BMS Particlesin Three Dimensions

BMS Particles
in Three Dimensions
Blagoje Oblak∗
arXiv:1610.08526v3 [hep-th] 26 Jan 2018
Abstract
This thesis is devoted to the group-theoretic aspects of three-dimensional

quantum gravity on Anti-de Sitter and Minkowskian backgrounds. In partic-
ular we describe the relation between unitary representations of asymptotic
symmetry groups and gravitational perturbations around a space-time met-
ric. In the asymptotically flat case this leads to BMS particles, representing
standard relativistic particles dressed with gravitational degrees of freedom
accounted for by coadjoint orbits of the Virasoro group. Their thermody-
namics are described by BMS characters, which coincide with gravitational
one-loop partition functions. We also extend these considerations to higher-
spin theories and supergravity.
∗
Current e-mail: [email protected]
Université Libre de Bruxelles
faculté des sciences
service de physique théorique et mathématique
BMS Particles
in Three Dimensions
Blagoje Oblak
Thèse présentée en vue de l’obtention du titre de Docteur en Sciences
Réalisée sous la direction du Prof. Glenn Barnich
Année académique 2015-2016
— electronic version —
This thesis was defended behind closed doors on June 22nd, and publicly on June
24th, in front of the following jury at Université Libre de Bruxelles:
President: Prof. Marc Henneaux
Secretary: Prof. Petr Tinyakov
Advisor: Prof. Glenn Barnich
Prof. Guillaume Bossard,

Ecole Polytechnique, Paris (France)
Prof. Stéphane Detournay,
Université Libre de Bruxelles (Belgium)
Prof. Axel Kleinschmidt,
Albert Einstein Institut, Potsdam (Germany)
The work presented in this thesis was carried out in the group of Mathematical Physics of Fun-
damental Interactions at Université Libre de Bruxelles (Belgium), and at the Department
of Applied Mathematics and Theoretical Physics of the University of Cambridge (United
Kingdom). It was supported by the Fonds de la Recherche Scientifique – F.N.R.S., of which
the author is a Research Fellow (aspirant), by the Wiener-Anspach Foundation, and by the
International Solvay Institutes.
Printed in June 2016 by Presses Universitaires de Bruxelles ASBL — Avenue Franklin Roo-
sevelt 50, C.P. 149, B-1050, Belgium.
Typeset in LATEX with the memoir class.

Mojoj najmilijoj porodici,
s mnogo ǉubavi.
Contents
Note: sections marked with an asterisk may be skipped on a first reading.
Foreword xiii
How to read this thesis ...................................................................................... xiii
Acknowledgements............................................................................................. xiv
1 Introduction 1
1.1 Asymptotic BMS symmetry ...................................................................... 1
1.2 Global BMS and extended BMS ............................................................... 2
1.3 Holography ................................................................................................ 6
1.4 BMS particles and soft gravitons .............................................................. 8
1.5 Plan of the thesis....................................................................................... 10
I Quantum symmetries 13
2 Quantum mechanics and central extensions 15
2.1 Symmetries and projective representations ............................................... 15
2.1.1 Quantum mechanics .......................................................................... 15
2.1.2 Symmetry representation theorem...................................................... 16
2.1.3 Projective representations ................................................................. 18
2.1.4 Central extensions ............................................................................ 19
2.1.5 Topological central extensions............................................................ 20
2.1.6 Classifying projective representations ................................................. 21
2.2 Lie algebra cohomology ............................................................................. 22
2.2.1 Cohomology ..................................................................................... 22
2.3 Group cohomology .................................................................................... 27
2.3.1 Cohomology ..................................................................................... 27
3 Induced representations 31
3.1 Wavefunctions and measures ..................................................................... 32
3.1.1 Measures .......................................................................................... 32
vii
Contents viii
3.1.2 Hilbert spaces of wavefunctions.......................................................... 34

3.1.3 Equivalent measures and Radon-Nikodym derivatives ......................... 36
3.2 Quasi-regular representations .................................................................... 38
3.2.1 Quasi-invariant measures on homogeneous spaces ............................... 38
3.2.2 The simplest induced representations ................................................. 41
*3.2.3 Radon-Nikodym is a cocycle .............................................................. 42
3.3 Defining induced representations............................................................... 43
3.3.1 Standard boosts ................................................................................ 44
3.3.2 Induced representations ..................................................................... 45
3.3.3 Properties of induced representations ................................................. 46
3.3.4 Plane waves ...................................................................................... 48
3.4 Characters ................................................................................................. 51
3.4.1 Characters are partition functions ...................................................... 51
3.4.2 The Frobenius formula ...................................................................... 52
3.4.3 Characters and fixed points ............................................................... 54
*3.5 Systems of imprimitivity ........................................................................... 54
3.5.1 Projections and imprimitivity ............................................................ 55
3.5.2 Imprimitivity theorem ....................................................................... 57
4 Semi-direct products 61
4.1 Representations and particles.................................................................... 61
4.1.1 Semi-direct products ......................................................................... 61
4.1.2 Momenta .......................................................................................... 63
4.1.3 Orbits and little groups ..................................................................... 64
4.1.4 Particles ........................................................................................... 66
4.1.5 Exhaustivity theorem ........................................................................ 70
4.2 Poincaré particles ...................................................................................... 71
4.2.1 Poincaré groups ................................................................................ 71
4.2.3 Particles ........................................................................................... 77
4.2.4 Massive characters ............................................................................ 78
4.2.5 Massless characters ........................................................................... 81
*4.2.6 Wigner rotations and entanglement.................................................... 83
4.3 Poincaré particles in three dimensions ...................................................... 85
4.3.1 Poincaré group in three dimensions .................................................... 85
4.3.2 Particles in three dimensions ............................................................. 87
4.3.3 Characters ........................................................................................ 89
*4.4 Galilean particles....................................................................................... 90
4.4.1 Bargmann groups ............................................................................. 90
4.4.3 Particles ........................................................................................... 94
4.4.4 Characters ........................................................................................ 95
5 Coadjoint orbits and geometric quantization 97

5.1 Symmetric phase spaces ............................................................................ 98
5.1.1 Lie groups ........................................................................................ 98
5.1.2 Adjoint and coadjoint representations ................................................ 99
Contents ix
5.1.3 Poisson structures ............................................................................. 100

5.1.4 Symplectic structures ........................................................................ 102
5.1.5 Kirillov-Kostant structures ................................................................ 104
5.1.6 Momentum maps .............................................................................. 106
5.2 Geometric quantization ............................................................................. 111
5.2.1 Line bundles and wavefunctions ......................................................... 111
5.2.2 Quantization of cotangent bundles ..................................................... 113
*5.2.3 Quantization of arbitrary symplectic manifolds ................................... 115
5.2.4 Symmetries and representations ......................................................... 117
5.3 World lines on coadjoint orbits.................................................................. 118
5.3.1 World lines and quantization conditions ............................................. 118
5.3.2 Interlude: the Maurer-Cartan form .................................................... 120
5.3.3 Coadjoint orbits and Sigma models .................................................... 122
*5.3.4 Coadjoint orbits and characters of SL(2, R) ........................................ 124
5.4 Coadjoint orbits of semi-direct products ................................................... 127
5.4.1 Adjoint representation of G n A ........................................................ 127
5.4.2 Coadjoint representation of G n A ..................................................... 128
5.4.3 Coadjoint orbits ................................................................................ 130
5.4.4 Geometric quantization and particles ................................................. 134
5.4.5 World lines ....................................................................................... 137
5.5 Relativistic world lines .............................................................................. 138
5.5.1 Coadjoint orbits of Poincaré .............................................................. 138
5.5.2 Scalar world lines .............................................................................. 138
*5.5.3 Galilean world lines........................................................................... 140
II Virasoro symmetry and AdS3 gravity 143

6 The Virasoro group 145
6.1 Diffeomorphisms of the circle .................................................................... 146
6.1.1 Infinite-dimensional Lie groups .......................................................... 146
6.1.2 The group of diffeomorphisms of the circle ......................................... 146
6.1.3 Topology of Diff(S 1 ) ......................................................................... 148
6.1.4 Adjoint representation and vector fields.............................................. 150
6.1.5 Primary fields on the circle ................................................................ 152
6.1.6 Coadjoint representation of Diff(S 1 ) ................................................... 153
6.1.7 Exponential map and vector flows ...................................................... 155
6.2 Virasoro cohomology ................................................................................. 156
6.2.1 The Gelfand-Fuks cocycle .................................................................. 156
6.2.2 The Bott-Thurston cocycle ................................................................ 159
6.2.3 Primary cohomology of Vect(S 1 ) ....................................................... 162
6.2.4 Primary cohomology of Diff(S 1 ) ........................................................ 163
6.3 On the Schwarzian derivative .................................................................... 164
6.3.1 The Schwarzian derivative is a cocycle ............................................... 164
6.3.2 Projective invariance of the Schwarzian .............................................. 166
6.4 The Virasoro group ................................................................................... 171
6.4.1 Centrally extended groups revisited ................................................... 171
Contents x
6.4.2 Virasoro group.................................................................................. 173

6.4.3 Adjoint representation and Virasoro algebra ....................................... 173
6.4.4 Coadjoint representation ................................................................... 174
6.4.5 Kirillov-Kostant bracket .................................................................... 176
7 Virasoro coadjoint orbits 179

7.1 Coadjoint orbits of the Virasoro group...................................................... 179
7.1.1 Centreless coadjoint orbits ................................................................. 179
7.1.2 Basic properties of centrally extended orbits ....................................... 181
7.1.3 Hill’s equation and monodromy ......................................................... 183
7.1.4 Winding number ............................................................................... 189
7.2 Virasoro orbit representatives.................................................................... 191
7.2.1 Prelude: Conjugacy classes of SL(2, R) ............................................... 191
7.2.2 Elliptic orbits ................................................................................... 193
7.2.3 Degenerate parabolic orbits ............................................................... 196
7.2.4 Hyperbolic orbits without winding ..................................................... 197
7.2.5 Hyperbolic orbits with winding .......................................................... 199
7.2.6 Non-degenerate parabolic orbits ......................................................... 202
7.2.7 Summary: a map of Virasoro orbits ................................................... 204
7.3 Energy positivity ....................................................................................... 206
7.3.1 Energy functional ............................................................................. 207
7.3.2 The average lemma ........................................................................... 208
7.3.3 Orbits with constant representatives .................................................. 210
7.3.4 Orbits without constant representatives ............................................. 213
7.3.5 Summary: a new map of Virasoro orbits ............................................ 214
8 Symmetries of gravity in AdS3 217

8.1 Generalities on three-dimensional gravity ................................................. 217
8.1.1 Einstein gravity in three dimensions ................................................... 218
8.1.2 Boundary conditions and boundary terms .......................................... 219
8.1.3 Asymptotic symmetries ..................................................................... 221
8.2 Brown-Henneaux metrics in AdS3 ............................................................. 226
8.2.1 Geometry of AdS3 ............................................................................. 226
8.2.2 Brown-Henneaux boundary conditions ............................................... 228
8.2.3 Asymptotic Killing vector fields ......................................................... 230
8.2.4 On-shell Brown-Henneaux metrics ..................................................... 232
8.2.5 Surface charges and Virasoro algebra ................................................. 233
8.2.6 Zero-mode solutions .......................................................................... 235
8.3 The phase space of AdS3 gravity............................................................... 238
8.3.1 AdS3 metrics as CFT2 stress tensors .................................................. 238
8.3.2 Boundary gravitons and Virasoro orbits ............................................. 239
8.3.3 Positive energy theorems ................................................................... 241
8.4 Quantization and Virasoro representations ............................................... 242
8.4.1 Highest-weight representations of sl(2, R) ........................................... 242
8.4.2 Virasoro modules .............................................................................. 244
8.4.3 Virasoro characters ........................................................................... 249
8.4.4 Dressed particles and quantization ..................................................... 251
Contents xi
III BMS3 symmetry and gravity in flat space 255

9 Classical BMS3 symmetry 257
9.1 BMS metrics in three dimensions .............................................................. 257
9.1.1 Three-dimensional Minkowski space ................................................... 258
9.1.2 Poincaré symmetry at null infinity ..................................................... 259
9.1.3 BMS3 fall-offs and asymptotic symmetries .......................................... 261
9.1.4 On-shell BMS3 metrics ...................................................................... 265
9.1.5 Surface charges and BMS3 algebra ..................................................... 266
9.1.6 Zero-mode solutions .......................................................................... 268
9.2 The BMS3 group ....................................................................................... 270
9.2.1 Exceptional semi-direct products ....................................................... 270
9.2.2 Defining BMS3 ................................................................................. 275
9.2.3 Adjoint representation and bms3 algebra ............................................ 278
9.2.4 Coadjoint representation ................................................................... 279
*9.2.5 Some cohomology ............................................................................. 281
9.3 The BMS3 phase space.............................................................................. 283
9.3.1 Phase space as a coadjoint representation ........................................... 283
9.3.2 Boundary gravitons and BMS3 orbits ................................................. 285
9.3.3 Positive energy theorem .................................................................... 285
9.4 Flat limits.................................................................................................. 286
9.4.1 From Diff(S 1 ) to BMS3 ..................................................................... 286
9.4.2 From Witt to bms3 ........................................................................... 288
9.4.3 Stress tensors and central charges ...................................................... 290
9.4.4 The Galilean conformal algebra ......................................................... 291
10 Quantum BMS3 symmetry 293

10.1 BMS3 particles ......................................................................................... 293
10.1.1 Orbits and little groups .................................................................... 294
10.1.2 Mass, supermomentum, central charge .............................................. 296
10.1.3 Measures on superrotation orbits ...................................................... 300
10.1.4 States of BMS3 particles .................................................................. 302
10.1.5 Dressed particles and quantization .................................................... 304
10.1.6 The BMS3 vacuum .......................................................................... 307
10.1.7 Spinning BMS3 particles .................................................................. 309
10.1.8 BMS particles in four dimensions? .................................................... 310
10.2 BMS modules and flat limits.................................................................... 313
10.2.1 Poincaré modules in three dimensions ............................................... 314
10.2.2 Induced modules for bms3 ................................................................ 318
10.2.3 Representations of the Galilean conformal algebra ............................. 321
10.3 Characters of the BMS3 group ................................................................. 324
10.3.1 Massive characters ........................................................................... 324
10.3.2 Comparison to Poincaré and Virasoro ............................................... 327
10.3.3 Vacuum character ............................................................................ 328
11 Partition functions and characters 331

11.1 Rotating canonical partition functions..................................................... 332
11.1.1 Heat kernels and method of images ................................................... 332
Contents xii
11.1.2 Bosonic higher spins ........................................................................ 334

11.1.3 Partition functions and BMS3 characters........................................... 340
11.1.4 Relation to Poincaré characters ........................................................ 341
11.2 Representations and characters of flat WN .............................................. 343
11.2.1 Higher spins in AdS3 and WN algebras ............................................. 343
11.2.2 Flat W3 algebra ............................................................................... 346
11.2.3 Flat W3 characters .......................................................................... 349
11.2.4 Flat WN algebras ............................................................................ 350
11.3 Flat W3 modules ...................................................................................... 353
11.3.1 Ultrarelativistic and non-relativistic limits of W3 ............................... 353
11.3.2 Induced modules for the flat W3 algebra ........................................... 355
11.4 Super-BMS3 and flat supergravity ........................................................... 357
11.4.1 Fermionic higher spin partition functions .......................................... 358
11.4.2 Supersymmetric BMS3 groups .......................................................... 361
11.4.3 Supersymmetric BMS3 particles ....................................................... 365
*11.A From mixed traces to bosonic characters ................................................. 369
11.A.1 Mixed traces and symmetric polynomials .......................................... 369
11.A.2 Symmetric polynomials and SO(D) characters ................................... 372
11.A.3 Differences of SO(D) characters ........................................................ 374
11.A.4 From SO(D) to SO(D − 1) ............................................................... 376
*11.B From mixed traces to fermionic characters .............................................. 377
11.B.1 Mixed traces and symmetric polynomials .......................................... 377
11.B.2 Symmetric polynomials and SO(D) characters ................................... 378
12 Conclusion 381
Bibliography 385
Index 407
Foreword
This thesis collects thoughts and results that originate from a four-year long research
project in theoretical physics. The main topic is representation theory and its appli-
cation to quantum gravity, in particular in the context of BMS symmetry.
How to read this thesis

The text consists of three parts:
Part I: Group theory;
Part II: Virasoro symmetry and AdS3 /CFT2 ;
Part III: BMS symmetry in three dimensions.
It is written in such a way that each part can be read more or less independently of
the others, although the later parts do depend on background material presented in
the earlier ones; the logical flow of chapters is explained in section 1.5 below. A few
sections are marked with an asterisk; they contain somewhat more advanced material
that may be skipped without affecting the reading of the main track.
The original contributions of this thesis are based on the following publications:
• G. Barnich and B. Oblak, “Holographic positive energy theorems in three-di-
mensional gravity,” Class. Quant. Grav. 31 (2014) 152001, 1403.3835.
• G. Barnich and B. Oblak, “Notes on the BMS group in three dimensions: I.
Induced representations,” JHEP 06 (2014) 129, 1403.5803.
• G. Barnich and B. Oblak, “Notes on the BMS group in three dimensions: II.
Coadjoint representation,” JHEP 03 (2015) 033, 1502.00010.
• B. Oblak, “Characters of the BMS Group in Three Dimensions,” Commun. Math.
Phys. 340 (2015), no. 1, 413–432, 1502.03108.
• G. Barnich, H. A. González, A. Maloney, and B. Oblak, “One-loop partition
function of three-dimensional flat gravity,” JHEP 04 (2015) 178, 1502.06185.
• B. Oblak, “From the Lorentz Group to the Celestial Sphere,” Notes de la Sep-
tième BSSM, U.L.B. (2015). 1508.00920.
• A. Campoleoni, H. A. González, B. Oblak, and M. Riegler, “Rotating Higher
Spin Partition Functions and Extended BMS Symmetries,” JHEP 04 (2016)
034, 1512.03353.
xiii
Foreword xiv
• H. Afshar, S. Detournay, D. Grumiller, and B. Oblak, “Near-Horizon Geometry

and Warped Conformal Symmetry,” JHEP 03 (2016) 187, 1512.08233.
• A. Campoleoni, H. A. González, B. Oblak, and M. Riegler, “BMS Modules

in Three Dimensions,” Int. J. Mod. Phys. A31 (2016), no. 12, 1650068,
1603.03812.
Acknowledgements
Professional community
This thesis could not have been completed without the help and support of a number
of people. First and foremost I am indebted to my supervisor, Prof. Glenn Barnich, for
suggesting the topic in the first place and guiding me through its completion. Working
with Glenn has been both a delight and a challenge; the former for his contagious pas-
sion for physics, and the latter for his sharp critical mind and unwavering skepticism,
leading to numerous lively discussions about the nature of our work and of science
altogether. Our complementary approaches have made our collaboration all the more
fruitful; I am grateful to him for preserving my complete freedom while guiding me
along the thesis. He has been a teacher, a friend, and a model of independence and
scientific integrity that I hope to emulate myself.
I also wish to thank my other collaborators. The earliest ones in the history of
my Ph.D. are Sophie De Buyl, Stéphane Detournay and Antonin Rovai, with whom I
spent a few weeks in Harvard in the Spring of 2013. I am grateful to them for their
friendship and our many relaxed discussions about physics and other matters, includ-
ing world economy, Belgian movies and American food. Later on my path crossed that
of Hernán González, starting with several ultra local journal clubs on quantum gravity,
and ending with common projects. I am grateful to him for sharing his enthusiasm
and almost convincing me that three-dimensional gravity is, in fact, gravity. My grat-
itude also goes to Hamid Afshar, Daniel Grumiller, Alexander Maloney, Max Riegler,
and especially Andrea Campoleoni, for fruitful and enjoyable scientific collaborations;
I hope there will be many more in the future.
The friendly environment at the physics department of the Université Libre de

Bruxelles has also been a great help in completing this thesis. In particular I am grate-
ful to my office neighbour, Laura Donnay, as well as Pujian Mao and Marco Fazzi, for
enjoyable discussions during coffee breaks or lunches. I also thank the other students,
postdocs and professors in the service of mathematical physics for making my days
as a Ph.D. student in Brussels as pleasant as can be; this includes Riccardo Argurio,
Andrés Collinucci, Geoffrey Compère, Laure-Anne Douxchamps, Simone Giacomelli,
Gaston Giribet, Paolo Gregori, Marc Henneaux, Victor Lekeu, Arnaud Lepage-Jutier,
Andrea Marzolla, Roberto Oliveri, Arash Ranjbar, Waldemar Schulgin, Shyam Sunder
Gopalakrishnan, and Céline Zwikel. In addition I wish to thank my colleagues out-
side of Brussels — Harold Erbin, Lucien Heurtier, Jules Lamers, Ruben Monten, Ali
Seraj, and Ellen Van der Woerd — for our numerous delightful interactions. Finally,
the logistics and organization of pretty much anything at the service of mathematical
Foreword xv
physics would be impossible without the precious help and efficiency of the adminis-
trative staff of the service and the Solvay Institutes — Dominique Bogaerts, Fabienne
De Neyn, Marie-France Rogge, Isabelle Van Geet, and Chantal Verrier. To them,
thank you for your unshakeable goodwill in the face of the fiercest of administrative
challenges.
Since October 2015 I have had the chance to meet a number of new colleagues at
the Deparment of Applied Mathematics and Theoretical Physics of the University of
Cambridge. This would not have been possible without the support of Harvey Re-
all, whom I wish to thank warmly for this incredible opportunity. I am also grateful
to Siavash Golkar, Shahar Hadar, Sasha Hajnal-Corob, Kai Roehrig, Arnab Rudra,
Joshua Schiffrin and Piotr Tourkine for making my time there a daily enjoyment.
Still on the international side, I would like to thank Matthias Gaberdiel and Bianca
Dittrich for giving me the opportunity to give seminars at their respective institutions.
Both of these visits were a delight, and I had great pleasure in sharing some thoughts
about my research with them and their colleagues.
Finally, I am grateful to Guillaume Bossard, Axel Kleinschmidt and Petr Tinyakov

(in addition to Glenn Barnich, Stéphane Detournay and Marc Henneaux who have
already been cited) for accepting to be part of my thesis jury, and for their many
questions and suggestions at the private and public defences.
Family and friends

On a more private side I must mention the help and unconditional support provided
by my family, and in particular my parents, Tijana and Dušan Oblak. Their contri-
bution to this work is invisible to the naked eye, but it is actually so all-pervasive
that it is hard to tell what would have become of me if I hadn’t had such an amazing
team behind my back. I am grateful to them for all the love they have given me and
in particular for their support (both moral and practical) during the last year. I also
wish to thank my cousins Sofija and Milena Stevanović, my aunt Ksenija Stevanović
and my grandmother Vera Vujadinović for always being there for me when it counts
and for the many beautiful moments we have had the chance to share.
My friends have also contributed greatly to this thesis, often unknowingly. Above
all I wish to thank David Alaluf, Mitia Duerinckx, Jihane Elyahyioui, Geoffrey Mullier,
Olmo Nieto-Silleras, and Roxane Verdikt for our many adventures together, includ-
ing holidays and unforgettable parties (in the Balkans and elsewhere). As a matter of
fact, Mitia Duerinckx has been a great help in understanding some of the mathematics
used in this thesis, always patiently replying to my many e-mails with questions about
Hilbert spaces, measures and the like. More generally, I am grateful to my friends from
the mathematics department of ULB for their friendship and sometimes professional
interactions, including Charel Antony, Cédric De Groote, Julien Meyer and Patrick
Weber. Finally, I am indebted to Thierry Maerschalk for helping me out long ago with
some cute LATEX tricks, many of which were used in this thesis.
Foreword xvi
On the other side of the Channel I wish to thank the flurry of people whom I
have had the chance to meet during the last year of my Ph.D. and who have allowed
me to take helpful breaks away from physics. This includes Abhimanyu Chandra,
Robert Cochrane, Marius Leonhardt, Dušan Perović, Frank Schindler, Bianca Schor,
and Aaron Wienkers. I hope to meet them again in the future, despite our living in
different corners of the world.
Finally, I wish to thank Vanessa Drianne, who has been in an entangled state
between Brussels and Cambridge for most of the past year and whose support has
been critical for my well-being in the last stages of writing the thesis.
Physics and art

At this point I would like to thank two professors who have been most important for
my development, both as a scientist and as a person. The first is my high school
physics teacher, Emmanuel Thiran, who first managed to show me a glimpse of the
beauty of Nature and the thrill of lifting its veil. While indirect, his influence pervades
the entirety of my approach to physics and guides me to this very day. The second
is Michel Laurent, my piano teacher. Since more than a decade he has been showing
me the subtleties of music, but in truth his teaching extends far beyond that. While
difficult to express in words, the conceptions of art and beauty that he has conveyed
to me have had a great influence on me, and hence on this thesis.
Financial and logistic support

To conclude I wish to thank the institutions who have provided me with financial
and logistic support throughout my Ph.D. Above all I am grateful to the Fonds de la
Recherche Scientifique – F.N.R.S. for the grant (number FC-95570) that has allowed
me to make a living out of the most delightful of occupations. On the other hand the
financial support allowing me to work at the university of Cambridge during the last
academic year has been granted to me by the Fondation Wiener-Anspach. I gratefully
acknowledge their support in this marvelous experience, and wish to thank particularly
Nicole Bosmans for her help with logistics.
Chapter 1
Introduction
The quantization of gravity is one of the long-standing puzzles of theoretical physics.

The purpose of this thesis is to study certain aspects of the problem that can be
studied on the sole basis of symmetries, without any assumptions on the underlying
microscopic theory. In this introduction we describe this strategy in some more detail,
starting in section 1.1 with a broad overview of asymptotic symmetries in general and
Bondi-Metzner-Sachs (BMS) symmetry in particular. We then introduce the distinc-
tion between global and extended BMS groups in section 1.2. Section 1.3 is devoted to
a lightning review of AdS/CFT and its putative Minkowskian counterpart. Finally, in
section 1.4 we describe the relation between BMS symmetry and soft graviton degrees
of freedom. Section 1.5 contains a general presentation of the upcoming chapters and
describes their logical flow.
1.1 Asymptotic BMS symmetry

The notion of symmetry is a cornerstone of physics and mathematics. A system is
symmetric if there exists a set of transformations that leave it invariant, i.e. that pre-
serve its structure. In physical terms, saying that a system has symmetries is really
saying that there exist certain transformations that can be performed without affecting
the outcome of experiments. For instance, translational symmetry is the statement
that the result of an experiment does not depend on where one carries it out. By
construction, the set of symmetry transformations of a system forms a group, so the
mathematical tool used in the study of symmetries is group theory.
In this thesis we shall be concerned with symmetries of gravitational systems, that

is, changes of coordinates that can be applied to space-time and that leave invariant
the large-distance behaviour of the gravitational field. They are known as asymp-
totic symmetries and can be thought of as a generalization of Poincaré symmetry for
systems endowed with a weak gravitational field. In other words, these symmetries
are those one would observe by looking at a gravitational system “from far away”. In
that context, the general type of question that we will ask is the following: given the
asymptotic symmetries of a gravitational system, what are their physical implications?
In particular, how do these symmetries affect one’s intuition about particle physics?
1
Chapter 1. Introduction 2
Asymptotic symmetries of gravitational systems have been studied for about fifty
years by now. Their first appearance in the literature is also the one that motivates
the present work. Indeed, it was observed in the sixties by Bondi, van der Burg,
Metzner [1, 2] and Sachs [3, 4] that the presence of gravitation in an asymptotically
flat space-time leads to a symmetry group that is much, much larger than standard
Poincaré. The group that they found turned out to be an infinite-dimensional exten-
sion of the Poincaré group, and is known today as the Bondi-Metzner-Sachs group, or
BMS group for short.
The BMS group considered by the authors of [2–4] consists of two pieces: the
first is the standard Lorentz group of special relativity, and the second is an infinite-
dimensional Abelian group of so-called supertranslations.1 In abstract mathematical
notation, its structure can be written symbolically as
BMS = Lorentz n Supertranslations. (1.1)
The notation n used here means that elements of the BMS group are pairs consisting
of a Lorentz transformation and a supertranslation, and that Lorentz transformations
act non-trivially on supertranslations. In the same way, the Poincaré group is
Poincaré = Lorentz n Translations. (1.2)
The latter is a subgroup of BMS: the group of space-time translations is contained in

the infinite-dimensional group of supertranslations.
Groups of the form (1.1) or (1.2) are known as semi-direct products. They are
ubiquitous in physics, and many of the conclusions of this thesis rely on this structure.
1.2 Global BMS and extended BMS

In this section we introduce Bondi coordinates to explain briefly how BMS symme-
try emerges from an asymptotic analysis. We then describe the distinction between
“global” and “extended” BMS transformations.
Bondi coordinates
Consider Minkowski space-time, endowed with inertial coordinates xµ in terms of which
the metric reads
ds2 = ηµν dxµ dxν , with (ηµν ) = diag(−1, +1, +1, +1). (1.3)
Now suppose we wish to study, say, outgoing massless particles sent by an observer
located at the spatial origin. For this purpose we introduce retarded Bondi coordinates
1/2 x1 + ix2
r ≡ xi xi u ≡ x0 − r.

, z≡ , (1.4)
r + x3
1
The terminology of “super-things” here has nothing to do with supersymmetry: “super-object”
simply means that a certain object, which one is familiar with in the finite-dimensional context of
special relativity, gets extended in an infinite-dimensional way in the BMS group.
Here r is a space-like radial coordinate, z is a stereographic coordinate on the sphere

of radius r (such that the north and south poles respectively correspond to z = 0 and
z = ∞), and u is known as retarded time. In these coordinates the Minkowski metric
(1.3) reads
4dzdz̄
ds2 = −du2 − 2 dudr + r2 (1.5)
(1 + z z̄)2
and the world line of an outgoing massless particle (moving away from the origin) is
of the form u = const., z = const.:
Figure 1.1: The coordinates u and r in space-time. The time coordinate x0 points
upwards. The wavy red line represents an outgoing radial massless particle emitted at
r = 0 and moving to some non-zero distance r away from the observer at r = 0; the
particle moves along one of the generators of the light cone given by u = const. The
drawing is three-dimensional, so the circle of radius r in this picture would actually
be a sphere (spanned by the coordinate z) in a four-dimensional space-time.
In terms of Bondi coordinates, the region reached by massless particles emitted at

some moment from the origin r = 0 is a sphere at null infinity (r → +∞) spanned by
the complex coordinate z, called a (future) celestial sphere. There is one such sphere
for each value of retarded time u; the succession of all possible celestial spheres is a
manifold R × S 2 located at r → +∞ and known as future null infinity. It is the region
where all outgoing massless radiation “escapes” out of space-time; it is the upper null
cone of the Penrose diagram of Minkowski space-time.
Figure 1.2: A representation of celestial spheres on the Penrose diagram of Minkowski

space. As in fig. 1.1, the wavy red line represents an outgoing radial light ray. The
drawing is three-dimensional, so the red circle at the top of the picture would really be
a sphere — a celestial sphere — in a four-dimensional space-time. Future null infinity
is the cone R × S 2 on the upper half of the image, spanned by u and z.
There exists a parallel construction of Bondi coordinates which is convenient for

the study of past null infinity. These are advanced Bondi coordinates (r, z, v), defined
in terms of inertial coordinates xµ exactly as in (1.4) up to a sign difference in retarded
time: v = x0 + r. The spheres at r → +∞ then are past celestial spheres and they
foliate past null infinity (spanned by v and z) in slices of constant time. In particular,
all incoming massless particles originate from past null infinity.
Asymptotic flatness and the BMS group

We now have the tools needed to introduce BMS symmetry. First, one declares that
a space-time manifold is asymptotically flat at, say, future null infinity, if it admits
local coordinates (r, z, u) such that, as r goes to infinity with u finite, the metric takes
the form (1.5) up to subleading corrections. These coordinates need not be defined
globally — all that is needed is that they span a neighbourhood of future null infinity.
Also, there is a precise definition of what is meant by “subleading corrections”; these
are alterations of the Minkowski metric (1.5) that typically decay as inverse powers
of r at infinity, but the allowed powers themselves are constrained in a specific way.
These constraints are motivated by physical considerations and they are part of the
definition of “asymptotic flatness”. (We will not deal with these subtleties for now, but
we shall display them in section 9.1 in the three-dimensional case.)
The notion of asymptotic flatness allows one to define the associated asymptotic
symmetry group. The latter consists, roughly speaking, of diffeomorphisms of space-
time that preserve the asymptotic behaviour of the metric.2 Bondi et al. [2–4] found
2
More precisely, the asymptotic symmetry group is the quotient of the group of diffeomorphisms
that preserve the asymptotic behaviour of the metric by its normal subgroup consisting of so-called
that there are two families of such diffeomorphisms:

• The first family consists of Lorentz transformations. Their effect at null infinity
is that of conformal transformations of celestial spheres (i.e. Möbius transforma-
tions) given in terms of the stereographic coordinate z of (1.4) by

az + b a b
z 7→ + O(1/r), ∈ SL(2, C). (1.6)
cz + d c d
This property relies on the isomorphism3 SO(3, 1)↑ ∼ = SL(2, C)/Z2 , which ex-
presses Lorentz transformations in terms of SL(2, C) matrices. Lorentz trans-
formations also act on the coordinates r and u at infinity by angle-dependent
rescalings, but this subtlety is unimportant at this stage.
• The second family consists of angle-dependent translations of retarded time,
u 7→ u + α(z, z̄), (1.7)
where α(z, z̄) is any (smooth) real function on the sphere. In this language,
Poincaré space-time translations are reproduced by functions α which are linear
combinations of the functions
1 − z z̄ z + z̄ i(z − z̄)
1, , , .
1 + z z̄ 1 + z z̄ 1 + z z̄
(In terms of polar coordinates θ and ϕ, this corresponds to the spherical har-
monics Y00 (θ, ϕ) and Y1,m (θ, ϕ) with m = −1, 0, 1.) This is the main surprise
discovered by Bondi et al. It states that asymptotic symmetries (as opposed to
isometries) enhance the Poincaré group to an infinite-dimensional group with
an infinite-dimensional Abelian normal subgroup consisting of transformations
(1.7). These transformations are the supertranslations alluded to in eq. (1.1).
Extended BMS
The group of asymptotic symmetry transformations (1.6) and (1.7) is the original
BMS group discovered in [2–4]. It consists of globally well-defined, invertible trans-
formations of null infinity, so from now on we call it the global BMS group. This
slight terminological alteration is rooted in one of the most intriguing aspects of BMS
symmetry. Indeed, in their work, Bondi et al. observed that asymptotic symmetries
include conformal transformations (1.6), but in principle one may even include trans-
formations generated by arbitrary (generally singular) conformal Killing vector fields
on the celestial spheres. Only six of those vector fields generate the invertible Möbius
transformations (1.6); the remaining ones are singular. Upon including these extra
generators, the global conformal transformations (1.6) are enhanced to arbitrary local
conformal transformations
z 7→ f (z) + O(1/r), (1.8)
where f (z) is any meromorphic function. Despite their singularities, these transfor-
mations do preserve the asymptotic behaviour of the metric and may therefore qualify
trivial diffeomorphisms. We will return to this in section 8.1.
3
O(3, 1) is the Lorentz group in four dimensions and SO(3, 1)↑ is its largest connected subgroup.
as asymptotic symmetries, at least infinitesimally.
The extension of the BMS group obtained by replacing Lorentz transformations by

local conformal transformations (1.8) is called the extended BMS group.4 In that con-
text, local conformal transformations of celestial spheres are known as superrotations
and should be thought of as an infinite-dimensional extension of Lorentz transforma-
tions, in the same way that supertranslations extend space-time translations. In the
notation of (1.1) and (1.2), the extended BMS group looks like
Extended BMS = Superrotations n Supertranslations (1.9)
where now both factors of the semi-direct product are infinite-dimensional.
It was recently suggested by Barnich and Troessaert [5, 6] that extended (as op-
posed to global) BMS symmetry is the true, physically relevant symmetry of asymp-
totically flat gravitational systems in four dimensions (see also footnote 17 of [7]).
This proposal is motivated by a similar symmetry enhancement occurring in two-
dimensional conformally-invariant systems: while their global symmetry algebra is
finite-dimensional, they turn out to enjoy a much richer infinite-dimensional symme-
try. This observation first appeared in a seminal paper by Belavin, Polyakov and
Zamolodchikov [8] and triggered the development of two-dimensional conformal field
theory (CFT).
Thus, the truly thrilling aspect of extended BMS symmetry is the prospect of
applying conformal field-theoretic techniques to gravitational phenomena in four di-
mensions. This reduction from four to two dimensions is reminiscent of holograms, and
indeed the notion of “holography” in quantum gravity is one of the main motivations
that led to these considerations.
1.3 Holography
The elementary concept of holography in quantum gravity is simple: it is the state-
ment that gravitational phenomena occurring in a certain space-time manifold can be
described equivalently in terms of some lower-dimensional, “dual” theory. This idea is
originally due to ’t Hooft [9] and Susskind [10], who were led to it by model-independent
considerations. In particular, holography is compatible with the Bekenstein-Hawking
entropy formula [11,12], according to which the entropy of a black hole is proportional
to the area of its horizon. (The keyword here is “area”, as opposed to the “volume”
expected on the basis of standard thermodynamics.)
AdS/CFT
In practice, the first genuine illustration of holography in a concrete model of quantum
gravity — namely string theory — was exhibited by Maldacena [13], initiating what
4
Strictly speaking, local conformal transformations do not span a group but a semi-group, and
the same applies to extended BMS. This abuse of terminology is pretty common, and it will be
inconsequential for the discussion of this introduction.
has come to be known as the Anti-de Sitter/Conformal Field Theory (AdS/CFT)

correspondence. The latter states, in a nutshell, that (quantum) gravity on a D-
dimensional asymptotically Anti-de Sitter space-time is dual to a (D − 1)-dimensional
conformal field theory. The CFT may be seen as living on the boundary of AdS, that
is, at spatial infinity, and is supposed to capture all the information on gravitational
observables. This is a statement of duality, where two completely different theories
contain the same physical information. While there is (as yet) no proof of the full
equivalence, a substantial amount of checks have been carried out to confirm that
gravity on AdS and a suitable CFT on its boundary do indeed produce the same phys-
ical predictions.
The case of a three-dimensional bulk space-time (D = 3) is especially important

for our purposes. In that context the first hint of a holographic duality actually dates
back to the eighties, when Brown and Henneaux [14] noticed that the asymptotic
symmetries of AdS3 gravity are infinite-dimensional. Analogously to the Minkowskian
setting studied two decades earlier by Bondi et al., Brown and Henneaux found that
asymptotic symmetries enhance the usual AdS3 isometry algebra so(2, 2) in an infinite-
dimensional way and span the algebra of local conformal transformations (1.8) in two
dimensions. In addition, the conserved charges generating these symmetries turn out
to satisfy a centrally extended algebra, with a central charge proportional to the AdS
radius measured in Planck units, now known as the Brown-Henneaux central charge.
The latter is the one key parameter specifying the putative two-dimensional CFT dual
to gravity on AdS3 . For instance, it was used in [15] to show that the entropy of
black holes in three dimensions [16,17] can be reproduced by a purely conformal field-
theoretic computation.
The proposal that BMS symmetry might account for gravitational physics in
asymptotically flat space-times is similar in spirit to AdS/CFT. The problem is that
most known holographic constructions rely on the key assumption that the bulk space-
time is endowed with a negative cosmological constant, i.e. that it is of the AdS type.
This leads to a natural question: how should one deal with holography in flat space?
Flat space holography

In the context of AdS/CFT, the dual theory of gravity is a CFT; in particular, even
without full knowledge of the dual theory, one can at least hope to make sense of it by
relying on the well understood consequences of conformal invariance. By contrast, in
asymptotically flat space-times, the concept of a “dual theory” is unclear, partly due to
poorly understood symmetries and partly because (in dimension four or higher) grav-
itational waves cross the null boundary of space-time; in fact, one may ask whether
flat space holography makes any sense to begin with. In view of this pessimistic omen,
a safe approach to the problem is to avoid unnecessary assumptions and rely solely
on the one known property of asymptotically flat gravity, namely BMS symmetry.
Indeed, whatever flat space holography means, if a dual theory exists, then it must be
invariant under a certain version of BMS.
The most interesting incarnation of BMS symmetry is the four-dimensional one,

since it is relevant to macroscopic gravitational waves. Unfortunately, the structure of
the extended BMS (semi-)group in four dimensions is very poorly understood (despite
recent progress [18, 19]). In short, this structure appears to be such that standard
group theory fails to apply. One is thus led to study toy models that capture the key
features of BMS symmetry without the complications of a four-dimensional world.
In this thesis we argue that the BMS group in three dimensions [20], or BMS3 ,
provides such a toy model. We shall see that it displays the extended structure (1.9) in
a simplified and controlled setting, and successfully accounts for many aspects of three-
dimensional asymptotically flat gravity, both classically and quantum-mechanically.
The BMS3 group is the main actor of this work and we will use it to develop our
intuition on flat space holography in general, including the four-dimensional case.
Holography as an Erlangen programme

Aside from the study of quantum gravity in asymptotically flat space-times, this the-
sis puts a strong accent on the relation between group theory and physics. Most,
if not all, of the topics that we will encounter in both AdS3 /CFT2 and flat space
holography follow from the properties of suitable groups — Virasoro and BMS3 , re-
spectively. For instance, the phase space of gravity will turn out to coincide with
the space of the coadjoint representation of its asymptotic symmetry group, and its
quantization will produce families of unitary representations of that group. In this
sense, three-dimensional gravity and its “holographic” properties can be reformulated
as statements in group theory.
In hindsight this observation is not too surprising. Indeed, Klein’s Erlangen pro-
gramme [21] posits that geometric statements can be recast in the language of group
theory. This point of view has led to numerous developments in mathematics through-
out the twentieth century (including e.g. the work of Poincaré on special relativity).
Since general relativity is essentially the dynamics of pseudo-Riemannian geometry,
it is natural that the programme should apply to it as well provided one identifies
the correct symmetry group. In particular, holography may sometimes be seen as an
Erlangen programme in disguise.5
1.4 BMS particles and soft gravitons

The symmetry representation theorem of Wigner [22] states that the symmetry group
of any quantum-mechanical system acts on the corresponding Hilbert space by unitary
transformations. Accordingly, a natural first step in the study of BMS symmetry is
the construction and classification of its irreducible, unitary representations.
Since the Poincaré group is a subgroup of BMS, it provides a first rough picture
of what one should expect from BMS representations. Indeed, irreducible unitary
representations of Poincaré are, by definition, particles: they are classified by their
5
See e.g. the Wikipedia page https://en.wikipedia.org/wiki/Erlangen_program.
mass and spin [23] and their Hilbert space accounts for the available one-particle states.
The BMS group is expected to generalize this notion in a way that incorporates certain
gravitational effects. It should describe the quantum states of a particle, plus some
extra degrees of freedom accounting for the fact that BMS is only an asymptotic, rather
than an exact, symmetry group. Guided by this picture, we introduce the following
terminology:
A BMS particle is an irreducible, unitary representation of the BMS group.
This thesis is devoted to the description and classification of such particles in three
space-time dimensions.
Recent developments in the study of BMS symmetry provide a simple interpreta-

tion for BMS particles. Indeed, it was observed in [19,24] that, in four dimensions, the
statement of supertranslation-invariance of the gravitational S-matrix is equivalent
to Weinberg’s soft graviton theorem [25]. This result was subsequently generalized to
include superrotations [26,27], producing a subleading term in the soft graviton expan-
sion of the S-matrix. The bottom line of these considerations is that BMS symmetry
describes the soft sector of gravity, that is, the one consisting of infinite-wavelength
gravitational degrees of freedom. The interpretation of BMS particles follows: they
are particles (in the standard sense) dressed with soft gravitons. Dressed particles are
indeed ubiquitous in the quantization of gauge theories [28–36], and this in itself is not
a new result. What is new, however, is the fact that this dressing is accounted for by
a symmetry principle that generalizes Poincaré; this is the key content of the relation
between BMS symmetry and soft theorems.
Accordingly, the classification of BMS particles that we expose in this thesis may
be thought of as a classification of all possible ways to dress a Poincaré particle with
soft gravitons. A word of caution is in order: since we will be working in three space-
time dimensions, the gravitational field will have no local degrees of freedom so there
will be no genuine gravitons. In particular, the name “soft graviton” is ambiguous, as
there is no actual graviton whose zero-energy limit would be a soft particle. However,
asymptotic symmetries precisely account for soft graviton degrees of freedom, so we
shall adopt the viewpoint that any system with non-trivial asymptotic symmetries
does indeed have non-trivial soft degrees of freedom. This amounts to using the words
“soft graviton” as a synonym for the more standard “topological” or “boundary degree
of freedom”. In particular, three-dimensional gravitational systems generally do have
highly non-trivial asymptotic symmetries [14, 37] and therefore possess soft degrees of
freedom in this sense. In this language, the statement that three-dimensional gravity
has no bulk degrees of freedom turns into the fact that the only non-trivial degrees of
freedom of three-dimensional gravity are soft.
Remark. Unitary representations of the globally well-defined BMS group (1.1) have
already been classified by McCarthy and others in [38–40], and it was indeed suggested
in [41] that BMS symmetry is relevant to particle physics in that it provides a better
definition of the notion of “particle”. However, these representations appear to miss the
fact that supertranslations create soft gravitons when acting on the vacuum, which
is crucial for the application of BMS symmetry to soft theorems. In this sense the
understanding of BMS particles in four dimensions is still an open problem; it suggests

that some extension of (1.1) is necessary if representations of BMS are to reflect reality.
We shall comment further on this issue in section 10.1.
1.5 Plan of the thesis

We now describe the topics studied in this thesis. The latter is divided in three parts,
devoted respectively to group theory in quantum mechanics, to the Virasoro group,
and to the BMS3 group.
Quantum symmetries
The first part of the thesis deals with the implementation of symmetries in quantum
mechanics through projective unitary representations, which are worked out in detail
for the Poincaré groups and the Bargmann groups. It consists of four chapters.
Quantum symmetries generally act in a projective way, which is to say that the
group operation of the underlying symmetry group is represented up to certain con-
stant phases. The presence of such phases is captured by central extensions of the
symmetry group. Accordingly, chapter 2 is devoted to central extensions and to the
more general notion of group and Lie algebra cohomology. Chapter 3 then explains
how one can build Hilbert spaces of wavefunctions on a homogeneous space endowed
with a unitary action of a symmetry group. This involves the important notion of
induced representations, which we discuss in detail.
As an application, in chapter 4 we describe the irreducible unitary representations

of semi-direct products of the general form (1.1) or (1.2). As it turns out, all these
representations are induced representations and consist of wavefunctions on a mo-
mentum orbit. This provides an exhaustive classification of unitary representations
for such groups. We illustrate these considerations with the Poincaré group (in any
space-time dimension) and with its non-relativistic counterpart, the Bargmann group,
corresponding respectively to relativistic and Galilean particles.
Finally, chapter 5 describes the relation between classical and quantum symmetries
through geometric quantization. In a nutshell this relation is obtained by defining a
space of wavefunctions on what is known as a coadjoint orbit of a symmetry group. For
semi-direct products this approach reproduces the classification of representations by
momentum orbits and leads to a group-theoretic version of the world line formalism.
Remark. The tools used in chapters 2 to 4 rely on elementary group theory; we

refer for instance to [42] for an introduction. The language of chapter 5, on the other
hand, relies more heavily on differential and symplectic geometry; see e.g. [43, 44] for
some background material.
Virasoro symmetry and AdS3 gravity

The second part of the thesis deals with the Virasoro group and its application to
three-dimensional gravity on Anti-de Sitter backgrounds. It consists of three chap-
ters. The material exposed in part II relies in a crucial way on chapter 2 and to a
lesser extent on chapter 5, but is independent of the considerations of chapters 3 and 4.
Chapter 6 is devoted to the construction of the Virasoro group as a central ex-

tension of the group of diffeomorphisms of the circle and introduces its coadjoint
representation. The latter coincides with the transformation law of stress tensors in
two-dimensional conformal field theory. In chapter 7 we classify the orbits of this ac-
tion, i.e. the coadjoint orbits of the Virasoro group, and observe that they look roughly
like infinite-dimensional cousins of Poincaré momentum orbits.
In chapter 8 we show how Virasoro symmetry emerges in AdS3 gravity with Brown-
Henneaux boundary conditions, after explaining some basic notions on asymptotic
symmetries in general. We also show that the phase space of AdS3 gravity is embedded
as a hyperplane at constant central charge in the space of the coadjoint representa-
tion of two copies of the Virasoro group. As an application we relate highest-weight
representations of the Virasoro algebra to the quantization of gravitational boundary
degrees of freedom.
BMS3 symmetry and gravity in flat space

The third and last part of the thesis is devoted to three-dimensional BMS symme-
try and contains most of the original contributions of this work. It consists of three
chapters, plus a conclusion. The material presented in part III relies crucially on the
content of parts I and II.
In chapter 9 we introduce BMS3 symmetry by way of an asymptotic analysis of

Brown-Henneaux type applied to Minkowskian backgrounds, and show that the re-
sulting algebra of surface charges has a classical central extension. We then put this
observation on firm mathematical ground by defining rigorously the BMS3 group and
its central extension. We also show that the phase space of asymptotically flat gravity
is a hyperplane at fixed central charges embedded in the space of the coadjoint repre-
sentation of BMS3 [45, 46].
Chapter 10 is devoted to the quantization of BMS3 symmetry, i.e. to its irreducible

unitary representations [47]. In the language introduced above, each representation is
a BMS3 particle. We show that the supermomentum orbits that classify these particles
coincide with coadjoint orbits of the Virasoro group and describe the resulting Hilbert
spaces of one-particle states. This leads in particular to the interpretation of BMS3
particles as particles dressed with gravitational degrees of freedom.
Finally, chapter 11 deals with rotating one-loop partition functions of quantum

fields in flat space at finite temperature. Each partition function takes the form of an
exponential of Poincaré characters. In three space-time dimensions and for a massless
field with spin two, the combination of characters is precisely such that the whole
partition function coincides with the character of a unitary representation of the BMS3
group [48, 49]. For higher spins in three dimensions we similarly obtain characters of
flat non-linear WN algebras [50]. Along the way we describe unitary representations
of these algebras [51] and show that they differ qualitatively from earlier proposals in
the literature. We end by describing certain supersymmetric extensions of the BMS3
group, their representations, and their characters.
Remark. The group-theoretic methods developed in this thesis apply to essentially

any symmetry group involving the Virasoro group. In particular one can use this
approach to derive the transformation laws of the stress tensor of a warped conformal
field theory for all values of its three central charges. Since these considerations are
somewhat out of our main line of thought we will not review them in this thesis and
refer instead to [52], where they were used to derive a Cardy-like formula for the
entropy of Rindler backgrounds.
Part I
Quantum symmetries
In this part we describe symmetry groups in quantum mechanics along

three related lines of thought. First we argue that the action of symmetry
transformations in quantum mechanics is unitary up to phases, which
leads to central extensions. Then we show how to build concrete unitary
representations using the method of induced representations, which we apply
to the description of relativistic particles. Finally we describe the general
relation between unitary representations and homogeneous spaces through
geometric quantization.
Chapter 2
Quantum mechanics and central

extensions
In this short chapter we discuss the implementation of symmetries in a quantum-

mechanical context. For definiteness and simplicity we assume throughout that these
symmetries span a Lie group. We start in section 2.1 with a brief review of the sym-
metry representation theorem of Wigner and show how quantum mechanics gives rise
to projective unitary representations. The problem of classifying such representations
then leads to sections 2.2 and 2.3, respectively devoted to Lie algebra cohomology and
group cohomology. The presentation is inspired by [53–56]; see also [57].
2.1 Symmetries and projective representations

In this section we review the interplay between quantum mechanics and symmetries.
After a brief general reminder on the formalism of quantum theory, we state the
symmetry representation theorem which justifies the study of unitary representations
of groups and Lie algebras. We also show how the fact that quantum states are rays
in a Hilbert space (rather than individual vectors) leads to projective representations,
hence to central extensions. We end with a discussion of topological central extensions,
while algebraic central extensions are postponed to section 2.2.
2.1.1 Quantum mechanics

Definition. A (complex) Hilbert space H is a vector space over C endowed with a
Hermitian form
h·|·i : H × H → C : (Φ, Ψ) 7→ hΦ|Ψi, (2.1)
p
such that the norm of a vector Ψ be hΨ|Ψi, and such that the resulting normed
vector space be complete.1 We take the scalar product (2.1) to be linear in its second
argument and antilinear in the first one.
Note that our notation is not the standard Dirac notation of bras and kets: a
vector in H is denoted as Ψ (not |Ψi), and its dual is the linear form hΨ|·i on H .
1
Recall that a metric space is complete if any Cauchy sequence converges.
15
Chapter 2. Central extensions 16
Accordingly, the Hermitian conjugate Â† of a linear operator Â is defined by
hΦ|Â† Ψi ≡ hÂΦ|Ψi for all Φ, Ψ ∈ H . (2.2)
An operator Â is Hermitian (or self-adjoint)2 if Â† = Â.
Now consider a quantum system whose space of states is a Hilbert space H . A

pure quantum state of the system is a ray in H , that is, a one-dimensional subspace
[Ψ] = {zΨ|z ∈ C} (2.3)
where Ψ is some non-zero state vector. The vanishing vector does not represent a
quantum state, so the set of mutually inequivalent pure states is the projective space
PH = (H \0)/C. It is the set of one-dimensional subspaces of H . Stated differently,
the set of distinct states in H is the quotient of the unit sphere in H by the equivalence
relation
Ψ ∼ eiθ Ψ for all θ ∈ R. (2.4)
We shall denote by [Ψ] the resulting equivalence class of Ψ. For example, in a two-level
system where H = C2 , the set of inequivalent states is CP 1 ∼ = S 2.
Now let the system be in a state [Ψ]. If Â is an observable and if λ is one of its
eigenvalues with eigenvector Φ say, the probability of finding the value λ is
|hΦ|Ψi|2
Prob(λ, Â, [Ψ]) = . (2.5)
hΦ|ΦihΨ|Ψi
(We are assuming for simplicity that the eigenvalue λ is not degenerate.) Note that
this expression is independent, as it should, of the choice of both the representative Ψ
of the state [Ψ], and the eigenvector Φ.
Remark. In quantum mechanics, one generally assumes that the Hilbert space is
separable, i.e. that it admits a countable basis. Any such space is isometric to the space
`2 (N) of square-integrable sequences of complex numbers — so there really exists only
one infinite-dimensional separable Hilbert space. This is not to say that all separable
Hilbert spaces describe the same quantum system, because the definition of a system
also involves the set of observables that act on it — and identical Hilbert spaces may
well come with very different operator algebras.
2.1.2 Symmetry representation theorem

Symmetry groups
A symmetry is a transformation of a system that leaves it invariant. In particular,
the set of symmetries of a system always contains the identity transformation, and
any symmetry transformation is invertible. In addition the composition of any two
2
We will not take into account issues related to the domains of operators.
symmetry transformations is itself a symmetry, and composition is associative. Put

together, these properties imply that
the set of symmetries of any system forms a group.
Accordingly, the framework suited for the study of symmetries is group theory.
In this thesis we will be concerned with Lie groups, consisting of symmetry transfor-
mations that depend smoothly on a certain number of real parameters. This number is
the dimension of the group. In part I of the thesis, all Lie groups are finite-dimensional.
Remark. The notion of symmetry can be relaxed in such a way that not all pairs of
symmetry transformations are allowed to be composed together. The resulting set of
symmetry transformations then spans a groupoid rather than a group (see e.g. [58,59]).
This relaxed notion of symmetry is relevant to gauge theories [60], and in particular to
BMS symmetry in four dimensions [18]. However, standard group theory suffices for
all symmetry considerations in three-dimensional gravity (and in particular for BMS3 ),
so we will not deal with groupoids in this thesis.
Symmetries in quantum mechanics

Consider a quantum Hilbert space of states H . In these terms a symmetry is a bijec-
tion PH → PH : [Ψ] 7→ S([Ψ]) that preserves the probabilities (2.5). Equivalently,
if we represent rays in H by normalized vectors subject to the identification (2.4), a
symmetry transformation S must be such that
|hΦ|Ψi| = |hΦ0 |Ψ0 i| (2.6)
for all normalized vectors Φ, Ψ, Φ0 , Ψ0 such that Φ0 ∈ S([Φ]) and Ψ0 ∈ S([Ψ]). The
key result on symmetries in quantum mechanics is the following [22]:
Symmetry representation theorem. Let S : PH → PH be an invertible trans-

formation satisfying property (2.6). Then it takes the form S([Ψ]) = [Û · Ψ], where Û
is either a linear, unitary operator so that
Û · (λΦ + µΨ) = λ Û · Φ + µ Û · Ψ and hÛ · Φ|Û · Ψi = hΦ|Ψi ,
or an antilinear, antiunitary operator so that
Û · (λΦ + µΨ) = λ̄ Û · Φ + µ̄ Û · Ψ and hÛ · Φ|Û · Ψi = hΨ|Φi
for all λ, µ ∈ C and all Φ, Ψ ∈ H . A proof of this theorem can be found in chapter 2
(appendix A) of [53].
Note that symmetries represented by antiunitary operators only arise when the
symmetry group is disconnected. For example, in Lorentz-invariant theories, time-
reversal is always represented in an antiunitary way (see e.g. [61]). In this work
we will restrict attention to connected symmetry groups, in which case all symmetry
operators are linear and unitary. In particular they satisfy Û † = Û −1 , where Hermitian
conjugation is defined by (2.2).
2.1.3 Projective representations

The symmetry representation theorem implies that all (connected) symmetry groups
are represented unitarily in a quantum-mechanical system, and thus motivates the
study of unitary representations in general. Let us first recall the basics:
Definition. A representation of a group G in a vector space H is a homomorphism3
T : G → GL(H ) : g 7→ T [g]
where GL(H ) is the group of invertible linear transformations of H . When H is a

Hilbert space, the representation is unitary if T [g] is a unitary operator for each g ∈ G.
In quantum mechanics the notion of symmetry as a transformation that satisfies

(2.6) leads to a key subtlety. Let us call T [f ] the unitary operator that represents a
symmetry transformation f belonging to some group G. Then, because a quantum
state is really an equivalence class (2.3) of vectors in H , there is no need to require T
to be a homomorphism; rather, all we need is that the ray of T [f ] · T [g] · Φ coincides
with that of T [f · g] · Φ (for all f, g ∈ G and any Φ ∈ H ). Accordingly, T must really
be a unitary representation up to a phase,
T [f ] · T [g] = eiC(f,g) T [f · g] for f, g ∈ G, (2.7)
where C is some real function on G × G. In more abstract terms, T must define a

group action on the projective space PH , which is to say that the map
[T ] : G → GL(H )/C∗ : f 7→ T [f ]

(2.8)
is a homomorphism. Here GL(H )/C∗ is the projective group of H , i.e. the quotient
of the linear group of H by its normal subgroup consisting of multiples of the iden-
tity. For any operator O in GL(H ), the symbol [O] denotes its class in the projective
group. Throughout this thesis, any map T satisfying this property will be called a pro-
jective representation. In quantum mechanics, symmetries are represented by unitary
projective representations, i.e. projective representations whose operators are unitary.
From now on, if we wish to stress that a representation is not projective, we will
call it exact. Quantum mechanics tells us that exact representations are overrated: the
truly important ones are generally projective. This seemingly anecdotal observation is
at the core of the richest aspects of the representation theory of the Virasoro algebra,
and it will also play a key role for BMS3 particles. For instance, all interesting two-
dimensional conformal field theories are such that the conformal group is represented
projectively in their Hilbert space, and how exactly this phenomenon takes place is
measured by the central charge. For this reason, this whole chapter is devoted to the
various ways in which projective effects occur; they are accounted for by group and
Lie algebra cohomology.
3
Throughout this thesis representations of groups are denoted by the letters R, S, T , etc. The
letter G denotes a group whose elements are written f , g, h, etc. The identity in G is denoted e.
Remark. Since we are focussing on Lie groups, the representations of interest are
continuous in the sense that the map G × H → H : (f, Ψ) 7→ T [f ] · Ψ is continuous.
From now on it is understood that all representations are continuous.
2.1.4 Central extensions

The function C appearing in (2.7) is not completely arbitrary. Indeed, the product
(2.7) must be associative in the sense that T [f ] · (T [g] · T [h]) = (T [f ] · T [g]) · T [h] for
all group elements f, g, h, so that
C(f, gh) + C(g, h) = C(f g, h) + C(f, g) for all f, g, h ∈ G. (2.9)
Any function C : G × G → R satisfying this requirement is known as a (real) two-

cocycle, and the condition itself is known as the cocycle condition. Given any such
function one can define a new group
b ≡G×R
G (2.10)
whose elements are pairs (f, λ), endowed with a group operation

(f, λ) · (g, µ) = f · g, λ + µ + C(f, g) . (2.11)
The group (2.10) is called a central extension of the group G. We will study this
notion in much greater detail in section 2.3. For now let us only work out the basic
consequences of this structure and its relation to representation theory.
Projective versus exact representations

Property (2.7) says that T is an exact unitary representation of the centrally extended
group (2.10), provided one represents the pair (f, λ) by eiλ T [f ]. In other words, exact
representations are not overrated after all: we may view any projective representation
of G as an exact (i.e. non-projective) representation of a central extension G b of G, and
the problem of classifying projective unitary representations of G boils down to that
of classifying exact unitary representations of its central extensions.
The question then is whether G admits central extensions to begin with. For any
group, an obvious type of central extension always exists. Namely, suppose K is a real
function on G and define C : G × G → R by
C(f, g) ≡ K(f g) − K(f ) − K(g). (2.12)
This automatically satisfies condition (2.9). A two-cocycle of that form is said to be

trivial. In particular, if the cocycle in (2.7) is trivial, it can be absorbed by defining
T̃ [f ] ≡ eiK(f ) T [f ], which is an exact representation of G. Thus, what we wish to know
is not quite whether G admits two-cocycles at all (since trivial ones are always avail-
able), but rather whether it admits non-trivial two-cocycles. If yes, it admits genuine
projective representations, whose phases cannot be absorbed by a mere redefinition.
This question leads to group (and Lie algebra) cohomology, studied in detail in
sections 2.2 and 2.3. For now we simply point out that central extensions may arise
via two distinct mechanisms. The first is algebraic in that it follows from the local
group structure of G, or equivalently from the commutation relations of its Lie algebra.
In short, in some cases, the Lie algebra g of G can be enlarged into a bigger algebra
g which contains extra generators commuting with those of g (see eq. (2.27) below).
b
The group corresponding to this enlarged algebra then is a central extension of G.
The second mechanism is topological in the sense that it is due to the global structure
of G. We now describe this topological mechanism in some more detail.
2.1.5 Topological central extensions

If the group G is not simply connected (i.e. its fundamental group is non-trivial),
there exist closed paths in G that cannot be continuously deformed into a point.
Let γ : [0, 1] → G be such a path, starting and ending at some group element f
so that γ(0) = γ(1) = f . Suppose we are given a (continuous) projective unitary
representation T of G, and consider the path
T ◦ γ : [0, 1] → GL(H ) : t 7→ T [γ(t)]
in the space of unitary operators on H . Since T is projective, the fact that γ is a

closed path does not imply that T ◦ γ is closed: in general T [γ(0)] and T [γ(1)] differ
by a γ-dependent phase, T [γ(1)] = eiφ(γ) T [γ(0)].
Owing to the fact that the map T is continuous, the phase φ(γ) only depends on
the homotopy class of γ. In addition, if γ1 and γ2 are two closed paths starting at f , we
can concatenate them into a single path γ1 · γ2 (which is γ1 at double speed followed by
γ2 at double speed); the phase φ must be compatible with this operation in the sense
that eiφ(γ1 ) · eiφ(γ2 ) = eiφ(γ1 ·γ2 ) . Thus, any one-dimensional unitary representation of the
fundamental group of G, multiplying an exact unitary representation of G, produces
a projective unitary representation of G.
This is the topological notion of central extensions that we wanted to exhibit: if

G is multiply connected, it admits genuine projective representations (whose phases
cannot be removed by redefinitions) due to one-dimensional unitary representations
of its fundamental group.4 Projective representations of that type may equivalently
be seen as exact representations of the universal cover G e of G, which is the unique
connected and simply connected group locally isomorphic to G.
Remark. One might be worried by the fact that only one-dimensional unitary repre-
sentations of the fundamental group are allowed to appear in this construction. Indeed,
if the fundamental group was non-Abelian, it would generally admit no non-trivial
one-dimensional unitary representation. Fortunately, it turns out that the fundamen-
tal group of any finite-dimensional Lie group is a discrete commutative group, whose
irreducible unitary representations are necessarily one-dimensional.
4
Beware: a manifold being multiply connected means that it has a non-trivial fundamental group,
and not that it has several connected components.
Figure 2.1: The group U(1) is diffeomorphic to a circle S 1 , whose universal cover is the
real line R. The projection R → S 1 ∼= R/Z is obtained by identifying points of R that
differ by some periodicity, typically θ ∼ θ + 2π. In particular, paths in R which are
not closed may be projected on closed paths in S 1 . As an application we can picture
topological projective representations: if T is projective and if γ is a closed path in
the circle, the sequence T [γ(t)] may not be a closed path in the space of operators.
Rotations and anyons

The simplest example of topological projective representations arises with the group
U(1). The latter is diffeomorphic to a circle and has a fundamental group isomorphic
to Z (see fig. 2.1). Any exact irreducible, unitary representation of U(1) takes the form
T : U(1) → C∗ : θ 7→ eisθ (2.13)
where θ is identified with θ + 2π, as a consequence of which the “spin” s is an integer.
For example, when s = 2, a rotation by θ = π is represented by the identity. (We will
see in section 4.3 that the label s actually is the spin of a particle in certain representa-
tions of the Poincaré groups.) But there is a subtlety: U(1) is multiply connected and
admits topological projective representations, which from the viewpoint of quantum
mechanics are just as acceptable as exact ones. For example, the map (2.13) with
s = 1/2 definitely isn’t an exact representation because a full rotation by 2π is now
represented by an inversion, T [2π] = eiπ = −1. Nevertheless, in quantum mechanics,
the vectors Ψ and T [2π] · Ψ define the same state by virtue of the identification (2.4),
so in this sense T [2π] acts as an “almost-identity” operator. More generally, formula
(2.13) is a projective representation of U(1) for any real value of the spin s.
The example just described occurs in Nature. Indeed, fermions provide a well-
known example of projective representations, as already suggested above by the case
s = 1/2. By the spin-statistics theorem, all fermions have half-integer spins, and
therefore transform according to a projective representation of the Lorentz group.
The latter is multiply connected (its fundamental group is Z2 ), which is why it admits
projective representations in the first place. We will return to the representation theory
of the Lorentz group (as a subgroup of Poincaré) in much greater detail in section 4.2.
In the cases where arbitrary real values of spin are allowed by quantum mechanics,
as for example in three space-time dimensions, the particles whose spin is neither an
integer nor a half-integer are known as anyons. We will encounter this phenomenon
in section 10.1 when dealing with BMS3 particles.
2.1.6 Classifying projective representations

Given a group G, suppose we wish to find all its projective unitary representations.
The above considerations provide an algorithm that allows us, in principle, to solve
that problem:
• First find the universal cover G

e of G to take care of topological central extensions.
• Then find the most general central extension G of G

e in order to take care of
be
differentiable central extensions. (We will deal with the actual definition of
these extensions in the next section.)
• Finally, consider an exact unitary representation of G; any projective unitary

be
representation of G may be seen as a representation of that type.
Thus we now have a systematic procedure allowing us to build arbitrary projective

unitary representations of symmetry groups in quantum mechanics. We will apply it
later to the Virasoro algebra (section 8.4) and the BMS3 group (section 10.1), where
central extensions play a crucial role.
2.2 Lie algebra cohomology

This section is devoted to a thorough investigation of the concept of central extensions
at the Lie-algebraic level. In fact, we shall describe the more general framework of Lie
algebra cohomology and we will show how statements on algebraic central extensions
can be recast in that language. The group-theoretic analogue of this construction is
relegated to section 2.3.
2.2.1 Cohomology
Let g be a Lie algebra with Lie bracket [·, ·]. We recall that a representation of g in
V is a linear map T : g → End(V) such
a vector space that T [X] ◦ T [Y ] − T [Y ] ◦
T [X] = T [X, Y ] for all Lie algebra elements X, Y . 5
Definition. Let k be a non-negative integer, T a representation of g in V. Then a

V-valued k-cochain on g is a continuous, multilinear, completely antisymmetric map6
c : g × · · · × g → V : (X1 , ..., Xk ) 7→ c(X1 , ..., Xk ). (2.14)

| {z }
k times
In other words, a V-valued k-cochain on g is a k-form on g with values in V; note

that 0 ≤ k ≤ dim(g). A zero-cochain on g is a vector in V while a dim(g)-cochain
is a volume form on g. We denote the space of V-valued k-cochains on g by C k (g, V)
dim(g)
and we define the associated cochain complex C ∗ (g, V) ≡ ⊕k=0 C k (g, V). The latter
is sometimes called the Chevalley-Eilenberg complex.
5
Throughout this thesis the elements of a Lie algebra g will be denoted as X, Y , etc. Represen-
tations of Lie algebras will be denoted by script capital letters such as R, S , T .
6
Cochains on Lie algebras will be denoted by lowercase sans serif letters such as c, s, etc.
Definition. The Chevalley-Eilenberg differential d : C ∗ (g, V) → C ∗ (g, V) is defined

by dim(g) linear maps
dk : C k (g, V) → C k+1 (g, V) : c 7→ dk c
where k runs from 0 to dim(g) − 1 and the (k + 1)-cochain dk c is given by

X
(−1)i+j−1 c [Xi , Xj ], X1 , ..., X

(dk c)(X1 , ..., Xk+1 ) ≡ ci , ..., X
cj , ..., Xk+1
1≤i<j≤k+1
X
(−1)i T [Xi ] · c X1 , ..., X

+ ci , ..., Xk+1 (2.15)
1≤i≤k+1
for all X1 , ..., Xk+1 in g; the hat denotes omission. Note that the representation T of
g in V appears explicitly in this definition. In particular, when T is trivial, formula
(2.15) simplifies since its last line disappears.
Cocycles and coboundaries

Using the fact that T is a representation, one can verify that the Chevalley-Eilenberg
differential (2.15) is nilpotent:
dk ◦ dk−1 = 0 ∀ k = 0, ..., dim(g) (2.16)
where it is understood that the “extreme differentials” are d−1 : 0 → V : 0 7→ 0 and

ddim g : C dim g (g, V) → 0 : c 7→ 0. Accordingly, one adapts the standard terminology
of differential forms to cochains on a Lie algebra: a k-cocycle is a k-cochain c such
that dk c = 0; a k-coboundary is a k-cochain c of the form c = dk−1 b, where b is some
(k − 1)-cochain. By virtue of property (2.16), one has Im(dk−1 ) ⊆ Ker(dk ) for each
k (any coboundary is a cocycle). One can therefore define the k th cohomology space
of g with coefficients in V as the quotient of the space of k-cocycles by the space of
k-coboundaries:
Hk (g, V) ≡ Ker(dk )/Im(dk−1 ). (2.17)
A k-cocycle is said to be trivial if its equivalence class vanishes in Hk , i.e. if the cocycle
is a coboundary; the cocycle is non-trivial otherwise. When V = R with T the trivial
representation of g, we write Hk (g, R) ≡ Hk (g).
Isomorphic Lie algebras have the same cohomology for any choice of the represen-
tation T . Thus, cohomology is a way to associate invariants with Lie algebras: if two
algebras have different cohomology spaces, then they cannot be isomorphic. This is
analogous to, say, de Rham cohomology in differential geometry, as manifolds with
different de Rham cohomologies cannot be diffeomorphic.
Low degree cohomologies

There is a simple interpretation for the lowest cohomology spaces. For example, zero-
cocycles are vectors v ∈ V that are invariant under g in the sense that
T [X] · v = 0 for all X ∈ g , (2.18)

so the zeroth cohomology space of g classifies the invariants of the representation T .

Similarly, one-cocycles are known as derivations of g and are classified by the first
cohomology space H1 (g, V). In the particular case where T is trivial and V = R,
a one-cocycle is a linear map c : g → R such that c([X, Y ]) = 0 for all Lie algebra
elements X, Y . Hence the first real cohomology space of g can be written as
H1 (g) ∼
= g/[g, g] , (2.19)
which motivates the following definition:
Definition. A Lie algebra g is perfect if g = [g, g], i.e. if any Lie algebra element can
be written as the bracket of two other elements.
It follows from (2.19) that g is perfect if and only if H1 (g) vanishes. We will use
this property in section 2.2.2 when defining central extensions.
By the definitions above, a two-cochain is an antisymmetric map c : g × g → V. It

is a coboundary if
(2.15)
c(X, Y ) = (d1 k)(X, Y ) = k([X, Y ]) − T [X] · k(Y ) + T [Y ] · k(X) (2.20)
for some one-cochain k; and it is a cocycle if
c([X, Y ], Z) + c([Y, Z], X) + c([Z, X], Y ) =

(2.21)
= T [X] · c(Y, Z) + T [Y ] · c(Z, X) + T [Z] · c(X, Y ).
As we shall see shortly, when T is trivial, a two-cocycle defines a central extension of g.

Thus the second cohomology of g classifies its extensions. More generally, cohomology
may be seen as a measure of flexibility: Lie algebras with high-dimensional cohomology
groups can be “deformed” in many inequivalent ways; by contrast, Lie algebras with
trivial cohomology are “rigid” in the sense that any deformation is equivalent to no
deformation at all.
Remark. Here we have been using the word “deformation” in a vague way, but there
is an exact definition of the notion of deformations. Namely, a (true) deformation of
a Lie algebra g is a Lie algebra g̃ that coincides with g as a vector space, but whose
brackets are
˜[X, Y ˜] = [X, Y ] + c(X, Y ) (2.22)
where [·, ·] is the bracket in g while c is a g-valued two-cocycle on g,7 such that the
image of c belongs to its kernel. The latter condition means that c X, c(Y, Z) = 0 for
all Lie algebra elements X, Y, Z; together with the fact that c is a cocycle, this ensures
that (2.22) is a Lie bracket.
Examples
For finite-dimensional semi-simple Lie algebras, cohomology is trivial:
7
It is understood that the relevant representation of g in this case is the adjoint, T [X]·Y ≡ [X, Y ].
Whitehead’s lemma. Let g be a finite-dimensional semi-simple Lie algebra, T an

irreducible, finite-dimensional representation of g in a space V. Then
Hk (g, V) = 0 for all k > 0. (2.23)
Despite this result, examples of non-trivial cohomologies do exist in physics. For

instance, let c be an arbitrary non-vanishing antisymmetric bilinear form on R2 , and
view the latter as an Abelian Lie algebra. Then c defines a non-trivial, real-valued
two-cocycle on R2 , so the real-valued second cohomology of R2 is non-trivial; in fact
one can prove that
H2 (R2 ) ∼
= R. (2.24)
We shall see below that this property is related to the (three-dimensional) Heisenberg
algebra, which is crucial for quantum mechanics. Other important examples of alge-
bras with non-trivial cohomology spaces include the Galilei algebra (section 4.4), the
Virasoro algebra (chapter 6) and the bms3 algebra (chapter 9).

Definition. Let g be a (real) Lie algebra and let c ∈ C 2 (g, R) be a real two-cocycle
on g. Then c defines a central extension b
g of g, which is a Lie algebra whose underlying
vector space
g=g⊕R
b (as vector spaces) (2.25)
is endowed with the centrally extended Lie bracket

(X, λ), (Y, µ) ≡ [X, Y ], c(X, Y ) . (2.26)
In particular, elements of bg are pairs (X, λ) where X ∈ g and λ ∈ R, so that R is an

Abelian subalgebra of b g. The bracket (2.26) satisfies the Jacobi identity on account
of the fact that c is a two-cocycle with respect to a trivial representation of g (so that
the right-hand side of eq. (2.21) vanishes).
In (2.26) we displayed the definition of central extensions in intrinsic terms thanks

to the two-cocycle c. The same definition can be written in terms of Lie algebra
generators: let {ta |a = 1, ..., n} be a basis of g with brackets [ta , tb ] = fab c tc . Then a
g of g is a Lie algebra generated by the basis elements Ta ≡ (ta , 0)
central extension b
together with a central element Z = (0, 1), whose Lie brackets read
[Ta , Tb ] = fab c Tc + cab Z (2.27)
where cab = c(ta , tb ), while all brackets with Z vanish. The cocycle condition on c then
becomes the requirement
fab d cdc + fbc d cda + fca d cdb = 0
for the coefficients cab . Note that this construction can be readily generalized to
g = g ⊕ RN , in which case there are N central generators
multiple central extensions b
Z1 , ..., ZN .
Non-trivial central extensions

When the two-cocycle c is trivial in the sense of Lie algebra cohomology, it takes the
form (2.20) in terms of some one-cocycle k and the map

g→b g : X 7→ X, k(X) (2.28)
is an injective homomorphism of Lie algebras. The central extension is then said to be

trivial : the cocycle c can be absorbed by the “redefinition” (2.28), and b
g is isomorphic
to the direct sum g ⊕ R as a Lie algebra. By contrast, when c is non-trivial, it defines
a non-zero element in the second cohomology space H2 (g); such a two-cocycle cannot
be removed by a mere redefinition, and the central extension is non-trivial.
For example, as on page 25, consider the Abelian Lie algebra R2 and let c be a
non-zero antisymmetric bilinear form on R2 . We then define the three-dimensional
Heisenberg algebra as the algebra R3 = R2 ⊕ R whose elements are pairs (X, λ),
endowed with the Lie bracket (2.26). Since c is non-trivial, so is the central extension.
If we choose a basis {Q, P } of R2 such that c(Q, P ) = 1 and if we call Z the central
element (0, 1), the commutation relations of the Heisenberg algebra take the form
[Q, P ] = Z. (2.29)
Property (2.24) says that there is only one linearly independent central extension of
R2 , i.e. that Heisenberg algebras built using different (non-zero) two-cocycles c are
mutually isomorphic. This can be generalized to higher dimensions: by seeing R2n as
an Abelian Lie algebra and taking c an arbitrary non-zero 2n-form on R2n , the Lie
algebra defined by the bracket (2.26) is the (2n + 1)-dimensional Heisenberg algebra.
Universal central extensions

It is important to know how many inequivalent central extensions an algebra may
possess. This leads to the following notion:
Definition. A central extension b g of g is universal if, for any other central extension
g0 of g, there exists a unique isomorphism of Lie algebras b
b g0 ∼
=bg.
As it turns out, any perfect Lie algebra admits a universal central extension. By
virtue of (2.19), this is to say that any algebra such that H1 (g) = 0 admits a universal
central extension. For example we will see in chapter 6 that the Virasoro algebra is
the universal central extension of the Lie algebra of vector fields on the circle.
2.3 Group cohomology

This section is devoted to the group-theoretic analogue of the considerations of the pre-
vious pages. We start by discussing generalities on group cohomology before focussing
on central extensions of groups.
2.3.1 Cohomology
Let G be a Lie group, T : G → GL(V) a representation of G in a vector space V.
Definition. Let k ≥ 0 be an integer. A V-valued k-cochain on G is a smooth map8
C : G × · · · × G → V : (g1 , ..., gk ) 7→ C(g1 , ..., gk ). (2.30)

| {z }
k times
Note that, in contrast to the Lie-algebraic definition (2.14), there is no restriction

on k. The new ingredient in the group-theoretic context is the requirement that the
map (2.30) be smooth. As in the case of Lie algebras, we denote by C k (G, V) the
vector space of V-valued k-cochains on G and we let C ∗ (G, V) = ⊕+∞ k
k=0 C (G, V) be the
associated cochain complex. The space of zero-cochains is just V.
Definition. The differential d : C ∗ (G, V) → C ∗ (G, V) is defined by the maps
dk : C k (G, V) → C k+1 (G, V) : C 7→ dk C
where k ∈ N and the (k + 1)-cochain dk C is given by
(dk C)(g1 , ..., gk+1 ) ≡ T [g1 ] · C(g2 , ..., gk+1 ) + (−1)k+1 C(g1 , ..., gk )
k
X (2.31)
+ (−1)i C(g1 , ..., gi gi+1 , ..., gk+1 )
i=1
for all g1 , ..., gk+1 in G.
The differential (2.31) satisfies the key property (2.16), so the usual machinery of
homological algebra applies: one defines a k-cocycle as a closed k-cochain, that is,
a cochain C such that dk C = 0. One also defines a k-coboundary to be an exact k-
cochain, i.e. one that can be written as the differential of a (k − 1)-cochain. As before
any coboundary is trivially a cocycle, so one defines the k th cohomology space of G with
values in V as the quotient of the space of k-cocycles by the space of k-coboundaries:
Hk (G, V) ≡ Ker(dk )/Im(dk−1 ).
A k-cocycle is trivial if its class in Hk (G, V) vanishes; it is non-trivial otherwise. When

V = R with T the trivial representation, we write Hk (G, R) ≡ Hk (G).
8
Cochains on a group will be denoted by capital sans serif symbols such as C, S, etc.
Interpretation
As in the case of Lie algebras, cohomology spaces are invariants that measure the flex-
ibility of a group structure; isomorphic Lie groups have the same cohomology. This
interpretation is simplest to illustrate with the cohomology spaces of lowest degree.
A V-valued zero-cocycle on G is a vector v ∈ V such that (d0 v)(f ) = T [f ]·v −v = 0

for any group element f . Accordingly, the zeroth cohomology space of G classifies vec-
tors v ∈ V that are left invariant by G. This is the group-theoretic analogue of (2.18).
A V-valued one-cocycle is a (smooth) map S : G → V satisfying the property
S(f g) = T [f ] · S(g) + S(f ) ∀ f, g ∈ G. (2.32)
Given a one-cocycle S, one defines the associated affine module as the space V ⊕ R
acted upon by the following representation Tb of G:

Tb [f ] · (v, λ) ≡ T [f ] · v + λ S(f ), λ . (2.33)
The cocycle condition (2.32) ensures that Tb is indeed a representation. In addition

one can show that affine modules defined using different one-cocycles are equivalent
if (and only if) their cocycles differ by a coboundary. Thus H1 (G, V) classifies affine
G-modules based on V. For example, in section 6.3 we will see that the Schwarzian
derivative is a one-cocycle on the group of diffeomorphisms of the circle; this is why we
denote the cocycle in (2.33) by S. The corresponding affine module will be the coad-
joint representation of the Virasoro group and the parameter λ left invariant by (2.33)
will be a Virasoro central charge. More generally one can think of the term λS[f ] in
(2.33) as an anomaly that adds an inhomogeneous term to the otherwise homogeneous
transformation law of v under G.
Two-cocycles lead to the notion of group extensions; in particular, when V = R

with T the trivial representation, H2 (G) classifies central extensions of G. Indeed,
when C is a real two-cocycle on G, the requirement d2 C = 0 becomes the cocycle
condition (2.9); the central extension is trivial when C is a coboundary, i.e. if it takes
the form (2.12) for some one-cochain K. We will return to central extensions of groups
in section 2.3.2.
Relation to Lie algebra cohomology

One may ask how group and Lie algebra cohomology are related. The following result
provides a first answer:
Proposition. Let G be a Lie group, g its Lie algebra. Let V be a vector space, T a
smooth representation of G in V, and T the representation of g corresponding to T
by differentiation. Then, for any non-negative integer k, there is a homomorphism
Hk (G, V) → Hk (g, V) : [C] 7→ [δC] (2.34)

given by
" #
∂k X
i1 ...ik C eti1 Xi1 , ..., etik Xik

δC(X1 , ..., Xk ) ≡
∂t1 ...∂tk 1≤i1 <...<ik ≤k t1 =0,...,tk =0
for all X1 , ..., Xk in g, with eX the exponential of X ∈ g and i1 ...ik the Levi-Civita
symbol with k indices (and 12...k ≡ +1). For k = 2 this can be rewritten as
∂2 h tX sY
sY tX
i
δC(X, Y ) = C e ,e − C e ,e . (2.35)
∂t ∂s t=0, s=0
The fact that (2.34) is a homomorphism ensures that, if δC is a non-trivial cocycle,

then C itself is non-trivial. The converse is not true since the map need not be injec-
tive: a non-trivial cocycle C may well be such that δC is trivial.
We will use formula (2.35) in section 6.2 to relate the Virasoro algebra to the
Virasoro group. The key point here is that any differentiable group cocycle C admits
an algebraic analogue δC. The converse problem is to start from a Lie algebra cocycle,
say c, and ask whether there exists a group cocycle whose differential is c. This is
the problem of integrating Lie algebra cocycles to group cocycles, and it is generally
much more complicated than differentiation. However, for “sufficiently connected” Lie
groups, the Van Est theorem states that integration is trivial because group and Lie
algebra cohomologies coincide (see e.g. [56]). In particular, when the universal cover
of a group is homotopic to a point, the cohomology of the universal cover coincides
with that of the Lie algebra.

Here we return in more detail to the notion of centrally extended groups, already
outlined around (2.11). For simplicity we deal only with simply connected groups, so
as to avoid the topological complications of section 2.1.5. Including these subtleties
would lead to a definition of central extensions somewhat more general (see e.g. [56])
than the one given here:
Definition. Let G be a Lie group, C a real two-cocycle on G. Then the associated

b is topologically a product G × R whose elements are pairs
centrally extended group G
(f, λ) with f ∈ G and λ ∈ R, endowed with a group operation (2.11).
It is straightforward to generalize this definition to the case where R is replaced by

an arbitrary (additive) Abelian group such as RN .
Non-trivial central extensions

As in the Lie-algebraic case, a central extension of G is trivial if the two-cocycle C
defining the group operation (2.11) is a coboundary (2.12) for some one-cochain K.
Then the map G → G b : f 7→ f, K(f ) is an injective homomorphism whose Lie-
algebraic analogue is (2.28), and G
b is isomorphic, as a group, to the direct product
G × R. Thus any trivial central extension can be absorbed by a redefinition of the
group, and is irrelevant as regards projective representations. By contrast, when the

cohomology class of C is a non-zero vector in H2 (G), the central extension cannot be
removed by a redefinition and is said to be non-trivial.
Example. Let us find the group corresponding to the (2n + 1)-dimensional Heisen-
berg algebra. Consider the Abelian additive group G = Rn × Rn (whose elements are
pairs of column vectors (α, β)) and define the Heisenberg group as
  
 1 αt λ 
n
G≡
b  0 In β α, β ∈ R , λ ∈ R
 (2.36)
0 0 1
 
where In denotes the n × n identity matrix and αt is the transpose of α. The group
operation in G
b is given by matrix multiplication and can be written as
(α, β, λ) · (α0 , β 0 , λ0 ) = α + α0 , β + β 0 , λ + λ0 + αt · β 0

(2.37)
where αt · β 0 ≡ αi β 0i is the Euclidean scalar product of α and β 0 . Thus the Heisenberg

group is a central extension of R2n defined by the two-cocycle
C (α, β), (α0 , β 0 ) = αt · β 0 .

(2.38)
By differentiation, one can associate with C a Lie algebra cocycle given by (2.35). For
example, when n = 1 (and writing elements of the Lie algebra R2 as pairs X = (x, y)),
0 0
(2.35) ∂ 2
δC (x, y), (x , y ) = (tx · sy 0 − sx0 · ty)|t=0, s=0 = xy 0 − yx0 .
∂t ∂s
This is a non-zero antisymmetric bilinear form on R2 , hence defining the Heisenberg
algebra of (2.29). Note that this is an example of “cocycle integration”: we have found
the explicit group two-cocycle whose differential defines the Heisenberg Lie algebra.
Universal central extensions

Universal central extensions of groups can be defined exactly as for Lie algebras. A
central extension G b0 of G by A,
b of G is universal if, for any other central extension G
there exists a unique isomorphism G b→G b0 .
As in the algebraic case, there is a simple criterion for knowing when a group
admits a universal central extension. A group is said to be perfect if it coincides with
the group of its commutators, i.e. if any f ∈ G can be written as f = ghg −1 h−1
for some g, h ∈ G. It turns out that any perfect group admits a universal central
extension. In chapters 6 and 9 we will see that both Diff(S 1 ) and BMS3 are perfect
groups, so that their central extensions are universal.
Chapter 3
Induced representations
In the previous chapter we learned how to deal with projective representations: given
a symmetry group, we are to find its universal cover and its most general central ex-
tension. Exact representations of this central extension then account for all projective
representations of the original group. The remaining problem then is to write down
explicit representations, so our goal in this chapter is to build Hilbert spaces of wave-
functions acted upon by a group of unitary transformations. Guided by group actions
on homogeneous spaces, we will be led to the method of induced representations. Their
basic principle is very simple: starting from a representation of some subgroup H of a
group G, one induces a representation of G that acts on wavefunctions which live on
the quotient space G/H.
Induced representations are ubiquitous in mathematics and physics:
• The irreducible unitary highest-weight representations of any compact, simple

Lie group are induced from those of its maximal torus, i.e. its largest Abelian
subgroup (whose Lie algebra is the Cartan subalgebra).
• Highest-weight representations of sl(2, R) and of the Virasoro algebra are induced

from representations of their u(1) subalgebra generated by L0 (see section 8.4).
• All irreducible unitary representations of the Euclidean groups, the Bargmann

groups, the Poincaré groups and the BMS3 group are induced from those of their
translation subgroups combined with “little groups” (see chapters 4 and 10).
The plan of this chapter is as follows. In section 3.1 we review some basics of
measure theory and the ensuing construction of Hilbert spaces of square-integrable
wavefunctions. Section 3.2 is concerned with measures on homogeneous spaces and
introduces quasi-regular representations — the simplest examples of induced represen-
tations. In section 3.3 we display the basic formulas of induced representations and
list some of their elementary properties. Along the way we define a basis of plane
waves, later to be interpreted as particles with definite momentum. This basis is then
used in section 3.4 to compute characters. Finally, section 3.5 is devoted to systems
of imprimitivity. All these notions are crucial prerequisites for chapter 4.
31
Chapter 3. Induced representations 32
It would be illusory to present a complete account of the rich theory of induced rep-
resentations, so we refer to Barut and Raczka [62] or Mackey [63] for a more thorough
exposition. For some background on measure theory, see e.g. [64, 65].
3.1 Wavefunctions and measures

Here we start with general considerations on measure theory before reviewing the
construction of Hilbert spaces of square-integrable wavefunctions, independently of
group theory. We also define Radon-Nikodym derivatives and show that Hilbert spaces
of wavefunctions built with equivalent measures are isomorphic. For the record, our
approach will not be mathematically rigorous, and is merely intended to give a rough
picture of the actual mathematical theory.
3.1.1 Measures
When defining a quantum-mechanical system, one of the key ingredients is a prescrip-
tion for computing scalar products. For the spaces of wavefunctions that we wish
to consider, this requires being able to evaluate integrals of functions on a manifold.
Integration, in turn, relies on the existence of a measure.
Measures
Let M be a set. Roughly speaking, a measure is a function µ that associates a non-
negative number with essentially any subset U of M. That number, denoted µ(U ),
“measures” the size of U . Strictly speaking, not all subsets of M can be measured:
there exists a family of subsets of M, called “measurable sets”, and only those subsets
can actually be measured. The measure µ then is a map
µ : {measurable subsets of M} → R̄+ : U 7→ µ(U ) (3.1)
where R̄+ denotes the set of non-negative real numbers supplemented with +∞. In
order to qualify as a measure, this map needs to satisfy certain conditions; in particular,
it must be σ-additive: if U1 , U2 , etc. are disjoint measurable sets, then
+∞
! +∞
[ X
µ Ui = µ(Ui ) when Ui ∩ Uj = ∅ ∀ i, j. (3.2)
i=1 i=1
In other words, the total measure of a set consisting of several disconnected compo-
nents must be the sum of the measures of the individual components. A measure µ
on M is said to be finite if µ(M) is finite; it is σ-finite if M is a countable union of
measurable sets with finite measure (any finite measure is trivially σ-finite).
For instance, the standard translation-invariant Lebesgue measure on the real line
R is defined so that µ([a, b]) = b − a for any closed interval [a, b] ⊂ R; the measure
takes the same value for open or half-open intervals. In particular, R is a countable
union of intervals of finite length, so the Lebesgue measure is σ-finite. This definition
is readily generalized to Rn .
If N is a topological space, a function F : M → N is said to be measurable

if F −1 (V ) is a measurable subset of M for any open set V in N . In other words,
measurable functions are those that “preserve the structure of measurable sets”. Those
are the functions that we will be allowed to integrate later on.
Borel measures
Throughout this chapter and the next ones, we systematically endow M with a topol-
ogy. One can take advantage of this structure when defining a measure:
Definition. Let M be a topological space. A Borel set in M is a subset U ⊆ M

which is either an open set, or a closed set, or a union or an intersection of countably
many open or closed sets. A Borel measure on M is a measure whose measurable sets
are the Borel sets of M.
Thus, Borel measures are compatible with the topology of M. In particular, any
continuous function F : M → N is Borel-measurable. From now on, all measures are
understood to be Borel. When M is a smooth manifold, the data of a Borel measure
is equivalent to that of a volume form on M. For simplicity, we always assume that
M is a manifold.
Integrals
Measures can be used to integrate functions.1 Let µ be a Borel measure on M and
U ⊆ M a Borel set. When V is a topological vector space and F : M → V : q 7→ F(q)
is a measurable function, the (Lebesgue) integral of F over U relative to the measure
µ is written as2 Z Z
F(q) dµ(q) or F dµ .
U U
In these terms, the measure µ(U ) of a Borel set U is the integral of the function
F(q) = 1 over U : Z
µ(U ) = dµ(q). (3.3)
U
The word “measure” often also refers to the quantity dµ appearing in this expression.
For example, the standard translation-invariant Lebesgue measure on Rn is denoted

dµ(x) ≡ dn x, with the usual rules for integration. One can generate infinitely many
other measures on Rn by multiplying the Lebesgue measure by an arbitrary function:
for any non-negative measurable map ρ : Rn → R : x 7→ ρ(x), the quantity dµ(x) =
ρ(x)dn x is a Borel measure on Rn . Another example is provided by the sphere S 2 ,
which admits the rotation-invariant measure sin θ dθ dϕ in terms of polar coordinates
θ, ϕ. Finally, in section 4.2 we will use the Lorentz-invariant measure
dD−1 q
dµ(q) = p (3.4)
M 2 + q2
1
The concrete definition of integrals relies on a limiting procedure where the integrand is approx-
imated by a sequence of locally constant functions, but we will not review these details here.
2
We denote points of M as p, q, etc. to suggest thinking of them as possible momenta of a particle.
where M 2 is a positive parameter (the mass squared) while q = (q1 , ..., qD−1 ) is the
spatial momentum in D space-time dimensions.
3.1.2 Hilbert spaces of wavefunctions

We now have the tools needed to define Hilbert spaces of square-integrable functions.
For the sake of generality we consider wavefunctions taking values in a complex Hilbert
space E endowed with a scalar product
(·|·) : E × E → C : v, w 7→ (v|w) ,
which we take to be linear in its second argument and antilinear in the first one. When
E = C we simply set (v|w) = v ∗ w.
Wavefunctions
Definition. Let M be a topological space, µ a Borel measure on M, E a complex
Hilbert space with scalar product (·|·). Then an E-valued square-integrable wavefunc-
tion is a measurable map Ψ : M → E such that
Z

dµ(q) Ψ(q)|Ψ(q) < +∞.
M
We denote by L2 (M, µ, E) the vector space of such functions.
It is tempting to turn L2 (M, µ, E) into a Hilbert space by declaring that the scalar
product of two wavefunctions is the integral of their product over M, but there is a
problem: wavefunctions need not be continuous. In particular, functions that vanish
everywhere on M except at some countable number of points, are strictly speaking
non-zero vectors in L2 even though all their would-be scalar products vanish. In
the language of conformal field theory, those are “null states”. In order to cure this
pathology, one introduces the following notion:
Definition. Let µ be a Borel measure on M. A property is said to be true almost

everywhere on M if there exists a Borel set U ⊂ M such that µ(U ) = 0 and such that
the property be true on each point of M\U .
For example, when F and G are functions M → N , we say that F = G almost

everywhere on M and write F ∼ G if F and G differ only on a set of measure zero.
The relation ∼ is an equivalence relation. This solves the pathology of L2 spaces, as
one can show that integrals of functions that coincide almost everywhere are equal. In
particular, any function F ∼ 0 is said to vanish almost everywhere; such a function
belongs to L2 (the integral of its square vanishes) and can now be identified with the
function that vanishes identically on M. More precisely, let us denote by N (M, µ, E)
the space of E-valued measurable functions on M that vanish almost everywhere; it
is a subspace of L2 and may be seen as the set of null states (hence the notation N )
in L2 . This leads to the following notion:
Definition. The space of square-integrable wavefunctions on M with values in E

relative to the measure µ is the quotient of L2 by N :
L2 (M, µ, E) ≡ L2 (M, µ, E) N (M, µ, E).

(3.5)
This space is also simply called the (E-valued) L2 space on M relative to the measure µ.
Elements of L2 are thus equivalence classes of functions Ψ : M → E, two functions

being identified if they coincide almost everywhere. With this identification, one can
endow L2 with a norm k · k defined by
Z
2

kΨk ≡ dµ(q) Ψ(q)|Ψ(q) . (3.6)
M
Strictly speaking we should write the left-hand side of this definition as k[Ψ]k2 , where
[Ψ] ∈ L2 is the class3 of Ψ ∈ L2 . However, we will systematically abuse notation
by choosing arbitrarily a representative Ψ of a class [Ψ], and we use the word “wave-
function” to refer both to actual functions Ψ : M → E and to the corresponding
equivalence classes in L2 .
Formula (3.6) is a well-defined norm on L2 : it is independent of the chosen represen-

tative for the class [Ψ], and it satisfies the properties required for a norm. In particular,
a function has zero norm if it vanishes almost everywhere, i.e. if its class is the zero
vector in L2 . This is indeed the solution of the pathology we encountered in L2 spaces.
It can be shown that the space L2 is a complete normed vector space, i.e. a Banach
space, with respect to the norm (3.6). In addition the space of (equivalence classes of)
smooth functions with compact support is dense in L2 (M, µ, E), so any wavefunction
can be approximated with arbitrary precision by a smooth function.
Hilbert spaces of wavefunctions

Definition. Let µ be a Borel measure on M, E a Hilbert space with scalar product
(·|·). Let Φ and Ψ be two E-valued square-integrable wavefunctions on M. Then the
scalar product of Φ and Ψ is
Z

hΦ|Ψi ≡ dµ(q) Φ(q)|Ψ(q) , (3.7)
M
where the integrand reduces to Φ∗ (q)Ψ(q) when E = C. The space L2 (M, µ, E) is a

Hilbert space with respect to this scalar product.
With this definition we can start interpreting L2 (M, µ, E) as the space of states of
some quantum system. In Dirac notation we would write wavefunctions as Ψ ≡ |Ψi,
which is indeed suggested by the notation (3.7). The quantum state defined by such
a wavefunction is a ray (2.3)consisting of all functions M → E that are equal almost
everywhere to some constant multiple of Ψ. (Again, the notation [·] in (2.3) does not
3
This class has nothing to do with the ray (2.3) despite the identical notation.
mean the same thing as the class of a wavefunction in (3.5)!)
To interpret E-valued wavefunctions, we note the isomorphism

L2 (M, µ, E) ∼
= L2 (M, µ, C) ⊗ E. (3.8)
For example suppose E = C2 is the Hilbert space of a two-state system (as will be
the case, say, for the spin-1/2 representation of the Poincaré group in section 4.2). In
Dirac notation, we can define an orthonormal basis {|+i, |−i} of E such that a generic
(normalized) state of L2 (M, µ, E) takes the form
1
|Φi = √ |φi ⊗ |+i + |ψi ⊗ |−i (3.9)
2
where |φi and |ψi are normalized complex-valued wavefunctions on M. If we think
of E as a space of spin degrees of freedom and if M is a space of momenta, then
the wavefunction (3.9) describes the propagation of two spin states with generally
different momentum distributions accounted for by φ and ψ. Note that the state (3.9)
is typically entangled with respect to the splitting (3.8); it is unentangled if and only if
|ψi = eiλ |φi for some real number λ. The generalization of (3.9) to higher-dimensional
spaces E is straightforward.
Remark. When dealing with unitary representations of the BMS3 group in part
III, we will need to describe square-integrable wavefunctions on infinite-dimensional
manifolds (see section 10.1). Until then we will not discuss this issue.
3.1.3 Equivalent measures and Radon-Nikodym derivatives

The definition of Hilbert spaces of square-integrable wavefunctions relies on the mea-
sure µ used to define the scalar product (3.7). Naively, one might therefore expect
that the Hilbert spaces L2 (M, µ, E) and L2 (M, ν, E) differ if the measures µ and ν
do not coincide. However it is easy to show that the space L2 (M, µ, E) is essentially
independent of the measure µ. Here we prove this statement while introducing the
notion of equivalent measures and their Radon-Nikodym derivative.
Definition. Let µ, ν be two Borel measures on a manifold M. We say that µ and

ν are equivalent if they have the same sets of zero measure.
Equivalent measures can be vastly different, yet they are still pretty much the same
with regard to measure theory:
Radon-Nikodym theorem. Let µ and ν be equivalent σ-finite measures. Then

there exists a measurable function ρ : M → R+ such that
Z
ν(U ) = ρ(q)dµ(q) for any Borel set U . (3.10)
U
This relation is often written in infinitesimal form

dν(q)
dν(q) = ρ(q)dµ(q) or ρ(q) = . (3.11)
dµ(q)
In addition, any other function ρ̃ satisfying this property coincides with ρ almost ev-
erywhere on M. The function ρ is called the Radon-Nikodym derivative of ν with
respect to µ.4 A proof of this theorem can be found in [66].
For example, we mentioned below (3.3) that when dn x is the Lebesgue measure
on Rn , any non-negative function ρ gives rise to a new measure ρ(x)dn x. The Radon-
Nikodym derivative of that measure with respect to the Lebesgue measure then coin-
cides with the function ρ. In particular, when ρ(x) only vanishes on a set of Lebesgue
measure zero, the measures dn x and ρ(x)dn x are equivalent.
Remark. When µ and ν are equivalent measures, one has dµ(q)/dν(q) ∼ [dν(q)/dµ(q)]−1 ,
i.e. the Radon-Nikodym of µ with respect to ν is (almost everywhere) the inverse of
the Radon-Nikodym of ν with respect to µ.
Isomorphic L2 spaces
The notion of equivalent measures allows us to address the question raised above,
namely whether the Hilbert spaces L2 (M, µ, E) and L2 (M, ν, E) differ if the measures
µ and ν differ.
Proposition. Let µ and ν be equivalent Borel measures on M, E a Hilbert space;

we write L2 (M, µ, E) ≡ L2 (µ) and similarly for ν. Then there is an isometry
s
dµ(q)
U : L2 (µ) → L2 (ν) : Ψ 7→ U · Ψ with (U · Ψ)(q) ≡ Ψ(q) (3.12)
dν(q)
so the spaces L2 (M, µ, E) and L2 (M, ν, E) are isomorphic as Hilbert spaces.

Proof. The map (3.12) is manifestly linear and invertible, since the measures µ and
ν are equivalent so that the Radon-Nikodym derivative ρ = dν/dµ is strictly positive
almost everywhere. It only remains to prove that U preserves the scalar products (3.7);
let us denote them by h·|·iµ and h·|·iν in L2 (µ) and L2 (ν), respectively. For any two
µ-square-integrable wavefunctions Φ and Ψ, the definitions (3.11) and (3.12) readily
yield hU · Φ|U · Ψiν = hΦ|Ψiµ , which proves that U is an isometry.
This proposition says that the structure of the Hilbert space L2 (M, µ, E) does
not depend on the measure µ, since any other equivalent measure would give rise
to an isomorphic Hilbert space. A similar phenomenon will occur in section 3.2.2,
where induced representations built with different scalar products will turn out to be
equivalent.
4
There exist infinitely many functions that all represent equally well the Radon-Nikodym deriva-
tive; the theorem ensures that these functions agree, except possibly on a set of zero measure. Ac-
cordingly, we call “the” Radon-Nikodym derivative any function that satisfies (3.10).
3.2 Quasi-regular representations

In the previous pages we have seen how to build spaces of wavefunctions. Our goal
now is to endow such Hilbert spaces with a unitary group action. The strategy will
be to take the manifold M (on which wavefunctions are defined) to be homogeneous
with respect to some group action, then use this action to define unitary operators.
We now describe this approach after recalling some basic properties of group actions
and measures on homogeneous spaces. This will lead to the notion of quasi-regular
representations, which provides the simplest example of induced representations.
3.2.1 Quasi-invariant measures on homogeneous spaces

Group actions and orbits
Definition. Let M be a manifold, G a Lie group.5 An action of G on M is a smooth
map G×M → M : (f, q) 7→ f ·q such that e·q = q and f ·(g ·q) = (f g)·q for all group
elements f, g and any q ∈ M. Equivalently, an action of G on M is a homomorphism
from G to the group Diff(M) of diffeomorphisms of M.
There exist many important examples of group actions in physics: the space Rn
can be seen as an Abelian group acting on itself by the addition of vectors; the sphere
S 2 is acted upon by rotations. More generally, any group representation is a linear
action of a group on a vector space; in particular the energy-momentum of a particle
in Minkowski space is acted upon linearly by the Lorentz group.
Consider an action of G on M, and pick a point p ∈ M. The orbit of p is the

submanifold of M consisting of all points that can be reached by acting on p with G:
Op ≡ {f · p|f ∈ G} . (3.13)
The orbit is independent of the choice of p in the sense that, whenever q ∈ Op , we

have Op = Oq . The stabilizer of p ∈ M is the subgroup of G that leaves it invariant,
Gp ≡ {f ∈ G|f · p = p} . (3.14)
If q is another point in Op , and if g ∈ G is such that g · p = q, then the stabilizer of

q is g Gp g −1 , which is isomorphic to Gp . (In particular, one often abuses terminology
by saying “the stabilizer of an orbit” instead of the stabilizer of a point on the orbit.)
The stabilizer is a (closed) subgroup of G and the orbit (3.13) is diffeomorphic to the
coset space
Op ∼= G/Gp . (3.15)
This diffeomorphism is explicitly given by the bijection G/Gp → Op : f Gp 7→ f · p.
Homogeneous spaces
Definition. An action of a group G on M is said to be transitive when for any two
points p, q ∈ M there exists a group element f such that f · p = q. The space M is
5
As before elements of G are written as f , g, etc. and the identity is denoted e.
then said to be a homogeneous space for this action.
In particular a homogeneous space coincides with the orbit of any of its points
under the group action: M = Op for any p ∈ M. It follows that any homogeneous
space can be written as a coset space (3.15).
The simplest example of a G-homogeneous space is the group G itself, with the
action given by left multiplication:
g 7−→ Lf (g) = f g. (3.16)
The stabilizer in that case is trivial. Note that right multiplication
g 7−→ Rf (g) = gf (3.17)
is not quite a group action since Rf ◦Rg = Rgf does not coincide with Rf g . This can be
cured by considering right multiplication by inverse elements, i.e. g 7→ Rf −1 g = gf −1 .
In the aforementioned example of Rn , seen as an Abelian group acting on itself by
the addition of vectors, left and right multiplications coincide. (This is true for any
Abelian group.) The sphere S 2 is a more interesting example of homogeneous space,
since it is acted upon transitively by the group of rotations SO(3) but has a non-trivial
stabilizer SO(2), and is therefore diffeomorphic to the quotient SO(3)/SO(2). More
generally, one has a family of diffeomorphisms S n ∼ = SO(n + 1)/SO(n). Homogeneous
spaces will play a central role in representation theory, so we will encouter many more
examples of transitive actions later in this thesis.
The Haar measure

We now initiate the study of measure theory on homogeneous spaces.
Definition. Let M be a homogeneous space with respect to the action of a group

G; let µ be a Borel measure on M. We say that the measure is invariant under G if
µ(f · U ) = µ(U ) for all g ∈ G and for any Borel set U .
For example, the measure sin θ dθ dϕ on a sphere S 2 is invariant under rotations,

while the momentum measure (3.4) is invariant under Lorentz transformations. When
the homogeneous space M is the group manifold G itself, one has the following result:
Haar’s theorem. Let G be a finite-dimensional Lie group. Then, up to a positive

multiplicative constant, there exists a unique Borel measure on G invariant under left
multiplication (3.16), known as the left Haar measure on G.
Proof. Any left-invariant volume form on G is the pull-back by left multiplication of
a volume form on the tangent space Te G at the identity. Since G is finite-dimensional
the volume form on Te G is unique up to a positive multiplicative constant, so the
theorem follows. (See e.g. [43] for details.)
The same theorem would hold for right multiplications, although the resulting right
Haar measure generally differs from the left one. If the group G is Abelian, any left-
invariant measure is also right-invariant. For instance, the standard measure dn x on
Rn is the translation-invariant Haar measure for the Abelian group Rn .
Quasi-invariant measures
Given a homogeneous space M, we wish to integrate functions over it and we ask
whether there exists an invariant measure. It turns out that this is not always the
case (see e.g. [62]), so one introduces the following weaker notion of invariance:
Definition. Let M be a homogeneous space with respect to the action of a group

G. A Borel measure µ on M is said to be quasi-invariant under G if, for any group
element f , the measure µf defined by
µf (U ) ≡ µ(f · U ) for any Borel set U (3.18)
is equivalent to µ.
Equivalent measures are related through (3.11) by their Radon-Nikodym derivative.

Accordingly, when µ is a quasi-invariant measure on a homogeneous space M, we
denote the Radon-Nikodym derivative of µf with respect to µ by
dµf (q) dµ(f · q)

ρf (q) ≡ = (3.19)
dµ(q) dµ(q)
for any f ∈ G and any q ∈ M. We shall refer to it as “the” Radon-Nikodym derivative
of µ under the action of G. Note that it satisfies the important property
ρf g (q) = ρf (g · q)ρf (q). (3.20)
The measure µ is invariant if and only if its Radon-Nikodym derivative ρf is equal ot

one (almost everywhere) for any group element f .
Intuitively one can think of the Radon-Nikodym derivative (3.19) as an anomaly:

since µ is defined on a homogeneous space, one naively expects it to be invariant
under the group action. The Radon-Nikodym derivative measures the extent to which
invariance is spoiled. Taking again the example of Rn , the Lebesgue measure dn x is
invariant under translations and rotations, but not under arbitrary diffeomorphisms.
Indeed, for f : x 7→ f (x) a diffeomorphism of Rn , the Lebesgue measure transforms as
∂f n
dµf (x) = dn [f (x)] = d x, (3.21)
∂x
where |∂f /∂x| is the Jacobian of f . Thus the Radon-Nikodym derivative of a quasi-
invariant measure can also be seen as a generalization of the Jacobian.
Since we motivated quasi-invariant measures by the observation that invariant mea-

sures do not always exist, one might worry that a similar problem arises for quasi-
invariant measures. Fortunately one can show that, in contrast to invariant measures,
quasi-invariant measures do always exist on any finite-dimensional homogeneous space
(see e.g. [62]). We shall discuss the infinite-dimensional generalization of that state-
ment in section 10.1. Note that the existence of one quasi-invariant measure µ on M
implies the existence of infinitely many of them, since multiplying µ by any positive
function yields another quasi-invariant measure.
3.2.2 The simplest induced representations

We are now in position to describe quasi-regular representations. Let M be a manifold
acted upon by a group G, and consider a vector space of wavefunctions Ψ : M → C.
We then readily define a representation T of G in that space by writing
(T [f ] · Ψ) (q) ≡ Ψ(f −1 · q). (3.22)
Each operator T [f ] is manifestly linear, and the fact that this is indeed a representation
follows from the fact that the map q 7→ f · q is a group action. The interpretation of
formula (3.22) is simple: if the wavefunction Ψ is sharply centred around some point k
of M, then the operator T [f ] maps Ψ on a new wavefunction, now centred around the
point f · k. In chapter 4 the space M will consist of the allowed momenta of a particle,
Ψ will be the particle’s wavefunction (in momentum space), and the map q 7→ f · q
will be an action by boosts or rotations.
Figure 3.1: A wavefunction Ψ on M = R centred around some point k is acted upon by

a unitary operator T [f ] that implements the transformation k 7→ f · k. The resulting
transformed wavefunction T [f ] · Ψ is the old one, translated by f .
In order to interpret formula (3.22) as the action of a symmetry group on a Hilbert

space of wavefunctions, we need to make sure that each operator T [f ] is unitary. If M
is a homogeneous space and µ is a quasi-invariant measure on M, the scalar product
of wavefunctions is (3.7) with (Φ(q)|Ψ(q)) = Φ∗ (q)Ψ(q). Now it is easy to verify that
the representation (3.22) is generally not unitary for this scalar product:
Z Z
∗ −1 −1 (3.19)
hT [f ]Φ|T [f ]Ψi = dµ(q)Φ (f ·q)Ψ(f ·q) = dµ(q)ρf (q)Φ∗ (q)Ψ(q). (3.23)
M M
The far right-hand side generally does not coincide with the original scalar product
(3.7) because it involves the Radon-Nikodym derivative (3.19). Thus, in order to
ensure unitarity, we need to correct formula (3.22) by a factor that compensates the
non-trivial transformation law of µ:
Definition. Let G be a Lie group acting transitively on a manifold M. Let µ be a

quasi-invariant measure on M and let L2 (M, µ, C) be the space of square-integrable
wavefunctions on M. The quasi-regular representation T of G acts on this space
according to q
(T [f ] · Ψ) (q) ≡ ρf −1 (q) Ψ(f −1 · q) (3.24)
for any wavefunction Ψ, where ρf is the Radon-Nikodym derivative (3.19) of µ. If µ
is invariant, the quasi-regular representation boils down to (3.22).
Proposition. The quasi-regular representation defined by (3.24) is a unitary repre-

sentation of G in L2 (M, µ, C).
Proof. First we need to check that (3.24) actually defines a representation, i.e. that
T [f · g] = T [f ] ◦ T [g] for all f, g ∈ G. As in (3.22), linearity of T [f ] is obvious. Now
pick a wavefunction Ψ ∈ L2 (M, µ, C). At some point q ∈ M, we find
(3.24)
(T [f g]Ψ)(q) = [ρg−1 f −1 (q)]1/2 Ψ g −1 · (f −1 · q)

where we relied on the fact that q 7→ f · q is a group action. Now using (3.20) and the
definition (3.24), we can rewrite this as
1/2
(T [f g]Ψ)(q) = ρg−1 (f −1 · q)ρf −1 (q) Ψ g −1 · (f −1 · q)

(3.25)

1/2 −1
= [ρf −1 (q)] (T [g]Ψ) (f · q) = T [f ] ◦ T [g] · Ψ (q) , (3.26)
which proves that (3.24) is indeed a representation. To complete the proof we also have
to show that T is unitary for the scalar product (3.7) with (Φ(q)|Ψ(q)) = Φ∗ (q)Ψ(q).
Repeating the computation (3.23) we now find that the Radon-Nikodym derivative in
(3.24) yields an extra term in the integrand. Using (3.20) and the fact that ρe = 1, this
term cancels the Radon-Nikodym derivative in (3.23) so (3.24) is indeed unitary.
Remark. When the homogeneous space M coincides with the group G and is en-
dowed with the invariant Haar measure, formula (3.22) defines a unitary representation
of G known as the regular representation. Quasi-regular representations extend this
concept by trading the base manifold G for an arbitrary homogeneous space M.
Equivalence of quasi-regular representations

Recall that L2 spaces defined with equivalent measures are isometric via the map
(3.12). One may wonder how that statement affects quasi-regular representations: is
it true that two representations of the form (3.24) are equivalent if they are defined
using different but equivalent measures? The answer is yes: if µ and ν are equiva-
lent quasi-invariant measures on M and if we denote the corresponding quasi-regular
representations by Tµ and Tν respectively, then the isometry (3.12) is an intertwiner :
U ◦ Tµ [f ] = Tν [f ] ◦ U for all f ∈ G. (3.27)
Accordingly, the representations Tµ and Tν are unitarily equivalent. This is to say that
the quasi-regular representation (3.24) is essentially independent of the measure µ.
*3.2.3 Radon-Nikodym is a cocycle

Here we show that property (3.20) is a cohomological statement: it says that the
Radon-Nikodym derivative is a one-cocycle with respect to the representation (3.22).
This is an anecdotal observation, so the hasty reader may go directly to section 3.3.
Proposition. Let µ be a quasi-invariant measure on a homogeneous space M and

let (3.19) be its Radon-Nikodym derivative. Then the map
log ρ : G → C ∞ (M) : f 7→ log(ρf −1 ) (3.28)
is a one-cocycle with respect to the representation (3.22), with

the understanding that
log(ρf −1 ) is the function on M mapping q on log ρf −1 (q) .
Proof. We need to show that the Chevalley-Eilenberg differential (2.31) of the map
(3.28) vanishes. Using (3.22) we find
(d log ρ)f g (q) = log ρg−1 (f −1 · q) + log ρf −1 (q) − log ρg−1 f −1 (q) ,
which vanishes by virtue of property (3.20).

