𝜄(O1 ) for an 𝑅-order O; by comparison
under maps of finite index, we may suppose Γ = Γ1 (O).
Second, we claim that it suffices to show that O1 \𝐵R1 is compact. Indeed, recall
(38.3.9) that
O1 ↩→ 𝐵R1 ' (H1 ) 𝑟 −𝑡 × SL2 (R) 𝑡 × SL2 (C) 𝑐 .
From the symmetric space models (34.6.3) and (36.3.14), we have homeomorphisms
∼ H2
SL2 (R)/SO(2) −→
SL (C)/SU(2) −→∼ H3 ;
2
so the fibers of the continuous projection map
O1 \𝐵R1 → Γ\H = 𝑌 (Γ)
are given by
(H1 ) 𝑟 −𝑡 × SO2 (R) 𝑡 × SU(2) 𝑐
and therefore compact, and the claim follows.
Now from (38.2.2), we have
O ↩→ 𝐵R ' M2 (R) 𝑡 × H𝑟 −𝑡 × M2 (C) 𝑐 ;

choosing an R-basis for 𝐵R , we have 𝐵R ' R4𝑛 and under the standard metric, this
gives O ' Z4𝑛 the (non-canonical) structure of a Euclidean lattice. Let 𝑋 ⊆ 𝐵R
be a compact, convex, symmetric subset with volume vol(𝑋) > 24𝑛 covol(O). (The
precise value of this covolume will not figure in the argument, and anyway would
depend on the choice of Euclidean structure; all such structures induce the same
topology.) Therefore, by Minkowski’s convex body theorem (Theorem 17.5.5), there
exists nonzero 𝛼 ∈ O ∩ 𝑋. Moreover, for all 𝑔 = (𝑔 𝑣 ) 𝑣 ∈ 𝐵R1 we have
vol(𝑔𝑋) = vol(𝑋)
Î
since 𝑣 det(𝑔 𝑣 ) = 1, and 𝑔𝑋 is again compact, convex, and symmetric, so similarly
there exists nonzero 𝛼𝑔 ∈ O ∩ 𝑔𝑋.
We will show that the quotient space O1 \𝐵R1 is sequentially compact. To this end,
let 𝑔𝑛 be a sequence from 𝐵R1 . By the previous paragraph, there exist 𝛼𝑛 ∈ O such that
𝛼𝑛 = 𝑔𝑛 𝑥 𝑛 with 𝑥 𝑛 ∈ 𝑋 nonzero. Since 𝑋 is compact, we can restrict to a subsequence
such that 𝑥 𝑛 → 𝑥 ∈ 𝑋 converges.
The reduced norm nrd : 𝐵 → 𝐹 extends by scaling to a continuous function
nrd : 𝐵R → 𝐹R . Since 𝑋 ⊆ 𝐵R is bounded so too is nrd(𝑋) ⊆ 𝐹R bounded. But
nrd(𝛼𝑛 ) = nrd(𝑔𝑛 ) nrd(𝑥 𝑛 ) = nrd(𝑥 𝑛 ) ∈ nrd(𝑋),
and the values nrd(𝛼𝑛 ) ∈ nrd(O) ⊆ Z𝐹 lie in a discrete subset, so there are only
finitely many possibilities for nrd(𝛼𝑛 ). Moreover, the left ideals 𝐼𝑛 = O𝛼𝑛 have
N(𝐼𝑛 ) = Nm𝐹 |Q (nrd(𝐼𝑛 )) 2 = Nm𝐹 |Q (nrd(𝛼𝑛 )) 2 (38.4.4)
bounded, so there are only finitely many possibilities for 𝐼𝑛 by Lemma 17.7.26. Thus,
by the pigeonhole principle and restricting to a subsequence, we may suppose 𝐼𝑛 =
O𝛼𝑛 = O𝛼1 and nrd(𝛼𝑛 ) = nrd(𝛼1 ) for all 𝑛. Therefore 𝛼𝑛 = 𝛾𝑛 𝛼1 with 𝛾𝑛 ∈ O1 for
all 𝑛.
To conclude, we note that since 𝐵 is a division algebra,
nrd(𝑥 𝑛 ) = nrd(𝛼𝑛 ) = nrd(𝛼1 ) ≠ 0
so
𝑥 𝑛−1 = 𝑥 𝑛 nrd(𝑥 𝑛 ) −1 = 𝑥 𝑛 nrd(𝛼1 ) −1 → 𝑥 nrd(𝛼1 ) −1 = 𝑥 −1
converges. Therefore
𝛾𝑛−1 𝑔𝑛 = 𝛾𝑛−1 𝛼𝑛 𝑥 𝑛−1 = 𝛼1 𝑥 𝑛−1 → 𝛼1 𝑥 −1
converges as well. Therefore the quotient O1 \𝐵R1 is sequentially compact, and therefore
compact.
Remark 38.4.5. Main Theorem 38.4.3 was proven by Hey [Hey29, Hilfssatz 4] in her
1929 Ph.D. thesis (see also Remark 29.10.24) in the case where 𝐵 is a division algebra
over Q. In particular, there is no need to suppose that 𝐵 is central, so it contains the
Dirichlet unit theorem as a consequence: see Exercise 38.6.
After treating the decomposition of the adelic coset space, we will also see Hey’s
theorem as an essentially direct consequence of Fujisaki’s lemma (Main Theorem
27.6.14): see Theorem 38.7.21.
38.5. ∗ ARITHMETIC GROUPS, MORE GENERALLY 709
Theorem 38.4.6. The group Γ is finitely generated.
Proof. Let 𝑥 0 ∈ 𝑋 (Γ) satisfy StabΓ (𝑥0 ) = {1}. Then by Theorem 37.4.18, the
Dirichlet domain ◊ = ◊(Γ; 𝑥0 ) is a locally finite fundamental set for Γ. By Main
Theorem 38.4.3, 𝑋 (Γ) is compact, so too is ◊. Then since ◊ is locally finite, we
conclude that 𝛾◊ ∩ ◊ ≠ ∅ for only finitely many 𝛾 ∈ Γ. But then by Theorem 37.4.2,
the group Γ is generated by such elements and Γ is finitely generated.
Remark 38.4.7. In fact, Γ is finitely presented, and an argument similar to Proposition

37.3.14 shows that orbits of nonempty subsets
◊ ∩ 𝛾◊ ∩ 𝛾 0◊
with 𝛾 ≠ 𝛾 0 provide a finite generating set of relations (generalizing the vertex cycle
relations): see e.g. Raghunathan [Rag72, Theorem 6.15].
38.5 ∗ Arithmetic groups, more generally

In this section, we briefly discuss more general definitions of arithmetic group and
show that they reduce to the “working” one given above (Definition 38.3.4). As above,
let 𝐹 be a number field.
38.5.1. Let G ≤ GL𝑛,𝐹 be a linear algebraic group, a subgroup variety of GL𝑛,𝐹

defined by polynomial equations in the entries and the inverse of the determinant, with
coefficients in 𝐹. Equivalently, G is an affine variety over 𝐹 equipped with identity,
multiplication, and inversion morphisms giving it the structure of a group variety.)
We say a subgroup
Ö Ö
Γ ≤ G (𝐹∞ ) = G (𝐹𝑣 ) ≤ GL𝑛 (𝐹𝑣 )
𝑣 |∞ 𝑣 |∞
is arithmetic (as a subgroup of G (𝐹∞ )) if it is commensurable with G (Z𝐹 ). This

notion of arithmetic group was developed significantly by Borel [Bor62, Bor69].
38.5.2. Let 𝐵 be a quaternion algebra over 𝐹. Then there is an embedding 𝜌 : 𝐵 ↩→

M4 (𝐹) as in Exercise 2.11 by the regular representation. Thus 𝐵× ≤ GL4 (𝐹) and the
image is described by explicit polynomial equations (Exercise 38.5). Therefore there
exists a linear algebraic group G ≤ GL4,𝐹 such that G (𝐹) ' 𝐵× .
Similar statements hold for 𝐵1 ≤ SL4 (𝐹).
Lemma 38.5.3. A group Γ commensurable with Γ1 (O) for a quaternion algebra 𝐵

and order O ⊆ 𝐵 is arithmetic in the sense of 38.5.1.
Proof. Applying 38.5.2, we have 𝜌 : 𝐵 ↩→ M4 (𝐹) realizing 𝐵1 ' G (𝐹) ≤ GL4 (𝐹) as
a linear algebraic group by appropriate polynomial equations. Under
Î this embedding,
O := 𝜌(𝐵) ∩ M4 (Z𝐹 ) is a Z𝐹 -order, and thus O1 ' G (Z𝐹 ) and 𝑣 G (𝐹𝑣 ) ' 𝑣 𝐵1𝑣 ,
Î
as required in the definition.
In view of Definition 38.3.4, we consider now the converse: when does the more
general definition give rise to discrete subgroups of two-by-two matrices?
38.5.4. Let G be a linear algebraic group over 𝐹, and suppose that G𝐹 ' GL2,𝐹 , so
there is a chance to obtain discrete groups of symmetries like the ones considered
above. (For a complete treatment, we should consider SL2 as well as the group PGL2 ,
but the arguments are similar.) We say then that G is an 𝐹-form of the algebraic group
GL2 .
Lemma 38.5.5. Let G be an 𝐹-form of GL2 . Then there exists a quaternion algebra
𝐵 over 𝐹, unique up to 𝐹-algebra isomorphism, such that G (𝐹) ' 𝐵× .
Proof. This is a basic result in non-abelian Galois cohomology, and it would take us
too far afield to prove it here: see e.g. Milne [Milne2017, Theorem 20.3.5] and more
generally Serre [Ser79, Chapter X].
Lemma 38.5.5 explains that more general notions of arithmetic groups do not
create anything new beyond our quaternionic definition.
Remark 38.5.6. In this context, there is a criterion for compactness, generalizing
Main Theorem 38.4.3 (conjectured by Godement): A discrete subgroup of G (R)
is cocompact if and only if the reductive part of the connected component of G is
anisotropic over 𝐹. If G is semisimple, then cocompactness is equivalent to asking
that every element of G (𝐹) is semisimple. This criterion was proven by Borel–
Harish-Chandra [BHC62] and Mostow–Tamagawa [MT62]; Godement [God62] (with
Weil) extended the method of Mostow–Tamagawa and simplified the proof by working
directly on adele groups. See also Platonov–Rapinchuk [PR94, §4.5].
38.6 ∗ Modular curves, seen idelically

We have already seen how idelic methods can be both a conceptual and a computa-
tional simplification. The quaternion groups defined above naturally also fit into this
perspective, and we describe this in the final two sections. As motivation, we begin in
this section by reconsidering the classical modular curves from an idelic point of view.
38.6.1. Recall that the adeles of Q decompose as
b×R
Q=Q
into finite and infinite parts. Let 𝐵 = M2 (Q). Then
𝐵 = M2 (Q) = M2 ( Q)
b × M2 (R) = 𝐵
b × 𝐵∞ .
The order O = M2 (Z) is maximal in 𝐵, and we have the adelic order O b = M2 ( bZ) ⊂
M2 ( Q). (We have seen that 𝐵 = M2 (Q) ≤ 𝐵 = M2 (Q) sits discretely and the quotient
b
𝐵\𝐵 is compact; like the adelic quotient 𝐹\𝐹 itself, this is not very interesting.)
Similarly, we have
𝐵× = GL2 (Q) = 𝐵 ×
b × 𝐵∞ = GL2 ( Q)
b × GL2 (R). (38.6.2)
38.6. ∗ MODULAR CURVES, SEEN IDELICALLY 711
It was a key consequence of strong approximation—but easy to establish in this case

(Lemma 28.2.4)—that
𝐵× \ 𝐵
b×/O
b× = GL2 (Q)\GL2 ( Q)/GL
b 2 (b
Z) = {1} (38.6.3)
is trivial: every nonzero right (invertible) fractional M2 (Z)-ideal is principal, or

equivalently, every b b2 has a basis in Q2 .
Z-lattice in Q
As lovely as this is, this description leaves out the real place, and by putting it back
we restore archimedean structure.
38.6.4. The projection map
GL2 (Q) = GL2 ( Q)

b × GL2 (R) → GL2 ( Q)
b
yields a continuous projection
GL2 (Q)\GL2 (Q)/GL2 (b

Z) → GL2 (Q)\GL2 ( Q)/GL
b 2 (b
Z) = {1}
by (38.6.3). Therefore, every element of GL2 (Q) is represented in the double coset by
an element of the form (1, 𝛼∞ ) with 𝛼∞ ∈ GL2 (R); and the element 𝛼∞ is well-defined
up the action of the group of pairs (𝛾, 𝜇
b) ∈ GL2 (Q) × GL2 (b
Z) satisfying
0
𝛾(1, 𝛼∞ ) 𝜇
b = (𝛾 𝜇
b, 𝛾𝛼∞ ) = (1, 𝛼∞ )
so 𝛾 = 𝜇b−1 ∈ GL2 (Q) ∩ GL2 (b

Z) = GL2 (Z), acting on the left. In other words, we
have a bĳection
GL2 (Q)\GL2 (Q)/GL2 (b

Z) ↔ GL2 (Z)\GL2 (R). (38.6.5)
38.6.6. At this point, we are no stranger to the quotient GL2 (Z)\ GL2 (R)! We have
studied in detail the related quotient SL2 (Z)\ SL2 (R): by the symmetric space de-
scription, we have an isometry
∼ H2
SL2 (R)/SO(2) −→
(38.6.7)
𝑔 SO(2) ↦→ 𝑔𝑖.
We similarly obtain a bĳection

∼ H2± = C r R
GL2 (R)/R× SO(2) −→ (38.6.8)
(any matrix of negative determinant interchanges the upper and lower half-planes, so
we maintain a bĳection). Then we can take the quotient on the left by GL2 (Z) to get
an identification
𝑌 (1) = PSL2 (Z)\H2 = GL2 (Z)\GL2 (R)/R× SO(2) (38.6.9)
from (38.6.7), and we find again the classical modular curve we considered in section
35.1. (In section 40.1, we will see another version of this in that the space 𝑌 (1)
parametrizes complex lattices Λ ⊆ C up to scaling by C× .)
Putting (38.6.9) and (38.6.5) together, we have
𝑌 (1) ↔ GL2 (Z)\H2±

↔ GL2 (Z)\GL2 (R)/R× SO(2)
Z))/(R× SO(2))
↔ (GL2 (Q)\GL2 (Q)/GL2 (b (38.6.10)
= GL2 (Q)\(GL2 ( Q)/GL
b Z) × GL2 (R)/R× SO(2))
2 (b
= 𝐵× \( 𝐵
b× /O
b× × 𝐵× /𝐾∞ ) = 𝐵× \𝐵× /𝐾;
∞
in the last line, we introduce the notation 𝐾∞ := R× SO(2) and 𝐾 := O

b× × 𝐾∞ to help in
grasping this double coset. Chasing down all of the maps, the bĳection is obtained by
sending an element 𝑧 ∈ H2± to the class of (1, 𝛼) ∈ GL2 ( Q)
b × GL2 (R) where 𝑧 = 𝛼𝑖.
Note there is a nice symmetry in the expression on the right-hand side of (38.6.10).
Remark 38.6.11. In this way, we just wrapped the classical quotient in ideles. This
double cosetification (a beautiful monster of a word) provides a uniform way to describe
the orbifold quotients obtained from quaternionic arithmetic groups more generally:
in particular, class number issues are made more transparent in the language of double
cosets. On the other hand, geometric structures are not always visible in this language,
which is why we have treated both approaches in this text.
38.7 ∗ Double cosets

In this section, we give a description of quaternionic orbifolds in terms of idelic double
cosets. We retain the notation from 38.2.1, 38.2.6, and 38.3.1. In particular, 𝐹 is a
number field with 𝑟 real places and 𝑐 complex places, 𝐵 is a quaternion algebra over
𝐹 that is split at 𝑡 real places with 𝑡 + 𝑐 > 0, we have an embedding
𝜄 : 𝐵 → M2 (R) 𝑡 × M2 (C) 𝑐 .
We have defined
H = (H2 ) 𝑡 × (H3 ) 𝑐
where we may also write
H2 ↔ SL2 (R)/SO(2) and H3 ↔ SL2 (C)/SU(2)
by the symmetric space decomposition. We similarly define
H ± = (H2± ) 𝑡 × (H3 ) 𝑐
where similarly
H2± ↔ GL2 (R)/R× SO(2) and H3 ↔ GL2 (C)/C× SU(2). (38.7.1)
In this chapter we have been working with groups Γ ≤ Isom+ (H ) commensurable

with Γ1 (O) = P𝜄(O1 ). To work idelically, we restrict our attention to such groups that
can be defined idelically: these are the congruence subgroups.
38.7. ∗ DOUBLE COSETS 713
Definition 38.7.2. Let 𝔐 ⊆ 𝑅 be a nonzero ideal. Let

O(𝔐) := {𝛼 ∈ O : 𝛼 ∈ 𝑅 + 𝔐O}.
The principal congruence subgroup of level 𝔐 ⊆ 𝑅 is the group
Γ(𝔐) = Γ1 (O(𝔐)) := P𝜄(O(𝔐) 1 ) ≤ Isom+ (H )
obtained from the units of norm 1 in O(𝔐).
A group Γ commensurable with Γ1 (O) is congruence if it contains Γ(𝔐) for some
𝔐.
A congruence subgroup is defined by (finitely many) congruence conditions. Let
Γ be the closure of Γ with respect to the topology on 𝐵
b b× , where observe that a
fundamental system of open neighborhoods of 1 is given by the images of the principal
congruence subgroups.
Γ ∩ 𝐵× = Γ if
Lemma 38.7.3. Suppose that Γ is commensurable with Γ1 (O). Then b
and only if Γ is congruence.
Proof. The group b b× commensurable with the compact open
Γ is a closed subgroup of 𝐵
d b b× , the
subgroup Γ(1) and so Γ is also compact open. By definition of the topology on 𝐵
×
closure of Γ := Γ ∩ 𝐵 is the smallest congruence group containing Γ, and Γ contains
b
Γ with finite index. So Γ is congruence if and only if Γ = Γ.
From now on, suppose that Γ is a congruence subgroup.
Remark 38.7.4. In general, we may work with the congruence closure b Γ ∩ 𝐵× ≥ Γ of
an arithmetic group Γ.
Î
38.7.5. In view of (38.6.10) and (38.7.1), we let 𝐾∞ := 𝑣 |∞ 𝐾 𝑣 where


 R× H1 = H× , if 𝑣 is real and ramified in 𝐵;


𝐾 𝑣 := R× SO(2), if 𝑣 is real and split in 𝐵; (38.7.6)

 C× SU(2),

if 𝑣 is complex.

The groups 𝐾 𝑣 are the extension of a maximal compact subgroup by the center. We
Γ × 𝐾∞ and consider the double coset space
then let 𝐾 = b
𝑌 Sh (Γ) := 𝐵× \𝐵× /𝐾 (38.7.7)
where 𝐵× 𝐵×
acts on by left multiplication (under the diagonal embedding) and 𝐾 acts
× × ×
on 𝐵 = 𝐵 × 𝐵∞ by right multiplication.
b
The expression (38.7.7) is tidy, and generalizes well, but we also want to know
what it looks like. Plugging back in (38.7.1), we have 𝐵∞ × /𝐾 = H ± , so
∞
𝑌 Sh (Γ) = 𝐵× \(( 𝐵
b× /b
Γ) × H ± ) (38.7.8)
where 𝐵× acts on H ± via 𝜄 and on 𝐵 b× /b
Γ by left multiplication. Since there can
hopefully no confusion about this action, removing parentheses we will write
𝑌 Sh (Γ) = 𝐵× \( 𝐵
b× × H ± )/b
Γ.
Definition 38.7.9. We call 𝑌 Sh (Γ) the quaternionic Shimura orbifold of level Γ.

Definition 38.7.9 explains the subscript Sh . We immediately proceed to justify the
name orbifold; moreover, we will see that the space is possibly disconnected, and we
write it as a union of connected components.
38.7.10. By weak approximation, there exist elements 𝛼 ∈ 𝐵× with nrd(𝛼) having all
possible real signs at the split real places of 𝐵. Therefore
𝑌 Sh (Γ) = 𝐵×>0 \( 𝐵
b× × H )/b
Γ. (38.7.11)
38.7.12. There is a natural (continuous) projection map
𝑌 Sh (Γ) → 𝐵×>0 \ 𝐵
b× /b
Γ. (38.7.13)
Recall that 𝑡 + 𝑐 > 0, and 𝐵 is indefinite. Therefore, strong approximation (as in
Corollary 28.6.8) implies that there is a bĳection
nrd : 𝐵×>0 \ 𝐵
b× /b ∼ 𝐹 × \𝐹
Γ −→ b× /nrd(b
Γ) =: Cl+𝐺 (Γ) 𝑅. (38.7.14)
>0
Therefore Cl+𝐺 (Γ) 𝑅 is a (narrow) class group of 𝐹 associated to the group Γ; as such,
it is a finite abelian group that surjects onto the strict class group Cl+ 𝑅.
38.7.15. By 38.7.12, the space 𝑌 Sh (Γ) is the disjoint union of finitely many connected
components indexed by the group Cl+𝐺 (Γ) 𝑅. We identify these connected components
explicitly as follows.
Let the ideals 𝔟 ⊆ 𝑅 form a set of representatives for Cl+𝐺 (Γ) 𝑅, and let b
𝔟 = 𝔟 ⊗Z b
Z
be their adelification; then for each 𝔟, there exists 𝑏 ∈ 𝑅 generating 𝔟, so
b b b
𝑏𝑅
b b ∩ 𝑅 = 𝔟.
For simplicity, choose 𝔟 = 𝑅 and b𝑏 =b1 for the representatives of the trivial class.
By surjectivity of the map (38.7.14), for each b b× such that
𝑏 there exists 𝛽b ∈ 𝐵
nrd( 𝛽) = 𝑏. Therefore
b b
Ä
𝑌 Sh (Γ) = 𝐵×>0 ( 𝛽bb
Γ × H ). (38.7.16)
𝔟
For each 𝔟, let
Γ 𝛽b−1 ∩ 𝐵×>0
Γ𝔟 := 𝛽bb (38.7.17)
Then we have a natural bĳection
𝐵×>0 ( 𝛽bb
Γ × H ) ↔ Γ𝔟 \H
(38.7.18)
( 𝛽bb
Γ, 𝑧) ↦→ 𝑧.
Letting
𝑌 (Γ𝔟 ) := Γ𝔟 \H
we see that each 𝑌 (Γ𝔟 ) is a connected orbifold of dimension 2𝑡 + 3𝑐. We abbreviate
𝑌 (Γ) = 𝑌 (Γ (1) ) for the trivial class. Putting these together, we have
Ä
𝑌 Sh (Γ) = 𝑌 (Γ𝔟 ) (38.7.19)
𝔟
as a disjoint union of connected orbifolds.

38.7. ∗ DOUBLE COSETS 715
Remark 38.7.20. When 𝑐 = 0, i.e., 𝐹 is totally real, then 𝑌 Sh (Γ) can be canonically
given the structure of an algebraic variety defined over a number field, by work of
Shimura [Shi67] and Deligne [Del71]; in this case we upgrade 𝑌 Sh (Γ) to a quater-
nionic Shimura variety. The theory of Shimura varieties is both broad and deep—see
Milne [Milne-SV] and the references therein. We will begin the study quaternionic
Shimura varieties of dimension 1 in Chapter 43, touching upon the theory of canonical
models in section 43.8.
The above description has hopefully provided a more transparent way to understand
arithmetic orbifolds. For example, we can prove an idelic version of Hey’s theorem
(Main Theorem 38.4.3) as follows.
Theorem 38.7.21. Suppose that 𝐵 is a division algebra. Then 𝑌 Sh (Γ) is compact.
Proof. We appeal to Fujisaki’s lemma (Main Theorem 27.6.14): the quotient 𝐵× \𝐵 (1)
is cocompact, under the important hypothesis that 𝐵 is a division algebra. Let 𝐾∞ be
as in 38.7.6, and consider the inclusion followed by the projection
𝐵× \𝐵 (1) ↩→ 𝐵× \𝐵× → 𝐵× \𝐵× /𝐾 = 𝑌 Sh (Γ). (38.7.22)
These maps are continuous, and the composition is surjective as R× ≤ 𝐾∞ (embedded

diagonally) and R× 𝐵 (1) = 𝐵× . So the target is compact, proving the statement.
Exercises
In these exercises, we maintain the notation in this chapter: let 𝐹 be a number field
with 𝑟 real places and 𝑐 complex places, degree 𝑛 = [𝐹 : Q], and ring of integers 𝑅,
and let O be an 𝑅-order in a quaternion algebra 𝐵 over 𝐹.
√
1. Let 𝛼 ∈ R r Q. Show that Z[𝛼] is not discrete in R. (Taking e.g. 𝛼 = 𝑑, this
gives a reason to worry about discreteness of number fields when we project.)
2. Embed O1 diagonally in SL2 (R) 𝑟 −𝑡 × SL2 (C) 𝑐 . Show that a (further) projection
to a proper factor is not discrete.

𝑎, 𝑏
3. Let 𝐵 = , and let 𝑣 be a split real place of 𝐵. Show that O1 ↩→ 𝐵1𝑣 '
𝐹
SL2 (R) if and only if 𝐹 is totally real and for all nonidentity real places 𝑣 0, we
have 𝑣 0 (𝑎) < 0 and 𝑣(𝑏) < 0.
4. In this exercise, we give a direct argument for the discreteness of an arithmetic
Fuchsian group. Suppose 𝐹 is totally real, let 𝑣 be a split place of 𝐵, consider
𝐹 ↩→ 𝑣(𝐹) ⊆ R as a subfield of R, and suppose that 𝐵 is ramified at all other
(nonidentity) real places. Prove that O1 ⊆ SL2 (R) is discrete.
5. Consider the regular representation 𝜌 : 𝐵 ↩→ M4 (𝐹) (Exercise 2.11). Describe
the image explicitly in terms of polynomial equations in matrix entries. Conclude
that 𝐵× and 𝐵1 are also described by polynomial equations.
(a) Suppose not: then there exists a sequence 𝛼𝑛 = 𝑡 𝑛 + 𝑥 𝑛 𝑖 + 𝑦 𝑛 𝑗 + 𝑧 𝑛 𝑖 𝑗 → 1

with 𝑡 𝑛 , 𝑥 𝑛 , 𝑦 𝑛 , 𝑧 𝑛 ∈ 𝐹 with bounded denominators. Multiplying through,
suppose that all coordinates are integral. Show for 𝑛 sufficiently large that
all of the coordinates are integral and bounded.
(b) Show that for all nonidentity 𝑣, the coordinates of 𝑣(𝛼𝑛 ) are also bounded
using compactness.
(c) Finally, prove that there are only a finite number of elements in 𝑅 that are
bounded in each coordinate (all conjugates are bounded). [Hint: look at
the coefficients of a minimal polynomial, and derive a contradiction.]
6. Let 𝐹 be a number field and 𝑅 = Z𝐹 its ring of integers. In this exercise, we
give a proof of Dirichlet’s unit theorem using the same method as in the proof
of Main Theorem 38.4.3.
(a) Show that [𝑅 × : 𝑅 1 ] ≤ 2.
(b) Show (following the proof of Main Theorem 38.4.3) that 𝑅 1 \𝐹R1 is compact.
(c) Under the usual logarithmic embedding log : 𝐹R1 → ( 𝑣 R) 0 , conclude
Î
that log 𝑅 1 \ log 𝐹R1 is compact, and therefore log 𝑅 1 has rank 𝑟 + 𝑐 − 1 as
an abelian group (written additively).
(d) Conclude that 𝑅 × has rank 𝑟 + 𝑐 − 1 as an abelian group (written multi-
plicatively).
Chapter 39
Volume formula
In this chapter, we exhibit a formula for the covolume of a quaternionic group, a

formula with many applications.
39.1 ⊲ Statement
We saw in (35.1.5) that the hyperbolic area of the quotient SL2 (Z)\H2 can be computed
directly from the fundamental domain
◊ = {𝑧 ∈ H2 : |Re 𝑧| ≤ 1/2 and | 𝑧| ≥ 1}
as ∫
d𝑥 d𝑦 𝜋
area(SL2 (Z)\H2 ) = area(◊) = = . (39.1.1)
◊ 𝑦2 3
In fact, given a Fuchsian or Kleinian group Γ, the hyperbolic area or volume of the
quotient Γ\H (where H = H2 , H3 , respectively) can be computed without recourse
to a fundamental domain: it is given in terms of the arithmetic invariants of the order
and quaternion algebra that give rise to Γ.
To begin, we consider the already interesting case where the quaternion algebra is
defined over Q.
Theorem 39.1.2. Let 𝐵 be a quaternion algebra over Q of discriminant 𝐷 and let

O ⊆ 𝐵 be a maximal order. Let Γ10 (𝐷) ≤ PSL2 (R) be the Fuchsian group associated
to the group PO1 = O1 /{±1} of units of reduced norm 1.
Then
𝜋
area(Γ10 (𝐷)\H2 ) = 𝜑(𝐷) (39.1.3)
3
where
Ö Ö 1

𝜑(𝐷) := ( 𝑝 − 1) = 𝐷 1− (39.1.4)
𝑝
𝑝 |𝐷 𝑝 |𝐷
Recall that the discriminant 𝐷 ∈ Z is a squarefree positive integer, so the function

𝜑 is just the Euler totient function. Theorem 39.1.2 recovers (39.1.1) with 𝐷 = 1.
717
718 CHAPTER 39. VOLUME FORMULA
Example 39.1.5. Recall the case 𝑋01 (6) from section 38.1. We confirm that the
hyperbolic area computed from the fundamental domain agrees with the formula in
Theorem 39.1.2:
𝜋 2𝜋
area(𝑋01 (6, 1)) = 𝜑(6)𝜓(1) = .
3 3
The expression (39.1.3) is quite similar to the Eichler mass formula (Theorem
25.3.15). Indeed, the method of proof is the same, involving the zeta function of the
order O. For the case of a definite quaternion algebra, when the unit group was finite,
we used Theorem 26.2.12 to relate the mass to a residue of the zeta function (see
in particular Proposition 26.5.10); for an indefinite quaternion algebra, to carry this
out in general would involve a multivariable integral whose evaluation is similar to
the proof of Proposition 26.2.18 (the commutative case)—not an appealing prospect.
That being said, this type of direct argument with the zeta function was carried out by
Shimizu [Shz65, Appendix] over a totally real field 𝐹.
We prefer instead to use idelic methods; we already computed the normalized
Tamagawa measure 𝜏 1 (𝐵1 \𝐵1 ) = 1 (see Theorem 29.11.3), and this number has
everything we need! It is a much simpler computation to relate this to the volume of
the hyperbolic quotient: we carry this out in section 39.3.
The following notation will be used throughout.
39.1.6. Let 𝐹 be number field of degree 𝑛 = [𝐹 : Q] with 𝑟 real places and 𝑐 complex
places, so 𝑟 + 2𝑐 = 𝑛. Let 𝐵 be a quaternion algebra over 𝐹 of discriminant 𝔇 that is
split at 𝑡 real places. Suppose that 𝐵 is indefinite (i.e., 𝑡 + 𝑐 > 0).
Let 𝑅 = Z𝐹 be the ring of integers of 𝐹. Let O ⊆ 𝐵 be an 𝑅-order of reduced
discriminant 𝔑 that is locally norm-maximal. Let
H := (H2 ) 𝑡 × (H3 ) 𝑐
and let
Γ1 (O) ≤ PSL2 (R) 𝑡 × PSL2 (C) 𝑐 H
be the discrete group associated to the group PO1 = O1 /{±1} of units of O of reduced
norm 1.
O
For a prime 𝔭 | 𝔑 with Nm(𝔭) = 𝑞, let ∈ {−1, 0, 1} be the Eichler symbol
𝔭
(Definition 24.3.2), and let
 1 + 1/𝑞, if (O | 𝑝) = 1;
1 − Nm(𝔭) −2




𝜆(O, 𝔭) := = 1 − 1/𝑞, if (O | 𝑝) = −1; (39.1.7)
O
Nm(𝔭) −1 

1−  1 − 1/𝑞 2 , if (O | 𝑝) = 0.
𝔭 
Main Theorem 39.1.8. With notation as in 39.1.6, we have
2(4𝜋) 𝑡 Ö
vol(Γ1 (O)\H ) = 2 2
𝜁 𝐹 (2)𝑑 𝐹3/2 Nm(𝔑) 𝜆(O, 𝔭). (39.1.9)
(4𝜋 ) (8𝜋 )
𝑟 𝑐
𝔭 |𝔑
39.1. ⊲ STATEMENT 719
Main Theorem 39.1.8 is proven as Main Theorem 39.3.1. Since 𝜁 𝐹 (2) ≈ 1, for
algebras 𝐵 with fixed signature 𝑟, 𝑠, 𝑡, we can roughly estimate
vol(Γ1 (O)) ≈ 𝑑 𝐹3/2 Nm(discrd O). (39.1.10)
39.1.11. The constant factor in (39.1.9) is written to help in remembering the formula;
multiplying out, we have
2(4𝜋) 𝑡 1
2 2
= 2𝑟 +3𝑐−2𝑡−1 2𝑟 +2𝑐−𝑡 .
(4𝜋 ) (8𝜋 )
𝑟 𝑐 2 𝜋
Note that the right-hand side of (39.1.9) is independent of the choice of order O in
the genus of O, as it depends only on O.
b
Remark 39.1.12. If O is not locally norm-maximal, then one corrects the formula
by inserting factors [𝑅𝔭× : nrd(O×𝔭 )] as in (39.2.10); since 𝑅𝔭× ⊆ O×𝔭 we have 𝑅𝔭×2 ⊆
nrd(O×𝔭 ), so these factors are at most 2 for each 𝔭.
An important special case of Main Theorem 39.1.8 is the case where O is an
Eichler order, generalizing Theorem 39.1.2.
Theorem 39.1.13. Suppose that O = O0 (𝔐) is an Eichler order of level 𝔐 and
disc 𝐵 = 𝔇, so 𝔑 = 𝔇𝔐. Write Γ10 (𝔐) = Γ1 (O) and Γ1 (1) for the group associated
to a maximal order O(1) ⊇ O0 (𝔐).
Then
2(4𝜋) 𝑡
vol(Γ10 (𝔐)\H ) = 𝜁 𝐹 (2)𝑑 𝐹3/2 𝜑(𝔇)𝜓(𝔐)
(4𝜋 2 ) 𝑟 (8𝜋 2 ) 𝑐
where
Ö 1

×
𝜑(𝔇) = #(Z𝐹 /𝔇) = Nm 𝔇 1−
Nm 𝔭
𝔭 |𝔇
Ö (39.1.14)
1 1 1
𝜓(𝔐) = [Γ (1) : Γ0 (𝔐)] = Nm 𝔐 1+ .
Nm 𝔭
𝔭 k𝔐
𝑒
We write 𝔭𝑒 k 𝔐 to mean 𝔭𝑒 exactly divides 𝔐, i.e., ord𝔭 (𝔐) = 𝑒, and we take

the product over all prime power divisors of 𝔐 in the definition of 𝜓.
In particular, when O is a maximal order, then 𝔐 = 𝑅 and 𝜓(𝔐) = 1.
Remark 39.1.15. Theorem 39.1.13 is often attributed to Borel [Bor81], who derived
it under the (highly nontrivial!) assumption that 𝜏 1 (𝐵1 \𝐵1 ) = 1.
Example 39.1.16. Let Γ1 = PSL2 (Z[𝑖]). Then
2 3/2 2
vol(𝑋 1 ) = 4 𝜁 𝐹 (2) = 2 𝜁 𝐹 (2) = 0.3053 . . . .
8𝜋 2 𝜋
This agrees with the computation we did with a fundamental domain (see 36.6.8),
since for 𝜒 the character 36.6.14 of conductor 4,
2 2 1
2
𝜁 𝐹 (2) = 2 𝜁 (2)𝐿(2, 𝜒) = 𝐿 (2, 𝜒).
𝜋 𝜋 3
39.2 Volume setup

In this section, we setup a few calculations in preparation for the volume formula.
Let 𝐹 be number field with discriminant 𝑑 𝐹 , let 𝐵 be a quaternion algebra over 𝐹
of discriminant 𝔇 = disc 𝐵, and let O ⊆ 𝐵 be an order. The key input to the proof is
the following ingredient: from Theorem 29.11.3,
𝜏 1 (𝐵1 \𝐵1 ) = 1. (39.2.1)
We will convert the volume (39.2.1) into the desired form by separating out the
contribution from the finite places and the infinite (real) ramified places: what remains
is the volume of the orbifold we seek, which we then renormalize from the adelic to
the standard hyperbolic volume.
We define
O := O b × 𝐵∞ ⊆ 𝐵.
Then
O1 = {𝛾 ∈ O× : nrd(𝛾) = 1} = O× ∩ 𝐵1
and
O1 = O
b1 × 𝐵1 .
∞
Now we apply the important assumption: we suppose that 𝐵 is indefinite. The

hypothesis that 𝐵 is indefinite is necessary even to get a nontrivial space H = (H2 ) 𝑡 ×
(H3 ) 𝑐 for the group to act upon! The case where 𝐵 is definite was handled in the proof
of the Eichler mass formula (Main Theorem 25.3.19).
39.2.2. By strong approximation (Corollary 28.5.12) we have 𝐵1 = 𝐵1 O1 . Therefore,

the natural inclusion
O1 \O1 ↩→ 𝐵1 \𝐵1
is also surjective, hence an isomorphism. Thus by (39.2.1)
𝜏 1 (O1 \O1 ) = 𝜏 1 (𝐵1 \𝐵1 ) = 1. (39.2.3)
We have an embedding O1 ↩→ 𝐵∞
1 , so
1 = 𝜏 1 (O1 \O1 ) = b
𝜏1 (O
b1 )𝜏 1 (O1 \𝐵1 ).
∞ ∞ (39.2.4)
39.2.5. If O is maximal, then from (29.7.25) we have

Ö Ö
𝜏1 (O
b b1 ) = 𝜏𝔭1 (O1𝔭 ) = | 𝑑 𝐹 | −3/2 𝜁 𝐹 (2) −1 (Nm 𝔭 − 1) −1
𝔭 𝔭∈Ram(𝐵)
so that
b b1 ) −1 = | 𝑑 𝐹 | 3/2 𝜁 𝐹 (2)𝜑(𝔇).
𝜏1 (O (39.2.6)
So by (39.2.6), we conclude that
1
𝜏∞ (O1 \𝐵∞
1
) = 𝑑 𝐹3/2 𝜁 𝐹 (2)𝜑(𝔇). (39.2.7)
39.2. VOLUME SETUP 721
39.2.8. In general, if O ⊆ O0 with O0 maximal, then
𝜏1 (O
b c01 ) = [O
c01 : O
b1 ]b
𝜏1 (O
b1 )
so similarly
1
𝜏∞ (O1 \𝐵∞
1
) = 𝑑 𝐹3/2 𝜁 𝐹 (2)𝜑(𝔇) [ O
c01 : O
b1 ]. (39.2.9)
If O is locally norm-maximal, then further
c01 : O
[O c0× : O
b1 ] = [O b× ] (39.2.10)
and by Lemma 26.6.7, with 𝔑 = discrd O we have

Ö Ö
c0× : O
𝜑(𝔇) [ O b× ] = [O𝔭0 : O𝔭 ]𝜆(O, 𝔭) = Nm(𝔑) 𝜆(O, 𝔭) (39.2.11)
𝔭 |𝔑 𝔭 |𝔑
(as in (26.6.8)).
Next, we relate the measures on SL2 (R) and SL2 (C) to the corresponding measures
on H2 and H3 .
39.2.12. Recall the symmetric space identification
H2 → SL2 (R)/SO(2) (39.2.13)
The hyperbolic plane H2 is equipped with the hyperbolic measure 𝜇, the unique
measure invariant under the action of PSL2 (R); the group SL2 (R) has the Haar measure
𝜏 1 , also invariant under the left action of SL2 (R). Therefore, the identification (39.2.13)
relates these two measures up to a constant 𝑣(SO(2)) that gives a total measure to SO(2)
(normalizing its Haar measure). Ditto for H3 and SU(2), with a constant 𝑣(SU(2)).
𝑣(SO(2)) = 𝜋 and 𝑣(SU(2)) = 8𝜋 2 .
Proof. One could compute the relevant constant by doing a (compatible) integral, but
we prefer just to refer to an example where both sides are computed. We consider a
fundamental domain for the action of SL2 (Z) SL2 (R) that is invariant under SO(2):
for example, we can lift a fundamental domain for PSL2 (Z) H2 under (39.2.13).
The difference between SL2 (Z) and PSL2 (Z) = SL2 (Z)/{±1} is annoyingly relevant
here! We have
PSL2 (Z)\H2 = PSL2 (Z)\ SL2 (R)/SO(2);
if we let SO(2)2 ' SO(2)/{±1} be the rotation group acting by 2𝜃 instead of 𝜃, then
we can lift from PSL2 (Z) to SL2 (Z) and
PSL2 (Z)\ SL2 (R)/SO(2) = SL2 (Z)\ SL2 (R)/SO(2)2
which under compatible metrics gives
𝜏 1 (SL2 (Z)\ SL2 (R))

𝑣(SO(2)) = 2𝑣(SO(2)2 ) = 2 . (39.2.15)
𝜇(PSL2 (Z)\H2 )
On the bottom we have 𝜇(SL2 (Z)\H2 ) = 𝜋/3 by (35.1.5) and on the top we have
𝜏 1 (SL2 (Z)\ SL2 (R)) = 𝜁 (2) = 𝜋 2 /6 by (39.2.9). Plugging in, we compute
𝜋 2 /6
𝑣(SO(2)) = 2 = 𝜋. (39.2.16)
𝜋/3
Similarly,
𝜏 1 (SL2 (Z[𝑖])\ SL2 (C)) 43/2 𝜁Q(𝑖) (2)
𝑣(SU(2)) = 2 =2 = 8𝜋 2 (39.2.17)
𝜇(PSL2 (Z[𝑖])\H3 ) (2/𝜋 2 )𝜁Q(𝑖) (2)
by Example 39.1.16 and again (39.2.9).
39.3 Volume derivation

We now establish the volume formula (Main Theorem 39.1.8) using the computation
of the Tamagawa measure (Theorem 29.11.3, (39.2.1)), following Borel [Bor81, 7.3].
We continue with the notation from 39.1.6, so in particular 𝑛 = [𝐹 : Q] and 𝐹
has 𝑟 real places, 𝑐 complex places, and 𝐵 is split at 𝑡 real places; 𝐵 is indefinite, so
𝑡 + 𝑐 > 0; and O ⊆ 𝐵 is a locally norm-maximal order. We restate the formula for
convenience.
Main Theorem 39.3.1. We have
2(4𝜋) 𝑡 3/2
Ö
vol(Γ1 (O)\H ) = 𝜁 𝐹 (2)𝑑 Nm(𝔑) 𝜆(O, 𝔭). (39.3.2)
(4𝜋 2 ) 𝑟 (8𝜋 2 ) 𝑐 𝐹
𝔭 |𝔑
Proof. To summarize the previous section, we started with 𝜏 1 (𝐵1 \𝐵1 ) and concluded
𝜏 1 (O1 \O1 ) = 1 by strong approximation; we factored this into finite and infinite parts,
with the finite part computed in terms of the order and the infinite part. Then
Ö Ö Ö
1 1 1 1 1 1
O \𝐵∞ = O \ 𝐵𝑣 ' 𝐵𝑣 × O \ 𝐵1𝑣 . (39.3.3)
𝑣 |∞ 𝑣 ∈Ω 𝑣 |∞
𝑣∉Ω
Each term contributes to the volume. For the first product, for each of the 𝑟 − 𝑡 places
𝑣 ∈ Ω we have 𝐵1𝑣 ' H1 and we computed in Lemma 29.5.9 that 𝜏 1 (H1 ) = 4𝜋 2 . For
the remaining terms, we employ the comparison formula between measures (Lemma
39.2.14), and are plagued by the same factor 2 coming from the fact that Γ1 (O) arises
from PO1 /{±1}. Putting these together, the decomposition (39.3.3) yields a volume
1
1
𝜏∞ (O1 \𝐵∞
1
) = (4𝜋 2 ) 𝑟 −𝑡 𝜋 𝑡 (8𝜋 2 ) 𝑐 vol(Γ1 \H )
2
(39.3.4)
(4𝜋 2 ) 𝑟 (8𝜋 2 ) 𝑐
= vol(Γ1 \H ).
2(4𝜋) 𝑡
From (39.2.9) and (39.3.4) we conclude
2(4𝜋) 𝑡
vol(Γ1 \H ) = 𝜇(O1 \𝐵∞ 1
)
(4𝜋 2 ) 𝑟 (8𝜋 2 ) 𝑐
(39.3.5)
2(4𝜋) 𝑡
= 𝑑 3/2 𝜁 𝐹 (2)𝜑(𝔇) [ O
c01 : O
b1 ].
(4𝜋 2 ) 𝑟 (8𝜋 2 ) 𝑐 𝐹
39.4. GENUS FORMULA 723
Finally, the computation (39.2.11) of the local index completes the proof.
Remark 39.3.6. A similar proof works for the case where 𝐹 is a function field or
where Γ is an S-arithmetic group, but in both cases still under the hypothesis that 𝐵 is
S-indefinite for an eligible set S (playing the role of the archimedean places above).
Example 39.3.7. Suppose 𝐹 is totally real, and we take 𝐵 = M2 (𝐹) and O = M2 (Z𝐹 ).
Then H = (H2 ) 𝑛 and
2𝜁 𝐹 (2) (4𝜋) 𝑛 3/2 2𝜁 𝐹 (2) 3/2

vol(Γ1 (O)\H ) = 𝑑𝐹 = 𝑑𝐹 .
(4𝜋 2 ) 𝑛 𝜋𝑛
39.4 Genus formula

In this section, we take the volume formula (Main Theorem 39.3.1) in the special case
of a Fuchsian group and extend it to a formula for the genus of a Shimura curve.
We maintain our notation but now specialize to the case where 𝑐 = 0 and 𝑡 = 1: in
particular, 𝐹 is a totally real field, and 𝐵 is indefinite. Thus Γ = Γ1 (O) ≤ PSL2 (R) is
a Fuchsian group.
We suppose that O is an Eichler order of level 𝔐.
39.4.1. Recalling 37.7, let (𝑔; 𝑒 1 , . . . , 𝑒 𝑘 ; 𝛿) be the signature of Γ. Then 𝑌 (Γ) has
genus 𝑔; has 𝑘 elliptic cycles of orders 𝑒 1 , . . . , 𝑒 𝑘 ∈ Z ≥2 , corresponding to cone points
on 𝑌 (Γ) with given order; and has 𝛿 parabolic cycles, corresponding to the punctures
of 𝑌 (Γ). We have 𝛿 = 0 unless 𝐵 = 𝑀2 (Q), corresponding to the case of classical
modular curves.
By Proposition 37.7.4,
𝑘 !
∑︁ 1
𝜇(𝑌 (Γ)) = 2𝜋 (2𝑔 − 2) + 1− +𝛿 .
𝑖=1
𝑒𝑖
Rewriting this slightly, for 𝑞 ∈ Z ≥2 , let 𝑚 𝑞 be the number of elliptic cycles of order 𝑞.
Then
𝜇(𝑌 (Γ)) ∑︁ 1
= 2𝑔 − 2 + 𝑚𝑞 1 − +𝛿 (39.4.2)
2𝜋 𝑞 ≥2
𝑞
and the sum is finite.

The numbers 𝑚 𝑞 are determined by embedding numbers of quadratic orders into
the quaternion order O, as studied in chapter 30.
39.4.3. Let 𝑞 ∈ Z ≥2 . Suppose that 𝑚 𝑞 > 0, so that O1 has a maximal finite subgroup
h𝛾i ≤ O1 of order 2𝑞. Then the field 𝐾𝑞 = 𝐹 (𝜁2𝑞 ) ⊃ 𝐹 is a quadratic field extension
and 𝐾𝑞 ↩→ 𝐵 embeds, where 𝜁2𝑞 is a primitive 2𝑞th root of unity, and we have two
optimal embeddings 𝑆 = 𝐹 (𝛾) ∩ O ↩→ O given by 𝛾 and 𝛾. Conversely, to every
embedding 𝜙 : 𝐾𝑞 ↩→ 𝐵, we associate the order 𝑆 = 𝜙(𝐾𝑞 ) ∩ O and the finite subgroup
× ⊂ O1 .
𝑆tors
Thus there is a two-to-one map
× = 2𝑞}
{O1 -conjugacy classes of optimal embeddings 𝜙 : 𝑆 ↩→ O with 𝑆tors
↓
{Elliptic cycles of Γ of order 2𝑞}.
In the notation of 30.3.10, we have shown that
1 ∑︁
𝑚𝑞 = 𝑚(𝑆, O; O1 ). (39.4.4)
2
𝐾𝑞 ⊃𝑆 ⊇𝑅 [𝜁2𝑞 ]
× =2𝑞
#𝑆tors
Our next major ingredient is the theory of selectivity, treated in chapter 31.
39.4.5. We claim that 𝐾𝑞 does not satisfy the selectivity condition (OS), defined in
31.1.6. If O is an Eichler order, we may appeal to Proposition (31.2.1) and condition
(a): since 𝐹 is totally real, 𝐾𝑞 is totally complex, and 𝐵 is split at a real place, condition
(a) fails.
Therefore by Main Theorem 31.1.7(a), Gen O is not optimally selective. By
Corollary 31.1.10, for every 𝑅-order 𝑆 ⊆ 𝐾𝑞 , we have
ℎ(𝑆)
𝑚(𝑆, O; O× ) = b O;
𝑚( 𝑆, b× )
b O (39.4.6)
# ClΩ 𝑅
where we have substituted # Cls O = # ClΩ 𝑅 (by Corollary 28.5.17).
The adelic embedding numbers 𝑚( 𝑆, b O;
b Ob× ), a product of (finitely many) local
embedding numbers by 30.7.1, are computed in section 30.6.
We will need one lemma relating units to class numbers.
Lemma 39.4.7. We have [𝑅>× Ω 0 : 𝑅 ×2 ] = 2[ClΩ 𝑅 : Cl 𝑅].
Proof. The index [𝑅>× Ω 0 : 𝑅 ×2 ], which does not depend on 𝑆, is related to class
numbers as follows. For each real place 𝑣, define sgn 𝑣 : 𝐹 × → {±1} by the real sign
sgn 𝑣 (𝑎) = sgn(𝑣(𝑎)) at 𝑣. Let
sgnΩ : 𝐹 × → {±1}Ω
𝑎 ↦→ (sgn 𝑣 (𝑎)) 𝑣
collect the signs at the places 𝑣 ∈ Ω. Then we have an exact sequence
1 → {±1}Ω /sgnΩ 𝑅 × → ClΩ 𝑅 → Cl 𝑅 → 1 (39.4.8)
where the map on the left is induced by mapping a tuple of signs in {±1}Ω to the
principal ideal generated by any 𝑎 ∈ 𝐹 × with the given signs. We have a second
(tautological) exact sequence
sgnΩ
1 → 𝑅>× Ω 0 /𝑅 ×2 → 𝑅 × /𝑅 ×2 −−−→ {±1}Ω → {±1}Ω /sgnΩ 𝑅 × → 1 (39.4.9)
of elementary abelian 2-groups (or F2 -vector spaces). Combining (39.4.9) with
(39.4.8), and noting that [𝑅 × : 𝑅 ×2 ] = 2𝑟 by Dirichlet’s unit theorem and #Ω = 𝑟 − 1
by hypothesis, we conclude that
[𝑅>× Ω 0 : 𝑅 ×2 ] = 2𝑟 −#Ω [ClΩ 𝑅 : Cl 𝑅] = 2[ClΩ 𝑅 : Cl 𝑅].
Definition 39.4.10. For a quadratic 𝑅-order 𝑆 ⊆ 𝐾, the Hasse unit index is defined
by
𝑄(𝑆) := [Nm𝐾 |𝐹 (𝑆 × ) : 𝑅 ×2 ].
We have 𝑄(𝑆) < ∞ because 𝑆 × and 𝑅 × have the same Z-rank.

Remark 39.4.11. Hasse [Hass52, Sätze 14–29] proved numerous theorems about
𝑄(Z𝐾 ), including that 𝑄(Z𝐾 ) ≤ 2: see also Washington [Was97, Theorem 4.12].
We are now ready to write in a simplified way the count 𝑚 𝑞 of elliptic cycles.
1 ∑︁ ℎ(𝑆) b O; b× )
b O
𝑚𝑞 = 𝑚( 𝑆,
ℎ(𝑅) 𝑆 ⊂𝐾𝑞
𝑄(𝑆)
× =2𝑞
#𝑆tors
where ℎ(𝑅) = # Pic 𝑅 and ℎ(𝑆) = # Pic 𝑆.
Proof. Beginning with (39.4.4), we have
1 ∑︁
𝑚𝑞 = 𝑚(𝑆, O; O1 ). (39.4.13)
2 𝑆 ⊂𝐾𝑞
× =2𝑞
#𝑆tors
𝑚(𝑆, O; O1 ) = 𝑚(𝑆, O; O× ) [nrd(O× ) : nrd(𝑆 × )]. (39.4.14)
Since 𝐵 is indefinite, by Corollary 31.1.11 we have nrd(O× ) = 𝑅>× Ω 0 . By Lemma

39.4.7, we have [𝑅>× Ω 0 : 𝑅 ×2 ] = 2[ClΩ 𝑅 : Cl 𝑅]. Thus
[nrd(O× ) : nrd(𝑆 × )] = [𝑅>× Ω 0 : 𝑅 ×2 ] [𝑅 ×2 : nrd(𝑆 × )]

2[ClΩ 𝑅 : Cl 𝑅] (39.4.15)
= .
𝑄(𝑆)
Substituting (39.4.6) and (39.4.15) into (39.4.14), we find
𝑚(𝑆, O; O1 ) = 𝑚(𝑆, O; O× ) [nrd(O× ) : nrd(𝑆 × )]

ℎ(𝑆) [𝑅>× Ω 0 : 𝑅 ×2 ]
×
= 𝑚( 𝑆, O; O )
b b b
(39.4.16)
𝑄(𝑆) # ClΩ 𝑅
2 ℎ(𝑆) b O; b× ).
b O
= 𝑚( 𝑆,
ℎ(𝑅) 𝑄(𝑆)
Finally, plugging (39.4.16) into (39.4.13) and cancelling a factor 2 gives the result.
Corollary 39.4.17. The signature of a Shimura curve depends only on the discriminant
𝔇 and level 𝔐.
Proof. The ambiguity corresponds to a choice of Eichler order of level 𝔐 and choice
of split real place; when 𝐹 = Q, there is no ambiguity in either case. So we may
suppose that Γ has no parabolic cycles. Then we simply observe that the formula
(Proposition 39.4.12) for the number of elliptic cycles depends only on O.
b
Example 39.4.18. As a special case of Proposition 39.4.12, suppose that 𝔑 = 𝔇𝔐 is

coprime to 𝑞. Then as in Example 30.7.5, we have
Ö ∑︁
1 Ö 𝐾𝑞 𝐾𝑞 ℎ(𝑆)
𝑚𝑞 = 1− 1+ .
ℎ(𝑅) 𝔭 𝔭 𝑆 ⊂𝐾
𝑄(𝑆)
𝔭 |𝔇 𝔭 |𝔐 𝑞
× =2𝑞
#𝑆tors
We now have the ingredients to give a formula for a Shimura curve.
Theorem 39.4.19. Let 𝑌 1 (O) = Γ1 (O)\H2 . Then 𝑌 1 (O) is an orbifold with genus 𝑔
where
4 3/2
∑︁ 1
2𝑔 − 2 = 𝜁 𝐹 (2)𝑑 𝐹 𝜑(𝔇)𝜓(𝔐) − 𝑚𝑞 1 − −𝛿
(4𝜋 2 ) 𝑟 𝑞 ≥2
𝑞
where 𝑚 𝑞 are given in Proposition (39.4.12).
Proof. Combine the volume formula (Main Theorem 39.3.1) with (39.4.2).
The special case where 𝐹 = Q is itself important.
Theorem 39.4.20. Let 𝐷 = disc 𝐵 > 1 and let O ⊆ 𝐵 be an Eichler order of level 𝑀,
so 𝑁 = 𝐷 𝑀 = discrd O with 𝐷 squarefree and gcd(𝐷, 𝑀) = 1.
Then 𝑋 1 (O) = Γ1 (O)\H2 is an orbifold with genus 𝑔 where
𝜑(𝐷)𝜓(𝑀) 𝑚 2 2𝑚 3
2𝑔 − 2 = − −
6 2 3
where the embedding numbers were computed in Example 30.7.7:
 Ö Ö
−4

−4
1− 1+ , if 4 - 𝑀;


×


𝑚 2 = 𝑚(Z[𝑖], O; O ) = 𝑝 |𝐷 𝑝
𝑝 |𝑀
𝑝

 0, if 4 | 𝑀.


Ö
 Ö −3 −3
1− 1+ , if 9 - 𝑀;


×


𝑚 3 = 𝑚(Z[𝜔], O; O ) = 𝑝 |𝐷 𝑝
𝑝 |𝑀
𝑝

 0, if 9 | 𝑀;


Example 39.4.21. Suppose 𝐷 = 6 and 𝑀 = 1, so we are in the setting of Then
𝑚 2 = 𝑚 3 = 2 and
𝜙(6) 4 1
2𝑔 − 2 = − 1 − = = −2
6 3 3
so 𝑔 = 0; this confirms that 𝑋 1 (O) has signature (0; 2, 2, 3, 3) as in 37.9.10.
Exercises
1. Let 𝐹 be the function field of a curve 𝑋 over F𝑞 of genus 𝑔. Let 𝐵 be a quaternion
algebra over 𝐹.
(a) Let 𝑣 ∈ Pl 𝐹 be place that is split in 𝐵. Let S = {𝑣}, let 𝑅 = 𝑅 (S) , and
let O ⊆ 𝐵 be an 𝑅-order. Let T be the Bruhat–Tits tree associated to
𝐵 𝑣 ' M2 (𝐹𝑣 ). Via the embedding 𝜄 : 𝐵 ↩→ 𝐵 𝑣 , show that the group
Γ1 (O) = P𝜄(O1 ) T acts on T by left multiplication as a discrete group
acting properly.
(b) Continuing as in (a), compute the measure of Γ1 (O)\T using the methods
of section 39.3.
2. Generalizing the previous exercise, let 𝐹 be a global field, let 𝐵 be a quaternion
algebra over 𝐹, let S be an eligible set and suppose that 𝐵 is S-indefinite. Let
𝑅 = 𝑅 (S) and let O ⊆ 𝐵 be an 𝑅-order. Define a symmetric space H on
which PO1 acts as a discrete group acting properly, and compute the measure of
Γ1 (O)\H .
Part V
Arithmetic geometry
729
Chapter 40
Classical modular forms
In this chapter, we introduce modular forms on the classical modular group. This
chapter will provide motivation as well as important examples for generalizations in
this last part of the text.
40.1 ⊲ Functions on lattices

In this section, we pursue the interpretation of the quotient Γ\H2 with Γ = PSL2 (Z) as
a moduli space of lattices, and we study functions on the quotient. There are a wealth
of references for the classical modular forms, including Apostol [Apo90, Chapters
1–2], Diamond–Shurman [DS2005], Miyake [Miy2006, Chapter 4], Lang [Lang95,
§1], and Serre [Ser73, Chapter VII]. For this section, see Silverman [Sil2009, Chapter
VI] for the complex analytic theory of elliptic curves and the relationship to Eisenstein
series.
Recall from 35.3.3 that 𝑌 = Γ\H2 parametrizes complex lattices up to homothety,
i.e., there is a bĳection
𝑌 = Γ\H2 → {Λ ⊂ C lattice}/∼
(40.1.1)
Γ𝜏 ↦→ [Z + Z𝜏].
In particular, the set of homothety classes has a natural structure of a Riemann surface,
and we seek now to make this explicit. We show that there are natural, holomorphic
functions on the set of lattices that allow us to go beyond the bĳection 40.1.1 to realize
the complex structure on 𝑌 explicitly.
Let Λ ⊂ C be a lattice. To write down complex moduli, we average over Λ in a
convergent way, as follows.
Definition 40.1.2. The Eisenstein series of weight 𝑘 ∈ Z>2 for Λ is
∑︁ 1
𝐺 𝑘 (Λ) = .
𝜆∈Λ
𝜆𝑘
𝜆≠0
If 𝑘 is odd, then 𝐺 𝑘 (Λ) = 0, so let 𝑘 ∈ 2Z ≥2 .
731
732 CHAPTER 40. CLASSICAL MODULAR FORMS
Lemma 40.1.3. The series 𝐺 𝑘 (Λ) converges absolutely.
Proof. Up to homothety (which does not affect convergence), we may suppose Λ =

Z + Z𝜏, with 𝜏 ∈ H2 . Then we consider the corresponding absolute sum
∑︁ 1 ∑︁ 1
= . (40.1.4)
𝜆∈Z+Z𝜏
|𝜆| 𝑘 𝑚,𝑛∈Z
|𝑚 + 𝑛𝜏| 𝑘
𝜆≠0 (𝑚,𝑛)≠(0,0)
The number of pairs (𝑚, 𝑛) with 𝑟 ≤ |𝑚𝜏 + 𝑛| < 𝑟 + 1 is the number of lattice points in
an annulus of area 𝜋(𝑟 + 1) 2 − 𝜋𝑟 2 = 𝑂 (𝑟), so there are 𝑂 (𝑟) such points; and thus the
series (40.1.4) is majorized by (a constant multiple of) 𝑟∞=1 𝑟 1−𝑘 , which is convergent
Í
for 𝑘 > 2.
40.1.5. For 𝑧 ∈ H2 and 𝑘 ∈ 2Z ≥2 , define

∑︁ 1
𝐺 𝑘 (𝑧) = 𝐺 𝑘 (Z + Z𝑧) = . (40.1.6)
𝑚,𝑛∈Z
(𝑚 + 𝑛𝑧) 𝑘
(𝑚,𝑛)≠(0,0)
Lemma 40.1.7. 𝐺 𝑘 (𝑧) is holomorphic for 𝑧 ∈ H2 , and
𝐺 𝑘 (𝛾𝑧) = (𝑐𝑧 + 𝑑) 𝑘 𝐺 𝑘 (𝑧)

𝑎 𝑏
for all 𝛾 = ∈ PSL2 (Z).
𝑐 𝑑
Proof. This is true for 𝑧 ∈ ◊ since then
|𝑚 + 𝑛𝑧| 2 = 𝑚 2 + 2𝑚𝑛 Re 𝑧 + 𝑛2 |𝑧| 2 ≥ 𝑚 2 − 𝑚𝑛 + 𝑛2 = |𝑚 + 𝑛𝜔| 2
thus |𝐺 𝑘 (𝑧)| ≤ |𝐺 𝑘 (𝜔)| and so by the Weierstrass 𝑀-test, 𝐺 𝑘 (𝑧) is holomorphic for
𝑧 ∈ ◊: by Morera’s
theorem, uniform convergence implies holomorphicity. But now
𝑎 𝑏
for all 𝛾 = ∈ Γ, we claim that
𝑐 𝑑
𝐺 𝑘 (𝛾𝑧) = (𝑐𝑧 + 𝑑) 𝑘 𝐺 𝑘 (𝑧) (40.1.8)
(and note this does not depend on the choice of sign): indeed,
1 𝑐𝑧 + 𝑑
= (40.1.9)
𝑚 + 𝑛(𝛾𝑧) (𝑏𝑛 + 𝑑𝑚) + (𝑎𝑛 + 𝑐𝑚)𝑧
and the map

𝑎 𝑏
(𝑛, 𝑚) ↦→ (𝑛, 𝑚) = (𝑎𝑛 + 𝑐𝑚, 𝑏𝑛 + 𝑑𝑚)
𝑐 𝑑
40.1. ⊲ FUNCTIONS ON LATTICES 733
is a permutation of Z2 − {(0, 0)}, so by absolute convergence we may rearrange the

sum to get
∑︁ 1
𝐺 𝑘 (𝛾𝑧) = (𝑐𝑧 + 𝑑) 𝑘
𝑚,𝑛∈Z
((𝑏𝑛 + 𝑑𝑚) + (𝑎𝑛 + 𝑐𝑚)𝑧) 𝑘
(𝑚,𝑛)≠(0,0)
(40.1.10)
∑︁ 1
= (𝑐𝑧 + 𝑑) 𝑘 = (𝑐𝑧 + 𝑑) 𝑘 𝐺 𝑘 (𝑧).
𝑚,𝑛∈Z
(𝑚,𝑛)≠(0,0)
By transport, since Γ◊ = H2 , we see that 𝐺 𝑘 (𝑧) is holomorphic on all of H2 .
40.1.11. In this (somewhat long) paragraph, we connect the theory of Eisenstein series
above to the theory of elliptic curves.
Let Λ ⊂ C be a lattice. We define the Weierstrass ℘-function (relative to Λ) by

1 ∑︁ 1 1
℘(𝑧) = ℘(𝑧; Λ) = 2 + − . (40.1.12)
𝑧 𝜆∈Λ
(𝑧 − 𝜆) 2 𝜆2
𝜆≠0
We have
1 1 |𝑧|(2|𝜆| + |𝑧|) 1
(𝑧 − 𝜆) 2 − 𝜆2 ≤ |𝜆| 2 (|𝜆| − |𝑧|) 2 = 𝑂 |𝜆| 3 (40.1.13)

so as above we see that ℘(𝑧) is absolutely convergent for all 𝑧 ∈ C r Λ and uniformly
convergent on compact subsets, and so defines a holomorphic function on CrΛ. Since
∞
1 1 ∑︁ 𝑧𝑛
− = (𝑛 + 1) (40.1.14)
(𝑧 − 𝜆) 2 𝜆2 𝑛=1 𝜆 𝑛+2
by differentiating the geometric series, we find
∞
1 ∑︁ 1
℘(𝑧) = 2
+ (𝑘 − 1)𝐺 𝑘 (Λ)𝑧 𝑘 = 2 + 3𝐺 4 (Λ)𝑧 4 + 5𝐺 6 (Λ)𝑧6 + . . . . (40.1.15)
𝑧 𝑘=3
𝑧
Differentiating with respect to 𝑧 and squaring, we find that

2
d℘ 4 24𝐺 4 (Λ)
(𝑧) = − − 80𝐺 6 (Λ) + . . . . (40.1.16)
d𝑧 𝑧6 𝑧2
Expanding out the first few terms, we find that

2
d℘
𝑓 (𝑧) = (𝑧) − 4℘(𝑧) 3 + 60𝐺 4 (Λ)℘(𝑧) + 140𝐺 6 (Λ) = 𝑂 (𝑧2 )
d𝑧
is holomorphic at 𝑧 = 0 and satisfies 𝑓 (𝑧 + 𝜆) = 𝑓 (𝑧) for all 𝜆 ∈ Λ. By periodicity,

𝑓 (𝑧) takes its maximum in a fundamental parallelogram for Λ; then by Liouville’s
theorem, 𝑓 is bounded on C so constant. Since 𝑓 (0) = 0, we conclude that 𝑓 (𝑧) is

identically zero.
Following convention, write
𝑔4 = 𝑔4 (Λ) = 60𝐺 4 (Λ) and 𝑔6 = 𝑔6 (Λ) = 140𝐺 6 (Λ)
and
d℘
𝑥(𝑧) = ℘(𝑧; Λ) and 𝑦(𝑧) = (𝑧; Λ).
d𝑧
Then the image of the map
C/Λ → P2 (C)
(40.1.17)
𝑧 ↦→ (𝑥(𝑧) : 𝑦(𝑧) : 1)
is cut out by the affine equation
𝑦 2 = 4𝑥 3 − 𝑔4 𝑥 − 𝑔6 ;
the map (40.1.17) is an isomorphism of Riemann surfaces. Looking ahead to Definition

42.1.1, this map exhibits C/Λ as an elliptic curve over C.
To produce holomorphic functions that are well-defined on the quotient Γ\H2 , we

can take ratios of Eisenstein series; soon we will exhibit a map
𝑗 : H2 → C (40.1.18)
∼ C (Theorem
obtained in this way that defines a bĳective holomorphic map Γ\H2 −→
40.3.8).
40.1.19. Eisenstein series can also be thought of as weighted averages over the (cosets
of the) group PSL2 (Z) as follows.
Let Γ∞ ≤ Γ = PSL2 (Z) be the stabilizer of ∞; then Γ∞ is the infinite cyclic
group

1 1 𝑎 𝑏
generated by 𝑇 = . We consider the cosets Γ∞ \Γ: for 𝑡 ∈ Z and 𝛾 = we
0 1 𝑐 𝑑

𝑎 + 𝑡𝑐 𝑏 + 𝑡𝑑
have 𝑇 𝑡 𝛾 = with the same bottom row. Thus the function (𝑐𝑧 + 𝑑) 2
𝑐 𝑑
is well-defined on the coset Γ∞ 𝛾. Thus we can form the sum
∑︁ 1 ∑︁ 1
𝐸 𝑘 (𝑧) = (𝑐𝑧 + 𝑑) −𝑘 = , (40.1.20)
2 𝑐,𝑑 ∈Z
(𝑐𝑧 + 𝑑) 𝑘
Γ∞ 𝛾 ∈Γ∞ \Γ
gcd(𝑐,𝑑)=1
the factor 2 coming from the choice of sign in PSL2 (Z). Since every nonzero (𝑚, 𝑛) ∈
Z2 can be written (𝑚, 𝑛) = 𝑟 (𝑐, 𝑑) with 𝑟 = gcd(𝑚, 𝑛) > 0 and gcd(𝑐, 𝑑) = 1, we find
that
𝐺 𝑘 (𝑧) = 𝜁 (𝑘)𝐸 𝑘 (𝑧).
40.2. ⊲ EISENSTEIN SERIES AS MODULAR FORMS 735
40.2 ⊲ Eisenstein series as modular forms

In the previous section, we saw that natural sums (Eisenstein series) defined functions
on H2 that transformed with respect to PSL2 (Z) with a natural invariance. In this
section, we pursue this more systematically.
Definition 40.2.1. Let 𝑘 ∈ 2Z and let Γ ≤ PSL2 (R) be a Fuchsian group. A map
𝑓 : H2 → C ∪ {∞} is weight 𝑘-invariant under Γ if

𝑎 𝑏
𝑓 (𝛾𝑧) = (𝑐𝑧 + 𝑑) 𝑘 𝑓 (𝑧) for all 𝛾 = ∈ Γ. (40.2.2)
𝑐 𝑑
40.2.3. If 𝑓 is weight 𝑘 invariant and 𝑓 0 is weight 𝑘 0 invariant, then 𝑓 𝑓 0 is weight
𝑘 + 𝑘 0 invariant, and if 𝑘 0 = 𝑘 then 𝑓 + 𝑓 0 is weight 𝑘 invariant. Therefore, the set of
weight 𝑘-invariant functions has the structure of a C-vector space.
40.2.4. Weight 𝑘 invariance under Γ can be checked on a set of generators for Γ lifted
𝑎 𝑏
to SL2 (Z), as follows. For 𝛾 = ∈ SL2 (R), define 𝚥 (𝛾; 𝑧) = 𝑐𝑧 + 𝑑. Then
𝑐 𝑑
(40.2.2) can be rewritten 𝑓 (𝛾𝑧) = 𝚥 (𝛾; 𝑧) 𝑓 (𝑧).
𝑘
For 𝛾 0 ∈ Γ, we compute that 𝚥 satisfies the cocycle relation

𝚥 (𝛾𝛾 0; 𝑧) = 𝚥 (𝛾; 𝛾 0 𝑧) 𝚥 (𝛾 0; 𝑧) (40.2.5)
𝑎0 0

𝑏
because if 𝛾 0 = , then
𝑐0 𝑑0
0
𝑎 𝑧 + 𝑏0

0 0 0 0
(𝑎 𝑐 + 𝑐 𝑑)𝑧 + 𝑏(𝑏 𝑐 + 𝑑𝑑 ) = 𝑐 0 + 𝑑 (𝑐 0 𝑧 + 𝑑 0). (40.2.6)
𝑐 𝑧 + 𝑑0
Therefore, if 𝑓 is a map with 𝑓 (𝛾𝑧) = 𝚥 (𝛾; 𝑧) 𝑘 𝑓 (𝑧) and 𝑓 (𝛾 0 𝑧) = 𝚥 (𝛾 0; 𝑧) 𝑘 𝑓 (𝑧), then
𝑓 (𝛾(𝛾 0 𝑧)) = 𝚥 (𝛾; 𝛾 0 𝑧) 𝑘 𝑓 (𝛾 0 𝑧) = 𝚥 (𝛾; 𝛾 0 𝑧) 𝑘 𝚥 (𝛾 0; 𝑧) 𝑘 𝑓 (𝑧)
(40.2.7)
= 𝚥 (𝛾𝛾 0; 𝑧) 𝑘 𝑓 (𝑧).
Since PSL2 (Z) is generated by 𝑆, 𝑇, it follows from (40.2.7) that a map 𝑓 is weight
𝑘 invariant for PSL2 (Z) if and only if both equalities
𝑓 (𝑧 + 1) = 𝑓 (𝑧)
(40.2.8)
𝑓 (−1/𝑧) = 𝑧 𝑘 𝑓 (𝑧)
hold for all 𝑧 ∈ H2 .

40.2.9. Since
d(𝛾𝑧) 1
= (40.2.10)
d𝑧 (𝑐𝑧 + 𝑑) 2
the weight 𝑘 invariance (40.2.2) of a map 𝑓 can be rewritten
𝑓 (𝛾𝑧) d(𝛾𝑧) ⊗𝑘/2 = 𝑓 (𝑧) d𝑧 ⊗𝑘/2 (40.2.11)
so equivalently, the differential 𝑓 (𝑧) d𝑧 ⊗𝑘/2 is (straight up) invariant under Γ.
40.2.12. Let 𝑓 : H2 → C be a meromorphic map that is weight 𝑘 invariant under a

1 1
Fuchsian group Γ 3 . Then 𝑓 (𝑧 + 1) = 𝑓 (𝑧). If 𝑓 admits a Fourier series
0 1
expansion in 𝑞 = exp(2𝜋𝑖𝑧) of the form
∞
∑︁
𝑓 (𝑧) = 𝑎 𝑛 𝑞 𝑛 ∈ C((𝑞)) (40.2.13)
𝑛=−∞
with 𝑎 𝑛 ∈ C and 𝑎 𝑛 = 0 for all but finitely many 𝑛 < 0, then we say that 𝑓 is
meromorphic at ∞; if further 𝑎 𝑛 = 0 for 𝑛 < 0, we say 𝑓 is holomorphic at ∞.
More generally, let Γ ≤ PSL2 (Z) be a subgroup of finite index. For 𝛾 ∈ PSL2 (Z),
we define
𝑓 [𝛾] 𝑘 (𝑧) := 𝚥 (𝛾; 𝑧) −𝑘 𝑓 (𝛾𝑧). (40.2.14)
Then 𝑓 [𝛾] 𝑘 (𝑧) is weight 𝑘 invariant under the group 𝛾 −1 Γ𝛾. We say that 𝑓 is mero-
morphic at the cusps if for every 𝛾 ∈ PSL2 (Z), the function 𝑓 [𝛾] 𝑘 is meromorphic
at ∞, in the above sense. Since 𝑓 is weight 𝑘 invariant, to check if 𝑓 is meromorphic
at the cusps, it suffices to take representatives of the finite set of cosets Γ\ PSL2 (Z).
(The name cusp comes from the geometric description at ∞ coming from the parabolic
stabilizer group, recalling Definition 33.4.5.)
Finally, we say that 𝑓 is holomorphic at the cusps if 𝑓 [𝛾] 𝑘 (𝑧) is holomorphic at
∞ for all 𝛾 ∈ PSL2 (Z), and vanishes at the cusps if 𝑓 [𝛾] 𝑘 (∞) = 0 for all 𝛾.
Definition 40.2.15. Let 𝑘 ∈ 2Z and let Γ ≤ PSL2 (Z) be a subgroup of finite index. A
meromorphic modular form of weight 𝑘 is a meromorphic map 𝑓 : H2 → C that is
weight 𝑘 invariant under Γ and meromorphic at the cusps. A meromorphic modular
function is a meromorphic modular form of weight 0.
A (holomorphic) modular form of weight 𝑘 is a holomorphic map 𝑓 : H2 → C
that is weight 𝑘 invariant under Γ and holomorphic at the cusps. A cusp form of
weight 𝑘 is a holomorphic modular form of weight 𝑘 that vanihses at the cusps.
Let 𝑀𝑘 (Γ) be the C-vector space of modular forms of weight 𝑘 for Γ, and let
𝑆 𝑘 (Γ) ⊆ 𝑀𝑘 (Γ) be the subspace of cusp forms.
Lemma 40.2.16. The Eisenstein series 𝐺 𝑘 (𝑧) is a holomorphic modular form of

weight 𝑘 ∈ 2Z ≥2 for PSL2 (Z) with Fourier expansion
∞
(2𝜋𝑖) 𝑘 ∑︁
𝐺 𝑘 (𝑧) = 2𝜁 (𝑘) + 2 𝜎𝑘−1 (𝑛)𝑞 𝑛 (40.2.17)
(𝑘 − 1)! 𝑛=1
where
∞
∑︁ 1
𝜁 (𝑘) =
𝑛=1
𝑛𝑘
and ∑︁
𝜎𝑘−1 (𝑛) = 𝑑 𝑘−1 .
𝑑 |𝑛
𝑑>0
40.2. ⊲ EISENSTEIN SERIES AS MODULAR FORMS 737
Proof. We start with the formula

∞ 𝑀
∑︁ 1 ∑︁ 1
𝜋 cot(𝜋𝑧) = = lim (40.2.18)
𝑚=−∞
𝑧 + 𝑚 𝑀 →∞ 𝑚=−𝑀 𝑧 + 𝑚
(Exercise 40.2); with 𝑞 = exp(2𝜋𝑖𝑧),

cos(𝜋𝑧) 𝑞+1 2
cot(𝜋𝑧) = =𝑖 =𝑖 1+
sin(𝜋𝑧) 𝑞−1 𝑞−1
and so we obtain the Fourier expansion
∞
∑︁
𝜋 cot(𝜋𝑧) = 𝜋𝑖 − 2𝜋𝑖 𝑞𝑛 . (40.2.19)
𝑛=0
Equating (40.2.18)–(40.2.19) and differentiating 𝑘 − 1 times, we find that

∞ ∞
∑︁ 1 1 𝑘
∑︁
= (2𝜋𝑖) 𝑛 𝑘−1 𝑞 𝑛 (40.2.20)
𝑚=−∞
(𝑚 + 𝑧) 𝑘 (𝑘 − 1)! 𝑛=1
(since 𝑘 is even). Thus

∞ ∑︁∞
∑︁ 1 ∑︁ 1
𝐺 𝑘 (𝑧) = = 2𝜁 (𝑘) + 2
(𝑚,𝑛)≠(0,0)
(𝑚 + 𝑛𝑧) 𝑘 𝑛=1 𝑚=−∞
so replacing 𝑛 ← 𝑎 and then substituting 𝑧 ← 𝑛𝑧 in (40.2.20), summing over 𝑛 we

obtain
∞ ∞
(2𝜋𝑖) 𝑘 ∑︁ ∑︁ 𝑘−1 𝑎𝑛
𝐺 𝑘 (𝑧) = 2𝜁 (𝑘) + 2 𝑎 𝑞
(𝑘 − 1)! 𝑛=1 𝑎=1
∞ (40.2.21)
(2𝜋𝑖) 𝑘 ∑︁ ∑︁ 𝑘−1 𝑛
= 2𝜁 (𝑘) + 2 𝑑 𝑞
(𝑘 − 1)! 𝑛=1
𝑑 |𝑛
grouping together terms in the second step. The fact that 𝐺 𝑘 is holomorphic at ∞ then
follows by definition.
40.2.22. We accordingly define the normalized Eisenstein series by
1
𝐸 𝑘 (𝑧) = 𝐺 𝑘 (𝑧)
2𝜁 (𝑘)
(see also 40.1.19). We have
∞
2𝑘 ∑︁
𝐸 𝑘 (𝑧) = 1 − 𝜎𝑘−1 (𝑛)𝑞 𝑛 (40.2.23)
𝐵 𝑘 𝑛=1
where 𝐵 𝑘 ∈ Q× are the Taylor coefficients of

∞
𝑥 ∑︁ 𝑥 𝑘 1 1 𝑥2 1 𝑥4 1 𝑥6
= 𝐵 𝑘 = 1 − 𝑥 + − + +...
𝑒 𝑥 − 1 𝑘=0 𝑘! 2 6 2! 30 4! 42 6!
(Exercise 40.3): the numbers 𝐵 𝑘 ∈ Q (with 𝐵 𝑘 ≠ 0 for 𝑘 ∈ 2Z ≥0 ) are Bernoulli

numbers. Expanding, we find
𝐸 4 (𝑧) = 1 + 240𝑞 + 2160𝑞 2 + 6720𝑞 3 + 17520𝑞 4 + . . .

𝐸 6 (𝑧) = 1 − 504𝑞 − 16632𝑞 2 − 122976𝑞 3 − 532728𝑞 4 − . . . .
Remark 40.2.24. The notion of Eisenstein series extends in a natural way to the Bianchi
groups PSL2 (Z𝐹 ) where 𝐹 is an imaginary quadratic field: see Elstrodt–Grunewald–
Mennicke [EGM98, Chapter 3].
40.3 ⊲ Classical modular forms

In this section, we study modular forms for PSL2 (Z); to ease notation, we abbreviate
Γ := PSL2 (Z).
40.3.1. Let 𝑓 be a meromorphic modular form of weight 𝑘 for Γ = PSL2 (Z). If 𝑘 = 0,
so 𝑓 is a modular function, then 𝑓 descends to a meromorphic function on 𝑌 := Γ\H2 .
Although this is not true for (nonzero) forms of weight 𝑘 ≠ 0, the order of zero or
pole ord𝑧 ( 𝑓 ) is well defined on the orbit Γ𝑧 by weight 𝑘 invariance (40.2.2). With the
Fourier expansion (40.2.13), we define
∑︁
ord∞ ( 𝑓 ) := ord𝑞 𝑎 𝑛 𝑞 𝑛 = min({𝑛 : 𝑎 𝑛 ≠ 0}).
𝑛
The form 𝑓 has only finitely many zeros or poles in 𝑌 , i.e., only finitely many
Γ-orbits of zeros or poles: since 𝑓 is meromorphic at ∞, there exists 𝜖 > 0 such that
𝑓 has no zero or pole with 0 < |𝑞| < 𝜖, so with
log(1/𝜖)
Im 𝑧 > 𝑀 = ;
2𝜋
but the part of ◊ with Im 𝑧 ≤ 𝑀 is compact, and since 𝑓 is meromorphic in H2 , it has
only finitely many zeros or poles in this part as well.
40.3.2. In a similar way, the order of the stabilizer 𝑒 𝑧 := # StabΓ (𝑧) is well defined on
the orbit Γ𝑧, since points in the same orbit have conjugate stabilizers. By 35.1.14,


 3, if Γ𝑧 = Γ𝜔;


𝑒 𝑧 = 2, if Γ𝑧 = Γ𝑖; (40.3.3)

 1,

otherwise.

Proposition 40.3.4. Let 𝑓 : H2 → C be a meromorphic modular form of weight 𝑘 for
Γ = PSL2 (Z), not identically zero. Then
∑︁ 1 𝑘
ord∞ ( 𝑓 ) + ord𝑧 ( 𝑓 ) = (40.3.5)
2
𝑒𝑧 12
Γ𝑧 ∈Γ\H
where 𝑒 𝑧 := # StabΓ (𝑧).

40.3. ⊲ CLASSICAL MODULAR FORMS 739
The sum (40.3.5) has only finitely many terms, by 40.3.1, and the stabilizers are
given in 40.3.2.
Proof. See Serre [Ser73, §3, Theorem 3]: the proof consists of performing a contour
1 d𝑓
integration on the boundary of ◊. Alternatively, this statement can be seen
2𝜋𝑖 𝑓
as a manifestation of the Riemann–Roch theorem: see Diamond–Shurman [DS2005,
§3.5].
40.3.6. We have 𝐸 4 (𝑆𝑇 𝑧) = (𝑧 + 1) 4 𝐸 4 (𝑧), so since (𝑆𝑇) (𝜔) = 𝜔,
𝐸 4 (𝜔) = (𝜔 + 1) 4 𝐸 4 (𝜔) = 𝜔2 𝐺 4 (𝜔)
so 𝐸 4 (𝜔) = 0. Since 𝐸 4 is holomorphic in H2 , we have ord𝑧 (𝐸 4 ) ∈ Z ≥0 for all 𝑧 ∈ H2 ,

and thus by Proposition 40.3.4, we must have that 𝐸 4 (𝑧) has no other zeros in ◊.
Similarly,
𝐸 6 (𝑖) = 𝐸 6 (𝑆𝑖) = 𝑖 6 𝐸 6 (𝑖) = −𝐸 6 (𝑖)
so 𝐸 6 (𝑖) = 0, and 𝐸 6 (𝑧) has no other zeros.
For the same reason, the function
𝐸 4 (𝑧) 3 − 𝐸 6 (𝑧) 2
Δ(𝑧) = = 𝑞 − 24𝑞 2 + 252𝑞 3 − 1472𝑞 4 + . . . (40.3.7)
1728
is a modular form of weight 12 with no zeros in H2 with ord∞ (Δ) = 1.
We give two applications of Proposition 40.3.4. First, we obtain the identification

promised in (40.1.18).
Theorem 40.3.8. The function
𝐸 4 (𝑧) 3 1
𝑗 (𝑧) = = + 744 + 196884𝑞 + 21493760𝑞 2 + . . . (40.3.9)
Δ(𝑧) 𝑞
is a meromorphic modular function for PSL2 (Z), holomorphic in H2 , defining a

bĳection
𝑌 = Γ\H2 → C.
Proof. The function 𝑗 is weight 0 invariant under Γ as the ratio of two forms that are
weight 12 invariant. Since 𝐸 4 is holomorphic in H2 , and Δ is holomorphic and has
no zeros in H2 , the ratio is holomorphic in H2 ; and 𝑗 (𝑧) has a simple pole at 𝑧 = ∞,
corresponding to a simple zero of Δ at 𝑧 = ∞. From 40.3.6, we have 𝑗 (𝑖) = 1728, and
𝑗 (𝑧) − 1728 has a double zero at 𝑧 = 𝑖, and 𝑗 (𝑧) has a triple zero at 𝑧 = 𝜔.
To conclude that 𝑗 is bĳective, we show that 𝑗 (𝑧) − 𝑐 has a unique zero Γ𝑧 ∈ 𝑌 .
If 𝑐 ≠ 0, 1728, this follows immediately from Proposition 40.3.4; if 𝑐 = 0, 1728, the
results follow for the same reason from the multiplicity of the zero.
Remark 40.3.10. The definition of 𝑗 (𝑧) is now standard, but involves some choices.
In some circumstances (including the generalization to abelian surfaces, see 43.5.7),
it is more convenient to remember the values of the Eisenstein series themselves, as
follows. To 𝑧 ∈ H , we associate the pair (𝐸 4 (𝑧), 𝐸 6 (𝑧)) ∈ C2 ; if 𝛾 ∈ Γ and 𝑧 0 = 𝛾𝑧,
then
(𝐸 4 (𝑧 0), 𝐸 6 (𝑧 0)) = (𝛿4 𝐸 4 (𝑧), 𝛿6 𝐸 6 (𝑧))
where 𝛿 = 𝚥 (𝛾; 𝑧) ∈ C× . We therefore define the weighted projective (4, 6)-space by
P(4, 6) (C) := (C2 r {(0, 0)})/∼
where
(𝐸 4 , 𝐸 6 ) ∼ (𝛿4 𝐸 4 , 𝛿6 𝐸 6 )
for 𝛿 ∈ C× . We write equivalence classes (𝐸 4 : 𝐸 6 ) ∈ P(4, 6) (C). The map
𝑗 : P(4, 6) (C) → P1 (C)

1728𝐸 43
(𝐸 4 : 𝐸 6 ) ↦→ 𝑗 (𝐸 4 : 𝐸 6 ) =
𝐸 43 − 𝐸 62
is well-defined and bĳective—see Silverman [Sil2009, Proposition III.1.4(b)].

To conclude, we give a complete description of the ring of (holomorphic) modular
forms. By 40.2.3, the C-vector space
Ê
𝑀 (Γ) = 𝑀𝑘 (Γ)
𝑘 ∈2Z
under multiplication has the structure of a (graded) C-algebra; we call 𝑀 (Γ) the ring
of modular forms for Γ.
Theorem 40.3.11. We have 𝑀 (Γ) = C[𝐸 4 , 𝐸 6 ], i.e., every modular form for Γ =
PSL2 (Z) can be written as a polynomial in 𝐸 4 , 𝐸 6 .
In particular, 𝑀𝑘 (Γ) = {0} for 𝑘 < 0.
Proof. We have 𝑀 (Γ) ⊇ C[𝐸 4 , 𝐸 6 ], so we prove the reverse inclusion. We refer to

Proposition 40.3.4, and ask for solutions 𝑎 1 , 𝑎 2 , 𝑎 3 ∈ Z ≥0 to 𝑎 1 + 𝑎 2 /2 + 𝑎 3 /3 = 𝑘/12.
When 𝑘 < 0, there are no such solutions; when 𝑘 = 0, 2, 4, 6, 8, 10, there is a unique
solution, and we find that 𝑀𝑘 (Γ) is spanned by 1, 0, 𝐸 4 , 𝐸 6 , 𝐸 42 , 𝐸 4 𝐸 6 , respectively.
For all even 𝑘 ≥ 4, there exist 𝑎, 𝑏 ∈ Z ≥0 such that 4𝑎 + 6𝑏 = 𝑘 (if 𝑘 ≥ 10 and
𝑘 ≡ 2 (mod 4), then 𝑘 − 6 ≥ 4 and 4 | 𝑘), so 𝐸 4𝑎 𝐸 6𝑏 ∈ 𝑀𝑘 (Γ) and (𝐸 4𝑎 𝐸 6𝑏 ) (∞) = 1
(by 40.2.22). Let 𝑆 𝑘 (Γ) ⊆ 𝑀𝑘 (Γ) be the subspace of forms that vanish at ∞. Then
𝑀𝑘 (Γ) = C𝐸 4𝑎 𝐸 6𝑏 ⊕ 𝑆 𝑘 (Γ) by linear algebra, and by the previous paragraph, 𝑆 𝑘 (Γ) =
{0} for 𝑘 ≤ 10.
We claim that multiplication by Δ furnishes an isomorphism 𝑀𝑘 (Γ) −→ ∼ 𝑆
𝑘+12 (Γ)
of C-vector spaces for all 𝑘: division by Δ defines an inverse because Δ has a simple
zero at ∞ by 40.3.6 and no zeros in H2 . The result now follows by induction on
𝑘 ≥ 0.
40.4. THETA SERIES 741
More generally, one can study modular forms for congruence subgroups (section
35.4) of PSL2 (Z) in an explicit way, as the following example illustrates.
Example 40.3.12. At the end of section 35.4, we examined a fundamental domain for
the group Γ(2), defined by (35.4.8). As with 𝑋 (1), the homeomorphism (35.4.11) can
be given by a holomorphic map
∼ P1 (C)
𝜆 : 𝑋 (2) −→
obtained from Eisenstein series for Γ(2), analogous to 𝑗 (𝑧). The map 𝜆 satisfies
𝜆(𝛾𝑧) = 𝜆(𝑧) for all 𝛾 ∈ Γ(2) and in particular is invariant under 𝑧 ↦→ 𝑧 + 2. One can
compute its Fourier expansion in terms of 𝑞 1/2 = 𝑒 𝜋𝑖𝑧 as:
𝜆(𝑧) = 16𝑞 1/2 − 128𝑞 + 704𝑞 3/2 − 3072𝑞 2 + 11488𝑞 5/2 − 38400𝑞 3 + . . . . (40.3.13)
Since 𝑗 (𝑧) induces a degree 6 = [Γ(2) : Γ(1)] map 𝑋 (2) → 𝑋 (1), we find the
relationship
(𝜆2 − 𝜆 + 1) 3
𝑗 = 256 2 . (40.3.14)
𝜆 (𝜆 − 1) 2
From (40.3.14) (and the first term), the complete series expansion (40.3.13) can be
obtained recursively.
As a uniformizer for a congruence subgroup of PSL2 (Z), the function 𝜆(𝑧) has
a moduli interpretation (cf. 40.1.11): there is a family of elliptic curves over 𝑋 (2)
equipped with extra structure. Specifically, given 𝜆 ∈ P1 (C) \ {0, 1, ∞}, the corre-
sponding elliptic curve with extra structure is given by the Legendre curve
𝐸 𝜆 : 𝑦 2 = 𝑥(𝑥 − 1) (𝑥 − 𝜆),
equipped with the isomorphism (Z/2Z) 2 −→ ∼ 𝐸 [2] determined by sending the standard
generators to the 2-torsion points (0, 0) and (1, 0). The map 𝑗 is the map that forgets
this additional torsion structure on a Legendre curve and remembers only isomorphism
class.
40.4 Theta series

An important class of classical modular forms arise via theta series, which count the
number of representations of an integer by a positive definite quadratic form. We
present only a small fraction of the general theory here.
Let 𝑄 : Z𝑚 → Z be a positive definite integral quadratic form in 𝑚 variables and
suppose that 𝑚 = 2𝑘 is a positive even integer. We define the theta series of 𝑄 by
Θ𝑄 : H2 → C
∞
∑︁ ∑︁ (40.4.1)
Θ𝑄 (𝑧) = 𝑒 2 𝜋𝑖𝑄 ( 𝑥) 𝑧 = 𝑟 𝑄 (𝑛)𝑞 𝑛
𝑥 ∈Z𝑚 𝑛=0
where 𝑞 = 𝑒 2 𝜋𝑖𝑧 and

𝑟 𝑄 (𝑛) = #{𝑥 ∈ Z𝑚 : 𝑄(𝑥) = 𝑛} < ∞
√
counts the number of lattice points on the sphere of radius 𝑛.
Lemma 40.4.2. Θ𝑄 (𝑧) is a holomorphic function.
Proof. Since 𝑄 is positive definite, there exists 𝑐 ∈ R>0 such that
𝑄(𝑥) ≥ 𝑐(𝑥12 + · · · + 𝑥 𝑚
2
).
Thus 𝑟 𝑄 (𝑛) = 𝑂 (𝑛 𝑘 ), and the series Θ𝑞 (𝑧) is majorized by (a constant multiple of)
Í ∞
𝑛=1 𝑛 𝑞 , so converges to a holomorphic function.
𝑘 𝑛
Let [𝑇] be the Gram matrix for the symmetric bilinear form associated to 𝑄;
then [𝑇] ∈ M𝑚 (Z) is an integral symmetric matrix with even diagonal entries. Let
𝑑 = det 𝑄 = det[𝑇] ∈ Z. Then 𝑑𝐴−1 ∈ M𝑚 (Z) is the adjugate matrix: it is again
symmetric.
Definition 40.4.3. The least positive integer 𝑁 ∈ Z>0 such that 𝑁 𝐴−1 is integral with
even diagonal entries is called the level of 𝑄.
We recall the definition of the congruence subgroups 35.4.5.
Theorem 40.4.4. The theta series Θ𝑄 (𝑧) is a modular form of weight 𝑘 for Γ1 (𝑁).
Proof. Unfortunately, in this generality the proof would take us too far afield. Fun-
damentally, the transformation formula for Θ𝑄 follows from Poisson summation and
careful computations: see Eichler [Eic73, §I.3, Proposition 2], Miyake [Miy2006,
Corollary 4.9.5], or Ogg [Ogg69, Chapter VI].
40.4.5. We can be a bit more specific about the transformation group for Θ𝑄 (𝑧) as

(−1) 𝑘 det 𝑄
follows. To 𝑄, we associate the character 𝜒 defined by 𝜒(𝑛) = . Then
𝑛

𝑎 𝑏
Θ𝑄 (𝛾𝑧) = 𝜒(𝑑) (𝑐𝑧 + 𝑑) 𝑘 Θ𝑄 (𝑧) for all 𝛾 = ∈ Γ0 (𝑁);
𝑐 𝑑
accordingly, we say Θ𝑄 is a modular form of level 𝑁 with character 𝜒.
40.5 Hecke operators

Significantly, the C-vector space 𝑀𝑘 (Γ0 (𝑁)) of modular forms of weight 𝑘 for Γ0 (𝑁)
carries with it an action of commuting semisimple operators, called Hecke operators.
These operators may be interpreted as averaging modular forms over sublattices of
a fixed index; for efficiency, we work with a more explicit definition. For further
reference, see e.g. Diamond–Shurman [DS2005, Chapter 5] or Miyake [Miy2006,
§2.7, §4.5].
Throughout, let 𝑁 ∈ Z ≥1 . Let

𝑎 𝑏
O = O0 (𝑁) := ∈ M2 (Z) : 𝑁 | 𝑐 ⊆ M2 (Z)
𝑐 𝑑
be the standard Eichler order of level 𝑁 in M2 (Q), so that Γ = Γ0 (𝑁) = O1 /{±1}.

40.5. HECKE OPERATORS 743
Let 𝑛 ∈ Z ≥1 with gcd(𝑛, 𝑁) = 1. We consider the set of matrices
O𝑛 = {𝛼 ∈ O : det(𝛼) = 𝑛}. (40.5.1)
Visibly, there is a left (and right) action of O1 on O𝑛 by multiplication.
Lemma 40.5.2.
A system of representatives of O1 \O𝑛 is given by the set of matrices
𝑎 𝑏
of the form with 𝑎𝑑 = 𝑛, 𝑎 > 0, and 0 ≤ 𝑏 < 𝑑.
0 𝑑
Proof. The lemma follows as in Lemma 26.4.1(b) using the theory of elementary
divisors, but applying row operations (acting on the left).
Example 40.5.3. When 𝑝 - 𝑁 is prime, the set O1 \O 𝑝 is represented by the 𝑝 + 1

matrices
𝑝 0 1 0 1 1 1 𝑝−1
, , ,··· , .
0 1 0 𝑝 0 𝑝 0 𝑝
Definition 40.5.4. For 𝑛 ∈ Z ≥1 with gcd(𝑛, 𝑁) = 1, we define the Hecke operator
𝑇 (𝑛) : 𝑀𝑘 (Γ) → 𝑀𝑘 (Γ)

∑︁
(𝑇 (𝑛) 𝑓 ) (𝑧) = 𝑛 𝑘/2−1 𝚥 (𝛼; 𝑧) −𝑘 𝑓 (𝛼𝑧). (40.5.5)
O1 𝛼∈O1 \O 𝑛
By the condition of automorphy 𝑓 (𝛾𝑧) = 𝚥 (𝛾; 𝑧) 𝑘 𝑓 (𝑧) and the cocycle relation
(40.2.5), the Hecke operators are well-defined and preserve weight 𝑘 invariance.
40.5.6. By Lemma 40.5.2, we have more explicitly

𝑑−1
𝑘−1
∑︁ 1 ∑︁ 𝑎𝑧 + 𝑏
(𝑇 (𝑛) 𝑓 ) (𝑧) = 𝑛 𝑓 . (40.5.7)
𝑎𝑑=𝑛
𝑑 𝑘 𝑏=0 𝑑
𝑎>0
Í∞
𝑎 𝑛 𝑞 𝑛 , then (𝑇 (𝑛) 𝑓 ) (𝑧) = ∞
Í
Accordingly, if 𝑓 (𝑧) = 𝑛=0 𝑚=0 𝑏 𝑚 𝑞 where
𝑚
∑︁
𝑏𝑚 = 𝑑 𝑘−1 𝑎 𝑚𝑛/𝑑 2 (40.5.8)
𝑑>0
so in particular for 𝑛 = 𝑝 prime we have

(
𝑝 𝑘−1 𝑎 𝑚/ 𝑝 , if 𝑝 | 𝑚;
𝑏 𝑚 = 𝑎 𝑝𝑚 + (40.5.9)
0, if 𝑝 - 𝑚.
Applying (40.5.9), we see that if 𝑓 ∈ 𝑀𝑘 (Γ) has 𝑓 (∞) = 0 (equivalently, 𝑎 0 = 0), then
the same is true for 𝑇 (𝑛) 𝑓 . Repeating this for the functions 𝑓 [𝛾] 𝑘 with 𝛾 ∈ PSL2 (Z)
(as in (40.2.14)), we conclude that the operators 𝑇 (𝑛) act on the space of cusp forms
𝑆 𝑘 (Γ) ⊂ 𝑀𝑘 (Γ).
Proposition 40.5.10. For 𝑚, 𝑛 ∈ Z ≥1 , we have

∑︁
𝑇 (𝑚)𝑇 (𝑛) = 𝑑 𝑘−1𝑇 (𝑚𝑛/𝑑 2 ). (40.5.11)
In particular, if gcd(𝑚, 𝑛) = 1 then
𝑇 (𝑚)𝑇 (𝑛) = 𝑇 (𝑛)𝑇 (𝑚) = 𝑇 (𝑚𝑛)
and if 𝑝 is prime and 𝑟 ≥ 1 then
𝑇 ( 𝑝)𝑇 ( 𝑝 𝑟 ) = 𝑇 ( 𝑝 𝑟 +1 ) + 𝑝 𝑘−1𝑇 ( 𝑝 𝑟 −1 ).
Proof. These statements follow directly from the expansion (40.5.8).
Theorem 40.5.12. The Hecke operators 𝑇 (𝑛) for gcd(𝑛, 𝑁) = 1 on 𝑀𝑘 (Γ0 (𝑁))
generate a commutative, semisimple Z-algebra.
Proof. See e.g. Diamond–Shurman [DS2005, Theorem 5.5.4]. Briefly, we treat Eisen-
stein series separately and work with cusp forms 𝑆 𝑘 (Γ0 (𝑁)). To prove that the operators
are semisimple, we would need to show that the Petersson inner product
∫
h 𝑓 , 𝑔i = 𝑓 (𝑧)𝑔(𝑧)𝑦 𝑘 d𝜇(𝑧)
Γ\H2
is well-defined, positive, and nondegenerate, and then verify that the operators are
normal with respect to this inner product.
By Theorem 40.5.12 and linear algebra, there exists a C-basis 𝑓𝑖 (𝑧) of 𝑀𝑘 (Γ0 (𝑁))
consisting of simultaneous eigenfunctions for all 𝑇 (𝑛).
Exercises
I 1. Let 𝑓 : 𝑈 → C be a function that is meromorphic in an open neighborhood
𝑈 ⊇ C with 𝑧 ∈ 𝑈, and let 𝐶 be a contour along an arc of a circle of radius 𝜖 > 0
centered at 𝑧 contained in 𝑈 with total angle 𝜃. Show that
∫
d𝑓
lim = 𝜃𝑖 ord𝑧 ( 𝑓 ).
𝜖 →0 𝐶 𝑓
I 2. Prove the formula

∞
∑︁ 1
𝜋 cot(𝜋𝑧) =
𝑚=−∞
𝑧+𝑚
for 𝑧 ∈ C. [Hint: the difference ℎ(𝑧) of the left- and right-hand sides is bounded
away from Z, invariant under 𝑧 ↦→ 𝑇 (𝑧) = 𝑧 + 1, and in a neighborhood of 0
is holomorphic (both sides have principal part 1/𝑧) so bounded; thus ℎ(𝑧) is
bounded in C and hence constant.]
40.5. HECKE OPERATORS 745
I 3. In this exercise, we give Euler’s evaluation of 𝜁 (𝑘) in terms of Bernoulli numbers.

As in Exercise 36.12, define the series
∞
𝑥 ∑︁ 𝑥 𝑘 𝑥 1 𝑥2 1 𝑥4
= 𝐵𝑘 =1− + − + . . . ∈ Q[[𝑥]]. (40.5.13)
𝑒 − 1 𝑘=0
𝑥 𝑘! 2 6 2! 30 4!
(a) Plug in 𝑥 = 2𝑖𝑧 into (40.5.13) to obtain

∞
∑︁ (2𝑖𝑧) 𝑘
𝑧 cot 𝑧 = 1 + 𝐵𝑘 .
𝑘=2
𝑘!
(b) Take the logarithmic derivative of

∞
𝑧2
Ö
sin 𝑧 = 𝑧 1−
𝑛=1
𝑛 𝜋2
2
to show
∞ ∑︁∞
∑︁ 𝑧 𝑘
𝑧 cot 𝑧 = 1 − 2 .
𝑘=2 𝑛=1
𝑛𝜋
𝑘 even
(c) Conclude that

∞
∑︁ 1 1 (2𝜋𝑖) 𝑘
𝜁 (𝑘) = = − 𝐵𝑘
𝑛=1
𝑛𝑘 2 𝑘!
for 𝑘 ∈ 2Z ≥1 .
4. We defined Eisenstein series 𝐺 𝑘 (𝑧) for 𝑘 ≥ 4, and found 𝐺 𝑘 (𝑧) ∈ M 𝑘 (SL2 (Z))
are modular forms of weight 𝑘 for SL2 (Z). The case 𝑘 = 2 is also important,
though we must be a bit more careful in its analysis. Let
∑︁ ∑︁ 1
𝐺 2 (𝑧) := .
𝑐 ∈Z 𝑑 ∈Z
(𝑐𝑧 + 𝑑) 2
(𝑐,𝑑)≠(0,0)
(a) Show that 𝐺 2 (𝑧) converges (conditionally) and satisfies

∞
∑︁
𝐺 2 (𝑧) = 2𝜁 (2) − 8𝜋 2 𝜎(𝑛)𝑞 𝑛
𝑛=1
where 𝑞 = 𝑒 2 𝜋𝑖𝑧 .
(b) Show that 𝐺 2 (𝑧 + 1) = 𝐺 2 (𝑧) and

−1 2𝜋𝑖
𝐺2 = 𝐺 2 (𝑧) − .
𝑧 𝑧
[Hint: use a telescoping series and rearrange terms.]

(c) Conclude that

2𝜋𝑖𝑐 𝑎 𝑏
𝐺 2 (𝛾𝑧) = 𝐺 2 (𝑧) − for all 𝛾 = .
𝑐𝑧 + 𝑑 𝑐 𝑑
[Hint: use the cocycle relation.]
(d) Define
𝜋
𝐺 ∗2 (𝑧) = 𝐺 2 (𝑧) − .
Im 𝑧
Show that 𝐺 ∗2 (𝑧) is weight 2 invariant under SL2 (Z).
5. In this exercise, we give a proof using modular forms of the formula for the
number of ways of representing an integer as the sum of four squares, due to
Jacobi. Consider the function
∞
∑︁ 2
𝜗(𝑧) := 𝑞𝑛
𝑛=−∞
where 𝑞 := 𝑒 2 𝜋𝑖𝑧 and 𝑧 ∈ H2 . [The function 𝜗(𝑧) is a theta series for the univari-
ate quadratic form 𝑥 ↦→ 𝑥 2 .] Let 𝑟 4 (𝑛) be the number of ways of representing
𝑛 ≥ 0 as the sum of 4 squares.
(a) Show that
∞
∑︁
Θ𝑄 (𝑧) := 𝜗(𝑧) 4 = 1 + 𝑟 4 (𝑛)𝑞 𝑛 = 1 + 8𝑞 + 12𝑞 2 + . . . .
𝑛=1
(b) Show that Θ𝑄 (𝑞) ∈ M2 (Γ0 (4)) is a modular form of weight 2 on Γ0 (4).
(c) Show that dimC 𝑀2 (Γ0 (4)) = 2 and dimC 𝑆2 (Γ0 (4)) = 0.
(d) Let
𝐺 2,2 (𝑧) = 𝐺 2 (𝑧) − 2𝐺 2 (2𝑧)
𝐺 2,4 (𝑧) = 𝐺 2 (𝑧) − 4𝐺 2 (4𝑧).
Show that 𝐺 2,2 , 𝐺 2,4 are a basis for M2 (Γ0 (4)). [Hint: use Exercise
40.4(c).]
(e) Show that
∞
3 ∑︁
:=
𝐸 2,2 (𝑧) − 2 𝐺 2,2 (𝑧) = 1 + 24 𝜎 (2) (𝑛)𝑞 𝑛
𝜋 𝑛=1
∞
1 ∑︁
𝐸 2,4 (𝑧) := − 𝐺 2,4 (𝑧) = 1 + 8 𝜎 (4) (𝑛)𝑞 𝑛
𝜋2 𝑛=1
where ∑︁
𝜎 (𝑚) (𝑛) = 𝑑.
𝑚-𝑑 |𝑛
(f) Matching the first few coefficients, show that
Θ𝑄 (𝑧) = 𝐸 2,4 (𝑧).
Conclude that 𝑟 4 (𝑛) = 8𝜎 (4) (𝑛) for all 𝑛 > 0.
Chapter 41
Brandt matrices
In this chapter, we revisit classes of quaternion ideals: organizing ideals of given norm
in terms of their classes, we find modular forms.
41.1 ⊲ Brandt matrices, neighbors, and modular forms

Let 𝐵 be a quaternion algebra over Q. A major theme of this text has been the study of
classes of quaternion ideals, beginning with chapter 17. When 𝐵 is indefinite, we saw
(Theorem 17.8.3, treated broadly in chapter 28) that strong approximation applies: via
the reduced norm, very often the conclusion is that the class set is trivial.
We are left with the case that 𝐵 is definite. By the geometry of numbers (see section
17.5) we found that the number of ideal classes of an order is finite, generated by ideals
of small reduced norm. (Studying the zeta function we found a mass formula in
chapter 25, and then studying quadratic embeddings we found a class number formula
in section 30.8.) We now pursue this further: there is an exquisite arithmetic and
combinatorial structure to be found by counting right ideals of given norm by their
classes as follows. We begin in this section by an introduction and survey (working
over Q).
Let O ⊂ 𝐵 be an order. Let Cls O be the right class set of O, keeping track of the
isomorphism classes of invertible right O-ideals in 𝐵. Let ℎ := # Cls O be the (right)
class number of O, and let 𝐼1 , . . . , 𝐼 ℎ ⊆ 𝐵 represent the distinct classes in Cls O.
Let 𝑛 ∈ Z ≥1 . We define an ℎ × ℎ-matrix 𝑇 (𝑛) ∈ Mℎ (Z) with nonnegative integer
entries, called the 𝑛-Brandt matrix, by
𝑇 (𝑛)𝑖 𝑗 := #{𝐽 ⊂ 𝐼 𝑗 : nrd(𝐽) = 𝑛 nrd(𝐼 𝑗 ) and [𝐽] = [𝐼𝑖 ]}

(41.1.1)
= #{𝐽 ⊂ 𝐼 𝑗 : [𝐼 𝑗 : 𝐽] = 𝑛2 and [𝐽] = [𝐼𝑖 ]}.
(The notation 𝑇 (𝑛) deliberately overloads that of the Hecke operators defined in section
40.5: keep reading to see why!) The Brandt matrix 𝑇 (𝑛) depends on O, but for brevity
we do not include this in the notation. In the 𝑗th column of the Brandt matrix 𝑇 (𝑛),
we look at the subideals of 𝐼 𝑗 with index 𝑛2 and count them in the 𝑖 𝑗th entry according
to the class [𝐼𝑖 ] they belong to. If 𝑛 = 𝑝 is prime and 𝑝 - 𝑁 = disc O, then there are
747
748 CHAPTER 41. BRANDT MATRICES
exactly 𝑝 + 1 such ideals, so the sum of the entries in every column in 𝑇 ( 𝑝) is equal to
𝑝 + 1.

−1, −23
Example 41.1.2. We continue with Example 17.6.3. We have 𝐵 = of
Q
discriminant 23 and a maximal order O with three ideal classes [𝐼1 ], [𝐼2 ], [𝐼3 ]. In
(17.6.5), we found three ideals in 𝐼1 = O: two belong to the class [𝐼2 ] and the third
is principal, belonging to [𝐼1 ]. This gives the first column of the matrix as (1, 2, 0) t .
Computing further, we find
1 1 0
𝑇 (2) = 2 1 3® .
© ª
«0 1 0¬
In a similar manner, we compute
0 1 3 30 28 24
𝑇 (3) = 2 3 0® , 𝑇 (101) = 56 54 60® .
© ª © ª
«2 0 1¬ «16 20 18¬
41.1.3. There is a second and computationally more efficient way to define the Brandt
matrix using representation numbers of quadratic forms. Let 𝑞 𝑖 = nrd(𝐼𝑖 ), let O𝑖 =
OL (𝐼𝑖 ), and let 𝑤 𝑖 = #O×𝑖 /{±1} < ∞. Then
1
𝑇 (𝑛)𝑖 𝑗 = #{𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 : nrd(𝛼)𝑞 𝑖 /𝑞 𝑗 = 𝑛} :
2𝑤 𝑖
indeed, 𝛼𝐼𝑖 = 𝐽 ⊆ 𝐼 𝑗 with nrd(𝐽) = 𝑛 nrd(𝐼 𝑗 ) if and only if 𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 = (𝐼 𝑗 : 𝐼𝑖 ) L and

nrd(𝛼)𝑞 𝑖 = 𝑝𝑞 𝑗 , and 𝛼 is well defined up to right multiplication by 𝜇 ∈ O×𝑖 . Now
𝑄 𝑖 𝑗 : 𝐼 𝑗 𝐼𝑖−1 → Z
𝑞𝑖 (41.1.4)
𝑄 𝑖 𝑗 (𝛼) = nrd(𝛼)
𝑞𝑗
is a positive definite quadratic form, so it suffices to enumerate lattice points!
Example 41.1.5. Returning to our example, we have
1
𝑇 (𝑛)𝑖𝑖 = #{𝛾 ∈ O𝑖 : nrd(𝛾) = 𝑛}.
2𝑤 𝑖
For 𝑖 = 1, we have 𝑤 1 = 2 and with 𝛾 = 𝑡 + 𝑥𝛼 + 𝑦𝛽 + 𝑧𝛼𝛽 and 𝑡, 𝑥, 𝑦, 𝑧 ∈ Z, by (17.6.6)
nrd(𝛾) = 𝑡 2 + 𝑡𝑦 + 𝑥 2 + 𝑥𝑧 + 6𝑦 2 + 6𝑧2
so 𝑇 ( 𝑝)11 counts half the number of representations of 𝑛 by this positive definite

quaternary quadratic form.
There is a third way to understand Brandt matrices which is visual and combina-
torial.
41.1. ⊲ BRANDT MATRICES, NEIGHBORS, AND MODULAR FORMS 749
Definition 41.1.6. Let 𝐼, 𝐽 ⊆ O be invertible right O-ideals. We say 𝐽 is a 𝑛-neighbor

of 𝐼 if 𝐽 ⊆ 𝐼 and 𝑛 nrd(𝐼) = nrd(𝐽).
The 𝑛-Brandt graph is the directed graph with vertices Cls O and a directed edge
from [𝐼𝑖 ] to [𝐽] for each 𝑛-neighbor 𝐽 ⊆ 𝐼𝑖 .
There is no extra content here, just a reinterpretation: the 𝑛-Brandt matrix is simply
the adjacency matrix of the 𝑛-Brandt graph.
Example 41.1.7. Returning a third time to our example, we have the 2-Brandt graph,
I1 I2
I3
Figure 41.1.8: The 2-Brandt graph for discriminant 23

41.1.9. For 𝑛 = 𝑝 prime, there is another way to think of the 𝑝-Brandt graph. Consider
the directed graph whose vertices are invertible right O-ideals whose reduced norm is
a power of 𝑝, and draw a directed edge from 𝐼 to 𝐽 if 𝐽 is a 𝑝-neighbor of 𝐼. If 𝑝 - 𝑁,
then this graph is a (𝑝 + 1)-regular directed tree, that is to say, from each vertex there
are 𝑝 + 1 directed edges. The notion of belonging to the same ideal class induces an
equivalence relation on this graph, and the quotient is the 𝑝-Brandt graph.
It is helpful to think of the matrices 𝑇 (𝑛) as operators on a space, so we define the
Brandt module 𝑀2 (O) to be the C-vector space with basis Cls O and equipped with
the action of Brandt matrices 𝑇 (𝑛) for 𝑛 ∈ Z ≥0 on the right.
The Brandt matrices have two important properties. First, they commute: by a
quaternionic version of the Sun Zi theorem (CRT), if gcd(𝑚, 𝑛) = 1 then
𝑇 (𝑚)𝑇 (𝑛) = 𝑇 (𝑛)𝑇 (𝑚). (41.1.10)
Informally, we might say that the process of taking 𝑚-neighbors commutes with the
process of taking 𝑛-neighbors, when 𝑚, 𝑛 are coprime. Second, they are self-adjoint
for the pairing on 𝑀2 (O) given by
(
1/𝑤 𝑖 , if 𝑖 = 𝑗;
h[𝐼𝑖 ], [𝐼 𝑗 ]i =
0, else.
The proof of self-adjointness is contained in the equality 𝑤 𝑖 𝑇 (𝑛)𝑖 𝑗 = 𝑤 𝑗 𝑇 (𝑛) 𝑗𝑖 , and

this follows from a bĳection induced by the standard involution.
Therefore the matrices 𝑇 (𝑛) are semisimple (diagonalizable) and 𝑀2 (O) has a
simultaneous basis of eigenvectors, which we call a eigenbasis. The row 𝑒 =
(1, 1, . . . , 1) is always an eigenvector (by the sum of columns) with eigenvalue
𝑎 𝑝 (𝑒) = 𝑝 + 1 for 𝑝 - 𝑁.
Example 41.1.11. We check that 𝑇 (2)𝑇 (3) = 𝑇 (3)𝑇 (2) from Example 41.1.2; and
𝑤 1 , 𝑤 2 , 𝑒 3 = 2, 1, 3, so we verify that
2 0 0 2 2 0
0 1 0® 𝑇 (2) = 2 1 3®
© ª © ª
«0 0 3¬ «0 3 0¬
is symmetric. The characteristic polynomial of 𝑇 (2) is (𝑥−3) (𝑥 2 +𝑥−1),

√ and for√𝑇 (3) it
is (𝑥 −4) (𝑥 2 −5). We find the eigenbasis 𝑒 = (1, 1, 1) and 𝑓± = (4, ± 5−3, ∓3 5+3),
and observe the required orthogonality
1 √ 1 √
h𝑒, 𝑓± i = · 4 + (± 5 − 3) + (∓3 5 + 3) = 0.
2 3
By now, hopefully the reader is convinced that the Brandt matrices capture in-
teresting arithmetic information about the order O and that they are not difficult to
compute.
Now comes the modular forms: the second way of viewing Brandt matrices shows
that we should be thinking of a generating series for the representation numbers of the
quadratic forms 𝑄 𝑖 𝑗 defined in (41.1.4). As in section 40.4, we define the theta series
for the quadratic form 𝑄 𝑖 𝑗
∞
∑︁ 1 ∑︁
Θ𝑖 𝑗 (𝑞) := 𝑇 (𝑛)𝑖 𝑗 𝑞 𝑛 = 𝑞 𝑄𝑖 𝑗 (𝛾) ∈ Z[[𝑞]].
2𝑤 𝑖
𝑛=0 𝛾 ∈𝐼 𝑗 𝐼𝑖−1
Letting 𝑞 := 𝑒 2 𝜋𝑖𝑧 for 𝑧 ∈ H2 , by Theorem 40.4.4 (a consequence of Poisson sum-

mation), the function Θ𝑖 𝑗 (𝑧) : H2 → C is a modular form of weight 2 for an explicit
congruence subgroup of SL2 (Z): for example, if O is an Eichler order with reduced
discriminant 𝑁, then 𝜃 𝑖 𝑗 (𝑞) ∈ 𝑀2 (Γ0 (𝑁)) (trivial character).
There is something enduringly magical about the fact that the entries of Brandt ma-
trices (arithmetic) give Fourier coefficients of holomorphic modular forms (geometric,
analytic).
Example 41.1.12. Returning one last time to our example, the space 𝑀2 (Γ0 (23)) of
modular forms of weight 2 and level Γ0 (23) has eigenbasis 𝑒 23 , 𝑓+ , 𝑓− where
∞
11 ∑︁ ∗ ∑︁
𝑒 23 (𝑧) = + 𝜎 (𝑛)𝑞 𝑛 , 𝜎 ∗ (𝑛) = 𝑑
12 𝑛=1
𝑑 |𝑛
23-𝑑
and √
± 5+1 2 √ 3
𝑓± (𝑧) := 𝑞 − 𝑞 + 5𝑞 + . . .
2
are cusp forms matching the eigenbasis in Example 41.1.11.
One of the main applications of Brandt matrices is to express the trace of the Hecke
operator in terms of arithmetic data, as follows.
41.2. BRANDT MATRICES 751
Theorem 41.1.13. Let 𝐵 be a definite quaternion algebra over Q of discriminant 𝐷

and O ⊆ 𝐵 a maximal order. For 𝑑 < 0 a fundamental discriminant, define

ℎ(𝑑) Ö 𝑑
ℎ 𝐷 (𝑑) := 1−
𝑤𝑑 𝑝
𝑝 |𝐷
√

𝑑
where ℎ(𝑑) is the class number of Q( 𝑑), 𝑤 𝑑 its number of roots of unity, and is
𝑝
the Kronecker symbol. For 𝑑 0 < 0 a discriminant with 𝑑 0 = 𝑑𝑓 2 and 𝑑 fundamental,
define ℎ 𝐷 (𝑑 0) = ℎ 𝐷 (𝑑). Then the trace of the 𝑛th Brandt matrix associated to O is
(
∑︁
2 𝜑(𝐷)/12, if 𝑛 is a square;
tr 𝑇 (𝑛) = ℎ 𝐷 (𝑡 − 4𝑛) +
𝑡 ∈Z
0, otherwise
𝑡 2 <4𝑛
where 𝜑 is the Euler totient function.