Let us discuss the measure-theoretic interpretation of this cohomological statement.
For example, suppose the map (3.28) is a trivial one-cocycle. Then
(2.31) (3.22)
log ρf −1 (q) = (dΨ)f (q) = (T [f ] · Ψ)(q) − Ψ(q) = Ψ(f −1 · q) − Ψ(q)
for some function Ψ(q). Equivalently,
dµ(f · q)
= eΨ(f ·q)−Ψ(q) , i.e. e−Ψ(q) dµ(q) = e−Ψ(f ·q) dµ(f · q) , (3.29)
dµ(q)
which says that the quasi-invariant measure µ is actually an invariant measure in
disguise! Indeed, the measure ν defined by dν(q) = e−Ψ(q) dµ(q) is invariant by virtue
of (3.29). In other words, the first cohomology of G with values in the space of func-
tions on M classifies the inequivalent quasi-invariant measures on M, two measures
being equivalent if they are related to one another by a function that multiplies them.
In particular, the first cohomology vanishes if all quasi-invariant measures on M are
equivalent to an invariant measure.
We can also rephrase this in the language of representation theory: the quasi-
regular representation (3.24) is a (multiplicative) affine module (2.33) on top of the
original representation (3.22). Two such modules are equivalent if the corresponding
measures are equivalent.
Remark. The cohomological properties of the Radon-Nikodym derivative have ap-

plications in physics: we mentioned below (3.19) that the Radon-Nikodym derivative
may be thought of as an anomaly, and indeed anomalies in quantum field theory are
one-cocycles for the BRST differential, valued in a suitable space of functions [67].
Our observation on the Radon-Nikodym derivative may be seen as a baby version of
that general statement.
3.3 Defining induced representations

We now extend the construction of quasi-regular representations. Suppose we have a
group G with some (closed) subgroup H. Given a representation S of H, we wish to
induce a corresponding representation T of G. To describe this mechanism we first

need to study in more detail the homogeneous manifold
M∼
= G/H . (3.30)
We will then define an action T of G on wavefunctions that live on M.
3.3.1 Standard boosts

In section 3.2.1 we introduced homogeneous spaces and observed that they can be
written as coset spaces G/Gp , where Gp is the stabilizer of some point p ∈ M. Now
note that, conversely, any coset space G/H determines a homogeneous manifold (pro-
vided H is a closed subgroup of G). Indeed the elements of G/H are left cosets
gH, where g spans G, and the action of G on G/H is given by left multiplication:
gH 7→ f · (gH) = (f g)H. In particular one can think of G/H as a manifold M where
the coset gH corresponds to the point q = g · p, where p is identified with the coset
eH at the identity. The stabilizer of g · p is gHg −1 , as noted below (3.14). Thus from
now on we describe the space G/H with the same notation as in section 3.2.1.
Now consider the point p ∈ M, identified with the identity coset eH in G/H. Since
M is a homogeneous space one can map p on any other point q ∈ M, with a group
element g ∈ G such that g · p = q. Given q, this group element is only defined up to
multiplication from the right by an element of H, since h · p = p for any h ∈ H.
Definition. Let G act transitively on M ∼

= G/H and let p ∈ M. Then a family of
standard boosts for p on M is a map
M → G : q 7→ gq such that gq · p = q. (3.31)
Any homogeneous manifold admits a family of standard boosts. For example,

in special relativity, the map (3.31) would typically be an assignment of a Lorentz
boost gq for each possible energy-momentum vector q of a massive particle. If p is
the energy-momentum of the particle at rest, gq would map this momentum p on the
boosted momentum q. In fact, in the latter case this assignment can be chosen in
such a way that gq depends continuously on q, so the family of standard boosts is
continuous. In all cases of interest below, continuous families of standard boosts will
exist, so from now on we always assume that the map (3.31) is continuous.
Remark. The existence of a continuous family of standard boosts is equivalent to

that of a global section for the principal bundle G → G/H, which in turn amounts to
saying that this bundle is trivial (see e.g. [68]). In general this is not the case — the
typical example is the non-trivial bundle SO(n + 1) → S n , which is relevant to the
Euclidean group in (n + 1) dimensions. However, all relativistic symmetry groups as
well as BMS3 are such that continuous families of standard boosts do exist,6 so we do
not need to dwell on this subtlety any further.
6
This actually follows from the fact that typical momentum orbits for such groups are homotopic
to a point, which then implies that the corresponding bundles G → G/H are trivial.
3.3.2 Induced representations

Definition. Let G be a group acting transitively on a manifold M endowed with
a quasi-invariant measure µ. Let p ∈ M be stabilized by a closed subgroup H of
G and let the map (3.31) be a continuous family of standard boosts on M. Then
the representation T of G induced by S acts in the Hilbert space H = L2 (M, µ, E)
according to
q
(T [f ] · Ψ) (q) ≡ ρf −1 (q) S[gq−1 f gf −1 ·q ] Ψ(f −1 · q) (3.32)
for any wavefunction Ψ, where ρ denotes the Radon-Nikodym derivative (3.19). It is

common to write
T = IndGH (S). (3.33)
As given here, formula (3.32) comes a bit out of the blue, so it is worth analysing
its elementary features. First note that, if we denote the trivial representation of a
group by the symbol 1, then IndG H (1) is the quasi-regular representation (3.24) of G
G
on G/H while Ind{e} (1) is the regular representation. Formula (3.32) extends these
constructions by including a non-trivial action of H in an internal space E. Before
interpreting (3.32) any further, we now verify that it is a consistent definition.
Consistency
Up to the term involving S, the right-hand side of (3.32) coincides with the quasi-
regular representation (3.24), so the only potential problem could arise from the in-
sertion of S. But since the map (3.31) is a family of standard boosts, we have

gq−1 f gf −1 ·q · p = gq−1 · f · gf −1 ·q · p = gq−1 · f · f −1 · q = gq−1 · q = p ,

so the combination gq−1 f gf −1 ·q belongs to the stabilizer H of p, as it should. Since

by assumption Ψ takes its values in the carrier space E of S, we conclude that the
right-hand side of (3.32) is well-defined. For each f ∈ G, it defines a linear operator
T [f ] acting on E-valued functions on M.
Proposition. Let H = L2 (M, µ, E). Then the map T : G → GL(H ) defined by

(3.32) is a unitary representation of G.
Proof. If it were not for the representation S, formula (3.32) would coincide with
(3.24); since the latter is a unitary representation of G, we only have to convince
ourselves that this feature is not spoiled by the presence of S. Noting that
gq−1 f g g(f g)−1 ·q = gq−1 f gf −1 ·q · gf−1

−1 ·q g g(f g)−1 ·q (3.34)
and using the fact that S is a representation of H, one can mimick the sequence of equa-
tions (3.25)-(3.26) for T given by (3.32), which implies that it is indeed a representa-

tion. As forunitarity, it follows from the fact that S is unitary: S[h]Φ(q) S[h]Ψ(q) =
Φ(q) Ψ(q) for all wavefunctions Φ, Ψ, any point q ∈ M and any h ∈ H.
Interpretation
The basic interpretation of the induced representation (3.32) is the same as for (3.24):
Ψ(f −1 · q) represents the fact that the wavefunction Ψ is “boosted” by f , while the
√
factor ρf −1 ensures unitarity. The new ingredient is the combination
S gq−1 f gf −1 ·q ≡ Wq [f ].

(3.35)
Its appearance represents the fact that, in contrast to the quasi-regular representation,
wavefunctions take their values not in C, but in some more general “internal” Hilbert
space E carrying a representation S of H.
When interpreting induced representations as particles, H is typically a group

of spatial rotations combined with space-time translations, the space E consists of
spin degrees of freedom, and S determines the value of spin. In that context the
operator (3.35) is known as the Wigner rotation associated with f at momentum
q. The quasi-regular representation (3.24) can thus be seen as a “scalar” induced
representation, as opposed to the spinning case (3.32). Wigner rotations are trivial for
scalar particles. Note that because S is a representation, Wigner rotations satisfy the
property Wq [f g] = Wq [f ] Wf −1 ·q [g].
Remark. We mentioned on page 44 that generic homogeneous manifolds do not

admit continuous families of standard boosts, which invalidates the global well-defi-
niteness of the Wigner rotation (3.35). This problem can be cured by reformulating
induced representations in terms of wavefunctions defined on the group manifold G
rather than G/H (see e.g. [62]). In this thesis we systematically use the homogeneous
space viewpoint (3.32), as it will suffice for all cases of interest below. This being said,
note that the reformulation in terms of wavefunctions on G is useful for certain appli-
cations of three-dimensional higher-spin theories [69, 70] due to the relation between
induced representations and harmonic analysis on homogeneous spaces [71, 72].
3.3.3 Properties of induced representations

Induced representations have a number of important properties that we now explore.
We first show that the definition (3.32) is “robust” in that it depends neither on the
choice of the measure µ, nor on the choice of standard boosts (3.31). Then we turn
to the behaviour of induced representations under operations such as direct sums and
tensor products.
Robustness
Formula (3.32) depends not only on the inducing data (the group G, its subgroup H
and a spin representation S), but also on the measure µ on M ∼ = G/H and on the
choice of a family of standard boosts gq . Naively, one expects all these parameters to
affect the representation. We now show that this is not the case.
As far as the measure is concerned, one readily verifies that two induced repre-
sentations defined with the same inducing data and the same standard boosts but
different, though equivalent, quasi-invariant measures, are unitarily equivalent. The

proof is essentially the same as for quasi-regular representations (see eq. (3.27)), and
the intertwiner is the map (3.12). As regards standard boosts, a similar result holds:
Proposition. Let M ∼ = G/H, S a spin representation of H, H = L2 (M, µ, E). Let

g : M → G : q 7→ gq and g 0 : M → G : q 7→ gq0 be two continuous families of standard
boosts and call T , T 0 (respectively) the corresponding induced representations of G.
Then there is a unitary operator
V : H → H : Ψ 7→ V · Ψ (V · Ψ) (q) = S gq−1 · gq0 Ψ(q)

with (3.36)
that intertwines T and T 0 , which are therefore unitarily equivalent:
T [f ] ◦ V = V ◦ T 0 [f ] ∀ f ∈ G. (3.37)
Proof. The fact that V is a unitary operator follows from unitarity of S. Property
(3.37) then follows from the definitions (3.32) and (3.36).
A corollary of these observations on robustness is that one may unambiguously say
“the” representation of G induced by the representation S of H, without any reference
to the measure or to the choice of standard boosts.
Remark. The transformation (3.36) may be seen as a gauge transformation with

gauge group H. Indeed the combination gq−1 gq0 ∈ H can depend on q in an arbitrary
way, owing to one’s freedom in the choice of standard boosts. It acts on wavefunctions
as a momentum-dependent transformation Ψ 7→ V · Ψ given by the representation S,
and each such transformation maps the system on a unitarily equivalent one. In fact,
the differentiation of the operator S[gq−1 f gf −1 ·q ] defines a gauge field on G/H valued
in the Lie algebra h of H, and the Wigner rotation itself may be seen as a holonomy
(see e.g. [73, 74]).
Operations on induced representations

We now study the behaviour of induced representations under standard operations
such as conjugation, direct sums and the like. The proofs are omitted and we refer
to [62] for details.
Let E be a Hilbert space with scalar product (·|·). We call conjugation the map
∗ ∗
C : E → Ects : v 7→ (v|·), where Ects denotes the space of continuous linear functionals7
on E. Then, if S is a unitary representation acting on E, its conjugate representation
is S ≡ C ◦ S ◦ C −1 . In the context of induced representations, one can then show that8
IndG G
H (S) ∼ IndH (S), i.e. the representation induced by the conjugate of S is unitarily
equivalent to the conjugate of the representation induced by S.
7
Recall that any continuous linear functional on a Hilbert space E is a scalar product (v|·) for
some fixed vector v ∈ E.
8
The symbol ∼ denotes unitary equivalence of representations.
One can similarly show that induced representations behave well under direct sums
and tensor products thanks to the unitary equivalences
IndG G G
H (S1 ⊕ S2 ) ∼ IndH (S1 ) ⊕ IndH (S2 ) ,
(3.38)
IndG G G
H (S1 ⊗ S2 ) ∼ IndH (S1 ) ⊗ IndH (S2 ) .
As a corollary, if S is reducible, then IndG

H (S) is reducible. The converse is not true; for
instance, if H = 1 is the trivial subgroup with S the irreducible trivial representation,
then T is the regular representation, which is generally reducible.
One should also check that induction itself is a “good” operation on representations.
This is guaranteed by the theorem of induction in stages: let H1 be a closed subgroup
of H2 , which itself is a closed subgroup of G. Let S be a unitary representation of H1 .
Then one has the following unitary equivalence of representations:
H2
IndG G

H1 (S) ∼ IndH2 IndH1 (S) .
In other words, inducing directly from H1 to G, or from H1 to H2 and then to G, are

the same operations.
3.3.4 Plane waves

For practical purposes it is convenient to rewrite formula (3.32) in a basis of plane wave
states. In the relativistic context, they represent particles with definite momentum.
Delta functions
Let M ∼ = G/H be a homogeneous space, µ a quasi-invariant measure on M. Pick
a point k ∈ M. We define the Dirac distribution δk at k associated with µ as the
distribution such that hδk , ϕi ≡ ϕ(k) for any test function ϕ on M. Equivalently, we
introduce a “Dirac delta function” δ(k, ·) such that
Z
hδk , ϕi = dµ(q)δ(k, q)ϕ(q) ≡ ϕ(k). (3.39)
M
Thus the distribution δk acts on a test function ϕ(·) by integrating it against the delta
function δ(k, ·).
Note that the definition of the delta function δ relies on the measure µ since the
combination dµ(q)δ(k, q) is G-invariant by design. To make this explicit, let us denote
by δµ the delta function associated with µ. If ρ is some positive function on M and
dν(q) = ρ(q)dµ(q) is a new measure, then the delta function δν associated with ν
differs from δµ by a factor ρ:
δµ (k, q)
δν (k, q) = . (3.40)
ρ(q)
In particular, since µ is quasi-invariant under G, for any f ∈ G we have
δµ (k, q)
δµ (f · k, f · q) = (3.41)
ρf (q)
where ρf is the Radon-Nikodym derivative (3.19). Thinking of the latter as a kind of

Jacobian, eq. (3.41) is a restatement of the transformation law of the Dirac distribu-
tion under changes of coordinates. In what follows we shall not indicate explicitly the
dependence of δ on the measure µ.
The best known Dirac distribution is the one associated with the translation-
invariant Lebesgue measure on Rn . We will encounter this delta function repeatedly
so we denote it by δ (n) to distinguish it from other Dirac distributions. With that
notation, the delta function associated with the Lorentz-invariant measure (3.4) is
p
δ(k, q) = M 2 + q2 δ (D−1) (k − q). (3.42)
Plane wave states

Definition. Let {e1 , ..., eN } be a countable orthonormal basis9 of E. For ` ∈ {1, ..., N }
and k ∈ M, we call plane wave state with spin ` and momentum k the wavefunction
Ψk,` (q) ≡ e` δ(k, q) (3.43)
where δ is the Dirac distribution associated with the measure µ on M.
In non-relativistic quantum mechanics, if we were describing a particle on the line

M = R, a plane wave would typically be one of the states Ψx = |xi representing a
particle located at the point x ∈ R (with infinite momentum uncertainty). In the
dual, momentum-space picture, a plane wave would be a state Ψk = |ki with definite
momentum k (but infinite position uncertainty). Our terminology is motivated by the
latter viewpoint. In the Poincaré case a wavefunction (3.43) will describe a particle
with definite spin projection ` and energy-momentum k, i.e. a typical asymptotic state
in a scattering experiment. With the notation of (3.9), for instance, the space E of spin
degrees of freedom is two-dimensional and has a basis {|+i, |−i} = {e1 , e2 } consisting
of states with definite spin along the vertical axis.
Scalar products of plane waves can be evaluated thanks to the definition (3.7).
Using the fact that the basis of ek ’s is orthonormal and the definition (3.39) of the
delta function, one finds
hΨk,` |Ψk0 ,`0 i = δ(k, k 0 )δ``0 . (3.44)
This property allows us to see in which sense plane waves form a “basis” of the Hilbert
space H = L2 (M, µ, E). Indeed, any wavefunction Φ : M → E can be written as
N
X
Φ(q) = Φ` (q) e` where Φ` (k) = hΨk,` |Φi .
`=1
Removing the argument q, this says that any wavepacket Φ is a superposition of plane
waves:
Z XN Z N
X
Φ= dµ(k) Φ` (k)Ψk,` = dµ(k) hΨk,` |Φi Ψk,` . (3.45)
M `=1 M `=1
9
We are assuming that E is a separable Hilbert space; N may be infinite.
Note that this can be interpreted as the completeness relation

Z N
X
I= dµ(k) hΨk,` |·i Ψk,` (3.46)
M `=1
where I is the identity operator. In the more common (but less precise) Dirac notation
this would be a sum of projectors |Ψk,` ihΨk,` |. For example, for a Rparticle on the real
line, the Dirac formR of this completeness relation would read I = R dx|xihx| in posi-
tion space, or I = R dk|kihk| in momentum space. We will show in section 3.5 that
the existence of a family of projectors associated with M is one of the key properties
of induced representations.
From the construction of plane waves we see that induced representations are just
an upgraded version of one-particle quantum mechanics, with extra freedom in the
choice of the space M, the group G, and the spin states contained in E. This obser-
vation will guide us in developing our intuition of induced representations, especially
in part III of the thesis.
Remark. The quantity (3.43) is not a square-integrable function on M and therefore

does not, strictly speaking, belong to the Hilbert space. The same problem arises in
standard quantum mechanics, where the states |xi or |ki form a “basis” only in a weak
sense. Intuitively one can think of plane waves (3.43) as idealizations of Gaussian
wavefunctions centred at k in the limit where their spread goes to zero. A more
rigorous way to include such states is to work with so-called rigged Hilbert spaces
(see e.g. [75, 76]), which are designed so as to include both standard square-integrable
functions and distributions.
Boosted plane waves

We can now rewrite formula (3.32) for induced representations in terms of plane waves.
Choosing a plane wave (3.43), for f ∈ G and q ∈ M we find
Ψk,` (f −1 · q) = ρf (k) Ψf ·k,` (q) (3.47)
where we used (3.41) and the property (3.20). Note what we have achieved: in the
original definition (3.32) the argument of Ψ changes between the left and the right-
hand sides; here, by contrast, the argument will be the same, but what changes is the
label specifying the momentum of the plane wave. Indeed, using (3.47) in formula
(3.32), we find
q
T [f ] · Ψk,` (q) = ρf −1 (q)ρf (k) S gq−1 f gf −1 ·q · Ψf ·k,` (q).

Since the plane wave on the right-hand side contains a delta function δ(f · k, q), we
can replace all q’s in this expression
p by f · k and
p remove the argument from both sides.
Using once more (3.20) in ρf −1 (q)ρf (k) = ρf (k), we end up with
q
T [f ] · Ψk,` = ρf (k) S gf−1

·k f gk · Ψf ·k,` (3.48)
This formula is the simplest rewriting of the induced representation (3.32). The only
extra improvement we could still add is to write as (S[· · · ])``0 the matrix element of
the operator S[· · · ] between the states e` and e`0 , whereupon (3.48) becomes
q
T [f ] · Ψk,` = ρf (k) S gf−1

·k · f · gk `0 ,` · Ψf ·k,`0 (3.49)
with implicit summation over `0 .
Formula (3.48) gives a geometric picture of the states of an induced representation.

Indeed, the label k spans all points of M, so we can now view each point of M
as a quantum state. (More precisely, a point of M is a family of dim(E) linearly
independent states.) Two different points of M, say k and k 0 , then correspond to
two linearly independent states Ψk and Ψk0 , and a transformation f of M mapping
k on k 0 = f · k gives rise to a unitary operator T [f ] relating the corresponding plane
waves. This is a geometrization of representation theory: we can “see” each linearly
independent state of the representation T as a point of G/H. This observation is at the
core of the orbit method, which consists in quantizing suitable homogeneous manifolds
to obtain unitary group representations (see chapter 5). In three-dimensional gravity,
the phase space of gravitational perturbations will turn out to be precisely such a
homogeneous manifold, and its quantization will produce a Hilbert space of “soft” or
“boundary gravitons”. We will address these questions in chapter 8 and in part III of
the thesis.
3.4 Characters
In this section we describe the characters associated with induced representations.
We start by motivating and defining characters in general terms, before proving the
Frobenius formula. We end by discussing the relation between characters and fixed
point theorems.
3.4.1 Characters are partition functions

Unitary representations may be seen as general models of symmetric quantum systems:
any system invariant under a certain symmetry group G forms a (generally reducible,
generally projective) unitary representation of G. Accordingly, symmetry generators
provide natural observables in the system, and one may ask about the properties of
these observables — typically, about their spectrum.
When a system is invariant under time translations, for instance, the correspond-
ing symmetry generator is the Hamiltonian operator H. The information about its
spectrum is captured by the canonical partition function 10
Z(β) = Tr e−βH ,

(3.50)
where β is the inverse of the temperature. If the system admits extra symmetries such
as, say, rotations, one can look for the maximal set of mutually commuting symmetry
10
The notation “Z” stands for the German word Zustandssumme, meaning “sum over states.”
generators Qa and switch on their chemical potentials µa .11 The spectrum of these
new operators, together with H, is then contained in the grand canonical partition
function !
Xr
Z(β, µ1 , ..., µr ) = Tr exp − β H − µa Q a . (3.51)
a=1
Now suppose we take β to be purely imaginary (while keeping the µa ’s real) in this
expression. Then the operator inside the trace is unitary, since it is an exponential
of anti-Hermitian operators. In fact, it is a symmetry transformation acting in the
Hilbert space according to some unitary representation T , so we can write
Z(β, µ1 , ..., µr ) = Tr (T [f ])
for some element f belonging to the symmetry group G. This motivates the following
definition:
Definition. Let T be a representation of a group G in a complex vector space H .

The character of that representation is the map12
χ : G → C : f 7→ χ[f ] ≡ Tr (T [f ]) . (3.52)
This definition ensures that χ[f ] is independent of the basis of H used to evaluate
it. As an application, recall that two group elements f and f 0 are conjugate if there
exists an element g ∈ G such that f 0 = gf g −1 , and that the conjugacy class of f is
[f ] ≡ g f g −1 g ∈ G .

Thus, formula (3.52) ensures that characters are class functions in the sense that χ[f ]
only depends on the conjugacy class of f , and not on f itself:
χ[f ] = χ[g f g −1 ]. (3.53)
Remark. The definition (3.52) suggests that characters are functions on G. While
this is true for finite-dimensional representations, it is not true in infinite-dimensional
ones. In fact, characters should not be seen as functions, but rather as distributions
[79]. Similarly to our dealing with Dirac distributions as if they were “delta functions”,
we will not take such mathematical subtleties into account.
3.4.2 The Frobenius formula

Our derivation of the character formula for induced representations is inspired by [80],
although the formula itself appears in many textbooks on group theory; see e.g. [81].
11
Here the index a runs from one to r, the latter being essentially the rank of the symmetry group.
12
The terminology of “characters” is due to Weber and Frobenius, and stems from the fact that
irreducible representations of finite groups are wholly characterized by their character (see e.g. [77,
p.117] or [78, p.783]).
Theorem. The character of the induced representation T = IndG

H (S) defined by
(3.32) is given by the Frobenius formula
Z
χ[f ] = Tr (T [f ]) = dµ(k) δ(k, f · k) χS [gk−1 f gk ] (3.54)
M
where µ is a quasi-invariant measure on M ∼ = G/H, δ is the associated Dirac distri-

bution, the gk ’s are standard boosts, and χS is the character of S.
Proof. Let f ∈ G and let T [f ] be the associated unitary operator (3.32). We work in
the basis of plane wave states (3.43) so that the trace of T [f ] reads
Z N
X
χ[f ] = Tr (T [f ]) = dµ(k) Ψk,` T [f ] · Ψk,`
M `=1
where we “sum over momenta” thanks to the measure µ on M. Now using (3.48) and
the scalar products (3.44), we find
Z q N
X
χ[f ] = dµ(k) ρf (k) hψk,` |S[gf−1
·k · f · gk ]ψf ·k,` i (3.55)
M `=1
Z q N
(3.44) X
S[gf−1

= dµ(k) ρf (k) δ(k, f · k) ·k · f · gk ] ``
.
M `=1
Here the delta function δ(k, f ·k) allows us to trade f ·k for k. In particular the Radon-
Nikodym derivative ρf (k) = dµ(f · k)/dµ(k) reduces to unity. One then recognizes the
sum N
P
`=1 (S[· · · ])`` ≡ χS [· · · ] as the character of S, and eq. (3.54) follows.
The Frobenius formula (3.54) embodies the geometrization of representation the-
ory mentioned at the end of section 3.3.4: the trace of an operator has now become
an integral over a (subset of a) homogeneous space. That integral can be interpreted
as a sum of characters of S. Before studying this formula further, we need to check
that it satisfies the basic properties of a character.
First, since it is the character of an induced representation and since the latter
is independent (up to unitary equivalence) of the choice of the measure µ, the same
should be true of expression (3.54). To see that this is indeed the case, recall that
the combination dµ(k)δ(k, ·) is invariant under changes of measures (as follows from
the definition (3.39) of the Dirac distribution), which then implies invariance of the
character. Note in particular that the Radon-Nikodym derivative of the measure µ
does not appear in (3.54). Secondly, induced representations are independent of the
choice of standard boosts gq ; using the fact that the character of S is a class function,
one readily verifies that (3.54) is also independent of that choice. Finally, recall from
(3.53) that characters are class functions; using (3.20) one verifies that this is indeed
the case with formula (3.54). Note one crucial implication of this fact: because of
the term χS [gk−1 f gk ] in (3.54), the character χ[f ] vanishes if f is not conjugate to an
element of H. In other words the character of the induced representation IndG H (S) is
supported on the points of G whose conjugacy class intersects H.
Remark. Since the character χS of the spin representation is a class function, one is
naively tempted to pull the term χS [gk−1 f gk ] out of the integral (3.54), as the notation
suggests that f is conjugate to gk−1 f gk . This is not true, because for generic g, g 0 ∈ G
one has χS [g −1 f g] 6= χS [g 0−1 f g 0 ]. As a consequence, the integral (3.54) is generally
non-trivial.
3.4.3 Characters and fixed points

Formula (3.54) is one of the key results of this chapter. Its two most salient features
are (i) the fact that the character of T is completely specified by that of S and the
space M, and (ii) the fact that it is an integral over the points of M that are left fixed
by f . At first sight the latter observation is a surprise: there is no obvious reason why
a sum over all states of the induced representation would collapse to an integral over
fixed points of f , though in practice this is due to the scalar products hΨk |Ψf ·k i in
the trace (3.55). This collapse is an instance of localization: an integral localizes to a
small subset of points in M, so that the evaluation of (3.54) becomes child’s play. We
shall encounter this situation with the BMS3 group in section 10.3.
Figure 3.2: A manifold M acted upon by a rotation around some axis. The points
that belong to the axis are the only ones left fixed by the rotation, and are therefore
the only ones that contribute to the integral of formula (3.54).
Remark. The relation between characters of group representations and fixed point
theorems is much deeper and more general than the superficial description given here.
Indeed one can show [82] (see also [79]) that (3.54) coincides with the Lefschetz number
of T [f ] when the latter is seen as an endomorphism acting on a space of E-valued
sections on M. In turn, the fact that T [f ] is derived by (3.32) from a diffeomorphism
action of f on M turns out to imply that its Lefschetz number is given by the Atiyah-
Bott fixed point theorem [83].
*3.5 Systems of imprimitivity

This technical section is for advanced reading: other than for a key corollary that
implies the exhaustivity of induced representations for semi-direct products, it is in-
consequential to the remainder of the thesis and may be skipped in a first reading.
We saw in eq. (3.46) that the identity operator I can be written as an integral of
projectors Ψk hΨk |·i = |Ψk ihΨk |. This leads to a seemingly random idea: why not
combine these projectors into more general operators? For example, if U is any Borel
subset of M, we can associate with it a projection operator
Z N
X
PU ≡ dµ(k) hΨk,` |·i Ψk,` . (3.56)
U `=1
In that language the identity operator is I = PM . As it turns out this idea is one of
the key properties of induced representations, and sparked the whole development of
the theory by Mackey in the fifties [84–86]. In particular it leads to the imprimitivity
theorem, which roughly states that any representation that admits a suitable family
of projectors (3.56) is necessarily induced. An important corollary of that result is the
fact that all irreducible unitary representations of semi-direct products are induced.
The plan of this section is the following. We first define the notion of systems of
imprimitivity as suitable families of projection operators, and show that any induced
representation admits such a family. We then state (without proof) the imprimitivity
theorem, which we eventually use to define a restricted notion of equivalence for in-
duced representations. The presentation is based on [62, 63], but our approach will be
heuristic at times; we refer to [87, 88] for a mathematically rigorous presentation.
3.5.1 Projections and imprimitivity

Here we describe the operators (3.56) in the framework of projection-valued measures
and show that they form a system of imprimitivity.
Projection-valued measures
Let us put (3.56) in a more general context. Observe that, given the Borel set U , the
projector PU acts on wavefunctions by setting them to zero everywhere outside of U :

Ψ(q) if q ∈ U,
PU · Ψ (q) = (3.57)
0 otherwise.
The construction of such projectors motivates the following definition:
Definition. Let M be a manifold, H a Hilbert space, End(H ) the space of linear

operators in H . Then a projection-valued measure on M with respect to H is a map
P : {Borel subsets of M} → End(H ) : U 7→ PU (3.58)
satisfying the following properties:

• PM is the identity operator I in H .
• For any pair of Borel sets U and V , we have PU ∩V = PU PV ; in particular each
PU is a projector.
• The map P is σ-additive in the sense that, if U1 , U2 , etc. are disjoint Borel sets,
PU1 ∪U2 ∪··· = PU1 + PU2 + · · · . (3.59)
The terminology here is inspired by measure theory: the map (3.58) is an operator
analogue of (3.1) and property (3.59) corresponds to (3.2). Thus a projection-valued
measure measures the “size” of a subset U not by a real number µ(U ), but by the rank
of a projector PU .
It is easy to verify that the projectors (3.56) define a projection-valued measure on

M with respect to H . One can build such a family for any induced representation.
In terms of plane waves (3.43), this measure infinitesimally reads
N
X N
X
dP (k) = dµ(k) Ψk,` hΨk,` |·i = dµ(k) |Ψk,` ihΨk,` | ≡ dµ(k)Ik (3.60)
`=1 `=1
at any k ∈ M. Here we have used both our Pnotation and the standard Dirac one;
N
we have also introduced an operator Ik = `=1 |Ψ k,` ihΨk,` | such that the identity
operator in H is an integral I = M dµ(k)Ik . Analogously to (3.3), the operator PU is
R
the integral of dP over U . For a particle on the real line, for example, the projection-
valued measure in momentum space would read dP = dk |kihk| with k ∈ R.
Systems of imprimitivity
There is one property that makes the projection-valued measure (3.56) very special.
Namely, the transitive action of G on M gives rise to an action (3.48) on wavefunctions;
the latter, in turn, yields an action on the projectors |Ψk ihΨk |. So the fact that (3.56)
acts in the space of a representation provides a relation between the geometry of M
and the action of G on the projection-valued measure, which motivates the following
definition:
Definition. Let M be a manifold, G a Lie group acting on M. Let T be a unitary

representation of G in a Hilbert space H . Then a system of imprimitivity for T based
on M is a projection-valued measure P on M with respect to H such that
Pf ·U = T [f ] ◦ PU ◦ T [f ]−1 (3.61)
for all f ∈ G and any Borel subset U of M. The system is said to be transitive if the
action of G on M is transitive.
In this language the projectors (3.56) imply that any induced representation has a
transitive system of imprimitivity:
Proposition. Let S be a unitary representation of a closed subgroup H of G,

T = IndG ∼
H (S) the corresponding induced representation, and M = G/H. Then the
associated projection-valued measure (3.56) is a transitive system of imprimitivity for
T based on M. In the notation (3.60) this is to say that
dP (f · k) = T [f ] ◦ dP (k) ◦ T [f ]−1 (3.62)
for all k ∈ M and any f ∈ G. We shall refer to (3.56) as the canonical system of
imprimitivity of the induced representation T .
Proof. Transitivity is obvious, so the only subtlety is proving (3.61). Let us pick a
group element f ∈ G and a Borel set U ⊆ M. We start from the definition (3.56) and
relate T [f ] ◦ PU ◦ T [f ]−1 to Pf ·U using formula (3.48) and the fact that T is unitary:
Z N
X
T [f ] ◦ PU ◦ T [f ] −1
= dµ(k)ρf (k) S[gf−1 −1
·k f gk ]Ψf ·k,` hS[gf ·k f gk ]Ψf ·k,` |·i.
U `=1
Here the sum over ` allows us to cancel the two S[· · · ]’s by unitarity. Using also
dµ(k)ρf (k) = dµ(f · k) and renaming the integration variable, the right-hand side
boils down to Pf ·U .
Remark. The word “imprimitive” means “which is not primitive” and was introduced
by Galois [89] in the context of permutation groups. The action of a group on a set
shuffles the elements of this set, and the action is imprimitive if these permutations
preserve some (non-trivial) partition of the set. In the present case the group G acts
on the Hilbert space H by the induced representations (3.48), and property (3.62)
says that this action preserves the partition of H into isomorphic subspaces Ek ∼ =E
with definite momentum k.
3.5.2 Imprimitivity theorem

The considerations of the previous pages open the door to a highly non-trivial state-
ment, namely the fact that any representation that admits a transitive system of
imprimitivity is an induced representation:
Imprimitivity theorem. Let G be a finite-dimensional Lie group, H a closed sub-

group of G. Let T be a continuous, unitary representation of G in some Hilbert
space H and let P be a system of imprimitivity for T on M = G/H. Then there
exists a unitary representation S of H in some Hilbert Sspace E such that the pair
(T , P ) is unitarily equivalent to IndG
H (S), P S
where P is the canonical system of
G
imprimitivity (3.56) associated with IndH (S). More precisely, there exists an isom-
etry U : L2 (G/H, µ, E) → H , where µ is a quasi-invariant measure on G/H, that
intertwines the representations T and IndG H (S) and that satisfies
U ◦ PUS ◦ U −1 = PU (3.63)
for any Borel set U ⊆ M.
The complete proof of this theorem can be found in [90] and is reproduced in [62].
Given the representation T and the system of imprimitivity P , the key subtlety is to
construct a Hilbert space E and a representation S of H in E. We will not dwell on
this proof any further, but turn now to some of its applications.
Equivalent induced representations

Here we describe a restricted notion of equivalence for induced representations, culmi-
nating with the observation that two induced representations T1 , T2 are “equivalent” if
and only if they are induced from equivalent representations S1 , S2 . The proofs rely
crucially on the imprimitivity theorem, but we omit them; they can be found in [62].
Definition. Let T1 and T2 be two representations of G induced by some represen-

tations S1 , S2 (respectively) of a subgroup H. Let their respective carrier spaces be
H1 , H2 , and let P1 , P2 (respectively) be their canonical systems of imprimitivity. Then
a linear map A : H1 → H2 intertwines the pairs (T1 , P1 ) and (T2 , P2 ) if
A ◦ T1 [f ] = T2 [f ] ◦ A and A ◦ (P1 )U = (P2 )U ◦ A (3.64)
for any f ∈ G and any Borel set U in G/H.
Equivalence theorem. Let S1 and S2 be unitary representations of H in the Hilbert

spaces E1 and E2 (respectively). Let T1 and T2 be the corresponding induced repre-
sentations, and P1 , P2 the associated canonical systems of imprimitivity. Then there
exists a (continuous) vector space isomorphism between the space of operators inter-
twining S1 and S2 and the space of intertwiners between (T1 , P1 ) and (T2 , P2 ).
As a corollary, the space of intertwiners between (T1 , P1 ) and (T2 , P2 ) contains an

isometry if and only if the space of intertwiners between S1 and S2 does. In other
words, if we declare that the pairs (T1 , P1 ) and (T2 , P2 ) are equivalent once there exists
an isometry A satisfying (3.64), then we have
(T1 , P1 ) ∼ (T2 , P2 ) if and only if S1 ∼ S2 . (3.65)
In particular, if S1 and S2 are equivalent (in the usual sense), then so are the induced
representations T1 and T2 (regardless of their systems of imprimitivity). The converse
is not true, since the “if and only if” of (3.65) also involves the systems of imprimitivity
associated with T1 and T2 . In other words, saying that two induced representations
are equivalent without saying anything about their systems of imprimitivity is not
sufficient to conclude that they are induced from the same spin representation S.
As mentioned below eq. (3.38), irreducibility of S does not generally imply irre-
ducibility of the corresponding induced representation. In the next chapter we shall
state a stronger result for semi-direct products, but for now we display a theorem that
provides a slightly weaker criterion for the irreducibility of induced representations.
Definition. Let T be an induced representation, P the associated canonical sys-

tem of imprimitivity. We call the pair (T , P ) irreducible if the space of operators
intertwining it with itself consists of multiples of the identity.
Irreducibility theorem. Let T be induced by S and let P be its system of imprim-

itivity. Then the pair (T , P ) is irreducible if and only if S is irreducible.
The latter theorem shows that a suitable notion of irreducibility is preserved along
the induction process, since an irreducible S will lead to an induced representation T
and a system of imprimitivity P which, together, will be considered irreducible in the
above sense. But the theorem does not say that an induced representation T on its
own is irreducible if it is induced from an irreducible S.
Chapter 4
Semi-direct products
In this chapter we introduce semi-direct products such as the Poincaré group, the
Galilei group and the Bargmann group. We describe their irreducible unitary rep-
resentations, which are induced from representations of their translation subgroup
combined with a so-called little group. We interpret these representations as particles
propagating in space-time and having definite transformation properties under the
corresponding symmetry group. This picture will be instrumental in our study of the
BMS3 group.
The plan is as follows. In section 4.1 we define semi-direct products and intro-
duce the key notions of momentum orbits, little groups and particles. We also explain
why irreducible unitary representations are always induced, and describe these rep-
resentations in general terms. The remaining sections are devoted to applications of
these considerations. In section 4.2 we describe relativistic particles, i.e. unitary rep-
resentations of the Poincaré group, with a particular emphasis in section 4.3 on the
three-dimensional setting (which will be useful when dealing with BMS3 ). Section
4.4 is devoted to non-relativistic particles, i.e. unitary representations of Bargmann
groups. Useful references include [62, 63] for the general theory, and [53, 91, 92] for its
application to Poincaré.
4.1 Representations and particles

In short, a semi-direct product group consists of two pieces: a non-Abelian group G of
transformations that can be interpreted as “rotations” or “boosts”, and another group
A that consists of transformations analogous to translations that are acted upon by
rotations and boosts. This structure is denoted G n A and is common to the Poincaré
groups (1.2) as well as the BMS groups (1.1)-(1.9). In this section we define such
groups in abstract terms, define the associated notion of “momenta” and describe their
irreducible unitary representations, which we interpret as particles.
4.1.1 Semi-direct products

Definition. Let G and A be Lie groups; we denote elements of G as f , g, etc. and
those of A as α, β, etc. Let σ : G × A → A : (f, α) 7→ σf (α) be a smooth action of G
61
Chapter 4. Semi-direct products 62
on A where each σf is an automorphism of A. Then the semi-direct product of G and

A with respect to σ is the group denoted
G nσ A or GnA (4.1)
whose elements are pairs (f, α) where f ∈ G and α ∈ A, with a group operation

(f, α) · (g, β) = f · g, α · σf (β) . (4.2)
This definition implies for instance that the inverse of (f, α) is

(f, α)−1 = f −1 , [σf −1 (α)]−1 .

(4.3)
It follows that A is a normal subgroup of G n A: identifying A with the set of elements
(e, α) ∈ G n A (where e is the identity in G), one finds
(4.2)
(g, β) · (e, α) · (g, β)−1 = e, β · σg (α) · β −1 ∈ A.

(4.4)
It is equally easy to verify that G is a subgroup of G n A, though it is generally not a
normal subgroup. Indeed, upon identifying G with the subgroup of G n A consisting
of elements (f, eA ) (where eA is the identity in A), we find
(4.2)
(g, β) · (f, eA ) · (g, β)−1 = gf g −1 , β · σgf g−1 (β −1 ) .

(4.5)
For this to be an element of G, we must require that β · σgf g−1 (β −1 ) coincides with eA ,
which is the statement that σf (α) = α for any α ∈ A. Thus G is a normal subgroup of
A if and only if its action σ is trivial, in which case G n A is isomorphic to the direct
product G × A. From now on we always take the action σ to be non-trivial.
Rotations and translations

A case of great interest, both for the general theory and for our specific purposes,
occurs when A is a vector group. By this we mean a vector space endowed with the
Abelian group operation given by the addition of vectors: α · β ≡ α + β. In that case
the identity in A is the vanishing vector eA = 0.
Definition. Let G be a Lie group, A a vector space, σ a representation of G in A,

and consider the semi-direct product G nσ A whose elements are pairs (f, α) with
group operation

(f, α) · (g, β) = f · g, α + σf (β) . (4.6)
Elements of G are then called rotations or boosts while elements of A are translations.
Note that, since A is a vector group, the inverse (4.3) of (f, α) is (f, α)−1 =
f , −σf −1 (α) . Relation (4.4) also simplifies to (g, β) · (e, α) · (g, β)−1 = (e, σg α).
−1
From now on the words “semi-direct product” and the notation G n A will always refer
to a group (4.1) with A a vector group. (This is why the second factor in (4.1) was
denoted “A” in the first place.)
The terminology of “rotations” and “translations” is justified by the semi-direct

products commonly encountered in physics:
• The Euclidean group in n space dimensions takes the form (4.1) where rotations
span the group G = O(n) while translations belong to A = Rn , with the action
σ of rotations on translations given by the vector representation of O(n).
• The Poincaré group in D space-time dimensions takes the form (4.1) where rota-
tions and boosts span the Lorentz group O(D−1, 1) while space-time translations
span A = RD (which is sometimes written RD−1,1 ); the action σ is the vector
representation of the Lorentz group.
• The BMS groups (1.1)-(1.9) all take the form (4.1) with G a specific non-Abelian
group and A an Abelian vector group of so-called “supertranslations”. A similar
structure will hold in three space-time dimensions.
Note that the definition of G n A singles out the normal subgroup A, so G and A
live on unequal footings. In particular the Lie algebra of G n A contains a non-trivial
Abelian ideal and is not semi-simple. This implies that, in contrast to simple Lie
groups, the representations of G n A must somehow distinguish the roles of G and A
by making them act on the carrier space in radically different ways. We will illustrate
this in the pages that follow (see e.g. formula (4.23)).
4.1.2 Momenta
Suppose we wish to build unitary representations of a semi-direct product G nσ A.
Where should we start? A simple approach is to note that the restriction to A of any
unitary representation of GnA is a (reducible) unitary representation of A. So instead
of directly looking for representations of G n A, let us consider the simpler problem of
building unitary representations of the group of translations, A.
We denote by A∗ the vector space dual to A. It consists of linear forms p : A →

R : α 7→ hp, αi, which motivates the definition of a bilinear pairing
h·, ·i : A∗ × A → R : (p, α) 7→ hp, αi. (4.7)
Since A is Abelian, any one of its irreducible unitary representations is one-dimensional

and takes the form
R : A → C : α 7→ eihp,αi (4.8)
for some fixed element p of A∗ . Indeed, any representation R of A is such that
R[α + β] = R[α] R[β] (4.9)
where the right-hand side is a composition of linear operators. Assuming that A has
a countable basis so that α and β have components αi , β i , the derivative of (4.9) with
respect to β i yields
∂j R[α] = i − i∂j R[0] R[α] (4.10)
where each (−i∂j R[0]) is Hermitian by unitarity. Hence R[α] = exp[i(−i∂j R[0])αj ],
which can be diagonalized into a direct sum of multiplicative operators (4.8).
For example, for the Euclidean group in n dimensions, α = (α1 , ..., αn ) is an n-

component vector and hp, αi = pi αi where p = (p1 , ..., pn ) is a “covector”. For the
Poincaré group in D space-time dimensions, α = (α0 , ..., αD−1 ) is a D-vector and

hp, αi = pµ αµ for some energy-momentum covector (p0 , ..., pD−1 ). When interpreting
the corresponding unitary representations as “particles”, the quantity p represents the
particle’s momentum vector.1 Accordingly, from now on the dual space A∗ will be
called the space of momenta, and its elements will be denoted as p, q or k. In the
BMS3 groups, translations and momenta are vectors with infinitely many components.
Note that two irreducible representations of the form (4.8) are equivalent if and only
if their momenta coincide.
Remark. In proving that all irreducible unitary representations of A takes the form
(4.8), we relied crucially on eq. (4.9). The latter assumes that R is an exact repre-
sentation of A, which is not a restrictive assumption as long as there exists no central
extension of G n A that turns A into a non-Abelian group. The Poincaré groups, the
Bargmann groups and the BMS3 group all satisfy this property, so one may safely
assume that A is Abelian even upon switching on central extensions. By contrast, the
symmetry group of warped conformal field theories [93] is a semi-direct product whose
central extension makes translations non-Abelian [52].
4.1.3 Orbits and little groups

We now ask how irreducible, unitary representations of the Abelian group A are em-
bedded in unitary representations of the larger group G nσ A. Let T be a unitary
representation of the latter; then its restriction to A is, in general, reducible. It is
typically a direct sum, or rather a direct integral, of irreducible representations (4.8):2
Z
T [(e, α)] = dµ(q) eihq,αi Iq ∀ α ∈ A. (4.11)
O
∗
Here O is a certain subset of A , µ is some measure on O, and each Iq is an identity
operator acting in a suitable Hilbert space at momentum q. The question then is:
What is the minimal set O of momenta
(4.12)
appearing in the decomposition (4.11)?
The answer will lead to the notion of orbits; hence the notation “O” in (4.11).
Call H the Hilbert space of the representation T . Suppose there exists a subspace
E of H where translations are represented by multiplicative operators (4.8) with a
certain momentum p:
T [(e, α)] = eihp,αi IE ∀ α ∈ A, (4.13)

E
where IE is the identity operator in E. We shall refer to this property by saying that the
representation T “contains the momentum p”. Now pick some group element f ∈ G.
By virtue (4.6) and since T is a representation, one has
T [(e, α)] · T [(f, 0)] = T [(f, 0)] · T [(e, σf −1 α)]. (4.14)
1
More precisely the momentum vector is obtained by raising the indices of the covector p thanks
to some metric on A left invariant by G, but we will keep referring to p as the “momentum vector”.
2
Our notation here is not mathematically precise; we refer to [62] for a more rigorous treatment.
One can then act with both sides of this equation on the space E; the last term on the
right-hand side produces a multiplicative operator (4.13) with α replaced by σf −1 α.
This operator is a c-number and therefore commutes with T [(f, 0)]. We conclude that,
on the space T [(f, 0)] · E ≡ E 0 , all translations are again represented by multiplicative
operators, but now with an additional insertion of σf −1 in the phase hp, αi:
T [(e, α)] = eihp,σf −1 αi IE 0 ∀ α ∈ A. (4.15)

E0
This motivates the following definition for the action of boosts on momenta:
Definition. For any momentum p ∈ A∗ and any f ∈ G, we write
σf∗ (p) ≡ p ◦ σf −1 , (4.16)
i.e. hσf∗ (p), αi ≡ hp, σf −1 αi for all translations α. This defines a representation σ ∗ of
G in the space of momenta, known as the dual representation corresponding to σ. To
reduce clutter, we will often denote it by
σf∗ (p) ≡ f · p . (4.17)
In terms of the dual representation (4.17) we can rewrite (4.15) as T [(e, α)] E 0 =
ihf ·p,αi
e IE 0 , where E 0 = T [(f, 0)] · E. Thus, whenever the representation T contains a
momentum p, compatibility with the structure of G n A implies that it also contains
the boosted momentum f · p, where f is any element of G. This is the answer to
the question (4.12): if there exists a momentum p such that (4.13) holds, then the
representation also contains all momenta that belong to the orbit (3.13) of p under G,
Op ≡ {f · p | f ∈ G} . (4.18)
This orbit is the minimal set of momenta needed to cook up a representation of G n A;

we will see below that it is also sufficient. In fact, the whole classification of irre-
ducible unitary representations of G n A will be provided by a partition of the space
of momenta into G-orbits. Note that this partition is scale-invariant in the following
sense: since the action of boosts on momenta is linear, the orbits Op and Oλp are
diffeomorphic for any real number λ 6= 0.
Each orbit Op is a homogeneous space for the action (4.16) of G. Accordingly we

define the little group of a momentum p as the set of rotations that leave it fixed,
Gp ≡ {f ∈ G | f · p = p} . (4.19)
It is the stabilizer (3.14) for the action of G on the homogeneous space Op . As in

(3.15) there is a diffeomorphism Op ∼ = G/Gp . Note that the little group of the vanish-
ing momentum p = 0 is the whole group G.
The notion of orbits is perhaps the one most important concept needed to under-
stand representations of semi-direct products. We will encounter it repeatedly later
on. Orbits hint at a geometrization of representation theory analogous to the one men-
tioned at the end of section 3.3, and therefore suggest that representations of G n A
are closely related to induced representations. In the next pages we will confirm this
intuition by showing how to associate representations of G n A with a given orbit.
Note that the action (4.16) of G on the space of momenta leaves the pairing (4.7) in-
variant in the sense that hf ·p, σf αi = hp, αi. This has an important implication: when
A is finite-dimensional it is isomorphic to its dual, so (4.7) defines a non-degenerate
bilinear form on A and the action σ ∗ of G on momenta is equivalent to σ. We shall see il-
lustrations of this in the Poincaré groups. By contrast, when A is infinite-dimensional,
σ ∗ may not be equivalent to σ despite the property hf · p, σf αi = hp, αi. This obser-
vation will be relevant to the BMS3 group in part III.
4.1.4 Particles
We now explain how to build irreducible unitary representations of GnA starting from
a momentum orbit Op . Inspired by the Poincaré group, we refer to such representations
as particles. We start by describing scalar particles and identify them with induced
representations of G n A. This identification will then allow us to introduce spin.
Scalar particles
Let p ∈ A∗ be a momentum with orbit (4.18). The latter is a homogeneous space
and therefore admits a quasi-invariant measure µ. Let then H = L2 (Op , µ, C) be the
Hilbert space of square-integrable wavefunctions in momentum space,
Ψ : Op → C : q 7→ Ψ(q). (4.20)
The scalar product of wavefunctions is (3.7), with (Φ(q)|Ψ(q)) = Φ∗ (q)Ψ(q).
Now let us endow H with a unitary action T of G n A. In other words, if Ψ ∈ H

is a wavefunction, we wish to define the object
T [(f, α)] · Ψ (4.21)
where (f, α) belongs to G n A and where T [(f, α)] is some unitary operator. Linearity
implies that the result should be proportional to Ψ, so

T [(f, α)] · Ψ (q) = (some number) × Ψ(some point on Op )
where the unknown quantities may depend on q, f and α. Note that the quantity
multiplying Ψ(· · · ) on the right-hand side must be a number, as opposed to an operator,
because Ψ takes its values in C (this will change upon adding spin). Now recall that
the reason for introducing orbits in the first place was to represent translations by
multiplicative operators (4.8). Accordingly the translation α in (4.21) should produce
a momentum-dependent phase factor:
T [(f, α)] · Ψ (q) = eihq,αi × Ψ(some point on Op ) .

Finally, since Ψ is a wavefunction in momentum space, its argument on the right-hand

side should represent the fact that a boost f maps a particle with momentum k on
a particle with momentum f · k. This is exactly the situation encountered in the
quasi-regular representation (3.22) so we can borrow that construction:
T [(f, α)] · Ψ (q) = eihq,αi Ψ(f −1 · q) .

(4.22)
In particular, the intuition depicted in fig. 3.1 remains valid.
Formula (4.22) defines a representation T of G n A, as can be verified by following

the same steps as for the quasi-regular representation (3.22). It is also irreducible by
virtue of the fact that the orbit Op is a homogeneous space. Finally, it is unitary if
the measure µ in (3.7) is invariant under G. If the measure has a non-trivial Radon-
Nikodym derivative (3.19), the representation (4.22) can be made unitary by inserting
a compensating term in front of the exponential, as in (3.24):
q
T [(f, α)] · Ψ (q) = ρf −1 (q) eihq,αi Ψ(f −1 · q).

(4.23)
We call this representation a scalar particle with momentum orbit Op . Note how
translations and rotations have radically different roles: translations multiply wave-
functions by momentum-dependent phase factors, while boosts move them around on
the orbit by changing their argument. In particular, pure translations act as
T [(e, α)] · Ψ(q) = eihq,αi Ψ(q). (4.24)
Thinking of wavefunctions as sections of a complex line bundle over Op with fibres

Eq ∼
= C, formula (4.24) can be rewritten symbolically as
Z
T [(e, α)] = dµ(q) eihq,αi Iq (4.25)
Op
where Iq is the identity operator in the fibre at q. This is precisely the anticipated
expression (4.11).
Figure 4.1: A momentum orbit Op crossed by one-dimensional fibres isomorphic to C.

(For simplicity the fibres are depicted as if they were real rather than complex.) The
fibre at q ∈ Op is denoted Eq and the disjoint union of such fibres is a (complex) line
bundle over Op . A wavefunction Ψ is a section of that bundle. Translations (4.25) act
by complex multiplication z 7→ eihq,αi z in each fibre Eq .
Particles are induced representations

Formula (4.23) is almost identical to the quasi-regular representation (3.24), and more
generally to the induced representation (3.32). To investigate this relation, let Gp be
the little group (4.19) of p and consider the subgroup Gp n A of G n A. Define a map
S : Gp n A → C : (f, α) 7→ S[(f, α)] ≡ eihp,αi , (4.26)
which is a one-dimensional representation of Gp n A. Indeed, for all (f, α) and (g, β)

belonging to Gp n A, S preserves the group structure in the sense that
(4.26) f ∈Gp −1 ·p,βi
S[(f, α)] · S[(g, β)] = eihp,αi+ihp,βi = eihp,αi+ihf
(4.16) (4.26) (4.6)
= eihp,(α+σf β)i = S[(f g, α + σf β)] = S[(f, α) · (g, β)].
Furthermore, (4.26) is unitary so we can use it to induce a unitary representation
T = IndGnA
Gp nA (S) (4.27)
of G n A. Using the general formula (3.32) and the diffeomorphisms Op ∼ = G/Gp ∼ =

(G n A)/(Gp n A), we see that the induced representation (4.27) acts on wavefunctions
exactly in the way displayed in eq. (4.23). Note that the little group Gp is represented
trivially in the “spin” representation (4.26). This is why we say that the particle (4.23)
is scalar : its states are essentially unaffected by the rotations that span Gp . The
picture (4.27) suggests a simple generalization of this behaviour, as we now explain.
Spinning particles
To generalize (4.23), let R be an irreducible, unitary representation of Gp in some
space E and consider the spin representation
S : Gp n A → GL(E) : (f, α) 7→ eihp,αi R[f ] . (4.28)
This reduces to (4.26) when R is trivial, and the corresponding induced representation
of G n A is
GnA GnA ihp,·i

T = IndG p nA
(S) = Ind Gp nA e R . (4.29)
Its action on wavefunctions is analogous to (3.32) and generalizes (4.23):
q
(T [(f, α)] · Ψ) (q) = ρf −1 (q) eihq,αi R[gq−1 f gf −1 ·q ] · Ψ(f −1 · q) , (4.30)
where the map g : Op → G : q 7→ gq is a continuous family of standard boosts (3.31).

In contrast to (4.23), Ψ now takes its values in E rather than C.
We call the representation (4.30) a spinning particle with spin R and momenta
belonging to Op . It is an irreducible unitary representation of G n A acting on the
Hilbert space H = L2 (Op , µ, E). The operator
R[gq−1 f gf −1 ·q ] ≡ Wq [f ] (4.31)
is the Wigner rotation (3.35) associated with f and q. It is the transformation that
corresponds to f in the space of internal degrees of freedom E at q and it entangles
momentum and spin degrees of freedom. The decomposition (4.25) still holds in the
spinning case, with Iq the identity operator in the fibre Eq ∼
= E at q.
From this point on, the whole machinery of induced representations applies to
unitary representations (4.30) of G n A. In particular they are independent of the
choice of the quasi-invariant measure µ and of the family of standard boosts gq . The
plane waves (3.43) provide a basis of the Hilbert space and represent one-particle states
with definite momentum and definite spin. They transform under G n A according to
q
T [(f, α)] · Ψk,` = ρf (k) eihf ·k,αi R gf−1

·k · f · gk · Ψf ·k,` . (4.32)
This is just formula (3.48) applied to (4.29); it is the plane wave analogue of (4.30).
Using this, one can go on and evaluate characters along the lines that led to the
Frobenius formula (3.54). One finds
Z
dµ(k) δ(k, f · k) eihk,αi χR [gk−1 f gk ] ,

χ[(f, α)] = Tr T [(f, α)] = (4.33)
Op
where χR is the character of the representation R of Gp . As before one can check

that this formula defines a class function and that χ[(f, α)] vanishes when f is not
conjugate to an element of the little group. The delta function localizes the integral
to the momenta that are left fixed by the action of f on Op .
4.1.5 Exhaustivity theorem

Eq. (4.30) is an irreducible unitary representations of G n A. As it turns out, all
irreducible representations of G n A take this form for some momentum orbit Op and
some spin R. We refer to this property as the exhaustivity theorem for induced repre-
sentations.
This theorem has enormous practical value: it provides the classification of all
irreducible unitary representations of a semi-direct product G nσ A when A is a vector
group. This classification can be performed thanks to the following algorithm:
1. Consider the space of momenta, A∗ . For each p ∈ A∗ , determine the orbit Op

given by (4.18). This foliates A∗ into disjoint momentum orbits, and each point
of A∗ belongs to exactly one orbit.
2. We call set of orbit representatives a set of momenta that exhaust all orbits
in a non-redundant way, in the sense that (i) each orbit contains one of the
representatives, and (ii) different representatives belong to different orbits. Find
a set of orbit representatives, compute the little group of each representative, and
find standard boosts connecting each representative to the points of its orbit.
3. For each representative p with little group Gp , classify all irreducible unitary
representations of Gp . Given such a representation R, the associated induced
representation of G n A is (4.30).
We will illustrate this classification for the Poincaré groups in section 4.2 and for the
Bargmann groups in section 4.4, and of course for BMS3 in chapter 10.
The proof of the exhaustivity theorem is essentially an upgraded version of our

arguments in section 4.1.3 and relies on two crucial ingredients: the first is the com-
mutativity of the vector group A, and the second is the imprimitivity theorem of
section 3.5. Thanks to commutativity, any unitary representation of A can be written
as a direct integral (4.11) of irreducible representations specified by certain momenta
q ∈ A∗ . (This is known as the snag theorem.) This implies that any unitary repre-
sentation T of G n A is imprimitive. Indeed, relation (4.14) can be rewritten as
T [(f, 0)] · T [(e, α)] · T [(f, 0)]−1 = T [(e, σf α)]
whereupon the direct integral representation (4.11) yields
T [(f, 0)] · dµ(q)Iq · T [(f, 0)]−1 = dµ(f · q)If ·q , (4.34)
which is precisely the statement (3.62) that the projection-valued measure dµ(q)Iq is
a system of imprimitivity for T on A∗ . The imprimitivity theorem then implies that
the representation T is induced. The last step of the proof consists in showing that,
if T is irreducible, then the measure µ in (4.34) localizes to a single momentum orbit.
We refer to [62] for details.
Remark. The exhaustivity theorem relies on an extra technical assumption that we

haven’t mentioned so far. Namely, one says that G n A is regular if the space of
momenta A∗ and the action (4.16) of G are such that A∗ contains a countable family
of Borel sets, each a union of momentum orbits, such that each orbit is the limit of
a decreasing sequence of such sets. As it turns out regularity is necessary for the
measure µ in (4.34) to be localized on a momentum orbit. All semi-direct products
treated in part I of this thesis are regular. As for the BMS3 group of part III, the issue
of regularity will be discussed briefly in section 10.1.
4.2 Poincaré particles

In this section and the next ones we study examples of semi-direct products to illus-
trate induced representations. Here we deal with the Poincaré group — the isometry
group of Minkowski space — whose representations describe relativistic particles. Fol-
lowing the algorithm of page 70 we will find that these particles are classified by two
parameters: their mass and their spin. In view of treating the BMS3 group in part
III, we relegate the detailed description of relativistic particles in three dimensions to
section 4.3.
The plan is the following. First we define the Poincaré group as a semi-direct
product of the Lorentz group with the group of space-time translations. Then we turn
to the classification of its momentum orbits and describe the corresponding particles.
We also compute their characters and end with the observation that Lorentz transfor-
mations generally entangle momentum and spin degrees of freedom.
The classification of relativistic particles was first performed by Wigner [23], and
their relation to wave equations was worked out in [94]. These results are among the
foundations of quantum mechanics and field theory; see e.g. [53, 62, 91, 92].
4.2.1 Poincaré groups

Lorentz transformations
We consider the vector space RD ; its elements are column vectors α, β, etc. with
components αµ , β µ where µ = 0, 1, ..., D − 1. Here RD is to be interpreted as a space-
time manifold with dimension D ≥ 2. We endow RD with a non-degenerate bilinear
form given by the Minkowski metric,
 
−1 0 0 · · · 0
 0 1 0 · · · 0
 
µ ν  0 0 1 · · · 0
(α, β) ≡ ηµν α β , (ηµν ) =  . (4.35)
 .. .. .. . . .. 
 . . . . .
0 0 0 ··· 1
We write (α, α) ≡ α2 for any α. The sign of α2 determines whether α is time-like, null
or space-like, corresponding respectively to α2 < 0, α2 = 0 or α2 > 0.
Definition. The Lorentz group O(D − 1, 1) in D dimensions is the group of linear

transformations RD → RD : α 7→ f · α that preserve (4.35) in the sense that
(f · α, f · β) = (α, β). (4.36)
It consists of D × D matrices f = (f µ ν ) such that
f t · η · f = η, i.e. f λ µ ηλρ f ρ ν = ηµν (4.37)
where the dot denotes matrix multiplication. In particular, the Minkowskian norm is
left invariant by Lorentz transformations.
Topology of Lorentz groups

The Lorentz group O(D − 1, 1) is disconnected. Indeed, any Lorentz matrix f has
determinant det(f ) = ±1. This cuts the group in two pieces consisting of matrices
with positive and negative determinant, corresponding to transformations that pre-
serve or break (respectively) the orientation of the spatial coordinates. The subgroup
of O(D − 1, 1) consisting of Lorentz matrices with positive determinant is the proper
Lorentz group, SO(D − 1, 1). Any improper Lorentz matrix is the product of a proper
Lorentz transformation with parity. In addition, one can show that any Lorentz ma-
trix f satisfies |f 0 0 | ≥ 1, which again cuts the Lorentz group in two pieces: matrices
with positive or negative f 0 0 , corresponding to transformations that preserve or invert
(respectively) the orientation of the arrow of time. The subgroup consisting of Lorentz
transformations with positive f 0 0 is the orthochronous Lorentz group O(D − 1, 1)↑ .
Any Lorentz matrix that reverts the arrow of time is the product of an orthochronous
Lorentz matrix with time reversal. The situation is depicted in fig. 4.2.
In this section we focus on the connected Lorentz group, i.e. the proper orthochro-
nous Lorentz group SO(D −1, 1)↑ . The latter satisfies an important property known as
standard decomposition: any proper, orthochronous Lorentz transformation is a prod-
uct f = R1 · Λ · R2 , where R1 and R2 are spatial rotations and Λ is a pure boost [95].
In what follows we often refer to SO(D − 1, 1)↑ simply as “the Lorentz group”.
The Lorentz group is not simply connected: in space-time dimension D ≥ 4, its

fundamental group is isomorphic to Z2 . The universal cover of the connected Lorentz
group is then called the spin group, so that
SO(D − 1, 1)↑ ∼= Spin(D − 1, 1)/Z2 (4.38)
where the Z2 subgroup of Spin(D−1, 1) consists of the identity matrix and its opposite.
In four dimensions, Spin(3, 1) = SL(2, C). In D = 3 dimensions the situation is a bit
different; we will return to it in the next section. In any case the Lorentz group is always
multiply connected, and therefore admits topological projective representations; this
will be important for representations of Poincaré.
Poincaré groups
Definition. The Poincaré group or inhomogeneous Lorentz group in D space-time
dimensions is the semi-direct product
IO(D − 1, 1) ≡ O(D − 1, 1) n RD (4.39)
Figure 4.2: The four connected components of the Lorentz group. The upper left
component is the proper orthochronous Lorentz group SO(D−1, 1)↑ . It can be mapped
on the other components using parity and time reversal. In particular the proper
Lorentz group is generated by SO(D − 1, 1)↑ together with time reversal, while the
orthochronous Lorentz group is generated by SO(D − 1, 1)↑ together with parity.
whose elements are pairs (f, α) where f is a Lorentz transformation, α a space-time

translation. The group operation is (f, α) · (g, β) = (f · g, α + f · β) where the dots on
the right-hand side denote matrix multiplication and the action of matrices on column
vectors. The connected Poincaré group is the largest connected subgroup of (4.39),
ISO(D − 1, 1)↑ ≡ SO(D − 1, 1)↑ n RD , (4.40)
and its universal cover is

Spin(D − 1, 1)↑ n RD (4.41)
D
where spin transformations act on R according to the composition of the homomor-
phism given by (4.38) with the vector representation of the Lorentz group.
The Poincaré group turns out to have no algebraic central extensions, so its only
non-trivial projective transformations are of topological origin. The Poincaré Lie al-
gebra is generated by D(D − 1)/2 Lorentz generators and D translation generators;
we will not display their brackets here.

From now on we focus on the connected Poincaré group (4.40), to which we refer
simply as “the Poincaré group”.
Momenta and orbits

The space of Poincaré momenta is (RD )∗ = RD ; its elements are D-dimensional cov-
ectors p = (p0 , p1 , ..., pD−1 ), where p0 is to be interpreted as the energy of a relativistic
particle, while p = (p1 , ..., pD−1 ) is its spatial momentum.3 Given a momentum p and
3
Strictly speaking the energy of the particle is p0 = −p0 , but this detail will not affect our
discussion so we neglect it for simplicity.
a space-time translation α, the pairing (4.7) is hp, αi = pµ αµ .
The Minkowski metric (4.35) provides a Lorentz-invariant pairing between trans-

lation vectors and can be used to define an isomorphism
I : RD → (RD )∗ : α 7→ (α, ·) (4.42)
where the components of (α, ·) are those of α lowered with the Minkowski metric.
Using I, one verifies that the action σ ∗ of Lorentz transformations on momenta is
equivalent to their action σ on translations:
σf∗ = I ◦ σf ◦ I −1 . (4.43)
As a consequence, momentum orbits coincide with orbits of translations under Lorentz

transformations, and consist of momenta q with constant Minkowskian norm squared
q 2 . We thus conclude that
the orbits of momenta of relativistic particles are connected hyperboloids
specified by an equation of the form q02 − q2 = const. in RD .
The only exception to this rule is the trivial orbit of the vanishing momentum p = 0,
which contains only one point. The word “connected” appears here because we are
dealing with the connected Poincaré group (4.40). By contrast the momentum orbits
of (4.39) are generally disconnected.
The connected Poincaré group has six distinct families of momentum orbits, which
we now describe. Further details can be found e.g. in [91, 92].
• Let p = 0 be the vanishing momentum. Its orbit O0 = {0} contains a single
point. Its little group is the whole Lorentz group.
• Let p be a timelike momentum with positive energy, p0 > 0. Its orbit Op is
massive with positive energy and consists of momenta q satisfying
q02 − q2 = M 2 > 0, q0 > 0, (4.44)
where we have introduced the mass squared M 2 ≡ −p2 . We can choose as orbit
representative the rest frame momentum
p = (M, 0, ..., 0), M > 0. (4.45)
The little group of (4.45) is the group of spatial rotations
Gp = SO(D − 1) (4.46)
consisting of proper Lorentz transformations that leave the time coordinate fixed.
In particular, the orbit is diffeomorphic to the quotient
Op ∼
= SO(D − 1, 1)↑ /SO(D − 1) ∼
= RD−1 (4.47)
and its points can be labelled by the spatial components of momentum (since
the zeroth component is then determined by eq. (4.44)). Massive orbits with
different masses are disjoint.
• Let p be a time-like momentum with negative energy, p0 < 0. Its orbit is massive
with negative energy and consists of momenta q satisfying (4.44) with q0 < 0.
A typical orbit representative is (4.45) with M < 0, and orbits with different
masses are disjoint. The little group is a again SO(D − 1).
• Let p be a null momentum with positive energy, p0 > 0. Its orbit Op is massless
with positive energy. It consists of momenta q satisfying (4.44) with M 2 = 0. A
typical orbit representative is
p = (E, E, 0, ..., 0) (4.48)
where the energy E is positive; different values of E yield the same orbit. Note
that there is no rest frame for massless particles. The little group of (4.48) is
isomorphic to the Euclidean group
Gp ∼
= SO(D − 2) n RD−2 = ISO(D − 2). (4.49)
In particular, the orbit is diffeomorphic to the quotient
Op ∼
= SO(D − 1, 1)↑ /ISO(D − 2) ∼
= R × S D−2 (4.50)
and its points can be labelled by the spatial components of momentum (since
the zeroth component is then determined by q 2 = 0).
• Let p be a null energy-momentum vector with negative energy. Its orbit is

massless with negative energy and consists of null momenta q with q0 < 0. A
typical orbit representative is (4.48) with negative E. The little group is (4.49)
and the orbit can be represented as a quotient (4.50).
• Let p be a space-like momentum. Its orbit is tachyonic and consists of momenta

q satisfying (4.44) with M 2 < 0. A typical orbit representative is
√
p = 0, 0, ..., 0, −M 2 . (4.51)
The little group is the lower-dimensional Lorentz group SO(D − 2, 1)↑ consisting
of transformations that leave the spatial coordinate xD−1 fixed. In particular the
orbit is diffeomorphic to the quotient
O∼
= SO(D − 1, 1)↑ /SO(D − 2, 1)↑ ∼
= R × S D−2 . (4.52)
Tachyonic orbits with different negative values of M 2 are disjoint. Note that
rotations always allow us to map p on −p, which is why any tachyonic orbit
representative can be written as (4.51).
This enumeration exhausts all Poincaré momentum orbits. Among the six families of
orbits, three contain only one orbit: the trivial orbit and the two massless orbits. The
remaining three families all contain infinitely many orbits labelled by a non-vanishing
mass squared, corresponding to massive particles and tachyons. These orbits and their
representatives are schematically depicted in fig. 4.3.
(a) (b)
Figure 4.3: On the left, fig. (a) represents a few momentum orbits of the Poincaré
group in three dimensions, embedded in R3 with the vertical axis corresponding to p0
and the two horizontal axes (not represented in the figure) corresponding to spatial
components of momentum. Orbits can be massive, massless or tachyonic depending
on whether M 2 is positive, vanishing or negative, respectively. The cross in the middle
is the trivial orbit of p = 0, consisting of a single point. On the right, fig. (b) is a
schematic representation of momentum orbits: each point of the diagram corresponds
to an orbit representative, where massive orbits are represented by a vertical line,
tachyonic ones by a horizontal line, and discrete orbits (the two massless ones and the
trivial one) by dots. This schematic representation will be useful in parts II and III
for the interpretation of BMS3 supermomentum orbits.
To complete the description of orbits we now display standard boosts for massive
particles (the other cases are less important for our purposes so we skip them). We
take as orbit representative the momentum
p (4.45) of a particle at rest, and look for a
family of boosts gq such that gq · p = ( M 2 + q2 , q) that depend continuously on q.
One readily verifies that the matrices [96]
p !
1 + q2 /M 2 qj /M
gq = qq
p (4.53)
qi /M δij + qi 2j 1 + q2 /M 2 − 1
satisfy these requirements. Here i, j = 1, ..., D − 1 are spatial

p indices. Each such
matrix is a boost in the direction q/|q| with rapidity arccosh[ 1 + q2 /M 2 ].
Remark. The little groups displayed in (4.46) and (4.49) hold for the connected
Poincaré group (4.40). If we replace the latter by its universal cover (4.41), then the
little groups are replaced by their double covers (assuming that D ≥ 4). In particular
the little group of massive particles becomes Spin(D−1) while that of massless particles
becomes Spin(D − 2) n RD−2 , with the convention that Spin(2) is the double cover of
SO(2). Note that Spin(3) = SU(2).
4.2.3 Particles
According to the exhaustivity theorem of section 4.1.5, the momentum orbits in fig.
4.3 roughly classify relativistic particles. The states of each particle are wavefunc-
tions on its momentum orbit, valued in a spin representation of the little group and
transforming under Poincaré transformations according to formula (4.30). Provided
we know all irreducible unitary representations of all little groups, we have effectively
classifed all irreducible unitary representations of the Poincaré group.
Vacuum
Vacuum representations of Poincaré are those whose orbit O0 = {0} is trivial and is
left invariant by the whole Lorentz group. In that case a spin representation is a (pro-
jective) irreducible unitary representation R of SO(D − 1, 1)↑ . The latter is simple but
non-compact, so its only finite-dimensional irreducible unitary transformation is the
trivial one; the corresponding induced representation of Poincaré is trivial as well. All
other irreducible unitary representations of the Lorentz group are infinite-dimensional;
the corresponding induced representations of Poincaré are such that translations act
trivially, while Lorentz transformations act non-trivially on an infinite-dimensional
Hilbert space E of spin-like degrees of freedom. These representations can be inter-
preted as “vacua with spin” but are generally discarded as unphysical.
Massive particles
The momenta of a massive particle with mass M span an orbit (4.44) with little group
(4.46). The spin representation R then is a finite-dimensional, irreducible, generally
projective unitary representation of SO(D − 1) specified by some highest weight λ.
For example, when D = 4, R is a highest-weight representation of SO(3) = SU(2)/Z2
with spin s ≥ 0; the latter is either an integer or a half-integer. The carrier space
of R has dimension 2s + 1 and is generated by states | − si, | − s + 1i, ..., |s − 1i, |si
with definite spin projection along a prescribed axis. In that case the highest weight
λ coincides with s. The higher-dimensional case is analogous except that the number
of coefficients specifying λ is the rank b(D − 1)/2c of SO(D − 1). We will illustrate
this point in section 11.1 when dealing with partition functions of higher-spin fields in
Minkowski space.
Given a spin representation, the remainder of the construction is straightforward:

formula (3.4) yields a Lorentz-invariant measure that can be used to define scalar
products (3.7) of wavefunctions, and the Poincaré representation acts according to
(4.30) with the Radon-Nikodym derivative set to ρf = 1 thanks to the choice of
measure. For example, when D = 4 and s = 1/2, any state takes the form (3.9).
Massless particles
The description of massless particles is analogous to that of massive ones, up to the
key difference that the massless little group is the Euclidean group (4.49). It is a
semi-direct product (4.1) with an Abelian normal subgroup, so the exhaustivity theo-
rem ensures that its irreducible unitary representations are induced and classified by
momentum-like orbits of their own. From the Poincaré viewpoint each induced repre-
sentation of (4.49) is a spin representation for a massless particle.
In that context one makes the distinction between two types of massless particles:
particles with discrete spin are those given by spin representations of (4.49) with van-
ishing Euclidean momentum. These are Euclidean analogues of the “spinning vacua”
described earlier, except that they are finite-dimensional. They amount to making
the action of RD−2 in (4.49) trivial, and coincide with (projective) irreducible uni-
tary representations of SO(D − 2). Thus massless particles with discrete spin have
a finite-dimensional space of spin degrees of freedom. By contrast, massless particles
with continuous or infinite spin are those whose spin representations of (4.49) have
non-trivial Euclidean momentum. The space of spin degrees of freedom is infinite-
dimensional in that case, since it consists of wavefunctions on a Euclidean momentum
orbit SO(D − 2)/SO(D − 3) ∼ = S D−3 . Particles with continuous spin are generally
discarded on the grounds that they are unphysical, although they have recently been
described in a field-theoretic framework [97–99].
Tachyons
Tachyons are particles moving faster than light. Their little group is SO(D − 2, 1)↑ . It
is simple and non-compact, so tachyons either have no spin at all, or have continuous
spin. They are generally considered as unphysical.
4.2.4 Massive characters

Having completed the enumeration of relativistic particles, we now evaluate characters
of massive irreducible unitary representations of the Poincaré group. Massless charac-
ters are relegated to section 4.2.5. These computations are important for our purposes,
as we will rely on them in chapter 11. To our knowledge, Poincaré characters were
first studied in [80, 100] before reappearing more recently in [48, 50, 101].
Setting the stage

By virtue of the Frobenius formula (4.33), the character of an induced representation
vanishes when evaluated on a transformation f that does not belong to the little group.
Since the character is a class function, only the conjugacy class of f matters for the
final result. Accordingly, for a massive particle in D dimensions we let f be a rotation

 
1 0 0 ··· 0 0 0
0 cos θ1 − sin θ1 · · · 0 0 0
 
0 sin θ1 cos θ1 · · · 0 0 0
 
 .. .
. .
. . . .
. .
. .
.
f = . (4.54)

. . . . . . 
 
0
 0 0 · · · cos θr − sin θr 0
0 0 0 · · · sin θr cos θr 0
0 0 0 ··· 0 0 1
written here for even D with r = b(D − 1)/2c; if D is odd we erase the last row and
the last column. We assume for simplicity that all angles θ1 , ..., θr are non-zero.
Now let µ be a quasi-invariant measure on a momentum orbit with mass M , and

let δ be the corresponding delta function. Given an arbitrary space-time translation
α, our goal is to evaluate the character of (f, α) using eq. (4.33). To do this we treat
separately odd and even space-time dimensions.
Odd dimensions
For odd D we erase the last row and column of (4.54). Then the Frobenius formula
(4.33) localizes to the unique rotation-invariant point of the momentum orbit, namely
the momentum at rest p = (M, 0, 0, ..., 0). This allows us to simplify (4.33) by setting
k = p in the exponential and the little group character, and pulling them out of the
integral. Denoting by λ the spin of the particle (it is a highest weight for SO(D − 1)),
we find Z
iM α0 (D−1)
χ[(f, α)] = e χλ [f ] dµ(k) δ(k, f · k) (4.55)
Op
where the replacement of k by p has projected the translation α on its time component
(D−1)
α0 . The little group character χλ [f ] is some function of the angles θ1 , ..., θr that we
do not need to write down at this stage (in practice it follows from the Weyl character
formula and is displayed in eq. (11.151) below). To obtain (4.55) it only remains to
evaluate the integral of the delta function. As coordinates on the orbit we choose the
spatial components of momentum, in terms of which the Lorentz-invariant measure on
Op is (3.4) and the corresponding delta function is (3.42). We thus get
dD−1 k p 2
dµ(k) δ(k, q) = p M + k2 δ (D−1) (k − q) = dD−1 k δ (D−1) (k − q) , (4.56)
2 2
M +k
where the multiplicative factors of the measure and its delta function cancel out. Note
that the same cancellation would have taken place for any measure µ proportional to
dD−1 k, in accordance with the fact that induced representations are insensitive to the
choice of measure. Applied to (4.55), the cancellation (4.56) allows us to write
Z
iM α0 (D−1)
χ[(f, α)] = e χλ [f ] dD−1 k δ (D−1) (k, f · k) (4.57)
RD−1
where f · k denotes the action of the spatial submatrix of (4.54) on k. The integral
can be written as
Z
1
dD−1 k δ (D−1) (I − f ) · k =

(4.58)
RD det(I − f )
where I is the (D − 1)-dimensional identity matrix. In terms of angles θi we find

r r r
Y 1 − cos θj sin θj Y
2
Y
det(I − f ) = = 4 sin θj = |1 − eiθj |2 . (4.59)
− sin θj 1 − cos θj
j=1 j=1 j=1
Plugging this into (4.58), the character (4.57) finally becomes

r
iM α0 (D−1)
Y 1
χ[(f, α)] = e χλ [f ] . (4.60)
j=1
|1 − eiθj |2
Note that for a Euclidean time translation α0 = iβ, this quantity may be seen as
the partition function of a relativistic particle in a rotating frame (albeit with purely
imaginary angular velocity).
Remark. The localization effect (4.58) is a restatement of the Atiyah-Bott fixed

point theorem. In that context the term
1
(4.61)
det(I − f )
is the Lefschetz number of the operator T [(f, α)]. If f was a number e−βω , (4.61)
would coincide with the partition function of a harmonic oscillator with frequency ω
at temperature 1/β.
Even dimensions
For even D the rotation f is exactly given by (4.54). Then the situation is more
complicated because the integral (4.33) localizes to a line rather than a point, as in
fig. 3.2. To make things simple we take α = (α0 , 0, ..., 0) to be a pure time translation.
Formula (4.55) is then replaced by
Z √
(D−1) D−1 iα0 M 2 +k2 (D−1)
χ[(f, α)] = χλ [f ] d ke δ (k − f · k) (4.62)
RD−1
where we have already implemented the simplification (4.56). The SO(D−1) character
(D−1)
χλ has been pulled out of the integral because, for D even, boosts along the
direction kD−1 commute with rotations (4.54). It remains once more to integrate the
delta function in (4.62). As far as the first D − 2 components of k are concerned, the
computation is the same as in the odd-dimensional case and results in a factor (4.58)
given by (4.59). But the last component of k is untouched by (4.54), so (4.62) becomes
r Z +∞ √
(D−1)
Y 1 0 M 2 +k2
χ[(f, α)] = χλ [f ] dk eiα δ (1) (k − k) , (4.63)
j=1
|1 − eiθj |2 −∞
where k ≡ kD−1 . Here the last term is an infrared-divergent factor

Z +∞
1 L
δ(k − k) = dz ≡ (4.64)
2π −∞ 2π
where the length scale L is an infrared regulator. The integral in (4.63) then gives
Z +∞ √
0 2 2
dk eiα M +k = 2M K1 (−iM α0 )
−∞
where K1 is the first modified Bessel function of the second kind. In conclusion we get
r
ML 0 (D−1)
Y 1
χ[(f, α)] = K1 (−iM α ) χλ [f ] , (4.65)
π j=1
|1 − eiθj |2
whose Wick-rotated version can now be seen as the rotating partition function of a
particle trapped in a box of height L.
Time translations
All characters written above diverge when one of the angles θj goes to zero. These
divergences are infrared since they are due to delta functions evaluated at zero in
momentum space, and can be regularized as in (4.64). A case of particular interest is
the character of a pure time translation, whose Wick rotation is a canonical partition
function (3.50). Using once more the Frobenius formula (4.33) and the cancellation
(4.56), and letting α = (α0 , 0, ..., 0) be a pure time translation, we find
Z √
D−1 iα0 M 2 +k2 (D−1)
χ[(e, α)] = N d ke δ (0) (4.66)
RD−1
where N ≡ dim(E) is the number of spin degrees of freedom of the particle. The
infrared-divergent delta function can be seen as the spatial volume of the system,
Z
(D−1) 1 V
δ (0) = D−1
dD−1 x = .
(2π) RD−1 (2π)D−1
Using spherical coordinates we can then rewrite (4.66) as
Z +∞
NV 2π (D−1)/2 D−2
√
iα0 M 2 +k2
χ[(e, α)] = k dk e
(2π)D−1 Γ((D − 1)/2) 0
(D−2)/2
2N V 2πM
= M KD/2 (−iM α0 ) , (4.67)
(2π)D−1 −iα0
where KD/2 denotes once more a modified Bessel function of the second kind. This
is the character of a pure time translation in a massive Poincaré representation. For
α0 = iβ purely imaginary, it becomes the canonical partition function of a massive
relativistic particle,
(D−2)/2
−βH
2N V 2πM
Tr e massive particle
= M KD/2 (βM ). (4.68)
(2π)D−1 β
4.2.5 Massless characters

Characters of massless Poincaré representations can be evaluated along the same lines
as massive ones, but there are subtleties due to the little group (4.49). The latter
admits both finite- and infinite-dimensional irreducible unitary representations, corre-

sponding to massless particles with discrete or continuous spin, respectively. Here we
focus on the discrete case. As in the massive case we treat separately even and odd
dimensions, this time starting with the former. For simplicity we take α to be a pure
time translation.
Even dimensions
For even D the Lorentz transformation (4.54) belongs to the little group of a massless
particle since it leaves invariant the momentum vector (E, 0, ..., 0, E). The character
computation then is the same as in the even-dimensional massive case; formula (4.63)
still holds with M = 0 and χ(D−1) replaced by the character χ(D−2) of a representation
of SO(D − 2) instead of SO(D − 1). Note that for even D these two groups have
the same rank r = b(D − 1)/2c, so there is no restriction on the values of the angles
θ1 , ..., θr (this will change for odd D). Using the regulator (4.64) and the fact that
Z +∞
0 2
dk ei|k|(α +iε) = − 0 ,
−∞ iα
one finds the character
r
iL (D−2) Y 1
χ[(f, α)] = 0
χ λ [f ] iθj |2
. (4.69)
πα j=1
|1 − e
Up to the replacement of D − 1 by D − 2, this is the limit M → 0 of the massive

character (4.65).
Odd dimensions
For odd D the transformation (4.54) (with the last row and column suppressed) is
no longer an element of the little group of (E, 0, ..., 0, E) so its character vanishes if
θr 6= 0. This is consistent with the fact that SO(D − 2) has lower rank than SO(D − 1)
when D is odd. Accordingly we now take θr = 0 in (4.54), being understood that
the last row and column are suppressed. From there on the character computation is
identical to the cases treated above, except that the infrared divergence of the integral
becomes worse and requires two regulators L, L0 :
Z
0
√ LL0
2 2
dkdq eiα k +q δ (1) (k − k)δ (1) (q − q) = − . (4.70)
R2 2π(α0 )2
Massless characters in odd dimension D thus read
r−1
LL0
Y
(D−2) 1
χ[(f, α)] = χλ [f ] − (4.71)
2π(α0 )2 j=1 |1 − eiθj |2
(D−2)
where it is understood that θr = 0 in (4.54) and χλ is a character of SO(D − 2).
Note that this expression is not the massless limit of (4.60) because in general θr 6= 0 in
the latter formula. However, upon setting θr = 0 in (4.57) and regulating the resulting
double infrared divergence as in (4.70), the limit M → 0 does produce an expression of
the form (4.71), albeit with a reducible representation of SO(D − 2). We shall return
to this in section 11.1. Characters of time translations can be treated as in the massive
case and coincide, up to spin multiplicity, with the massless limit of (4.67).
*4.2.6 Wigner rotations and entanglement

We now analyse the Wigner rotation (4.31) and show that, for generic spinning parti-
cles, it entangles momentum and spin degrees of freedom. This phenomenon was first
investigated in [102] (see also [103] and the related considerations in [104, 105]).
Wigner rotations
Consider a particle with mass M and spin representation R. We wish to understand
the action of the Wigner rotation (4.31) for an arbitrary momentum q belonging to its
orbit, and for a boost
 
cosh γ − sinh γ 0 · · · 0
− sinh γ cosh γ 0 · · · 0
 
f =
 0 0 1 · · · 0 
 (4.72)
 .. .. .. . . .. 
 . . . . . 
0 0 0 ··· 1
with rapidity γ in the direction x1 . Since the gq ’s are standard boosts (4.53), the
combination gq−1 f gf −1 ·q is a sequence of three pure boosts in a plane, so we may
safely take D = 3 without affecting the outcome of the computation. The momentum
q then reads p 
M 2 + Q2
q =  Q cos ϕ  (4.73)
Q sin ϕ
for some angle ϕ and some positive number Q. After a mildly cumbersome but straight-
forward computation, one finds a Wigner rotation matrix
 
1 0 0
gq−1 f gf −1 ·q = 0 cos θ − sin θ (4.74)
0 sin θ cos θ
whose entries are given by
M sin(ϕ) sinh(γ)
sin θ = − q p ,
Q2 sin2 (ϕ) + (Q cos(ϕ) cosh(γ) + M 2 + Q2 sinh(γ))2
p (4.75)
Q cosh(γ) + M 2 + Q2 cos(ϕ) sinh(γ)
cos θ = q p .
Q2 sin2 (ϕ) + (Q cos(ϕ) cosh(γ) + M 2 + Q2 sinh(γ))2
This is a pure rotation, as it should. It represents the fact that a boost acting on a
particle with non-zero momentum is seen, from the rest frame of the particle, as a
boost combined with a rotation (4.74) rather than a pure boost. The rotation only
affects spin degrees of freedom; scalar particles are insensitive to it.
The Wigner rotation (4.75) is responsible for the phenomenon of Thomas precession
[106] (see also [96], section 11.8). The latter is visible in atomic physics, where the spin
of an electron orbiting around a nucleus undergoes a slow precession due to the fact
that the electron’s acceleration is a sequence of boosts directed towards the nucleus.
Momentum/spin entanglement
In (3.8) we saw that a space of E-valued wavefunctions on Op is a tensor product of
E with the scalar space L2 (Op , µ, C). For a relativistic particle, the former consists
of spin degrees of freedom while the latter accounts for momenta (or positions after
Fourier transformation). For example, any state of a massive particle with spin 1/2
takes the form (3.9) and describes the separate propagation of the two spin states |+i
and |−i. (For simplicity we use the Dirac notation until the end of this section.)
Hilbert space factorizations such as (3.8) are seldom preserved by unitary maps.
Indeed, if H = A ⊗ B and |Ψi ∈ H is a state with unit norm, the reduced density
matrix associated with |Ψi and acting in B is ρ ≡ TrA |ΨihΨ|. When U is a unitary
operator in H , it is generally not true that the reduced density matrix of U · |Ψi is
unitarily equivalent to ρ. In particular, U does not preserve the degree of entanglement
between A and B. Accordingly one may ask [102] whether Poincaré representations
spoil the splitting (3.8). To answer this, consider for definiteness a massive spin 1/2
particle in four dimensions. We start from a normalized E-valued wavefunction
Ψ(q) = ψ(q)|+i, i.e. |Ψi = |ψi ⊗ |+i (4.76)
where ψ is some complex-valued wavefunction while |+i is one of the two members of
an orthonormal basis |+i, |−i of E. This state represents a particle with spin up (say
along the x3 axis) propagating with a momentum probability distribution dµ(q)|ψ(q)|2 .
(For definiteness we take the measure µ to be the Lorentz-invariant expression (3.4).)
The corresponding reduced density matrix obtained by tracing over spin degrees of
freedom acts on the scalar Hilbert space L2 (Op , µ, C) and reads

ρ = h+| |ψi|+ihψ|h+| |+i + h−| |ψi|+ihψ|h+| |−i = |ψihψ| ,
which is a pure state. Now let us act on (4.76) with a Lorentz transformation f .
According to (4.30), and writing U ≡ T [(f, 0)], the resulting wavefunction is
(4.76)
(U · Ψ) (q) = Wq [f ] · Ψ(f −1 · q) = ψ(f −1 · q) Wq [f ]|+i (4.77)
where Wq [f ] is the Wigner rotation (4.31). Denoting φ(q) ≡ ψ(f −1 · q) we now find
that the entries of the reduced density matrix of (4.77) are
∗ ∗
ρ̃(q, q 0 ) = φ(q)χ+ (q) φ(q 0 )χ+ (q 0 ) + φ(q)χ− (q) φ(q 0 )χ− (q 0 )

(4.78)
where we have defined χ± (q) ≡ h±|Wq [f ]|+i. In general expression (4.78) is not equal
to a product ψ̃(q)ψ̃ ∗ (q 0 ) (for some complex wavefunction ψ̃), so the state (4.78) is not
pure! In particular the boosted state (4.77) is generally entangled with respect to the
splitting (3.8), even though the original state (4.76) was not. The reason for this is
that the Wigner rotation (4.74) generally has non-vanishing +− entries. Note that
the functions χ± satisfy |χ+ |2 + |χ− |2 = 1 by virtue of the fact that Wigner rotations
are unitary, so formula (4.78) indeed defines a density matrix.
These arguments can be generalized to any unitary representation of a semi-direct

product (4.1). The only exceptions arise (i) if the spin representation R is one-
dimensional so that E = C and the tensor product (3.8) is trivial, or (ii) if f is
such that Wq [f ] does not depend on q. In both situations the splitting (3.8) is robust
against symmetry transformations. An example of momentum-independent Wigner
rotations will be provided by the Bargmann group below. Thus the entanglement of
spin and momentum due to Wigner rotations is a purely relativistic effect.
4.3 Poincaré particles in three dimensions

Here we apply the considerations of the previous section to the Poincaré group in
D = 3 space-time dimensions. This exercise will be a helpful guide for the description
of BMS3 particles in part III. To our knowledge, representations of Poincaré in three
dimensions have previously been studied in [107, 108].
4.3.1 Poincaré group in three dimensions

Prelude: the group SL(2, R)
Many properties of the Poincaré group in three dimensions rely on the group SL(2, R),
so we start by describing the latter. SL(2, R) is the group of linear transformations of
the plane R2 that preserve volume and orientation. It consists of real 2 × 2 matrices
with unit determinant:
a b
, ad − bc = 1. (4.79)
c d
The centre of SL(2, R) consists of the identity matrix and its opposite, thus spanning
a group Z2 . Furthermore:
Lemma. The group SL(2, R) is connected, but not simply connected. It is homotopic
to a circle and its fundamental group is isomorphic to the group of integers Z:
π1 (SL(2, R)) ∼
= Z. (4.80)
Proof. Since the determinant of (4.79) is non-zero, the vectors (a, b) and (c, d) in R2
are linearly independent. We can thus find linear combinations of these vectors that
span an orthonormal basis of R2 . In other words there exists a real matrix

ᾱ 0
K̄ =
β̄ γ̄
such that, for any SL(2, R) matrix S of the form (4.79), the product

ᾱa ᾱb
O ≡ K̄S =
β̄a + γ̄c β̄b + β̄d
belongs
√ to the orthogonal group O(2). We can make ᾱ positive by setting ᾱ−1 =
a2 + b2 and we can set γ̄ = 1/ᾱ so that O ∈ SO(2). Any matrix S ∈ SL(2, R) can
therefore be decomposed uniquely as

−1 x 0
S = K̄ O ≡ KO, with O ∈ SO(2) and K = (4.81)
y 1/x
for some y ∈ R and x ∈ R strictly positive.4 This shows that SL(2, R) is connected
and homotopic to its maximal compact subgroup consisting of rotations

cos θ − sin θ
, θ ∈ R. (4.82)
sin θ cos θ
In particular, the fundamental group of SL(2, R) is isomorphic to Z.
Lorentz transformations in three dimensions

The definitions of section 4.2.1 remain valid in three dimensions. In particular the
Lorentz group O(2, 1) still has four connected components as in fig. 4.2, and it is still
multiply connected. However, in contrast to the higher-dimensional case, the Lorentz
group is now homotopic to a circle and therefore has a fundamental group isomorphic
to Z. This is a consequence of the following result:
Proposition. There is an isomorphism
SO(2, 1)↑ ∼
= SL(2, R)/Z2 ≡ PSL(2, R) (4.83)
where the Z2 subgroup of SL(2, R) consists of the identity matrix and its opposite. In
particular, the fundamental group of the connected Lorentz group in three dimensions
is isomorphic to Z.
Proof. Our goal is to build a homomorphism
φ : SL(2, R) → O(2, 1) : f 7→ φ[f ] (4.84)
and then use the property
Im(φ) ∼
= SL(2, R)/Ker(φ). (4.85)
Let A be the Lie algebra sl(2, R). Each matrix α ∈ sl(2, R) can be written as a linear
combination
α = αµ tµ (4.86)
where the αµ ’s are real coefficients and the matrices

1 0 1 1 0 1 1 1 0
t0 ≡ , t1 ≡ , t2 ≡ (4.87)
2 −1 0 2 1 0 2 0 −1
form a basis of sl(2, R). With these conventions,

1 1
det(α) = − ηµν αµ αν = − α2 . (4.88)
4 4
Now SL(2, R) naturally acts on A according to the adjoint representation,
A → A : α 7→ f αf −1 . (4.89)
4
This is a rewriting of the Iwasawa decomposition.
This action preserves the determinant since det(f ) = 1, so according to (4.88) it may
be seen (for each f ) as a Lorentz transformation. This motivates the definition of a
homomorphism (4.84) given by
f tµ f −1 = tν φ[f ]ν µ ∀ µ = 0, 1, 2. (4.90)
The entries of φ[f ] are quadratic combinations of those of f , so φ is a continuous map.

Since SL(2, R) is connected, the image Im(φ) is contained in the connected Lorentz
group SO(2, 1)↑ . In fact one has Im(φ) = SO(2, 1)↑ , which follows from the standard
decomposition theorem for Lorentz transformations (see e.g. [95, 109]). The kernel of
φ coincides with the centre of SL(2, R), i.e. Ker(φ) = {I, −I}. The isomorphism (4.83)
follows upon using (4.85).
Remark. For future reference note that the homomorphism (4.90) explicitly reads
1 2
(a + b2 + c2 + d2 ) 21 (a2 − b2 + c2 − d2 ) −ab − cd


a b 2
φ =  21 (a2 + b2 − c2 − d2 ) 12 (a2 − b2 − c2 + d2 ) −ab + cd , (4.91)
c d
−ac − bd bd − ac ad + bc
where the argument of φ is an SL(2, R) matrix. This will be useful in section 9.1.
Poincaré group in three dimensions

The Poincaré group for D = 3 is defined as in (4.39) and its connected subgroup is
(4.40). Owing to the isomorphism (4.83), its double cover can be written as
SL(2, R) n R3 (4.92)
where the action of SL(2, R) on R3 is given by (4.89). The latter is in fact the adjoint
representation so we can also rewrite (4.92) as
double cover of ISO(2, 1)↑ = SL(2, R) nAd sl(2, R)Ab (4.93)
where sl(2, R)Ab is the Lie algebra of SL(2, R), seen as an Abelian vector group. This
observation will turn out to be crucial in part III of this thesis. We stress that (4.93)
is not the universal cover of the Poincaré group in three dimensions, since SL(2, R) is
homotopic to a circle. This implies that (4.93) admits topological projective represen-
tations (which are equivalent to exact representations of its universal cover). There
are no algebraic central extensions.
4.3.2 Particles in three dimensions

Here we describe projective irreducible unitary representations of the connected Poin-
caré group in three dimensions and point out a few differences with respect to the
higher-dimensional case described in section 4.2.
Orbits and little groups

The classification of Poincaré momentum orbits in three dimensions is the same as in
section 4.2.2 and is summarized in fig. 4.3. Considering the double cover (4.93) for
definiteness, the little groups are as follows:
Orbit Little group

Trivial SL(2, R)
Massive U(1) ∼= SO(2)
Massless R × Z2
Tachyonic R × Z2
Table 4.1: Orbits and little groups for Poincaré in three dimensions.
Let us prove that these are the correct little groups. We shall use the fact that the
action of Lorentz transformations on momenta is equivalent to its action on transla-
tions, which in turn is equivalent to the adjoint representation of SL(2, R) according to
the definition (4.93). From that point of view a momentum (p0 , p1 , p2 ) is represented
by a matrix η µν pµ tν where ηµν is the Minkowski metric in D = 3 dimensions and the
tµ ’s are given by (4.87). Explicitly the matrix is

1 p2 −p0 + p1
p= . (4.94)
2 p0 + p1 −p2
In that language the little group of p is the set of SL(2, R) matrices that commute with
(4.94). It immediately follows that the little group of p = 0 is SL(2, R). For massive
orbits we move to a rest frame where p0 = M and p1 = p2 = 0; the only matrices
leaving p fixed then are rotations (4.82). For tachyons we take p0 = p1 = 0, p2 6= 0
and find that the little group consists of matrices of the form
x
e 0
± , x ∈ R, (4.95)
0 e−x
spanning a group R × Z2 . Finally, for massless particles we take p0 = −p1 6= 0 and

p2 = 0; the resulting little group R × Z2 is spanned by matrices of the type

1 x
± , x ∈ R. (4.96)
0 1
This reproduces all little groups listed in table 4.1.
Note that the little groups listed in table 4.1 are sensitive to the cover chosen in
(4.93). Had we chosen the standard connected Poincaré group SO(2, 1)↑ n R3 , the
little groups for massless particles and tachyons would be quotiented by Z2 and would
reduce to R. For the universal cover of the Poincaré group, the little groups would
instead get decompactified, so e.g. U(1) would be replaced by R. This has important
implications for the spin of relativistic particles in three dimensions.
Massive particles
The properties of massive particles in three dimensions are the same as in section 4.2.3.
In particular their momentum orbits take the form
Op ∼
= SO(2, 1)↑ /U(1) ∼
= SL(2, R)/S 1 . (4.97)
The only subtlety is that the group of spatial rotations now is U(1) ∼ = SO(2), so the
spin of a massive particle is a one-dimensional irreducible unitary representation of the
form (2.13) labelled by some number s. If the double cover (4.93) was the universal
cover of the Poincaré group, that number would be restricted to integer or half-integer
values. However the fact that (4.93) is homotopic to a circle implies that s may take
any real value. Thus massive particles in three dimensions can be anyons [110, 111].
The same phenomenon will occur with massive BMS3 particles.
Remark. Wigner rotations do occur in three dimensions, but they do not lead
to momentum/spin entanglement when the space of spin degrees of freedom is one-
dimensional.
Massless particles
The spin properties of massless particles in three dimensions are also somewhat pecu-
liar compared to those of their higher-dimensional cousins. Their little group R × Z2
can be seen as a Euclidean group in one dimension, where Z2 plays the role of rotations
while R is spanned by Euclidean translations. If the latter is represented non-trivially,
one obtains an analogue of “continuous spin” particles in three dimensions, although
in the present case the space of spin degrees of freedom is actually finite-dimensional.
By contrast, when R is represented trivially, the spin representation boils down to an
irreducible unitary representation of Z2 . The latter has exactly two irreducible unitary
representations (the trivial one and the fundamental one), so we conclude that “dis-
crete spin” massless particles in three dimensions can only be distinguished by their
statistics (bosonic or fermionic); they have no genuine spin. This is consistent with
the fact that massless field theories in three dimensions either have no local degrees
of freedom at all (such as in gravity or Chern-Simons theory), or have only scalar or
Weyl fermion degrees of freedom.
4.3.3 Characters
For future reference, we now list characters of irreducible unitary representations of
the Poincaré group in three dimensions. The results of sections 4.2.4 and 4.2.5 apply,
so the character of a rotation by θ combined with an arbitrary translation α in a
Poincaré representation with mass M and spin s is given by formula (4.60),
0 +isθ 1 0 1
χ[(rotθ , α)] = eiM α iθ 2
= eiM α +isθ 2 , (4.98)
|1 − e | 4 sin (θ/2)
where we have replaced the little group character by χλ [f ] = eisθ . In part III we shall
encounter the BMS3 generalization of this expression. Similarly the character (4.68)
of Euclidean time translations becomes

V
Tr e−βH (1 + βM )e−βM .

massive particle
= 2
(4.99)
2πβ
Characters of massless particles with discrete spin are given by formula (4.71) with
D = 3, r = 1 and χλ = ±1.
*4.4 Galilean particles

In this section we classify irreducible unitary representations of the Bargmann groups,
i.e. non-relativistic or Galilean particles. This example will be useful as a comparison
to the relativistic case, and will also involve a dimensionful central charge that makes it
similar to the centrally extended BMS3 group of part III. This being said, the material
exposed in this section is not crucial for our later considerations, so it may be skipped
in a first reading. The plan is similar to that of section 4.2: after defining Bargmann
groups, we classify their orbits and litte groups, describe non-relativistic particles and
compute their characters. We refer to [112, 113] for further reading on the Bargmann
groups and to [114] for their representations.
4.4.1 Bargmann groups

Galilei groups
Definition. The Galilei group in D space-time dimensions is a nested semi-direct
product
O(D − 1) n RD−1 n RD−1 × R

(4.100)
whose elements are quadruples (f, v, α, t) where f ∈ O(D − 1) is a rotation, v is a
boost belonging to the first RD−1 , α is a spatial translation belonging to the second
RD−1 , and t ∈ R is a time translation. The group operation is

(f, v, α, s) · (g, w, β, t) = f · g, v + f · w, α + f · β + vt, s + t (4.101)
where the dots on the right-hand side denote either matrix multiplication, or the ac-
tion of a matrix on a column vector. The largest connected subgroup of (4.100) is
obtained upon replacing O(D − 1) by SO(D − 1); its universal cover is obtained by
replacing SO(D − 1) by its universal cover, Spin(D − 1).
The intricate structure (4.100) translates the fact that space and time live on
different footings in Galilean relativity. Thus the analogue of a Lorentz transformation
now is a pair (f, v), while space-time translations are pairs (α, t). Boosts and rotations
span a group SO(D − 1) n RD−1 while space-time translations span an Abelian group
RD . In particular each boost is a velocity vector v acted upon by rotations according
to the matrix representation of O(D − 1). Since time is absolute in Galilean relativity,
the last entry on the right-hand side of (4.101) is a sum s + t without influence of
boosts. The term vt of the third entry is a time-dependent translation at velocity v.

Finally, there is an Abelian subgroup R2D consisting of pairs
(e, v, α, 0) (4.102)
where e is the identity in O(D − 1).
The Lie algebra of the Galilei group is generated by (D − 1)(D − 2)/2 rotation
generators, (D − 1) boost generators, (D − 1) spatial translation generators, and one
generator of time translations. We will not display their Lie brackets here.
Bargmann groups
The Galilei group turns out to admit a non-trivial algebraic central extension:
Definition. The Bargmann group in D space-time dimensions is a centrally extended

semi-direct product
Bargmann(D) ≡ O(D − 1) n RD−1 n RD−1 × R × R ,

(4.103)
whose elements are 5-tuples (f, v, α, t, λ) where (f, v, α, t) belongs to the Galilei group
(4.100) while λ is a real number. The group operation is
1 2
(f, v, α, s, λ)·(g, w, β, t, µ) = (f, v, α, s)·(g, w, β, t), λ+µ+v·f ·β + v t (4.104)
2
where the first entry on the right-hand side is given by (4.101) while v · β ≡ v i β i is
the Euclidean scalar product of v and β; in particular, v2 ≡ v i v i .
This central extension says that the Abelian subgroup of boosts and translations
(4.102) gets extended into a Heisenberg group (2.37):

(e, v, α, 0, λ) · (e, w, β, 0, µ) = e, v + w, α + β, 0, λ + µ + v · β .
In other words, in quantum mechanics, spatial translations and boosts do not com-
mute. Note that, even in the Bargmann group, the normal subgroup of (centrally
extended) space-time translations
(e, 0, α, t, λ) (4.105)
remains Abelian. Hence the exhaustivity theorem of section 4.1.5 applies to the
Bargmann group: all Galilean particles are induced representations.
For D ≥ 4 space-time dimensions, eq. (4.104) is the only algebraic central exten-
sion of the Galilei group. But for D = 3, the Galilei group admits three non-trivial
differentiable central extensions [113], one of which is the one displayed in (4.104).
We will not take these extra central extensions into account. As regards topological
central extensions, the Galilean situation is identical to that of the Poincaré group.
Thus Bargmann(3) has a fundamental group Z and admits infinitely many topological
projective representations, while for D ≥ 4 the fundamental group of Bargmann(D)
is Z2 , leading either to exact representations or to representations up to a sign.
Remark. The Bargmann group is a limit of the Poincaré group as the speed of light
goes to infinity (see e.g. [53]), known more accurately as an Inönü-Wigner contraction
[115]. We will not describe this procedure here, although we will encounter a very
similar one in part III when showing that the BMS3 group is an ultrarelativistic limit
of two Virasoro groups.

We now classify the orbits and little groups of the Bargmann group (4.103). We follow
the same strategy as in section 4.2.2.
Generalized momenta
The Abelian normal subgroup of (4.103) consists of centrally extended translations
(4.105). Its dual space consists of generalized momenta
(p, E, M ) (4.106)
paired with translations according to5
h(p, E, M ), (α, t, λ)i = hp, αi − Et − M λ (4.107)
where hp, αi ≡ pi αi , i = 1, ..., D − 1. Accordingly, p is dual to spatial translations
and represents the actual momentum of a particle; E is dual to time translations and
represents the particle’s energy; finally M is a central charge dual to the central entries
λ in (4.105). Working in units such that ~ = 1, eq. (4.104) says that λ has dimensions
[distance] × [velocity] so the pairing (4.107) implies that M is a mass scale:
[energy] × [time]
[M ] = = [mass]. (4.108)
[distance] × [velocity]
In fact we will see below that M is the mass of a non-relativistic particle. Note that
M lives on a different footing than p and E, which is why the relation E = M c2 is
invisible in Galilean relativity.
According to the structure (4.103), the group G acting on translations is the Eu-
clidean group spanned by rotations and boosts. Its action is given by
1 2
σ(f,v) (α, t, λ) = f · α + vt, t, λ + v · f · α + v t (4.109)
2
by virtue of (4.104). The pairing (4.107) then yields the action σ ∗ of boosts and
rotations on generalized momenta:
∗
σ(f,v) (p, E, M ), (α, t, λ) =
(4.16)
= (p, E, M ), σ(f −1 ,−f −1 ·v) (α, t, λ)
D 1 E
= (p, E, M ), f −1 · α − f −1 · vt, t, λ − v · α + v2 t
2
(4.107)
1
= p, f −1 · α − f −1 · vt − Et − M λ − v · α + v2 t , (4.110)
2
5
The minus signs are conventional, and included for later convenience.
where we have used the fact that rotations preserve Euclidean scalar products. We can
then use the Euclidean analogue of the isomorphism (4.42) to identify (RD−1 )∗ with
RD−1 and rewrite the pairing hp, αi = pi αi as a scalar product p · α = pi αi = pi αi ,
where indices are raised and lowered thanks to the Euclidean metric. This allows us
to rewrite v · α as hv, αi in (4.110), and leads to
∗
1 2

σ(f,v) (p, E, M ) = f · p + M v, E + v · f · p + M v , M . (4.111)
2
One may recognize here the non-relativistic transformations laws of momentum and
energy under rotations and boosts. The mass M is left unchanged, as was to be
expected for a central charge.
Orbits
Let us classify orbits of generalized momenta under the transformations (4.111). Since
the mass M is invariant, it is a constant quantity specifying each orbit; orbits with dif-
ferent masses are disjoint. In particular, the orbits differ greatly depending on whether
M vanishes or not.
A massless non-relativistic particle is one for which M = 0, whereupon (4.111)

simplifies to
∗
σ(f,v) (p, E, 0) = f · p, E + v · f · p, 0 . (4.112)
This implies that the norm of the momentum p of a massless particle is invariant
under rotations and boosts. If p = 0 the particle is static (in all reference frames) and
the momentum orbit is trivial. If on the other hand p 6= 0, then the particle moves
(in all references frames); its momentum orbit is
O(p,E,0) = (f · p, E + v · f · p, 0) f ∈ SO(D − 1), v ∈ RD−1 ∼ = S D−2 × R

where the sphere S D−2 is spanned by all momenta f · p while R is spanned by the
values of energy. The little group is
G(p,E,0) = SO(D − 2) n RD−2 (4.113)
and consists of rotations leaving p invariant together with boosts that are orthogonal
to p. Note that this is the same little group (4.49) as for relativistic massless particles.
A massive non-relativistic particle is such that M 6= 0. Let (p, E) be its momentum

and energy. Then the boost v = −p/M plugged in (4.111) maps (p, E, M ) on
∗
p2
σ(e,−p/M ) (p, E, M ) = 0, E + , M (4.114)
2M
so any massive particle admits a rest frame. If we call E0 ≡ E + p2 /2M , the orbit of
(4.114) under rotations and boosts is a parabola
M v2

D−1
O(0,E,M ) = M v, E0 + ,M v ∈ R ⊂ RD−1 × R. (4.115)
2
As orbit representative we can take the generalized momentum in the rest frame,
(0, E0 , M ) (4.116)
where E0 is an arbitrary real number; at fixed M , representatives with different values

of E0 define distinct orbits. The little group is the group of rotations
G(0,E0 ,M ) = SO(D − 1) (4.117)
in accordance with the fact that the orbit (4.115) is diffeomorphic to the quotient
space SO(D − 1) n R D−1
/SO(D − 1) = R∼ D−1
. Note again that this is exactly the
same little group (4.46) as for relativistic massive particles. Finally, pure boosts
gq = (e, q/M ) (4.118)
provide a continuous family of standard boosts on the orbit (4.115) of (4.116). Note
that energy is bounded from below on the orbit if and only if M > 0.
4.4.3 Particles
According to the exhaustivity theorem of section 4.1.5, all irreducible unitary repre-
sentations of Bargmann groups are induced, and they are classified by momentum
orbits. Each such representation consists of wavefunctions on an orbit, representing
the quantum states of a non-relativistic particle.
For example, the spin of a massive Galilean particle is an irreducible unitary repre-
sentation of SO(D − 1). The space of states of the particle then consists of wavefunc-
tions on the orbit (4.115) taking values in the space of the spin representation. Scalar
products of wavefunctions are defined as usual by (3.7), where µ is some measure on
the orbit. For convenience one can pick the standard Lebesgue measure dD−1 q, which
is left invariant by both rotations and boosts since (4.111) says that they act on (4.115)
as Euclidean transformations q 7→ f · q + M v.
In order to write down formula (4.30) explicitly for a non-relativistic particle, we

still need to understand the Wigner rotation (4.31). Let us evaluate it for a pair (f, v)
at a point q belonging to the momentum orbit. We have (f, v) · q = f · q + M v, so the
−1
standard boost (4.118) for the momentum (f, v)−1 · q is g(f,v)−1 ·q = e, f M·q − f −1 · v .

Using the group operation (4.101) we read off the Wigner rotation
gq−1 · (f, v) · g(f,v)−1 ·q = (f, 0) . (4.119)
Surprise: the Wigner rotation is blind to boosts! In fact it is momentum-independent

and simply coincides with the rotation f . Thus formula (4.30) for the transformation
law of non-relativistic one-particle states becomes
2
T [(f, v, α, t, λ)] · Ψ (q) = e−iM λ eiq·α−iq t/2M R[f ] · Ψ (f, v)−1 · q ,

(4.120)
where we have also used the fact that the measure dD−1 q is invariant to cancel its
Radon-Nikodym derivative. This result differs from the Poincaré transformations of
relativistic particles in two key respects. First, Galilean Wigner rotations (4.119) are
momentum-independent, so in contrast to (4.75) they do not entangle momentum and
spin. In fact, there is no Thomas precession for non-relativistic particles. The second
difference is the presence of the mass M : formula (4.120) is an exact representation of
the Bargmann group (4.104), but because M 6= 0 it is a projective representation of
the centreless Galilei group (4.100). This can be seen by noting that for a pure boost
v and a spatial translation α, eq. (4.120) gives
T [v] · T [α] = e−iM v·α T [α] · T [v], (4.121)
which says that boosts and spatial translations do not commute. In part III we will
encounter a similar phenomenon with the BMS3 group, whose dimensionful central
charge will coincide with the Planck mass.
4.4.4 Characters
We now evaluate characters of massive non-relativistic particles. Let M > 0 and
choose a spin λ, specifying an irreducible unitary representation of the little group
SO(D − 1). For definiteness we take the rest frame energy E0 = 0 in (4.115). In order
for the character (4.33) to be non-zero we must set v = 0. Eq. (4.119) then allows us
(D−1)
to pull the little group character χR = χλ out of the momentum integral:
Z
(D−1) 2
χ[(f, 0, α, t, λ)] = e−iM λ χλ [f ] dD−1 k δ (D−1) (k − f · k) eik·α−ik t/2M . (4.122)
RD−1
For simplicity we set λ = 0 from now on and neglect writing this entry. We take f to
be a rotation (4.54) with the first row and column suppressed and all angles θ1 , ..., θr
non-zero, r = b(D − 1)/2c. If D is odd we also erase the last row and column. We
treat separately even and odd dimensions.
If D is odd, then the only fixed point of f on (4.115) is the tip k = 0. The integral
of (4.122) localizes and (4.58) yields
r
Y 1
χ[(f, 0, α, t)] = χλ [f ] . (4.123)
j=1
|1 − eiθj |2
Note that translations do not contribute to this result. Up to the normalization of

energy, it coincides with the relativistic character (4.60).
If D is even, then f leaves fixed the whole axis kD−1 as in fig. 3.2. Integrating first
over the rotated coordinates k1 , ..., kD−2 in (4.122) and writing kD−1 ≡ k, we find
r Z +∞
(D−1)
Y 1 D−1 2
χ[(f, 0, α, t)] = χλ [f ] iθ 2
dk δ(0) eikα −ik t/2M . (4.124)
j=1
|1 − e | −∞
j
Here the term δ(0) = δ(k − k) is an infrared divergence that we regularize as in (4.64)
with a length scale L. Denoting αD−1 ≡ x, we are left with the integral
Z +∞ 1/2
ikx−ik2 t/2M 2πM 2
dk e = eiM x /2t (4.125)
−∞ it
and thus conclude

r 1/2
L (D−1) Y 1 2πM 2
χ[(f, 0, α, t)] = χλ [f ] iθ 2
eiM x /2t . (4.126)
2π j=1
|1 − e |j it
The only dependence of this expression on α appears through the component αD−1 ≡
x, because we picked a rotation f leaving fixed the direction kD−1 . For a general ro-
tation, the component of α appearing in the character would be its projection on the
axis left fixed by f .
Two comments are in order. First note that in D = 2 space-time dimensions,

(4.126) boils down to the quantum propagator of a free non-relativistic particle at
time t and separation x, up to an infrared-divergent factor L. This is because the
character of a pure spatial translation in two space-time dimensions is
Z +∞ Z +∞
dy +∞
Z
ikx−itk2 /2M 2
dkeikx−itk /2M

Tr T [(e, x, t)] = dk δ(0) e =
−∞ −∞ 2π −∞
which we can interpret as the trace of the operator eiP x−iHt in the Hilbert space of a
free massive particle on the real line:
Z +∞
iP x−iHt
dy y + x|e−iHt |y .