Theorem 41.1.13 is a special case of Main Theorem 41.5.2, see Example 41.5.8.
In a surprising way, it exhibits a relationship between traces of Hecke operators and
(modified) class numbers of imaginary quadratic fields!
41.2 Brandt matrices

To begin, we define the all-important Brandt matrices in the generality considered in
this text.
Let 𝑅 be a global ring with eligible set S ⊆ Pl 𝐹, let 𝐵 be an S-definite quaternion
algebra over 𝐹 of discriminant disc 𝐵 = 𝔇 and let O ⊂ 𝐵 be an 𝑅-order in 𝐵 with
reduced discriminant discrd O = 𝔑. The reader will probably have in mind the case
where 𝐹 is a totally real (number) field, S the set of real (archimedean) places of 𝐹,
and 𝐵 a definite quaternion algebra over 𝐹; but the arguments hold just as well with a
larger set S of ramified places or in the function field case.
41.2.1. Let Cls O be the right class set of O. By Corollary 27.6.20, the class number
ℎ = # Cls O < ∞ is finite. Let 𝐼1 , . . . , 𝐼 ℎ be a set of representative invertible right
O-ideals for Cls O. For 𝑖 = 1, . . . , ℎ, let O𝑖 = OL (𝐼𝑖 ) be the left order of 𝐼𝑖 ; then O𝑖
depends on the choice of 𝐼𝑖 but its isomorphism class (i.e., type) is independent of the
choice of 𝐼𝑖 . (Due to the possible presence of two-sided ideals, there may be repetition
of types among the orders O𝑖 .)
41.2.2. Let 𝔫 ⊂ 𝑅 be a nonzero ideal. For each 𝑗, we consider the set of right invertible
O-ideals 𝐽 ⊆ 𝐼 𝑗 with nrd(𝐽) = 𝔫 nrd(𝐼 𝑗 ), and we count them according to their class
in Cls O:
𝑇 (𝔫)𝑖 𝑗 := #{𝐽 ⊆ 𝐼 𝑗 : nrd(𝐽) = 𝔫 nrd(𝐼 𝑗 ) and [𝐽] = [𝐼𝑖 ]} ∈ Z ≥0 . (41.2.3)
A containment 𝐽 ⊆ 𝐼 𝑗 of right O-ideals yields a compatible product 𝐽 0 = 𝐽 𝐼 −1𝑗 and

thus an invertible right O 𝑗 -ideal with reduced norm nrd(𝐽 0) = nrd(𝐽 𝐼 −1
𝑗 ) = 𝔫, and
conversely. So equivalently
𝑇 (𝔫)𝑖 𝑗 = #{𝐽 0 ⊆ O 𝑗 : nrd(𝐽 0) = 𝔫 and [𝐽 0 𝐼 𝑗 ] = [𝐼𝑖 ]}.
(We could also rewrite [𝐽 0 𝐼 𝑗 ] = [𝐼𝑖 ] in terms of classes of right ideals of O 𝑗 .)

Definition 41.2.4. The 𝔫-Brandt matrix for O is the matrix 𝑇 (𝔫) ∈ Mℎ (Z) whose
(𝑖, 𝑗)th entry is equal to 𝑇 (𝔫)𝑖, 𝑗 .
41.2.5. To make the definition more canonical, we define
𝑀2 (O) := Map(Cls O, Z)
to be the set of maps from Cls O to Z (as sets). Then 𝑀2 (O) has the structure of an
abelian group under addition of maps, and it is a free Z-module on the characteristic
functions for Cls O. The 𝔫-Hecke operator is defined to be
𝑇 (𝔫) : 𝑀2 (O) → 𝑀2 (O)

∑︁
(𝑇 (𝔫) 𝑓 ) ( [𝐼]) = 𝑓 ( [𝐽]) (41.2.6)
𝐽 ⊆𝐼
nrd( 𝐽 )=𝔫 nrd(𝐼 )
again the sum over all invertible right O-ideals 𝐽 ⊆ 𝐼 with condition on the reduced
norm. Visibly, this definition does not depend on the choice of representative 𝐼 in its
right ideal class. And in the basis of characteristic functions for 𝐼𝑖 , the matrix of 𝑇 (𝔫)
is precisely the 𝔫-Brandt matrix.
Brandt matrices may be given in terms of elements instead of ideals. Let 𝑤 𝑖 =
[O×𝑖 : 𝑅 × ]. By Proposition 32.3.7, since 𝐵 is S-definite, the unit index 𝑤 𝑖 < ∞ is
finite.
Lemma 41.2.7. Let 𝔫𝑖 𝑗 = 𝔫 nrd(𝐼 𝑗 )/nrd(𝐼𝑖 ) for 𝑖, 𝑗 = 1, . . . , ℎ. Then following
statements hold.
(a) We have
𝑇 (𝔫)𝑖 𝑗 = # 𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 : nrd(𝛼)𝑅 = 𝔫𝑖 𝑗 /O×𝑖

1 (41.2.8)
# 𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 : nrd(𝛼)𝑅 = 𝔫𝑖 𝑗 /𝑅 ×

=
𝑤𝑖
where we count orbits under right multiplication by O×𝑖 and 𝑅 × , respectively.
(b) If the class of 𝔫𝑖 𝑗 in Cl+ 𝑅 is nontrivial, then 𝑇 (𝔫)𝑖 𝑗 = 0.
× totally positive. Then
(c) Suppose that 𝔫𝑖 𝑗 = 𝑛𝑖 𝑗 𝑅 with 𝑛𝑖 𝑗 ∈ 𝐹>0
1 ∑︁
# 𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 : nrd(𝛼) = 𝑢𝑛𝑖 𝑗

𝑇 (𝔫)𝑖 𝑗 = (41.2.9)
2𝑤 𝑖 × /𝑅 ×2
𝑢𝑅 ×2 ∈𝑅>0
where the sum is over a choice of representatives of totally positive units of 𝑅

modulo squares.
41.2. BRANDT MATRICES 753
Proof. We claim that there is a bĳection
{𝐽 ⊆ 𝐼 𝑗 : nrd(𝐽) = 𝔫 nrd(𝐼 𝑗 ) and [𝐽] = [𝐼𝑖 ]}

(41.2.10)
↔ 𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 : nrd(𝛼) = 𝔫𝑖 𝑗 /O×𝑖

with orbits under right multiplication by O×𝑖 = OR (𝐼𝑖−1 ) × . Indeed, a containment

𝐽 ⊆ 𝐼 𝑗 of invertible right O-ideals with [𝐼𝑖 ] = [𝐽] corresponds to 𝛼 ∈ 𝐵× such that
𝛼𝐼𝑖 = 𝐽, so in fact 𝛼 ∈ (𝐽 : 𝐼𝑖 ) L = 𝐽 𝐼𝑖−1 , and nrd(𝐽) = 𝔫 nrd(𝐼 𝑗 ) translates into
nrd(𝛼) nrd(𝐼𝑖 ) = nrd(𝐽) = 𝔫 nrd(𝐼 𝑗 ) or nrd(𝛼)𝑅 = 𝔫𝑖 𝑗 . Writing 𝐽 𝐼𝑖−1 = 𝛼O𝑖 , we see
that 𝛼 is unique up to multiplication on the right by O×𝑖 . To finish (a), we note that
the right action by O×𝑖 is free; and (b) follows from (a), since nrd(𝐵× ) ≤ 𝐹>0 × as 𝐵 is
S-definite.
For (c), we just need to organize our generators; the sum in (41.2.9) is finite by
the Dirichlet S-unit theorem. If nrd(𝛼)𝑅 = 𝔫𝑖 𝑗 then nrd(𝛼) = 𝑢𝑛𝑖 𝑗 for some 𝑢 ∈ 𝑅>0 × .
Multiplying by an element of 𝑅>0 × , we may suppose that nrd(𝛼)/𝑛 = 𝑢 belongs in a set

𝑖𝑗
× /𝑅 ×2 , and for 𝑣 ∈ 𝑅 × , we have nrd(𝑣𝛼) = 𝑣 2 nrd(𝛼) = nrd(𝛼)
of representatives for 𝑅>0
if and only if 𝑣 = ±1, which gives us an extra factor 2.
41.2.11. The advantage of (41.2.9) is that it can be expressed simply in terms of

a quadratic form. Suppose that 𝐹 is a number field and 𝑅 = Z𝐹 , when this ob-
servation is especially clean. Since 𝐵 is totally definite, as in 17.7.10, the space
𝐵 ↩→ 𝐵 ⊗Q R H𝑛 R4𝑛 comes equipped with the positive definite quadratic form
𝑄 = Tr𝐹 /Q nrd : 𝐵 → R, and if 𝐽 is a 𝑅-lattice, then 𝐽 Z4𝑛 embeds as a Euclidean
lattice 𝐽 ↩→ R4𝑛 with respect to this quadratic form. Therefore,
{𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 : nrd(𝛼) = 𝑢𝑛𝑖 𝑗 } ⊆ {𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 : 𝑄(𝛼) = Tr𝐹 /Q 𝑢𝑛𝑖 𝑗 }
where the latter set is finite and effectively computable.
Finally, Brandt matrices are adjacency matrices.
Definition 41.2.12. Let 𝐼, 𝐽 ⊆ O be invertible right O-ideals. We say 𝐽 is a 𝔫-neighbor

of 𝐼 if 𝐽 ⊆ 𝐼 and nrd(𝐽) = 𝔫 nrd(𝐼).
The 𝔫-Brandt graph is the directed graph with vertices Cls O and a directed edge
from [𝐼𝑖 ] to [𝐽] for each 𝔫-neighbor 𝐽 ⊆ 𝐼.
By definition, the adjacency matrix of the 𝔫-Brandt graph is the 𝔫-Brant matrix
𝑇 (𝔫).
41.2.13. Let 𝔭 - 𝔑 be prime and suppose that the class of 𝔭 generates Cl𝐺 (O) 𝑅. Then
by Proposition 28.5.18, we may take the ideals 𝐼𝑖 to have reduced norm a power of
𝔭. Consider the directed graph whose vertices are the right O-ideals whose reduced
norm is a power of 𝔭 with directed edges for each 𝔭-neighbor relation. (This graph is
a regular directed tree by Proposition 41.3.1 below, every vertex has out degree equal
to N𝔭 + 1.) The equivalence relation of belonging to the same right ideal class (left
equivalent by an element of 𝐵× ) respects edges, and the quotient by this equivalence
relation is the 𝔭-Brandt graph.
Remark 41.2.14. The Brandt graphs have interesting graph theoretic properties: they
are Ramanujan graphs (also called expander graph), having high connectivity and
are potentially useful in communication networks. In the simplest case where 𝐹 = Q
and 𝐵 is the quaternion algebra of discriminant 𝑝, they were first studied by Ihara,
then studied in specific detail by Lubotzky–Phillips–Sarnak [LPS88] and Margulis
[Marg88]; for further reading, see the books by Lubotzky [Lub2010] and Sarnak
[Sar90]. Over totally real fields, see work of Livné [Liv2001] as well as Charles–
Goren–Lauter [CGL2009]. The proof that Brandt graphs are Ramanujan relies on the
Ramanujan–Petersson conjecture, a deep statement proven by Deligne [Del74], giving
bounds on coefficients of modular forms.
Remark 41.2.15. The space of functions on Cls O can itself be understood as a space
of modular forms, a special case of the theory of algebraic modular forms due to Gross
[Gro99]. This general formulation harmonizes with the double coset description given
in section 38.7, via the canonical bĳection Cls O ↔ 𝐵× \ 𝐵 b× /O
b× , but without the
geometry!
41.3 Commutativity of Brandt matrices

In this section, we examine basic properties of Brandt matrices—including that they
commute.
Í
(a) The sum of the entries 𝑖 𝑇 (𝔫)𝑖 𝑗 in every column of 𝑇 (𝔫) is constant; if 𝔫 is
Í
coprime to 𝔑, then this constant is equal to 𝔡 |𝔫 N (𝔡), where N (𝔞) = #(𝑅/𝔞) is
the absolute norm.
(b) If 𝔪, 𝔫 are relatively prime, then
𝑇 (𝔪𝔫) = 𝑇 (𝔪)𝑇 (𝔫) = 𝑇 (𝔫)𝑇 (𝔪). (41.3.2)
Proof. First we prove (a). The orders O 𝑗 are locally isomorphic, so by the local-
global dictionary for lattices (Theorem 9.4.9), the number of invertible right O 𝑗 -ideals
with given reduced norm is independent of 𝑗, giving the first statement. For the
second statement, under the hypothesis that 𝔫 is coprime to 𝔑, for all 𝔭 | 𝔫 we
have O𝔭 ' M2 (𝑅𝔭 ), and we counted right ideals in our pursuit of the zeta function:
by Proposition 26.3.9, these counts are multiplicative, and by Lemma 26.4.1(b), the
number of reduced norm 𝔭𝑒 is 1 + 𝑞 + · · · + 𝑞 𝑒 where 𝑞 = N (𝔭).
Statement (b) follows similar logic but with “unique factorization” of right ideals.
As above, an invertible right O 𝑗 -ideal of reduced norm 𝔪𝔫 by Lemma 26.3.6 factors
uniquely into a compatible product of invertible lattices of reduced norm 𝔪 and 𝔫:
organizing by classes, this says precisely that
ℎ
∑︁
𝑇 (𝔪𝔫)𝑖 𝑗 = 𝑇 (𝔪)𝑖𝑘 𝑇 (𝔫) 𝑘 𝑗 (41.3.3)
𝑘=1
which gives the matrix product 𝑇 (𝔪𝔫) = 𝑇 (𝔪)𝑇 (𝔫). Repeating the argument inter-
changing the roles of 𝔪 and 𝔫, the result is proven.
41.3. COMMUTATIVITY OF BRANDT MATRICES 755
For prime powers coprime to 𝔑, we have a recursion for the 𝔭𝑟 -Brandt matrices
that is a bit complicated: the uniqueness of factorization fails when the product is a
two-sided ideal, so we must account for this extra term. To this end, we need to keep
track of the effect of multiplication by right ideals of 𝑅 on the class set.
41.3.4. For an ideal 𝔞 ⊆ 𝑅, let 𝑃(𝔞) ∈ Mℎ (Z) be the permutation matrix given by
𝐼𝑖 ↦→ 𝔞𝐼𝑖 . In other words, we place a 1 in the (𝑖, 𝑗)th entry according as [𝔞𝐼 𝑗 ] = [𝐼𝑖 ]
(with 0 elsewhere). The matrix 𝑃(𝔞) only depends on the class [𝔞] ∈ Cl 𝑅: in
particular, if 𝔞 is principal then 𝑃(𝔞) is the identity matrix. Therefore we have a
homomorphism
𝑃 : Cl 𝑅 → GLℎ (Z)
[𝔞] ↦→ 𝑃(𝔞).
We have
𝑃(𝔞𝔟) = 𝑃(𝔞)𝑃(𝔟) = 𝑃(𝔟)𝑃(𝔞)
and in particular 𝑃(𝔞)𝑃(𝔞−1 )
= 1 and the image 𝑃(Cl 𝑅) ⊆ GLℎ (Z) is an abelian
subgroup; however, this map need not be injective. Moreover, for all 𝔞, 𝔫 we have
𝑃(𝔞)𝑇 (𝔫) = 𝑇 (𝔫)𝑃(𝔞) (41.3.5)
by commutativity of multiplication by 𝔞.
As in 26.4.3, we say an integral right O-ideal 𝐼 is primitive if we cannot write
𝐼 = 𝔞𝐼 0 with 𝐼 0 integral and 𝔞 ( 𝑅.
Proposition 41.3.6. Let 𝔭 - 𝔑 be prime. Then for 𝑟, 𝑠 ∈ Z ≥0 ,
min(𝑟
∑︁,𝑠)
𝑇 (𝔭𝑟 )𝑇 (𝔭𝑠 ) = N (𝔭) 𝑖 𝑇 (𝔭𝑟 +𝑠−2𝑖 )𝑃(𝔭) 𝑖 . (41.3.7)
𝑖=0
In particular, for all 𝑟 ≥ 0,
𝑇 (𝔭𝑟 +2 ) = 𝑇 (𝔭𝑟 +1 )𝑇 (𝔭) − N (𝔭)𝑇 (𝔭𝑟 )𝑃(𝔭). (41.3.8)
Proof. When 𝑠 = 0, the matrix 𝑇 (1) is the identity and the result holds. We next
consider the case 𝑠 = 1, and will then proceed by induction, and consider the product
𝑇 (𝔭𝑟 )𝑇 (𝔭): its 𝑖 𝑗th entry
ℎ
∑︁
𝑟
(𝑇 (𝔭 )𝑇 (𝔭))𝑖 𝑗 = 𝑇 (𝔭𝑟 )𝑖𝑘 𝑇 (𝔭) 𝑘 𝑗
𝑘=1
counts the number of compatible products of right ideals 𝐽𝑟0 𝐽 0 where 𝐽𝑟0 is an invertible
O𝑖 , O 𝑘 -ideal with nrd(𝐽𝑟0 ) = 𝔭𝑟 and 𝐽 0 is an invertible O 𝑘 , O 𝑗 -ideal with nrd(𝐽 0) = 𝔭.
The issue: these products may not all be distinct when they are imprimitive. If the
product 𝐽𝑟0 𝐽 0 is imprimitive, then we rewrite it as a compatible product
0 0
𝐽𝑟0 𝐽 0 = (𝐽𝑟0 (𝐽 ) −1 )𝐽 0 𝐽 0 = 𝔭(𝐽𝑟0 (𝐽 ) −1 )
where now 𝐽𝑟0 −1 = 𝔭−1 𝐽𝑟0 𝐽 0 has reduced norm 𝔭𝑟 −1 . This procedure works in reverse
as well.
With apologies for the temporarily annoying notation, define 𝑇prim (𝔭𝑟 +1 ) and
𝑇imprim (𝔭𝑟 +1 ) to be the 𝔭𝑟 -Brandt matrix counting classes of primitive or imprimi-
tive, accordingly. Then
𝑇 (𝔭𝑟 +1 ) = 𝑇prim (𝔭𝑟 +1 ) + 𝑇imprim (𝔭𝑟 +1 ). (41.3.9)
Under multiplication by 𝔭, we have
𝑇imprim (𝔭𝑟 +1 ) = 𝑇 (𝔭𝑟 −1 )𝑃(𝔭). (41.3.10)
Since there are N (𝔭) + 1 right O-ideals of reduced norm 𝔭, with the previous paragraph
we obtain
𝑇 (𝔭𝑟 )𝑇 (𝔭) = 𝑇prim (𝔭𝑟 +1 ) + ( N (𝔭) + 1)𝑇imprim (𝔭𝑟 +1 )
= 𝑇 (𝔭𝑟 +1 ) + N (𝔭)𝑇imprim (𝔭𝑟 +1 ) (41.3.11)
𝑟 +1 𝑟 −1
= 𝑇 (𝔭 ) + N (𝔭)𝑇 (𝔭 )𝑃(𝔭).
This proves the result for 𝑠 = 1, and it gives (41.3.8) upon rearrangement and shifting
indices.
We now proceed by (an ugly but harmless) induction on 𝑠:

𝑇 (𝔭𝑟 ) 𝑇 (𝔭𝑠+1 ) + N (𝔭)𝑇 (𝔭𝑠−1 )𝑃(𝔭)
𝑠
∑︁
= N (𝔭) 𝑖 𝑇 (𝔭𝑟 +𝑠+1−2𝑖 )𝑃(𝔭) 𝑖 (41.3.12)
𝑖=0

+N (𝔭) 𝑖+1 𝑃(𝔭𝑟 +𝑠+1−2(𝑖+1) )𝑃(𝔭) 𝑖+1
so
𝑠
∑︁
𝑇 (𝔭𝑟 )𝑇 (𝔭𝑠+1 ) = N (𝔭) 𝑖 𝑇 (𝔭𝑟 +𝑠+1−2𝑖 𝑃(𝔭) 𝑖 )
𝑖=0

+N (𝔭) 𝑖+1 𝑃(𝔭) 𝑖+1 𝑃(𝔭𝑟 +𝑠+1−2(𝑖+1) )
𝑠
∑︁ (41.3.13)
− N (𝔭) 𝑖+1 𝑃(𝔭) 𝑖+1𝑇 (𝔭𝑟 +𝑠+1−2(𝑖+1) )
𝑖=0
𝑠+1
∑︁
= N (𝔭) 𝑖 𝑃(𝔭) 𝑖 𝑇 (𝔭𝑟 +𝑠+1−2𝑖 )
𝑖=0
as claimed.
Definition 41.3.14. The Hecke algebra T(O) is the subring of Mℎ (Z) generated by
the matrices 𝑇 (𝔫) with 𝔫 coprime to 𝔑.
Corollary 41.3.15. The ring T(O) is a commutative Z-algebra.
41.4. SEMISIMPLICITY 757
Proof. By Proposition 41.3.1(b), we reduce to showing that 𝑇 (𝔭𝑟 )𝑇 (𝔭𝑠 ) = 𝑇 (𝔭𝑠 )𝑇 (𝔭𝑟 )
for all 𝑟, 𝑠 ≥ 0, and this holds by Proposition 41.3.6: the right-hand side of (41.3.7) is
symmetric under interchanging 𝑟, 𝑠.
√ √
Example 41.3.16. Let 𝐹 = Q( 10) and 𝑅 = Z𝐹 = Z[ 10] its ring of integers.
Then the√class group Cl 𝑅 ' Z/2Z is nontrivial, represented by the class of the ideal
+
√ and the narrow class group Cl 𝑅 ' Cl 𝑅 is no bigger: the fundamental
𝔭2 = (2, 10),
unit is 3 + 10 of norm −1.
Let 𝐵 = (−1, −1 | 𝐹). Since 2 is not split in 𝐹, the ramification set Ram 𝐵 is the
set of real places of 𝐹. A maximal order is given by
1 + 𝑖 + 𝑗 + 𝑖𝑗
O = 𝑅 ⊕ 𝔭−1 −1
2 (1 + 𝑖) ⊕ 𝔭2 (1 + 𝑗) ⊕ 𝑅 .
2
We find that # Cls O = 4, and
0 0 0 1
0 0 3 2®
© ª
𝑇 (𝔭2 ) = ®.
0 2 0 0®
«3 1 0 0¬
In this case, the matrix 𝑃(𝔭2 ) is the identity matrix: for example, we have 𝔭2 O =
(1 + 𝑖)O. Thus
1 1 0 0
6 6 0 0®
© ª
𝑇 (𝔭22 ) 2
= 𝑇 (2𝑅) = 𝑇 (𝔭2 ) − 2 = ®.
0 0 4 4®
«0 0 3 3¬
41.4 Semisimplicity
We now equip the space 𝑀2 (O) = Map(Cls O, Z) with a natural inner product, and
we show that the Hecke operators are normal with respect to this inner product.
41.4.1. For [𝐼] ∈ Cls O, we define 𝑤 [𝐼 ] := [OL (𝐼) × : 𝑅 × ]; this is well-defined, as a

different choice of representative gives an isomorphic (conjugate) order.
Let 1 [𝐼 ] be the characteristic function of [𝐼] ∈ Cls O; then 1 [𝐼 ] for [𝐼] ∈ Cls O
form a basis for 𝑀2 (O). We define the bilinear form
h , i : 𝑀2 (O) × 𝑀2 (O) → Z
(41.4.2)
h1 [𝐼 ] , 1 [𝐽 ] i := 𝑤 [𝐼 ] 𝛿 [𝐼 ], [𝐽 ]
where 𝛿 [𝐼 ], [𝐽 ] = 1, 0 according as [𝐼] = [𝐽] or not, and extend linearly. The matrix
of this pairing in the basis of characteristic functions is the diagonal matrix diag(𝑤 𝑖 )𝑖 ,
where 𝑤 𝑖 = [O×𝑖 : 𝑅 × ]. The pairing is symmetric and nondegenerate.
The inner product (41.4.2) defines an adjoint 𝑇 ↦→ 𝑇 ∗ .

𝑃(𝔫) ∗ = 𝑃(𝔫−1 ) (41.4.4)

∗ −1
𝑇 (𝔫) = 𝑃(𝔫 )𝑇 (𝔫). (41.4.5)
The Hecke operators 𝑇 (𝔫) are normal with respect to the inner product (41.4.2), and
for 𝔫 trivial in Cl+ 𝑅 the operators 𝑇 (𝔫) are self-adjoint.
Proof. We may show the proposition for the Brandt matrices. Let 𝑊 = diag(𝑤 𝑖 )𝑖 define
the inner product on Zℎ with the Brandt matrices acting on the right on row vectors.
Then the inner product is h𝑥, 𝑦i = 𝑥𝑊 𝑦 t and accordingly the adjoint h𝑥𝑇, 𝑦i = h𝑥, 𝑇 ∗ 𝑦i
is defined by
𝑇 ∗ = 𝑊 −1𝑇 t𝑊 (41.4.6)
The transpose of a permutation matrix is its inverse and that OL (𝔫𝐼𝑖 ) = OL (𝐼𝑖 ), so
that the unit groups match up, whence
𝑃(𝔫) ∗ = 𝑃(𝔫) −1 = 𝑃(𝔫−1 ) (41.4.7)
giving (41.4.4).
For the Brandt matrices, we refer to Lemma 41.2.7(a), giving
1
# 𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 : nrd(𝛼)𝑅 = 𝔫𝑖 𝑗 /𝑅 ×

𝑇 (𝔫)𝑖 𝑗 =
𝑤𝑖
where 𝔫𝑖 𝑗 = 𝔫 nrd(𝐼 𝑗 )/nrd(𝐼𝑖 ). Let
Θ(𝔫)𝑖 𝑗 = 𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 : nrd(𝛼)𝑅 = 𝔫𝑖 𝑗 /𝑅 × .

By (41.4.6),
𝑊𝑇 (𝔫) = (Θ(𝔫)𝑖 𝑗 )𝑖, 𝑗 =: Θ(𝔫).
We extend the definition of Θ(𝔫) to include all fractional ideals 𝔫. For each 𝑖, let
𝑖 0 be such that [𝔫−1 𝐼𝑖 ] = [𝐼𝑖0 ], so that 𝔫−1 𝐼𝑖 = 𝛽𝑖 𝐼𝑖0; the induced action is given by the
permutation map 𝑃(𝔫−1 ).
Θ(𝔫)𝑖 𝑗 → Θ(𝔫) 𝑗𝑖0

(41.4.8)
𝛼 ↦→ (𝛼𝛽𝑖 ) −1 = 𝛽𝑖−1 𝛼−1
is well-defined and bĳective.

Indeed, if 𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 then
nrd(𝐼 𝑗 )
𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 = 𝐼𝑖−1 𝐼 𝑗 = 𝐼𝑖 𝐼 −1
𝑗 (41.4.9)
nrd(𝐼𝑖 )
since 𝐼 𝐼 = nrd(𝐼) for an invertible 𝑅-lattice 𝐼, so
𝛼−1 ∈ 𝔫−1 𝐼𝑖 𝐼 −1 −1
𝑗 = 𝛽𝑖 𝐼 𝑖 0 𝐼 𝑗
41.5. EICHLER TRACE FORMULA 759
and therefore 𝛽𝑖−1 𝛼−1 ∈ 𝐼𝑖0 𝐼 −1

𝑗 as claimed. And nrd(𝛼) = 𝔫𝑖 𝑗 implies nrd(𝛼 ) =
−1
𝔫−2 𝔫 𝑗𝑖 so nrd(𝛽𝑖−1 𝛼−1 ) = 𝔫 𝑗𝑖0 . We can run the argument in the other direction to
produce an inverse, and we thereby conclude the map is bĳective.
The map (41.4.8) together with the action by permutation and 𝑊 t = 𝑊 yields
𝑊𝑇 (𝔫) = Θ(𝔫) = 𝑃(𝔫−1 )Θ(𝔫) t = 𝑃(𝔫−1 )𝑊𝑇 (𝔫) ∗
and thus 𝑇 (𝔫) ∗ = 𝑃(𝔫) ∗𝑇 (𝔫), and substituting (41.4.7) gives (41.4.5).
For the final statement, by (41.3.5) we have 𝑇 (𝔫) commuting with 𝑃(𝔫), so 𝑇 (𝔫)
commutes with 𝑇 (𝔫) ∗ ; and when 𝔫 is narrowly principal, then 𝑃(𝔫−1 ) is the identity
matrix so 𝑇 (𝔫) ∗ = 𝑇 (𝔫).
By the spectral theorem in linear algebra, we have the following corollary.
Corollary 41.4.10. T(O) is a semisimple commutative ring, and there exists a basis
of common eigenvectors (eigenfunctions) for the Hecke operators. Each 𝑇 (𝔫) with 𝔫
narrowly principal has real eigenvalues.
41.5 Eichler trace formula

In this section, we compute the trace of the Brandt matrices in terms of embedding
numbers. We continue notation from the previous section.
We begin by recalling the main ingredients. Let 𝐾 ⊃ 𝐹 be a separable quadratic
field extension and let 𝑆 ⊆ 𝐾 be a quadratic 𝑅-order. Let ℎ(𝑆) = # Pic 𝑆. Let
𝑚(𝑆, O, O× ) be the number of O× -conjugacy classes of optimal embeddings 𝑆 ↩→ O.
Then by Theorem 30.4.7,
∑︁
𝑚(𝑆, OL (𝐼); OL (𝐼) × ) = ℎ(𝑆)𝑚( 𝑆,
b O; b× )
b O (41.5.1)
[𝐼 ] ∈Cls O
We also recall
ℎ
∑︁ 1
mass(Cls O) :=
𝑖=1
𝑤 𝑖
and that the Eichler mass formula (Main Theorem 26.1.5) gives an explicit formula for
this mass in terms of the relevant arithmetic invariants.
Main Theorem 41.5.2 (Trace formula). If 𝔫 is not narrowly principal, then tr 𝑇 (𝔫) = 0.
× , then
If 𝔫 = 𝑛𝑅 is narrowly principal with 𝑛 ∈ 𝐹>0
(
1 ∑︁ ℎ(𝑆) b b b× mass(Cls O), if 𝔫 = 𝑐2 𝑅, 𝑐 ∈ 𝐹 × ;
tr 𝑇 (𝔫) = 𝑚( 𝑆, O; O ) +
2 𝑤𝑆 0, otherwise
(𝑢,𝑡 ,𝑆)
where 𝑤 𝑆 := [𝑆 × : 𝑅 × ] and the sum is over finitely many triples (𝑢, 𝑡, 𝑆) where:
• 𝑢 belongs to a set of representatives of 𝑅>0

× /𝑅 ×2 ;
• 𝑡 ∈ 𝑅 satisfies 𝑡 2 − 4𝑢𝑛 ∈ 𝐹<0

× ; and
• 𝑆 ⊇ 𝑅[𝑥]/(𝑥 2 − 𝑡𝑥 + 𝑢𝑛).
Í𝑘
Proof. We have tr 𝑇 (𝔫) = 𝑖=1 𝑇 (𝔫)𝑖𝑖 . By Lemma 41.2.7, since 𝔫𝑖𝑖 = 𝔫 we conclude
tr 𝑇 (𝔫) = 0 if 𝔫 is not narrowly principal. So suppose 𝔫 = 𝑛𝑅 is narrowly principal,
with 𝑛 ∈ 𝐹>0× . Then by (41.2.9) we have
1 ∑︁
𝑤 𝑖 𝑇 (𝔫)𝑖𝑖 = #{𝛼 ∈ O𝑖 : nrd(𝛼) = 𝑢𝑛}. (41.5.3)
2 × /𝑅 ×2
𝑢𝑅 ×2 ∈𝑅>0
We are free to organize by reduced trace, giving
1 ∑︁ ∑︁
𝑤 𝑖 𝑇 (𝔫)𝑖𝑖 = #{𝛼 ∈ O𝑖 : trd(𝛼) = 𝑡, nrd(𝛼) = 𝑢𝑛}. (41.5.4)
2 𝑢 𝑡 ∈𝑅
Since 𝐵 is definite, we have disc(𝛼) = 𝑡 2 − 4𝑢𝑛 either zero or totally negative, so the
inner sum is over finitely many 𝑡 ∈ 𝑅 either satisfying 𝑡 2 = 4𝑢𝑛 or 𝑡 2 − 4𝑢𝑛 ∈ 𝐹<0 × .
If 𝑡 2 = 4𝑢𝑛 (equivalently 𝛼 = 𝑡/2 ∈ 𝐹), then 𝔫 = 𝑛𝑅 = 𝑐2 𝑅 with 𝑐 = ±𝑡/2;

conversely, if 𝔫 = 𝑛𝑅 = 𝑐2 𝑅 for some 𝑐 ∈ 𝐹 × , then there exists a unique representative
𝑢 ∈ 𝑅>0× /𝑅 ×2 such that 𝑢𝑛 = 𝑐 2 . Consequently, exactly when 𝔫 = 𝑐 2 𝑅 is a square of a
principal ideal, there is a contribution of (1/2) (2) = 1 to the sum.

For the remaining terms, we have 𝛼 ∉ 𝐹 and 𝑅[𝛼] ' 𝑅[𝑥]/(𝑥 2 −𝑡𝑥+𝑢𝑛) is a domain.
The embedding 𝑅[𝛼] ↩→ O𝑖 need not be optimal, but nevertheless corresponds to the
optimal embedding 𝑆 ↩→ O𝑖 for a unique superorder 𝑆 ⊇ 𝑅[𝛼], and conversely. We
count these up to units: the action of conjugaton by 𝜇 ∈ O×𝑖 centralizes such an
embedding if and only if 𝜇 ∈ 𝑆 × , so letting 𝑤 𝑆 := [𝑆 × : 𝑅 × ] we have
∑︁ 𝑤𝑖
#{𝛼 ∈ O𝑖 : trd(𝛼) = 𝑡, nrd(𝛼) = 𝑢𝑛} = 𝑚(𝑆, O𝑖 ; O×𝑖 ) .
2
𝑤 𝑆
𝑆 ⊇𝑅 [𝑥 ]/( 𝑥 −𝑡 𝑥+𝑢𝑛)
Plugging these into (41.5.4), we obtain
1 ∑︁ ∑︁ 𝑤𝑖
𝑤 𝑖 𝑇 (𝔫)𝑖𝑖 = 𝑚(𝑆, O𝑖 ; O×𝑖 )
2 𝑢,𝑡 𝑤𝑆
𝑆 ⊇𝑅 [𝑥 ]/( 𝑥 2 −𝑡 𝑥+𝑢𝑛)
( (41.5.5)
1, if 𝔫 = (𝑐𝑅) 2 ;
+
0, otherwise.
Dividing through by 𝑤 𝑖 and summing (41.5.5), we have

ℎ
1 ∑︁ ∑︁ 1 ∑︁ 1
tr 𝑇 (𝔫) = 𝑚(𝑆, O𝑖 ; O×𝑖 ) + 𝛿 (41.5.6)
2 𝑤𝑆 𝑤𝑖
(𝑢,𝑡 ,𝑆)
𝑖=1 𝑖
where 𝛿 = 1, 0 according as 𝔫 is a square of a principal ideal or not. Now substituting

(41.5.1) and the definition of mass, the theorem is proven.
1 ∑︁ ℎ(𝑆) b b b×
# Cls O = mass(Cls O) + 𝑚( 𝑆, O; O )
2 𝑤𝑆
(𝑢,𝑡 ,𝑆)
where the sum is over (𝑢, 𝑡, 𝑆) as in Main Theorem 41.5.2 with 𝑛 = 1.
Proof. For 𝔫 = 𝑅, we have 𝑇 (1) the identity matrix, so tr 𝑇 (1) = # Cls O.
Corollary 41.5.7 gives a different way to prove (and interpret) the Eichler class
number formula (Main Theorem 30.8.6): for the exact comparison, see Exercise 41.2.
Example 41.5.8. Suppose √ 𝐹 = Q and 𝑅 = Z. For 𝑑 ∈ Z a nonsquare discriminant,

we write 𝑆 𝑑 := Z[(𝑑 + 𝑑)/2] for the unique quadratic ring of discriminant 𝑑, so
𝑤 𝑆𝑑 = [𝑆 ×𝑑 : Z× ] = 1 except for 𝑑 = −3, −4. In the trace formula (Main Theorem
41.5.2), the quaternion algebra that appears is definite, and so the only quadratic orders
that embed are imaginary quadratic, with discriminant 𝑑 < 0. Simplifying in this way,
the trace formula then becomes
1 ∑︁ ∑︁ ℎ(𝑆 𝑑 ) b b b×
tr 𝑇 (𝑛) = 𝑚( 𝑆 𝑑 , O; O ) (41.5.9)
2 𝑡 ∈Z 𝑤 𝑆𝑑
𝑑 𝑓 2 =𝑡 2 −4𝑛<0
for 𝑛 not a square (adding a mass term for 𝑛 a square).

To notationally simplify a bit further, we define modified Hurwitz class numbers
ℎ(𝑆) b b b×
ℎO (𝑆) := 𝑚( 𝑆, O; O )
𝑤𝑆
b O;
where the factor 𝑚( 𝑆, b Ob× ) is defined by purely local data, given in section 30.5 for
maximal orders and section 30.6 for Eichler orders. Writing ℎO (𝑑) = ℎO (𝑆 𝑑 ) for the
order of discriminant 𝑑, we arrive at a pleasing formula:
1 ∑︁ ∑︁
tr 𝑇 (𝑛) = ℎO (𝑑) (41.5.10)
2 𝑡 ∈Z
𝑑 𝑓 2 =𝑡 2 −4𝑛<0
again for 𝑛 not a square.

Taking 𝑛 = 1 as in Corollary 41.5.7 (and adding back the mass term) gives
1 ∑︁
# Cls O = mass(Cls O) + ℎO (𝑡 2 − 4)
2 𝑡 ∈Z
𝑡 2 <4 (41.5.11)
1
= mass(Cls O) + ℎO (−4) + 2ℎO (−3).
2
For O an Eichler order, after substitution we recover the Eichler class number formula
(Theorem 30.1.5).
Example 41.5.12. As an illustration of Example 41.5.8, we compute tr 𝑇 (3) for the

Hurwitz order. We compute the following table of values:
𝑡 𝑑 ℎ(𝑆 𝑑 ) 𝑤 𝑆𝑑 𝑚( 𝑆b𝑑 , O;
b Ob× ) ℎO (𝑑)
0 −12 1 1 0 0
0 −3 1 3 2 2/3
±1 −11 1 1 1 − (−11 | 2) = 2 2
±2 −8 1 1 1 − (−8 | 2) = 1 1
±3 −3 1 3 1 − (−3 | 2) = 2 2/3
Summing then gives
tr 𝑇 (3) = (1/2) (0 + 2/3) + (2 + 1 + 2/3) = 4.
Indeed, more generally since the Hurwitz order has # Cls O = 1, the matrix 𝑇 (𝑛) is
a 1 × 1-matrix with 𝑇 (𝑛) = [𝜎(𝑛)] for 𝑛 odd. This observation implies a nontrivial
(and otherwise surprising) relationship between class numbers of imaginary quadratic
orders!
Remark 41.5.13. Brandt [Bra43, §III] defined Brandt matrices in the same paper as
his groupoid; he called them Hecke matrices, as he claimed to follow parallels with
certain operators defined by Hecke. Indeed, Hecke [Hec40, §9, Satz 53] conjectured
that the space of cusp forms of weight 2 on Γ0 ( 𝑝) for 𝑝 prime was spanned by certain
linear combinations of theta series, and it was this observation that motivated Brandt.
(Eichler [Eic56a, footnote 16] says that Brandt should not have named them after
Hecke, since it was really Brandt who interpreted function-theoretic results of Hecke
using pure arithmetic.)
Eichler [Eic56a] proved that the ring generated by the Brandt matrices was a com-
mutative, semisimple ring and proved the trace formula for Brandt matrices [Eic56a,
§6]. In this early work, he already foresaw the application of Brandt matrices to other
base fields: as an application, he used Brandt matrices to give class number relations
between imaginary quadratic fields, and in the function field case these become rela-
tions among divisor class groups for hyperelliptic curves. Eichler [Eic77, Chapter II]
presented the generalization to totally real fields, giving a treatment of Hecke opera-
tors, Brandt matrices, and theta series, and he proved that the Brandt matrices realize
Hecke operators in certain spaces of Hilbert modular forms.
Eichler then later gave a self-contained presentation [Eic73, Chapter II] of the
theory of Brandt matrices over Q, with the intended application the solution to Hecke’s
conjecture (suitably corrected), now known as the basis problem for Γ0 ( 𝑝): to give
bases of linearly independent forms of spaces of modular forms in terms of theta series
of quadratic forms coming from quaternion algebras. This line of work was followed
by generalizations by Hĳikata [Hĳ74] and Hĳikata–Saito [HS73] for general Eichler
orders, Pizer [Piz76b, Piz76c] for residually split orders, culminating in a solution over
the rational numbers to the basis problem by Hĳikata–Pizer–Shemanske [HPS89a].
The method of proof for the solution to the basis problem is the use of the trace
formula, for which a key ingredient is the theory of optimal embeddings: see Remark
30.6.18.
Indeed, it is much more involved analytically, but one can similarly compute
the trace of the Hecke operator acting on classical spaces of modular forms or more
generally spaces of Hilbert modular forms. These trace formulae are quite complicated,
but one notices that they have a similar shape as the above trace formula; and in fact,
under certain hypotheses and after restricting to an appropriate new subspace, the traces
are equal. But since both rings are semisimple, this implies that the same systems of
eigenvalues for the Hecke operators arise! Such a correspondence was first given by
Eichler, as above; it was generalized to totally real fields by Shimizu [Shz72] using theta
series, and the most general formulation given by Jacquet–Langlands [JL70]. This
correspondence was conjectured to generalize to the principle of Langlands functorial
transfer: for an introduction to this vast area, see Gelbart [Gel84].
In light of the preceding epic remark, we hope we have inspired the reader to pursue
the relationship between Brandt matrices and modular forms! Unfortunately, it would
require another book to respectfully develop this subject.
Remark 41.5.14. Pizer [Piz80a] was the first to give an algorithm for computing clas-
sical modular forms using Brandt matrices (on Γ0 (𝑁) for 𝑁 not a perfect square); see
also the work of Kohel [Koh2001] over Z. This algorithm was generalized to compute
Hilbert modular forms over a totally real field of narrow class number 1 by Socrates–
Whitehouse [SW2005], with algorithmic improvements by Dembélé [Dem2007]. The
assumption on the class number was removed by Dembélé–Donnelly [DD2008]. A
survey of these methods are given by Dembélé–Voight [DV2013, §4, §8].
Exercises
Unless otherwise specified, in these exercises let 𝑅 be a global ring with eligible set
S ⊆ Pl 𝐹, let 𝐵 be an S-definite quaternion algebra over 𝐹 and let O ⊂ 𝐵 be an 𝑅-order
in 𝐵.
1. Extend the definition of the Brandt matrix to include the case 𝔫 = (0) of the zero
ideal, following (41.2.8): define 𝑇 (0)𝑖 𝑗 = 1/𝑤 𝑖 for 𝑖, 𝑗 = 1, . . . , ℎ. Conclude
tr 𝑇 (0) = mass(Cls O).
2. Show that Corollary 41.5.7 agrees with Main Theorem 30.8.6. [Hint: organize
by 𝑞 := [𝑆 × : 𝑅 × ], observe that in 𝑆 ⊇ 𝑅[𝑥]/(𝑥 2 − 𝑡𝑥 + 𝑢) we have necessarily
[𝑆 × : 𝑅 × ] ≥ 2 and each such 𝑆 contains 𝑞 − 1 orders of the form 𝑅[𝑥]/(𝑥 2 −
𝑡𝑥 + 𝑢).]
3. Refine Lemma 41.2.7(c) in a special case as follows. Suppose Cl+ 𝑅 is trivial.
Show that
1
𝑇 (𝔫)𝑖 𝑗 = #{𝛼 ∈ 𝐼 𝑗 𝐼𝑖−1 : nrd(𝛼) = 𝑛𝑖 𝑗 }
𝑤 𝑖,1
where 𝑤 𝑖,1 := #O1𝑖 .

4. Suppose 𝑅 = Z and 𝐹 = Q, suppose disc 𝐵 = 𝑝 is prime and O is a maximal

order. Show that
(
1, if 𝑝 = 2, 3;
tr 𝑇 ( 𝑝) =
ℎO (−4𝑝), if 𝑝 > 3.
where


 ℎ(−4𝑝)/2, if 𝑝 ≡ 1 (mod 4);


ℎO (−4𝑝) = ℎ(−𝑝), if 𝑝 ≡ 7 (mod 8);

 2ℎ(−𝑝), if 𝑝 ≡ 3 (mod 8) and 𝑝 > 3.


What does this say about the number of maximal orders in 𝐵 up to isomorphism
such that every two-sided ideal is principal?
5. Give another proof of Proposition 41.3.1(b) using the local-global dictionary for
lattices.
6. Prove (41.3.12) using induction and then expand to verify (41.3.13).

−1, −11
7. Let 𝐵 = with disc 𝐵 = 11 and let O = Zh𝑖, 12 ( 𝑗 + 1)i.
Q
(a) Show that O is a maximal order with #O× = 4.
(b) Show that the ternary quadratic form associated to O is similar to 𝑥 2 − 𝑥𝑧 +
𝑦 2 + 3𝑧2 .
(c) Show that Cl O = {[O], [𝐼2 ]} where 𝐼2 = 2O + 12 (1 + 2𝑖 + 𝑗)O. [Hint:
Follow Example 17.6.3.] Along the way, show that

1 3
𝑇 (2) = .
2 0
(d) Pause and show that O2 := OL (𝐼2 ) has #O×2 = 6 and associated ternary
quadratic form 𝑥 2 − 𝑥𝑦 − 𝑥𝑧 + 𝑦 2 + 𝑦𝑧 + 4𝑧2 .
(e) Show that 𝑀2 (O) has two eigenspaces for the Hecke algebra, one spanned
by a form 𝑒 with 𝑇 ( 𝑝) (𝑒) = ( 𝑝 + 1)𝑒 for all 𝑝 ≠ 11, and the other spanned
by a form 𝑓 with 𝑇 (2) ( 𝑓 ) = −2 𝑓 .
(f) Verify the trace formula (41.5.9) for tr 𝑇 (2) = 1 by computing class num-
bers.
[There is a unique normalized cusp form 𝑓 ∈ 𝑆2 (Γ0 (11)) of weight 2 and level
11 with
∞
Ö ∞
Ö ∞
∑︁
𝑓 (𝑞) = 𝑞 (1 − 𝑞 𝑛 ) 2 (1 − 𝑞 11𝑛 ) 2 = 𝑞 − 2𝑞 2 − 𝑞 3 + · · · = 𝑎𝑛 𝑞𝑛
𝑛=1 𝑛=1 𝑛=1
matching 𝑓 in the sense that 𝑇 (𝑛) 𝑓 = 𝑎 𝑛 𝑓 for all 11 - 𝑛.]

8. Let O(3, 5) be an Eichler order of level 5 and reduced discriminant 15 (in a
quaternion algebra 𝐵 of discriminant 5) and similarly let O(5, 3) be an Eichler
order of level 3 and reduced discriminant 15.
(a) Show that # Cls O(3, 5) = # Cls O(5, 3) = 2.

(b) Show that there is a unique eigenvector for the Brandt matrix 𝑇 (2) for
both orders with eigenvalue −1. As far as you can compute, show that this
eigenvector shares the same eigenvalues for 𝑇 (𝑛) both orders.
9. We consider
an example
of Brandt matrices not restricted to maximal orders.
−1, −1
Let 𝐵 = and let
Q
O = Zh2𝑖, 2 𝑗i = Z ⊕ Z(2𝑖) ⊕ Z(2 𝑗) ⊕ Z(4𝑖 𝑗).
(a) Show that O is an order with discrd O = 𝑁 = 64.

(b) Compute that # Cls O = 4.
(c) Under the action of the Brandt matrices 𝑇 (𝑛), show that there are 3 irre-
ducible factors of dimensions 1, 1, 2. In a basis of characteristic functions,
identify the one-dimensional factors as:
∞
1 ∑︁ ∗
(1, 1, 1, 1) ↔ 𝑒(𝑞) := + 𝜎 (𝑛)𝑞 𝑛
24 𝑛=1
∞
1 ∑︁ ∗
(1, −1, −1, 1) ↔ 𝑒 𝜒 (𝑞) := + 𝜎 (𝑛) 𝜒(𝑛)𝑞 𝑛
24 𝑛=1

∑︁ −1
where now 𝜎 ∗ (𝑛) := 𝑑 and 𝜒(𝑛) = .
𝑛
2-𝑑 |𝑛
[The two-dimensional space has basis (1, 0, 0, −1), (0, 1, −1, 0) ↔ 𝑓1 , 𝑓2
and 𝑎 𝑝 ( 𝑓1 ) = 𝑎 𝑝 ( 𝑓2 ) for all 𝑝, with
𝑓𝑖 = 𝑞 + 2𝑞 5 − 3𝑞 9 − 6𝑞 13 + 2𝑞 17 + . . .
corresponding to the isogeny class of the elliptic curve 𝐸 : 𝑦 2 = 𝑥 3 + 𝑥 of

conductor 64.]
Chapter 42
Supersingular elliptic curves
In the previous chapter, we showed that Brandt matrices for an order in a definite
quaternion algebra 𝐵 contain a wealth of arithmetic. In the special case where disc 𝐵 =
𝑝 is prime, there is a further beautiful connection between Brandt matrices and the
theory of supersingular elliptic curves, arising from the following important result:
there is an equivalence of categories between supersingular elliptic curves over F 𝑝 and
right ideals in a (fixed) maximal order O ⊂ 𝐵. We pursue this important connection
in this chapter for the reader who has a bit more background in algebraic curves.
42.1 Supersingular elliptic curves

In this section, we briefly review what we will need from the theory of elliptic curves;
see Silverman [Sil2009] for further general reference. Let 𝐹 be a field with algebraic
closure 𝐹 al .
Definition 42.1.1. An elliptic curve is a smooth projective curve (variety of dimension
1) of genus 1 equipped with a rational point. Every elliptic curve 𝐸 is isomorphic over
𝐹 to the projective curve associated to the affine equation
𝐸 : 𝑦 2 + 𝑎 1 𝑥𝑦 + 𝑎 3 𝑦 = 𝑥 3 + 𝑎 2 𝑥 2 + 𝑎 4 𝑥 + 𝑎 6
with 𝑎 𝑖 ∈ 𝐹.
Definition 42.1.2. An isogeny 𝜙 : 𝐸 → 𝐸 0 is a nonconstant morphism of pointed
curves; such a map is automatically surjective and a group homomorphism, with the
marked point as origin.
Let Hom(𝐸, 𝐸 0) be the collection of isogenies from 𝐸 to 𝐸 0 defined over 𝐹; if we
need to allow isogenies defined over a larger field, we will similarly extend the field
of definition of our elliptic curves. Then Hom(𝐸, 𝐸 0) is a torsion-free Z-module of
rank at most four. Let End(𝐸) := Hom(𝐸, 𝐸) be the endomorphism ring of 𝐸 and
let End(𝐸)Q := End(𝐸) ⊗Z Q be the endomorphism algebra.
42.1.3. For each nonzero isogeny 𝜙 : 𝐸 → 𝐸 0, there exists a dual isogeny 𝜙∨ : 𝐸 0 → 𝐸
such that 𝜙∨ ◦ 𝜙 and 𝜙 ◦ 𝜙∨ are equal to multiplication by the degree deg 𝜙 ∈ Z>0 on 𝐸
767
768 CHAPTER 42. SUPERSINGULAR ELLIPTIC CURVES
and 𝐸 0, respectively. In particular, the dual ∨ is a standard involution on End(𝐸) that

is positive (see 8.4.1); the Q-algebra End(𝐸)Q := End(𝐸) ⊗Z Q is therefore a division
ring. In particular, we have the equality
1
𝜙−1 = 𝜙∨ (42.1.4)
deg 𝜙
in End(𝐸)Q for all nonzero 𝜙 ∈ End(𝐸).

From now on, let 𝐸 be an elliptic curve over 𝐹.
Lemma 42.1.5. The endomorphism algebra End(𝐸)Q of 𝐸 is either Q, an imaginary
quadratic field 𝐾, or a definite quaternion algebra over Q.
Proof. We apply Theorem 3.5.1 to conclude that End(𝐸)Q is either Q, a quadratic field,
or a division quaternion algebra. Then by Example 8.4.2, the involution is positive if
and only if End(𝐸)R is R, C, or H, so in the second case we must have an imaginary
quadratic field and in the third case we must have a definite quaternion algebra.
42.1.6. Among the possibilities in Lemma 42.1.5, if End(𝐸 𝐹 al )Q is a quaternion

algebra, then we say 𝐸 is supersingular.
See Silverman [Sil2009, §V.3] for a treatment of supersingular elliptic curves.
Proposition 42.1.7. If 𝐸 is supersingular, then char 𝐹 = 𝑝 > 0. Moreover, the
following are equivalent:
(i) 𝐸 is supersingular;
(ii) 𝐸 [ 𝑝] (𝐹 al ) = {0}; and
(iii) the map [ 𝑝] : 𝐸 → 𝐸 is purely inseparable and 𝑗 (𝐸) ∈ F 𝑝2 ;
If 𝐹 is a finite field, then these are further equivalent to
(iv) trd(𝜙) = 𝜙 + 𝜙∨ ≡ 0 (mod 𝑝), where 𝜙 : 𝐸 → 𝐸 is the Frobenius endomor-
phism.
Proof. See Silverman [Sil2009, V.3.1].
42.1.8. One can often reduce questions about supersingular elliptic curves to ones
where the base field 𝐹 is F 𝑝2 as follows: by Proposition 42.1.7(iii), if 𝐸 is supersingular
then 𝐸 is isomorphic over 𝐹 al to a curve 𝐸 defined over F 𝑝2 .
The following fundamental result is due to Deuring [Deu41]; we give a proof due
to Lenstra [Len96, §3].
Theorem 42.1.9. Let 𝐸 be an elliptic curve over 𝐹 and suppose that rkZ End(𝐸) = 4.
Then 𝐵 = End(𝐸)Q is a quaternion algebra over Q ramified at 𝑝 = char 𝐹 and ∞, and
End(𝐸) is a maximal order in 𝐵.
In particular, if over 𝐹 we have dim End(𝐸) = 4, then automatically 𝐸 has all of
its endomorphisms defined over 𝐹.
42.1. SUPERSINGULAR ELLIPTIC CURVES 769
Proof. Let O = End 𝐸 ⊆ 𝐵 = End(𝐸)Q . Let 𝑛 > 0 be prime to 𝑝. Then there is an

isomorphism [Sil2009, Corollary III.6.4(b)]
𝐸 [𝑛] = 𝐸 [𝑛] (𝐹 al ) ' Z/𝑛Z ⊕ Z/𝑛Z
as abelian groups, and the endomorphism ring of this abelian group is End 𝐸 [𝑛] '
M2 (Z/𝑛Z).
We claim that the structure map O/𝑛O → End 𝐸 [𝑛] is injective, which is to say,
𝐸 [𝑛] is a faithful module over O/𝑛O. Indeed, suppose 𝜙 ∈ O annihilates 𝐸 [𝑛]; then
since multiplication by 𝑛 is separable, by the homomorphism theorem for elliptic curves
[Sil2009, Corollary III.4.11] there exists 𝜓 ∈ O such that 𝜙 = 𝑛𝜓, so 𝜙 ≡ 0 ∈ O/𝑛O,
proving injectivity. But further, since #O/𝑛O = # End 𝐸 [𝑛] = 𝑛4 , the structure map is
an isomorphism.
Since O is a free Z-module, we have
Oℓ := O ⊗Z Zℓ = O ⊗Z lim Z/ℓ 𝑛 Z ' lim O/ℓ 𝑛 O.

← −−
𝑛
← −−
𝑛
The structure isomorphisms in the previous paragraph are compatible with respect to
powers of ℓ, so with the previous line they provide an isomorphism
∼ lim End 𝐸 [ℓ 𝑛 ] = End 𝐸 [ℓ ∞ ] ' M (Z )
Oℓ −→ Zℓ 2 ℓ
←−−
𝑛
of Zℓ -algebras, and in particular Oℓ is maximal and 𝐵ℓ ' M2 (Qℓ ) so 𝐵 is split at ℓ.

Since 𝐵 is definite, it follows from the classification theorem (Main Theorem
14.1.3, equivalent to quadratic reciprocity) that Ram(𝐵) = {𝑝, ∞}, so 𝐵 𝑝 is a division
algebra over Q 𝑝 .
To conclude, we show that O 𝑝 is maximal. For 𝜙 ∈ O an isogeny, let deg𝑖 𝜙 be
the inseparable degree of 𝜙, which is a power of 𝑝. We put deg𝑖 0 = ∞. Then deg𝑖 𝜙
is divisible by 𝑞 = 𝑝 𝑟 if and only if 𝜙 factors via the 𝑞th power Frobenius morphism
𝐸 → 𝐸 (𝑞) . The map
𝑣 : End(𝐸)Q → Q ∪ {∞}
1 (42.1.10)
𝑣(𝑎𝜙) = ord 𝑝 (𝑎) + ord 𝑝 (deg𝑖 𝜙)
2
for 𝑎 ∈ Q and 𝜙 ∈ End(𝐸) is well-defined (since deg𝑖 [ 𝑝] = deg[ 𝑝] = 𝑝 2 ). Factoring
an isogeny into its separable and inseparable parts shows that
ord 𝑝 (deg𝑖 𝜙) = ord 𝑝 (deg 𝜙) = ord 𝑝 (nrd 𝜙)
so (42.1.10) is precisely the valuation (13.3.1) on 𝐵 = End(𝐸)Q extending the 𝑝-adic

valuation on Q. (See also Exercise 42.2.)
To conclude, we show that O ( 𝑝) is the valuation ring (13.3.3) of 𝐵 and is therefore
maximal (Proposition 13.3.4). If 𝛼 ∈ O ( 𝑝) = O ⊗Z Z ( 𝑝) then deg 𝛼 ∈ Z ( 𝑝) so 𝛼 is
in the valuation ring. Conversely, let 𝛼 ∈ 𝐵 be a rational isogeny with 𝑣(𝛼) ≥ 0,
and write 𝛼 = 𝑎𝜙 where 𝜙 is an (actual) isogeny not divisible by any integer. Then
𝑣(𝛼) = ord 𝑝 (𝑎) + 𝑣(𝜙) ≥ 0 and 0 ≤ 𝑣(𝜙) ≤ 1/2, since the multiplication by 𝑝 is
purely inseparable; so ord 𝑝 (𝑎) ≥ −1/2 and therefore 𝑎 ∈ Z ( 𝑝) , and hence 𝛼 ∈ O ( 𝑝) .
Finally, since an order is maximal if and only if it is locally maximal, O itself is a
maximal order in the quaternion algebra 𝐵.
In light of 42.1.8, we now let 𝐹 = Fal𝑝 be an algebraic closure of F 𝑝 . Let 𝐸, 𝐸 0 be

elliptic curves over 𝐹. If 𝐸 is isogenous to 𝐸 0, then 𝐸 is supersingular if and only if
𝐸 0 is supersingular (see Exercise 42.1). The converse is also true, as follows.
Lemma 42.1.11. Let 𝐸, 𝐸 0 be supersingular elliptic curves over 𝐹. Then Hom(𝐸, 𝐸 0)

is a Z-module of rank 4 that is invertible as a right End(𝐸)-module under precompo-
sition and a left End(𝐸 0)-module under postcomposition.
In particular, if 𝐸, 𝐸 0 are supersingular elliptic curves over Fal𝑝 , then there exists a
separable isogeny 𝐸 → 𝐸 0.
Proof. We may suppose 𝐸 is defined over a finite field F𝑞 such that 𝐸 has all of its
endomorphisms defined over F𝑞 . Let 𝜋 ∈ O = End(𝐸) be the 𝑞-power Frobenius
endomorphism. Then 𝐵 = O ⊗Z Q is a quaternion algebra over Q. Since End(𝐸) is
defined over F𝑞 , the endomorphism 𝜋 commutes with every isogeny 𝛼 ∈ O, and
so 𝜋 lies in the center of O; since 𝑍 (𝐵) = Q, we have 𝜋 ∈ Z = 𝑍 (O). But
√ √
deg 𝜋 = 𝜋𝜋 = 𝜋 2 = 𝑞 so 𝜋 = ± 𝑞 ∈ Z. Therefore #𝐸 (F𝑞 ) = 𝑞 + 1 ∓ 2 𝑞. Therefore
2
#𝐸 (F𝑞2 ) = 𝑞 + 1 − 2𝑞 = (𝑞 − 1) . 2
Continuing to enlarge F𝑞 , we may repeat the above argument with 𝐸 0 to conclude

that #𝐸 (F𝑞 ) = #𝐸 0 (F𝑞 ). It then follows that 𝐸, 𝐸 0 are isogenous over F𝑞 [Sil2009,
Exercise III.5.4(b)], but we will show this and more. Let ℓ ≠ 𝑝 be prime and let
𝑇ℓ (𝐸) = lim 𝐸 [ℓ 𝑛 ] ' Zℓ2 is the ℓ-adic Tate module of 𝐸. By the Isogeny Theorem
←−−𝑛
[Sil2009, Theorem III.7.7(a)], for every prime ℓ ≠ 𝑝, the natural map
HomF𝑞 (𝐸, 𝐸 0) ⊗ Zℓ → HomF𝑞 (𝑇ℓ (𝐸), 𝑇ℓ (𝐸 0))
is an isomorphism, where HomF𝑞 (𝐸, 𝐸 0) denotes the group of isogenies 𝐸 → 𝐸 0

defined over F𝑞 and HomF𝑞 (𝑇ℓ (𝐸), 𝑇ℓ (𝐸 0)) denotes the group of Zℓ -linear maps that
commute with the action of the 𝑞-power Frobenius Galois automorphism. In the first
paragraph, we showed that this Frobenius action is scalar so commuting is automatic,
and
HomF𝑞 (𝑇ℓ (𝐸), 𝑇ℓ (𝐸 0)) = Hom(Zℓ2 , Zℓ2 ) ' M2 (Zℓ );
and HomF𝑞 (𝐸, 𝐸 0) = Hom(𝐸, 𝐸 0) has rkZ Hom(𝐸, 𝐸 0) = 4.
Finally, we can precompose by endomorphisms of O so Hom(𝐸, 𝐸 0) is a torsion-
free Z-module with a right action by O. Let 𝜓 ∈ Hom(𝐸, 𝐸 0) be nonzero and let
𝜓 ∨ : 𝐸 0 → 𝐸 be the dual isogeny. Then 𝐼 := 𝜓 ∨ Hom(𝐸, 𝐸 0) ⊆ O is an integral
right O-ideal; since O is a maximal order by Theorem 42.1.9, the right O-ideal 𝐼
is necessarily invertible (see 23.1.1), and the same then holds for Hom(𝐸, 𝐸 0) as a
right O-module. The same is true as a left End(𝐸 0)-module, and these two actions
commute.
42.2. SUPERSINGULAR ISOGENIES 771
42.2 Supersingular isogenies

We now investigate the quaternionic endomorphism rings of supersingular elliptic
curves in more detail; we use Waterhouse [Wate69, §3] as our main reference. Let 𝐸
be a supersingular elliptic curve over 𝐹 := Fal𝑝 , let O := End(𝐸), and let 𝐵 := O ⊗ Q.
By Theorem 42.1.9, we have Ram(𝐵) = {𝑝, ∞}, and O ⊆ 𝐵 is a maximal order. Thus
disc 𝐵 = 𝑝 = discrd O (recalling section 15.1).
We temporarily and briefly need the language of group schemes, at the level of
Waterhouse [Wate69]. The reader who is unfamiliar with this language is advised to
skip to 42.2.4 and restrict consideration to left ideals 𝐼 with nrd(𝐼) coprime to 𝑝 (or
equivalently, separable isogenies).
42.2.1. Let 𝐼 ⊆ O be a nonzero integral left O-ideal. Since O is maximal, necessarily
𝐼 is locally principal (in particular, invertible) by Proposition 16.1.2.
We define 𝐸 [𝐼] ⊆ 𝐸 to be the scheme-theoretic intersection
Ù
𝐸 [𝐼] := 𝐸 [𝛼] (42.2.2)
𝛼∈𝐼
where 𝐸 [𝛼] = ker 𝛼 as a group scheme over 𝐹.

Accordingly, there exists an isogeny 𝜙 𝐼 : 𝐸 → 𝐸 𝐼 where 𝐸 𝐼 = 𝐸/𝐸 [𝐼].
42.2.3. We will not need much about the theory of group schemes except that we
can measure the degree of an isogeny via the rank of its kernel. Let 𝐻 ≤ 𝐸 (𝐹) be a
finite 𝐹-subgroup scheme, for example, 𝐻 := ker 𝜙 for 𝜙 : 𝐸 → 𝐸 0 an isogeny. Then
𝐻 = Spec 𝐴 𝐻 is affine and 𝐴 𝐻 is a finite 𝐹-algebra; we define the rank of 𝐻 by
rk 𝐻 := dim𝐹 𝐴 𝐻 . In all cases, we have rk ker 𝜙 = deg 𝜙, even when 𝜙 is inseparable.
This general, scheme-theoretic construction can usually be given plainly, as follows.
42.2.4. If there is a nonzero 𝛼 ∈ 𝐼 giving a separable isogeny 𝛼 : 𝐸 → 𝐸, then
𝐸 [𝐼] (𝐹) = {𝑃 ∈ 𝐸 (𝐹) : 𝛼(𝑃) = 0 for all 𝛼 ∈ 𝐼}. (42.2.5)
We then have the more familiar separable isogeny 𝜙 𝐼 : 𝐸 → 𝐸/𝐸 [𝐼] with ker(𝜙 𝐼 ) =
𝐸 [𝐼] [Sil2009, Proposition III.4.12], and rk ker 𝜙 𝐼 = # ker 𝜙 𝐼 (𝐹) = deg 𝜙 𝐼 .
What remains are inseparable isogenies. Since Ram 𝐵 = {𝑝, ∞} and O is maximal,
by Theorem 18.1.3 (more generally, see 23.3.19), there is a unique two-sided O-ideal
𝑃 ⊆ O of reduced norm 𝑝. Then the map 𝐸 → 𝐸 𝑃 ' 𝐸 ( 𝑝) is the 𝑝-Frobenius map,
and rk ker 𝜙 𝑃 = 𝑝 = deg 𝜙 even though (ker 𝜙) (𝐹) = {0}. The equality 𝑃2 = 𝑝O
corresponds to the fact that 𝐸 [𝑃2 ] = 𝐸 [ 𝑝], and this lies behind the fact that 𝑗 (𝐸) ∈ F 𝑝2
as in Proposition 42.1.7.
Accordingly, a left O-ideal 𝐼 can be written uniquely as 𝐼 = 𝑃𝑟 𝐼 0 with nrd(𝐼 0)
coprime to 𝑝, and this corresponds to a factorization
𝜙 𝐼 : 𝐸 → 𝐸 𝑃𝑟 → 𝐸 𝐼 (42.2.6)
with 𝐸 𝐼 ' 𝐸 𝑃𝑟 /𝐸 𝑃𝑟 [𝐼 0], the isogeny 𝐸 → 𝐸 𝑃𝑟 purely inseparable and the isogeny
𝐸 𝑃𝑟 → 𝐸 𝐼 separable. (This corresponds to the factorization of the extension of
function fields into first a separable extension, then a purely inseparable extension
[Sil2009, Corollary II.2.12].)
Lemma 42.2.7. The pullback map
𝜙∗𝐼 : Hom(𝐸 𝐼 , 𝐸) → 𝐼
(42.2.8)
𝜓 ↦→ 𝜓𝜙 𝐼
is an isomorphism of left O-modules.
Proof. The image of Hom(𝐸 𝐼 , 𝐸) under precomposition by 𝜙 𝐼 lands in End(𝐸) = O

and factors through 𝜙 𝐼 so lands in 𝐼 by definition. The map 𝜙∗𝐼 is an injective
homomorphism of abelian groups. It compatible with the left O-action, given by
postcomposition on Hom(𝐸 𝐼 , 𝐸) and left multiplication on 𝐼.
To conclude, we show it is also surjective. Let 𝛼 ∈ 𝐼; then 𝛼(𝐸 [𝐼]) = {0}
by construction. If 𝛼 is separable, then 𝛼 factors through 𝜙 𝐼 : 𝐸 → 𝐸 𝐼 [Sil2009,
Corollary III.4.11]. In general, we factor 𝜙 𝐼 as in (42.2.6), and then combine the
separable case in the previous sentence with the 𝑝 𝑟 -Frobenius map.
Finally, we can identify the right module structure as follows.
Lemma 42.2.9. The ring homomorphism
𝜄 : End(𝐸 𝐼 ) ↩→ 𝐵
1 (42.2.10)
𝜄(𝛽) = 𝜙−1
𝐼 𝛽𝜙 𝐼 = (𝜙∨ 𝛽𝜙 𝐼 )
deg 𝜙 𝐼 𝐼
is injective and 𝜄(End(𝐸 𝐼 )) = OR (𝐼).
Proof. The equality in (42.2.10) is justified in (42.1.4). The content of the lemma
follows from the identification in the previous Lemma 42.2.7, by transporting structure:
for 𝛽 ∈ End(𝐸 𝐼 ) acting by precomposition, we fill in the diagram
𝜙 𝐼∗
Hom(𝐸 𝐼 , 𝐸) /𝐼
𝛽 𝛽∗ (42.2.11)
𝜙 𝐼∗
Hom(𝐸 𝐼 , 𝐸) /𝐼
to find that
𝛽∗ (𝜓𝜙 𝐼 ) = 𝜓 𝛽𝜙 𝐼 = 𝜓𝜙 𝐼 (𝜙−1
𝐼 𝛽𝜙 𝐼 ) (42.2.12)
and so 𝜄 defines the induced action on 𝐼 by right multiplication, giving an inclusion
𝜄(End(𝐸 𝐼 )) ⊆ OR (𝐼). But End(𝐸 𝐼 ) is a maximal order and OR (𝐼) is an order, so
equality holds.
Next, we show that the isomorphism class of 𝐸 𝐼 depends only on the left ideal
class of 𝐼.
Lemma 42.2.13. If 𝐽 = 𝐼 𝛽 ⊆ O with 𝛽 ∈ 𝐵× , then 𝐸 𝐼 ' 𝐸 𝐽 .

42.2. SUPERSINGULAR ISOGENIES 773
Proof. First, suppose 𝛽 ∈ O. Then
𝐸 [𝐼 𝛽] = {𝑃 ∈ 𝐸 (𝐹) : 𝛼𝛽(𝑃) = 0 for all 𝛼 ∈ 𝐼}.
We claim that 𝛽𝐸 [𝐼 𝛽] = 𝐸 [𝐼]. The containment (⊆) is immediate. For the contain-
ment (⊇), let 𝑄 ∈ 𝐸 [𝐼]. Since 𝛽 is surjective (it is nonconstant), there exists 𝑃 ∈ 𝐸 (𝐹)
such that 𝛽(𝑃) = 𝑄. Thus for all 𝛼 ∈ 𝐼, we have (𝛼𝛽) (𝑃) = 𝛼(𝑄) = 0 so 𝑃 ∈ 𝐸 [𝐼 𝛽].
By the claim, we conclude that 𝜙 𝐼 𝛽 = 𝜙 𝐼 𝛽 and 𝐸 𝐼 𝛽 ' 𝐸 𝐼 .
In general, there exists nonzero 𝑚 ∈ Z such that 𝑚𝛽 ∈ O. By the previous
paragraph, we have isomorphisms 𝐸 𝐼 ' 𝐸 𝐼 (𝑚𝛽) = 𝐸 (𝐼 𝛽)𝑚 ' 𝐸 𝐼 𝛽 .
So far, we have shown how to pass from classes of left O-ideals to (isogenous)
supersingular elliptic curves via kernels. We can also go in the other direction.
42.2.14. Given a finite subgroup scheme 𝐻 ≤ 𝐸 (𝐹), we define
𝐼 (𝐻) := {𝛼 ∈ O : 𝛼(𝑃) = 0 for all 𝑃 ∈ 𝐻} ⊆ O;
then 𝐼 (𝐻) is a left O-ideal, nonzero because #𝐻 ∈ 𝐼 (𝐻).

If 𝐻1 ≤ 𝐻2 ≤ 𝐸 (𝐹) are two such subgroups, then 𝐼 (𝐻1 ) ⊇ 𝐼 (𝐻2 ).
Lemma 42.2.15. If 𝐻1 ⊆ 𝐻2 and 𝐼 (𝐻1 ) = 𝐼 (𝐻2 ), then 𝐻1 = 𝐻2 .
Proof. Let 𝜙1 : 𝐸 → 𝐸/𝐻1 . Factoring, without loss of generality we may assume that
𝜙1 is either separable or purely inseparable. Suppose first that 𝜙1 is separable, and let
𝑛 = #𝐻2 (𝐹). By the proof of Theorem 42.1.9, the structure map O/𝑛O → End 𝐸 [𝑛]
is faithful. So if 𝐻2 > 𝐻1 , then there exists 𝛼 ∈ O such that 𝛼(𝐻1 ) = {0} but
𝛼(𝐻2 ) ≠ {0}, so 𝐼 (𝐻2 ) ≠ 𝐼 (𝐻1 ). Second, suppose that 𝜙1 is purely inseparable: then
𝐻1 = ker 𝜙𝑟𝑝1 is the kernel of the 𝑟 1 -power Frobenius for some 𝑟 1 > 0, and 𝐼 (𝐻1 ) = 𝑃𝑟1
as in 42.2.4. Then 𝑝 𝑟1 ∈ 𝐼 (𝐻1 ) = 𝐼 (𝐻2 ), so 𝐸 → 𝐸/𝐻2 is also purely inseparable,
and 𝐻2 = ker 𝜙𝑟𝑝2 and 𝐼 (𝐻2 ) = 𝑃𝑟2 . We conclude 𝑟 1 = 𝑟 2 , and then 𝐻1 = 𝐻2 .
(a) deg 𝜙 𝐼 = nrd(𝐼).

(b) 𝐼 (𝐸 [𝐼]) = 𝐼.
Proposition 42.2.16 justifies the use of overloaded notation. Our proof follows
Waterhouse [Wate69, Theorem 3.15].
Proof. We begin with (a). We first prove it in an illustrative special case. Suppose
𝐼 = O𝛽 is a principal left O-ideal. Then 𝐸 [𝐼] = 𝐸 [𝛽] where 𝜙 𝐼 = 𝛽 : 𝐸 → 𝐸, and
deg 𝛽 = 𝛽𝛽 = nrd(𝛽) = nrd(𝐼)
is the constant term of the (reduced) characteristic polynomial of 𝛽.

We now return to the general case. We first show that deg(𝐼) | deg 𝜙 𝐼 . Let O0 :=
OR (𝐼) = OL (𝐼 −1 ). By Exercise 17.5, there exists 𝛼 ∈ 𝐵× such that 𝐼 0 = 𝐼 −1 𝛼 ⊆ O0
is in the same right O0-ideal class as 𝐼 −1 and with nrd(𝐼 0) coprime to deg 𝜙 𝐼 . Thus
𝐼 𝐼 0 = O𝛼 ⊆ O is a compatible product and 𝛼 ∈ O. By the previous paragraph, we

have
deg 𝜙 𝐼 𝐼 0 = nrd(𝐼 𝐼 0) = nrd(𝐼) nrd(𝐼 0).
The map 𝜙 𝐼 𝐼 0 factors through 𝜙 𝐼 , so deg 𝜙 𝐼 | deg 𝜙 𝐼 𝐼 0 . Since nrd(𝐼 0) is coprime to
deg 𝜙 𝐼 we have deg 𝜙 𝐼 | nrd(𝐼).
Repeating this argument, we have deg 𝜙 𝐼 0 | nrd(𝐼 0) as well. We combine these to
conclude (a). We have
(deg 𝜙 𝐼 ) (deg 𝜙 𝐼 0 ) | nrd(𝐼) nrd(𝐼 0) = deg 𝜙 𝐼 𝐼 0 . (42.2.17)
But
deg 𝜙 𝐼 𝐼 0 = rk 𝐸 [𝐼 𝐼 0] | (rk 𝐸 [𝐼]) (rk 𝐸 [𝐼 0]) = (deg 𝜙 𝐼 ) (deg 𝜙 𝐼 0 ). (42.2.18)
Putting together (42.2.17)–(42.2.18), we see that equality holds, so deg 𝜙 𝐼 = nrd(𝐼).

Now we prove (b). Let 𝐽 = 𝐼 (𝐸 [𝐼]). Then 𝐼 ⊆ 𝐽 since 𝐼 𝐸 [𝐼] = {0}; thus
𝐸 [𝐼] ⊇ 𝐸 [𝐽]. At the same time, we have
Ù Ù Ù
𝐸 [𝐽] = 𝐸 [𝛼] = 𝐸 [𝛼] ⊇ 𝐸 [𝛼] = 𝐸 [𝐼] (42.2.19)
𝛼∈𝐽 𝛼∈O 𝛼∈𝐼
𝛼𝐸 [𝐼 ]={0}
so equality holds and 𝐸 [𝐼] = 𝐸 [𝐽]. Thus deg 𝜙 𝐼 = deg 𝜙 𝐽 . By (a), we have
nrd(𝐼) = deg 𝜙 𝐼 = deg 𝜙 𝐽 = nrd(𝐽). (42.2.20)
Since 𝐼 ⊆ 𝐽, we have O ⊆ 𝐽 𝐼 −1 . From Proposition 16.4.3 and (42.2.20), we have

nrd(𝐽 𝐼 −1 ) = [O : 𝐽 𝐼 −1 ] = 1 so O = 𝐽 𝐼 −1 and therefore 𝐼 = 𝐽.
Corollary 42.2.21. For every isogeny 𝜙 : 𝐸 → 𝐸 0, there exists a left O-ideal 𝐼 and
an isomorphism 𝜌 : 𝐸 𝐼 → 𝐸 0 such that 𝜙 = 𝜌𝜙 𝐼 . Moreover, for every maximal order
O0 ⊆ 𝐵, there exists 𝐸 0 such that O0 ' End(𝐸 0).
Proof. Let 𝐻 be the scheme-theoretic kernel of 𝜙. Then 𝐻 ⊆ 𝐸 [𝐼 (𝐻)], so 𝜙 𝐼 (𝐻 )

factors through 𝜙 with 𝜙 𝐼 (𝐻 ) = 𝜌𝜙 for some isogeny 𝜌 : 𝐸 𝐼 (𝐻 ) → 𝐸 0. But 𝐼 (𝐻) =
𝐼 (𝐸 [𝐼 (𝐻)]) by Proposition 42.2.16, so 𝐻 = 𝐸 [𝐼 (𝐻)] by Lemma 42.2.15. Thus
deg 𝜙 𝐼 (𝐻 ) = deg 𝜙, and so deg 𝜌 = 1 and 𝜌 is an isomorphism, with 𝜙 = 𝜌 −1 𝜙 𝐼 (𝐻 ) .
The second statement follows similarly using a connecting ideal between orders (see
section 17.4).
We may now compare endomorphisms analogously to Lemma 42.2.7.

Lemma 42.2.22. Let 𝐼, 𝐼 0 ⊆ O be nonzero integral left O-ideals. Then the natural
map
Hom(𝐸 𝐼 , 𝐸) Hom(𝐸 𝐼 0 , 𝐸 𝐼 ) → Hom(𝐸 𝐼 0 , 𝐸)
is bĳective, giving a further bĳection
Hom(𝐸 𝐼 0 , 𝐸 𝐼 ) → (𝐼 0 : 𝐼) R = 𝐼 −1 𝐼 0
(42.2.23)
𝜓 ↦→ 𝜙−1
𝐼 𝜓𝜙 𝐼 0 .
42.3. EQUIVALENCE OF CATEGORIES 775
Proof. By Lemma 42.2.7, we have Hom(𝐸 𝐼 , 𝐸)𝜙 𝐼 = 𝐼. By Proposition 42.2.16, we

have 𝑚 = deg 𝜙 𝐼 = nrd(𝐼). The left ideal 𝐼 ⊆ O is invertible thus 𝑚 = nrd(𝐼) ∈ 𝐼 𝐼,
hence there exist finitely many 𝛼𝑖 , 𝛽𝑖 ∈ Hom(𝐸 𝐼 , 𝐸) such that
∑︁ ∑︁ ∑︁
[𝑚] = (𝛼𝑖 𝜙 𝐼 ) (𝛽𝑖 𝜙 𝐼 ) ∨ = (𝛼𝑖 𝜙 𝐼 ) (𝜙∨𝐼 𝛽𝑖∨ ) = 𝛼𝑖 [𝑚] 𝛽𝑖∨ (42.2.24)
𝑖 𝑖 𝑖
therefore ∑︁
[1] = 𝛼𝑖 𝛽𝑖∨ ∈ End(𝐸). (42.2.25)
𝑖
For 𝜓 ∈ Hom(𝐸 𝐼 0 , 𝐸), postcomposing with (42.2.25) gives

∑︁
𝜓= 𝛼𝑖 (𝛽𝑖∨ 𝜓) ∈ Hom(𝐸 𝐼 , 𝐸) Hom(𝐸 𝐼 0 , 𝐸 𝐼 ) (42.2.26)
𝑖
so the natural injective map is bĳective. This gives
𝐼𝜙−1
𝐼 Hom(𝐸 𝐼 0 𝐸 𝐼 )𝜙 𝐼 0 = 𝐼
0
(42.2.27)
and thereby the bĳective map (42.2.23), using Exercise 16.11 for the relationship to
the colon ideal.
42.3 Equivalence of categories

We now show that the association from supersingular elliptic curves to right ideals is
an equivalence of categories. We recall that 𝐹 is an algebraically closed field with
𝑝 = char 𝐹 prime.
42.3.1. Let 𝐸 0 be a supersingular elliptic curve over 𝐹 := Fal𝑝 ; it will serve the role as
a base object. Let O0 := End(𝐸 0 ) and 𝐵0 := O0 ⊗ Q.
Theorem 42.3.2. The association 𝐸 ↦→ Hom(𝐸, 𝐸 0 ) is functorial and defines and
equivalence between the category of
supersingular elliptic curves over 𝐹, under isogenies
and
invertible left O0 -modules, under nonzero left O0 -module homomorphisms.
Remark 42.3.3. Written this way, the functor Hom(−, 𝐸 0 ) is contravariant. One can
equally well take Hom(𝐸 0 , −) to get a covariant functor with right O0 -modules; using
the standard involution, these are seen to contain the same content.
Proof. To begin, we need to show Hom(−, 𝐸 0 ) is a functor. The association 𝐸 ↦→

Hom(𝐸, 𝐸 0 ) makes sense on objects by Lemma 42.1.11. On morphisms, to an isogeny
𝜙 : 𝐸 → 𝐸 0 we associate
𝜙∗ : Hom(𝐸 0, 𝐸 0 ) → Hom(𝐸, 𝐸 0 )
(42.3.4)
𝜓 ↦→ 𝜓𝜙.
The map 𝜙∗ is a homomorphism of left O0 -modules, since it is compatible with

postcomposition with O0 = End(𝐸), so Hom(−, 𝐸 0 ) is functorial.
Next, we claim that Hom(−, 𝐸 0 ) is essentially surjective. Let 𝐼 be an invertible
left O0 -module. Tensoring with Q we get an injection 𝐼0 ↩→ 𝐼0 ⊗ Q ' 𝐵0 , so up to
isomorphism of left O0 -modules, we may suppose 𝐼 ⊆ 𝐵0 . Scaling by an integer, we
may suppose 𝐼 ⊆ O0 is a left O0 -ideal. Let 𝐸 𝐼 = 𝐸/𝐸 [𝐼]. By Lemma 42.2.7, we have
Hom(𝐸 𝐼 , 𝐸 0 ) ' 𝐼 as left O0 -modules, as desired.
Finally, we show that Hom(−, 𝐸 0 ) is fully faithful, i.e., the map
Hom(𝐸, 𝐸 0) → Hom(Hom(𝐸 0, 𝐸 0 ), Hom(𝐸, 𝐸 0 ))
𝜙 ↦→ 𝜙∗
(the former as isogenies, the latter as homomorphisms of left O0 -modules) is bĳective.
This bĳectivity is made plain by an application of Corollary 42.2.21: there is a left O0 -
ideal 𝐼 such that 𝐸 ' 𝐸 0,𝐼 ; applying this isomorphism, we may suppose without loss
of generality that 𝐸 = 𝐸 0,𝐼 . Then by Lemma 42.2.7, we have 𝐼 = Hom(𝐸 0,𝐼 , 𝐸 0 )𝜙0,𝐼 .
Repeat with 𝐸 0 and 𝐼 0. Then after these identifications, we are reduced to the setting
of Lemma 42.2.22 (with the location of the prime swapped): the map
Hom(𝐸 0,𝐼 , 𝐸 0,𝐼 0 ) → (𝐼 : 𝐼 0) R = 𝐼 0−1 𝐼

(42.3.5)
𝜓 ↦→ 𝜙−1
0,𝐼 0 𝜓𝜙0,𝐼 .
is indeed bĳective.
Remark 42.3.6. See also Kohel [Koh96, Theorem 45], where the categories are en-
riched with a Frobenius morphism.
Corollary 42.3.7 (Deuring correspondence). There is a bĳection between isomor-
phism classes of supersingular elliptic curves over 𝐹 and the left class set ClsL O0 .
Under this bĳection, if 𝐸 ↔ [𝐼], then End(𝐸) ' OR (𝐼) and Aut(𝐸) ' OR (𝐼) × .
Proof. Take isomorphism classes on both sides of the equivalence in Theorem 42.3.2,
and compare endomorphism groups and automorphism groups. (We had to work
with left O0 -modules in the equivalence of categories, but each isomorphism class of
objects is represented by a left O0 -ideal 𝐼 ⊆ 𝐵.)
42.3.8. From the Eichler mass formula and Corollary 42.3.7 (swapping left for right,
as in Remark 42.3.3), we conclude that
∑︁ 1 ∑︁ 1 𝑝−1
= = (42.3.9)
# Aut(𝐸) #OL (𝐼) × 24
[𝐸 ] [𝐼 ] ∈Cls R O
where the sum on the left is over isomorphism classes of supersingular elliptic curves
over 𝐹 = Fal𝑝 .
Similarly, from the Eichler class number formula (Theorem 30.1.5), the number of
isomorphism classes of supersingular elliptic curves over 𝐹 is equal to

𝑝 − 1 𝜖2 −4 𝜖3 −3
+ 1− + 1− .
12 4 𝑝 3 𝑝
42.4. SUPERSINGULAR ENDOMORPHISM RINGS 777
Remark 42.3.10. We can generalize this setup slightly as follows. Let 𝑀 ∈ Z>0
be coprime to 𝑝, and let 𝐶0 ≤ 𝐸 0 (𝐹) be a cyclic subgroup of order 𝑀. Then
End(𝐸 0 , 𝐶0 ) ' O0 (𝑀) is an Eichler order of level 𝑀 and reduced discriminant
𝑝𝑀 in 𝐵0 . In a similar way as above, one can show that Hom(−, (𝐸 0 , 𝐶0 )) defines
an equivalence of categories between the category of supersingular elliptic curves
equipped with a cyclic 𝑀-isogeny (under isogenies identifying the cyclic subgroups),
to the category of left invertible O0 (𝑀)-modules (under homomorphisms). The mass
formula now reads
∑︁ 1 ∑︁ 1 𝑝−1
= = 𝜓(𝑀).
# Aut(𝐸, 𝐶) #OL (𝐼) 24
[ (𝐸 ,𝐶) ] [𝐼 ] ∈Cls O0 ( 𝑀 )
One can also consider instead the category of cyclic 𝑀-isogenies 𝜙 : 𝐸 → 𝐸 0.

−1, −11
Example 42.3.11. Consider 𝑝 = 11. The algebra 𝐵 = has discriminant 11
Q
and the maximal order O = Zh𝑖, (1 + 𝑗)/2i. We have # Cls O = 2, with the nontrivial
class represented by the ideal 𝐼 generated by 2 and 1 + 𝑖(1 + 𝑗)/2.
We have O× = h𝑖i of order 4 and OL (𝐼) = h1/2 − 𝑖(1 + 𝑗)/4i of order 6, and indeed
1/4 + 1/6 = 10/24 = 5/12. The two supersingular curves modulo 11 are the ones with
𝑗-invariants 0 and 1728 ≡ 1 (mod 11), and End(𝐸) ' O if 𝑗 (𝐸) = 1728 whereas for
End(𝐸 0) ' O0 we have Hom(𝐸, 𝐸 0) ' 𝐼, in other words, 𝐸 0 ' 𝐸/𝐸 [𝐼].
Example 42.3.12. We return to Example 41.1.2. The order O = O1 is the endomor-
phism ring of the elliptic curve 𝐸 1 : 𝑦 2 = 𝑥 3 − 𝑥 with 𝑗 (𝐸 1 ) = 1728 ≡ 3 (mod 23),
and similarly 𝐸 2 : 𝑦 2 = 𝑥(𝑥 − 1) (𝑥 + 2) with 𝑗 (𝐸 2 ) = 19 and 𝐸 3 : 𝑦 2 = 𝑥 3 + 1 with
𝑗 (𝐸 3 ) = 0. We have 2𝑤 𝑖 = # Aut(𝐸 𝑖 ) is the order of the automorphism group of 𝐸 𝑖 .
And the 𝑝-Brandt graph is the graph of 𝑝-isogenies among the three supersingular
elliptic curves over F23 .
42.3.13. Finally, and most importantly, in the above correspondence the entries of
the Brandt matrix 𝑇 (𝑛) have meaning as counting isogenies. For 𝑛 coprime to 𝑝, the
entry 𝑇 (𝑛)𝑖 𝑗 is equal to the number of subgroups 𝐻 ≤ 𝐸 𝑖 (𝐹) such that 𝐸 𝑖 /𝐻 ' 𝐸 𝑗 .
This statement is just a translation of Lemma 42.2.22. Gross [Gro87] gives a beautiful
and essentially self-contained presentation of the results of the previous chapter in the
special case that disc 𝐵 = 𝑝.
Remark 42.3.14. The approach via supersingular elliptic curves connects back in
another way: Serre [Ser96] gives an alternative approach to modular forms modulo
𝑝 in a letter to Tate: one evaluates classical modular forms at supersingular elliptic
curves and then relates these to quaternionic modular forms modulo 𝑝.
42.4 Supersingular endomorphism rings

In this section, we give a second categorical perspective, giving a base-object free
refinement of Corollary 42.3.7 following Ribet [Rib1989, p. 360–361] (who credits
Mestre–Oesterlé). To get there, we need to deal with a small subtlety involving the
field of definition (fixed by keeping track of extra data). Recall that 𝐹 = Fal𝑝 .
Lemma 42.4.1. Let O be a maximal order. Then there exist one or two supersingular
elliptic curves 𝐸 up to isomorphism over 𝐹 such that End(𝐸) ' O. There exist two
such elliptic curves if and only if 𝑗 (𝐸) ∈ F 𝑝2 r F 𝑝 if and only if the unique two-sided
ideal of O of reduced norm 𝑝 is not principal.
Proof. In Corollary 42.2.21, we proved that there is always at least one supersingular
elliptic curve 𝐸 with End(𝐸) ' O using a connecting ideal. We now elaborate on this
point, refining our count.
By Corollary 42.3.7, the isomorphism classes of supersingular elliptic curves are
in bĳection with the left class set ClsL O0 ; their endomorphism rings are then given by
End(𝐸) ' OR (𝐼) for [𝐼] ∈ ClsL O0 . By Lemma 17.4.13 (interchanging left for right),
the map
ClsL O0 → Typ O0
(42.4.2)
[𝐼] L ↦→ class of OR (𝐼)
is a surjective map of sets. The connecting ideals are precisely the fibers of this
map, and by the bĳection of Corollary 42.3.7, there is a bĳection between the set of
supersingular elliptic curves 𝐸 with End(𝐸) ' O and the fiber of this map over the
isomorphism class of O.
We now count these fibers. We recall Theorem 18.1.3 with 𝐷 = 𝑝 and the
text that follows (interchanging left for right): the fibers are given by the quotient
group PIdl O\Idl O of the two-sided invertible fractional two-sided O-ideals by the
subgroup of principal such ideals. There is a surjection Pic(O) → PIdl O\Idl O and
Pic(O) ' Z/2Z is generated by the unique maximal two-sided ideal 𝑃 of reduced norm
𝑃. The class of 𝑃 in the quotient is trivial if and only if 𝑃 = O𝜋 is principal.
To conclude we recall 42.2.4: the Frobenius map is the map 𝐸 → 𝐸 𝑃 ' 𝐸 ( 𝑝) . So
𝑃 is principal if and only if 𝐸 ( 𝑝) ' 𝐸 if and only if 𝑗 (𝐸) = 𝑗 (𝐸 ( 𝑝) ) = 𝑗 (𝐸) 𝑝 if and
only if 𝑗 (𝐸) ∈ F 𝑝 .
We dig into this issue a bit further.
42.4.3. Let 𝐸 be a supersingular elliptic curve over Fal𝑝 . Let 𝜔 be a nonzero invariant
differential on 𝐸. Then there is a ring homomorphism
𝑎 : End(𝐸) → Fal𝑝
(42.4.4)
𝜙 ↦→ 𝑎 𝜙 , where 𝜙∗ 𝜔 = 𝑎 𝜙 𝜔
(see Silverman [Sil2009, Corollary III.5.6]) independent of the choice of 𝜔.
In light of 42.4.3, we make the following definitions. Let 𝐵 be a quaternion algebra
over Q of discriminant disc 𝐵 = 𝑝; such an algebra 𝐵 is unique up to isomorphism.
Let O ⊆ 𝐵 be a maximal order in 𝐵; then discrd O = 𝑝.
Definition 42.4.5. An orientation of O is a ring homomorphism O → Fal𝑝 .
42.4.6. We claim that there are two possible orientations of O. In fact, 𝑃 = [O, O]
is the commutator ideal (cf. Exercise 13.8) and an orientation factors through the
commutator. Since nrd(𝑃) = 𝑝, localizing we have O/𝑃 ' O 𝑝 /𝑃 𝑝 ' F 𝑝2 by Theorem
13.3.11(c). The claim follows as there are two possible inclusions F 𝑝2 ↩→ Fal𝑝 .
The notion of isomorphism of oriented maximal orders is evident.

Definition 42.4.7. An isomorphism of oriented maximal orders from (O, 𝜁) to (O0, 𝜁 0)
is an isomorphism of orders 𝜙 : O → O0 such that 𝜁 0 𝜙 = 𝜁.
We define the set of reduced isomorphisms to be Isom(𝐸, 𝐸 0)/{±1}.
Proposition 42.4.8. The association 𝐸 ↦→ (End(𝐸) ⊆ End(𝐸)Q , 𝑎) is functorial and
induces an equivalence from the category of
supersingular elliptic curves over Fal𝑝 , under reduced isomorphisms
to the category of
oriented maximal orders (O ⊆ 𝐵, 𝜁)
in a quaternion algebra 𝐵 of discriminant 𝑝,
under isomorphisms.
In the latter category, we do not choose the representative of the isomorphism
class of quaternion algebra of discriminant 𝑝; it is tagging along only to provide a
quaternionic wrapper for the order.
Proof. The association has the right target by Theorem 42.1.9 for the order and 42.4.3
for the orientation. This association is (covariantly) functorial with respect to isomor-
phisms. Indeed, if 𝜓 : 𝐸 −→∼ 𝐸 0 is an isomorphism of elliptic curves, then we have an
induced isomorphism
End(𝐸) → End(𝐸 0)
(42.4.9)
𝜙 ↦→ 𝜓𝜙𝜓 −1 .
that is compatible with composition. The isomorphism (42.4.9) is also compatible
with orientations, as follows. Let 𝜔 0 be a nonzero invariant differential on 𝐸 0; then
𝜓 ∗ 𝜔 0 is so on 𝐸. Thus for all 𝜙 ∈ End(𝐸), we have
𝑎 𝜙 𝜓 ∗ 𝜔 0 = 𝜙∗ 𝜓 ∗ 𝜔 0 = 𝜓 ∗ (𝜓 ∗ ) −1 𝜙∗ 𝜓 ∗ 𝜔 0 = 𝜓 ∗ (𝜓𝜙𝜓 −1 ) ∗ 𝜔 0
(42.4.10)
= 𝜓 ∗ (𝑎 0𝜓 𝜙 𝜓−1 𝜔 0) = 𝑎 0𝜓 𝜙 𝜓−1 𝜓 ∗ 𝜔 0
so 𝑎 0𝜓 𝜙 𝜓−1 = 𝑎 𝜙 , which is the desired compatibility.

The functor is essentially surjective, which is to say that every oriented maximal
order arises up to isomorphism: that every maximal order arises is a consequence of
Corollary 42.2.21, and that the orientation may be so chosen corresponds to applying
the Frobenius morphism, by 42.2.4 and 42.4.6.
Finally, we show the map is fully faithful, which is to say the map of finite sets
Isom(𝐸, 𝐸 0)/{±1} → Isom((End 𝐸, 𝜁), (End 𝐸 0, 𝜁 0))
from (42.4.9) is bĳective. Any two reduced isomorphisms on the left differ by an
automorphism of 𝐸, and the same on the right, so it suffies to show that the map
Aut(𝐸)/{±1} → Aut((End(𝐸), 𝜁))

(42.4.11)
𝜈 ↦→ (𝜙 ↦→ 𝜈𝜙𝜈 −1 )
is bĳective. Let O = End(𝐸), so Aut(𝐸) ' O× . Then Aut(O) ' 𝑁 𝐵× (O)/Q× , and
1 → O× /{±1} → Aut(O) → AL(O) → 1
where the Atkin–Lehner group AL(O) is nontrivial (and isomorphic to Z/2Z) if and
only if 𝑃 = 𝜋O is principal; but conjugation by 𝜋 acts nontrivially on O/𝑃 and fails to
commute with 𝜁 so does not act by an automorphism of (End(𝐸), 𝜁); thus
Aut(𝐸)/{±1} ' O× /{±1} ' Aut(O, 𝜁).
(We did not need to choose a base object in order to define this equivalence!)
Corollary 42.4.12. There is a bĳection between isomorphism classes of supersingular

elliptic curves over 𝐹 and oriented maximal orders in a quaternion algebra 𝐵 of
discriminant 𝑝.
Proof. Again we take isomorphism classes of objects in Proposition 42.4.8.
Exercises
1. Let 𝐸, 𝐸 0 be elliptic curves over 𝐹 and suppose 𝐸, 𝐸 0 are isogenous (necessarily
by a nonzero isogeny). Show that 𝐸 is supersingular if and only if 𝐸 0 is
supersingular. [Hint: show dimQ End(𝐸 𝐹 al )Q = dimQ End(𝐸 𝐹0 al )Q ; or show
that deg𝑖 ( [ 𝑝]) = deg𝑖 ( [ 𝑝] 0) where [ 𝑝], [ 𝑝] 0 are multiplication by 𝑝 on 𝐸, 𝐸 0.]
2. Let 𝐸 be an elliptic curve over 𝐹 with char 𝐹 = 𝑝. Show that for all 𝜙, 𝜓 ∈
End(𝐸), we have
deg𝑖 (𝜙𝜓) = deg𝑖 (𝜙) deg𝑖 (𝜓)

deg𝑖 (𝜙 + 𝜓) ≥ min{deg𝑖 𝜙, deg𝑖 𝜓}.
Conclude that |𝜙| = 1/deg𝑖 (𝜙) defines a nonarchimedean absolute value on

End(𝐸) ( 𝑝) .
3. In this exercise, we give an alternate “hands on” proof of Lemma 42.2.9.
Let 𝐸 be a supersingular elliptic curve over 𝐹 = Fal𝑝 , let O = End(𝐸) and let
𝐵 = O ⊗ Q. Let 𝐼 ⊆ O be a nonzero integral left O-ideal, and let 𝜙 𝐼 : 𝐸 → 𝐸 𝐼
where 𝐸 𝐼 = 𝐸/𝐸 [𝐼]. Consider the pullback isomorphism 𝜙∗𝐼 : Hom(𝐸 𝐼 , 𝐸) →
𝐼 by 𝜓 ↦→ 𝜓𝜙 𝐼 in (42.2.8).
(a) Show that 𝜙∗ induces an isomorphism of O-module endomorphism rings
∼ End(𝐼)
𝜌 : End(Hom(𝐸 𝐼 , 𝐸)) −→
𝛼 ↦→ 𝜙∗ 𝛼𝜙∗−1 .
(b) Show that End(𝐼) = OR (𝐼) op .

(c) Show that End(Hom(𝐸 𝐼 , 𝐸)) = End(𝐸 𝐼 ) op . [Hint: End(𝐸 𝐼 ) is a maximal
order, so the natural inclusion is an equality.]
(d) Conclude that 𝜌 op induces an isomorphism End(𝐸 𝐼 ) −→∼ O (𝐼).
R
(e) Show that 𝜌 op is the map 𝜄 in (42.2.10).

4. Let 𝐵 be a quaternion algebra over Q with disc 𝐵 = 𝑝 prime. Recall the definition
of the Brandt matrix (e.g. (41.1.1), and more generally 41.2.2).
Let 𝐼𝑖 for 𝑖 = 1, . . . , ℎ be the representatives of the left class set ClsL O, with
ℎ = # ClsL O. Let 𝐸 𝑖 be the supersingular elliptic curves over Fal𝑝 corresponding
to 𝐼𝑖 in Corollary 42.3.7.
(a) For every 𝑚 ≥ 1, show that 𝑇 (𝑚)𝑖 𝑗 is equal to the number of subgroup
schemes 𝐶 of order 𝑚 in 𝐸 𝑗 such that 𝐸 𝑗 /𝐶 ' 𝐸 𝑖 .
(b) Show that 𝑇 ( 𝑝) is a permutation matrix of order dividing 2 and that
𝑇 ( 𝑝 𝑟 ) = 𝑇 ( 𝑝) 𝑟 for all 𝑟 ≥ 1.
(c) Show that 𝐸 𝑖 is conjugate by an element of Aut(Fal𝑝 ) to 𝐸 𝑗 if and only if
𝑖 = 𝑗 or 𝑇 ( 𝑝)𝑖 𝑗 = 1. Conclude that the number of elliptic curves 𝐸 𝑖 defined
over F 𝑝 is equal to tr 𝑇 ( 𝑝).
Chapter 43
QM abelian surfaces
In this final chapter, we consider quotients of the upper half-plane by quaternionic

unit groups as generalizations of such quotients from the matrix group (Chapter 40),
realizing them as moduli spaces for abelian surfaces with quaternionic multiplication.
This chapter can be seen as a culminating application of all of the parts of this book,
and for that reason, is necessarily more advanced. Concepts are reviewed in the attempt
to be self-contained, but additional background in algebraic and arithmetic geometry
is suggested.
43.1 ⊲ QM abelian surfaces

Recall (40.1.1) that the curve SL2 (Z)\H2 parametrizes complex elliptic curves up to
isomorphism: to 𝜏 ∈ H2 , we associate the lattice Λ 𝜏 := Z + Z𝜏 and the elliptic curve
𝐸 𝜏 := C/Λ 𝜏 , and the association
SL2 (Z)\H2 ↔ {Complex elliptic curves up to isomorphism}

(43.1.1)
SL2 (Z)𝜏 ↦→ [𝐸 𝜏 ]
is bĳective. Moreover, we have a biholomorphic map 𝑗 : SL2 (Z)\H2 → C, which is

to say, two complex elliptic curves are isomorphic if and only if they have the same
𝑗-invariant. We compactify to 𝑋 := SL2 (Z)\H2∗ by including the cusp at ∞.
As in section 38.1, we are led to seek a generalization of (43.1.1), replacing
𝐵 = M2 (Q) with a quaternion algebra. To this end, let 𝐵 be an indefinite quaternion
algebra over Q of discriminant 𝐷, let O ⊂ 𝐵 be a maximal order, and let
𝜄∞ : 𝐵 → 𝐵 ⊗Q R ' M2 (R)
be an embedding (explicitly, we may take (38.1.1)). The order O is unique up to

conjugation in 𝐵 (by strong approximation) and similarly the embedding 𝜄∞ is unique
up to conjugation in M2 (R), so these choices are harmless. Let
Γ1 (O) := 𝜄∞ (O1 )/{±1} ≤ PSL2 (R).
783
784 CHAPTER 43. ABELIAN SURFACES WITH QM
The quotient Γ1 (O)\H2 is compact when 𝐵 ; M2 (Q); for uniformity, we define
𝑋 1 := Γ1 (O)\H2(∗) ,
where H2(∗) = H2∗ , H2 according as 𝐷 = 1 or 𝐷 > 1. Then 𝑋 1 is a good (compact)

complex 1-orbifold.
We may then ask: what does 𝑋 1 parametrize? The answer is, roughly: 𝑋 1
parametrizes complex abelian surfaces with endomorphisms by O. The correspon-
dence itself is as pleasingly simple as for elliptic curves (43.1.1). To a point 𝜏 ∈ H2 ,
we associate
𝜏
Λ 𝜏 := 𝜄∞ (O) ⊆ C2
1
(43.1.2)
𝐴 𝜏 := C2 /Λ 𝜏
𝜄 𝜏 := 𝜄∞ : O ↩→ End( 𝐴 𝜏 )
Then 𝐴 𝜏 is a complex torus of dimension 2 and 𝜄 𝜏 is an injective ring homomorphism,
realizing endomorphisms of 𝐴 𝜏 by O.
However, there are a number of technical points required to make this completely
precise. We quickly survey the theory of complex abelian varieties in section 43.4. One
basic fact of life is that not every complex torus has enough meromorphic functions to
give it the structure of a complex abelian variety embedded in projective space. One
needs a polarization given by a Riemann form, and the simplest polarizations are the
principal polarizations. (One can think of this rigidification as the difference between
a genus 1 curve and an elliptic curve, where the genus 1 curve is equipped with a
point.) A principal polarization defines positive involution on the endomorphism ring,
called the Rosati involution.
This rigidification is matched on the quaternion order: a principal polarization
on O is an element 𝜇 ∈ O such that 𝜇2 + 𝐷 = 0. Every (maximal) order has a principal
polarization, and the involution 𝛼 ↦→ 𝛼∗ = 𝜇−1 𝛼𝜇 is a positive involution on O. A
quaternionic multiplication (QM) structure by (O, 𝜇) on a principally polarized
complex abelian surface 𝐴 is an injective ring homomorphism O ↩→ End( 𝐴) that
respects the positive involutions on 𝐵 and End( 𝐴)Q .
The happy fact is that 𝐴 𝜏 as defined in (43.1.2) has via 𝜇 a principal polarization
and thereby QM by (O, 𝜇). In other words, the choice of the QM structure determines a
canonical principal polarization: but it gives a finite amount of additional information,
as there will in general be more than one QM structure on a principally polarized
abelian surface. In many cases, these structures can be understood in terms of the
Atkin–Lehner group
Ö
AL(O) = 𝑁 𝐵× (O)/Q× O× ' Z/2Z (43.1.3)
𝑝 |𝐷
acting by automorphisms of 𝑋 1 .
In any event, the main result of this chapter (Main Theorem 43.6.14) is that this
association is bĳective.
43.1. ⊲ QM ABELIAN SURFACES 785
Main Theorem 43.1.4. The map

 ( 𝐴, 𝜄) principally polarized 

complex abelian surfaces 


 

1 2
Γ (O)\H ↔

 with QM by (O, 𝜇) 
 (43.1.5)
up to isomorphism
 
 
1
Γ (O)𝜏 ↦→ [( 𝐴 𝜏 , 𝜄 𝜏 )]
is a bĳection.
This main theorem

generalizes (43.1.1): indeed, we may take 𝐵 = M2 (Q) ⊃ O =
0 1
M2 (Z) and 𝜇 = , and we find that 𝐴 𝜏 ' 𝐸 𝜏2 as principally polarized abelian
1 0
surfaces.
One feature that makes this theory even more appealing is that abelian surfaces
arise naturally as Jacobians of genus 2 curves via the Abel–Jacobi map: this motivates
much of the theory, so we begin with it in section 43.3. In particular, there are
functions called Igusa invariants analogous to the elliptic 𝑗-function that record the
isomorphism class of a principally polarized abelian surface.
43.1.6. We then define modular forms as for the classical modular group. Let 𝑘 ∈
2Z ≥0 . A map 𝑓 : H2 → C is weight 𝑘-invariant under Γ = Γ1 (O) if

𝑎 𝑏
𝑓 (𝛾𝑧) = (𝑐𝑧 + 𝑑) 𝑘 𝑓 (𝑧) for all 𝛾 = ∈ Γ. (43.1.7)
𝑐 𝑑
A modular form for Γ of weight 𝑘 is a holomorphic function 𝑓 : H2 → C that is

weight 𝑘 invariant and is holomorphic at ∞, if Γ = PSL2 (Z). Let 𝑀𝑘 (Γ) be the
C-vector space of modular forms for Γ. Then 𝑀𝑘 (Γ) is a finite-dimensional C-vector
space, and by a similar contour integration as in the proof of Proposition 40.3.4,
dimC 𝑀𝑘 (Γ) can be expressed in terms of 𝑘 and the signature of Γ. And
Ê
𝑀 (Γ) := 𝑀𝑘 (Γ) (43.1.8)
𝑘 ∈2Z≥0
has the structure of a graded C-algebra under multiplication. (When 𝐷 > 1, there are
no cusps, so vacuously all modular forms are cusp forms.)
It would not be unreasonable for us to have started the book here, with this topic
at front and center. In this chapter, we will do our best to treat the complex analytic
theory in as complete and self-contained a manner as possible, but this is really just the
beginning of the subject, one that is rich, deep, and complicated—worthy of a book
all to itself. For example, the following result is fundamental.
Theorem 43.1.9 (Shimura [Shi67, p. 58]). There exists a projective nonsingular curve
𝑋 1 defined over Q and a biholomorphic map
∼ 𝑋 1 (C).
𝜑 : Γ1 (O)\H2 −→
The curve 𝑋 1 over Q coarsely represents the functor from schemes over Q to
sets whose values are isomorphism classes of QM abelian schemes, suitably defined.
Moreover, the map 𝜑 respects the field of definition and Galois action on certain
special points called CM points on Γ1 (O)\H2 obtained as fixed points of elements
𝜈 ∈ 𝐵× with Q(𝜈) an imaginary quadratic field. As a result, the curve 𝑋 1 is canonical,
uniquely characterized up to isomorphism, and is so called the canonical model. We
give some indications of this result by example in the next section and more generally
in section 43.8.
43.2 ⊲ QM by discriminant 6
For concreteness, before embarking on our general treatment, we consider in this
section an illustrative example and one of special interest; it is well-studied and
beloved by quaternionic
practitioners, see Remark 43.2.21 for further reference.
−1, 3
Let 𝐵 = be the quaternion algebra of discriminant 6 studied in sections
Q
37.8–37.9. As in 37.8.12, we have a maximal order
1 + 𝑖 + 𝑗 + 𝑖𝑗
O = Z ⊕ Z𝑖 ⊕ Z 𝑗 ⊕ Z𝑘, 𝑘=
2
with 𝑘 2 − 𝑘 − 1 = 0, and an embedding
𝜄∞ : 𝐵 ↩→ M2 (R)
√
0 −1 3 0
√
𝑖, 𝑗 ↦→ ,
1 0 0 − 3
Let Γ1 = 𝜄∞ (O1 )/{±1} ≤ PSL2 (R) and 𝑋 1 = Γ1 \H2 . We computed a compact
Dirichlet fundamental domain ◊ for Γ1 in 37.9.4, with 𝜇(◊) = 2𝜋/3. Further, we saw
explicitly in 37.9.10 (and again by formula in Example 39.4.21) that Γ1 has signature
(0; 2, 2, 3, 3); that is, 𝑋 1 has topological genus 𝑔 = 0 and there are 4 cone points, two
points 𝑧 2 , 𝑧20 ∈ ◊ with stabilizer of order 2 and two 𝑧3 , 𝑧30 ∈ ◊ with order 3 stabilizer.
As in Chapter 40, to exhibit a model for 𝑋 1 we seek modular forms, indeed, we
now describe the full graded ring of (even weight) modular forms (43.1.8). We will
use the following essential proposition.
(a) Let 𝑓 : H2 → C be a nonzero meromorphic modular form of weight 𝑘 for Γ1 ,
not identically zero. Then
∑︁ 1 𝑘
ord𝑧 ( 𝑓 ) = .
1 1 2
# StabΓ 1 (𝑧) 6
Γ 𝑧 ∈Γ \H
(b) We have


 1, if 𝑘 = 0;
1


dimC 𝑀𝑘 (Γ ) = 0, if 𝑘 = 2; (43.2.2)

 1 − 𝑘 + 2b𝑘/4c + 2b𝑘/3c,

if 𝑘 ≥ 4.

43.2. ⊲ QM BY DISCRIMINANT 6 787
Proof. See Theorem 43.9.4; for the purposes of this introduction, we provide a sketch
to tide the reader over. For (a), we argue just as in Proposition 40.3.4: we integrate
dlog 𝑓 = d 𝑓 / 𝑓 over the boundary of the fundamental domain ◊ and use the identifica-
tion of sides provided by rotation at their fixed points (elliptic vertices), reversing the
direction of the path so the contributions cancel, and we are left again to sum angles.
The details are requested in Exercise 43.7. For (b), we can get upper bounds on the
dimension using (a), but to provide lower bounds we need to exhibit modular forms,
and these are provided by the Riemann–Roch theorem. For example, for 𝑘 = 2, we
have dimC 𝑀2 (Γ1 ) = 𝑔 = 0 by (40.2.11).
43.2.3. We are now in a position to prove an analogous statement to Theorem 40.3.11.