Tr T [(e, x, t)] = Tr e = (4.127)
−∞
The integrand of this expression is the propagator of a free non-relativistic particle

evaluated between y and y + x at time t and coincides with (4.125).
The second comment concerns the relation between Bargmann characters and
Poincaré characters. For even D, (4.126) is the non-relativistic analogue of (4.65)
but the functions appearing in the two results are different. By contrast, for odd
D, the Bargmann character (4.123) coincides with its Poincaré analogue (4.60). This
may be seen as a consequence of the phenomenon (4.58), whose effect is to localize
the computation of the character to the region of momentum space surrounding the
momentum at rest, that is, the non-relativistic region. By contrast, when the localiza-
tion is not complete as is the case for even D, the momenta in the integral (4.125) are
arbitrarily large and relativistic effects become important. This produces a difference
between Bargmann and Poincaré characters. It is particularly apparent for characters
of Euclidean time translations, which in the non-relativistic case are given by
Z (D−1)/2
NV D−1 −βk2 /2M M
χ[(e, 0, 0, −iβ)] = d ke = NV
(2π)D−1 RD−1 2πβ
where N is the dimension of the spin representation. This is the non-relativistic version
of (4.68). For D = 3 (and N = 1) it reduces to V M/(2πβ), which is the non-relativistic
limit of (4.99).
Chapter 5
Coadjoint orbits and geometric

quantization
In the previous chapters we have seen how representation theory leads to geometric
objects such as orbits. The purpose of this chapter is to describe the opposite phe-
nomenon: starting from a coadjoint orbit of a group G, we will obtain a representation
by quantizing the orbit. This construction will further explain why orbits of momenta
classify representations of semi-direct products. In addition it will turn out to be a
tool for understanding gravity in parts II and III.
The plan is as follows. We start in section 5.1 with basic reminders on symplectic
manifolds with symmetries, including their momentum maps. Along the way we intro-
duce the notion of coadjoint orbits, which will turn out to be crucial for the remainder
of this thesis. Section 5.2 is then devoted to the quantization of symplectic manifolds,
and describes in particular the relation between representation theory and symplectic
geometry. In section 5.3 we reformulate geometric quantization in terms of action
principles that describe the propagation of a point particle on a group manifold. The
two last sections of the chapter are concerned with applications of these considerations
to semi-direct products: in section 5.4 we describe the coadjoint orbits and world line
actions of such groups in general, while in section 5.5 we illustrate these results with
the Poincaré group and the Bargmann group.
Our language in this chapter will be slightly different than in the previous ones,
as we rely on differential-geometric tools that were unnecessary for our earlier con-
siderations. Useful references include [43, 55] for symplectic geometry, [116, 117] for
quantization, as well as the (sadly unpublished) Modave lecture notes [118].
Remark. The presentation adopted here is self-contained, but fairly dense. We urge
the reader who is not acquainted with differential geometry to only read sections 5.1.1
and 5.1.2, then go directly to part II of the thesis. In doing so one will miss the
symplectic aspects of our later considerations, but the other points of our presentation
should remain accessible.
97
Chapter 5. Coadjoint orbits and quantization 98
5.1 Symmetric phase spaces

In this section we study classical systems with symmetries, that is, homogeneous sym-
plectic manifolds. We start by recalling a few basic facts about Lie groups and we
define their adjoint and coadjoint representations. We then describe in general terms
Poisson and symplectic structures, and show how such structures arise in the case of
coadjoint orbits. Finally we discuss the notion of momentum maps associated with the
symmetries of a symplectic manifold. We use the notational conventions of chapter 3.
5.1.1 Lie groups

A Lie group is a group G which also has a structure of smooth manifold such that
multiplication and inversion are smooth maps. In particular the operations of left and
right multiplication defined in (3.16) and (3.17) are diffeomorphisms. We denote by e
the identity in G, and generic group elements are denoted f , g, etc.
Definition. A vector field ξ on G is left-invariant if (Lf )∗ ξ = ξ for all f ∈ G, i.e. if

(Lf )∗g ξg = ξf g for all f, g ∈ G.1
One can verify that any left-invariant vector field is given by ξg = (Lg )∗e X for some
tangent vector X ∈ Te G. Thus the space of left-invariant vector fields is isomorphic to
the tangent space of G at the identity. We shall denote by ζX the left-invariant vector
field on G given by (ζX )g = (Lg )∗e X.
Definition. The Lie algebra of G is the vector space g = Te G endowed with the Lie
bracket
[X, Y ] ≡ [ζX , ζY ]e (5.1)
where the bracket on the right-hand side is the usual Lie bracket of vector fields eval-
uated at the identity.
One can show that the bracket (5.1) is such that ζ[X,Y ] = [ζX , ζY ]. As a corollary,
any smooth homomorphism of Lie groups F : G → H is such that its differential
F∗e at the identity is a homomorphism of Lie algebras. When interpreting G as a
symmetry group, the elements of its Lie algebra are seen as “infinitesimal” symmetries,
i.e. transformations near the identity. In practice the Lie algebra structure of g is often
displayed in terms of a basis {ta |a = 1, ..., dim g} of g with Lie brackets
[ta , tb ] = fab c tc . (5.2)
In that context the coefficients fab c ∈ R are known as the structure constants of g in
the basis {ta }.
1
Recall that the differential of a smooth map F : M → N at p ∈ M is the map F∗p : Tp M →
d

TF (p) N : γ̇(0) 7→ dt F(γ(t)) t=0 , where γ(t) is a path in M such that γ(0) = p.
Exponential map
Definition. Let X ∈ g, and let γX be the integral curve2 of the corresponding left-
invariant vector field ζX such that γX (0) = e. Then the exponential map of G is
exp : g → G : X 7→ exp[X] ≡ γX (1). (5.3)
One can
P verifynthat, for matrix groups, this definition reduces to the standard Taylor
X
series n∈N X /n!. We often denote exp[X] ≡ e .
Since the exponential map is defined by a vector flow, it automatically satisfies

exp[(s + t)X] = exp[sX] exp[tX] for all s, t ∈ R. In particular any X ∈ g determines
a one-dimensional subgroup of G consisting of elements exp[tX], t ∈ R. Note that
left-invariant vector fields are complete, which ensures the existence of exp[tX] for all
t ∈ R. Finally, for any smooth homomorphism F : G → H, one can show that
F ◦ expG = expH ◦F∗e . (5.4)
5.1.2 Adjoint and coadjoint representations

Definition. Let G be a Lie group with Lie algebra g. Then the adjoint representation
of G is the homomorphism
Ad : G → GL(g) : g 7→ Adg (5.5)
where Adg is the linear operator that acts on g according to
d
g etX g −1

Adg (X) = t=0
. (5.6)
dt
Here one may freely replace etX by any path γ(t) in G such that γ(0) = e and γ̇(0) = X.
For matrix groups, eq. (5.6) reduces to Adg (X) = gXg −1 .
One can verify that this is indeed a representation of G. Using (5.4), one also
shows that it satisfies the identity
eAdf X = f eX f −1 (5.7)
where eX is the exponential map of G. Note that the adjoint representation of any
Abelian Lie group is trivial. Finally, the adjoint representation of the Lie algebra g is
defined as the differential of (5.5) at the identity:
d
adX (Y ) ≡ AdetX (Y ) t=0 = [X, Y ] . (5.8)
dt
In (4.16) we saw how to define dual representations. Let us apply this to the
adjoint representation (5.5): we write the dual space of g as g∗ , which consists of
linear forms p : g → R : X 7→ hp, Xi. When interpreting G as a symmetry group,
the elements of the dual of g can be seen as “momenta”, or more generally conserved
vectors, associated with the symmetries. In particular the number hp, Xi then is
the Noether charge associated with the symmetry generator X when the system has
“momentum” p.
2
An integral curve of a vector field ξ on a manifold M is a path γ(t) on M such that γ̇(t) = ξγ(t) .
Definition. Let G be a Lie group with Lie algebra g. Then the coadjoint represen-
tation of G is the homomorphism
Ad∗ : G → GL(g∗ ) : f 7→ Ad∗f (5.9)
which is dual to the adjoint representation in the sense that
Ad∗f (p) ≡ p ◦ (Adf )−1 , (5.10)
i.e. Ad∗f (p), X ≡ hp, Adf −1 (X)i for all p ∈ g∗ and any X ∈ g. From now on we refer
to elements of g and g∗ as adjoint and coadjoint vectors, respectively.
The coadjoint representation is a linear action of G on g∗ . In particular one can

foliate the space g∗ into disjoint G-orbits. We call the set
Wp ≡ Ad∗g (p) g ∈ G

the coadjoint orbit of p. It is a homogeneous space for the coadjoint action of G. Note
that the coadjoint representation of any Abelian group is trivial, so its coadjoint orbits
are single points. By contrast, coadjoint orbits of non-Abelian groups are generally
non-trivial (except if p = 0). We will see in section 5.4.3 that the coadjoint orbits of
semi-direct products contain their momentum orbits.
The dual of the infinitesimal adjoint representation (5.8) is the differential of (5.9)
at the identity, i.e. the coadjoint representation of the Lie algebra g:
d (5.10)
ad∗X (p) ≡ (Ad∗etX (p))|t=0 = −p ◦ adX = −p ◦ [X, ·] . (5.11)
dt
Remark. The adjoint and coadjoint representations of a group G are generally in-
equivalent. In fact they are equivalent if and only if g admits a non-degenerate bilinear
form (which is the case e.g. for semi-simple Lie groups).
5.1.3 Poisson structures

The phase space of a system is the set of its classical states. In the previous pages
we have reviewed some basic concepts of group theory, and our goal is to eventually
apply them to phase spaces with symmetries. Accordingly we now investigate Poisson
structures and symplectic structures in more detail.
Definition. Let M be a manifold. A Poisson structure on M is an antisymmetric

bilinear map3
{·, ·} : C ∞ (M) × C ∞ (M) → C ∞ (M) : F, G 7→ {F, G}
which satisfies the Jacobi identity and the Leibniz rule:

{F, {G, H}} + {G, {H, F}} + {H, {F, G}} = 0 (Jacobi),
{F, GH} = {F, G}H + G{F, H} (Leibniz).
3
From now on, real functions on M are denoted as F, G, H, etc.

This map is called the Poisson bracket on M, and the pair M, {·, ·} is a Poisson
manifold or a phase space.
The Poisson bracket endows the space of functions C ∞ (M) with a structure of Lie
algebra; the Leibniz identity implies in addition that the map
{F, ·} : C ∞ (M) → C ∞ (M) : G 7→ {F, G}
is a derivation for any function F ∈ C ∞ (M).4 These properties together endow

the space C ∞ (M) with the structure of a Poisson algebra. Note that the existence
of a Poisson structure sets no restrictions on the dimension of M. In particular,
odd-dimensional manifolds admit Poisson structures, e.g. M = R3 with the bracket
{F, G} = ∂x F∂y G − ∂y F∂x G. This will change once we turn to symplectic structures.
Definition. Let M, {·, ·} be a Poisson manifold; let H ∈ C ∞ (M). We call Hamil-

tonian vector field associated with H the (unique) vector field ξH on M such that
ξH = −{H, ·}. (5.12)
The existence of ξH is ensured by the one-to-one correspondence between derivations

of C ∞ (M) and vector fields on M.
The Hamiltonian vector field associated with a function H is a differential operator

acting on functions on M. Its integral curves are the paths γ(t) in M that satisfy
γ̇(t) = (ξH )γ(t) , which in local coordinates on M corresponds to a set of dim(M)
first-order differential equations ẋi (t) = ξH (x(t)). These are the equations of motion
associated with the Hamiltonian H. The definition (5.12) ensures that {H, G} =
−ξH (G), which implies that the equations of motion can be written locally as ẋi =
{xi , H} in terms of the Poisson bracket. In particular one has {H, G} = 0 if and only
if G is constant along integral curves of ξH . Note also that
[ξF , ξG ] = −ξ{F ,G} , (5.13)
so the Lie brackets of Hamiltonian vector fields are Hamiltonian.
Now consider the set of all Hamiltonian vector fields on M; at a point p ∈ M, they
span a subspace of the tangent space Tp M. By taking this span for all p ∈ M, one
obtains a subbundle of the tangent bundle T M (i.e. a distribution on M). Because
brackets of Hamiltonian vector fields are Hamiltonian, Frobenius’ theorem implies
that Hamiltonian vector fields yield a foliation of M into so-called symplectic leaves.
Two points belong to the same leaf if they can be joined by the integral curve of a
Hamiltonian vector field. In the example of R3 mentioned above, symplectic leaves are
planes z = const. This leads to the definition of symplectic manifolds.
4
A derivation of an algebra A is a linear map D : A → A : a 7→ D(a) that satisfies the Leibniz
rule D(a · b) = D(a) · b + a · D(b).
5.1.4 Symplectic structures

Definition. Let M be a manifold. A symplectic form on M is a closed, non-
degenerate two-form ω on M.5 The pair (M, ω) is a symplectic manifold.
Non-degeneracy means that, in local coordinates, the components ωij of ω form

an invertible antisymmetric matrix. This implies that all symplectic manifolds are
even-dimensional. Note that any symplectic manifold admits a Liouville volume form
µ≡ ω
| ∧ {z
... ∧ ω} . (5.14)
dim(M)/2 times
The symplectic leaves described above are prime examples of symplectic mani-
folds: they are endowed with a symplectic form ω such that ω(ξF , ξG ) ≡ {F, G}; this
condition determines ω unambiguously because symplectic leaves are, by definition,
spanned by the integral curves of Hamiltonian vector fields. Another common exam-
ple is the phase space M = R2n of a non-relativistic particle in Rn , with coordinates
(q 1 , ..., q n , p1 , ..., pn ) and symplectic form
ω = dq i ∧ dpi (implicit sum over i = 1, ..., n). (5.15)
Canonical symplectic form

The symplectic structure (5.15) has an important generalization: consider a manifold
Q describing the configuration space of a classical system (so dim Q is the number
of Lagrange variables). The corresponding phase space is the cotangent bundle T ∗ Q,
which consists of pairs (q, α) where q ∈ Q and α ∈ Tq∗ Q. These pairs are generally
interpreted as describing a “position” q and a “momentum” α, but we will see below
that the interpretation stemming from semi-direct products is different: q will in fact
be a momentum (with Q a momentum orbit), while α will essentially be a position (or
rather a translation vector). The symplectic form on T ∗ Q is defined as follows. We
let
π : T ∗ Q → Q : (q, α) 7→ q (5.16)
be the natural projection and define the Liouville one-form θ on T ∗ Q by
θ(q,α) , V ≡ α, π∗(q,α) V (5.17)
for any vector V ∈ T(q,α) T ∗ Q. Then ω ≡ −dθ is the canonical symplectic form on
T ∗ Q. In the example (5.15), Q = Rn .
Let us verify that ω = −dθ is indeed symplectic. We choose local coordinates

(q , ..., q n ) on some open set U ⊂ Q and denote by (q i , pj ) the corresponding local
1
coordinates on π −1 (U ), so that the form α ∈ Tq∗ Q reads α = pj (dq j )q . Given a vector

V ∈ T(q,α) T ∗ Q, one can write
∂ ∂ ∂
V = ai i
+ bj ⇒ π∗(q,α) V = ai .
∂q ∂pj ∂q i
5
Closedness means dω = 0, where d is the exterior derivative. Non-degeneracy means that for all
p ∈ M, any vector v ∈ Tp M such that ωp (v, w) = 0 for all w ∈ Tp M necessarily vanishes.
Thus the differential of the projection (5.16) projects V on its part tangent to Q. The
definition (5.17) then implies that θ = pi dq i , so
ω = −dθ = dq i ∧ dpi (5.18)
is definitely a closed, non-degenerate two-form. It coincides locally with (5.15).
Remark. The Darboux theorem states that any point of a symplectic manifold has
a neighbourhood with local coordinates (q i , pj ) such that the symplectic form reads
(5.18). Thus any symplectic manifold is locally equivalent to a cotangent bundle.
Hamiltonian vector fields revisited

Any symplectic manifold can be endowed with a Poisson structure by mimicking the
symplectic leaves described at the end of section 5.1.3. This relies on a new definition
of Hamiltonian vector fields:
Definition. Let (M, ω) be a symplectic manifold, F ∈ C ∞ (M). The Hamiltonian

vector field ξF associated with F is defined by
!
iξF ω = ω(ξF , ·) = dF. (5.19)
Conversely, a vector field ζ is Hamiltonian if there exists a function F such that ζ = ξF .
Hamiltonian vector fields can be used to define Poisson brackets in the same way
as on symplectic leaves of Poisson manifolds: for any two functions F, G we write
{F, G} ≡ ω(ξF , ξG ). (5.20)
In terms of this bracket the definition (5.19) is equivalent to our earlier definition of
Hamiltonian vector fields in (5.12). In local coordinates the definition (5.19) reads
ωij ξFi = ∂j F ⇔ ξFi = ∂j Fω ji (5.21)
where ωij are the components of ω and ω ij is the matrix inverse of ωij . Accordingly,
the bracket (5.20) can be written as {F, G} = −ω ij ∂i F∂j G.
Note that symplectic manifolds only contain kinematical data: they tell us the
available combinations of “positions” and “momenta” — those combinations are clas-
sical states. Classical observables then are real-valued functions on phase space. Once
we declare that a certain observable H is the Hamiltonian, time evolution is given
locally by the equations of motion ẋ = {x, H}.
Symplectomorphisms
Definition. Let (M, ω) and (N , Ω) be symplectic manifolds. A symplectomorphism
(or canonical transformation) from M to N is a diffeomorphism φ : M → N that
preserves the symplectic structure in the sense that6
φ∗ Ω = ω. (5.22)
6
Recall that the pullback of a tensor field T of rank k on a manifold N by a map φ : M → N is
defined by (φ∗ T )p (v1 , ..., vk ) ≡ Tφ(p) (φ∗p v1 , ..., φ∗p vk ) for any p ∈ M and all v1 , ..., vk ∈ Tp M.
Then (M, ω) and (N , Ω) are said to be symplectomorphic.
When φ : M → N is a symplectomorphism, it preserves Poisson brackets in the

sense that {φ∗ F, φ∗ G} = φ∗ {F, G} for all functions F, G on N , where the brackets
on the left and on the right-hand side are those of M and N , respectively. Note
that the flow of any Hamiltonian vector field on M defines a one-parameter family of
symplectomorphisms of M.
5.1.5 Kirillov-Kostant structures

We now describe phase spaces whose structure is entirely determined by group theory.
They are prototypes for all phase spaces with symmetries.
Kirillov-Kostant Poisson bracket

Definition. Let G be a Lie group with Lie algebra g. Then the Kirillov-Kostant
Poisson bracket on g∗ is defined as
{F, G} (p) ≡ hp, [F∗p , G∗p ]i (5.23)
where F, G ∈ C ∞ (g∗ , R) and F∗p denotes the differential of F at p ∈ g∗ .7
One can associate a Hamiltonian vector field (5.12) with any function F on a
Poisson manifold. In the present case one has the following result:
Proposition. Let F be a real function on g∗ , ξF the corresponding Hamiltonian

vector field. The associated evolution equation is the Euler-Poisson equation on g∗ ,
γ̇(t) = (ξF )γ(t) = ad∗F∗γ(t) γ(t) .

(5.24)
Proof. Let G be a function on g∗ and take p ∈ g∗ . We are going to compute ξF (G) at

p in two different ways. First, (ξF )p is a vector tangent to g∗ at p and may therefore
be seen as an element of g∗ (since g∗ is a vector space). But (ξF )p (G) only depends on
the differential of G at p, so we may write
(ξF )p (G) = h(ξF )p , G∗p i. (5.25)
Secondly, using (5.12) and the bracket (5.23), one has

(5.11)
(ξF )p (G) = −{F, G} = − hp, [F∗p , G∗p ]i = − p, adF∗p (G∗p ) = ad∗F∗p (p), G∗p .
Comparing this with (5.25), eq. (5.24) follows.

7
F∗p is a linear map from Tp g∗ ∼
= g∗ to TF (p) R ∼
= R and therefore belongs to (g∗ )∗ ∼
= g.
Corollary. The symplectic leaves of the Kirillov-Kostant bracket are the coadjoint
orbits of G. In particular, all (finite-dimensional) coadjoint orbits have even dimension.
Proof. By the above proposition, the Hamiltonian vector field ξF associated with a
function F and evaluated at a point p ∈ g∗ is
(ξF )p = ad∗F∗p (p). (5.26)
Now, given an adjoint vector X ∈ g, we can always find a real function F on g∗ such
that F∗p = X. Accordingly eq. (5.26) implies that the integral curves of all possible
Hamiltonian vector fields going through p span the coadjoint orbit of p.
For future reference it is useful to rewrite the Kirillov-Kostant bracket in terms
of a basis {ta } of g with Lie brackets (5.2). Any adjoint vector can then be written
as X = X a ta . If {(ta )∗ |a = 1, ..., n} denotes the corresponding dual basis of g∗ , so
that h(ta )∗ , tb i = δba , any coadjoint vector can be written as p = pa (ta )∗ with real
components pa . This defines global coordinates {pa |a = 1, ...n} on g∗ , where each pa
is a real function on g∗ that associates with a coadjoint vector p its component pa .
To apply (5.23) we note that the differential (pa )∗ of pa acts on the basis vectors ∂p∂ c
according to
∂ ∂p
a
(pa )∗ = = δac . (5.27)
∂pc ∂pc
But since g∗ is a vector space we can identify Tp g∗ with g∗ by declaring that ∂/∂pc
coincides with (tc )∗ , so in fact the differential satisfies (pa )∗ ((tc )∗ ) = δac . With this
identification the differential (pa )∗ belongs to the dual of the dual, (g∗ )∗ = g, and may
be seen as an adjoint vector. Property (5.27) says that this adjoint vector is precisely
ta . The Poisson bracket follows:
{pa , pb } = fab c pc . (5.28)
In parts II and III we will see that the Poisson brackets of three-dimensional gravity
coincide with Kirillov-Kostant brackets for suitable asymptotic symmetry groups.
Remark. The Euler-Poisson equation (5.24) has numerous applications in physics,

particularly when there exists an invertible and self-adjoint “inertia operator” I : g →
g∗ .8 Indeed one can then consider a quadratic Hamiltonian function F(p) = 12 hp, I −1 pi
and eq. (5.24) becomes γ̇(t) = ad∗I −1 γ(t) γ(t). For G = SO(3) this coincides with
the equations of motion of a free rigid body; for the Virasoro group, it leads to the
Korteweg-de Vries equation (see [55] for details).
Kirillov-Kostant symplectic form

Since the coadjoint orbits of G are symplectic leaves of the Kirillov-Kostant Poisson
bracket, they have a symplectic structure given by (5.20):
8
Here self-adjointness means that hI(X), Y i = hI(Y ), Xi for any two adjoint vectors X, Y .
Definition. Let G be a Lie group, p ∈ g∗ a coadjoint vector with orbit Wp . Then

the Kirillov-Kostant(-Souriau) symplectic form at q ∈ Wp is given by
ωq (ad∗X q, ad∗Y q) = hq, [X, Y ]i (5.29)
where X, Y ∈ g.
Here ad∗X q and ad∗Y q are “infinitesimal displacements” of q and represent generic
tangent vectors of Wp at q. One can verify that (5.29) is closed and non-degenerate on
Wp , so it is indeed a symplectic form. In addition it is invariant under the coadjoint
action of G in the sense that (Ad∗f )∗ (ω) = ω for all f ∈ G. Thus each coadjoint orbit
of G is a homogeneous space equipped with a G-invariant symplectic structure. In
this sense it is a symmetric phase space. We will see below that, for instance, each
coadjoint orbit of the Poincaré group coincides with the space of classical states of a
relativistic particle with definite mass and (classical) spin.
5.1.6 Momentum maps

Noether’s theorem states that any classical system with a Lie group of symmetries
possesses conserved quantities. Here we investigate this statement in the framework of
symplectic geometry. Until the end of this section M is understood to be a manifold
acted upon by some group G according to q 7→ f · q.
Group actions and infinitesimal generators

Definition. Let G × M → M : (f, q) 7→ f · q be a smooth action of a Lie group
G on a manifold M. Then the infinitesimal generator of the action corresponding to
X ∈ g is the vector field ξX on M defined by
d tX
(ξX )q ≡ e ·q t=0
. (5.30)
dt
One can show (see e.g. [43]) that this definition implies
[ξX , ξY ] = −ξ[X,Y ] (5.31)
where the Lie bracket on the left-hand side is that of vector fields, while the bracket
on the right is that of g, given by (5.1).
For example, the representations (5.8) and (5.11) are infinitesimal generators of the
adjoint and coadjoint representations of G, respectively.9 In this language the tangent
space at q of an orbit (3.13) consists of all vectors of the form (ξX )q , where X spans
the Lie algebra g. The flow of ξX is R × M → M : (t, q) 7→ etX · q. In what follows
we study group actions where the manifold M is symplectic.
9
Property (5.31) does not contradict the fact that the adjoint and coadjoint representations of
g are actual representations, i.e. for example that adX adY − adY adX = ad[X,Y ] . Indeed, the vector
fields in (5.31) are derivations acting on functions on M, while adX and ad∗X are generally understood
as linear operators acting on g and g∗ , respectively.
Momentum maps
Let (M, ω) be a symplectic manifold. An action of G on M is symplectic if each map
q 7→ f · q is a symplectomorphism, in which case G is a symmetry group of M. If ξX
denotes the infinitesimal generator (5.30) of a symplectic action, then LξX ω = 0.
Definition. Let G × M → M : (f, q) 7→ f · q be a symplectic group action. A

momentum map for this action is a smooth map
J : M → g∗ : p 7→ J (p) (5.32)
such that, for any X ∈ g,

iξX ω = d hJ (·), Xi (5.33)
where ξX is the infinitesimal generator (5.30) and hJ (·), Xi is the real function on M
that associates with q ∈ M the value hJ (q), Xi. In the sequel we write hJ (·), Xi ≡
JX , to which we also refer as a “momentum map”.
The definition (5.33) can be compared to that of Hamiltonian vector fields, eq.
(5.19), and is equivalent to the statement
ξX = ξJX = −{JX , ·}. (5.34)
Here ξX is the infinitesimal generator (5.30), while ξJX is the Hamiltonian vector field
associated with the function JX and {·, ·} is the Poisson bracket (5.20). Thus the
function JX generates the transformation corresponding to X ∈ g in phase space, in
the sense that for any function F ∈ C ∞ (M) we have {JX , F} = −ξX (F) = −δX F.
From this observation we can derive an important corollary: if X, Y ∈ g and if we
consider the corresponding functions JX and JY , their Poisson bracket acts on classical
observables according to

{JX , JY }, F = JX , {JY , F} − JY , {JX , F}
(5.34) (5.31) (5.34)
= [ξX , ξY ](F) = −ξ[X,Y ] (F) = J[X,Y ] , F (5.35)
where in the first equality we have used the Jacobi identity. Since this property is
true for any function F, it is tempting to remove it from both ends of the equation
and conclude that the momentum map provides a representation of the Lie algebra
g. However, this hasty argument overlooks one crucial possibility, namely the fact
that brackets of momentum maps may include a central term that commutes with all
functions on phase space (see [43] or appendix 5 of [119]). Thus we conclude that:
Proposition. Provided phase space is connected, the Poisson algebra of momentum

maps is a representation of the Lie algebra g up to a (classical) central extension:
{JX , JY } = J[X,Y ] + c(X, Y ) ∀ X, Y ∈ g , (5.36)
for some real two-cocycle c on g. If the phase space has several connected components,
there may be several different cocycles (one for each connected component).
This statement is equivalent to saying that momentum maps provide a projective

representation of g, or alternatively an exact representation of a central extension of g.
It is the classical analogue of the symmetry representation theorem of section 2.1; it will
be crucial once we use geometric quantization to produce unitary representations. In
parts II and III we will see that the surface charges generating asymptotic symmetries
in gravity provide a shining illustration of this phenomenon.
Remark. Not all symplectic group actions have a momentum map, as there may be
no function JX such that (5.34) holds. If such a function exists for each X ∈ g, then
the action does admit a momentum map and is said to be Hamiltonian. Note that,
if J and J 0 are momentum maps for the same group action, then (5.33) implies that
they differ by a constant coadjoint vector (provided M is connected).
Noether’s theorem
The momentum map gives the conserved quantity J (p) ∈ g∗ associated with each
classical state p ∈ M. As anticipated earlier, coadjoint vectors may thus be seen as
“conserved vectors” for symmetric phase spaces. This interpretation stems from the
following fundamental result:
Noether’s theorem. Let q 7→ f · q be a Hamiltonian action of G on (M, ω) with

momentum map J . Let also H ∈ C ∞ (M) be a classical observable invariant under
G in the sense that H(f · q) = H(q) for all f ∈ G and all q ∈ M, and let ξH be
the associated Hamiltonian vector field (5.12). Then, for any integral curve γ(t) of ξH
with initial condition γ(0), one has

J γ(t) = J γ(0) (5.37)
for any time t belonging to the domain of the curve. In other words the dim g com-
ponents of the coadjoint vector J (γ(t)) are conserved when the equations of motion
γ̇ = (ξH )γ are satisfied.
Proof. The Hamiltonian H is invariant under G, so dtd H(etX · p) = 0 for any p ∈ M.
Since any integral curve of ξX takes the form etX · p for some initial condition p, this
is to say that ξX (H) = ξJX (H) = 0, so H is constant along integral curves of JX ;
equivalently, JX is constant along integral curves of ξH .
In a translation-invariant system the momentum map associates a momentum vec-
tor with any point in phase space. Similarly, in a rotation-invariant system it coincides
with angular momentum. Finally, in a two-dimensional conformal field theory, it co-
incides with the stress tensor of a given field configuration. We will illustrate these
statements below. In the remainder of this section we build momentum maps for
specific families of symplectic manifolds.
Momentum maps for coadjoint orbits

Let us build a momentum map (5.33) for a coadjoint orbit Wp of some group G,
endowed with the Kirillov-Kostant symplectic form (5.29). First we note that any
path γ(t) in Wp can be written as γ(t) = Ad∗f (t) p for some path f (t) in G. If γ(0) = q
and γ̇(0) = ad∗Y q for some Y ∈ g, then we find
(5.29)
ωq ad∗X q, γ̇(0) = ωq ad∗X q, ad∗Y q = hq, [X, Y ]i = had∗Y q, Xi

for any X ∈ g. Since ad∗X q is the infinitesimal generator ξX of the coadjoint action
of
G on Wp , the far left-hand side of this equation coincides with (iξX ω)q γ̇(0) . As a
consequence the momentum map (5.33) should be such that
d
had∗Y q, Xi = = hJ∗q ad∗Y q, Xi

hJ (γ(t)), Xi t=0
dt
for all X ∈ g. This implies that the differential J∗q : Tq Wp → g∗ is just the inclusion,
and determines J up to a constant coadjoint vector. In particular:
Proposition. The inclusion of the coadjoint orbit Wp in g∗ ,
→ g∗ : q 7→ q,
J : Wp ,− (5.38)
is a momentum map for the coadjoint action of G on (Wp , ω) when ω is the Kirillov-
Kostant symplectic form (5.29).
This result implies that the action of a Lie group on its coadjoint orbits is always
Hamiltonian. In fact one can show that any symplectic manifold endowed with a tran-
sitive Hamiltonian action of some group G is a covering space of a coadjoint orbit of
G. In this sense coadjoint orbits are “universal” homogeneous phase spaces.
Note that the momentum map (5.38) automatically realizes g symmetry without
central extensions. Indeed, if X, Y belong to g and if JX , JY are the corresponding
momentum maps, then at a point p ∈ g∗ we find
(5.23) (5.38)
{JX , JY }(p) = hp, [(JX )∗p , (JY )∗p ]i = hp, [X, Y ]i = J[X,Y ] (p).
This is exactly the statement (5.36) with a vanishing cocycle c. However, one should
keep in mind that the group G itself may be centrally extended.
Momentum maps for cotangent bundles

Proposition. Consider a symplectic action of G on M. Let ω = −dθ for some
symplectic potential θ. If the group action leaves θ invariant, then
JX ≡ hθ, ξX i (5.39)
defines a momentum map (5.32) that satisfies (5.36) with a vanishing cocycle c = 0.
Proof. Since the action leaves θ invariant, one has LξX θ = 0 for any X ∈ g. Writing
the Lie derivative as Lξ = d ◦ iξ + iξ ◦ d and using ω = −dθ, this is equivalent
to dhθ, ξX i = −iξX dθ = iξX ω. One may recognize this as the definition (5.33) of a
momentum map given by (5.39). In order to prove that (5.36) is satisfied with a
vanishing cocycle c, we evaluate the Poisson bracket
1 1 1
{JX , JY } = [ξY hθ, ξX i − ξX hθ, ξY i] = ω(ξX , ξY ) − hθ, [ξX , ξY ]i.
2 2 2
Here ω(ξX , ξY ) = {JX , JY } by virtue of (5.20) and (5.34), while eqs. (5.31) and (5.39)
imply that hθ, [ξX , ξY ]i = −J[X,Y ] . Eq. (5.36) follows with c = 0.
Let us now apply this result to the cotangent bundle T ∗ Q of a manifold Q. Let
φ : Q → Q be a diffeomorphism. We define the associated point transformation as
φ̄ : T ∗ Q → T ∗ Q : (q, α) 7→ φ−1 (q), α ◦ φ∗φ−1 (q) .

(5.40)
As one can verify, this definition ensures that
φ ◦ π ◦ φ̄ = π (5.41)
where π : T ∗ Q → Q is the canonical projection (5.16). Thanks to this property, one

can show (see e.g. [43]) that point transformations are symplectomorphisms:
Proposition. Consider T ∗ Q with the symplectic form ω = −dθ, where θ is the

canonical one-form (5.17). Let φ : Q → Q be a diffeomorphism and let φ̄ be the
associated point transformation (5.40). Then
(φ̄)∗ θ = θ . (5.42)
In particular, point transformations are symmetries of T ∗ Q.
Suppose now that there is an action q 7→ f · q of a Lie group G on the manifold Q.

For clarity we will also write f · q ≡ σf∗ (q), where the notation is purposely the same
as in eq. (4.17). Then one can define an action of G on T ∗ Q by
f · (q, α) ≡ f · q, α ◦ (σf∗−1 )∗f ·q .

(5.43)
Proposition (5.42) ensures that this action is symplectic and even preserves the Liou-
ville one-form. Accordingly we can apply (5.39) to build its momentum map:
Proposition. A momentum map for (5.43) is provided by the prescription
hJ (q, α), Xi ≡ hα, (ξX )q i (5.44)
where ξX is the infinitesimal generator of the action q 7→ f · q = σf∗ (q).

Proof. Applying (5.39) to the case at hand, we find a momentum map J given by
hJ (q, α), Xi = hθ(q,α) , (ξ¯X )(q,α) i where ξ¯X denotes the infinitesimal generator of (5.43).
Then (5.41) implies that π∗(q,α) (ξ¯X )(q,α) = (ξX )q , so (5.44) follows.
As an application of these results one can show for instance that the momentum
map given by (5.44) for a translation-invariant system is just the standard momentum
vector, while for a rotation-invariant system it yields the angular momentum. See [43].
5.2 Geometric quantization

Given a symmetric phase space (M, ω), one would like to understand how “quan-
tizing” that system produces unitary representation of the symmetry group. This
section is devoted to an overview of that problem, particularly as applied to coadjoint
orbits. In short, the quantum Hilbert space associated with (M, ω) will consist of
“wavefunctions”, or rather sections of suitable line bundle over M, and will indeed
provide unitary representations provided certain quantization conditions are satisfied.
Our plan is to start by reviewing the technology of line bundles and their connec-
tions, before explaining how it applies to the quantization of symplectic manifolds and
discussing the realization of unitary group representations by geometric quantization.
The presentation is condensed and superficial; we refer to [116–118] for a much more
detailed account of the subject.
5.2.1 Line bundles and wavefunctions

The basic idea of geometric quantization is to consider wavefunctions on a symplectic
manifold M. These wavefunctions are sections of a complex line bundle over M.
Recall that a fibre bundle is a quadruple (L, M, F, π) where π : L → M is a projection
and L is locally diffeomorphic to the product M × F, where M is known as the base
space and F is known as the fibre. A vector bundle is a fibre bundle whose fibres are
diffeomorphic to a vector space.
Definition. A complex line bundle L over M is a vector bundle π : L → M whose

fibres are isomorphic to C.
Thus a complex line bundle consists of infinitely many copies of the complex plane
C, one at each point of M, glued together in a smooth way (see fig. 4.1 with Op
replaced by M). The bundle locally looks like the direct product of M with C. When
this is true globally, i.e. when L is diffeomorphic to M × C, the line bundle is said to
be trivial.
Definition. Let π : L → M be a complex line bundle. Then a section of L is a map

Ψ : M → L such that π ◦ Ψ = IdM . The space of sections is denoted Γ(M, L).
When π : L → M is trivial, the space of sections coincides with the space of

complex-valued functions on M. For example, when M = R2 is interpreted as the
phase space of a particle on a line, complex functions Ψ(x, p) on R2 would provide
sections of the trivial line bundle R2 × C.
Remark. If M is a symplectic manifold and L is a line bundle over M, not all

sections of L are eligible as quantum wavefunctions because they depend on too many
arguments. In the example M = R2 just given, sections Ψ(x, p) depend both on
positions and on momenta, which violates the uncertainty principle. This will be
remedied much later by so-called polarization (see sections 5.2.2-5.2.3).
Connections and curvature

In order to define quantum operators acting on wavefunctions seen as sections of a line
bundle, we need a prescription for computing derivatives of sections. (For instance the
momentum operator typically reads P = −i∂x .) This requires a notion of covariant
differentiation:
Definition. Let L be a complex line bundle over M, VectC (M) the space of complex
vector fields on M. A connection for L is a map
∇ : VectC (M) × Γ(M, L) → Γ(M, L) : (ξ, Ψ) 7→ ∇ξ Ψ
which is linear on Γ(M, L) and VectC (M), and satisfies the property ∇F ξ Ψ = F ∇ξ Ψ
as well as the Leibniz rule ∇ξ (F Ψ) = ξ(F) Ψ + F ∇ξ Ψ for any F ∈ C ∞ (M, C). The
linear operator ∇X is known as the covariant derivative along ξ.
A connection defines a notion of parallel transport and allows one to connect, or

identify, fibres at different points. The extent to which these identifications deform
the fibres as one moves around on the base manifold is measured by the curvature
R : VectC (M) × VectC (M) × Γ(M, L) → Γ(M, L)
which is a two-form defined for all ξ, ζ ∈ VectC (M) and any section Ψ by

R(ξ, ζ)Ψ ≡ ∇ξ ∇ζ − ∇ζ ∇ξ − ∇[ξ,ζ] Ψ. (5.45)
When the curvature vanishes the connection is said to be flat. Any trivial vector
bundle admits a flat connection, but the converse is not true: there exist non-trivial
bundles with flat connections.
Hermitian structures
Since we eventually wish to interpret sections as wavefunctions, we need to define their
scalar products.
Definition. A Hermitian structure on a line bundle L → M is a smooth map

M × Γ(M, L) × Γ(M, L) → C : (q, Φ, Ψ) 7→ Φ(q)|Ψ(q) . (5.46)
which is linear in Ψ and antilinear in Φ.10 Provided M is endowed with a measure

µ, the Hermitian structure can be used to define a space of square-integrable sections
with scalar product (3.7).
Now let L be a complex line bundle over M endowed with a connection ∇ and a
Hermitian structure (5.46). We say that ∇ is Hermitian if it is compatible with the
Hermitian structure in the sense that
ξ · (Φ|Ψ) = (∇ξ Φ|Ψ) + (Φ|∇ξ Ψ) (5.47)

10
The map being “smooth” means that, given any two smooth sections Φ, Ψ, the assignment
q 7→ (Φ(q)|Ψ(q)) is smooth.

where (Φ|Ψ) is the function M → C whose value at q is Φ(q) Ψ(q) . Condition (5.47)
is the Hermitian analogue of the condition of metric-compatibility for connections on
the tangent bundle. In the realm of quantum mechanics, property (5.47) will allow us
to define self-adjoint operators.
5.2.2 Quantization of cotangent bundles

Having introduced the setup, we now return to our original problem of defining a
quantum Hilbert space associated with a symplectic manifold (M, ω). In order for
this definition to qualify as a consistent quantization prescription, the Hilbert space
H must be endowed with an operator algebra that is somehow associated with the
Poisson algebra of classical observables. This association must be a linear map that
sends a function F ∈ C ∞ (M) on a linear operator F̂ in H , in such a way that

\
F̂, Ĝ = i~ {F, G} . (5.48)
Furthermore the constant function F(p) = 1 must be mapped on the identity opera-
tor, i.e. 1̂ = I. Thus, the problem is to find a quantum/classical correspondence that
fulfills these criteria.
The solution turns out to be given by so-called geometric quantization and consists
of two steps: prequantization and polarization. Here we describe these steps for the
simple case of cotangent bundles endowed with the symplectic form (5.18), so that
ω = −dθ. More general symplectic manifolds are treated in section 5.2.3.
Prequantization
As a first attempt at quantization, let us consider the space of complex wavefunctions
on M. Their scalar products are then given by (3.7) where one may choose µ to be
the Liouville volume form (5.14). To define a linear correspondence between classical
and quantum observables, one can try to use the Hamiltonian vector fields (5.12):
? ?
F 7−→ F̂ = −i~ ξF . (5.49)
Here ~ is an arbitrary (positive) constant, to be identified with Planck’s constant.

Indeed, using (5.13) one verifies that (5.49) satisfies the basic consistency requirement
(5.48), and is thus at first sight a satisfactory quantization prescription. However,
the problem with (5.49) is that the trivial observable F(p) = 1 is mapped on the
zero operator F̂ = 0 instead of the identity. This inconsistency can be remedied by
improving (5.49) as
? ?
F 7−→ F̂ = −i~ ξF + F , (5.50)
where the second term on the right-hand side multiplies wavefunctions by F. This
modification ensures that F = 1 is represented by the identity operator, but now re-
lation (5.48) no longer holds.
We seem to be stuck: how are we to define F̂ in such a way that both condition
(5.48) and the requirement 1̂ = I be satisfied? The way out turns out to be the further
improvement that consists in adding to (5.50) the momentum map (5.39):
F 7−→ F̂ = −i~ ξF − hθ, ξF i + F, (5.51)
where θ is such that ω = −dθ. Indeed, using (5.13) one can verify that the com-
mutators of operators (5.51) close according to (5.48), and furthermore the constant
observable F = 1 is represented, as it should, by the identity operator F̂ = I. Thus,
provided θ exists, the prescription (5.51) is a consistent quantization of the algebra of
classical observables on M.
In the present case we are assuming that M = T ∗ Q is a cotangent bundle, so θ

is just the Liouville one-form (5.17) and (5.51) is a globally well-defined differential
operator that quantizes the classical observable F. One says that cotangent bundles
are quantizable. For example, on R2n with the symplectic form (5.15) the position and
momentum operators given by prequantization are
∂ ∂
q̂ j = i~ + qj , p̂j = −i~ .
∂pj ∂q j
Note that (5.51) may be seen as a differential operator
F̂ = −i~∇ξF + F (5.52)
where ∇ξ = ξ − ~i hθ, ξi is a covariant derivative determined by the connection whose

connection one-form is θ. From this viewpoint the symplectic potential is seen as an
Abelian gauge field on M = T ∗ Q, and the corresponding field strength/curvature is
the symplectic form ω = −dθ.
Polarization
Since the symplectic form is exact, the map (5.51) provides a globally well-defined
quantization prescription and our job here is almost done. But there is still a problem:
the would-be wavefunctions Ψ : M → C depend at this stage on all coordinates of
M = T ∗ Q. For example, on R2n we would have Ψ = Ψ(q i , pj ). In particular, in
the current situation one could easily devise a wavefunction with arbitrarily accurate
values of position and momentum, violating Heisenberg uncertainty. The purpose of
polarization is to cure this pathology by cutting in half the number of coordinates on
which wavefunctions are allowed to depend.
In the case of cotangent bundles it is common to declare that polarized wavefunc-

tions only depend on the coordinates of Q, and not on the transverse coordinates in
each fibre Tq∗ Q. On R2n this would correspond to saying that polarized wavefunctions
Ψ(q i ) do not depend on the coordinates pj , which is generally interpreted by saying
that wavefunctions are written “in position space” — although we shall see below that
the analogue of this polarization for semi-direct products leads instead to the “mo-
mentum space” picture of chapter 4. The scalar product of wavefunctions is obtained
by endowing the manifold Q with a measure, resulting in a Hilbert space of polarized
wavefunctions.
Polarization also affects quantum observables since they must preserve the polar-
ization while still satisfying the commutation relations (5.48). As a result, the space
of quantizable classical observables is a subset of the full space C ∞ (M). For instance,
in R2n with Darboux coordinates q i , pj (i, j = 1, ..., n), the classical observables whose
quantization preserves the polarization ∂p Ψ = 0 all take the form
F(q, p) = pj F j (q) + G(q) (5.53)
for some functions F j , G. Observables which are not of this form do not preserve
the polarization and are therefore not quantizable in this sense. One should keep
in mind, however, that this does not mean that all quantum operators acting in the
polarized Hilbert space are forced to take the form (5.53). Rather, quantizable classical
observables give rise to a vector space of Hermitian quantum operators, and the full
algebra of quantum observables is generated by sums and products of these operators.
For example, the non-relativistic Hamiltonian p̂2 is obtained by squaring the operator
that quantizes the classical observable p, although there exists no quantizable classical
observable whose quantization would yield the operator p̂2 . In this way one essentially
recovers standard quantum mechanics from the quantization of the phase space T ∗ Q.
*5.2.3 Quantization of arbitrary symplectic manifolds

We now describe geometric quantization without assuming that the symplectic
form is exact. As it turns out, relaxing that assumption leads to serious complica-
tions. Since these subtleties will have very few immediate effects on the remainder of
our exposition, we urge the hasty reader to go directly to section 5.2.4.
As before, the requirement that the commutators of quantum observables satisfy

(5.48) leads to the quantization prescription (5.51), where the one-form θ is such that
ω = −dθ. However, in contrast to cotangent bundles, there is in general no such one-
form on M because ω need not be exact. Thus the best one can do is to treat (5.51)
locally: if {Ui |i ∈ I} is a contractible open cover of M, the Poincaré lemma ensures
that there exist one-forms θi such that
ω|Ui = −dθi ∀i ∈ I (5.54)
since ω is closed. Then, locally on each Ui , one can define operators
F̂ i ≡ −i~ξF − hθi , ξF i + F (5.55)
that provide a linear correspondence between classical and quantum observables. The
problem then is to glue together operators defined on different open sets. On any
non-empty intersection Uj ∩ Uk one has dθj = dθk so there exists a function Gjk on
Uj ∩ Uk such that
θj − θk = dGjk . (5.56)
Using (5.55) one can then show that the multiplicative operator
`kj ≡ eiGkj /~ (5.57)

(acting on functions on Uj ∩ Uk ) is such that
F̂ k
= `kj ◦ F̂ j
◦ `−1
kj on Uj ∩ Uk (5.58)
for any classical observable F ∈ C ∞ (M). This result indicates that the action of F̂
on functions depends on whether one defines it on Uj or on Uk . It is an ambiguity in
the definition of the operator corresponding to F, which threatens the consistency of
the construction based on (5.55). The way out is think of F̂ as a differential operator
acting not on functions, but on sections of a complex line bundle over M. Indeed,
if the line bundle is chosen properly, one may hope that its transition functions for
some local trivialization associated with the open covering {Ui } coincide with the mul-
tiplication maps (5.57), so that the local formula (5.55) provides globally well-defined
differential operators acting on sections.
One is thus led to the problem of determining whether there exists a line bundle
whose transition functions take the form (5.57) for the covering {Ui |i ∈ I}, in such
a way that the operator (5.55) can be written globally as (5.52) for a connection ∇
whose local connection one-forms are the θi ’s. This can be addressed in the framework
of Čech cohomology, which we will not describe here. The bottom line is that such a
line bundle with such a connection exists if and only if the cohomology class of ω/2π~
2
is integral in the cohomology space Hde Rham (M, R), i.e. if
h ω i
2
∈ Hde Rham (M, Z). (5.59)
2π~
This quantization condition is equivalent to demanding that the integral of ω/2π~ over
any closed two-surface be an integer.11 The only quantizable symplectic manifolds are
those that satisfy this requirement.
The reason why we did not see this condition in the case of cotangent bundles is
that their symplectic form is globally exact, so that its cohomology class vanishes and
the requirement (5.59) is trivially satisfied. In fact one can show that the curvature
two-form (5.45) of the connection determined by (5.55) is R = iω/~, consistently with
the fact that the curvature of any line bundle is integral. In particular the connection
used to define quantum operators (5.52) for cotangent bundles is flat.
Provided the quantization condition (5.59) is satisfied, one can endow the space of
sections of the line bundle with a Hermitian structure and use it to define the scalar
product (3.7) thanks to the Liouville volume form (5.14). One can show that the Her-
mitian structure can always be chosen in a way (5.47) compatible with the connection
determined by ω, so that all operators (5.55) are Hermitian.This completes the first
step of geometric quantization, i.e. prequantization.
As in the case of cotangent bundles, the Hilbert space of sections produced by

prequantization is “too large” in the sense that wavefunctions depend on too many
arguments. Polarization corrects this problem by “cutting in half” the number of co-
ordinates on which wavefunctions are allowed to depend. Since this procedure will
11
Here “closed” means “compact without boundary”.
not be directly visible in our later considerations, we skip its presentation and refer
instead to [116, 118] for a much more thorough discussion.
Remark. When M is a coadjoint orbit, there exists a simple reformulation of the

integrality condition (5.59). Namely, if Wp ∼ = G/Gp is the coadjoint orbit of p ∈
g∗ with stabilizer Gp , the Kirillov-Kostant symplectic form (5.29) is integral if and
only if there exists a character χ of Gp whose differential (at the identity e ∈ G)
satisfies dχe = ~i j gp , with gp the Lie algebra of Gp . The textbook example of this
phenomenon is provided by coadjoint orbits of SU(2), which are spheres embedded
in su(2)∗ : the quantization condition requires that the radius of such a sphere be
an integer or half-integer multiple of ~, corresponding to the statement that highest-
weight representations of su(2) have integer or half-integer “spin”.
5.2.4 Symmetries and representations

We now combine the results of section 5.1 with the tools of geometric quantization to
address the following question: given a symplectic manifold (M, ω) acted upon by a
group G, does quantization produce a unitary representation of G?
We will assume that the action of G is Hamiltonian, with a momentum map (5.32).
We also assume that we have chosen a certain value for Planck’s constant ~ and that
ω/2π~ is integral in the sense (5.59). Then (M, ω) is quantizable and prequantization
can be carried out independently of the group action. In particular, for each adjoint
vector X ∈ g there is a classical observable JX given by (5.33), and the corresponding
operator (5.52) is
JˆX = −i~∇ξX + JX (5.60)
where we have used property (5.34) to replace ξJX by the infinitesimal generator (5.30).
By virtue of (5.36) the map X 7→ JˆX is a homomorphism, possibly up to a central
extension. Thus the assignment (5.60) provides a (generally projective) representation
of the Lie algebra g, acting on a space of sections on M.
The subtlety arises with polarization, since then the wavefunctions of the system
satisfy extra conditions which may not be preserved by (5.60). To avoid such patholo-
gies one has to choose a G-invariant polarization. In that case each operator (5.60)
is a well-defined Hermitian operator acting on polarized wavefunctions, and one ob-
tains a projective, unitary representation of g. It was shown by Kostant [120] that,
when the action of G on M is transitive, the homomorphism X 7→ JˆX exponentiates
to a unitary representation of the group G. This is true in particular when M is a
coadjoint orbit [117]. In addition, when G is semi-simple, compact or solvable, the
representations obtained in this way are irreducible. Thus geometric quantization does
produce unitary representations of groups, which is the conclusion we were hoping to
obtain.
Remark. One can discuss semi-classical approximations in symplectic terms, and

this applies in particular to coadjoint orbits. Indeed, the Liouville volume form (5.14)
measures the “size” of portions of phase space and can be used to compare identical
manifolds endowed with different symplectic structures. If ω and λω (with λ > 0) are
two symplectic forms on M, then large λ assigns a larger measure to a given portion
of (M, λω) than to the same portion in (M, ω). In this sense large λ is a semi-classical
regime with respect to (M, ω), with 1/λ playing the role of the coupling constant. In
the case of coadjoint orbits, by linearity, Wp is diffeomorphic to Wλp for any λ 6= 0,
but the definition (5.29) ensures that the symplectic form on Wλp is “larger” (for λ > 1
say) than that on Wp . Thus, for λ large enough the quantization of Wλp can be treated
semi-classically. Note that this intuition breaks down if the orbit is invariant under
scalings, i.e. Wλp = Wp .
5.3 World lines on coadjoint orbits

In this section we reformulate the observations of the previous pages in terms of action
principles and path integrals. In doing so we will develop a group-theoretic world line
formalism, which will eventually allow us (in section 5.4) to interpret representations
of semi-direct products as actual quantized point particles propagating in space-time.
We will start with general geometric considerations explaining how to associate

an action principle with any quantizable symplectic manifold. After a group-theoretic
interlude on the Maurer-Cartan form, we will focus on coadjoint orbits and describe
their world line actions as gauged non-linear Sigma models. Useful references are
[122–124]; see also [125].
5.3.1 World lines and quantization conditions

Our approach here is similar to [126]. Let (M, ω) be a symplectic manifold, p ∈ M.
Since ω is closed, there exists a neighbourhood U of p such that ω|U = −dθ for some
one-form θ on U . Now let γ : [0, 1] → U : t 7→ γ(t) be a path in phase space contained
in U . We can associate with it an action
Z Z
S U [γ] ≡ θ = − d−1 ω , (5.61)
γ γ
where the notation −d−1 ω means “whatever one-form θ such that ω = −dθ”. This is
a purely kinematical Hamiltonian action associated with the symplectic form ω. For
example, when M = R2n with ω = dq i ∧ dpi = −d(pi dq i ), expression (5.61) is globally
well-defined and reads Z 1
S[q i (t), pj (t)] = dt pj (t)q̇ j (t) (5.62)
0
which is the standard reparameterization-invariant kinetic term of any Hamiltonian
action. There is no term involving p2 or any other combination of q’s and p’s because
there is no Hamiltonian at this stage.
For a generic symplectic form ω the definition (5.61) is not enough: one needs an
action principle that makes sense for any path in M, regardless of exactness. So let
{Ui |i ∈ I} be a contractible open covering of M such that ω|Ui = −dθi for each i ∈ I.
We can then write an action (5.61) on each Ui , but we can also attempt to define S[γ]
for any path γ by Z
S[γ] ≡ − d−1 ω . (5.63)
γ
We refer to this functional as the geometric action for (M, ω) evaluated on the path
γ; its definition follows from the geometry of M. In particular, when a group G acts
on M by symplectomorphisms, the action automatically has global G symmetry. In
section 5.5 we will interpret (5.63) as the action of a point particle in space-time.
The action (5.63) can be evaluated as follows. Given a path γ, we can cover its
image by open sets Uj , with j ∈ J ⊂ I. If only one Uj suffices we can simply use the
original definition (5.61) to evaluate the action. If there are two open sets, say U1 and
U2 , then we call γj the portion of the path γ contained in Uj (for j = 1, 2) and γ12 the
portion contained in U1 ∩ U2 . We can then define
Z Z Z
S[γ] ≡ θ1 + θ2 − θ1 (5.64)
γ1 γ2 γ12
where the last term removes the overcounting due to a double integration on U1 ∩ U2 .
There is a subtlety in this expression: we chose to write ω|U1 ∩U2 = −dθ1 in the last
term, but we could equally well have chosen ω = −dθ2 ; this would have given a different
compensating term in (5.64), hence a different value for the action! This is a problem
at first sight, but one may recall that the action as such need not be a single-valued
functional on the space of paths in phase space. The truly important quantity is the
complex number
eiS[γ]/~ (5.65)
which determines the path integral measure and leads to transition amplitudes in the
quantum theory. Thus we are free to have a multivalued action as long as all ambi-
guities are integer multiples of 2π~. This is in effect a quantization condition on the
parameters of the action.
A simple reformulation of this condition is obtained by considering a closed path

γ (so γ(0) = γ(1)) and evaluating the action along that path. Using Stokes’ theorem
one can write I Z
−1
S[γ] = − d ω = − ω (5.66)
γ Σγ
where Σγ is a two-surface with boundary γ. As expected this is a multivalued func-

tional of γ. Requiring the exponential eiS[γ] to be single-valued then implies that the
integral of ω over any closed two-surface must be an integer multiple of 2π~, which
is the old Bohr-Sommerfeld quantization condition and coincides with the integral-
ity requirement (5.59) mentioned above. One can also show, more generally, that this
condition is sufficient to ensure that (5.65) is single-valued on the space of paths. Thus
the quantization condition determined by the action functional (5.63) coincides with
the condition that follows from geometric quantization. This applies, in particular, to
the coadjoint orbits of any Lie group.
Given an action (5.63) that satisfies the quantization condition, one can choose
a Hamiltonian H ∈ C ∞ (M) and compute transition amplitudes using path integrals
with the action functional
Z Z T
−1
S[γ] = − d ω − dt H(γ(t)). (5.67)
γ 0
Note that this expression is no longer invariant under time reparameterizations for
generic choices of the Hamiltonian function.
Remark. The geometric actions (5.63) associated with coadjoint orbits of centrally
extended loop groups describe certain families of Wess-Zumino-Witten models [124].
In that context the single-valuedness of (5.65) leads to the quantization of the Kac-
Moody level [127–129].
5.3.2 Interlude: the Maurer-Cartan form

When the phase space M is a coadjoint orbit Wp of a group G, any path γ on Wp can
be written as
γ(t) = Ad∗f (t) p (5.68)
for some path f (t) in G. Geometric actions such as (5.63) can then be seen as func-
tionals of paths on a group manifold. This reformulation turns out to rely on the
Maurer-Cartan form of G, which we now study.
Definition. Let G be a Lie group and let Lf : G → G : g 7→ f · g denote left

multiplication by f ∈ G. Then the (left) Maurer-Cartan form on G is
Θf ≡ (Lf −1 )∗f . (5.69)
At any point f , the map Θf is the differential of left multiplication by f −1 .
It follows from (5.69) that the Maurer-Cartan form at f is an isomorphism beween

the tangent spaces Tf G and Te G, the latter being identified as usual with the Lie
algebra of G. Thus Θ is a g-valued one-form on G and may be seen as a section of the
vector bundle T ∗ G ⊗ g. It is also left-invariant in the sense that
L∗g (Θ) = Θ (5.70)
for any group element g. When G is a matrix group, any f can be written as a
matrix and the entries of f define local coordinates on G. One can then think of
df as the matrix whose entries are the differentials of these coordinates, and the left
Maurer-Cartan form can be written as
Θf = f −1 · df . (5.71)
One can similarly define a right Maurer-Cartan form (Rf −1 )∗f , where R denotes right
multiplication (3.17). Its expression for a matrix group is df · f −1 .
Proposition. The Maurer-Cartan form (5.69) satisfies the Maurer-Cartan equation
(dΘ)(ξ, ζ) + [Θ(ξ), Θ(ζ)] = 0 (5.72)
for all vector fields ξ, ζ on G, where [·, ·] denotes the Lie bracket (5.1) in g.
Proof. Recall that the exterior derivative of Θ is such that, for all vector fields ξ, ζ,
(dΘ)(ξ, ζ) ≡ ξ · Θ(ζ) − ζ · Θ(ξ) − Θ([ξ, ζ]) , (5.73)
where [·, ·] is the Lie bracket of vector fields. If ξ and ζ are left-invariant, they can be
written as ξg = (Lg )∗e X and ζg = (Lg )∗e Y for some adjoint vectors X, Y . Then (5.69)
implies that Θ(ξ) = X is constant on g, and (5.73) reduces to
(dΘ)(ξ, ζ) + Θ([ξ, ζ]) = 0. (5.74)
By left-invariance we may write Θ([ξ, ζ]) = [Θ(ξ), Θ(ζ)] where the bracket on the
right-hand side now is the Lie bracket (5.1) of g. Eq. (5.74) then takes the form (5.72)
save for the fact that ξ and ζ are left-invariant. This condition can be relaxed upon
recalling that the span of left-invariant vector fields at a point g ∈ G is the whole
tangent space Tg G.
Kirillov-Kostant from Maurer-Cartan

Thanks to (5.68), the Maurer-Cartan form provides a convenient rewriting of the
Kirillov-Kostant symplectic form (5.29) in terms of vectors tangent to a group mani-
fold. Indeed, let
π : G → Wp : g 7→ Ad∗g (p) (5.75)
be the natural projection. We then define a two-form ω on G by
ω ≡ π∗ω (5.76)
where ω is the Kirillov-Kostant symplectic form (5.29). One may think of ω as the
analogue of (5.29) on the group G.
Lemma. Let g ∈ G, and consider tangent vectors v, w ∈ Tg G. Then
ωg (v, w) = p, [Θg (v), Θg (w)] (5.77)
where the bracket on the right-hand side is that of g.

Proof. The definition of the two-form (5.76) explicitly reads
!
(π ∗ ω)g (v, w) = ωπ(g) (π∗g v, π∗g w) = ωg (v, w). (5.78)
We can represent the vector v by a path γ in G such that γ̇(0) = v, so that
d d
Ad∗γ(t) (p)

π∗g (v) = π(γ(t)) t=0
= . (5.79)
dt dt t=0
In turn we can write γ = g · γ0 (t) where γ0 (0) = e is the identity. Then γ̇0 (0) belongs
to the Lie algebra Te G = g of G and is given by
d d (5.69)
g −1 · γ(t) t=0 =

γ̇0 (0) = Lg−1 γ(t) = (Lg−1 )∗g (v) = Θg (v) (5.80)
dt dt t=0
where we used v = γ̇(0). We can now use this in (5.79) to obtain

d (5.11) (5.80)
Ad∗g (Ad∗γ0 (t) (p)) = Ad∗g (ad∗γ̇0 (0) p) = Ad∗g (ad∗Θ(v) p).

π∗g (v) =
dt t=0
Eq. (5.77) follows upon plugging this result (and its analogue for w) in (5.78).
Formula (5.77) is sometimes rewritten as
1
ω= hp, [Θ ∧, Θ]i (5.81)
2
where [Θ ∧, Θ] is the g-valued two-form such that [Θ ∧, Θ]g (v, w) ≡ 2[Θg (v), Θg (w)]
for all tangent vectors v, w ∈ Tg G. In what follows we call ω the symplectic form on
G since it is related by (5.76) to the Kirillov-Kostant symplectic form (in particular
dω = 0), but one should keep in mind that this terminology is actually incorrect:
Lemma. The two-form ω is degenerate. Its kernel consists of left-invariant vector

fields ζX for which X belongs to the Lie algebra of the stabilizer of p.
Proof. Let v, w ∈ Tg G; there are two (unique) adjoint vectors X, Y ∈ Te G = g such
that v = (Lg )∗e X and w = (Lg )∗e Y , so that Θg (v) = X and similarly for w. Formula
(5.77) can then be rewritten as
ωg (v, w) = hp, [X, Y ]i = −had∗X p, Y i. (5.82)
The kernel of ω consists of vectors v = (Lg )∗e X such that (5.82) vanishes for any
Y ∈ g, which is to say that ad∗X p = 0. The latter property holds if and only if X
belongs to the Lie algebra of the stabilizer of p.
This lemma confirms that ω is not a symplectic form because its components do
not form an invertible matrix. The rank of ω is dim(G) − dim(Gp ), where Gp is the
stabilizer of p. This number coincides (as it should) with the dimension of the coadjoint
orbit of p, which proves by the way that the original form (5.29) on G/Gp ∼ = Wp is
invertible.
5.3.3 Coadjoint orbits and Sigma models

The “symplectic form” (5.81) is a closed two-form, and is therefore locally exact. As
such it can be used to define a kinetic action functional analogous to (5.63),
Z
S[f (t)] ≡ − d−1 ω, (5.83)
f (t)
whose argument is a path f (t) in G. Using the Maurer-Cartan equation (5.72) in

(5.77), one can write
ωf = − hp, dΘf i = −d hp, Θi f
where the exterior derivative goes through the coadjoint vector p by linearity. Thus
the action (5.83) becomes
Z Z T
dt p, Θf (t) f˙(t) .

S[f (t)] = hp, Θi = (5.84)
f (t) 0
It describes the dynamics of paths f (t) ∈ G and may be seen as the (kinetic piece of
the) action of a non-linear Sigma model. When G is a simple matrix group, adjoint
and coadjoint vectors can be identified so that hp, ·i = Tr[X·] for some X ∈ g, and
(5.71) allows us to recast the integrand of (5.84) in the form Tr Xf −1 f˙ .
Note that the global G symmetry of (5.84) is manifest: if f (t) is a path in G

and g ∈ G is an arbitrary constant group element, then left-invariance of Θ readily
implies S[g · f (t)] = S[f (t)]. In addition (5.84) is the integral of a one-form and is
thus invariant under redefinitions of the time parameter t. As in (5.67) one can include
a Hamiltonian in the action, at the cost of breaking time reparameterization invariance.
A key subtlety with (5.84) is that the group variable f (t) is the group element
that appears in a coadjoint action Ad∗f (t) p, as in (5.68). The latter coadjoint vector is
invariant under multiplication of f (t) from the right by any (generally time-dependent)
group element h(t) belonging to the stabilizer of p. This means that (5.84) should be
invariant under gauge transformations f (t) 7→ f (t) · h(t), and therefore describes a
gauged non-linear Sigma model. Let us check that (5.84) does indeed admit such a
symmetry. Using the Leibniz rule we find
Z TD E Z T D E
S[f ·h] = ˙
p, Θf (t)h(t) (Rh )∗f (t) f (t) + p, Θf (t)h(t) (Lf (t) )∗h(t) ḣ(t) (5.85)
0 0
where the adjoint vector paired with p in the first term can be rewritten as

Θf (t)h(t) (Rh )∗f (t) f˙(t) = Adh−1 Θf (t) f˙(t)
thanks to the definitions (5.69) and (5.6). This implies that the first term of (5.85)
coincides with the original action (5.84). As for the second term in (5.85), we use
left-invariance of Θ to rewrite it as a Sigma model action evaluated on a path wholly
contained in the stabilizer Gp :
Z T D E
S[h(t)] = dt p, Θh(t) ḣ(t) . (5.86)
0
The counterpart of h(t) in the coadjoint orbit of p is the constant path Ad∗h(t) p = p,
but in the Sigma model it carries a generally non-vanishing action (5.86). Thus gauge-
invariance of (5.84) may be true, but is not obvious at this stage since the gauge-
transformed action (5.85) differs from (5.84) by the extra term (5.86). To reconcile
this observation with the much desired gauge-invariance of (5.84), we note that the
exterior derivative of the integrand of (5.86) vanishes. Indeed, for all v, w ∈ Th Gp the
Maurer-Cartan equation (5.72) yields
(5.11)
d hp, Θih (v, w) = − hp, [Θh (v), Θh (w)]i = ad∗Θh (v) p, Θh (w) = 0
where the last equality follows from the fact that Θh (v) belongs to the Lie algebra
of the stabilizer of p. Thus the integrand of (5.86) is closed, and is therefore locally
exact. In particular, for a path h(t) located in a sufficiently small neighbourhood of
the identity in H, there exists a function F(t) such that
D E
p, Θh(t) ḣ(t) = Ḟ(t) (5.87)
for any t ∈ [0, T ]. The integral (5.86) of this quantity is a boundary term, so the action
functional (5.84) is indeed gauge-invariant, albeit up to boundary terms that can be
cancelled by requiring for instance that initial and final configurations be fixed.
We stress that this gauge symmetry is unavoidable if (5.84) is interpreted as the

Sigma model version of the action (5.63) on a coadjoint orbit. In particular the inclu-
sion of a Hamiltonian is now subject to a constraint: in order to reproduce (5.67), the
Hamiltonian expressed in terms of group variables must be invariant under stabilizer
gauge transformations.
*5.3.4 Coadjoint orbits and characters of SL(2, R)

As an application of the above considerations, we now classify the coadjoint orbits
of SL(2, R) and quantize some of them, showing along the way that they are equiva-
lent to one-dimensional harmonic oscillators. As an application we evaluate SL(2, R)
characters by geometric quantization. We refer e.g. to [130, 131] for more details on
the coadjoint orbits of SL(2, R), and to [122, 124] for similar computations in more
general cases. This section is not crucial for the remainder of the thesis and may be
skipped in a first reading.
Coadjoint orbits of SL(2, R)

For the basic properties of the group SL(2, R) we refer to section 4.3. Its Lie algebra
sl(2, R) consists of real, traceless 2 × 2 matrices. Any such matrix is a real linear
combination X = xµ tµ of basis elements (4.87) whose brackets read
[tµ , tν ] = µν ρ tρ . (5.88)
Here µνρ is the completely antisymmetric tensor such that 012 = +1, and indices
are raised and lowered using the Minkowski metric ηµν = diag(− + +). For future
reference we also note that, in the complex basis
`0 ≡ −t0 , `1 ≡ t2 − it1 , `−1 ≡ t2 + it1 , (5.89)
the Lie brackets (5.88) take the form
i[`m , `n ] = (m − n)`m+n (5.90)
for m, n = −1, 0, 1. On account of the isomorphism (4.83) this can also be seen as the
Lorentz algebra in three dimensions.
The sl(2, R) algebra has a non-degenerate bilinear form

(X, Y ) ≡ 2 Tr(XY ) = ηµν xµ xν (5.91)
which is left invariant by the adjoint action (4.89) of SL(2, R). We can then use the
isomorphism (4.42) to intertwine the adjoint and coadjoint representations of SL(2, R)
as in (4.43). In particular we may identify coadjoint with adjoint vectors, and coad-
joint orbits coincide with adjoint orbits under that identification. Those are exactly
the momentum orbits of the Poincaré group in three dimensions, which were described
in sections 4.2 and 4.3. This provides the classification of coadjoint orbits of SL(2, R)
and an exhaustive family of orbit representatives is depicted schematically in fig. 4.3.
Note that the fact that coadjoint orbits of SL(2, R) coincide with Poincaré mo-
mentum orbits in three dimensions follows from the structure G nAd gAb of the double
cover (4.93) of the Poincaré group. We will encounter a similar structure in the BMS3
group, albeit with an infinite-dimensional group G.
Kirillov-Kostant symplectic form

We can write any coadjoint vector of SL(2, R) as q = qµ (tµ )∗ where (tµ )∗ = η µν (tν , ·)
is the dual basis corresponding to (4.87). The components qµ are global coordinates
on sl(2, R)∗ and their Kirillov-Kostant Poisson brackets read
{pµ , pν } = µν ρ pρ (5.92)
on account of (5.88) and the general result (5.28). Now consider a “massive” orbit
n q o
Wp = qµ (tµ )∗ q0 = h2 + q12 + q22 ∼ = SL(2, R)/U(1) (5.93)
with orbit representative p = h(t0 )∗ and stabilizer U(1). We denote the “mass” of
the orbit by h rather than M because its quantization will eventually correspond to
a representation of sl(2, R) with highest weight h (or more precisely h + 1/2). The
restriction of (5.92) to the orbit gives rise to the Kirillov-Kostant symplectic form
(5.29), which we now evaluate.
We can label the points of (5.93) by their “spatial components” (q1 , q2 ). In order to
write down (5.29) in these coordinates, we need a dictionary between the components
xµ of X and those of the corresponding vector field ad∗X q in terms of the coordinates q1 ,
q2 . We first evaluate ad∗X q for X = xµ tµ ; using (5.88), for any adjoint vector Y = y µ tµ
we find had∗X q, Y i = −hq, xµ y ν µν ρ tρ i. For q belonging to (5.93) one obtains
q
∗
adX q = − q1 X + q2 X (t ) + − h2 + q12 + q22 X 2 − q2 X 0 (t1 )∗
2 1 0 ∗
q (5.94)
2 2 2 1 0 2 ∗
+ h + q1 + q2 X + q1 X (t ) .
This is an infinitesimal variation of q tangent to Wp . Any such variation can be

expressed in terms of the coordinates (q1 , q2 ): for an infinitesimal variation (δq1 , δq2 )
of (q1 , q2 ), the variation of q0 on the orbit (5.93) is
q1 δq1 + q2 δq2
δq0 = p . (5.95)
h2 + q12 + q22
The variation of q produced by a vector v = v1 ∂q1 + v2 ∂q2 tangent to Wp takes the

same form with δqi replaced by vi . Given such a vector v at q one may ask what Lie
algebra element X is such that v = ad∗X q. Owing to (5.94) and (5.95) we may choose
V2 −V1
x0 = 0 , x1 = p , x2 = p . (5.96)
h2 + q12 + q22 h2 + q12 + q22
This solution to v = ad∗X q is not unique for a given v due to the non-trivial stabilizer
U(1), but it is all we need for evaluating the Kirillov-Kostant symplectic form. Indeed,
using (5.29) and the fact that the orbit (5.93) is two-dimensional, we find
dq2 ∧ dq1
ω=p , (5.97)
h2 + q12 + q22
which coincides (up to sign) with the Lorentz-invariant volume form (1.6) on the mass
shell Wp . One can rewrite it in global Darboux coordinates
p !1/2 p !1/2
2 h2 + q12 + q22 − 2h 2 h2 + q12 + q22 − 2h
P ≡ q1 , Q≡ q2 (5.98)
q12 + q22 q12 + q22
such that (5.97) simply becomes
ω = dQ ∧ dP. (5.99)
Hence the Kirillov-Kostant symplectic form on the orbit (5.93) is globally exact and
the quantization condition (5.59) is trivially satisfied for any value of h.
Characters as path integrals

We can now quantize the orbit Wp with the symplectic form (5.99). The associated line
bundle is trivial and its sections are just complex-valued functions on Wp ; polarized
sections can be chosen to depend only on the coordinate Q. One can then evaluate
characters of suitable unitary representations of SL(2, R) by computing traces of op-
erators in the resulting Hilbert space, as follows.
The character of a representation is the trace of the exponential of a certain Lie

algebra generator. When interpreting the latter as a Hamiltonian, the character may
be seen as a partition function. Here we take the Hamiltonian to be the generator
of rotations, corresponding to the basis element t0 in (4.87). As a function on phase
space the Hamiltonian maps the point qµ (tµ )∗ on its component q0 , so on the orbit
(5.93) we have
1
q
H= h2 + q12 + q22 (5.100)
`
where we have included a prefactor12 ~c/` ≡ 1/` to ensure that H has dimensions of
energy (we think of h, qµ as being dimensionless). In Darboux coordinates Q, P , we
find the Hamiltonian of a harmonic oscillator:
h 1
H= + (P 2 + Q2 ). (5.101)
` 2`
12
We denote by c the speed of light in the vacuum.
Thus the quantization of the orbit Wp with the Hamiltonian (5.100) is a quantum
harmonic oscillator on the line!
This tremendous simplification allows us to evaluate characters. In principle we

could use the path integral formalism, but the operator approach is much simpler since
we know the spectrum of the Hamiltonian. Its eigenvalues are
h + 1/2 h + 3/2 h + 5/2 h + 1/2 + n
, , , ..., , ...
` ` ` `
each with unit multiplicity. In particular the partition function at temperature 1/β
is that of a harmonic oscillator, e−βh/` /(2 sinh[β/`]). For future reference we rewrite
it as follows: we call L0 the operator that generates rotations so that Ĥ = 1` L0 , and
we write e−β/` ≡ q. We also allow β to be complex as long as its real part is positive.
Then the partition function can be written as
q h+1/2
Tr q L0 = . (5.102)
1−q
In section 8.4.1 we will show that this is the character of a unitary representation of
the sl(2, R) Lie algebra with highest weight h + 1/2. In the present case one can think
of the “1/2” as a quantum correction to the classical weight h.
5.4 Coadjoint orbits of semi-direct products

We now apply the considerations of the previous sections to the semi-direct products13
described in chapter 4. In particular we explain how the induced representations of
section 4.1 emerge from geometric quantization of coadjoint orbits. The plan is as
follows. We first work out general expressions for the adjoint representation, the Lie
bracket and the coadjoint representation of any semi-direct product.14 Then we expose
a general classification of coadjoint orbits, seen as fibre bundles over cotangent bundles
of momentum orbits. Finally we turn to geometric quantization and describe the world
line actions associated with coadjoint orbits. The considerations of this section can be
found e.g. in [132], and also in more recent works [46,133,134]. The textbooks [135,136]
contain detailed computations and examples.
5.4.1 Adjoint representation of G n A

We consider a semi-direct product (4.1) with A a vector group. Then the Lie algebra
of G n A is a semi-direct sum
g AΣ A, (5.103)
where g is the Lie algebra of G and A is its own Lie algebra since it is a vector group.
The symbol Σ denotes the differential of the action σ at the identity, Σ : g → End(A) :
X 7→ ΣX , where ΣX is the infinitesimal generator (5.30) associated with X:
d
ΣX : A → A : α 7→ ΣX α ≡ (σ tX α)|t=0 . (5.104)
dt e
13
Here the words “semi-direct product” refer to a group (4.1) with an Abelian vector group A.
14
The sequence “group adjoint coadjoint” will be ubiquitous in this thesis.
We will denote elements of (5.103) as pairs (X, α) where X ∈ g and α ∈ A; in the

terminology of (4.6), X is an infinitesimal rotation/boost while α is a translation.
The adjoint representation of G n A is given by (5.6), which yields

(4.6) d

Ad(f,α) (X, β) = f etX f −1 , α + tσf β − σf etX f −1 α
dt t=0
= Adf X, σf β − ΣAdf X α (5.105)
where the symbol “Ad” on the right-hand side denotes the adjoint representation of
G. (More generally, in case of ambiguous notations, the argument of a group action
determines which group it refers to.) In particular, rotation generators transform ac-
cording to the adjoint representation of G, while translations are subject to mixed
transformations involving both the finite action σ and its differential Σ.
From (5.105) one can read off the Lie bracket in g A A upon using (5.8):

(X, α), (Y, β) = [X, Y ], ΣX β − ΣY α . (5.106)
The presence of Σ on the right-hand side justifies calling g A A a semi -direct sum.
Note that, if A was non-Abelian, the second entry on the right-hand side would include
a bracket of generators of A.
The structure of the algebra (5.106) can be made more transparent by choosing a
basis. Let ta be a basis of g satisfying the brackets (5.2), and let αi be a basis of A
(here a = 1, ..., dim g and i = 1, ..., dim A). Introducing the basis elements
ja ≡ (ta , 0), pi ≡ (0, αi )
that generate the semi-direct sum g A A, the Lie bracket (5.106) yields
[ja , jb ] = fab c jc , [ja , pi ] = gai k pk , [pi , pj ] = 0 (5.107)
where gai k pk ≡ Σta pi so that the coefficients (ga )i k are the entries of the matrix
representing the linear operator Σta : A → A in the basis αi . The brackets (5.107)
make the semi-direct structure manifest since the bracket [j, p] gives p’s while the
bracket [p, p] vanishes on account of the fact that A is Abelian. This structure will
appear repeatedly in this thesis.
5.4.2 Coadjoint representation of G n A

The space of coadjoint vectors of G n A is the dual of the semi-direct sum (5.103),
g∗ ⊕ A∗ . (5.108)
Its elements are pairs (j, p) where j ∈ g∗ and p ∈ A∗ , paired with adjoint vectors
according to
(j, p), (X, α) = hj, Xi + hp, αi (5.109)
where the first pairing h·, ·i on the right-hand side is that of g∗ with g while the
second one pairs A∗ with A. Note that A∗ is precisely the space of momenta (see
section 4.1), while g∗ is dual to infinitesimal rotations and may be seen as a space of
angular momentum vectors. This is consistent with the general interpretation of coad-
joint vectors as conserved quantities (see section 5.1.2) and justifies the notation (j, p).
The coadjoint representation of G n A acts on the space (5.108). In order to write

it down, it is convenient to introduce the following notation:
Definition. The cross product of translations and momenta is the bilinear map A ×
A∗ → g∗ : (α, p) 7→ α × p given for any X ∈ g by
hα × p, Xi ≡ hp, ΣX αi . (5.110)
The notation is justified by the fact that × coincides with the vector product when
G n A is the Euclidean group in three dimensions.
With this notation the coadjoint action of G n A is given by

(5.10)
Ad∗(f,α) (j, p), (X, β) = (j, p), Ad(f,α)−1 (X, β)
(5.105)
D E
= (j, p), Adf −1 X, σf −1 β + ΣAdf −1 X σf −1 α . (5.111)
We can use ΣX (σf α) = σf (ΣAdf −1 X α) to rewrite this as
(5.109)
Ad∗(f,α) (j, p), (X, β) = hj, Adf −1 Xi + hp, σf −1 β + σf −1 ΣX αi . (5.112)
In the first term of the right-hand side we recognize the coadjoint representation of G;
the part of the second term involving β is the transformation law (4.16) of momenta;
the last term involves the cross product (5.110) of σf∗ p with α. Collecting all these terms
and removing the argument (X, β), we conclude that the coadjoint representation of
G n A is
Ad∗(f,α) (j, p) = Ad∗f j + α × σf∗ p , σf∗ p

(5.113)
where we keep the notation σf∗ p instead of the simpler f · p to avoid confusion. Note
that the coadjoint action of the translation group A affects only angular momenta,
since the transformation of p only involves f ∈ G. The translation α contributes a
term α×σf∗ p, which for trivial f boils down to the cross product α×p; this contribution
can be identified with a combination of orbital angular momentum and the centre of
mass vector, while the spin angular momentum is contained in Ad∗f j. We will return
to this interpretation below.
From (5.113) we obtain the coadjoint representation (5.11) of g AΣ A:
ad∗(X,α) (j, p) = ad∗X j + α × p , Σ∗X p ,

(5.114)
where Σ∗X p ≡ −p ◦ ΣX . We will use this formula below when dealing with the Kirillov-
Kostant symplectic form.
Remark. We shall see in chapter 9 that the space of asymptotically Minkowskian

solutions of Einstein’s equations in three dimensions spans (a subset of) the space of
the coadjoint representation of the BMS3 group. Each metric will then be labelled by
a pair (j, p), where j and p are certain functions on the celestial circle that can be
interpreted as the angular momentum aspect and the Bondi mass aspect, respectively.
5.4.3 Coadjoint orbits

Let us now classify the coadjoint orbits of a semi-direct product. This may be seen as
a classification of all classical particles, analogous to the quantum classification worked
out in section 4.1. The coadjoint orbit of (j, p) is the set
W(j,p) = Ad∗(f,α) (j, p) (f, α) ∈ G n A

(5.115)
embedded in g∗ ⊕ A∗ , with Ad∗(f,α) (j, p) given by (5.113). In order to classify all

such orbits, we will assume that the orbits (4.18) and little groups (4.19) of induced
representations are known. Due to the second entry of the right-hand side of (5.113),
involving only σf∗ p, each W(j,p) is a fibre bundle over the orbit Op . The fibre above
q = σf∗ p is the set
n o
Ad∗g Ad∗f j + α × q, q g ∈ Gq , α ∈ A .

It remains to understand the geometry of these fibres and the relation between fibres
at different points. Note that in the degenerate case p = 0 the orbit W(j,0) is simply the
coadjoint orbit of j under G; in particular W(0,0) contains only one point. Accordingly
we take p 6= 0 until the end of this section.
Warm-up: Scalar orbits

We start by describing scalar orbits, that is, coadjoint orbits that contain points with
vanishing angular momentum j = 0. The terminology is justified by the fact that each
orbit is a homogeneous phase space invariant under G n A, whose quantization yields
the Hilbert space of a particle transforming under a unitary representation of G n A.
Saying that an orbit contains points with j = 0 then means that there exists a frame
where the particle’s spin vanishes, i.e. that the particle is scalar.
So let us describe an orbit W(0,p) . With j = 0 the first entry of the right-hand side
of (5.113) reduces to
α × σf∗ p. (5.116)
Keeping q = σf∗ p fixed, the set spanned by angular momenta of this form is
A × q ≡ {α × q|α ∈ A} ⊂ g∗ (5.117)
and coincides with the set of orbital angular momenta that can be reached by acting
with translations on a particle with momentum q. The geometric interpretation of
(5.117) is as follows. Recall first that the tangent space of Op at q can be identified
with the space of “small displacements” of q generated by infinitesimal boosts:
Tq Op = Σ∗X q X ∈ g ⊂ A∗ .

(5.118)
Here Σ∗X q = 0 if and only if X belongs to the Lie algebra gq of the little group Gq ,
so (5.118) is isomorphic to the coset space g/gq . It follows that the cotangent space
Tq∗ Op at q is the annihilator of gq in g∗ ,
Tq∗ Op = g0q ≡ j ∈ g∗ hj, Xi = 0 ∀ X ∈ gq ⊂ g∗ ,

(5.119)
which provides the sought-for interpretation:
Lemma. The cotangent space (5.119) coincides with the set (5.117):
Tq∗ Op = g0q = A × q. (5.120)
Proof. Let X ∈ g be an infinitesimal rotation leaving q invariant, i.e. Σ∗X q = 0. One

then has hα × q, Xi = 0 for any translation α, so α × q belongs to the annihilator g0q .
By (5.119) this implies that the span A × q is contained in Tq∗ Op . To prove (5.120) we
need to show the opposite inclusion, i.e. that any element of the annihilator g0q can be
written as α × q for some α ∈ A. To see this, consider the linear function
τq : A → g0q : α 7→ α × q (5.121)
mapping a translation on the associated orbital angular momentum. The rank of this
map is dim(A) − dim[Ker(τq )], where
Ker(τq ) = α ∈ A hΣ∗X q, αi = 0 ∀ [X] ∈ g/gq .

(5.122)
The elements of this kernel are translations constrained by dim(g) − dim(gq ) inde-
pendent conditions (the subtraction of dim(gq ) comes from the quotient by gq ). This
implies that dim[Ker(τq )] = dim(A) − dim(g) + dim(gq ), from which we conclude that
the rank of τq is
dim[Im(τq )] = dim(g) − dim(gq ) = dim(g0q ).
It follows that τq is surjective, which was to be proven.
We have just shown that the span (5.117) at each q ∈ Op is the cotangent space of
Op at q. Since j = 0, this analysis exhausts all points of W(0,p) and we conclude that
the scalar coadjoint orbits of G n A

(5.123)
are cotangent bundles of momentum orbits.
In mathematical terms we would write W(0,p) = T ∗ Op . In particular, if we have

classified all momentum orbits of G n A, then we already know the classification of
all scalar coadjoint orbits W(0,p) . Note that the map (5.121) allows us to express the
stabilizer of (0, p) in a compact way: it is a semi-direct product
Stabilizer of (j, p) = Gp n Ker(τp ) (5.124)
where Gp is the little group of p.

Spinning orbits
We now turn to spinning orbits, which generally contain no point with vanishing
total angular momentum. To begin, we pick a coadjoint vector (j, p) and restrict our
attention to rotations f that belong to the little group Gp . The resulting span is
n o
Ad∗f j + α × p, p f ∈ Gp , α ∈ A

(5.125)
and is a subset of the full orbit (5.115). In general Ad∗f (j) 6= j because the little group
Gp need not be included in the stabilizer of j for the coadjoint action of G. Noting
that the cross product (5.110) satisfies the property Ad∗f (α × p) = σf α × σf∗ p, and
using the fact that f fixes p, we rewrite (5.125) as
Ad∗f (j + β × p) , p f ∈ Gp , β ∈ A

(5.126)
where β is related to the α of (5.125) by β = σf −1 α. In particular we have
Stabilizer of (j, p) = (Gj ∩ Gp ) n Ker(τp ) (5.127)
where Gj is the stabilizer of j for the coadjoint action of G and all the remaining
6 0.
notation is as before. This extends (5.124) to the case j =
The rewriting (5.126) allows us to see that translations along β can modify at will
all components of j that point along directions in the annihilator g0p . The only piece
of j that is left unchanged by the action of translations is its restriction to gp ,
j gp
≡ jp . (5.128)
Accordingly the set (5.126) is diffeomorphic to a product

∗
Adf jp |f ∈ Gp × {α × p|α ∈ A} , (5.129)
| {z } | {z }
Wjp Tp∗ Op
where we recognize the cotangent space (5.120) and where Wjp ⊂ g∗p denotes the coad-
joint orbit of jp ∈ g∗p under the little group Gp . This is in fact our main conclusion:
when W(j,p) is seen as a fibre bundle over Op , the fibre at p is a product (5.129) of the
cotangent space of Op at p with the coadjoint orbit of the projection jp of j under the
action of the little group of p.
Inspecting (5.129), note in particular how the little group orbit Wjp factorizes from
the cotangent space A×p due to translations. This splitting is reminiscent of the repre-
sentation (4.28) of Gp n A, where the operators representing f ∈ Gp and α ∈ A live on
very different footings (and actually commute). Recall that we used this representation
to induce an irreducible representation (4.29) of the full group G n A. What we see in
(5.129) is the classical analogue of this little group representation; upon quantization,
the sub-orbit (5.129) will precisely produce a representation of the form (4.28), and its
extension to the full orbit W(j,p) will correspond to the induction (4.29). In particu-
lar the projection (5.128) is a classical definition of spin. We shall return to this below.
The arguments that led from (5.125) to the result (5.129) can be run at any other
point q on Op , except that the little group is Gq instead of Gp . Thus the fibre above
any point q = σf∗ p ∈ Op is a product of the cotangent space of Op at q with the
Gq -coadjoint orbit W(Ad∗f j)q , where (Ad∗f j)q denotes the restriction of Ad∗f j to gq . But
little groups at different points of Op are isomorphic: if one chooses standard boosts
gq ∈ G such that σg∗q (p) = q, then Gq = gq · Gp · gq−1 and gq = Adgq gp . Therefore
W(Ad∗f j)q is diffeomorphic to Wjp for any q = σf∗ p ∈ Op ; the relation between the fibres
above q and p is given by the coadjoint action of G n A.
Classification of coadjoint orbits

The conclusions of the previous paragraph can be used to classify the orbits of G n A.
We start with some terminology:
Definition. Let (j, p) be a coadjoint vector of the semi-direct product G nσ A. The

corresponding bundle of little group orbits is
n o
B(j,p) ≡ (Ad∗f j)σf∗ p , σf∗ p f ∈ G . (5.130)
According to our earlier observations, the bundle of little group orbits associated
with (j, p) is really the same as the coadjoint orbit W(j,p) , except that the cotangent
spaces at each point of Op are “neglected” since translations do not appear in (5.130).
Thus B(j,p) is a fibre bundle over Op , the fibre Fq at q ∈ Op being a coadjoint orbit of
the little group Gq . The relation between fibres at different points of Op is given by
the coadjoint action of G n A, or explicitly
(k, q) ∈ Fq iff ∃ f ∈ G such that k = Ad∗f j q and q = σf∗ p.

Conversely, suppose that two elements p ∈ A∗ and j0 ∈ g∗p are given. The group G
can be seen as a principal Gp -bundle over Op , equipped with a natural Gp -action by
multiplication from the left in each fibre. In addition Gp acts on the coadjoint orbit
Wj0 , so one can define an action of Gp on G × Wj0 by
g∈Gp
(f, k) ∈ G × Wj0 7−→ g · f, Ad∗g (k) .

The corresponding bundle of little group orbits is defined as the associated bundle
B(j0 ,p) ≡ (G × Wj0 ) /Gp . (5.131)
Thus one can associate a bundle of little group orbits (5.130) with each coadjoint or-
bit of G n A; conversely, starting from any bundle of little group orbits as defined in
(5.131), one can build a coadjoint orbit of G n A by choosing any j ∈ g∗ such that
jp = j0 and taking the orbit W(j,p) . In other words the classification of coadjoint or-
bits of GnA is equivalent to the classification of bundles of little group orbits [132,133].
These arguments yield the complete picture of coadjoint orbits of G n A:

the coadjoint orbit W(j,p) is a fibre bundle over Op , where
the fibre at q ∈ Op is a product of the cotangent space Tq∗ Op (5.132)
with a coadjoint orbit of the little group Gq .
Equivalently, W(j,p) is a fibre bundle over the cotangent bundle T ∗ Op , the fibre above
(q, α × q) ∈ T ∗ Op being a coadjoint orbit of Gq . To exhaust all coadjoint orbits of
G n A, one can proceed as follows:
1. Pick an element p ∈ A∗ and compute its momentum orbit Op under the action
σ ∗ of G; let Gp be the corresponding little group.
2. Pick jp ∈ g∗p and compute its coadjoint orbit under the action of Gp .
The set of all orbits Op and of all coadjoint orbits of the corresponding little groups
classifies the coadjoint orbits of G n A. Put differently, suppose one has classified the
following objects:
1. The orbits of G for the action σ ∗ , with an exhaustive set of orbit representatives
pλ ∈ A∗ and corresponding little groups Gλ , with λ ∈ I some index such that
Opλ and Opλ0 are disjoint whenever λ 6= λ0 ;
2. The coadjoint orbits of each Gλ , with an exhaustive set of orbit representatives
jλ,µ ∈ g∗λ , µ ∈ Jλ being some index such that Wjλ,µ and Wjλ,µ0 are disjoint
whenever µ 6= µ0 .
Then the set
(jλ,µ , pλ ) λ ∈ I, µ ∈ Jλ ⊂ g∗ ⊕ A∗

(5.133)
is an exhaustive set of orbit representatives for the coadjoint orbits of G n A. The
(generally continuous) indices λ, µ label the orbits uniquely. This algorithm is a clas-
sical analogue of the classification of representations described in section 4.1, since it
classifies the phase spaces of all “particles” associated with G n A.
5.4.4 Geometric quantization and particles

We now describe the quantization of coadjoint orbits of semi-direct products and argue
that it yields Hilbert spaces of one-particle states as described in chapter 4.
A remark on cotangent bundles

Before studying quantization we briefly digress on cotangent bundles and their canon-
ical symplectic form ω = −dθ, where θ is the Liouville one-form (5.17). Our goal
is to rewrite the symplectic form on T ∗ Q in a simpler way. For a sufficiently small
open neighbourhood U of q ∈ Q, the preimage π −1 (U ) is diffeomorphic to the product
U × Tq∗ Q. Hence the tangent space T(q,α) T ∗ Q can be written as a direct sum
T(q,α) T ∗ Q ∼
= Tq Q ⊕ Tα T ∗ Q ∼
q = Tq Q ⊕ T ∗ Q q (5.134)
which justifies writing its elements as V = (v, β), where v ∈ Tq Q and β ∈ Tq∗ Q. The
differential of (5.16) at (q, α) then reads π∗(q,α) (v, β) = v and the Liouville one-form
(5.17) reduces to
θ(q,α) (v, β) = hα, vi. (5.135)
Accordingly one finds that the canonical symplectic form ω = −dθ is
ω(q,α) ((v, β), (w, γ)) = hγ, vi − hβ, wi (5.136)
which is just a more intrinsic rewriting of the standard ω = dq ∧ dp. As we now show,
this reformulation is useful for coadjoint orbits of semi-direct products.
Quantization
Suppose we wish to quantize a coadjoint orbit W(j,p) of G n A; let ω be its Kirillov-
in g A A is (5.106), the sym-
Kostant symplectic form (5.29). Since the Lie bracket
∗
plectic form evaluated at the point Adf j + α × q, q ≡ (κ, q) in W(j,p) reads
ω(κ,q) ad∗(X,β) (κ, q), ad∗(Y,γ) (κ, q) =

(5.137)
= Ad∗f j, [X, Y ] + hα × q, [X, Y ]i + hγ × q, Xi − hβ × q, Y i .
In the two last terms of this expression we recognize the Liouville symplectic form
(5.136) on the cotangent bundle T ∗ Op when α × q is seen as an element of Tq∗ Op
thanks to (5.120). On the other hand the first term of (5.137) looks like the natural
symplectic form (5.29) on the G-coadjoint orbit of j. In particular, when X and Y
belong to the Lie algebra gq of the little group at q, the second term in (5.137) vanishes
and the first one reduces to
Ad∗f j, [X, Y ] = (Ad∗f j)q , [X, Y ]
where we use the notation (5.128). This is the natural symplectic form on the coadjoint
orbit W(Ad∗f j)q , so if we see W(j,p) as a fibre bundle over T ∗ Op with typical fibre Wjp ,
restricting the symplectic form (5.137) to a fibre gives back the symplectic form on the
little group’s coadjoint orbit. This observation actually follows from a more general
result, which states that the coadjoint orbits of a semi-direct product are obtained
by symplectic induction from the coadjoint orbits of its little groups. Symplectic in-
duction is the classical analogue of the method of induced representations that yields
irreducible unitary representations of semi-direct products. We will not dwell on the
details of this construction and refer e.g. to [133,137] for a much more thorough treat-
ment.
For quantization to be possible, the symplectic form (5.137) must be integral in the
sense (5.59). But the Liouville two-form (5.136) is exact, so its de Rham cohomology
class vanishes and demanding that (5.137) be integral reduces to demanding integrality
of the symplectic form on the coadjoint orbit of the little group. We conclude (see
e.g. [134] for the proof):
Theorem. Let G n A be a semi-direct product, (j, p) a coadjoint vector with coad-

joint orbit W(j,p) . Then W(j,p) is prequantizable if and only if the corresponding Gp -
coadjoint orbit Wjp is prequantizable.
Provided the little group orbit Wjp is quantizable, one obtains a unitary repre-
sentation R of the little group Gp acting on polarized sections of a line bundle over
Wjp . These sections are spin states; as in section 4.1, we denote their Hilbert space
by E. Upon declaring that the polarized sections on T ∗ Op depend only on the coor-
dinates of the momentum orbit Op , polarized sections on the whole orbit W(j,p) can
be seen as E-valued wavefunctions in momentum space. Assuming that there exists a
quasi-invariant measure µ on Op , the Hilbert space H obtained by quantizing W(j,p)
becomes a tensor product (3.8) of E with the space of square-integrable functions
Op → C. This exactly reproduces the construction of section 4.1.
Recovering induced representations

As the last step of quantization, we now need to understand how the group G n A
acts on polarized sections, or equivalently what differential operators represent the Lie
algebra g A A on sections. Recall that these operators take the general form (5.60)
where J is a momentum map (5.33) while ξX is an infinitesimal generator (5.30) for
the Lie algebra element X. In the present case X is replaced by a pair (X, α) ∈ g A A.
Furthermore, since the phase space is a coadjoint orbit, the momentum map is an
inclusion (5.38) and the infinitesimal generator is ξ(X,α) = ad∗(X,α) .
Let us describe this in more detail in the scalar case j = 0, so that W(0,p) = T ∗ Op .
Then the Kirillov-Kostant symplectic form coincides with the canonical symplectic
form on T ∗ Op and the operator (5.60) representing a Lie algebra element (X, α) is
Jˆ(X,α) = −i~ ad∗(X,α) (β × q, q) + hq, αi

(β×q,q)
when evaluated at a point (β×q, q) belonging to T ∗ Op . Polarized sections are functions

Ψ : W(0,p) → C : (β × q, q) 7→ Ψ(q) since they only depend on momenta q ∈ Op . Upon
acting on such a function the operator Jˆ(X,α) yields
Jˆ(X,α) · Ψ(q) = −i~ (Σ∗X q) · Ψ + hq, αiΨ(q) (5.138)
where Σ∗X q ∈ Tq Op acts on Ψ according to
d
(Σ∗X q) · Ψ ≡ − Ψ(σe∗−tX q) .
dt t=0
Thus all observables Jˆ(X,α) are polarized and can be quantized so as to satisfy (5.48).
Formula (5.138) describes the action of Hermitian operators Jˆ(X,α) on wavefunc-

tions Ψ : Op → C, provided the measure µ on Op is invariant under G. It can be
rewritten as d h −ihq,tαi/~ i
Jˆ(X,α) · Ψ (q) = i~ e Ψ(e−tX · q)
dt t=0
and thus corresponds by differentiation to the finite transformation law
T [(f, α)]Ψ (q) = e−ihq,αi/~ Ψ(f −1 · q)

(5.139)
where the map T is a representation of G n A such that

tX it ˆ
T (e , tα) = exp − J(X,α) .
~
When the measure µ on Op defining the scalar product of wavefunctions is invariant

under G, formula (5.139) is a unitary representation of G n A that coincides (up
to a sign due to different conventions) with a scalar induced representation (4.30).
We have thus recovered induced representations by quantization! The argument can
be generalized to spinning representations and to quasi-invariant measures [134, 137],
although we will not prove it here. Thus we conclude:
Theorem. Let G n A be a semi-direct product, W(j,p) one of its coadjoint orbits.

Then the unitary representation of G n A obtained by geometric quantization of W(j,p)
is an induced representation of the form (4.30) with momentum orbit Op and spin jp .
Remark. This theorem says nothing about the exhaustivity of the procedure: it does
not guarantee that all induced representations can be obtained by quantization. In fact
it is easy to work out explicit examples where certain induced representations cannot
follow from geometric quantization, for instance if the little group is not connected. In
this sense geometric quantization is somewhat weaker than the full theory of induced
representations exposed in section 4.1.
5.4.5 World lines

Geometric actions for semi-direct products can be obtained following the general
method described in section 5.3. As we now show they can be interpreted as world line
actions describing the motion of a point particle (generally with spin) in “space-time”
A. We will rely on the Sigma model picture (5.84).
We start by evaluating the left Maurer-Cartan form (5.69) for a semi-direct product
with multiplication (4.6). In order to describe a vector tangent to G n A at the point
(f, α), consider a path in G n A given by

γ(t) = g(t), β(t) (5.140)
with g(0) = f , β(0) = α and γ̇(0) ≡ v. Using the group operation (4.6) in G n A, we
then find
(5.69) d
h i
f −1 g(t), σf −1 β(t) − σf −1 α

Θ(f,α) (v) = = Θf ⊕ σf −1 (v) , (5.141)
dt t=0
where on the far right-hand side Θ denotes the Maurer-Cartan form on G. The direct
sum refers to the fact that the tangent space T(f,α) (G n A) is isomorphic to Tf G ⊕ A.
Using (5.141) we can now write the Sigma model action (5.84) associated with the
orbit of a coadjoint vector (j, p) ∈ g∗ ⊕ A∗ :
Z T Z T
(5.109)
S[f (t), α(t)] = dt j, Θf (t) (f˙(t)) + dt σf∗(t) p, α̇(t) . (5.142)
0 0
This can be recast in intrinsic terms as

Z Z
S[f (t), α(t)] = hj, Θi + hσ ∗ p, dαi (5.143)
f (t) (f (t),α(t))
where hσ ∗ p, dαi is the one-form on G n A that gives hσf∗ p, βi when evaluated at (f, α)
and acting on a vector (v, β). Note that this is just the sum of the Sigma model action
(5.84) on G with a purely kinetic scalar action functional
Z
Sscalar [f (t), α(t)] = hσ ∗ p, dαi (5.144)
(f (t),α(t))
describing a point particle propagating in A along a path α(t) with momentum q(t) =
σf∗(t) p. In particular the group A is now interpreted as “space-time”. Expression (5.144)
also has a gauge symmetry with gauge group (5.124), and it is invariant under redef-
initions of the time parameter. As in (5.67), adding a Hamiltonian generally spoils
reparameterization symmetry. In the example of the Poincaré group below the con-
dition p(t) ∈ Op will be a constraint generating time reparameterizations. Note that
this condition only applies to momenta q(t) ∈ A∗ , while the position of the particle,
α(t) ∈ A, is completely unconstrained.
5.5 Relativistic world lines

In this section we study coadjoint orbits of Poincaré groups and show that the corre-
sponding geometric actions describe world lines of relativistic particles. At the end we
also turn to Galilean world lines and show that the corresponding partition functions
coincide with Bargmann characters. These topics have been studied previously in a
number of references. The papers [138, 139] deal with the classification problem (see
also [140]); the books [141–143] describe particles in terms of quantization of Poincaré
coadjoint orbits; finally the papers [144–147] describe the relation between world line
actions and propagators of relativistic fields.
5.5.1 Coadjoint orbits of Poincaré

The classification of coadjoint orbits of the Poincaré group is an application of the gen-
eral algorithm described in section 5.4.3: all of them are fibre bundles over momentum
orbits, the fibre being a coadjoint orbit of the corresponding little group. Since mo-
mentum orbits have been classified in section 4.2, the classification of coadjoint orbits
is straightforward. Quantizing any coadjoint orbit yields an irreducible, unitary rep-
resentation of the Poincaré group, i.e. the Hilbert space of a relativistic particle.
As an example consider the (double cover of the) Poincaré group in three di-
mensions, (4.93). Its momentum orbits coincide with SL(2, R) coadjoint orbits, and
the little groups are stabilizers of SL(2, R) coadjoint vectors. All stabilizers are one-
dimensional and Abelian, except for the trivial orbit whose little group is SL(2, R).
This implies that all Poincaré coadjoint orbits are cotangent bundles of momentum
orbits, except in the case p = 0 for which W(j,0) coincides with the coadjoint orbit
of j under SL(2, R). The set of coadjoint orbit representatives for Poincaré can be
obtained by following the algorithm outlined above (5.133).
5.5.2 Scalar world lines

Let us consider a massive scalar coadjoint orbit of the Poincaré group in space-time
dimension D. We wish to work out the corresponding Sigma model action (5.144).
We refer to [144–147] for a similar approach and for spinning generalizations.
The action principle describing a scalar world line is (5.144). We choose a basis
eµ of RD such that each translation can be written as α = αµ eµ . The dual basis
consists of momenta (eµ )∗ such that hp, αi = pµ αµ for p = pµ (eµ )∗ . The argument of
the action functional (5.144) is a path in G n A, which we denote (f (τ ), x(τ )) in order
to distinguish the time parameter τ along the world line from the time coordinate
t = x0 . With the coordinates pµ just described we have σf∗(τ ) p µ ≡ pµ (τ ) for some
orbit representative p, and the action becomes
Z T
S[p(τ ), x(τ )] = dτ pµ (τ )ẋµ (τ ) with a constraint pµ (τ )pµ (τ ) = −M 2 ∀ τ,
0
where indices are raised and lowered using the Minkowski metric. The constraint
accounts for the fact that momenta must belong to a massive orbit. It can be incor-
porated in the action thanks to a Lagrange multiplier N (τ ):
Z T h i
S[p(τ ), x(τ ), N (τ )] = dτ pµ (τ )ẋµ (τ ) − N (τ ) pµ (τ )pµ (τ ) + M 2 . (5.145)
0
The equations of motion enforce the constraint
φ ≡ pµ pµ + M 2 = 0 (5.146)
and describe a point particle propagating in space-time with constant momentum:
ṗµ = 0, ẋµ = 2N pµ . (5.147)
Note how the non-trivial dynamics emerges from the fact that momenta span an orbit,
even though we haven’t included any Hamiltonian.
To rewrite (5.145) in Lagrangian form, we use the second equation of motion in

(5.147) to express momenta in terms of velocities:
ẋµ
pµ = . (5.148)
2N
Contracting this with pµ and using the mass shell constraint (5.146) then gives
ẋµ ẋµ
2
−M = . (5.149)
4N 2
Since our goal is to describe a massive particle, its trajectory must be time-like so we
require that ẋµ remains inside the light-cone at any time τ , which gives ẋµ ẋµ < 0.
This implies that (5.149) has two real solutions N ; we choose the positive one,
p
−ẋµ ẋµ
N= . (5.150)
2M
Together with (5.148) this defines an invertible Legendre transformation from the
space of positions and velocities {(xµ , ẋµ )} to the space of positions and constrained
momenta supplemented with a Lagrange multiplier,
n o
(xµ , pµ , N ) x ∈ RD , p ∈ RD such that p2 = −M 2 , N > 0 .
Upon expressing p and N in terms of ẋ thanks to this correspondence, the Hamiltonian

action (5.145) can be rewritten as
Z T
p
S[x(τ ), ẋ(τ )] = −M dτ −ẋµ ẋµ . (5.151)
0
This is an action functional

p describing the dynamics of a scalar relativistic particle,
with the Lagrangian −M −ηµν ẋµ ẋν . We have thus recovered the metric structure of
space-time from the coadjoint orbits of its isometry group.
One can also run the argument in reverse and recover the Hamiltonian action from
the Lagrangian one. In doing so one discovers that the mass shell condition is a
primary constraint generating time reparameterizations while N (τ ) is a lapse function
along the world line. The canonical Hamiltonian then reads H = 2N (p2 + M 2 ) and
vanishes on the constraint surface, as usual for generally covariant systems.
Remark. Starting from the Hamiltonian action (5.145), one can evaluate the as-
sociated transition amplitude as a path integral. This computation was performed
in [144–147] and the result turns out to coincide with the Feynman propagator of a
free scalar field with mass M . This observation is one of the starting points of the
world line formalism of quantum field theory [148–150], where scattering amplitudes
are reformulated in terms of point particles propagating in space-time.
*5.5.3 Galilean world lines

Here we study coadjoint orbits of the Bargmann group (4.103) and write down
world line actions for scalar non-relativistic particles. We also show how these actions
account for Bargmann characters.
Scalar world lines

The classification of coadjoint orbits of the Bargmann group follows from the general
considerations of section 5.4.3, combined with the classification of momentum orbits
and little groups described in section 4.4. In what follows we study the geometric
action associated with one such coadjoint orbit with mass M > 0 and spin j = 0.
The orbit then is a cotangent bundle T ∗ Op , where Op is a massive momentum orbit
(4.115). The corresponding representation describes a scalar non-relativistic particle.
In order to write down the action (5.144) we use the same trick as in (5.145) to
express the integrand in components and absorb the constraint p(τ ) ∈ Op with a
Lagrange multiplier N (τ ). (The time parameter along the world line is once again
denoted as τ , in order to distinguish it from the time coordinate t.) Using the pairing
(4.107) the world line action reads
Z T 2
i p
S x(τ ), t(τ ), p(τ ), E(τ ), N (τ ) = dτ pi ẋ − E ṫ − N −E (5.152)
0 2M
where i = 1, ..., D − 1. In principle we should also include a time-dependent central
term λ(τ ) (recall the last entry of (4.105)), but one readily verifies that its contribution
to the action is a boundary term so we neglect it from now on. This being said, note
that the presence of the central extension is crucial in giving rise to the constraint
E ≈ p2 /2M obtained by varying N . The equations of motion obtained by varying E
give N = ṫ, so we can once more interpret N (τ ) as a lapse function along the world
line. Plugging the solution of the equations of motion of N and E into (5.152), we get
Z T
p2

i

S x(τ ), t(τ ), p(τ ) = dτ pi ẋ − ṫ ,
0 2M
which we recognize as the action of a free non-relativistic particle moving in RD−1 ,
written in a reparameterization-invariant way (see e.g. chapter 4 of [151]). By express-
ing the action as an integral over the “real time” t = t(τ ), we find
Z T
p2

i
S[x(t), p(t)] = dt pi ẋ − (5.153)
0 2M
where the dot now denotes differentiation with respect to t.
Path integrals and characters

From now on we take D = 3 for simplicity. Our goal is to plug the action (5.153)
into a path integral so as to recover the Bargmann character (4.123) for r = 1. Note
that the steps leading from the original Hamiltonian action (5.152) to the quadratic
action (5.153) all go through in the path integral since they amount to integrating out
variables on which the action depends linearly.
We wish to evaluate the rotating partition function of a massive Galilean particle,

−β Ĥ+iθJˆ
Z(β, θ) = Tr e , (5.154)
where Ĥ = p̂2 /2M is the Hamiltonian and Jˆ is the angular momentum operator
Jˆ = x̂1 p̂2 − x̂2 p̂1 . (5.155)
The trace (5.154) can be interpreted as the partition function of a free non-relativistic
particle in a frame that rotates at imaginary angular velocity iθ/β. There are at least
two equivalent ways to evaluate it. The first is to compute a time-sliced path integral
Z β
p2
Z
j 1 2
Z(β, θ) = DxDp exp − dτ −ipj ẋ + − iθ(x p2 − x p1 ) (5.156)
0 2M
x(β)=x(0)
where DxDp is the standard path integral measure of quantum mechanics. In the ar-
gument of the exponential we recognize the Euclidean section of (5.153) supplemented
by a term proportional to θJ. Expression (5.156) may thus be seen as the canonical
partition function (3.50) of a system with effective Hamiltonian Ĥeff = Ĥ − iθβ J. ˆ The
second way is to realize that the operator Jˆ generates rotations in the plane. Thus
if we introduce a basis of states |x1 , x2 i localized at (x1 , x2 ), the trace (5.154) is a
(finite-dimensional) integral
Z
Z(β, θ) = dx1 dx2 Rθ · (x1 , x2 ) e−βH x1 , x2 (5.157)
R2
where Rθ ·(x1 , x2 ) denotes the action of a rotation by θ on the vector (x1 , x2 ). From this
second viewpoint, the partition function is a trace over transition amplitudes between
initial and final states that are rotated with respect to each other. Since transition
amplitudes can be written as path integrals, expression (5.157) is a path integral in
disguise and takes the same form as (5.156) up to two key differences: (i) the term
iθJ no longer appears in the exponential, and (ii) the periodicity condition on paths
is x(β) = Rθ · x(0) instead of x(β) = x(0).
The two methods just described give identical results, but we pick the second one
for simplicity. Recall that the propagator of a free massive particle on a plane is (in
Dirac notation)
iM |x0 − x|2

0 −iHt M
x ,t e x, 0 = exp (5.158)
2πit 2t
where | · | is the Euclidean norm. From this we find the Euclidean propagator

−βH M M 2
Rθ · x, t e x, 0 = exp − (1 − cos θ)x (5.159)
2πβ 2β
where x2 ≡ |x|2 . To obtain the partition function (5.157) we integrate (5.159):

Z
2 M M 2
Z(β, θ) = dx exp − (1 − cos θ)x . (5.160)
R2 2πβ 2β
For θ 6= 0 (modulo 2π) this is just a Gaussian integral and the result is precisely a
character (4.123) with r = 1. We conclude that the space obtained by quantizing a
massive coadjoint orbit of the Bargmann group coincides with the Hilbert space of a
free, massive, non-relativistic particle.
Remark. Having seen the computation of the trace of e−βH+iθJ in Bargmann rep-
resentations, one may wonder if the result can be analytically continued to the grand
canonical partition function
Z(β, Ω) = Tr e−β(H−ΩJ)

(5.161)
where Ω is a real angular velocity, describing the thermodynamics of a system in

a real rotating frame. This corresponds to taking θ = −iβΩ purely imaginary in
(5.154). If we were to evaluate (5.161), we would be led to expression (5.159) with
1 − cos θ = 1 − cosh(βΩ) < 0, which is a serious problem: the integral (5.160) would
diverge. Intuitively this divergence is due to the fact that free particles move all over
space without any potential that prevents them from escaping to infinity when put in
a rotating frame. This divergence is typical of rotating characters in flat space and
can also be seen in the Poincaré characters of section 4.2. By contrast, the partition
function (5.161) of a two-dimensional harmonic oscillator is well-defined as long as the
angular velocity Ω is smaller than the oscillator’s natural frequency.
Part II
Virasoro symmetry and AdS3 gravity
In this part we initiate the study of infinite-dimensional symmetry groups by

analysing the group of diffeomorphisms of the circle, whose central extension
is the Virasoro group. Upon defining the latter, we classify its coadjoint
orbits, i.e. orbits of CFT stress tensors under conformal transformations
in two dimensions. As an application we show how Virasoro symmetry is
realized in asymptotically Anti-de Sitter gravity in three dimensions and
interpret unitary representations of the Virasoro algebra from a gravitational
perspective. Note that Virasoro coadjoint orbits will play a key role for BMS3
particles in part III, as they will coincide with their supermomentum orbits.
Chapter 6
The Virasoro group
In the first part of this thesis we have introduced some general tools for dealing with
symmetries in quantum mechanics. Our goal is to eventually apply these tools to the
BMS3 group in three dimensions. Accordingly, in this chapter and the two next ones
we address a necessary prerequisite for these considerations by studying the central
extension of the group of diffeomorphisms of the circle, i.e. the Virasoro group. The
latter is part of the asymptotic symmetry group of many gravitational systems, where
it essentially consists of conformal transformations of celestial circles. It also accounts
for the symmetries of two-dimensional conformal field theories and thus illuminates
certain aspects of holography in general, and AdS3 /CFT2 in particular.
A word of caution is in order at the outset regarding the interpretation of the Vira-
soro group from a gravitational viewpoint. While diffeomorphisms in general relativity
are generally thought of as gauge redundancies, the group Diff(S 1 ) that we shall study
here should by no means be understood in that sense. On the contrary, it should
be interpreted as a global space-time symmetry group on a par with SL(2, R) or the
Poincaré group. In fact, in the BMS3 case, Diff(S 1 ) will be an infinite-dimensional
extension of the Lorentz group in three dimensions. Accordingly this chapter and the
next one may be seen as a detailed investigation of a group that extends Lorentz sym-
metry in an infinite-dimensional way.
Our plan for this chapter is the following. In section 6.1 we define the group Diff(S 1 )
of diffeomorphisms of the circle as an infinite-dimensional Lie group, and we describe
its adjoint representation, its Lie algebra Vect(S 1 ), and its coadjoint representation.
Section 6.2 is devoted to its cohomology; in particular we introduce the Gelfand-Fuks
cocycle and its integral, the Bott-Thurston cocycle, which respectively define the Vira-
soro algebra and the Virasoro group. In section 6.3 we study the Schwarzian derivative,
which will lead to a unified picture of Virasoro cohomology. Finally, in section 6.4 we
define the Virasoro group and work out its adjoint and coadjoint representations; the
latter coincides with the transformation law of two-dimensional CFT stress tensors
under conformal transformations.
Regarding references, the holy book on the Virasoro group is [56] while [55] is a
pedagogical introduction to infinite-dimensional group theory. Some familiarity with
two-dimensional CFT may come in handy at this stage; see e.g. [152–154].
145
Chapter 6. The Virasoro group 146
6.1 Diffeomorphisms of the circle

In this section we study the elementary properties of the group Diff(S 1 ). We first briefly
mention issues related to infinite-dimensional Lie groups, then define Diff(S 1 ) and show
that its Lie algebra consists of vector fields on the circle. We also introduce densities
on the circle, i.e. primary fields, display the coadjoint representation of Diff(S 1 ), and
discuss certain properties of the exponential map.
6.1.1 Infinite-dimensional Lie groups

The diffeomorphisms of any manifold depend on an infinity of parameters and there-
fore span an infinite-dimensional group. One would like this group to be smooth in a
certain sense, which leads to the problem of defining infinite-dimensional Lie groups
and manifolds. Here we review this question in broad terms; we refer e.g. to [155] for
a much more complete presentation.
In the same way that any finite-dimensional manifold looks locally like Rn , one
would like to find the prototypical infinite-dimensional topological vector space V
such that infinite-dimensional manifolds be locally homeomorphic to V. As it turns
out, taking V to be a Fréchet space leads to a well-defined theory of differentiation
and smoothness, which can then be used to define Fréchet manifolds. Roughly speak-
ing, Fréchet spaces are vector spaces that generalize Banach spaces. For example the
space C ∞ (M) of smooth functions on a finite-dimensional manifold M is a Fréchet
space (but not a Banach space). A Lie-Fréchet group then is a group endowed with
a structure of Fréchet manifold such that multiplication and inversion are smooth.
For instance the group Diff(M) of diffeomorphisms of a compact finite-dimensional
manifold M is a Lie-Fréchet group. From now on we refer to infinite-dimensional
Lie-Fréchet groups simply as “infinite-dimensional groups”.
Infinite-dimensional manifolds are strikingly different from finite-dimensional ones

in many respects. For example the notion of “tangent vectors” is ambiguous in infi-
nite dimension, and the lack of existence/uniqueness theorems makes other seemingly
obvious definitions fail, such as the notion of integral curves. We will encounter a
similarly counter-intuitive phenomenon below, when explaining that the exponential
that maps vector fields on diffeomorphisms is not locally surjective, in contrast with
its finite-dimensional counterpart.
In the remainder of this section we deal with the group of diffeomorphisms of the
circle as an infinite-dimensional Lie(-Fréchet) group. In particular we will think of its
Lie algebra as its tangent space at the identity, identified with the space of left-invariant
vector fields, from which the remaining definitions will follow.
6.1.2 The group of diffeomorphisms of the circle

We consider the unit circle S 1 = eiϕ ∈ C ϕ ∈ [0, 2π[ . Its fundamental group is
isomorphic to Z and its universal cover is the real line R, with a projection
p : R → S 1 : ϕ 7→ eiϕ (6.1)
depicted in fig. 2.1. This allows us to think of S 1 as the quotient R/2πZ of the real line
by the equivalence relation ϕ ∼ ϕ + 2π, since the kernel of p consists of translations
of R by integer multiples of 2π.
Diffeomorphisms in the complex plane

A diffeomorphism of the circle is a smooth bijection F : S 1 → S 1 whose inverse is also
smooth. We denote the group of all such maps by Diff(S 1 ), with the group operation
given by composition:
F ·G≡F ◦G ∀ F, G ∈ Diff(S 1 ). (6.2)
Diff(S 1 ) is an infinite-dimensional Lie group that inherits its smooth structure from
that of the Fréchet manifold of smooth maps S 1 → S 1 . Given an orientation on S 1 ,
diffeomorphisms may preserve it or break it. In particular the set of diffeomorphisms
that preserve orientation is a subgroup of Diff(S 1 ), denoted Diff+ (S 1 ) and called the
group of orientation-preserving diffeomorphisms of the circle. We will prove below
that Diff+ (S 1 ) is connected.
For practical purposes it is useful to describe diffeomorphisms of the circle in terms

of the 2π-periodic coordinate ϕ of (6.1). A diffeomorphism then is a map F : eiϕ 7→
F (eiϕ ) where F (eiϕ ) has unit norm. As an example one can verify that the set of
transformations of the form
Aeiϕ + B
F (eiϕ ) = , |A|2 − |B|2 = 1 (6.3)
B̄eiϕ + Ā
is a subgroup of Diff+ (S 1 ) isomorphic to the connected Lorentz group in three dimen-
(4.83)
sions, SO(2, 1)↑ ∼
= PSL(2, R). Rigid rotations are given by A = eiθ/2 and B = 0:
F (eiϕ ) = ei(ϕ+θ) = eiθ eiϕ . (6.4)
Similarly, the typical orientation-changing diffeomorphism is the parity transformation
F (eiϕ ) = e−iϕ . (6.5)
Any parity-changing diffeomorphism of the circle can be written as the composition of
(6.5) with an orientation-preserving transformation. There appears to be no analogue
of time-reversal in Diff(S 1 ). All in all, Diff+ (S 1 ) is an infinite-dimensional cousin of
the connected Lorentz group in three dimensions, SO(2, 1)↑ , while Diff(S 1 ) extends
the orthochronous Lorentz group O(2, 1)↑ . See also fig. 6.1 below.
Diffeomorphisms in real coordinates

Given a diffeomorphism F : S 1 → S 1 , there exists a diffeomorphism f : R → R of the
real line such that
F (eiϕ ) = eif (ϕ) , i.e. F ◦p=p◦f (6.6)
in terms of the projection (6.1). In order for f to be compatible with the periodicity
of ϕ, we must require that f (ϕ + 2π) = f (ϕ) ± 2π, where the plus sign corresponds
to an orientation-preserving diffeomorphism while the minus sign corresponds to an
orientation-changing one. In this language the rotation (6.4) corresponds to f (ϕ) =
ϕ + θ while the parity transformation (6.5) is f (ϕ) = −ϕ.
Definition. A smooth map f : R → R is 2πZ-equivariant if f (ϕ + 2π) = f (ϕ) + 2π.