Referring to Proposition (43.2.1), by part (a) we seek 𝑎 1 , 𝑎 2 , 𝑎 20 , 𝑎 3 , 𝑎 30 ∈ Z ≥0 with
𝑎 2 + 𝑎 20 𝑎 3 + 𝑎 30 𝑘
𝑎1 + + = . (43.2.4)
2 3 6
By part (b), we have dimC 𝑀𝑘 (Γ1 ) = 0 for 𝑘 < 0, and indeed, there are no solutions.
For 𝑘 = 0, there is a unique solution corresponding to the constant functions. For
𝑘 = 2, there are no solutions, as follows. Let 𝑓 (𝑧) ∈ 𝑀2 (Γ1 ). Let 𝛾3 be a generator
for the stabilizer at 𝑧 3 . Then
𝑓 (𝑧3 ) = 𝑓 (𝛾3 𝑧3 ) = 𝚥 (𝛾3 ; 𝑧3 ) 2 𝑓 (𝑧 3 );
by the cocycle relation, we have 1 = 𝚥 (𝛾33 ; 𝑧3 ) = 𝚥 (𝛾3 ; 𝑧3 ) 3 a nontrivial cube root of

unity, so 𝑓 (𝑧3 ) = 0 and 𝑎 3 > 0. Similarly 𝑎 30 > 0, and this contradicts (43.2.4).
Arguing in the same way, we find that the unique solution for 𝑘 = 4 is 𝑎 2 = 𝑎 20 = 0
and 𝑎 3 = 𝑎 30 = 1; thus 𝑀4 (Γ1 ) = C 𝑓4 , and 𝑓4 necessarily vanishes at 𝑧 2 , 𝑧20 . Similarly,
for 𝑘 = 6 we have only 𝑎 3 = 𝑎 30 = 0 and 𝑎 2 = 𝑎 20 = 1, with 𝑀6 (Γ1 ) = C𝑔6 .
Continuing as in 43.2.3, we collect dimensions and spanning functions as in Table

43.2.5.
𝑘 dimC 𝑀𝑘 (Γ1 ) Spanning functions

0 1 1
2 0 -
4 1 𝑓4
6 1 𝑔6
8 1 𝑓42
10 1 𝑓4 𝑔6
12 3 𝑓43 , 𝑔62 , ℎ12
.. .. ..
. . .
24 5 𝑓46 , 𝑓43 𝑔62 , 𝑓43 ℎ12 , 𝑔64 , 𝑔62 ℎ12 , ℎ212
Table 43.2.5: Generators for 𝑀 (Γ)
43.2.6. In weights 8, 10, we have products of forms seen previously. In weight 𝑘 = 12,
we find a third function ℎ12 ∈ 𝑀12 (Γ1 ) spanning together with 𝑓43 , 𝑔62 . Continuing in
this way, finally in weight 𝑘 = 24, we find 6 functions in a 5 dimensional space, and so
they must satisfy an equation 𝑟 ( 𝑓4 , 𝑔6 , ℎ12 ) ∈ C[ 𝑓4 , 𝑔6 , ℎ12 ], homogeneous of degree
24 if we give 𝑓4 , 𝑔6 , ℎ12 the weights 4, 6, 12.
C[ 𝑓4 , 𝑔6 , ℎ12 ]
𝑀 (Γ1 ) ' .
h𝑟 ( 𝑓4 , 𝑔6 , ℎ12 )i
Proof. The bound on the degrees of generators and relations in Theorem 43.9.6 makes
this proposition immediate. It is also possible to give a proof with bare hands: see
Exercise 43.9.
43.2.8. We do not have Eisenstein series available in this setting, but the notion of
taking averages 40.1.19 is still quite sensible:
we
find what are known as Poincaré
𝑎 𝑏
series. Recall 𝚥 (𝛾; 𝑧) = 𝑐𝑧 + 𝑑 for 𝛾 = ∈ SL2 (R). The square 𝚥 (𝛾; 𝑧) 2 is
𝑐 𝑑
well-defined on 𝛾 ∈ PSL2 (R).
For 𝑘 ∈ 2Z ≥2 , we define the Poincaré series
∑︁
𝑃 𝑘 (𝑧) = 𝚥 (𝛾; 𝑧) −𝑘 .
𝛾 ∈Γ1
Then 𝑃 𝑘 (𝑧) is nonzero, absolutely convergent on H2 , uniformly on compact subsets,

and satisfies
𝑃 𝑘 (𝛾𝑧) = 𝚥 (𝛾; 𝑧) 𝑘 𝑃 𝑘 (𝑧). (43.2.9)
The convergence statement is fairly tame, because the fundamental domain ◊ is com-
pact: it implies that the total integral
∫
(Im 𝑧) 𝑘
d𝜇(𝑧) < ∞
◊ | 𝚥 (𝛾; 𝑧)|
is finite, and the Poincaré series converges (absolutely) by comparison [Kat85, §1,
Proposition 1]. The equality (43.2.9) follows from the cocycle relation (40.2.5).
Therefore 𝑃 𝑘 (𝑧) ∈ 𝑆 𝑘 (Γ), and in particular we may take 𝑓4 = 𝑃4 and 𝑔6 = 𝑃6 ; with a
bit more computation, one can also show that 𝑃43 , 𝑃62 , 𝑃12 are linearly independent, so
that we may take ℎ12 = 𝑃12 as well.
43.2.10. A convenient and meaningful normalization of the functions above is given
by Baba–Granath [BG2008, §3.1]. √
√ embeddings 𝑆 = Z[ −6] ↩→ O
First, there are exactly two (necessarily optimal)
by Example 30.7.5: we have √ # Cls O = 1 and 𝐾 ( −6) is ramified at 𝑝 = 2, 3 | 𝐷 = 6,
so 𝑚(𝑆, O; O× ) = ℎ(Z[ −6]) = 2. The fixed points of these two embeddings are
distinct points 𝑧6 , 𝑧60 ∈ ◊. Explicitly, we note that
𝜇 = 3𝑖 + 𝑖 𝑗 = −1 + 2𝑖 − 𝑗 + 2𝑘 (43.2.11)
has 𝜇2 + 6 = 0, and choose its fixed point as 𝑧6 .

43.2. ⊲ QM BY DISCRIMINANT 6 789
√
We rescale 𝑔6 so that 𝑓43 (𝑧 6 )/𝑔62 (𝑧6 ) = −3, and we choose ℎ12 such that ℎ12 (𝑧6 ) =
ℎ12 (𝑧60 ) = 0, and rescale so that
𝑟 ( 𝑓4 , 𝑔6 , ℎ12 ) = ℎ212 + 3𝑔64 + 𝑓46 = 0. (43.2.12)
Corollary 43.2.13. The holomorphic map
Γ1 (O) → P2
(43.2.14)
𝑧 ↦→ ( 𝑓43 (𝑧) : 𝑔62 (𝑧) : ℎ12 (𝑧))
has image the conic 𝑋 1 : 𝑥 2 + 3𝑦 2 + 𝑧 2 = 0, defining the canonical model over Q.
Proof. This result is attributed to Ihara by Kurihara [Kur79, Theorem 1-1(1)]; it is

proven by Baba–Granath [BG2008, Theorem 3.10] along the lines above.
We note that 𝑋 1 (R) = ∅; this is a general feature, see Proposition 43.7.2.
43.2.15. The Atkin–Lehner group
AL(O) = 𝑁 𝐵× (O)/Q× O× ' (Z/2Z) 2
has three nontrivial involutions 𝑤 2 , 𝑤 3 , 𝑤 6 having positive reduced norm. Explicitly,

we have 𝑤 6 = 𝜇 by (43.2.11) and 𝑤 2 = 1 + 𝑖 and 𝑤 3 = 𝑤 6 /𝑤 2 = 1 + 𝑖 − 𝑗 + 𝑘. These
involutions act on the space of modular forms as follows [BG2008, §3.1]:
𝑤2 𝑤3 𝑤6
𝑓4 − − +
𝑔6 + − −
ℎ12 − + −
So for example 𝑓4 (𝑤 2 𝑧) = − 𝑓4 (𝑧).

These involutions act on the canonical model 𝑋 1 by 𝑤 2 (𝑥 : 𝑦 : 𝑧) = (𝑥 : −𝑦 : 𝑧),
𝑤 3 (−𝑥 : 𝑦 : 𝑧), and 𝑤 6 (𝑥 : 𝑦 : 𝑧) = (𝑥 : 𝑦 : −𝑧).
We choose a principal polarization (see Definition 43.6.4) on O by 𝜇 in (43.2.11).

In this way, Main Theorem 43.1.4 provides that the curve 𝑋 1 parametrizes abelian
surfaces with QM by (O, 𝜇).
43.2.16. The forgetful map [( 𝐴 𝜏 , 𝜄 𝜏 )] ↦→ [ 𝐴 𝜏 ] which forgets the QM structure is the

map [BG2008, Proposition 3.9]
𝑗 : 𝑋 1 → P1
16𝑦 2 (43.2.17)
(𝑥 : 𝑦 : 𝑧) ↦→
9𝑥 2
generically 4-to-1. The map 𝑗 can fruitfully be thought of as an analogue of the classical
elliptic 𝑗-invariant, mindful of the above technicalities: it parametrizes principally
polarized complex abelian surfaces that can be equipped with a QM structure.
The Igusa invariants 43.3.5 of 𝐴 𝑗 where 𝑗 = 𝑗 (𝜏) are given by [BG2008, Proposi-
tion 3.6]
(𝐼2 : 𝐼4 : 𝐼6 : 𝐼10 )
= (12( 𝑗 + 1) : 6( 𝑗 2 + 𝑗 + 1) : 4( 𝑗 3 − 2 𝑗 2 + 1) : 𝑗 3 ) (43.2.18)
∈ P(2, 4, 6, 10).
There exists a genus 2 curve with these Igusa invariants if and only if 𝑗 = 0, −16/27
or the Hilbert symbol
(−6 𝑗, −2(27 𝑗 + 16))Q = 1
is trivial.
Example 43.2.19. The two points with 𝑗 = 0, ∞ are exactly those points which
√ are
not Jacobians of genus 2 curves: these correspond to points with CM by Z[ −1] and
Z[𝜔], and these abelian surfaces are the squares of the corresponding CM elliptic
curves (with the product polarization). Elkies [Elk98, §3] computes equations and
further CM points for discriminant 6.
Example 43.2.20. The case 𝑗 = −16/27 corresponds to a CM point with discriminant
𝐷 = −24 [BG2008, §3.3]: it is the Jacobian of the curve
√ √ √ √
𝑦 2 = (1 + 2)𝑥 6 − 3(7 − 3 2)𝑥 4 − 3(7 + 3 2)𝑥 2 + (1 − 2)
√
isomorphic to the product of the two elliptic curves with CM by Z[ −6] (but not with
the product polarization).
Remark 43.2.21. For further reading to connect some of the dots above, see the article
by Baba–Granath [BG2008], refining the work by Hashimoto–Murabayashi [HM95,
Theorem 1.3] who give an explicit family of genus 2 curves whose Jacobians have QM
by O.
43.3 Genus 2 curves

We begin in the concrete setting of genus 2 curves. Let 𝐹 be a perfect field with
char 𝐹 ≠ 2 and let 𝐹 al be an algebraic closure of 𝐹. Let 𝑋 be a smooth projective
curve of genus 2 over 𝐹.
43.3.1. Using Riemann–Roch in a manner analogous to the proof for elliptic curves
(see e.g. Silverman [Sil2009, Proposition III.3.1(a)]), 𝑋 is given by a Weierstrass
equation of the form
𝑦 2 = 𝑓 (𝑥) (43.3.2)
where 𝑓 (𝑥) ∈ 𝐹 [𝑥] is squarefree of degree 5 or 6. It follows that 𝑋 is hyperelliptic
over 𝐹, with map 𝑥 : 𝑋 → P1 of degree 2.
If (𝑦 0) 2 = 𝑓 0 (𝑥 0) is another Weierstrass equation for 𝑋, then it is related by a
change of variables of the form
𝑎𝑥 + 𝑏 𝑒𝑦
𝑥0 = , 𝑦0 = (43.3.3)
𝑐𝑥 + 𝑑 (𝑐𝑥 + 𝑑) 3
43.3. GENUS 2 CURVES 791
with 𝑎𝑑 − 𝑏𝑐, 𝑒 ∈ 𝐹 × . After such a change of variable, we may suppose without loss
of generality that deg 𝑓 = 6.
Example 43.3.4. Let 𝑋 al be the base change of 𝑋 to 𝐹 al . The automorphism group
Aut(𝑋 al ) is a finite group containing the hyperelliptic involution (𝑥, 𝑦) ↦→ (𝑥, −𝑦).
The possibilities for this group were classified by Bolza [Bol1887, p. 70]: when
char 𝐹 ≠ 2, 3, 5, the group Aut(𝑋 al ) is isomorphic to one of the groups
𝐶2 , 𝑉4 , 𝐷 8 , 𝐶10 , 𝐷 12 , 2𝐷 12 , 𝑆e4
of orders 2, 4, 8, 10, 12, 24, 48. A generic genus 2 curve over 𝐹 al has Aut(𝑋 al ) ' 𝐶2 .
43.3.5. We now seek invariants of the curve defined in terms of a model to classify
isomorphism classes. We factor
6
Ö
𝑓 (𝑥) = 𝑐 (𝑥 − 𝑎 𝑖 )
𝑖=1
with 𝑎 𝑖 ∈ 𝐹 al the roots of 𝑓 . We abbreviate 𝑎 𝑖 − 𝑎 𝑗 by (𝑖 𝑗), and we define

∑︁
𝐼2 := (4𝑐) 2 (1 2) 2 (3 4) 2 (5 6) 2 ,
∑︁
𝐼4 := (4𝑐) 4 (1 2) 2 (2 3) 2 (3 1) 2 (4 5) 2 (5 6) 2 (6 4) 2 ,
∑︁ (43.3.6)
𝐼6 := (4𝑐) 6 (1 2) 2 (2 3) 2 (3 1) 2 (4 5) 2 (5 6) 2 (6 4) 2 (1 4) 2 (2 5) 2 (3 6) 2 ,
Ö
𝐼10 := (4𝑐) 10 (1 2) 2 ,
where each sum and product runs over the distinct expressions obtained by permuting
the index set {1, . . . , 6}; by Galois theory, we have 𝐼2 , 𝐼4 , 𝐼6 , 𝐼10 ∈ 𝐹. In particular, we
have Ö
𝐼10 = (4𝑐) 10 (𝑎 𝑖 − 𝑎 𝑗 ) 2 = disc(4 𝑓 ) ≠ 0
1≤𝑖< 𝑗 ≤6
is the discriminant of the polynomial 4 𝑓 . The invariants 𝐼2 , 𝐼4 , 𝐼6 , 𝐼10 , defined by Igusa

[Igu60, p. 620] by modifying a set of invariants due to Clebsch, are known as the
Igusa-Clebsch invariants.
Under a change of variable of the form (43.3.3), we have
(𝐼20 , 𝐼40 , 𝐼60 , 𝐼10

0
) = (𝛿2 𝐼2 , 𝛿4 𝐼4 , 𝛿6 𝐼6 , 𝛿10 𝐼10 )
where 𝛿 = 𝑒 2 /(𝑎𝑑 −𝑏𝑐) 3 . Accordingly, we define the weighted (2, 4, 6, 10)-projective

space
P(2, 4, 6, 10) (𝐹 al ) := ((𝐹 al ) 4 r {(0, 0, 0, 0)})/∼
under the equivalence relation
(𝐼2 , 𝐼4 , 𝐼6 , 𝐼10 ) ∼ (𝛿2 𝐼2 , 𝛿4 𝐼4 , 𝛿6 𝐼6 , 𝛿10 𝐼10 )
for all 𝛿 ∈ (𝐹 al ) × ; we write equivalence classes (𝐼2 : 𝐼4 : 𝐼6 : 𝐼10 ) ∈ P(2, 4, 6, 10) (𝐹 al ).

Proposition 43.3.7. The genus 2 curves 𝑋 and 𝑋 0 over 𝐹 are isomorphic over 𝐹 al if
and only if
(𝐼2 : 𝐼4 : 𝐼6 : 𝐼10 ) = (𝐼20 : 𝐼40 : 𝐼60 : 𝐼10

0
) ∈ P(2, 4, 6, 10) (𝐹 al ).
Proof. See Igusa [Igu60, Corollary, p. 632].
43.3.8. For arithmetic reasons (in particular to deal with problems in characteristic 2),
Igusa [Igu60, pp. 617ff] defined the invariants [Igu60, pp. 621–622]
𝐽2 := 𝐼2 /8,
𝐽4 := (4𝐽22 − 𝐼4 )/96,
𝐽6 := (8𝐽23 − 160𝐽2 𝐽4 − 𝐼6 )/576, (43.3.9)
𝐽8 := (𝐽2 𝐽6 − 𝐽42 )/4,
𝐽10 := 𝐼10 /4096,
now called the Igusa invariants, with (𝐽2 : 𝐽4 : 𝐽6 : 𝐽8 : 𝐽10 ) ∈ P(2, 4, 6, 8, 10) (𝐹 al ).
Visibly, the Igusa–Clebsch invariants determine the Igusa invariants and vice versa.
Remark 43.3.10. One can also take ratios of these invariants with the same weight and
define (three) absolute invariants analogous to the classical 𝑗-invariant of an elliptic
curve, following Cardona–Nart–Pujolas [CNP2005] and Cardona–Quer [CQ2005].
Example 43.3.11. The locus of genus 2 curves with given automorphism group (cf.
Example 43.3.4) can be described explicitly by the vanishing of polynomials in the
Igusa(–Clebsch) invariants. For example, the unique genus 2 curve up to isomorphism
over 𝐹 al with automorphism group 𝐶10 (when char 𝐹 ≠ 5) is the curve defined by the
equation 𝑦 2 = 𝑥(𝑥 5 − 1) with (𝐼2 : 𝐼4 : 𝐼6 : 𝐼10 ) = (0 : 0 : 0 : 1), with automorphism
group generated by (𝑥, 𝑦) ↦→ (𝜁5 𝑥, −𝜁53 𝑦), where 𝜁5 is a primitive fifth root of unity.
43.3.12. The group Aut𝐹 (𝐹 al ) acts on P(2, 4, 6, 10) (𝐹 al ) in each coordinate:
𝜎(𝐼2 : 𝐼4 : 𝐼6 : 𝐼10 ) = (𝜎(𝐼2 ) : 𝜎(𝐼4 ) : 𝜎(𝐼6 ) : 𝜎(𝐼10 ))
for 𝜎 ∈ Aut𝐹 (𝐹 al ). Given a point 𝑃 ∈ P(2, 4, 6, 10) (𝐹 sep ), we define its field of
moduli 𝑀 (𝑃) to be the fixed field of 𝐹 sep under the stabilizer of 𝑃 under this action.
Just as in the case of ordinary projective space, the field 𝑀 (𝑃) is the minimal field
over which 𝑃 is defined.
43.3.13. In this way, given a genus 2 curve, we have associated invariants of the curve
that determine it up to isomorphism over 𝐹 al . We may also ask the inverse problem:
given Igusa invariants (𝐽 𝑘 ) 𝑘 with 𝐽10 ≠ 0, find a genus 2 curve with the desired
invariants. This problem has been solved explicitly by work of Mestre [Mes91] and
Cardona–Quer [CQ2005].
We give a sketch of the generic case of curves whose only automorphism over 𝐹 al
is the hyperelliptic involution, due to Mestre [Mes91]: in brief, the field of moduli may
not be a field of definition for the desired genus 2 curve, but a quadratic extension will
43.4. COMPLEX ABELIAN VARIETIES 793
always suffice. Abbreviate Q[𝐽] = Q[𝐽2 , 𝐽4 , 𝐽6 , 𝐽8 , 𝐽10 ]. First, Mestre constructs an

explicit ternary quadratic form 𝐿(𝐽) and ternary cubic form 𝑀 (𝐽) defined over Q[𝐽].
Under substitution of generic invariants, the quadratic form 𝐿 (𝐽) defines a quaternion
algebra 𝐵(𝐽) over the field of moduli 𝐹 of the point, and Mestre proves that there exists
a curve 𝑋 over a field 𝐾 ⊇ 𝐹 with the desired Igusa invariants if and only if 𝐾 is a
splitting field for 𝐵(𝐽). The quaternion algebra 𝐵(𝐽) is accordingly called the Mestre
obstruction. Over a field 𝐾 where 𝐵(𝐽) splits, equivalently over a field 𝐾 where the
conic defined by 𝐿 (𝐽) = 0 has a 𝐾-rational point, we can parametrize 𝐿(𝐽) and by
substituting into 𝑀 (𝐽) we obtain a binary sextic form 𝑓 (𝑥, 𝑧) with the property that
𝑦 2 = 𝑓 (𝑥, 1) has the desired invariants.
43.4 Complex abelian varieties

Shifting gears, we pause to briefly recall some basic properties of complex abelian
varieties, needed for our discussion of abelian surfaces. For further reference, see
Birkenhake–Lange [BL2004], Mumford [Mum70], or Swinnerton–Dyer [Swi74].
Definition 43.4.1. A complex torus of dimension 𝑔 ∈ Z ≥1 is a complex manifold of

the form 𝐴 = 𝑉/Λ where 𝑔 = dimC 𝑉 and Λ ⊆ 𝑉 is a lattice of rank 2𝑔. A morphism
of complex tori 𝑉/Λ → 𝑉 0/Λ0 is a C-linear map 𝜙 : 𝑉 → 𝑉 0 such that 𝜙(Λ) ⊆ Λ0.
Let 𝐴 = 𝑉/Λ be a complex torus of dimension 𝑔. Then 𝑉 ' C𝑔 and Λ ' Z2𝑔 so
𝑉/Λ ' (R/Z) 2𝑔 as smooth real manifolds.
43.4.2. Suppose for concreteness (choosing a basis) that 𝑉 = C𝑔 , working with

column vectors. Choose a basis {𝜆 𝑗 } 𝑗=1,...,2𝑔 for Λ with 𝜆 𝑗 = (𝜆𝑖 𝑗 )𝑖t ∈ C𝑔 . The matrix
Π = (𝜆𝑖 𝑗 )𝑖, 𝑗 ∈ Mat𝑔×2𝑔 (C) is called the big period matrix of the lattice Λ (with
respect to the basis {𝜆 𝑗 } 𝑗 ).
A change of basis of C𝑔 corresponds to left multiplication by an element of GL𝑔 (C)
on Π and induces an isomorphism of complex tori. Writing

Π = 𝑃1 𝑃2 , with 𝑃1 , 𝑃2 ∈ GL𝑔 (C)
we have 𝑃2−1 Π = Ω 1 , and Ω = 𝑃2−1 𝑃1 ∈ GL𝑔 (C). Therefore every complex torus

is isomorphic to a torus of the form C𝑔 /(ΩZ𝑔 + Z𝑔 ) for some Ω ∈ GL𝑔 (C), called the
small period matrix.
Definition 43.4.3. A complex torus 𝐴 is a complex abelian variety if there exists a

holomorphic embedding 𝐴 ↩→ P𝑛 (C) for some 𝑛 ≥ 1.
Remark 43.4.4. Every complex torus of dimension 𝑔 = 1 is an abelian variety, indeed

an elliptic curve, by the theory of classical Eisenstein series (see 40.1.11). But the case
𝑔 = 1 is quite special! For a general lattice Λ ⊆ C𝑔 with 𝑔 ≥ 2, there will probably be
no meromorphic functions on C𝑔 /Λ and in particular there will be no way to realize
the torus as a projective algebraic variety.
The conditions under which a complex torus is a complex abelian variety are given
by the following conditions, due to Riemann.
Definition 43.4.5. A matrix Π ∈ Mat𝑔×2𝑔 (C) is a Riemann matrix if there is a

alternating matrix 𝐸 ∈ M2𝑔 (Z)alt with det 𝐸 ≠ 0 such that:
(i) Π𝐸 −1 Πt = 0, and
√ t
(ii) −1Π𝐸 −1 Π∗ is a positive definite Hermitian matrix, where ∗ = denotes
conjugate transpose.
Conditions (i) and (ii) are called the Riemann relations.
Theorem 43.4.6. Let 𝐴 = C𝑔 /(ΠZ2𝑔 ) be a complex torus with Π ∈ Mat𝑔×2𝑔 (C).
Then 𝐴 is a complex abelian variety if and only if Π is a Riemann matrix.
Example 43.4.7. Let 𝑓 (𝑥) ∈ C[𝑥] be a squarefree polynomial of degree 2𝑔 + 1 or
2𝑔 + 2 with 𝑔 ≥ 1. The equation 𝑦 2 = 𝑓 (𝑥) defines an algebraic curve with projective
closure 𝑋 a nonsingular curve over C. A basis of the holomorphic differential 1-forms
on 𝑋 is 𝜔𝑖 = 𝑥 𝑖−1 d𝑥/𝑦 for 𝑖 = 1, . . . , 𝑔.
The set of points 𝑋 (C) has naturally the structure of a compact (connected) Rie-
mann surface of genus 𝑔 ≥ 1. Let 𝛼1 , 𝛽1 , . . . , 𝛼𝑔 , 𝛽𝑔 be a basis of the homology
𝐻1 (𝑋, Z) and suppose that this basis is symplectic: each closed loop 𝛼𝑖 intersects 𝛽𝑖
with (oriented) intersection number 1 and all other intersection numbers are 0, as in
the following standard picture in Figure 43.4.8.
β1
α1 α2 β2
Figure 43.4.8: A standard surface with a symplectic homology basis

The integration pairing
Ω1 × 𝐻1 (𝑋, Z) → C
∫
(𝜔, 𝜐) ↦→ 𝜔
𝜐
is nondegenerate, giving a map 𝐻1 (𝑋, Z) ↩→ HomC (Ω1 , C). Let

( ∫ t )
Λ= 𝜔𝑖 : 𝜐 ∈ 𝐻1 (𝑋, Z) ⊆ C𝑔 .
𝜐
A Z-basis of Λ is given by the integrals with 𝜐 = 𝛼𝑖 , 𝛽𝑖 for 𝑖 = 1, . . . , 𝑔. Let
Jac 𝑋 := HomC (Ω1 , C)/𝐻1 (𝑋, Z) ' C𝑔 /Λ
be the Jacobian of 𝑋. Then Jac 𝑋 is a complex torus of dimension 𝑔. It has big period
matrix Π = 𝑃1 𝑃2 , where
∫ ∫
𝑃1 = 𝜔𝑗 , 𝑃2 = 𝜔𝑗 .
𝛼𝑖 𝑖, 𝑗 𝛽𝑖 𝑖, 𝑗
By cutting open the Riemann surface along the given paths and applying Green’s
theorem, we verify that the big period matrix Π is indeed a Riemann matrix. Therefore
the Jacobian Jac(𝑋) is an abelian variety of genus 𝑔.
We now upgrade the above to a basis-free formulation.
Definition 43.4.9. Let

𝐸: Λ×Λ → Z
be an alternating Z-bilinear map. Let 𝐸 R : 𝑉 × 𝑉 → R is the scalar extension of 𝐸 over

R obtained from RΛ = 𝑉.
We say 𝐸 is a Riemann form for (𝑉, Λ) if the following conditions hold:
(i) 𝐸 R (𝑖𝑥, 𝑖𝑦) = 𝐸 R (𝑥, 𝑦) for all 𝑥, 𝑦 ∈ 𝑉; and

(ii) The map
𝑉 ×𝑉 → R
(𝑥, 𝑦) ↦→ 𝐸 R (𝑖𝑥, 𝑦)
defines a symmetric, positive definite R-bilinear form on 𝑉.
43.4.10. Let 𝐸 be a Riemann form for (𝑉, Λ). If we choose a Z-basis for Λ, we get a
period matrix Π and a matrix for 𝐸 which is a Riemann matrix (satisfying conditions
(i)–(ii) of Definition 43.4.5), and conversely.
Proposition 43.4.11. If 𝐸 is a Riemann form for (𝑉, Λ), then the map
𝐻: 𝑉 ×𝑉 → C
(43.4.12)
𝐻 (𝑥, 𝑦) = 𝐸 R (𝑖𝑥, 𝑦) + 𝑖𝐸 R (𝑥, 𝑦)
is a positive definite Hermitian form on 𝑉 with Im 𝐻 = 𝐸 R .

Conversely, if 𝐻 is a positive definite Hermitian form on 𝑉 such that Im 𝐻 (Λ) ⊆ Z,
then Im 𝐻| Λ is a Riemann form for (𝑉, Λ).
Proof. This proposition can be checked directly, a bit of linear algebra fun: see
Exercise 43.6.
Example 43.4.13. For the torus C/(Z𝑖 + Z), the forms
𝐸 (𝑥 1 + 𝑖𝑥2 , 𝑦 1 + 𝑖𝑦 2 ) = 𝑥2 𝑦 1 − 𝑥 1 𝑦 2
𝐸 R (𝑖(𝑥1 + 𝑖𝑥2 ), 𝑦 1 + 𝑖𝑦 2 ) = 𝑥1 𝑦 1 − (−𝑥2 𝑦 2 ) = 𝑥1 𝑦 1 + 𝑥2 𝑦 2
𝐻 (𝑥, 𝑦) = 𝑥𝑦.
define a Riemann form 𝐸, its associated (symmetric, positive definite) real part, and
its associated (positive definite) Hermitian form 𝐻.
Definition 43.4.14. A complex torus 𝐴 = 𝑉/Λ equipped with a Riemann form is said
to be polarized.
A homomorphism of polarized complex tori is a homomorphism 𝜙 : 𝐴 → 𝐴 0 of
complex tori that respects the polarizations in the sense that the diagram
Λ×Λ
𝐸 /; Z
( 𝜙, 𝜙)
𝐸0
Λ0 × Λ0
commutes.
By Theorem 43.4.6 and Proposition 43.4.11, a polarized complex torus is an
abelian variety, and accordingly we call it a polarized abelian variety.
We now seek to classify the possibilities for Riemann forms.
43.4.15. There is a normal form for alternating matrices, analogous to the Smith
normal form of an integer matrix, called the Frobenius normal form. Let 𝑀 be a free
Z-module of rank 2𝑔 equipped with an alternating form 𝐸 : 𝑀 × 𝑀 → Z. Then there
exists a basis of 𝑀 such that the matrix of 𝐸 in this basis is

0 𝐷
[𝐸] =
−𝐷 0
where 𝐷 = diag(𝑑1 , . . . , 𝑑 𝑔 ) is a diagonal matrix with diagonal entries 𝑑𝑖 ∈ Z ≥0 and

𝑑1 | 𝑑2 | · · · | 𝑑 𝑔 . The integers 𝑑1 , . . . , 𝑑 𝑔 are uniquely determined by 𝐸, and are called
the elementary divisors of 𝐸 when all 𝑑𝑖 > 0 (equivalently, 𝐸 is nondegenerate).
Definition 43.4.16. A Riemann form 𝐸 with elementary divisors 1, . . . , 1 in its Frobe-
nius normal is called a principal Riemann form.
Lemma 43.4.17. Let 𝐴 = 𝑉/Λ be a polarized abelian variety, and suppose the
Riemann form 𝐸 has elementary divisors 𝑑1 , . . . , 𝑑 𝑔 . Then
there is a basis for 𝑉 and a
basis for Λ such that the big period matrix of Λ is Ω 𝐷 where 𝐷 = diag(𝑑1 , . . . , 𝑑 𝑔 ),
and Ω is symmetric and Im Ω is positive definite.
In particular, if 𝐴 is principally polarized, then the conclusion of Lemma 43.4.17
holds for Ω, the small period matrix.
Proof. Compute the period matrix with respect to a basis in which the Riemann form
is in Frobenius normal form.
Example 43.4.18. Let 𝐴1 = 𝑉1 /Λ1 and 𝐴2 = 𝑉2 /Λ2 be two polarized abelian varieties,
with Riemann forms 𝐸 1 , 𝐸 2 . Let 𝐴 = 𝐴1 × 𝐴2 = 𝑉/Λ, where 𝑉 = 𝑉1 ⊕ 𝑉2 and
Λ = Λ1 ⊕ Λ2 ⊆ 𝑉1 ⊕ 𝑉2 = 𝑉. Then 𝐴 can be equipped with the product polarization
𝐸 = 𝐸 1 𝐸 2 , defined by
𝐸 (𝑥 1 + 𝑥2 , 𝑦 1 + 𝑦 2 ) = 𝐸 1 (𝑥 1 , 𝑦 1 ) + 𝐸 2 (𝑥2 , 𝑦 2 ).
If 𝐸 1 , 𝐸 2 are principal, then the product 𝐸 is also principal.

43.4.19. Polarizations can be understood in terms of duality, as follows.

Let 𝐴 = 𝑉/Λ be a complex torus. A C-antilinear functional on 𝑉 is a function
𝑓 : 𝑉 → C such that 𝑓 (𝑥 + 𝑥 0) = 𝑓 (𝑥) + 𝑓 (𝑥 0) for all 𝑥, 𝑥 0 ∈ 𝑉 and 𝑓 (𝑎𝑥) = 𝑎 𝑓 (𝑥)
for all 𝑎 ∈ C and 𝑥 ∈ 𝑉. Let 𝑉 ∗ = HomC (𝑉, C) be the C-vector space of C-
antilinear functionals on 𝑉. Then 𝑉 ∗ is a C-vector space with dimC 𝑉 = dimC 𝑉 ∗ and
the underlying R-vector space of 𝑉 ∗ is canonically isomorphic to HomR (𝑉, R). The
canonical R-bilinear form
𝑉∗ × 𝑉 → R
( 𝑓 , 𝑥) ↦→ Im 𝑓 (𝑥)
is nondegenerate, so
Λ∗ := { 𝑓 ∈ 𝑉 ∗ : Im 𝑓 (Λ) ⊆ Z}
is a lattice in 𝑉 ∗ , called the dual lattice of Λ, and the quotient 𝐴∨ := 𝑉 ∗ /Λ∗ is a
complex torus. Double antiduality and nondegeneracy gives a canonical identification
(𝑉 ∗ ) ∗ 𝑉, giving a canonical identification ( 𝐴∨ ) ∨ 𝐴.
Now suppose 𝐴 is polarized with 𝐸 a Riemann form for (𝑉, Λ), and let 𝐻 be the
associated Hermitian form (43.4.12). Then double duality induces a Riemann form
𝐸 ∗ on (𝑉 ∗ , Λ∗ ), so 𝐴∨ is a polarized abelian variety. We have a C-linear map
𝜆 : 𝑉 → 𝑉∗
(43.4.20)
𝑥 ↦→ 𝐻 (𝑥, −)
with the property that 𝜆(Λ) ⊆ Λ∗ . Since the form is nondegenerate, the induced
homomorphism 𝜆 : 𝐴 → 𝐴∨ is an isogeny of polarized abelian varieties. The degree
of the isogeny 𝜆 is equal to the product 𝑑1 · · · 𝑑 𝑔 of the elementary divisors of 𝐸, so in
particular if 𝐸 is principal then 𝜆 is an isomorphism of principally polarized abelian
varieties.
43.4.21. Let 𝐴 = 𝑉/Λ be a principally polarized complex abelian variety with Riemann
∼ 𝐴∨ be the isomorphism of principally polarized abelian varieties
form 𝐸. Let 𝜙 : 𝐴 −→
induced by (43.4.20). Then we define the Rosati involution associated to 𝐸 (or 𝜙) by
†
: End( 𝐴) → End( 𝐴)
(43.4.22)
𝛼 ↦→ 𝛼† = 𝜙−1 𝛼∨ 𝜙
where 𝛼∨ : 𝐴∨ → 𝐴∨ is the isogeny induced by pullback. The Rosati involution is

uniquely defined by the condition
𝐸 (𝑥, 𝛼𝑦) = 𝐸 (𝛼† 𝑥, 𝑦) (43.4.23)
for all 𝑥, 𝑦 ∈ Λ.
Proposition 43.4.24. The Rosati involution † is a positive involution on the Q-algebra

End( 𝐴) ⊗ Q.
Proof. Let 𝐵 := End( 𝐴) ⊗ Q, and let 𝛼 ∈ 𝐵 with 𝛼 ≠ 0. Let 𝛽 := 𝛼† 𝛼. Since 𝛽† = 𝛽,

and † is the adjoint involution with respect to the positive definite Hermitian form 𝐻,
the action of 𝛽 on 𝑉 is diagonalizable with positive real eigenvalues: indeed, if 𝑥 ∈ 𝑉
is an eigenvector with eigenvalue 𝜆, then
𝜆𝐻 (𝑥, 𝑥) = 𝐻 (𝛽𝑥, 𝑥) = 𝐻 (𝛼𝑥, 𝛼𝑥) ∈ R>0
and 𝐻 (𝑥, 𝑥) ∈ R>0 , so 𝜆 ∈ R>0 . The eigenvalues of 𝛽 on End( 𝐴) by left multiplication
are its eigenvalues with some multiplicity, and accordingly the trace tr(𝛽) = tr(𝛼† 𝛼)
is a nonempty sum of these eigenvalues, hence is positive.
43.4.25. Following Lemma 43.4.17, we define the Siegel upper-half space
ℌ𝑔 = {𝜏 ∈ M𝑔 (C) : 𝜏 t = 𝜏 and Im 𝜏 is positive definite}.
To 𝜏 ∈ ℌ𝑔 , we associate the lattice Λ 𝜏 = 𝜏Z𝑔 ⊕ Z𝑔 ⊂ C𝑔 and the abelian variety
𝐴 𝜏 = C𝑔 /Λ 𝜏 with principal polarization

𝑦
𝐸 𝜏 (𝜏𝑥1 + 𝑥 2 , 𝜏𝑦 1 + 𝑦 2 ) := 𝑥 1 𝑦 2 − 𝑥2 𝑦 1 = (𝑥 1 , 𝑥2 )𝐽 1
t t t t
𝑦2

0 1
where 𝐽 = . By Lemma 43.4.17, every principally polarized complex abelian
−1 0
variety arises in this way.
Two elements 𝜏, 𝜏 0 ∈ ℌ𝑔 give rise to isomorphic abelian varieties if and only if
they arise from a symplectic change of basis of Λ if and only if they are in the same
orbit under the group
Sp2𝑔 (Z) = {𝛾 ∈ M2𝑔 (Z) : 𝛾 t 𝐽𝛾 = 𝐽}
where Sp2𝑔 (Z) ℌ𝑔 acts by

𝑎 𝑏
𝜏 ↦→ (𝑎𝜏 + 𝑏) (𝑐𝜏 + 𝑑) −1 , for ∈ Sp2𝑔 (Z).
𝑐 𝑑
These maps give a bĳection between the set of principally polarized complex
abelian varieties of dimension 𝑔 and the quotient
A 𝑔 (C) := Sp2𝑔 (Z)\ℌ𝑔 .
By the theory of theta functions, A 𝑔 (C) is the set of complex points of a quasi-
projective variety defined over Q of dimension (𝑔 2 + 𝑔)/2.
43.5 Complex abelian surfaces

We now specialize to the case 𝑔 = 2 of principally polarized abelian surfaces; in this
section, we describe their moduli and the relationship with genus 2 curves, in analogy
with elliptic curves (𝑔 = 1).
We recall Example 43.4.7, where abelian varieties were obtained from Jacobians
of curves—we now specialize this to the case 𝑔 = 2. The following theorem links
complex genus 2 curves, via their Jacobians, to complex abelian surfaces.
43.5. COMPLEX ABELIAN SURFACES 799
Theorem 43.5.1. Let 𝐴 be a principally polarized abelian surface over C. Then

exactly one of the two holds:
(i) 𝐴 ' Jac 𝑋 is isomorphic as a principally polarized abelian surface to the
Jacobian of a genus 2 curve 𝑋 equipped with its natural polarization; or
(ii) 𝐴 ' 𝐸 1 × 𝐸 2 is isomorphic as a principally polarized abelian surface to the
product of two elliptic curves equipped with the product polarization.
43.5.2. In case (i) of Theorem 43.5.1, we say that 𝐴 is indecomposable (as a prin-
cipally polarized abelian surface, up to isomorphism). It is possible for 𝐴 to be
indecomposable as a principally polarized abelian surface and yet 𝐴 is not simple, so
𝐴 is isogenous to the product of elliptic curves. In case (ii), we say 𝐴 is decomposable,
and this case arises if and only if there is a basis 𝑒 1 , 𝑒 2 for C2 such that
Λ = Λ1 𝑒 1 ⊕ Λ2 𝑒 2
where Λ1 , Λ2 ⊆ C.
We now pursue an explicit version of Theorem 43.5.1, linking the algebraic descrip-
tion (section 43.3) to the analytic description (section 43.4), in a manner analogous to
the construction of Eisenstein series for elliptic curves (𝑔 = 1) in 40.1.11 and 40.1.19.
43.5.3. For brevity, let Γ := Sp4 (Z), let

𝑎 𝑏
𝚥 (𝛾; 𝜏) = 𝑐𝜏 + 𝑑 ∈ M2 (C), for 𝛾 = ∈ Γ and 𝜏 ∈ ℌ2 ,
𝑐 𝑑
where 𝑎, 𝑏, 𝑐, 𝑑 ∈ M2 (Z), and let

𝑎 𝑏
Γ∞ = 𝛾 = ∈Γ:𝑐=0 .
𝑐 𝑑
We define for 𝑘 ∈ 2Z ≥2 the Eisenstein series
𝜓 𝑘 : ℌ2 → C
∑︁
𝜓 𝑘 (𝜏) := det 𝚥 (𝛾; 𝜏) −𝑘 .
Γ∞ 𝛾 ∈Γ∞ \Γ
As for classical Eisenstein series, 𝜓 𝑘 (𝜏) is absolutely convergent on compact domains.

By design, the function 𝜓 𝑘 has a natural invariance under Γ:
𝜓 𝑘 (𝛾𝜏) = (det 𝚥 (𝛾; 𝜏)) 𝑘 𝜓 𝑘 (𝜏) (43.5.4)
for all 𝛾 ∈ Γ and 𝜏 ∈ ℌ2 .
We define two further functions:
43867
𝜒10 := − 12 5 2 1 1 (𝜓4 𝜓6 − 𝜓10 )
2 3 5 7 53
131 · 593
𝜒12 := 13 7 3 2 (32 72 𝜓43 + 21 53 𝜓62 − 691𝜓12 )
2 3 5 7 3371
(The constants are taken so that the Fourier expansion is appropriately normalized;
their precise nature can be safely ignored on a first reading.)
The function 𝜒10 is somewhat analogous to the classical function Δ, in the following
sense.
Lemma 43.5.5. Let 𝜏 ∈ ℌ2 . Then 𝜒10 (𝜏) = 0 if and only if 𝐴 𝜏 is decomposable (as a
principally polarized abelian variety).
In other words, the vanishing locus of 𝜒10 is precisely where case (ii) of Theorem
43.5.1 holds and the abelian surface is not isomorphic to the Jacobian of a genus 2
curve (as a principally polarized abelian surface).
Remark 43.5.6. More generally, a (classical) Siegel modular form of weight 𝑘 ∈ 2Z
for the group Γ = Sp4 (Z) is a holomorphic function 𝑓 : ℌ2 → C such that
𝑓 (𝛾𝜏) = (det 𝚥 (𝛾; 𝜏)) 𝑘 𝑓 (𝜏)
for all 𝛾 ∈ Sp4 (Z) and 𝜏 ∈ ℌ2 . (By the Koecher principle, such a function is
automatically holomorphic at infinity in a suitable sense, and so the conditions at
cusps 40.2.12 for classical modular forms do not arise here.)
Let 𝑀𝑘 (Γ) be the C-vector space of Siegel modular forms of weight 𝑘; then 𝑀𝑘 (Γ)
is finite-dimensional, 𝑀𝑘É (Γ) = {0} for 𝑘 < 0, and 𝑀0 (Γ) = C consists of constant
functions. Let 𝑀 (Γ) = 𝑘 ∈2Z≥0 𝑀 𝑘 (Γ); then 𝑀 (Γ) has the structure of a graded
C-algebra under pointwise multiplication of functions. Igusa proved that
𝑀 (Γ) = C[𝜓4 , 𝜓6 , 𝜒10 , 𝜒12 ]
in analogy with Theorem 40.3.11. Extending the analogy, Igusa also proved that
𝜓4 , 𝜓6 , −4𝜒10 , 12𝜒12 have integer Fourier coefficients with content 1.
43.5.7. The Igusa–Clebsch invariants 43.3.5 can be expressed in terms of the functions
above. The precise relationship was worked out by Igusa [Igu60, p. 620]: we have
𝜒12
𝐼2 = −23 31
𝜒10
𝐼4 = 22 𝜓4
23 𝜓4 𝜒12
𝐼6 = − 𝜓6 − 25
3 𝜒10
𝐼10 = −214 𝜒10
The functions 𝐼4 , 𝐼6 , 𝜒10 , 𝜒12 are holomorphic, but 𝐼2 is meromorphic. In other words,
if 𝑋 is a complex genus 2 curve with Jac 𝑋 = 𝐴 𝜏 for 𝜏 ∈ ℌ2 , then the algebraic
invariants of 𝑋 can be computed in terms of the values of these functions evaluated
at 𝜏. This description is again analogous to the case of elliptic curves (cf. Remark
40.3.10).
Proposition 43.5.8. Two indecomposable principally polarized abelian surfaces 𝐴 𝜏 ,

𝐴 𝜏0 are isomorphic (as principally polarized abelian surfaces) if and only if
(𝐼2 : 𝐼4 : 𝐼6 : 𝐼10 ) (𝜏) = (𝐼2 : 𝐼4 : 𝐼6 : 𝐼10 ) (𝜏 0) ∈ P(2, 4, 6, 10) (C).

43.6. ABELIAN SURFACES WITH QM 801
In other words, the Igusa(–Clebsch) invariants are naturally defined coordinates on

the moduli space A2 (C) of abelian surfaces, a complex threefold by 43.4.25.
Proof. Combine 43.5.7 and Proposition 43.3.7.
43.5.9. Let 𝐴 be a principally polarized complex abelian surface. Let End( 𝐴) be the
ring of endomorphisms of 𝐴, and let 𝐵 = End( 𝐴) ⊗Z Q. By the classification theorem
of Albert (Theorem 8.5.4), the Q-algebra 𝐵 is exactly one of the following:
(i) 𝐵 = Q, and we say 𝐴 is typical;
(ii) 𝐵 = 𝐹 a real quadratic field, and we say 𝐴 has real multiplication (RM) by 𝐹;
(iii) 𝐵 is an indefinite quaternion algebra over Q, and we say 𝐴 has quaternionic
multiplication (QM) by 𝐵;
(iv) 𝐵 = 𝐾 is a quartic CM field 𝐾, and we say 𝐴 has complex multiplication (CM)
by 𝐾; or
(v) 𝐵 = M2 (𝐾) where 𝐾 is an imaginary quadratic field.
Each one of these 5 cases is interesting in its own right—but given the subject
of this book, we concern ourselves essentially with only case (iii), where quaternion
algebras play a defining role.
43.6 Abelian surfaces with QM

In this section, we consider moduli spaces of abelian surfaces with quaternionic
multiplication. For further reference, see Lang [Lang82, §IX.4–5]. Throughout, let 𝐴
be a principally polarized complex abelian surface with Riemann form 𝐸. Let 𝐵 be an
indefinite quaternion algebra over Q with disc 𝐵 = 𝐷, let
𝜄∞ : 𝐵 ↩→ 𝐵R ' M2 (R) (43.6.1)
be a splitting over R.
43.6.2. The Rosati involution † (defined in 43.4.21) is a positive involution on End( 𝐴)Q .
We classified positive involutions in section 8.4: specifically, when End( 𝐴)R ' M2 (R),
by Example 8.4.15 there exists 𝜇 ∈ End( 𝐴)R× with 𝜇2 ∈ R<0 such that
𝛼† = 𝜇−1 𝛼𝜇 (43.6.3)
for all 𝛼 ∈ End( 𝐴). The map † defines a Q-antiautomorphism of End( 𝐴), so by the
Skolem–Noether theorem, we must have 𝜇 ∈ End( 𝐴) × .
From now on, let O be a maximal order in 𝐵. (One can relax this hypothesis with
some additional technical complications, but there is enough to wrangle with here
already!) In light of 43.6.2 we make the following definition (cf. Rotger [Rot2004,
§3]).
Definition 43.6.4. A polarization on O is an element 𝜇 ∈ O such that 𝜇2 ∈ Z<0 ; a
polarization is principal if 𝜇2 + 𝐷 = 0.
An isomorphism of polarized orders (O, 𝜇) ' (O0, 𝜇 0) is an isomorphism of orders
∼ O0 such that 𝜙(𝜇) = 𝜇 0.
𝜙 : O −→
43.6.5. By Lemma 17.4.2 (employing the Skolem–Noether theorem), an isomorphism

∼ O0 of rings is induced by conjugation by an element of 𝐵× . In particular, the
𝜙 : O −→
polarized orders (O, 𝜇) and (O, 𝜇 0) are isomorphic as polarized orders if and only if
there exists 𝜈 ∈ 𝑁 𝐵× (O) such that 𝜇 0 = 𝜈 −1 𝜇𝜈.
43.6.6. Every O has a principal polarization. Indeed, by the local-global

√ principle
for splitting/embeddings
√ (Proposition 14.6.7), the field 𝐾 = Q( −𝐷) embeds in 𝐵
because 𝐾 𝑝 = Q 𝑝 ( −𝐷) is a field for all 𝑝 | 𝐷. Therefore there exists 𝜇 0 ∈ 𝐵 with
(𝜇 0) 2 + 𝐷 = 0, and 𝜇 0 ∈ O0 for some maximal order O0. But by a consequence of strong
approximation (Theorem 28.2.11), O is conjugate to O0, so there exists a conjugate
𝜇 ∈ O with still 𝜇2 + 𝐷 = 0. (By the theory of optimal embeddings, this is also
immediately implied by Example 30.7.5, which counts the number of O× -equivalence
classes of optimal embeddings Z𝐾 ↩→ O.)
Lemma 43.6.7. Let 𝜇 be a principal polarization on O. Then 𝜇 ∈ 𝐷O♯ , i.e., trd(𝜇O) ⊆

𝐷Z.
Proof. We may check the desired equality locally. If 𝑝 - 𝐷, then 𝜇 ∈ O×𝑝 and
trd(𝜇O 𝑝 ) = trd(O 𝑝 ) = Z 𝑝 . Otherwise, if 𝑝 | 𝐷, then 𝜇 generates the normalizer
group 𝑁 𝐵×𝑝 (O 𝑝 )/(Q×𝑝 O×𝑝 ) by Exercise 23.4, and trd(𝜇O 𝑝 ) ⊆ 𝑝Z 𝑝 as desired.
For a principally polarized order (O, 𝜇), we define the positive involution
∗
:𝐵→𝐵
(43.6.8)
𝛼∗ = 𝜇−1 𝛼𝜇
From now on, let (O, 𝜇) be a principally polarized order.
Definition 43.6.9. A quaternionic multiplication (QM) structure by (O, 𝜇) on 𝐴

is an injective ring homomorphism 𝜄 : O ↩→ End( 𝐴) such that the induced homomor-
phism 𝜄 : 𝐵 ↩→ End( 𝐴)Q respects involutions, i.e., the diagram
𝐵
𝜄 / End( 𝐴)Q
∗ † (43.6.10)

𝐵
𝜄 / End( 𝐴)Q
commutes, where † is the Rosati involution defined in (43.4.22).

We say 𝐴 has quaternionic multiplication (QM) by (O, 𝜇) if 𝐴 can be equipped
with a QM structure by (O, 𝜇), and we say that 𝐴 is a QM abelian surface if it has a
QM structure for some (O, 𝜇).
Writing out (43.6.10), for a QM structure 𝜄 we require that 𝜄(𝛼) † = 𝜄(𝛼∗ ) for all
𝛼 ∈ 𝐵.
Definition 43.6.11. A homomorphism ( 𝐴, 𝜄) → ( 𝐴 0, 𝜄0) of principally polarized

complex abelian surfaces with QM by (O, 𝜇) is a homomorphism 𝜙 : 𝐴 → 𝐴 0 of
polarized abelian surfaces (respecting the polarization) that also respects 𝜄, 𝜄0 in the
sense that the diagram
𝜄0 /
𝐵 End( 𝐴 0)Q
𝜄
𝜙∗
#
End( 𝐴)Q
commutes; an isogeny is a surjective homomorphism with finite kernel.
QM abelian surfaces can be constructed as follows.
43.6.12. Extend 𝜄∞ to a map 𝜄∞ : 𝐵 ↩→ 𝐵C ' M2 (C). Let 𝜏 ∈ H2 . Let

𝜏
Λ 𝜏 := 𝜄∞ (O) ⊂ C2 .
1
Then Λ 𝜏 is a lattice in C2 , since rkZ O = 4; let 𝐴 𝜏 := C2 /Λ 𝜏 be the associated complex

torus. The map 𝜄∞ induces a natural injective ring homomorphism 𝜄 𝜏 : O → End( 𝐴 𝜏 )
since 𝜄∞ (O)Λ 𝜏 ⊆ Λ 𝜏 as O itself is closed under multiplication. Define the form
𝐸 𝜏 : Λ𝜏 × Λ𝜏 → Z
(43.6.13)

𝜏 𝜏 1
𝑥 ,𝑦 ↦→ trd(𝜄∞ (𝜇)𝑥𝑦)
1 1 𝐷
with 𝑥, 𝑦 ∈ 𝜄∞ (O). The form 𝐸 𝜏 takes values in Z by Lemma 43.6.7.
The main result of this section is then the following theorem.
Main Theorem 43.6.14. Let Γ = 𝜄∞ (O1 )/{±1} ⊆ PSL2 (R). Then the map

 ( 𝐴, 𝜄) principally polarized 

2
 complex abelian surfaces 

 

Γ\H ↔

 with QM by (O, 𝜇) 
 (43.6.15)
up to isomorphism
 
 
Γ𝜏 ↦→ [( 𝐴 𝜏 , 𝜄 𝜏 )]
is a bĳection.
The proof of this theorem occupies the rest of this section; it amounts to checking
that various conditions and compatibilities are satisfied. The reader who is willing to
take these as verified can profitably move along to the next section.
We begin by verifying the Riemann relations.
Lemma 43.6.16. The form 𝐸 = 𝐸 𝜏 defined in (43.6.13) or its negative 𝐸 = −𝐸 𝜏 is a
Riemann form.
This sign ambiguity was already present above, as the involution (43.6.8) is the
same replacing 𝜇 by −𝜇; working backwards with Lemma 43.6.16 in hand, we may
fix the sign by replacing 𝜇 by −𝜇 once and for all so that 𝐸 𝜏 itself is a Riemann form
for all 𝜏.
Proof. 𝐸 is alternating since

trd(𝜄∞ (𝜇)𝑥𝑥) = nrd(𝑥) trd(𝜇) = 0
for all 𝑥 ∈ 𝜄∞ (O).
Next, let
−|𝜏| 2

1 Re 𝜏
𝜚= ∈ M2 (R).
Im 𝜏 1 − Re 𝜏
Then
−(Re 𝜏) 2 + |𝜏| 2
det( 𝜚) = 𝜚 𝜚 = =1
(Im 𝜏) 2
so 𝜚 ∈ SL2 (R), and moreover
1 𝜏 Re 𝜏 − |𝜏| 2

𝜏 𝑖 Re 𝜏 − Im 𝜏 𝜏
𝜚 = = =𝑖 . (43.6.17)
1 Im 𝜏 𝜏 − Re 𝜏 𝑖 1
Therefore, for all 𝑥, 𝑦 ∈ 𝜄∞ (O),

𝜏 𝜏 𝜏 𝜏 𝜏 𝜏
𝐸 𝑖𝑥 , 𝑖𝑦 = 𝐸 𝑥𝑖 , 𝑦𝑖 = 𝐸 𝑥𝜚 , 𝑦𝜚
1 1 1 1 1 1
= trd(𝜄∞ (𝜇) (𝑥 𝜚)𝑦 𝜚) = trd(𝜄∞ (𝜇)𝑥( 𝜚 𝜚) 𝑦) (43.6.18)

𝜏 𝜏
= trd(𝜄∞ (𝜇)𝑥𝑦) = 𝐸 𝑥 ,𝑦 .
1 1
We now show that (𝑥, 𝑦) ↦→ 𝐸 R (𝑖𝑥, 𝑦) is a symmetric, positive definite R-bilinear
form on 𝑉. It is enough to verify this for 𝑥, 𝑦 ∈ 𝜄∞ (O). In this calculation, to avoid
clutter we write 𝜇 for 𝜄∞ (𝜇). Following as above, first we show symmetry:

𝜏 𝜏
𝐸 𝑖𝑥 ,𝑦 = trd(𝜇(𝑥 𝜚)𝑦) = trd(𝑦 𝜚 𝑥 𝜇)
1 1
(43.6.19)
𝜏 𝜏
= trd(𝜇(𝑦 𝜚)𝑥) = 𝐸 𝑖𝑦 , 𝑥
1 1
using that 𝜇 = −𝜇 and 𝜌 = −𝜌 since they have trace zero. For positivity, we replace 𝜚
with an expression
√ in 𝜇 in order
√ to simplify, and then apply positivity. Since 𝜇2 = −𝐷,
if we let 𝜇1 = 𝜇/ 𝐷 with 𝐷 ∈ R>0 , then 𝜇1 = −1. Since also 𝜚2 = −1, there exists
2
𝛿 ∈ GL2 (R) such that 𝛿−1 𝜇1 𝛿 = 𝜚. From the calculation that

√
𝜇1 = 𝜇1−1 = 𝐷 𝜇−1 ,
we obtain
√
√ 𝛿 𝐷
𝜚 = 𝛿𝜇1 𝛿−1 = 𝛿( 𝐷 𝜇−1 ) = 𝛿𝜇−1 𝛿 (43.6.20)
nrd(𝛿) nrd(𝛿)
and hence
𝜏 𝜏
𝐸 𝑖𝑥 ,𝑥 = trd(𝜇(𝑥 𝜚)𝑥) = − trd(𝜇𝑥 𝜚 𝑥)
1 1
√ (43.6.21)
𝐷
=− trd(𝜇𝑥𝛿𝜇−1 𝛿𝑥).
nrd(𝛿)
This may look worse, but now we use positivity of ∗ applied to 𝑥𝛿:

trd (𝑥𝛿) (𝑥𝛿) ∗ = trd(𝑥𝛿𝜇−1 𝛿𝑥𝜇) = trd(𝜇𝑥𝛿𝜇−1 𝛿𝑥) > 0.

𝜏 𝜏
It follows that 𝐸 𝑖𝑥 ,𝑥 is always either positive definite or negative definite
1 1
(depending on the sign of nrd(𝛿)).
Lemma 43.6.22. The polarization induced by 𝐸 is principal.
Proof. Let 𝜙 : 𝐴 → 𝐴∨ be the isogeny induced by 𝐸. Then the degree of 𝜙 is

𝜏 𝜏
deg 𝜙 = det 𝐸 𝑥𝑖 ,𝑥 = det(trd((𝜄∞ (𝜇)/𝐷)𝑥 𝑖 𝑥 𝑗 ))𝑖, 𝑗
1 𝑗 1 𝑖, 𝑗
where 𝑥𝑖 are a Z-basis for 𝜄∞ (Λ). Thus
Nm(𝜇)
det(trd((𝜄∞ (𝜇)/𝐷)𝑥𝑖 𝑥 𝑗 ))𝑖, 𝑗 = det(trd(𝑥 𝑖 𝑥 𝑗 ))𝑖, 𝑗
𝐷4
nrd(𝜇) 2 1
= (discrd O) 2 = 2 𝐷 2 = 1.
𝐷4 𝐷
Lemma 43.6.23. The homomorphism 𝜄 𝜏 : O → End( 𝐴 𝜏 )Q satisfies the compatibility
(43.6.10), and 𝐸 𝜏 is the unique compatible principal polarization on 𝐴 𝜏 compatible
with 𝜄 𝜏 .
Proof. Abbreviate 𝜄 = 𝜄 𝜏 . Let 𝛼 ∈ O. Then the Rosati involution is uniquely defined

by the condition

𝜏 𝜏 † 𝜏 𝜏
𝐸 𝑥 , 𝜄(𝛼)𝑦 = 𝐸 𝜄(𝛼) 𝑥 ,𝑦 (43.6.24)
1 1 1 1
for all 𝑥, 𝑦 ∈ 𝜄∞ (O), i.e.,
trd(𝜄∞ (𝜇)𝑥𝜄(𝛼)𝑦) = trd(𝜄∞ (𝜇)𝜄(𝛼) † 𝑥𝑦). (43.6.25)
Let 𝑧 = 𝜄(𝛼) and 𝑥, 𝑦 ∈ 𝜄∞ (O). We verify that (43.6.25) holds by
trd(𝜄∞ (𝜇)𝑧† 𝑥𝑦) = trd(𝜄∞ (𝜇) (𝜄∞ (𝜇) −1 𝑧𝜄∞ (𝜇))𝑥𝑦)

(43.6.26)
= trd(𝜄∞ (𝜇)𝑥𝑦 𝑧) = trd(𝜄∞ (𝜇)𝑥𝑧𝑦)
as desired.
To conclude, any other polarization corresponds to another positive involution of
the form 𝛼 ↦→ 𝜈 −1 𝛼𝜈 as in 43.6.2; scaling, we may take 𝜈 ∈ O such that trd(𝜈O) ⊂
𝐷O♯ . In the proof of Lemma 43.6.22, we see that the degree of the polarization is
equal to nrd(𝜈) 2 /𝐷 2 , so it is principal if and only if nrd(𝜈) = 𝐷, i.e., 𝜈 2 + 𝐷 = 0. But
then by compatibility trd(𝜄∞ (𝜇)𝑥𝑦) = trd(𝜄∞ (𝜈)𝑥𝑦) for all 𝑥, 𝑦 ∈ O, which implies
𝜇 = 𝜈.
Remark 43.6.27. Lemma 43.6.23 shows that one could be more relaxed in the definition
of QM abelian surface in the following sense. Let 𝐴 be a (not yet polarized) complex
abelian surface, and let 𝜄 : O ↩→ End( 𝐴) be a ring homomorphism. Then there is a
unique principal polarization on 𝐴 such that the induced Rosati involution is compatible
with 𝜇 in the sense of (43.6.10).
Proposition 43.6.28. Every principally polarized complex abelian surface with QM

by (O, 𝜇) is isomorphic as such to one of the form ( 𝐴 𝜏 , 𝜄 𝜏 ) for some 𝜏 ∈ H2 .
Proof. Let ( 𝐴, 𝜄) be a principally polarized complex abelian surface with QM by

(O, 𝜇), and let 𝐴 = C2 /Λ. Then
End( 𝐴) = {𝛼 ∈ M2 (C) : 𝛼Λ ⊆ Λ}.
Therefore 𝜄 : O ↩→ End( 𝐴) extends to an inclusion 𝐵 ↩→ M2 (C). By the Skolem–

Noether theorem, every two inclusions are conjugate by an element of GL2 (C), acting
by an isomorphism of 𝐴; so without loss of generality, we may suppose that 𝜄 extends
to 𝜄∞ .
We claim that Λ = 𝜄∞ (O)𝑥 for some 𝑥 ∈ C2 . Indeed, Λ ⊗ Q has the structure of a
left 𝐵-module with the same dimension as 𝐵 as a Q-vector space; by Exercise 7.6, it
follows that Λ ⊗ Q = 𝐵𝑥 is free as a left 𝐵-module with 𝑥 ∈ Λ ⊆ C2 ; thus Λ = 𝐼𝑥 where
𝐼 ⊆ 𝐵 is a left O-ideal. Since O is maximal and therefore hereditary, 𝐼 is invertible
as a left O-ideal, and in particular 𝐼 is sated. Now comes strong approximation: by
Corollary 28.5.17, since 𝐵 is indefinite and Cl+ Z is trivial, we conclude that 𝐼 is
principal, and therefore we can rewrite Λ = 𝜄∞ (O)𝑥 with 𝑥 ∈ C2 . By Lemma 43.4.17,
we may suppose that 𝑥 = 𝜏 1 with Im 𝜏 > 0, so 𝜏 ∈ H2 .

Finally, the polarization agrees by the uniqueness in Lemma 43.6.23.
We are now ready to finish the proof of Main Theorem 43.6.14.
Proof of Main Theorem 43.6.14. By Lemmas 43.6.16, 43.6.22, and 43.6.23, the asso-
ciation 𝜏 ↦→ ( 𝐴 𝜏 , 𝜄 𝜏 ) yields a principally polarized complex abelian surface with QM
by (O, 𝜇). By Proposition 43.6.28, the map is surjective.
𝑎 𝑏
Next, we show the map is well-defined up to the action of Γ. Let 𝛾 = ∈Γ
𝑐 𝑑
0
and 𝜏 = 𝛾𝜏. Then

𝛾𝜏 −1 𝑎𝜏 + 𝑏
Λ 𝜏0 = 𝜄∞ (O) = 𝜄∞ (O) (𝑐𝜏 + 𝑑)
1 𝑐𝜏 + 𝑑
(43.6.29)
−1 𝜏 −1
= (𝑐𝜏 + 𝑑) 𝜄∞ (O)𝛾 = (𝑐𝜏 + 𝑑) Λ 𝜏 .
1
Therefore scalar multiplication by (𝑐𝜏 + 𝑑) −1 induces an isomorphism 𝐴 𝜏 → 𝐴 𝛾 𝜏 of

abelian surfaces; this map preserves the polarization by writing
0
𝜏 𝜏
(𝑐𝜏 + 𝑑)𝑥 = 𝑥𝛾
1 1
43.7. REAL POINTS, CM POINTS 807
for 𝑥 ∈ 𝜄∞ (O), and then verifying that

trd(𝜇𝑥𝑦) = trd(𝜇(𝑥𝛾) (𝑦𝛾)) = trd(𝜇𝑥𝛾𝛾 𝑦) = trd(𝜇𝑥𝑦).
The induced map End( 𝐴 𝛾 𝜏 ) → End( 𝐴 𝜏 ) obtained from conjugation by a scalar matrix
is the identity, and the compatibility for the QM is then verified by the fact that for
𝛼 ∈ 𝐵,
𝜄0 (𝛼) = 𝜄∞ (𝛼) = 𝜄(𝛼).
To conclude, suppose that ( 𝐴 𝜏 , 𝜄 𝜏 ) ' ( 𝐴 𝜏0 , 𝜄 𝜏0 ) with 𝜏, 𝜏 0 ∈ H2± ; then there exists
𝜙 ∈ GL2 (C) such that 𝜙Λ 𝜏 = Λ 𝜏0 and 𝜙 commutes with 𝜄∞ (𝛼). Since 𝜄∞ (𝛼) ⊗ C =
M2 (C), we conclude that 𝜙 is central in M2 (C) and hence scalar. From
0
𝜏 𝜏
𝜙 =𝛾
1 1
we conclude 𝛾 ∈ Γ and then that 𝜙 = 𝑐𝜏 0 + 𝑑 so 𝜏 = 𝛾𝜏 0.
Remark 43.6.30. Analogous to the case SL2 (Z), one may similarly define congruence
subgroups of Γ1 ; the objects then parametrized are QM abelian surfaces equipped with
a subgroup of torsion points.
Remark 43.6.31. The “forgetful” map which forgets the QM structure 𝜄 gives a map of
moduli from Γ\H2 to A2 (C), but this map is not injective: it factors via the quotient by
the larger group Γh𝜇i. In other words, the bĳection of Main Theorem 43.6.14 induces a
bĳection between Γh𝜇i\H2 and the set of isomorphism classes of principally polarized
abelian surfaces 𝐴 such that 𝐴 has QM by (O, 𝜇): accordingly, generically there will
be two choices of QM structure on a surface 𝐴 that can be given QM by (O, 𝜇).
Remark 43.6.32. The results above for 𝐹 = Q extend, but not in a straightforward
way, to totally real fields 𝐹 of larger degree 𝑛 = [𝐹 : Q]. If 𝐴 is an abelian variety
with dim( 𝐴) = 𝑔 such that 𝐴 has QM by 𝐵 over 𝐹, then 4𝑛 | 2𝑔, so we must consider
abelian varieties of dimension at least 2𝑛. If equality 𝑔 = 2𝑛 holds, then 𝐴 is simple,
and by Albert’s classification of the possible endomorphism algebras of an abelian
variety, 𝐵 is either totally definite or totally indefinite. So if 𝑡 = 1, then 𝐹 = Q and 𝐵
is totally indefinite.
Consequently, we must consider abelian varieties of larger dimension. The basic
construction works as follows. First, one chooses an element 𝜇 ∈ O such that 𝜇2 ∈ Z𝐹
is totally negative. (If Z𝐹 has strict class number 1, then one has disc 𝐵 = 𝔇 =√ 𝐷Z𝐹
with 𝐷 totally positive and one can choose 𝜇 satisfying 𝜇2 = −𝐷.) Let 𝐾 = 𝐹 ( −𝐷);
note that since 𝐾 ↩→ 𝐵 we have 𝐵 𝐾 = 𝐵 ⊗𝐹 𝐾 𝑀2 (𝐾). Then the complex space
𝑋 𝐵 (1)C parametrizes complex abelian 4𝑛-folds with endomorphisms (QM) by 𝐵 𝐾 and
equipped with a particular action on 𝐹 on the complex differentials of 𝐵. Amazingly,
this moduli interpretation does not depend on the choice of 𝐾; but because of this
choice, it is not canonical as for the case 𝐹 = Q.
43.7 Real points, CM points

Let 𝑋 1 = Γ\H2 . Then 𝑋 1 has the structure of a complex 1-orbifold. We conclude this
chapter with some discussion about real structures.
43.7.1. By Example 28.6.5, there exists 𝜖 ∈ O× such that nrd(𝜖) = −1. Then 𝜖 2 ∈ O1
and 𝜖 normalizes O1 , so the action of 𝜖 (as in (33.3.12)) defines an anti-holomophic
involution on 𝑋 (Γ) that is independent of the choice of 𝜖: this gives the natural action
of complex conjugation on 𝑋 (Γ).
With respect to this real structure, and in view of Main Theorem 43.6.14, we may
ask if there are any principally polarized abelian surfaces with QM by (O, 𝜇) with
and the QM defined over R. When 𝐵 ' M2 (Q), then the element
both the surface
1 0
𝜖 = acts by complex conjugation, and the real points are those points on
0 −1
the imaginary axis (the points with real 𝑗-invariant). More generally, the answer is
provided by the following special case of a theorem of Shimura [Shi75, Theorem 0].
Proposition 43.7.2 (Shimura). If 𝐵 is a division algebra, then 𝑋 1 (R) = ∅.

𝑎 𝑏
Proof. We follow Ogg [Ogg83, §3]. Let 𝜄∞ (𝜖) = . If 𝑋 1 (R) ≠ ∅, then by
𝑐 𝑑
43.7.1 there exists 𝑧 ∈ H2 such that
𝑎𝑧 + 𝑏
𝑧 = 𝜄∞ (𝜖) · 𝑧 = .
𝑐𝑧 + 𝑑
Then 𝑎𝑧 + 𝑏 = 𝑐| 𝑧| 2 + 𝑑𝑧; taking imaginary parts we find −𝑎 Im 𝑧 = 𝑑 Im 𝑧 so 𝑎 + 𝑑 =

trd 𝜖 = 0 and 𝜖 2 − 1 = 0. Since 𝐵 is a division algebra, we conclude 𝜖 = ±1, a
contradiction.
It may nevertheless happen that certain quotients of 𝑋 1 by Atkin–Lehner involu-

tions may have real points.
Remark 43.7.3. A similar issue arises for CM elliptic curves: such a curve, together
with all its endomorphisms, cannot be defined over R.
Just as 𝑌 (1) = SL2 (Z)\H2 parametrizes isomorphism classes of elliptic curves,
among them are countably many elliptic curves whose endomorphism algebra is larger
than Z: these are the elliptic curves with complex multiplication, and they correspond
to points in H2 that are quadratic irrationalities, so Q(𝜏) = 𝐾 is an imaginary quadratic
field and 𝑆 = Z[𝜏] ⊆ 𝐾 is an imaginary quadratic order. By the theory of complex
multiplication, the corresponding 𝑗-invariants are defined over the ring class field
𝐻 ⊇ 𝐾 with Gal(𝐻 | 𝐾) ' Pic(𝑆), and there is an explicit action of Gal(𝐻 | 𝐾) on this
set.
In a similar way, on 𝑋 1 we have CM points, given by complex abelian surfaces
with extra endomorphisms, defined more precisely as follows.
43.7.4. Let 𝐾 ⊇ Q be an imaginary quadratic field and suppose that 𝜄𝐾 : 𝐾 ↩→ 𝐵

embeds. Let 𝑆 = 𝐾 ∩ O, so that 𝑆 ↩→ O is optimally embedded. Suppose the image
of this embedding is given by 𝑆 = Z[𝜈] where 𝜈 ∈ O. Let 𝜏 = 𝜏𝜈 be the unique fixed
point of 𝜄∞ (𝜈) in H2 ; a point of H2 that arises this way is called a CM point.
The corresponding abelian surface 𝐴 𝜏 visibly has M2 (𝑆) ↩→ End( 𝐴 𝜏 ), and in
particular End( 𝐴 𝜏 )Q ' M2 (𝐾) as this is as large as possible.
43.8. ∗ CANONICAL MODELS 809
43.8 ∗ Canonical models

In this section, we sketch the theory of canonical models for the curves 𝑋 1 .
Theorem 43.8.1 (Shimura [Shi67, Main Theorem I (3.2)]). There exists a projective,
nonsingular curve 𝑋Q1 defined over Q and a holomorphic map
𝜑 : H2 → 𝑋Q1 (C)
that induces an analytic isomorphism
𝜑 : Γ1 \H2 → 𝑋Q1 (C).
43.8.2. The curve 𝑋Q1 is made canonical (unique up to isomorphism) according to the
field of definition of CM points (see 43.7.4).
Let 𝑧 ∈ H2 be a CM point with 𝑆 of discriminant 𝐷. Let 𝐻 ⊇ 𝐾 be the ring class
extension 𝐻 ⊇ 𝐾 with Gal(𝐻 | 𝐾) ' Pic(𝑆). Then 𝜙(𝑧) ∈ 𝑋Q1 (𝐻). Moreover, there is
an explicit law, known as the Shimura reciprocity law, which describes the action of
Gal(𝐻 | 𝐾) on them: to a class [𝔠] ∈ Pic 𝑆, we have
𝜄𝐾 (𝔠)O = 𝜉O (43.8.3)
for some 𝜉 ∈ O, and if 𝜎 = Frob𝔠 ∈ Gal(𝐻 | 𝐾) under the Artin isomorphism, then
𝜎(𝜙(𝑧)) = 𝜙(𝜉 −1 𝑧). (43.8.4)
For more detail, see Shimura [Shi67, p. 59]; and for an explicit, algorithmic version,
see Voight [Voi2006].
43.8.5. Before continuing, we link back to the idelic, double-coset point of view,
motivated in section 38.6 and given in general in section 38.7.
The difference between a lattice in R2 and a lattice in C is an identification of R2
with C, i.e., an injective R-algebra homomorphism 𝜓 : C → EndR (R2 ); we call 𝜓 a
complex structure.
There is a bĳection between the set of complex structures and the set C r R = H2±
as follows. A complex structure 𝜓 : C → EndR (R2 ), by R-linearity is equivalently
given by the matrix 𝜓(𝑖) ∈ GL2(R) satisfying
𝜓(𝑖) 2 = −1. By rational canonical form,
0 1
every such matrix is similar to , i.e., there exists 𝛽 ∈ GL2 (R) such that
−1 0

−1 0 1
𝜓(𝑖) = 𝛽 𝛽,
−1 0

0 1
and 𝛽 is well-defined up to the centralizer of ; this matrix acts by fixing
−1 0
2
𝑖 ∈ H , and it follows that this centralizer is precisely R× SO(2). Finally, we have
GL2 (R)/R× SO(2) −→ ∼ H2± under 𝛽 ↦→ 𝛽𝑖.
In this way,
𝑋 1 = Γ1 \H2 ↔ O1 \H2±
↔ O1 \(GL2 (R)/R× SO(2)) (43.8.6)
↔ 𝐵× \( 𝐵
b× /O
b× × GL2 (R)/R× SO(2)).
(The final line is just an expression of the fact that 𝐵× \ 𝐵

b× /O
b× is a set with one element,
by strong approximation; it is placed there for comparison with other settings, where
class numbers may add spice.) So the bĳection (43.8.6) says that 𝑋 1 parametrizes
O-lattices in 𝐵 with a complex structure up to homothety. In the previous few sections,
we showed how such lattices equipped with complex structure can be interpreted as a
moduli space for abelian surfaces with quaternion multiplication.
We conclude this section with some more parting comments on representability.

Remark 43.8.7. Let SchQ denote the category of schemes over Q under morphisms
of schemes, and let Set denote the category of sets under all maps of sets. Let
F : SchQ → Set be a contravariant functor. Then 𝑋 ∈ SchQ is a coarse moduli
space for F (or 𝑋 coarsely represents F ) if there exists a natural transformation
Φ : F (−) → Hom(−, 𝑋) which satisfies:
∼ Hom(𝑘, 𝑋) is bĳective if 𝑘 is algebraically closed (and char 𝑘 = 0);
(i) Φ : F (𝑘) −→
and
(ii) Φ is universal: if (𝑍, Ψ) is another such coarse moduli space, then there is a
unique commutative diagram
F (−) / Hom(−, 𝑍)
∃!
%
Hom(−, 𝑋)
By Yoneda’s lemma, condition (ii) is equivalent to a unique (commuting) morphism

𝑋 → 𝑍.
We then define a functor FO : SchQ → Set which associates to 𝑆 the set of
isomorphism classes of abelian schemes 𝐴 over 𝑆—which can be thought of families
of abelian surfaces parametrized by 𝑆—together with a map 𝜄 : O ↩→ End𝑆 ( 𝐴).
Deligne [Del71] has shown that the functor FO is coarsely representable by a curve
𝑋Q1 over Q. By uniqueness and the solution to the moduli problem over C, there is a
map Γ 𝐵 (1)\H2 −→ ∼ 𝑋 1 (C) which is in fact an analytic isomorphism. Together with
Q
the field of definition of CM points, we recover the canonical model (Theorem 43.8.1).
43.9 ∗ Modular forms

In this final section, we sketch aspects of the theory of modular forms for arithmetic
Fuchsian groups.
We restore a bit more generality: let 𝐹 be a totally real field of degree 𝑛 = [𝐹 : Q],
let 𝐵 be a quaternion algebra over 𝐹 that split at exactly one real place corresponding
43.9. ∗ MODULAR FORMS 811
to an embedding 𝜄∞ : 𝐵 ↩→ M2 (R), let O ⊂ 𝐵 be an order, and let Γ ≤ PSL2 (R) be a

group commensurable with Γ1 (O) = 𝜄∞ (O1 )/{±1}.
Let 𝑌 = Γ\H2 , let 𝑋 = Γ\H2(∗) be its completion, and call the set 𝑋 r 𝑌 the set of
cusps. We recall the notion of orbifold from section 34.8. Then 𝑋 has the structure
of a good complex 1-orbifold with signature (𝑔; 𝑒 1 , . . . , 𝑒𝑟 ; 𝛿): 𝑋 has genus 𝑔, there
are exactly 𝑟 elliptic points 𝑃𝑖 of orders 𝑒 𝑖 , and 𝛿 cusps 𝑄 1 , . . . , 𝑄 𝛿 . The hyperbolic
area of 𝑋 is written 𝜇(𝑋), and can be computed as the area of a suitable fundamental
domain.
As in the introduction 43.1.6, we make the following definition.
Definition 43.9.1. Let 𝑘 ∈ 2Z ≥0 . A map 𝑓 : H2 → C is weight 𝑘-invariant under Γ
if
𝑘 𝑎 𝑏
𝑓 (𝛾𝑧) = (𝑐𝑧 + 𝑑) 𝑓 (𝑧) for all 𝛾 = ∈ Γ. (43.9.2)
𝑐 𝑑
A modular form for Γ of weight 𝑘 is a holomorphic function 𝑓 : H2 → C that is
weight 𝑘 invariant and is holomorphic at the cusps.
Let 𝑀𝑘 (Γ) be the C-vector space of modular forms for Γ and let
Ê
𝑀 (Γ) := 𝑀𝑘 (Γ); (43.9.3)
𝑘 ∈2Z≥0
then 𝑀 (Γ) is a graded C-algebra under multiplication.

We can understand the degree of divisors of 𝑀𝑘 (Γ) as follows.
Theorem 43.9.4. For 𝑓 ∈ 𝑀𝑘 (Γ), we have
∑︁ 1 𝑘
ord𝑧 ( 𝑓 ) = 𝜇(𝑋).
# StabΓ (𝑧) 4𝜋
Γ𝑧 ∈Γ\H2
Proof. See Shimura [Shi71, Proposition 2.16, Theorem 2.20].
By an application of the theorem of Riemann–Roch and the description of modular

forms behind Theorem 43.9.4, we find that dimC 𝑀𝑘 (Γ) can be expressed in terms of
𝑘 and the signature of Γ as follows.


 1, if 𝑘 = 0;

 𝑔 + max(0, 𝛿 − 1),

 if 𝑘 = 2;
dimC 𝑀𝑘 (Γ) = 𝑟
 𝑘 ∑︁ 𝑘 1
 (𝑘 − 1) (𝑔 − 1) + 2 𝛿 + 1− if 𝑘 > 2.

 ,
2

 𝑖=1
𝑒𝑖
Proof. See Shimura [Shi71, Theorem 2.23].
The above formulas can be proven in a different and straightforward way in the
language of stacky curves. This description gives the following further information on
𝑀 (Γ).
Theorem 43.9.6 (Voight–Zureick-Brown [VZB2015]). Let 𝑒 = max(1, 𝑒 1 , . . . , 𝑒𝑟 ).

Then the ring 𝑀 (Γ) is generated as a C-algebra by elements of weight at most 6𝑒 with
relations in weight at most 12𝑒.
(The case of signature (0; 2, 2, 3, 3; 0) from section 43.2 is described [VZB2015,

Table (IVa-3)] as a weighted plane curve of degree 12 in P(6, 3, 2).)
Remark 43.9.7. An appealing mechanism for working explicitly with modular forms
in the absence of cusps is provided by power series expansions: for an introduction
with computational aspects, see Voight–Willis [VW2014] and the references therein.
Exercises

1. Let 𝑔 = 1 and consider a period matrix Π = 𝜔1 𝜔2 with 𝜔1 , 𝜔2 ∈ C. Let
0 1
𝐸 = . Show that in Definition 43.4.5 for 𝐸 that the condition (i) is
−1 0
automatic and condition (ii) is equivalent to Im(𝜔2 /𝜔1 ) > 0.
2. Let Π ∈ Mat
𝑔×2𝑔 (C). Show that Π is a period matrix for a complex torus if and
Π
only if , the matrix obtained by vertically stacking Π on top of its complex
Π
conjugate Π, is nonsingular.
3. Let 𝐸 : Λ × Λ → Z be a Z-bilinear form that satisfies conditions (i) and (ii)
of Definition 43.4.9 (so a Riemann form but without the condition that 𝐸 is
alternating). Show that 𝐸 R (and 𝐸) are alternating.
4. Show that the symmetry (43.6.19) implies the equality

𝜏 𝜏 𝜏 𝜏
𝐸 𝑖𝑥 , 𝑖𝑦 =𝐸 𝑥 ,𝑦
1 1 1 1
from (43.6.18) directly using Exercise 43.3 (without 𝜚).

I 5. Let 𝐴 = 𝑉/Λ be a complex abelian variety.
(a) Let 𝜙 : 𝐴 → 𝐴 0 be an isogeny with 𝑛 = # ker(𝜙). Show that there exists a
unique isogeny 𝜓 : 𝐴 0 → 𝐴 such that 𝜓 ◦ 𝜙 = 𝑛 𝐴 is multiplication by 𝑛 on
𝐴 and similarly 𝜙 ◦ 𝜓 = 𝑛 𝐴0 on 𝐴 0. [Hint: ker 𝜙 ⊆ 𝐴[𝑛].]
(b) Suppose that 𝐴 is (not necessarily principally) polarized and let 𝜆 : 𝐴 →
𝐴∨ be the map induced by (43.4.20). Using (a), show that (43.4.22) gives
a well-defined involution on End( 𝐴) ⊗ Q, still called the Rosati involution
(attached to the polarized abelian variety 𝐴).
I 6. Let 𝑉 be a finite-dimensional vector space over C.
(a) Let 𝐻 : 𝑉 ×𝑉 → C be a nondegenerate Hermitian form on 𝑉. Let 𝑆 := Re 𝐻
and 𝐸 := Im 𝐻, so that 𝐻 = 𝑆 + 𝑖𝐸 for R-bilinear forms 𝑆, 𝐸 : 𝑉 × 𝑉 → R.
Show that 𝑆 is symmetric (i.e., 𝑆(𝑦, 𝑥) = 𝑆(𝑥, 𝑦)), 𝐸 is alternating (i.e.,
𝐸 (𝑦, 𝑥) = −𝐸 (𝑥, 𝑦)), and 𝑆(𝑥, 𝑦) = 𝐸 (𝑖𝑥, 𝑦) for all 𝑥, 𝑦 ∈ 𝑉, and moreover
that 𝐸 (and 𝑆) are nondegenerate.
(b) Let 𝐸 be a nondegenerate alternating form on 𝑉 (as an R-vector space).

Show there exists a unique nondegenerate Hermitian form 𝐻 on 𝑉 wth
Im 𝐻 = 𝐸 if and only if 𝐸 (𝑖𝑥, 𝑖𝑦) = 𝐸 (𝑥, 𝑦) for all 𝑥, 𝑦 ∈ 𝑉.
(c) Show that a Hermitian form on 𝑉 is positive definite if and only if the
corresponding symmetric form 𝑆 := Re 𝐻 is positive definite.
I 7. Prove Proposition 43.2.1.
I 8. In the following exercise, we do a few manipulations with generating functions,
applied to understand the presentation of the ring of modular forms in the next
exercise.
(a) Prove that
∑︁ 𝑘 𝑡4
𝑡𝑘 =
𝑘 ∈2Z≥2
4 (1 − 𝑡 2 ) 2 (1 + 𝑡 2 )
∑︁ 𝑘 𝑡4 + 𝑡6
𝑡𝑘 =
𝑘 ∈2Z
3 1 − 𝑡2 − 𝑡6 + 𝑡8
≥2
[Hint: break up the sum by congruence class according to the floor and
then use geometric series.]
(b) Let 𝑚 2 , 𝑚 3 ∈ Z ≥0 . For 𝑘 ∈ 2Z ≥0 , define


 1, if 𝑘 = 0;


𝑐 𝑘 = 𝑔, if 𝑘 = 2;

 (𝑘 − 1) (𝑔 − 1) + 𝑚 2 b𝑘/4c + 𝑚 3 b𝑘/3c,

if 𝑘 ≥ 4.

Show that
∑︁ 1 + 𝑔𝑡 2 + 𝑎 4 𝑡 4 + 𝑎 6 𝑡 6 + 𝑎 4 𝑡 8 + 𝑔𝑡 10 + 𝑡 12
𝑐𝑘 𝑡 𝑘 =
𝑘 ∈2Z≥0
(1 − 𝑡 4 ) (1 − 𝑡 6 )
where
𝑎 4 = 3𝑔 + 𝑚 2 + 𝑚 3 − 4
𝑎 6 = 4𝑔 + 𝑚 2 + 2𝑚 3 − 6.
I 9. Prove Proposition 43.2.7 as follows.

(a) Show that the functions 𝑓4 , 𝑓6 are algebraically independent. [Hint: reduce
to the case where the relation is weighted homogeneous, and plug in 𝑧2 to
show the relation reduces to one of smaller degree.]
(b) Show that any relation between 𝑓4 , 𝑔6 , ℎ12 is a multiple of 𝑟. [Hint: Without
loss of generality, we may suppose that 𝑟 is of the form ℎ212 ∈ C[ 𝑓4 , 𝑔6 ] 24 .
Therefore, another relation expresses ℎ12 as a rational function in 𝑓4 , 𝑔6 .
Use (a) and unique factorization to show that this purported relation is in
fact polynomial, and obtain a contradiction from the linear independence
of 𝑓43 , 𝑔62 , ℎ12 .]
(c) Show that the subring of 𝑀 (Γ1 ) generated by 𝑓4 , 𝑔6 , ℎ12 is isomorphic to
𝑀 0 = C[ 𝑓4 , 𝑔6 , ℎ12 ]/h𝑟 ( 𝑓4 , 𝑔6 , ℎ12 )i.
(d) Show that

∑︁ 1 + 𝑡 12
(dimC 𝑀𝑘 (Γ1 ))𝑡 𝑘 =
𝑘 ∈2Z≥0
(1 − 𝑡 4 ) (1 − 𝑡 6 )
and dimC 𝑀𝑘 = dimC 𝑀𝑘0 for all 𝑘. [Hint: use Exercise 43.8.]
(e) Conclude that the subring of 𝑀 (Γ1 ) generated by 𝑓4 , 𝑔6 , ℎ12 is equal to
𝑀 (Γ1 ). [Hint: Suppose that equality does not hold, and consider the
minimal degree with a new generator; by dimensions, there must be a new
relation, but argue that this relation must be among 𝑓4 , 𝑔6 , ℎ12 , contradict-
ing (b).]
Symbol Definition List
kk idelic norm, 452

k𝜙k
modulus of an automorphism 𝜙 of a topological group, 498
𝐾, 𝑏
, (𝐾, 𝑏 | 𝐹) quaternion algebra defined by 𝐾 ⊇ 𝐹 and 𝑏 ∈ 𝐹 × , 84
𝐹

𝐾
splitting (Legendre) symbol, 540
𝔭

O
, (O | 𝔭) Eichler symbol of a local , 388
𝔭
(𝑎,
𝑏)𝐹 Hilbert symbol on 𝑎, 𝑏 ∈ 𝐹 × , 187
𝑎, 𝑏
, (𝑎, 𝑏 | 𝐹) quaternion algebra defined by 𝑎, 𝑏 ∈ 𝐹 × (char 𝐹 ≠ 2), 22
𝐹
involution, 36
∗, † positive involution, 126
◊ fundamental set, 617
∗ conjugate transpose (of a complex matrix), 27
h𝑎 1 , . . . , 𝑎 𝑛 i diagonal quadratic form on 𝑎 1 , . . . , 𝑎 𝑛 , 51
h i additive homomorphism of a local field, 504
𝑎, 𝑏
, [𝑎, 𝑏 | 𝐹) quaternion algebra defined by 𝑎, 𝑏 ∈ 𝐹 × , char 𝐹 = 2, 84
𝐹
|| absolute value, 179
| |𝑝 𝑝-adic absolute value, 175
2𝐷 𝑛 binary dihedral (dicyclic) group of order 2𝑛, 170
2𝐼 binary icosahedral group, 169
2𝑂, 𝑆e4 binary octahedral group, 168
2𝑇, 𝐴 e4 binary tetrahedral group, 162
[𝑀 : 𝑁] 𝑅 index of 𝑅-lattices 𝑀, 𝑁 ⊆ 𝑉, 143
𝛼, 𝛽 elements (of an algebra), 23
𝐴 𝐹-algebra, 39
𝔞, 𝔟 fractional ideals (of a noetherian domain), 136
𝑎 𝔫 (O) count of right O-ideals of reduced norm 𝔫, 424
AL(O) the Atkin–Lehner group of the 𝑅-order O, 485
𝛼, 𝛽b
b adelic or idelic elements of an algebra, 450
𝐴𝑛 alternating group on 𝑛 letters, 161
815
ann 𝑉 annihilator of a (left) module 𝑉, 102

Aut𝐹 𝐵 automorphism group of an 𝐹-algebra 𝐵, 22
𝐵 adele ring of the algebra 𝐵 over a global field, 450
𝐵 algebra, 21
𝐵0 reduced trace zero elements of 𝐵, 38
𝐵1 reduced norm one elements of 𝐵, 38
𝐵× idele group of an algebra 𝐵 over a global field, 450
bd boundary, 596
𝐵
b finite adele ring of the algebra 𝐵 over a global field, 450
b×
𝐵 finite idele group of an algebra 𝐵 over a global field, 450
𝐵op opposite algebra of the algebra 𝐵, 36
orthogonal direct sum, 51
Br 𝐹 Brauer group of a field 𝐹, 125
BrtCl O Brandt class groupoid of the order O, 303
C field of complex numbers, 21
𝐶 𝐵 ( 𝐴) centralizer of a subalgebra 𝐴 ⊆ 𝐵, 108
𝐶𝐹 idele class group of the global field 𝐹, 452
char 𝐹 characteristic of a field 𝐹, 22
cl closure, 617
Clf 0 𝑄 even Clifford algebra of a quadratic form 𝑄, 69
Clf 1 𝑄 odd Clifford bimodule of a quadratic form 𝑄, 69
Clf 𝑄 Clifford algebra of a quadratic form 𝑄, 67
Cl𝐺 𝑁 (O) 𝑅 class group of 𝑅 attached to the normalizer of O, 474
Cl𝐺 (O) 𝑅 class group of 𝑅 attached to the order O, 461
Cl 𝑄 class set of a quadratic module 𝑄, 146
ClsR O, Cls O (right) class set of an 𝑅-order O, 273
𝐶𝑛 cyclic group of order 𝑛, 170
covol covolume, 276
CSA(𝐹) set of isomorphism classes of central simple 𝐹-algebras, 124
𝐷 division ring, 22
D2 hyperbolic unit disc, 604
D3 hyperbolic unit ball, 655
𝑑 (𝛼1 , . . . , 𝛼𝑛 ) discriminant of elements in a semisimple algebra 𝐵, 234
D (𝐵) absolute discriminant of the algebra 𝐵, 505
dim𝐹 𝑉 dimension of an 𝐹-vector space 𝑉, 21
disc 𝐵 discriminant of a quaternion algebra 𝐵, 210
disc(𝐼) discriminant of an 𝑅-lattice 𝐼, 234
disc(𝑄) discriminant of a quadratic form 𝑄, 52
disc𝑅 𝐵 𝑅-discriminant of a quaternion algebra 𝐵, 222
𝐷𝑛 dihedral group of (even) order 𝑛, 170
𝑒 quadratic nonresidue, 184
𝐸 = 𝐸 𝑆,O elements conjugating a fixed embedding into optimal embeddings
𝑆 ↩→ O, 534
b= 𝐸
𝐸 bb b
𝑆, O
elements conjugating a fixed embedding into adelic optimal em-
beddings 𝑆b ↩→ O,
b 536
𝐸𝑘 , 𝐺 𝑘 Eisenstein series, 731
b O;
Emb( 𝑆, bbΓ) Γ-conjugacy classes of optimal adelic embeddings 𝑆b ↩→ O
set of b b
of the quadratic order 𝑆 into the quaternion order O, 536
b b
Emb𝑅 (𝑆, O) set of optimal embeddings 𝑆 ↩→ O of the quadratic order 𝑆 into
the quaternion order O, 534
E𝑛 Euclidean 𝑛-space, 592
Emb(𝑆, O; Γ) set of Γ-conjugacy classes of optimal embeddings 𝑆 ↩→ O, 534
𝐹 field, 21
𝐹 (1) product norm 1 ideles, 452
𝐹 al algebraic closure of the field 𝐹, 21
𝑓∨ Fourier transform of Schwartz function 𝑓 , 492
𝔣(O) conductor of the quaternion 𝑅-order O, 385
𝐹>×Ω 0 elements of 𝐹 positive at the real places 𝑣 ∈ Ω, 225
𝐹 sep separable closure of 𝐹, 91
Gal(𝐾 | 𝐹) Galois group of the Galois extension of fields 𝐾 ⊇ 𝐹, 24
Γ0 (𝑁), Γ1 (𝑁) standard congruence subgroups of level 𝑁, 644
Γ(𝑁) full congruence subgroup of level 𝑁, 644
ΓR (𝑠), ΓC (𝑠) standard archimedean Γ-factors, 436
Gen O genus of an 𝑅-order O, 274
Gen 𝑄 genus of a quadratic module 𝑄, 146
𝐺 𝑁 (O) idelic reduced norm group attached to the normalizer of O, 474
𝐺 (O) idelic reduced norm group attached to O, 461
GO(𝑄) (𝐹) general orthogonal group of 𝑄, 49
Gor(O) Gorenstein saturation of the quaternion 𝑅-order O, 385
G linear algebraic group, 709
GSO(𝑄) (𝐹) general special orthogonal group of a quadratic form 𝑄, 55
H (real) Hamiltonians, 23
H0 pure Hamiltonians, 28
H1 unit Hamiltonians, 27
H2 upper half-plane, 594
H2± union of upper and lower half-planes, 595
H2∗ completed upper half-plane, 596
H3 upper half-space, 649
H product of hyperbolic planes and spaces, 704
H𝑛 hyperbolic 𝑛-space, 656
𝐻 hyperbolic plane, 71
ℎ(𝑆) class number of 𝑆, 536
𝐼 lattice (over a domain, in an algebra), 152
𝐼2 , 𝐼4 , 𝐼6 , 𝐼10 Igusa invariants, 800
𝐼 (𝛼) axis of 𝛼 ∈ H1 , 29
Idl(O) group of invertible two-sided fractional O-ideals, 290

𝐼 −1 quasi-inverse of a lattice 𝐼, 257
𝑖, 𝑗 standard generators for a quaternion algebra, 23
int interior, 617
𝐼♯ dual of the 𝑅-lattice 𝐼 with respect to the reduced trace, 242
𝚥 automorphy factor, 735
𝐾 𝐾 ⊇ 𝐹 quadratic algebra, 39
𝑘 residue field, 180
𝐿 projective 𝑅-module of rank 1, 145
L2 Lorentz hyperboloid, 606
L3 Lorentz model of hyperbolic 3-space, 656
Λ lattice, 643
𝜆 left regular representation, 24
𝜆(O, 𝔭) mass correction term at a prime 𝔭 for the quaternion order O, 411
𝐿∨ dual of the the invertible 𝑅-module 𝐿, 237
L Lobachevsky function, 658
ℓ(𝜐) length of the path upsilon, 592
𝔪 maximal ideal (of a domain), 139
𝑀, 𝑁 lattices (over a domain, inside a vector space), 137
𝑚(𝛼1 , 𝛼2 , 𝛼3 ) box product of elements in a quaternion algebra, 239
mass(Cls O) mass of the class set of a quaternion order O, 420
M𝑛 ( 𝐴) ring of 𝑛 × 𝑛-matrices over a ring 𝐴, 21
𝑀 (𝔭) localization of an 𝑅-lattice 𝑀 at 𝔭, 139
𝑀𝔭 completion of an 𝑅-lattice 𝑀 at 𝔭, 142
𝑚( 𝑆,b O;
bb Γ) number of b Γ-conjugacy classes of optimal adelic embeddings 𝑆b ↩→
O, 536
b
𝑚(𝑆, O; Γ) number of Γ-conjugacy classes of optimal embeddings 𝑆 ↩→ O,
535
𝜇 adelic (product) measure, 509
𝜇 polarization (on a quaternion order), 801
𝜇× idelic (product) measure on Q× , 494
𝜇𝑝 standard Haar measure on Q 𝑝 , 493
𝜇×𝑝 normalized Haar measure on Q×𝑝 , 494
N absolute norm, 256
𝑁 𝐵× normalizer in 𝐵× , 171
Nm algebra norm, 39
nrd reduced norm, 38
O order (in an algebra), 151
O
b profinite completion of the order O over a global ring, 450
O×
b finite idele group of the order O over a global ring, 450
OL (𝐼), OR (𝐼) left, right order of a lattice 𝐼 (in an algebra), 151
Ω set of real, ramified places of a quaternion algebra 𝐵, 225
O(𝑛) orthogonal group, 29

O♮ radical idealizer of the local 𝑅-order O, 391
O(𝑄) (𝐹) orthogonal group of 𝑄, 49
𝔭 prime ideal (of a domain), 139
𝑃 maximal (two-sided) ideal of valuation ring, 195
P1 ( 𝐴) projective line over a commutative ring 𝐴, 427
P𝐺 the group 𝐺 modulo scalars, 580
𝜑 Euler totient function, 403
Π big period matrix, 793
𝜋 uniformizer, 180
Pic Picard group, 290
Pic(O, O0) set of homothety classes of lattices with left, right order O, O0, 303
Pic𝑅 (O) Picard group of the 𝑅-order O, 295
PIdl(O) group of principal two-sided fractional O-ideals, 290
𝜛 nontrivial normalizer of an Eichler order, 366
Pl 𝐹 set of places of a global field 𝐹, 219
𝔭 maximal ideal (of a valuation ring), 180
℘ Artin–Schreier group, 185
Ψ standard function on the algebra 𝐵, the characteristic function of a
maximal order, 506
𝜓 Dedekind 𝜓-function, 411
𝜓 standard unitary character, 501
Q× rational idele group, 448
Q field of rational numbers, 21
QS product of completions of Q at S, 447
𝑄 quadratic form, 49
𝑄8 quaternion group of order 8, 160
Q
b profinite completion of Q, 448
Q6S projection of rational adele ring away from S, 447
Q𝑝 field of 𝑝-adic numbers, 175
Q rational adele ring, 446
R field of real numbers, 21
𝑅 (commutative) noetherian domain, 136
𝑅 valuation ring, 180
rad radical of a quadratic form, 53
rad 𝐴 Jacobson radical of a ring 𝐴, 316
rad 𝐵 Jacobson radical of an algebra 𝐵, 101
Ram 𝐵 ramification set of a quaternion algebra 𝐵, 210
rev reversal map (on a Clifford algebra), 69
𝜌 distance (in a metric space), 592
𝑅 (𝔭) localization of a domain 𝑅 at 𝔭, 139
𝑅𝔭 completion of a domain 𝑅 at 𝔭, 142
𝑅 (S) ring of S-integers of a global field, 221
S set of places of a global field, 221

𝑆 quadratic 𝑅-algebra, 200
sgndisc(𝑄) signed discriminant of a quadratic form 𝑄, 52
𝑆𝑛 symmetric group on 𝑛 letters, 160
SO(𝑛) special orthogonal group, 29
SO(𝑄) (𝐹) special orthogonal group of the quadratic form 𝑄, 55
Stab𝐺 (𝑥) stabilizer of 𝑥 ∈ 𝑋 under the action by the group 𝐺, 616
StCl O stable (or locally free) class group, 327
SU(𝑛) special unitary group, 27
T Bruhat–Tits tree, 378
𝑇 associated (symmetric) bilinear form, 49
𝜏 Tamagawa measure, 510
𝜏 self-dual measure, 499
𝜏𝑥 reflection in 𝑥, 56
Ten0 𝑉 even tensor algebra of a vector space 𝑉, 70
Ten 𝑉 tensor algebra of a vector space 𝑉, 67
Θ Jacobi theta function, 492
𝜗(◊, 𝑣), 𝜗(◊, ℓ) interior angle of ◊ at the vertex 𝑣 or edge ℓ, 686
𝑇 (𝔫) 𝔫-Brandt matrix, 747
Tr algebra trace, 39
trd reduced trace, 38
Typ O type set of an 𝑅-order O, 274
t
transpose (of a matrix), 27
𝜐 path, 592
𝑈S open subset of the adeles, 446
𝑉 finite-dimensional vector space, 137
𝑣, 𝑤 valuation, 180
◊(Γ; 𝑧0 ) Dirichlet domain for Γ centered at 𝑧 0 , 671
𝑤𝐽 unit group modular scalars of the left order of the 𝑅-lattice 𝐽, 403
𝜔 Hurwitz element, 159
℘ Weierstrass ℘-function, 733
𝑋 metric space, 592
𝜉 𝐵 (𝑠) completed zeta function of the quaternion algebra 𝐵, 437
Z ring of integers, 21
𝑍 (𝐵) center of an algebra 𝐵, 21
𝑍 𝐵Φ (𝑠) local zeta function of the algebra 𝐵 with respect to the function Φ,
506
𝜁 orientation of a quadratic form, 76
𝜁 𝐵 (𝑠) zeta function of the quaternion algebra 𝐵, 409
𝜁 𝐾 , [𝔟] (𝑠) partial zeta function of the number field 𝐾, 405
𝜁 𝐾 (𝑠) Dedekind zeta function of the number field 𝐾, 405
𝜁O (𝑠) zeta function of the quaternion order O, 408
𝜁 (𝑠) Riemann zeta function, 404
Z𝐹 ring of integers of a number field 𝐹, 221

Zh𝑖, 𝑗i Lipschitz order, 159
Z𝑝 ring of 𝑝-adic integers, 175
𝜁 ∗ (𝑎) leading coefficient in Laurent series expansion of 𝜁 at 𝑠 = 𝑎, 422
Z
b profinite completion of Z, 445
Bibliography
[AB2008] Peter Abramenko and Kenneth S. Brown, Buildings: theory and applications,
Grad. Texts in Math., vol. 248, Springer, New York, 2008. 23.5.17
[AA66] J. F. Adams and M. F. Atiyah, 𝐾-theory and the Hopf invariant, Quart. J. Math.
Oxford (2), 17 (1966), 31–38. 1.2
[Alb34] A. A. Albert, Integral domains of rational generalized quaternion algebras,
Bull. Amer. Math. Soc. 40 (1934), 164–176. 1.3, 14.2.8
[Alb39] A. Adrian Albert, Structure of algebras, AMS Colloquium Publications, vol. 24,
Amer. Math. Soc., Providence, RI, 1939. 1.3, 8.2.9
[Alb72] A. A. Albert, Tensor products of quaternion algebras, Proc. Amer. Math. Soc. 35
(1972), no. 1, 65–66. 8.2.9
[AH32] A. Adrian Albert and Helmut Hasse, A determination of all normal division
algebras over an algebraic number field, Trans. Amer. Math. Soc. 34 (1932), no.
3, 722–726. 1.3, 14.6.6
[Alp93] Roger C. Alperin, 𝑃𝑆𝐿 2 (Z) = Z2 ∗ Z3 , Amer. Math. Monthly 100 (1993), no. 4,
385–386. 35.1.20
[AB2004] Montserrat Alsina and Pilar Bayer, Quaternion orders, quadratic forms, and
Shimura curves, CRM Monograph Series, vol. 22, American Math. Soc., Provi-
dence, RI, 2004. iv (document), 1.4, 14.2.13, 37.9.13
[Alt89] Simon L. Altmann, Hamilton, Rodrigues, and the quaternion scandal, Math.
Magazine 62 (1989), no. 5, 291–308. 1.1
[AO2005] Simon L. Altmann and Eduardo L. Ortiz, eds., Mathematics and social utopias in
France: Olinde Rodrigues and his times, Amer. Math. Society, Providence, RI,
2005. 1.1
[ART79] S. A. Amitsur, L. H. Rowen, and J. P. Tignol, Division algebras of degree 4 and
8 with involution, Israel J. Math. 33 (1979), 133–148. 8.3.6
[Apo76] Tom M. Apostol, Introduction to analytic number theory, Undergraduate Texts in
Math., Springer-Verlag, New York-Heidelberg, 1976. 14.2
[Apo90] Tom M. Apostol, Modular functions and Dirichlet series in number theory,
Springer-Verlag, New York, 1990. 35, 40.1
[And2005] James W. Anderson, Hyperbolic geometry, 2nd. ed., Springer-Verlag, London,
2005. 33.1
[A-C2012] Luis Arenas-Carmona, Maximal selectivity for orders in fields, J. Number Theory
132 (2012), no. 12, 2748–2755. 31.7.7
[A-C2013] Luis Arenas-Carmona, Eichler orders, trees and representation fields, Int. J.
Number Theory 9 (2013), no. 7, 1725–1741. 31.7.7
[Art26] Emil Artin, Zur Theorie der hyperkomplexen Zahlen, Abh. Math. Sem. Hambur-
gischen Univ. 5 (1926), 251–260. 7.3.11
823
824 BIBLIOGRAPHY
[AT2008] Alexander Arhangel’skii and Mikhail Tkachenko, Topological groups and related
structures, Atlantist Studies in Math., vol. 1, World Scientific Publishing Co.,
Hackensack, NJ, 2008. 34.1
[Arm88] M. A. Armstrong, Groups and symmetry, Undergrad. Texts in Math., Springer-
Verlag, New York, 1988. 11.5
[Art50] Emil Artin, The influence of J. H. M. Wedderburn on the development of modern
algebra, Bull. Amer. Math. Soc. 56 (1950), no. 1, 65–72. 1.2
[Asl96] Helmer Aslaksen, Quaternionic determinants, Math. Intelligencer 18 (1996), no.
3, Summer 1996, 57–65. 2.20
[AM69] Michael Francis Atiyah and Ian Grant Macdonald, Introduction to commutative
algebra, Reading, Addison-Wesley, 1969. 9.2, 9.5
[Ati67] M. F. Atiyah, 𝐾-theory, W. A. Benjamin, Inc., New York, 1967.
[ABB2014] Asher Auel, Marcello Bernardara, and Michele Bolognesi, Fibrations in complete
intersections of quadrics, Clifford algebras, derived categories, and rationality
problems, J. Math. Pures Appl. (9) 102 (2014), no. 1, 249–291. 22.2.14
[ABGV2006] Asher Auel, Eric Brussel, Skip Garibaldi, and Uzi Vishne, Open problems on
central simple algebras, Transform. Groups 16 (2011), no. 1, 219–264. 7.5.5
[Auel2011] Asher Auel, Clifford invariants of line bundle-valued quadratic forms, MPIM
preprint series, no. 33, January 2011. 22.2.14, 22.2.15
[Auel2015] Asher Auel, Surjectivity of the total Clifford invariant and Brauer dimension, J.
Algebra, 443 (2015), 395–421. 22.2.14
[AG60] Maurice Auslander and Oscar Goldman, Maximal orders, Trans. Amer. Math.
Soc. 97 (1960), 1–24. 15.17, 18.3
[BG2008] Srinath Baba and Håkan Granath, Genus 2 curves with quaternionic multiplica-
tion, Canad. J. Math. 60 (2008), no. 4, 734–757. 43.2.10, 43.2, 43.2.15, 43.2.16,
43.2.16, 43.2.20, 43.2.21
[Bae2002] John C. Baez, The octonions, Bull. Amer. Math. Soc. (N.S.) 39 (2002), 145–205;
Errata for “The octonions”, Bull. Amer. Math. Soc. (N.S.) 42 (2005), 213. 1.2
[Bae2005] John C. Baez, On quaternions and octonions: their geometry, arithmetic, and
symmetries by John H. Conway and Derek A. Smith, Bull. Amer. Math. Soc. 42
(2005), 229–243. 11.1
[Bak71] A. Baker, Linear forms in the logarithms of algebraic numbers, Mathematika 13
(1966), 204–216. 25.2.19
[Bal2007] Venkata Balaji Thiruvalloor Eesanaipaadi, Line-bundle-valued ternary quadratic
forms over schemes, J. Pure Appl. Algebra 208 (2007), 237–259. 22.5.13
[Bas62] Hyman Bass, Injective dimension in Noetherian rings, Trans. Amer. Math. Soc.
102 24.2.24
[Bas63] Hyman Bass, On the ubiquity of Gorenstein rings, Math. Z. 82 (1963), 8–28.
24.2.24
[BHV2008] Bachir Bekka, Pierre de la Harpe, and Alain Valette, Kazhdan’s Property (T),
Cambridge University Press, Cambridge, 2008. 29.3
[BCK93] F. Beukers, E. Calabi, and J. Kolk, Sums of generalized harmonic series and
volumes, Nieuw Arch. Wisk. 11 (1993), 217–224. 25.1
[Bea95] A. Beardon, The geometry of discrete groups, Grad. Texts in Math., vol. 91,
Springer-Verlag, New York, 1995. 33.1, 34.7, 37.3
[BP92] R. Benedetti and C. Petronio, Lectures on hyperbolic geometry, Universitext,
Springer, Berlin, 1992. 36.3.17
[BO2013] Grégory Berhuy and Frédérique Oggier, An introduction to central simple alge-
bras and their applications to wireless communication, Math. Surveys Monogr.,
BIBLIOGRAPHY 825
vol. 191, American Math. Soc., Providence, RI, 2013. 1.5

[BK2000] A. J. Berrick and M. E. Keating, An introduction to rings and modules with 𝐾-
theory in view, Cambridge Studies in Adv. Math., vol. 65, Cambridge University
Press, Cambridge, 2000. 20.2
[Bha2004a] Manjul Bhargava, Higher composition laws. I. A new view on Gauss composition,
and quadratic generalizations, Ann. of Math. (2) 159 (2004) 217–250. 19.1
[Bia1892] Luigi Bianchi, Sui gruppi di sostituzioni lineari con coefficienti appartenenti a
corpi quadratici immaginari, Math. Ann. 40 (1892), 332–412. 36.4.4
[Bic85] Walter Bichsel, Quadratische raeume mit werten in invertierbaren moduln, Ph.D.
thesis, ETH Zürich, 1985. 22.2.14
[BK94] W. Bichsel and M.-A. Knus, Quadratic forms with values in line bundles, Re-
cent advances in real algebraic geometry and quadratic forms (Berkeley, CA,
1990/1991; San Francisco, CA, 1991), Contemp. Math., vol. 155, Amer. Math.
Soc., Providence, RI, 1994, 293–306. 22.2.13, 22.2.14
[BL2004] Christina Birkenhake and Herbert Lange, Complex abelian varities, second aug-
mented ed., Springer-Verlag, Berlin, 2004. 8.5, 43.4
[Bol1887] Oskar Bolza, On binary sextics with linear transformations into themselves, Amer.
J. Math. 10 (1887), no. 1, 47–70. 43.3.4
[Bor62] Armand Borel, Arithmetic properties of linear algebraic groups, Proc. Internat.
Congr. Mathematicians (Stockholm, 1962), Inst. Mittag-Leffler, Djursholm, 1963,
10–22. 38.5.1
[Bor69] Armand Borel, Introduction aux groupes arithmétiques, Publications de l’Institut
de Mathématique de l’Université de Strasbourg, vol. XV, Actualités Scientifiques
et Industrielles, No. 1341, Hermann, Paris, 1969. 38.5.1
[Bor81] A. Borel, Commensurability classes and volumes of hyperbolic 3-manifolds, Ann.
Scuola Norm. Sup. Pisa Cl. Sci. (4) 8 (1981), no. 1, 1–33. 39.1.15, 39.3
[BHC62] Armand Borel and Harish-Chandra, Arithmetic subgroups of algebraic groups,
Ann. of Math. (2) 75 (1962), 485–535. 38.5.6
[BS66] Z.I. Borevich and I.R. Shafarevich, Number theory, Academic Press, New York,
1966. 26.2, 26.2, 26.2
[BM58] R. Bott and J. Milnor, On the parallelizability of the spheres, Bull. Amer. Math.
Soc. 64 (1958), 87–89. 1.2
[Bou60] Nicolas Bourbaki, Elements of mathematics: General topology, Part 1, Addison-
Wesley, Reading, 1966. 34.5.4
[Bou98] Nicolas Bourbaki, Commutative algebra, Chapters 1–7, Springer-Verlag, Berlin,
1998. 9.2, 9.5
[Bra13] H. Brandt, Zur Komposition der quaternären quadratischen Formen, J. Reine
Angew. Math. 143 (1913), 106–129. 19.1
[Bra24] H. Brandt, Der Kompositionsbegriff bei den quaternären quadratischen Formen,
Math. Ann. 91 (1924), no. 3–4, 300–315. 19.1
[Bra25] H. Brandt, Die Hauptklassen in der Kompositionstheorie der quaternären
quadratischen Formen, Math. Ann. 94 (1925), no. 1, 166–175. 19.1
[Bra27] H. Brandt, Über eine Verallgemeinerung des Gruppenbegriffes, Math. Ann. 96
(1927), 360–366. 19.3.11
[Bra28] H. Brandt, Idealtheorie in Quaternionenalgebren, Math. Annalen 99 (1928), 1–
29. 1.4, 17.1, 19.1, 25.1
[Bra37] H. Brandt, Zur Zahlentheorie der quadratische Formen, Jber. Deutsch. Math.-
Verein. 47 (1937), 149–159. 19.1
[Bra40] H. Brandt, Uber die Axiome des Gruppoids, Vierteljschr. naturforsch. Ges. Zurich
826 BIBLIOGRAPHY
85 (1940), Beiblatt (Festschrift Rudolph Fueter), 95–104. 1.4, 19.3.11

[Bra41] H. Brandt, Uber die Zerlegungsgesetze der rationalen Zahlen in Quaternionen-
Körpern, Math. Ann. 117 (1941), 758–763. 1.4
[Bra43] H. Brandt, Zur Zahlentheorie der Quaternionen, Jahresbericht der Deutschen
Mathematiker-Vereinigung 53 (1943), 23–57. 1.4, 19.1, 22.5.13, 41.5.13
[BHN31] R. Brauer, H. Hasse, and E. Noether, Beweis eines Hauptsatzes in der Theorie
der Algebren, J. Reine Angew. Math. 167 (1932), 399–404. 1.3, 14.6.6
[Bre2010] Matej Brešar, An elementary approach to Wedderburn’s structure theory, Expo-
sition. Math. 28 (2010), 79–83. 7.2
[BH99] Martin R. Bridson and André Haefliger, Metric spaces of non-positive curva-
ture, Grundlehren der Mathematischen Wissenschaften, vol. 319, Springer-Verlag,
Berlin, 1999. 33.2
[Bro61] Bancroft H. Brown, Mathematics at Dartmouth: 1769–1961, Dedicatory Con-
ference, Albert Bradley Center for Mathematics, Dartmouth College, Hanover,
November 3, 1961. v (document)
[Bro87] Ronald Brown, From groups to groupoids: a brief survey, Bull. London
Math. Soc. 19 (1987), 113–134. 19.3.11
[Bruc71] Richard Herbert Bruck, A survey of binary systems, third printing, Springer-
Verlag, Berlin, 1971. 19.3.11
[Brum63a] Armand Brumer, The structure of hereditary orders, Ph.D. thesis, Princeton,
1963. 21.4
[Brum63b] Armand Brumer, Structure of hereditary orders, Bull. Amer. Math. Soc. 69
(1963), 721–724; Addendum to “Structure of hereditary orders”, Bull. Amer.
Math. Soc. 70 (1964), 185. 21.4
[Brz80] J. Brzezinski, Arithmetical quadratic surfaces of genus 0, I, Math. Scand. 46
(1980), 183–208. 22.5.13, 24.3.11
[Brz82] J. Brzezinski, A characterization of Gorenstein orders in quaternion algebras,
Math. Scand. 50 (1982), no. 1, 19–24. 15.11, 16.8, 22.5.13, 24.2.24
[Brz83a] Julius Brzezinski, On orders in quaternion algebras, Comm. Algebra 11 (1983),
no. 5, 501–522. 23.5, 24.1.5, 24.2.18, 24.5, 24.9
[Brz83b] Julius Brzezinski, Spinor class groups of orders, J. Algebra 84 (1983), 468–481.
20.7.23, 22.5.13
[Brz85] Julius Brzezinski, Brauer–Severi schemes of orders, Orders and their applications
(Oberwolfach, 1984), Lecture Notes in Math., vol. 1142, Springer, Berlin, 1985,
18–49. 22.5.13, 24.3.11
[Brz87] J. Brzezinski, Riemann–Roch theorem for locally principal orders, Math. Ann.
276 (1987), 529–536. 24.2.24
[Brz89] J. Brzezinski, A combinatorial class number formula, J. Reine Angew. Math. 402
(1989), 199–210. 30.4.22
[Brz90] J. Brzezinski, On automorphisms of quaternion orders, J. Reine Angew. Math.
403 (1990), 166–186. 24.1.5, 24.5.7, 24.5, 24.5, 25.3, 26.1.14, 30.6.18
[Brz91] J. Brzezinski, On embedding numbers into quaternion orders, Comment. Math.
Helvetici 66 (1991), 302–318. 30.6.18
[Brz95] Juliusz Brzezinski, Definite quaternion orders of class number one, J. Théorie
Nombres Bordeaux 7 (1995), 93–96. 25.4
[Brz97] Juliusz Brzezinski, A generalization of Eichler’s trace formula, Collect. Math. 48
(1997), 53–61. 30.9.12
[Brz98] Juliusz Brzezinski, On traces of the Brandt–Eichler matrices, J. Théorie Nombres
Bordeaux 10 (1998), no. 2, 273–285. 25.4
BIBLIOGRAPHY 827
[B+2003] D. Bump, J.W. Cogdell, E. de Shalit, D. Gaitsgory, E. Kowalski, S. S. Kudla,