Any such map can be written as f (ϕ) = ϕ + u(ϕ), where u is 2π-periodic.
In these terms, any orientation-preserving diffeomorphism F of the circle is a pro-

jection (6.6) of a 2πZ-equivariant diffeomorphism f of the real line, that is, a smooth
function f : R → R such that
f 0 (ϕ) > 0 , f (ϕ + 2π) = f (ϕ) + 2π (6.7)
for any ϕ ∈ R, where prime denotes differentiation with respect to ϕ. The group
operation (6.2) then becomes
f ·g =f ◦g (6.8)
where f, g correspond to F, G according to (6.6). From now on we always describe
Diff(S 1 ) in terms of diffeomorphisms of R satisfying the properties (6.7). By the way,
this is why we have kept writing group elements as “f ” throughout this thesis.
Note that the diffeomorphism F does not determine f uniquely: one can add to
f (ϕ) an arbitrary constant multiple of 2π without affecting F = eif . This ambiguity
can be removed by requiring e.g. that f (0) belongs to the interval [0, 2π[. One says that
f is a lift of F , and there are infinitely many lifts for a given F . For our purposes it is
only important that giving f determines F = eif uniquely, so that we can consistently
write all orientation-preserving diffeomorphisms of the circle in the form (6.7).
6.1.3 Topology of Diff(S 1 )

Lemma. The group Diff+ (S 1 ) of orientation-preserving diffeomorphisms is connec-
ted, and Diff(S 1 ) has two connected components related by parity.
Proof. Let f (ϕ) be a diffeomorphism of R that satisfies (6.7), and consider the cor-
responding diffeomorphism of the circle given by (6.6). We wish to show that there
exists a continuous path that connects f to the identity. Consider therefore the one-
parameter family of functions
ft (ϕ) = (1 − t)f (ϕ) + tϕ , t ∈ [0, 1] . (6.9)
For each t, ft satisfies (6.7) and therefore defines a diffeomorphism of the circle. At
t = 0 it coincides with f while at t = 1 it is the identity. See fig. 6.2.
For the purposes of representation theory it is important to know the fundamental
group of Diff+ (S 1 ), as it determines whether Diff+ (S 1 ) has topological central exten-
sions. In that context the key result is the following:
Lemma. Diff+ (S 1 ) is homotopic to a circle, so its fundamental group is
π1 Diff+ (S 1 ) ∼

= Z. (6.10)
g + (S 1 ) is the group of 2πZ-equivariant diffeomorphisms of R,

Its universal cover Diff
with the projection given by (6.6).
Figure 6.1: The two connected components of Diff(S 1 ) are related by parity. Compare
with the connected components of the Lorentz group in fig. 4.2.
Proof. We follow [156]. The key to the proof is to realize that Diff+ (S 1 ) is homotopic
to its subgroup Isom+ (S 1 ) of orientation-preserving isometries of the circle (for the
standard flat metric). Since Isom+ (S 1 ) is a group U(1) of rigid rotations, it will follow
that Diff+ (S 1 ) has the homotopy type of a circle and therefore has a fundamental
group Z. So let us prove the homotopy equivalence Diff+ (S 1 ) ∼ Isom+ (S 1 ). Call
Diff+ 1
0 (S ) the group of orientation-preserving diffeomorphisms of the circle leaving the
point ϕ = 0 fixed. Since isometries of S 1 are rotations, there exists a decomposition
Diff+ (S 1 ) = Diff+ 1 + 1
0 (S ) · Isom (S ). (6.11)
Indeed, any diffeomorphism of the circle is the composition of a rigid rotation with a
diffeomorphism leaving ϕ = 0 fixed; both Diff+ 1 + 1
0 (S ) and Isom (S ) are groups and their
intersection only contains the identity. Now note that any diffeomorphism preserving
ϕ = 0 admits a unique lift f such that f (0) = 0 and f (2π) = 2π, so we can think
of Diff+ 1
0 (S ) as the set of 2πZ-equivariant diffeomorphisms of R that fix the point
ϕ = 0; this identification is one-to-one. It only remains to observe that the maps
(6.9) define a homotopy whose effect at t = 1 is to retract the whole Diff+ 1
0 (S ) on the
identity. As a result Diff0 (S ) is homotopic to a point, and so by (6.11) Diff+ (S 1 ) is
+ 1
homotopic to a circle. Unwinding this circle gives rise to the group of 2πZ-equivariant
diffeomorphisms of R, which therefore span the universal cover of Diff+ (S 1 ).
This lemma confirms the interpretation of Diff+ (S 1 ) as an infinite-dimensional ana-
logue of PSL(2, R), since the latter is also homotopic to a circle (see section 4.3). In
particular formula (6.11) is the Diff(S 1 ) analogue of the Iwasawa decomposition (4.81)
of SL(2, R). Since Diff+ 1
0 (S ) has the homotopy type of a point, the group Diff (S )
+ 1
may be seen as an infinite-dimensional cylinder S 1 × R∞ where S 1 consists of rigid

rotations while R∞ is spanned by infinite-dimensional generalizations of boosts. Note
that property (6.10) implies the existence of topological projective representations of
Diff(S 1 ). Applied to BMS3 , it will imply that the spin of massive particles is not
quantized (as in the Poincaré group in three dimensions).
In what follows we focus on the universal cover Diff g + (S 1 ) rather than Diff(S 1 )
or Diff+ (S 1 ), except if explicitly stated otherwise. To reduce clutter we will abuse
notation by writing Diff(S 1 ) for the universal cover, instead of the more accurate
g + (S 1 ). Accordingly, from now on elements of Diff(S 1 ) are diffeomorphisms
notation Diff
f , g, etc. of the real line satisfying the properties (6.7). The inverse of f will be denoted
f −1 and is such that f (f −1 (ϕ)) = f −1 (f (ϕ)) = ϕ.
Figure 6.2: The homotopy (6.9) turns a diffeomorphism f (ϕ) (here leaving the point
ϕ = 0 fixed) into the identity. It implies both that the group Diff+ (S 1 ) is connected
and that it is homotopic to a circle.
6.1.4 Adjoint representation and vector fields

We can now look for the Lie algebra of Diff(S 1 ), which is identified with the tangent
space at the identity and corresponds to infinitesimal diffeomorphisms. It is intuitively
clear that this algebra is a space of functions, since a diffeomorphism close to the
identity can be written as
f (ϕ) = ϕ + X(ϕ) (6.12)
where is “small” and X(ϕ) is a function on the circle. A more subtle problem
is to determine the adjoint action of diffeomorphisms on this Lie algebra, and to
deduce the expression of the Lie bracket. In order to work this out we pick a path
γ : R → Diff(S 1 ) : t 7→ γt such that γ0 is the identity and
γt (ϕ) = ϕ + tX(ϕ) + O(t2 ) (6.13)
for small t. The adjoint representation is defined by (5.6), so we find
dh i (6.13) dh i
f γt f −1 (ϕ) f f −1 (ϕ) + tX f −1 (ϕ)

Adf (X) (ϕ) = = .
dt t=0 dt t=0
(6.14)
Since t is “small” in this expression, we can Taylor expand
f f −1 (ϕ) + tX f −1 (ϕ) = ϕ + tX f −1 (ϕ) f 0 f −1 (ϕ) + O(t2 )

(6.15)
where we have used f (f −1 (ϕ)) = ϕ. The derivative of the latter equation implies
1
f 0 (f −1 (ϕ)) = (6.16)
(f −1 )0 (ϕ)
which can be plugged into (6.15) and thus provides the adjoint representation
X (f −1 (ϕ))
Adf (X) (ϕ) = . (6.17)
(f −1 )0 (ϕ)
This formula is the transformation law of X(ϕ) under a diffeomorphism ϕ 7→ f (ϕ).

It shows in particular that X(ϕ) in (6.12) is not just a function on the circle, due to
the derivative of f −1 in (6.17). Using (6.16), we can also rewrite it by evaluating the
left-hand side at f (ϕ) rather than ϕ:
Adf (X) (f (ϕ)) = f 0 (ϕ)X(ϕ)

(6.18)
We recognize here the transformation law of the component X(ϕ) of a vector field
∂
X(ϕ) (6.19)
∂ϕ
under a diffeomorphism f . We shall denote by Vect(S 1 ) the space of smooth vector

fields on S 1 , whose elements will be written as X, Y , etc. We have just shown that
Vect(S 1 ) is the Lie algebra of Diff(S 1 ); it now remains to find the Lie bracket.
Take once more a path γt in Diff(S 1 ) that satifies (6.13). Picking a vector field
Y ∈ Vect(S 1 ) and a point ϕ ∈ [0, 2π], let us evaluate
Y (γt−1 (ϕ))

(5.8) d (6.17) d
(adX (Y )) (ϕ) = Adγt (Y ) t=0 (ϕ) = . (6.20)
dt dt (γt−1 )0 (ϕ) t=0
Here (6.13) implies (γt−1 )0 (ϕ) = 1 − tX 0 (ϕ) as well as Y (γt−1 (ϕ)) = Y (ϕ) − tX(ϕ)Y 0 (ϕ)
to first order in t. Plugging these expressions in (6.20) we obtain adX Y = −XY 0 +Y X 0 ,
where it is understood that both sides are evaluated at the same point ϕ. We conclude
that the Lie bracket of Vect(S 1 ), seen as the Lie algebra of the group Diff(S 1 ), is the
opposite of the standard Lie bracket of vector fields:
[X, Y ]Lie algebra = −[X, Y ]Vector fields . (6.21)
This is in fact a common phenomenon: as a consequence of (5.31), the group Diff(M) of

diffeomorphisms of a (compact, finite-dimensional) manifold M is a Lie-Fréchet group
whose Lie algebra is the space Vect(M) endowed with the opposite of the standard
Lie bracket of vector fields [43, 55, 56]. Thus:
Proposition. The Lie algebra of the group Diff(S 1 ) is the space Vect(S 1 ) of vector
fields on the circle, with the Lie bracket (6.21) given by the opposite of the standard
Lie bracket of vector fields.
In what follows we will bluntly neglect the sign subltety and endow Vect(S 1 ) with
the usual bracket
∂
[X, Y ] = X(ϕ)Y 0 (ϕ) − Y (ϕ)X 0 (ϕ) . (6.22)
∂ϕ
This is a harmless abuse of conventions and may be seen as an alternative definition
of the Lie bracket for groups of diffeomorphisms. With that abuse the Lie algebra of
Diff(S 1 ) becomes Vect(S 1 ) with the usual Lie bracket of vector fields.
Witt algebra
Since all functions on the circle can be expanded in Fourier series, any vector field is
a (generally infinite) complex linear combination of generators
`m ≡ eimϕ ∂ϕ , m ∈ Z. (6.23)
The brackets (6.22) of these generators can be written as
i[`m , `n ] = (m − n)`m+n , (6.24)
where one may recognize the Witt algebra of conformal field theory. It is an infinite-
dimensional extension of the sl(2, R) algebra (5.90) spanned by `−1 , `0 , `1 . The latter
consists of vector fields X(ϕ)∂ϕ with X(ϕ) = X0 + X1 cos ϕ + X2 sin ϕ for Xµ ∈ R, and
generates diffeomorphisms of the form (6.3). In particular the constant vector field `0
generates rigid rotations of the circle.
6.1.5 Primary fields on the circle

Formula (6.18) gives the transformation law of vector fields on the circle under diffeo-
morphisms. Similarly, a one-form α(ϕ)dϕ would transform as α 7→ f · α, where
α(ϕ)
(f · α)(f (ϕ)) = , (6.25)
f 0 (ϕ)
while a function α(ϕ) would simply transform as (f · α)(f (ϕ)) = α(ϕ). Vector fields,
one-forms and functions can all be seen as sections of suitable vector bundles on the
circle, which suggests that they can be generalized to sections of tensor product bundles
such as T S 1 ⊗ · · · ⊗ T S 1 or T ∗ S 1 ⊗ · · · ⊗ T ∗ S 1 .
Definition. A density of weight h ∈ R on the circle is an expression of the form

α = α(ϕ)(dϕ)h (6.26)
where α(ϕ) is a smooth function on the circle; it acts on the tangent space Tϕ S 1
according to α(ϕ)dϕh , V ∂ϕ ≡ α(ϕ)V h . We denote by Fh (S 1 ) the vector space of
densities of weight h.
When h is an integer, a density of weight h is a section of

T ∗ S 1 ⊗ · · · ⊗ T ∗ S 1 if h ≥ 0, or T S 1 ⊗ · · · ⊗ T S 1 if h < 0.
| {z } | {z }
|h| times |h| times
In particular, a density is a vector field when h = −1, a one-form when h = 1, and a

function when h = 0. The definition (6.26) generalizes these notions to arbitrary real
values of h. The notation α is justified by the fact that in the BMS3 group, super-
translations will be densities with weight −1.
Expression (6.26) suggests that the density α(ϕ)(dϕ)h is a coordinate-independent

quantity. Indeed, under a diffeomorphism f : ϕ 7→ f (ϕ), it transforms as
h
(f · α)(ϕ) ≡ (f −1 )0 (ϕ) α(f −1 (ϕ)) (6.27)
or equivalently as
α(ϕ)
(f · α)(f (ϕ)) ≡ . (6.28)
(f 0 (ϕ))h
This reduces to (6.18) for h = −1 and to (6.25) for h = 1. If we think of f (ϕ) as
a “conformal transformation” of the circle, then eq. (6.28) coincides with the trans-
formation law of a (chiral) primary field of weight h. It provides a representation of
Diff(S 1 ) in the space Fh (S 1 ). This representation is infinite-dimensional and generally
non-unitary because the would-be “scalar product”
Z 2π
dϕ α(ϕ)β(ϕ) , α, β ∈ Fh (S 1 ) (6.29)
0
is not left invariant by (6.28) for generic values of h. The only exception is the case
of spinor fields, h = 1/2. One can think of (6.28) as a Diff(S 1 ) generalization of the
various finite-dimensional (but non-unitary) irreducible representations of the Lorentz
group. The number h can then be thought of as a spin label, in the same way that
finite-dimensional Lorentz representations correspond to transformation laws of rela-
tivistic fields with definite spin. (Beware: the word “spin” here does not refer to the
notion of “spin” encountered in representations of semi-direct products. These two no-
tions are related in that the Lorentz spin of a quantum field determines the Poincaré
spin of the corresponding particles, but they are nevertheless different concepts.)
From (6.28) one can read off the transformation law of densities under infinitesimal
diffeomorphisms, that is, under vector fields on the circle. Taking f (ϕ) = ϕ + X(ϕ)
in (6.28) one finds, to first order in ,
(f · α)(ϕ) = α(ϕ) − X(ϕ)α0 (ϕ) + h α(ϕ)X 0 (ϕ) .

(6.30)
We then define
(f · α)(ϕ) − α(ϕ)
X · α(ϕ) ≡ − (6.31)

and obtain
X · α = Xα0 + h αX 0 . (6.32)
As before, one may recognize here the infinitesimal transformation law of a primary
field α of weight h under an conformal transformation generated by X.
6.1.6 Coadjoint representation of Diff(S 1 )

Dual spaces
We mentioned around (6.29) that the integral of the product of two densities with the
same weight is generally not invariant under diffeomorphisms. However, there does
exist a Diff(S 1 )-invariant pairing of densities. Indeed, consider the space Fh (S 1 ) of
densities of weight h. Its dual space consists of all linear forms
p : Fh (S 1 ) → R : α 7→ hp, αi . (6.33)
Since Fh (S 1 ) is infinite-dimensional, its dual space is pathological: the map (6.33)

need not be continuous and therefore does not preserve the differentiable structure
of Fh (S 1 ). Accordingly one generally restricts attention to the space of continuous

linear forms (6.33); the latter coincides with the space of distributions on the circle.
In addition, for concrete computations it is much more convenient to consider only
regular distributions, that is, distributions that can be written in the form
Z 2π
1
hp, αi = dϕ p(ϕ)α(ϕ) (6.34)
2π 0
where p(ϕ) is a smooth function on the circle. We will call the space of such distribu-
tions the smooth or regular dual of Fh (S 1 ). As a vector space, it is isomorphic to the
space C ∞ (S 1 ) of smooth functions on the circle. Note that any distribution can be
obtained as the limit of a sequence of regular distributions, so in this sense we are not
missing anything even when restricting attention to regular distributions. The regular
dual of Fh (S 1 ) will be denoted as Fh (S 1 )∗ . The notation in (6.33) is justified by the
fact that, in BMS3 , p will be an infinite-dimensional supermomentum vector dual to
supertranslations.
Since Fh (S 1 ) carries a representation (6.28), it is natural to ask how the dual repre-
sentation (4.16) acts on the regular dual. By definition, one has hf · p, αi = hp, f −1 · αi
for any diffeomorphism f , any density α ∈ Fh (S 1 ) and any smooth distribution
p ∈ Fh (S 1 )∗ . Using (6.27) and the pairing (6.34), we get
Z 2π
1
hf · p, αi = dϕ p(ϕ)(f 0 (ϕ))h α(f (ϕ)) . (6.35)
2π 0
If now we rewrite hf · p, αi as an integral (6.34) with the integration variable ϕ replaced

by f (ϕ), the condition that hf · p, αi matches the right-hand side of (6.35) for any α
readily provides
p(ϕ)
(f · p)(f (ϕ)) = 0 .
(f (ϕ))1−h
This is the transformation law (6.28) with h replaced by 1 − h:
Proposition. There is an isomorphism Fh (S 1 )∗ ∼ = F1−h (S 1 ) which is compatible

1
with the natural action of Diff(S ) on these spaces. In addition, the pairing between
Fh (S 1 ) and F1−h (S 1 ) is Diff(S 1 )-invariant in the sense that hf · p, f · αi = hp, αi for
all f ∈ Diff(S 1 ) and all densities p ∈ F1−h (S 1 ), α ∈ Fh (S 1 ).
Thus the duals of densities with weight h are densities with weight 1 − h, and
vice-versa. One can apply this to the examples encountered above:
• the duals of vector fields (h = −1) are quadratic densities (h = 2);
• the duals of functions (h = 0) are one-forms (h = 1);
• the duals of spinor fields (h = 1/2) are spinor fields (i.e. F1/2 (S 1 ) is self-dual).
Note that, for all values of h except h = 1/2, the conformally invariant pairing (6.34)
is not a scalar product since its arguments are densities whose transformation laws
under Diff(S 1 ) differ. This should be contrasted with the finite-dimensional examples
encountered in chapter 4, where the existence of an invariant bilinear form led to the
equivalence of σ ∗ and σ. We shall see below that this difference is crucial for coadjoint
orbits of the Virasoro group (section 7.1) and hence for the supermomentum orbits of
the BMS3 group (see part III).
Coadjoint representation
Since the adjoint representation of Diff(S 1 ) is the transformation law (6.18) of vector
fields, we now know that the coadjoint representation of Diff(S 1 ) is the transformation
law of quadratic densities:
p(ϕ)
(Ad∗f p)(f (ϕ)) = (6.36)
(f 0 (ϕ))2
This can also be written infinitesimally as
ad∗X p = Xp0 + 2X 0 p. (6.37)
In CFT terminology, vector fields are infinitesimal conformal transformations and their
duals are (quasi-)primary fields with weight h = 2, that is, CFT stress tensors. Indeed
formula (6.36) is the transformation law of a stress tensor p(ϕ) if we think of the map
ϕ 7→ f (ϕ) as a conformal transformation. Similarly, the duals of functions are primary
fields with weight h = 1, i.e. currents. From now on we sometimes refer to Diff(S 1 )-
invariance as “conformal invariance”. Note that at this point we haven’t included any
central charge yet; this will change once we turn to the Virasoro group.
6.1.7 Exponential map and vector flows

We mentioned above that each vector field X(ϕ)∂ϕ may be seen as an infinitesimal
diffeomorphism; let us make this more precise. Given a vector field X(ϕ)∂ϕ , its integral
curves are paths ϕ(t) on the circle that satisfy the evolution equation
ϕ̇(t) = X(ϕ(t)). (6.38)
In particular, when t = is “small” one finds that ϕ() takes the form (6.12) with
initial condition ϕ(0) = ϕ. Equation (6.38) is an ordinary differential equation in one
dimension and X(ϕ) is smooth, so given an initial condition ϕ(0), the solution exists
and is unique. We define the flow of X as the one-parameter family of diffeomorphisms
that maps a “time” t and an initial condition ϕ on the point ϕ(t) obtained by solving
(6.38) with this initial condition. If we call this solution ϕ̃(t, ϕ), then the flow of X is
φX : R × S 1 → S 1 : (t, ϕ) 7→ ϕ̃(t, ϕ). (6.39)
For example the flow of the constant vector field X(ϕ) = 1 is given by ϕ̃(t) = ϕ + t
and consists of rigid rotations by t, as already anticipated above. Using the notion of
flow, one can define an exponential map for Diff(S 1 ):
Definition. The exponential map of the group Diff(S 1 ) is
exp : Vect(S 1 ) → Diff(S 1 ) : X(ϕ)∂ϕ 7→ exp[X] ≡ φX (1, ·) (6.40)
where φX is the flow (6.39) of X. In other words the diffeomorphism exp[X](ϕ) is

obtained by requiring that the equality
Z exp[X](ϕ)
dφ
=1 (6.41)
ϕ X(φ)
holds for any initial condition ϕ.
In any finite-dimensional Lie group, the exponential map (5.3) is a local diffeomor-
phism, so any group element belonging to a suitable neighbourhood of the identity can
be written as the exponential of an element of the Lie algebra. However, this is not
so for groups of diffeomorphisms: one can show that the exponential map (6.40) does
not define a local chart on Diff(S 1 ) in that it is neither locally injective, nor locally
surjective. The idea of the proof is to build and explicit family of diffeomorphisms that
are arbitrarily close to the identity but cannot be written as exponentials of vector
fields. This being said, the exponential map is always well-defined on a Lie-Fréchet
group, even when it is not locally surjective. See [55, 56] for details.
6.2 Virasoro cohomology

As emphasized in chapter 2, cohomology is crucial for quantum-mechanical appli-
cations: it measures the possible “deformations” of a group structure (e.g. central
extensions), which typically do occur in quantum mechanics. When an algebra is
finite-dimensional and semi-simple, Whitehead’s lemma (2.23) ensures that there are
essentially no deformations; the same is true of the Poincaré group (4.40). By contrast,
we have seen how crucial cohomology is for the Galilei group (4.103), since its central
extension gives rise to the notion of mass.
With this motivation, the present section is devoted to the cohomology of Diff(S 1 )
and its Lie algebra. These considerations will eventually lead to the definition of
the Virasoro algebra, so we refer to them as “Virasoro cohomology”. We will start
by describing the real cohomology groups of Vect(S 1 ) and of Diff(S 1 ), then turn to
cohomologies whose cochains are primary fields on the circle. The results summarized
here are discussed at greater length in [56].
Remark. We use the notation and conventions of chapter 2, and all cochains are
required to be smooth. Lie algebra cochains are denoted by lowercase sans serif letters
such as b, c, s, etc. while group cochains are denoted by uppercase letters B, C, S, etc.
6.2.1 The Gelfand-Fuks cocycle

Here we derive the first and second real cohomology groups of Vect(S 1 ); in particular
we introduce the Gelfand-Fuks cocycle, which will eventually give rise to the Virasoro
algebra. We also describe higher-degree real cohomology groups.
Cohomology in degrees one and two

The computation of the first cohomology of Vect(S 1 ) is immediate: since any vector
field can be written as the bracket (6.22) of two other vector fields, the first cohomology
group (2.19) of Vect(S 1 ) vanishes:
H1 Vect(S 1 ) = 0.

(6.42)
In other words there is no non-trivial real one-cocycle on Vect(S 1 ). The second coho-
mology of Vect(S 1 ) is far more interesting:
Theorem. The second real cohomology space of Vect(S 1 ) is one-dimensional. It is

generated by the class of the Gelfand-Fuks cocycle
Z 2π
1
c(X, Y ) ≡ − dϕX(ϕ)Y 000 (ϕ) (6.43)
24π 0
whose expression in the basis (6.23) is
m3
c(`m , `n ) = −i δm+n,0 . (6.44)
12
Proof. Let c be a real two-cocycle on Vect(S 1 ). Then dc = 0 where d is the Chevalley-
Eilenberg differential (2.15) for the trivial representation T . In terms of the basis
(6.23), the statement dc = 0 is tantamount to
c([`m , `n ], `p ) + c([`n , `p ], `m ) + c([`p , `m ], `n ) = 0 (6.45)
for all integers m, n, p. Taking p = 0 and using the antisymmetry of c we get
c(`0 , [`m , `n ]) = c([`0 , `m ], `n ) + c(`m , [`0 , `n ]). (6.46)
Here we can interpret the left-hand side as the differential of the one-cochain k =
c(`0 , ·), so the left-hand side is exact while the right-hand side is a Lie derivative1
c([`0 , `m ], `n ) + c(`m , [`0 , `n ]) = ((i`0 ◦ d + d ◦ i`0 ) · c) (`m , `n ) = (L`0 c)(`m , `n ) (6.47)
where we used dc = 0. Since the left-hand side of (6.46) is exact, we conclude that Lie
derivation with respect to `0 leaves the cohomology class of c invariant. (In geometric
terms `0 generates rotations, so this says that the cohomology class of c is invariant
under rotations.) This allows us to turn c into a rotation-invariant cocycle. Indeed,
let b be a one-cochain and define c̃ ≡ c + db, which has the same cohomology class as
c. The Lie derivative of c̃ with respect to `0 now is
(6.46)
L`0 c̃ = L`0 c + L`0 db = dk + d(i`0 (db)) = d k + db(`0 , ·) . (6.48)
In order to make c̃ invariant under rotations, we need to choose b such that (6.48)
vanishes. One verifies that the definition
i
b(`m ) = c(`0 , `m ) for m 6= 0 (6.49)
m
1
We denote by i the interior product of cochains.
satisfies this requirement for any b(`0 ). Thus, from now on we work only with the
rotation-invariant cocycle c̃ and we rename it into c for simplicity. Then we have
c(`0 , `m ) = 0, and eq. (6.46) becomes
c([`0 , `m ], `n ) + c(`m , [`0 , `n ]) = 0 (6.50)
for all integers m, n. The Lie brackets (6.24) then yield
(m + n) c(`m , `n ) = 0 (6.51)
and thus imply that c(`m , `n ) = 0 whenever m + n is non-zero. Writing c(`m , `n ) =
cm δm+n,0 for some coefficients cm = −c−m , we are left with the task of determining the
cm ’s with m > 0. Returning to the cocycle identity (6.45) with p = −m − 1 and using
once more the brackets (6.24), we find
(2 + m)cm − (2m + 1)c1
cm+1 = (6.52)
m−1
for m ≥ 2. This shows that all cm ’s are determined recursively by c1 and c2 . In
particular, we now know that the cohomology space H2 Vect(S 1 ) is at most two-
dimensional; the choices cm = m3 and cm = m are indeed two linearly independent
solutions of the recursion relations (6.52). Now note that, if c is a coboundary c = dk
for some one-cochain k, then
(2.15) (6.24)
c(`m , `n ) = dk(`m , `n ) = k([`m , `n ]) = −i(m − n)k(`m+n )
so that c(`m , `−m ) = −2imk(`0 ) always depends linearly on m. Accordingly, the
solution cm = m of the recursion relations (6.52) yields a trivial cocycle, while cm = m3
is non-trivial. We conclude that, up to a coboundary, any non-trivial two-cocycle on
Vect(S 1 ) reads
c(`m , `n ) = N m3 δm+n,0 (6.53)
2 1

for some normalization N 6= 0. In particular, H Vect(S ) is one-dimensional.
Higher degree cohomologies

The real cohomology groups of Vect(S 1 ) increase in complexity as their degree be-
comes higher. Since we will not need any degree higher than two, we restrict ourselves
here to a qualitative description of the result (details can be found in [56]).
The first step is to fix the kind of cochains one wants to study. For Vect(S 1 ) it is
natural to consider local real cochains
c : Vect(S 1 )k → R : (X1 , ..., Xk ) 7→ c(X1 , ..., Xk ) (6.54)
that take the form of an integral over S 1 of some “cochain density” C :
Z 2π
(n ) (n )
c(X1 , ..., Xk ) = dϕ C X1 (ϕ), X10 (ϕ), ..., X1 1 (ϕ), ..., Xk (ϕ), ..., Xk k (ϕ) .
0
Here the word local is used in the same sense as in field theory. The Gelfand-Fuks
cocycle (6.43) is of that form, with a density C (X, Y ) ∝ XY 000 . With this restriction on

k
the allowed cochains, one can study the resulting cohomology groups Hloc Vect(S 1 ) .
The result is as follows:
k

Proposition. The real local cohomology groups Hloc Vect(S 1 ) are all trivial except
if k is equal to 0, 2 or 3, in which case the cohomology group is one-dimensional:
(
k R if k ∈ {0, 2, 3}
Vect(S 1 ) =

Hloc (6.55)
0 otherwise.
0
The generator of Hloc is the class of any non-zero constant function on Vect(S 1 ); that
2 3
of Hloc is the Gelfand-Fuks cocycle (6.43). Finally Hloc is generated by the class of
the Godbillon-Vey cocycle
 
Z 2π X Y Z
dϕ det  X 0 Y 0 Z 0  (6.56)
0 X 00 Y 00 Z 00
where it is understood that the integrand is evaluated at ϕ.
Note that the full, generally non-local, cohomology groups of Vect(S 1 ) do not
coincide with (6.55) because they contain classes of wedge products of the Gelfand-Fuks
and Godbillon-Vey cocycles. For example c ∧ c is a non-trivial, non-local four-cocycle
on Vect(S 1 ) when c is the Gelfand-Fuks cocycle.
Remark. In this work the Godbillon-Vey cocycle (6.56) will be unimportant. How-
ever, it does play a key role in a specific context, as it was shown in [157] that it
is responsible for the unique non-trivial gauge-invariant deformation of a higher-spin
Chern-Simons action in three dimensions with gauge algebra Vect(S 1 ) Aad∗ Vect(S 1 )∗ .
6.2.2 The Bott-Thurston cocycle

g + (S 1 )
We now turn to the low-degree real cohomology groups of the universal cover Diff
of the group of orientation-preserving diffeomorphisms of the circle. We show in par-
ticular how one can build a non-trivial two-cocycle corresponding to Gelfand-Fuks by
integration, and known as the Bott-Thurston cocycle. The latter will eventually lead
to the definition of the Virasoro group. As before, we abuse notation by denoting the
g + (S 1 ) simply as Diff(S 1 ).
universal cover Diff
Cocyclic recipes
We start by describing a general algorithm for building two-cocycles on a group [54,56].
Let M be an orientable manifold endowed with a volume form µ. For any orientation-
preserving diffeomorphism f : M → M, we define a function T[f −1 ] on M by
−1 ]
f ∗ µ ≡ eT[f µ. (6.57)
This function can be thought of as a modified derivative of f . It appears to have

no standard name in the literature but we will use it repeatedly below in the case
M = S 1 , so from now on we refer to T[f −1 ] as the twisted derivative of f .
Lemma. The map T : Diff+ (M) → C ∞ (M) = f 7→ T[f ] defined by (6.57) is

a C ∞ (M)-valued one-cocycle on Diff+ (M), where the action of diffeomorphisms on
functions is given by
(f · F)(p) = F(f −1 (p)) (6.58)
for f ∈ Diff+ (M), F ∈ C ∞ (M) and p ∈ M.
Proof. We need to show that dT = 0 with the Chevalley-Eilenberg differential (2.31)
and the representation T given by the action (6.58) of diffeomorphisms on functions.2
If f, g are orientation-preserving diffeomorphisms, one readily verifies from the defini-
tion (6.57) that T satisfies the cocycle property (2.32).
Now let us consider another recipe, seemingly unrelated to and just as random as
the previous one. Take two vector spaces V and W acted upon by a group G according
to representations S and T , respectively, and let Ω : V × V → W : (v, v 0 ) 7→ Ω(v, v 0 )
be an antisymmetric bilinear map such that
Ω S[f ]v, S[f ]v 0 = T [f ] Ω(v, v 0 )

(6.59)
for any group element f ∈ G and all v, v 0 ∈ V. Finally, let T : G → V be a V-valued

one-cocycle on G with respect to the representation S.
Lemma. The map

C : G × G → W : (f, g) 7→ C(f, g) ≡ Ω T[f ], T[f g] (6.60)
is a W-valued two-cocycle on G.
Proof. We need to show that dC = 0 for the Chevalley-Eilenberg differential (2.31),
given that W is acted upon by G according to the representation T . Using the fact
that Ω is bilinear and antisymmetric together with property (6.59), one readily verifies
by brute force that this is indeed the case.
The Bott-Thurston cocycle

The two constructions just described can be used to define a non-trivial two-cocycle
on the group Diff(S 1 ). We will first use (6.57) to define a one-cocycle on Diff(S 1 ), then
plug it into (6.60) for a well chosen map Ω to obtain the desired two-cocycle.
We consider the circle S 1 endowed with the flat volume form µ = dϕ. Under a
0
diffeomorphism ϕ 7→ f (ϕ) we have (f ∗ µ)ϕ = d(f (ϕ)) = f 0 (ϕ)dϕ = elog(f (ϕ)) µ, so
(6.57) provides a C ∞ (S 1 )-valued twisted derivative
T[f ](ϕ) ≡ log (f −1 )0 (ϕ) ,

(6.61)
which is a one-cocycle. We use square brackets to denote the argument of T because

the latter is a functional on Diff(S 1 ); then T[f ](ϕ) is the function T[f ] evaluated at ϕ.
2
The fact that the same letter denotes the cocycle T and the representation T is merely a nota-
tional coincidence.
To apply the construction (6.60), we also need to find a bilinear antisymmetric map
Ω : C ∞ (S 1 ) × C ∞ (S 1 ) → R which is invariant under Diff(S 1 ) in the sense that (6.59)
holds when T is the trivial representation while S is the action (6.58) of Diff(S 1 ) on
functions. A natural guess is
Z 2π Z
∞ 1 ∞ 1 0
Ω : C (S ) × C (S ) → R : (F, G) 7→ dϕ F(ϕ)G (ϕ) = FdG , (6.62)
0 S1
which is manifestly antisymmetric (integrate by parts) and reparameterization-inva-

riant (the integrand is analogous to the pq̇ of Hamiltonian actions). Applying the pre-
scription (6.60) with the one-cocycle (6.61), we obtain the following real two-cocycle:
Definition. The Bott-Thurston cocycle on Diff(S 1 ) is [158]

Z
1
C(f, g) ≡ − T[f ] dT[f ◦ g] (6.63)
48π S 1
Z 2π
(6.61) 1 0
dϕ log (f −1 )0 (ϕ) log ((f ◦ g)−1 )0 (ϕ)

= − (6.64)
48π 0
where d denotes the exterior derivative on the circle.
By construction, the Bott-Thurston cocycle satisfies the cocycle identity (2.9),
C(f, gh) + C(g, h) = C(f g, h) + C(f, g) . (6.65)
This will be instrumental in ensuring that C yields a well-defined centrally extended

group. For future applications it is useful to rewrite (6.64) in a slightly simpler way,
which relies on the following result:
Lemma. If T is the twisted derivative (6.61), then for all f, g ∈ Diff(S 1 ) one has
Z Z
T[f ]dT[f ◦ g] = T[(f ◦ g)−1 ]dT[g −1 ] . (6.66)
S1 S1
Proof. We use two key properties: the first is the fact that T is a one-cocycle with
respect to the action of Diff(S 1 ) on C ∞ (S 1 ), so
T[f ◦ g] = T[f ] + T[g] ◦ f −1 , (6.67)
and the second is a property that follows from the definition (6.61) and eq. (6.16):
T[f ] ◦ f = −T[f −1 ]. (6.68)
We then find that (6.63) can be rewritten as

Z Z Z
−1
T[f ] dT[f ◦ g] = − T[f ] dT[g] = T[(f ◦ g)−1 ] dT[g −1 ] ,
S1 S1 S1
which was to be proven.

Thanks to this lemma we can write the Bott-Thurston cocycle (6.64) in a more
convenient way, without f −1 ’s all around the place:
Z Z
1 0 0 1
C(f, g) = − log(f ◦ g) d log(g ) = − T[(f ◦ g)−1 ]dT[g −1 ]. (6.69)
48π S 1 48π S 1
This is the definition that we will be using from now on.
At this stage the Bott-Thurston cocycle seems to be coming out of the blue. How-
ever it turns out that (6.69) is, in fact, a very natural quantity. We will explain this
in greater detail in section 6.3, but for now we simply note the following relation:
Proposition. The Bott-Thurston cocycle (6.69) is the integral of the Gelfand-Fuks

cocycle (6.43) in the sense of formula (2.35):
d2 h i
C etX , esY − C esY , etX

c(X, Y ) = − . (6.70)
dt ds t=0, s=0
In particular, the Bott-Thurston cocycle is non-trivial.

Proof. We consider infinitesimal diffeomorphisms f (ϕ) = ϕ+tX(ϕ)+O(t2 ) and g(ϕ) =
ϕ + sY (ϕ) + O(s2 ). Then log (f 0 ◦ g) = tX 0 and log(g 0 ) = sY 0 so that
Z 2π
tX sY (6.69) 1
C(e , e ) = − dϕ tX 0 (ϕ)sY 00 (ϕ)
48π 0
to first order in t, s. Relation (6.70) follows. It also follows that the Bott-Thurston
cocycle is non-trivial, since the Gelfand-Fuks cocycle is non-trivial.
Remark. The bilinear map (6.62) is a non-trivial two-cocycle on the Abelian Lie
algebra C ∞ (S 1 ) of smooth functions on the circle. It defines a central extension of
C ∞ (S 1 ) that can be interpreted in several ways: either as an infinite-dimensional
Heisenberg algebra, or as a u(1) Kac-Moody algebra. This kind of central extension
occurs for instance in the realm of warped conformal field theories [52, 93].
6.2.3 Primary cohomology of Vect(S 1 )

Here we study some of the cohomology groups of Vect(S 1 ) in spaces of densities (i.e.
primary fields). As in the real-valued case described earlier we consider the cohomology
defined by local cochains, which in the present case take the form
(n ) (n )
c[X1 , ..., Xk ] = C X1 (ϕ), X10 (ϕ), ..., X1 1 (ϕ), ..., Xk (ϕ), ..., Xk k (ϕ) (dϕ)h

for some weight h. The functional C depends on the Xi ’s and finitely many of their
derivatives, all evaluated at the same point ϕ. We denote the corresponding cohomol-
ogy spaces by Hk Vect(S 1 ), Fλ (S 1 ) . In order to avoid technical considerations we
state the results without proof and refer to [56] for details.
Theorem. If the weight h is not a non-negative integer, then
Hk Vect(S 1 ), Fh (S 1 ) = 0 for all k ∈ N.

(6.71)

In particular H2 Vect(S 1 ), Vect(S 1 ) = 0, so there exists no non-trivial deformation
of Vect(S 1 ).
The result (6.71) implies that the non-trivial primary cohomology of Vect(S 1 ) is
localized only on non-negative integers with weights h ∈ N. Here we briefly describe
the non-trivial first cohomology groups (k = 1) for the cases h = 0, 1, 2 that will be
useful below.
The case h = 0 corresponds to one-cochains taking values in the space of functions

on the circle. It turns out that there are exactly two linearly independent, non-trivial
one-cocycles in that case, namely c̃[X](ϕ) = X(ϕ) and
t[X](ϕ) = X 0 (ϕ) . (6.72)
The latter may be recognized as the infinitesimal cocycle corresponding to the twisted
derivative (6.61) by differentiation.
At weight h = 1 we are in the realm of cochains taking values in the space Ω1 (S 1 )

of one-forms; in particular one can show that the corresponding first cohomology is
one-dimensional, generated by the (class of the) one-cocycle
w[X](ϕ) = X 00 (ϕ)dϕ . (6.73)
The notation w is because this cocycle is relevant to certain aspects [52] of warped
conformal symmetry [93].
Finally, when h = 2, cochains take their values in the space F2 (S 1 ) of quadratic

densities on the circle. In particular one can show that the first cohomology space is
one-dimensional, generated by the (class of the) infinitesimal Schwarzian derivative
s[X](ϕ) = X 000 (ϕ)dϕ2 . (6.74)
One can go on and similarly classify all cohomology groups with higher weight h. Since
we will not need these results here, we refrain from displaying them (see e.g. [159,160]).
Instead, we now relate the cocycles (6.72), (6.73) and (6.74) to one-cocycles on Diff(S 1 ).
6.2.4 Primary cohomology of Diff(S 1 )

The complete classification of density-valued cohomology groups of Diff(S 1 ) is beyond
the scope of this presentation, so we refer to [56] for a more detailed discussion. Here
we simply note that the Lie algebra one-cocycles mentioned above can be integrated
to non-trivial group one-cocycles:
• The cocycle (6.72) can be integrated to the twisted derivative (6.61), which we
used to build the Bott-Thurston cocycle. Indeed, for f (ϕ) = ϕ + X(ϕ), formula
(6.61) reduces to T[f ] = − t[X].
• The cocycle (6.73) can be integrated to
W[f ](ϕ) = d log[(f −1 )0 (ϕ)] = d T[f ](ϕ) (6.75)
where d denotes the exterior derivative on the circle. As mentioned above this
“warped derivative” has been used recently [52] to describe certain aspects of
warped conformal field theories.
R Note that in these terms the Bott-Thurston
cocycle (6.63) is C(f, g) ∝ T[f ] ⊗ W[f ◦ g].
The one-cocycle (6.74) can similarly be related to the F2 -valued Schwarzian derivative
on Diff(S 1 ), although the integration is somewhat less trivial than in the two cases
just described. The Schwarzian derivative is crucial for our upcoming considerations,
so the whole next section is devoted to it.
6.3 On the Schwarzian derivative

Definition. Let f ∈ Diff(S 1 ). Then the Schwarzian derivative 3 of f at ϕ is
2
f 000 (ϕ) 3 f 00 (ϕ)

S[f ](ϕ) ≡ 0 − . (6.76)
f (ϕ) 2 f 0 (ϕ)
Many references use the notation {f ; ϕ}, but we will stick to S[f ](ϕ) instead.
In this section we investigate the many properties of the Schwarzian derivative.

We will start by showing that it is (related to) a one-cocycle on Diff(S 1 ) taking its
values in the space F2 (S 1 ) of quadratic densities, and that it corresponds to the Lie
algebra cocycle (6.74) by differentiation. We will then show that it is related to the
Bott-Thurston cocycle by the so-called Souriau construction. We will also describe
the remarkable symmetry properties of the Schwarzian derivative under the PSL(2, R)
subgroup of Diff(S 1 ), and obtain as a by-product the expression of Lorentz transfor-
mations in terms of diffeomorphisms of the circle. (In chapter 9 these transformations
will turn out to be actual Lorentz transformations on the celestial circle.)
6.3.1 The Schwarzian derivative is a cocycle

Here we show that the Schwarzian derivative is the one-cocycle corresponding to (6.74)
by integration. Note that the relation between (6.76) and (6.74) is obvious: upon
taking f (ϕ) = ϕ + X(ϕ) in (6.76), one finds S[f ] = X 000 to first order in . The
non-trivial problem is showing that the Schwarzian derivative is actually a cocycle:
Proposition. The Schwarzian derivative (6.76) defines a one-cocycle
Diff(S 1 ) → F2 (S 1 ) : f 7→ S[f −1 ](ϕ)dϕ2
valued in the space of quadratic densities on the circle.

3
The name refers to H. Schwarz, who first introduced the object (6.76); it is the same Schwarz as
in the Cauchy-Schwarz inequality.
Proof. We start by noting that the definition (6.76) implies
S[f ◦ g] = Ad∗g−1 S[f ] + S[g] = S[g] + (g 0 )2 S[f ] ◦ g , (6.77)
where Ad∗ denotes the coadjoint representation (6.36) of Diff(S 1 ). Upon defining
S̃[f ] ≡ S[f −1 ]dϕ2 , one obtains a map that associates a quadratic density with any
diffeomorphism f , and which satisfies
S̃[f ◦ g] = S̃[f ] + ((f −1 )0 )2 S̃[g] ◦ f −1 = S̃[f ] + Ad∗f S̃[g] (6.78)
by virtue of (6.77). This is precisely the cocycle identity (2.32).
The Souriau construction

We now study the relation between the Schwarzian derivative and the Bott-Thurston
cocycle, which follows from the so-called Souriau construction.
Definition. Let G be a Lie group with Lie algebra g, C : G × G → R a real two-

cocycle on G. Then the Souriau cocycle associated with C is the map G 7→ g∗ : f 7→
S[f −1 ] defined by
d h i 1
C(f, etX ) + C(f etX , f −1 ) ≡− hS[f ], Xi (6.79)
dt t=0 12
for any f ∈ G and any adjoint vector X ∈ g. (The normalization is chosen so that S
eventually coincides with the Schwarzian derivative.)
Proposition. The Souriau cocycle is a one-cocycle on G valued in the space of

coadjoint vectors.
Proof. Since C is a real two-cocycle, it is clear that the left-hand side of (6.79) defines
a real linear function of X ∈ g, that is, a coadjoint vector. The latter only depends
on f so we can certainly write it as S[f ], which defines the map S. The problem is to
show that the map f 7→ S[f −1 ] is actually a one-cocycle. For this we let X ∈ g, pick
two group elements f, g ∈ G, and write
−1 d h
(6.79) −1 −1 tX −1 −1 tX
i
S[(f g) ], X = C(g f , e ) + C(g f e , f g) . (6.80)
dt t=0
On the other hand, if Ad∗ denotes the coadjoint representation of G, we have
Ad∗f S[g −1 ] + S[f −1 ], X = (6.81)

(6.79) d
h i
= C(g −1 , etAdf −1 X ) + C(g −1 etAdf −1 X , g) + C(f −1 , etX ) + C(f −1 etX , f ) .
dt t=0
Using the cocycle identity (6.65) together with property (5.7), one can then show by
brute force that (6.81) coincides with the right-hand side of (6.80).
In the case of the group Diff(S 1 ), the Souriau construction yields the Schwarzian
derivative from the Bott-Thurston cocycle. Let us check this explicitly: taking g(ϕ) =
ϕ + tX(ϕ) in (6.69), one finds
Z 2π " 00 2 #
000
t f f
C(f, etX ) = − dϕ − X(ϕ),
48π 0 f0 f0
Z 2π " 00 2 #
000
t f f
C(f ◦ etX , f −1 ) = (t-independent) − dϕ − 2 X(ϕ)
48π 0 f0 f0
to first order in t. It then follows that relation (6.79) holds when S is the Schwarzian
derivative (6.76) and h·, ·i is the pairing (6.34) of Vect(S 1 ) with its dual. Note that by
taking f (ϕ) = ϕ + sY (ϕ) with small s, the Schwarzian derivative reduces to S[f ] =
sY 000 . Upon differentiating with respect to s in the right-hand side of (6.79), we recover
precisely the Gelfand-Fuks cocycle (6.43).
Virasoro universality
At this point the cohomological constructions of the previous pages are starting to
fit in a global picture of Virasoro cohomology: eq. (6.70) relates the Bott-Thurston
cocycle to the Gelfand-Fuks cocycle (6.43), while (6.79) relates it to the Schwarzian
derivative, which in turn is the integral of the infinitesimal cocycle (6.74). In addition
the integral of the latter with a vector field on the circle reproduces the Gelfand-Fuks
cocycle. The common feature of all these expressions is the occurrence of third deriva-
tives such as f 000 or X 000 , which will indeed play a key role in the sequel (and give rise
to the term m3 in (6.44)).
In this sense, all these cocycles are really one and the same quantity, albeit ex-
pressed in very different ways. Depending on one’s viewpoint, one may decide that
the most fundamental quantity is the Gelfand-Fuks cocycle, or the Schwarzian deriva-
tive, or Bott-Thurston. Our point of view will be that the Bott-Thurston cocycle is
the most fundamental of them all, since it yields the other ones by differentiation:
Bott-Thurston
 C

SouriauSouriau

y
Schwarzianderivative S

differentiatedif f erentiate

y
infinitesimalSchwarzian s

pair with Vect(S 1 )pair with Vect(S 1 )

y
Gelfand-Fuks c
6.3.2 Projective invariance of the Schwarzian

There exists a deep relation between the circle, the projective line and the Schwarzian
derivative [54], which in turn leads to powerful symmetry properties under the group
SL(2, R)/Z2 = PSL(2, R). Our goal here is to explore this relation. Accordingly
we start with a short detour through one-dimensional projective geometry, before
recovering the Schwarzian derivative as a quantity that measures the extent to which
diffeomorphisms deform the projective structure. Along the way we will encounter the
expression of Lorentz transformations in terms of diffeomorphisms of the circle.
The projective line

Consider the plane R2 and define the projective line RP 1 to be the space of its one-
dimensional subspaces. Equivalently RP 1 is the space of straight lines in R2 going
through the origin, i.e. a quotient of R2 \{(0, 0)} by the equivalence relation
(x, y) ∼ (x0 , y 0 ) if ∃ λ ∈ R∗ such that (x, y) = λ(x0 , y 0 ).
Denoting by [(x, y)] the equivalence class of (x, y) ∈ R2 , the projective line is thus
RP 1 = [(x, y)] (x, y) ∈ R2 \{(0, 0)} .

(6.82)
In topological terms the projective line is a circle centred at the origin in R2 with
antipodal points identified. This is to say that RP 1 ∼ = S 1 /Z2 , where Z2 acts on S 1 by
rotations. Since any group Zn acting on the circle by rotations of 2π/n is such that
S 1 /Zn ∼
= S 1 , the projective line is actually diffeomorphic to a circle:
RP 1 ∼
= S 1. (6.83)
As a result, all considerations concerning the group of diffeomorphisms of S 1 can be

recast in terms of projective geometry, and vice-versa.
The diffeomorphism (6.83) can be made explicit in terms of well chosen coordinates.
Indeed, in terms of (6.82), the projective line is a union RP 1 = {[(ζ, 1)]|ζ ∈ R} ∪
{[(1, 0)]}, so the projective coordinate
ζ ≡ x/y (6.84)
is a local coordinate on RP 1 that misses only one point, namely the class of (1, 0). In
this sense the projective line is a real line R with an extra “point at infinity”.
This is exactly the same situation as with the stereographic coordinate on a circle.
For later convenience we define this coordinate in terms of an angular coordinate ϕ on
the circle by
eiϕ + 1
ζ = −cot(ϕ/2) = iϕ (6.85)
ie − i
(see fig. 6.3). The diffeomorphism (6.83) is then obtained by identifying this stereo-
graphic coordinate with the projective coordinate (6.84).
Figure 6.3: The stereographic coordinate (6.85) is obtained by projecting the points of
a unit circle in the (X, Y ) plane on the Y axis, along a straight line that goes through
the “east pole” (1, 0). Upon writing X = cos ϕ and Y = sin ϕ and declaring that
the projective coordinate ζ is minus the Y coordinate of the projection, one obtains
(6.85). The minus sign is included so as to preserve the orientation of the circle in the
sense that dζ/dϕ > 0.
Projective transformations
The projective line inherits a symmetry from the linear action of GL(2, R) on R2 .
Explicitly, an invertible matrix
a b
(6.86)
c d
acts on the coordinate (6.84) as a projective transformation
aζ + b
ζ 7→ . (6.87)
cζ + d
Any such transformation is independent of the determinant of (6.86), which we can

therefore set to one without loss of generality. Furthermore the overall sign of the
matrix is irrelevant, so the transformations (6.87) span a projective group
PGL(2, R) ≡ GL(2, R)/R∗ ∼

= SL(2, R)/Z2 ≡ PSL(2, R).
According to (4.83) this is the connected Lorentz group in three dimensions.
Upon identifying the projective coordinate (6.84) with the stereographic coordinate
(6.85), the transformation (6.87) can be reformulated in terms of the angular coordi-
nate ϕ. Using (6.87) and the inverse of (6.85), we find that projective transformations
act on eiϕ according to
Aeiϕ + B
eiϕ 7→ eif (ϕ) = (6.88)
B̄eiϕ + Ā
where the complex coefficients
1 1
A = (a + ib − ic + d), B = (a − ib − ic − d) (6.89)
2 2
are such that |A|2 − |B|2 = 1. The family of transformations (6.88) spans a subgroup
PSL(2, R) of Diff(S 1 ) that we already anticipated in (6.3). In section 9.1 we shall
interpret that subgroup as the Lorentz group acting on null infinity. For infinitesimal
parameters A = 1 + i and B = ε (with ∈ R and ε ∈ C), formula (6.88) becomes an
infinitesimal diffeomorphism
f (ϕ) = ϕ + 2 + 2 Re(ε) cos ϕ − 2 Re(ε) sin(ϕ).
This is an sl(2, R) transformation generated by the vector fields `0 , `1 and `−1 men-
tioned below (6.24). Conversely, any transformation (6.88) belongs to the flow of an
sl(2, R) vector field.
Remark. The relation between S 1 and RP 1 discussed here has a complex analogue
CP 1 ∼= S 2 , where CP 1 is the complex projective line. In this generalization the
projective coordinate (6.84) becomes a complex coordinate z and coincides with the
stereographic coordinate (1.4) of S 2 . The projective transformations of CP 1 then are
Möbius transformations (1.6), i.e. Lorentz transformations in four dimensions.
Cross ratios and the Schwarzian derivative

Given the projective line RP 1 , one may look for projective invariants, i.e. quantities
that are left invariant by the transformations (6.87). For example, consider four points
on RP 1 whose projective coordinates are ζ1 , ζ2 , ζ3 , ζ4 and define their cross ratio
(ζ1 − ζ3 )(ζ2 − ζ4 )
[ζ1 , ζ2 , ζ3 , ζ4 ] ≡ .
(ζ1 − ζ2 )(ζ3 − ζ4 )
This number is a projective invariant, as one can verify by direct computation. Now
take a diffeomorphism f : S 1 → S 1 , where we think of S 1 as a projective line (6.83).
This diffeomorphism can be written in terms of the projective coordinate (6.84), giving
rise to a map ζ 7→ f(ζ). The explicit correspondence between f and f follows from
(6.85) and reads
f −cot(ϕ/2) = −cot f (ϕ)/2 . (6.90)
In general, f is not a projective transformation (6.87) and therefore spoils the projective
structure of S 1 . This can be measured by taking a point with coordinate ζ on RP 1
together with three other nearby points, then evaluating the change in their cross-ratio
under the action of f. Let therefore ζ1 = ζ + , ζ2 = ζ + 2 and ζ3 = ζ + 3 be three
points close to ζ. These points move under the action of f, and one can show that
f(ζ), f(ζ1 ), f(ζ2 ), f(ζ3 ) = [ζ, ζ1 , ζ2 , ζ3 ] − 22 S[f](ζ) + O(||3 )

where S is the Schwarzian derivative (6.76). Thus we have recovered the Schwarzian
derivative as a measuring device that tells us how much the diffeomorphism f spoils
the projective structure of RP 1 . From this one concludes:
Proposition. Let f, g be diffeomorphisms of the projective line. Then
S[f ◦ g](ζ) = S[g](ζ) (6.91)

if and only if f is a projective transformation (6.87). In particular, S[f](ζ) = 0 if and

only if f(ζ) is a projective transformation of the form (6.87).
In technical terms, a one-cocycle S on a group G with a subgroup H is said to be H-

relative if S[h] = 0 for all h ∈ H. Thus (6.91) says that the Schwarzian derivative is a
PSL(2, R)-relative cocycle. In conformal field theory this corresponds to the statement
that the Schwarzian derivative is blind to Möbius transformations (1.6).
Schwarzians on the circle

All the above considerations can be reformulated in terms of the angular coordinate ϕ
using the correspondence (6.85). Here we work out this rewriting for projective trans-
formations (6.88). To begin, note that (6.88) precisely takes the form of a projective
transformation (6.87) in terms of the coordinate eiϕ . Accordingly,
S eif (ϕ) (eiϕ ) = 0.

(6.92)
In order to go from (6.92) to S[f ](ϕ) we use the cocycle identity (6.77) repeatedly.
First we write
(6.77) 2
S[f ](ϕ) = S[log(eif (ϕ) )](ϕ) = S[eif (ϕ) ](ϕ) + if 0 (ϕ)eif (ϕ) S[log](eif (ϕ) ) , (6.93)
(6.76) 1
where S[log](x) = 2x2
. The first term on the far right-hand side of (6.93) involves
(6.77) 2 (6.92) (6.76) 1
S[eif (ϕ) ](ϕ) = S[eiϕ ](ϕ) + (eiϕ )0 S[eif (ϕ) ](eiϕ ) = S[eiϕ ](ϕ) = ,
2
which finally gives
1 2
1 − f 0 (ϕ)
S[f ](ϕ) = (6.94)
2
when f (ϕ) is given by (6.88). We will put this formula to use in the next chapter.
Note that these observations can be generalized to infinitely many other families of
diffeomorphisms of the circle. Indeed, pick a positive integer n ∈ N∗ and take formula
(6.88) with ϕ replaced by nϕ and f (ϕ) replaced by nf (ϕ):
Aeinϕ + B
einf (ϕ) = , |A|2 − |B|2 = 1. (6.95)
B̄einϕ + Ā
This defines a diffeomorphism of the circle, and the family of such diffeomorphisms
also spans a subgroup of Diff(S 1 ) which is locally isomorphic to SL(2, R). The dif-
ference with respect to the case n = 1 discussed above is that the corresponding Lie
algebra sl(2, R) is generated by `0 , `n and `−n , and that the actual group spanned by
such transformations is an n-fold cover of PSL(2, R); we shall denote this cover by
SL(n) (2, R)/Z2 ≡ PSL(n) (2, R). One can also verify that the Schwarzian derivative of
the diffeomorphism f defined by (6.95) satisfies
n2
1 − (f 0 (ϕ))2 ,

S[f ](ϕ) = (6.96)
2
generalizing the case n = 1 of (6.94).
Remark. One can show that the restriction of the Bott-Thurston cocycle (6.69)
to the PSL(2, R) subgroup (6.88) coincides with the unique non-trivial two-cocycle on
PSL(2, R). The latter acts on the hyperbolic plane H2 by isometries of the form (6.87),
where ζ ∈ C has positive imaginary part, and the two-cocycle associates with two
such transformations f, g the area of the triangle with corners i, f −1 (i) and g−1 ◦ f −1 (i).
See [56] for details.
6.4 The Virasoro group

We are now in position to describe the central extension of Diff(S 1 ). This discus-
sion is crucial for our purposes, as all symmetry groups of the later chapters will be
variations on the basic Virasoro pattern. As announced above, our viewpoint is that
the fundamental Virasoro structure is that of the group, from which the rest follows.
Accordingly we start this section by reviewing general properties of centrally extended
groups, which we then apply to the Virasoro group whose adjoint and coadjoint rep-
resentations follow by differentiation. We also define the Virasoro algebra and end by
displaying the Kirillov-Kostant Poisson bracket on its dual.
6.4.1 Centrally extended groups revisited

Let Gb be a central extension of some Lie group G, with group operation (2.11) in terms
of some two-cocycle C. Here we work out its adjoint and coadjoint representations.
Adjoint representation
Since the group G b consists of pairs (f, λ) where f ∈ G and λ ∈ R, its Lie algebra b
g
consists of pairs (X, λ) where X ∈ g. The adjoint representation of G b then follows
from (5.6): for X ∈ g, f ∈ G and λ, µ ∈ R we find
c (f,µ) (X, λ) = d f ◦ etX ◦ f −1 , tλ + C(f, etX ) + C(f ◦ etX , f −1 )

h i
Ad (6.97)
dt t=0
where the hat in Ad

c stresses that this is the adjoint representation of a centrally ex-
tended group, as opposed to that of G. Note that µ acts trivially, so we can lighten
the notation by writing Ad c f instead of Adc (f,µ) . This follows from the fact that G
b is
a central extension of G so that “central elements” (i.e. the real numbers that enter
in the second slot of (f, µ)) act trivially on everything, which is a general property of
centrally extended groups.
It then remains to compute the two entries on the right-hand side (6.97). The
first entry yields the adjoint representation of G, while the second is precisely the
expression (6.79) defining the Souriau cocycle S associated with C. We conclude that
the adjoint representation of Gb reads
c f (X, λ) = Adf X, λ − 1 hS[f ], Xi

Ad (6.98)
12
where the “Ad” on the right-hand side is the adjoint representation of G.
Centrally extended algebra

From the adjoint representation of a group one can read off the Lie brackets (5.8) of its
algebra. Let therefore (X, λ) and (Y, µ) belong to the centrally extended Lie algebra
g. Using (6.98) we find
b

d 1 tX
(X, λ), (Y, µ) = AdetX Y, µ − S[e ], Y . (6.99)
dt 12 t=0
The first entry on the right-hand side is the same as expression (5.8) in G; accordingly
it boils down to the standard Lie bracket of g, which we denote as [X, Y ]. The second
entry involves the differential of the Souriau cocycle,
d
s[X] ≡ S[etX ] t=0
, (6.100)
dt
paired with Y ∈ g. We therefore define a two-cocycle c on g by
1
c(X, Y ) ≡ − hs[X], Y i (6.101)
12
and the bracket of bg takes the centrally extended form (2.26):
h i 1
(X, λ), (Y, ρ) = [X, Y ], c(X, Y ) = [X, Y ], − hs[X], Y i . (6.102)
12
The fact that (6.101) is indeed a two-cocycle is inherited from the Souriau cocycle.
Note that the central terms λ, µ commute with everything, as they should. In terms
of Lie algebra generators the bracket (6.102) takes the general form (2.27).
The Lie algebra bg is spanned by pairs (X, λ), so its dual consists of pairs (p, c) where
p belongs to g∗ while c ∈ R is a real number, paired with adjoint vectors according to
(p, c), (X, λ) = hp, Xi + cλ (6.103)
where the pairing h·, ·i on the right-hand side is that of g∗ with g. The number c is
known as a central charge. The coadjoint transformation law of (p, c) follows from
the definition (5.10). In particular, since central elements act trivially in the adjoint
c∗ c∗
representation (6.97), we can safely write Ad (f,λ) ≡ Adf for any (f, λ) ∈ G, where the
b
∗
hat on top of Ad indicates that we refer to a representation of the centrally extended
group. If then (X, λ) ∈ b g and (p, c) ∈ bg∗ , one obtains
D E
c ∗ (p, c), (X, λ) (6.98)
D 1 −1
E
Ad f = (p, c), Ad f −1 , λ − S[f ], X
12
(6.103) c
= hp, Adf −1 Xi + cλ − S[f −1 ], X
12
where the pairing h·, ·i on the right-hand side of the last equation is the centreless
pairing of g∗ with g. In particular the first term is simply the coadjoint representation
of G. Removing the dependence on X, we conclude that
∗

∗ c −1

Adf (p, c) = Adf p − S[f ], c .
c (6.104)
12
Note that the central charge c is left invariant by the coadjoint representation, as it
should. Crucially, it also appears in the first entry and thus affects the transformation
law of p. In abstract terms, formula (6.104) is the affine G-module (2.33) associated
with the Souriau cocycle.
The coadjoint action can be differentiated, which yields a representation (5.11) of

g. Using (6.102) with the two-cocycle (6.101) one finds
the Lie algebra b
c∗ (p, c) = ad∗ p + c s[X], 0

adX X (6.105)
12
where the ad∗ on the right-hand side is the coadjoint representation of g while s is the
infinitesimal Souriau cocycle (6.100). In the remainder of this section we apply these
considerations to the Virasoro group.
6.4.2 Virasoro group

Definition. The Virasoro group is the universal central extension of Diff(S 1 ). It
is diffeomorphic to the product Diff(S 1 ) × R and its elements are pairs (f, λ) where
f ∈ Diff(S 1 ) and λ ∈ R, with a group operation (2.11) where C is the Bott-Thurston
cocycle (6.69). Explicitly:

(f, λ) · (g, µ) = f ◦ g, λ + µ + C(f, g) . (6.106)
d 1 ).
We shall denote the Virasoro group by Diff(S
As in the previous sections we abuse notation and terminology by simply calling

“Virasoro group” what is really the universal cover of the maximal connected subgroup
g + (S 1 ), while we denote it by
of the Virasoro group. It should in fact be written as Diff
d
d 1 ) to reduce clutter.
Diff(S
6.4.3 Adjoint representation and Virasoro algebra

As a vector space, the Lie algebra of the Virasoro group is equivalent to the direct
sum Vect(S 1 ) ⊕ R. In particular, Virasoro adjoint vectors are pairs (X, λ) where
X = X(ϕ)∂ϕ is a vector field on the circle and λ is a real number. The adjoint rep-
resentation of the Virasoro group follows from the group operation (6.106) and the
definition (5.6). Thus the adjoint representation takes the form (6.98), where the ad-
joint representation of Diff(S 1 ) is the transformation law (6.18) of vector fields, while
S is the Schwarzian derivative (6.76). This result will be instrumental in our definition
of the centrally extended BMS3 group in section 9.2.
The adjoint representation of a group yields the Lie brackets (5.8) of its algebra.
In the present case this definition leads to an awkward sign (6.21), which we absorb
by declaring that the Lie bracket of the Virasoro algebra is defined by
d hc i
(X, λ), (Y, µ) ≡ − AdetX (Y, µ) . (6.107)
dt t=0
With this definition formula (6.99) holds up to an overall minus sign on the right-hand
side. Using then the infinitesimal Schwarzian derivative (6.74), the pairing (6.34)
allows us to recognize the Gelfand-Fuks cocycle (6.43) in hs[X], Y i. Thus the Lie
bracket of the algebra of the Virasoro group takes the form (6.102), or explicitly

(X, λ), (Y, µ) = [X, Y ], c(X, Y ) (6.108)
where [X, Y ] is the usual Lie bracket of vector fields.
d 1 ) = Vect(S 1 ) ⊕ R en-
Definition. The Virasoro algebra is the Lie algebra Vect(S
dowed with the Lie bracket (6.108). It is the universal central extension of Vect(S 1 ).4
In the physics literature it is customary to rewrite the Virasoro algebra in a form

analogous to (6.24). Let us therefore define the basis elements
Lm ≡ (`m , 0), Z ≡ (0, 1) (6.109)
where the `m ’s are given by (6.23). The bracket (6.108) then yields [Z, Z] = [Z, Lm ] =
0, as well as
(6.108)
i[Lm , Ln ] = i[(`m , 0), (`n , 0)] = (i[`m , `n ], ic(`m , `n )) .
Using the Witt algebra (6.24) and eq. (6.44), we can rewrite this as
Z 3
i[Lm , Ln ] = (m − n)Lm+n + m δm+n,0 , (6.110)
12
which is indeed the standard expression of the Virasoro algebra [152–154]. In this form
it can be seen as a central extension of the Witt algebra (6.24), with a central term
involving the celebrated m3 δm+n,0 . As mentioned below (6.44), the latter is a remnant
of the third derivative of Y in the Gelfand-Fuks cocycle (6.43), while the δm+n,0 follows
from the integration over the circle and reflects the fact that the cocycle is invariant
under rotations.
Remark. The generator Z of eq. (6.109) should rightfully be called the “central
charge” of the Virasoro algebra, since it is a Lie algebra element that commutes with
everything. However, in keeping with the standard physics terminology, we will also
use the word “central charge” to refer to the dual of Z, which is just a real number c
(see the coadjoint representation below). This ambiguous terminology should not lead
to any confusion.
6.4.4 Coadjoint representation

Coadjoint vectors
Virasoro adjoint vectors are pairs (X, λ) where X is a vector field and λ a real number.
Accordingly the smooth dual of the Virasoro algebra consists of pairs (p, c) where
4
Universality follows from the fact that the first real cohomology of Vect(S 1 ) vanishes.
p = p(ϕ)dϕ2 is a quadratic density and c ∈ R is a real number, paired with adjoint

vectors according to the centrally extended generalization (6.103) of (6.34):
Z 2π
1
(p, c), (X, λ) ≡ dϕ p(ϕ)X(ϕ) + cλ. (6.111)
2π 0
d 1 )∗ .
We refer to pairs (p, c) as Virasoro coadjoint vectors; they span the space Vect(S
It is worth mentioning that coadjoint vectors are crucial physical quantities in all
theories enjoying Diff(S 1 ) symmetry, and in particular all conformal field theories in
two dimensions. Indeed the function p(ϕ) is nothing but (the chiral component of) a
CFT stress tensor, while c is a CFT central charge. Expression (6.111) then coincides
(up to central terms) with the Noether charge associated with a symmetry generator
X(ϕ)∂ϕ , seen as an infinitesimal conformal transformation. More precisely, in a CFT
on a Lorentzian cylinder, the coordinate ϕ would be replaced by one of the two light-
cone coordinates x± and p(ϕ) would become p(x+ ) or p̄(x− ). This is consistent with
the interpretation of coadjoint vectors as Noether currents, thanks to the momentum
maps of section 5.1.
Remark. Our notation is somewhat non-standard in that we denote by p(ϕ) what

would normally be written as T (ϕ), where T stands for the stress tensor. This choice
has to do with our motivations: we shall see in chapter 9 that Virasoro coadjoint
vectors play the role of supermomentum vectors for the BMS3 group. They will be
infinite-dimensional generalizations of the Poincaré momenta pµ , so the notation p(ϕ)
is introduced here to suggest thinking of coadjoint vectors as quantities related to
energy and momentum. In fact this interpretation also holds in CFT, since a stress
tensor is nothing but an energy-momentum density.
The transformation law of Virasoro coadjoint vectors follows from the definition (5.10).
In particular, since central elements act trivially in the adjoint representation (6.97),
c∗ c∗
we may write Ad (f,λ) ≡ Adf for any (f, λ) belonging to the Virasoro group. If then we
d 1 ) be an adjoint vector and (p, c) ∈ Vect(S
let (X, λ) ∈ Vect(S d 1 )∗ be a coadjoint one,
formula (6.104) still holds upon letting S be the Schwarzian derivative. The central
charge c is left invariant by the coadjoint representation, as it should, but it also affects
the transformation law of p(ϕ). Accordingly, from now on we often write the coadjoint
representation of the Virasoro group without including a second slot for the central
charge, since the latter is invariant. With this simplified notation formula (6.104) boils
down to
c ∗ p = Ad∗ p − c S[f −1 ].
Ad (6.112)
f f
12
For future reference it will be useful to rewrite this in detail, in terms of functions on
the circle. Evaluating both sides of the equation at a point ϕ on the circle, we obtain
c ∗ p (ϕ) = (f −1 )0 (ϕ) 2 p(f −1 (ϕ)) − c S[f −1 ](ϕ)

Adf (6.113)
12
by virtue of the centreless coadjoint action (6.36). The formulas are much simpler if
we evaluate eq. (6.112) at f (ϕ); using the cocycle identity (6.77) we find
c ∗ p (f (ϕ)) =
1 h c i
Adf p(ϕ) + S[f ](ϕ) . (6.114)
(f 0 (ϕ))2 12
This is a transparent expression of the coadjoint representation of the Virasoro group,

with S[f ](ϕ) given by (6.76). It is the most important equation of this chapter. We
will sometimes refer to the two terms on the right-hand side as the “homogeneous
term” and the “central” or “inhomogeneous term”, respectively. The formula can also
be recognized as the transformation law of a CFT stress tensor p(ϕ) with a central
charge c; in that context p(ϕ) is said to be a quasi-primary field with weight two. In
the next chapter we will classify the orbits of this action, which in chapter 9 will turn
out to be supermomentum orbits labelling BMS particles in three dimensions.
The differential (6.105) of formula (6.114) is a representation of the Virasoro alge-

bra. Taking an infinitesimal diffeomorphism f (ϕ) = ϕ + X(ϕ), we treat the homo-
geneous term Ad∗f p as in (6.30) and find Ad∗f p = p − (Xp0 + 2X 0 p) to first order in
(both sides of the equation are evaluated at the same point). For the Schwarzian
derivative we use S[f −1 ] = −X 000 . Defining
c ∗ p)(ϕ) − p(ϕ)
(Ad
c∗ p(ϕ) ≡ − f
adX

as in (6.31), we end up with the coadjoint representation of the Virasoro algebra:
c∗ p = Xp0 + 2X 0 p − c X 000 ,
ad (6.115)
X
12
where both sides are evaluated at the same point.5 This is the Virasoro version of
eq. (6.105). In the homogeneous term we recognize the primary transformation law
(6.32), while the central term involves the infinitesimal Schwarzian (6.74).
6.4.5 Kirillov-Kostant bracket

In order to make contact with physics, let us describe the Kirillov-Kostant Poisson
bracket (5.23) for the Virasoro group. In that case the bracket eats functions on
d 1 )∗ , i.e. functionals F[ p(ϕ), c]. In practice, since any quadratic density p(ϕ)dϕ2
Vect(S
can be Fourier-expanded as X
p(ϕ) = pm e−imϕ , (6.116)
m∈Z
the Fourier modes pm = p∗−m

define global coordinates on Vect(S 1 )∗ . Any functional
F[ p(ϕ), c] can then be seen as a function of the variables pm and c, so it suffices to
know the Poisson brackets of these variables in order to find the Poisson brackets of
5 c∗ (p, c) ≡ −(p, c) ◦ ad
Eq. (6.115) can also be written as ad cX , where the infinitesimal adjoint rep-
X
cX (Y, µ) = [X, Y ], c(X, Y )
resentation of the Virasoro algebra is defined with a sign such that that ad
coincides with the bracket (6.108).
d 1 )∗ .
functions on Vect(S
Now recall the basis (6.109) of the Virasoro algebra and let {(Lm )∗ , Z ∗ } denote
the corresponding dual basis, such that hL∗m , Ln i = δmn and hZ ∗ , Zi = 1. Using the
pairing (6.111) we find that, as coadjoint vectors,
(Lm )∗ = (`m )∗ , 0 = e−imϕ dϕ2 , 0 , Z ∗ = (0, 1).

(6.117)
Thus, when writing a quadratic density as a Fourier series (6.116), the parameters
d 1 )∗ defined with respect to the basis (6.117):
pm , c are actually coordinates on Vect(S
X
p(ϕ)dϕ2 , c = pm L∗m + c Z ∗ .
m∈Z
Accordingly, eq. (5.28) implies that the Kirillov-Kostant Poisson bracket of these co-
ordinates reproduces the Lie brackets (6.110):
c 3
i{pm , pn } = (m − n)pm+n + m δm+n,0 , (6.118)
12
while all Poisson brackets involving the central charge c vanish. The key difference
between (6.118) and (6.110) is that the latter is an abstract Lie bracket, while the
former is its phase space realization.
The bracket (6.118) is well-known to physicists. Indeed, the standard way to intro-
duce the Virasoro algebra in CFT textbooks is to expand the stress tensor in modes as
in (6.116), and then compute their Poisson brackets. Upon quantization, the operator
[
i{·, ·} coincides with the commutator [·, ·] and the resulting quantum commutators
span a Virasoro algebra (6.110)-(6.118), generally with a non-zero central charge c.
d 1 )∗ that maps
Note that each coordinate pm can be seen as the function on Vect(S
(p, c) on h(p, c), Lm i. As mentioned below (6.111), the object h(p, c), Lm i may be
thought of as the Noether charge associated with the symmetry generator Lm , so the
Poisson bracket (6.118) can be interpreted as a Poisson bracket of Noether charges.
We shall see in chapters 8 and 9 that the Poisson brackets of surface charges in three-
dimensional gravity coincide with the Kirillov-Kostant brackets on the dual of suitable
asymptotic symmetry algebras (albeit with definite values of the central charges).
Chapter 7
Virasoro coadjoint orbits
In this chapter we classify the coadjoint orbits of the Virasoro group. Aside from
their usefulness in the study of conformal symmetry, they are crucial for our purposes
because they will turn out to coincide with the supermomentum orbits that classify
BMS3 particles. As we shall see, despite being infinite-dimensional, these orbits be-
have very much like the finite-dimensional coadjoint orbits of SL(2, R).
The plan is as follows. In section 7.1 we describe the problem and explain how
it can be addressed in terms of two invariant quantities, namely the conjugacy class
of a certain monodromy matrix and the winding number of a related curve taking its
values in a circle. We then use this approach in section 7.2 to display explicit orbit
representatives. Finally, section 7.3 is devoted to a discussion of energy positivity in
the Virasoro context.
Coadjoint orbits of the Virasoro group were first classified in [161, 162] and are
described in many later papers [121,163–165] and textbooks [55,56]. The presentation
of this chapter follows [166].
7.1 Coadjoint orbits of the Virasoro group

In this section we explain the methods used to classify coadjoint orbits of the Virasoro
group. We start by describing the simple (but pathological) classification of orbits
at zero central charge, before discussing certain basic aspects of centrally extended
orbits. We then turn to the correspondence between Virasoro coadjoint vectors and
Hill’s operators, which yields two invariant quantities that can be used to classify the
orbits. These invariants are (i) the conjugacy class of a monodromy matrix and (ii)
the winding number of a curve on the real line whose target space is a circle.
7.1.1 Centreless coadjoint orbits

We start our investigation with a problem that is much simpler than the full classi-
d 1 ), namely the classification
fication of coadjoint orbits of the Virasoro group Diff(S
of orbits at vanishing central charge, c = 0. Those are orbits of the centreless group
179
Chapter 7. Virasoro orbits 180
Diff(S 1 ) = Diff
g + (S 1 ), whose coadjoint representation is given by eq. (6.36).

Let us pick a coadjoint vector p(ϕ)dϕ2 , c = 0 and denote its coadjoint orbit by
W(p,0) . For now, suppose for simplicity that p(ϕ) is strictly positive for all ϕ ∈ [0, 2π].
One can then verify that the integral
√ 1
Z 2π p
M≡ dϕ p(ϕ) (7.1)
2π 0
is invariant under the action (6.36) of Diff(S 1 ) on p. This actually follows from the
fact that p is a quadratic density, so its square root is a one-form and can be integrated
on the circle in a Diff(S 1 )-invariant way. With this notation the diffeomorphism
Z ϕ r
p(φ)
f (ϕ) ≡ dφ (7.2)
0 M
maps p(ϕ) on the constant coadjoint vector f · p = M . Thus any strictly positive
coadjoint vector p belongs to the orbit of a constant M given by (7.1), which is the
“mass” associated with p(ϕ). The stabilizer of p(ϕ) = M is the set of diffeomorphisms
f such that M = M/(f 0 (ϕ))2 . Since we set f 0 > 0 to preserve orientation, the only
solution is f 0 = 1 and the stabilizer of p = M is the group U(1) of rigid rotations
f (ϕ) = ϕ + θ. Thus the orbit of any strictly positive coadjoint vector is diffeomorphic
to the quotient space Diff(S 1 )/S 1 . The same analysis applies, up to signs, to strictly
negative coadjoint vectors. Note that Diff(S 1 )/S 1 has codimension one in Diff(S 1 ).
Of course, coadjoint vectors may well vanish at certain points of the circle, and
in particular they can change sign; the previous analysis must then be modified. For
example, suppose p(ϕ) is everywhere non-negative, but vanishes at the point ϕ = 0.
We will say that p(ϕ) has a “double zero” at ϕ = 0, since both p(ϕ) and p0 (ϕ) vanish
there. Then the integral (7.1) is still invariant on the orbit of p, but it is no longer
true that p(ϕ) can be mapped on a constant because the corresponding would-be
diffeomorphism (7.2) is degenerate: its derivative vanishes at ϕ = 0. We conclude
that the orbit of p is now specified by two invariant statements: first, the fact that
√
the integral of p takes the value (7.1), and second, the fact that p(ϕ) has one double
zero. More generally, if p(ϕ) is everywhere non-negative but has N double zeros at
the points ϕ = ϕ1 , ..., ϕN , then the N integrals
Z ϕi+1 p
dϕ p(ϕ)
ϕi
(where ϕN +1 ≡ ϕ1 ) are invariants specifying the orbit of p. The orbit is labelled by

the values of these integrals together with their ordering (which is Diff+ (S 1 )-invariant)
and the statement that all its elements have exactly N double zeros. In particular, the
orbit has codimension N in Diff(S 1 ) since it is specified by N continuous parameters.
A similar treatment can be applied to coadjoint vectors that change sign on the
circle, i.e. functions p(ϕ) having simple zeros (where p0 does not vanish). The number
of such points is always even since p(ϕ) p is 2π-periodic, so let us suppose p(ϕ) has
2N 0 simple zeros. Then the integral of |p(ϕ)| between any two consecutive zeros is
Diff(S 1 )-invariant as before, so the orbit of p is specified by the 2N 0 values of these

integrals, by their ordering and by the sign of p(ϕ) on one of the intervals where it does
not vanish. From this we also deduce the general classification of orbits for quadratic
densities with a finite number of zeros: if p(ϕ) has N double zeros and 2N 0 simple
zeros, its orbit is specified by the values of N + 2N 0 integral invariants of the form
Z ϕi+1 p
dϕ |p(ϕ)| (7.3)
ϕi
(where ϕi and ϕi+1 are any two consecutive zeros), together with the ordering of these
invariants, the specification of whether the points ϕi and ϕi+1 are simple or double
zeros, and the sign of p on a given interval, say [ϕ1 , ϕ2 ]. The orbit of p then has
codimension N + 2N 0 in Diff(S 1 ); in particular there exist orbits with arbitrarily high
codimension.
As we can see here, centreless coadjoint orbits are somewhat messy: they can be
specified by an arbitrarily large number of parameters. Besides, we haven’t even dis-
cussed the case of coadjoint vectors p(ϕ) that vanish on a whole open set in S 1 — these
have an infinite-dimensional little group and their orbits have infinite codimension in
Diff(S 1 ). In particular, coadjoint orbits can have arbitrary (even or odd) codimension
in Diff(S 1 ). This is in sharp contrast with finite-dimensional Lie groups, where all
coadjoint orbits are even-dimensional since they are symplectic manifolds. In the case
of Diff(S 1 ), coadjoint orbits are still symplectic, but they need not satisfy “codimension
parity”: the fact that a given orbit has codimension N does not imply that there are
no orbits with codimension N ± 1. We shall see that this pathology does not occur
when the Virasoro central charge is non-zero, where all orbits have codimension one
or three.
7.1.2 Basic properties of centrally extended orbits

Let us turn to coadjoint orbits of the Virasoro group at non-zero central charge c 6= 0.
From now on we pick some non-zero value for c and we stick to it; for definiteness we
take c > 0, although all our considerations also apply to c < 0 after a few straightfor-
ward sign modifications. In principle our goal is to address the following problems:
1. Classify all Virasoro coadjoint orbits with central charge c.
2. Find a non-redundant, exhaustive set of orbit representatives.
3. Given a coadjoint vector p(ϕ)dϕ2 (at central charge c), write down the diffeo-
morphism f ∈ Diff(S 1 ) that maps it on one of the orbit representatives.
If we manage to satisfy these criteria, we will have fully classified the orbits of the
Virasoro group (at non-zero central charge).
While this task was relatively easy in the centreless case thanks to the integral in-
variants (7.3), it turns out to be much more complicated in the centrally extended case;
the remainder of this chapter is devoted to orbits at non-zero central charge, where
the classification will rely on elaborate techniques involving monodromy matrices. For
now we simply describe the most elementary aspects of some of these orbits.
Stabilizers

Suppose we are given a coadjoint vector p(ϕ)dϕ2 , c . Since the central charge is
invariant, the orbit of (p, c) under the Virasoro group can be represented as
n o
W(p,c) = Adc ∗ p f ∈ Diff(S 1 ) , (7.4)
f
∗
c p is given by (6.114). It is an infinite-dimensional manifold, so obtaining
where Ad f
information on its geometry sounds at first like an impossible task. Accordingly,
instead of actually trying to picture the orbit as such, let us look for the stabilizer Gp
of p, which is a subgroup of Diff(S 1 ) such that
W(p,c) ∼ d 1 )/(Gp × R) ∼
= Diff(S = Diff(S 1 )/Gp . (7.5)
The stabilizer consists of diffeomorphisms f (ϕ) such that

1 h c i
p(f (ϕ)) = p(ϕ) + S[f ](ϕ) . (7.6)
(f 0 (ϕ))2 12
Given p(ϕ), this is a highly non-linear differential equation for f (ϕ); if we could actu-
ally solve it, we would know the stabilizer.
To make things simpler let us look only for the Lie algebra of the stabilizer, rather
than the stabilizer itself. This algebra is spanned by vector fields X that leave p(ϕ)
invariant, which according to (6.115) amounts to the requirement
c 000
Xp0 + 2X 0 p − X = 0. (7.7)
12
This is already a lot easier than eq. (7.6): it is a linear third order equation for the
function X(ϕ), assuming that the function p(ϕ) is known. A number of important
consequences follow from this equation. The first is that, for non-zero c, it admits at
most three linearly independent solutions:
Lemma. The stabilizer of p(ϕ)dϕ2 at non-zero central charge is at most three-

dimensional.
This is already a sharp difference with respect to the centreless case, where sta-
bilizers had arbitrarily high dimension. If we were on a line rather than a circle, we
would actually conclude from (7.7) that the stabilizer is always three-dimensional; but
the requirement of periodicity restricts the space of allowed solutions X for a given p,
as we shall see momentarily.
Constant coadjoint vectors

It is worth exploring the solutions of (7.7) in the simple case where p(ϕ) = p0 is a
constant. The equation then reduces to
24p0 0
X 000 − X = 0,
c
whose general solution is a sum of exponentials

q q
24p0 24p0
ϕ − ϕ
X(ϕ) = A + B e c +Ce c (7.8)
wherepA is real while B and C are generally complex coefficients, being understood
that 24p0 /c is purely imaginary when p0 < 0. For generic values of p0 , the only 2π-
periodic solution of this type is a constant X(ϕ) = const. In that case the stabilizer
is one-dimensional, and consists of rigid rotations of the circle. But there also exist
exceptional values of p0 whose stabilizer is larger, namely
n2 c
p0 = − (7.9)
24
where n ∈ N∗ is a positive integer. At such values the exponentials in (7.8) are e±inϕ
and the corresponding vector field X is automatically 2π-periodic (and real upon set-
ting C = B ∗ ). Thus, for exceptional constants (7.9), the stabilizer is three-dimensional.
Its Lie algebra is isomorphic to sl(2, R); we will see below that the stabilizer itself is
an n-fold cover of PSL(2, R). In particular, orbits of generic constants p0 are radically
different from orbits of exceptional constants (7.9). The situation is depicted in fig. 7.1.
If one thinks of Diff(S 1 ) as a group of conformal transformations and identifies p(ϕ)

with the stress tensor of a CFT, one can recognize in (7.9) with n = 1 the vacuum
value of a stress tensor on the cylinder:
c
pvac = − . (7.10)
24
We shall see below that this interpretation is indeed correct, as the coadjoint orbit of
pvac turns out to be the lowest-lying orbit with energy bounded from below and has
its energy minimum at pvac . By contrast, the points (7.9) with n ≥ 2 belong to orbits
with energy unbounded from below.
7.1.3 Hill’s equation and monodromy

At this point we have seen the most basic features of Virasoro orbits at c 6= 0; we
now describe the first step of the complete classification by explaining the relation
between orbits and monodromy matrices for solutions of Hill’s equations. To lighten
c ∗ p ≡ f · p.
the notation, from now on we write Ad f
Until the next section it will be convenient to think of the coordinate ϕ as span-
ning the real line R, without identification ϕ ∼ ϕ + 2π (this will be justified below).
Accordingly we now reinstate the notation Diff g + (S 1 ) for the universal cover of the
group of orientation-preserving diffeomorphisms of the circle and we think of it as a
subgroup of Diff+ (R). Functions on the circle then are 2π-periodic functions on R.
We also sometimes use the words “conformally invariant” or “conformally equivalent”
g + (S 1 )-invariant or Diff
to refer to objects that are Diff g + (S 1 )-equivalent, respectively.
Figure 7.1: The map of Virasoro coadjoint orbits with constant representatives. The
open dots labelled by n = 1, 2, 3, ... indicate the location of the exceptional points
−c/24, −4c/24, −9c/24, etc.
Virasoro symmetry of Hill’s equation

The key idea of the classification is the following: given a coadjoint vector (p, c), we
can associate with it a differential operator
c ∂2
∆(p,c) ≡ − + p(ϕ) (7.11)
6 ∂ϕ2
where ϕ is a coordinate on the real line, p(ϕ) is 2π-periodic, and the operator ∆(p,c)
acts on suitable densities on the real line. The normalization in front of ∂ϕ2 is chosen
g + (S 1 ),
so as to ensure that the operator has good transformation properties under Diff
as we shall see below. Note that the crucial term ∂ϕ2 disappears if c = 0, which is why
the considerations that follow apply exclusively to orbits with non-zero central charge.
Definition. Let (p, c) be a Virasoro coadjoint vector with c 6= 0. The associated

Hill’s equation is the second-order, linear differential equation
c
− ψ 00 (ϕ) + p(ϕ)ψ(ϕ) = 0 (7.12)
6
for the real-valued function ψ(ϕ) on the real line. With the notation (7.11) this is just
the statement ∆(p,c) · ψ = 0.
Hill’s equation may be seen as a non-relativistic Schrödinger equation on the real

line for a “wavefunction” ψ(ϕ) with a periodic “potential energy” p(ϕ), up to the fact
that ψ is real and need not be square-integrable.1 Thus we can associate an equation
(7.12) with each coadjoint vector (p, c), and vice-versa.
g + (S 1 ) determines that of Hill’s operator

The transformation law of (p, c) under Diff
(7.11). Using (6.112) and the centreless transformation law (6.36), we find
c 2 c
∆(f ·p,c) (ϕ) = − ∂ϕ2 + p(f −1 (ϕ)) (f −1 )0 (ϕ) − S[f −1 ](ϕ) , (7.13)
6 12
which is indeed very different from the original operator (7.11). The key point, how-
ever, is that the associated Hill’s equation (7.12) can be made conformally invariant
by choosing a suitable transformation law for ψ(ϕ):
Lemma. If ψ(ϕ) is a density with weight −1/2 on the real line, then Hill’s equation
g + (S 1 ).
(7.12) is invariant under Diff
Proof. To simplify formulas, let us act on Hill’s operator with a diffeomorphism f −1
rather than f so that f −1 · ψ (ϕ) = ψ(f (ϕ))(f 0 (ϕ))−1/2 . Then f −1 maps the left-hand
side of Hill’s equation (7.12) on
c
− ∂ϕ2 ψ(f (ϕ))(f 0 (ϕ))−1/2 + p(f (ϕ))ψ(f (ϕ))(f 0 (ϕ))3/2

6 (7.14)
c
− S[f ](ϕ) (f 0 (ϕ))−1/2 ψ(f (ϕ))
12
where the term with a second derivative can be written as
1
∂ϕ2 ψ(f (ϕ))(f 0 (ϕ))−1/2 = ψ 00 (f (ϕ))(f 0 (ϕ))3/2 − ψ(f (ϕ))(f 0 (ϕ))−1/2 S[f ](ϕ) .

2
Here the term involving the Schwarzian derivative cancels that of (7.14); the latter
expression can therefore be rewritten as
3/2 h c 00 i
f 0 (ϕ) − ψ (ϕ) + p(ϕ)ψ(ϕ) .
6
Provided ψ solves (7.12), this vanishes.
As a consequence of this lemma, the map that associates Hill’s equations with
g + (S 1 )-invariant. Thus, Hill’s equation is an
Virasoro coadjoint vectors (p, c) is Diff
invariant associated with each coadjoint orbit of the Virasoro group, and classifying
g + (S 1 )-inequivalent Hill’s equations.
Virasoro orbits is equivalent to classifying all Diff
1
Note that Hill’s operator (7.11) coincides with a Sturm-Liouville operator with periodic potential.
Monodromy
Hill’s equation (7.12) is a second-order linear differential equation, so its solutions span
a two-dimensional vector space. Let ψ1 and ψ2 be two linearly independent solutions.
We define their Wronskian as

ψ1 ψ2
W ≡ det = ψ1 ψ20 − ψ2 ψ10 . (7.15)
ψ10 ψ20
The Wronskian is constant on the real line (W 0 = 0) by virtue of Hill’s equation.
Furthermore W does not vanish since ψ1 and ψ2 are linearly independent. (Conversely,
if the Wronskian does not vanish, then the solutions ψ1 , ψ2 are linearly independent.)
Thus we can always choose
W [ψ1 , ψ2 ] = −1 . (7.16)
We will refer to this equality as the “Wronskian condition” and to the solutions that
satisfy it as being “normalized”. Note that the Wronskian (7.15) is invariant under
g + (S 1 ) when the ψi ’s transform as densities of weight −1/2, regardless of them
Diff
solving Hill’s equation:
W [f · ψ1 , f · ψ2 ](ϕ) = W [ψ1 , ψ2 ](f −1 (ϕ)) .
In particular, for solutions of Hill’s equation, the Wronskian is constant:
W [f · ψ1 , f · ψ2 ] = W [ψ1 , ψ2 ] when ψ1 , ψ2 solve (7.12). (7.17)
Hill’s equation is a differential equation on the real line ϕ ∈ R with a 2π-periodic

potential p(ϕ). Its solutions need not be periodic, but they do satisfy certain con-
straints due to the periodicity of p(ϕ):
Lemma. Let p(ϕ) be 2π-periodic and let ψ1 , ψ2 be linearly independent solutions of

Hill’s equation (7.12). Then there exists a monodromy matrix M ∈ SL(2, R) such that,
for any ϕ ∈ R,
ψ1 (ϕ + 2π) ψ1 (ϕ)
=M· . (7.18)
ψ2 (ϕ + 2π) ψ2 (ϕ)
Proof. Let ψ1 , ψ2 be two linearly independent solutions of (7.12), and define ψ̃i (ϕ) ≡
ψi (ϕ + 2π) for i = 1, 2. Then the Wronskian associated with ψ̃1,2 takes the same value
as that of ψ1,2 ; furthermore, the functions ψ̃i solve the same Hill’s equation as the
functions ψi since p(ϕ) is 2π-periodic. This implies that there exists a real matrix
M such that (7.18) holds for any ϕ ∈ R. Since the ψi ’s and the ψ̃i ’s have the same
Wronskian, M must have unit determinant.
Thus we can associate a monodromy matrix with any Virasoro coadjoint vector
and any pair of (normalized) solutions of the corresponding Hill’s equation. From now
on we use the notation
ψ1
Ψ≡ (7.19)
ψ2
for the “solution vector” associated with the basis of solutions ψ1 , ψ2 . Relation (7.18)
then becomes Ψ(ϕ + 2π) = M · Ψ(ϕ). If we were to choose another normalized basis
of solutions, say (φ1 , φ2 ) = Φt , there would be a linear relation Φ = S · Ψ between

solution vectors, for some matrix S ∈ SL(2, R). Accordingly the monodromy matrix
MΦ associated with Φ would be related to the monodromy MΨ of Ψ by MΦ = SMΨ S −1 .
Thus the monodromy matrix changes by conjugation in SL(2, R) under changes of
bases of normalized solutions. In particular, the conjugacy class of M,
[M] ≡ SMS −1 S ∈ SL(2, R) ,

is invariant under changes of bases. It depends only on the function p(ϕ), and not on
the choice of solutions Ψ.
We have shown above that Hill’s equation is invariant under Diffg + (S 1 ) in the sense
that, if ψ solves the equation with a potential p(ϕ), then f · ψ solves the same equation
with a potential f · p. In addition we have seen in (7.17) that this transformation
preserves the Wronskian condition, so that normalized solutions remain normalized
g + (S 1 ). Accordingly, if Ψ is a normalized solution vector for the potential p,
under Diff
then f · Ψ is a normalized solution vector for f · p. And now comes the key argument:
g + (S 1 ) on Ψ are linear, the monodromy
since both Hill’s equation and the action of Diff
matrix of f · Ψ coincides with that of Ψ. We therefore conclude:
Theorem. Let c 6= 0 and denote by [M](p,c) the conjugacy class of any monodromy
matrix M associated with the Hill’s equation (7.12) specified by p(ϕ) and c. Then
there is a well-defined map
{Virasoro orbits at central charge c} → {Conjugacy classes of SL(2, R)} (7.20)
that associates with a coadjoint orbit W(p,c) the equivalence class [M](p,c) of the corre-
sponding monodromy matrix. In particular, Virasoro coadjoint vectors with the same
central charge but non-conjugate monodromy matrices do not belong to the same orbit.
This result illustrates the power of Hill’s operators. It provides a rough classi-
fication of Virasoro orbits by allowing us to distinguish orbits with non-conjugate
monodromies and may be seen as an infinite-dimensional analogue of the classification
of coadjoint orbits of SL(2, R) according to the value of the “mass squared”. In partic-
ular the trace Tr(M) is a conformally invariant quantity. However, the classification is
not precise in that two orbits whose monodromy matrices are conjugate may well be
different: the map (7.20) need not be injective (and we shall see below that it is not).
To make further progress we need to investigate Hill’s equation in more detail.
For future reference, note the following: thanks to the fact that the conjugacy class
of the monodromy matrix is independent of the choice of a solution vector Ψ for Hill’s
equation associated with (p, c), one can write its trace as a Wilson loop
Z 2π
0 1
Tr(M) = Tr P exp dϕ (7.21)
0 6p(ϕ)/c 0
where P denotes path ordering. This quantity is conformally invariant, so one can
replace (p, c) by any coadjoint vector (q, c) belonging to its orbit without affecting the
value of (7.21). In particular, if (p, c) belongs to the orbit of a constant coadjoint

vector (p0 , c) with positive p0 , the trace reads
r
6p0
Tr(M) = 2 cosh 2π . (7.22)
c
The same formula holds for negative p0 , with cosh(ix) = cos(x). We will put it to use
in section 10.1 when defining the mass of BMS3 particles.
Hill’s equation and stabilizers

The stabilizer of a coadjoint vector (p, c) consists of diffeomorphisms that satisfy (7.6).
Let us see how this information is related to Hill’s equation (7.12). First note that, if
ψ1 and ψ2 are linearly independent solutions of Hill’s equation, then the combinations
ψ12 , ψ1 ψ2 , ψ22 (7.23)
all solve the stabilizer equation (7.7). These products are generally not 2π-periodic
and therefore do not represent vector fields on the circle, but one can show that there
always exist either one or three 2π-periodic linear combinations of these products.
This confirms our earlier observation that the stabilizer of all orbits is either one- or
three-dimensional. Note that, being −1/2-densities on the circle, the products (7.23)
were bound to be densities of weight −1, i.e. vector fields.
Let us now see how the stabilizer Gp of (p, c) is described in the Hill language. If
f ∈ Gp and if Ψ is a normalized solution vector of Hill’s equation associated with (p, c),
the action of Gp on Ψ is such that f · Ψ provides another normalized solution vector
for the same equation. Accordingly there exists some (constant) SL(2, R) matrix Af
such that
f · Ψ = A−1
f Ψ. (7.24)
In addition we have seen that the action of Diff(S 1 ) leaves the monodromy matrix
invariant, so the monodromy of f ·Ψ coincides with the monodromy M of Ψ. Combining
this statement with (7.24) we conclude that A−1 f MAf = M, which is to say that Af
belongs to the stabilizer GM of M with respect to conjugation. In addition the inversion
A−1
f in (7.24) ensures that Af g = Af Ag , so we conclude:
Lemma. Let (p, c) be a Virasoro coadjoint vector with c 6= 0, Ψ a normalized solution

vector of the associated Hill’s equation. Let Gp be the stabilizer of p for the coadjoint
action (6.114) and let GM be the stabilizer of M for conjugation. Then the map
A : Gp → GM : f 7→ A(f ) ≡ Af (7.25)
defined by (7.24) is a homomorphism.
This map relates the stabilizer of p to that of the corresponding monodromy matrix.
In particular it allows us to classify the conformally inequivalent solutions of Hill’s
equation at fixed (p, c). Indeed, the set of normalized solution vectors of Hill’s equation
at p with fixed monodromy M is in one-to-one correspondence with the elements of

GM , so the set of orbits of the stabilizer Gp in that set of solutions is a quotient
GM /Im(A) (7.26)
where Im(A) is the image of (7.25). Two solution vectors are conformally equivalent if
and only if they belong to the same orbit under Gp , i.e. if they define the same point
in (7.26).
Remark. The fact that the products of half-densities (7.23) solving Hill’s equation
produce integer densities solving the stabilizer equation (7.7) is reminiscent of the fact
that the “square” of two Killing spinors is a Killing vector. This correspondence is
exactly realized in three-dimensional gravity: eq. (7.7) turns out to coincide with the
Killing equation expressed in terms of a suitable component X of a vector field on
space-time, while Hill’s equation (7.12) corresponds to the Killing spinor equation for
a suitable spinor component (see e.g. eq. (16) in [167]).
7.1.4 Winding number

The conjugacy class of monodromy matrices provides a continuous parameter that
roughly classifies Virasoro orbits. We now describe a second invariant quantity which,
combined with monodromies, will provide a precise classification of orbits. This second
invariant turns out to be the discrete winding number of a path in the circle.
Let ψ1 and ψ2 be normalized solutions of Hill’s equation (7.12). They have non-zero
g + (S 1 ), but their ratio
weight under Diff
ψ1 (ϕ)
η(ϕ) ≡ (7.27)
ψ2 (ϕ)
g + (S 1 ) as a function (i.e. a zero-weight density). It blows up at
transforms under Diff
the zeros of ψ2 , so it is more convenient to think of it as a curve
η : R → RP 1 : ϕ 7→ η(ϕ) (7.28)
whose expression is (7.27) in terms of the projective coordinate (6.84). The points
where η diverges are then mapped by η on the “point at infinity” in RP 1 . Since RP 1
is diffeomorphic to the circle (6.83), we can also think of η as a path in S 1 whose
expression in stereographic coordinates is (7.27).
Coadjoint vectors from projective curves

As in (7.19) we denote the basis of solutions ψ1,2 by Ψ. Then the quasi-periodicity
(7.18) of Ψ implies a similar “projective” monodromy for η(ϕ),
a η(ϕ) + b
η(ϕ + 2π) = (7.29)
c η(ϕ) + d
where a, b, c, d are the entries of the monodromy matrix M. If we let Φ = AΨ be another
normalized basis of solutions with A ∈ SL(2, R), the curve η̃ = φ1 /φ2 corresponding
to Φ by (7.27) is related to η by a projective transformation of the form (6.87). In

particular eq. (6.91) implies that the Schwarzian derivative of η with respect to ϕ is left
unchanged by such a transformation. Thus the Schwarzian derivative of η is invariant
under changes of (normalized) bases of solutions of Hill’s equation, which is consistent
with the following observation:
Lemma. Let ψ1 and ψ2 be normalized solutions of Hill’s equation (7.12) and η ≡

ψ1 /ψ2 . Then the function p(ϕ) is specified by the solutions of its Hill’s equation:
12
S[η](ϕ) = − p(ϕ) . (7.30)
c
Proof. By virtue of the Wronskian condition (7.16),
1
η0 = . (7.31)
(ψ2 )2
ψ 00 (ϕ)
It then follows from the definition (6.76) that S[η](ϕ) = −2 ψ22 (ϕ) , which coincides with
the right-hand side of (7.30) upon using Hill’s equation (7.12).
This lemma says that the correspondence between Virasoro coadjoint vectors and
solutions of Hill’s equation goes bothways: Hill’s equation specifies certain solutions,
which in turn uniquely determine the periodic potential p(ϕ) via (7.30). In particular
the coadjoint transformation law (6.114) of p can be rewritten in terms of the scalar
transformation law of (7.27) plugged into (7.30).
Winding numbers
One can think of η(ϕ) as a path in the circle with a “time parameter” ϕ. Eq. (7.31)
then says that η 0 (ϕ) > 0, so η(ϕ) always spins around the circle in the same direction.
We therefore introduce the following terminology:
Definition. The winding number n ∈ N of η(ϕ) is the number of laps around the
circle performed by η in a “time interval” of length 2π.
We will illustrate the computation of the winding number in the next section, when
describing explicit Virasoro orbit representatives. For now note that η(ϕ) transforms
g + (S 1 ) as a function, so its winding number is conformally invariant:
under Diff
Proposition. Let c 6= 0 and let n(p,c) ∈ N be the winding number of the curve η
associated with the Hill’s equation (7.12) specified by p(ϕ) and c. Then there is a
well-defined map
{Virasoro orbits at central charge c} → N : W(p,c) 7→ n(p,c) . (7.32)
In particular, Virasoro coadjoint vectors with the same central charge but different
winding numbers do not belong to the same orbit.
This supplements our previous observation (7.20) that the conjugacy classes of
monodromy matrices yield a rough classification of Virasoro coadjoint orbits. In fact,
these two invariants together provide the complete classification of Virasoro orbits.
Indeed one can show that the map that associates a pair ([M], n) with each Virasoro
coadjoint orbit is injective, provided [M] is the conjugacy class of the monodromy
matrix and n is the winding number. Note however that the map is not surjective, as
some pairs ([M], n) do not belong to its image. We now verify this by brute force by
describing orbit representatives.
7.2 Virasoro orbit representatives

Virasoro coadjoint orbits are classified by two parameters, one of them continuous (the
conjugacy class of the monodromy M), the other discrete (the winding number n). In
this section we display explicit orbit representatives for all admissible pairs ([M], n),
after a brief review of conjugacy classes in SL(2, R). We end with a picture of orbits
that extends fig. 7.1. As before, we assume that the central charge c is positive.
7.2.1 Prelude: Conjugacy classes of SL(2, R)

In order to classify the conjugacy classes of SL(2, R), we note that the trace of an
SL(2, R) matrix is invariant under conjugation; matrices with different traces cannot
be conjugate. This motivates the following terminology:

elliptic
 if |Tr(M)| < 2;
An SL(2, R) matrix M is parabolic if |Tr(M)| = 2;

hyperbolic if |Tr(M)| > 2.