An introduction to the Langlands program, eds. Joseph Bernstein and Stephen
Gelbart, Birkhäuser, Boston, 2003. 1.4
[BBI2001] Dmitri Burago, Yuri Burago, and Sergei Ivanov, A course in metric geometry,
Grad. Studies in Math., vol. 33, Amer. Math. Soc., 2001. 33.2.8
[Bus55] Herbert Busemann, The geometry of geodesics, Academic Press Inc., New York,
1955. 33.2.8
[BFS2008] Peter Buser, Nicole Flach, and Klaus-Dieter Semmler, Spectral rigidity and
discreteness of 2233-groups, with an appendix by Colin MacLachlan and Gerhard
Rosenberger, Math. Proc. Cambridge Philos. Soc. 144 (2008), no. 1, 145–180,
31.8
[BR85] Colin J. Bushnell and Irving Reiner, A survey of analytic methods in noncommu-
tative number theory, Orders and their applications (Oberwolfach, 1984), Lecture
Notes in Math., vol. 1142, Springer, Berlin, 1985, 50–87. 26.3.14
[CFKP97] James W. Cannon, William J. Floyd, Richard Kenyon, and Walter R. Parry,
Hyperbolic geometry, Flavors of geometry, ed. Silvio Levy, MSRI Publ., vol. 31,
Cambridge University Press, Cambridge, 1997, 59–115. 33.1, 36.3.17
[CNP2005] G. Cardona, E. Nart, and J. Pujolas, Curves of genus two over fields of even
characteristic, Math. Z. 250 (2005), 177–201. 43.3.10
[CQ2005] G. Cardona and J. Quer, Field of moduli and field of definition for curves of genus
2, Computational aspects of algebraic curves, Lecture Notes Ser. Comput. 13,
World Sci. Publ. (2005), Hackensack, NJ, 71–83. 43.3.10, 43.3.13
[Cas78] J.W.S. Cassels, Rational quadratic forms, Academic Press, London, 1978. 4.2,
14.4.4, 14.4.12
[Cas86] J.W.S. Cassels, Local fields, London Math. Soc. Student Texts, vol. 3, Cambridge
University Press, Cambridge, 1986. 12.2, 12.2, 13.2
[Cas97] J.W.S. Cassels, An introduction to the geometry of numbers, Springer, Berlin,
1997. 17.5
[Cas2010] J.W.S. Cassels, Global fields, Algebraic number theory, 2nd. ed., J.W.S. Cassels
and A. Frölich, eds., London Mathematical Society, London, 2010, 42–84. 14.4,
27.4, 27.4
[Cay1845a] Arthur Cayley, On certain results relating to quaternions, Phil. Mag. 26 (1845),
141–145. 1.2
[Cay1845b] Arthur Cayley, On Jacobi’s elliptic functions, in reply to the Rev. B. Bronwin; and
on quaternions, Phil. Mag. 26 (1845), 208–211. 1.2
[CCL2013] Jean-Paul Cerri, Jérôme Chaubert, and Pierre Lezowski, Euclidean totally definite
quaternion fields over the rational field and over quadratic number fields, Inter.
J. Number Theory 9 (2013), no. 3, 653–673. 26.7.4
[CX2004] Wai Kiu Chan and Fei Xu, On representations of spinor genera, Compositio
Math. 140 (2004), no. 2, 287–300. 31.7, 31.7.7
[CDL2015] Adam Chapman, Andrew Dolphin, and Ahmed Laghribi, Total linkage of quater-
nion algebras and Pfister forms in characteristic two, J. Pure Appl. Algebra, 220
(2016), 3676–3691. 8.2
[Cha2020] Sara Chari, Metacommutation of primes in central simple algebras, J. Number
Theory 206 (2020), 296–309. 11.4.11
[CGL2009] D. Charles, E. Goren, and K. Lauter, Families of Ramanujan graphs and quater-
nion algebras, Groups and Symmetries, CRM Proc. Lecture Notes 47 (2009),
53–80. 41.2.14
[Chev34] Claude Chevalley, Sur certains idéaux dans un algèbre simple, Abh. Math. Sem.
828 BIBLIOGRAPHY
Hamburger Univ. 10 (1934), 83–105. 30.3.17

[Chev36] Claude Chevalley, L’arithmétique dans les algèbres de matrices, Actualités Sci.
Ind. no. 323 (1936). 31.1.8
[CSV2019] Sara Chari, Daniel Smertnig, and John Voight, On basic and Bass quaternion
orders, 2019, arXiv:1903.00560. 24.5, 24.5
[Chi2009] Nancy Childress, Class field theory, Universitext, Springer, New York, 2009. 27
[CF99] Ted Chinburg and Eduardo Friedman, An embedding theorem for quaternion
algebras, J. London Math. Soc. (2) 60 (1999), no. 1, 33–44. 31.1.8, 31.7, 31.7.7
[CF2000] Ted Chinburg and Eduardo Friedman, The finite subgroups of maximal arithmetic
Kleinian groups, Ann. Inst. Fourier (Grenoble) 50 (2000), no. 6, 1765–1798.
32.5, 32.6, 32.6.5
[Cli1878] William K. Clifford, Applications of Grassmann’s extensive algebra, Amer. Jour.
Math. 1 (1878), 350–358. 1.2
[Coh93] Henri Cohen, A course in computational algebraic number theory, Grad. Texts in
Math., vol. 138, Springer-Verlag, Berlin, 1993. 11.15
[CK2015] Henry Cohn and Abhinav Kumar, Metacommutation of Hurwitz primes, Proc.
Amer. Math. Soc. 143 (2015), no. 4, 1459–1469. 11.4.11
[Con] Keith Conrad, A multivariable Hensel’s lemma, preprint, http://kconrad.
math.uconn.edu/blurbs/gradnumthy/multivarhensel.pdf. 12.12a
[CG93] J. B. Conrey and A. Ghosh, On the Selberg class of Dirichlet series: small
degrees, Duke Math. J. 72 (1993), no. 3, 673–693. 26.3.15
[Con2012] John B. Conway, A course in abstract analysis, Grad. Studies in Math., vol. 141,
Amer. Math. Soc., Providence, RI, 2012. 3.2.11
[CSl88] J. H. Conway and N. J. A. Sloane, Sphere packings, lattices and groups,
Grundlehren der Mathematischen Wissenschaften, vol. 290, Springer-Verlag,
New York, 1988. 11.5.7
[CSm2003] John H. Conway and Derek A. Smith, On quaternions and octonions: their
geometry, arithmetic, and symmetry, A. K. Peters, Ltd., Natick, MA, 2003. 1.2,
11.1, 11.2.4, 11.4.11, 11.5
[Cox89] David A. Cox, Primes of the form 𝑥 2 + 𝑛𝑦 2 : Fermat, class field theory, and
complex multiplication, John Wiley & Sons, Inc., New York, 1989. 16.1, 16.5.19,
17.1, 30.9.9
[Coxtr40] H.S.M. Coxeter, The binary polyhedral groups and other generalizations of the
quaternion group, Duke Math. J. 7 (1940), 367–379. 32.4
[Cro64] Michael J. Crowe, A history of vector analysis, Mineola, NY, Dover Press, 1964.
1.1, 1.2
[CR62] Charles W. Curtis and Irving Reiner, Representation theory of finite groups and
associative algebras, John Wiley & Sons, New York, 1962. 9.3
[CR81] Charles W. Curtis and Irving Reiner, Methods of representation theory, Volume
I: with applications to finite groups and orders, John Wiley & Sons, New York,
1981. 7.2, 7.2.20, 7.9, 9.2, 10.2, 17.3.7, 17.7.27, 18.3.9, 20.1.4, 20.2, 20.3.6,
20.4, 20.6, 22.2, 24.4, 24.5
[CR87] Charles W. Curtis and Irving Reiner, Methods of representation theory, Volume
II: with applications to finite groups and orders, John Wiley & Sons, New York,
1987. 12.4.4, 18.4.11, 20.1.4, 20.7, 20.7.13, 20.7
[DT2008] Samit Dasgupta and Jeremy Teitelbaum, The 𝑝-adic upper half plane, 𝑝-adic
geometry, Univ. Lecture Ser., vol. 45, Amer. Math. Soc., Providence, RI, 2008,
65–121. 23.5
[Dei2005] Anton Deitmar, A first course in harmonic analysis, 2nd. ed., Universitext,
BIBLIOGRAPHY 829
Springer-Verlag, New York, 2005. 29.3

[DE2009] Anton Deitmar and Siegfried Echterhoff, Principles of harmonic analysis, Uni-
versitext, Springer-Verlag, New York, 2009. 29.3, 29.3, 29.3.10, 29.4
[Del71] Pierre Deligne, Travaux de Shimura, Séminaire Bourbaki, 23ème année
(1970/71), Exp. No. 389, 123–165, Lecture Notes in Math., vol. 244, Springer,
Berlin, 1971. 38.7.20, 43.8.7
[Del74] Pierre Deligne, La conjecture de Weil. I, Publ. Math. l’IHÉS, 43 (1974), 273–307.
41.2.14
[DR80] Pierre Deligne and Kenneth A. Ribet, Values of abelian 𝐿-functions at negative
integers over totally real fields, Inventiones Math. 59 (1980), 227–286. 26.1.9
[Dem2007] Lassina Dembélé, Quaternionic Manin symbols, Brandt matrices and Hilbert
modular forms, Math. Comp. 76 (2007), no. 258, 1039–1057. 41.5.14
[DD2008] Lassina Dembélé and Steve Donnelly, Computing Hilbert modular forms over
fields with nontrivial class group, Algorithmic number theory (Banff, 2008),
Lecture Notes in Comput. Sci., vol. 5011, Springer, Berlin, 2008, 371–386.
41.5.14
[DV2013] Lassina Dembélé and John Voight, Explicit methods for Hilbert modular forms,
Elliptic curves, Hilbert modular forms and Galois deformations, Birkhauser,
Basel, 2013, 135–198. 41.5.14
[dR71] G. de Rham, Sur les polygons générateurs de groups Fuchsiens, Enseign. Math. 17
(1971), 49–61. 37.6.5
[DG89] Dennis M. Deturck and Carolyn S. Gordon, Isospectral deformations II: Trace
formulas, metrics, and potentials, with an appendix by Kyung Bai Lee, Comm.
Pure. Applied Math. 42 (1989), 1067–1095. 31.8
[Deu41] Max Deuring, Die Typen der Multiplikatorenringe elliptischer Funktionenkörper,
Abh. Math. Sem. Hansischen Univ. 14 (1941), 197–272. 42.1
[Deu51] Max Deuring, Die Anzahl der Typen von Maximalordnungen einer definiten
Quaternionenalgebra mit primer Grundzahl, Über. Deutsch. Mat. Verein. 54
(1951), 24–41. 30.9
[Deu68] Max Deuring, Algebren, Ergebnisse der Mathematik und ihrer Grenzgebiete, vol.
41, Springer-Verlag, Berlin-New York, 1968. (document), 29.10.24
[DS2005] Fred Diamond and Jerry Shurman, A first course in modular forms, Grad. Texts
in Math., vol. 228, Springer-Verlag, New York, 2005. 35, 40.1, 40.3, 40.5, 40.5
[Dic03] Leonard Eugene Dickson, Definitions of a linear associative algebra by indepen-
dent postulates, Trans. Amer. Math. Soc. 4 (1903), no. 1, 21–26. 1.2
[Dic12] Leonard Eugene Dickson, Linear algebras, Trans. Amer. Math. Soc. 13 (1912),
no. 1, 59–73. 1.2
[Dic14] Leonard Eugene Dickson, Linear associative algebras and abelian equations,
Trans. Amer. Math. Soc. 15 (1914), 31–46. 1.2, 1.3
[Dic19] L. E. Dickson, On quaternions and their generalization and the history of the
eight square theorem, Ann. Math. (2) 20 (1919), no. 3, 155–171. 1.1, 1.3
[Dic23] L. E. Dickson, Algebras and their arithmetics, Univ. of Chicago Press, Chicago,
1923. 1.2, 1.3
[Dic24] L. E. Dickson, On the theory of numbers and generalized quaternions,
Amer. J. Math. 46 (1924), no. 1, 1–16. 1.3
[Dic71] Leonard Eugene Dickson, History of the theory of numbers, Volume II: Diophan-
tine analysis, Chelsea, New York, 1971. 1.4
[DS2014] Joe Diestel and Angela Spalsbury, The joys of Haar measure, Grad. Studies in
Math., vol. 150, Amer. Math. Soc., 2014. 29.3
830 BIBLIOGRAPHY
[Die48] Jean Dieudonné, Sur les groupes classiques, Actualitiés scientifique et indus-
trielles, no. 1040, Paris, Hermann, 1948. 3.5.3
[Die53] Jean Dieudonné, On the structure of unitary groups (II), Amer. J. Math. 75 (1953),
665–678. 3.5.3
[DK95] Adel Diek and R. Kantowski, Some Clifford algebra history, Clifford Algebras
and Spinor Structures, Math. Appl. 321 (1995), 3–12. 1.2
[DM89] Radoslav Dimitrić and Brendan Goldsmith, The mathematical tourist: Sir William
Rowan Hamilton, Math. Intelligencer 11 (1989), no. 2, 29–30. 1.1
[Dix77] Jacques Dixmier, 𝐶 ∗ -algebras, Translated from the French by Francis Jellett,
North-Holland Mathematical Library, vol. 15, North-Holland, Amsterdam, 1977.
3.2.11
[Don2011] Simon Donaldson, Riemann surfaces, Oxford University Press, Oxford, 2011.
33.8.6
[Dra83] P.K. Draxl, Skew fields, Cambridge University Press, Cambridge, 1983. 7.5
[DK68] Yu. A. Drozd and V. V. Kirichenko, Hereditary orders, Math. Inst. Acad. Sci.
Ukranian SSR, translated from Ukrainskii Mat. Zhurnal, 20 (1968), no. 2, 246–
248. 21.4
[DKR67] Yu. A. Drozd, V. V. Kirichenko, and A. V. Roĭter, On hereditary and Bass orders,
Math. USSR-Izvestĳa, translated from Izv. Akad. Nauk SSSR, 1 (1967), no. 6,
1357–1375. 21.4, 24.2.24, 24.4, 24.5
[DK94] Yurĳ A. Drozd and Vladimir V. Kirichenko, Finite dimensional algebras, with
an appendix by Vlastimil Dlab, translated by Vlastimil Dlab, Springer-Verlag,
Berlin, 1994. 7.2, 7.9
[EGM2008] Bas Edixhoven, Gerard van der Geer, and Ben Moonen, Modular forms, Modular
forms on Schiermonnikoog, Cambridge University Press, Cambridge, 2008. 1.4
[Eic36] M. Eichler, Untersuchungen in der Zahlentheorie der rationalen Quaternione-
nalgebren, J. Reine Angew. Math. 174 (1936), 129–159. 1.4, 23.4.22, 24.1.5,
24.5
[Eic37] Martin Eichler, Bestimmung der Idealklassenzahl in gewissen normalen einfachen
Algebren, J. Reine Angew. Math. 176 (1937), 192–202. 28.6.11
[Eic38a] M. Eichler, Allgemeine Kongruenzklasseneinteilungen der Ideale einfacher Alge-
bren über algebraischen Zahlkörper und ihre 𝐿-Reihen, J. Reine Angew. Math.
179 (1938), 227–251. 26.8, 28.6.11
[Eic38b] M. Eichler, Über die Idealklassenzahl total definiter Quaternionenalgebren,
Math. Z. 43 (1938), no. 1, 102–109. 26.1.14, 30.6.18
[Eic38c] M. Eichler, Über die Idealklassenzahl hyperkomplexer Systeme, Math. Z. 43
(1938), no. 1, 481–494. 28.6.11
[Eic53] Martin Eichler, Quadratische Formen und orthogonale Gruppen, Grundlehren
Math. Wiss., vol. 63, Springer, New York, 1974. 1.4, 4.2, 4.2.19, 22.5.13
[Eic55-56] Martin Eichler, Lectures on modular correspondences, Tata Institute of Funda-
mental Research, Bombay, 1955–56. 14.5.2, 25.3
[Eic56a] Martin Eichler, Zur Zahlentheorie der Quaternionen-Algebren, J. Reine Angew.
Math. 195 (1956), 127–151; Berichtigung zu der Arbeit “Zur Zahlentheorie der
Quaternionen-Algebren”, J. Reine Angew. Math. 197 (1957), 220. 1.4, 23.4.22,
23.1, 25.3, 26.1.14, 30.3.17, 30.6.18, 30.8.8, 30.9.12, 41.5.13
[Eic56b] Martin Eichler, Über die Darstellbarkeit von Mdoulformen durch Thetareihen,
J. Reine Angew. Math. 195 (1956), 156–171. 1.4
[Eic56c] Martin Eichler, Modular correspondences and their representations, J. Indian
Math. Soc. (N.S.) 20 (1956), 163–206. 23.4.22
BIBLIOGRAPHY 831
[Eic58] Martin Eichler, Quadratische Formen und Modulfunktionen, Acta Arith. 4 (1958),
217–239. 1.4
[Eic73] Martin Eichler, The basis problem for modular forms and the traces of the Hecke
operators, Modular functions of one variable, I (Proc. Internat. Summer School,
Univ. Antwerp, Antwerp, 1972), Lecture Notes in Math., vol. 320, Springer,
Berlin, 1973, 75–151; Correction to: “The basis problem for modular forms and
the traces of the Hecke operators”, Modular functions of one variable, IV (Proc.
Internat. Summer School, Univ. Antwerp, Antwerp, 1972), Lecture Notes in
Math., vol. 476, Springer, Berlin, 1975, 145–147. 1.4, 23.4.22, 30.3.17, 30.6.18,
40.4, 41.5.13
[Eic77] Martin Eichler, On theta functions of real algebraic number fields, Acta Arith. 33
(1977), 269–292. 41.5.13
[EM45] Samuel Eilenberg and Saunders Mac Lane, General theory of natural equiva-
lences, Transactions Amer. Math. Soc. 58 (1945), 231–294. 19.3.11
[Elk98] Noam D. Elkies, Shimura curve computations, Algorithmic number theory (Port-
land, OR, 1998), Lecture Notes in Comput. Sci., vol. 1423, Springer, Berlin,
1998, 1–47. 43.2.19
[EKM2008] Richard S. Elman, Nikita Karpenko, and Alexander Merkurjev, The algebraic
and geometric theory of quadratic forms, vol. 56, Amer. Math. Soc., 2008. 6.3
[EGM98] Jürgen Elstrodt, Fritz Grunewald, and Jens Mennicke, Groups acting on hyper-
bolic space: harmonic analysis and number theory, Springer-Verlag, Berlin,
1998. 36.1, 40.2.24
[EP94] David B.A. Epstein and Carlo Petronio, An exposition of Poincaré’s polyhedron
theorem, Enseign. Math. 40 (1994), 113–170. 37.6, 37.6.5
[Ewe82] John A. Ewell, A simple derivation of Jacobi’s four-square formula, Proc. Amer.
Math. Soc. 85 (1982), no. 3, 323–326. 1.4
[Fad65] D.K. Faddeev, Introduction to multiplicative theory of modules of integral repre-
sentations, Algebraic number theory and representations, Proc. Steklov Institute
of Math., vol. 80 (1965), American Math. Soc., Providence, RI, 1968, 164–210.
15.6, 16.8, 20.3.6
[FD93] Benson Farb and R. Keith Dennis, Noncommutative algebra, Grad. Texts in Math.,
vol. 144, Springer-Verlag, New York, 1993. 7.2
[Fen98] Della Dumbaugh Fenster, Leonard Eugene Dickson and his work in the Arith-
metics of Algebras, Arch. Hist. Exact Sci. 52 (1998), no. 2, 119–159. 1.2
[FS2007] Della D. Fenster and Joachim Schwermer, Beyond class field theory: Helmut
Hasse’s arithmetic in the theory of algebras in 1931, Arch. Hist. Exact Sci. 61
(2007), 425–456. 1.3, 14.6
[Fin89] Benjamin Fine, The algebraic theory of the Bianchi groups, Marcel Dekker, New
York, 1989. 35.5, 36.4.4, 36.8
[Fit2011] Robert W. Fitzgerald, Norm Euclidean quaternionic orders, INTEGERS 11
(2011), no. A58. 17.10
[Fol95] Gerald Folland, A course in abstract harmonic analysis, Studies in Advanced
Mathematics, CRC Press, Boca Raton, FL, 1995. 29.4
[For72] L. R. Ford, Automorphic functions, 2nd. ed., Chelsea, New York, 1972. 33.1,
37.2.9
[FGS2016] Adam Forsyth, Jacob Gurev, and Shakthi Shrima, Metacommutation as a group
action on the projective line over F 𝑝 , Proc. Amer. Math. Soc. 144 (2016), no. 11,
4583–4590.
[FK1890-2] Robert Fricke and Felix Klein, Vorlesungen über die Theorie der elliptischen
832 BIBLIOGRAPHY
Modulfuctionen, vols. 1 and 2, B.G. Teubner, Leipzig, 1890, 1892. 1.4, 1.4.4
[FK1897] Robert Fricke and Felix Klein, Vorlesungen über die Theorie der automorphen
Functionen. Erster Band: Die gruppentheoretischen Grundlagen, B.G. Teubner,
Leipzig, 1897. 1.4
[FK12] Robert Fricke and Felix Klein, Vorlesungen über die Theorie der automorphen
Functionen. Zweiter Band: Die funktionentheoretischen Ausführungen und die
Anwendungen, B.G. Teubner, Leipzig, 1912. 1.4
[Fro1878] F.G. Frobenius, Über lineare Substitutionen und bilineare Formen, J. Reine
Angew. Math. 84 (1878), 1–63. 1.2
[Frö75] A. Fröhlich, Locally free modules over arithmetic orders, J. Reine Angew. Math.
274/275 (1975), 112–124. 20.7, 20.7.23
[Frö73] A. Fröhlich, The Picard group of noncommutative rings, in particular of orders,
Trans. Amer. Math. Soc. 180 (1973), 1–45. 18.4.11
[FRU74] A. Fröhlich, I. Reiner and S. Ullom, Class groups and Picard groups of orders,
Proc. London Math. Soc. (3) 29 (1974), 405–434. 20.7.23
[FT91] A. Fröhlich and M.J. Taylor, Algebraic number theory, Cambridge Stud. Adv.
Math., vol. 27, Cambridge University Press, Cambridge, 1991. 9.3
[Fuj58] Genjiro Fujisaki, On the zeta-function of the simple algebra over the field of
rational numbers, J. Fac. Sci. Univ. Tokyo. Sect. I 7 (1958), 567–604. 27.6,
29.10.24
[FH91] William Fulton and Joe Harris, Representation theory: a first course, Graduate
Texts in Math., vol. 129, Springer-Verlag, New York, 1991. 2.17
[GV2018] Steven D. Galbraith and Frederik Vercauteren, Computational problems in su-
persingular elliptic curve isogenies, Quantum Inf. Process. 17 (2018), no. 10,
Art. 265, 22 pp. 1.5
[Gar2004] Skip Garibaldi, The characteristic polynomial and determinant are not ad hoc
constructions, American Math. Monthly 2004 (111), no. 9, 761–778. 7.8
[GS2010] Skip Garibaldi and David J. Saltman. Quaternion algebras with the same subfields,
Quadratic forms, linear algebraic groups, and cohomology, Springer, New York,
2010, 225–238. 8.2.6, 14.17
[Gau00] Carl Friedrich Gauss, Mutation des Raumes, in Carl Friedrich Gauss Werke, Band
8, König. Gesell. Wissen., Göttingen, 1900, 357–361. 1.1
[Gau86] Carl Friedrich Gauss, Disquisitiones arithmeticae, Translated and with a preface
by Arthur A. Clarke, revised by William C. Waterhouse, Cornelius Greither, and
A. W. Grootendorst, and with a preface by Waterhouse. Springer-Verlag, New
York, 1986. 25.1, 25.2.19, 30.1
[GMMR97] F. W. Gehring, C. Maclachlan, G. J. Martin, and A. W. Reid, Arithmeticity,
discreteness and volume, Trans. Amer. Math. Soc. 349 (1997), 3611–3643. 32.7
[Gel84] Stephen Gelbart, An elementary introduction to the Langlands program,
Bull. Amer. Math. Soc. (N.S.) 10 (1984), no. 2, 177–219. 1.4, 41.5.13
[GM2004] Stephen S. Gelbart and Stephen D. Miller, Riemann’s zeta function and beyond,
Bull. Amer. Math. Soc. (N.S.) 41 (2004), no. 1, 59–112. 26.3.15
[GG90] I.M. Gel’fand, M.I. Graev, and I.I. Pyatetskii-Shapiro, Representation theory
and automorphic functions, trans. K.A. Hirsch, Generalized Functions, vol. 6,
Academic Press, Boston, 1990. 37.7
[Gib1891] Josiah Willard Gibbs, On the rôle of quaternions in the algebra of vectors, Nature
(1891), vol. 43, 511–513. 1.1
[Gir83] P.R. Girard, The quaternion group and modern physics, Eur. J. Phys. 5 (1984),
25–32. 1.5
BIBLIOGRAPHY 833
[GT2010] Olivier Giraud and Koen Thas, Hearing the shape of drums: Physical and math-
ematical aspects of isospectrality, Rev. Modern Phys. 82 (2010), 2213–2255.
31.8
[God58a] Roger Godement, Les fonctions 𝜁 des algèbres simples, I, Séminaire N. Bourbaki,
1958–1960, exp. no. 171, 27–49. 29.10.24
[God58b] Roger Godement, Les fonctions 𝜁 des algèbres simples, II, Séminaire N. Bourbaki,
1958–1960, exp. no. 176, 109–128. 29.10.24
[God62] Roger Godement, Domaines fondamentaux des groupes arithmétiques, 1964
Séminaire Bourbaki, 1962/63, Fasc. 3, No. 257, Secrétariat mathématique, Paris.
38.5.6
[GJ72] Roger Godement and Hervé Jacquet, Zeta functions of simple algebras, Lecture
Notes in Math., vol. 260, Springer-Verlag, Berlin, 1972. 29.10.24
[Gol85] Dorian Goldfeld, Gauss’ class number problem for imaginary quadratic fields,
Bull. Amer. Math. Soc. 13 (1985), 23–37. 25.2.19
[Gor2000] Carolyn S. Gordon, Survey of isospectral manifolds, Handbook of differential
geometry, Vol. I, North-Holland, Amsterdam, 2000, 747–778. 31.8
[Gor2009] Carolyn S. Gordon, Sunada’s isospectrality technique: two decades later, Spectral
analysis in geometry and number theory, Contemp. Math., vol. 484, Amer. Math.
Soc., Providence, RI, 2009, 45–58. 31.8
[Gol71] Larry Joel Goldstein, Analytic number theory, Prentice-Hall, Inc., Englewood
Cliffs, N.J., 1971. 27.4
[Gor2012] Carolyn S. Gordon, Orbifolds and their spectra, Spectral geometry, Proc. Sympos.
Pure Math., vol. 84, Amer. Math. Soc., Providence, RI, 2012, 49–71. 34.8
[GWW92] Carolyn Gordon, David L. Webb, and Scott Wolpert, One cannot hear the shape
of a drum, Bull. Amer. Math. Soc. (N.S.) 27 (1992), no. 1, 134–138. 31.8
[Gor52] D. Gorenstein, An arithmetic theory of adjoint plane curves, Trans. Amer. Math.
Soc. 72 (1952), 414–436. 24.2.24
[Gou97] Fernando Q. Gouvêa, 𝑝-adic numbers: an introduction, Universitext, 2nd. ed.,
Springer-Verlag, Berlin, 1997. 12.1
[Gras1862] Hermann Grassmann, Die Ausdehnungslehre. Vollstaändig und in strenger Form
begründet [Extension theory], Berlin, Wiegand, 1862. 1.1
[Grav1882] Robert Perceval Graves, Life of Sir William Rowan Hamilton, Volume I, Dublin
University Press, 1882. 1.1
[Grav1885] Robert Perceval Graves, Life of Sir William Rowan Hamilton, Volume II, Dublin
University Press, 1885. 1.1, 1.1
[Grav1889] Robert Perceval Graves, Life of Sir William Rowan Hamilton, Volume III, Dublin
University Press, 1889. 1.1, 1.1
[Gray94] Jeremy Gray, On the history of the Riemann mapping theorem, Rendiconti del
Circolo Matematico di Palermo, Serie II, Supplemento 34 (1994), 47–94. 34.8.3
[Gray2000] Jeremy Gray, Linear differential equations and group theory from Riemann to
Poincaré, 2nd ed., Birkhäuser, Boston, 2000. 1.4
[Gray2013] Jeremy Gray, Henri Poincaré: a scientific biography, Princeton University Press,
Princeton, 2013. 1.4
[Gri28] Lois W. Griffiths, Generalized quaternion algebras and the theory of numbers,
Amer. J. Math. 50 (1928), no. 2, 303–314. 1.3
[Gro87] Benedict H. Gross, Heights and the special values of 𝐿-series, Number theory
(Montreal, Que., 1985), CMS Conf. Proc., vol. 7, Amer. Math. Soc., Providence,
RI, 1987, 115–187. 42.3.13
[Gro99] Benedict Gross, Algebraic modular forms, Israel J. Math. 113 (1999), 61–93.
834 BIBLIOGRAPHY
41.2.15
[GrLu2009] Benedict H. Gross and Mark W. Lucianovic, On cubic rings and quaternion rings,
J. Number Theory 129 (2009), no. 6, 1468–1478. 3.15, 22.4, 22.4.3, 22.5.13,
24.2
[Gro85] Emil Grosswald, Representations of integers as sums of squares, Springer-Verlag,
New York, 1985. 30.2
[GB2008] L. C. Grove and C. T. Benson, Finite reflection groups, 2nd ed., Grad. Texts in
Math, vol. 99, Springer-Verlag, New York, 1985. 11.5
[GrLe87] P.M. Gruber and C.G. Lekkerkerker, Geometry of numbers, 2nd ed., North-
Holland, New York, 1987. 17.5
[GQ2004] Xuejun Guo and Hourong Qin, An embedding theorem for Eichler orders, J.
Number Theory 107 (2004), no. 2, 207–214. 31.7, 31.7.7
[Grov2002] Larry C. Grove, Classical groups and geometric algebra, Grad. Studies in Math.,
vol. 39, Amer. Math. Soc., 2002. 4.2, 6.3, 28.8
[Hae84] A. Haefliger, Groupoides d’holonomie et classifiants, Structure transverse des
feuilletages, Toulouse 1982, Astérisque 116 (1984), 70–97. 34.8.15
[HM2006] Emmanuel Hallouin and Christian Maire, Cancellation in totally definite quater-
nion algebras, J. Reine Angew. Math. 595 (2006), 189–213. 20.7
[Ham1843] William R. Hamilton, On a new species of imaginary quantities connected with
a theory of quaternions, Proc. Royal Irish Acad. 2 (1843), 424–434. 1.1
[Ham1844] William R. Hamilton, On quaternions, or on a new system of imaginaries in
algebra: copy of a letter from Sir William R. Hamilton to John T. Graves, Esq.
on quaternions, Philosophical Magazine 25 (1844), 489–495. 1.2
[Ham1853] W. R. Hamilton, Lectures on quaternions, Cambridge, Cambridge University
Press, 1853. 1.1
[Ham1866] W. R. Hamilton, Elements of quaternions, Cambridge, Cambridge University
Press, 1866. 1.1.5, 1.1
[Ham1899] Hamilton’s quaternions, review of Elements of quaternions, Nature, August 24,
1899, 387. 1.1
[Ham67] W. R. Hamilton, The mathematical papers of Sir William Rowan Hamilton, Vol.
III: Algebra, eds. H. Halberstam and R. E. Ingram, Cambridge University Press,
1967. 1.1, 1.1
[Hanlon81] Phil Hanlon, Applications of quaternions to the study of imaginary quadratic
class groups, Ph.D. thesis, California Institute of Technology, 1981. 30.2
[Hankin80] Thomas L. Hankins, William Rowan Hamilton, Johns Hopkins University Press,
1980. 1.1, 1.1
[Hans2006] Andrew J. Hanson, Visualizing quaternions, Morgan Kaufmann, San Francisco,
2006. 1.5, 2.4.22
[Hap80] Dieter Happel, Klassifikationstheorie endlich-dimensionaler Algebren in der Zeit
von 1880 bis 1920, Enseign. Math. (2) 26 (1980), no. 1–2, 91–102. 1.2
[Har63a] Manabu Harada, Hereditary orders, Trans. Amer. Math Soc. 107 (1963), no. 2,
273–290. 20.3, 21.4
[Har63b] Manabu Harada, Structure of hereditary orders over local rings, J. Math. Osaka
City Univ. 14 (1963), 1–22. 21.4
[Har63c] Manabu Harada, Hereditary orders in generalized quaternions 𝐷 𝜏 , J. Math. Os-
aka City Univ. 14 (1963), 71–81. 21.4
[Har1881] A. S. Hardy, Elements of quaternions, Ginn, Heath, and Company, Boston, 1881.
v (document)
[HFK94] John C. Hart, George K. Francis, Louis H. Kauffman, Visualizing quaternion
BIBLIOGRAPHY 835
rotation, ACM Trans. Graphics 13 (1994), no. 3, 256–276. 1.5

[Har77] Robin Hartshorne, Algebraic geometry, Grad. Texts in Mathematics, vol. 52,
Springer-Verlag, New York, 1977. 4.13, 9.15
[Hash77] Ki-ichiro Hashimoto, Twisted trace formula of the Brandt matrix, Proc. Japan
Acad., Ser. A 53 (1977), 98–102. 30.6.18
[HM95] Ki-ichiro Hashimoto and Naoki Murabayashi, Shimura curves as intersections of
Humbert surfaces and defining equations of QM-curves of genus two, Tôhoku
Math. J. 47 (1995), 271–296. 43.2.21
[Hass34] Helmut Hasse, Über gewisse Ideale in einer einfachen Algebra, Act. Sci. Ind.
Paris 109 (1934), 12–16. 30.3.17
[Hass52] Helmut Hasse, Über die Klassenzahl abelscher Zahlkörper, Akademie-Verlag,
Berlin, 1952. 39.4.11
[Hass67] Helmut Hasse, History of Class Field Theory, in Algebraic number theory,
J. W. S. Cassels and A. Fröhlich (eds.), Academic Press, New York, 1967, 266–
279. 1.3
[Haw2000] Thomas Hawkins, The emergence of the theory of Lie groups: an essay in the
history of mathematics, 1869–1926, Springer, New York, 2000. 1.2
[Hec40] Erich Hecke, Analytische Arithmetik der positiven quadratischen Formen, Danske
Vid. Selsk. Math.-Fys. Medd. 17, (1940). no. 12. 1.4, 41.5.13
[Heeg52] Kurt Heegner, Diophantische Analysis und Modulfunktionen, Math. Z. 56 (1952),
227–253. 25.2.19
[Hel2001] Sigurdur Helgason, Differential geometry, Lie groups, and symmetric spaces,
Grad. Studies in Math., vol. 34, Amer. Math. Soc., Providence, RI, 2001. 34.6.7
[Herm1854] Charles Hermite, Sur la théorie des formes quadratiques: Premier mémoire, J.
Reine Angew. Math. 47 (1854), 307–342. 22.5.13
[Hey29] Käte Hey, Analytische Zahlentheorie in Systemen hyperkomplexer Zahlen, Ph.D.
dissertation, Universität Hamburg, 1929. 25.1, 26.8, 29.10.24, 38.4.5
[Hĳ74] Hiroaki Hĳikata, Explicit formula of the traces of Hecke operators for Γ0 (𝑁), J.
Math. Soc. Japan 26 (1974), no. 1, 56–82. 23.4, 23.4.22, 30.6, 30.6.18, 41.5.13
[HPS89a] Hiroaki Hĳikata, Arnold K. Pizer, and Thomas R. Shemanske, The basis problem
for modular forms on Γ0 (𝑁), Mem. Amer. Math. Soc. 82 (1989), no. 418. 1.4,
24.5.7, 30.6.18, 41.5.13
[HPS89b] H. Hĳikata, A. Pizer, and T. Shemanske, Orders in quaternion algebras, J. Reine
Angew. Math. 394 (1989), 59–106. 24.5.7, 30.6.18
[HN94] H. Hĳikata and K. Nishida, Bass orders in non semisimple algebras, J. Math. Ky-
oto Univ. 34 (1994), no. 4, 797–837. 21.4, 21.5
[HS73] H. Hĳikata and H. Saito, On the representability of modular forms by theta series,
Number Theory, Algebraic Geometry, and Commutative Algebra in honor of Y.
Akizuki, Kinokuniya, Tokyo, 1973, 13–21. 41.5.13
[Hil32] David Hilbert, Gesammelte Abhandlungen, Bd. 1: Zahlentheorie, Springer,
Berlin, 1932. 14.6.4
[Hir53] Friedrich Hirzebruch, Über vierdimensionale Riemannsche Flächen mehrdeutiger
analytischer Funktionen von zwei komplexen Veränderlichen, Math. Ann. 126
(1953), 1–22. 35.4
[Hir91] Friedrich Hirzebruch, Division algebras and topology, Numbers, Grad. Texts in
Math., vol. 123, Springer-Verlag, New York, 1991, 281–302. 1.2
[HK2004] H.-J. Hoehnke and M.-A. Knus, Algebras, their Invariants, and 𝐾-forms: a
tribute to [the work of] HEINRICH BRANDT on the 50th anniversary of his
death, preprint, 2004, available at
836 BIBLIOGRAPHY
https://people.math.ethz.ch/~knus/papers/brandt04.pdf. 1.4
[Hun99] Craig Huneke, Hyman Bass and ubiquity: Gorenstein rings, Algebra, 𝐾-theory,
groups, and education (New York, 1997), Contemp. Math., vol. 243, Amer. Math.
Soc., Providence, RI, 1999, 55–78. 24.2.24
[Hur1898] Adolf Hurwitz, Über die Komposition der quadratischen Formen von beliebig
vielen Variablen, Nach. der köbig. Gesell. Wissen. Göttingen, Math.-Physik.
Klasse, 1898, 309–316. 1.3
[Hur1896] Adolf Hurwitz, Über die Zahlentheorie der Quaternionen, Nachrichten der
Gesellschaft der Wissenschaften zu Göttingen, 1896, 314–340. 1.3
[Hur19] Adolf Hurwitz, Vorlesungen über die Zahlentheorie der Quaternionen, Springer-
Verlag, Berlin, 1919. 1.3, 11.1
[Ibu82] Tomoyoshi Ibukiyama, On maximal orders of division quaternion algebras over
the rational number field with certain optimal embeddings, Nagoya Math. J. 88
(1982), 181–195. 15.5.7
[Igu60] Jun-Ichi Igusa, Arithmetic variety of moduli for genus two, Ann. Math. 72 (1960),
no. 3, 612–649. 43.3.5, 43.3, 43.3.8, 43.5.7
[Ive92] Birger Iversen, Hyperbolic geometry, London Math. Soc. Student Texts, vol. 25,
Cambridge University Press, Cambridge, 1992. 33.1, 36.1
[Iwa92] K. Iwasawa, Letter to J. Dieudonné, dated April 8, 1952, Zeta functions in
geometry, N. Kurokawa and T. Sunada, eds., Adv. Stud. in Pure Math. 21 (1992),
445–450. 29.10.24
[Jaci68] H. Jacobinski, Genera and decompositions of lattices over orders, Acta Math.
121 (1968), 1–29. 20.7.23
[Jaci71] H. Jacobinski, Two remarks about hereditary orders, Proc. Amer. Math. Soc 28
(1971), no. 1, 1–8. 21.2, 21.4
[Jacn74] Nathan Jacobson, Abraham Adrian Albert (1905–1972), Bull. Amer. Math.
Soc. 80 (1974) 1075–1100. 1.3
[Jacn2009] Nathan Jacobson, Finite-dimensional division algebras over fields, Springer-
Verlag, Berlin, 2009. 7.2, 8.2
[JL70] Hervé Jacquet and Robert Langlands, Automorphic forms on GL(2), Lecture
Notes in Math., vol. 114, Springer, New York, 1970. 1.4, 41.5.13
[JKS97] William C. Jagy, Irving Kaplansky, and Alexander Schiemann, There are 913
regular ternary forms, Mathematika 44 (1997), no. 2, 332–341. 25.4
[Jah2010] Majid Jahangiri, Generators of arithmetic quaternion groups and a Diophantine
problem, Int. J. Number Theory 6 (2010), no. 6, 1311–1328. 32.1
[Jan96] Gerald J. Janusz, Algebraic number fields, 2nd ed., Grad. Studies in Math., vol. 7,
Amer. Math. Soc., Providence, RI, 1996. 14.4, 14.4.4, 14.4.12, 27.5
[JS87] Gareth A. Jones and David Singerman, Complex functions: an algebraic and
geometric viewpoint, Cambridge University Press, 1987. 33.1, 34.7
[Jun97] Sungtae Jun, On the certain primitive orders, J. Korean Math. Soc. 34 (1997),
no. 4, 791–807. 24.5.7
[Jun08] Heinrich W.E. Jung, Darstellung der Funktionen eines algebraischen Körpers
zweier unabhängigen Veränderlichen 𝑥, 𝑦 in der Umgebung einer Stelle 𝑥 = 𝑎, 𝑦 =
𝑏, J. reine angew. Math. 133 (1908), 289–314. 35.4
[Kac66] Mark Kac, Can one hear the shape of a drum?, Amer. Math. Monthly 73 (1966),
no. 4, part II, 1–23. 31.8
[Kap69] Irving Kaplansky, Submodules of quaternion algebras, Proc. London Math. Soc.
(3) 19 (1969), 219–232 16.6, 16.6, 16.6, 16.8, 16.18, 19.5.8
[Kar2010] Max Karoubi, 𝐾-theory, an elementary introduction, Cohomology of groups and
BIBLIOGRAPHY 837
algebraic 𝐾-theory, Adv. Lect. Math. (ALM), vol. 12, Int. Press, Somerville, MA,
2010, 197–215. 12.4.4
[Kat85] Svetlana Katok, Closed geodesics, periods and arithmetic of modular forms, Inv.
Math. 80 (1985), 469–480. 43.2.8
[Kat92] Svetlana Katok, Fuchsian groups, University of Chicago Press, Chicago, 1992.
33.1, 33.6, 37.3, 37.7
[Kat2007] Svetlana Katok, 𝑝-adic analysis compared with real, Student Math. Library,
vol. 37, Amer. Math. Soc., Providence, RI, 2007. 12.1
[Ker58] M. Kervaire, Non-parallelizability of the n-sphere for 𝑛 > 7, Proc. Nat. Acad. Sci.
44 (1958) 280–283. 1.2
[Kil1888] Wilhelm Killing, Die Zusammensetzung der stetigen endlichen Transformations-
gruppen, Math. Ann. 31 (1888), 252–290; Math. Ann. 33 (1888), 1–48; Math.
Ann. 34 (1889), 57–122; Math. Ann. 36 (1890), no. 2, 161–189. 1.2
[Kir2005] Markus Kirschmer, Konstruktive Idealtheorie in Quaternionenalgebren, Ph.D.
thesis, Universität Ulm, 2005. 17.7.25
[Kir2014] Markus Kirschmer, One-class genera of maximal integral quadratic forms,
J. Number Theory 136 (2014), 375–393 25.4.7
[KL2016] Markus Kirschmer and David Lorch, Ternary quadratic forms over number fields
with small class number, J. Number Theory 161 (2016), 343–361. 26.7, 26.7.4
[KLwww] Markus Kirschmer and David Lorch, The one-class genera of ternary quadratic
forms, website
http://www.math.rwth-aachen.de/~kirschme/orders/. 26.7.4
[Kle1878] Felix Klein, Über die Transformation der elliptischen Funktionen und die Au-
flösung der Gleichungen fünften Grades, Math. Ann. 14 (1878–1879), 111–172.
1.4
[Kle56] Felix Klein, Lectures on the icosahedron and the solution of equations of the
fifth degree, translated by George Gavin Morrice, 2nd. revised edition, Dover
Publications, New York, 1956. 32.4
[Kle79] Felix Klein, Development of mathematics in the 19th century, translated by M.
Ackerman, Lie Groups: History, frontiers and applications, vol. IX, Math Sci
Press, Brookline, Mass., 1979. 34.3.10
[Klt2000] Ernst Kleinert, Units in skew fields, Progress in Math., vol. 186, Birkhäuser
Verlag, Basel, 2000. 28.8, 38.4
[Kna2016] Anthony W. Knapp, Advanced algebra, digital 2nd. edition, East Setauket, New
York, 2016. 27.4
[Kne66a] Martin Kneser, Starke approximation in algebraischen gruppen. I, J. Reine Angew.
Math. 218 (1965), 190–203. 28.7.10, 28.8
[Kne66b] Martin Kneser, Strong approximation, Algebraic Groups and Discontinuous Sub-
groups (Boulder, Colo., 1965), Proc. Sympos. Pure Math., Amer. Math. Soc.,
Providence, RI, 1966, 187–196. 28.7.10, 28.8
[KKOPS86] M. Kneser, M.-A. Knus, M. Ojanguren, R. Parimala, and R. Sridharan, Composi-
tion of quaternary quadratic forms, Compositio Math. 60 (1986), no. 2, 133–150.
19.5.8
[Knu88] Max-Albert Knus, Quadratic forms, Clifford algebras and spinors, Seminários
de Matemática, 1, Universidade Estadual de Campinas, Instituto de Matemática,
Estatística e Ciência da Computaç ã o, Campinas, 1988. 9.7
[Knu93] Max-Albert Knus, Sur la forme d’Albert et le produit tensoriel de deux algebres
de quaternions, Bull. Soc. Math. Belg. 45 (1993), 333–337. 8.2.6
[KMRT98] Max-Albert Knus, Alexander Merkurjev, Markus Rost, and Jean-Pierre Tignol,
838 BIBLIOGRAPHY
The book of involutions, Amer. Math. Soc. Colloquium Publications, vol. 44,
Amer. Math. Soc., Providence, RI, 1998. 4.4.6, 5.6, 5.6.11
[Kob84] Neal Koblitz, 𝑝-adic numbers, 𝑝-adic analysis, and zeta-functions, 2nd edition,
Grad. Texts in Math., vol. 58, Springer-Verlag, New York, 1984. 12.1
[Koh96] David Kohel, Endomorphism rings of elliptic curves over finite fields, Ph.D. thesis,
University of California, Berkeley, 1996. 42.3.6
[Koh2001] David Kohel, Hecke module structure of quaternions, Class field theory: its
centenary and prospect (Tokyo, 1998), ed. K. Miyake, Adv. Stud. Pure Math.,
vol. 30, Math. Soc. Japan, Tokyo, 2001, 177–195. 41.5.14
[KV2003] David R. Kohel and Helena A. Verrill, Fundamental domains for Shimura curves,
J. Théorie des Nombres de Bordeaux 15 (2003), 205–222. 37.9.13
[Kör85] Otto Körner, Über die zentrale Picard-Gruppe und die Einheiten lokaler Quater-
nionenordnungen, Manuscripta Math. 52 (1985), 203–225. 26.6
[Kör87] Otto Körner, Traces of Eichler–Brandt matrices and type numbers of quaternion
orders, Proc. Indian Acad. Sci. (Math. Sci.) 97 (1987), no. 1–3, 189–199. 25.3,
26.1.14, 30.8.8, 30.9.12
[Kur79] Akira Kurihara, On some examples of equations defining Shimura curves and the
Mumford uniformization, J. Fac. Sci. Univ. Tokyo Sect. IA Math. 25 (1979), no.
3, 277–300. 43.2
[Lam99] Tsit-Yuen Lam, Lectures on modules and rings, Grad. Texts in Math., vol. 189,
Springer-Verlag, New York, 1999. 7.2.20, 20.2, 20.2, 20.3.6
[Lam2001] Tsit-Yuen Lam, A first course in noncommutative rings, 2nd. ed., Grad. Texts in
Math., vol. 131, Springer-Verlag, New York, 2001. 7.2
[Lam2002] Tsit-Yuen Lam, On the linkage of quaternion algebras, Bull. Belg. Math. Soc. 9
(2002), 415–418. 8.2.6
[Lam2003] Tsit-Yuen Lam, Hamilton’s quaternions, Handbook of algebra, vol. 3, North
Holland, Amsterdam, 2003, 429–454. 1
[Lam2005] Tsit-Yuen Lam, Introduction to quadratic forms over fields, Grad. Studies in
Math., vol. 67, Amer. Math. Soc., Providence, RI, 2005. 3.5.3, 4.2, 4.2, 4.5, 5.2,
5.4, 8.2.9, 26.8
[Lam2006] Tsit-Yuen Lam, Serre’s problem on projective modules, Springer Monographs in
Mathematics, Springer-Verlag, Berlin, 2006. 20.7.24
[Lamo86] Klaus Lamotke, Regular solids and isolated singularities, Adv. Lectures in Math.,
Friedr. Vieweg and Sohn, Braunschweig, 1986. 32.4
[Lanc67] C. Lanczos, William Rowan Hamilton—An appreciation, Amer. Scientist 55
(1967), 129–143. 1.1
[Lang94] Serge Lang, Algebraic number theory, 2nd ed., Grad. Texts in Math., vol. 110,
Springer-Verlag, Berlin, 1994. 26.2, 26.2, 26.8, 27.4, 27.5
[Lang95] Serge Lang, Introduction to modular forms, appendixes by D. Zagier and Walter
Feit, Grundlehren der Mathematischen Wissenschaften, vol. 222, Springer-Verlag,
Berlin, 1995. 40.1
[Lang82] Serge Lang, Introduction to algebraic and abelian functions, 2nd ed., Grad. Texts
in Math., vol. 89, Springer-Verlag, New York, 1982. 43.6
[Lat26] Claiborne G. Latimer, Arithmetics of generalized quaternion algebras, Amer. J.
Math. (2) 27 (1926), 92–102. 1.3
[Lat35] Claiborne G. Latimer, On the fundamental number of a rational generalized
quaternion algebra, Duke Math. J. 1 (1935), no. 4, 433–435. 14.2.8
[Lat37] Claiborne G. Latimer, The classes of integral sets in a quaternion algebra, Duke
Math. J. 3 (1937), 237–247. 22.5.13
BIBLIOGRAPHY 839
[Lee2011] John M. Lee, Introduction to topological manifolds, 2nd ed., Grad. Texts in Math.,
vol 202, Springer-Verlag, New York, 2011. 34.5
[Lem2011] Stefan Lemurell, Quaternion orders and ternary quadratic forms, 2011,
arXiv:1103.4922. 16.12, 22.5.13, 24.1.5, 24.6
[Len96] H.W. Lenstra, jr., Complex multiplication structure of elliptic curves, J. Number
Theory 56 (1996), 227–241. 42.1
[LS91] H.W. Lenstra and Peter Stevenhagen, Primes of degree one and algebraic cases
of Cebotarev’s theorem, Enseign. Math. 37 (1991), 17–30. 14.2.10
[Lev2013] Alex Levin, On the classification of algebras, M.Sc. thesis, University of Vermont,
2013. 3.5.11
[LR2011] Lawrence S. Levy and J. Chris Robson, Hereditary Noetherian prime rings and
idealizers, Math. Surveys Monogr., vol. 174, American Math. Soc., Providence,
RI, 2011. 21.4.11
[Lew2006a] David W. Lewis, Quaternion algebras and the algebraic legacy of Hamilton’s
quaternions, Irish Math. Soc. Bull. 57 (2006), 41–64. 1
[Lew2006b] David W. Lewis, Involutions and anti-automorphisms of algebras, Bull. London
Math. Soc. 38 (2006), 529–545. 1.2
[LL2012] Wen-Ching Winnie Li and Ling Long, Atkin and Swinnerton-Dyer congruences
and noncongruence modular forms, Algebraic number theory and related topics
2012, RIMS Kôkyûroku Bessatsu, B51, Res. Inst. Math. Sci. (RIMS), Kyoto,
2014, 269–299, arXiv:1303.6228. 35.4.4
[Lin2012] Benjamin Linowitz, Selectivity in quaternion algebras, J. Number Theory 132
(2012), no. 7, 1425–1437. 31.7.7
[LS2012] Benjamin Linowitz and Thomas R. Shemanske, Embedding orders into central
simple algebras, J. Théor. Nombres Bordeaux 24 (2012), no. 2, 405–424. 31.7.7
[LV2015] Benjamin Linowitz and John Voight, Small isospectral and nonisometric orbifolds
of dimension 2 and 3, Math. Z. 281 (2015), no. 1–2, 523–569. 31.7.7, 31.8, 31.4
[Liv2001] Ron Livné, Communication networks and Hilbert modular forms, Applications
of algebraic geometry to coding theory, physics and computation (Eilat, 2001),
NATO Sci. Ser. II Math. Phys. Chem., vol. 36, Kluwer Acad. Publ., Dordrecht,
2001, 255–270. 41.2.14
[LP2010] Nikolai Ivanovich Lobachevsky, Pangeometry, ed. Athanase Papadopoulos, Eu-
ropean Mathematical Society, 2010. 33.1
[Loo53] Lynn H. Loomis, An introduction to abstract harmonic analysis, D. Van Nostrand
Company, Inc., Princeton, 1953. 29.3
[LK2013] David Lorch and Markus Kirschmer, Single-class genera of positive integral
lattices, LMS J. Comput. Math. 16 (2013), 172–186. 25.4, 25.4.7
[Lub2010] Alexander Lubotzky, Discrete groups, expanding graphs and invariant measures,
appendix by Jonathan D. Rogawski, Birkhäuser Verlag, Basel, reprint edition,
2010. 41.2.14
[LPS88] A. Lubotzky, R. Phillips, and P. Sarnak, Ramanujan graphs, Combinatorica 8
(1988), 261–277. 41.2.14
[Luc2003] Mark Lucianovic, Quaternion rings, ternary quadratic forms, and Fourier coef-
ficients of modular forms on PGSp(6), Ph.D. thesis, Harvard University, 2003.
22.5.13
[Macf1891] Alexander Macfarlane, Principles of the algebra of physics, Proc. Amer. Assoc.
Adv. Sci. 15 (1891), 65–112. 1.2
[Macf00] Alexander Macfarlane, Hyperbolic quaternions, Proc. Royal Soc. Edinburgh 23
(1900) 169–180. 1.2
840 BIBLIOGRAPHY
[Macl2008] C. Maclachlan, Optimal embeddings in quaternion algebras, J. Number Theory

128 (2008) 2852–2860. 31.1.8, 31.6
[MR2003] Colin Maclachlan and Alan W. Reid, The arithmetic of hyperbolic 3-manifolds,
Grad. Texts in Math., vol. 219, Springer-Verlag, New York, 2003. iv (document)
[MacRos94] C. Maclachlan and G. Rosenberger, Small volume isospectral, nonisometric,
hyperbolic 2-orbifolds, Arch. Math. (Basel) 62 (1994), no. 1, 33–37. 31.8
[Mard2007] Albert Marden, Outer circles: an introduction to hyperbolic 3-manifolds, Cam-
bridge University Press, Cambridge, 2007. 36.1
[Marg88] G. Margulis, Explicit group theoretic constructions of combinatorial schemes and
their application to the design of expanders and concentrators, J. Prob. of Info.
Trans. (1988), 39–46. 41.2.14
[Mas88] Bernard Maskit, Kleinian groups, Springer-Verlag, 1988. 37.6.5
[Mas71] B. Maskit, On Poincaré’s theorem for fundamental polygons, Advances in Math.
7 (1971), 219–230. 37.6, 37.6.5
[Mat46] Margaret Matchett, On the zeta function for ideles, Ph.D. thesis, Indiana Univer-
sity, 1946. 29.10.24
[Mat89] H. Matsumura, Commutative ring theory, translated by M. Reid, 2nd ed., Cam-
bridge Stud. Adv. Math., vol. 8, Cambridge University Press, Cambridge, 1989.
9.2, 9.5, 22.2
[Mat69] Hideya Matsumoto, Sur les sous-groupes arithmétiques des groupes semi-simples
déployés, Ann. Sci. École Norm. Sup. (4) 2 (1969), 1–62. 12.4.4
[Max1869] J. C. Maxwell, Remarks on the mathematical classification of physical quantities,
Proc. London Math. Soc. 3 (1869), 224–232. 1.1
[May66] Kenneth O. May, The impossibility of a division algebra of vectors in three
dimensional space, Amer. Math. Monthly 73 (1966), no. 3, 289–291. 1.1
[McC2011] George McCarty, Topology: an introduction with application to topological
groups, Dover, New York, 2010. 34.1, 34.3.11
[McCR87] J. C. McConnell and J. C. Robson, Noncommutative Noetherian rings, Grad. Stud.
in Math., vol. 30, American Math. Soc., Providence, RI, 1987. 18.3.9
[Mer82] Alexander Merkurjev, On the norm residue symbol of degree 2, Soviet
Math. Dokl. 24 (1982), 1546–1551. 8.3.6
[Mes91] Jean-Francois Mestre, Construction de courbes de genre 2 à partir de leurs mod-
ules, Effective methods in algebraic geometry, Progr. Math. 94 (1991), Birkhäuser,
Boston, 313–334. 43.3.13
[Milne80] James S. Milne, Étale cohomology, Princeton Mathematical Series, vol. 33,
Princeton University Press, Princeton, 1980. 15.17
[Milne2017] James S. Milne, Algebraic groups: The theory of group schemes of finite type
over a field, Cambridge Studies in Advanced Mathematics, vol. 170, Cambridge
University Press, Cambridge, 2017. 38.5
[Milne-CFT] J.S. Milne, Class field theory, available at
http://www.jmilne.org/math/CourseNotes/cft.html. 14.6.10, 26.8
[Milne-SV] J.S. Milne, Introduction to Shimura varieties, available at
https://www.jmilne.org/math/xnotes/svi.html. 38.7.20
[Mil64] J. Milnor, Eigenvalues of the Laplace operator on certain manifolds, Proc. Nat.
Acad. Sci. U.S.A. 51 (1964), 542.
[Milno71] John Milnor, Introduction to algebraic 𝐾-theory, Annals of Math. Studies, no. 72,
Princeton University Press and University of Tokyo Press, Princeton, 1971. 12.4.4
[Milno82] John Milnor, Hyperbolic geometry: the first 150 years, Bull. Amer. Math. Soc. 6
(1982), no. 1, 9–24. 33.1, 36.5, 36.5
BIBLIOGRAPHY 841
[Mir95] Rick Miranda, Algebraic curves and Riemann surfaces, Grad. Studies in Math.,
vol. 5, Amer. Math. Soc., Providence, RI, 1995. 33.8.6
[Miy2006] Toshitsune Miyake, Modular forms, Translated from the 1976 Japanese original
by Yoshitaka Maeda, Springer Monographs in Math., Springer-Verlag, Berlin,
2006. 28.7, 40.1, 40.4, 40.5
[Moe2002] Ieke Moerdĳk, Orbifolds as groupoids: an introduction, Orbifolds in mathematics
and physics (Madison, WI, 2001), Contemp. Math., vol. 310, Amer. Math. Soc.,
Providence, RI, 2002, 205–222. 34.8.15
[MP1997] I. Moerdĳk and D. A. Pronk, Orbifolds, sheaves and groupoids, 𝐾-Theory 12
(1997), no. 1, 3–21. 34.8.15
[Mol1893] Theodor Molien, Über Systeme höherer complexer Zahlen, Math. Ann. 41 (1893),
83–156; Berichtigung zu dem Aufsatze ‘Über Systeme höherer complexer Zahlen’,
Math. Ann. 42 (1893), 308–312. 1.2
[MW77] Hugh L. Montgomery and Peter J. Weinberger, Real quadratic fields with large
class number, Math. Ann. 225 (1977), no. 2, 173–176. 26.7.1
[Moore35] Eliakim Hastings Moore, General analysis, Part I, Memoirs Amer. Phil. Soc., vol.
1, Amer. Phil. Soc., Philadelphia, 1935. 3.5.3
[Mor69] L. J. Mordell, Diophantine equations, Pure Applied Math., vol. 30, Academic
Press, London, 1969. 11.4.4
[MT62] G. D. Mostow and T. Tamagawa, On the compactness of arithmetically defined
homogeneous spaces, Ann. of Math. (2) 76 (1962), 446–463. 38.5.6
[Mum70] David Mumford, Abelian varieties, Tata Institute of Fundamental Research Stud-
ies in Math., no. 5, Oxford University Press, Oxford, 1970. 8.5, 43.4
[Mun2004] W. D. Munn, Involutions on finite-dimensional algebras over real closed fields,
J. Austral. Math. Soc. 77 (2004) 123–128. 8.4.17
[Mur88] M. Ram Murty, Primes in certain arithmetic progressions, J. Madras Univ.,
Section B 51 (1988), 161–169. 14.2.10
[Neu99] Jürgen Neukirch, Algebraic number theory, Grundlehren Math. Wiss., vol. 322,
Springer-Verlag, Berlin, 1999. 12.2, 12.2, 13.2, 14.2.10, 14.4, 14.4.4, 14.4.12,
15.1, 26.2, 26.2, 26.8, 27.4, 27.5, 28.7, 28.7
[New72] Morris Newman, Integral matrices, Pure Appl. Math., vol. 45, Academic Press,
New York, 1972. 17.2
[NSW2008] Jürgen Neukirch, Alexander Schmidt, and Kay Wingberg, Cohomology of number
fields, 2nd. ed., Grundlehren Math. Wiss., vol. 323, Springer, Berlin, 2008.
14.6.10
[Nip74] Gordon L. Nipp, Quaternion orders associate with ternary lattices, Pacific J.
Math. 53 (1974), no. 2, 525–537. 22.5.13
[Noe34] Emmy Noether, Zerfallende verschränkte Produkte und ihre Maximalordnungen,
Act. Sci. Ind. Paris 148 (1934), 15 pp. 30.3.17
[ÓCa2000] Fiacre Ó Cairbre, William Rowan Hamilton (1805–1865), Ireland’s greatest math-
ematician, Ríocht na Midhe (Meath Archaeological and Historical Society) 11
(2000), 124–150. 1.1
[ÓCa2010] Fiacre Ó Cairbre, Twenty years of the Hamilton walk, Irish Math. Soc. Bulletin
65 (201), 33–49. 1.1
[O’Do83] Sean O’Donnell, William Rowan Hamilton, Boole Press Limited, 1983. 1.1
[Ogg69] Andrew Ogg, Modular forms and Dirichlet series, W.A. Benjamin, Inc., New
York, 1969. 40.4
[Ogg83] Andrew Ogg, Real points on Shimura curves, Arithmetic and geometry, Vol. I,
Progr. Math., vol. 35, Birkhäuser Boston, Boston, 1983, 277–307. 43.7
842 BIBLIOGRAPHY
[O’Me73] O. T. O’Meara, Introduction to quadratic forms, Springer-Verlag, Berlin, 1973.

4.2, 4.2, 4.5, 9.3, 9.3, 9.7, 14.6, 28.7, 28.7
[PS2014] Ariel Pacetti and Nicolás Sirolli, Computing ideal classes representatives in
quaternion algebras, Math. Comp. 83 (2014), no. 289, 2479–2507. 24.6, 26.6.1
[Pall46] Gordon Pall, On generalized quaternions, Trans. Amer. Math. Soc. 59 (1946),
280–332. 22.5.13
[Pap2014] Athanese Papadopoulos, Metric spaces, convexity and non-positive curvature,
2nd edition, IRMA Lectures in Mathematics and Theoretical Physics, vol. 6,
European Mathematical Society, Zürich, 2014. 33.2.8
[PS2018] Ori Parzanchevski and Peter Sarnak, Super-golden-gates for 𝑃𝑈 (2), Adv. Math.
327 (2018), 869–901. 1.5
[Pei1882] Benjamin Peirce, Linear associative algebra, Amer. J. Math. 4 (1881), no. 1–4,
97–229. 1.2
[Pet69] M. Peters, Ternäre und quaternäre quadratische Formen und Quaternionenalge-
bren, Acta Arith. 15 (1969), 329–365. 22.5.13, 30.9.12
[Pie82] Richard S. Pierce, Associative algebras, Springer-Verlag, New York, 1982. 7.2,
7.9
[Piz73] Arnold K. Pizer, Type numbers of Eichler orders, J. Reine Angew. Math. 264
(1973), 76–102. 23.4.22, 30.9.12
[Piz76a] Arnold Pizer, On the arithmetic of quaternion algebras, Acta Arith. 31 (1976),
no. 1, 61–89. iv (document), 30.6.18, 30.8.8, 30.9.12
[Piz76b] Arnold Pizer, On the arithmetic of quaternion algebras. II, J. Math. Soc. Japan
28 (1976), no. 4, 676–688. 24.3.7, 30.6.18, 41.5.13
[Piz76c] Arnold Pizer, The representability of modular forms by theta series, J. Math. Soc.
Japan 28 (1976), 689–698. 30.6.18, 41.5.13
[Piz80a] Arnold Pizer, An algorithm for computing modular forms on Γ0 (𝑁), J. Algebra
64 (1980), 340–390. 15.5.7, 24.3.7, 41.5.14
[Piz80b] Arnold Pizer, Theta series and modular forms of level 𝑝 2 𝑀, Compositio Math.
40 (1980), 177–241. 30.6.18
[Pla69] V. P. Platonov, Strong approximation in algebraic groups and the Kneser-Tits
conjecture, Dokl. Akad. Nauk BSSR 13 (1969), no. 7, 585–587. 28.7.10
[Pla69-70] V. P. Platonov, The strong approximation problem and the Kneser-Tits conjec-
ture for algebraic groups, Izv. Akad. Nauk SSSR, Ser. Math. 33 (1969), no. 6,
1211–1219; Additions to ‘The strong approximation problem and the Kneser-Tits
conjecture for algebraic groups’, Izv. Akad. Nauk SSSR, Ser. Math. 33 (1970),
no. 4, 775–777. 28.7.10, 28.8
[PR94] Vladimir Platonov and Andrei Rapinchuk, Algebraic groups and number theory,
Pure and Applied Mathematics, vol. 139, Academic Press, Inc., Boston, MA,
1994. 28.7.10, 38.5.6
[Pl83] Wilhelm Plesken, Group rings of finite groups over 𝑝-adic integers, Lect. Notes
in Math., no. 1026, Springer, 1983. 23.4.23
[Poi1882] Henri Poincaré, Théorie des groups fuchsiens, Acta Math. 1 (1882), 1–62. 37.6.5
[Poi1883] Henri Poincaré, Mémoire sur les groupes Kleinéens, Acta Math. 3 (1882), 49–92.
37.6.5
[Poi1908] Henri Poincaré, Science et méthode, Paris, Flammarion, 1908. 1.4
[Pol2010] Paul Pollack, Hypothesis 𝐻 and an impossibility theorem of Ram Murty,
Rend. Semin. Mat. Univ. Politec. Torino 68 (2010), no. 2, 183–197. 14.2.10
[Pra77] Gopal Prasad, Strong approximation for semi-simple groups over function fields,
Annals of Math. (2) 105 (1977), no. 3, 553–572. 28.7.10, 28.8
BIBLIOGRAPHY 843
[Pre68] Alexander Prestel, Die elliptischen Fixpunkte der Hilbertschen Modulgruppen,

Math. Ann. 177 (1968), 181–209. 30.6.18
[Puj2012] Jose Pujol, Hamilton, Rodrigues, Gauss, quaternions, and rotations: a historical
reassessment, Comm. Math. Anal. 13 (2012), no. 2, 1–14. 1.1
[Puj2014] Jose Pujol, On Hamilton’s nearly-forgotten early work on the relation between
rotations and quaternions and on the composition of rotations, Amer. Math.
Monthly 121 (2014), June–July, 515–522. 1.1
[Pyn2006] Thomas Pynchon, Against the Day, Penguin Press, 2006. 1.1
[Rag72] M. S. Raghunathan, Discrete subgroups of Lie groups, Ergebnisse der Mathematik
und ihrer Grenzgebiete, Band 68, Springer-Verlag, New York-Heidelberg, 1972.
38.4.7
[Raj93] A. R. Rajwade, Squares, Lecture Note Series, vol. 171, Cambridge University
Press, Cambridge, 1993. 3.5.10
[RM99] Dinakar Ramakrishnan and Robert J. Valenza, Fourier analysis on number fields,
Grad. Texts in Math., vol. 186, Springer-Verlag, New York, 1999. 27.4, 29.3
[Ran2011] Andrew Ranicki, Commentary on “On the parallelizability of the spheres” by R.
Bott and J. Milnor and “On the nonexistence of elements of Hopf invariant one”
by J. F. Adams, Bull. Amer. Math. Soc. (N.S.) 48 (2011), no. 4, 509–511. 1.2
[Rap2014] Andrei S. Rapinchuk, Strong approximation for algebraic groups, Thin groups
and superstrong approximation, Math. Sci. Res. Inst. Publ., vol. 61, Cambridge
Univ. Press, Cambridge, 2014, 269–298. 28.7.10
[Rat2006] John G. Ratcliffe, Foundations of hyperbolic manifolds, 2nd. ed., Grad. Texts in
Math., vol. 149, Springer-Verlag, New York, 2006. 34.8, 36.1, 36.3.17, 36.5,
36.5, 37.6
[Reh76] Hans Peter Rehm, On a theorem of Gauss concerning the number of integral
solutions of the equation 𝑥 2 + 𝑦 2 + 𝑧 2 = 𝑚, Selected topics on ternary forms and
norms (Sem. Number Theory, California Inst. Tech., Pasadena, Calif., 1974/75),
Paper No. 4, California Inst. Tech., Pasadena, Calif., 1976, 11 pp. 30.2
[Rei70] Irving Reiner, A survey of integral representation theory, Bull. Amer. Math. Soc.
76 (1970), no. 2, 159–227. 20.1.4
[Rei76] Irving Reiner, Class groups and Picard groups of group rings and orders, Confer-
ence Board of the Math. Sci., Regional Conf. Ser. Math., no. 26, Amer. Math. Soc.,
Providence, RI, 1976. 20.1.4
[Rei2003] Irving Reiner, Maximal orders, London Math. Soc. Monogr. (N.S.), vol. 28,
Clarendon Press, Oxford University Press, Oxford, 2003. iv (document), 7.2.20,
7.5.5, 7.8.1, 7.8.12, 7.9, 9.2, 10.2, 13.4.3, 15.2, 16.3, 17.7.27, 17.8, 17.12, 18.2,
18.3, 18.3, 18.3, 20.4, 20.5, 20.5, 20.6, 20.7, 21.2.8, 21.4, 21.5.3, 21.6, 21.8,
23.3, 28.6.11
[Rib1989] Kenneth A. Ribet, Bimodules and abelian surfaces, Algebraic number theory: in
honor of K. Iwasawa, Adv. Pure Math. 17 (1989), 359–407. 42.4
[Roq2006] Peter J. Roquette, The Brauer–Hasse–Noether theorem in historical perspective,
Schr. Math.-Nat.wiss. Kl., Springer, Berlin–Heidelberg, 2006. 1.3, 29.10.24
[Rod1840] Olinde Rodrigues, Des lois geometriques qui regissent les déplacements d’un
système solide dans l’espace, et la variation des coordonnées provenant de ses
déplacements considérés indépendamment des causes qui peuvent les produire,
J. de Mathematiques Pures et Appliqués 5 (1840), 380–440. 1.1
[Ros2002] Michael Rosen, Number theory in function fields, Grad. Texts in Math., vol. 210,
Springer-Verlag, New York, 2002. 29.13
[Rot2004] Victor Rotger, The field of moduli of quaternionic multiplication on abelian
844 BIBLIOGRAPHY
varieties, Int. J. Math. Math. Sci. 2004, no. 49–52, 2795–2808. 43.6
[Rud87] Walter Rudin, Real and complex analysis, 3rd. ed., McGraw-Hill, New York,
1987. 29.3.2
[Sah72] Chih-Han Sah, Symmetric bilinear forms and quadratic forms, J. Algebra 20
(1972), 144–160. 8.2.6
[Sal99] David J. Saltman, Lectures on division algebras, CBMS regional conf. series in
math., vol. 94, Amer. Math. Soc., Providence, RI, 1999. 7.5, 7.5.6
[San2017] Jonathan Sands, Zeta-functions and ideal classes of quaternion orders, Rocky
Mountain J. Math. 47 (2017), no. 4, 1277–1300. 26.5.22
[Sar90] Peter Sarnak, Some applications of modular forms, Cambridge Tracts in Math.,
vol. 99, Cambridge University Press, 1990. 41.2.14
[Scha85] Winfried Scharlau, Quadratic and Hermitian forms, Springer-Verlag, Berlin,
1985. 4.2, 4.2, 4.5, 9.7, 9.8
[Scha2009] Rudolf Scharlau, Martin Kneser’s work on quadratic forms and algebraic groups,
Quadratic forms: algebra, arithmetic, and geometry, Contemporary Math.,
vol. 493, Amer. Math. Soc., Providence, RI, 2009, 339–357. 1.3, 28.7.10, 29.11.9
[Schi35] Otto Schilling, Über gewisse Beziehungen zwischen der Arithmetik hyperkom-
plexer Zahlsysteme und algebraischer Zahlkörper, Math. Ann. 111 (1935), no. 1,
372–398. 30.3.17
[Schn75] Volker Schneider, Die elliptischen Fixpunkte zu Modulgruppen in Quaternionen-
schiefkörpern, Math. Ann. 217 (1975), no. 1, 29–45. 30.6.18
[Schn77] Volker Schneider, Elliptische Fixpunkte und Drehfaktoren zur Modulgruppe in
Quaternionenschiefkörpern über reellquadratischen Zahlkörpern, Math. Z. 152
(1977), no. 2, 145–163. 30.6.18
[Sco83] Peter Scott, The geometries of 3-manifolds, Bull. London Math. Soc. 15 (1983),
401–487. 33.1, 34.8
[Ser73] Jean-Pierre Serre, A course in arithmetic, Grad. Texts in Math., vol. 7, Springer-
Verlag, New York, 1973. 14.1, 14.2, 14.3, 25.2.19, 35, 40.1, 40.3
[Ser79] Jean-Pierre Serre, Local fields, Grad. Texts in Math., vol. 67, Springer-Verlag,
New York, 1979. 8.3.7, 12.2, 12.2, 13.2, 13.4.7, 38.5
[Ser96] J-P. Serre, Two letters on quaternions and modular forms (mod p), Israel J. Math.
95 (1996), 281–299. 42.3.14
[Ser2003] Jean-Pierre Serre, Trees, Translated by John Stillwell, corrected 2nd printing,
Springer Monographs in Math., Springer-Verlag, Berlin, 2003. 23.5
[Sha90] Daniel B. Shapiro, Compositions of quadratic forms, de Gruyter Expositions in
Math., vol. 33, Walter de Gruyter, Berlin, 2000. 1.3
[Shem86] Thomas R. Shemanske, Representations of ternary quadratic forms and the class
number of imaginary quadratic fields, Pacific J. Math. 122 (1986), no. 1, 223–250.
30.1
[Shz63] Hideo Shimizu, On discontinuous groups operating on the product of the upper
half planes, Ann. of Math. (2) 77 (1963), no. 1, 33–71. 30.6.18
[Shz65] Hideo Shimizu, On zeta functions of quaternion algebras, Ann. of Math. (2) 81
(1965), 166–193. 1.4, 30.6.18, 39.1
[Shz72] Hideo Shimizu, Theta series and automorphic forms on GL2 , J. Math. Soc. Japan
24 (1972), 638–683. 41.5.13
[Shi67] Goro Shimura, Construction of class fields and zeta functions of algebraic curves,
Ann. of Math. (2) 85 (1967), 58–159. 38.7.20, 43.1.9, 43.8.1, 43.8.2
[Shi71] Goro Shimura, Introduction to the arithmetic theory of automorphic functions,
Kanô Memorial Lectures, no. 1, Publications of the Mathematical Society of
BIBLIOGRAPHY 845
Japan, no. 11, Iwanami Shoten, Tokyo, Princeton University Press, Princeton,
1971. 28.2, 43.9, 43.9
[Shi75] Goro Shimura, On the real points of an arithmetic quotient of a bounded symmetric
domain, Math. Ann. 215 (1975), 135–164. 43.7
[Shi80] Goro Shimura, On some problems of algebraicity, Proceedings of the International
Congress of Mathematicians (Helsinki, 1978), Acad. Sci. Fennica, Helsinki, 1980,
373–379. 1.4
[Shi96] Goro Shimura, in 1996 Steele Prizes, Notices Amer. Math. Soc. 43 (1996), no.
11, 1340–1347.
[Sho85] Ken Shoemake, Animating rotation with quaternion curves, ACM SIGGRAPH
Computer Graphics 19 (1985), 245–254. 1.5
[Sie45] Carl Ludwig Siegel, Some remarks on discontinuous groups, Ann. of Math. (2)
46 (1945), 708–718. 37.7
[Sie65] Carl Ludwig Siegel, Vorlesungen über ausgewählte Kapitel der Functionentheo-
rie, vol. 2, Lecture notes, Universität Göttingen, Germany, 1965. 37.6.5
[Sie69] Carl Ludwig Siegel, Berechnung von Zetafunktionen an ganzzahligen Stellen,
Nachr. Akad. Wiss. Göttingen Math.-Phys. Kl. II 1969 (1969), 87–102. 26.1.9
[Sie89] Carl Ludwig Siegel, Lectures on the geometry of numbers, Springer, Berlin, 1989.
17.5
[Sil2009] Joseph H. Silverman, The arithmetic of elliptic curves, 2nd ed., Grad. Texts in
Math., vol. 106, Springer-Verlag, New York, 2009. 29.13, 40.1, 40.3.10, 42.1,
42.1, 42.1, 42.1, 42.1, 42.2.4, 42.2.4, 42.2, 42.4.3, 43.3.1
[Sim2010] Peter Simons, Vectors and beyond: geometric algebra and its philosophical
significance, dialectica 63 (2010), no. 4, 381–395. 1.1
[Sme2015] Daniel Smertnig, A note on cancellation in totally definite quaternion algebras,
J. Reine Angew. Math. 707 (2015), 209–216. 20.7, 20.7.24
[SV2019] Daniel Smertnig and John Voight, Definite orders with locally free cancellation,
Trans. Lond. Math. Soc. 6 (2019), no. 1, 53–86. 20.7, 20.7.24
[SW2005] Jude Socrates and David Whitehouse, Unramified Hilbert modular forms, with
examples relating to elliptic curves, Pacific J. Math. 219 (2005), no. 2, 333–364.
41.5.14
[Sol77] Louis Solomon, Zeta functions and integral representation theory, Adv. in Math.
26 (1977), 306–326. 26.3.14
[Sta67] H. M. Stark, A complete determination of the complex quadratic fields of class-
number one, Michigan Math. J. 14 (1967), 1–27. 25.2.19
[Sta2007] H. M. Stark, The Gauss class-number problems, Analytic number theory: a tribute
to Gauss and Dirichlet, Clay Math. Proc., no. 7, Amer. Math. Soc., Providence,
RI, 2007, 247–256. 25.2.19
[Sun85] Toshikazu Sunada, Riemannian coverings and isospectral manifolds, Ann. of
Math. (2) 121 (1985), no. 1, 169–186. 31.8
[Swa60] Richard G. Swan, Induced representations and projective modules, Ann. of Math.
(2) 71 (1960), 552–578. 20.7.13
[Swa62] Richard G. Swan, Projective modules over group rings and maximal orders, Ann.
of Math. (2) 76 (1962), 55–61. 20.7.13
[Swa80] Richard G. Swan, Strong approximation and locally free modules. Ring theory
and algebra, III, Lecture Notes in Pure and Appl. Math., vol. 55, Dekker, New
York, 1980, 153–223. 20.7, 20.7.23, 28.8
[Swi74] H.P.F. Swinnerton-Dyer, Analytic theory of abelian varieties, London Math. Soc.
Lecture Note Ser., no. 14, Cambridge University Press, London-New York, 1974.
846 BIBLIOGRAPHY
43.4
[Syl1883] J. J. Sylvester, Lectures on the principles of universal algebra, American J. Math.
6 (1883–1884), no. 1, 270–286. 1.2
[Tai1890] P. G. Tait, An elementary treatise on quaternions, 3rd ed., Cambridge University
Press, Cambridge, 1890. 1.1
[Tam66] Tsuneo Tamagawa, Adèles, Algebraic groups and discontinuous subgroups, Proc.
Sympos. Pure Math., vol. 9, Amer. Math. Soc., Providence, RI, 113–121. 29.11.9
[Tate2010] J. T. Tate, Global class field theory, Algebraic number theory, 2nd. ed., J.W.S.
Cassels and A. Frölich, eds., London Mathematical Society, London, 2010, 163–
203. 27.5, 27.5, 27.5.13
[Tate67] J. T. Tate, Fourier analysis in number fields and Hecke’s zeta functions (Thesis,
1950), Algebraic number theory, 2nd. ed., J.W.S. Cassels and A. Frölich, eds.,
London Mathematical Society, London, 2010, 305–347. 29.2, 29.10.24
[Tho10] Silvanus P. Thompson, The Life of Lord Kelvin, Baron Kelvin of Largs, Vol. II,
Macmillan, London, 1910. 1.1
[Tit79] J. Tits, Reductive groups over local fields, Automorphic forms, representations
and 𝐿-functions (Proc. Sympos. Pure Math., Oregon State Univ., Corvallis, Ore.,
1977), Part 1, Proc. Sympos. Pure Math., vol. 33, Amer. Math. Soc., Providence,
RI, 1979, 29–69. 23.5.17
[Thu97] William P. Thurston, Three-dimensional geometry and topology, vol. 1, ed. Silvio
Levy, Princeton University Press, Princeton, 1997. 33.8.3, 34.8, 36.5, 36.5
[Tig98] Jean-Pierre Tignol, Algebras with involution and classical groups. European
Congress of Mathematics, Vol. II (Budapest, 1996), Progr. Math., vol. 169,
Birkhäuser, Basel, 1998, 244–258. 4.4.6
[vdB60] F. van der Blĳ, History of the octaves, Simon Stevin 34 (1960/1961), 106–125.
1.2
[vdW76] B. L. van der Waerden, Hamilton’s discovery of quaternions, Math. Magazine 49
(1976), 227–234. 1.1
[vdW85] B. L. van der Waerden, A history of algebra, Springer-Verlag, Berlin, 1985. 1.2
[vPr2002] Paul van Praag, Quaternions as reflexive skew fields, Adv. Appl. Clifford Algebr.
12 (2002), no. 2, 235–249. 3.5.3
[Ven22] B.A. Venkov, On the arithmetic of quaternion algebras, Izv. Akad. Nauk (1922),
205–220, 221–246. 1.3, 30.1, 30.2
[Ven29] B.A. Venkov, On the arithmetic of quaternion algebras, Izv. Akad. Nauk (1929),
489–509, 532–562, 607–622. 1.3, 30.1, 30.2
[VG74] Marie-France Vignéras-Guého, Simplification pour les ordres des corps de
quaternions totalement définis, C. R. Acad. Sci. Paris Sér. A 279 (1974), 537–540.
32.8
[Vig76a] Marie-France Vignéras, Invariants numériques des groupes de Hilbert, Math.
Ann. 224 (1976), no. 3, 189–215. 30.6.18
[Vig76b] Marie-France Vignéras, Simplification pour les ordres des corps de quaternions
totalement définis, J. Reine Angew. Math. 286/287 (1976), 257–277. 20.7
[Vig80a] Marie-France Vignéras, Arithmétique des algèbres de quaternions, Lecture Notes
in Math., vol. 800, Springer, Berlin, 1980. iii (document), 21.4.10, 28.7, 28.9.23,
29.3, 30.8.8, 30.9.12, 31.1.8, 31.3, 31.8
[Vig80b] Marie-France Vignéras, Variétés riemanniennes isospectrales et non isomét-
riques, Ann. of Math. (2) 112 (1980), no. 1, 21–32. 31.8, 31.8
[Voi2006] John Voight, Computing CM points on Shimura curves arising from cocompact
arithmetic triangle groups, Algorithmic number theory (ANTS VII, Berlin, 2006),
BIBLIOGRAPHY 847
eds. Florian Hess, Sebastian Pauli, Michael Pohst, Lecture Notes in Comp. Sci.,
vol. 4076, Springer, Berlin, 2006, 406–420. 43.8.2
[Voi2011a] John Voight, Characterizing quaternion rings over an arbitrary base, J. Reine
Angew. Math. 657 (2011), 113–134. 22.2.14, 22.5.13, 22.5.14
[Voi2011b] John Voight, Rings of low rank with a standard involution, Illinois J. Math. 55
(2011), no. 3, 1135–1154. 3.5.5, 3.15
[Voi2013] John Voight, Identifying the matrix ring: algorithms for quaternion algebras and
quadratic forms, Quadratic and higher degree forms, Developments in Math., vol.
31, Springer, New York, 2013, 255–298. 9.8
[VW2014] John Voight and John Willis, Computing power series expansions of modular
forms, Computations with modular forms, eds. Gebhard Boeckle and Gabor
Wiese, Contrib. Math. Comput. Sci., vol. 6, Springer, Berlin, 2014, 331–361.
43.9.7
[VZB2015] John Voight and David Zureick–Brown, The canonical ring of a stacky curve,
2015, arXiv:1501.04657. 43.9.6, 43.9
[Vos2009] S.V. Vostokov, The classical reciprocity law for power residues as an analog of
the abelian integral theorem, St. Petersburg Math. J. 20 (2009), no. 6, 929–936.
14.6.4
[Wad86] A. Wadsworth, Merkurjev’s elementary proof of Merkurjev’s theorem, Ap-
plications of algebraic 𝐾-theory to algebraic geometry and number theory,
Part II, (Boulder, Colorado, 1983), Contemporary Mathematics, vol. 55,
Amer. Math. Soc., Providence, RI, 1986, 741–776. 8.3.6
[War89] Seth Warner, Topological fields, North-Holland Math. Studies, vol. 157, North-
Holland Publishing Co., Amsterdam, 1989. 13.5.13
[Was97] Lawrence C. Washington, Introduction to cyclotomic fields, 2nd ed., Grad. Texts
in Math., vol. 83, Springer, New York, 1997. 39.4.11
[Wate69] William C. Waterhouse, Abelian varieties over finite fields, Ann. Scient. École
Norm. Sup., 4th series, 2 (1969), 521–560. 42.2, 42.2
[Wate79] William C. Waterhouse, Introduction to affine group schemes, Grad. Texts in
Math., vol. 66, Springer-Verlag, New York-Berlin, 1979.
[Wats62] G. L. Watson, Transformations of a quadratic form which do not increase the
class-number, Proc. Lond. Math. Soc. (3) 12 (1962), 577–587. 25.4.7
[Wats75] G. L. Watson, One-class genera of positive ternary quadratic forms, II, Mathe-
matika 22 (1975), 1–11. 25.4
[Wed08] J.H. Maclagan Wedderburn, On hypercomplex numbers, Proc. London Math. Soc.
2 (1908), vol. 6, 77–118. 1.2, 7.3.11
[WW2018] Anne van Weerden and Steven Wepster, A most gossiped about genius: Sir
William Rowan Hamilton, BSHM Bull. (2018) 33, no. 1, 2–20. 1.1
[Weib2013] Charles A. Weibel, The 𝐾-book: An introduction to algebraic 𝐾-theory, Grad.
Studies in Math., vol. 145, Amer. Math. Soc., Providence, RI, 2013. 12.4.4
[Weil60] André Weil, Algebras with involution and the classical groups, J. Indian Math.
Soc. 24 (1960) 589–623. 4.4.6, 8.4.17
[Weil66] André Weil, Fonction zêta et distributions, Séminaire Bourbaki (1964-1966), vol.
9, 523–531. 29.10.24
[Weil74] André Weil, Basic number theory, 3rd ed., Springer-Verlag, Berlin, 1974. 12.2,
29.3, 29.10.24
[Weil82] André Weil, Adeles and algebraic groups, Progress in Math., Birkhäuser, Boston,
1982. 27.1.11, 27.6, 29.5, 29.8.8, 29.10, 29.10.4, 29.10.24, 29.11.9
[Wein96] A. Weinstein, Groupoids: unifying internal and external symmetry, Notices
848 BIBLIOGRAPHY
Amer. Math. Soc. 45 (1996), 848–859. 19.3.11

[Wes] Tom Weston, Lectures on the Dirichlet class number formula for imaginary
quadratic fields, http://people.math.umass.edu/~weston/oldpapers/
cnf.pdf. 25.2.19
[Whe62] Lynde Phelps Wheeler, Josiah Willard Gibbs: The history of a great mind, Yale
University Press, New Haven, 1962. 1.1
[Will2001] Dana P. Williams, A primer for the Brauer group of a groupoid, Groupoids in
analysis, geometry, and physics (Boulder, CO, 1999), Contemp. Math., vol. 282,
Amer. Math. Soc., Providence, RI, 2001, 21–34. 19.3.11
[Wils2009] Robert A. Wilson, The finite simple groups, Grad. Texts in Math., Springer-Verlag,
London, 2009. 11.10c
[Zag94] Don Zagier, Values of zeta functions and their applications, First European
Congress of Mathematics, Vol. II (Paris, 1992), Progr. Math., vol. 120, Birkhäuser,
Basel, 1994, 497–512. 25.2
[Zas72] Hans Zassenhaus, On the units of orders, J. Algebra 20 (1972), 368–395. 38.4
Index
S-Eichler condition, 473 Albert–Brauer–Hasse–Noether theorem,

𝜎-compact, 496 11, 224
S-indefinite, 473 algebra, 21, 152
S-integers, 221 absolute discriminant, 506, 510
𝑝-adic absolute value, 175 absolutely semisimple, 114
𝑝-adic integers, 176 algebraic, 42
𝑝-adic valuation, 176 automorphism, 21
center, 21
abelian variety central, 21
complex, 793 degree, 36, 40
polarization, 129 dimension, 21
Rosati involution, 129 endomorphism, 21
simple, 129 homomorphism, 21
absolute discriminant, 506 inseparable, 83
absolute norm, 251, 256 quadratic, 39
absolute reduced norm, 281, 282 reduced, 54
absolute value, 179 semisimple, 99
𝑝-adic, 175 separable, 83, 113
archimedean, 179 simple, 97
equivalent, 179 algebraic 𝐾-theory, 188
nonarchimedean, 179 analytic class number formula, 422
trivial, 180 anisotropic, 51
absolutely semisimple, 114 annihilator
action map, 622 module, 102, 148
adele ring anti-automorphism, 36
finite, 451, 457 antilinear functional, 797
global field, 451 archimedean
quaternion, 450 absolute value, 179
rationals, 446 place, 219
simple algebra, 456 archimedean property, 180
adjoint representation, 28 arithmetic group, 709
adjugate, 37 congruence closure, 713
adjugation, 27 Fuchsian, 703
admissible, 620 Artin isomorphism, 455
Albert form, 124 Artin–Schreier group, 185
849
850 INDEX
associated bilinear form, 49 relative, 132

associated bilinear map, 144 Brauer–Severi variety, 126
Atkin–Lehner group, 485 Bruhat–Tits tree, 365, 378
atlas
manifold, 609 canonical divisor, 527
orbifold, 629 canonical model, 786
smooth, 609 Cayley–Hamilton theorem
atomic, 147 reduced, 39
augmentation ideal, 116 center, 21
automorphism central, 21
algebra, 21 centralizer, 108
axis, 29, 654 chain lemma, 60
Azumaya chart, 609
order, 247 orbifold, 629
circle at infinity, 596, 604
basic class group
order, 381, 395 narrow, 284
basis problem, 762 narrow S, 454
Bass strict, 284
order, 381, 394 strict S, 454
Bernoulli numbers, 662, 668, 738 class number
Bianchi group, 657 analytic formula, 422
big period matrix, 793 order, 270, 279
bilinear form class set
nondegenerate, 53 narrow, 477
orthogonal, 50 order, 273
symmetric, 49 quadratic module, 146
bimodule, 320 classical modular group, 635
binary dihedral group, 170 Clausen’s integral, 658
binary icosahedral group, 169 Clifford algebra, 67
binary octahedral group, 168 reversal involution, 69
binary quadratic form, 301 CM point, 786, 808
binary tetrahedral group, 162 cocycle relation, 735
biquaternion algebra, 11, 121, 122 codifferent, 244
biquaternions, 8 coefficient ideals
Borel measure, 496 pseudobasis, 138
Bourbaki proper, 625 commensurable, 703
box product, 238 compatible, 252
Brandt class groupoid, 303, 309 Haar measure, 498
Brandt graph, 749, 753 lattice, 250, 252
Brandt groupoid, 303, 308 complete
Brandt invariant, 382, 386 flag, 337
Brandt matrix, 747, 752 completed
Brandt module, 749 Dedekind zeta function, 436
Brauer equivalent, 121, 125 completed upper half-plane, 596
Brauer group, 121, 125 completed upper half-space, 650
INDEX 851
completely decomposable, 138, 156 cycle condition, 686, 687

completely reducible, 99 cyclic algebra, 105
completion, 142 cyclic algebras, 116
complex 𝑛-manifold, 610
complex 𝑛-orbifold, 630 decays rapidly, 492, 502
complex multiplication (CM), 801 decomposable, 147, 799
complex place, 219 completely, 138
complex structure, 809 Dedekind domain, 137
complex torus, 793 Dedekind scheme, 150
morphism, 793 Dedekind zeta function, 405, 421
composable, 252 completed, 436
composition factor, 325 functional equation, 436
composition law, 301 definite
composition series, 325 quaternion algebra, 169, 210
composition factor, 325 degree, 36, 40
length, 325 descends, 231
conductor dicyclic group, 170
quadratic order, 250, 267 diffeomorphism, 610
quaternion order, 386 different, 265
cone, 423 digits, 175
cone point, 630 dilogarithm function, 668
congruence, 713 dimension, 21
congruence closure, 713 Dirichlet
congruence subgroup, 644 class number formula, 403
level, 644 Dirichlet domain, 671, 682
conic, 74 Dirichlet series, 405
isomorphism, 74 disconnected
conjugate transpose, 126 totally, 176
conjugation, 27 discrete, 180
connected, 270, 274, 306 discrete valuation, 199
connected components, 306 discriminant, 234, 237, 249
connecting ideal, 270, 274 fundamental, 250
continued fraction quadratic form, 47, 52
Hirzebruch–Jung, 646 quaternion algebra, 210, 222
negative-regular, 646 quaternion order, 234
continuous reduced, 239
group action, 616 discriminant form, 59
convex, 603 distance, 380
set, 276 division algebra, 22
counting norm, 256 division ring, 22
covering space action, 618 divisor group, 527
covolume, 276 double centralizer property, 108
cross product, 28 dual
curve, 767 isogeny, 767
cusp form, 736 lattice, 242, 797
cusps, 811 dual basis lemma, 319
852 INDEX
duplication formula, 659 finite idele group, 457

flag, 336
Eichler condition, 473 complete, 337
Eichler mass formula, 271 football, 630
Eichler order, 366, 372 Ford domain, 676
level, 375 form, 710
Eichler symbol, 388, 391 integral binary quadratic, 641
eigenbasis, 749 formally real, 42
Eisenstein series, 731, 799 Fourier transform, 492, 499
elementary, 628 fractional O-ideal, 250, 253
elementary divisors, 796 fractional O, O0-ideal, 253
elementary matrix, 235, 470 fractional ideal, 136
eligible invertible, 249
set of places, 221 locally principal, 250
elliptic, 599, 654 fractional ideals, 249
elliptic curve, 767 free
isogeny, 767 quadratic form, 145
embedding quadratic module, 145
normalized, 543 Frobenius norm, 126
endomorphism Frobenius normal form, 796
algebra, 21 Fuchsian group, 618, 628
endomorphism algebra, 767 arithmetic, 703
endomorphism ring, 767 elementary, 628
equidistant hyperplane, 682 first kind, 628
equivalent, 180 second kind, 628
absolute value, 179 full, 137
Euclidean, 417, 489 full congruence subgroup, 644
Euclidean lattice, 276 function
Euler product, 404, 409, 426 integrable, 497
even Clifford algebra, 70, 344, 346 nonnegative, 497
even Clifford functor, 347 Schwartz, 502
exactly divides, 719 step, 497
exceptional, 167, 589 function field, 219
exceptional algebra, 44 functional
extends, 196 antilinear, 797
exterior power, 349 functional equation, 436
extremal, 334, 335 fundamental discriminant, 250
fundamental domain, 673, 685
face pairing fundamental exact sequence, 225
complete, 687 fundamental group, 313, 628
faithful, 25 fundamental groupoid, 313
field fundamental set, 617
separably closed, 118 fundamental solution, 578
field of moduli, 792 fundamental unit, 578
finite, 325
finite adele ring, 451, 457 general linear groupoid, 312
INDEX 853
general orthogonal group, 49 Haar measure

general special orthogonal group, 55 compatible, 498
generalized Kronecker symbol, 388 Hamilton’s quaternions, 23
generated, 251 Hamiltonians, 23
genial, 559 Hasse invariant, 187
genus, 146, 270, 274, 527 Hasse principle, 225
genus 2 curve Hasse unit index, 588, 725
Weiestrass equation, 790 Hecke algebra, 756
geodesic, 592 Hecke operator, 743, 752
geodesic plane, 650 hereditary, 333, 338, 366
geodesic segment, 592 hereditary closure, 391
geodesic space, 593 Hermite normal form, 138, 273
geometry of numbers, 269 Hermite ring, 330
global Artin homomorphism, 455 Hermitian, 27
global field, 219 Hilbert equation, 187, 190
preferred embeddings, 220 Hilbert reciprocity, 223
global ring, 221 Hilbert symbol, 187
good, 630 Hilbert’s criterion, 187
good basis, 355 Hirzebruch–Jung
continued fraction, 646
good pseudobasis, 362
holomorphic at ∞, 736
Gorenstein, 381, 383
homeomorphism
Gorenstein conductor, 382
local, 619
Gorenstein saturation, 382, 386
homogeneous, 608, 616
Gram matrix, 49
homology
Grothendieck group, 328
symplectic basis, 794
group
homomorphism, 179, 305, 802
arithmetic, 705, 709 algebra, 21
topological, 179 homothetic, 303, 309, 376, 643
group action homotopic, 312
continuous, 616 Hurwitz order, 10, 160
covering space, 618 irreducible, 165
free, 616 primitive, 165
orbit, 616 unit migration, 166
proper, 622 hyperbolic, 599, 629, 654
properly discontinuous, 620 hyperbolic 𝑛-space, 656
quotient, 616 hyperbolic area, 602
quotient map, 616 hyperbolic area form, 602
regular, 615 hyperbolic length element, 594, 649
wandering, 619 hyperbolic metric, 604, 655
groupoid, 270, 302, 305 hyperbolic plane, 61, 71, 90, 594
Brandt, 303, 309 hyperbolic polygon, 603
connected, 306 sides, 603
homomorphism, 305 vertex, 603
left identity, 305 hyperbolic space, 649
right identity, 305 hyperbolic triangle, 603
854 INDEX
hyperbolic unit ball, 655 𝑝-adic, 176

hyperbolic unit disc, 604 integrable, 497
hyperbolic upper half-space, 656 integral, 151, 154, 252, 301
hyperboloid, 606 integral closure, 154
hyperboloid model, 606 integral representation, 316
hypercharacteristic, 393 integrally closed, 154
hypercomplex numbers, 7 invariant factors, 139
hyperelliptic, 790 inverse
hyperelliptic involution, 791 lattice, 257
hyperplane bisector, 682 left, 263
right, 263
icosian group, 169 inversion, 666
ideal invertible, 259
𝑅-primitive, 486 fractional ideal, 249
augmentation, 116 lattice, 250, 257
Euclidean, 417 lattice class, 272
invertible as a fractional O, O0- left, 263
ideal, 259 right, 263
maximal, 291 involution, 36
odd, 229 canonical, 37
prime, 291 conjugation, 37
proper, 260 first kind, 36
regular, 260 main, 37
topologically nilpotent, 323 positive, 122, 126
ideal tetrahedron, 659 second kind, 36
ideal vertex, 677 standard, 36
idele class group, 452 irreducible, 165, 320
idele group, 450, 452, 456 isogeny, 767, 803
finite, 457 dual, 767
idempotents, 98 rank, 771
identity elements, 306 isometric circle, 675
Igusa invariants, 792 isometry, 49, 145, 592
Igusa-Clebsch invariants, 791 oriented, 76
in the same class, 269 proper, 55
in the same narrow right class, 477 special, 55
in the same right class, 269 isomorphism, 49, 74
indecomposable, 96, 147, 320, 799 algebra, 21
indefinite, 222, 473 isotropic, 51, 608
quaternion algebra, 210 isotropy group, 630
index, 143 Iwasawa, 598
inertial degree, 197, 202 Iwasawa decomposition, 598
infinite vertex, 687
infinite vertex sequence, 687 Jacobian, 794
infinite vertex transformation, 687 Jacobson radical, 101, 316, 321
inseparable, 83 Jacobson semisimple, 321
integers Jordan–Zassenhaus theorem, 283
INDEX 855
Kleinian group, 657 linear algebraic group, 709

linearly disjoint, 228
lattice, 135, 137, 626 linked, 123
compatible, 250, 252 Lipschitz order, 159
complex, 643 Lobachevsky, 658
dual, 242, 797 Lobachevsky function, 658
full, 137 duplication formula, 659
generated, 251 local field, 181
homothetic, 376 local homeomorphism, 619
in the same right class, 269 local homeomorphism modulo stabi-
integral, 252 lizers, 620
invertible, 257 local property, 136, 141
left colon, 243 local ring
left inverse, 263 noncommutative, 200
left invertible, 263 local-global principle, 211
level, 244 localization, 135
locally free, 318 locally compact, 181
locally principal, 252 locally finite, 673, 681
mutiplicator ring, 260 locally free
order, 260 module, 140
principal, 251 locally free cancellation, 329
right colon, 243 locally free class group, 327
right inverse, 263 locally Hausdorff, 632
right invertible, 263 locally isomorphic, 270, 274
two-sided inverse, 257 locally norm-maximal, 475
left O-lattice, 315, 318 locally of the same type, 270, 274
left action, 615 locally principal, 140, 250, 252
left colon lattice, 243 locally residually inert, 391
left conductor, 342 log sine integral, 658
left Haar measure, 496 Lorentz hyperboloid, 606
left hereditary, 333, 338 Lorentz metric, 606
left identity, 305 Lorentz model, 606, 656
left inverse, 263 loxodromic, 654
left invertible, 263
left order, 151, 153 main involution, 37
left principal, 252 majorizes, 280
Legendre curve, 741 manifold, 609
Leibniz half-space, 682 atlas, 609
length, 325 isomorphism, 610
length element, 593 smooth, 609
length metric space, 592 manifolds
level, 644, 742 morphism, 609
Eichler order, 375 map
lattice, 244 equivariant, 616
Shimura orbifold, 714 Maschke’s theorem, 117
level 𝑁 with character 𝜒, 742 mass, 420, 429, 442
856 INDEX
matrix continued fraction, 646

elementary, 470 neighbor, 749, 753
matrix ring nilpotent, 102
Frobenius norm, 126 noetherian reduction, 153
maximal, 151, 155, 291 nonarchimedean, 179, 219
measurable, 497 noncommutative local ring, 200
measure nondegenerate, 47, 48, 53, 86, 145
normalized, 505 nonsingular, 61, 145
meromorphic, 736 norm, 255
meromorphic modular form of weight absolute, 256
𝑘, 736 counting, 256
meromorphic modular function, 736 normalized, 52, 87, 147, 180, 505, 543
Mestre obstruction, 793 normalized absolute values, 220
metric normalized Eisenstein series, 737
intrinsic, 592 number field, 219
mildly ramified, 209
minimal polynomial odd
universal, 111 ideal, 229
Minkowski, 269 odd Clifford bimodule, 70, 348
Minkowski space, 8 opposite algebra, 36
mixed product, 238 optimal embedding, 534
modular form optimal selectivity condition, 559
cusp form, 736 optimally selective, 559
holomorphic, 736, 785, 811 orbifold, 629
Siegel, 800 atlas, 629
modular function, 497 atlas axiom, 629
module chart, 629
annihilator, 102, 148 good, 630
completely reducible, 99 orbifold point, 630
indecomposable, 96, 320 orbifold set, 630
irreducible, 96, 320 orbifold point, 630
locally free, 140 isotropy group, 630
projective, 136, 317 stabilizer group, 630
semisimple, 99 orbifold set, 630
simple, 96, 320 orbit, 616
modulus, 498 order, 151, 152
Morita equivalent, 98 Azumaya, 247
morphism basic, 381, 395
complex torus, 793 Bass, 381, 394
multiplicative, 61 Brandt invariant, 382, 386
multiplicator ring, 260 class number, 270, 279
class set, 273
narrow class group, 284, 301 codifferent, 244
narrow class set connected, 270
order, 477 Gorenstein conductor, 382
negative-regular Hurwitz, 160
INDEX 857
lattice, 260 pointwise convergence, 631

left hereditary, 333, 338 polarization, 129, 801
Lipschitz, 159 polarized, 796
locally norm-maximal, 475 positive, 122, 126
maximal, 151, 155 positive definite, 128
projective, 152 power norm residue algebras, 116
reduced discriminant, 239 Prüfer domain, 260
simple, 247 preferred embeddings
special, 395 global field, 220
split, 376 prime, 289, 291
standard Eichler, 374 primitive, 146, 165, 301, 381, 427,
type, 270, 273 486, 530, 755
orders primitive orthogonal idempotents, 98
connected, 274 principal, 251, 796, 801
orientation, 76, 778 principal congruence subgroup, 713
oriented, 76, 78 principal ideal ring, 489
oriented basis, 643 principal polarization, 784
orthogonal, 50 product formula, 212, 220
orthogonal direct sum, 51 product polarization, 796
orthogonal group, 49 projective, 315, 317, 320
module, 136
parabolic, 599, 654 projective class group, 328
parallelizable, 608 projective line, 427
parity condition, 209, 210 proper, 55, 260, 621
partial function, 305 group action, 622
path, 312, 592 properly discontinuous, 620
length, 592 prosolvable, 206
rectifiable, 592 pseudobasis, 138
path metric space, 592 coefficient ideals, 138
Pauli spin matrices, 30 good, 362
Pell’s equation, 578 pseudogenerating set, 138
fundamental solution, 578 pseudosphere, 597
Petersson inner product, 744 pure, 27, 64
Pfister form, 61
Picard group, 290, 295 QM abelian surface, 802
Picard modular group, 663 homomorphism, 802
place isogeny, 803
archimedean, 219 quadratic, 39
complex, 219 quadratic field
nonarchimedean, 219 fundamental unit, 578
real, 219 mildly ramified, 209
places, 209, 219 parity condition, 209
eligible, 221 quadratic form, 47, 48, 136, 145
Poincaré extension, 654 anisotropic, 51
Poincaré model, 604 associated bilinear form, 49
Poincaré series, 788 binary, 301
858 INDEX
diagonal, 52 similarity, 49
discriminant, 47, 52, 59 quadric, 61
free, 145 quasi-inverse, 257
Gram matrix, 49 quasi-proper, 621
hyperbolic plane, 61, 71, 90 quaternion
integral, 301 imaginary, 27
isometry, 49 pure, 27, 64
isotropic, 51 real, 27
level, 742 scalar, 64
modular, 146 quaternion algebra, 22, 84
multiplicative, 61 S-definite, 581
nondegenerate, 53, 86, 145 absolute discriminant, 506, 510
nonsingular, 61, 145 definite, 169, 210
normalized, 52 descends, 231
orientation, 76 discriminant, 210, 222
orthogonal, 50 indefinite, 210, 222
orthogonal sum, 51 parity condition, 210
Pfister, 61 ramified, 210, 222
primitive, 301 split, 64, 73, 210, 222
radical, 53 splitting field, 73
reduced, 301 standard generators, 23
regular, 145 totally definite, 222
represents, 51 unramified, 210, 222
signed discriminant, 52 quaternion group, 160, 582
totally hyperbolic, 61 quaternion order
totally isotropic, 61 conductor, 386
universal, 51 discriminant, 234
Witt cancellation, 52 good basis, 355
Witt extension, 52 zeta function, 408, 424
Witt index, 61 quaternion ring, 361
quadratic map, 144 quaternionic multiplication (QM), 801
quadratic module, 145 quaternionic multiplication (QM) struc-
associated bilinear map, 144 ture, 784, 802
class set, 146 quaternionic projective line, 651
free, 145 quaternionic Shimura orbifold, 714
isometry, 145 level, 714
primitive, 146 quaternionic Shimura variety, 715
quadratic module in 𝑉, 345 quotient map, 616
quadratic order quotient set, 616
conductor, 250, 267 quotient topology, 617
narrow class group, 301
quadratic reciprocity, 211 radical, 53
number field, 223 radical idealizer, 382, 391
supplement, 211 radically covers, 335
quadratic space, 49 Radon measure, 496
isomorphism, 49 ramification degree, 197
INDEX 859
ramification index, 202 Riemannian metric, 607, 610

ramified, 210, 222 right absolute norm, 251
rational idele group, 448 right class, 272
real, 27 right class set, 269
real closed, 42 right colon lattice, 243
real multiplication (RM), 801 right common divisor, 164
real place, 219 right divides, 164
rectifiable, 592 right greatest common divisor, 164
reduced, 54, 301 right ideal
reduced Cayley–Hamilton theorem, 39 neighbor, 749
reduced characteristic polynomial, 39, right identity, 305
110, 113 right inverse, 250, 263
reduced discriminant, 234, 239 right invertible, 250, 263
reduced isomorphisms, 779 right multiple, 164
reduced norm, 36, 38, 110, 113, 253 right order, 151, 153
reduced projective class group, 328 right principal, 252
reduced trace, 36, 38, 110, 113 ring of integers, 221
reduction algorithm, 637 ring of modular forms, 740
reflection, 56, 62 Rodrigues’s rotation formula, 33
regular, 145, 260, 496 Rosati involution, 129, 797, 812
regular group action, 615
regular representation, 25 sated, 253
relative Brauer group, 132 satisfies strong approximation, 481
representation, 95 satisfies the S-Eichler condition, 478,
faithful, 25 489
regular, 25 satisfies the Eichler condition, 283
represents, 51 scalar, 64
residually inert, 388 scalar triple product, 30, 238
residually ramified, 388 Schwartz, 492, 502
residually split, 388 Schwartz–Bruhat, 506, 511
residue field, 181 selective, 569
restricted direct product, 446 self-dual measure, 499
reversal involution, 69 self-isometries, 49
Clifford algebra, 69 semi-order, 260
reversal map, 347 semisimple, 94
Riemann, 630 absolutely, 114
Riemann form, 795 separable, 83, 113
Riemann matrix, 794 separably closed, 118
Riemann relations, 794 separator, 682
Riemann sphere, 610 set
Riemann surface, 610 convex, 276
Riemann zeta function, 404 symmetric, 276
Riemann–Roch theorem, 527 sheaf
Riemannian manifold 𝒪𝑋 -lattices, 149
homogeneous, 608 𝒪𝑋 -orders, 231
isotropic, 608 side, 677
860 INDEX
side pairing, 677 standard tetrahedron, 661

complete, 687 standard unitary character, 501, 504,
sides 505
hyperbolic polygon, 603 star-shaped, 682
Siegel modular form, 800 Steinberg symbol, 188
Siegel upper-half space, 798 Steinitz class, 139
signature, 631, 689 step function, 497
signed discriminant, 52 strict class group, 284
similarity, 49, 145 strong approximation, 271, 468
oriented, 78 supermodule, 323
similitude factor, 49 supersingular, 768
similarity group, 49 symbol
similitude factor, 49 Eichler, 388
simple, 96, 97, 129, 247, 320 symmetric
ring, 93 set, 276
simplification property, 329 symmetric space, 626
skew field, 22
skew-Hermitian, 27 Tamagawa measure, 510, 511, 525
Skolem–Noether theorem, 95, 106 Tamagawa number, 525
small period matrix, 793 teardrop, 630
Smith–Minkowski–Siegel mass formula, tensor algebra, 67
525 theta series, 741, 750
smooth, 630 topological field, 179
atlas, 609 topological group, 179
manifold, 609 homomorphism, 179
smoothly compatible, 609 topological ring, 179
solenoid, 448 homomorphism, 179
special, 55 topologically nilpotent, 323
order, 395 torsion free, 136
special orthogonal group, 55, 76 totally definite, 222
special similarity group, 55 totally disconnected, 176
specialization, 111 totally hyperbolic, 61
sphere at infinity, 649, 655 totally isotropic, 61
spin group, 33 totally positive, 226
split, 64, 73, 210, 222 trace form, 60
splitting field, 31, 73 trace formula, 539
stabilizer, 616 transition map, 609
stabilizer group, 630 translation-invariant, 496
stable cancellation, 329 transvection, 470
stable class group, 327 triangle
stably isomorphic, 327 hyperbolic, 603
standard Eichler order, 366, 374 triangle inequality, 179
standard function, 502, 506 twist, 358
standard generators, 23 twisted similarity, 345, 358
standard involution transpose, 126 type
standard picture, 678 order, 270, 273
INDEX 861
type set, 270, 274 genus 2 curve, 790

typical, 801 weight 𝑘-invariant, 735, 785, 811
weighted projective space, 740, 791
ultrametric inequality, 179 Witt cancellation, 52
uniformizer, 181 Witt extension, 52
unimodular, 497 Witt index, 61
uniquely geodesic space, 593
unit ball zeta function
hyperbolic, 655 Dedekind, 405
unit disc idelic, 512
hyperbolic, 604 local, 502, 506
unit Hamiltonians, 27 quaternion order, 408, 424
unit migration, 166 Riemann, 404
unit tangent bundle, 608
unitary character, 499
unitary character group, 499
universal, 51, 287
universal (half-)discriminant, 86
universal element, 110
universal minimal polynomial, 111
unramified, 197, 210, 222
upper half-plane, 594
upper half-space, 649
valuation, 180
discrete, 180
equivalent, 180
extends, 196
normalized, 180
trivial, 180
value group, 180
valuation ring, 181, 199
value group
valuation, 180
vertex, 677
hyperbolic polygon, 603
vertex cycle relation, 679
volume element, 649
Voronoi domains, 685
wandering
group action, 619
Wedderburn’s little theorem, 44, 92
Wedderburn–Artin theorem, 93, 100
Weierstrass ℘-function, 733
Weierstrass equation

Quat Book

Uploaded by

Copyright:

Available Formats

Quat Book

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Quat Book

Uploaded by

Copyright:

Available Formats

Quaternion algebras

How to use this book

Program on Computational Aspects of the Langlands Program: in particular, thanks

an adequate, if not inspiring text. It was something for Dartmouth to offer

5 Ternary quadratic forms 63

8 Simple algebras and involutions 121

9 Lattices and integral quadratic forms 135

9.7 Quadratic forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

11 The Hurwitz order 159

12 Ternary quadratic forms over local fields 175

13 Quaternion algebras over local fields 195

14 Quaternion algebras over global fields 209

15.1 ⊲ Discriminantal notions . . . . . . . . . . . . . . . . . . . . . . . . 233

16 Quaternion ideals and invertibility 249

17 Classes of quaternion ideals 269

18 Picard group 289

19 Brandt groupoids 301

20 Integral representation theory 315

21 Hereditary and extremal orders 333

22 Ternary quadratic forms 343

23 Quaternion orders 365

24 Quaternion orders: second meeting 381

III Analysis 401

25 The Eichler mass formula 403

26 Classical zeta functions 419

27 Adelic framework 445

28 Strong approximation 465

29 Idelic zeta functions 491

29.3 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496

30 Optimal embeddings 529

IV Geometry and topology 575

32 Unit groups 577

33 Hyperbolic plane 591

34 Discrete group actions 615

35 Classical modular group 635

36 Hyperbolic space 649

37 Fundamental domains 671

37.8 The (6, 4, 2)-triangle group . . . . . . . . . . . . . . . . . . . . . . 690

38 Quaternionic arithmetic groups 701

39 Volume formula 717

V Arithmetic geometry 729

40 Classical modular forms 731

41 Brandt matrices 747

42 Supersingular elliptic curves 767

43 Abelian surfaces with QM 783

Symbol Definition List 815

1.1 Hamilton’s quaternions

Figure 1.1.2: William Rowan Hamilton

Figure 1.1.3: William Rowan Hamilton, a sand sculpture by Daniel Doyle,

events on the part of others, if I should even be allowed to live long

which contains the Solution of the Problem, but of course, as an inscrip-

Figure 1.1.4: The Broom Bridge plaque (author’s photo)

160 BLBHENTS OF QUATBRNIOliS. (BOOK. U.

the law of i, j, Aagree with usual and algebraic law: namely~

Figure 1.1.5: A page from Hamilton’s Elements of quaternions [Ham1866]

Leonhard Euler (1707–1783):

(𝑎 21 + 𝑎 22 + 𝑎 23 + 𝑎 24 ) (𝑏 21 + 𝑏 22 + 𝑏 23 + 𝑏 24 ) = 𝑐21 + 𝑐22 + 𝑐23 + 𝑐24 =

Indeed, the full multiplication law for quaternions reads precisely

with 𝑐 1 , 𝑐 2 , 𝑐 3 , 𝑐 4 as defined in (1.1.6); the four-square identity corresponds to taking