Each conjugacy class of SL(2, R) is contained in one of these three families, but each
family contains several conjugacy classes. The elliptic and hyperbolic families contain
infinitely many conjugacy classes since they depend on a continuous parameter (the
trace of M). Note that the trace of M determine the properties of its eigenvalues:
M is elliptic ↔ distinct complex eigenvalues;
M is parabolic ↔ degenerate real eigenvalue ±1;
M is hyperbolic ↔ distinct real eigenvalues.
We now determine the conjugacy classes contained in each family. The computations
are very similar to those of section 4.3 where we determined the orbits of momenta for
the Poincaré group in three dimensions.
Lemma (elliptic family). Let M be elliptic. Then it is conjugate to a unique

rotation matrix
cos(2πω) sin(2πω)
(7.33)
− sin(2πω) cos(2πω)
where ω belongs to the set ]0, 1/2[ ∪ ]1/2, 1[ . The stabilizer of (7.33) is the U(1)
rotation subgroup (4.82) of SL(2, R).
Proof. In the elliptic family, the eigenvalues of M are complex conjugates of one another
with non-zero imaginary part. Since det(M) = 1, they can be written as e±2πiω where
ω belongs to the open interval ]0, 1[ without loss of generality, but differs from 1/2.
Let v ∈ C2 be an eigenvector of M such that M · v = e2πiω v. This vector is complex
and linearly independent of its complex conjugate v̄; the latter is an eigenvector of M
with eigenvalue e−2πiω . Then v + v̄ and i(v − v̄) are linearly independent real vectors;
we can choose the norm of v in such a way that the (real) matrix S expressing M in
the basis {v + v̄, i(v − v̄)} has unit determinant. Then SMS −1 takes the form (7.33).
The stabilizer consists of all matrices that commute with (7.33) and is readily seen to
consist of rotations.
Lemma (parabolic family). Let M be parabolic. Then it is conjugate to exactly

one of the following six matrices:

1 0 1 1 1 −1
± , ± , ± . (7.34)
0 1 0 1 0 1
The stabilizer of the first two matrices is the whole group SL(2, R), while the stabilizer
R × Z2 of the four remaining ones consists of triangular matrices (4.96).
Proof. When M is parabolic, its eigenvalues are either both 1 or both −1. Let λ be
the eigenvalue of M and let v ∈ R2 be a (real) eigenvector of M. Let v 0 be another
vector such that {v, v 0 } is a basis of R2 , and choose the normalization of v and v 0 in
such a way that the matrix S expressing M in this basis has unit determinant. Then

−1 λ x
SMS = (7.35)
0 λ
where λ = ±1 and x is an arbitrary real number. For x = 0 we find the first two
matrices in the list (7.34), each of which is alone in its conjugacy class. For non-zero
x, note that
y 0 1 ±1 1/y 0 1 ±y 2
= , (7.36)
0 1/y 0 1 0 y 0 1
so for λ = +1, M is conjugate to the second matrix in (7.34) if x > 0 and to the
third one if x < 0, in both cases with an overall plus sign. The situation is similar
when λ = −1, but with the minus sign. The proof ends with the observation that all
matrices in (7.34) belong to disjoint conjugacy classes. The stabilizer is obtained by
direct computation.
Lemma (hyperbolic family). Let M be hyperbolic. Then it is conjugate to a

unique matrix of the form 2πω
e 0
± (7.37)
0 e−2πω
where ω is a strictly positive real number. Its stabilizer is the group R × Z2 consisting
of matrices (4.95) of the same form as (7.37) but without restriction on ω ∈ R.
Proof. Since M is hyperbolic, it has two distinct real eigenvalues λ and 1/λ, where
λ ∈ R∗ . Let v and v 0 be two eigenvectors of M for these eigenvalues; we can normalize
them so that the matrix S expressing M in the basis {v, v 0 } has unit determinant.
Then SMS −1 takes the form (7.37) with e2πω = λ or e2πω = 1/λ. The ordering of
0 1
eigenvalues can be changed thanks to the SL(2, R) matrix , so we are free to
−1 0
pick ω > 0, and this specifies uniquely the conjugacy class of the matrix M. Finding
the stabilizer is straightforward.
From now on we say that a Virasoro orbit is elliptic, parabolic or hyperbolic if the
associated monodromy matrix is of one of those three types, respectively. In addition
we will distinguish parabolic orbits associated with ±I from parabolic orbits associated
with the four other matrices in (7.34) by referring to the former as “degenerate” and
to the latter as “non-degenerate”.
7.2.2 Elliptic orbits

Here we initiate the classification of Virasoro coadjoint orbits, by studying those whose
monodromy is elliptic. Parabolic and hyperbolic orbits will be investigated in sections
7.2.3 to 7.2.6.
Finding orbit representatives

Let c > 0 and suppose that q(ϕ)dϕ2 is a quadratic density such that the monodromy
of Hill’s equation (7.12) is elliptic. (We denote the quadratic density by q rather than
p, because the latter will eventually be the “representative” of the orbit of q.) Then
we can choose a solution vector Ψ whose monodromy matrix takes the form (7.33) for
some angle 2πω which is not an integer multiple of π. The function
Xq (ϕ) ≡ ψ12 (ϕ) + ψ22 (ϕ) (7.38)
is strictly positive and 2π-periodic; it is a vector field on the circle, since it is a quadratic
combination of −1/2-densities such as (7.23). In fact, it belongs to the Lie algebra of
the stabilizer of q since it solves equation (7.7). In addition it is invariant under the
action of the stabilizer of M and is therefore a well-defined functional of q(ϕ), which
justifies the notation Xq . Conversely, Xq determines q(ϕ) since Hill’s equation implies
" 0 2 #
00 00 00
c ψ1 ψ1 + ψ2 ψ2 (7.38) c 1 qX 1 X q 1
q= = − − 2 . (7.39)
6 Xq 6 2 Xq 4 Xq Xq
We would have obtained the same formula upon using eq. (7.30) with η = ψ1 /ψ2 . Our
g + (S 1 ) such that q is obtained by acting
goal now is to build a diffeomorphism gq ∈ Diff
with gq on a suitable orbit representative p. In the language of induced representa-
tions, the maps gq will be “standard boosts” on the orbit of p.
Let us define the negative number

Z 2π 2
c dϕ
p0 ≡ − , (7.40)
6 0 Xq (ϕ)
where the notation “p0 ” will be justified below. This number is well-defined since
g + (S 1 ) since Xq is a vector field. We
Xq (ϕ) never vanishes, and it is invariant under Diff
can then define a diffeomorphism f ∈ Diffg + (S 1 ) by
Z ϕ
2π dφ
f (ϕ) ≡ p . (7.41)
6|p0 |/c 0 Xq (φ)
This quantity is the inverse of the sought-for standard boost since eq. (7.39) can be
written as
c
q(ϕ) = p0 (f 0 (ϕ))2 − S[f ](ϕ) , (7.42)
12
which we recognize as the coadjoint action (6.113) of
gq ≡ f −1 (7.43)
on the constant coadjoint vector p(ϕ) = p0 < 0. In conclusion:
Proposition. Let (q, c) with c > 0 be a Virasoro coadjoint vector with elliptic
monodromy. Then it belongs to the orbit of a constant coadjoint vector (p, c) with
p(ϕ) = p0 , where the value of p0 is determined by q(ϕ) according to (7.40) with Xq
given by (7.38) in terms of normalized solutions of the Hill’s equation of (q, c). In
addition, the diffeomorphism gq defined as the inverse of (7.41) is a standard boost for
the orbit of p in the sense that
gq · p = q (7.44)
where the dot denotes the coadjoint action (6.114).
Monodromy and winding number

Let us now see how the parameter (7.40) is related to the monodromy matrix. At
p = p0 , Hill’s equation (7.12) reads
c
− ψ 00 − |p0 |ψ = 0 (7.45)
6
where we write p0 = −|p0 | to emphasize that this is a harmonic oscillator equation
with frequency p
ω = 6|p0 |/c . (7.46)
A basis of solutions satisfying the Wronskian condition (7.16) is provided by
1 1
ψ1 (ϕ) = √ sin(ωϕ), ψ2 (ϕ) = √ cos(ωϕ). (7.47)
ω ω
The corresponding monodromy matrix M is readily seen to take the form (7.33) with
ω given by (7.46) in terms of p0 /c. The fact that the monodromy matrix is elliptic
implies that ω is not an integer multiple of 1/2, which is equivalent to saying that
n2 c
p0 6= − . (7.48)
24
In the language of section 7.1.2, the constant orbit representative p0 must be generic
in order for its orbit to be elliptic. By contrast, the exceptional orbit representatives
(7.9) will turn out to have degenerate parabolic monodromy (see below).
Thus different values of p0 generally define disjoint orbits since their monodromy
matrices (7.33) are not conjugate. However, at this stage we cannot tell whether
p q 2
c
p0 and − |p0 | + 6
N (7.49)
belong to different orbits when N ∈ N since their monodromy matrices coincide (their
angles differ by 2πN ). This issue is settled by the winding number (7.32): the curve
(7.27) associated with the solutions (7.47) is
η(ϕ) = tan(ωϕ) , (7.50)
which can be seen as a path on a circle written in terms of a stereographic coordinate

η = tan(θ/2), where the coordinate θ ∈ R is identified as θ ∼ θ + 2π. In terms of θ
the path (7.50) is a rotation around the circle at constant velocity, θ(ϕ) = 2ωϕ. The
number of laps performed by this path around the circle when ϕ goes from zero to 2π
is the winding number2 $r %
(7.46) 24|p 0 |
np = b2ωc = (7.51)
c
where b·c denotes the integer part. Thus the winding number associated with p0 < 0
2c 2
takes a definite value in each interval ] − (n+1)
24
, − n24c [ , and jumps by one unit every
time p0 takes one of the exceptional values (7.9). For instance np = 0 when p0 belongs
to ] − c/24, 0[ , while np = 1 when p0 ∈ ] − c/6, −c/24[ , and so on.
In conclusion, the orbits of two generic constants p0 and p̃0 are disjoint if and only
if these constants differ. We have thus recovered the lower part (p0 < 0) of fig. 7.1. As
a bonus we can now assign a monodromy matrix determined by (7.46), and a winding
number (7.51), with each point on that part. In particular the integers n written on
the left of the p0 axis can be interpreted as winding numbers for constants p0 located
between −(n + 1)2 c/24 and −n2 c/24.
Stabilizers
To conclude the description of orbits of generic constants p0 < 0, it remains to find
their stabilizer. As anticipated in (7.8), one shows that the stabilizer of p0 is the group
U(1) of rigid rotations f (ϕ) = ϕ + θ (or more precisely its universal cover R when
g + (S 1 )). The coadjoint orbit of (p0 , c) can thus be written as
dealing with Diff
W(p0 ,c) ∼ g + (S 1 )/R ∼

= Diff = Diff+ (S 1 )/S 1 , (7.52)
which may be seen as an infinite-dimensional generalization of the orbit SL(2, R)/S 1 of

SL(2, R). The latter coincides with the momentum orbit (4.97) of a massive Poincaré
particle in three dimensions; in the same way, we shall see in section 10.1 that (7.52)
is the supermomentum orbit of a massive BMS3 particle.
2
We denote the winding number by np instead of n(p0 ,c) to reduce clutter.
Remark. The stabilizer U(1) coincides with the stabilizer of the monodromy matrix
(7.33), so the quotient (7.26) consists of a single point. This implies that all conformally
inequivalent normalized solutions of the Hill equation associated with (p0 , c) can be
obtained by acting with rotations on the solution (7.47).
7.2.3 Degenerate parabolic orbits

Orbit representatives
We now turn our attention to coadjoint vectors (q, c) whose monodromy matrix is
of the “degenerate” parabolic type (7.34), i.e. coincides with ±I. We proceed as in
the elliptic case. In particular the monodromy matrix still ensures that (7.38) is a
positive, 2π-periodic vector field belonging to the Lie algebra of the stabilizer of q(ϕ).
g + (S 1 )-invariant, and formula
The negative number (7.40) is still well-defined and Diff
(7.41) provides a diffeomorphism of S 1 such that eq. (7.42) holds. Then (7.43) is a
standard boost that maps the constant coadjoint vector p0 on q(ϕ); in particular the
proposition surrounding (7.44) still holds up to the replacement of the word “elliptic”
by “degenerate parabolic”.
As in the elliptic case we can choose constant coadjoint vectors as orbit representa-
tives. The corresponding Hill’s equation then reads (7.45) and admits the normalized
solutions (7.47), but the monodromy matrix is ±I. Such a monodromy matrix M only
occurs when p0 takes the exceptional form (7.9) for some strictly positive integer n, in
which case
n2 c
p0 = − and M = (−1)n I. (7.53)
24
By contrast, elliptic orbits never contain an exceptional constant; this is a sharp dif-
ference between elliptic and degenerate parabolic orbits.
The monodromy matrix (7.53) implies that two exceptional constants specified by
integers n, n0 can belong to the same orbit only if n and n0 have the same parity; but
at this stage we cannot tell if two orbits with the same parity are disjoint. As in the
elliptic case we can address this question by studying the winding number of the curve
(7.27) associated with the solutions (7.47). One can verify that the winding number
coincides with the number n specified by p0 = −n2 c/24, which implies that any two
orbits of exceptional constants specified by different values of n > 0 are disjoint. In
conclusion, we have now recovered the dots in the lower part of fig. 7.1, and the values
of n displayed there coincide with winding numbers. In particular the orbit at n = 1
will be called the vacuum orbit from now on; in the context of Riemann surfaces, it is
known as universal Teichmüller space [56, 168]. Note that the orbit of p0 = 0 does not
have degenerate parabolic monodromy, and so has not yet been accounted for by our
classification of Hill’s equations.
Stabilizers
We now study the stabilizers of orbits of exceptional constants p0 = −n2 c/24. We saw
below (7.8) that the stabilizer is three-dimensional for such values, and is generated
by the vector fields

∂ ∂ ∂
, sin(nϕ) , cos(nϕ) . (7.54)
∂ϕ ∂ϕ ∂ϕ
The Lie algebra of the stabilizer is therefore isomorphic to sl(2, R), but different values
of n define non-conjugate embeddings of sl(2, R) in Vect(S 1 ). In fact one can verify
using (6.96) that the finite diffeomorphisms that span the stabilizer of p0 = −n2 c/24
(and that reduce to (7.54) close to the identity) are projective transformations (6.95)
spanning a group PSL(n) (2, R) (the n-fold cover of PSL(2, R)). In conclusion:
Lemma. The stabilizer of p0 = −n2 c/24 for the coadjoint action of Diff g + (S 1 ) (resp.
Diff+ (S 1 )) is the group PSL
g (n) (2, R) (resp. PSL(n) (2, R)) spanned by diffeomorphisms
g (n) (2, R) is the universal cover of the n-fold cover of
f (ϕ) given by (6.95), where PSL
PSL(2, R) = SL(2, R)/Z2 . The coadjoint orbit of (p0 , c) can be written as
W 2
∼
= Diff g (n) (2, R) ∼
g + (S 1 )/PSL = Diff+ (S 1 )/PSL(n) (2, R) (7.55)
− n24c ,c
The Lie algebra of the stabilizer is generated by the vector fields (7.54).
In section 9.3 we will interpret the orbit of p0 = −c/24 as the set of gravitational
perturbations around Minkowski space. In that context the little group PSL(2, R) will
be seen as the Lorentz group in three dimensions. The remaining exceptional values
p0 = −n2 c/24 (with n ≥ 2) will be interpreted as conical excesses where one turn
around the origin of space spans an angle 2πn.
Remark. An important difference between elliptic and degenerate parabolic orbits

is that, in the latter case, the stabilizer of the monodromy matrix (7.53) is the whole
group SL(2, R), which does not leave the combination (7.38) invariant. Nevertheless,
the integral (7.40) is still independent of the choice of the normalized solution vector
Ψ because, in that specific case, any SL(2, R) transformation Ψ 7→ SΨ is equivalent to
the action of a diffeomorphism of the circle belonging to the stabilizer of p0 ; since the
integral (7.40) is invariant under diffeomorphisms, it follows that it is also invariant
under Ψ 7→ SΨ for any S ∈ SL(2, R).
7.2.4 Hyperbolic orbits without winding

Consider a Virasoro coadjoint vector (q, c) whose monodromy matrix is of the hyper-
bolic type (7.37) with some ω > 0. We shall see that hyperbolic orbits differ greatly
depending on the winding number of the curve (7.27), so we focus here on the case of
zero winding; the non-zero case will be treated in section 7.2.5.
Finding orbit representatives

Let ψ1 and ψ2 be normalized solutions of Hill’s equation associated with (q, c) and let
η = ψ1 /ψ2 . Since the winding number of η is zero, we can choose our solution vector
such that ψ2 has no zeros on the real line. Then η(ϕ) is smooth and, by virtue of
(7.37), we have
η(ϕ + 2π) = e4πω η(ϕ) (7.56)
so η(ϕ) is monotonically increasing on R (since ω > 0). As in the case of elliptic orbits,
our goal is to find a “standard boost” gq whose inverse gq−1 ≡ f will map q(ϕ) on a
suitably chosen orbit representative. To do so we define

1 (7.27) 1 ψ1 (ϕ)
f (ϕ) ≡ log(η(ϕ)) = log (7.57)
2ω 2ω ψ2 (ϕ)
g + (S 1 ) by virtue of (7.56). Now if we set
which belongs to Diff
c ω2
p0 ≡ , (7.58)
6
we can use (7.30) and the cocycle identity (6.77) to write
c
q(ϕ) = p0 (f 0 (ϕ))2 − S[f ] . (7.59)
12
As in (7.42) we recognize the coadjoint action of gq = f −1 on the constant coadjoint
vector p(ϕ) = p0 > 0, and thus conclude:
Proposition. Let (q, c) with c > 0 be a Virasoro coadjoint vector with hyperbolic
monodromy and zero winding number. Then it belongs to the orbit of a constant
coadjoint vector (p0 , c), where p0 > 0 is determined by the monodromy matrix accord-
ing to (7.58). In addition the diffeomorphism gq defined as the inverse of (7.57) is a
standard boost for the orbit of p in the sense (7.44).
Note that the definition (7.58) coincides with eq. (7.46) for p0 > 0. Roughly
speaking, “hyperbolic orbits are an analytic continuation of elliptic orbits to imaginary
values of the monodromy parameter ω”. This is analogous to the fact that tachyonic
momentum orbits may be seen as massive orbits with imaginary mass.
Stabilizers
At p = p0 , Hill’s equation (7.12) reads − 6c ψ 00 + |p0 |ψ = 0 where we stress that the sign
of the potential term is opposite to the one in (7.45). A basis of solutions satisfying
the Wronskian condition (7.16) is provided by
1 1
ψ1± (ϕ) = ± √ eωϕ , ψ2± (ϕ) = ± √ e−ωϕ (7.60)
2ω 2ω
where ω > 0 is given by (7.46). The corresponding monodromy matrix is (7.37).
We have seen in (7.8) that the stabilizer is one-dimensional for p0 > 0, and that
it consists of rotations of the circle. Thus the stabilizer of p0 is a group U(1) of rigid
rotations (or more precisely its universal cover R when dealing with Diff g + (S 1 )). In
particular the orbit of (p0 , c) for p0 > 0 and c > 0 is diffeomorphic to
W(p0 ,c) ∼ g + (S 1 )/R ∼

= Diff = Diff+ (S 1 )/S 1 . (7.61)
As in (7.52) this orbit may be seen as an infinite-dimensional generalization of SL(2, R)

orbits of the type SL(2, R)/S 1 . However, the orbit differs from those of negative p0 ’s
in that the two choices of signs in (7.60) are conformally inequivalent. Indeed, the
stabilizer GM of the matrix (7.37) under conjugation is isomorphic to R × Z2 while
the universal cover of the little group of p0 is just R. Accordingly the quotient (7.26)
contains two points, indicating that there are two inequivalent normalized families of
solutions to Hill’s equation at (p0 , c); these two families are labelled by the sign ± in
(7.60).
In terms of fig. 7.1, we have now completed our understanding of almost the whole
real line p0 ∈ R, since we now know that the orbits that pass through p0 > 0 are of
hyperbolic type without winding. The only remaining mystery is the orbit of p0 = 0,
and of course all the orbits that do not contain constant representatives.
7.2.5 Hyperbolic orbits with winding

Building orbit representatives
We now consider a Virasoro coadjoint vector (q, c) with hyperbolic monodromy (7.37)
but strictly positive winding number n > 0. The classification of orbits of such vectors
is more involved than in the previously encountered cases, so we proceed in a “back-
wards” fashion. Namely, suppose we are given a pair of smooth real functions ψ1 , ψ2
on R, chosen in such a way that they satisfy the Wronskian condition (7.16). Then it
is automatically true that the function p(ϕ) defined by
c ψ100 c ψ200
p≡ = (7.62)
6 ψ1 6 ψ2
is smooth for any constant c > 0. If in addition there exists a monodromy matrix M
such that (7.18) holds, then p(ϕ) is 2π-periodic and ψ1 , ψ2 are solutions of the corre-
sponding Hill’s equation. This procedure provides a way to build Virasoro coadjoint
vectors out of functions ψi ; in particular, in order to prove that there exist Virasoro
orbits with hyperbolic monodromy and non-zero winding number, it suffices to find
two normalized functions ψi satisfying these criteria, and the identification of the cor-
responding Virasoro coadjoint vectors will follow.
Thus, let ω > 0 be a strictly positive real number and let n > 0 be a positive
integer. Let us define the positive function
2
2 2ω
Fn,ω (ϕ) ≡ cos (nϕ/2) + sin(nϕ/2) + cos(nϕ/2) (7.63)
n
as well as
r
eωϕ 2 ω
ψ1 (ϕ) ≡ p sin(nϕ/2) + cos(nϕ/2) , (7.64)
Fn,ω (ϕ) n n
r
e−ωϕ 2
ψ2 (ϕ) ≡ p cos(nϕ/2). (7.65)
Fn,ω (ϕ) n
Since Fn,ω is strictly positive, the ψi ’s are smooth functions. They satisfy the Wron-
skian condition (7.16) and their monodromy matrix is (7.37). Their ratio is
ω
η(ϕ) = e2ωϕ tan(nϕ/2) +
n
and describes a path on the circle with varying velocity and winding number n. It
follows that the function p(ϕ) defined by (7.62) is a Virasoro coadjoint vector with
hyperbolic monodromy (7.37) and winding number n > 0. It is explicitly given by
c ω2 c n2 + 4ω 2 c n2
p(ϕ) = + − 2 (ϕ)
(7.66)
6 12 Fn,ω (ϕ) 8 Fn,ω
in terms of the function (7.63). We have thus built explicit orbit representatives with
hyperbolic monodromy and non-zero winding number.
It is worth spending some time to interpret formula (7.66). Let us take ω small
and expand p around ω = 0. To first order in ω, we get
n2 c nc
p(ϕ) = − + ω sin(nϕ) + O(ω 2 ) . (7.67)
24 3
The leading term in p is an exceptional constant −n2 c/24, so we can think of (7.66) as
a deformation of that constant. The term of order one in ω in (7.67) is proportional to
sin(nϕ), which is one of the elements of the Lie algebra of the stabilizer of −n2 c/24.
This ensures that the deformation does not belong to the orbit of −n2 c/24. Indeed,
all deformations that do belong to that orbit take the form
2
∗ n c (6.115) c
n2 X 0 + X 000

adX −
c = −
24 12
for some vector field X, where the term n2 X 0 + X 000 annihilates the contribution of the
modes sin(nϕ) or cos(nϕ).
In section 10.1 we will interpret (7.66) as the supermomentum of a BMS3 tachyon

with imaginary mass proportional to ω 2 . Accordingly, from now on we refer to hy-
perbolic Virasoro orbits with non-zero monodromy as tachyonic orbits. They are our
first example of orbits that do not admit any constant representative, so they are not
accounted for by fig. 7.1. In order to include them in our “map of coadjoint orbits”,
we think of them as orbits of deformations (7.67) of exceptional constants. With this
viewpoint and the “tachyonic” terminology, it is natural to identify this kind of defor-
mation with the horizontal line in fig. 4.3b that represents tachyonic orbits of Poincaré.
Accordingly we represent tachyonic Virasoro orbits by a horizontal line to the right of
the point labelled “n” in fig. 7.1. With this convention our schematic representation of
Virasoro orbits becomes the one displayed in fig. 7.2. It remains to understand which
orbit contains the point p0 = 0, and to find the remaining orbits that have no constant
representative. Before doing so, we address a few minor points regarding tachyonic
orbits:
• The construction that led from (7.63) to (7.66) did produce Virasoro coadjoint
vectors with the desired monodromy and winding number, but it is not clear at
Figure 7.2: A partial map of Virasoro orbits, including orbits of constant coadjoint
vectors together with tachyonic orbits. Compare to fig. 7.1.
this stage that any coadjoint vector satisfying these properties can be mapped
on (7.66). However, this turns out to be the case; in this sense, the orbit rep-
resentatives (7.66) exhaust all orbits with hyperbolic monodromy and non-zero
winding. See [166] for the proof.
• The Lie algebra of the stabilizer of (7.66) is spanned by the periodic linear com-
binations of the functions (7.23). As it turns out, the only periodic combination
in this case is the product ψ1 ψ2 . The latter has 2n simple zeros inside [0, 2π[
and generates a non-compact group R. In addition the function (7.66) is peri-
odic with period 2π/n, so the stabilizer must contain a group Zn consisting of
rotations by integer multiples of 2π/n. In fact, one can show (see [166]) that the
stabilizer of p in Diff+ (S 1 ) is isomorphic to a product R × Zn , while its stabilizer
g + (S 1 ) is R × T2π/n where T2π/n is the group of trans-
in the universal cover Diff
lations of the real line by integer multiples of 2π/n. We conclude that the orbit
of (7.66) is diffeomorphic to
W(p,c) ∼ g + (S 1 )/(R × T2π/n ) ∼

= Diff = Diff+ (S 1 )/(R × Zn ). (7.68)
7.2.6 Non-degenerate parabolic orbits

Here we include the last missing pieces of our description of Virasoro orbits. When
the monodromy matrix is non-degenerate parabolic, it is conjugate to one of the four
last elements in the list (7.34). As in the hyperbolic case we discuss zero and non-zero
windings separately.
Zero winding
At zero winding we proceed as in the elliptic and n = 0 hyperbolic cases, i.e. we
look for standard boosts. Let therefore (q, c) be a Virasoro coadjoint vector such that
a normalized solution vector Ψ = (ψ1 ψ2 )t associated with the corresponding Hill’s
equation has non-degenerate parabolic monodromy and zero winding number. The
monodromy matrices in (7.34) imply that

ψ1 (ϕ + 2π) = ± ψ1 (ϕ) + εψ2 (ϕ) , ψ2 (ϕ + 2π) = ±ψ2 (ϕ) (7.69)
where ε is a priori +1 or −1. The corresponding curve (7.27) satisfies
η(ϕ + 2π) = η(ϕ) + ε (7.70)
and (7.31) implies that ε must actually be equal to +1. The opposite sign corresponds
to changing the orientation in the space of solutions of Hill’s equation, so with our
choice of orientation for ψ1 , ψ2 , only the value ε = +1 gives rise to an admissible
monodromy matrix. Then the function
f (ϕ) ≡ 2π η(ϕ) (7.71)
is a 2πZ-equivariant diffeomorphism of the real line, and property (7.30) implies that
c
q(ϕ) = − S[f ](ϕ) .
12
As in eqs. (7.42) and (7.59), we recognize the coadjoint action of gq ≡ f −1 :
Proposition. Let (q, c) with c > 0 be a Virasoro coadjoint vector with non-dege-
nerate parabolic monodromy and vanishing winding number. Then it belongs to the
orbit of (0, c) and the inverse of the diffeomorphism (7.71) is a standard boost in the
sense of eq. (7.44).
Thus we have finally found the orbit of p0 = 0 ! It was the only point of fig. 7.1 that
was still eluding us. We now know that its orbit has parabolic type. The corresponding
stabilizer is the group U(1) of rigid rotations (as for all positive or generic constants
p0 ), and there are two conformally inequivalent solutions of Hill’s equation at p0 = 0,
namely ψ1± = ±ϕ, ψ2± (ϕ) = ±1. The orbit can be represented as a quotient space
W(0,c) ∼ g + (S 1 )/R ∼
= Diff = Diff+ (S 1 )/S 1
and is diffeomorphic to the orbits (7.52)-(7.61) of generic or positive constants.

Non-zero winding
At non-zero winding our strategy will be similar to that used in the hyperbolic case
with winding: we rely on the fact that formula (7.62) always defines a 2π-periodic func-
tion p(ϕ) when ψ1 and ψ2 satisfy the Wronskian condition and admit a well-defined
monodromy, which allows us to build orbit representatives.
Thus, pick a number ε ∈ {±1} and let n ∈ N∗ be a non-zero winding number. Let
us define the positive function
ε
Hn,ε (ϕ) ≡ 1 + sin2 (nϕ/2) (7.72)
2π
as well as

1 εϕ 2
ψ1 (ϕ) ≡ p sin(nϕ/2) − cos(nϕ/2) , (7.73)
Hn,ε (ϕ) 2π n
1
ψ2 (ϕ) ≡ p sin(nϕ/2) . (7.74)
Hn,ε (ϕ)
Since the function Hn,ε is strictly positive, the ψi ’s are smooth functions. They satisfy
the Wronskian condition (7.16) and their monodromy matrix is one of the four matrices
on the right in the list (7.34), with the off-diagonal entry coinciding with ε and the
overall ±1 = (−1)n . The curve (7.27) corresponding to this basis of solutions is
εϕ 2
η(ϕ) = − cot(nϕ/2)
2π n
and has winding number n. This is all as in the hyperbolic case below eq. (7.65). It
follows that the function p(ϕ) defined by (7.62) is a Virasoro coadjoint vector with
non-degenerate parabolic monodromy and winding number n > 0, explicitly given by
c n2 c n2 (1 + ε/2π)
p(ϕ) = − 2 (ϕ)
. (7.75)
12 Hn,ε (ϕ) 8 Hn,ε
As in the hyperbolic case, one can think of (7.75) as a deformation of a suitable

constant. However, in contrast ot (7.66), expression (7.75) seemingly contains no
continuous parameter that one could tune to “small” values since ε is only allowed to
take the values ±1. In order to solve this problem, recall from (7.36) that the matrices

1 ε 1 λε
and (7.76)
0 1 0 1
are conjugate in SL(2, R) for any positive real number λ. Accordingly we could just
as well have chosen the representatives of non-degenerate parabolic conjugacy classes
to involve an arbitrary positive parameter ε; the limit ε → 0 then may be taken since
it does not affect the conjugacy class of the monodromy matrix. The corresponding
coadjoint vector is (7.75) and its expansion to first order in ε reads
n2 c ε
p(ϕ) = − 1+ (1 + 2 cos ϕ) + O(ε2 ). (7.77)
24 2π
As in (7.67), the leading term is an exceptional constant (7.9) and we can think of
(7.77) as a deformation thereof. The deformation is designed so that it does not belong
to the orbit of −n2 c/24. When dealing with BMS3 supermomentum orbits in section
10.1, we will interpret (7.75) as the supermomentum of a massless BMS3 particle.
Accordingly, from now on we refer to non-degenerate parabolic orbits with non-zero
winding as massless orbits. Note that the statement that the matrices (7.76) are con-
jugate is tantamount to saying that massless orbits are scale-invariant.
To conclude our analysis we state (without proof) a few features of massless orbits:
• One can show that the orbit representatives (7.75) are exhaustive in that any
coadjoint vector belonging to a massless orbit can be brought in that form by a
suitable diffeomorphism. See appendix C of [166].
• The Lie algebra of the stabilizer of (7.75) is generated by the vector field X = ψ22 ,
which has n double zeros. In fact, as in the hyperbolic case, the stabilizer is
isomorphic to R × Zn , but the generator of the R part of that group is not
the same as in the hyperbolic case. The orbit is diffeomorphic to a quotient of
Diff+ (S 1 ) by this stabilizer, or equivalently a quotient of Diff
g + (S 1 ) by R × T2π/n
where T2π/n is the same discrete translation group as in (7.68).
g + (S 1 ) transformations, the solution (7.73)-(7.74) is the unique solution
• Up to Diff
of Hill’s equation with non-degenerate parabolic monodromy

n 1 ε
(−1) (7.78)
0 1
and winding number n.
7.2.7 Summary: a map of Virasoro orbits

The above analysis exhausts all coadjoint orbits of the Virasoro group. Since these
orbits will play a key role in the remainder of this thesis, we now briefly summarize
the salient features of the classification.
The schematic drawings of figs. 7.1 and 7.2 represent Virasoro orbits. The only
orbits which are not accounted for by these pictures are massless ones; in order to in-
clude them we use the same trick as in fig. 4.3b, where massless orbits are represented
by two dots near the origin (one with positive energy, the other with negative energy).
We will use the same notation here, except that such a pair of massless orbits occurs
for all positive integers n ∈ N∗ . With this convention, fig. 7.2 turns into the complete
map of Virasoro coadjoint orbits displayed in fig. 7.3.
Each point in that map represents an orbit representative; different points corre-
spond to different representatives and define disjoint orbits. All orbits are now ac-
counted for since the orbit representatives are exhaustive. The vertical line represents
orbits that contain a constant orbit representative:
Figure 7.3: The map of Virasoro coadjoint orbits at positive central charge. Note the
similarity with fig. 4.3b. Roughly speaking, the map consists of an infinity of copies
of Poincaré momentum orbits glued together and labelled by the winding number n.
Locally (near a node n), the two pictures look identical. This is not surprising given
that Poincaré momentum orbits in three dimensions coincide with SL(2, R) coadjoint
orbits, which in turn are classified similarly to the conjugacy classes of SL(2, R) that
were instrumental for Virasoro coadjoint orbits. This hints that there exists a relation
between Virasoro and Poincaré symmetry; we shall see in part III that this relation is
embodied by the BMS3 group.
• For generic p0 < 0 the orbit has elliptic monodromy determined by eq. (7.46).
Its winding number is given by (7.51), so the points of fig. 7.3 located between n
and n + 1 have winding number n (while points such that −c/24 < p0 < 0 have
zero winding number).
• For exceptional values p0 = −n2 c/24 with n ∈ N∗ , the orbit has degenerate
parabolic monodromy determined by (7.53). Its winding number is n. In partic-
ular, the orbit at n = 1 is the vacuum orbit.
• For p0 > 0, the orbit has hyperbolic monodromy with zero winding, and the
conjugacy class of the monodromy matrix is determined by (7.46).
• The orbit of p0 = 0 has non-degenerate parabolic monodromy with zero winding.

On the other hand, the points of fig. 7.3 that do not belong to the vertical axis represent
orbits that do not contain any constant representative:
• Each horizontal line starting at a point labelled by n represents a family of tachy-

onic orbits with winding number n. The orbit representatives are given by (7.66)
and involve a continuous parameter ω > 0 that determines the corresponding
monodromy matrix (7.37).
• Each pair of dots surrounding a tachyonic line at n represents the two massless
orbits with winding number n. The orbit representatives are given by (7.75)
and involve a discrete parameter ε = ±1 that determines the corresponding
monodromy matrix (7.78).
Focussing for definiteness on the multiply connected group Diff+ (S 1 ), the stabilizers
of Virasoro orbits are as follows:
Orbit Stabilizer
Vacuum-like p0 = −n2 c/24, n ≥ 1 PSL(n) (2, R)
Elliptic U(1)
Hyperbolic, zero winding U(1)
Non-degenerate parabolic, zero winding U(1)
Massless, winding n ≥ 1 R × Zn
Tachyonic, winding n ≥ 1 R × Zn
Table 7.1: Virasoro coadjoint orbits and their stabilizers.
In the universal cover of the Virasoro group the first four entries of the right column
would be replaced by their universal covers, while the two last ones would be replaced
by R×T2π/n . This should be compared with (and is very similar to) the list of Poincaré
little groups in table 4.1. Note that, at n = 1, the Virasoro stabilizers are quotients
by Z2 of their Poincaré counterparts. This is because table 4.1 lists the little groups
given by the double cover (4.93) of the Poincaré group.
Remark. Fig. 7.3 may be misleading since it suggests that all Virasoro orbits of
constant coadjoint vectors are of a similar type, which is clearly not the case since
orbits of constants p0 > 0 are hyperbolic while those of (generic) constants p0 < 0
are elliptic. In this sense, the map of orbits would have been more accurate if we had
represented the orbits of p0 > 0 by a horizontal line to suggest that they have the same
type of monodromy as the tachyonic orbits; see e.g. fig. 1 of [166]. Our convention
in fig. 7.3 is motivated instead by the fact that the value of p0 essentially measures
energy (see below), so that higher points in fig. 7.3 have higher energy.
7.3 Energy positivity

In this section we investigate the boundedness properties of an energy functional on Vi-
rasoro orbits. This question is motivated both by its use in two-dimensional conformal
field theory, and by its applications in three-dimensional gravity. We start by defining
the Virasoro energy functional, before showing that the Schwarzian derivative satisfies
an “average lemma” which will play a key role for this functional’s boundedness. We
then show that the only orbits with energy bounded from below are either orbits of
constants p0 ≥ −c/24, or the massless orbit at winding n = 1 and monodromy ε = −1.
To reduce clutter we return to our earlier abusive notation by writing as Diff(S 1 ) the
universal cover of the group of orientation-preserving diffeomorphisms of the circle.
Relevant references include [56, 166] as usual.
7.3.1 Energy functional

The group Diff(S 1 ) can be interpreted as (part of) the symmetry group of a two-
dimensional conformal field theory. In that context the quadratic density p(ϕ)dϕ2 is
(a component of) the stress tensor of the theory, and its zero-mode
Z 2π
1
E[p] ≡ dϕ p(ϕ) (7.79)
2π 0
is the associated energy. We shall refer to this quantity as the Virasoro energy func-
tional evaluated at p. If the theory admits a configuration whose stress tensor is p(ϕ),
then consistency with conformal symmetry requires that it also admits configurations
with stress tensor f · p, where f ∈ Diff(S 1 ) and the dot denotes the coadjoint action
(6.114) for some definite value of the central charge. The energy functional varies
under conformal transformations, since
Z 2π
1 dϕ h c i
E[f · p] = p(ϕ) + S[f ](ϕ) (7.80)
2π 0 f 0 (ϕ) 12
generally differs from (7.79).
Now consider a CFT with central charge c > 0 and let Ω be the space of its
stress tensors p(ϕ); in general Ω is a certain subset of the space F2 (S 1 ) of quadratic
densities. Since any quantum system with a well-defined vacuum is expected to have
energy bounded from below, the map
Ω → R : p 7→ E[p] (7.81)
should be bounded from below. In addition, consistency with conformal symmetry
implies that Ω is a union of Virasoro coadjoint orbits. One is thus led to the following
question:
Which of the Virasoro coadjoint orbits of fig. 7.3 have
(7.82)
energy bounded from below under conformal transformations?
In the sequel we will refer to orbits with energy bounded from below as orbits “with
positive energy”, although their energy (7.79) may actually be negative for some field
configurations p(ϕ).
Note that all orbits have energy unbounded from above. Indeed the term involving
the Schwarzian derivative in (7.80) can be written as
Z 2π Z 2π
c dϕ c
S[f ](ϕ) = − dϕ S[f −1 ](ϕ) (7.83)
24π 0 f 0 (ϕ) 24π 0
where we have renamed the integration variable from ϕ to f −1 (ϕ), then used (6.16)
and the cocycle identity (6.77). Since the Schwarzian derivative can be written as
00 0 2
f 1 f 00
S[f ](ϕ) = − , (7.84)
f0 2 f0
we can also recast (7.83) in the form
Z 2π −1 00 2
c (f )
dϕ .
48π 0 (f −1 )0
This can be made arbitrarily large for suitable choices of f , which proves that the
energy functional E is unbounded from above on any Virasoro orbit.
7.3.2 The average lemma

As a first step towards the answer of the question (7.82), we focus on the piece of eq.
(7.80) that involves the Schwarzian derivative. The result that we shall describe was
first derived in [169] and was based on projective geometry (see also [170, 171]). The
elementary proof given here is borrowed from [166].
Average lemma. Let f ∈ Diff g + (S 1 ) and let S[f ](ϕ) be its Schwarzian derivative
(6.76) at ϕ. Then the average of the Schwarzian derivative satisfies the inequality
Z 2π Z 2π
1
dϕ 1 − (f 0 (ϕ))2 ,

dϕ S[f ](ϕ) ≤ (7.85)
0 0 2
with equality if and only if f (ϕ) is a projective transformation of the form (6.88).
Proof. We consider the functional
Z 2π
1 0 2
I[f ] ≡ − dϕ (f (ϕ)) + S[f ](ϕ) . (7.86)
0 2
Our goal is to show that this quantity is bounded from below and that its minimum
value is −π. By (7.84), it only depends on f 0 and f 00 . A convenient way to express
this dependence is to define
(6.16) 1
Y (ϕ) ≡ f 0 (f −1 (ϕ)) = . (7.87)
(f −1 )0 (ϕ)
Since f is a 2πZ-equivariant, orientation-preserving diffeomorphism, Y (ϕ) is strictly
positive and 2π-periodic. In terms of Y we can rewrite (7.86) as
1 2π
0
(Y (ϕ))2
Z
I[Y ] = dϕ − Y (ϕ) (7.88)
2 0 Y (ϕ)
where the integrand is well-defined since Y > 0. Let us denote the minimum and
maximum of Y (ϕ) by
m ≡ min Y (ϕ) , M ≡ max Y (ϕ) . (7.89)

ϕ∈[0,2π] ϕ∈[0,2π]
With this notation the function

" 2 2 #
mM 1 M −m M +m
m + M − Y (ϕ) − = − Y (ϕ) −
Y (ϕ) Y (ϕ) 2 2
is non-negative and vanishes only at the points where Y reaches its minimum or its
maximum. Now consider the obvious inequality
r !2
|Y 0 | mM
√ − m+M −Y − ≥ 0. (7.90)
Y Y
Integrating this over the circle and using (7.88), we obtain

Z 2π r
|Y 0 | M − m 2 M + m 2
I[Y ] ≥ −π(m + M − mM ) + dϕ − Y − . (7.91)
0 Y 2 2
If there was no absolute value in the integrand on the right-hand side, we could just
change the integration variable from ϕ to Y using dϕ Y 0 (ϕ) = dY ; the absolute value
prevents us from doing this globally, but we can do it locally between two consecutive
extrema of the function Y (ϕ) (since the sign of Y 0 is constant in such an interval). We
can then express the right-hand side of (7.91) in terms of the primitive function of the
integrand,
s 2 2 Z Y
Z Y
dz M −m M +m
F(Y ) = − z− ≡ dz G(z), (7.92)
m z 2 2 m
where we have introduced the function G(z) to reduce clutter below. To see the
use of this, consider a function Y (ϕ) of the following shape (the general case follows
straightforwardly):
Figure 7.4: The function Y (ϕ) is 2π-periodic and strictly positive. Here we choose
it with four local extrema, the global minimum being Y (ϕ1 ) = m and the global
maximum Y (ϕ4 ) = M .
This function has two local minima at ϕ1 and ϕ3 and two local maxima at ϕ2
and ϕ4 (the numbers of local minima and maxima coincide since Y (ϕ) is smooth and
2π-periodic). Then the integral in (7.91) can be written as

s 2 2
Z 2π
|Y 0 | M −m M +m
dϕ − Y − =
0 Y 2 2
Z Y2 Z Y3 Z Y4 Z Y1
= dY G(Y ) − dY G(Y ) + dY G(Y ) − dY G(Y )
Y1 Y2 Y3 Y4
(7.92)
= 2 [F(Y2 ) + F(Y4 ) − F(Y1 ) − F(Y3 )]
with the shorthand notation Y (ϕi ) ≡ Yi . The same computations would work for
arbitrarily many minima and maxima of Y (ϕ), with the same result: the integral is
twice the sum of F’s evaluated at the maxima minus twice the sum of F’s evaluated
at the minima. Thus the inequality (7.91) can be written as
(7.89)
I[Y ] ≥ −π(m + M − mM ) + 2[F(M ) − F(m)] + 2[F(Y2 ) − F(Y3 )]
≥ −π(m + M − mM ) + 2F(M ) (7.93)
where we also used the fact that F(m) = 0 by virtue of the definition (7.92). Now it
√ √ 2
turns out that F(M ) = π2 M − m , which allows us to rewrite (7.93) as
√ √ 2
I[Y ] ≥ −π(m + M − mM ) + π M − m ≥ −π. (7.94)
We conclude that I[Y ] is bounded from below by the value −π, which is exactly the
inequality (7.85). It only remains to find the conditions under which (7.85) becomes
an equality. For this to be the case, the inequalities (7.90), (7.93) and (7.94) must all
be saturated; this occurs when Y (ϕ) satisfies the following three conditions:
• In order to saturate (7.90), it satisfies the differential equation
Y 02 = (m + M )Y − Y 2 − mM. (7.95)
• In order to saturate (7.93), Y (ϕ) has only one minimum and one maximum,
where it takes the values m and M , respectively.
• In order to saturate the second inequality of (7.94), M = 1/m.
To solve (7.95) we use (7.87) and rewrite the equation in terms of f −1 . Using M = 1/m
the derivative of (7.95) becomes
Y 0 1 − ((f −1 )0 )2 − 2 S[f −1 ] = 0,

(7.96)
which is equivalent to (6.94). We have shown below (7.54) that the only f ’s satisfying
this property are those that belong to the group of projective transformations (6.88),
which concludes the proof.
7.3.3 Orbits with constant representatives

The average lemma allows us to investigate the boundedness properties of the energy
functional (7.79) on Virasoro orbits. For now we limit ourselves to orbits that admit
a constant representative.
Proposition. The vacuum orbit, containing the point pvac = −c/24, has energy
bounded from below:
c
E[f · pvac ] ≥ E[pvac ] = − . (7.97)
24
The minimum of energy is located at pvac .
Proof. We consider formula (7.80) with p(ϕ) = pvac = −c/24. Renaming the integra-
tion variable from ϕ to f −1 (ϕ) and using eqs. (6.16) and (6.77), we find
Z 2π
c 1 −1 0
2 −1
E[f · pvac ] = dϕ − (f ) (ϕ) − S[f ](ϕ)
24π 0 2
which we recognize as the functional (7.86) evaluated at f −1 . The average lemma (7.85)
then implies that E[f · pvac ] ≥ −c/24, with equality if and only if f is a projective
transformation (6.88). Our earlier result (7.55) ensures that such transformations
precisely span the stabilizer of pvac , so the minimum of energy is reached at pvac .
Let us turn to other orbits containing a constant representative p(ϕ) = p0 . The
key will be to rewrite their energy functional as the vacuum energy functional, plus
another term. Starting from formula (7.80) we obtain
p0 + c/24 2π dϕ
Z
E[f · p0 ] = + E[f · pvac ] (7.98)
2π 0 f 0 (ϕ)
where the integral of 1/f 0 can be rewritten as

Z 2π Z 2π
1 dϕ (6.16) 1 2
dϕ (f −1 )0 (ϕ) − 1

0
= 1+
2π 0 f (ϕ) 2π 0
R 2π
as follows from 0 dϕ(f −1 )0 (ϕ) = 2π. Plugging this into (7.98) and using (7.97), we
obtain
p0 + c/24 2π
Z
2
dϕ (f −1 )0 (ϕ) − 1 .

E[f · p0 ] ≥ p0 +
2π 0
The right-hand side here is the sum of p0 and an integral whose integrand is manifestly
non-negative. This implies the following result:
Proposition. If p0 ≥ −c/24, then the orbit of (p0 , c) has energy bounded from
below, with the energy minimum located at p0 :
c
p0 ≥ − ⇒ E[f · p0 ] ≥ E[p0 ] = p0 .
24
Figure 7.5: Schematic representation of the Virasoro orbit of a constant p0 located

above the vacuum value −c/24, here understood to be the origin of the coordinate
system. The coordinates qm , m ∈ Z are the Fourier modes of coadjoint vectors q(ϕ);
in particular the zero-mode q0 = E[q] is their energy, which is bounded from below on
the orbit. Compare to the massive Poincaré orbit with positive energy in fig. 4.3a.
Now what happens when p0 is lower than −c/24? In that case energy is un-
bounded, as can be shown by finding a family of diffeomorphisms that lower the energy
indefinitely. Indeed, consider the matrix

cosh(γ/2) sinh(γ/2)
∈ SL(2, R) (7.99)
sinh(γ/2) cosh(γ/2)
where γ ∈ R (the normalization is chosen for later convenience). The corresponding

projective transformation (6.88) is
eiϕ cosh(γ/2) + sinh(γ/2)

eif (ϕ) = , (7.100)
−eiϕ sinh(γ/2) + cosh(γ/2)
and one verifies that

1
0
= |eiϕ cosh(γ/2) + sinh(γ/2)|2 = cosh γ + sinh γ cos ϕ . (7.101)
f (ϕ)
The Schwarzian derivative of f is given by (6.94), so we find that

Z 2π
(7.80) p0 + c/24 dϕ c
E[f · p0 ] = 0
−
2π 0 f (ϕ) 24
where we have used the fact that the integral of f 0 over S 1 is normalized to 2π. The
integral of (7.101) then yields
c
E[f · p0 ] = (p0 + c/24) cosh γ − , (7.102)
24
and this can become arbitrarily negative when p0 < −c/24. In conclusion:
The coadjoint orbit W(p0 ,c) of a constant p0

(7.103)
has energy bounded from below if and only if p0 ≥ −c/24.
Thus, when p0 < −c/24, fig. 7.5 is no longer valid because the energy functional can
reach arbitrarily low values in certain directions. The orbit then looks like an infinite-
dimensional saddle instead of the hyperboloid represented in fig. 7.5.
Note that the matrix (7.99) can be interpreted as the SL(2, R) group element
that represents a Lorentz boost with rapidity γ in three dimensions3 thanks to the
isomorphism (4.83), which also explains our choice of normalization. In that context,
formula (7.102) is the transformation law of the energy of a particle with mass p0 +c/24
under Lorentz boosts. We will return to this interpretation in part III.
7.3.4 Orbits without constant representatives

We now describe the boundedness properties of the energy functional on Virasoro
coadjoint orbits that do not admit a constant representative. As it turns out there is
only one orbit with energy bounded from below, while all other ones have unbounded
energy.
Consider the non-degenerate parabolic orbit with winding number n = 1 and

monodromy ε = −1; a typical orbit representative is given by (7.75). One can then
prove the following result:
Proposition. The energy functional on the massless orbit specified by n = 1 and

ε = −1 is bounded from below by −c/24. There exist infinitely many points on the
orbit whose energy is arbitrarily close to that value, but there is no orbit representative
that realizes this value of energy.
We will not prove this proposition here and refer instead to [166]. Roughly speak-
ing, the proof follows from a construction very similar to the one used in the proof
of the average lemma (7.85), except that it crucially relies on the parameters n = 1,
ε = −1. The proof of the fact that the infimum of energy is never reached on the
orbit follows from the construction of a one-parameter family of points belonging to
the orbit in such a way that they converge to the constant −c/24 without ever quite
reaching it.
3
Rapidity is related to velocity v by γ = arctanh(v).
Figure 7.6: Schematic representation of the Virasoro orbit with non-degenerate

parabolic monodromy ε = −1 and winding number n = 1. The origin of the co-
ordinate system is the vacuum coadjoint vector, qm = −(c/24)δm0 , which does not
belong to the orbit. Energy is bounded from below on the orbit but its infimum is
never quite reached, in contrast to fig. 7.5. Compare to the massless Poincaré orbit
with positive energy in fig. 4.3a.
One might think that the other orbits without constant representatives behave
in a similar way, i.e. that they also have energy bounded from below. However, for
any such orbit, it is possible to build a one-parameter family of orbit elements whose
energy can be arbitrarily low, similarly to constant representatives p0 < −c/24. We
refer again to [166] for explicit constructions. Thus one concludes that
all tachyonic or massless Virasoro orbits have unbounded energy,

except the one with non-degenerate parabolic monodromy (7.104)
ε = −1 and winding n = 1.
7.3.5 Summary: a new map of Virasoro orbits

The considerations of the last few pages allow us to include more information in
the map of Virasoro orbits of fig. 7.3; see fig. 7.7. Its two striking features are the
occurrence of a single orbit without constant representatives and positive energy, and
the fact that the lowest-lying orbit with positive energy is that of pvac = −c/24. This
observation justifies referring to the latter orbit as the “vacuum orbit” and to pvac as
the “vacuum stress tensor”. Note that the exact same situation occurs with relativistic
particles, as the only ones with energy bounded from below are either massive (with
non-negative mass) or massless.
Figure 7.7: The map of Virasoro coadjoint orbits at positive central charge. Orbits
with energy bounded from below are coloured in red. Those are orbits of constants
p0 ≥ −c/24, plus the unique massless orbit with monodromy (7.78) such that ε = −1
and winding number n = 1. All other orbits have energy unbounded from below.
Chapter 8
Symmetries of gravity in AdS3
In this chapter we explore a physical model where the Virasoro group plays a key role,
namely three-dimensional gravity on Anti-de Sitter (AdS) backgrounds and its puta-
tive dual two-dimensional conformal field theory (CFT). These considerations will be
a basis and a guide for our study of asymptotically flat space-times in part III.
The plan is the following. Section 8.1 is a prelude where we recall a few basic
facts about (three-dimensional) gravity, in particular regarding the notion of asymp-
totic symmetries. Section 8.2 is then devoted to three-dimensional space-times whose
metric approaches that of Anti-de Sitter space at spatial infinity; this includes Brown-
Henneaux boundary conditions and their asymptotic symmetries, which will turn out
to consist of two copies of the Virasoro group. In section 8.3 we describe the phase
space of AdS3 gravity as a hyperplane at fixed central charges in the space of the
coadjoint representation of two Virasoro groups. Finally, in section 8.4 we describe
unitary highest-weight representations of the Virasoro algebra and relate them to the
quantization of the AdS3 phase space.
Bibliographical remarks. This chapter is based on several combined references.

Perhaps the most important one is the original paper by Brown and Henneaux [14],
which triggered the development of the field as a whole. In that paper the authors
relied on the methods of [172–175] to build surface charges associated with asymp-
totic symmetries, but our approach will be led by their Lagrangian (or “covariant”)
reformulation [176–178]. In particular, our presentation of Brown-Henneaux bound-
ary conditions and of the associated asymptotic Killing vector fields follows [6]. The
general solution of the equations of motion first appeared in [179, 180]. It contains
in particular the BTZ black hole, which was discovered and studied in [16, 17]. Fi-
nally, the group-theoretic approach to the gravitational phase space first appeared
in [45, 46, 181], which is also where the AdS3 positive energy theorem was derived.
(See also [182–184] for earlier related considerations.)
8.1 Generalities on three-dimensional gravity

Here we recall a few basic facts about classical general relativity in three dimensions.
We start by explaining that three-dimensional Einstein gravity has no local degrees of
217
Chapter 8. Symmetries of gravity in AdS3 218
freedom, then turn to a discussion of boundary conditions and the ensuing boundary
terms that one adds to the action in order to make the variational principle well-
defined. This finally leads to the concept of asymptotic symmetries and the important
observation that the Poisson brackets of surface charges that generate these symmetries
generally contain central extensions.
8.1.1 Einstein gravity in three dimensions

We consider an orientable three-dimensional space-time manifold M endowed with
coordinates xµ (µ = 0, 1, 2) on which we put a metric gµν with signature (− + +).
The equations of motion are determined by the Einstein-Hilbert action,
√
Z
1
SEH [gµν , Φ] = d3 x −g (R − 2Λ) . (8.1)
16πG M
Here G is the Newton constant in three dimensions, R is the Ricci scalar associated
with gµν and Λ ∈ R is a cosmological constant. In three dimensions, and using units
such that c = ~ = 1, Newton’s constant G is a length scale. Equivalently 1/G is an
energy scale that coincides with the Planck mass.
Upon varying the action (8.1) and neglecting all boundary terms (which we shall
talk about later), one obtains the vacuum Einstein’s equations with a cosmological
constant:
1
Rµν − Rgµν + Λgµν = 0 , i.e. Rµν = 2Λgµν . (8.2)
2
What is special about three-dimensional manifolds is that their Ricci curvature wholly
determines their Riemann tensor independently of the equations of motion:
1
Rλµνρ = gλν Rµρ − gλρ Rµν − gµν Rλρ + gµρ Rλν − R(gλν gµρ − gλρ gµν ) . (8.3)
2
Then the Einstein equations (8.2) imply that, at each point of space-time, the on-shell
Riemann tensor is that of a maximally symmetric manifold with curvature determined
by the cosmological constant:
Rλµνρ = Λ(gλν gµρ − gλρ gµν ) . (8.4)
In other words, any solution of Einstein’s equations in three dimensions is locally

isometric to three-dimensional de Sitter, Minkowski or Anti-de Sitter space depending
on whether Λ is positive, vanishing or negative respectively. This is strikingly different
from higher-dimensional general relativity and relies on the relation (8.3) expressing
Riemann in terms of Ricci, valid only in two and three dimensions.1 In technical terms
it is the statement that
there are no local degrees of freedom in three-dimensional Einstein gravity.
It follows in particular that there are no gravitational waves, hence no gravitons.

Equivalently, all configurations of the metric are locally gauge-equivalent to empty
1
In two dimensions one has in addition Rµν = 21 Rgµν .
space. Importantly, this is not to say that the only solution of three-dimensional
gravity is empty space. For example, any quotient of Minkowski space by some dis-
crete group solves Einstein’s equations, but is not globally isometric to Minkowski.
Thus global aspects are essential: even though all solutions of Einstein’s equations
are locally isometric, they are generally not globally isometric and therefore represent
physically distinct field configurations. In this sense the absence of local degrees of
freedom in three-dimensional gravity does not prevent the overall absence of degrees
of freedom: it only means that the actual, physical degrees of freedom of the theory
cannot be captured by a local analysis, but require instead a global one, taking into
account topological properties of the space-time manifold. Field theories of this type,
having no local degrees of freedom but still globally non-trivial, are called topological
field theories.
Note that the absence of local degrees of freedom is confirmed by the Hamiltonian
formalism [185]: picking a time direction in M, one can split the metric field into a
lapse N , a shift N i and a spatial metric gij with conjugate momenta π ij , the indices
i, j ∈ {1, 2} labelling spatial directions. The lapse and shift play the role of Lagrange
multipliers enforcing the constraints that generate reparameterizations of time and
spatial diffeomorphisms, respectively. One thus obtains three dynamical Lagrange
variables gij with three conjugate momenta π ij , subject to three first-class constraints.
These constraints can be solved by choosing three gauge-fixing conditions (this is
the statement that “first-class constraints count twice”), which reduces the number of
physical degrees of freedom of three-dimensional Einstein gravity to 21 (3×2−3−3) = 0,
as expected.
Remark. Since three-dimensional Einstein gravity has no local degrees of freedom,

it is an unrealistic model of the world (where gravitational waves do exist [186]). This
motivates the construction of alternative theories of three-dimensional gravity that do
contain local degrees of freedom, such as topologically massive gravity [187] or new
massive gravity [188]. In this thesis we shall be concerned only with Einstein gravity,
although many of our considerations also apply to such modified theories.
8.1.2 Boundary conditions and boundary terms

A field theory, as a Hamiltonian system, is defined by (i) its field content and Poisson
brackets, and (ii) boundary conditions on fields and momenta. The second point is
crucial for gauge theories such as gravity. Here we explain certain generalities on
boundary conditions, leaving specific definitions in three-dimensional gravity for later.
In general terms, given a set of fields living on a manifold M, one chooses coordinates
(r, x) on M and calls “infinity” the region where r goes to infinity while all other
coordinates are kept finite. One then specifies certain fall-off conditions for fields and
their derivatives on that region, typically of the form
Φ(r, x) = O(r# ) as r → +∞
where Φ is some field and the coefficient # depends on the choice of fall-off conditions.
In writing this it is understood that ∂r Φ is of order O(r#−1 ) at infinity.2
The influence of fall-offs is visible at the level of the action principle. Indeed, it
is understood that the action of the theory should be plugged in an exponential eiS ,
which is then to be integrated over field configurations in a path integral so as to
produce quantum-mechanical transition amplitudes. In the classical limit, the leading
contribution to the path integral should be due to on-shell field configurations; but
for this to be true the integrand must be differentiable, which is to say that the
functional derivative δS/δΦ(x) is a local quantity. This, in turn, is only true provided
the variation of the action contains no boundary terms. For instance, the variation of
the Einstein-Hilbert action (8.1) is given by
√
Z
1 1
δSEH = d3 x −g Rµν − Rgµν + Λgµν δg µν
16πG M 2
(8.5)
√ √
Z
1 3 µν α µα λ

+ d x ∂α −gg δΓµν − −gg δΓλµ .
16πG M
The first term of this expression is the integral of the variation of the metric multiplying
the vacuum Einstein equations, as expected. The second term is the integral of a total
divergence and is therefore equal, by Stokes’ theorem, to the flux of a vector field
through the boundary ∂M of M. Depending on one’s choice of fall-off conditions for
the metric, this boundary term may or may not vanish. If it does vanish, then the
pure bulk action (8.1) can be legally plugged into a path integral. If it does not, then
(8.1) is not differentiable and cannot be inserted as such in a path integral, which is
to say that the semi-classical limit of a path integral involving only the action (8.1)
is not given by on-shell field configurations. Accordingly, in order for the theory to
have a well-defined semi-classical limit given by the equations of motion (8.2), one is
generally forced to modify the pure bulk action (8.1) as
Z
S[gµν ] = SEH [gµν ] + d2 x L(gµν , ∂gµν , ...) . (8.6)
∂M
Here L is a certain Lagrangian density on the boundary of M, chosen so as to cancel

the possibly non-vanishing boundary terms coming from the variation (8.5). Provided
one can find a suitable L, the variation of the improved action (8.6) only involves the
first term of (8.5) and the theory is classically consistent.
This explains, in terms of the action, how boundary conditions affect the definition
of the theory. A few remarks are in order:
• We have been sloppy in our discussion of the notion of “boundary”. Indeed we

claimed that the fields of our theory live on a manifold M and called ∂M its
boundary, which we identified with the region r → +∞ in terms of some radial
coordinate r. But typical space-time manifolds (such as R3 ) actually have no
boundary in the strict sense, so we should have been more precise: when we say
2
This is not a trivial requirement; for instance the function sin(r42 )/r is of order O(1/r) as
r → +∞ but its derivative is not of order O(1/r2 ).
that the region r → +∞ is the boundary ∂M of M, we really mean that we

complete M into a larger manifold, say M, which now has a boundary, and in
terms of the original coordinate r that boundary is located at r = +∞. This
completed manifold M is known as a conformal compactification of M [189].
• Aside from fall-off conditions, there is a second reason for adding boundary terms
to the Einstein-Hilbert action. Namely, the Ricci scalar contains second-order
derivatives of the metric, so in order to insert legally the gravity action in a path
integral when the metric satisfies Dirichlet boundary conditions, boundary terms
must be added to the Einstein-Hilbert action to cancel these second-order terms.
This is the origin of the Gibbons-Hawking-York boundary term [190, 191].
• The discussion of boundary terms clarifies in which sense a theory having no
local degrees of freedom can still have non-trivial topological degrees of freedom:
even though the bulk dynamics is trivial, that of the boundary is highly non-
trivial! In particular topological field theories, such as three-dimensional gravity,
are the simplest examples of holographic systems since all their physical degrees
of freedom live on the boundary of space-time. In higher space-time dimensions,
the discussion of boundary terms remains the same but it is complicated by the
presence of local, bulk degrees of freedom.
• Three-dimensional Einstein gravity can be reformulated as a Chern-Simons the-
ory whose gauge group is determined by the sign of the cosmological constant
[192, 193] (see also [194–196]). This allows one to rewrite the Einstein-Hilbert
action (plus boundary terms) as a purely two-dimensional action describing a
field theory on the boundary of space-time, as follows from the relation between
Chern-Simons theory, Wess-Zumino-Witten models and Liouville theory. It is
often referred to as “dimensional reduction”, and was first worked out in [197]
for Brown-Henneaux boundary conditions, while flat boundary conditions were
studied in [198].
8.1.3 Asymptotic symmetries

Having justified the necessity of boundary terms for field theories, we now turn to
gauge theories and explain qualitatively how one can find their global symmetries.
These symmetries turn out to depend in a crucial way on the choice of fall-off con-
ditions. We will first argue that the conserved charges associated with rigid global
symmetries are strikingly different from those of gauge theories, then describe the en-
suing notion of asymptotic symmetries. We conclude with the observation that the
canonical generators of these symmetries generally satisfy a centrally extended Poisson
algebra.
The problem of gauge symmetries

Suppose we are given some gauge-invariant field theory living on a manifold M, with
some bulk action S[Φ]. The system has gauge redundancies, i.e. gauge symmetries,
and one expects that there exist corresponding conserved quantities. The question is:
how to build such conserved charges? To answer this we follow [176].
A naive guess is to simply apply the Noether procedure. For a field theory which
is left invariant by certain symmetry transformations generated by some parameters
a , a = 1, ..., N , with field and space-time transformations of the general form
x 7→ x + δ x, Φ 7→ Φ + δ Φ ,
the N Noether currents jaµ can be obtained by “gauging” the symmetry, that is, replac-
ing the rigid parameters a by arbitrary functions a (x) on space-time. The variation
of the action then takes the form
Z
δS = − dD x jaµ ∂µ a (8.7)
M
from which one can read off the definition of the currents jaµ . Their conservation follows
from the fact that δS ≈ 0 on-shell, and the corresponding conserved Noether charges
are the fluxes of these currents through a space-like slice Σ of space-time:
Z
Qa = (dD−1 x)µ jaµ , (8.8)
Σ
where (dD−1 x)µ is proportional to µα1 ...αD−1 dxα1 ...dxαD−1 . Equivalently, (dD−1 x)µ ∝
dD−1 x · nµ where nµ is the future-pointing time-like unit vector field orthogonal to Σ,
and indices are moved thanks to the space-time metric.
Figure 8.1: A space-time manifold M with an embedded space-like slice Σ and future-
pointing time-like normal vector n.
The problem with gauge symmetries now becomes apparent. Indeed, in that case
the symmetry parameters a are already gauged, which is to say that the right-hand
side of (8.7) vanishes. This in turn implies that the Noether currents associated with
gauge transformations all vanish! In particular, there seems to be no way of defining
conserved charges of the form (8.8) for a gauge symmetry; this problem is the key
difference between gauge symmetries and rigid symmetries.
The solution is provided by the following observation: the Noether current defined
by (8.7) is not unique, as one can add to it the divergence of a two-form without
affecting the left-hand side. In other words eq. (8.7) does not specify the Noether
current jaµ uniquely, since the modified current j̃aµ = jaµ + ∂ν kaµν , where kaµν = −kaνµ ,
satisfies the same property provided the antisymmetric tensor k falls off fast enough at
infinity. The corresponding Noether charge (8.8) is left unaffected by this modification
provided the integral of k on the boundary of Σ vanishes; if that integral does not
vanish, however, the charge receives an additional surface contribution of the form
Z
Qsurface = (dD−2 x)µν k µν (8.9)
∂Σ
where (dD−2 x)µν is proportional to µνα1 ...αD−2 dxα1 ...dxαD−2 . As we have just argued,
the would-be Noether charges of a gauge theory can only receive surface contributions
such as (8.9) since the corresponding Noether current vanishes up to the divergence
of a two-form.
At first sight this means that the situation is even worse than expected, since the
Noether charges of gauge theories are apparently ill-defined: there is no a priori way
to associate a k µν with a given symmetry generator, so the surface integral (8.9) can
take any value. But in fact, this also suggests a solution to the problem: instead of
trying to build a conserved current j µ , one can associate, with a gauge symmetry, a
(D − 2)-form k µν and define the corresponding charge by (8.9). If k µν is conserved on-
shell in the sense that ∇µ k µν ≈ 0, then the corresponding charge (8.9) is conserved by
time evolution. In that context, the field k µν is called a superpotential and its integral
(8.9) over the boundary of Σ is known as the associated surface charge.3 For example,
in electrodynamics, the superpotential coincides with the strength tensor F µν and the
corresponding surface charge is the flux of the electric field at infinity, that is, the total
electric charge. Its conservation follows from the fact that ∂µ F µν vanishes on-shell by
virtue of Maxwell’s equations.
Thus the computation of conserved charges for gauge symmetries boils down to the
problem of associating a conserved superpotential with a given gauge transformation,
and understanding to what extent that superpotential is unique.
Asymptotic symmetries
While the definition (8.7) of the Noether current associated with a global symmetry
transformation is straightforward, that of the superpotential associated with a gauge
transformation is much more involved; see e.g. [176,178,199]. Here we simply summa-
rize the main ideas so as to apply them later to the specific case of three-dimensional
gravity. The contruction consists of several steps:
1. Define the theory by choosing a bulk action, imposing certain fall-off conditions
on the field content, and possibly adding a boundary term to the bulk action
such that the full action is differentiable.
2. Find, among all possible gauge transformations, those that preserve the fall-off
conditions. Such gauge transformations are said to be allowed, as opposed
to the gauge transformations that spoil the fall-off conditions and are therefore
3
The term “superpotential” here has nothing to do with supersymmetry.
“forbidden”. Allowed gauge transformations should then be thought of as the

symmetries (global or gauge) of the theory.
3. Associate, with each allowed gauge transformation, a conserved superpotential

k µν ; the latter depends linearly on the gauge parameters, while its dependence
on the field content depends on the model under study. We will not write down
that dependence explicitly here and refer to [176, 199] for details.
4. For each superpotential k µν , define a surface charge Q by (8.9). If all surface

charges associated with allowed gauge transformations are finite, then the bound-
ary conditions are consistent. The allowed gauge transformations whose surface
charges vanish are said to be trivial, while those whose surface charges do not
vanish are non-trivial.
This construction provides a distinction between three families of gauge transfor-

mations — forbidden, allowed and trivial — and is illustrated in fig. 8.2. It is not
just a matter of terminology; different classes of gauge transformations truly represent
physically distinct notions of symmetries:
• Trivial gauge transformations are genuine (allowed) gauge transformations, that

is, redundancies in the description of the theory.
• Non-trivial gauge transformations are global symmetries that map a field config-
uration on a physically different one. They fall off at infinity much slower than
trivial gauge transformations and change the state of the system when acting on
it. For example, in electrodynamics, non-trivial gauge transformations at spatial
infinity take the form δAµ (x) = ∂µ (x) with (x) = const.4 This corresponds to
a global U(1) symmetry and the associated charge is the electric charge.
• Forbidden gauge transformations are neither gauge transformations, nor even

global symmetries: they are literally excluded from the theory since they do not
leave its phase space invariant.
Note that infinitesimal gauge transformations are always endowed with a Lie bra-
cket. Accordingly, they span a Lie algebra. The notions introduced above then lead
to the following terminology:
Definition. The asymptotic symmetry algebra of a theory is the quotient of the alge-
bra of allowed gauge transformations by its ideal consisting of trivial transformations.
In the context of gravity, gauge transformations are diffeomorphisms of the space-

time manifold, generated by certain vector fields. Allowed gauge transformations
are generated by so-called asymptotic Killing vector fields. Their Lie bracket is the
standard Lie bracket of vector fields and the asymptotic symmetry algebra coincides
with the global symmetry algebra of the putative dual theory. In section 8.2 we will
illustrate these notions in the case of AdS3 gravity with Brown-Henneaux boundary
conditions, while section 9.1 will be devoted to their asymptotically flat analogue.
4
In practice, for constant this gives δAµ = 0, but for fields with non-zero electric charge the
transformation given by constant ’s is non-trivial.
Figure 8.2: Gauge transformations fall in three classes: forbidden transformations are
those that do not preserve the fall-off conditions of the theory; allowed transformations
are those that do, although they generally change the state of the system; trivial
transformations are those that preserve the fall-off conditions and leave the state of
the system unchanged. In this sense trivial gauge transformations are actual gauge
redundancies, and the global symmetry group of the system is the quotient of the
group of allowed transformations by its subgroup of trivial gauge transformations.
Central extensions in the surface charge algebra

The surface charges associated with asymptotic symmetries are designed in such a way
that they implement asymptotic symmetry transformations on the fields of the theory.
Explicitly, if we call ξ some infinitesimal gauge parameter generating an allowed non-
trivial gauge transformation and if we denote the associated surface charge by Q[ξ],
then the Poisson bracket of this charge with any field Φ takes the form
{Q[ξ], Φ} = −δξ Φ (8.10)
where the right-hand side is (minus) the variation of Φ under the transformation gen-
erated by ξ. This is a restatement of eq. (5.34), where we noted that Poisson brackets
with momentum maps generate symmetry transformations.
Since Poisson brackets satisfy the Jacobi identity, eq. (5.35) still holds: for any two
infinitesimal gauge transformations ξ, ζ and any field configuration Φ, we have

{Q[ξ], Q[ζ]} , Φ = Q [ξ, ζ] , Φ . (8.11)
It is tempting to remove the Poisson brackets from both sides of this equality and
conclude that surface charges provide an exact representation of the asymptotic sym-
metry algebra. However, this naive removal would overlook the crucial point (5.36)
that surface charges generally close according to a (classical) central extension of the
algebra of asymptotic symmetry generators:

{Q[ξ], Q[ζ]} = Q [ξ, ζ] + c(ξ, ζ) . (8.12)
Here c(ξ, ζ) is a real-valued two-cocycle that acts trivially on any field and is therefore
invisible in eq. (8.11). The point of the seminal paper [14] was to show that such non-
trivial central extensions do arise in asymptotic symmetries of gravitational systems.
8.2 Brown-Henneaux metrics in AdS3

In this section we analyse Brown-Henneaux boundary conditions for Einstein gravity in
AdS3 . After recalling some elementary geometric aspects of three-dimensional Anti-de
Sitter space, we introduce Brown-Henneaux fall-offs and work out the corresponding
asymptotic Killing vector fields. We also display the general solution of Einstein’s
equations satisfying these boundary conditions and use it to derive the algebra of
surface charges associated with asymptotic symmetries, resulting in a direct sum of two
Virasoro algebras with non-zero central charges. We end by describing an important
family of Brown-Henneaux metrics that includes BTZ black holes.
8.2.1 Geometry of AdS3

Anti-de Sitter space in three dimensions
Consider the space R4 = R2,2 endowed with coordinates (x, y, u, v) and the metric
ds2 = dx2 + dy 2 − du2 − dv 2 . (8.13)
Then three-dimensional Anti-de Sitter space (or simply AdS3 ) is the submanifold of
R2,2 given by
AdS3 ≡ (x, y, u, v) ∈ R2,2 u2 + v 2 = `2 + x2 + y 2

(8.14)
for some parameter `2 > 0, equipped with the induced metric of R2,2 . The parameter
` is called the AdS radius. The manifold (8.14) is diffeomorphic to a product S 1 × R2
where the circle is time-like; in particular it contains closed time-like curves. Its
isometry group is O(2, 2) and acts transitively according to xµ 7→ Λµ ν xν , where xµ
denotes the coordinates (x, y, u, v) and Λ is a 4×4 matrix that preserves the “Minkowski
metric” (8.13). The stabilizer for this action is isomorphic to O(2, 1), so there is a
diffeomorphism
AdS3 ∼= O(2, 2)/O(2, 1) ∼ = SO(2, 2)/SO(2, 1).
Figure 8.3: Two-dimensional Anti-de Sitter space-time embedded in R3 as the sub-

manifold u2 +v 2 = x2 +`2 in terms of coordinates u, v, x such that the mock-Minkowski
metric of R3 reads −du2 − dv 2 + dx2 . Circles at constant x are closed time-like curves
in AdS2 . The spatial boundary of AdS2 consists of two circles at |x| → +∞. For AdS3 ,
the boundary is a time-like torus S 1 × S 1 .
In practice, physical models of space-time are manifolds without closed time-like

curves. It is therefore customary to unwind AdS3 into its universal cover, AdS
g 3 , which
3 g3 ∼
is diffeomorphic to R as a manifold. (Of course the metric on AdS 3
= R is not flat!)
To describe AdS3 , we introduce new coordinates (r, ϕ, t) given on (8.14) by
g
p √
r = x2 + y 2 = u2 + v 2 − `2 ,
ϕ = arctan(y/x), (8.15)
t = ` arctanh(v/u).
On AdS3 the coordinate t ∈ R is subject to the identification t ∼ t + 2π`, while on
the universal cover AdS
g 3 it takes all real values, without identification; see fig. 8.4. In
terms of these coordinates the AdS3 metric induced by (8.13) is
dr2
ds2 = −(1 + r2 /`2 )dt2 + 2 2
+ r2 dϕ2 . (8.16)
1 + r /`
From now on we always refer to the universal cover R3 of (8.14) simply as AdS3 ,
without tilde. With this notation the coordinates t ∈ R, r ∈ [0, +∞[ , ϕ ∈ R with
ϕ ∼ ϕ + 2π, are global coordinates on AdS3 . In general-relativistic terms, AdS3 is the
(universal cover of the) maximally symmetric solution of Einstein’s vacuum equations
in three dimensions with a negative cosmological constant Λ = −1/`2 . Note at the
outset that gravitation on an AdS3 background is determined by two independent
length scales ` and G. In particular the dimensionless coupling constant of the theory
is G/`, so that the semi-classical regime corresponds to `/G → +∞.
Killing vectors
The Killing vectors that generate isometries of AdS3 can be found thanks to the em-
bedding (8.14), where “Lorentz” transformations are generated by the six independent
vector fields
ξ1 = u∂v − v∂u , ξ2 = x∂y − y∂x , ξ3 = u∂y + y∂u ,
ξ4 = v∂x + x∂v , ξ5 = u∂x + x∂u , ξ6 = v∂y + y∂v .
The combinations of signs appearing here are due to the metric (+ + − −) in (8.13).
Upon defining
`0 ≡ 12 (ξ1 + ξ2 ), `¯0 ≡ 12 (ξ1 − ξ2 ),
`1 ≡ 21 (ξ3 + ξ4 − iξ5 + iξ6 ), `¯1 ≡ 12 (−ξ3 + ξ4 − iξ5 − iξ6 ), (8.17)
`−1 ≡ 12 (ξ3 + ξ4 + iξ5 − iξ6 ), `¯−1 ≡ 12 (−ξ3 + ξ4 + iξ5 + iξ6 ),
one finds the following Lie brackets for m, n = −1, 0, 1:
i[`m , `n ] = (m − n)`m+n , i[`¯m , `¯n ] = (m − n)`¯m+n , i[`m , `¯n ] = 0 . (8.18)
This exhibits the isomorphism so(2, 2) ∼ = sl(2, R) ⊕ sl(2, R),5 upon identifying the Lie
brackets (5.90). Note that the generator of time translations is ∂t = 1` (`0 + `¯0 ) while
the generator of rotations is ∂ϕ = `0 − `¯0 .
5
Strictly speaking we have displayed this isomorphism here for the complexification of so(2, 2),
but it also holds for real Lie algebras.
Figure 8.4: The universal cover of three-dimensional Anti-de Sitter space-time, diffeo-
morphic to R3 . It is equivalent to the interior of a solid cylinder, which may be seen
as the Penrose diagram of AdSg 3 . The time coordinate t is directed along the axis of
the cylinder while r is a radial coordinate, and ϕ is a 2π-periodic coordinate on the
circle. The spatial boundary r → +∞ is a two-dimensional time-like cylinder spanned
by the coordinates (ϕ, t), or equivalently by the light cone coordinates x± .
Spatial infinity
The region r → +∞ is a cylinder spanned by coordinates (ϕ, t) at space-like infinity.
It is the spatial boundary ∂M of AdS3 . In that region the metric (8.16) is
`2 2
2
`2 2

dt
2
ds ∼ 2 dr − r 2
− dϕ 2
= dr − r2 dx+ dx− (8.19)
r `2 r2
where we have introduced the light-cone coordinates
t
x± ≡ ± ϕ. (8.20)
`
For large r the Killing vector fields (8.17) are asymptotic to
+ 1 + − 1 −
`m ∼ eimx ∂+ − imeimx r∂r , `¯m ∼ eimx ∂− − imeimx r∂r
2 2
where m = −1, 0, 1. They generate global conformal transformations of the cylinder
at infinity, including time translations `0 + `¯0 = `∂t and rotations `0 − `¯0 = ∂ϕ . These
expressions have the general form
1 1
ξ ∼ X(x+ )∂+ − ∂+ X(x+ )r∂r , ξ¯ ∼ X̄(x− )∂− − ∂− X̄(x− )r∂r (8.21)
2 2
where the functions X and X̄ are 2π-periodic. Brown-Henneaux boundary conditions
will be such that vector fields of the form (8.21) are asymptotic symmetry generators
for arbitrary functions X, X̄.
8.2.2 Brown-Henneaux boundary conditions

We now wish to define a family of metrics on R3 that are “asymptotically Anti-de
Sitter” in the sense that they take the form of a pure AdS3 metric (8.16) at infinity.
As a starting point we ask what is the minimum amount of metrics that we wish to
include. A natural choice is to take pure AdS3 together with conical deficits, which
are obtained by cutting out a wedge out of the middle of AdS3 and identifying its two
sides. Concretely, consider the manifold described by coordinates r ∈ [0, +∞[ , ϕ ∈ R,
t ∈ R subject to the identifications

(r, ϕ, t) ∼ r, ϕ + 4πω, t − 2πA (8.22)
for some A ∈ R and ω > 0. (The normalization of ω is chosen for later convenience.)
For ω = 1/2 and A = 0 this reduces to the identifications that define pure AdS3 .
For 0 < ω < 1/2 it is a conical deficit; for ω > 1/2 it is a conical excess. Since
this is a global (topological) identification, the resulting pseudo-Riemannian manifold
still solves Einstein’s vacuum equations everywhere, except at the origin. In fact, the
metric (8.16) with identifications (8.22) is the solution of Einstein’s equations coupled
to the stress tensor of a point mass at the origin. Using the change of coordinates
A ϕ
t0 ≡ t + ϕ, r0 ≡ r , ϕ0 ≡ , (8.23)
2ω 2ω
the space-time metric can be rewritten as
r02 dr02

ds = − 1 + 2 (dt0 − Adϕ0 )2 +
2
+ 4ω 2 r02 dϕ02 (8.24)
` 1 + r02 /`2
where now there are no identifications on t0 , while ϕ0 is 2π-periodic. The term A dt0 dϕ0
suggests that A is proportional to angular momentum, as will indeed be the case below.
Note that the integral curves of ∂ϕ0 contain closed time-like curves unless
02 A2 `2
|A| ≤ 2ω` and r ≥ . (8.25)
4ω 2 `2 − A2
Thus, the space-time manifold has no pathologies only in the region where r0 is large
enough (and in particular in the asymptotic region r0 → +∞), and provided the pa-
rameter A is not too large compared to ω`. Accordingly, from now on we refer to the
solutions (8.24) with 0 < ω < 1/2 and |A| = 2ω` as extreme conical deficits.
In order to find boundary conditions that genuinely describe AdS3 space-times, one
would like the asymptotic symmetry algebra to at least include so(2, 2). If in addition
the phase space is to contain conical deficits (8.24), one is led to act with so(2, 2)
transformations on such conical deficit metrics so that, if ξ is an AdS3 Killing vector
and gµν is the metric of a conical deficit, the fall-off conditions are satisfied by the
infinitesimally transformed metric
gµν + Lξ gµν . (8.26)
Here Lξ gµν generally does not vanish because ξ may not be a Killing vector for the
conical deficit. In terms of cylindrical coordinates (r, ϕ, t), one thus obtains metrics
that satisfy the fall-off conditions [14]
  `2
+ O(r−4 ) O(r−3 ) O(r−3 )
 
grr grϕ grt r2
(gµν ) = gϕr gϕϕ gϕt  ∼  O(r−3 ) r2 + O(1) O(1)  . (8.27)
2
gtr gtϕ gtt −3 r
O(r ) O(1) − `2 + O(1)
In practice, we will impose an extra gauge-fixing condition that simplifies the compu-
tation of asymptotic symmetries. Namely, it turns out that the mixed components grϕ
and grt can always be set to zero (identically) by applying a trivial differomorphism —
2
one whose surface charges all vanish. The subleading corrections to grr = r`2 + O(r−4 )
can similarly be set to zero. We refer to this gauge choice as the Fefferman-Graham
gauge. It leads to the following definition [6]:
Definition. Let M be a three-dimensional manifold with a pseudo-Riemannian met-

ric ds2 . Suppose there exist local coordinates (r, xa ) on M (with a = 0, 1), defined for
r larger than some lower limit, such that the region r → +∞ is a time-like cylinder at
spatial infinity where the asymptotic behaviour of the metric is
r→+∞ `2 2
ds2 dr + r2 ηab + O(1) dxa dxb

∼ 2
(8.28)
r
with ηab dxa dxb the two-dimensional Minkowski metric on the cylinder. Then we say
that (M, ds2 ) is asymptotically Anti-de Sitter in the sense of Brown-Henneaux (in the
Fefferman-Graham gauge), with a cosmological constant Λ = −1/`2 .
From now on, when dealing with AdS3 gravity, we always restrict our attention to
metrics satisfying the Brown-Henneaux boundary conditions (8.28). For practical pur-
poses we will mostly describe the time-like cylinder in terms of light-cone coordinates
x± , in which case the label a in (8.28) takes the values ± and the Minkowski metric
on the cylinder is ηab dxa dxb = −dx+ dx− . Note that asymptotically AdS3 space-times
need not be (and generally are not) globally diffeomorphic to AdS3 ; in particular there
may be singularities in the bulk, as the definition (8.28) only requires r to be larger
than some lower limiting value. In the following pages we establish the main properties
of this family of metrics, including their asymptotic symmetry algebra.
Remark. The fact that one is allowed to choose the Fefferman-Graham gauge with-
out losing any information is a general property of locally asymptotically Anti-de Sitter
space-times [200]. It is related to the Fefferman-Graham expansion of AdS metrics and
the ambient construction of conformal structures [201], where conformal manifolds are
built as boundaries, or celestial spheres, of higher-dimensional bulk manifolds.
8.2.3 Asymptotic Killing vector fields

The asymptotic Killing vector fields associated with Brown-Henneaux boundary condi-
tions are vector fields that generate diffeomorphisms which preserve the fall-offs (8.28).
If gµν is a Brown-Henneaux metric and if ξ is such a vector field, this is to say that
Lξ grr = Lξ gr± = 0, Lξ gab = O(1) (a, b = ±) (8.29)
in terms of light-cone coordinates (8.20). Here the first condition follows from the fact
that the components grr = `2 /r2 and gr± = 0 are fixed, while the components gab are
allowed to fluctuate by terms of order r0 at infinity.
Lemma. Let gµν be a metric that is asymptotically AdS3 in the sense (8.28) and let
ξ be a vector field that satisfies the properties (8.29). Then
1
ξ = X(x+ )∂+ + X̄(x− )∂− − ∂+ X(x+ ) + ∂− X̄(x− ) r∂r + (subleading)

(8.30)
2
where X(x+ ) and X̄(x− ) are two arbitrary (smooth) 2π-periodic functions while the
subleading terms take the form
Z +∞ 0
`2 dr ab 0 ±
− ∂a (∂+ X + ∂− X̄) g (r , x )∂b =
2 r r0
(8.31)
`2 −4
= 2 ∂− (∂+ X + ∂− X̄)∂+ + ∂+ (∂+ X + ∂− X̄)∂− + O(r ) .
2r
These formulas associate an asymptotic Killing vector ξ with an asymptotically AdS3
metric gµν and a vector field X(x+ )∂+ + X̄(x− )∂− on the cylinder; the dependence of
ξ on the latter is linear.
Proof. Let gµν be an asymptotically AdS3 metric (8.28). We first note that the re-
quirement Lξ grr = 0 imposes ∂r ξ r = ξ r /r, whose solution is
ξ r (r, x± ) = rF(x± ) (8.32)
for some function F on the cylinder. On the other hand the condition Lξ gr± = 0 yields
2
∂r ξ c = −g ca `r ∂a F, which is solved by
+∞
dr0 ab 0 ±
Z
a a ± 2 ±
ξ = X (x ) + ` ∂b F(x ) g (r , x ) (8.33)
r r0
where X a ∂a is an arbitrary vector field on the cylinder. Note that the integral over
r0 converges since gab (r, x± ) = r2 ηab + O(1) by virtue of (8.28), so that the inverse is
ab
g ab = ηr2 + O(r−4 ). Plugging this in the integral of (8.33) we find explicitly
`2 ab
ξ a = X a (x± ) + η ∂b F(x± ) + O(r−4 ) . (8.34)
2r2
In light-cone coordinates (8.20), the two-dimensional Minkowski metric reads

η++ η+− 0 −1/2 ab 0 −2
(ηab ) = = and (η ) =
η−+ η−− −1/2 0 −2 0
so that (8.34) becomes

`2
ξ± = X ± − ∂∓ F + O(r−4 ) . (8.35)
r2
Finally one finds Lξ gab = r2 (2Fηab + LX ηab ) + O(1) where LX denotes the Lie deriva-
tive on the cylinder with respect to the vector field X a ∂a . The requirement that this ex-
pression be of order one yields the conformal Killing equation for X, LX ηab = −2Fηab .
Contracting this with η ab one finds F = − 12 (∂+ X + + ∂− X − ) and the remaining con-
straints set ∂− X + = ∂+ X − , which implies X + = X(x+ ) and X − = X̄(x− ). Formula
(8.30) follows, while the subleading terms (8.31) are produced by (8.33).
Note that the asymptotic Killing vector (8.30) takes the anticipated form (8.21)
and thus provides the generalization we were hoping to find. We will denote by ξ(X,X̄)
the asymptotic Killing vector field determined by the functions X(x+ ) and X̄(x− ).
One can decompose these functions in Fourier modes and define the vector fields
`m ≡ ξ(eimx+ ,0) , `¯m ≡ ξ(0,eimx− ) , (8.36)
whose Lie brackets take the form (8.18) up to subleading corrections, with indices
m, n ranging over all integer values. Thus, asymptotically, the finite-dimensional
isometry algebra so(2, 2) of AdS3 is enhanced to two commuting copies of the infinite-
dimensional Witt algebra (6.24). In fact we can already anticipate the result:
Theorem. The asymptotic symmetry group of AdS3 gravity with Brown-Henneaux

g + (S 1 ) × Diff
boundary conditions is a direct product Diff g + (S 1 ) whose elements are
diffeomorphisms
(x+ , x− ) 7→ f (x+ ), f¯(x− )

(8.37)
acting as conformal transformations on the cylinder at spatial infinity.
At this stage, we have not yet proven this claim since we do not know whether
all asymptotic Killing vector fields (8.30) have non-vanishing surface charges on the
phase space; this will be done in the following pages. Also note that we are being
slightly sloppy in (8.37), since the diffeomorphisms generated by (8.30) affect the radial
coordinate. Hence formula (8.37) only holds up to 1/r corrections; it is accompanied
by transformations of the radial coordinate that we do not bother writing down, but
that do preserve the limit r → +∞ in that they map r on a positive O(1) multiple of
itself.
Remark. In our description of asymptotic symmetries we mentioned that the al-

gebra of vector fields (8.36) is a direct sum of two Witt algebras up to subleading
corrections which we did not take into account. This is because these corrections are
unimportant: starting from the standard Lie bracket of vector fields, one can define
a “modified bracket” that coincides with the standard one at infinity but ensures that
the asymptotic symmetry algebra is satisfied everywhere in the bulk; see e.g. [60,176].
8.2.4 On-shell Brown-Henneaux metrics

In order for the equations of motion to provide a true extremum of the action func-
tional, the latter must be differentiable in the space of fields subject to the chosen
boundary conditions. In the case of Brown-Henneaux fall-offs, one can show that the
improved action
Z
1 2
p 1
S[gµν ] ≡ SEH [gµν ] − d x −det(gab ) K +
8πG ∂M `
is differentiable in the space of metrics satisfying Brown-Henneaux boundary condi-
tions, where SEH is the Einstein-Hilbert action (8.1) while K is the trace of the extrinsic
curvature at the boundary [191, 202] (see also [203–206]).
With this improved action it makes sense to solve Einstein’s equations in the space
of metrics (8.28). We will not review this computation here and refer to [6, 179, 180]
for details. The bottom line is that the general solution of the equations of motion
with Brown-Henneaux boundary conditions in the Fefferman-Graham gauge reads
`2 2 4G` − −

− 4G`
ds2 = dr − rdx +
− p̄(x )dx rdx − p(x +
)dx +
(8.38)
r2 r r
where p(x+ ) and p̄(x− ) are arbitrary, 2π-periodic functions of their arguments. The
factors of 4G` are introduced for later convenience. We will study this space of so-
lutions in greater detail below. For now, we only note that it is endowed with a
well-defined action of asymptotic symmetry transformations. Indeed, we define the
variation of p and p̄ under the action of an asymptotic Killing vector (8.30) by
Lξ(X,X̄) ds2 ≡ 4G` δX p(x+ ) (dx+ )2 + 4G` δX̄ p̄(x− ) (dx− )2 + (subleading),
and this variation preserves the structure of the solution (8.38). In particular, observe
that ξ(X,X̄) is an exact Killing vector if the variations δX p and δX̄ p̄ vanish. Using (8.30)
one finds
c 3 c̄ 3
δX p = X∂+ p + 2p∂+ X − ∂ X, δX̄ p̄ = X̄∂− p̄ + 2p̄∂− X̄ − ∂ X̄ (8.39)
12 + 12 −
where c = c̄ is the Brown-Henneaux central charge
3`
c = c̄ = . (8.40)
2G
The transformations (8.39) are exactly those of the components of a CFT stress ten-
sor under conformal transformations; they coincide with the coadjoint representation
(6.115) of the Virasoro algebra when seeing p(x+ ) and p̄(x− ) as Virasoro coadjoint
vectors. In that context the condition for ξ(X,X̄) to be an exact Killing vector is equiv-
alent to the statement that (X, X̄) belongs to the stabilizer of (p, p̄). We refrain from
interpreting these results any further for now; we will return to them in section 8.3.
Note that at this stage there is actually no reason to call (8.40) a central charge: even
though it does appear in (8.39) exactly as the inhomogeneous term of the coadjoint
representation (6.115), the specific value (8.40) is irrelevant since changing the normal-
ization of p or p̄ would change the value of c and c̄. The importance of the parameter
(8.40) will become apparent only from the algebra of surface charges.
8.2.5 Surface charges and Virasoro algebra

Surface charges
Take an asymptotic Killing vector field (8.30) specified by the functions X(x+ ), X̄(x− ),
and choose an on-shell metric (8.38) specified by p(x+ ), p̄(x− ). We wish to evaluate the
surface charge associated with the symmetry transformation generated by ξ(X,X̄) on
the background specified by p, p̄. As explained around eq. (8.9), this charge depends
linearly on the components of ξ(X,X̄) . In addition we need to choose a “background”
solution for which all surface charges vanish, which we take to be the degenerate conical
deficit at p = p̄ = 0,
`2
ḡ = 2 dr2 − r2 dx+ dx− . (8.41)
r
With this normalization one can show that the conserved superpotentials correspond-
ing to Brown-Henneaux asymptotic symmetries are such that the surface charge (8.9)
associated with the vector field ξ(X,X̄) on the solution (p, p̄) is
Z 2π
1
dϕ p(x+ )X(x+ ) + p̄(x− )X̄(x− )

Q(X,X̄) [p, p̄] = (8.42)
2π 0
where ϕ = 12 (x+ − x− ). (See e.g. [6] for an explicit computation.)
This charge can be interpreted in two ways: first, as the Noether charge associated
with a conformal transformation (X, X̄) in a two-dimensional CFT on the cylinder
with stress tensor (p, p̄); second, as the pairing (6.111) between the direct sum of two
Virasoro algebras and its dual. This is consistent with the fact that the transformation
law (8.39) coincides with the coadjoint representation of Virasoro. In particular, the
charge associated with time translations corresponds to the asymptotic Killing vector
∂t = (∂+ + ∂− )/`; it is the ADM mass of the system, or equivalently the Hamiltonian
Z 2π
1
dϕ p(x+ ) + p̄(x− )

M [p, p̄] = (8.43)
2π` 0
and it coincides (up to a factor 1/`) with the sum of two Virasoro energy functionals
(7.79). Similarly the charge associated with rotations, generated by the asymptotic
Killing vector ∂ϕ = ∂+ − ∂− , is the angular momentum
Z 2π
1
dϕ p(x+ ) − p̄(x− )

J= (8.44)
2π 0
and coincides with the difference of two Virasoro energy functionals. With this nor-
1
malization, pure AdS3 (8.47) has mass M = − 8G ; all its other surface charges vanish.
Surface charge algebra

We now compute the algebra satisfied by the surface charges (8.42) under Poisson
brackets. Recall that these brackets are such that they generate symmetry transfor-
mations according to (8.10). We can apply this property here: if we let (p, p̄) be an
on-shell metric (8.38), then the bracket of two charges Q(X,0) [p, p̄] and Q(Y,0) [p, p̄] is
Z 2π
(8.39) 1 c 3
dϕ X∂+ p + 2p∂+ X − ∂+ X Y (x+ )

Q(X,0) [p, p̄], Q(Y,0) [p, p̄] = −
2π 0 12
= Q([X,Y ],0) [p, p̄] + c c(X, Y ) . (8.45)
In the last line we have introduced the bracket [X, Y ] defined as the usual Lie bracket
of vector fields on the line, while c(X, Y ) is the Gelfand-Fuks cocycle (6.43) expressed
in the coordinate x+ . This is a Virasoro algebra (6.108), with a classical central
extension! The same computation would hold in the barred (antichiral) sector, while
chiral and antichiral charges commute. Thus we conclude:
Theorem. The algebra of surface charges associated with asymptotic symmetries of

AdS3 space-times in the sense of Brown-Henneaux is the direct sum of two Virasoro
algebras with central charges (8.40).
The Poisson bracket algebra (8.45) can also be rewritten in terms of more conven-
tional Virasoro generators. If we define the charges
Lm ≡ Q(eimx+ ,0) [p, p̄] , L̄m ≡ Q(0,eimx− ) [p, p̄] , (8.46)
their Poisson brackets close according to two copies of the Virasoro algebra (6.118),
up to the renaming p → L. The central charges take the definite value c = c̄ = 3`/2G.
In particular, the normalization of the homogeneous term of the bracket fixes the
normalization of the Brown-Henneaux central charge, confirming the fact that it is an
unambiguous parameter specifying the phase space. In this language the mass (8.43)
and the angular momentum (8.44) are
1
M = (L0 + L̄0 ) , J = L0 − L̄0 ,
`
as in a two-dimensional conformal field theory. In particular, pure AdS3 has L0 =
L̄0 = −c/24, as does a CFT vacuum on the cylinder. Note that the Brown-Henneaux
central charge is essentially the Planck mass measured in units of the inverse of the
AdS3 radius. Equivalently, it is the inverse of the coupling constant of the system, so
the semi-classical limit corresponds to c → +∞.
Remark. Brown-Henneaux boundary conditions are the “standard” boundary con-

ditions for gravity on AdS3 but other fall-off conditions exist as well, both in pure
Einstein gravity and in modified theories of gravity. For instance, in the Einstein case,
free boundary conditions [207] extend those of Brown-Henneaux by allowing the con-
formal factor of the metric on the boundary to fluctuate, resulting in an even larger
asymptotic symmetry algebra. Many other families of boundary conditions exist, such
as the chiral boundary conditions of [208] or the AdS3 boundary conditions of topo-
logically massive gravity [200, 209, 210] and new massive gravity [211, 212], but we will
have very little to say about these alternative possibilities.
8.2.6 Zero-mode solutions

In order to interpret the metrics (8.38), let us study zero-mode solutions, where
p(x+ ) = p0 and p̄(x− ) = p̄0 are constants. In that case the only non-zero surface
charges (8.42) are the Virasoro zero-modes L0 = p0 and L̄0 = p̄0 .
(8.40)
At p0 = p̄0 = −c/24 = −`/16G the space-time metric is that of pure AdS3 ,
`2 2 `2 − − `2 +
ds2AdS +
= 2 dr − rdx + dx rdx + dx . (8.47)
r 4r 4r
To verify that this is indeed pure AdS3 , note that the change of coordinates
`
r = earcsinh(r̄/`) (8.48)
2
brings this metric into the manifest AdS3 form (8.16) (up to the bar on the coordinate
r̄) by virtue of the identity
dr dr̄
=√ .
r `2 + r̄2
The angular momentum vanishes while the ADM mass of the solution is
1 c 1
Mvac = (L0 + L̄0 ) = − =− . (8.49)
` 12` 8G
This is the energy of the vacuum state of a two-dimensional CFT on the cylinder.
Recall that Brown-Henneaux boundary conditions were designed so as to include

conical deficits. One can show that the zero-mode solution specified by
2 2
` A ` A
L 0 = p0 = − 2ω − , L̄0 = p̄0 = − 2ω + (8.50)
16G ` 16G `
is precisely a conical deficit (8.24) written in Fefferman-Graham coordinates provided

c
|A|/` < 2ω < 1. In terms of Virasoro charges (8.46), conical deficits have − 24 <
L0 , L̄0 ≤ 0. The angular momentum is J = p0 − p̄0 = ωA/2G, while the ADM mass is
A2

p0 + p̄0 1 2
M= =− 4ω + 2 .
` 8G `
Extreme conical deficits are solutions of this type for which either p0 or p̄0 vanishes,
or equivalently for which |A| = 2ω`. Conical excesses are solutions for which p0 , p̄0 are
of the form (8.50) with |A| ≤ 2ω but ω > 1/2, and the line separating deficits from
excesses is a section of parabola
` 2G 2
`M = − − J , |J| ≤ `/4G (8.51)
8G `
whose endpoints are tangent to the lines `M = |J|. The solution at p0 = p̄0 = 0 is the
degenerate conical deficit (8.41) that we used to normalize charges. Note that conical
excesses with an angle of 2πn around the origin correspond to ω = n/2; for fixed n,
the set of such excesses is again a section of parabola in the (J, `M ) plane specified by
` 2 2G J 2
`M = − n − ,
8G ` n2
which generalizes (8.51). Note that, for vanishing angular momentum (A = 0), eq.
(8.50) yields p0 = p̄0 = −c ω 2 /6 in terms of the Brown-Henneaux central charge. This
is precisely the relation (7.46) between constant elliptic Virasoro coadjoint vectors and
their monodromy matrix.
When p0 and p̄0 are positive constants, the metric (8.38) turns out to be that of
a BTZ black hole with mass M = (p0 + p̄0 )/` and angular momentum J = p0 − p̄0
written in Fefferman-Graham coordinates [179]:
`2 2 2G` −

− 2G`
ds2BTZ +
= 2 dr − rdx − (`M − J)dx rdx − +
(`M + J)dx . (8.52)
r r r
Figure 8.5: The zero-mode solutions of AdS3 gravity with Brown-Henneaux boundary
conditions. The origin of the coordinate system (J, `M ) is the degenerate conical
deficit (8.41); the AdS3 metric is located below, on the `M axis, at the lower tip of
the shaded square. BTZ black holes are located in the wedge |J| ≤ `M . Conical
deficits and excesses are located in the lower wedge |J| ≤ −`M , respectively above
and below the parabola (8.51). All metrics such that |J| > `|M | contain closed time-
like curves at arbitrarily large radius. Anticipating the results of section 8.3.3, we
have shaded the solutions whose orbit has energy bounded from below under Brown-
Henneaux transformations; those are all BTZ black holes, the AdS3 metric, and all
conical deficits such that p0 , p̄0 ≥ −c/24. Certain solutions with energy bounded from
below are pathological in that they contain closed time-like curves at infinity — those
are the two diagonal strips surrounding the BTZ wedge.
In that context the requirement p0 , p̄0 ≥ 0 is interpreted as a cosmic censorship con-

dition |J| ≤ `M , which is saturated by extremal black holes. Beyond that barrier, all
zero-mode metrics for which |J| > `|M | contain closed time-like curves at arbitrarily
large r.
The lightest BTZ black hole at M = J = 0 is the degenerate conical deficit (8.41).
This is strikingly different from four-dimensional black holes: in the latter case, the
lightest black hole is typically empty space, whereas in three dimensions the lightest
black hole is separated from AdS3 by a classical mass gap. The metrics that fill this
gap are conical deficits, i.e. metrics of point particles, so one can loosely say that a
particle turns into a black hole when its mass is higher than the threshold c/24`, which
is essentially the Planck mass. We will encounter a similar phenomenon in flat space,
though in that case black holes will be replaced by cosmological space-times.
Remark. Since three-dimensional gravity has no local degrees of freedom, all solu-
tions of Einstein’s equations in three dimensions are locally isometric to AdS3 and can
therefore be realized as quotients of AdS3 . In particular, the BTZ metric (8.52) has
no curvature singularity at r = 0, where, as everywhere else, it is locally isometric
to AdS3 . So how can it be a black hole? The answer to this question was clarified
in [17], where it was noted that the point r = 0 is a singularity in the causal sense
even though it is a regular point in the metric sense. Note that black holes obtained
as regular identifications of AdS also exist in higher dimensions [213, 214].
8.3 The phase space of AdS3 gravity

From a Hamiltonian perspective, a phase space is a manifold consisting of “positions
and momenta” endowed with a Poisson structure, and time evolution is generated by
Poisson brackets with a Hamiltonian function H. This time evolution is in fact the one-
parameter group of diffeomorphisms generated by a Hamiltonian vector field (5.12); it
follows that phase space trajectories corresponding to different initial conditions never
cross, so one is free to think of phase space as the set of possible initial conditions of
the equations of motion. In other words, one can identify the phase space of a system
with the space of solutions of its equations of motion [215]. This reformulation is
at the core of the covariant approach to Hamiltonian mechanics, which is sometimes
stressed by referring to the phase space as being covariant. (In contrast to the stan-
dard Hamiltonian approach, the covariant one treats space and time coordinates on
an equal footing.)
According to this viewpoint, the space of solutions (8.38) is really the phase space
of AdS3 gravity with Brown-Henneaux boundary conditions. The purpose of this sec-
tion is to analyse some of its properties and to relate them with holography. Thus,
we interpret points of phase space as CFT stress tensors, describe and interpret their
transformation law under Brown-Henneaux transformations, and derive a positive en-
ergy theorem for AdS3 gravity. Quantization is relegated to section 8.4.
8.3.1 AdS3 metrics as CFT2 stress tensors

According to (8.39), the functions p, p̄ that specify an on-shell metric transform un-
der asymptotic symmetry transformations as the components of the stress tensor of a
two-dimensional CFT with central charges (8.40). The corresponding surface charges
(8.42) generate two Virasoro algebras (8.45). Accordingly, from now on we interpret
Brown-Henneaux asymptotic symmetries as the global conformal symmetries of a two-
dimensional CFT “dual” to AdS3 gravity. One can think of that theory as living on
the cylindrical boundary of AdS3 . Its central charges are (8.40) and the components
of its stress tensor should be operators whose one-point functions are the functions
p(x+ ) and p̄(x− ) appearing in the metric (8.38), which is in fact a general feature of
the AdS/CFT correspondence [13, 205, 216]. Thus the covariant phase space of AdS3
gravity coincides with the space of CFT stress tensors on the cylinder at fixed central
charges. The finite transformation laws of these stress tensors under conformal trans-
formations are given by the coadjoint representation of the Virasoro group, eq. (6.114).
The Poisson structure on the phase space of AdS3 gravity is determined by the
requirement (8.10) ensuring that surface charges generate the correct transformation
laws when acting on the fields of the theory. For Brown-Henneaux boundary conditions
this leads to the Poisson brackets of charges (8.45), which coincides with the Kirillov-
Kostant Poisson bracket (6.118). Hence we conclude:
Theorem. The covariant phase space of AdS3 gravity with Brown-Henneaux bound-
ary conditions is a hyperplane at fixed central charges (8.40) embedded in the space
of the coadjoint representation of the direct product of two Virasoro groups, and en-
dowed with the corresponding Kirillov-Kostant Poisson structure.
A loose way to interpret this theorem is to say that AdS3 gravity is group theory:
the whole phase space of the system is determined by the structure of the Virasoro
group, save for the fact that the value of the central charge is fixed by the coupling
constant. We will encounter a similar phenomenon in the next chapter when dealing
with asymptotically flat gravity. This being said, the occurrence of a Virasoro coad-
joint representation should not come as too big a surprise. Indeed, it is always true
that the charges associated with certain symmetries transform under the coadjoint
representation of the symmetry group (since these charges are nothing but momen-
tum maps). Accordingly, the surface charges of AdS3 gravity were bound to involve
the coadjoint representation of the Virasoro group. The only surprise is that Virasoro
coadjoint vectors exactly coincide with the functions specifying the metric, instead of
being some complicated non-linear combinations of the entries of the metric and their
derivatives. In particular, note that the set of on-shell Brown-Henneaux metrics (8.38)
is a vector space.
Remark. It is not strictly true that the whole phase space of AdS3 gravity coincides
with the dual of two Virasoro algebras. Indeed, this conclusion is entirely based
on the asymptotic solutions (8.38), but completely overlooks the fact that some of
these solutions cannot be extended arbitrarily far into the bulk. This subtlety leads
to (finitely many) additional directions in the complete phase space of the theory,
as discussed in [217]. We will ignore this detail since it plays a minor role for our
purposes.
8.3.2 Boundary gravitons and Virasoro orbits

The covariant phase space (8.38) is spanned by pairs of functions p(x+ ), p̄(x− ) . The

zero-mode solutions were described in section 8.2.6, but a generic par (p, p̄) is definitely
not a zero-mode since p and p̄ may have some non-trivial profile on the circle. If we pick
one such solution at random, we can generate infinitely many other ones by acting on
it with asymptotic symmetry transformations. The resulting manifold is the product
of two Virasoro coadjoint orbits at central charges (8.40),
W(p,c) × W(p̄,c̄) . (8.53)
If we think of the asymptotic symmetry group as a generalization of the space-time

isometry group O(2, 2), and of (p, p̄) as an infinite-dimensional generalization of space-
time momentum, then the orbit (8.53) is an infinite-dimensional generalization of the
standard orbits of momenta under, say, Lorentz transformations. In particular the
metrics spanning the orbit (8.53) should be seen as boosts of the metric (p, p̄).
This is a good point to introduce a terminology which has come to be more or less
standard [181,218,219]: a metric (p, p̄) obtained by acting on pure AdS3 with a certain
asymptotic symmetry transformation is known as a (classical) boundary graviton. This
nomenclature is then extended to any metric obtained from a zero-mode solution by
an asymptotic symmetry transformation. The name is justified by the fact that three-
dimensional gravity has no local (bulk) degrees of freedom, but does have non-trivial
topological (boundary) degrees of freedom visible in the arbitrariness of the pair (p, p̄)
that specifies a solution of the equations of motion.
If our goal is to classify all solutions (8.38), then orbits provides a natural organiz-
ing criterion: rather than classifying the solutions, we can classify their orbits under
asymptotic symmetries. Since we know the classification of Virasoro coadjoint orbits,
we may claim to control the full covariant phase space of AdS3 gravity. In particu-
lar, the classification of zero-mode solutions in fig. 8.5 is a first step towards the full
classification: each point in the plane (J, `M ) defines the orbit of the corresponding
zero-mode solution, and different points define distinct orbits. However, as we have
seen in section 7.2, not all orbits have constant representatives: there exist infinitely
many conformally inequivalent on-shell metrics that cannot be brought into zero-mode
solutions by asymptotic symmetry transformations. Thus the complete classification
of AdS3 metrics is essentially a product of two copies of fig. 7.3, where zero-mode
solutions are those where both p and p̄ belong to the vertical axis of the figure. This
classification foliates the covariant phase space of AdS3 gravity into disjoint orbits of
the asymptotic symmetry group.
Figure 8.6: A schematic representation of the AdS3 phase space foliated into orbits
of the asymptotic symmetry group. Solutions belonging to the same symplectic leaf
are related to one another by asymptotic symmetry transformations, i.e. “boosts”, but
there are no boosts that connect different leaves.
Remark. The relation between AdS3 gravity and orbits of the Virasoro group has
recently been the object of renewed interest, as it was realized that a similar structure
arises in many other contexts. To the author’s knowledge, the first explicit mention of
that relation appears in [182, 183, 220]; it is also hidden between the lines in [184, 221].
The relation was later studied in [45, 181] due to its implications for positive energy
theorems, while [222] (see also [223]) is devoted to the geometric properties of metrics
corresponding to non-constant pairs (p, p̄).
8.3.3 Positive energy theorems

As in section 7.3 one may ask which solutions of AdS3 gravity belong to Virasoro
orbits on which the energy functional (8.43) is bounded from below. These solutions
can then be considered as “physical”, in contrast to the pathological solutions whose
energy can be made arbitrarily low by suitable asymptotic symmetry transformations.
Since the transformation law of the components (p, p̄) is the coadjoint representation
of the Virasoro group, the results of section 7.3 are directly applicable to the problem
at hand. Thus the only solutions (p, p̄) with energy bounded from below are those in
which both p and p̄ belong to one of the orbits highlighted in red in fig. 7.7. More ex-
plicitly, zero-mode solutions (p0 , p̄0 ) belong to orbits with energy bounded from below
if and only if both p0 and p̄0 are larger than (or equal to) the vacuum value −c/24.
For solutions that do not admit a rest frame, either p or p̄ (or both) must belong to
the unique massless orbit with energy bounded from below.
These arguments can be interpreted as a positive energy theorem for AdS3 gravity
[45, 181]. They imply in particular that, in the diagram of zero-mode solutions of fig.
8.5, all BTZ black holes belong to orbits with energy bounded from below, while all
conical excesses belong to orbits with unbounded energy. The pure AdS3 metric also
belongs to an orbit with energy bounded from below, while the only conical deficits
whose energy is bounded from below under asymptotic symmetries are those located in
the square −c/24 ≤ p0 , p̄0 ≤ 0. The absolute minimum of energy among all solutions
with energy bounded from below is realized by AdS3 space-time.
Remark. Positive energy theorems in general relativity have a long history; in short,
the problem is to show that energy is bounded from below in a suitably defined phase
space of metrics. This problem is classically addressed in four-dimensional asymp-
totically flat space-times, where positivity of energy was first proved in [224]. A
supersymmetry-based proof can also be found in [225], while the case of the Bondi
mass was settled shortly thereafter by various authors — see e.g. [226] for a list of
references. Note that a relation between positive energy theorems [227] and Virasoro
orbits was already suggested in footnote 8 of [121], albeit in a very different con-
text. In section 9.3 we will encounter a positive-energy theorem in three-dimensional
asymptotically flat space-times.
8.4 Quantization and Virasoro representations

Recall from section 5.2 that the quantization of coadjoint orbits yields unitary group
representations. Assuming that this applies to the Virasoro group, unitary repre-
sentations can be interpreted as quantized orbits of Brown-Henneaux metrics under
asymptotic symmetry transformations. Accordingly, we now investigate the relation
between unitary representations of the Virasoro algebra and AdS3 quantum gravity.
We start with an overview of sl(2, R) and Virasoro highest-weight representations,
which we interpret as particles dressed with gravitational degrees of freedom in AdS3 .
We then conclude with the observation that Virasoro characters coincide with quantum
gravity partition functions in AdS3 .
Remark. This section is our first encounter with representations of Lie algebras in
this thesis, so our language and notations will be somewhat different from those of
part I. The link between the language of part I and Lie algebra representations will be
established through induced modules, in section 10.2.
8.4.1 Highest-weight representations of sl(2, R)

Here we describe highest-weight unitary representations of the Lie algebra sl(2, R),
which will be useful guides for studying Virasoro representations. From a space-time
perspective, sl(2, R) is half of the isometry algebra of AdS3 , so tensor products of
highest-weight representations of sl(2, R) are particles propagating in AdS3 . Their
Minkowskian analogue are the Poincaré representations of section 4.3.
Highest weights and descendants

The Lie algebra sl(2, R) of SL(2, R) consists of traceless real 2 × 2 matrices; its basis
can be chosen as in (4.87), but for the purpose of describing unitary representations
it is convenient to use the complex basis
L0 ≡ −it0 , L1 ≡ t1 + it2 , L−1 ≡ −t1 + it2 . (8.54)

Equivalently, Lm = i`m in terms of the basis (5.89) and the brackets (5.90) become
[Lm , Ln ] = (m − n)Lm+n (8.55)
with m, n = −1, 0, 1. In any unitary representation T of sl(2, R), the real generators
tµ are represented by anti-Hermitian operators acting in a suitable Hilbert space. In
terms of the generators (8.54), this is to say that the Hermiticity conditions
T [Lm ]† = T [L−m ] (8.56)
hold in a unitary representation. From now on we will abuse notation and neglect
writing the representation T , so that T [Lm ] ≡ Lm . This abuse is common in physics,
so it should not lead to any misunderstanding. Until the end of this chapter we also
use the Dirac notation instead of the less standard notation of part I.
Since the group SL(2, R) is simple but non-compact, all its non-trivial unitary
representations are infinite-dimensional. Fortunately, the complexification of sl(2, R)
coincides with that of su(2), and we definitely know how to build unitary highest-
weight representations of the latter. Let us therefore use the same approach for sl(2, R):
we start from a (normalized) highest-weight state |hi belonging to the Hilbert space
of the representation, such that
L0 |hi = h|hi, L1 |hi = 0 , (8.57)
where h is the highest weight. The conditions (8.56) imply that the operator repre-
senting L0 is Hermitian, so h must be real. The interpretation of (8.57) is that |hi
has energy h if we think of L0 as the Hamiltonian, while L1 is an annihilation operator.6
In order to produce a representation we must also act with operators L−1 on the
highest-weight state. This leads to descendant states of the form
(L−1 )N |hi, (8.58)
where the non-negative integer N is the level of the descendant. Each descendant
state is an eigenstate of L0 with eigenvalue h + N , so L−1 is analogous to a creation
operator. We then declare that the carrier space H of the representation is spanned by
all linear combinations of descendant states. Since we want H to be a Hilbert space,
descendant states with different levels must be orthogonal because their eigenvalues
under L0 differ. Furthermore, all descendant states must have non-negative norm
squared. Using the Hermiticity conditions (8.56), this amounts to the requirement
2 (8.56)
0 ≤ (L−1 )N |hi = h [(L−1 )N ]† (L−1 )N h = h (L1 )N (L−1 )N h .
To ensure that this condition holds, we evaluate scalar products of descendant states.
Thanks to the commutation relations (8.55) one finds
N
Y −1
N N
h (L1 ) (L−1 ) h = N! (2h + k) (8.59)
k=0
6
The terminology is somewhat backwards, since the highest weight h is actually the lowest eigen-
value of L0 in the space of the representation; this terminological clash is standard.
where we have used hh|hi = 1. Thus, all descendant states have strictly positive norm
squared if and only if h > 0. If h = 0, then the representation is trivial.
Note that here the only restriction imposed on h by unitarity is h ≥ 0. However, if

we were to integrate a highest-weight representation of sl(2, R) into an exact unitary
representation of the group SL(2, R), then h would have to be an integer in order
to ensure that a rotation by 2π is represented by the identity. On the other hand,
projective representations of SL(2, R) allow h to be an arbitrary positive real number
since the fundamental group of SL(2, R) is isomorphic to Z.
Representations by quantization
The representations just described can be identified with representations obtained by
quantizing suitable coadjoint orbits of SL(2, R). Indeed, note that the Lie algebra
(8.55) admits a quadratic Casimir operator
1 (8.54)
C = L20 − (L1 L−1 + L−1 L1 ) = −t20 + t21 + t22 . (8.60)
2
This operator is proportional to the identity in any irreducible representation of
sl(2, R). In the highest-weight representation (8.58) it takes the value
C = h(h − 1), (8.61)
which proves by the way that the representation is irreducible. The fact that (8.60)
takes a constant value is reminiscent of the “mass shell” condition defining the coad-
joint orbit (5.93), and indeed the highest-weight representation just displayed is the
quantization of such an orbit for h > 0. The only subtlety is that the value of the
Casimir (8.61) is not quite h2 , but h(h − 1); the two numbers coincide in the semi-
classical limit h → +∞, and the h − 1 of (8.61) may be seen as a quantum correction
of the classical result.
This identification is confirmed by the computation of SL(2, R) characters. Indeed,

the counting argument of the end of section 5.3 can now be applied to the space H
spanned by the highest-weight state |hi and its descendants. As a result one finds
that Tr(q L0 ) is precisely given by formula (5.102) up to the replacement of h + 1/2 by
h. From a space-time perspective, the product of two such characters is the character
of an irreducible representation of the isometry algebra so(2, 2) of AdS3 . If the two
representations have weights h and h̄ say, the product of their characters can be
interpreted as the partition function of a particle with mass (h + h̄)/` and spin h − h̄
propagating in AdS3 .
8.4.2 Virasoro modules

We now generalize the representation theory of sl(2, R) to the Virasoro algebra. Be-
fore doing so, a word of caution is in order: in the case of SL(2, R), we were able
to interpret highest-weight representations as quantized coadjoint orbits of the type
(5.93); this identification was supported by the matching of the Casimir operator
(8.61) with the definition of the orbit. In the case of the Virasoro algebra the situation
is complicated by the fact that all coadjoint orbits (at non-zero central charge) are
infinite-dimensional; in addition, the only Casimir operators of the Virasoro algebra
are functions of its central charge [228]. Accordingly, we start with a few comments
regarding Virasoro geometric quantization, before turning to the construction of its
highest-weight representations and the evaluation of the associated characters.
Semi-classical regime
If one believes in the orbit method, geometric quantization applied to the coadjoint
orbits of the Virasoro algebra should produce unitary Virasoro representations. This
viewpoint was adopted in [121, 130, 131], with the conclusion that the quantization of
orbits with positive energy and constant representatives indeed provides highest-weight
representations in the large c limit. By contrast, the limit of small c is much more
elusive, and at present it is not known if the discrete series of Virasoro representations
at c ≤ 1 can be obtained by geometric quantization (see e.g. [229]). From a gravita-
tional point of view, the Virasoro central charge (8.40) is the AdS radius in units of the
Planck length, so large c corresponds to the semi-classical regime. This is confirmed
by symplectic geometry: the Kirillov-Kostant symplectic form (5.29) evaluated at a
constant Virasoro coadjoint vector (p0 , c) is
(6.111) c
ω(p0 ,c) (ξm )p0 , (ξn )p0 = −im 2p0 + m2 δm+n,0 (8.62)
12
where ξm = ad∗Lm is the vector field on W(p0 ,c) that generates the coadjoint action
of the Virasoro generator Lm given by (6.109). The occurrence of c confirms that
the regime of large c is semi-classical in the sense that a large volume is assigned to
any portion of phase space. Conversely, small c corresponds to the non-perturbative
regime, where quantum corrections may alter classical results in a radical way.
This heuristic argument is consistent with the fact that geometric quantization is
relatively well established at large c, but poorly understood at small c. Since appli-
cations to three-dimensional gravity rely on the semi-classical limit anyway, from now
on we restrict ourselves to the regime of large c. This assumption turns out to greatly
simplify representation theory, and allows us to think of highest-weight Virasoro rep-
resentations as quantizations of Virasoro orbits with constant representatives.
This being said, to our knowledge there is as yet no strict mathematical proof of
the fact that geometric quantization of Virasoro orbits produces highest-weight repre-
sentations, despite numerous attempts in the literature (see e.g. [230,231]). Our view-
point here will be pragmatic, and we shall assume that the representations obtained by
quantizing such orbits are indeed highest-weight representations. This assumption will
be supported, among other observations, by the fact that Virasoro characters match
suitable gravitational partition functions (see section 8.4.4).
Highest-weight representations
The basis of the Virasoro algebra given by (6.109) is such that, in any unitary repre-
sentation, the operators representing the generators Lm + L−m , i(Lm − L−m ) and Z
are anti -Hermitian. A more standard basis is given by

Z
Lm ≡ iLm + i δm,0 , Z ≡ iZ, (8.63)
12
where the constant shift in L0 ensures that the vacuum state has zero eigenvalue
under L0 . According to this definition the operators representing Lm and Z in a
unitary representation satisfy the Hermiticity conditions
L†m = L−m , Z† = Z (8.64)
where we abuse notation by denoting the basis element Lm and the operator that
represents it with the same symbol. In any irreducible representation the Hermitian
central operator Z is proportional to the identity with a coefficient c ∈ R, so we may
write the commutation relations of the operators representing the generators (8.63) as
c
[Lm , Ln ] = (m − n)Lm+n + m(m2 − 1)δm+n,0 . (8.65)
12
Here the existence of the sl(2, R) subalgebra (8.55) is manifest, since the contribution
of the central extension vanishes for m = −1, 0, 1.
Highest-weight representations of the Virasoro algebra (8.65) seem to have first

appeared in [232–234], and were then further studied in [163, 235, 236]. They are built
in direct analogy to the highest-weight representations of sl(2, R). In accordance with
geometric quantization, the parameters that specify these representations coincide with
those that determine the corresponding coadjoint orbits. In the present case, taking
the orbit of a constant coadjoint vector (p0 , c), one defines a real number h by
c
p0 = h − . (8.66)
24
This ensures that h = 0 for the vacuum configuration, while h ≥ 0 for orbits with
energy bounded from below. With this notation, the representation obtained by quan-
tizing the orbit of (p0 , c) is obtained as follows.
To begin, in analogy with (8.57), one defines the highest-weight state of the repre-
sentation to be a normalized state |hi such that7
L0 |hi = h|hi, Lm |hi = 0 for m > 0. (8.67)
The state |hi is also called a primary state. Its definition ensures that it has energy
h under the Hamiltonian L0 , while the operators Lm with m > 0 are annihilation
operators. In analogy with (8.58) one also defines descendant states
L−k1 ...L−kn |hi, 1 ≤ k1 ≤ k2 ≤ · · · ≤ kn . (8.68)
Thus one can interpret the operators L−m with m > 0 as creation operators. We will
discuss the gravitational interpretation of this representation in section 8.4.4.
7
The actual value of the quantum weight h may differ from the classical parameter defined in
(8.66) by corrections of order O(1/c), so from now on it is understood that h refers to the quantum
value. This subtlety will have very little effect on our discussion.
Using the commutation relations (8.65) of the Virasoro algebra, one verifies that
each descendant (8.68) is an eigenstate of L0 with eigenvalue
n
X
h+ ki ≡ h + N
i=1
where the non-negative integer N is the level of the descendant. One then declares
that the space H of the Virasoro representation is the Verma module spanned by all
linear combinations of descendant states. As in the case of sl(2, R), descendants with
different levels have different eigenvalues under L0 . According to (8.64) the latter must
be Hermitian if the representation is to be unitary, so scalar products of descendants
with different levels vanish. However, in contrast to sl(2, R), there are in general many
different descendant states with the same level. More precisely, at large central charge
c, the number of different descendants at level N is the number p(N ) of partitions of
N in distinct positive integers (e.g. p(0) = p(1) = 1, p(2) = 2, p(3) = 3, p(4) = 5, etc.).
Note that the representation whose carrier space is spanned by the descendant
states (8.68) is an induced representation of the Virasoro algebra, i.e. an induced
module. Indeed, the conditions (8.67) define a one-dimensional representation of the
subalgebra generated by L0 and the Lm ’s with m > 0, and the prescription (8.68) is
the algebraic analogue of the statement that wavefunctions live on a quotient space
G/H (recall section 3.3). By the way, a similar interpretation holds for the highest-
weight representations of sl(2, R) displayed in section 8.4.1. We will return to this
observation in section 10.2.
Unitarity for Virasoro representations

We now ask whether the vector space spanned by |hi and its descendants is a Hilbert
space, given that Hermitian conjugation is defined by (8.64). Working only with
low-level descendant states, one can easily derive some basic necessary conditions for
unitarity. For instance, the only descendant at level one is L−1 |hi and its norm squared
is hh|L1 L−1 |hi = 2h, so a necessary condition for unitarity is h ≥ 0. Similarly, at level
N there is a state L−N |hi with norm squared
(8.65) c
hh|LN L−N |hi = 2N h + N (N 2 − 1)
12
whose positivity for large N requires c ≥ 0. Thus, the only values of c and h that
give rise to unitary representations are positive. In terms of coadjoint orbits of the
Virasoro group, this is to say that only orbits with positive energy can produce unitary
representations under quantization. In order to go further one generally relies on
the so-called Gram matrix of the module, whose entries are the scalar products of
descendants. Demanding unitarity then boils down to the requirement that the Gram
matrix be positive-definite. One can show that this condition is always verified by
Virasoro highest-weight representations at large c. Since this is a standard result in
two-dimensional conformal field theory (see e.g. [153]), we simply state it here without
proof:
Proposition. Highest-weight representations of the Virasoro algebra with c > 1 and

h > 0 are unitary and irreducible in the sense that all descendant states have strictly
positive norm squared.
Note that, by contrast, unitary highest-weight representations at central charge

c ≤ 1 have a very intricate structure due to null states, i.e. states with vanishing norm
that are modded out of the Hilbert space as in the definition (3.5) of L2 spaces. In
that case, not all descendant states (8.68) are linearly independent, since some of them
are effectively set to zero — in this sense the Verma module is reducible. We will not
take such subtleties into account here because we are interested only in the limit of
large central charge, where null states are absent.
Remark. It was recently shown [237] that all irreducible unitary representations
of the Virasoro group with a spectrum of L0 bounded from below are highest-weight
representations of the type described here. In this sense, highest-weight representations
exhaust all unitary representations of the Virasoro algebra.
Vacuum representation
In sl(2, R), the representation with highest weight h = 0 is trivial since all descendant
states are null by virtue of eq. (8.59). We now describe the Virasoro representation
obtained by setting h = 0, to which we refer as the vacuum representation. It is ob-
tained by quantizing the vacuum Virasoro orbit, containing the point pvac = −c/24.
The highest weight state |0i of that representation satisfies the properties
L0 |0i = L−1 |0i = Lm |0i = 0 for all m > 0, (8.69)
which ensures that the vacuum state |0i is invariant under the sl(2, R) subalgebra
generated by L−1 , L0 and L1 ; it is the quantum counterpart of the statement that the
stabilizer of the vacuum orbit is the group PSL(2, R). Crucially, the vacuum is not
invariant under the higher-mode generators L−2 , L−3 etc. so that the representation
whose carrier space is spanned by all descendant states
L−k1 ...L−kn |0i, 2 ≤ k1 ≤ · · · ≤ kn (8.70)
is non-trivial. This representation is unitary for all c > 0, and it is free of null states
(i.e. irreducible) whenever c > 1. It is in many ways analogous to the standard
highest-weight representation generated by the descendant states (8.68), but the con-
dition L1 |0i = 0 makes it slightly smaller than generic highest-weight representations.
It may seem puzzling that the vacuum state is not left invariant by all Virasoro
generators but only by a subset thereof as in (8.69). The reason is that, at non-zero
central charge, it is impossible to define a non-zero state that is annihilated by all
Virasoro generators. Indeed, if there was such a state |0̃i, then we would have
(8.65)
c c
0 = h0̃|Ln L−n |0̃i = h0̃| 2nL0 + n(n2 − 1) |0̃i = n(n2 − 1) 0̃|0̃ ,
12 12
which is a contradiction when c 6= 0 and n 6= −1, 0, 1. Hence the vacuum state of

a Virasoro-invariant theory is always sl(2, R)-invariant but never Virasoro-invariant.
One should appreciate the counter-intuitive nature of this phenomenon: it means
that even the absolute simplest Virasoro-invariant quantum theory is described by
an infinite-dimensional Hilbert space of vacuum descendants. In section 10.1 we will
encounter a similar non-trivial vacuum representation for the BMS3 group.
8.4.3 Virasoro characters

The relation between AdS3 quantum gravity and Virasoro representations is most eas-
ily expressed in terms of partition functions, so as a preliminary we now evaluate
characters of highest-weight representations of the Virasoro algebra at c > 1.
Let H be the Hilbert space of an irreducible, unitary representations of the Vira-

soro algebra with central charge c and highest-weight h (if h = 0 we take the vacuum
representation). We define the character of the representation as
χ(τ ) ≡ TrH q L0 −c/24 , q ≡ e2πiτ

(8.71)
where the notation is almost the same as in eq. (5.102). The parameter τ is a complex
number with positive imaginary part so that, if L0 is interpreted as the Hamiltonian,
then Im(τ ) is the inverse temperature and the character itself is a canonical partition
function. The normalization factor −c/24 is conventional.
Generic highest-weight representations

Let h > 0 be a highest weight and c > 1 a large central charge. Consider the Verma
module spanned by all descendant states (8.68). Then there are no null states and
distinct descendants are linearly independent, so the spectrum of L0 consists of all
values h + N , with multiplicity p(N ). Accordingly, the character (8.71) reads
X +∞
X
h+N −c/24 h−c/24
χ(τ ) = p(N )q =q p(N )q N . (8.72)
N =0 N =0
We now rewrite this in a more convenient way thanks to the following result:
Lemma. The series (8.72) can be rewritten as an infinite product

+∞ +∞
X
N
Y 1
p(N )q = . (8.73)
N =0 n=1
1 − qn
Proof. We follow [238], to which we refer for a careful treatment of the convergence
issues that will not be addressed here. To prove (8.73) we consider its right-hand side
and expand each individual term of the infinite product as a geometric series:
+∞ +∞ X+∞
Y 1 Y
nk 2 3
2 4

n
= q = 1 + q + q + q + · · · 1 + q + q + · · · ···
n=1
1 − q n=1 k=0
+∞ X
X N X +∞
X
k1 +k2 +···+kn
=1+ q = p(N )q N .
| {z }
N =1 n=1 k1 ,k2 ,...,kn N =0
k1 +k2 +···+kn =N qN
1≤k1 ≤k2 ≤···≤kn
| {z }
p(N )
This concludes the argument.

Thus the character (8.72) can be rewritten as
+∞
h−c/24
Y 1 q h−(c−1)/24
χ(τ ) = q = (8.74)
n=1
1 − qn η(τ )
where in the second equality we have introduced the Dedekind eta function
+∞
Y
1/24
η(τ ) ≡ q (1 − q n ). (8.75)
n=1
The result (8.74) can be seen as an infinite product of sl(2, R) characters (5.102) with
parameters nτ . Equivalently, since sl(2, R) representations coincide with harmonic
oscillators whose partition functions were written in (5.102), one can think of a Virasoro
representation as an infinite collection of harmonic oscillators. This suggests that
Virasoro characters can be interpreted as quantum field theory partition functions,
which will be confirmed in section 8.4.4 below. Note also the presence of the ubiquitous
factor 1 − q n in the denominator.
Remark. Our derivation of (8.74) relied on the fact that the eigenvalue h + N of L0
has degeneracy p(N ). This is only true provided there are no null states, i.e. provided
c > 1. By contrast, for c ≤ 1, null states generally do exist and are modded out of the
Hilbert space of the representation. This leads to a smaller degeneracy of eigenvalues
of L0 , hence to a character that is strikingly different from (8.74). We will not display
characters at c ≤ 1 here; see e.g. [239, 240] for explicit formulas.
Vacuum representation
The character of the vacuum Virasoro representation can be evaluated in the same way
as for generic highest-weight representations. The only subtlety is that the vacuum is
sl(2, R)-invariant, leading to a reduced number of descendant states (8.70). Explicitly,
let ∆(N ) denote the degeneracy of the eigenvalue N in the space spanned by the
vacuum descendants. For N ≥ 2, this degeneracy is the number of partitions of N
in positive integers which are strictly greater than one (thus ∆(0) = 1 by convention
but ∆(1) = 0, ∆(2) = ∆(3) = ∆(4) = ∆(5) = 1, ∆(6) = 2, etc.). Then the vacuum
character is
+∞
X +∞
X
N −c/24 −c/24
χ(τ ) = ∆(N )q =q ∆(N )q N . (8.76)
N =0 N =0
In order to relate ∆(N ) to the usual partition p(N ), we note that

number of partitions of N
∆(N ) = p(N ) − = p(N ) − p(N − 1)
containing at least one “1”
which allows us to rewrite the character (8.76) as

+∞ +∞
X (8.73) Y 1
χvac,c (τ ) = q −c/24 p(N )q N (1 − q) = q −c/24 n
. (8.77)
N =0 n=2
1 − q
Note how the product in the denominator of (8.77) is truncated (no term n = 1) owing
to the sl(2, R) symmetry of the vacuum state. In particular the vacuum character
(8.77) is not just the limit h → 0 of the generic character (8.74).
8.4.4 Dressed particles and quantization

Since AdS3 gravity has Virasoro symmetry, its quantization is expected to produce
unitary representations of the direct sum of two Virasoro algebras. We now investi-
gate to what extent this is the case. For notational simplicity we denote the Virasoro
d 1 )), so that the asymptotic symmetry algebra of AdS3
algebra by vir (instead of Vect(S
gravity with Brown-Henneaux boundary conditions is vir ⊕ vir.
First let us make the proposal more precise: the orbit of a metric (p, p̄) is a coad-
joint orbit (8.53) of the product of two Virasoro groups with central charges (8.40). For
simplicity let us assume that the metric is a zero-mode and that its energy is bounded
from below under asymptotic symmetry transformations, so p(x+ ) = p0 ≥ −c/24 and
p̄(x− ) = p̄0 ≥ −c̄/24. The non-zero modes belonging to the orbit (8.53) may be seen
as classical analogues of the descendant states (8.68). Upon defining h ≡ p0 + c/4 and
h̄ ≡ p̄0 + c̄/24, one expects that the geometric quantization of the orbit (8.53) pro-
duces the tensor product of two highest-weight representations of the Virasoro algebra
labelled by (h, c) and (h̄, c̄).8 The same would be true by quantizing the orbit of AdS3 ,
except that the result would be the tensor product of two vacuum representations.
It is worth comparing these representations to those of so(2, 2) ∼= sl(2, R)⊕sl(2, R),

the isometry algebra of AdS3 . Since the latter is a subalgebra of vir ⊕ vir, any Virasoro
representation with highest weights (h, h̄) contains many so(2, 2) subrepresentations
with weights increasing from (h, h̄) to infinity. Thus a Virasoro representation is an
so(2, 2) representation dressed with (infinitely many) extra directions in the Hilbert
space obtained by acting with the Virasoro generators L−2 , L−3 , etc.
Now recall that a particle propagating in AdS3 is an irreducible unitary repre-

sentation of so(2, 2), while asymptotic symmetries generalize isometries by including
gravitational fluctuations. Accordingly one is led to interpret Virasoro representa-
tions as particles in AdS3 dressed with some extra gravitational degrees of freedom
accounted for by the modes Ln that do not appear in the isometry algebra. These are
8
As before, the quantum values of (h, h̄) may differ from their classical counterparts by 1/c
corrections.
quantum analogues of the classical “boundary gravitons” described in section 8.3.2.

Thus a Virasoro representation is a particle in AdS3 dressed with boundary gravitons.
In particular the vacuum representation of vir ⊕ vir is identified with the Hilbert space
of quantum boundary gravitons around pure AdS3 .
As a verification of the fact that Virasoro symmetry is realized in AdS3 quantum

gravity, one may wonder whether the quantum partition function of gravity reproduces
a (combination of) Virasoro character(s). This computation was carried out in [241],
where the authors evaluated the one-loop partition function of gravity on AdS3 at finite
temperature 1/β and angular potential θ (both taken to be real). Upon combining
these numbers into a modular parameter
1 β
τ≡ θ+i , (8.78)
2π `
it was found that the gravitational one-loop partition function reads
+∞

L0 −c/24 L̄0 −c/24

c/12
Y 1
Zgrav (β, θ) = Tr q q̄ = |q| (8.79)
n=2
|1 − q n |2
where q ≡ e2πiτ and q̄ is its complex conjugate. This is precisely the character of the
tensor product of two Virasoro vacuum representations, which confirms that quantized
boundary gravitons around AdS3 span an irreducible highest-weight representation of
vir ⊕ vir. The same computation can be performed for orbifolds of AdS3 obtained by
imposing identifications of the form ϕ ∼ ϕ + 2π/N in terms of cylindrical coordinates,
where N ∈ N∗ . The corresponding metric is that of a conical deficit labelled by the
Virasoro coadjoint vectors p0 = p̄0 = −c/(24N 2 ). For N ≥ 2 the resulting one-loop
partition function is found to be
+∞
2h
Y 1
Zgrav,N (β, θ) = |q| (8.80)
n=1
|1 − q n |2
c
where h = 24 (1 − 1/N 2 ). This is again the character of the tensor product of two
highest-weight representations of the Virasoro algebra with weights h = h̄, which
confirms the interpretation of Virasoro modules as particles dressed with boundary
gravitons.
In chapter 10 we will develop a similar interpretation for BMS3 particles, which will
then be confirmed in chapter 11 by the matching of BMS3 characters with one-loop
partition functions for asymptotically flat gravity and higher-spin theories.
Remark. In [184] it was conjectured that the one-loop partition function (8.79) is
exact because it is the only expression compatible with Virasoro symmetry. Higher
loop corrections would then renormalize the Brown-Henneaux central charge but would
leave the q-dependent one-loop determinant unaffected. This conjecture was used to
evaluate a Farey tail sum representing the putative full, non-perturbative, partition
function of AdS3 gravity (see also [242]). To our knowledge the one-loop exactness of
(8.79) is still an unproven statement.
A note on the Fabri-Picasso theorem

Quantizing the orbit of the AdS3 metric under Brown-Henneaux transformations yields
the vacuum representation of vir ⊕ vir, whose highest-weight state |0i is annihilated
by all so(2, 2) generators. But the Virasoro generators Lm , L̄m with m ≤ −2 do not
leave the vacuum invariant. Thus Virasoro symmetry is spontaneously broken in AdS3
gravity, and the descendant states
L−k1 ...L−kn L̄−`1 ...L̄−`m |0i (8.81)
can be loosely interpreted as Goldstone bosons obtained by acting on the vacuum

with broken symmetry generators. This interpretation has come to be standard in the
realm of asymptotic symmetries; in four dimensions it leads to the identification of soft
graviton states with Goldstone bosons for spontaneously broken BMS symmetry [19].
This being said, one should keep in mind that the comparison to Goldstone bosons
should be handled with care. Indeed, spontaneously broken internal symmetries are
always such that states obtained by acting with broken symmetry generators on the
vacuum do not belong to the Hilbert space. This statement is the Fabri-Picasso the-
orem (see e.g. [243]), and it follows from the fact that the norm of the state Q|0i has
an infrared (volume) divergence whenever Q generates a broken global internal sym-
metry. If the Fabri-Picasso theorem was to hold for asymptotic symmetries, then the
descendant states (8.81) would make no sense. Fortunately the situation of asymptotic
symmetries is different because the charges that generate them are surface charges (8.9)
rather than Noether charges. As a result, when Q is a broken asymptotic symmetry
generator, the norm squared of Q|0i is an integral over a compact manifold, and is
therefore finite. In this sense spontaneously broken asymptotic symmetries behave in
a way radically different from spontaneously broken internal symmetries.
Part III
BMS3 symmetry
and gravity in flat space
This part contains the original contributions of the thesis and is de-
voted to Bondi-Metzner-Sachs (BMS) symmetry in three dimensions. It
starts with an introductory chapter where the definition of the BMS3 group
is motivated by asymptotic symmetry considerations. We then move on
to the quantization of BMS3 symmetry and show that irreducible unitary
representations of the BMS3 group, i.e. BMS3 particles, are classified by
supermomentum orbits that coincide with coadjoint orbits of the Virasoro
group. We also evaluate the associated characters and show that they
coincide with one-loop partition functions of the gravitational field at finite
temperature and angular potential. Finally we extend this matching to
higher-spin theories and supergravity in three dimensions.
Chapter 9
Classical BMS3 symmetry
The Bondi-Metzner-Sachs (BMS) group is an infinite-dimensional symmetry group of

asymptotically flat gravity at null infinity, that extends Poincaré symmetry. It was
originally discovered in four space-time dimensions in the seminal work of Bondi, Van
der Burg, Metzner [1, 2] and Sachs [3, 4]. In this chapter we introduce BMS symmetry
in three dimensions [20] and describe its classical aspects, i.e. those that do not rely
on its realization as the quantum symmetry group of a Hilbert space. We will show in
particular that the phase space of asymptotically flat gravity coincides with (a hyper-
plane in) the coadjoint representation of the centrally extended BMS3 group.
The structure is as follows. In section 9.1 we show how BMS3 symmetry emerges
from an asymptotic symmetry analysis. Section 9.2 is devoted to the abstract mathe-
matical definition of the BMS3 group and its central extension, including their adjoint
and coadjoint representations. In section 9.3 we describe the phase space of three-
dimensional asymptotically flat gravity embedded in the space of the coadjoint repre-
sentation of BMS3 . Finally, in section 9.4 we show how BMS3 symmetry can be seen
as a flat limit of Virasoro symmetry.
This chapter is mostly based on [45–47], although the first section follows the earlier
references [6, 37]. As usual, more specialized references will be cited in due time.
9.1 BMS metrics in three dimensions

The purpose of this section is to explain how the BMS3 group (and its central exten-
sion) emerges as an asymptotic symmetry of three-dimensional Minkowskian space-
times at null infinity. In particular we describe the embedding of Poincaré transfor-
mations and the action of BMS3 on the covariant phase space of the system, and
observe the appearance of a classical central extension. We refer to section 8.1 for
some background on three-dimensional gravity and asymptotic symmetries in general.
257
Chapter 9. Classical BMS3 symmetry 258
9.1.1 Three-dimensional Minkowski space

Minkowski space in three dimensions is the manifold R3 endowed with a metric whose
expression in inertial coordinates (t, x, y) = (x0 , x1 , x2 ) is
ds2 = −dt2 + dx2 + dy 2 = ηµν dxµ dxν (9.1)
where (ηµν ) is the Minkowski metric (4.35) in D = 3 dimensions. In general-relativistic

terms Minkowski space-time is the maximally symmetric solution of Einstein’s equa-
tions with vanishing cosmological constant.
The isometry group of Minkowski space is the Poincaré group (4.39) with D = 3:
IO(2, 1) = O(2, 1) n R3 . Its elements are pairs (f, α) acting transitively on R3 ,
xµ 7→ f µ ν xν + αµ , (9.2)
where (f µ ν ) is a Lorentz transformation while αµ is a space-time translation. The

stabilizer of the origin xµ = 0 is the Lorentz group, confirming the obvious diffeomor-
phism R3 ∼ = IO(2, 1)/O(2, 1).
While inertial coordinates are the most common in Minkowski space, a different set
of coordinates will be more convenient for the description of BMS3 symmetry. Namely,
as in (1.4), we define retarded Bondi coordinates (r, ϕ, u) by
p x + iy
r≡ x2 + y 2 , eiϕ ≡ , u ≡ t−r, (9.3)
r
whose range is r ∈ [0, +∞[ , u ∈ R, and ϕ ∈ R with the identification ϕ ∼ ϕ + 2π.
In that context the coordinate u is known as retarded time. We will also refer to the
coordinates (r, ϕ, t) as cylindrical coordinates; they are analogous to (8.15) in AdS3 .
In terms of cylindrical and Bondi coordinates, the Minkowski metric (9.1) reads
ds2 = −dt2 + dr2 + r2 dϕ2 = −du2 − 2dudr + r2 dϕ2 (9.4)
which is the three-dimensional analogue of (1.5). Bondi coordinates are represented on

the Penrose diagram of Minkowski space in fig. 1.1. Note that parity acts as ϕ 7→ −ϕ.
Killing vectors
The Killing vector fields that generate Poincaré transformations (9.2) are simplest to
write down in inertial coordinates, where they have the general form
ξ(x) = αρ + X µ xν µν ρ ∂ρ .

(9.5)
Here αµ and X µ are two arbitrary, constant vectors generating translations and Lorentz
transformations, respectively, while µνρ is the completely antisymmetric tensor such
that 012 = 1 (indices are raised and lowered with the Minkowski metric). In partic-
ular, the component α0 is responsible for time translations, while α1 and α2 generate
translations in the directions x = x1 and y = x2 , respectively. The component X 0 is
responsible for spatial rotations while X 1 and X 2 give rise to boosts in the directions
x1 and x2 , respectively.
For later comparison with asympotic symmetries it is convenient to rewrite the

Killing vectors (9.5) in Bondi coordinates (9.3). For pure translations we find
α0 (ϕ)
ξTranslation = α(ϕ)∂u − ∂ϕ + α00 (ϕ)∂r (9.6)
r
where the function α(ϕ) is related to the translation vector αµ by
α(ϕ) = α0 − α1 cos ϕ − α2 sin ϕ . (9.7)
For pure Lorentz transformations we similarly obtain

u
ξLorentz = X(ϕ) − X 00 (ϕ) ∂ϕ + uX 0 (ϕ)∂u − rX 0 (ϕ) − uX 000 (ϕ) ∂r

(9.8)
r
where the function X(ϕ) is related to the boost vector X µ by
X(ϕ) = X 0 − X 1 cos ϕ − X 2 sin ϕ . (9.9)
Note that both (9.6) and (9.8) depend on functions on the circle; already at this stage
it is tempting to speculate that there exist boundary conditions such that asymptotic
symmetry generators take that form with arbitrary functions (X, α) on the circle. The
BMS boundary conditions below will do just that.1
The structure of the algebra spanned by the vector fields (9.6) and (9.8) can be
made more transparent by a suitable choice of basis. Thus we define the complexified
Poincaré generators
jm ≡ ξLorentz , pm ≡ ξTranslation
X(ϕ)=eimϕ α(ϕ)=eimϕ
where m, n = −1, 0, 1. The resulting Lie brackets read
i[jm , jn ] = (m − n)jm+n , i[jm , pn ] = (m − n)pm+n , i[pm , pn ] = 0 , (9.10)
with m, n = −1, 0, 1. The Lie algebra of the BMS3 group will extend these brackets
by allowing arbitrary integer values of m, n, in the same way that the Witt algebra
(6.24) extends sl(2, R). Note that the structure G n g of the Poincaré group (4.93) is
manifest in these relations, as the bracket of j’s with p’s takes exactly the same form
as the bracket of j’s with themselves.
9.1.2 Poincaré symmetry at null infinity

Null infinities and celestial circles
In preparation for the asymptotic analysis to come, note that in Bondi coordinates the
region r → +∞ at finite ϕ and u is a cylinder spanned by coordinates (ϕ, u) at future
1
We are cheating in (9.8), since for now there is no way to distinguish X 000 (ϕ) from −X 0 (ϕ); the
justification for this combination of derivatives will be provided by asymptotic symmetries.
null infinity. It is the upper cone on the boundary of the Penrose diagram of fig. 1.1.
This is due to our choice of coordinates: instead of (9.3) we could have defined ad-
vanced Bondi coordinates, with advanced time given by v = t + r instead of u = t − r.
As a result we would have found that the region r → +∞ is past null infinity, but up
to this difference the whole construction would have been the same. In this thesis we
use retarded Bondi coordinates throughout, but it is always understood that a parallel
construction exists in terms of advanced Bondi coordinates. In particular the region
r → +∞ will always be future null infinity, denoted I + .
Future null infinity is the region of space-time where all light rays eventually escape;
similarly past null infinity is the origin of all incoming light rays. In optical terms, if we
were living in a three-dimensional space-time, the region that we would see around us
would be a circle on our past light-cone. As the distance from us to the circle increases,
the latter approaches past null infinity. A similar (time-reversed) interpretation holds
for future null infinity, and justifies the following terminology:
Definition. The future celestial circle at retarded time u associated with the Bondi
coordinates (r, ϕ, u) is the circle spanned by the coordinate ϕ on future null infinity,
and at fixed time u. Similarly the past celestial circle at advanced time v is the circle
at fixed time v on past null infinity.
This definition is illustrated in fig. 1.2. From now on the words “celestial circle”
always refer to a future celestial circle. As we shall see, BMS3 symmetry will refor-
mulate and generalize the action of Poincaré transformations on celestial circles, and
more generally on null infinity.
Poincaré transformations on I +
Since we have rewritten Minkowski Killing vectors in Bondi coordinates, it is worth
asking whether one can write finite Poincaré diffeomorphisms (9.2) (as opposed to
infinitesimal vector fields) in Bondi coordinates. The answer is obviously yes, but the
result is not particularly illuminating because the linear nature of the transformations
is hidden when writing them in terms of (r, ϕ, u). Fortunately, Bondi coordinates are
designed so that things simplify at null infinity; in particular it turns out that Poincaré
transformations preserve the limit r → +∞ in the sense that (i) they map r on a pos-
itive multiple of itself and (ii) they affect ϕ and u but leave them finite. Accordingly
Poincaré transformations are well-defined at null infinity and one may ask how they
act on the coordinates (ϕ, u) spanning I + . The procedure for finding this action is
explained in greater detail in [109].
For definiteness we focus on the connected Poincaré group (4.40). We can use the
isomorphism (4.83) to describe Lorentz transformations in terms of SL(2, R) matrices,
the correspondence being given explicitly by (4.91). One then finds that a pure space-
time translation xµ 7→ xµ + αµ acts on null infinity according to

(u, ϕ) 7→ u + α(ϕ), ϕ (translation) (9.11)
where the function α(ϕ) is related to the components αµ by (9.7). In particular,
pure time translations act on Bondi coordinates as u 7→ u + α0 , without affecting the
other coordinates. Similarly, a Lorentz transformation specified by an SL(2, R) matrix

(6.86) acts on I according to (u, ϕ) 7→ f (ϕ)u, f (ϕ) where f (ϕ) is a projective
0
+

transformation (6.88) of the celestial circle, with parameters A, B given by (6.89). For
instance, spatial rotations act as ϕ 7→ ϕ + θ, leaving all other coordinates untouched.
This is analogous to the four-dimensional situation described in eqs. (1.6)-(1.7). Upon
performing simultaneously a translation α and a Lorentz transformation f , the trans-
formation of (u, ϕ) reads

(u, ϕ) 7→ f 0 (ϕ)u + α(f (ϕ)) , f (ϕ) (9.12)
where f takes the form (6.88) while α is given by (9.7). As we shall see below, the
BMS3 group acts on I + in the same way, except that f (ϕ) will be an arbitrary dif-
feomorphism of the circle and that α(ϕ) will be an arbitrary function on the circle.
Analogous results hold at past null infinity.
Note that in (9.12) we are abusing notation slightly. Indeed, a Poincaré transfor-
mation is a diffeomorphism of the whole space-time (not just null infinity) and acts
on all three coordinates r, ϕ, u. In particular the transformation law (9.12) only holds
up to corrections of order 1/r. These corrections vanish in the limit r → +∞ and
leave out only the leading piece displayed in (9.12), but they matter for the extension
of Poincaré (or BMS) transformations from the boundary into the bulk.
9.1.3 BMS3 fall-offs and asymptotic symmetries

We now wish to define a family of metrics on R3 that are “asymptotically flat” at
future null infinity in the sense that they approach the Minkowski metric (9.4) near
the boundary of space-time. As in the AdS3 case above, a good starting point is to
ask what is the minimum amount of metrics that one wants to include; clearly, pure
Minkowski space should be there, but in addition one may include conical deficits.
These are defined by cutting out a wedge of angular opening 2π(1 − 2ω) out of the
middle of space and quotienting Minkowski space-time with identifications of the type
(8.22) in terms of cylindrical coordinates. The change of coordinates (8.23) then turns
the metric of (quotiented) Minkowski space into
ds2 = (dt0 − Adϕ0 )2 + dr02 + 4 ω 2 r02 dϕ02 , (9.13)
which is the flat limit (` → +∞) of eq. (8.24). In these terms there are no identifica-
tions on t0 , and ϕ0 is 2π-periodic. As in the AdS3 case the cross-term Adt0 dϕ0 suggests
that A is proportional to angular momentum, as will indeed be the case below. In
contrast to AdS3 , however, the region of large r0 is always free of closed time-like
curves since the condition for the integral curves of ∂ϕ0 to be space-like now simply
yields r02 > A2 /(4ω 2 ), without condition on the ratio of A and ω. This is a flat limit
of the more stringent conditions (8.25) encountered in AdS3 . We refer for instance
to [244, 245] for a more thorough study of conical deficits.
Now suppose we wish to find boundary conditions that include such conical deficits.
If we want the asymptotic symmetry group to contain the Poincaré group, we are
forced to include in the phase space all metrics obtained by performing rotations,
translations and boosts of conical deficits. This is the same argument as in section
8.2, where we derived Brown-Henneaux boundary conditions. It leads to a class of
metrics with prescribed asymptotic behaviour at null infinity, analogous to eq. (8.27)
in the AdS3 case. Some of the subleading components of the metric can then be set
to zero identically as a gauge choice, which leads to the following definition:
Definition. Let M be a three-dimensional manifold with a pseudo-Riemannian met-

ric ds2 . Suppose there exist local Bondi coordinates (r, ϕ, u) on M, defined for r larger
than some lower limit, such that the region r → +∞ be a two-dimensional cylinder at
future null infinity and such that the asymptotic behaviour of the metric be
r→+∞
ds2 O(1)du2 − 2 + O(1/r) dudr + r2 dϕ2 + O(1)dudϕ .

∼ (9.14)
Then we say that (M, ds2 ) is asymptotically flat at future null infinity in the BMS
gauge. A parallel construction exists at past null infinity.
The BMS gauge condition is the flat space analogue of the Fefferman-Graham gauge
used in (8.28). We stress that it is truly a gauge condition in the sense of asymptotic
symmetries: the diffeomorphism used to bring a metric from a general asymptotically
flat form into the BMS gauge is trivial, as it does not affect the surface charges of
the metric. By contrast, the diffeomorphisms that change the physical state of the
system are generated by non-zero surface charges and span the asymptotic symmetry
group of the system, which will turn out to be the BMS3 group. From now on, when
dealing with asymptotically flat gravity, we always restrict our attention to metrics
satisfying the BMS boundary conditions (9.14). Note that asymptotically flat space-
times need not be (and generally are not) globally diffeomorphic to Minkowski space;
the definition (9.14) only requires r to be larger than some lower limiting value. Note
also that there is no restriction on the sign of the fluctuating components in the metric
(9.14); in particular the term of order r0 multiplying du2 may be positive.
Asymptotic Killing vectors

The asymptotic Killing vector fields associated with flat boundary conditions (in BMS
gauge) are vector fields that generate diffeomorphisms which preserve the fall-off con-
ditions (9.14). This is to say that, if gµν is an asymptotically flat metric, its Lie
derivative under such a vector field ξ must satisfy
Lξ grr = Lξ grϕ = Lξ gϕϕ = 0 (9.15)
together with
Lξ guu = O(1), Lξ guϕ = O(1), Lξ gur = O(1/r) (9.16)
in terms of retarded Bondi coordinates. Here (9.15) follows from the fact that the
components grr = grϕ = 0 and gϕϕ = r2 are fixed in the BMS gauge (9.14); by
contrast the components guu , guϕ and gur are allowed to fluctuate by terms of order
r0 , r0 and r−1 respectively.
Lemma. Let gµν be an asymptotically flat metric in the sense (9.14) and let ξ be a
vector field that satisfies (9.15) and (9.16). Then
ξ = X(ϕ)∂ϕ + α(ϕ) + uX 0 (ϕ) ∂u − rX 0 (ϕ)∂r + (subleading)

(9.17)
where X(ϕ) and α(ϕ) are two arbitrary (smooth) 2π-periodic functions, while the
subleading terms take the form
Z +∞ 0
0 00 dr
(α + uX ) gur ∂ϕ
r r02
Z +∞ 0
0 00 dr 1 0 00 (9.18)
+ ∂ϕ (α + uX ) gur + 2 (α + uX )guϕ ∂r =
r r02 r
1 1
= (α0 + uX 00 )∂ϕ + (α00 + uX 000 )∂r + O(r−2 ) .
r r
These formulas uniquely associate an asymptotic Killing vector
field ξ with an asymp-
totically flat metric gµν and two functions X(ϕ), α(ϕ) on the celestial circle; the
dependence of ξ on these functions is linear.
Proof. Let gµν be an asymptotically flat metric (9.14). First note that the condition
Lξ grr = 0 yields ∂r ξ u = 0, so ξ u is r-independent. On the other hand the condition
Lξ grϕ = 0 gives a differential equation ∂r ξ ϕ = − r12 gru ∂ϕ ξ u , which is solved by
+∞
dr0
Z
ϕ u
ξ = X(u, ϕ) + ∂ϕ ξ gr0 u (9.19)
r r02
where X(u, ϕ) is an arbitrary function on the cylinder at null infinity. The integral
over r0 converges since gru = −1 + O(1/r) by virtue of (9.14), so that ξ ϕ = X(u, ϕ) +
1
∂ ξ u + O(1/r2 ). At this point we introduce a function α(u, ϕ) defined by
r ϕ
ξ u = α(u, ϕ) + uX 0 (u, ϕ) (9.20)
(prime denotes partial differentiation with respect to ϕ), which is allowed by virtue of
the fact that ξ u is r-independent. In these terms the condition Lξ gϕϕ = 0 gives
1
ξ r = −r∂ϕ ξ ϕ − guϕ (α0 + uX 00 ) (9.21)
r
where ξ ϕ is given by (9.19). Since we now know that the most general solution ξ
of (9.15) is determined by two functions X(u, ϕ) and α(u, ϕ) on null infinity, we can
use the remaining conditions (9.16) to constrain these functions. Using first Lξ gur =
O(1/r), we find
∂u ξ u = X 0 , (9.22)
which upon rewriting ξ u as (9.20) says that the combination ∂u α+u∂u X 0 vanishes. The
requirement Lξ guϕ = O(1) then yields ∂u X = 0, which is to say that X(u, ϕ) = X(ϕ)
only depends on the coordinate ϕ on the celestial circle. Plugging this back into (9.22)
then yields ∂u α = 0 as well. Formula (9.17) follows, while the subleading terms (9.18)
are produced by (9.19) and (9.21).
Note that the asymptotic Killing vectors (9.17) precisely take the anticipated form
(9.6)-(9.8) and generalize Poincaré transformations in an infinite-dimensional way. In
particular the asymptotic symmetry group contains all space-time translations (cor-
responding to α(ϕ) of the form (9.7)) and all Lorentz transformations (corresponding
to X(ϕ) of the form (9.9)). We shall denote by ξ(X,α) the asymptotic Killing vector
determined by the functions X(ϕ) and α(ϕ). One verifies that the Lie brackets of such
vector fields read

ξ(X,α) , ξ(Y,β) = ξ([X,Y ],[X,β]−[Y,α]) + (subleading) (9.23)
where the brackets in the subscript on the right-hand side are understood to be stan-
dard Lie brackets on the circle, e.g. [X, α] ≡ Xα0 − αX 0 . The subleading terms can be
neglected because they will turn out not to contribute to the surface charges; alterna-
tively, as in the AdS3 case, they can be absorbed by a redefinition of the Lie bracket
such that the algebra is realized everywhere in the bulk [60, 176].
The structure of the algebra

(9.23) can be made more transparent by decomposing
the functions X(ϕ), α(ϕ) in Fourier modes and defining the vector fields
jm ≡ ξ(eimϕ ,0) , pm ≡ ξ(0,eimϕ ) . (9.24)
As one can verify, formula (9.23) implies that their Lie brackets take the form (9.10)
with arbitrary integer labels m, n, up to subleading corrections.
Thus we now know that the asymptotic symmetries of three-dimensional Min-

kowskian space-times span an algebra that contains the Witt algebra (extending
the Lorentz algebra) and an infinite-dimensional Abelian algebra (extending space-
time translations). The corresponding asymptotic symmetry transformations are re-
ferred to as superrotations and supertranslations, respectively;2 they span an infinite-
dimensional algebra known as the BMS algebra in three dimensions, that we shall
denote as bms3 . Finite BMS3 transformations act on null infinity according to formula
(9.12), where f (ϕ) is an arbitrary diffeomorphism of the celestial circle while α(ϕ) is
an arbitrary function on the circle.
We refrain from analysing the group-theoretic aspects of these symmetries at this

point — this will be the subject of all later sections in this chapter. Instead, we
now keep going in our study of asymptotically flat gravity; in particular we actually
still have to confirm that superrotations and supertranslations are indeed non-trivial
asymptotic symmetries, i.e. that the associated surface charges do not vanish.
Remark. One should keep in mind that Bondi coordinates are global, since the
definition (9.3) covers all points of Minkoswki space-time. Thus the fact that Bondi
coordinates allow one to describe either only future or only past null infinity (and not
both) does not mean that they cover only “half” of the space-time. A similar comment
applies to BMS symmetry, whose definition in terms of space-time relies on a choice
of coordinates that favours future over past null infinity (or vice-versa). Despite this
2
The prefix “super” has nothing to do with supersymmetry, but stresses the fact that special-
relativistic quantities are extended in an infinite-dimensional way.
asymmetry, it was recently realized that (for well-behaved asymptotically flat space-
times [246]) the two definitions of BMS can be related by an “antipodal identification”,
which leads to the application of BMS symmetry to scattering phenomena [19, 24, 26,
27, 247–258]. A related question (as yet unsolved) is whether BMS symmetry can be
defined at spatial infinity [259].
9.1.4 On-shell BMS3 metrics

In order for the equations of motion to provide an extremum of the action func-
tional, the latter must be differentiable in the space of fields satisfying certain fall-off
conditions. In the case of asymptotically flat three-dimensional gravity, it was shown
in [198], using the Chern-Simons formalism, that there exists a well-defined variational
principle. The same conclusion was obtained more recently in [260] in the metric for-
malism, with the observation that the pure Einstein-Hilbert action (8.1), without any
extra boundary term, is differentiable in the space of asymptotically flat metrics.
Accordingly, it makes sense to ask about the general solution of Einstein’s vacuum
equations in the BMS gauge. It was shown in [6] that this solution reads
ds2 = 8G p(ϕ)du2 − 2dudr + 8G j(ϕ) + up0 (ϕ) dudϕ + r2 dϕ2

(9.25)
where p(ϕ) and j(ϕ) are arbitrary, 2π-periodic functions of ϕ. Upon evaluating surface
charges we will see that p(ϕ) and j(ϕ) are densities of energy and angular momentum
at null infinity, respectively. As in the earlier AdS3 case (8.38), the normalization
factors involving Newton’s constant G are included for later convenience.
The transformation law of the solution (9.25) under the action of asymptotic Killing
vectors follows from the definition
Lξ(X,α) ds2 ≡ 8G δ(X,α) p(ϕ) du2 + 8G δ(X,α) j(ϕ) + u δ(X,α) p0 (ϕ) dudϕ

(9.26)
where the functions X(ϕ) and α(ϕ) determine the vector field (9.17). Evaluating the
Lie derivative (9.26) one finds
c2
δ(X,α) j = Xj 0 + 2X 0 j + αp0 + 2α0 p − α000 , (9.27)
12
0 0 c2 000
δ(X,α) p = Xp + 2X p − X (9.28)
12
where c2 is a dimensionful central charge proportional to the Planck mass [37]:
3
c2 = . (9.29)
G
In this language the asymptotic vector field ξ(X,α) is an exact Killing vector field for
the metric (j, p) if both variations (9.27)-(9.28) vanish. The subscript “2” in (9.29) will
be justified below.
The transformation law of p in (9.28) coincides with that of a CFT stress tensor
under a conformal transformation generated by X; it is the coadjoint representation
(6.115) of the Virasoro algebra. The transformation (9.27) of j is somewhat more

involved. We refrain from interpreting these results for now, as we will return to
them in much greater detail in the upcoming sections. Note that at this stage all
normalizations are arbitrary, and in particular the central charge (9.29) would take
another value if we chose to change the normalization of p.
9.1.5 Surface charges and BMS3 algebra

Surface charges

Take an asymptotic Killing vector field (9.17) specified by the functions X(ϕ), α(ϕ) ,
and choose an on-shell metric (9.25) specified by j(ϕ), p(ϕ) . We wish to evaluate
the surface charge associated with the symmetry transformation generated by ξ(X,α)
on the background specified by (j, p). This charge depends linearly on the components
of ξ(X,α) , as explained around eq. (8.9). In addition we must choose a normalization,
that is, a “background” solution for which we declare that all surface charges vanish.
Here we take it to be the null orbifold at j = p = 0,
ḡ = −2dudr + r2 dϕ2 . (9.30)
With this normalization one can show that the surface charge (8.9) associated with
the vector field ξ(X,α) on the solution (j, p) is [6]
Z 2π
1
Q(X,α) [j, p] = dϕ j(ϕ)X(ϕ) + p(ϕ)α(ϕ) . (9.31)
2π 0
It can be interpreted as the pairing of the bms3 algebra, consisting of pairs (X, α),
with its dual consisting of pairs (j, p). In particular, even though we haven’t defined
the BMS3 group at this stage, we already know that the space of solutions (9.25)
belongs to its coadjoint representation. The charge associated with time translations
correponds to the asymptotic Killing vector ∂u ; it is the Hamiltonian of the system,
Z 2π
1
M = P0 = dϕ p(ϕ) , (9.32)
2π 0
and it allows us to interpret p(ϕ) as the energy density carried by the gravitational
field at (future) null infinity. Thus p(ϕ) is the Bondi mass aspect associated with the
metric (9.25) and its zero-mode (9.32) is the Bondi mass. More generally the charges
associated with supertranslations (X = 0) take the form
Z 2π
1
Q(0,α) [j, p] = dϕ p(ϕ)α(ϕ) . (9.33)
2π 0
In the same way, the charge associated with rotations corresponds to the asymptotic
Killing vector ∂ϕ ; it is the angular momentum
Z 2π
1
J = J0 = dϕ j(ϕ) . (9.34)
2π 0
We can interpret j(ϕ) as the density of angular momentum carried by the gravitational
field at null infinity; it is the angular momentum aspect associated with the metric
(9.25). More generally all superrotation charges take the form
Z 2π
1
Q(X,0) [j, p] = dϕ j(ϕ)X(ϕ)
2π 0
and generalize centre of mass charges. With this normalization Minkowski space (9.4)
has energy M = −1/8G and all its other surface charges vanish.
Surface charge algebra

We now compute the Poisson brackets of surface charges for asymptotically flat space-
times. Recall that these brackets generate symmetry transformations (8.10), on ac-
count of the fact that conserved charges are momentum maps (5.34). We can apply
this property here to deduce the Poisson brackets of charges: if we let (j, p) be an
on-shell metric (9.25), then the bracket of charges is

Q(X,α) [j, p], Q(Y,β) [j, p] =
Z 2π
(9.31) 1 (9.35)
= − dϕ δ(X,α) j(ϕ)Y (ϕ) + δ(X,α) p(ϕ)β(ϕ) .
2π 0
Using the infinitesimal transformation laws (9.27)-(9.28) and integrating by parts one
can then show that

Q(X,α) [j, p], Q(Y,β) [j, p] = Q([X,Y ],[X,β]−[Y,α]) [j, p] + c2 [c(X, β) − c(Y, α)] , (9.36)
where as in (9.23) we denote by [X, Y ] ≡ XY 0 −Y X 0 the standard Lie bracket of vector

fields on the circle, while c(X, Y ) is the Gelfand-Fuks cocycle (6.43). Thus, the surface
charges of asymptotically flat space-times close under the Poisson bracket according
to a central extension of the BMS3 Lie algebra displayed in (9.23). Furthermore the
central extension is remarkably similar to that of the Virasoro algebra (6.108). Again,
we refrain from interpreting this result any further at this point, since we haven’t truly
defined the BMS3 group yet. For future reference we simply note that the Poisson
brackets (9.36) can be rewritten in terms of a discrete set of generators analogous to
(9.24). Namely, let us define the charges
Jm ≡ Q(eimϕ ,0) [j, p], Pm ≡ Q(0,eimϕ ) [j, p]
for all m ∈ Z, generalizing the Hamiltonian (9.32) and angular momentum (9.34).
Then the bracket (9.36) yields the algebra
i{Jm , Jn } = (m − n)Jm+n ,
c2 3
i{Jm , Pn } = (m − n)Pm+n + m δm+n,0 , (9.37)
12
i{Pm , Pn } = 0 .
This is an infinite-dimensional central extension of (9.10), with m, n ∈ Z.

Note that the central extension proportional to c2 in (9.37) pairs superrotation

generators Jm with supertranslation generators Pm . By contrast the Witt algebra
spanned by superrotations receives no central extension. This is why we wrote the
central charge (9.29) with an index “2”: the notation c1 will be kept for the central
charge pairing superrotation generators with themselves. Despite many similarities,
we stress that c2 is not a Virasoro central charge; in particular it is a dimensionful
quantity. This is consistent with the fact that the value of c2 varies when changing
the normalization of the charges Pm : if we were to define P̃m ≡ λPm with some
non-zero real number λ, the Poisson brackets of J ’s and P̃’s would take the form
(9.37) with the central charge c2 replaced by λc2 . Nevertheless, the value displayed in
(9.29) is canonical in the sense that it is the one provided by the normalization of the
Hamiltonian (9.32), which in turn is the surface charge associated with the vector field
∂u in terms of Bondi coordinates. In essence the central charge c2 is analogous to that
of the Bargmann group (4.103), which as we saw in (4.108) is also a mass scale. This
is radically different from the Virasoro algebra, where the value of the central charge
c in (6.118) is unambiguously fixed by the condition that the homogeneous structure
constants take the form (m − n).
Remark. The BMS boundary conditions given here are the flat analogue of Brown-
Henneaux boundary conditions. In this sense they are the “standard” fall-offs for
three-dimensional asymptotically flat gravity. However, it is likely that other consistent
boundary conditions exist in Einstein gravity — for instance adapting to flat space the
free AdS3 boundary conditions of [207]. In addition one can devise BMS-like boundary
conditions for other theories of gravity, such as topologically massive gravity [261,262],
bigravity [263], conformal gravity [264] or new massive gravity [265,266]. In particular,
in parity-breaking theories such as TMG, one typically finds that the Virasoro algebra
spanned by superrotations Jm develops a non-zero central charge c1 . Aside from this
comment we will have very little to say about these alternative possibilities.
9.1.6 Zero-mode solutions

We focus here on zero-mode metrics, with constant (j, p) = (j0 , p0 ) in eq. (9.25). The
only non-vanishing surface charges for such metrics are the Bondi mass (9.32) and the
angular momentum (9.34), which coincide with p0 and j0 respectively.
(9.29)
At j0 = 0, p0 = −c2 /24 = −1/8G, the metric is that of pure Minkowski space-
time (9.4). Solutions having p0 = −c2 /24 but non-zero j0 corresponding to “spinning
Minkowski space-time”. Note that, while the normalization of p0 and c2 is arbitrary,
the relation
c2
pvac = −
24
3
is a normalization-independent statement. It suggests that Minkowski space plays
the role of a classical vacuum for a putative dual theory; we will return to this later.
3
Indeed, changing the normalization of p would also change the value of the central charge that
ensures that the bracket {J , P} takes the canonical form in eq. (9.37).
Figure 9.1: The zero-mode solutions of asymptotically flat gravity with BMS3 fall-offs.
The origin of the coordinate system (J, M ) is the null orbifold (9.30); the Minkowski
metric is located below, on the M axis, right between conical deficits and conical
excesses. Flat space cosmologies are located in the region M > 0. Conical deficits are
such that −c2 /24 < M < 0 while excesses have M < −c2 /24. Anticipating section
9.3.3, we have shaded the solutions whose orbit has energy bounded from below under
BMS3 transformations; those are all flat space cosmologies and all conical excesses,
plus Minkowski space. Note that this figure is a flat limit of fig. 8.5, as the slope of
the curve `M = J in the plane (J, M ) goes to zero when ` → +∞.
Solutions having 0 > p0 > −c2 /24 are conical deficits for all values of j0 , with
a deficit angle 2π(1 − 2ω) given by (7.46). In particular, solutions with p0 = 0 are
degenerate conical deficits, and the solution p0 = j0 = 0 is the null orbifold (9.30) that
we used to normalize charges. Solutions having p0 < −c2 /24 are conical excesses with
an excess angle 2π(2ω − 1) given again by (7.46). For p0 = −c2 n2 /24 the excess angle
is 2π(n − 1).
Zero-mode solutions with positive p turn out to describe flat space cosmologies,
sometimes also called shifted boost orbifolds [267, 268]. They represent a (2 + 1)-
dimensional universe that undergoes a big crunch followed by a big bang, where the
transition between the contracting and expanding phases is smooth only if j 6= 0.
When j = 0 these solutions can be thought of as a compactification of the three-
dimensional Milne universe. They can also be seen as limits of the interior region of
BTZ black holes as the AdS3 radius goes to infinity. The lightest flat space cosmology
has p0 = 0 and is separated from Minkowski space-time pvac = −c2 /24 by a classical
mass gap; the latter is filled by conical deficits. This is very similar to the mass gap
separating BTZ black holes from AdS3 .
Note that, in contrast to AdS3 , no cosmic censorship is needed to ensure the absence
of closed time-like curves at infinity (although closed time-like curves generally do exist
in the bulk). In fact, the whole classification of flat zero-mode metrics may be seen as
a limit ` → +∞ of that of zero-mode metrics in AdS3 . The family of flat zero-mode
solutions is plotted in fig. 9.1.
9.2 The BMS3 group

This section is devoted to a detailed description of the BMS3 group and its central
extension. This will rely on a level of abstraction that may seem offputting at first
sight, but one should keep in mind that BMS3 is an extension of Poincaré symmetry
so that almost all statements on BMS have an analogue in special relativity. We
urge the reader to adopt this point of view whenever there is a risk of getting lost
in mathematical formulas. In particular, our notation will be consistent with the
analogies between Poincaré and BMS3 :
Notation Poincaré BMS3

f finite Lorentz tsf. finite superrotation
X infinitesimal Lorentz tsf. infinitesimal superrotation
α translation supertranslation
jm , Jm , Jm Lorentz generator superrotation generator
pm , Pm , Pm translation generator supertranslation generator
j relativistic angular momentum angular supermomentum
p energy-momentum supermomentum
Z1 , c 1 / superrotational central charge
Z2 , c 2 / supertranslational central charge
Table 9.1: Analogies between Poincaré and BMS3 .
The plan of this section is as follows. Motivated by the structure of the Poincaré
group (4.93), we start by defining a notion of “exceptional semi-direct products” (gen-
erally centrally extended) and work out their adjoint and coadjoint representations.
We then use asymptotic symmetries to motivate the definition of the BMS3 group and
its central extension, which turn out to be exceptional semi-direct products based on
the Virasoro group. Finally, we write down the adjoint representation, the Lie algebra
and the coadjoint representation of the (centrally extended) BMS3 group. Through-
out the section, these structures are compared to their Poincaré counterparts and to
three-dimensional asymptotically flat gravity. Note that the material presented here
relies heavily on chapters 2, 4 and 6.
9.2.1 Exceptional semi-direct products

Here we study a general family of semi-direct products, whose structure turns out to be
common to the Poincaré group (in three dimensions) and the BMS3 group. We start
by describing this structure and its central extension, then display the corresponding
adjoint and coadjoint representations.
Defining exceptional semi-direct products

Definition. Let G be a Lie group with Lie algebra g. The associated exceptional
semi-direct product is the group
G nAd gAb ≡ G n g (9.38)
where gAb denotes the Lie algebra of G seen as an Abelian vector group acted upon by
G according to the adjoint representation. Its group operation is given by (4.6) with
the action σ replaced by the adjoint.
As usual we denote elements of (9.38) as pairs (f, α) where f ∈ G is a “rotation”

while α ∈ A is a “translation”. For instance the (double cover of the) Poincaré group
(4.93) takes the exceptional form with G = SL(2, R). It is straightforward to obtain
central extensions of this structure: if G
b is a central extension of G with group opera-
tion (2.11) in terms of some two-cocycle C and if b g is its Lie algebra, one can consider
the exceptional semi-direct product
G Ad gAb
b nc b (9.39)
c denotes the adjoint representation (6.98) of G.

where Ad b Its elements are quadruples
(f, λ; α, µ) (9.40)
where λ, µ are real numbers, being understood that the pair (f, λ) belongs to G b while
(α, µ) belongs to bgAb . The notation emphasizes the fact that “centrally extended rota-
tions” (f, λ) play a role radically different
from “centrally extended translations” (α, µ).
In fact the notation (f, λ), (α, µ) would be more accurate, but to reduce clutter we
stick to (9.40).
The group operation in (9.39) is that of an exceptional semi-direct product based

on the centrally extended group G.
b Explicitly, using the centrally extended adjoint
representation (6.98), we have
(f, λ; α, µ) · (g, ρ; β, ν) =
(6.98)
1 (9.41)
= f · g, λ + ρ + C(f, g) ; α + Adf β, µ + ν − hS[f ], βi
12
where C is the two-cocycle that defines G, b S is the associated Souriau one-cocycle

∗
(6.79), and the pairing h·, ·i is that of g with g. This is exactly the structure that we
will find in the centrally extended BMS3 group below, but for now we first investigate
the adjoint and coadjoint representations of (9.39) in general terms.
Adjoint representation and Lie algebra
Consider a centrally extended exceptional semi-direct product G g. Owing to the

bnb
general form (5.103), its Lie algebra is a semi-direct sum
g Aad
b c bgAb (9.42)
where ad g, i.e. the Lie bracket (6.102). The elements of

c is the adjoint representation of b
this algebra are quadruples (X, λ; α, µ) where (X, λ) belongs to b g while (α, µ) belongs
to gAb .
b
The adjoint representation of the group (9.39) follows from formula (5.105). Start-
ing for simplicity with the centreless group (9.38), it is given by

Ad(f,α) (X, β) = Adf X, Adf β − adAdf X α = Adf X, Adf β − [Adf X, α] (9.43)
where the “Ad” on the right denotes the adjoint representation of G alone.4 In the
second equality we abuse notation by writing a bracket between Adf X ∈ g and α ∈
gAb , being understood that we use the Lie bracket of g and interpret the result as an
element of gAb . The Lie bracket of the centreless Lie algebra g Aad gAb follows:

(X, α), (Y, β) = [X, Y ], adX β − adY α = [X, Y ], [X, β] − [Y, α] , (9.44)
in accordance with the general formula (5.106). Note that this is precisely the form of
the Lie bracket (9.23) of BMS3 asymptotic Killing vectors.
The centrally extended adjoint representation corresponding to (9.43) can be ob-

tained in a similar fashion. Using (6.98) and (6.102) we find explicitly
c (f,α) (X, λ; β, µ) =
Ad
1 1 1
= Adf X, λ − hS[f ], Xi ; Adf β − [Adf X, α], µ − hS[f ], βi + hs[Adf X], αi
12 12 12
(9.45)
where Ad on the right-hand side denotes the adjoint representation of G and [·, ·] is
the Lie bracket of g. On the left-hand side we have neglected central terms in the
subscript of the adjoint representation, since they act trivially.
From the adjoint representation one can read off, by differentiation, the Lie bracket
of the centrally extended algebra (9.42). One expects the contribution of central terms
to include a cocycle c given by (6.101), and indeed one finds

(X, λ; α, µ), (Y, ρ; β, ν) = [X, Y ], c(X, Y ); [X, β] − [Y, α], c(X, β) − c(Y, α) (9.46)
where we abuse notation as in (9.44). Already note that the last entry precisely takes
the form of the central extension in the Poisson bracket (9.36) of flat surface charges.
The appearance of the same cocycle c in both central entries of (9.46) is due to the
exceptional semi-direct product structure of (9.39). It implies that, when written in
terms of generators, the brackets of rotations with translations take the same form as
the brackets of rotations with themselves, including central terms. Explicitly, suppose
gAb
we are given a basis of b gAb consisting of non-central generators
Ja ≡ (ja , 0; 0, 0) , Pa ≡ (0, 0; pa , 0)
4
In case of identical notations, the subscript indicates which group we are referring to.
where the ja ’s and pa ’s respectively generate g and gAb , together with two central
elements
Z1 ≡ (0, 1; 0, 0) , Z2 ≡ (0, 0; 0, 1) . (9.47)
Suppose also that the Lie brackets of Ja ’s take the form (2.27) with some structure
constants fab c and some central coefficients cab , and let us choose the basis elements
Pa such that their bracket with Ja ’s takes the same form as the bracket of Ja ’s with
themselves. This is allowed by the exceptional semi-direct product structure. Then
the bracket (9.46) implies that the commutation relations of b gAb gAb are
[Ja , Jb ] = fab c Jc + cab Z1 ,
[Ja , Pb ] = fab c Pc + cab Z2 , (9.48)
[Pa , Pb ] = 0 .
The fact that J ’s act on P’s according to the adjoint representation is now manifest
since the structure constants of the two first lines are identical. Note in particular that
the central generator Z1 pairs rotations with themselves, while Z2 pairs rotations with
translations. The centrally extended BMS3 algebra (9.37) illustrates this phenomenon,
as does the Poincaré algebra (9.10), albeit without central extension.
Note that the definition of (9.39) rules out all central extensions in the bracket
[P, P] of (9.48), and indeed we will show in section 9.2.5 that such central extensions
never take place in the centrally extended BMS3 algebra. However, for other semi-
direct product groups, such extensions may occur; an example is the symmetry group
of warped conformal field theories [93], Diff(S 1 ) n C ∞ (S 1 ).
g∗ ⊕ b
The space of coadjoint vectors dual to the algebra (9.42) is a direct sum b g∗ , or more
g∗ ⊕ b
accurately b g∗Ab . Following the notation of section 5.4, its elements are quadruples
(j, c1 ; p, c2 ) (9.49)
where (j, c1 ) is a centrally extended angular momentum dual to b g, while (p, c2 ) is a
centrally extended momentum dual to b gAb . The real numbers c1 , c2 are central charges;
the first pairs rotation generators with themselves, while the second pairs rotations
with translations. The pairing of (9.49) with b gAb gAb is
(j, c1 ; p, c2 ), (X, λ; α, µ) = hj, Xi + hp, αi + c1 λ + c2 µ , (9.50)
where the two pairings h·, ·i on the right-hand side are those of g∗ with g and g∗Ab with
gAb , respectively. This is a centrally extended generalization of (5.109).
Recall that the coadjoint representation of a semi-direct product involves a cross

product (5.110). For the centreless exceptional semi-direct product (9.38), we have
(5.110) (5.11)
hα × p, Xi = hp, adX αi = − hp, adα Xi = had∗α p, Xi
where ad and ad∗ denote the adjoint and coadjoint representations of g, respectively.
In other words,
α × p = ad∗α p (9.51)
where we abuse notation slightly by acting with an element of gAb on an element of

g∗Ab . Using (5.113), it readily follows that the coadjoint representation of the centreless
semi-direct product (9.38) is given by
Ad∗(f,α) (j, p) = Ad∗f j, +ad∗α Ad∗f p, Ad∗f p

(9.52)
where the Ad∗ on the right-hand side is the coadjoint representation of G. For exam-
ple, when G = SL(2, R), this formula is the transformation law of relativistic angular
momentum j and energy-momentum p under Poincaré transformations in three dimen-
sions. From (9.52) we also find that the coadjoint representation of the Lie algebra
g A gAb is
ad∗(X,α) (j, p) = ad∗X j + ad∗α p, ad∗X p

(9.53)
in accordance with eq. (5.114).
The centrally extended generalization of these considerations is straightforward, if

mildly technical. Using eq. (6.104) for the coadjoint action of G,
b formula (9.52) yields
the coadjoint representation of Gbnb gAb :
c ∗ (j, c1 ; p, c2 ) =
Ad(f,α)

∗ c1 −1 ∗
h
∗ c2 −1
i c
2 ∗ c2 −1

= Adf j − S[f ] + adα Adf p − S[f ] + s[α], c1 ; Adf p − S[f ], c2 .
12 12 12 12
(9.54)
Here it is understood that all Ad∗ ’s and ad∗ ’s on the right-hand side are centreless —
they are the coadjoint representations of G and g, respectively.
Formula (9.54) looks a bit scary but it is crucial for our purposes, so let us briefly
point out two of its important features. First, the central charges c1 , c2 are left invariant
by the action of the group, as expected. Second, note that the transformation law of
momentum is
c2
f · p = Ad∗f p − S[f −1 ] , (9.55)
12
where the Ad∗ on the right-hand side is that of G (not G). b This formula says that p is
invariant under translations (since it is unaffected by α) and that its transformation
law is blind to the central charge c1 , but not to c2 . In fact, eq. (9.55) is the coadjoint
representation (6.104) of the centrally extended group G b at central charge c2 . As
a corollary we can already conclude that the orbits of momenta labelling unitary
representations of (9.39) are coadjoint orbits of the group G b at fixed central charge c2 ;
there is no need to master the much more complicated transformation law of angular
momentum in (9.54) in order to classify such representations. This will have key
consequences for the BMS3 group below.
Remark. Property (9.51) explains why we refer to the map (5.110) as a cross product.
Indeed, the (double cover of the) Euclidean group in three dimensions is an exceptional
semi-direct product SU(2) nAd su(2)Ab . Since the coadjoint representation of SU(2)
is equivalent to the adjoint, one may identify vectors with covectors and the cross
product (9.51) for the Euclidean group can be rewritten as α × p = adα p = [α, p].
Here the Lie bracket is that of su(2), so in components one has (α × p)i = ijk αj pk ,
which is the standard definition of the cross product in mechanics.
9.2.2 Defining BMS3

Now that we are acquainted with exceptional semi-direct products, let us show how
this structure occurs in three-dimensional BMS symmetry.
Centreless BMS3 group

Our first task is to move backwards from the centreless BMS3 algebra (9.23) to the
corresponding group. The algebra consists of pairs X(ϕ), α(ϕ) , where X(ϕ)∂ϕ is a
vector field on the circle while α(ϕ) is a priori just a function on the celestial circle.
These two quantities were referred to above as infinitesimal superrotations and super-
translations, respectively. Together, they generate finite transformations (9.12) of the
cylinder at null infinity, where f (ϕ) is a diffeomorphism of the circle. Thus we already
know that the BMS3 group consists of pairs (f, α), where f is a diffeomorphism of
the circle while α is a function. It only remains to work out the group operation; the
latter is given by the composition of two transformations (9.12):
(g,β)
(u, ϕ) 7−→ g 0 (ϕ)u + β(g(ϕ)), g(ϕ)

(f,α)

7−→ f 0 (g(ϕ)) g 0 (ϕ)u + β(g(ϕ)) + α f (g(ϕ)) , f (g(ϕ)) .

Here the last result on the right-hand side can be rewritten as

(f ◦ g)0 (ϕ)u + [α + Adf β] , (f ◦ g)(ϕ) (9.56)
(f ◦g)(ϕ)
where Adf β denotes the adjoint representation (6.17) of Diff(S 1 ) acting on β, that is,
the transformation law of a vector field β(ϕ)∂ϕ under f (ϕ):
(Adf β) f (ϕ)
= f 0 (ϕ)β(ϕ). (9.57)
Expression (9.56) indicates three things:

1. The group operation of superrotations is given by composition (6.8); hence finite
(as opposed to infinitesimal) superrotations span a group Diff(S 1 ).5
2. If it wasn’t for superrotations, the group operation of supertranslations would
just be addition, α · β ≡ α + β. Thus supertranslations span an Abelian additive
group whose elements are certain functions on the circle.
3. The action of superrotations on supertranslations is that of diffeomorphisms on
vector fields, i.e. it is the adjoint representation (9.57) of Diff(S 1 ). In partic-
ular, supertranslations, which so far we thought of as functions α(ϕ) on the
circle, should better be seen as vector fields α(ϕ)∂ϕ . The only subtlety is that
these vector fields do not generate diffeomorphisms of celestial circles, but rather
angle-dependent translations (9.11) of retarded time u. Equivalently, each su-
pertranslation is a density α = α(ϕ)(dϕ)−1 on the circle.
These observations motivate the following definition:
5
As in chapter 6 we describe diffeomorphisms of the circle by their lifts belonging to the universal
g + (S 1 ), which we abusively denote simply as Diff(S 1 ).
cover Diff
Definition. The centreless BMS group in three dimensions is the exceptional semi-
direct product
BMS3 ≡ Diff(S 1 ) nAd Vect(S 1 )Ab (9.58)
where Diff(S 1 ) is the group of diffeomorphisms of the circle while Vect(S 1 )Ab is its
Lie algebra, seen as an Abelian vector group acted upon by Diff(S 1 ) according to the
adjoint representation. Its elements are pairs (f, α); its group operation follows from
the general definition (4.6) and is given by

(f, α) · (g, β) = f ◦ g, α + Adf β (9.59)
where Ad is the action (9.57) of Diff(S 1 ) on vector fields. With this definition the
action (9.12) of BMS3 on null infinity reproduces the group operation (9.59).
The BMS3 group is infinite-dimensional and has the announced form (9.38), with
G = Diff(S 1 ). Since we saw in section 6.1 that PSL(2, R) is a subgroup of Diff(S 1 ),
the Poincaré group is obviously a subgroup of BMS3 . We therefore introduce officially
the following terminology:
Definition. In the BMS3 group (9.58), elements of Diff(S 1 ) are known as superro-
tations while elements of Vect(S 1 )Ab are called supertranslations.
Remark. The name “superrotation” has come to be standard, but the geometric
interpretation of Diff(S 1 ) makes the terminology “superboosts” somewhat more ap-
propriate. Indeed, recall from section 6.1 that the group Diff(S 1 ) is homotopic to a
circle, so that the only superrotations spanning a compact group are those conjugate
to rigid rotations f (ϕ) = ϕ + θ. The other one-parameter subgroups of Diff(S 1 ) are
all non-compact and should be seen as boost groups.
Universal cover of BMS3

As in section 6.1 we should be careful about what we mean by Diff(S 1 ). Strictly
speaking, Diff(S 1 ) consists of all diffeomorphisms of the circle with the composition
law (6.2); its connected subgroup Diff+ (S 1 ) consists of orientation-preserving diffeo-
morphisms. Since the group of supertranslations is a vector space, it is also connected
and we define the connected BMS3 group as
BMS+ + 1 1
3 ≡ Diff (S ) nAd Vect(S )Ab . (9.60)
If we think of the group of superrotations as an extension of the Lorentz group
in three dimensions, then Diff(S 1 ) corresponds to the disconnected orthochronous
Lorentz group O(2, 1)↑ while Diff+ (S 1 ) corresponds to the connected (orthochronous
and proper) Lorentz group SO(2, 1)↑ . It appears that no Diff(S 1 ) transformation cor-
responds to time reversal (which sounds reasonable since BMS symmetry is defined
separately at future and past null infinity).
The group Diff+ (S 1 ) of orientation-preserving superrotations is homotopic to a

circle, so it admits topological projective transformations that can be dealt with by
g + (S 1 ). Since the vector group of supertranslations
trading it for its universal cover, Diff
is homotopic to a point, the BMS3 group has the homotopy type of a circle.
Definition. The universal cover of the BMS group in three dimensions is the excep-
tional semi-direct product
]+
BMS g+ 1 1
3 ≡ Diff (S ) nAd Vect(S )Ab (9.61)
g + (S 1 ) is the universal cover of the connected group Diff+ (S 1 ) and consists

where Diff
of 2πZ-equivariant superrotations (6.7).
]+
In particular, exact representations of BMS 3 generally correspond to projective rep-
resentations of BMS3 . The groups BMS3 , BMS+
+ ]+
3 and BMS3 are well-defined infinite-
dimensional Lie-Fréchet groups. In what follows, motivated by quantum-mechanical
applications, we always focus (implicitly) on the universal cover. Accordingly we abuse
notation and denote the universal cover simply by BMS3 , neglecting the superscript
“+” and the tilde.
Centrally extended BMS3 group

In order to define the central extension of BMS3 , we apply the prescription (9.39) for
centrally extended exceptional semi-direct products to the case G = Diff(S 1 ):
Definition. The centrally extended BMS group in three dimensions is the exceptional
semi-direct product
d 1 ) n c Vect(S
[ 3 ≡ Diff(S
BMS d 1) (9.62)
Ad
d 1 ) is the (universal cover of the) Virasoro group.

where Diff(S
[ 3 , let us make its definition a bit

Since this thesis is concerned with the group BMS

more explicit before going further. The elements of BMS [ 3 are quadruples f, λ; α, µ
where f is a superrotation, α a supertranslation, while λ, µ are real numbers, extend-
ing Poincaré transformations as before. In BMS [ 3 , centrally extended superrotations
(f, λ) span a Virasoro group while extended supertranslations (α, µ) span an infinite-
dimensional Abelian group acted upon by superrotations according to the Virasoro
adjoint representation. Explicitly, the group operation in BMS [ 3 takes the form (9.41)
where f · g = f ◦ g, while C is the Bott-Thurston cocycle (6.69) and S is the Schwarzian
derivative (6.76). The pairing h·, ·i is that of Vect(S 1 ) with its dual, given by (6.34).
The centreless BMS3 group (9.58) is perfect, in the same way as Diff(S 1 ); this
implies that it admits a universal central extension. As it turns out, this is precisely
[ 3 (see section 9.2.5 for the proof):
achieved by BMS
Theorem. The centrally extended BMS3 group (9.62) is the universal central exten-
sion of the centreless BMS3 group (9.58).
9.2.3 Adjoint representation and bms3 algebra

Lie algebra
Since BMS3 is an exceptional semi-direct product, its centreless Lie algebra takes the
form g A gAb , where g is the Lie algebra of Diff(S 1 ):
bms3 = Vect(S 1 ) Aad Vect(S 1 )Ab . (9.63)
Its elements are pairs (X, α) where X = X(ϕ)∂ϕ is an infinitesimal superrotation

and α = α(ϕ)(dϕ)−1 an infinitesimal supertranslation. These functions determine the
components of vector fields (9.17) generating asymptotic symmetries, so that elements
of bms3 can be seen as infinitesimal BMS3 transformations. In particular the Poincaré
subalgebra of bms3 consists of pairs (X, α) whose only non-vanishing Fourier modes are
the three lowest ones, as in (9.7)-(9.9). The centrally extended generalization (9.42)
of this definition is immediate:
Definition. The Lie algebra of BMS

[ 3 is an exceptional semi-direct sum
d 3 ≡ Vect(S
bms d 1 )Ab .
d 1 ) Ac Vect(S (9.64)
ad
Its elements are quadruples (X, λ; α, µ) where X = X(ϕ)∂ϕ is an infinitesimal superro-

tation, α = α(ϕ)(dϕ)−1 an infinitesimal supertranslation, while λ, µ are real numbers.
Adjoint representation
The adjoint representation of the centreless BMS3 group is given by formula (9.43),
where the adjoint action of Diff(S 1 ) is the transformation law of vector fields (6.18).
An important subtlety is that the Lie bracket appearing on the right-hand side is
that of the Lie algebra of Diff(S 1 ) and is therefore the opposite (6.21) of the standard
bracket of vector fields. Accordingly, in terms of the usual Lie bracket of vector fields
on the circle one would write the adjoint representation of BMS3 as

Ad(f,α) (X, β) = Adf X, Adf β + [Adf X, α] , (9.65)
with a plus sign instead of a minus sign in the second entry of (9.43). The centrally
extended generalization of that expression is provided by eq. (9.45), where S is the
Schwarzian derivative (6.76), s is its infinitesimal version (6.74), and h·, ·i is the stan-
dard pairing (6.34). Again, when writing the adjoint representation in terms of the
standard Lie bracket of vector fields, the sign in front of the bracket of the third entry
of (9.45) is a plus instead of a minus. Since we will not explicitly need the adjoint
[ 3 group, we do not display it here.
representation of the BMS
Lie brackets
From the adjoint representation one can read off the Lie bracket of the bms
d 3 algebra.
In order to absorb the minus sign of (6.21) we define the bracket to be
d c
(X, λ; α, µ), (Y, ρ; β, ν) ≡ − Ad (etX ,tα) (Y, ρ; β, ν) .
dt t=0
With this definition the Lie bracket in bms

d 3 takes the form (9.46) where the brackets
on the right-hand side are standard Lie brackets of vector fields while c is the Gelfand-
Fuks cocycle (6.43). This is consistent with the algebra of surface charges (9.36).
The Lie algebra structure can be made more apparent by writing the bracket
(9.46) in a suitable basis. As in (9.24) we define the complex superrotation and
supertranslation generators of the centreless bms3 algebra,
jm ≡ eimϕ ∂ϕ , 0 , pm ≡ 0, eimϕ (dϕ)−1 ,

(9.66)
where the index m runs over all integers. Their brackets take the form (9.10). The
corresponding basis of the centrally extended bms
d 3 algebra is
(9.66)
Jm ≡ jm , 0; 0, 0 = eimϕ ∂ϕ , 0; 0, 0 ,

(9.66) (9.67)
Pm ≡ 0, 0; pm , 0 = 0, 0; eimϕ (dϕ)−1 , 0 ,

together with two central elements (9.47), i.e. Z1 = (0, 1; 0, 0) and Z2 = (0, 0; 0, 1). In
these terms the centrally extended bracket (9.46) yields
Z1 3
i[Jm , Jn ] = (m − n)Jm+n + m δm+n,0 ,
12
Z2 3
i[Jm , Pn ] = (m − n)Pm+n + m δm+n,0 , (9.68)
12
i[Pm , Pn ] = 0 .
Up to central terms this is of the same form as the asymptotic symmetry algebra (9.10),
and it is consistent with the general form (9.48) for centrally extended exceptional
semi-direct products. In the first line we see that superrotations close according to
a Virasoro algebra (6.110) with central generator Z1 , while the second line shows
that brackets of superrotations with supertranslations take the Virasoro form with a
different central element Z2 . The algebra of surface charges (9.37) takes that form,
with definite values c1 = 0, c2 = 3/G for the central generators Z1 , Z2 .
Remark. The canonical Poincaré subgroup of BMS3 is the one spanned by super-
rotations (6.88) and supertranslations (9.7), or equivalently the one generated by ba-
sis elements jm , pm with m = −1, 0, 1. But in fact, BMS3 admits infinitely many
other Poincaré subgroups: each of them has a Lie algebra spanned by jn , j0 , j−n and
pn , p0 , p−n , consisting of superrotations of the form (6.95) and supertranslations
α(ϕ) = α0 − α1 cos(nϕ) − α2 sin(nϕ)
whose only non-vanishing Fourier modes are the zero-mode and the nth modes.
9.2.4 Coadjoint representation

Angular and linear supermomentum

[ 3 are quadruples j, c1 ; p, c2 where j = j(ϕ)dϕ2 and
The coadjoint vectors of BMS
p = p(ϕ)dϕ2 are quadratic densities on the circle, respectively dual to infinitesimal
superrotations and supertranslations. The coefficients c1 and c2 are central charges.

The pairing of (j, c1 ; p, c2 ) with the Lie algebra bms
d 3 is given by formula (9.50), or
explicitly
Z 2π
1
(j, c1 ; p, c2 ), (X, λ, α, µ) = dϕ j(ϕ)X(ϕ) + p(ϕ)α(ϕ) + c1 λ + c2 µ .
2π 0
The right-hand side of this expression coincides (up to central terms) with the surface
charge (9.31). Inspired by the terminology of superrotations and supertranslations,
we introduce the following nomenclature:
Definition. Let (j, p) be a coadjoint vector for the BMS3 group. Then p = p(ϕ)dϕ2
is called a supermomentum while j = j(ϕ)dϕ2 is an angular supermomentum.
The embedding of the Poincaré algebra in bms3 suggests an interpretation for the
lowest Fourier modes of
X X
j(ϕ) = jm e−imϕ and p(ϕ) = pm e−imϕ . (9.69)
m∈Z m∈Z
Indeed, p0 is dual to time translations and should be interpreted as the energy as-
sociated with p(ϕ); similarly, the components p1 and p−1 = p∗1 are complex linear
combinations of the spatial components of momentum. As for j0 , it is the angular
momentum associated with j(ϕ), while j1 and j−1 are centre of mass charges. More
generally, the function p(ϕ) should be seen as an energy density on the circle — es-
sentially a stress tensor — while j(ϕ) is an angular momentum density on the circle.
In particular, it is natural to give dimensions of energy to the function p(ϕ) and the
central charge c2 , while the function j(ϕ) and the central charge c1 are dimensionless.
This interpretation is confirmed by the surface charges (9.31), since p(ϕ) is a Bondi
mass aspect while j(ϕ) is an angular momentum aspect; furthermore the central charge
(9.29) is indeed a mass scale.
Remark. To our knowledge the terminology of “supermomentum” for duals of su-

pertranslations dates back to [269], and has subsequently been used throughout the
BMS literature (see e.g. [270–274]). The terminology of “angular supermomentum”,
on the other hand, seems to have first appeared in [275, 276] in relation to the prob-
lem of angular momentum, but apparently without direct relation to BMS symmetry.
It was later independently introduced in [46] in the BMS context. In [274], angular
supermomentum is referred to as “super centre of mass”.
As in the Virasoro case, one should think of the pair (j, p) as the stress tensor of a
BMS3 -invariant theory; its transformations under BMS3 then coincide with the coad-
joint representation, given for centrally extended exceptional semi-direct products by
formula (9.54). In that expression, the central charges are invariant (as they should)
while the transformation law of supermomentum coincides with the coadjoint repre-
sentation (6.114) of the Virasoro group at central charge c2 :
1 h c2 i
f · p f (ϕ) = 2 p(ϕ) + S[f ](ϕ) , (9.70)
f 0 (ϕ) 12
where S denotes the Schwarzian derivative (6.76). We stress once more that supermo-
mentum is left invariant by supertranslations, as it should.
The transformation law of angular supermomentum is a bit more involved and

translates the fact that j is sensitive both to superrotations and to supertranslations, as
it should since angular momentum and centre of mass charges are always defined with
respect to an arbitrarily chosen origin. We refrain from describing this transformation
law any further at this point, as we shall return to it in section 9.3.1 when showing that
the phase space of metrics (9.25) is a hyperplane at central charges c1 = 0, c2 = 3/G
embedded in the coadjoint representation of BMS [ 3.
Kirillov-Kostant bracket
A prerequisite for showing that the asymptotically flat phase space is a coadjoint
representation is to understand the Kirillov-Kostant Poisson bracket of the asymptotic
symmetry
∗ ∗ group.
[ 3 ; we proceed as in section 6.4. Thus let
Let us do this here for BMS
∗ ∗
Jm , Pm , Z1 , Z2 be the dual basis corresponding to (9.67) and (9.47). Writing any
coadjoint vector as
X
j(ϕ)dϕ2 , c1 ; p(ϕ)dϕ2 , c2 = jm Jm∗ + pm Pm
∗
+ c1 Z1∗ + c2 Z2∗ ,

m∈Z
d 3 ∗ . Their
the components {jm , pm , c1 , c2 } are global coordinates on the dual space bms
Poisson brackets (5.28) take the form
c1 3
i{jm , jn } = (m − n)jm+n + m δm+n,0 ,
12
c2
i{jm , pn } = (m − n)pm+n + m3 δm+n,0 , (9.71)
12
i{pm , pn } = 0 .
As is obvious here, c1 is a genuine Virasoro central charge for superrotations, while

c2 is the (generally dimensionful) central charge pairing superrotations with super-
translations. The surface charges of asymptotically flat gravity satisfy the exact same
algebra (9.37), with the values of central charges c1 = 0, c2 = 3/G. We will return to
this in section 9.3.
*9.2.5 Some cohomology

To conclude our abstract description of BMS3 symmetry, we now show that the cen-
trally extended group (9.62) is in fact the universal central extension of the BMS3
group (9.58). (In both cases Diff(S 1 ) is understood to denote the universal cover of
the group of orientation-preserving diffeomorphisms of the circle.) We use the notation
of section 2.2. Since the proof is very similar to the construction of the Gelfand-Fuks
cocycle (6.43), this section may be skipped in a first reading.
The bms3 algebra is perfect: it is equal to its Lie bracket with itself. This can
be seen, for instance, by noting that the right-hand sides of the brackets (9.10) span
all possible bms3 generators. Accordingly it follows from (2.19) that the first real
cohomology of bms3 vanishes: H1 (bms3 ) = 0. Since the same is true of the centreless
BMS3 group, its central extension is universal, and it only remains to establish the
second cohomology of bms3 .
Theorem. The second real cohomology space of bms3 is two-dimensional. It is

generated by the classes of the two-cocycles

c1 (X, α), (Y, β) = c(X, Y ) and c2 (X, α), (Y, β) = c(X, β) − c(Y, α) (9.72)
where c is the Gelfand-Fuks cocycle (6.43). Their expression in the basis (9.66) is
m3
c1 (jm , jn ) = c2 (jm , pn ) = −i δm+n,0 , (9.73)
12
while their other components vanish. As a consequence, the Lie algebra (9.64) is
the universal central extension of (9.63), and the group (9.62) is the universal central
extension of (9.58).
Proof. The fact that the cocycle c1 is the only non-trivial cocycle pairing superrotation
generators with themselves follows from the fact that infinitesimal superrotations span
a Witt subalgebra of bms3 . The considerations of section 6.2 then carry over directly
to bms3 . Now let us ask whether there exists a two-cocycle c such that c(pm , pn ) 6= 0.
The cocycle identity (2.21) with trivial T implies
!
c(j0 , [pm , pn ]) + c̃(pm , [pn , j0 ]) + c̃(pn , [j0 , pm ]) = (m + n)c̃(pm , pn ) = 0 ,
where we used the Lie brackets (9.10). This yields c̃(pm , pn ) = c̃m δm+n,0 where the
coefficients c̃m = −c̃−m are to be determined. We now attempt to find a recursion
relation for these coefficients; using the cocycle identity
c(p−1 , [j−m+1 , pm ]) + c(j−m+1 , [pm , p−1 ]) + c(pm , [p−1 , j−m+1 ]) = 0 ,
the bms3 algebra (9.10) implies (2m − 1)c̃1 + (m − 2)c̃m = 0. Since this must be true
for all integer values of m we conclude that c̃1 = 0, which in turn implies c̃m = 0
for all m ∈ Z. Thus, any two-cocycle on the bms3 algebra has vanishing components
c̃(pm , pn ) = 0. Finally, suppose that c is a two-cocycle on the bms3 algebra and let
us ask whether one can have c(jm , pn ) 6= 0. As in (6.46) we start by ensuring that
the cocycle c is rotation-invariant by adding to it a suitable coboundary. Consider
therefore the cocycle relation
c(j0 , [jm , pn ]) = c([j0 , jm ], pn ) + c(jm , [j0 , pn ]) .
The left-hand side can be interpreted as the differential of the one-cochain k = c(j0 , ·)
evaluated at (jm , pn ), while the right-hand side is the Lie derivative of c with respect
to j0 . Since the left-hand side is exact we know that the cohomology class of c is
left invariant by rotations; in particular we can add to c the differential db of the
one-cochain
i
b(jm ) ≡ 0 , b(pm ) ≡ c(j0 , pm ) ,
m
which is such that
Lj0 (c + db)(jm , pn ) = 0 . (9.74)
From now on we simply write c to denote c + db. Then, analogously to (6.50), eq.
(9.74) implies (m + n)c(jm , pn ) = 0 by virtue of the brackets (9.10). In particular we
can now write c(jm , pn ) = cm δm+n,0 and it only remains to find the coefficients cm . For
this we derive a recursion relation using the cocycle identity
c(p1 , [j−m−1 , jm ]) + c(j−m−1 , [jm , p1 ]) + c(jm , [p1 , j−m−1 ]) = 0 ,
which implies
(2m + 1)c−1 + (m − 1)c−m−1 + (m + 2)cm = 0 (9.75)
by virtue of the bms3 algebra (9.10). In particular we have c0 = 0 and c1 = −c−1 ,
which then gives cm = −c−m and the recursion relation (9.75) can be rewritten as
(m + 2)cm − (2m + 1)c1

cm+1 = .
m−1
This is the same relation as in the Virasoro case, eq. (6.52). In particular it is solved by
cm = m and cm = m3 , the former being a coboundary. The result (9.73) follows.
9.3 The BMS3 phase space

As explained in section 8.3, the space of solutions of a Hamiltonian system coincides
with its phase space. Accordingly the on-shell metrics (9.25) span the phase space
of asymptotically flat gravity in three dimensions. Here we show that this space is
a hyperplane at fixed central charges c1 = 0, c2 = 3/G embedded in the coadjoint
[ 3 group. We also discuss this result from a holographic
representation of the BMS
perspective and derive a positive energy theorem.
9.3.1 Phase space as a coadjoint representation

The space of on-shell metrics (9.25) is spanned by pairs (j, p) transforming under
BMS3 according to (9.27)-(9.28). We now show that these formulas coincide with
the coadjoint representation of the bms d 3 algebra at central charges c1 = 0, c2 =
3/G. This is trivially true for the transformation law of p(ϕ) since (9.28) coincides
with the coadjoint representation (6.115) of the Virasoro algebra, which in turn is
the infinitesimal version of the transformation law (9.70). As pointed out in section
9.2.4, the case of the angular momentum aspect is more intricate since its coadjoint
transformation law is the first entry on the right-hand side of (9.54). When applied
to BMS3 , the latter formula must be modified slightly to match our conventions for
Diff(S 1 ). Namely, due to the minus sign of the Lie bracket (6.21), the ad∗ of eq. (9.54)
should be replaced by −ad∗ . Taking this subtlety into account, the transformation
law of angular supermomentum is
c1 h c2 i c
2
(f, α) · j = Ad∗f j − S[f −1 ] − ad∗α Ad∗f p − S[f −1 ] + s[α] . (9.76)
12 12 12
Here Ad∗ denotes the coadjoint representation (6.36) of Diff(S 1 ), S is the Schwarzian
derivative (6.76), s is its infinitesimal cousin (6.74), and ad∗ is the infinitesimal coad-
joint representation (6.37) so that ad∗α p = αp0 + 2α0 p. In order to relate formula (9.76)
to the transformation law of the angular momentum aspect, we take an infinitesimal

superrotation f (ϕ) = ϕ + X(ϕ), an infinitesimal supertranslation α(ϕ), and define
the variation of j by
(f, α) · j − j
δ(X,α) j ≡ − .

As a result we obtain
c1 000 c2
δ(X,α) j = Xj 0 + 2X 0 j − X + αp0 + 2α0 p − α000 ,
12 12
which exactly coincides with (9.27) when c1 = 0, as expected.
Using the fact that the Poisson algebra of surface charges (9.37) coincides with
the Kirillov-Kostant bracket (9.71), we conclude that the (covariant) phase space of
three-dimensional asymptotically flat gravity with BMS boundary conditions is a hy-
perplane c1 = 0, c2 = 3/G embedded in the space of the coadjoint representation of the
[ 3 group and endowed with its Kirillov-Kostant Poisson bracket. This observation
BMS
is the flat space analogue of the statement that the subleading components of an AdS
space-time metric contain one-point functions of the dual CFT stress tensor. As in
AdS, this observation should not come as a surprise. Indeed, the coadjoint representa-
tion of BMS3 was bound to appear in the transformation law of the momentum map
of the system, and it just so happens that this map is determined by the entries of the
metric (9.25). The truly surprising aspect of this observation is the fact that it is the
entries of the metric, and not some non-linear combinations thereof, that determine the
momentum map. In particular, as in AdS3 , the set of solutions (9.25) is a vector space.
In view of these results, one may ask whether the subleading components of asymp-
totically flat metrics can be interpreted as the components of the stress tensor of some
dual theory, similarly to AdS/CFT. The notion of “dual theory” appears to be elusive
in the asymptotically flat case, essentially because the metric becomes degenerate at
null infinity, but the question can be answered regardless of this complication. In-
deed, whatever the dual theory is, it must be such that its stress tensor transforms
under the coadjoint representation of the BMS3 group (generally with some non-zero
central charges), by virtue of the very nature of momentum maps. Accordingly, the
stress tensor T of any BMS3 -invariant theory is necessarily such that Tuu = p(ϕ) is a
supermomentum generating supertranslations, while Tuϕ = j(ϕ) is an angular super-
momentum generating superrotations.
This being said, it would be reassuring to have explicit field-theoretic illustrations

of the fact that (j, p) actually is the stress tensor of some two-dimensional field the-
ory. Such an illustration is provided by [198] (see also [277]), where a two-dimensional
field theory invariant under BMS3 was obtained thanks to the “dimensional reduction”
of three-dimensional gravity through the Chern-Simons formalism. As expected, the
stress tensor of that theory is a pair (j, p) that coincides with the functions speci-
fying the metric (9.25), and whose BMS3 transformations exactly take the form of
the coadjoint representation (9.54) with central charges c1 = 0, c2 = 3/G [278]. The
higher-spin [279] and supersymmetric [167,280] generalizations of these considerations
confirm this statement, so known examples of BMS3 -invariant field theories do sup-
port our claim that the functions (j, p) coincide with the components of a “dual” stress
tensor.
9.3.2 Boundary gravitons and BMS3 orbits

If one picks a metric (9.25) at random, the pair j(ϕ), p(ϕ) is most likely to consist
of functions that are not constant on the circle. This is actually implied by BMS3
symmetry: if we let (j, p) be any seed solution (with j, p constant or not), the set of
metrics obtained from it by asymptotic symmetry transformations spans an infinite-
[ 3 group at central charges c1 = 0, c2 = 3/G,
dimensional coadjoint orbit of the BMS
W(j,c1 ;p,c2 ) . (9.77)
The metrics belonging to this orbit are infinite-dimensional analogues of Poincaré

transforms of the state of a particle with momentum p and angular momentum j.
As in section 8.3 one may refer to the orbit (9.77) as a space of classical “boundary
gravitons” around the background (j, p).
The fact that the phase space of flat gravity coincides with (a hyperplane in) the
coadjoint representation of BMS [ 3 allows us to use the orbits (9.77) as an organizing
principle. As in fig. 8.6, the space of solutions is foliated into disjoint BMS [ 3 orbits,
each of which is a symplectic manifold. Since the classification of coadjoint orbits of
[ 3 follows from the results of section 5.4, we may claim to control the full covariant
BMS
phase space of asymptotically flat gravity. In particular the classification of zero-mode
solutions in fig. 9.1 is a first step towards the full classification: each point in the plane
(J, M ) determines an orbit (9.77), and different points define disjoint orbits. Since
not all orbits have constant representatives, fig. 9.1 is an incomplete representation
of the full phase space of the system. The complete picture would involve the BMS3
analogue of fig. 7.3. Note that the relation between metrics and BMS [ 3 orbits hints
that the quantization of asymptotically flat gravity produces unitary representations
of BMS3 . We will investigate this proposal in chapters 10 and 11.
9.3.3 Positive energy theorem

Positive energy theorems in general relativity are commonly formulated in asymp-
totically flat space-times, so we are now in position to address the three-dimensional
version of that problem. The question that we wish to ask is the following: which
asymptotically flat metrics (9.25) have energy bounded from below under BMS3 trans-
formations?
The answer follows from the fact that asymptotically flat metrics transform un-
der BMS3 according to the coadjoint representation (9.54). For our purposes the key
property of that formula is the fact that the transformation law of p is blind to super-
translations. In this sense the positive energy theorem in three-dimensional flat space
is even simpler than in AdS3 :
Positive energy theorem. The asymptotically flat metric (j, p) has energy boun-
ded from below under BMS3 transformations if and only if p belongs to a Virasoro
coadjoint orbit (at central charge c2 = 3/G) with energy bounded from below. This is
to say that either p is superrotation-equivalent to a constant p0 ≥ −c2 /24, or p belongs
to the unique massless Virasoro orbit with bounded energy.
As a corollary, we now know that all conical deficits and all flat space cosmologies
have energy bounded from below under BMS3 transformations. By contrast, all conical
excesses have energy unbounded from below.
9.4 Flat limits

There are many similarities between the asymptotic symmetries of three-dimensional
Anti-de Sitter and flat space-times. Intuitively, this is because the limit ` → +∞ (i.e.
Λ → 0) of AdS3 is just Minkowski space. It is tempting to ask if the phenomenon can
be formulated in a mathematically precise way such that the conclusions of section
9.1 follow from those of section 8.2 by a suitably defined flat limit. This question
was addressed in [281], and the answer is yes. In short, upon reformulating Brown-
Henneaux boundary conditions in Bondi-like coordinates at null (rather than spatial)
infinity, the Minkowskian asymptotic Killing vectors (9.17), the on-shell metrics (9.25)
and the surface charges (9.31) are flat limits of their AdS3 counterparts displayed in
eqs. (8.30), (8.38) and (8.42) respectively. In particular, BMS3 symmetry may be seen
as a limit of two-dimensional conformal symmetry.
In this section we explore this flat limit from the point of view of group theory,
starting from the definition of the AdS3 asymptotic symmetry group as a set of con-
formal transformations of a time-like cylinder. We describe the limit at the level of
groups, then at the level of Lie algebras, and finally at the level of the coadjoint repre-
sentation. We end by pointing out a different contraction that produces the Galilean
conformal algebra in two dimensions. Considerations related to flat limits of unitary
representations are relegated to section 10.2.
9.4.1 From Diff(S 1 ) to BMS3

In Anti-de Sitter space, spatial infinity coincides with null infinity. This observation
allows one to reformulate Brown-Henneaux boundary conditions (originally defined at
spatial infinity) in terms of Bondi-like coordinates (r, ϕ, u) at null infinity [281]. The
conclusion of this reformulation is that AdS3 results take the same form as in the
standard Fefferman-Graham gauge, up to the replacement of the time coordinate t by
a retarded time coordinate u. In particular one can introduce light-cone coordinates
u
x± ≡ ±ϕ (9.78)
`
in terms of which the asymptotic symmetry group acts on the cylinder at (null) infinity
according to conformal transformations (8.37). Our goal here is to start from these
transformations and rediscover the BMS3 transformations (9.12).
The way to go is to expand everything in powers of a “small” parameter = 1/`.

In practice ` is dimensionful so it makes no sense to think of it as being “large”; a more
precise statement would be that the dimensionless Brown-Henneaux central charge,
proportional to `/G, must go to infinity. Despite this subtlety we will keep referring to
` as a “large” parameter, keeping in mind that there exists a more precise formulation
of the procedure.
In order to distinguish Virasoro elements from those of BMS3 , we denote elements

of the group Diff(S 1 ) × Diff(S 1 ) as pairs (F, F̄) where F and F̄ are lifts of orientation-
preserving diffeomorphisms of the circle satisfying the conditions (6.7). Let then (F, F̄)
be a conformal transformation of the cylinder with coordinates (9.78). In the large `
limit the transformation of the angular coordinate ϕ becomes
1 `→+∞ 1
F(x+ ) − F̄(x− ) →

ϕ 7→ F(ϕ) − F̄(−ϕ) , (9.79)
2 2
where the combination of F’s on the far right-hand side was obtained by Taylor-
expanding functions around ±ϕ in terms of the small parameter u/`, and neglecting
all terms of order O(1/`). The combination of diffeomorphisms in (9.79) is itself a (lift
of a) diffeomorphism of the circle. Indeed one readily verifies that
1
f (ϕ) ≡ (F(ϕ) − F̄(−ϕ)) (9.80)
2
satisfies the conditions (6.7) when F and F̄ do. Let us investigate what happens with
the time coordinate u in the same limit. Using (9.80) we find
` `→+∞ `
F(x+ ) + F̄(x− ) → F(ϕ) + F̄(−ϕ) + f 0 (ϕ)u,

u 7→
2 2
where all terms O(1/`) were neglected once more. The first term on the far right-
hand side is potentially divergent: typical diffeomorphisms are independent of `, so
the first term goes to infinity in the large ` limit. Note, however, that the combination
F(ϕ) + F̄(−ϕ) is 2π-periodic. Thus, in order for the limit ` → +∞ to work we require
that there be a finite, `-independent function α on the circle such that
2
F(ϕ) + F̄(−ϕ) ≡ α(f (ϕ)) + O(1/`2 ) (9.81)
`
where the argument of α is taken to be f (ϕ) for convenience. This is to say that the
diffeomorphisms F and F̄ are required to depend on ` in such a way that
F̄(−ϕ) = −F(ϕ) + O(1/`). (9.82)
For instance, if F(ϕ) = ϕ + θ and F̄(ϕ) = ϕ + θ̄ are rotations, this condition says that
θ − θ̄ goes to zero at least as fast as 1/` in the large ` limit. With this choice the trans-
formation law of u reduces to u 7→ f 0 (ϕ)u + α(f (ϕ)). Including (9.79), we have thus
reproduced the BMS3 transformations (9.12) from a flat limit of Diff(S 1 )×Diff(S 1 ). In
this sense the centreless BMS3 group (9.58) is a flat limit of the asymptotic symmetry
group of AdS3 with Brown-Henneaux boundary conditions. In particular superrota-
tions arise in the form (9.80) while supertranslations (9.81) measure how fast F̄(ϕ)
goes to −F(−ϕ) as ` goes to infinity. Similar considerations would reproduce the

centrally extended BMS3 group (9.62) as a contraction of the direct product of two
Virasoro groups.
Note that the condition (9.82) does not imply that there are less elements in the
BMS3 group than in the group Diff(S 1 ) × Diff(S 1 ). Indeed, both groups are infinite-
dimensional Lie groups consisting of two spaces of functions on the circle and have the
same cardinality in this sense.
9.4.2 From Witt to bms3

The limit from Diff(S 1 )×Diff(S 1 ) to BMS3 can be reformulated in terms of Lie algebras.
Again, our notation will be slightly different from that of the previous chapters so as to
distinguish AdS3 quantities from Minkowskian quantities. Thus we consider a vector
field X (x+ )∂+ + X̄ (x− )∂− on a two-dimensional cylinder and use ∂± = 21 (`∂u ± ∂ϕ ) to
rewrite it as
` 1
X (x+ ) + X̄ (x− ) ∂u + X (x+ ) − X (x− ) ∂ϕ .

(9.83)
2 2
In the flat limit ` → +∞ the angular component becomes
1 `→+∞ 1
X (x+ ) − X̄ (x− ) →

X (ϕ) − X̄ (−ϕ) ≡ X(ϕ) (9.84)
2 2
where X(ϕ) is some function on the circle, later to be interpreted as (the component
of) a superrotation generator. For the time component one finds
` `→+∞ `
X (x+ ) + X̄ (x− ) → X (ϕ) + X̄ (−ϕ) + uX 0 (ϕ) ≡ α(ϕ) + uX 0 (ϕ) (9.85)

2 2
where we have once more introduced a function α on the circle, later to be interpreted
as (a component of) a supertranslation generator. This time the requirement is
2
X (ϕ) + X̄ (−ϕ) = α(ϕ) (9.86)
`
with a finite, `-independent α, and is directly analogous to the condition (9.81). All
in all we find that, in the flat limit, the vector field (9.83) turns into
ξ(X,α) ≡ X(ϕ)∂ϕ + α(ϕ) + uX 0 (ϕ) ∂u

and thus coincides with the leading non-radial components of the asymptotic Killing
vector field (9.17). The Lie brackets of such vector fields satisfy the centreless bms3
algebra; the latter is thus a flat limit of the direct sum of two Witt algebras. Note
that from this perspective the fact that supertranslations have dimensions of length
follows from the fact (9.86) that α(ϕ) is proportional to `.
The limit from Witt to bms3 can also be formulated in terms of commutation
−
relations. Indeed, let `m = eimx ∂+ and `¯m = eimx ∂− denote the generators of two
+
commuting Witt algebras (6.24). Then the correspondence (9.84)-(9.86) instructs us

to define would-be superrotation and supertranslation generators
1
jm ≡ `m − `¯−m , pm ≡ (`m + `¯−m ) . (9.87)
`
The terminology here is consistent with the fact that, on the cylinder, `0 − `¯0 generates
rotations while `0 + `¯0 generates time translations. In the basis (9.87), the commutation
relations of the direct sum of two Witt algebras take the form
1
i[jm , jn ] = (m − n)jm+n , i[jm , pn ] = (m − n)pm+n , (m − n)jm+n .
i[pm , pn ] =
`2
(9.88)
In the limit ` → +∞ the last bracket vanishes and the algebra reduces to (9.10),
reproducing bms3 as expected. The same argument can be applied to the direct sum
of two Virasoro algebras and gives rise to the centrally extended bms
d 3 algebra (see eq.
(9.93) below).
This observation can be used to define “flat limits” of Lie algebras in general terms.
Consider indeed the Lie algebra g ⊕ g, whose generators we denote ta and t̄a , with
identical commutation relations (5.2) in both sectors:
[ta , tb ] = fab c tc , [t̄a , t̄b ] = fab c t̄c . (9.89)
Then consider the redefinitions

1
ja ≡ ta + t̄a , pa ≡ (ta − t̄a ) (9.90)
`
where ` is some length scale that we will eventually let go to infinity. In terms of j’s
and p’s the commutation relations (9.89) become
1 c
[ja , jb ] = fab c jc , [ja , pb ] = fab c pc , [pa , pb ] = fab jc (9.91)
`2
and the limit ` → +∞ reproduces the commutation relations (9.48) of exceptional
semi-direct sums (without central terms). Thus, the flat limit of any group G × G is
an exceptional semi-direct product G nAd gAb . The flat limit (9.87) giving rise to bms3
from two copies of the Witt algebra is a special case of that construction. Indeed, the
map
`m 7→ −`−m (9.92)
is a Lie algebra isomorphism when the `m ’s generate a Witt algebra (6.24), so the
redefinitions (9.87) precisely take the form (9.90) with the correspondence ta ↔ `m
and t̄a ↔ −`¯−m . As it turns out, all symmetry algebras found so far in the realm of
asymptotically flat field theories in three dimensions can be seen as flat limits of the
type just described when compared to their AdS3 counterparts.
The limiting procedure that turns the sum of two Witt algebras into bms3 is an ex-
ample of Inönü-Wigner contraction [115], similar to the relation between the Poincaré
group and the Galilei group. Conversely, the direct sum of two Witt algebras is a
deformation of bms3 . The same construction can be used to show that the Poincaré
algebra is a flat limit of the AdS3 isometry algebra, so(2, 2) ∼
= sl(2, R) ⊕ sl(2, R). We
will return to this in section 10.2.
9.4.3 Stress tensors and central charges

We now apply the flat limit to the coadjoint representation of two Virasoro groups,
generally with non-zero central charges. Let therefore T (x+ ) and T̄ (x− ) be CFT
stress tensors transforming under left and right conformal transformations as Virasoro
coadjoint vectors with central charges c and c̄, respectively. (In (8.38) we denoted
these stress tensors as p, p̄, but here we keep the letter p for supermomenta.) They are
paired with vector fields X (x+ )∂+ + X̄ (x− )∂− on the cylinder according to
Z 2π
1
dϕ T (x+ )X (x+ ) + T̄ (x− )X̄ (x− )

(T, T̄ ), (X , X̄ ) =
2π 0
which (up to notation) is just the AdS3 surface charge (8.42). As above we expand
the functions T and T̄ in powers of 1/` and we define
1
T (x+ ) + T̄ (x− ) , j(ϕ) + up0 (ϕ) ≡ lim T (x+ ) − T̄ (x− ) .

p(ϕ) ≡ lim
`→+∞ ` `→+∞
Using (9.85) and (9.86) one then verifies that, in the limit ` → +∞,
T (x+ )X (x+ ) + T̄ (x− )X̄ (x− ) = j(ϕ)X(ϕ) + p(ϕ)α(ϕ) + u(pX)0 (ϕ),
which coincides up to a total derivative with the integrand of the flat surface charge
(9.31). In other words the surface charges of flat space gravity are large ` limits of
those of AdS3 . In mathematical terms this is to say that the flat limit of the coadjoint
representation of the direct product of two Virasoro groups is the coadjoint represen-
tation of the (centrally extended) BMS3 group.
This phenomenon also allows us to relate the central charges of the Virasoro algebra
to those of bms
d 3 . (We could have done this in terms of abstract Lie algebra generators,
but for comparison with three-dimensional gravity we do it here in terms of coadjoint
vectors.) Let us consider two Virasoro algebras with central charges c and c̄ that
depend on ` as
c = A` + B + O(1/`), c̄ = A` + B̄ + O(1/`)
where A, B and B̄ are `-independent. Then the definitions
c + c̄
c1 ≡ lim (c − c̄), c2 ≡ lim (9.93)
`→+∞ `→+∞ `
allows us to write the flat limit of the algebra in the bms
d 3 form (9.71) in terms of
generators (jm , pm ) related to Virasoro generators (`m , `¯m ) by (9.87). This is the
centrally extended analogue of the flat limit described in (9.91). Note that for the
Brown-Henneaux central charges (8.40) the prescription (9.93) yields c1 = 0 and c2 =
3/G, which are indeed the standard values for asymptotically flat space-times.
Remark. The fact that flat space holography can be studied as a flat limit of the
AdS/CFT correspondence is an old idea [282, 283]; see also [284–286]. Here we have
described its group-theoretic formulation. It should be noted, however, that there is
no known limiting construction that yields BMS symmetry in four dimensions from
some corresponding asymptotic symmetry in AdS4 .
9.4.4 The Galilean conformal algebra

The bms3 algebra turns out to be isomorphic to the Galilean conformal algebra in
two dimensions. We now explain how the latter can be obtained as a non-relativistic
contraction of two Witt algebras and discuss the extent to which Galilean conformal
symmetry applies to asymptotically flat gravity in three dimensions. As in the earlier
sections of this chapter we work only at the classical level. The quantum version of
these considerations will be exposed in chapter 10.2.
The redefinitions (9.90) suggest a contraction of Witt algebras that differs from
the flat limit (9.87). Namely, instead of performing the involution (9.92) before taking
the limit ` → +∞, one can define
1
j̃m ≡ `¯m + `m , p̃m ≡ (`¯m − `m ) . (9.94)
`
In contrast to (9.87), this redefinition has nothing to do with the flat limit of AdS3 ,
but the limit ` → +∞ still gives rise to an algebra with commutation relations (9.10)
upon renaming jm → j̃m and pm → p̃m . The key difference is that now the generator
of time translations is j̃0 (since it coincides with `0 + `¯0 ) while p̃0 generates rotations
(since it is proportional to `0 − `¯0 ). More generally, with the redefinition (9.94), the
generators of would-be supertranslations do not commute while those of would-be su-
perrotations do commute. This is the opposite of the behaviour of superrotations and
supertranslations in three-dimensional Einstein gravity.
The redefinitions (9.94) can be interpreted as a non-relativistic contraction of the

direct sum of two Witt algebras in two dimensions, analogous to the usual Inönü-
Wigner contraction of the Poincaré algebra to the Galilei algebra. For this reason the
algebra spanned by j̃m ’s and p̃m ’s is known as the Galilean conformal algebra in two
dimensions [287, 288]. It is the non-relativistic limit of the conformal algebra in two
dimensions and, by a geometric coincidence, it is isomorphic to bms3 . The Galilean
conformal algebra has been extensively studied in its own right; see e.g. [289] for its
supersymmetric extension and [287, 290] for its highest-weight representations. It is a
fundamental tool in the non-relativistic limit of the AdS/CFT correspondence [291].
In what follows we denote it by gca2 .
At some point the isomorphism gca2 ∼ = bms3 led to the proposal that flat space
holography (in three space-time dimensions) is described by a Galilean conformal field
theory [288, 292]. In view of the geometric interpretation of superrotations and super-
translations described above, this sounds suspicious: the Galilean conformal algebra is
a version of the bms3 algebra “rotated by 90 degrees” where the roles of the Hamiltonian
and angular momentum are exchanged. In particular the flat limit of AdS3 /CFT2 , if it
exists, should not give rise to a Galilean conformal field theory since the gravitational
flat limit (9.87) of two Witt algebras gives rise to standard bms3 , in which p0 generates
time translations. Nevertheless, at the level of classical symmetries, there is essentially
no distinction between bms3 and gca2 ; the two are interchangeable. This coincidence
led to many publications concerned with flat space holography and attempting to
describe its dual theory as a Galilean conformal field theory; see e.g. [293–295] and
references therein. One of the goals of this thesis is to explain why the dual theory
of asymptotically flat gravity, if it exists at all, cannot be a Galilean conformal field

theory. The reason for this is rooted in the elementary observation that the corre-
spondence bms3 ↔ gca2 exchanges the Hamiltonian and the angular momentum, but
we will go much beyond that. In fact we shall see that the difference between bms3
and gca2 , while classically invisible, becomes apparent at the quantum level. This will
rely on the induced representations developed in the next chapter and will be studied
in much greater detail in section 10.2. As it turns out, the most striking illustration
of this distinction will arise in section 11.2 in the realm of quantum higher-spin theories.
This being said, we stress that discarding Galilean conformal field theories as pu-
tative duals for asymptotically flat gravity does not rule out all the conclusions of
the substantial literature on flat space holography approached from the Galilean side.
Rather, the point we wish to make is that those computations that did work in flat
space while relying on gca2 symmetry would have worked equally well in the language
of BMS3 symmetry. More precisely, any computation that holds for gca2 but does
not rely on its realization as a quantum symmetry algebra also holds for bms3 , and
therefore for asymptotically flat gravity.
Chapter 10
Quantum BMS3 symmetry
This chapter is devoted to irreducible unitary representations of the BMS3 group, i.e.
BMS3 particles, which we classify and interpret. As we shall see, the classification is
provided by supermomentum orbits that coincide with coadjoint orbits of the Virasoro
group. Upon identifying supermomentum with the Bondi mass aspect of asymptot-
ically flat metrics, we will be led to interpret BMS3 particles as relativistic particles
dressed with gravitational degrees of freedom.
The plan is as follows. In section 10.1 we classify BMS3 particles according to orbits
of supermomenta under superrotations. We also describe and interpret the resulting
Hilbert spaces of wavefunctions, which we relate to the quantization of (coadjoint)
orbits of asymptotically flat metrics under BMS3 . Section 10.2 is devoted to the de-
scription of BMS3 particles as representations of the (centrally extended) bms3 algebra
and their relation to highest-weight representations of the Virasoro algebra; we also
briefly touch upon Galilean representations. Finally, in section 10.3 we evaluate char-
acters of BMS3 particles. To lighten the notation, from now on the words “BMS3
group” or “bms3 algebra” implicitly refer to their centrally extended versions (except
if stated otherwise). We also abuse notation by writing Diff(S 1 ) to refer either to
Diff+ (S 1 ) or to Diff
g + (S 1 ), depending on the context.
Most of the results exposed in this chapter have been reported in the papers [46–
48, 51]. The relation between BMS3 particles and gravitational one-loop partition
functions [49, 50] will be described in the next chapter. Note that the considerations
that follow rely heavily on the material of chapter 4.
10.1 BMS3 particles

In high-energy physics a particle is usually defined as an irreducible unitary represen-
tation of the Poincaré group. If one takes BMS symmetry seriously, it is tempting to
apply the same terminology to representations of BMS. Accordingly, in this section
our goal is to answer the following question:
Replace the word “Poincaré” by “BMS” in the definition of a particle.
What new notion of particle does one then obtain?
What new quantum numbers describe its degrees of freedom?
293
Chapter 10. Quantum BMS3 symmetry 294
In principle this problem should be addressed in the realistic four-dimensional world.

However, as mentioned in the introduction of this thesis, BMS symmetry in four di-
mensions is very poorly understood at present, so we will content ourselves with the
more modest task of understanding irreducible unitary representations of BMS in three
dimensions — that is, BMS3 particles. Remarkably, we will discover that BMS3 par-
ticles are labelled by mass and spin, exactly as standard relativistic particles. As in
section 8.4, we will interpret their extra degrees of freedom as boundary gravitons, or
equivalently soft gravitons.
The plan of this relatively long section is the following. We first describe the su-
permomentum orbits and little groups that classify BMS3 particles. We shall see that
these orbits are in fact coadjoint orbits of the Virasoro group, which will allow us to
define massive, massless and tachyonic BMS3 particles. We also discuss the existence
of integration measures on supermomentum orbits, since such measures are required
to define scalar products of wavefunctions. We then describe the states represented
by such wavefunctions and interpret them as particles dressed with quantized gravita-
tional degrees of freedom, in accordance with the relation between asymptotically flat
metrics and the coadjoint representation of BMS3 . We also apply this interpretation
to the vacuum representation and to spinning BMS3 particles, and we conclude by
discussing the extension of our considerations to four space-time dimensions.

Our goal is to understand the quantum-mechanical implementation of BMS3 symme-
try, at least as far as irreducible representations are concerned. According to section
2.1 we should leave room for projective representations; to do this we consider exact
representations of the universal cover of the universal central extension of the con-
nected BMS3 group, that is, BMS [ 3 . The latter was defined in (9.62). Since BMS
[ 3 is
a semi-direct product, one expects all its irreducible unitary representations to be in-
duced à la Wigner. These representations are classified by the orbits and little groups
described in general terms in section 4.1. Here we perform that classification.
Supermomentum orbits
The key ingredient in the description of BMS3 particles is the dual of the space of
d 1 )∗ . Following the terminology of section 9.2, its elements
supertranslations, Vect(S Ab
are centrally extended supermomenta
p(ϕ)dϕ2 , c2

(10.1)
paired with centrally extended supertranslations (α, λ) according to (6.111) with the
replacements X → α and c → c2 . As mentioned below (9.69), p(ϕ) has dimensions
of energy; its three lowest Fourier modes form a Poincaré energy-momentum vector
(in particular the zero-mode is the energy of p). More generally p(ϕ) is an energy
density on the circle while the central charge c2 is an energy scale. Supermomentum
transforms as a Virasoro coadjoint vector (9.70) under superrotations, so the allowed
supermomenta of a BMS3 particle span a coadjoint orbit of the Virasoro group at
central charge c2 . This is the first key conclusion of this section:
Theorem. The orbit Op of a supermomentum (p, c2 ) under superrotations is a coad-

joint orbit of the Virasoro group at central charge c2 .
When interpreting p(ϕ) as the Bondi mass aspect of an asymptotically flat metric
(9.25), the orbit Op is a subset of the orbit (9.77) of the metric under BMS3 transfor-
mations. In that context the central charge c2 coincides with the Planck mass (9.29).
Accordingly, from now on we restrict our attention to centrally extended supermo-
menta whose central charge c2 is strictly positive.
Massive and massless BMS3 particles

The statement that supermomentum orbits are Virasoro coadjoint orbits is analogous
to the fact that coadjoint orbits of SL(2, R) classify the momenta of relativistic par-
ticles in three dimensions. In particular the map of Poincaré momenta in fig. 4.3b is
embedded in the larger picture of fig. 7.3, which is now interpreted as a map of BMS3
supermomenta. Thus, supermomentum orbits that contain a constant representative
(the vertical line in fig. 7.3) are the supermomenta of BMS3 particles that admit a rest
frame.
Definition. A massive BMS3 particle is a BMS3 particle whose supermomenta span

a Virasoro coadjoint orbit that admits a generic constant representative p0 .
In this definition the word “generic” refers to the fact that p0 should not take
one of the discrete exceptional values −n2 c2 /24. Indeed the orbits containing such
exceptional constants are better thought of as BMS3 generalizations of the trivial rep-
resentation of Poincaré; we will return to this interpretation below.
By contrast, supermomentum orbits that do not admit a constant representative

describe BMS3 particles that have no rest frame. For instance, the discrete dots that
do not belong to the vertical line in fig. 7.3 are BMS3 generalizations of massless
Poincaré particles, while the horizontal lines of fig. 7.3 generalize tachyons.
Definition. A massless BMS3 particle is a BMS3 particle whose supermomenta span

a Virasoro coadjoint orbit with non-degenerate parabolic monodromy and non-zero
winding number. A BMS3 tachyon is a BMS3 particle whose supermomenta span a
Virasoro coadjoint orbit with hyperbolic monodromy and non-zero winding number.
In these definitions the terms “monodromy” and “winding number” refer to the
Virasoro invariants defined in section 7.1. They are the BMS3 generalization of the
mass squared in the Poincaré group.
Little groups
The little groups of BMS3 particles coincide with the stabilizers of the corresponding
Virasoro coadjoint orbits. Here, for comparison with the Poincaré little groups of
section 4.3, we list the little groups obtained by using the central extension of the
multiply connected BMS3 group (9.60). The list of orbits is that of section 7.2 and
their little groups are summarized in table 7.1:
• For a massive BMS3 particle, the stabilizer is the group U(1) of spatial rotations.
• For a vacuum-like BMS3 particle whose supermomentum at rest takes the value
−n2 c2 /24, the little group is an n-fold cover of the Lorentz group in three di-
mensions, PSL(n) (2, R) (with n ≥ 1).
• For a massless particle with winding number n ≥ 1, the little group is R × Zn .
• For a BMS3 tachyon with winding number n ≥ 1, the little group is R × Zn .
This list should be compared with table 4.1. The representations of these little groups
will lead to a notion of BMS3 spin. Note that when dealing with the universal cover
(9.61) of BMS3 , all compact directions of the above little groups get decompactified
so that U(1) is replaced by R, PSL(n) (2, R) is replaced by its universal cover, and Zn
is replaced by the group T2π/n ∼
= Z of translations of R by integer multiples of 2π/n.
BMS3 particles with positive energy

It is natural to declare that physically admissible BMS3 particles have supermomentum
orbits such that the energy functional (7.79) is bounded from below under superrota-
tions. Finding these particles is the BMS3 analogue of the question (7.82) encountered
in the Virasoro context. The solution is provided by the earlier results (7.103)-(7.104):
Theorem. A BMS3 particle has energy bounded from below if and only if its super-
momenta span one of the Virasoro orbits coloured in red in fig. 7.7.
Recall that Poincaré particles with positive energy fall in exactly three classes,
two of which contain only one momentum orbit: massive particles, massless particles,
and the trivial orbit. The theorem tells us that essentially the same conclusion holds
for BMS3 particles, since all supermomentum orbits with energy bounded from below
belong to one of the three following classes:
• the unique vacuum orbit containing the supermomentum p0 = −c2 /24,
• one of the massive orbits located above the vacuum and containing a constant
supermomentum p0 > −c2 /24,
• the unique massless orbit with energy bounded from below.
From now on, when referring to BMS3 particles we always implicitly refer only to
particles with energy bounded from below (except if explicitly stated otherwise). Note
that, in contrast with Virasoro representations, BMS3 particles with unbounded energy
may provide unitary representations of BMS3 . Furthermore the energy spectrum of
any BMS3 particle is continuous.
10.1.2 Mass, supermomentum, central charge

In the list of physical BMS3 particles, the only family with infinitely many members
is the class of massive particles. Let us therefore describe these particles in some more
detail and interpret the labels (p, c2 ) that classify them.
Defining mass
The starting point is the observation that the vacuum supermomentum is pvac =
−c2 /24, while the supermomentum at rest of any massive BMS3 particle is located
above that vacuum value.
Definition. Consider a massive BMS3 particle with supermomentum at rest p0 >

−c2 /24. Then the mass of the particle is
M ≡ p0 + c2 /24. (10.2)
Massive BMS3 particles with energy bounded from below have positive mass.
The definition of mass in eq. (10.2) can be rewritten in a manifestly superrotation-

invariant way, without invoking any rest frame. Indeed, recall from (7.22) that the
value of p0 determines the trace of the monodromy matrix M. This relation can be
inverted and combined with the definition (10.2), which yields

c2 1 2
M= 1+ arccosh Tr(M/2) (10.3)
24 π
where we assume for definiteness that p0 ≥ 0, which is to say that M is hyperbolic and
M ≥ c2 /24. The same relation holds for elliptic M, hence M < c2 /24, upon replacing
arccosh by i arccos, with the convention arccos(1) = 0 and arccos(−1) = π.
Formula (10.3) is a superrotation-invariant definition of the mass of a BMS3 par-

ticle, since the trace of the monodromy matrix associated with Hill’s equation is
Virasoro-invariant. It is a BMS3 analogue of the relation
M 2 = E 2 − p2 = −pµ pµ
that determines the mass of a Poincaré particle from its energy-momentum pµ . As
a bonus, (10.3) allows us to distinguish massive particles with elliptic and hyperbolic
monodromy. This distinction is consistent with three-dimensional gravity, where met-
rics with a Bondi mass aspect p0 < 0 are conical deficits — i.e. classical particles —
while metrics with p0 > 0 are flat space cosmologies — Minkowskian analogues of
BTZ black holes. Furthermore, eq. (10.3) confirms that massless BMS3 particles (with
positive energy) are actually massless. Indeed, the corresponding monodromy matrix
is (7.78) with winding number n = 1. This implies that Tr(M) = −2 for physical
massless BMS3 particles, which can be plugged into (10.3) and yields M = 0 upon
using arccosh(−1) = i arccos(−1) = iπ. It would have been impossible to obtain this
result with the weaker definition of mass of eq. (10.2), since massless particles have no
rest frame.
Intepreting supermomentum
In order to develop our intuition about the supermomentum vector p(ϕ), it is use-
ful to rewrite standard Poincaré momenta in terms of functions on the circle. The
supermomentum of a Poincaré particle with mass M typically takes the form
q c2
p(ϕ) = M 2 + p2x + p2y + px cos ϕ + py sin ϕ − (10.4)
24
and represents a particle moving in space with a spatial momentum (px , py ). Note the
extra factor −c2 /24 at the end, which ensures that M actually coincides with the mass
of the particle.
One can then use (9.70) to act with superrotations on (10.4) and obtain various
boosted momenta. In particular,
p 2 Lorentz transformations take the form (6.88) and
act on the components ( M + px + p2y , px , py ) according to the vector representation.
2
This point can be verified by taking a supermomentum at rest, p0 = M − c2 /24, and

acting on it with a boost (7.100) in the direction ϕ = 0,
cosh(γ/2)eiϕ + sinh(γ/2)
eif (ϕ) = (10.5)
sinh(γ/2)eiϕ + cosh(γ/2)
where γ is the rapidity (in terms of standard velocity, γ = arctanh(v)). Since this
superrotation is of the projective form (6.88), its Schwarzian derivative satisfies (6.94)
and the corresponding transformation (9.70) of the supermomentum p can be rewritten
as
c2 1 h c2 i
f · p (f (ϕ)) + = 0 p(ϕ) + . (10.6)
24 (f (ϕ))2 24
This says that the combination p + c2 /24 transforms under Lorentz transformations
as a centreless coadjoint vector of Diff(S 1 ). In particular, the supermomentum of a
massive particle at rest transforms according to (10.6) with p(ϕ) = M − c2 /24. As a
result, since f 0 (ϕ) is given by (7.101), the energy E[γ] of the boosted particle is
M 2π dϕ
Z
E[γ] = = M cosh γ
2π 0 f 0 (ϕ)
while the boosted spatial momentum along the x direction is
M 2π dϕ
Z
px [γ] = cos(f (ϕ)) = M sinh γ.
2π 0 f 0 (ϕ)
The boosted spatial momentum along the y direction vanishes, as it should. This
confirms that Lorentz transformations act on a supermomentum at rest exactly as in
standard special relativity (albeit in three space-time dimensions).
From these considerations we can now draw a general conclusion on the supermo-
mentum of a (massive) BMS3 particle. Typically, the function p(ϕ) will have some
non-trivial profile on the circle; for instance the momentum of a particle moving fast
in the x direction is represented by a function p(ϕ) which is larger than −c2 /24, has a
bump around the point ϕ = 0, and almost vanishes in the neighbourhood of the oppo-
site point ϕ = π. If the particle is obtained by a pure Poincaré boost from a particle
at rest, the only non-vanishing components of its supermomentum are its three lowest
Fourier modes, p0 , p1 , p−1 (as in eq. (10.4)). Upon switching on superrotations, the
supermomentum of the particle acquires extra Fourier modes (p2 , p3 , etc.) that dress
the original Poincaré momentum with additional fluctuations. In the upcoming pages
we will interpret these extra degrees of freedom as being of gravitational origin.
(a) (b)
Figure 10.1: Two possible supermomenta of a massive BMS3 particle. In (a) the
supermomentum is that of a boosted Poincaré particle, given by eq. (10.4); it is a
function on the circle located above the line −c2 /24 and its only non-zero Fourier
modes are the three lowest ones. In (b) the function is dressed with extra non-vanishing
Fourier modes, which results in more wiggles. These extra Fourier modes account for
gravitational degrees of freedom that do not appear in the pure Poincaré case.
Interpreting the central charge

The central charge c2 is an energy scale; this is manifest in formula (10.3), where c2
converts the dimensionless trace of a monodromy matrix into a mass M . The actual
value of c2 is arbitrary in principle, but in Einstein gravity it is proportional to the
Planck mass: c2 = 3/G. In particular, note that c2 is not a Virasoro central charge,
even though it appears in the transformation law (9.70) of supermomentum as if p was
a CFT stress tensor.
It is worth stressing the crucial importance of c2 for the conclusions of the pre-
vious pages. For one thing, the whole classification of supermomentum orbits and
the ensuing definition of massive/massless BMS3 particles only makes sense because
c2 is non-zero. If c2 happened to vanish, none of these results would hold since the
corresponding supermomentum orbits would be Virasoro coadjoint orbits at vanishing
central charge, and we saw in section 7.1 that these orbits are radically different from
(and arguably much uglier than) their centrally extended peers. This is not to say that
c2 must be non-zero in order for the supermomentum (10.1) to yield a representation
of the BMS3 group; in principle, representations associated with orbits having c2 = 0
are just as acceptable as representations in which c2 6= 0. However the application to
gravity, and the ensuing interpretation of representations as particles, relies crucially
on the fact that c2 = 3/G does not vanish. Note that the change of monodromy
occurring at M = c2 /24 suggests that something radical happens with BMS3 particles
whose mass is higher than that bound. This bifurcation reflects the fact that the met-
ric of the gravitational field surrounding the particle changes from that of a conical
deficit (when M < 1/8G) to that of a flat cosmology (when M > 1/8G).
10.1.3 Measures on superrotation orbits

Suppose we actually want to describe the space of states of a BMS3 particle with
supermomentum orbit Op . If the particle is scalar, then its Hilbert space consists of
complex-valued wavefunctions (4.20) in supermomentum space whose scalar product
(3.7) involves an integral over Op with some measure µ. The latter needs to be quasi-
invariant under superrotations,1 which motivates the following question:
Let Op be a Virasoro coadjoint orbit at non-zero central charge;

(10.7)
is there a quasi-invariant Borel measure on it?
If the answer is affirmative, then the measure is a functional one since Op consists of
functions on the circle.
A conjecture
Our viewpoint regarding the problem (10.7) will be pragmatical: path integral mea-
sures are used on a daily basis in quantum mechanics, and their efficiency in correctly
predicting the values of physical observables is firmly established. Thus, if one is
willing to define Hilbert spaces of square-integrable functions thanks to functional
measures, their application to BMS3 particles is as acceptable as in quantum physics.
In particular one may hope that Virasoro coadjoint orbits do admit quasi-invariant
measures:
Conjecture. Let Op be a Virasoro coadjoint orbit with non-zero central charge (and
energy bounded from below). Then there exists a Borel measure dµ(q) on Op which
is quasi-invariant under the action of the Virasoro group (where q ∈ Op ).
In the remainder of this thesis we will rely on this conjecture in order to define the
Hilbert space of a BMS3 particle (at least one with bounded energy). The conjecture
does not say how the measure dµ(q) is actually defined, but this is not a problem
since the results of section 3.2 imply that induced representations based on different
quasi-invariant measures are unitarily equivalent. Thus, assuming that the conjecture
is true, we do not really need to know anything specific about the measure.2 In par-
ticular the character computation of section 10.3, although relying on an unknown
measure, will produce an unambiguous result.
Aside from these basic observations, we will have very few concrete things to say
about the measure. Nevertheless the lines that follow are devoted to a brief review of
the literature on Virasoro measure theory, with the intent of further motivating the
validity of the conjecture. The reader who is not interested in mathematical subtleties
is free to go directly to section 10.1.4.
1
We recall that the definition of quasi-invariant measures was given in section 3.2.
2
Note that the existence of a quasi-invariant measure dµ(q) on Op implies the existence of infinitely
many other ones, since one can always multiply the measure by a strictly positive smooth function
ρ(q) and obtain a new measure ρ(q)dµ(q).
Remark. Recall that all irreducible unitary representations of regular semi-direct

products are induced representations, where regularity refers to the property defined
at the end of section 4.1. Accordingly, in order to claim that all irreducible unitary
representations of BMS3 are BMS3 particles whose supermomenta span Virasoro orbits,
we would have to prove that the BMS3 group is a regular semi-direct product, which
in turn relies on the existence of a measure on the space of supermomenta. We will
not address this question here and assume instead that the standard results on finite-
dimensional semi-direct products carry over to BMS3 .
Measures on Virasoro orbits

The issue of rigorously defining path integral measures was first addressed by Wiener
about a century ago, in the context of stochastic processes. We refer e.g. to the bi-
ographical memoir in [296] for more references and a more accurate account of the
development of the subject. The problem of defining a Wiener-like quasi-invariant
measure on Virasoro coadjoint orbits is more recent, but well known. In the physics
literature, as in our conjecture above, the question of the measure is mostly treated in
a heuristic way motivated by quantum mechanics; see e.g. [297] for such an approach.
By contrast, there is a fair amount of mathematical literature that aims at solving the
problem in a rigorous way, and to our knowledge no definite, widely accepted solution
is known at present. As announced above, we do not claim to provide an answer here;
rather, we shall content ourselves with a brief literature review.
The main motivation for defining Virasoro measures comes from representations
of the Virasoro algebra and conformal field theory; the hope is that such measures
could provide a rigorous prescription for the geometric quantization of Virasoro or-
bits. This approach to the problem is adopted for instance in [230, 231, 298, 299].
In [300] the authors tackle the issue with a similar motivation, though with different
methods; in particular it is suggested there that a measure might be provided by an
infinite-dimensional version of the Liouville measure (5.14) obtained by taking the
Kirillov-Kostant symplectic form (8.62) to an infinite power.
A somewhat different approach consists in building measures on Virasoro orbits re-

gardless of their relation to conformal field theory and highest-weight representations.
This approach does not simplify the problem of geometric quantization of Virasoro
orbits, but it does have the virtue of producing the desired measure. One should keep
in mind that the measures needed for BMS3 particles generally have nothing to do
with those needed for geometric quantization of the Virasoro group, so the fact that
a measure is unrelated to Virasoro representations is not a problem for our purposes.
As it turns out, certain results due to Shavgulidze [301–304] precisely point in that
direction (see also [305,306]). Indeed it was shown in [301] that the group of diffeomor-
phisms of any compact manifold can be endowed with a quasi-invariant Borel measure.
This measure then plays a role analogous to the Haar measure (recall section 3.2) and
ensures that quotients of the group, such as Diff(S 1 )/S 1 , can also be endowed with
a quasi-invariant measure. As a corollary, Virasoro orbits (all of which are quotients
of Diff(S 1 )) should generically admit quasi-invariant Borel measures. This is precisely
what is needed for BMS3 particles, and it is in fact our main justification for the above
conjecture.
10.1.4 States of BMS3 particles

Under the assumption that there exist quasi-invariant measures on Virasoro coad-
joint orbits, we have all the ingredients required to build explicit unitary, irreducible,
projective representations of the BMS3 group. Here we describe and interpret the
wavefunctions that represent the quantum states of a scalar massive BMS3 particle
with mass M > 0. Spinning particles and the vacuum representation will be described
later.
The states of a BMS3 particle

The supermomenta of a massive particle span an orbit Op = Diff(S 1 )/S 1 (with implicit
central charge c2 > 0), which we assume to admit a quasi-invariant measure µ. For
convenience the orbit representative p is taken to be the supermomentum p(ϕ) = p0 =
M − c2 /24 at rest. As in (4.20) the particle’s Hilbert space H consists of complex-
valued wavefunctions
Ψ : Op → C : q 7→ Ψ(q) (10.8)
which are square-integrable with respect to µ in the usual sense that the integral
Z
dµ(q) |Ψ(q)|2 (10.9)
Op
is finite. As usual, Ψ should be thought of as a wavefunction in (super)momentum

space representing a wavepacket that propagates with fuzzy velocity. In contrast to
the finite-dimensional case, however, the map (10.8) is really a functional since it is
defined on a space of functions:
Ψ : q(ϕ) 7→ Ψ[q(ϕ)].
Similarly the integral (10.9) is actually a functional integral. Despite these compli-
cations we will keep using the simpler notation Ψ(q) for the wavefunctions of BMS3
particles. The scalar product on H = L2 (Op , µ, C) is then defined by (3.7) with
(Φ(q)|Ψ(q)) = Φ∗ (q)Ψ(q).
The space of states of a BMS3 particle carries an action of BMS3 by unitary trans-
formations. Since we are assuming that the particle is scalar, the action of BMS3 on
H is given by formula (4.23):
q
T [(f, α)] · Ψ (q) = ρf −1 (q) eihq,αi Ψ(f −1 · q) .

(10.10)
Here (f, α) is an element of the BMS3 group (9.58), with f (ϕ) a superrotation and
α(ϕ) a supertranslation. The point q ∈ Op is a supermomentum vector and its pairing
hq, αi with α is given by (6.34). The function ρf (q) is the (unknown) Radon-Nikodym
derivative (3.19) of the measure µ; if by chance the measure happens to be invariant,
then one can set ρf (q) = 1. Finally, the action f · q appearing in the argument of the
wavefunction on the right-hand side is the BMS3 generalization of the action of boosts
on momenta; it is given by formula (9.70), which is the coadjoint representation of the
Virasoro group at central charge c2 .
Remark. Wavefunctionals are common in quantum field theory. Indeed, the quan-
tum state of a typical field theory is a wavefunctional Ψ[φ(x)], where φ(x) is a spatial
field configuration. The truly striking aspect of (10.8) is not quite the fact that it
belongs to a space of wavefunctionals, but rather that it provides an irreducible rep-
resentation of the symmetry group.
BMS3 particles as projective representations

The central charge c2 6= 0 turns expression (10.10) into a projective representation of
the centreless BMS3 group. Indeed, as is manifest in (9.71), c2 is reponsible for an
extra term in the commutation relations of superrotations with supertranslations. At
the group-theoretic level this difference is due to the extra terms of the group operation
(9.41), as opposed to the centreless group operation (9.59). In the latter case we have
(f, 0) · (e, α) = (f, Adf α) = (e, Adf α) · (f, 0) (10.11)
where e is the identity in Diff(S 1 ) and Ad is the action of diffeomorphisms on vector

fields on the circle. By contrast, in the centrally extended case (9.41) there are two
extra slots for central terms and the analogue of (10.11) becomes
1
(f, 0; 0, 0) · (e, 0; α, 0) = f, 0; Adf α, − hS[f ], αi ,
12
(e, 0; α, 0) · (f, 0; 0, 0) = f, 0; Adf α, 0
where the two lines differ by a term proportional to hS[f ], αi, with S the Schwarzian
derivative (6.76). This phenomenon is analogous to the statement that boosts and
translations do not commute in the Bargmann group (4.103). In practice it means
that the wavefunction obtained by acting first with (e, α), then by (f, 0), differs from
the one obtained by acting first with (f, 0), then with (e, Adf α), by a constant complex
phase that can be evaluated using (9.55):
h c i
2
T [(f, 0)] · T [(e, α)] = exp −i hS[f ], αi T [(e, Adf α)] · T [(f, 0)] . (10.12)
12
This is indeed the statement (2.7) that the representation T is projective, when seen
as a representation of the centreless BMS3 group (9.58). It is the BMS3 analogue of
the Galilean result (4.121).
There is an important subtlety in (10.12): formula (10.10) is a projective repre-

sentation only if one insists on using the centreless group operation (9.59). If instead
one uses the centrally extended group (9.62), projectivity is absorbed by the definition
of the group operation (9.41). This is a restatement of our earlier observation in sec-
tion 2.1 that projective representations can be seen in two equivalent ways: either as
genuine projective representations of a centreless group, or as exact (non-projective)
representations of a centrally extended group. This same argument is the reason why
massive BMS3 particles can have arbitrary real values of spin, as we shall discuss
below.
Plane waves
As in section 3.3 we can describe the representation (10.10) in terms of a basis of
one-particle states with definite (super)momentum on the orbit Op . Let therefore
δ denote the Dirac distribution associated with the measure µ and defined by the
requirement (3.39). In the present case µ is a functional measure, so δ is a functional
delta distribution. For k ∈ Op , we define the plane wave state with supermomentum
k as
Ψk (q) ≡ δ(k, q) (10.13)
which is now a functional analogue of eq. (3.43). It is a typical asymptotic state in a
scattering experiment. The scalar products of plane waves are given by (3.44):
hΨk |Ψk0 i = δ(k, k 0 ). (10.14)
Strictly speaking, plane waves are not square-integrable, hence do not belong to the
space of states of a BMS3 particle. They should therefore be understood in the weaker
sense that any wavepacket (3.45) can be written as an infinite sum of plane waves.
With this word of caution, one may say that plane waves form a “basis” of the space
of states of a BMS3 particle. Their transformation law under BMS3 transformations
is given by eq. (4.32),
q
T [(f, α)] · Ψk = ρf (k) eihf ·k,αi Ψf ·k , (10.15)
except that we have removed the spin representation R since the particle considered
here has vanishing spin. This formula reflects the fact that a wavefunction with mo-
mentum k boosted by a superrotation f becomes a wavefunction with momentum f ·k.
In short, all results of chapters 3 and 4 remain valid, up to the fact that manifolds
become spaces of functions while functions become functionals.
10.1.5 Dressed particles and quantization

Asymptotically flat gravity enjoys BMS3 symmetry, so its quantization is expected
to produce unitary representations of the BMS3 group. More precisely, since the
phase space of metrics (9.25) is a hyperplane c1 = 0, c2 = 3/G in the space of the
[ 3 , the geometric quantization of the orbit (9.77) of a
coadjoint representation of BMS
metric (j, p) under asymptotic symmetry transformations is expected to produce an
irreducible unitary representation of BMS3 with supermomentum orbit Op and spin
(5.128) determined by the restriction of j to the Lie algebra of the little group of p.
Thus,
A BMS3 particle is the quantization
of the orbit of a metric under BMS3 transformations.
In the following pages we use this observation to compare BMS3 particles with standard
relativistic particles. For simplicity we focus on a massive scalar particle whose orbit
representative is taken to be the supermomentum p = M − c2 /24 at rest.
Leaking wavefunctions
The transformation law (10.10) is an infinite-dimensional generalization of a scalar
representation (with mass M ) of the Poincaré group in three dimensions. Indeed, by
restricting one’s attention to the Poincaré subgroup of BMS3 , one obtains a (highly
reducible) unitary representation of Poincaré. The latter contains the standard scalar
irreducible representation with mass M , but it also contains an uncountable infinity
of other representations with higher mass. These extra representations arise because
the action of Lorentz transformations on the supermomentum orbit Op ∼ = Diff(S 1 )/S 1
is not transitive; in fact, the set of Lorentz-inequivalent supermomenta is an infinite-
dimensional manifold
PSL(2, R)\Diff(S 1 )/S 1 , (10.16)
which is a double quotient of Diff(S 1 ). The quotient on the left is taken with respect
to the Lorentz group PSL(2, R).
This observation can be rephrased in a more intuitive way: it says that the super-
momentum orbit Op = Diff(S 1 )/S 1 contains infinitely many finite-dimensional sub-
manifolds SL(2, R)/S 1 obtained by acting on the orbit with Lorentz transformations.
Each submanifold is a standard momentum orbit (4.97) for massive Poincaré particles
in three dimensions. Any two of those submanifolds are mutually Lorentz-inequivalent;
the set of such inequivalent sub-orbits is the space (10.16).
Figure 10.2: A supermomentum orbit Op ∼ = Diff(S 1 )/S 1 with embedded Lorentz sub-
orbits represented as curvy lines. Lorentz transformations move points along these sub-
orbits. Transitions from one sub-orbit to another are only possible with superrotations
that do not belong to the Lorentz subgroup of Diff(S 1 ). One can define an equivalence
relation on the supermomentum orbit by declaring that two points are equivalent if
they belong to the same Lorentz sub-orbit. The quotient of Op by that equivalence
relation is the double quotient (10.16).
Now, a wavefunction Ψ(q) of a BMS3 particle is never suppported on just one

Lorentz sub-orbit of Op : if it was, some components of its supermomentum would
be sharply defined and the uncertainty principle would be violated. (By the way, we
stress again that the plane waves (10.13) do not actually belong to the Hilbert space.)
Rather, the wavefunction spreads over many Lorentz-inequivalent momentum orbits.
In other words, the wavefunction Ψ(q) “leaks” into the directions of supermomentum
obtained by acting with superrotations that are not Lorentz transformations:
Figure 10.3: The supermomentum orbit of fig. 10.2, now with a wavefunction on top.
The wavefunction is roughly supported on a Lorentz sub-orbit, but not quite: it leaks
into directions that cannot be achieved with Lorentz transformations.
Remark. The double quotient (10.16) is an application of the induction-reduction

theorem for induced representations. The latter roughly goes as follows: Let G be
a group with two (generally different) subgroups H, H 0 and let S be an irreducible
representation of H. Then the restriction to H 0 of the induced representation IndGH (S)
0
is a direct integral of irreducible representations of H labelled by the points of the
double quotient H 0 \G/H. In (10.16) we have applied this theorem to G = BMS3 ,
H = U(1) n Vect(S 1 ) and H 0 = PSL(2, R) n sl(2, R). The latter is the Poincaré group
in three dimensions while the U(1) of H is the little group for massive particles.
Gravitational dressing
We have just explained that any supermomentum orbit contains infinitely many Poin-
caré-inequivalent sub-orbits. It is natural to wonder how the extra directions of the
orbit — those that do not lie along Lorentz generators — are to be interpreted. In
other words: how should one think of the fact that the wavefunction of a BMS3 par-
ticle leaks into directions that are forbidden by Poincaré transformations?
A natural guess is suggested by the very origin of the BMS3 group: as we showed
in section 9.1, it is the asymptotic symmetry group of Minkowskian space-times (in
three dimensions). Asymptotic symmetries may be thought of as generalizations of
isometries that incorportate gravitational fluctuations. Now, the isometry group of
Minkowski space-time is the Poincaré group, and its unitary representations are par-
ticles in the usual sense. Accordingly,
A BMS particle is a Poincaré particle
(10.17)
dressed with gravitational degrees of freedom.
Let us be more precise about what we mean by “gravitational degrees of freedom”.

Classically, those are the classes of points on the orbit Op that cannot be obtained
from p by a Lorentz transformation; the set of these classes coincides with the double
coset space (10.16). Quantum-mechanically, we would like these gravitational degrees
of freedom to correspond to the set of quantum states that cannot be obtained from
the state of a particle at rest by Lorentz transformations. This can be rephrased in
precise terms: on the Hilbert space H of a BMS3 particle we define an equivalence
relation ∼ such that Ψ ∼ Ψ0 if there exists a Poincaré transformation (f, α) for which
Ψ0 = T [(f, α)]Ψ. Then the space of Poincaré-inequivalent states is the quotient
H/∼ . (10.18)
The latter can also be seen as the set of quantum states obtained by acting on a
state at rest with supertranslations and superrotations that do not belong to the
Poincaré subgroup. In this sense it is the three-dimensional analogue of the space of soft
gravitons in four dimensions, as follows from the recently discovered relation [19, 247]
between asymptotic symmetries and soft theorems in gauge theories. Thus,
A BMS particle is a particle dressed with soft gravitons.
Note that the terminology of “soft gravitons” is a bit dangerous here, since three-
dimensional Einstein gravity has no local degrees of freedom, hence no genuine (bulk)
gravitons. We already pointed out this subtlety in the introduction of the thesis, and
our point of view remains the same: owing to the relation between soft gravitons and
asymptotic symmetries, any theory with non-trivial asymptotic symmetries can be
interpreted as a theory containing soft degrees of freedom, regardless of the presence
of bulk degrees of freedom. In this sense three-dimensional gravity is a toy model for
soft gravitons.
10.1.6 The BMS3 vacuum

Having analysed massive scalar BMS3 particles, we now turn to some of their cousins.
Here we describe the vacuum BMS3 representation, while spinning particles are stud-
ied in section 10.1.7.
The vacuum representation of BMS3 is the scalar induced representation based on

the vacuum supermomentum orbit — the one containing the constant pvac = −c2 /24.
The corresponding little group is the Lorentz group PSL(2, R), so the orbit is
Ovac ∼
= Diff(S 1 )/PSL(2, R). (10.19)
As before we assume that it admits a measure which is quasi-invariant under superro-
tations. The Hilbert space of the representation is then spanned by square-integrable
wavefunctions (10.8) on Ovac transforming under BMS3 according to (10.10).
The interesting aspect of the vacuum representation is its interpretation. Indeed,

while massive particles exist in both the Poincaré group and the BMS3 group, the
vacuum representation is non-trivial only in the BMS3 context. This is analogous to
the observation of section 8.4 that the vacuum representation of the Virasoro group
is non-trivial. In particular, as in Virasoro, there is no fully BMS3 -invariant defini-
tion of the vacuum at non-zero central charge; the maximal possibility is Poincaré
invariance, which is indeed achieved by pvac = −c2 /24 (with j = 0). This reduced
symmetry is responsible for the non-triviality of the representation and for the fact
that wavefunctions of the vacuum representation “leak” into directions that cannot
be reached by Poincaré transformations. As in the massive case (10.16), the set of
Lorentz-inequivalent supermomenta on the vacuum orbit is a double coset space
PSL(2, R)\Diff(S 1 )/PSL(2, R) (10.20)
which parameterizes the decomposition of the vacuum BMS3 representation as a direct

integral of Poincaré sub-representations.
The presence of a non-trivial vacuum representation can be interpreted in gravita-

tional terms. Indeed, according to our heuristic proposal (10.17), the vacuum BMS3
representation consists only of gravitational degrees of freedom (since the correspond-
ing Poincaré particle is trivial). Classically these degrees of freedom span the BMS3
orbit of the Minkowski metric. In the terminology of section 8.3 those would be
“boundary gravitons”, or equivalently soft gravitons, around Minkowski space.
Dressed particles revisited

Let us return to the interpretation of BMS3 particles as dressed particles, now using
the vacuum representation as an extra input. We start with the following observation:
consider the space L2 (M × N , µ) of square-integrable functions on the product space
M × N ; suppose the measure µ factorizes as a product µ = µM × µN , where µM
is a measure on M and µN is a measure on N . Then one has a tensor product
decomposition
L2 (M × N , µM × µN ) ∼
= L2 (M, µM ) ⊗ L2 (N , µN ). (10.21)
Let us use this result to compare massive particles, with supermomentum orbits
Diff(S 1 )/S 1 , to the vacuum whose orbit is Diff(S 1 )/PSL(2, R). Since both of these
infinite-dimensional manifolds are homotopic to a point, we can relate them as
Diff(S 1 )/S 1 ∼
= PSL(2, R)/S 1 × Diff(S 1 )/PSL(2, R) .

(10.22)
This is to say that the massive supermomentum orbit is a direct product M × N . The
first factor of the product is the Poincaré momentum orbit (4.97) of a massive particle
in three dimensions, which suggests that a BMS3 particle is equivalent to a relativistic
particle “times” the vacuum BMS3 representation. This can be made precise using
(10.21): if the measure µ used to define the scalar product of wavefunctions for a mas-
sive BMS3 particle factorizes into a product on Diff(S 1 )/PSL(2, R) and PSL(2, R)/S 1 ,
then the Hilbert space of a massive BMS3 particle factorizes into
HBMS ∼
= HPoinc ⊗ Hvac (10.23)
where HPoinc is the space of states of a massive Poincaré particle and Hvac is that of
the BMS3 vacuum representation.
One can reformulate the statement (10.23) in the basis of plane wave states (10.13).
Indeed, on a massive supermomentum orbit Op = Diff(S 1 )/S 1 , the diffeomorphism
(10.22) allows us to write any supermomentum q as a pair q = (qPoinc , qvac ) where
qPoinc is a momentum vector with three components belonging to the Poincaré sub-
orbit OPoinc = PSL(2, R)/S 1 , while qvac is a supermomentum that belongs to the
vacuum BMS3 orbit Ovac = Diff(S 1 )/PSL(2, R). Then, under the assumption that the
measure µ on Op disintegrates into a product of measures on PSL(2, R)/S 1 and Ovac ,
any plane wave (10.13) can be written as a tensor product
Ψk = ΨkPoinc ⊗ Ψkvac
where the left and right factors of the product are plane waves on the orbits OPoinc and
Ovac , respectively. A generic state of a BMS3 particle is an infinite linear combination
of such factorized plane waves. Note that none of our upcoming conclusions rely on
this phenomenon, so in the sequel we will not necessarily assume that the measure µ
disintegrates into a product.
Remark. The decomposition (10.22) is similar to that of the Poincaré D-momentum

pµ = (E, p). In the latter case the truly covariant quantity is pµ but our non-relativistic
intuition splits it into an energy-momentum pair (E, p). In the same way, for BMS3
particles the truly covariant quantity is the full supermomentum p(ϕ), but our intuition
splits it into a pair (pPoinc , pvac ).
10.1.7 Spinning BMS3 particles

We finally turn to the spinning generalization of the BMS3 representations considered
above. We have already addressed almost all the subtleties of the construction, so we
display the inclusion of spin merely for completeness. In short, our main conclusion
will be that the spin of massive BMS3 particles is not quantized, exactly as in the
Poincaré group in three dimensions. The reader who is happy to accept this result, or
to deal only with scalar particles, may go to section 10.1.8.
We recall from (4.28) that spin is the label that specifies the representation of
the little group chosen for the description of a particle. In the case of BMS3 parti-
cles at non-zero c2 , the little groups were described at the end of subsection 10.1.1.
They are all either one-dimensional Abelian groups such as U(1) or R (possibly up
to discrete factors), or n-fold covers of the Lorentz group PSL(2, R). All these are
finite-dimensional Lie groups and their unitary representations are known, so writing
down generic spinning representations of BMS3 is mostly a technical problem.
For definiteness, let us consider a massive BMS3 particle. Its little group U(1) con-
sists of spatial rotations, exactly as for massive Poincaré particles in three dimensions.
All irreducible unitary representations of U(1) are of the form (2.13), with s an integer.
However, the BMS3 group (9.58) has the same homotopy type as Diff(S 1 ), which is
homotopic to a circle. This implies that it admits topological projective representa-
tions of the type described in section 2.1, which can be classified by considering exact
(non-projective) representations of the universal cover (9.61) of BMS3 . In the latter
case, the little group of massive particles gets unwrapped from U(1) to R, whose uni-
tary representations are now labelled by an arbitrary real spin s. Since those are the
physically relevant representations, we conclude that the spin of a massive BMS3 par-
ticle is generally an arbitrary real number. In particular, most massive BMS3 particles
are anyons. This is the same conclusion as in the Poincaré group in three dimensions.
Now suppose we fix a certain value of mass M and spin s ∈ R and ask what are
the states of the corresponding BMS3 particle. We denote the spin s representation of
the little group by R; it is a one-dimensional representation of the form (2.13). Thus
the Hilbert space of the BMS3 particle consists again of complex-valued wavefunctions
(10.8), but their transformation law under BMS3 contains an extra term with respect
to the scalar representation (10.10). That extra term involves the Wigner rotation
(4.31) associated with the superrotation f and the supermomentum q,
Wq [f ] = R[gq−1 f gf −1 ·q ] , (10.24)
where the superrotations gq (q ∈ Op ) are standard boosts such that gq · p = q. Thus we
run into the problem of finding standard boosts for a massive supermomentum orbit.
Incidentally we have already defined such standard boosts, though not in the same
language. Indeed, eq. (7.44) is precisely the definition of standard boosts for elliptic
Virasoro coadjoint orbits. These boosts are built as follows:
1. Take the supermomentum q(ϕ) with central charge c2 ; write down the associated
Hill’s equation (7.12) with the replacement (p, c) → (q, c2 ).
2. Find two linearly independent solutions ψ1 , ψ2 of Hill’s equation satisfying the
Wronskian condition (7.16).
3. Define a vector field Xq (ϕ) by (7.38).
4. Define a diffeomorphism f of S 1 by (7.41), with p0 = M − c2 /24
5. The standard boost associated with q is gq = f −1 .
This procedure is somewhat convoluted, but it does provide a family of standard boosts
on the orbit of a massive supermomentum, as desired. We refrain here from actually
computing these boosts.
Equipped with standard boosts one can write down the transformation law of
massive BMS3 particles with non-zero spin, given by eq. (4.30):
q
T [(f, α)] · Ψ (q) = ρf −1 (q) eihq,αi Wq [f ] · Ψ(f −1 · q) ,

where the notation is the same as in (10.10) up to the insertion of the Wigner rotation
(10.24). This can also be rewritten in terms of plane waves (10.13) as
q
T [(f, α)] · Ψk = ρf (k) eihf ·k,αi R gf−1

·k f gk · Ψf ·k
(recall eq. (4.32)). The interpretation of all these formulas is the same as before, up
to the extra Wigner rotation. In particular, a spinning BMS3 particle is a spinning
Poincaré particle dressed with soft gravitons.
10.1.8 BMS particles in four dimensions?

Having described BMS particles in three dimensions, we now ask to what extent our
observations apply to the realistic four-dimensional case. To answer this we first briefly
review some previous literature on BMS in four dimensions and its representations,
before exposing our viewpoint on the matter in light of the more recent developments
relating BMS symmetry to soft theorems.3 Our approach will be mostly qualitative
and heuristic.
3
To our knowledge there is, at present, no detailed review on BMS symmetry. Accordingly the
literature review provided here cannot fail to be biased by the author’s ignorance; we apologize in
advance for the references that we may have missed.
A history of BMS symmetry

BMS symmetry was discovered in the sixties by Bondi, Van der Burg, Metzner [1,2] and
Sachs [3,4], as a group (1.1) of globally well-defined diffeomorphisms of asymptotically
flat space-times. The notion of asymptotic flatness was put on firmer ground shortly
thereafter, thanks to the notion of conformal compactifications [307]. Arguably, at the
time of the discovery, the presence of an infinite-dimensional group of supertransla-
tions was seen as something of a pathology. Nevertheless it was quickly suggested that
full BMS symmetry (with the supertranslations turned on) could be used to discuss
aspects of both quantum gravity [308] and S-matrix physics [3]. This led to the study
of unitary representations of BMS.
The first suggestion that BMS representations might be relevant to particle physics
appeared in [3]. A little later McCarthy and collaborators set out to study the rep-
resentation theory of the global BMS group in full detail, with scattering amplitudes
as a motivation [41]. Owing to the semi-direct product structure of (1.1) the strategy
was to build induced representations à la Wigner in terms of orbits and little groups,
as described in this thesis in chapter 4. In particular it was shown in [38, 39] that all
little groups are compact, leading to the conclusion that BMS particles in four dimen-
sions cannot have continuous spin, in contrast to their Poincaré counterparts. This
was followed by the observation that the restriction of a BMS representation to its
Poincaré subgroup is reducible and consists of a tower of Poincaré particles with dif-
ferent spins [40]; in [309] this spin mixing was interpreted as being due to the presence
of the gravitational field. It was also shown in [310] that certain BMS representations
studied earlier in [311, 312] were in fact reducible induced representations.
Along the way it was realized that the absence of continuous-spin particles exhib-
ited in [38, 39] was due to a delicate choice of topology, and that different topologies
lead to radically different conclusions, including particles with continuous spin [313].
These continuous-spin particles were then interpreted as scattering states in [269,314].
Finally, the whole construction was put on firm mathematical ground in [315], where
the theory of induced representations was extended to groups of the type G n A with
an infinite-dimensional Abelian group A. As a corollary, it was shown in [316] that
induced representations à la Wigner exhaust all irreducible unitary representations of
the global BMS group (1.1). This analysis was later completed by the proof that the
global BMS group has no non-trivial central extensions, and therefore admits no pro-
jective representations other than those originating from its non-trivial topology [317].
BMS symmetry and holography

For about two decades after McCarthy’s work on BMS representations, the study of
BMS symmetry as such appears to have slowed down, with the exception of the dis-
covery of its supersymmetric version in [318]. Nevertheless, substantial progress was
made during that period in closely related areas, particularly in the study of the struc-
ture of gravity near null infinity (see e.g. [173, 270, 319, 320]). This led for instance
to the idea of asymptotic quantization [321, 322], which can roughly be thought of as
a quantization of bulk degrees of freedom obtained by quantizing a suitable radiative
phase space on the boundary — an idea that sounds prophetic given the development
of holography about ten years later.
As mentioned in the introduction of this thesis, holography emerged from general

considerations on the nature of quantum gravity [9, 10] guided by the seminal obser-
vation by Bekenstein and Hawking that black holes have entropy [11, 12]. Its best
known realization occurs in the AdS/CFT duality [13, 216, 323], but it was soon sug-
gested that a suitable notion of holography should hold for all families of space-times
(see e.g. [324, 325]), and in particular for asymptotically flat gravity. In that con-
text interest in BMS symmetry slowly re-emerged [326–330], leading in particular to
the proposal [5–7] that the globally well-defined BMS group (1.1) should be extended
to include a Virasoro-like semi-group of local conformal transformations of celestial
spheres. This proposal is the origin of the terminology of “superrotations” for asymp-
totic symmetry transformations that extend Lorentz transformations. The study of
asymptotically flat holography then took off, both in four dimensions [331–335] and in
three dimensions [46, 48, 198, 260, 263, 265, 266, 278, 281, 286, 336–339], although in the
latter case a substantial part of the literature is written in the language of Galilean
conformal symmetry [261, 264, 288, 292–295, 340–347].
Along the way, it was realized by Strominger and collaborators that BMS symmetry
does have highly non-trivial implications for the S-matrix, in the form of soft graviton
theorems [19, 24, 26, 27, 253]. This discovery sparked a flurry of papers discussing the
applications of the BMS group (and its gauge-theoretic generalizations [247–249, 251,
255–258]) to scattering amplitudes [348, 349], memory effects [250, 252, 254] and more
recently to black holes [350].
BMS4 particles?
Despite recent progress, we still seem to be quite far from having truly understood
BMS symmetry in four dimensions. Representation theory provides an easy way to
illustrate the problem. Indeed, one of the cornerstones of the relation between BMS
symmetry and soft theorems is the fact that supertranslations generate soft graviton
states when acting on the vacuum. Accordingly, if the global BMS group (1.1) is cor-
rect, then the representations considered by McCarthy in [38, 39] should account for
this effect: they should represent Poincaré particles dressed with soft gravitons.
However, it is easy to see that this is not the case. To illustrate this point, consider
the analogue of the global BMS group (1.1) in three dimensions,
gBMS3 ≡ PSL(2, R) n Vect(S 1 )Ab , (10.25)
where PSL(2, R) is the Lorentz subgroup of Diff(S 1 ) consisting of projective transfor-

mations (6.88) and acting on superrotations according to (6.17). The representations
of this group would be induced exactly in the same way as for the standard BMS3
group (9.58), but there would be two crucial differences:
• There would be no non-trivial central extensions.
• The supermomentum orbits would all be finite-dimensional (since PSL(2, R) is
finite-dimensional).
In particular the Hilbert space of any irreducible unitary representation of the group
(10.25) coincides with the space of states of a Poincaré particle; it consists of wave-
functions on a finite-dimensional momentum orbit. In fact, the only difference between
the representations of this group and those of Poincaré would be that translations are
paired with supermomenta according to the functional formula (6.34) rather than a
finite-dimensional product pµ αµ . In particular the vacuum representation would be
trivial and there would be no way for quantum supertranslations to create soft gravi-
ton states upon acting on the vacuum.
These observations exhibit an important point: the global BMS group (1.1) cannot
be the end of the story. It is too small to account for soft graviton degrees of freedom,
and it must be extended in some way. Unfortunately the argument does not tell us
how BMS symmetry should be extended. To our knowledge, two proposals have been
formulated so far, both suggesting a superrotational extension of the Lorentz group.
The first is the aforementioned idea of turning superrotations into a semi-group of
local conformal transformations of celestial spheres [5, 6, 18, 336, 351]; the second sug-
gests that superrotations should instead span a group Diff(S 2 ) of diffeomorphisms of
the sphere [352, 353]. It appears that there are currently no definitive arguments for
selecting one proposal over the other; see however [354], where it is argued that finite
singular conformal transformations of celestial spheres in four dimensions are patho-
logical.
The fact that BMS symmetry in four dimensions is ill-defined is a call for further
developments. This thesis is one of them: it aims at understanding a three-dimensional
toy model and using it as a guide for the realistic problem. Indeed, many properties
that we have encountered in our investigation of BMS3 should remain true in BMS4 .
In particular the semi-direct product structure G n A appears to be a robust feature
and implies that BMS particles are classified by supermomentum orbits that coincide
with orbits of the Bondi mass aspect under asymptotic symmetry transformations.
Furthermore the occurrence of a central extension pairing superrotations with super-
translations has also been observed in BMS4 [18]. However, a sharp difference between
BMS3 and BMS4 is that, in the former, supertranslations do not create new states when
acting on the vacuum. A possibly related difference is that the language suited to the
study of BMS4 appears to be that of groupoids rather than groups. We will not have
much more to say about this here, and return now to our study of BMS3 .
10.2 BMS modules and flat limits

In the previous pages we have described irreducible, unitary representations of BMS3 .
In order to make contact with the representation theory of the Virasoro algebra it is
useful to reformulate these representations in Lie-algebraic language, in the form of so-
called induced modules. This reformulation will also allow us to discuss the flat limit of
dressed particles in AdS3 and to understand the difference between unitarity in BMS3
and unitarity for Galilean conformal symmetry. The plan is as follows: we first describe
induced modules for the Poincaré algebra in three dimensions and interpret them as
ultrarelativistic limits of highest-weight representations of sl(2, R), before applying the
same construction to the bms3 algebra. The presentation is adapted from [51].
10.2.1 Poincaré modules in three dimensions

Our goal here is to rewrite the three-dimensional relativistic particles of section 4.3 in
Lie-algebraic language. This will allow us to relate Poincaré representations with the
ultrarelativistic limit of highest-weight representations of sl(2, R) ⊕ sl(2, R).
Poincaré algebra
In three dimensions, the Lie algebra of the Poincaré group is spanned by three Lorentz
generators jm and three translation generators pm (m = −1, 0, 1) with Lie brackets
(9.10). As in the case (8.54) of sl(2, R), it is more convenient to use a different
complexified basis Jm = ijm , Pm = ipm , in terms of which the brackets become
[Jm , Jn ] = (m − n) Jm+n , [Jm , Pn ] = (m − n) Pm+n , [Pm , Pn ] = 0 . (10.26)
These conventions are such that, in any unitary representation, the operators repre-
senting Poincaré generators satisfy Hermiticity conditions of the type (8.56):
(Pm )† = P−m , (Jm )† = J−m . (10.27)
As in section 8.4, we abuse notation by denoting with the same letter both the abstract
generators Jm , Pn and the operators that represent them.
The Poincaré algebra has two quadratic Casimir operators: the mass squared
M2 = P02 − P1 P−1 (10.28)
and the three-dimensional analogue of the square of the Pauli-Lubanski vector,

1
S = P0 J0 − (J1 P−1 + J−1 P1 + P1 J−1 + P−1 J1 ) . (10.29)
4
The eigenvalues of these operators classify irreducible representations according to
mass and spin, exactly as in section 4.3. See e.g. [62] for the proof of the fact that
the operator (10.28) actually takes the value M 2 in the space of states of a relativistic
particle with mass M .
Induced modules
Irreducible unitary representations of the Poincaré group are obtained by considering
the Lorentz orbit of a momentum p and building a Hilbert space of wavefunctions on
that orbit. A basis of this space is provided by plane waves (10.13), where the Dirac
distribution is determined by the choice of measure on the orbit. (For instance one
can take the Lorentz-invariant measure (3.4).) Their transformation laws are given
by (10.15). Here, in order to make the link with the standard Dirac notation, we
denote such plane waves by Ψk ≡ |k, si for any k ∈ Op , where s is the spin of the
representation.
For future comparison with bms3 , we focus on a relativistic particle with mass
M > 0. Its little group U(1) consists of spatial rotations generated by J0 . If we call p
the momentum of the representation in the rest frame, then the corresponding plane
wave Ψp ≡ |M, si satisfies
P0 |M, si = M |M, si , P−1 |M, si = P1 |M, si = 0 , J0 |M, si = s|M, si . (10.30)
From now on we call |M, si the rest frame state of the representation. Any other plane
wave Ψk = |k, si with boosted momentum k ∈ Op can be obtained by acting on |M, si
with a Lorentz transformation gk , where gk is a standard boost. In this sense the rest
frame state determines all the properties of the representation, in the same way that
highest-weight representations are determined by their highest-weight state. Note,
however, that the conditions (10.30) that define |M, si are not of the same form as the
highest-weight conditions (8.57) in that they involve both positive and negative modes.
Let us now understand how the conditions (10.30) induce a representation of the
Poincaré algebra. They define a one-dimensional representation of the subalgebra gen-
erated by {Pm , J0 }. This subalgebra consists of infinitesimal translations and spatial
rotations, i.e. it is a semi-direct sum u(1) A R3 where u(1) is generated by J0 while
R3 is generated by the Pm ’s. Thus the prescription (10.30) is a Lie-algebraic version
of the spin representation (4.28) for the case of a little group U(1) with R[θ] = eisθ .
Guided by our experience of induced representations, we can attempt to induce a
representation of the full Poincaré algebra out of the one-dimensional representation
(10.30); the result is known as an induced module (see e.g. section 10.7 of [355]). We
thus declare that the carrier space H of the representation is spanned by all states
obtained by acting on the rest frame state with operators that do not appear in the
conditions (10.30):
|k, l i = (J−1 )k (J1 )l |M, si , (10.31)
where k, l are non-negative integers. Such states are infinitesimally boosted states
analogous to the descendant states (8.58) that span Verma modules for the Virasoro
algebra. By definition, they form a basis of the space H . The latter provides a
Poincaré representation as it should, since acting from the left on the states (10.31)
yields linear operators on H whose commutators coincide with (10.26). Moreover, the
Casimir operators (10.28) and (10.29) have the same eigenvalue on each state (10.31),
since they commute by construction with all elements of the algebra. This readily
implies that the representation is irreducible.
Note that unitarity is not obvious in this picture: if one did not know that the in-
duced module follows from a manifestly unitary representation of the Poincaré group
in terms of wavefunctions, there would be no straightforward way to define a scalar
product on the space H spanned by the states (10.31), even after enforcing the Her-
miticity conditions (10.27). In fact, the norm squared of any plane wave state is strictly
infinite because of the delta function in (10.14). This is strikingly different from the
highest-weight representations of section 8.4, where the highest-weight conditions were
enough to evaluate the norm squared (8.59) of all descendant states.
Remark. The definition of the infinitesimally boosted states (10.31) follows from
the general construction of induced modules, as follows. Let g be a Lie algebra with
some subalgebra h. Let S be a one-dimensional representation of h. If U(g) denotes

the universal enveloping algebra of g, then the g-module T = Indgh (S ) induced by S
is the representation of g that acts in the space U(g) ⊗ C quotiented by the relations
Y ⊗ λ = 1 ⊗ S [Y ]λ for all Y ∈ h and all λ ∈ R. For any Lie algebra element
X ∈ g, the operator T [X] acts on the carrier space by hitting on vectors from the
left, in such a way that commutators of operators T [X] reproduce the Lie brackets
of the Lie algebra g. In this language the Poincaré module above is induced by the
representation (10.30) of the semi-direct sum gp A R3 , where R3 is the Lie algebra
of translations generated by the Pm ’s while gp is the Lie algebra of the little group
generated by J0 . The conditions (10.30) define a one-dimensional representation S of
gp A R3 , analogous to the spin representation (4.28).
Ultrarelativistic limit of sl(2, R) modules

In addition to being convenient for generalizations to infinite-dimensional extensions of
the Poincaré algebra, Poincaré modules can be seen as a limits of unitary representa-
tions of the AdS3 isometry algebra so(2, 2) ∼= sl(2, R) ⊕ sl(2, R). The generators of the
latter can be divided in two groups, Lm and L̄m with m = −1, 0, 1, whose Lie brackets
are two commuting copies of (8.55). In terms of these basis elements the quadratic
Casimir of each copy of sl(2, R) is (8.60). As usual our conventions are such that, in
any unitary representation, the Hermiticity conditions (8.56) hold in both sectors.
The Poincaré algebra (10.26) can be recovered from sl(2, R) ⊕ sl(2, R) as a flat
limit of the type described in section 9.4. Thus we introduce a length scale ` (to be
identified with the AdS radius) and define new generators as in (9.87):
1
Jm ≡ Lm − L̄−m , Pm ≡ (Lm + L̄−m ) . (10.32)
`
The resulting algebra is (9.88) without i’s on the left-hand side, and its limit ` → +∞
reproduces the Poincaré algebra (10.26). In addition the quadratic Casimir (8.60) can
be combined with its barred counterpart C, ¯ producing
2 ¯ = M2 + O(`−2 ), 1
C − C¯ = S,

C + C (10.33)
`2 `
where M2 and S are the Poincaré Casimirs (10.28) and (10.29).
The matching of Casimir operators suggests that the contraction also relates Poin-
caré modules to sl(2, R) ⊕ sl(2, R) representations. Concretely, consider the tensor
product of two highest-weight representations (8.57) of sl(2, R) with weights h, h̄:
L1 |h, h̄i = L̄1 |h, h̄i = 0 , L0 |h, h̄i = h|h, h̄i , L̄0 |h, h̄i = h̄|h, h̄i . (10.34)
This yields an irreducible representation of sl(2, R) ⊕ sl(2, R) whose carrier space is

spanned by descendant states (L−1 )m (L̄−1 )n |h, h̄i. Let us rewrite this representation
in terms of the operators (10.32). First we define the numbers
h + h̄
M≡ , s ≡ h − h̄ , (10.35)
`
which are eigenvalues of energy and angular momentum:
h + h̄
P0 |h, h̄i = |h, h̄i , J0 |h, h̄i = (h − h̄)|h, h̄i (10.36)
`
in terms of operators (10.32). Similarly, in terms of J’s and P ’s, the condition that
L1 and L̄1 annihilate the highest-weight state becomes
1
P±1 ± J±1 |h, h̄i = 0 . (10.37)
`
This allows us to reformulate the whole representation of sl(2, R) ⊕ sl(2, R) in terms
of operators Jm , Pn ; it results in expressions of the form
X (n) X (n)
Pn |k, li = Pk0 ,l0 ; k,l (M, s, `)|k 0 , l0 i , Jn |k, li = Jk0 ,l0 ; k,l (M, s)|k 0 , l0 i (10.38)
k0 ,l0 k0 ,l0
where the states |k, li take the form (10.31) with the identification |M, si ≡ |h, h̄i,
while P(n) and J(n) are infinite matrices. Owing to the definition (10.32) and property
(10.37), only negative powers of ` appear in (10.38). It follows that the matrix elements
(n) (n)
Pk0 ,l0 ; k,l and Jk0 ,l0 ; k,l have a well-defined limit ` → ∞. This limit coincides with the
result that one would find in a Poincaré module spanned by states (10.31), provided
that the conformal weights scale as
M` + s M` − s
h= + λ + O(1/`), h̄ = + λ + O(1/`), (10.39)
2 2
where λ is an arbitrary parameter independent of `. Thus, in the flat limit, the sl(2, R)
highest-weight conditions (10.37) are turned into a rest frame condition (10.30) and
the Poincaré Casimirs M2 and S take the values M 2 and M s, respectively. In short,
Poincaré modules are flat limits of sl(2, R) ⊕ sl(2, R) modules.
Note again that unitarity is subtle: starting from scalar products of sl(2, R) ⊕
sl(2, R) descendants, the flat limit gives rise to scalar products of states (10.31) that
diverge like positive powers of `. Indeed the norms (8.59) diverge when ` → +∞, ow-
ing to the fact that h is proportional to ` in (10.39). Equivalently, the wavefunctions
corresponding to states (10.31) become (derivatives of) delta functions in the flat limit;
from this point of view ` is an infrared regulator. Nevertheless, upon recognizing these
divergent scalar products as delta functions (10.14), one concludes that the Poincaré
module is a unitary representation in disguise.
Relation (10.39) shows that the contraction defined by (10.32) is an ultrarelativis-

tic/high-energy limit from the viewpoint of AdS3 . Poincaré modules are thus remnants
of so(2, 2) representations whose energy becomes large in the limit ` → ∞. In section
10.2.3 we shall see that the non-relativistic contraction from so(2, 2) to iso(2, 1) gives
rise to representations of a different type, that have been discussed in [287, 356].
We should mention that highest-weight representations of sl(2, R) can also be inter-

preted as induced modules. Indeed eq. (8.57) defines a one-dimensional representation
of the subalgebra spanned by {L0 , L1 }, while the vector space of descendant states
can be identified with a quotient of U(sl(2, R)) ⊗ C as discussed in the remark of page
315. The main difference with respect to Poincaré is the splitting of the algebra as
n− ⊕ h ⊕ n+ , where n± are nilpotent subalgebras, which allows one to evaluate scalar
products by enforcing the Hermiticity conditions (8.56).
10.2.2 Induced modules for bms3

Let us now apply the considerations of the previous pages to three-dimensional BMS
symmetry. The centrally extended bms3 algebra4 is spanned by superrotation gener-
ators Jm and supertranslation generators Pm (m ∈ Z) together with central charges
Z1 , Z2 , whose Lie brackets take the form (9.68). Following the Virasoro convention
(8.63), we change the normalization and define
Z1 Z2
Jm ≡ iJm + i δm,0 , Pm ≡ iPm + i δm,0 , (10.40)
24 24
as well as Z1 ≡ iZ1 and Z2 ≡ iZ2 . The constant shifts in P0 and J0 ensure that the
vacuum state has zero eigenvalues under these operators. According to this definition
the operators representing Jm and Pm in any unitary representation satisfy the Her-
miticity conditions (10.27). Furthermore, in any irreducible representation the central
charges Z1 and Z2 take definite values c1 , c2 , so we can write the commutation relations
of the bms3 algebra in a form analogous to (8.65):
c1
[Jm , Jn ] = (m − n)Jm+n + m(m2 − 1) δm+n,0 ,
12
c2
[Jm , Pn ] = (m − n)Pm+n + m(m2 − 1) δm+n,0 , (10.41)
12
[Pm , Pn ] = 0 .
In contrast to Poincaré, the quadratic operators (10.28)-(10.29) no longer commute

with the algebra (10.41). Nevertheless, the classification of BMS3 representations in
section 10.1 provides at least one obvious, yet non-trivial, Casimir operator.5 Indeed,
in the Hilbert space of any BMS3 particle, the “mass operator” (10.3) takes a definite
value when M is the monodromy matrix whose trace is given by the Wilson loop (7.21),
with c replaced by c2 and p(ϕ) replaced by the “supermomentum operator”
X c2
p̂(ϕ) = Pm e−imϕ −
m∈Z
24
where the Pm ’s are the supertranslation generators appearing in (10.41). Accordingly,

upon writing the right-hand side of (10.3) in terms of Pm ’s, one obtains a highly non-
linear combination of operators that commutes, by construction, with the entire bms3
algebra. (That it commutes with Pm ’s is trivial, since all supertranslations commute;
that it commutes with Jm ’s follows from the fact that (10.3) is invariant under super-
rotations.) The value of that Casimir operator can be used to classify BMS3 particles,
as we have done in section 10.1.
4
Recall that we use the same notation for both the bms3 algebra and its central extension.
5
I am indebted to Axel Kleinschmidt for this observation.
Aside from the mass operator (10.3), any function of the BMS3 central charges
c1 , c2 is clearly a Casimir. To our knowledge, whether this list exhausts all possible
bms3 Casimirs is an open question, though it seems plausible that it does since the
only Casimirs of the Virasoro algebra are functions of its central charges [228]. In
particular, it is not clear whether there exists a bms3 Casimir whose value specifies
the spin of a BMS3 particle, analogously to the Poincaré combination (10.29).
We now describe induced modules for the bms3 algebra (10.41), built analogously
to the Poincaré modules above and classified by their mass and spin (and central
charges). We discuss separately generic massive modules and the vacuum module,
and end by showing how they can all be obtained as ultrarelativistic limits of highest-
weight representations of Virasoro algebras.
Massive modules
Consider a BMS3 particle with mass M > 0 and spin s. Its supermomentum orbit
contains a constant p = M − c2 /24; the corresponding plane wave state Ψp ≡ |M, si
is such that
P0 |M, si = M |M, si, Pm |M, si = 0 for m 6= 0, J0 |M, si = s|M, si. (10.42)
Thus |M, si is a supermomentum eigenstate with vanishing eigenvalues under Pm ,
m 6= 0. In analogy with (10.30), we call |M, si the rest frame state of the module.
As in the Poincaré case, the conditions (10.42) define a one-dimensional represen-

tation of the subalgebra of (10.41) spanned by {Pn , J0 , c1 , c2 }. This representation can
be used to define an induced module H with basis vectors analogous to (10.31),
Jn1 Jn2 · · · JnN |M, si, (10.43)
where the ni ’s are non-zero integers such that n1 ≤ n2 ≤ ... ≤ nN . With this order-
ing, states (10.43) with different combinations of ni ’s are linearly independent within
the universal enveloping algebra of bms3 , and acting on them from the left with the
generators of the algebra provides linear operators on H whose commutators coincide
with (10.41). Thus one readily obtains a representation of the bms3 algebra.
As in the Poincaré case above, unitarity is hidden in this picture because there is
no straightforward way to compute scalar products of states (10.43). In fact, since
|M, si is a delta function, all such states strictly have infinite norm. This is because
realistic states of BMS3 particles are smeared wavefunctions that consist of infinite
linear combinations of plane waves. Unitarity can then be recognized in the fact that
acting with (finite) superrotations on |M, si produces a “basis” of plane waves that
generate a space of square-integrable wavefunctionals on the supermomentum orbit.
In particular the representation is automatically irreducible in the sense that all basis
states are obtained by acting with symmetry transformations on the single state |M, si.
Vacuum module
Recall that the BMS3 vacuum is the scalar representation whose supermomentum
orbit Diff(S 1 )/PSL(2, R) contains the vacuum configuration p = pvac = −c2 /24. The
corresponding induced module can be described similarly to massive ones. Owing to

the normalization (10.40), the Hilbert space of the vacuum representation contains a
plane wave Ψp ≡ |0i such that
Pm |0i = 0 for all m ∈ Z and Jn |0i = 0 for n = −1, 0, 1. (10.44)
Here the condition P0 |0i = 0 says that the vacuum has zero mass for the normalization
(10.40), while the extra conditions J±1 |0i = 0 enforce Lorentz-invariance. They reflect
the fact that the little group of the vacuum is the whole Lorentz group, rather than
the group of spatial rotations that occurs for massive particles.
If we were dealing with the Poincaré algebra, the requirements (10.44) would pro-
duce a trivial representation. Here, by contrast, there exist non-trivial “boosted vacua”
of the form (10.43), where now the ni ’s are integers different from −1, 0, 1. These vacua
are Lie-algebraic analogues of the boundary gravitons described earlier. The fact that
the vacuum is not invariant under the full BMS3 symmetry, but only under its Poincaré
subgroup, suggests that the boosted states (10.43) (with all ni ’s 6= −1, 0, 1) can be in-
terpreted as Goldstone-like states associated with broken symmetry generators; see the
discussion surrounding (8.81). Note that, in contrast to the realistic four-dimensional
case, BMS3 supertranslations do not create new states when acting on the vacuum.
Ultrarelativistic limit of Virasoro modules

In analogy with the observations of section 10.2.1, bms3 modules may be seen as limits
of tensor products of highest-weight representations of Virasoro. Let therefore Lm , L̄m
be generators of two commuting copies of the Virasoro algebra (8.65) with definite
central charges c, c̄. Highest-weight representations are then obtained starting from a
primary state |h, h̄i which satisfies (10.34) together with
Lm |h, h̄i = L̄m |h, h̄i = 0 for m > 0. (10.45)
The carrier space is spanned by descendant states
L−n1 ...L−nk L̄−n̄1 ...L̄−n̄l |h, h̄i (10.46)
with 1 ≤ n1 ≤ n2 ≤ ... ≤ nk and 1 ≤ n̄1 ≤ ... ≤ n̄l . Since we eventually wish to take
the ultrarelativistic limit of this representation, we will be interested in large values of
h and h̄, where the representation is irreducible and unitary thanks to the standard
Hermiticity conditions (8.64).
As in the Poincaré case, one can define new generators (10.32), now including
also the central charges c1 , c2 defined by (9.93). In particular, the space of Virasoro
descendants can be rewritten in the basis (10.43) with the identification |M, si ≡ |h, h̄i,
where M and s are the eigenvalues of P0 and J0 related to h and h̄ by (10.35). The
change of basis from descendant states (10.46) to infinitesimally boosted states (10.43)
is invertible because none of the Jn ’s annihilate the highest-weight state. The resulting
Virasoro representation takes a form analogous to (10.38), where now each state is
labelled by the quantum numbers ni of (10.43) and the matrices P(n) and J(n) also
depend on the central charges (9.93). As before, only negative powers of ` enter P(n)
via the highest-weight conditions (10.45) written in the new basis:
1
P±n ± J±n |h, h̄i = 0. (10.47)
`
A limit ` → ∞ performed at fixed M , s and c1 , c2 (rather than fixed h, h̄ say) then
yields a massive bms3 module of the type described above. In particular, the limit
maps the highest-weight state (10.45) on the rest frame state (10.42). In this sense
bms3 modules are high-energy limits of tensor products of Virasoro modules, since h
and h̄ go to infinity in the flat limit. By the way, this provides an intuitive picture of
why the energy spectrum of BMS3 particles is continuous: the typical distance between
two consecutive eigenvalues of P0 = (L0 + L̄0 )/` is 1/`, which shrinks to zero when `
goes to infinity.
10.2.3 Representations of the Galilean conformal algebra

Here we revisit the Galilean conformal algebra introduced in section 9.4. As explained
there, gca2 coincides with bms3 , which effectively makes them classically interchange-
able. We now argue that this equivalence does not hold at the quantum level; this
observation will be the basis of our arguments in section 11.3, when explaining why the
quantization of asymptotically flat gravity cannot be a Galilean conformal field the-
ory. The highest-weight representations that we shall describe here were first obtained
in [287], but their identification with induced representations of BMS3 is new.
Highest-weight representations of gca2

The gca2 algebra is isomorphic to bms3 , but their interpretations differ: in bms3 the
non-Abelian generators generalize the angular momentum operator and span superro-
tations, while the Abelian ones generalize the Hamiltonian and span supertranslations.
By contrast, in gca2 , the non-Abelian generators are the ones that generalize the Hamil-
tonian, and the Abelian ones generalize (angular) momentum. Due to this difference,
one is naturally led to look for unitary representations of gca2 where the operator J0 of
(10.41) is bounded from below. From the BMS3 perspective this is an awkward choice
(since it explicitly breaks parity by forcing all states to have the same sign of angular
momentum), but from the Galilean viewpoint it is perfectly well motivated.
To describe these representations we use the method of induced modules applied

to the algebra (10.41), with the Hermiticity conditions (10.27). To stress that we
are dealing with Galilean rather than relativistic representations, we denote all gca2
generators with a tilde on top, such as J˜m , P̃m , plus central charges c̃1 , c̃2 . In order
to obtain a representation where the spectrum of J˜0 is bounded from below, we start
from a state |M̃ , s̃i which has highest weight for the Virasoro subalgebra generated by
˜ in the sense that
J’s
J˜0 |M̃ , s̃i = s̃|M̃ , s̃i, P̃0 |M̃ , s̃i = M̃ |M̃ , s̃i (10.48)
and
J˜m |M̃ , s̃i = P̃m |M̃ , s̃i = 0 for m > 0 . (10.49)
We stress that s̃ is now interpreted as a (dimensionless) energy while M̃ is a (dimen-

sionful) momentum. In analogy with Virasoro representations, one can then define
descendant states of the form
P̃−k1 ...P̃−kn J˜−l1 ...J˜−lm |M̃ , s̃i (10.50)
with 1 ≤ k1 ≤ ... ≤ kn , 1 ≤ l1 ≤ ... ≤ lm , and declare that they form a basis of

the carrier space. The conditions (10.49) allow one to evaluate the would-be scalar
products of such descendants upon using the Hermiticity conditions (10.27), and one
finds that the representation is non-unitary whenever M̃ 6= 0 or c̃2 6= 0 [287]. When
M̃ = c2 = 0, unitarity requires in addition that s̃ > 0 and that the superrotation
central charge c̃1 be non-negative. Thus, unitary representations of gca2 boil down to
˜
highest-weight representations of its Virasoro subalgebra generated by the J’s.
A similar construction can be applied to a vacuum-like highest-weight representa-

tion of gca2 , whose highest-weight state |0̃i is annihilated by all Poincaré generators
J˜−1 , J˜0 , J˜1 , P̃−1 , P̃0 , P̃1 and by all positive modes as in (10.48):
P̃m |0̃i = 0 , J˜m |0̃i = 0 for m ≥ −1. (10.51)
Again, one concludes in that case that the representation is unitary if and only if
c̃2 = 0 and c̃1 ≥ 0.
The highest-weight representations of the type just described which are unitary (i.e.
have M̃ = 0 and c̃2 = 0) are special cases of induced representations of BMS3 as de-
scribed in section 10.1. Indeed, consider the vanishing supermomentum (p̃, c̃2 ) = (0, 0).
Its orbit under superrotations is trivial and its little group is the whole Virasoro group,
so the corresponding induced representation is entirely determined by its spin s̃. The
latter labels a unitary highest-weight representation of Virasoro, with central charge
c̃1 say. At the Lie-algebraic level this spin representation takes the form of a highest-
weight representation (10.48)-(10.49) with M̃ = 0 and c̃2 = 0. There is an analogue of
this construction in the Poincaré group: the vanishing momentum vector p = 0 has a
trivial orbit and its little group is the whole Lorentz group, so the corresponding in-
duced representation of Poincaré is just a unitary representation of the Lorentz group;
it is a “vacuum with spin” of the type mentioned in section 4.2.
The difference betwen the BMS3 vacuum (10.44) and the Galilean vacuum (10.51)
implies sharp differences for all quantum systems enjoying such symmetries, since it
affects the definition of normal ordering. For example, the normal-ordered product
: J2 P−3 : equals J2 P−3 in a BMS3 -invariant theory, while in a Galilean conformal field
theory one has : J˜2 P̃−3 : = P̃−3 J˜2 . We shall see explicit illustrations of this phenomenon
in section 11.3 below, when dealing with non-linear higher-spin symmetry algebras. It
suggests in particular that theories enjoying bms3 symmetry or gca2 symmetry differ
greatly at the quantum level, despite the isomorphism bms3 ∼ = gca2 .
Galilean limit of Virasoro modules

We now recover the (generally non-unitary) Galilean highest-weight representations
defined by (10.48)-(10.49) as a non-relativistic limit of Virasoro modules. As before
we let the generators Lm , L̄n satisfy the algebra (8.65) with central charges c, c̄ respec-
tively, and we consider a highest-weight representation of the type (10.34)-(10.45). In
order to take the non-relativistic limit (9.94), we introduce a length scale ` and define
1
J˜n ≡ L̄n + Ln ,

P̃n ≡ L̄n − Ln . (10.52)
`
We stress that the combinations of Lm ’s appearing here differ from those of the ultra-
relativistic limit (10.32). In particular, J˜0 now generates time translations while P̃0
generates spatial translations; the parameter ` should no longer be interpreted as the
AdS3 radius, and there is no mixing between positive and negative modes. In these
terms, the limit ` → +∞ of the direct sum of two Virasoro algebras reduces to a gca2
algebra (10.41) with tildes on top of all generators, including the central charges
c̄ − c
c̃1 = c̄ + c , c̃2 = . (10.53)
`
Note that, up to central charges, the same redefinitions applied to sl(2, R) ⊕ sl(2, R)
reproduce the Poincaré algebra iso(2, 1); this is a non-relativistic limit to be contrasted
with the ultrarelativistic limit described in section 10.2.1.
In analogy with section 10.2.2, let us rewrite the tensor product of two highest-
weight representations of Virasoro in terms of the operators (10.52). Given the weights
h, h̄ we define
h̄ − h
s̃ ≡ h̄ + h , M̃ ≡
`
which we stress is radically different from the ultrarelativistic redefinition (10.35).
Upon identifying |M̃ , s̃i ≡ |h, h̄i, the highest-weight state satisfies (10.48) and (10.49).
These conditions hold for any value of `, including the limit ` → +∞. The descendant
states (10.50) then provide a representation of the sum of two Virasoro algebras, which
in the non-relativistic limit ` → +∞ becomes a generically non-unitary representation
of gca2 . Unitarity is recovered if M̃ = c̃2 = 0 and s̃, c̃1 ≥ 0. Again, this is strikingly
different from the ultrarelativistic contraction described above.
Remark. The difference between gca2 modules and bms3 modules has been known,
albeit in disguise, ever since the nineties. Namely, the tensionless limit of string theory
gives rise to so-called null strings [357], whose worldsheet is a null surface and thus
provides a stringy generalization of null geodesics. It was observed in [358] that the
algebra of constraints arising from worldsheet reparameterization invariance of null
strings is the bms3 algebra, although the name “BMS” was not used at the time.6
In the same paper the authors observed that a suitable normal-ordering prescription
gives rise to a consistent quantization of the null string in any space-time dimension,
and systematically results in a continuous mass spectrum. This result is the stringy
analogue of the bms3 modules described in section 10.2.2. Only later was it realized
that a different normal ordering prescription [359, 360] gives rise to the same criti-
cal dimension as in standard string theory (26 for the bosonic string and 10 for the
superstring), but that the resulting spectrum is massless and discrete; in fact, the spec-
trum then coincides with the massless part of the spectrum of standard string theory.
6
This occurrence of the bms3 algebra predates its gravitational description [20] by a decade!
The latter result is the stringy analogue of the gca2 modules described here and the
critical dimension is analogous to the requirement c̃2 = 0 that ensures unitarity for
non-relativistic modules; see [361–363] for a recent account of these results.
10.3 Characters of the BMS3 group

Now that we are acquainted with BMS3 particles, we can start using them as a com-
putational tool. In this section we use the Frobenius formula (4.33),7
Z
dµ(k) δ(k, f · k) eihk,αi χR [gk−1 f gk ] ,

χ[(f, α)] = Tr T [(f, α)] = (10.54)
Op
to evaluate characters of rotations f and supertranslations α in induced representations

of the (centrally extended) BMS3 group. Remarkably, the localization effect due to
the delta function will allow us to compute characters despite the fact that we do
not know explicit measures on supermomentum orbits. We focus on massive BMS3
particles and on the BMS3 vacuum, and compare the results to the ultrarelativistic
limit of Virasoro characters. The results reviewed here were first reported in [48]; their
application to partition functions [49, 50] will be exposed in the next chapter.
10.3.1 Massive characters

We consider a BMS3 particle with mass M > 0 and spin s ∈ R, so that the little group
representation is R[θ] = eisθ . Our goal is to evaluate the character (10.54) for arbitrary
BMS3 transformations (f, α), along the lines described in section 4.2 for the Poincaré
group. The only subtlety is that now Op ∼ = Diff(S 1 )/S 1 is an infinite-dimensional su-
permomentum orbit; the gk ’s of eq. (10.54) are standard boosts and the pairing hk, αi
is given by (6.34). We assume as before that there exists a quasi-invariant measure
µ on the supermomentum orbit Op , but we stress again that different measures yield
equivalent representations so that the end result will be independent of µ. We shall
verify this point explicitly below.
The character (10.54) vanishes if f is not conjugate to an element of the little

group U(1); furthermore it is a class function, so we may take f (ϕ) = ϕ + θ to be a
pure rotation by some angle θ, without loss of generality. We assume for simplicity
that θ is non-zero; the case θ = 0 is radically different, and we shall briefly comment
about it below. When θ 6= 0, the delta function δ(k, f · k) of (10.54) localizes the
supermomentum integral to the unique point of Op that is left invariant by rotations,
namely the supermomentum at rest p = M − c2 /24. This allows us to pull the little
group character χR [f ] = eisθ out of the integral and to reduce the pairing hk, αi to a
product pα0 = M α0 − c2 α0 /24, so the whole character (10.54) boils down to a BMS3
analogue of eq. (4.55):
Z
isθ iα0 (M −c2 /24)
χ[(f, α)] = e e dµ(k) δ(k, f · k) . (10.55)
Op
7
Here δ denotes the delta function associated by (3.39) with the measure µ.
To evaluate the character it only remains to integrate the delta function. This
requires local coordinates on the orbit in a neighbourhood of p, which can be obtained
by Fourier-expanding each supermomentum k(ϕ) as in eq. (6.116). The Fourier modes
∗
kn = k−n then transform under rotations f (ϕ) = ϕ + θ according to
kn 7→ [f · k]n = kn einθ .
As we shall see, the character that follows from this transformation is divergent due
to the fact that the group is infinite-dimensional. To cure this divergence we consider
complex rotations rather than real ones and introduce a complex parameter
1
τ≡ (θ + i) (10.56)
2π
where > 0. We then define the transformation of supermomentum Fourier modes
under complex rotations to be

 kn e2πinτ if n > 0,
kn 7→ [f · k]n = k0 if n = 0, (10.57)
kn e2πinτ̄ if n < 0.

We will see below that this modification can be justified by thinking of BMS3 repre-
sentations as high-energy limits of Verma modules. Note that this prescription leaves
room for “Euclidean” rotations (i.e. rotations by an imaginary angle) while preserving
the reality condition (kn )∗ = k−n .
The problem now is to express the measure µ and the corresponding delta function
δ in terms of Fourier modes. On the massive supermomentum orbit Op ∼ = Diff(S 1 )/S 1 ,
the non-zero Fourier modes of k(ϕ) determine its energy k0 . This is analogous to the
statement that thepenergy of a relativistic particle is determined by its momentum
according to E = M 2 + k2 . Let us prove this in a neighbourhood of the supermo-
mentum at rest, p = M − c2 /24, by acting on it with an infinitesimal superrotation X
that we Fourier-expand as X
X(ϕ) = i Xn e−inϕ .
n∈Z
Since the action of superrotations on supermomenta is the coadjoint representation

(6.115) of the Virasoro algebra, we find a variation
X c2 X
(δX p) (ϕ) = 2n M + (n2 − 1) Xn e−inϕ ≡ δpn e−inϕ . (10.58)
n∈Z
24 n∈Z
Here the variation of the zero-mode, δp0 , vanishes for any choice of X. By contrast, all
other Fourier modes are acted upon in a non-trivial way and can therefore take arbi-
trary values by a suitable choice of X. This implies that (at least in a neighbourhood
of p) the non-zero modes of supermomenta provide local coordinates on Op . In terms
of the Fourier decomposition (6.116), this is to say that when k(ϕ) = p + ε(δX p)(ϕ),
the non-zero modes kn coincide with εδpn (while k0 = p to first order in ε).
It follows that in terms of kn ’s, the supermomentum measure µ of (10.55) reads

Y
dµ(k) = (Some k-dependent prefactor) × dkn (10.59)
n∈Z∗
where theQprefactor is unknown. In quantum mechanics one would write the infinite
product n∈Z∗ dkn as a path integral measure Dk, with the extra rule that the zero-
mode of k is not to be integrated over. The definition of the delta function (3.39)
associated with µ ensures that
Y
δ(q, k) = (Some k-dependent prefactor)−1 × δ(qn − kn ),
n∈Z∗
where the δ on the right-hand side is the usual Dirac distribution in one dimension.
Crucially, the prefactor appearing in front of the delta function is the inverse of the
prefactor of the measure (10.59). As in eq. (4.56) this implies that the combination
dµ(k)δ(k, ·) is invariant under changes of measures, and it allows us to rewrite the
character (10.55) as
Z
isθ iα0 (M −c2 /24)
Y Y
χ[(f, α)] = e e dkn δ (kn − [f · k]n )
R2∞ n∈Z∗ n∈Z∗
Z Y +∞ +∞ 2
(10.57) 0 (M −c
Y
eisθ eiα 2 /24)
δ kn (1 − e2πinτ )

= dkn , (10.60)
R∞ n=1 n=1
where we have replaced the real angle θ by its complex counterpart 2πτ given by
(10.56). Denoting q ≡ exp[2πiτ ] and evaluating the integral, the character of a massive
BMS3 particle finally reduces to
0 (M −c /24) 1
χ[(f, α)] = eisθ eiα 2
Q+∞ n 2
. (10.61)
n=1 |1 − q |
We stress that this holds only provided f is conjugate to a rotation by θ. We recognize

here the ubiquitous factor (1 − q)−1 arising from the Atiyah-Bott fixed point theorem
(4.61). The result can also be rewritten in terms of the Dedekind eta function (8.75),
|q|1/12 isθ iα0 (M −c2 /24)

χ[(f, α)] = e e ,
|η(τ )|2
with |q| = 1 in the (pathological) limit → 0.
Remark. At this stage, and in contrast to conformal field theory, the coefficient
τ should not be seen as a modular parameter. The small parameter in (10.56)
was merely introduced to ensure convergence of the determinant arising from the
integration of the delta function in (10.60). This being said, the occurrence of the
Dedekind eta function is compatible with the modular transformations used in [337,
341] to derive a Cardy-like formula reproducing the entropy of flat space cosmologies.
10.3.2 Comparison to Poincaré and Virasoro

Formula (10.61) extends the Poincaré character (4.98) in three dimensions. Indeed,
taking = 0 in (10.61) and forgetting about all convergence issues, one finds
+∞
isθ iα0 (M −c2 /24)
Y 1
χ[(f, α)] = e e 2 .
n=1
4 sin (nθ/2)
Here the term n = 1 coincides (4.98), while the contribution of higher Fourier modes
is due, loosely speaking, to the infinitely many Poincaré subgroups of BMS3 . This is
analogous to the fact that the Virasoro character (8.74) may be seen as a product of
infinitely many SL(2, R) characters (5.102) labelled by an integer n.
The divergence of the BMS3 character (10.61) as → 0 is identical to that of the

Virasoro character (8.74) as τ becomes real. In this sense, the divergence is not a
pathology of BMS3 , but rather a general phenomenon to be expected from infinite-
dimensional groups; here we have cured this divergence by adding an imaginary part
i to the angle. The origin of this imaginary part can be traced back to the fact
that BMS3 representations are ultrarelativistic limits of Virasoro representations, as
discussed at length in section 10.2.2. Indeed, suppose we are given a tensor product
of two Virasoro representations with highest weights h, h̄ and central charges c, c̄. The
corresponding character generalizes the partition function (8.80) as
(8.74) q h−c/24 q̄ h̄−c̄/24
L0 −c/24 L̄0 −c̄/24
Tr q q̄ = Q+∞ n 2
, q = e2πiτ . (10.62)
n=1 |1 − q |
Writing the modular parameter in the form (8.78) with an `-independent β and intro-
ducing a mass M and a spin s defined by (10.35), the large ` limit of the quantities
appearing in the right-hand side of (10.62) is
1
τ∼ (θ + i), q h−c/24 q̄ h̄−c̄/24 ∼ eiθ(s−c1 /24) e−β(M −c2 /24) . (10.63)
2π
(Here the imaginary part of τ goes to zero, but we keep writing it as > 0 to reproduce
the regularization used in (10.61).) Thus the flat limit of (10.62) is

L0 −c/24 L̄0 −c/24
1
lim Tr q q̄ = eiθ(s−c1 /24) e−β(M −c2 /24) Q+∞ n 2
,
n=1 |1 − q |
`→+∞
and coincides (up to a redefinition of spin) with the BMS3 character (10.61) for a
supertranslation whose zero-mode is a Euclidean time translation, α = iβ. The left-
hand side of this expression can be interpreted as a trace
(10.63)
lim Tr q L0 −c/24 q̄ L̄0 −c/24 = Tr eiθ(J0 −c1 /24) e−β(P0 −c2 /24) = χ[(f, α)], (10.64)

`→+∞
where f is a rotation by θ and the operators Jm , Pn are normalized so as to satisfy

the commutation relations (10.41). In this form, the matching between the flat limit
of the Virasoro character (10.62) and the BMS3 character (10.61) is manifest.
Universality of BMS3 characters

Even though BMS3 and Virasoro characters are related by the limit just described,
they are strikingly different in that the result (10.61) holds for any value of the central
charge c2 , any mass M , and any spin s. By contrast, the characters of irreducible,
unitary highest weight representations of the Virasoro algebra depend heavily on the
values of the central charge c and the highest weight h: when c ≤ 1, only certain
discrete values of c and h lead to unitary representations, and the resulting character
is not given by (10.62) [232, 239, 240]. In that sense, induced representations of the
BMS3 group are less intricate than highest weight representations of the Virasoro al-
gebra. Since the former are high-energy, high central charge limits of the latter, this
could have been expected: all complications occurring at small c vanish when ` goes
to infinity, since c scales linearly with ` by assumption.
We could also have guessed that such a simplification would occur thanks to di-
mensional arguments. Indeed, both M and c2 are dimensionful parameters labelling
BMS3 representations, so their values can be tuned at will by a suitable choice of
units. Accordingly, in contrast to Virasoro highest weight representations, one should
not expect to find sharp bifurcations in the structure of BMS3 particles as M and c2
vary. In this sense formula (10.61) is a universal character.
10.3.3 Vacuum character

We now turn to the character of the BMS3 vacuum, that is, the scalar representation
whose supermomentum orbit is that of pvac = −c2 /24. The computation is identical
to that of section 10.3.1, save for the fact that the little group is the Lorentz group
PSL(2, R) rather than U(1), so the orbit Ovac ∼ = Diff(S 1 )/PSL(2, R) has codimension
three rather than one in Diff(S 1 ).
As before, the quantity we wish to compute is χ[(f, α)], where α is any super-
translation. Since the little group is now larger than U(1), one can obtain non-trivial
characters even when f is not conjugate to a rotation. We will not consider such
cases here and stick instead to our earlier convention that f (ϕ) = ϕ + θ is a rotation
by θ 6= 0. (Equivalently we may take f to be merely conjugate to a rotation since
the character is a class function.) Then the integral of the Frobenius formula (10.54)
localizes to the unique rotation-invariant point pvac on the orbit and the character can
be written as Z
−iα0 c2 /24
χvac [(f, α)] = e dµ(k) δ(k, f · k). (10.65)
Ovac
Here µ is some quasi-invariant measure on the vacuum orbit. Using Fourier expan-
sions (6.116), we can think of Fourier modes as redundant coordinates on the orbit;
the subtlety is to understand which of these modes should be modded out so as to
provide genuine, non-redundant local coordinates on Ovac .
As in the case of massive characters we work in a neighbourhood of the rest frame

supermomentum pvac and rely on the action (10.58) of infinitesimal superrotations.
Taking M = 0 in that equation, we now find that all three modes δp1 , δp0 and δp−1
vanish for any choice of X. This is an infinitesimal restatement of the fact that the
little group is PSL(2, R). Thus, in a neighbourhood of pvac , we can use the higher
Fourier modes pn with |n| ≥ 2 as local coordinates. In particular the measure µ now
takes the form
+∞
Y
dµ(k) = (Some k-dependent prefactor) × dkn dk−n ,
n=2
where the prefactor is again unknown, but eventually irrelevant since it is cancelled
by the prefactor of the corresponding delta function. The vacuum character (10.65)
thus boils down to
Z +∞ +∞
−iα0 c2 /24
Y Y
χvac [(f, α)] = e dkn dk−n δ(kn − [f · k]n )δ(k−n − [f · k]−n )
R2∞−2 n=2 n=2
Z +∞ +∞ 2
(10.57) −iα0 c2 /24
Y Y
n
= e dkn δ (kn (1 − q )) ,
R∞−1 n=2 n=2
where q ≡ exp[2πiτ ] and τ = (θ + i)/2π contains an imaginary part that regularizes

the divergence of the infinite product. Integrating the delta functions and taking into
account the determinant, we finally obtain
0c 1
χvac [(f, α)] = e−iα 2 /24
Q+∞ n 2
. (10.66)
n=2 |1 − q |
Note the truncated product starting at n = 2, which reflects Lorentz-invariance. As

in the massive case above, this expression can be interpreted as a trace (10.64), now
taken in the Hilbert space of the vacuum representation. It can also be recovered as a
flat limit of the product of two Virasoro vacuum characters (8.77).
Remark. In this section we have systematically assumed that f is a rotation by

some non-zero angle θ. In doing so we have left aside the interesting problem of com-
puting characters of pure supertranslations. This includes for instance Euclidean time
translations, whose characters coincide with canonical partition functions of BMS3
particles. Analogously to the Poincaré results (4.68) or (4.99), all such characters are
infrared-divergent and rely on an integral taken over the whole supermomentum orbit,
due to the lack of a localizing delta function. We will not attempt to evaluate these
characters here.
Chapter 11
Partition functions and characters
The asymptotic symmetries described in chapter 9 suggest that the quantization of

asymptotically flat gravitational fields in three dimensions provides unitary represen-
tations of BMS3 . In particular, it should be possible to identify BMS3 particles with
quantized gravitational fluctuations around suitable background metrics. The purpose
of this chapter is to confirm this identification by matching one-loop partition func-
tions of gravity with BMS3 characters. As a by-product, the method of heat kernels
that we shall use for this computation also allows us to evaluate partition functions
for fields with arbitrary spin, which will lead us to higher-spin extensions of BMS3
symmetry. As we will show, the resulting irreducible unitary representations can be
classified analogously to the BMS3 particles of chapter 10, and their characters match
one-loop partition functions of combinations of higher-spin fields in three-dimensional
Minkowski space.
The plan is as follows. We start in section 11.1 by evaluating one-loop partition

functions of free fields with arbitrary mass and spin in D-dimensional Minkowski space
at finite temperature and angular potentials. We show that the result is an exponen-
tial of Poincaré characters which, for spin two in D = 3, coincides with the vacuum
BMS3 character (10.66). In section 11.2 we extend this matching to higher-spin the-
ories in three dimensions by describing a method for obtaining induced irreducible
unitary representations of the corresponding asymptotic symmetry groups. Section
11.3 is devoted to the Lie-algebraic counterpart of that method, which we compare to
earlier proposals in the literature [356]. We show in particular that ultrarelativistic
and non-relativistic limits of quantum W algebras differ, which singles out induced
representations as the correct approach to flat space holography. Finally, in section
11.4 we define supersymmetric extensions of the BMS3 group, describe their irreducible
unitary representations and show that their characters coincide with one-loop partition
functions of asymptotically flat hypergravity. Sections 11.A and 11.B are technical ap-
pendices that summarize computations related to SO(n) characters which are useful
for sections 11.1 and 11.4, respectively.
The results described in this chapter first appeared in [49–51]. They are Minkow-
skian analogues of earlier observations on partition functions in AdS3 [69,70, 241] that
we already referred to in section 8.4. Note that our language in this chapter will be
somewhat different than in the previous ones, as we will rely much more heavily on
331
Chapter 11. Partition functions and characters 332
quantum field theory. On the other hand the group-theoretic tools that we will be
using are essentially the same as in chapter 10.
11.1 Rotating canonical partition functions

We wish to study one-loop partition functions of higher-spin fields in D-dimensional
Minkowski space at finite temperature 1/β, and with non-zero angular potentials. As
in section 4.2, we denote these potentials by θ~ = (θ1 , ..., θr ), where r = b(D − 1)/2c is
the rank of SO(D−1), that is, the maximal number of independent rotations in (D−1)
space dimensions; we assume D ≥ 3. The computation involves a functional integral
over fields living on a quotient of RD , where the easiest way to incorporate one-loop
effects is the heat kernel method. Accordingly we now briefly review this approach,
before applying it to bosonic fields and rewriting the resulting partition function as an
exponential of Poincaré characters; for spin two and D = 3, the result coincides with
the vacuum BMS3 character. Fermions will be treated separately in section 11.4.1.
11.1.1 Heat kernels and method of images

Our goal is to evaluate partition functions of the form
Z
Z(β, θ~ ) = Dφ e−S[φ] (11.1)
where φ is some collection of fields (bosonic or fermionic) defined on a thermal quotient

RD /Z of flat Euclidean space, satisfying suitable (anti)periodicity conditions. (The
explicit action of Z on RD , with its dependence on β and θ, ~ will be displayed below —
see eq. (11.7).) The functional S[φ] is a Euclidean action for these fields. Expression
(11.1) can be evaluated perturbatively around a saddle point φc of S, leading to the
semi-classical (one-loop) result
"
2
##
δ S
Z(β, θ~ ) ∼ e−S[φc ] det (11.2)
δφδφ φc
where the exponent # depends on the nature of the fields that were integrated out. The
quantity δ 2 S/δφ(x)δφ(y) appearing in this expression is a differential operator acting
on sections of a suitable vector bundle over RD /Z. The evaluation of the one-loop con-
tribution to the partition function thus boils down to that of a functional determinant.
After gauge-fixing, such determinants reduce to expressions of the form det(−∆ +

M 2 ), where ∆ is a Laplacian operator on RD /Z. These, in turn, can be evaluated
thanks to the method of heat kernels. In short (see e.g. [241, 364] for details), one can
express det(−∆ + M 2 ) on RD as an integral
Z +∞ Z
2 dt
− log det(−∆ + M ) = dD x Tr [K(t, x, x)] , (11.3)
0 t RD
up to an ultraviolet divergence that can be regularized with standard methods. Here

K(t, x, x0 ) is a matrix-valued bitensor known as the heat kernel associated with (−∆ +
M 2 ). It satisfies the heat equation

∂
K(t, x, x0 ) − (∆x − M 2 ) K(t, x, x0 ) = 0 , (11.4)
∂t
with the initial condition
K(t = 0, x, x0 ) = δ (D) (x − x0 ) I (11.5)
where I is an identity matrix having the same tensor structure as K (here omitted for
brevity) while δ (D) is the Dirac delta function associated with the translation-invariant
Lebesgue measure on RD .
Heat kernels are well suited for the computation of functional determinants on
quotient spaces. Indeed, suppose Γ is a discrete subgroup of the isometry group of
RD , acting freely on RD . Introducing the equivalence relation x ∼ y if there exists a
γ ∈ Γ such that γ(x) = y, we define the quotient manifold RD /Γ as the corresponding
set of equivalence classes. Given a differential operator ∆ on RD , it naturally induces
a differential operator on RD /Γ, acting on fields that satisfy suitable (anti)periodicity
conditions. Because the heat equation (11.4) is linear, the heat kernel on the quotient
space can be obtained from the heat kernel on RD by the method of images:
D
X
K R /Γ (t, x, x0 ) = K t, x, γ(x0 ) .

(11.6)
γ ∈Γ
Here, abusing notation slightly, x and x0 denote points both in RD and in its quotient.
In writing (11.6) we are assuming, for simplicity, that the tensor structure of K is
trivial, but as soon as K carries tensor or spinor indices (i.e. whenever the fields under
consideration have non-zero spin), the right-hand side involves Jacobians that account
D
for the non-trivial transformation law of K. Once K R /Z is known, the determinant of
D
the operator −∆ + M 2 is given by (11.3) with K replaced by K R /Z and RD replaced
by RD /Z.
We shall be concerned with thermal quantum field theories on rotating Minkowski

space, so we define our fields on a quotient RD /Z of Euclidean space with the action
of Z obtained as follows. For odd D, we endow RD with Cartesian coordinates (xi , yi )
(where i = 1, ..., r) and a Euclidean time coordinate τ , so that an integer n ∈ Z acts
on RD according to (see fig. 11.1)

n xi cos(nθi ) − sin(nθi ) xi xi
γ = · ≡ R(nθi ) · , γ n (τ ) = τ + nβ .
yi sin(nθi ) cos(nθi ) yi yi
(11.7)
For even D we add one more spatial coordinate z, invariant under Z. In terms of
the coordinates {x1 , y1 , ..., xr , yr , τ } (and also z if D is even), the Euclidean Lorentz
transformation implementing the rotation (11.7) is the nth power of the rotation matrix
··· 0 0
 
R(θ1 ) 0
··· 0
 
R(θ1 ) 0 ... .. 
... ..   0

0 . 0
 0 0 .

 .
J = .  or  .. (11.8)

 .. 0 R(θ r ) 0 0 
0 R(θr ) 0  
 0 ··· 0 1 0
0 ··· 0 1
0 ··· 0 0 1
for D odd or D even, respectively. Being isometries of flat space, these transformations
are linear maps in Cartesian coordinates, so their nth power coincides with the Jacobian
matrix ∂γ n (x)µ /∂xν that will be needed later for the method of images. Throughout
this chapter we take all angles θ1 , ..., θr to be non-vanishing and combine them in a
vector θ~ = (θ1 , ..., θr ). We now display the computation of one-loop partition functions
on RD /Z for bosonic higher-spin fields.
Figure 11.1: The quotient space R3 /Z defined by identifications of R3 generated by

the group action (11.7); β is an inverse temperature while θ is an angular potential.
11.1.2 Bosonic higher spins

Here we study the rotating one-loop partition function of a free bosonic field with
spin s and mass M (including the massless case). For M > 0 its Euclidean action
can be presented either (i) using a symmetric traceless field φµ1 ...µs of rank s together
with a tower of auxiliary fields of ranks s − 2, s − 3, ..., 0 that do not display any gauge
symmetry [365]; or (ii) using a set of doubly traceless fields of ranks s, s−1, ..., 0 subject
to a gauge symmetry generated by traceless gauge parameters of ranks s−1, s−2, ..., 0
[366]. In the latter case, the action is a sum of Fronsdal actions [367] for each of the
involved fields, plus a set of cross-coupling terms with one derivative proportional to
M , and a set of terms without derivatives proportional to M 2 . In the massless limit, all
these couplings vanish and one can consider independently the (Euclidean) Fronsdal
action for the field of highest rank:
Z
1 D µ1 ...µs 1 λ
S[φµ1 ...µs ] = − d xφ Fµ1 ...µs − δ(µ1 µ2 Fµ3 ...µs )λ , (11.9)
2 2
where indices are raised and lowered thanks to the Euclidean metric, while
Fµ1 ···µs ≡ ∆ φµ1 ...µs − ∂(µ1 | ∂ λ φ|µ2 ...µs )λ + ∂(µ1 ∂µ2 φµ3 ...µs )λ λ . (11.10)
Parentheses denote the symmetrization of the indices they enclose, with the minimum
number of terms needed and without any overall factor. The massless action (11.9) has
a gauge symmetry φµ1 ...µs 7→ φµ1 ...µs + ∂(µ1 ξµ2 ...µs ) , where ξµ2 ...µs is a symmetric tensor
field. When s = 2 the action reduces to that of a metric perturbation hµν around a
flat background. We refer for instance to [368] for many more details on this topic.
Massive case
Applying e.g. the techniques of [70] to the presentation of the Euclidean action of a
massive field of spin s of [366], one finds that the partition function is given by
1 1
log Z = − log det(−∆(s) + M 2 ) + log det(−∆(s−1) + M 2 ) , (11.11)
2 2
where ∆(s) is the Laplacian ∂µ ∂ µ acting on periodic,1 symmetric, traceless tensor fields
with s indices on RD /Z. We denote the heat kernel associated with (−∆(s) + M 2 ) on
RD by Kµs ,νs (t, x, x0 ), where µs and νs are shorthands that denote sets of s symmetrized
indices. The differential equation (11.4) with initial condition (11.5) for Kµs ,νs (t, x, x0 )
then reads
(∆(s) − M 2 − ∂t )Kµs , νs = 0 , Kµs , νs (t = 0, x, x0 ) = Iµs , νs δ (D) (x − x0 ) , (11.12)
where Iµs ,νs is an identity matrix with the same tensor structure as Kµs ,νs . Sets of
repeated covariant or contravariant indices denote sets of indices that are symmetrized
with the minimum number of terms required and without multiplicative factors, while
contractions involve as usual a covariant and a contravariant index. For instance the
tracelessness condition on the heat kernel amounts to
δ µµ Kµs , νs = δ νν Kµs , νs = 0 . (11.13)
The unique solution of (11.12) fulfilling this condition is

1 1
−M 2 t− 4t |x−x0 |2
Kµs , νs (t, x, x0 ) = D/2
e Iµs , νs (11.14)
(4πt)
where |x − x0 | is the Euclidean distance between x and x0 , while the spin-s identity
matrix is
b 2s c
X (−1)n 2n n! [D + 2(s − n − 2)]!! n s−2n n
Iµs , νs = δµµ δµν δνν . (11.15)
n=0
s! [D + 2(s − 2)]!!
Note that the dependence of this heat kernel on the space-time points x, x0 and on
Schwinger proper time t is that of a scalar heat kernel, and completely factorizes from
its spin/index structure which is entirely accounted for by the matrix I.2 This simpli-
fication is the reason why heat kernel computations are simpler in flat space than in
AdS or dS.
To determine the heat kernel associated with the operator (−∆(s) + M 2 ) on RD /Z,
we use the method of images (11.6), taking care of the non-trivial index structure.
Denoting the matrix (11.8) by Jα β (it is the Jacobian of the transformation x 7→ γ(x)),
the spin-s heat kernel on the quotient space RD /Z is
D
X
KµRs , ν/Zs (t, x, x0 ) = (J n )α β ...(J n )α β Kµs , βs t, x, γ n (x0 ) ,

(11.16)
n∈Z
1
More precisely, the field at time τ + β is rotated by θ~ with respect to the field at time τ .
2
Note also that the scalar heat kernel coincides with the propagator of a free particle in RD ,
whose expression for D = 2 was written in eq. (5.158).
where we recall again that repeated covariant or contravariant indices are meant to
be symmetrized with the minimum number of terms required and without multiplica-
tive factors, while repeating a covariant index in a contravariant position denotes a
contraction. Accordingly, eq. (11.3) gives the determinant of (−∆(s) + M 2 ) on RD /Z:
Z +∞ Z
(s) 2 dt D
− log det(−∆ + M ) = dD x (δ µα )s KµRs , α/Zs (t, x, x)
0 t RD /Z
Z +∞ Z
X
n µβ n µβ dt 1 1
−M 2 t− 4t |x−γ n (x)|2
= (J ) · · · (J ) Iµs ,βs dD x D/2
e .
n∈Z 0 t RD /Z (4πt)
(11.17)
In this series the term n = 0 contains both an ultraviolet divergence (due to the singular
behaviour of the integrand as t → 0) and an infrared one (due to the integral of a
constant over RD /Z), proportional to the product βV where V is the spatial volume
of the system. This divergence is a quantum contribution to the vacuum energy, which
we ignore from now on. The only non-trivial one-loop contribution then comes from
the terms n 6= 0 in (11.17). Using
r
X
|x − γ n (x)|2 = n2 β 2 + 4 sin2 (nθi /2)(x2i + yi2 )
i=1
in terms of the coordinates introduced around (11.7), the integrals over t and x give
rise to a divergent series
(
X 1 χs [nθ~ ] e−|n|βM if D odd,
− log det(−∆(s) + M 2 ) = r ×
|n| Q ML
K1 (|n|βM ) if D even,
n ∈ Z∗ |1 − einθj |2 π
j=1
R +∞ (11.18)
where K1 is the first modified Bessel function of the second kind, L ≡ −∞ dz is an
infrared divergence (4.64) that arises in even dimensions because the z axis is left fixed
by the rotation (11.8), and
s
χs [nθ~ ] ≡ (J n )µβ ...(J n )µβ Iµs , βs ≡ (J n )µβ Iµs , βs

(11.19)
is the full mixed trace of Iµs ,νs . As such, expression (11.18) makes no sense because
the sum over n diverges. To cure this problem, one needs to choose a regularization
procedure. Motivated by the similar situation already encountered in eq. (10.56), for
now we choose to regulate the series by a naive replacement: we let j , j = 1, ..., r be
small positive parameters and replace θj by θj ± ij in all positive powers of e±iθj . As
a result, expression (11.18) is replaced by the convergent series
(
(s) 2
X 1 ~
χs [nθ,~ ] e−|n|βM if D odd,
− log det(−∆ + M ) = r ×
|n| Q ML
K1 (|n|βM ) if D even,
n ∈ Z∗ |1 − ein(θj +ij ) |2 π
j=1
(11.20)
~
where χs [nθ,~ ] is still given by (11.19), except that now all factors e ±iθj
appearing in
±i(θj ±ij )
the Jacobians are replaced by e .
The regularization described here is motivated by the fact that, for odd D, the
resulting expressions look very much like flat limits of AdS one-loop determinants, in
which case the parameters j ∝ β/` are remnants of the inverse temperature (with `
the AdS radius). The subtlety, however, is that the exact matching of the flat limit
of AdS with combinations such as (11.20) requires some of the j ’s to be multiplied
by certain positive coefficients; thus eq. (11.20) is not quite the same as the flat limit
of its AdS counterpart — we will illustrate this point for D = 3 in section 11.1.3. As
for even values of D, the situation is even worse since the flat limit of the AdS result
contains an infrared divergence; it is not obvious how this divergence can be regular-
ized so as to reproduce the combination L · K1 of (11.20), though apart from this the
other terms of the expression indeed coincide with the flat limit of their AdS counter-
parts. From now on we will use the i prescription systematically, often omitting to
indicate it explicitly. We will keep it only in the final results, and in section 11.1.3 we
will introduce a refined regularization such that partition functions in D = 3 exactly
reproduce characters of suitable asymptotic symmetry algebras, while also matching
the flat limit of their AdS peers.
In eq. (11.20), the divergence as j → 0 is the same as in the BMS3 charac-

ter (10.61). The new ingredient is the angle-dependent trace (11.19); in appendices
11.A.1 and 11.A.2 we show that the latter is the character of an irreducible, unitary
representation of SO(D) with highest weight λs ≡ (s, 0, ..., 0). More precisely, let
Hi denote the generator of rotations in the plane (xi , yi ), in the coordinates defined
around (11.7). Then the Cartan subalgebra h of so(D) is generated by H1 , ..., Hr , plus,
if D is even, a generator of rotations in the plane (τ, z). Denoting the dual basis of
h∗ by L1 , ..., Lr (plus possibly Lr+1 if D is even), we can consider the weight λs = sL1
whose only non-zero component (in the basis of Li ’s) is the first one. The character
of the corresponding highest-weight representation of so(D) coincides with expression
(11.127):
(D) (D)
χs [nθ~ ] = χλs [nθ1 , ..., nθr ] or χλs [nθ1 , ..., nθr , 0] , (11.21)
(n)
for D odd or even, respectively. From now on, χλ denotes a character of SO(n) with
highest weight λ.
We can now display the one-loop partition function (11.11). Using expression
(11.20) for the one-loop determinant together with property (11.21), we find
  
(D) (D)
X+∞
n−1  χλs [nθ~ ] − χλs−1 [nθ~ ] e−nβM

~

Z(β, θ ) = exp
r × 
Q (D) ~ 0] − χ (D) ~ 0] M L K1 (nβM )

|1 − einθj |2  χλs [nθ, [n θ,
n=1

λs−1 π
j=1
(11.22)
where the upper (resp. lower) line corresponds to the case where D is odd (resp. even).
Remarkably, the differences of SO(D) characters appearing here can be simplified:
according to eqs. (11.154a) and (11.155), the difference of two SO(D) characters with
weights (s, 0, ..., 0) and (s − 1, 0, ..., 0) is a (sum of) character(s) of SO(D − 1):
(D) (D)

χλs [θ~ ] − χλs−1 [θ~ ] (D odd) 
(D−1)
= χλs [θ~ ] . (11.23)
(D) ~ (D) ~
χλs [θ, 0] − χλs−1 [θ, 0] (D even) 
Since the rank of SO(D − 1) is r = b(D − 1)/2c, the right-hand side of this equality
makes sense regardless of the parity of D, and the partition function (11.22) boils
down to
 
e−nβM
+∞ (D−1) ~
(
X 1 χλs [nθ,~ ] (D odd) 
~
Z(β, θ ) = exp
 ×  . (11.24)
r
n Q in(θ +i )
ML
K1 (nβM ) (D even)

n=1 |1 − e j j |2 π
j=1
Note that the function of nθ~ and nβ appearing here in the sum over n is essentially
the character (4.60)-(4.65) of a Poincaré particle with mass M and spin λs ; we will
return to this observation in section 11.1.4. An analogous result holds in Anti-de Sitter
space [369–371].
Massless case
We now turn to the one-loop partition function associated with the Euclidean Fronsdal
action (11.9), describing a massless field with spin s. The extra ingredient with respect
to the massive case is the gauge symmetry φµ1 ...µs 7→ φµ1 ...µs + ∂(µ1 ξµ2 ...µs ) . This forces
one to fix a gauge and introduce ghost fields that absorb the gauge redundancy [70],
which adds two more functional determinants to the massive result (11.11) and leads
to the following expression for the one-loop term of the partition function:
1 1
log Z = − log det(−∆(s) ) + log det(−∆(s−1) ) − log det(−∆(s−2) ) . (11.25)
2 2
As before, ∆(s) is the Laplacian on RD /Z acting on periodic, traceless, symmetric fields
with s indices. The functional determinants can be evaluated exactly as in the massive
case, upon setting M = 0. In particular, using limx→0 xK1 (x) = 1, the massless version
of the functional determinant (11.20) is
(
X 1 ~ ]
χs [nθ,~ 1 if D odd,
(s)
− log det(−∆ ) = r × L (11.26)
|n| Q if D even,
n ∈ Z∗ |1 − ein(θj +ij ) |2 π|n|β
j=1
which has been regularized as in the massive case. The matching (11.21) between χs
and a character of SO(D) remains valid, but a sharp difference arises upon including
all three functional determinants in (11.25). Indeed, the combination of χs ’s now is
(11.21)-(11.23) (D−1) (D−1)
χs [nθ~ ] − 2χs−1 [nθ~ ] + χs−2 [nθ~ ] = χλs [nθ~ ] − χλs−1 [nθ~ ] . (11.27)
It is tempting to use (11.23) once more to rewrite this as a character of SO(D − 2),
and indeed this is exactly what happens for even D because in that case the rank of
SO(D − 1) equals that of SO(D − 2):
 
+∞ (D−2) ~
~
X 1 χλs [nθ,~ ] L 
Z(β, θ ) = exp 

r
 (even D). (11.28)
n Q in(θ +i ) πnβ 
n=1 |1 − e j j |2
j=1
If D is odd, however, the rank decreases by one unit in going from SO(D − 1) to
SO(D − 2), so expression (11.27) contains one angle too much to be a character of
SO(D − 2). In fact, when D = 3, the right-hand side of (11.27) is the best we can
hope to get; for s ≥ 2 it takes the form
(1) (2) (2)
χλs [nθ] ≡ χλs [nθ] − χλs−1 [nθ] = eisnθ − ei(s−1)nθ + c.c. (11.29)
where we have used the character χs [θ] = eisθ + e−isθ for parity-invariant unitary
representations of SO(2) and “c.c.” means “complex conjugate”. (For lower spins one
(1) (1)
has χλ0 [θ] ≡ 1 and χλ1 [θ] ≡ 2 cos θ − 1.) Hence the partition function given by (11.25)
becomes
" +∞ #
(0)
X1 1
Z(β, θ) = e−S exp eisn(θ+i) − ei(s−1)n(θ+i) + c.c.

in(θ+i) |2
(D = 3)
n=1
n |1 − e
(11.30)
upon using the crude regularization described below eq. (11.20). For the sake of
generality we have included a spin-dependent classical action S (0) , whose value is
a matter of normalization and is generally taken to vanish, except for spin two (see
below). In the more general case where D is odd and larger than three, a simplification
does occur on the right-hand side of (11.27): as we show in appendix 11.A.3, the
difference (11.27) can be written as a sum of SO(D−2) characters with angle-dependent
coefficients (see eq. (11.154b)). Indeed, let us define
| cos((r − i)θj )|θk =0

Ark (θ~ ) ≡ , k = 1, ..., r, (11.31)
| cos((r − i)θj )|
where |Aij | denotes the determinant of an r × r matrix. Then the rotating one-loop
partition function for a massless field with spin s in odd space-time dimension D ≥ 5
reads
 r 
P r ~ (D−2)
+∞ A (nθ,~ ) χλs [nθ1 , ..., nθk , ..., nθr ,~ ]
d
~
X 1 k=1 k 
Z(β, θ ) = exp 

r
 (odd D ≥ 5),
n Q in(θ +i ) 2

n=1 |1 − e j j |
j=1
(11.32)
where the hat on top of an argument denotes omission.
Note that the massless partition functions (11.28) and (11.32) are related to the
massless limit of (11.24). Indeed, as we show in appendix 11.A.4, it turns out that
P (D−2)
r r ~
k=1 Ak (θ ) χλj [θ1 , ..., θbk , ..., θr ] for odd D,
Xs 
(D−1) ~
χ λs [ θ ] = (11.33)
χ(D−2) [θ~ ] for even D.
j=0 λj
Accordingly, the massless limit of a massive partition function with spin s is a product
of massless partition functions with spins ranging from 0 to s,
s
Y
lim ZM,s = Zmassless,j , (11.34)
M →0
j=0
consistently with the structure of the massive action [366]. This result stresses again
the role of the functions Ark (θ~ ) defined in (11.31): when the space-time dimension
is odd, one needs angle-dependent coefficients because the rank of the little group of
massless particles is smaller than the maximum number of angular velocities, so that
a single SO(D − 2) character cannot account for all of them. By the way, the results
(11.33) and (11.34) also hold in dimension D = 3, provided one takes the “characters”
(1) (1) (1)
χλs [θ] to be of the form (11.29) with χλ0 [θ] = 1 and χλ1 [θ] = 2 cos θ − 1.
11.1.3 Partition functions and BMS3 characters

Let us rewrite the three-dimensional partition function (11.30) in a form more con-
venient for the group-theoretic discussion of the remainder of this chapter. As men-
tioned above, the only non-trivial step will be to slightly modify the i regularization.
Namely, instead of the combination of exponentials appearing in (11.30), consider the
expression
eisn(θ+i) − ei(s−1)nθ−(s+1)n + c.c. (11.35)
Writing q ≡ ei(θ+i) and plugging (11.35) into the sum over n of eq. (11.30), one obtains
the series
+∞ +∞ +∞
1 q ns − q ns q̄ n + c.c. X 1
ns
X q X
n |2
= n
+ c.c. = − log(1 − q j ) + c.c. (11.36)
n=1
n |1 − q n=1
n 1 − q j=s
where the new regularization (11.35) has ensured that the summand decomposes as
the sum of a chiral and an anti-chiral piece in q. (This was not the case with the
rough regularization of eq. (11.20)!) In order to write down the full partition function,
it only remains to assign a value to the classical action S (0) ; a convention that has
come to be standard in the realm of three-dimensional gravity is to set S (0) = 0 for any
spin s 6= 2 (vacuum expectation values are assumed to vanish), while S (0) = −β/8G
for spin two (with G the Newton constant in three dimensions). This choice ensures
covariance of the on-shell action under modular transformations [49, 372], in analogy
with the similar choice in AdS3 [241]. All in all, combining the value of S (0) with the
series (11.36) and renaming j into n, one finds that the three-dimensional partition
function (11.30) can be written as
+∞
δs,2
βc2 Y 1
Z(β, θ) = e 24 , c2 = 3/G. (11.37)
n=s
|1 − ein(θ+i) |2
This expression is the flat limit of the analogous higher-spin partition function in
AdS3 [70]. But most importantly for our purposes, taking s = 2 in this formula, we
recognize the vacuum BMS3 character (10.66). This is our first key conclusion in this
chapter: it confirms that boundary gravitons in flat space form an irreducible unitary
representation of the BMS3 group of the type described in chapter 10. The case s > 2
will be studied in section 11.2, with similar conclusions.
The result (11.37) can be generalized to orbifolds in flat space: upon declaring
that the angular coordinate ϕ of (9.4) is identified as ϕ ∼ ϕ + 2π/N with some integer
N > 1, one obtains a flat three-dimensional conical deficit. One can then evaluate heat
kernels on that background by computing a sum over images (11.6), where Γ is now a
group Z × ZN whose two generators enforce (i) the thermal identifications (11.7), and
(ii) the orbifolding ϕ ∼ ϕ + 2π/N . An important technical subtlety is that, in order
to evaluate a partition function with temperature 1/β and angular potential θ on that
background, the angle appearing in (11.7) must be θ/N rather than θ. The rest of the
computation is straightforward, and one finds that the one-loop partition function of
gravity can be written as
+∞
−βp0
Y 1
ZN (β, θ) = e ,
n=1
|1 − ein(θ+i) |2
where p0 = −c2 /(24N 2 ). Comparing with (10.61), we recognize the (Euclidean) char-
c2
acter of a BMS3 particle with mass M = 24 (1 − 1/N 2 ), which is indeed the mass one
would obtain by writing the conical deficit metric in BMS form (9.25). Note in par-
ticular that the sum over images of the
Q orbifolding group ZN converts the truncated
product of (11.37) into a full product +∞
n=1 (· · · ).
The result (11.37) first appeared in [49] and parallels earlier observations in AdS3
[241]. It is tempting to conjecture that formula (11.37) is one-loop exact, since it is the
only expression compatible with BMS3 symmetry. This being said we will not need
to assume one-loop exactness in this thesis and we will not attempt to prove it. The
remainder of this chapter is devoted to various extensions of this matching.
Remark. In [339] it was shown that the one-loop partition function (11.37) with s =
2, and hence the vacuum BMS3 character (10.66), can be reproduced using quantum
Regge calculus. In that context the truncation of the product over n = 2, 3, ... is a
consequence of triangulation-invariance in the bulk.
11.1.4 Relation to Poincaré characters

We now show that all one-loop partition functions displayed above can be written as
exponentials of (sums of) the Poincaré characters of section 4.2. Recall in particular
that massive characters are given by eq. (4.60), where f is the rotation (11.8) and χλ
is the character of an irreducible representation of the little group SO(D − 1) with
highest weight λ.
In order to represent a massive relativistic particle with spin s, we choose the

weight λ to be λs = (s, 0, ..., 0) in terms of the dual basis of the Cartan subalgebra
of so(D − 1) described above (11.21). With this choice, expressions (4.60) and (4.65)
actually appear in the exponent of (11.24): taking α0 = iβ, we can rewrite the rotating
one-loop partition function for a massive field with spin s as the exponential of a sum
of Poincaré characters:
" +∞ #
X1
ZM,s [β, θ~ ] = exp ~ inβ] .
χM,s [nθ, (11.38)
n=1
n
The series in the exponent diverges for real θi ’s, which can be cured by adding suitable
imaginary parts to these angles as explained above. This result holds for any space-
time dimension D (along with the infrared regularization (4.64)). From a physical
perspective, it is the statement that a free field is a collection of harmonic oscillators,

one for each value of momentum: the index n then labels the oscillator modes, while
the integral over momenta is the one in the Frobenius formula (10.54). In particu-
lar, standard, non-rotating one-loop partition functions are exponentials of sums of
characters of (Euclidean) time translations. This relation has also been observed in
AdS [369, 370, 373]; our partition functions are flat limits of these earlier results, up to
the even-dimensional regularization subtlety mentioned below eq. (11.19). Note that
this issue already occurs at the level of characters: although most of (4.65) is a flat
limit of an SO(D − 1, 2) character, it is not clear how to regularize the divergences
that pop up when one of the angles vanishes so as to recover the regulators (4.64).
This problem also appears for odd D when one or more angles are set to zero.
For massless fields, the situation is a bit more complicated. For even D the massless
Poincaré character (4.69) is the limit M → 0 of its massive counterpart (4.65), and
the one-loop partition function (11.28) can again be written as an exponential (11.38).
But in odd space-time dimensions, SO(D − 2) has lower rank than SO(D − 1), so the
rotation (11.8) is not, in general, conjugate to an element of the massless little group:
it has one angle too much, and whenever all angles θ1 , ..., θr are non-zero, the character
(4.33) vanishes. The only non-trivial irreducible character arises when at least one of
the angles θ1 , ..., θr vanishes, say θr = 0, in which case the massless character takes the
form (4.71). However, comparison with (11.32) reveals a mismatch: the partition func-
tion does not take the form (11.38) in terms of the massless characters (4.71); in field
theory, all r angles θi may be switched on simultaneously! To accommodate for this
one can resort to the angle-dependent coefficients Ark (θ) ~ introduced in (11.31), whose
origin can again be understood through the massless limit of the character (4.60).
Using relation (11.33), the product of massless partition functions with spins ranging
from zero to s can be written as (11.38), where the characters on the right-hand side
are massless limits of massive Poincaré characters. However, it is unclear whether the
quantities appearing in the exponent of (11.32) can be related directly to Poincaré
characters without invoking a massless limit.
The occurrence of Poincaré characters in (11.38) illuminates certain aspects of

gravity and higher-spin theories in three dimensions. Indeed, recall expression (11.30)
for the partition function of a field with spin s in three-dimensional thermal, rotating
Minkowski space. (The regularization is unimportant for our present argument.) Since
the space-time dimension is odd, the terms of the series in the exponential are not
quite massless Poincaré characters, but they can still be interpreted as contributions
of specific field excitations. Indeed, the terms eisnθ are due to the heat kernel with spin
s, while the terms −ei(s−1)nθ come from ghosts, with a minus sign due to their fermionic
statistics. The difference eisnθ − ei(s−1)nθ vanishes when θ = 0, in accordance with the
fact that ghosts cancel all would-be local degrees of freedom. However, when θ 6= 0,
the cancellation is incomplete because would-be local field excitations and ghosts have
different spins (s and s − 1, respectively). As a result the one-loop partition function
is non-trivial despite the absence of physical local degrees of freedom.
11.2 Representations and characters of flat WN

As an application of the results of the previous section, we now explain how certain
combinations of one-loop partition functions in three-dimensional flat space reproduce
characters of higher-spin asymptotic symmetry algebras at null infinity. As it turns
out, the coadjoint representation of standard WN algebras [374–376] will play a key
role in the analysis, so we first review briefly the analogous situation of higher-spin
fields in AdS3 . We then turn to the case of spin 3 in flat space and describe certain
irreducible unitary representations of the corresponding asymptotic symmetry group.
Upon evaluating their characters thanks to the Frobenius formula, we find that they
match suitable products of partition functions. We also extend these observations to
arbitrary spin N . The description of the induced modules and quantum algebras that
correspond to this construction are relegated to section 11.3.
11.2.1 Higher spins in AdS3 and WN algebras

As a preparation for flat space computations, we review here the asymptotic symme-
tries of higher-spin theories in AdS3 . We also describe the corresponding quantum
symmetry algebras, their unitary representations and their characters, which match
field-theoretic one-loop partition functions.
Asymptotic symmetries
Asymptotic symmetries of higher-spin theories in three dimensions were first studied
in AdS3 [377–380], and are similar to the Brown-Henneaux asymptotic symmetries of
gravity described in chapter 8. Here we focus on models including fields with spin
ranging from 2 to N ; this setup can be described as an sl(N, R) ⊕ sl(N, R) Chern-
Simons action with a principally embedded sl(2, R)⊕sl(2, R) gravitational subalgebra.
When N = 3, the asymptotic symmetries are generated by gauge transformations
specified by four arbitrary, 2π-periodic functions X(x+ ), ξ(x+ ) and X̄(x− ), ξ(x ¯ −)

of the light-cone coordinates x± on the boundary of AdS3 . In particular, the functions

X(x+ ) and X̄(x− ) generate Brown-Henneaux conformal transformations of the type
described in section 8.2. Since the results are left-right symmetric, we focus on the
left-moving sector. The surface charge associated with a transformation (X, ξ) then
generalizes the (left-moving half of the) gravitational expression (8.42) and takes the
form [378] Z 2π
1
dϕ X(x+ )p(x+ ) + ξ(x+ )ρ(x+ )

Q(X,ξ) [p, ρ] = (11.39)
2π 0
when the normalization is chosen so that pure AdS3 with all higher-spin fields switched
off has vanishing higher-spin charges and negative mass −1/8G. Here ϕ = (x+ −x− )/2,
while p(x+ ) and ρ(x+ ) are two arbitrary, 2π-periodic functions specifying a solution
of
+ + −
the field equations. In particular p(x ) is one of the two functions p(x ), p̄(x ) that
determine an on-shell AdS3 metric (8.38) while ρ(x+ ), together with its anti-chiral
counterpart ρ̄(x− ), specifies an on-shell higher-spin field configuration. The vacuum
field configuration (8.47) corresponds to pure AdS3 with all higher-spin fields set to
zero, and is given by ρ = ρ̄ = 0, p = p̄ = −`/16G.
One can think of the pair (X, ξ) as an element of the asymptotic symmetry algebra,
so the charge (11.39) is the pairing between this algebra and its dual space. This
generalizes the pairing (6.34) of the Virasoro algebra with CFT stress tensors, and
(p, ρ) may be seen as a coadjoint vector of the symmetry algebra. Its infinitesimal
transformation law extends (8.39) and turns out to be [378]
c 000
δ(X,ξ) p = Xp0 + 2X 0 p − X + 2 ξρ0 + 3 ξ 0 ρ , (11.40a)
12
δ(X,ξ) ρ = Xρ0 + 3X 0 ρ + 2ξp000 + 9ξ 0 p00 + 15ξ 00 p0 + 10ξ 000 p
c 192
− ξ (5) − ξpp0 + ξ 0 p2 ,

(11.40b)
12 c
where prime denotes differentiation with respect to x+ , and c = 3`/2G is the Brown-
Henneaux central charge. Analogous formulas hold in the anti-chiral sector. Since X
generates conformal transformations, this implies that p is a (chiral) quasi-primary
field with weight 2 while ρ is a primary with weight 3. Together with the surface
charges (11.39), these transformation laws yield the Poisson bracket (8.10):

Q(X,ξ) [ p, ρ], Q(Y,ζ) [ p, ρ] = − δ(X,ξ) Q(Y,ζ) [p, π] . (11.41)
Formula (11.40) turns out to coincide with the coadjoint representation of a Poisson
algebra known as the W3 algebra, and indeed one finds that the bracket (11.41) re-
produces the non-linear bracket of a W3 algebra with central charge c (see eq. (11.43)
below). Similar considerations apply to models including fields with spin ranging from
2 to N [378, 380]; the resulting asymptotic symmetry algebra is the direct sum of two
copies of WN .
Quantum W3 algebra
representation of W3 , the orbit of the

Owing to the fact that (11.40) is the coadjoint
AdS3 vacuum p = p̄ = −c/24, ρ = ρ̄ = 0 under asymptotic symmetry transfor-
mations is a direct product of two vacuum coadjoint orbits of WN ; these orbits are
well-defined infinite-dimensional manifolds even though the definition of finite sym-
metry transformations is more intricate than in the pure Virasoro case corresponding
to N = 2 [376]. Putting all mathematical subtleties under the rug, one thus expects
the quantization of that orbit to produce the vacuum highest-weight representation of
the quantum algebra WN ⊕ WN . Accordingly, we now describe the quantum version
of the W3 algebra.
As in the purely gravitational case (8.46), the classical asymptotic symmetry alge-
bra given by the surface charges (11.39) can be written in terms of modes
Lm ≡ Q(eimx+ ,0) , Wm ≡ Q(0,eimx+ ) (11.42)
and their barred counterparts in the right-moving sector. The normalization is such
that pure AdS3 has all charges vanishing except L0 = L̄0 = −c/24. Using (11.40),
one finds that the Poisson brackets (11.41) of the charges (11.42) take the form of a
classical W3 algebra:
c 3
i{Lm , Ln } = (m − n)Lm+n + m δm+n,0 ,
12
i{Lm , Wn } = (2m − n)Wm+n , (11.43)
96 c
i{Wm , Wn } = (m − n)(2m2 + 2n2 − mn)Lm+n + Λm+n + m5 δm+n,0 ,
c 12
P
where Λm ≡ p∈Z Lm−p Lp is a non-linear term and the first line is the usual Vira-
soro algebra. The same brackets hold in the right-moving sector, so as announced the
asymptotic symmetry algebra is a direct sum of two classical W3 algebras. Under quan-
tization the Poisson brackets are turned into commutators according to i{·,[ ·} = [ˆ·,ˆ·]
and the charges Lm , Wn become operators Lm , Wn which, in any unitary representa-
tion, satisfy the Hermiticity conditions
L†m = L−m , Wm† = W−m .
It is also customary to normalize the Virasoro generators Lm so that the vacuum state
c
has vanishing eigenvalue under L0 , i.e. to rename Lm + 24 δm,0 into Lm . As a result the
commutators of the operators Lm , Wn yield the quantum W3 algebra
c
[Lm , Ln ] = (m − n)Lm+n + (m3 − m)δm+n, 0 , (11.44a)
12
[Lm , Wn ] = (2m − n)Wm+n , (11.44b)
96
[Wm , Wn ] = (m − n)(2m2 + 2n2 − mn − 8)Lm+n + (m − n) : Λm+n :
c + 22/5
c
+ (m2 − 4)(m3 − m)δm+n,0 , (11.44c)
12
whose non-linear terms are normal-ordered according to the prescription
X X 3
: Λm : ≡ Lm−p Lp + Lp Lm−p − (m + 3)(m + 2)Lm . (11.45)
p≥−1 p<−1
10
Here the term linear in Lm ensures that the operator : Λm : is quasi-primary with
respect to the action of Lm ’s. Note how the denominator of the structure constant of
the non-linear term in (11.44c) involves a shifted central charge c + 22/5 instead of
the classical c in the last line of (11.43). Analogous commutation relations hold in the
barred sector.
Unitary representations and characters

Unitary representations of the quantum W3 algebra are obtained analogously to the
Virasoro highest-weight representations of section 8.4, and are spanned by the descen-
dants of a highest-weight state annihilated by operators Lm , Wn with m, n > 0. At
large c, such representations are irreducible. The same is true of WN algebras for
any finite N , and irreducible unitary representations of WN ⊕ WN are tensor prod-
ucts of individual highest-weight representations of the two WN algebras (at large
c, c̄). Characters of such representations can be evaluated by adapting the counting
argument that led to (8.74). In particular the character of a highest-weight representa-

tion of WN ⊕ WN with central charges (c, c̄), generic highest weights (h, h̄), vanishing
higher-spin weights and vanishing higher-spin chemical potentials, is
+∞
!N −1
Y 1
Tr q L0 −c/24 q̄ L̄0 −c̄/24 = q h−c/24 q̄ h̄−c̄/24 . (11.46)
n=1
|1 − q n |2
This reduces to the product of Virasoro characters (10.62) when N = 2. The vacuum
character of WN ⊕ WN similarly reads
N +∞
!
Y Y 1
Trvac q L0 −c/24 q̄ L̄0 −c̄/24 = n |2
, (11.47)
s=2 n=s
|1 − q
where the truncated product arises because the vacuum state is left invariant by the
wedge algebra sl(N, R). This reduces to the vacuum character (8.79) when N = 2.
As mentioned earlier, it was shown in [241] that the one-loop partition function of
gravitons in AdS3 at temperature 1/β and angular potential θ is a vacuum character
(8.79) with modular parameter (8.78). This result was later extended to higher-spin
theories [69, 70], whose one-loop partition functions on thermal AdS3 coincide with
vacuum WN ⊕ WN characters (11.47) upon including the contribution of fields with
spins s = 2, 3, ..., N . These results confirm the interpretation of irreducible unitary
representations of asymptotic symmetry groups as particles dressed with boundary
degrees of freedom. The purpose of the remainder of this chapter is to describe the
similar matching that occurs in asymptotically flat theories.
11.2.2 Flat W3 algebra

The asymptotic symmetries of higher-spin theories at null infinity in three-dimensio-
nal flat space were described in [356, 381, 382]. Here we focus on the model describing
the gravitational coupling of a field of spin 3, which is a three-dimensional Chern-
Simons theory whose gauge algebra sl(3, R) A sl(3, R)Ab is the flat limit of sl(3, R) ⊕
sl(3, R). The associated asymptotic symmetry generators turn out to be labelled by
four arbitrary, 2π-periodic functions X(ϕ), ξ(ϕ), α(ϕ) and a(ϕ) on the celestial circle at
(future) null infinity. Of these, X(ϕ) and α(ϕ) generate standard BMS3 superrotations
and supertranslations (respectively), while ξ and a are their higher-spin extensions.
The corresponding surface charges extend the gravitational formula (9.31) and read
Z 2π h
1 i
Q(X,ξ,α,a) [j, κ, p, ρ] = dϕ X(ϕ)j(ϕ) + ξ(ϕ)κ(ϕ) + α(ϕ)p(ϕ) + a(ϕ)ρ(ϕ) ,
2π 0
(11.48)
where the 2π-periodic functions j, κ, p and ρ determine a solution of the equations of
motion. In particular p(ϕ) is the Bondi mass aspect (supermomentum) and j(ϕ) is
the angular momentum aspect (angular supermomentum) appearing in an asymptot-
ically flat metric (9.25). The functions ρ and κ are analogous quantities for a spin-3
field. As usual, the quadruple (j, κ, p, ρ) may be seen as an element of the dual space
of the asymptotic symmetry algebra. In particular, the higher-spin supermomentum
(p, ρ) transforms under higher-spin superrotations (X, ξ) as a coadjoint vector of the
W3 algebra, i.e. according to (11.40), albeit with a central charge c2 = 3/G instead of c.
The Poisson brackets satisfied by the surface charges (11.48) are given as usual by
(11.41) and are most easily expressed in terms of generators
Jm ≡ Q(eimϕ ,0,0,0) , Km ≡ Q(0,eimϕ ,0,0) , Pm ≡ Q(0,0,eimϕ ,0) , Qm ≡ Q(0,0,0,eimϕ ) .
Note that with this normalization pure Minkowski space has all its charges vanishing,
except P0 = −1/8G. One then finds that the Jm ’s and Pm ’s close according to a
bms3 algebra (9.37) with central charge c2 = 3/G, while brackets involving higher-spin
charges take the form
i{Jm , Kn } = (2m − n)Km+n , (11.49a)

i{Jm , Qn } = (2m − n)Qm+n , (11.49b)
i{Pm , Kn } = (2m − n)Qm+n , (11.49c)
i{Pm , Qn } = 0 , (11.49d)
96
i{Km , Kn } = (m − n)(2m2 + 2n2 − mn)Jm+n + (m − n)Ωm+n , (11.49e)
c2
96 c2
i{Km , Qn } = (m − n)(2m2 + 2n2 − mn)Pm+n + Θm+n + m5 δm+n,0 (11.49f)
c2 12
where the non-linear terms Ω and Θ are given by
X X
Ωm ≡ (Pm−p Jp + Jm−p Pp ) , Θm ≡ Pm−p Pp . (11.50)
p∈Z p∈Z
These formulas show that, up to central terms, the brackets of (J , K)’s with (P, Q)’s
take the same form as the brackets of (J , K)’s with themselves. This is the situation
described in (9.48) and it implies that, similarly to (9.64), the asymptotic symmetry
algebra is an exceptional semi-direct sum
“flat W3 algebra” ≡ W3 Aad (W3 )Ab , (11.51)
where W3 is the classical W3 algebra (11.43) and (W3 )Ab denotes an Abelian Lie
algebra isomorphic, as a vector space, to W3 . This algebra is centrally extended, as
the bracket between generators of W3 and those of (W3 )Ab includes a central charge
c2 ; there is also a central charge c1 specific to the left (non-Abelian) W3 subalgebra of
(11.51), but it is not switched on in parity-preserving theories. We shall return to this
structure in section 11.3, upon describing its quantum version.
Induced representations and unitarity

Since the flat W3 algebra (11.51) has the exceptional form g A gAb , with g the standard
W3 algebra, its unitary representations should be induced representations labelled by
orbits of supermomenta under the coadjoint action of elements of a group whose tan-
gent space at the identity is the W3 algebra. However, the non-linearities that appear
in W algebras make this step subtle, so one can bypass the need to control the group
as follows. Generic W algebras define a Poisson manifold through (11.41) and one
can classify the submanifolds on which the Poisson structure is invertible, i.e. their
symplectic leaves in the terminology of section 5.1. In the case of the Virasoro algebra
(which corresponds to WN with N = 2) this concept coincides with that of a coadjoint
orbit of the Virasoro group. For higher N the symplectic leaves of WN algebras are
still well defined [376] despite the lack of a straightforward definition of the group
that corresponds to the WN algebra. These leaves may be seen as intersections of the
coadjoint orbits of sl(N )-Kac Moody algebras with the constraints that implement
the Hamiltonian reduction to WN algebras. Accordingly, it should be possible to build
unitary representations of flat WN algebras as Hilbert spaces of wavefunctions defined
on their symplectic leaves, which we assume as usual to admit quasi-invariant mea-
sures. In most of the remainder of this chapter, we test that proposal by showing how
it can be used to evaluate characters that coincide with higher-spin one-loop parti-
tion functions. Note that the non-linearities appearing in the brackets of the algebra
(11.51) imply an extra complication for representation theory in that one has to devise
a suitable normal-ordering prescription. In the standard W3 case we displayed this
normal ordering in (11.45). In the flat case we will address this issue in section 11.3.
The complete classification of the symplectic leaves of the W3 algebra has been
worked out in [375, 376]; according to our proposal this settles the classification of
irreducible unitary representations of the flat W3 algebra, in the same way that Vira-
soro coadjoint orbits classify BMS3 particles. Instead of describing the details of this
classification, we focus from now on on orbits of constant supermomenta, which can be
classified thanks to the infinitesimal transformation laws (11.40) given by the algebra.
To describe such an orbit, let us pick a pair (p, ρ) where p(ϕ) = p0 and ρ(ϕ) = ρ0 are
constants, and act on it with an infinitesimal higher-spin superrotation (X, ξ). Then,
all terms involving derivatives of p or ρ in eq. (11.40) vanish, and we find
c2 000
δ(X,ξ) p0 = 2 X 0 p0 − X + 3 ξ 0 ρ0 , (11.52a)
12
c2 (5) 192 0 2
δ(X,ξ) ρ0 = 3 X 0 ρ0 + 10 ξ 000 p0 − ξ − ξ p0 . (11.52b)
12 c2
The little group for (p0 , ρ0 ) consists of higher-spin superrotations leaving it invariant.
Its Lie algebra is therefore spanned by pairs (X, ξ) such that the right-hand sides of
eqs. (11.52) vanish:
c2
2 X 0 p0 − X 000 + 3 ξ 0 ρ0 = 0 , (11.53a)
12
c2 (5) 192 0 2
3 X 0 ρ0 + 10 ξ 000 p0 − ξ − ξ p0 = 0 . (11.53b)
12 c2
The solutions of these equations depend on the values of p0 and ρ0 . Here we take
ρ0 = 0 for simplicity, i.e. we only consider cases where all higher-spin charges are
switched off. Then, given p0 , eqs. (11.53) become two decoupled differential equations
for the functions X(ϕ) and ξ(ϕ), leading to three different cases:
• For generic values of p0 , the only pairs (X, ξ) leaving (p0 , 0) invariant are con-
stants, and generate a little group U(1) × R.
• For p0 = −n2 c2 /96 where n is a positive odd integer, the pairs (X, ξ) leaving
(p0 , 0) invariant take the form
X(ϕ) = A, ξ(ϕ) = B + C cos(nϕ) + D sin(nϕ), (11.54)
where A, B, C and D are real numbers. The corresponding little group is the
n-fold cover of GL(2, R).
• For p0 = −n2 c2 /24 = −(2n)2 c2 /96 where n is a positive integer, the Lie algebra
of the little group is spanned by
X(ϕ) = A + B cos(nϕ) + C sin(nϕ),
(11.55)
ξ(ϕ) = D + E cos(nϕ) + F sin(nϕ) + G cos(2nϕ) + H sin(2nϕ),
where A, B, ..., H are real coefficients. The little group is thus an n-fold cover
of SL(3, R). In particular, p0 = −c2 /24 realizes the absolute minimum of energy
among all supermomenta belonging to orbits with energy bounded from below.
It is thus the supermomentum of the vacuum state, and indeed, upon using
c2 = 3/G, the field configuration that corresponds to it is the metric of Minkowski
space (with the spin-3 field set to zero on account of ρ0 = 0).
These results extend our earlier observations on Virasoro orbits in section 7.1.
11.2.3 Flat W3 characters

The information on little groups turns out to be sufficient to evaluate certain characters
along the lines of section 10.3. For instance, consider an induced representation based
on the orbit Op of a generic pair (p0 , 0), and call (s, σ) the spin of the representation
R of the little group U(1) × R. Then take a superrotation which is an element of the
U(1) subgroup (i.e. a rotation f (ϕ) = ϕ + θ). The only point on the orbit that is left
invariant by the rotation is (p0 , 0), and the whole integral over the orbit in (10.54)
localizes to that point. Therefore, in analogy with the BMS3 example, the detailed
knowledge of the orbit is irrelevant to compute the character. In particular, including
a higher-spin supertranslation α(ϕ), a(ϕ) , the only components of α(ϕ) and a(ϕ)
that survive the integration are their zero-modes α0 and a0 . The character thus takes
the form Z
isθ ip0 α0
χ[(rotθ , α, a)] = e e dµ(k) δ(k, rotθ · k) (11.56)
Op
isθ
where the little group character e factors out as in (4.55). In writing this we assume
the existence of a quasi-invariant measure µ on the orbit, whose precise expression is
unimportant since different measures give representations that are unitarily equivalent.
Our remaining task is to integrate the delta function. To do so, we use local coordinates
on the orbit, which we choose to be the Fourier modes of higher-spin supermomenta
as we did in section 10.3. Since p0 is generic, the non-redundant coordinates on the
orbit are the non-zero modes. As in (10.60) the integral is thus
Z Y Z Y Z
inθ imθ
dµ(k)δ(k, rotθ · k) = dkn δ(kn − e kn ) dρm δ(ρm − e ρm ) ,
Op n ∈ Z∗ m ∈ Z∗
(11.57)
where we call kn the Fourier modes of the standard (spin 2) supermomentum, while
ρm are the modes of its higher-spin counterpart. Performing the integrals over Fourier
modes and adding small imaginary parts i to θ as in (10.56), we obtain
+∞
!2
0
Y 1
χ[(rotθ , α, a)] = eisθ eip0 α . (11.58)
n=1
|1 − ein(θ+i) |2
This is a natural spin-3 extension of the spin-2 (BMS3 ) massive character (10.61), in
the same way that (11.46) generalizes Virasoro characters. It is also a flat limit of
(11.46) for N = 3.
A similar computation can be performed for orbits of other higher-spin supermo-

menta (p0 , 0). The only subtlety is that, for the values of p0 for which the little group
is larger than U(1) × R, the orbit has higher codimension in W3∗ than the generic orbit
just discussed. Accordingly, there are fewer coordinates on the orbit and the products
of integrals (11.57) are truncated. For instance, when p0 = −n2 c2 = /24 with n a
positive integer, the little group is generated by pairs (X, ξ) of the form (11.55), so
that the Fourier modes providing non-redundant local coordinates on the orbit (in a
neighbourhood of (p0 , 0)) are the modes km with m ∈ / {−n, 0, n} and the higher-spin
modes ρm with m ∈ / {−2n, −n, 0, n, 2n}. Assuming that the representation R of the
little group is trivial, this produces a character
+∞
! +∞
!
2 0
Y 1 Y 1
χ[(rotθ , α, a)] = e−in c2 α /24 · . (11.59)
m=1,
|1 − eim(θ+i) |2 m=1,
|1 − eim(θ+i) |2
m6=n m6=n,
m6=2n
The choice n = 1 specifies the vacuum representation of the flat W3 algebra; taking α
to be a Euclidean time translation by iβ, we get
+∞
! +∞
!
Y 1 Y 1
χvac [(rotθ , α = iβ, a = 0)] = eβc2 /24 in(θ+i) |2
· in(θ+i) |2
.
n=2
|1 − e n=3
|1 − e
(11.60)
This is one of our key results in this chapter. Indeed, comparing with eq. (11.37), we
recognize the product of the (suitably regularized) rotating one-loop partition func-
tions of massless fields with spins two and three in three-dimensional flat space. It
provides a first non-trivial check of our proposal for the construction of unitary repre-
sentations of flat WN algebras.
All the induced representations described above are unitary by construction, pro-
vided one can define (quasi-invariant) measures on the corresponding orbits. In anal-
ogy with representations of the bms3 algebra, they can also be described as induced
modules that generalize those of section 10.2; we will turn to them in section 11.3.
11.2.4 Flat WN algebras

The considerations of the previous pages can be generalized to higher-spin theories
in flat space with spins ranging from 2 to N . In AdS3 the asymptotic symmetries of
models with this field content consist of two copies of a WN algebra, so it is natural
to anticipate that the corresponding theory in flat space will have an asymptotic
symmetry algebra of the exceptional form
“flat WN algebra” ≡ WN Aad (WN )Ab , (11.61)
in analogy with (11.51). The surface charges generating these symmetries should co-
incide with the pairing of the Lie algebra of (11.61) with its dual space, and they
are likely to satisfy a centrally extended algebra. Since the presence of higher-spin
fields does not affect the value of the central charge in three-dimensional AdS grav-
ity [377, 378], one expects the central charge in this case to be the usual c2 = 3/G
appearing in mixed brackets. This structure was indeed observed for N = 4 in [356].
We now argue that this proposal must hold for any N by showing that the vacuum
character of (11.61), computed along the lines followed above for flat W3 , reproduces
the product of one-loop partition functions of fields of spin 2, 3, ..., N .
According to our proposal for the characterization of the representations of semi-

direct sums of the type (11.61), unitary representation of flat WN algebras are classified
by their symplectic leaves, that is, by orbits of higher-spin supermomenta (p1 , ..., pN −1 ).
(Here p1 (ϕ) is the supermomentum that we used to write as p(ϕ), while p2 (ϕ) is
what we called ρ(ϕ) for N = 3.) The infinitesimal transformations that generalize
(11.40) and that define these orbits locally can be found for instance in [380]. Here
we focus on the vacuum orbit where we set all higher-spin charges to zero and take
only p1 = −c2 /24 to be non-vanishing. This particular supermomentum is left fixed
by higher-spin superrotations of the form
i
X
Xi (ϕ) = Ai + Bij cos(jϕ) + Cij sin(jϕ) , i = 1, ..., N − 1, (11.62)
j=1
where the coefficients Ai , Bij , Cij are real. In principle one can obtain such symmetry
generators by looking for the little group of the vacuum as in (11.53), using for instance
the explicit formulas of [380]. Yet, a simpler way to derive the same result is to look
for the higher-spin isometries of the vacuum in the first-order formulation, in which
the theory is described by a Chern-Simons action with gauge algebra sl(N, R) Aad
(sl(N, R))Ab (see e.g. [356,382,383]). In this language, and in terms of retarded Bondi
coordinates (r, ϕ, u), the vacuum field configuration takes the form
hr i
−1 −1
Aµ (x) = b(r) g(u, ϕ) ∂µ [g(u, ϕ)b(r)] , b(r) = exp P−1 , (11.63)
2
where g(u, ϕ) is a field valued in SL(N, R) n sl(N, R), given by

1 1
g(u, ϕ) = exp P1 + P−1 u + J1 + J−1 ϕ (11.64)
4 4
in terms of Poincaré generators that satisfy the commutation relations (10.26). The
isometries of this field configuration are generated by gauge parameters of the form
(g · b)−1 Ta (g · b), where Ta is any of the basis elements of the gauge algebra. Upon
expanding g −1 Ta g as a position-dependent linear combination of gauge algebra gener-
ators, the function multiplying the lowest weight generator coincides with the corre-
sponding asymptotic symmetry parameter (see e.g. [378] for details). The latter can
be obtained as follows.
For convenience, we diagonalize the Lorentz piece of the group element (11.64) as

1
exp J1 + J−1 ϕ = SeiJ0 ϕ S −1 (11.65)
4
where S is some SL(2, R) matrix. Then the gauge parameters that generate the little
group of the vacuum configuration can be written as
`
X
1 m (`) 1
exp − J1 + J−1 ϕ α Wm exp J1 + J−1 ϕ (11.66a)
4 m=−`
4
`
X
−iJ0 ϕ
=Se αm S −1 Wm(`) S eiJ0 ϕ S −1 , (11.66b)
m=−`
(`)
where the αm ’s are certain real coefficients, while the Wm ’s (with 2 ≤ ` ≤ N and
(2)
−` ≤ m ≤ `) generate the sl(N, R) algebra (including Jm ≡ Wm ). Note that the
matrix S preserves the conformal weight since it is an exponential of sl(2, R) generators,
so that
X ` X `
m (`) −1
α S Wm S = α̃m Wm(`) (11.67)
m=−` m=−`
for some coefficients α̃j obtained by acting on the αm ’s with an invertible linear map.
(`)
Since each generator Wm has weight m under J0 , eq. (11.66b) can be rewritten as
`
X `
X `
X
imϕ m −1 (`)
e α̃ S Wm(`) S = β mn
Wn(`) eijϕ = eimϕ β m` W` + · · · (11.68)
m=−` m,n=−` m=−`
(`)
for some coefficients β mn . In the last step we omitted all terms proportional to Wm ’s
with m < `; the important piece is the term that multiplies the highest-weight gen-
(`)
erator W` : it is the function on the circle that generates the asymptotic symmetry
(`)
corresponding to the generator `m=−` αm Wm that we started with in (11.66a). Since
P
the β m` ’s are related to the αm ’s by an invertible linear map, and since there are 2` + 1
linearly independent generators of this type, the isometries of the vacuum exactly span
the set of functions of the form (11.62). This is what we wanted to prove; there are
N 2 − 1 linearly independent asymptotic symmetry generators of this form, and they
span the Lie algebra sl(N, R).
The character associated with the vacuum representation of (11.61) can then be
worked out exactly as in the cases N = 2 and N = 3 discussed above: using the
Fourier modes of the N − 1 components of supermomentum as coordinates on the
orbit, we need to mod out the redundant modes. For the vacuum orbit, these are the
modes ranging from −(s − 1) to (s − 1) for the sth component. The integral over the
localizing delta function in the Frobenius formula (4.33) then produces a character
N +∞
!
Y Y 1
χ[(rotθ , a1 = iβ)] = eβc2 /24 (11.69)
s=2 n=s
|1 − ein(θ+i)2 |
where we implicitly set to zero all higher-spin supertranslations except the gravita-
tional one, a1 = iβ. Comparing with (11.37), we recognize the product of one-loop
partition functions of massless higher-spin fields with spins ranging from 2 to N , in-
cluding a classical contribution. This result confirms, on the one hand, our conjecture
(11.61) for the asymptotic symmetry algebras of generic higher-spin theories in three-
dimensional flat space, and on the other hand it provides another consistency check of
our proposal for the characterization of unitary representations of flat WN algebras.
It is also a flat limit of the vacuum WN ⊕ WN character displayed in (11.47).
11.3 Flat W3 modules

We now turn to the algebraic analogue of the above considerations, i.e. we describe
induced modules of flat WN algebras along the lines of section 10.2. For simplicity
we focus on the case N = 3 but the construction also applies to other higher-spin
extensions of bms3 . Due to the non-linearities of W3 , our plan in this section is slightly
different from that of section 10.2. Namely, we start by describing the quantum flat
W3 algebra as an ultrarelativistic limit of the direct sum of two quantum W3 algebras,
which produces a specific ordering of operators in the non-linear terms of commutators.
We then move on to the description of induced modules of the ultrarelativistic quantum
flat W3 algebra, and show that the ultrarelativistic normal ordering is defined with
respect to a rest frame vacuum. Along the way we compare our results to those of
the non-relativistic limit described in [356], and point out that the two limits lead to
different quantum algebras.
11.3.1 Ultrarelativistic and non-relativistic limits of W3

The flat W3 algebra (11.51) can be obtained as an Inönü-Wigner contraction of the
direct sum of two W3 algebras. This flat limit was discussed at the semiclassical
level in [381, 382], and a Galilean limit of the quantum algebra was described in [356].
Here we are interested instead in an ultrarelativistic limit of W3 ⊕ W3 . The key
difference between the Galilean and ultrarelativistic contractions is that the latter
mixes generators with positive and negative mode numbers, while the former does not.
For linear algebras such as Virasoro, this makes no difference and the two contractions
yield identical quantum algebras, namely bms3 ∼ = gca2 . When non-linear terms are
involved in the contraction, however, Galilean and ultrarelativistic limits generally give
different quantum algebras, as we now explain.
Ultrarelativistic contraction
The quantum W3 algebra is spanned by two sets of generators Lm and Wm (m ∈ Z)
whose commutation relations were displayed in (11.44). Consider now a direct sum
W3 ⊕ W3 , where the generators and the central charge of the other copy of W3 will
be denoted with a bar on top (L̄m , W̄m and c̄). Introducing a length scale ` to be
interpreted as the AdS3 radius, we define new generators Pm and Jm as in (10.32), as
well as
1
Km ≡ Wm − W̄−m , Qm ≡ Wm + W̄−m . (11.70)
`
We also define central charges c1 and c2 as in (9.93). In the limit ` → ∞, and provided
the central charges scale in such a way that both c1 and c2 are finite, one finds that
Jm and Pm satisfy the bms3 brackets (10.41) together with
[Jm , Kn ] = (2m − n)Km+n , [Jm , Qn ] = (2m − n)Qm+n , (11.71a)

[Pm , Kn ] = (2m − n)Qm+n , [Pm , Qn ] = 0 . (11.71b)
The remaining brackets involving higher-spin generators are

96
[Km , Kn ] = (m − n)(2m2 + 2n2 − mn − 8)Jm+n + (m − n)Ωm+n (11.71c)
c2
96 c1 c1
− 2
(m − n)Θm+n + (m2 − 4)(m3 − m)δm+n, 0 , (11.71d)
c2 12
96
[Km , Qn ] = (m − n)(2m2 + 2n2 − mn − 8)Pm+n + (m − n)Θm+n
c2
c2 2 3
+ (m − 4)(m − m)δm+n, 0 , (11.71e)
12
[Qm , Qn ] = 0 , (11.71f)
where the non-linear terms Ωm and Θm are quadratic operators given by (11.50), with
the exact same ordering (and calligraphic letters replaced by usual capital letters):
X X
Ωm ≡ (Pm−p Jp + Jm−p Pp ) , Θm ≡ Pm−p Pp . (11.72)
p∈Z p∈Z
The commutation relations (11.71) are quantum analogues of the Poisson brackets
(11.49), including a central charge c1 and with operators normalized so that the vac-
uum has zero eigenvalue under P0 . One can check that with the definition (11.50),
the brackets given by (10.41) and (11.71) satisfy Jacobi identities, so the generators
Jm , Km , Pn , Qn span a well-defined non-linear Lie algebra. We call it the quantum
flat W3 algebra. In any unitary representation, its generators satisfy the Hermiticity
conditions
(Qm )† = Q−m , (Km )† = K−m . (11.73)
supplemented with (10.27) for m ∈ Z.
The expressions (11.72) for the quadratic terms follow from the identities
`2 `
: Λm : + : Λ̄m : = Θm + O(`) , : Λm : − : Λ̄m : = Ωm + O(1) (11.74)
2 2
where : Λm : is the normal-ordered quadratic term (11.45) of the quantum W3 algebra,
while : Λ̄m : is its right-moving counterpart. Note, in particular, that both the linear
term in (11.45) and the mixing between positive and negative modes in (10.32)-(11.70)
are necessary to reorganize the sum of quadratic terms with the precise order of (11.72).
We shall see in section 11.3.2 that (11.50) is a normal-ordered polynomial with respect
to the natural vacuum in induced modules of the quantum flat W3 algebra.
Galilean contraction
In order to compare the quantum flat W3 algebra (11.71) with other results in the
literature [356], we now consider the non-relativistic limit of the quantum direct sum
W3 ⊕ W3 . It is obtained by defining central charges c̃1 and c̃2 as in (10.53), introducing

new generators J˜m and P̃m as in (10.52) and writing
1
K̃m ≡ W̄m + Wm , Q̃m ≡ W̄m − Wm . (11.75)
`
Note the difference with respect to (11.70). In the limit ` → +∞ one obtains brackets
of the same form as in (11.71) upon putting tildes on top of all generators, but there
are two important differences: (i) the coefficient in front of Θm+n in (11.71d) contains
a shifted central charge c̃1 + 44/5, and (ii) the quadratic term Ω̃m reads
X X 3
Ω̃m = P̃m−p J˜p + J˜m−p P̃p + P̃p J˜m−p + J˜p P̃m−p − (m + 3)(m + 2)P̃m
p≥−1 p<−1
5
(11.76)
instead of (11.72). The non-linear term Θ̃m remains the same (up to tildes) since the
generators P̃m commute in the large ` limit. The quadratic combinations Θ̃m and
Ω̃m can then be interpreted as normal-ordered operators with respect to a Galilean
highest-weight vacuum defined by conditions of the type (10.49) with M̃ = s̃ = 0.
These differences show that the two contractions lead to different quantum algebras,
despite the fact that the corresponding classical algebras coincide.3 Thus, in the pres-
ence of higher-spin fields, the difference between ultrarelativistic and Galilean limits
manifests itself directly in the symmetry algebras and not only at the level of the
representations that survive in the limit.
In the following we restrict attention to irreducible unitary representations of the

ultrarelativistic quantum algebra (11.71), built once again according to the induced
module prescription of section 10.2. On the other hand, highest-weight representations
of Galilean contractions of two copies of non-linear W algebras were discussed in [356],
where it was shown that unitary representations with higher-spin states do not exist.
11.3.2 Induced modules for the flat W3 algebra

According to our proposal of section 11.2.2, the Hilbert space of any unitary represen-
tation of the flat W3 algebra consists of wavefunctions on the orbit of a higher-spin
supermomentum p(ϕ), ρ(ϕ) . Assuming that the orbit admits a quasi-invariant mea-
sure, a basis of the Hilbert space is provided by plane waves (10.13) with definite
supermomentum. For definiteness, let us focus on an orbit containing a constant
higher-spin supermomentum (p0 , ρ0 ); this is to say that the representation admits a
rest frame. There is a corresponding plane wave Ψ(p0 ,ρ0 ) , and any other plane wave
can be obtained by acting on Ψ(p0 ,ρ0 ) with a higher-spin superrotation.
Massive modules
Let us take p0 = M − c2 /24 with M > 0; this corresponds to a massive representation
of the flat W3 algebra. Assuming also that ρ0 is generic, the little group is U(1) × R
3
An interesting problem is to understand if these algebras are merely different because of an
unfortunate choice of basis, or if they are genuinely distinct in the sense that they are not isomorphic.
We will not address this issue here.
and the spin of the representation is therefore a pair (s, σ) ∈ R2 . Now, the plane wave
at rest Ψ(p0 ,ρ0 ) ≡ |M, ρ0 i is a state that satisfies
Pm |M, ρ0 i = 0 , Qm |M, ρ0 i = 0 for m 6= 0 , (11.77a)
and is an eigenstate of zero-mode charges:
P0 |M, ρ0 i = M |M, ρ0 i , J0 |M, ρ0 i = s|M, ρ0 i , (11.77b)

Q0 |M, ρ0 i = ρ0 |M, ρ0 i , K0 |M, ρ0 i = σ|M, ρ0 i . (11.77c)
Here M and s are the mass and spin labels encountered in (10.42), while ρ0 and σ
are their spin-3 counterparts. As before we call |M, ρ0 i the rest frame state of the
representation and we normalize the operator P0 so that the vacuum has vanishing P0
eigenvalue.
The conditions (11.77) define a one-dimensional representation of the subalgebra

spanned by {Pm , Qm , J0 , W0 }. They can be used to define an induced module H with
basis elements
Kk1 ...Kkm Jl1 ...Jln |M, ρ0 i , (11.78)
where k1 ≤ ... ≤ km and l1 ≤ ... ≤ ln are non-zero integers. This provides an
explicit representation of the quantum flat W3 algebra. Note that the presence of non-
linearities in the commutators (11.71) does not affect the construction of the induced
module, which involves the universal enveloping algebra anyway.
As usual, unitarity is somewhat hidden in the induced module picture but can be
recognized in the fact that the state |M, ρ0 i is a plane wave, and that acting on it
with finite higher-spin superrotations generates an orthonormal basis of plane wave
states for the carrier space of the representation. Irreducibility can be inferred from
the same argument that we used for bms3 : by construction, a supermomentum orbit
is a homogeneous space for the action of superrotations, and this carries over to the
higher-spin setting. This implies that W3 superrotations can map any plane wave
state on any other one, which in turn implies that the space of the representation has
no non-trivial invariant subspace.
Vacuum module
The vacuum module of the flat W3 algebra can be built in direct analogy to its bms3
counterpart discussed around (10.44). The only subtlety is the enhancement of the
little group, which leads to additional conditions on superrotations. Indeed the vacuum
state |0i is now an eigenstate of all modes Pm and Qm with zero eigenvalue, and satisfies
in addition
Jn |0i = 0 for n = −1, 0, 1, Km |0i = 0 for m = −2, −1, 0, 1, 2. (11.79)
These conditions ensure that the vacuum is invariant under the sl(3, R) wedge algebra
of the W3 subalgebra (which includes in particular the Lorentz algebra). The corre-
sponding module can then be built as usual by acting with higher-spin superrotation
generators on the vacuum state and producing states of the form (11.78), where now
all li ’s must be different from −1, 0, 1 and all ki ’s must be different from −2, −1, 0, 1, 2.
We stress that the li ’s and ki ’s can be positive or negative, in sharp contrast to the
non-relativistic modules investigated in [356].
The definition of the flat W3 vacuum allows us to interpret the quadratic terms
(11.72) as being normal-ordered. Indeed, their expectation values vanish in the vacuum
|0i:
h0|Θn |0i = h0|Ωn |0i = 0. (11.80)
These considerations appear to be a robust feature of “flat W algebras”: ultrarelativis-
tic contractions of WN ⊕ WN algebras always take the form
“flat WN ” = WN Aad (WN )Ab (11.81)
and therefore contain an Abelian ideal, where the semi-direct sum ensures that the
structure constants of the non-linear terms are always proportional to inverse powers
of the central charge. Indeed, for a non-linear operator of nth order the structure
1
constants are of order cn−1 at large c. When expanding them in powers of the con-
traction parameter `, this implies that the leading term is proportional to `1−n thanks
to (9.93). In order to obtain a finite expression, it is thus necessary that the resulting
non-linear operator consists of at least n − 1 Abelian generators. Terms of this kind
always have a vanishing expectation value in the rest frame vacuum state, although
the precise ordering in the polynomial should be fixed by other means, e.g. by defining
the algebra via a contraction of the quantum algebra or by imposing Jacobi identities.
Thus the conditions (11.77) with M = ρ0 = s = σ = 0, together with (11.79), provide
a valid definition of the vacuum for all quantum flat WN algebras.
By contrast, for a highest-weight vacuum of the type (10.49), the quadratic oper-
ators Ωm given by (11.72) generally have non-vanishing vacuum expectation values.
Thus the extra non-linear structure introduced by higher spins exhibits the fact that
the natural representations in the ultrarelativistic limit are the induced ones discussed
above, rather than the highest-weight ones of [287,356]. This difference emphasizes the
physical distinction between ultrarelativistic and Galilean limits: the former is adapted
to gravity, and more generally to models of fundamental interactions, where unitarity
is a key requirement. In particular, flat space holography (at least in the framework
of Einstein gravity) is expected to rely on the unitary construction described in this
thesis. By contrast, the Galilean viewpoint is suited to condensed matter applications,
and more generally to situations where unitarity need not hold — as was indeed ar-
gued in [287]. We stress that this difference is a genuine quantum higher-spin effect:
it is not apparent at the classical level, and it does not occur in pure gravity either.
11.4 Super-BMS3 and flat supergravity

This section is devoted to supersymmetric extensions of the BMS3 group, to their
representations, and to their characters. Accordingly we start by describing rotating
one-loop partition functions of fermionic fields in flat space, along the same lines as
in section 11.1. Upon confirming that they take the form of exponentials of Poincaré
characters (11.38), we specialize to D = 3 space-time dimensions. There we describe
supersymmetric BMS3 groups and their unitary representations, and note that super
BMS3 multiplets contain towers of infinitely many particles with increasing spins.
Finally, we show that the resulting characters match suitable combinations of bosonic
and fermionic one-loop partition functions.
11.4.1 Fermionic higher spin partition functions

We wish to evaluate the partition function 11.1 of a free fermionic field ψ with spin
s + 1/2 (where s is a non-negative integer) and mass M > 0. Its Euclidean action
can be presented either (i) using a symmetric, γ-traceless spinor field with s space-
time indices and a set of auxiliary fields with no gauge symmetry [384] or (ii) using
a set of symmetric spinor fields with s, s − 1, ..., 0 space-time indices and vanishing
triple γ-trace, subject to a gauge symmetry generated by γ-traceless parameters with
s − 1, ..., 0 space-time indices [385]. In the latter case, just as for bosons, the action is
given by a sum of actions for massless fields of each of the involved spins, plus a set
of cross-coupling terms proportional to the mass. In the limit M → 0 the quadratic
couplings vanish and one is left with a sum of decoupled Fang-Fronsdal actions [386]
Z
D µ1 ...µs 1 1 λ
S[ψ, ψ̄] = d x ψ̄ Sµ1 ...µs − γ(µ1 6 S µ2 ···µs ) − δ(µ1 µ2 Sµ3 ···µs )λ + h.c. ,
2 2
(11.82)
where space-time indices are raised and lowered thanks to the Euclidean metric (and
“h.c.” means “Hermitian conjugate”). We use the same symmetrization conventions as
in section 11.1 and
Sµ1 ...µs = 6 ∂ ψµ1 ...µs − ∂(µ1ψ 6 µ2 ...µs ) . (11.83)
The slash notation means 6 V ≡ γ µ Vµ , where the γ µ ’s are Dirac matrices satisfying the
anticommutation relations {γ µ , γ ν } = δ µν .
To compute the partition function for ψ, ψ̄ one has to evaluate a path integral
(11.1) with the integration measure DψDψ̄ and S the action (11.82) or its massive
analogue. The fermionic fields live on RD /Z as defined by the group action (11.7), but
in contrast to bosons, they satisfy anti periodic boundary conditions along the thermal
cycle. For a massive field, one thus finds that the partition function is given by
1 1
log Z = log det(−∆(s+1/2) + M 2 ) − log det(−∆(s−1/2) + M 2 ) , (11.84)
2 2
where ∆(s+1/2) is the Laplacian acting on antiperiodic, symmetric, γ-traceless spinor
fields with s indices on RD /Z. For massless fields, the gauge symmetry enhancement
requires gauge-fixing and ghosts, leading to [387]
1 1
log Z = log det(−∆(s+1/2) ) − log det(−∆(s−1/2) ) + log det(−∆(s−3/2) ) . (11.85)
2 2
Eqs. (11.84) and (11.85) are fermionic analogues of the bosonic formulas (11.11) and
(11.25). To evaluate the functional determinants, we rely once more on heat kernels
and the method of images described in section 11.1.1.
The heat kernel KAB µs ,νs associated with the operator (−∆(s+1/2) + M 2 ) on RD is
the unique solution of
(∆(s+1/2) − M 2 − ∂t )KAB µs ,νs = 0 , KAB µs , νs (t = 0, x, x0 ) = I(F )

µs , νs 1 δ (x − x0 ) .
AB (D)
(11.86)
AB
Here K µs ,νs is a bispinor in the indices A and B, and a symmetric bitensor in the
indices µs and νs . (We use again the shorthand µs to denote a set of s symmetrized
indices.) It is also γ-traceless in the sense that
γ µ Kµs , νs = Kµs , νs γ ν = 0 . (11.87)
The solution of (11.86) satisfying this requirement is

1 2 1 0 2
Kµs , νs (t, x, x0 ) = D/2
e−M t− 4t |x−x | I(F )
µs , νs , (11.88)
(4πt)
(F )
where Iµs ,νs is the following bisymmetric, γ-traceless tensor:
s
b2c s s−2k−1 s
!
X (−1)k 2k k! [D + 2(s − k − 1)]!! δ µµ δµν δνν γµ γν
I(F )
µs , νs =
k s−2k s
δµµ δµν δνν − .
k=0
s! [D + 2(k − 1)]!! D + 2(s − k − 1)
(11.89)
(F )
Up to the replacement of I by I , the fermionic heat kernel (11.88) is the same as the
bosonic one in eq. (11.14). In particular, I(F ) carries all its tensor and spinor indices.
To evaluate the determinant of (−∆(s+1/2) + M 2 ) on RD /Z, we use once more the

method of images (11.6). As before, we need to keep track of the non-trivial index
structure of KAB µs ,νs , which leads to
D
X
KµRs , /Z 0
(−1)n (J n )α β ...(J n )α β U n Kµs , βs t, x, γ n (x0 ) ,

αs (t, x, x ) = (11.90)
n∈Z
where the factor (−1)n comes from antiperiodic boundary conditions, J is the matrix
(11.8), and U is a 2bD/2c × 2bD/2c matrix acting on spinor indices in such a way that
J α β γ β = U γ α U −1 . (11.91)
In other words, U is the matrix corresponding to the transformation (11.8) in the

spinor representation of SO(D), and it can be written as
 
b(D−1)/2c
1 X
U = exp  θj [γ2j−1 , γ2j ] .
4 j=1
In particular, a rotation by 2π around any given axis maps the field ψ on −ψ, in
accordance with the fact that spinors represent SO(D) up to a sign. Note that, using
an explicit D-dimensional representation of the γ matrices, one gets
r
Y
n bD/2c
Tr(U ) = 2 cos(nθi /2) . (11.92)
i=1
Now, plugging (11.90) into formula (11.3) for the determinant of −∆(s+1/2) , one obtains
a sum of integrals which can be evaluated exactly as in the bosonic case. The only
difference with respect to bosons comes from the spin structure, and the end result is
(
X (−1)n χ
(F ) ~
s [n θ,~
] e−|n|βM D odd
−log det(−∆(s+1/2) +M 2 ) = r ×
|n| Q M L
K1 (|n|βM ) D even
n ∈ Z∗ |1 − ein(θj +ij ) |2 π
j=1
(11.93)
where we have discarded a volume divergence independent of all chemical potentials
(as in eq. (11.20)), and where
) ~
χ(F ] = (J µα )s Tr I(F
)
s [nθ,~ µs ,αs (11.94)
is the fermionic analogue of (11.19), with the same rough regularization as in eq.
(11.20) (a more careful regularization will be described below for D = 3). This result
(F )
takes the same form as (11.20), up to the replacement of χs by χs and the occurrence
of (−1)n due to antiperiodicity. In appendices 11.B.1 and 11.B.2, we show that

χ(D) ~
(F ) [nθ ] for odd D,
(F ) ~ 11.B.1&11.B.2 λs
χs [nθ] = (11.95)
χ(D) ~
(F ) [nθ, 0] for even D,
λs
where the term on the right-hand side is the character of an irreducible representation
(F )
of SO(D) with highest weight λs = (s + 1/2, 1/2, ..., 1/2), written here in the dual
basis of the Cartan subalgebra of so(D) described above (11.21).
Having computed the required functional determinants on RD /Z, we can now write
down the partition functions given by (11.84) and (11.85). In the massive case, the
difference of Laplacians acting on fields with spins (s + 1/2) and (s − 1/2) produces
the difference of two factors (11.95), with labels s and s − 1. It turns out that formula
(F )
(11.23) still holds if we replace λs and λs−1 by their fermionic counterparts, λs and
(F )
λs−1 . (The proof of this statement follows the exact same steps as in the bosonic case
described in appendix 11.A.3, up to obvious replacements that account for the change
in the highest weight vector.) Accordingly, the rotating one-loop partition function of
a massive field with spin s + 1/2 is
 
∞ (D−1) ~ ] (
X (−1)n+1 χλ(F ) [nθ,~ e−nβM (D odd) 
Z(β, θ~ ) = exp  r
s
× ML
.
 n Q in(θ +i ) 2 K 1 (nβM ) (D even) 
n=1 |1 − e j j | π
j=1
(11.96)
In the massless case we must take into account one more difference of characters,
(F )
namely (11.27) with λs replaced by λs . For D ≥ 4, this difference can be written as
a combination of SO(D−2) characters (the proof is essentially the same as in appendix
11.A.3), and the partition function of a massless field with spin s + 1/2 exactly takes
the form (11.28) or (11.32) (for D even or odd, respectively) with an additional factor
(F )
of (−1)n+1 in the sum over n, and the replacement of λs by λs . One can also verify
that relation (11.34) remains true for fermionic partition functions.
For D = 3, differences of SO(2) characters cannot be reduced any further (recall

the discussion surrounding (11.29)), so the best one can do is to write the partition
function of a massless field with spin s + 1/2 as
" +∞ #
X (−1)n+1 1 i(s+1/2)n(θ+i) i(s−1/2)n(θ+i)

Z(β, θ) = exp e −e + c.c.
n=1
n |1 − ein(θ+i) |2
(11.97)
in(θ+i)/2
provided s ≥ 1. (For s = 0 the exponentials in the summand reduce to e +
c.c., without any negative contribution.) Here we are using once more the crude
regularization described around (11.20); a more careful prescription, motivated by the
bosonic combination (11.35), consists in regulating the sum of exponentials in the
summand according to
ei(s+1/2)n(θ+i) − ei(s−1/2)nθ−(s+3/2)n + c.c. (11.98)
Upon using this expression in the summand of (11.97) instead of the naive combination
of exponentials written there, the series in the exponential becomes
+∞ +∞
(−1)n+1 q n(s+1/2) − q n(s+1/2) q̄ n + c.c. (−1)n+1 q n(s+1/2)
X X
= + c.c.
n=1
n |1 − q n |2 n=1
n 1 − qn
+∞
X
= log(1 + q j+1/2 ) + c.c.
j=s
in terms of q = ei(θ+i) . As in the bosonic case (11.36), the regularization (11.98) has
ensured that log Z splits as the sum of a chiral and an anti-chiral function of q. After
renaming j into n, the end result is the following expression for the partition function
of a field with spin s + 1/2 in three dimensions:
+∞
2
Y
Z= 1 + ei(n+1/2)(θ+i) , (11.99)
n=s
which can also be recovered as the flat limit of the corresponding AdS result [387]. The
remainder of this chapter is devoted to relating this partition function to the vacuum
characters of various supersymmetric extensions of the BMS3 group.
11.4.2 Supersymmetric BMS3 groups

The supersymmetric BMS3 groups describe the symmetries of three-dimensional, a-
symptotically flat supergravity [167, 280, 388–390]. Here we briefly review some back-
ground on super Lie groups and the super Virasoro algebra, which we then use to
provide a definition of various supersymmetric extensions of BMS3 . The correspond-
ing unitary representations and characters will be investigated in section 11.4.3.
Supersymmetric induced representations

A super Lie group is a pair (Γ0 , γ) where Γ0 is a Lie group in the standard sense,
while γ is a super Lie algebra whose even part coincides with the Lie algebra of Γ0 ,
and whose odd part is a Γ0 -module such that the differential of the Γ0 action be the
bracket between even and odd elements of γ [391]. Then a super semi-direct product
is a super Lie group of the form [392, 393]

G nσ A, g A (A + A) , (11.100)
where GnA is a standard (bosonic) semi-direct product group with Lie algebra g A A,
while g A (A + A) is a super Lie algebra whose odd subalgebra A is a G-module such
that the bracket between elements of g and elements of A be the differential of the
action of G on A, and such that [A, A] = 0 and {A, A} ⊆ A. By virtue of this
definition, the action of G on A is compatible with the super Lie bracket:
{g · S, g · T } = σg {S, T } ∀ S, T ∈ A , (11.101)
where σ is the action of G on A.
It was shown in [392, 393] that all irreducible, unitary representations of a super
semi-direct product are induced in essentially the same sense as for standard, bosonic
groups. In particular, they are classified by the orbits and little groups of G nσ A, as
explained in section 4.1. However, there are two important differences with respect to
the purely bosonic case:
1. Unitarity rules out all orbits on which energy can be negative, so that the mo-
mentum orbits giving rise to unitary representations of the supergroup form a
subset of the full menu of orbits available in the purely bosonic case. More
precisely, given a momentum p ∈ A∗ , it must be such that
hp, {S, S}i ≥ 0 ∀S ∈ A . (11.102)
When this condition is not satisfied, the representations of (11.100) associated

with the orbit Op are not unitary. The momenta satisfying condition (11.102)
are said to be admissible. Note that admissibility is a G-invariant statement: if
f ∈ G and if p is admissible, then so is f · p, by virtue of (11.101). For instance,
the only admissible momenta for the super Poincaré group are those of massive
or massless particles with positive energy (and the trivial momentum p = 0).
2. Given an admissible momentum p, the odd piece A of the supersymmetric trans-

lation algebra produces a (generally degenerate) Clifford algebra
Cp = T (A)/ S 2 − hp, {S, S}i | S ∈ A ,

(11.103)
where T (A) is the tensor algebra of A. Quotienting this algebra by its ideal
generated by the radical of A, one obtains a non-degenerate Clifford algebra C¯p .
Since A is a G-module, there exists an action of the little group Gp on C¯p ; let us
denote this action by a 7→ g · a for a ∈ C¯p and g ∈ Gp . To obtain a representation
of the full supergroup (11.100), one must find an irreducible representation τ of
C¯p and a representation R0 of Gp acting in the same space, and compatible with
τ in the sense that
τ [g · a] = R0 [g] · τ [a] · (R0 [g])−1 . (11.104)
For finite-dimensional groups, the pair (τ, R0 ) turns out to be unique up to mul-
tiplication of R0 by a character of Gp (and possibly up to parity-reversal). Given
such a pair, we call it the fundamental representation of the supersymmetric little
group.
The Clifford algebra (11.103) leads to a replacement of the irreducible, “spin” rep-
resentations of the little group, by generally reducible representations R0 ⊗ R. This
is the multiplet structure of supersymmetry: the restriction of an irreducible unitary
representation of a supergroup to its bosonic subgroup is generally reducible, and the
various irreducible components account for the combination of spins that gives rise
to a susy multiplet. In the Poincaré group, an irreducible supermultiplet contains
finitely many spins; by contrast, we will see below that super-BMS3 multiplets contain
infinitely many spins. Apart from this difference, the structure of induced representa-
tions of super semi-direct products is essentially the same as in the bosonic case: they
consist of wavefunctions on an orbit, taking their values in the space of the represen-
tation R0 ⊗ R. In particular, the Frobenius formula (10.54) for characters remains
valid, up to the replacement of R by R0 ⊗ R.
Supersymmetric Virasoro algebra

As a preparation for super BMS3 , let us first recall the definition of the super Virasoro
algebra. The latter is built by adding to Vect(S 1 ) an odd subalgebra F−1/2 (S 1 ) of
−1/2-densities on the circle [56, 300]. This produces a Lie superalgebra, isomorphic
to Vect(S 1 ) ⊕ F−1/2 (S 1 ) as a vector space, which we shall write as sVect(S 1 ). Its
elements are pairs (X, S), where X = X(ϕ)∂/∂ϕ and S = S(ϕ)(dϕ)−1/2 , and the
super Lie bracket is defined as

(X, S), (Y, T ) ≡ [X, Y ] + S ⊗ T, X · T − Y · S . (11.105)
Here [X, Y ] is the standard Lie bracket of vector fields and the dot denotes the natural
action of vector fields on F−1/2 (S 1 ), so that X · T is the −1/2-density with component
1
X · T ≡ XT 0 − X 0 T. (11.106)
2
(This is formula (6.32) with h = −1/2.) Upon expanding the functions X(ϕ) and
S(ϕ) in Fourier modes, one recovers the standard N = 1 supersymmetric extension of
the Witt algebra. Choosing S(ϕ) to be periodic or antiperiodic leads to the Ramond
or the Neveu-Schwarz sector of the superalgebra, respectively.
The central extension of sVect(S 1 ) is the super Virasoro algebra, svir. Its elements
are triples (X, S, λ) where (X, S) ∈ sVect(S 1 ) and λ ∈ R, with a super Lie bracket

(X, S, λ), (Y, T, µ) ≡ [X, Y ] + S ⊗ T, X · T − Y · S, c(X, Y ) + h(S, T ) , (11.107)
where c is the Gelfand-Fuks cocycle (6.43) while h is its supersymmetric cousin,

Z 2π
1
h(S, T ) ≡ dϕ S 0 T 0 . (11.108)
12π 0
By expanding the functions X and S in Fourier modes, one obtains the usual commu-
tation relations of N = 1 super Virasoro. Explicitly, defining the generators
Lm ≡ eimϕ ∂ϕ , 0, 0 , Qr ≡ 0, eirϕ (dϕ)−1/2 , 0 ,

Z ≡ (0, 0, 1),
one finds that (11.107) yields the super Lie brackets

Z
i[Lm , Ln } = (m − n)Lm+n + m3 δm+n,0 ,
m 12
i[Lm , Qr } = − r Qm+r ,
2
Z
[Qr , Qs } = Lr+s + r2 δr+s,0 . (11.109)
6
The super Virasoro algebra is (half of) the asymptotic symmetry algebra of three-
dimensional supergravity with Brown-Henneaux boundary conditions [394, 395] (see
also [396]). In that context the vector field X is one of the components of an asymp-
totic Killing vector field (8.30), while S is one of the components of an asymptotic
Killing spinor. The fact that the quantization of three-dimensional supergravity pro-
duces super Virasoro representations was verified in [69] by showing that the one-loop
partition function of supergravity on thermal AdS3 coincides with the vacuum char-
acter of two super Virasoro algebras. In the remainder of this section our goal is to
describe the flat analogue of these results.
Supersymmetric BMS3 groups

Equipped with the definition of super semi-direct products and that of the super
Virasoro algebra, we can now define the N = 1 super BMS3 group [167, 280]: it is a
super semi-direct product (11.100) whose even piece is the BMS3 group (9.58), and
whose odd subspace is the space of densities F−1/2 (S 1 ) with the bracket {S, T } = S⊗T .
In other words, the (centreless) super bms3 algebra is a super semi-direct sum
sbms3 = Vect(S 1 ) A Vect(S 1 )Ab ⊕ F−1/2 ,

(11.110)
where Vect(S 1 )Ab ⊕ F−1/2 may be seen as an Abelian version of sVect(S 1 ). Again,
choosing periodic/antiperiodic boundary conditions for F−1/2 yields the Ramond/Ne-
veu-Schwarz sector of the theory (respectively). Central extensions can be included as
in (9.62) and lead to a supersymmetric version of the centrally extended algebra (9.64).
The elements of the resulting super Lie algebra sbmsd 3 are 5-tuples (X, λ; α, S, µ), where
(X, α, S) belongs to sbms3 and λ, µ are real numbers, with a super Lie bracket that
extends (9.46):
h o
(X, λ; α, S, µ), (Y, κ; β, T, ν) =

= [X, Y ], c(X, Y ); [X, β] − [Y, α], X · T − Y · S; c(X, β) − c(Y, α) + h(S, T ) .
(11.111)
Here c is again the Gelfand-Fuks cocycle (6.43) while h is given by (11.108). Upon
introducing generators analogous to (9.67), the central charges (9.47) and Qr ≡
(0, 0; 0, eirϕ (dϕ)−1/2 , 0), one finds the brackets (9.68) supplemented with
m
i[Jm , Qr } = − r Qm+r , (11.112a)
2
i[Pm , Qr } = 0 , (11.112b)
Z2 2
[Qr , Qs } = Pr+s + r δr+s,0 . (11.112c)
6
The indices r, s are integers/half-integers in the Ramond/Neveu-Schwarz sector, re-
spectively. Note that the centrally extended bracket of supercharges only involves the
central charge Z2 that pairs superrotations with supertranslations.
In the gravitational context, the functions X and α generate superrotations and

supertranslations, while S(ϕ) generates local supersymmetry transformations that be-
come global symmetries upon enforcing suitable boundary conditions on the fields.
The surface charge associated with (X, α, S) then takes the form [167]
Z 2π h
1 i
Q(X,α,S) [j, p, ψ] = dϕ X(ϕ)j(ϕ) + α(ϕ)p(ϕ) + S(ϕ)ψ(ϕ) , (11.113)
2π 0
where j and p are the angular momentum and Bondi mass aspects of (9.25), while
ψ(ϕ) is one of the subleading components of the gravitino at null infinity. The triple
(j, p, ψ) is a coadjoint vector for the (centrally extended) super BMS3 group. In partic-
ular (j, p) are quadratic densities, while ψ(ϕ) has weight 3/2 on the circle. Upon using
formula (11.41), the charges (11.113) satisfy the algebra (11.111) with definite values
Z1 = 0, Z2 = c2 = 3/G for the central charges. Note that the gravitino naturally
satisfies Neveu-Schwarz boundary conditions on the celestial circle, as it represents
Lorentz transformations up to a sign.
The construction of the super BMS3 group can be generalized in a straightforward

way. Indeed, let G be a (bosonic) group, g its Lie algebra, sg a super Lie algebra
whose even subalgebra is g. Then one can associate with G a (bosonic) exceptional
semi-direct product G n g — the even BMS [ 3 group (9.62) is of that form, with G
the Virasoro group. Now let sgAb denote the “Abelian” super Lie algebra which is
isomorphic to sg as a vector space, but where all brackets involving elements of g are
set to zero. One may then define a super semi-direct product

G n g, g A sgAb (11.114)
where we use the notation (11.100). This structure appears to be ubiquitous in three-
dimensional, asymptotically flat supersymmetric higher-spin theories.
11.4.3 Supersymmetric BMS3 particles

Unitary representations of the super BMS3 group can be classified along the lines
explained in chapter 10. In the remainder of this section we describe this classifica-
tion and use it to evaluate characters of the centrally extended super BMS3 group.
We conclude with the observation that these characters reproduce one-loop partition
functions of three-dimensional asymptotically flat supergravity and hypergravity.
Admissible super BMS3 orbits

The unitary representations of super BMS3 are classified by the same supermomen-
tum orbits as in the purely bosonic case, i.e. coadjoint orbits of the Virasoro group.
However, supermomenta that do not satisfy condition (11.102) are forbidden, so our
first task is to understand which orbits are admissible. To begin, recall that the ad-
missibility condition (11.102) is invariant under superrotations. Thus, if we consider
a supermomentum orbit containing a constant p0 say, the supermomenta on the orbit
will be admissible if and only if p0 is. Including the central charge c2 , we ask: which
pairs (p0 , c2 ) are such that
h(p0 , c2 ), {S, S}i ≥ 0 for any S ∈ F−1/2 (S 1 ) ? (11.115)
Here h·, ·i is the pairing (6.111) of centrally extended supermomenta with centrally
extended supertranslations. Using the super Lie bracket (11.111), we find
Z 2π
1 c2
h(p0 , c2 ), {S, S}i = dϕ p0 (S(ϕ))2 + (S 0 (ϕ))2 . (11.116)
2π 0 6
Since the term involving (S 0 )2 can be made arbitrarily large while keeping S 2 arbitrar-
ily small, a necessary condition for (p0 , c2 ) to be admissible is that c2 be non-negative.
Already note that this condition did not arise in the bosonic BMS3 group. The admis-
sibility condition on p0 , on the other hand, depends on the sector under consideration:
• In the Ramond sector, S(ϕ) is a periodic function on the circle. In particular,
S(ϕ) = const. is part of the supersymmetry algebra, so for expression (11.116)
to be non-negative for any S, we must impose p0 ≥ 0.
• In the Neveu-Schwarz sector, S(ϕ) is antiperiodic (i.e. S(ϕ + 2π) = −S(ϕ)) and
can be expanded in Fourier modes as
X
S(ϕ) = sn+1/2 ei(n+1/2)ϕ . (11.117)
n∈Z
Then expression (11.116) becomes

Xh c2 i
h(p0 , c2 ), {S, S}i = p0 + (n + 1/2)2 |sn+1/2 |2 , (11.118)
n∈Z
6
and the admissibility condition amounts to requiring all coefficients in this series
to be non-negative, which gives
c2
p0 ≥ − . (11.119)
24
These bounds are consistent with earlier observations in three-dimensional supergrav-

ity [167], according to which Minkowski space-time (corresponding to p0 = −c2 /24)
realizes the Neveu-Schwarz vacuum, while the Ramond vacuum is realized by the
null orbifold (corresponding to p0 = 0). Analogous results hold in AdS3 [396]. More
general admissibility conditions can presumably be worked out for non-constant su-
permomenta by adapting the proof of the positive energy theorem of section 7.3, but
we will not address this question here.
Super BMS3 multiplets

As explained around (11.103), any unitary representation of super BMS3 based on
a supermomentum orbit Op comes equipped with a representation τ of the Clifford
algebra
Cp = T F−1/2 (S 1 ) / S 2 − h(p, c2 ), {S, S}i .

(11.120)
Let us build such a representation. For definiteness we work in the Neveu-Schwarz
sector and take p to be a constant admissible supermomentum p0 = M − c2 /24 with
M > 0, whose little group is U(1). Then the bilinear form (11.118) is non-degenerate
and the representation τ of the Clifford algebra (11.120) must be such that
c
2 2
τ [Qr ] · τ [Qs ] + τ [Qs ] · τ [Qr ] = (r − 1/4) + M δr+s,0 , r, s ∈ Z + 1/2. (11.121)
6
In order to make τ irreducible, we start with a highest-weight state |0i such that
τ [Qr ]|0i = 0 for r > 0, and generate the space of the representation by its “descen-
dants” τ [Q−r1 ]...τ [Q−rn ]|0i, with 0 < r1 < ... <
Prnn . It follows from the Lie brackets
(11.112) that each descendant state has spin s+ i=1 ri , where s is the spin of the state
|0i; this observation uniquely determines the little group representation R0 satisfying
(11.104). Thus, a super BMS3 particle consists of infinitely many particles with spins
increasing from s to infinity.
A similar construction can be carried out for the vacuum supermomentum at M =

0, with the subtlety that the Clifford algebra (11.120) (or equivalently (11.121)) is
degenerate. As explained below (11.103), one needs to quotient (11.120) by the radical
of the bilinear form (11.118), resulting in a non-degenerate Clifford algebra C¯p . In the
case at hand this algebra is generated by supercharges Qr with |r| > 1, and the
representation τ must satisfy (11.121) with M = 0 and |r|, |s| > 1. The remainder
of the construction is straightforward: starting from a state |0i with, say, vanishing
spin, one generates the space of the representation by acting on it with τ [Q−r ]’s,
where r > 1. The vacuum representation of super BMS3 thus contains infinitely many
“spinning vacua” with increasing spins.
Characters
The Fock space representations just described can be used to evaluate characters. For
example, in the massive case with spin s one finds the supersymmetric little group
character
+∞
Y
iθJ0 isθ iθ/2 3iθ/2 2iθ isθ
1 + ei(n−1/2)(θ+i) ,

tr e =e 1+e +e +e + ··· = e
n=1
(11.122)
where we have added a small imaginary part to θ to ensure convergence of the prod-
uct; the trace is taken in the fermionic Fock space associated with the “highest-weight
state” |0i. The vacuum case is similar, except that the product would start at n = 2
rather than n = 1 (and s = 0). Note that (11.122) explicitly breaks parity invari-
ance; this can be fixed by replacing the parity-breaking Fock space representations τ
described above by parity-invariant tensor products τ ⊗ τ̄ , where τ̄ is the same as τ
with the replacement of Qr by Q−r . The trace of a rotation operator in the space of
τ ⊗ τ̄ then involves the norm squared of the product appearing in (11.122).
As explained at the beginning of section 11.4.2, the character of an induced repre-

sentation of a super semi-direct product takes the same form (4.33) as in the bosonic
case, but with the character of R replaced by that of a (reducible) representation
R0 ⊗ R compatible with the Clifford algebra representation τ . We thus find that the
character of a rotation by θ (together with a Euclidean time translation by β), in
the parity-invariant vacuum representation of the N = 1, Neveu-Schwarz super BMS3
group, reads
+∞
Y
χsuper
vac
BMS
[(rotθ , iβ)] = χBMS
vac [(rotθ , iβ)] · |1 + ei(n−1/2)(θ+i) |2
n=2
+∞
Y |1 + ei(n−1/2)(θ+i) |2
= eβc2 /24 . (11.123)
n=2
|1 − ein(θ+i) |2
Comparing with (11.37) and (11.99), we recognize the product of the (suitably reg-
ularized) partition functions of two massless fields with spins 2 and 3/2, that is, the
one-loop partition function of N = 1 supergravity in three-dimensional flat space.
Higher-spin supersymmetry and hypergravity

In [388, 389], the authors considered a three-dimensional hypergravity theory consist-
ing of a metric coupled to a single field with half-integer spin s + 1/2, with s larger
than one. Upon imposing suitable asymptotically flat boundary conditions, they found
that the asymptotic symmetry algebra spans a superalgebra that extends the bosonic
bms3 algebra by generators Qr of spin s + 1/2. The one-loop partition function of that
system is the product of the graviton partition function (see eq. (11.37) for s = 2)
with the fermionic partition function (11.99). We now show that this partition func-
tion coincides with the vacuum character of the corresponding asymptotic symmetry
group (in the Neveu-Schwarz sector).
The irreducible, unitary representations of the asymptotic symmetry group of [389]

are classified by the same orbits and little groups as for the standard BMS3 group. In
particular, we can consider the orbit of a constant supermomentum p0 = M − c2 /24;
the associated Clifford algebra representation τ mentioned below (11.103) then satisfies
a natural generalization of eq. (11.121) (see eq. (7.23) in [389]):
s−1
(2j + 1)2

Y c2 2
τ [Qr ]τ [Q` ] + τ [Q` ]τ [Qr ] = r − +M δr+`,0 , (11.124)
j=0
6 4
where r and ` are integers or half-integers, depending on the sector under consideration
(Ramond or Neveu-Schwarz, respectively). In order for the orbit to be admissible
in the sense of (11.102), the value of M must be chosen so as to ensure that all
coefficients on the right-hand side of (11.124) are non-negative. In particular, the
vacuum value M = 0 is admissible in the Neveu-Schwarz sector, in which case the
anticommutators {τ [Qr ], τ [Q−r ]} vanish for |r| = 1/2, ..., s − 1/2. Thus, in the Neveu-
Schwarz vacuum, the Clifford algebra (11.124) degenerates and τ must really be seen as
a representation of the non-degenerate subalgebra generated by the Qr ’s with |r| ≥ s.

The corresponding Fock space representation can be built as explained below (11.121),
and the spins of the basis states in this representation are uniquely determined by the
fact that the Qr ’s have spin s + 1/2. The corresponding Fock space character is thus
+∞
Y
tr eiθJ0 = 1 + ei(n−1/2)(θ+i) ,

(11.125)
n=s+1
which generalizes (11.122). The character for τ ⊗ τ̄ is the squared norm of this expres-
sion, and the resulting vacuum character of the hypersymmetric BMS3 group is
+∞
|1 + ei(n+1/2)(θ+i) |2
Q
n=s
χhyper
vac
BMS
[(rotθ , iβ)] = eβc2 /24 +∞
. (11.126)
Q
|1 − eim(θ+i) |2
m=2
As announced earlier, this coincides with the (suitably regularized) one-loop partition
function of asymptotically flat gravity coupled to a massless field with spin s + 1/2.
We have thus completed our overview of the relation between BMS3 characters and
one-loop partition functions in three dimensions.
*11.A From mixed traces to bosonic characters

This section and the next one are technical appendices that describe various com-
putations concerned with characters of highest-weight representations of SO(n). These
considerations are useful for sections 11.1.2 and 11.4.1. Other than that, they may be
skipped on a first reading.
11.A.1 Mixed traces and symmetric polynomials

In this part of the appendix we prove that the mixed trace (11.19) of Iµs ,αs in D
dimensions coincides with a certain difference of complete homogeneous symmetric
polynomials in the traces of J n as given by
χs [nθ~ ] = hs (J n ) − hs−2 (J n ) , (11.127)
where " m #
s
X Y Tr[(J n )k ] k
hs (J n ) = m
. (11.128)
m1 ,...,ms ∈ N k=1
m k !k k
m1 +2m2 +...+sms = s
By definition, the complete homogeneous symmetric polynomial of degree s in D com-

plex variables λ1 , ..., λD is
s
X X
hs (λ1 , ..., λD ) = λ`11 λ`22 ...λ`DD = λ`1 λ`2 ...λ`s . (11.129)
`1 ,...,`D = 0 1≤`1 ≤`2 ≤...≤`s ≤D
`1 +...+`D = s
Using the variant of Newton’s identities

s
1X
hs (λ1 , ..., λD ) = hs−N (λ1 , ..., λD )(λN N
1 + ... + λD ) , (11.130)
s N =1
one can show by recursion (see e.g. [397, p. 24f]) that the polynomial (11.129) can
equivalently be written as in (11.128):
s
X Y (λk + ... + λk )mk
1 D
hs (λ1 , ..., λD ) = . (11.131)
m1 ,...,ms ∈ N k=1
mk ! k mk
m1 +2m2 +...+sms = s
We shall use this relation later. To prove (11.127), we start with the following:
Lemma. Let J be a complex D × D matrix with eigenvalues λ1 , ..., λD . Then,

1
(δ µα )s (Jµα )s = hs (λ1 , λ2 , ..., λD ) , (11.132)
s!
where we use the same notation for contracting symmetrized indices as in (11.13).
Proof. The left-hand side of (11.132) can be seen as a trace over symmetric tensor
powers of J. Indeed, δ µα Jµα = Tr(J) is clear; as for 21 (δ µα )2 (Jµα )2 , one gets
2
1 µα 2 1 1X
(δ ) (Jµα )2 = Tr(J)2 + Tr J 2 = Tr S 2 (J) = Tr J i Tr S 2−i (J) ,

2 2 2 i=1
(11.133)
where S k (J) is the k th symmetric tensor power of J. One then defines recursively
s
1 µα s s s 1X
Tr J i Tr S s−i (J) ,

(δ ) (Jµα ) = Tr (S (J)) = (11.134)
s! s i=1
so that s!1 (δ µα )s (Jµα )s is just a trace in the sth symmetric tensor power of the D-
dimensional vector space V on which Jµα acts as a linear operator. Now consider an
eigenbasis {e1 , ..., eD } for Jµα , with J · ek = λk ek . Since s!1 (Jµα )s is the sth symmet-
ric tensor power of Jµα one can construct an eigenbasis for s!1 (Jµα )s by symmetrizing
ek1 ⊗ ek2 ⊗ ... ⊗ ekD , with k1 ≤ k2 ≤ ... ≤ kD . These eigenvectors have eigenvalues
λl1 λl2 ...λlD , and since (δ µα )s s!1 (Jµα )s is the trace of s!1 (Jµα )s , relation (11.132) follows
upon using the second expression of hs (λ1 , ..., λD ) in (11.129).
We can now turn to the proof of (11.127). To this end we fix conventionally the
number of terms entering the contraction of two symmetrized expressions as follows.
Objects with lower indices are symmetrized with the minimum number of terms re-
quired and without overall normalization factor, while objects with upper indices are
not symmetrized at all, since the symmetrization is induced by the contraction. This
specification is needed because terms with lower and upper indices in a contraction
may have a different index structure and therefore the number of terms needed for
their symmetrization may be different. For instance
Aµ B µ C µ Dµµ Eµ ≡ Aµ B ν C ρ (Dµν Eρ + Dνρ Eµ + Dρµ Eν )

1
= (Aµ B ν C ρ + Aν B ρ C µ + Aρ B µ C ν + Aµ B ρ C ν + Aρ B ν C µ + Aν B µ C ρ ) Dµν Eρ .
2
(11.135)
In order to simplify computations, we define
Tµs , αs ≡ Jµα ...Jµα , T [s] ≡ Tµs , αs (δ µα )s , (11.136)
which implies the contraction rules
δ µµ Tµs , αs = 2 δαα Tµs−2 , αs−2 , δ αα Tµs ,αs = 2 δµµ Tµs−2 ,αs−2 . (11.137)
In terms of the tensors Tµs ,αs , the mixed trace (11.19) can be written as
b 2s c
(−1)m s! [D + 2 (s − m − 2)]!!

1 s
~ = Tµs ,βs δ µβ +
X
χs [nθ] × (11.138)
s! m=1
2m m! (s − 2m)! [D + 2 (s − 2)]!!

µµ m µβ s−2m ββ m
× (δ ) (δ ) (δ )
[ 2s ]
(11.137) 1 [s] X (−1)m [D + 2 (s − m − 2)]!!
= T + × (11.139)
s! m=1
2m−1 m! (s − 2m)! [D + 2 (s − 2)]!!
× (δ µµ )m (δ µβ )s−2m (δ ββ )m−1 δµµ Tµs−2 ,βs−2 . (11.140)
To compute the trace of the (δ µµ )m (δ µβ )s−2m (δ ββ )m−1 terms, we first change our sym-
s!
metrization from δµµ Tµs−2 ,βs−2 (which contains 2(s−2)! terms) to the aforementioned
product of δ’s. In doing so one has to introduce a factor accounting for the number of
terms in each structure as
s!
δµµ Tµs−2 , βs−2 terms, (11.141a)
2(s − 2)!
s−2m m−1 s! (s − 2)!
(δ µµ )mu δ µβ δ ββ m
× m−1 terms, (11.141b)
2 m! 2 (m − 1)!(s − 2m)!
which implies
s
b2c
~ 1 [s] X (−1)m 2m−1 (m − 1)! [D + 2 (s − m − 2)]!!
χs [nθ] = T + ×
s! m=1
[(s − 2)!]2 [D + 2 (s − 2)]!! (11.142)
m s−2m m−1 µµ µs−2 ,βs−2
×δµµ δµβ δββ δ T .
Taking into account the correct combinatorial factors one obtains

m s−2m m−1 µµ m−1 s−2m m−1 m s−2m−2 m
δµµ δµβ δββ δ = [D + 2(s − m − 1)] δµµ δµβ δββ + 2m δµµ δµβ δββ , (11.143)
which then yields

 s
b2c
~ 1 [s]  X (−1)m 2m−1 (m − 1)! [D + 2 (s − m − 1)]!! m−1 s−2m m−1
χs [nθ] = T + δµµ δµβ δββ
s! m=1
[D + 2 (s − 2)]!!
s

b 2 c−1
X (−1)m 2m m! [D + 2 (s − m − 2)]!! 1
m s−2m−2 m  µs−2 ,βs−2
+ δµµ δµβ δββ 2 T .
m=1
[D + 2 (s − 2)]!! [(s − 2)!]
(11.144)
Shifting m → m + 1 in the upper sum one can see that both sums are identical apart
from the overall sign and the lower extremum. Thus (11.144) boils down to
~ = 1 [s] 1 s−2 µs−2 ,βs−2 1 1

χs [nθ] T − 2 δµβ T = T [s] − T [s−2] . (11.145)
s! [(s − 2)!] s! (s − 2)!
Now using (11.136) and (11.132) one obtains
~ = 1 T [s] −
χs [nθ]
1
T [s−2] = hs (λ1 , λ2 , ..., λD ) − hs−2 (λ1 , λ2 , ..., λD ) , (11.146)
s! (s − 2)!
where λ1 , ..., λD are the eigenvalues of J n . (These eigenvalues are e±inθj for j = 1, ..., r,
and one or two unit eigenvalues depending on whether D is odd or even, respectively.)
This leads to the desired result: since traces of powers of J n can be written as
Tr[(J n )k ] = λk1 + ... + λkD (11.147)
in terms of the eigenvalues of J n , the complete homogeneous symmetric polynomials

expressed as (11.131) exactly coincide with the combination (11.128), and equation
(11.146) coincides with (11.127).
11.A.2 Symmetric polynomials and SO(D) characters

In this part of the appendix we review the relation between complete homogeneous
symmetric polynomials and characters of orthogonal groups. Most of the explicit
proofs can be found in [398], chapter 24, to which we refer for details on our arguments
below. We study separately the cases of odd and even D and let r ≡ b(D − 1)/2c,
with θ1 , ..., θr the non-vanishing angles appearing in the rotations (11.8).
Odd D
We consider the Lie algebra so(D) = so(2r + 1), with rank r. Choosing a basis of
C2r+1 such that the Lie algebra so(2r + 1)C can be written in terms of complex ma-
trices, we may choose the Cartan subalgebra to be the subalgebra h of so(2r + 1)C
consisting of diagonal matrices. As a basis of h we choose the matrices Hi whose
entries all vanish, except the (i, i) and (r + i, r + i) entries which are 1 and −1,
respectively (with i = 1, ..., r). In our convention (11.7), the operator Hi generates ro-
tations in the plane (xi , yi ). Then, calling Li the elements of the dual basis (such that
hLi , Hj i = δij ), a dominant weight is one of the form λ = λ1 L1 + ...λr Lr ≡ (λ1 , ..., λr )
with λ1 ≥ ... ≥ λr ≥ 0.
Let λ be a dominant weight for so(2r + 1). According to formula (24.28) in [398],
the character of the irreducible representation of so(2r + 1) with highest weight λ is
λi +r−i+ 21 −(λi +r−i+ 12 )
qj − qj
(2r+1)
[q1 , ..., qr ] = Trλ q1H1 · · · qrHr =

χλ , (11.148)
r−i+ 12 −(r−i+ 12 )
qj − qj
where q1 , · · · qr are arbitrary complex numbers4 , Trλ denotes a trace taken in the space
of the representation, and |Aij | denotes the determinant of the matrix A with rows i
and columns j. This expression is a corollary of the Weyl character formula. Using
proposition A.60 and Corollary A.46 of [398], it can be rewritten as
(2r+1)
χλ [q1 , ..., qr ] = |hλi −i+j − hλi −i−j | , (11.149)
where hj = hj q1 , ..., qn , q1−1 , ..., qn−1 , 1 is a complete homogeneous symmetric polyno-

mial of degree j in 2r + 1 variables. In particular, for a highest weight λs = (s, 0, ..., 0)

(where s is a non-negative integer), the matrix appearing on the right-hand side of
(11.149) is upper triangular, with the entry at i = j = 1 given by hs − hs−2 and
all other entries on the main diagonal equal to one. Accordingly, the determinant in
(11.149) boils down to hs − hs−2 in that simple case. For the rotation (11.8) we may
identify qj = einθj , and we conclude that
sin λi + r − i + 21 nθj

(2r+1) ~
χλs [nθ] = = hs (J n ) − hs−2 (J n ) , (11.150)
sin r − i + 21 nθj

where λi = s δi1 . Thus for odd D the difference of symmetric polynomials in (11.127)
is just a character of SO(D).
Even D
We now turn to the Lie algebra so(2r + 2), with rank r + 1. As in the odd case
we choose a basis of C2r+2 such that we can write the Lie algebra so(2r + 2) in
terms of complex matrices and the Cartan subalgebra is generated by r + 1 diagonal
matrices Hi whose entries all vanish, except (Hi )ii = 1 and (Hi )r+1+i,r+1+i = −1.
We call Li the elements of the dual basis, and with these conventions a weight
λ = λ1 L1 + ... + λr+1 Lr+1 ≡ (λ1 , ..., λr+1 ) is dominant if λ1 ≥ λ2 ≥ ... ≥ λr ≥ |λr+1 |.
Let λ be a dominant weight for so(2r + 2). Then formula (24.40) in [398] gives the
character of the associated highest-weight representation as
h i
(2r+2) Hr+1
χλ [q1 , ..., qr+1 ] = Trλ q1H1 · · · qr+1
−(λi +r+1−i) −(λi +r+1−i)
qjλi +r+1−i + qj + qjλi +r+1−i − qj (11.151)
= ,
−(r+1−i)
qjr+1−i + qj
4
Eventually these numbers will be exponentials of angular potentials, so they are fugacities asso-
ciated with the rotation generators Hi .
where we use the same notations as in (11.148), except that now i, j = 1, ..., r +1. Note
that the second term in the numerator of this expression vanishes whenever λr+1 = 0
−(λ +r+1−i)
(because the (r + 1)th row of the matrix qjλi +r+1−i − qj i vanishes). Since this
is the case that we will be interested in, we may safely forget about that second term
from now on. Alternatively, for the mixed traces (11.19) that we need, we may take
qj = einθj for j = 1, ..., r and qr+1 = 1 without loss of generality, so that this second
term vanishes again. Using proposition A.64 of [398], one can then rewrite (11.151) as
(2r+2)
[q1 , ..., qr , 1] = |hλi −i+j − hλi −i−j | ,
χλ (11.152)
where hj = hj q1 , ..., qr , 1, q1−1 , ..., qr−1 , 1 . Finally, using the same arguments as for

odd D, one easily verifies that the determinant on the right-hand side of (11.152)
reduces once more to hs − hs−2 for a highest weight λs = (s, 0, ..., 0). Writing again
qj = einθj , one concludes that, for even D,
(2r+2) |cos [(λi + r + 1 − i) nθj ]|

χλs [nθ1 , ..., nθr , nθr+1 = 0] = = hs (J n ) − hs−2 (J n ),
|cos [(r + 1 − i) nθj ]|
θr+1 =0
(11.153)
where λi = s δi1 . This concludes the proof of (11.21). Note that, for non-vanishing
θr+1 , the quotient of denominators in the middle of (11.153) is actually the character
(2r+2)
χ λs (nθ1 , ..., nθr , nθr+1 ). This detail will be useful in appendix 11.A.3.
11.A.3 Differences of SO(D) characters

In this part of the appendix we prove the following relations between characters of
orthogonal groups:
(2r+1) (2r+1) (2r)
χ λs [θ~ ] − χλs−1 [θ~ ] = χλs [θ~ ] , (11.154a)
Xr
(2r) ~ (2r) ~ (2r−1)
χλs [θ ] − χλs−1 [θ ] = Ark [θ~ ]χλs [θ1 , ..., θbk , ..., θr ] . (11.154b)
k=1
Here θ~ = (θ1 , .., θr ), λs is the weight with components (s, 0, ..., 0) in the basis defined
above equations (11.148) and (11.151), and the hat denotes omission of an argument,
while the coefficients Ark are the quotients of determinants defined in (11.31). Note
that, when one of the angles θ1 , ..., θr vanishes, say θ` = 0, then Ark = δk` and relation
(11.154b) reduces to
(2r) (2r) (2r−1)
χλs [θ~ ] − χλs−1 [θ~ ] = χλs [θ1 , ..., θb` , ..., θr ] . (11.155)
θ` =0 θ` =0
Proof of (11.154). We start by defining the matrices
(Ar )ij = sin (r − i + 21 )θj , (B r )ij = cos [(r − i)θj ] ,

(11.156)
so that in particular
r
~ = |B |θk =0 .
Ark (θ) (11.157)
|B r |
We shall also use the shorthand notation
M r [θk ] ≡ |Mij (θ1 , ..., θk−1 , θk+1 , ..., θr+1 )| (11.158)
to denote the determinant of the r × r matrix missing the angle θk of any of the
matrices defined in (11.156). As a preliminary step towards the proof, we list the four
following identities:
|Ar |
Qr = 2r−1 |B r | , (11.159a)
j=1 sin (θj /2)
(r−1)(r−2) Y
| cos [(r − i)θj ] | = 2 2 (cos(θi ) − cos(θj )) , (11.159b)
1≤i<j≤r
r
|B r |θk =0 Y
= 2r−1 (−1) k+1
sin (θj /2) , (11.159c)
Ar−1 [θk ] j=1
j6=k
r
X
r
|B | = |B r |θk =0 . (11.159d)
k=1
Here (11.159a) can be proven by induction on r upon expanding the determinant

~ along the first line of the matrix Ar . Property (11.159b) can be shown by
|Ar (θ)|
observing that
r−i−1
X
r−i−1 r−i
cos[(r − i)θj ] = 2 cos (θj ) + ck cos(kθj ) (11.160)
k=1
with some irrelevant real coefficients ck , and that the contribution of the second term of
this expression to the determinant | cos[(r−i)θj ]| vanishes by linear dependence. Equa-
tion (11.159c) then follows from (11.159a) and (11.159b), while property (11.159d) can
again be proved by induction on r.
Thanks to eqs. (11.159), we can tackle the proof of (11.154). Equation (11.154a)
(2r+1)
is easy: using expression (11.150) for the character χλs , we can write the difference
of characters on the left-hand side of (11.154a) as
r
(−1)k+1 2 cos[(s + r − 1)θk ] sin (θk /2) Ar−1 [θk ]
P
(2r+1) (2r+1) k=1
χλs − χλs−1 = . (11.161)
|Ar |
Property (11.159a) then allows us to reduce this expression to the quotient of denom-
inators appearing in the middle of eq. (11.153) (with the replacement of r + 1 by r
(2r)
and all angles non-zero), which is indeed the sought-for character χλs [θ~ ].
Equation (11.154b) requires more work. Using once more the expression in the
middle of (11.153), we first rewrite the left-hand side of (11.154b) as
r
(−1)k+1 (−2 sin[(s + r − 23 )θk ] sin (θk /2) B r−1 [θk ]
P
(2r) (2r) k=1
χλs − χλs−1 = . (11.162)
|B r |
Let us now recover this expression as a combination of characters of SO(2r − 1): using
formula (11.150) and the identities (11.159), one finds
r
X (2r−1)
χ λs [θ1 , ..., θbk , ..., θr ]|B r |θk =0
k=1
r r
(11.159c) X
k+1 r−1
Y
= (−1) 2 sin (θj /2) ×
k=1 j=1
j6=k
k−1
X
× (−1)j+1 sin[(s + r − 32 )θj ]Ar−2 [θj , θk ]
j=1
r
X
j
+ (−1) sin[(s + r − 3
)θ ]Ar−2 [θj , θk ]
2 j
j=k+1
r
(11.159a) X
= (−1)k+1 22r−4 sin[(s + r − 23 )θk ] sin (θk /2) ×
k=1
k−1
X r
Y
j r−2
× (−1) B [θj , θk ] sin2 (θi /2)
j=1 i=1
i∈{j,k}
/
r
X Yr
j+1 r−2 2
+ (−1) B [θj , θk ] sin (θi /2)
j=k+1 i=1
i∈{j,k}
/
r
(11.159b) X
= (−1)k+1 (−2) sin[(s + r − 32 )θk ] sin (θk /2)
k=1
" k−1 r
#
X X
(−1)j B r−1 [θk ] θj =0
+ B r−1 [θk ] θj =0
j=1 j=k+1
r
(11.159d) X
= (−1)k+1 (−2) sin[(s + r − 23 )θk ] sin (θk /2) B r−1 [θk ]. (11.163)
k=1
This coincides with the numerator of the right-hand side of (11.162), so identity
(11.154b) follows with Ark given by (11.157).
11.A.4 From SO(D) to SO(D − 1)

In this appendix we prove relation (11.33) between characters of SO(D) and SO(D−1):
Lemma. One has the following relations:

s
X
(2r+1) (2r)
χ λs [θ1 , ..., θr ] = χλj (θ1 , ..., θr ), (11.164a)
j=0
s X
X r
(2r) ~ (2r−1)
χλs [θ1 , ..., θr ] = Ark (θ)χλj [θ1 , ..., θbk , ..., θr ]. (11.164b)
j=0 k=1
Here λj is the weight (j, 0, ..., 0) as explained above (11.21) or below (11.149), and
~ is the quotient (11.31) or (11.157). Since the proofs of these two identities are
Ark (θ)
very similar, we will only display the proof of (11.164a).
Proof of (11.164a). Eq. (11.164a) can be written as

r r
(−1)k+1 sin[(s + r − 21 )θk ]Ar−1 [θk ] (−1)k+1 cos[(j + r − 1)θk ]B r−1 [θk ]
P P
s
k=1 (11.159a) X k=1
= ,
|Ar | j=0
|B r |
(11.165)
where we used formulas (11.150) and (11.153) for the characters, as well as the def-
inition (11.156) of Ar and B r . One can then use identities (11.159a) and (11.159c)
to match the right-hand side of this expression with the left-hand side, proving the
desired identity.
*11.B From mixed traces to fermionic characters

This appendix is the fermionic (half-integer spin) analogue of section 11.A. It may
be skipped in a first reading.
11.B.1 Mixed traces and symmetric polynomials

Our goal here is to prove the first equality of (11.95), following the same method as in
appendix 11.A.1 for the bosonic case. First, using the definition (11.91) of U and the
contraction rules (11.137), one can write (11.94) as
" b 2s c
(F ) ~ 1 [s] X (−1)m [D + 2(s − m − 1)]!!
χs [nθ] = T + ×
s! m=1
2m−1 m!(s − 2m)![D + 2(s − 1)]!!
#
× (δ µµ )m (δ µβ )s−2m (δ ββ )m−1 δµµ Tµs−2 ,βs−2 Tr[U n ]
b s−1
2
c
X (−1)m+1 [D + 2(s − m − 2)]!!
+ ×
m=0
2m m!(s − 2m − 1)![D + 2(s − 1)]!!
× Tr[Tµs−1 ,βs−1 γµ γ µ (δ µµ )m (δ µβ )s−2m−1 (δ ββ )m U n ], (11.166)
where T [s] is the notation (11.136). In the first term of this expression, we shift the
symmetrization on the δ’s i.e. we exchange upper and lower indices while taking
into account the change in multiplicities of the terms involved; in all other terms, we
compute one contraction with δ ββ . Eq. (11.166) then simplifies to
) ~ 1 [s] 1 s−1 n
χ(F
s [nθ] = T Tr[U n ] − Tr[T µs−1 ,βs−1 γµ γ µ δµβ U ] (11.167)
s! [(s − 1)!]2 [D + 2(s − 1)]
b 2s c
X (−1)m 2m−1 (m − 1)![D + 2(s − m − 1)]!! µs−2 ,βs−2 µµ m s−2m m−1
+ 2 [D + 2(s − 1)]!!
T δ δµµ δµβ δββ Tr[U n ]
m=1
[(s − 2)!]
(−1)m+1 2m − 1(m − 1)![D + 2(s − m − 2)]!!

µs−2 ,βs−2 µµ m s−2m−1 m n
+ Tr[T δ δµµ δµβ δββ γµ γβ U ] .
[(s − 2)!]2 [D + 2(s − 1)]!!
The γ traces and mixed traces can now be evaluated using

s−1 s−1 s−2
γ µ γµ δµβ = [D + 2(s − 1)]δµβ − γµ γβ δµβ , (11.168a)
m s−2m m−1 m−1 s−2m m−1
δ µµ δµµ δµβ δββ = [D + 2(s − m − 1)]δµµ δµβ δββ
m s−2m−2 m
+ 2m δµµ δµβ δββ , (11.168b)
m s−2m−1 m−1 m−1 s−2m−1 m−1
δ µµ δµµ δµβ δββ γµ γβ = [D + 2(s − m − 1)]δµµ δµβ δββ γµ γβ
m s−2m−2 m m s−2m−3 m
+ 4m δµµ δµβ δββ + 2m δµµ δµβ δββ γµ γβ , (11.168c)
which yields

) ~ 1 [s] 1
χ(F
s [nθ] = T − T [s−1]
Tr[U n ]
s! (s − 1)!
1 s−2
+ 2
Tr[T µs−1 ,βs−1 δµβ γµ γβ U n ]
[(s − 1)!] [D + 2(s − 1)]
D + 2(s − 2) s−2
− T µs−2 ,βs−2 δµβ Tr[U n ]
[(s − 2)!]2 [D + 2(s − 1)]
1 s−3
+ 2
Tr[T µs−2 ,βs−2 δµβ γµ γβ U n ]. (11.169)
[(s − 2)!] [D + 2(s − 1)]
Using (11.146) and the definition (11.91) of U , together with some careful counting,
one verifies that this expression matches [hs (J n ) − hs−1 (J n )] Tr[U n ], which was to be
proven.
11.B.2 Symmetric polynomials and SO(D) characters

In this part of the appendix we prove the second equality in (11.95), following essen-
tially the same steps as in appendix 11.A.2. We refer again to [398] for details, and we
write the components of weights in the dual basis of the Cartan subalgebra described
above (11.148) and (11.151). We will consider separately odd and even space-time
dimensions.
Odd D
The character of a half-spin representation of so(2r + 1) with a dominant highest
weight λ = (λ1 + 21 , λ2 + 12 , ..., λr + 21 ) is [399, p.258f]
r
!
|sin [(λ + r − i + 1) θ ]| |sin [(λi + r − i + 1) θj ]|
i j =
(2r+1)
Y
2 cos θ2i

χλ [θ1 , ..., θr ] = 1
.
sin r − i + 2 θj i=1
|sin [(r − i + 1) θj ]|
(11.170)
Owing to expression (11.92) for the trace of U n , the second equality in (11.95) is
equivalent to
|sin [(λi + r − i + 1) θj ]|
hs (J) − hs−1 (J) = (11.171)
|sin [(r − i + 1) θj ]|
for λi = sδi1 . To prove this, consider the difference of the bosonic character (11.150)
and the right-hand side of (11.171):
sin[(λi + r − i + 12 )θj ] |sin[(λi + r − i + 1)θj ]|
1
− . (11.172)
sin[(r − i + 2 )θj ] |sin[(r − i + 1)θj ]|
Introducing the notation

(Ar )ij = 2 sin [(r − i + 1)θj ] , (B r )ij = 2 cos (r − i + 21 )θj

(11.173)
and in terms of (11.156), this difference can be written as
Pr Pr
k=1 (−1)
k+1
sin[(s + r − 21 )θk ]Ar−1 [θk ] k=1 (−1)
k+1
2 sin[(s + r)θk ]Ar−1 [θk ]
−
|Ar | |Ar |
(11.174)
upon expanding the determinants along the first row. Now it turns out that5
r
Y r
Y
r r r r−1 r
2 |A | 2 cos (θi /2) = |A |, 2 |B | 2 cos (θi /2) = |B r |, (11.175)
i=1 i=1
and plugging this property in (11.174) one sees that (11.172) is just hs−1 (J) − hs−2 (J).
Since the first term of (11.172) equals hs (J)−hs−2 (J) by virtue of (11.150), this proves
(11.171).
Even D
The character of an irreducible representation of so(2r + 2) with (dominant) highest-
weight λ = (λ1 + 1/2, ..., λr+1 + 1/2) can be written as [399, p.258-259]
3
r+1 3

(2r+2) cos λ i + r − i + θ j
Y cos λi + r − i + θj
2
2 cos θ2i 2 ,

χλ [θ1 , ..., θr ] = = 3
|cos [(r − i + 1) θj ]| i=1
cos r − i + 2 θj
(11.176)
where we are including the possibility of a non-zero angle θr+1 (while in (11.95) we
take θr+1 = 0). Taking into account (11.92), proving the second equality in (11.95)
amounts to showing that
cos λi + r − i + 32 θj

hs (J) − hs−1 (J) = (11.177)
cos r − i + 23 θj

θr+1 =0
for λi = sδi1 . To prove this we proceed as in the odd-dimensional case: the difference
of the bosonic character (11.153) and the right-hand side of (11.177),
|cos[(λi + r − i + 1)θj ]| cos[(λi + r − i + 32 )θj ]

− , (11.178)
|cos[(r − i + 1)θj ]| θr+1 =0 cos[(r − i + 32 )θj ]
θr+1 =0
can be written as
 r+1 r+1 
k+1 r k+1 1 r
P P
(−1) cos[(s + r)θk ]B [θk ] (−1) 2 cos[(s + r + 2 )θk ]B [θk ]
 k=1 k=1

 − 
 |B r+1 | |B r+1 | 
θr+1 =0
(11.179)
upon expanding the determinants along the first row and using the notation (11.156)-
(11.173). One can then verify that this reduces to hs−1 (J) − hs−2 (J) by the same
argument as in the odd-dimensional case. By virtue of the second equality in (11.153),
this proves (11.177).
5
See e.g. [399, p.259].
Chapter 12
Conclusion
We have now completed our survey of the group-theoretic aspects of three-dimensional

gravity, and in particular of BMS symmetry in three dimensions. In this conclusion
we take one last look at what we have achieved.
Quantum symmetries
The overarching theme of this thesis has been group theory and its application to
quantum systems with symmetries. Accordingly, parts I and II of this thesis were
devoted to a broad overview of group representations and geometry. In particular we
have motivated and introduced central extensions, defined induced representations,
applied them to semi-direct products and relativistic symmetry groups, and explained
how classical mechanical systems with symmetries become symmetric quantum sys-
tems upon “replacing Poisson brackets by commutators”. We have also applied some of
these tools to the Virasoro group — the symmetry group of two-dimensional conformal
field theories — and used it to analyse certain properties of three-dimensional gravity
on Anti-de Sitter backgrounds.
BMS3 particles
Part III of the thesis was devoted to the application of group theoretic methods to
the study of asymptotically flat quantum gravity in three dimensions. In that context
our weapon of choice has been the BMS group in three dimensions, which we have
introduced as an asymptotic symmetry group in chapter 9, before working out its ab-
stract definition independently of gravity. We have seen in particular that it enjoys
an exceptional structure of the type G n g, where G is the Virasoro group spanned
by superrotations while g is its Lie algebra, spanned by supertranslations. A crucial
implication of this structure was that irreducible unitary representations of the BMS3
group, i.e. BMS3 particles, have supermomenta that span coadjoint orbits of the Vira-
soro group. This observation has allowed us to classify all such representations thanks
to the classification of Virasoro orbits exposed earlier, in chapter 7.
We also observed that supermomentum orbits have a straightforward interpretation

in gravity, since supermomenta coincide with Bondi mass aspects of asymptotically flat
space-time metrics. As a result we interpreted BMS3 particles in two equivalent ways:
381
Chapter 12. Conclusion 382
(i) as quantizations of orbits of asymptotically flat metrics under BMS3 transforma-

tions, and (ii) as relativistic particles dressed with gravitational boundary degrees
of freedom. These topological degrees of freedom are the three-dimensional analogue
of soft gravitons, so a BMS3 particle is effectively a particle dressed with soft gravitons.
As a confirmation of this interpretation, we evaluated BMS3 characters and showed

that they coincide with gravitational one-loop partition functions. On the group-
theoretic side this computation involves the Frobenius character formula (4.33), which
is essentially an integral of little group characters over a supermomentum orbit. Re-
markably, for non-zero angular potentials, we found that the integral localizes to a
single point on the supermomentum orbit, which allowed us to evaluate characters ex-
actly. On the field-theoretic side the one-loop partition function was evaluated using
heat kernel methods which turn out, unsurprisingly, to be more tractable in flat space
than in Anti-de Sitter space.
Higher spins and supergravity

The study of thermodynamics has led us into higher-spin theories, whose partition
functions in flat space could be computed with little extra effort compared to the
gravitational case. We have used this computation as an excuse to investigate the
unitary representations of asymptotic symmetry algebras that occur in that context,
with methods and results very similar to those of the purely gravitational setting. A
striking aspect of these considerations was the fact that it enabled us to compare ul-
trarelativistic and non-relativistic limits of conformal higher spins in a way that pure
gravity does not allow. In doing so we uncovered a sharp difference between the two
limits at the quantum level. As a corollary we concluded that flat space holography
should not be described as a Galilean conformal field theory.
We have similarly studied the supersymmetric generalization of BMS3 symmetry,

whose unitary representations are essentially super BMS3 multiplets consisting of an
infinite tower of BMS3 particles with increasing spins.
Is this quantum gravity?

In the introduction of the thesis we motivated the study of BMS symmetry by pre-
senting it as a way to tackle the quantization of gravity. This is a good moment to
ask to what extent this proposal has succeeded. To begin, we should realize that what
we have done is, indeed, a partial quantization of gravity: we have described almost
explicitly a family of Hilbert spaces endowed with operator algebras inherited from
gravitational symmetries, and we have used these Hilbert spaces to compute concrete
quantities such as partition functions. In this sense we have actually worked with a
partial version of quantum gravity.
This being said, one should not be overly enthusiastic about what we have achieved:
in essence we have worked out the flat space analogue of results that were mostly al-
ready known in the framework of AdS3 /CFT2 . In fact, in many cases our results were
flat limits of their AdS peers, although the BMS3 approach often led us to consider
slightly different questions and use somewhat different methods than those suggested
by conformal symmetries. Thus we have indeed made progress in our understanding
of flat space holography, but whether this opens new doors towards quantum gravity
is a whole other matter.
There are two simple arguments that show why our work is not quite quantum grav-
ity. First, what we have studied are irreducible unitary representations of asymptotic
symmetry groups, while realistic gravitational systems are expected to form highly
reducible representations. In essence, saying that we have studied quantum gravity
by studying irreducible representations of BMS3 would be tantamount to saying that
relativistic one-particle quantum mechanics is the same as quantum field theory, which
is of course untrue. Secondly, one should realize that our study of the symmetries of
gravity hasn’t taught us anything about the microscopic details of gravity itself. For
instance, the fact that the phase space of gravity forms the coadjoint representation of
the asymptotic symmetry group is merely a restatement of the fact that momentum
maps belong to the coadjoint representation, which is a robust feature of all symmetric
phase spaces; it does not tell us anything about the details of gravity.
A look forward
Our observations on BMS particles in three dimensions have allowed us to describe
dressed particles in a group-theoretic framework. While we haven’t described inter-
acting particles in this thesis, it is likely that our methods do apply to such cases as
well. In particular, describing scattering phenomena in terms of BMS particles instead
of standard (naked) particles should incorporate soft graviton contributions. In three
dimensions this could presumably be used to describe, say, the merger of two parti-
cles into a flat space cosmology; optimistically, BMS3 representations might then even
account for gravitational quantum corrections to such amplitudes! In four dimensions
the situation is much less well understood, for reasons that we alluded to earlier. In
that case the problem is much more basic, since the very definition of BMS symmetry
is elusive — let alone its quantum representations.
A related project is the description of BMS world lines. The reader may recall that
we described in chapter 5 a general procedure for building world line actions associ-
ated with arbitrary Lie groups, and that the application of this method to the Poincaré
group resulted in relativistic world lines. It is tempting to ask what happens when
that approach is applied to the BMS3 group; the answer is very natural: the resulting
action principle describes world lines propagating in the space of supertranslations,
which can equivalently be seen as relativistic world lines dressed with gravitational
degrees of freedom. Yet another way to think of these world-lines is to interpret them
as two-dimensional field theories dual to three-dimensional asymptotically flat grav-
ity, and indeed one finds that their partition functions coincide with gravitational
(one-loop) partition functions. These considerations should appear soon in a separate
publication [400]. Note that, as before, the application of these ideas to BMS in four
dimensions is much more problematic due to the lack of a proper definition of BMS4
symmetry.
There is undoubtedly much more to be done in the future, both in toy models such
as three-dimensional gravity and in real-world, four-dimensional systems. In the wake
of the experimental observation of gravitational waves [186], it is likely that new tools
and methods will soon be required to understand and study gravity, both classically
and quantum-mechanically. On a more philosophical note, it is remarkable that a
question seemingly as simple as “what is a particle?” has an answer as intricate and
rich as what we have been attempting to describe in this thesis. We hope to have
contributed to a partial solution to the problem, and look forward to investigating
some of its future applications.
All nature is but art, unknown to thee;

All chance, direction, which thou canst not see;
All discord, harmony not understood.
Alexander Pope, An Essay on Man, 1733.

Bibliography
[1] H. Bondi, “Gravitational Waves in General Relativity,” Nature 186 (1960), no. 4724,
535–535. (Cited on pages 2, 257, and 311.)
[2] H. Bondi, M. G. J. van der Burg, and A. W. K. Metzner, “Gravitational waves in
general relativity. 7. Waves from axisymmetric isolated systems,” Proc. Roy. Soc. Lond.
A269 (1962) 21–52. (Cited on pages 2, 4, 5, 257, and 311.)
[3] R. Sachs, “Asymptotic symmetries in gravitational theory,” Phys. Rev. 128 (1962)
2851–2864. (Cited on pages 2, 4, 5, 257, and 311.)
[4] R. K. Sachs, “Gravitational waves in general relativity. 8. Waves in asymptotically flat
space-times,” Proc. Roy. Soc. Lond. A270 (1962) 103–126. (Cited on pages 2, 4, 5,
257, and 311.)
[5] G. Barnich and C. Troessaert, “Symmetries of asymptotically flat 4 dimensional space-
times at null infinity revisited,” Phys. Rev. Lett. 105 (2010) 111103, 0909.2617. (Cited
on pages 6, 312, and 313.)
[6] G. Barnich and C. Troessaert, “Aspects of the BMS/CFT correspondence,” JHEP 05
(2010) 062, 1001.1541. (Cited on pages 6, 217, 230, 233, 234, 257, 265, 266, 312,
and 313.)
[7] T. Banks, “A Critique of pure string theory: Heterodox opinions of diverse dimensions,”
hep-th/0306074. (Cited on pages 6 and 312.)
[8] A. A. Belavin, A. M. Polyakov, and A. B. Zamolodchikov, “Infinite Conformal Symme-
try in Two-Dimensional Quantum Field Theory,” Nucl. Phys. B241 (1984) 333–380.
(Cited on page 6.)
[9] G. ’t Hooft, “Dimensional reduction in quantum gravity,” in Salamfest 1993:0284-296,
pp. 0284–296. 1993. gr-qc/9310026. (Cited on pages 6 and 312.)
[10] L. Susskind, “The World as a hologram,” J. Math. Phys. 36 (1995) 6377–6396,
hep-th/9409089. (Cited on pages 6 and 312.)
[11] J. D. Bekenstein, “Black holes and entropy,” Phys. Rev. D7 (1973) 2333–2346. (Cited
on pages 6 and 312.)
[12] S. W. Hawking, “Particle Creation by Black Holes,” Commun. Math. Phys. 43 (1975)
199–220. [,167(1975)]. (Cited on pages 6 and 312.)
[13] J. M. Maldacena, “The Large N limit of superconformal field theories and supergravity,”
Int. J. Theor. Phys. 38 (1999) 1113–1133, hep-th/9711200. [Adv. Theor. Math. Phys.
2, 231 (1998)]. (Cited on pages 6, 238, and 312.)
[14] J. D. Brown and M. Henneaux, “Central Charges in the Canonical Realization of
Asymptotic Symmetries: An Example from Three-Dimensional Gravity,” Commun.
Math. Phys. 104 (1986) 207–226. (Cited on pages 7, 9, 217, 225, and 229.)
385
Bibliography 386
[15] A. Strominger, “Black hole entropy from near horizon microstates,” JHEP 02 (1998)
009, hep-th/9712251. (Cited on page 7.)
[16] M. Bañados, C. Teitelboim, and J. Zanelli, “The Black hole in three-dimensional space-
time,” Phys. Rev. Lett. 69 (1992) 1849–1851, hep-th/9204099. (Cited on pages 7
and 217.)
[17] M. Bañados, M. Henneaux, C. Teitelboim, and J. Zanelli, “Geometry of the (2+1) black
hole,” Phys. Rev. D48 (1993) 1506–1525, gr-qc/9302012. [Erratum: Phys. Rev.D88,
069902 (2013)]. (Cited on pages 7, 217, and 238.)
[18] G. Barnich and C. Troessaert, “BMS charge algebra,” JHEP 12 (2011) 105, 1106.0213.
(Cited on pages 8, 17, and 313.)
[19] A. Strominger, “On BMS Invariance of Gravitational Scattering,” JHEP 07 (2014)
152, 1312.2229. (Cited on pages 8, 9, 253, 265, 307, and 312.)
[20] A. Ashtekar, J. Bicak, and B. G. Schmidt, “Asymptotic structure of symmetry reduced
general relativity,” Phys. Rev. D55 (1997) 669–686, gr-qc/9608042. (Cited on pages 8,
257, and 323.)
[21] F. C. Klein, “A comparative review of recent researches in geometry,” ArXiv e-prints
(July, 2008) 0807.3161. English translation by M. W. Haskell. (Cited on page 8.)
[22] E. P. Wigner, Gruppentheorie und ihre Anwendung auf die Quantenmechanik der Atom-
spektren. Springer, 1931. (Cited on pages 8 and 17.)
[23] E. P. Wigner, “On Unitary Representations of the Inhomogeneous Lorentz Group,”
Annals Math. 40 (1939) 149–204. [Reprint: Nucl. Phys. Proc. Suppl. 6, 9 (1989)].
(Cited on pages 9 and 71.)
[24] T. He, V. Lysov, P. Mitra, and A. Strominger, “BMS supertranslations and Weinberg’s
soft graviton theorem,” JHEP 05 (2015) 151, 1401.7026. (Cited on pages 9, 265,
and 312.)
[25] S. Weinberg, “Infrared photons and gravitons,” Phys. Rev. 140 (1965) B516–B524.
(Cited on page 9.)
[26] F. Cachazo and A. Strominger, “Evidence for a New Soft Graviton Theorem,”
1404.4091. (Cited on pages 9, 265, and 312.)
[27] D. Kapec, V. Lysov, S. Pasterski, and A. Strominger, “Semiclassical Virasoro symmetry
of the quantum gravity S-matrix,” JHEP 08 (2014) 058, 1406.3312. (Cited on pages 9,
265, and 312.)
[28] F. Bloch and A. Nordsieck, “Note on the Radiation Field of the electron,” Phys. Rev.
52 (1937) 54–59. (Cited on page 9.)
[29] P. A. M. Dirac, “Gauge invariant formulation of quantum electrodynamics,” Can. J.
Phys. 33 (1955) 650. (Cited on page 9.)
[30] V. Chung, “Infrared Divergence in Quantum Electrodynamics,” Phys. Rev. 140 (1965)
B1110–B1122. (Cited on page 9.)
[31] T. W. B. Kibble, “Coherent states and infrared divergences,” Lect. Theor. Phys. 11D
(1969) 387–478. (Cited on page 9.)
[32] T. W. B. Kibble, “Coherent soft-photon states and infrared divergences. ii. mass-shell
singularities of green’s functions,” Phys. Rev. 173 (1968) 1527–1535. (Cited on page 9.)
[33] T. W. B. Kibble, “Coherent soft-photon states and infrared divergences. iii. asymptotic
states and reduction formulas,” Phys. Rev. 174 (1968) 1882–1901. (Cited on page 9.)
Bibliography 387
[34] T. W. B. Kibble, “Coherent soft-photon states and infrared divergences. iv. the scat-
tering operator,” Phys. Rev. 175 (1968) 1624–1640. (Cited on page 9.)
[35] P. P. Kulish and L. D. Faddeev, “Asymptotic conditions and infrared divergences in
quantum electrodynamics,” Theor. Math. Phys. 4 (1970) 745. [Teor. Mat. Fiz. 4, 153
(1970)]. (Cited on page 9.)
[36] D. Zwanziger, “Reduction formulas for charged particles and coherent states in quan-
tum electrodynamics,” Phys. Rev. D7 (1973) 1082–1099. (Cited on page 9.)
[37] G. Barnich and G. Compère, “Classical central extension for asymptotic symmetries at
null infinity in three spacetime dimensions,” Class. Quant. Grav. 24 (2007) F15–F23,
gr-qc/0610130. (Cited on pages 9, 257, and 265.)
[38] P. J. McCarthy, “Representations of the Bondi-Metzner-Sachs Group. I. Deter-
mination of the Representations,” Proceedings of the Royal Society of London A:
Mathematical, Physical and Engineering Sciences 330 (1972), no. 1583, 517–535,
http://rspa.royalsocietypublishing.org/content/330/1583/517.full.pdf.
[39] P. J. McCarthy, “Structure of the Bondi-Metzner-Sachs Group,” Journal of Mathemat-
ical Physics 13 (1972), no. 11, 1837–1842. (Cited on pages 9, 311, and 312.)
[40] P. J. McCarthy, “Representations of the Bondi-Metzner-Sachs Group. II. Properties
and Classification of the Representations,” Proceedings of the Royal Society of London
A: Mathematical, Physical and Engineering Sciences 333 (1973), no. 1594, 317–336,
[41] P. J. McCarthy, “Asymptotically flat space-times and elementary particles,” Phys. Rev.
Lett. 29 (1972) 817–819. (Cited on pages 9 and 311.)
[42] J. Rotman, An Introduction to the Theory of Groups. Graduate Texts in Mathematics.
Springer New York, 1999. (Cited on page 10.)
[43] R. Abraham and J. Marsden, Foundations of Mechanics. AMS Chelsea publishing.
AMS Chelsea Pub./American Mathematical Society, 1978. (Cited on pages 10, 39,
97, 106, 107, 110, and 151.)
[44] J. Lee, Introduction to Smooth Manifolds. Graduate Texts in Mathematics. Springer,
2003. (Cited on page 10.)
[45] G. Barnich and B. Oblak, “Holographic positive energy theorems in three-dimensional
gravity,” Class. Quant. Grav. 31 (2014) 152001, 1403.3835. (Cited on pages 11, 217,
241, and 257.)
[46] G. Barnich and B. Oblak, “Notes on the BMS group in three dimensions: II. Coadjoint
representation,” JHEP 03 (2015) 033, 1502.00010. (Cited on pages 11, 127, 217, 257,
280, 293, and 312.)
[47] G. Barnich and B. Oblak, “Notes on the BMS group in three dimensions: I. Induced
representations,” JHEP 06 (2014) 129, 1403.5803. (Cited on pages 11, 257, and 293.)
[48] B. Oblak, “Characters of the BMS Group in Three Dimensions,” Commun. Math. Phys.
340 (2015), no. 1, 413–432, 1502.03108. (Cited on pages 12, 78, 293, 312, and 324.)
[49] G. Barnich, H. A. González, A. Maloney, and B. Oblak, “One-loop partition function of
three-dimensional flat gravity,” JHEP 04 (2015) 178, 1502.06185. (Cited on pages 12,
293, 324, 331, 340, and 341.)
Bibliography 388
[50] A. Campoleoni, H. A. González, B. Oblak, and M. Riegler, “Rotating Higher Spin Par-
tition Functions and Extended BMS Symmetries,” JHEP 04 (2016) 034, 1512.03353.
(Cited on pages 12, 78, 293, 324, and 331.)
[51] A. Campoleoni, H. A. González, B. Oblak, and M. Riegler, “BMS Modules in Three
Dimensions,” Int. J. Mod. Phys. A31 (2016), no. 12, 1650068, 1603.03812. (Cited on
pages 12, 293, 314, and 331.)
[52] H. Afshar, S. Detournay, D. Grumiller, and B. Oblak, “Near-Horizon Geometry and
Warped Conformal Symmetry,” JHEP 03 (2016) 187, 1512.08233. (Cited on pages 12,
64, 162, 163, and 164.)
[53] S. Weinberg, The Quantum Theory of Fields. Cambridge University Press, 1995. (Cited
on pages 15, 17, 61, 71, and 92.)
[54] V. Ovsienko and S. Tabachnikov, Projective Differential Geometry Old and New: From
the Schwarzian Derivative to the Cohomology of Diffeomorphism Groups. Cambridge
Tracts in Mathematics. Cambridge University Press, 2004. (Cited on pages 15, 159,
and 166.)
[55] B. Khesin and R. Wendt, The Geometry of Infinite-Dimensional Groups. A series of
modern surveys in mathematics. Springer Berlin Heidelberg, 2008. (Cited on pages 15,
97, 105, 145, 151, 156, and 179.)
[56] L. Guieu and C. Roger, L’algèbre et le groupe de Virasoro. Publications du CRM,
Université de Montréal, 2007. (Cited on pages 15, 29, 145, 151, 156, 158, 159, 162,
163, 171, 179, 196, 207, and 363.)
[57] G. Tuynman and W. Wiegerinck, “Central extensions and physics,” Journal of Geom-
etry and Physics 4 (1987), no. 2, 207 – 258. (Cited on page 15.)
[58] A. Weinstein, “Groupoids: Unifying Internal and External Symmetry - A Tour Through
Some Examples,” Notices of the AMS 43 (1996) 744–752. (Cited on page 17.)
[59] M. Crainic and R. L. Fernandes, “Lectures on integrability of Lie brackets,” Lectures
on Poisson Geometry, Geom. Topol. Monogr. 17 (2011) 1–107. (Cited on page 17.)
[60] G. Barnich, “A Note on gauge systems from the point of view of Lie algebroids,” AIP
Conf. Proc. 1307 (2010) 7–18, 1010.0899. (Cited on pages 17, 232, and 264.)
[61] M. Peskin and D. Schroeder, An Introduction to Quantum Field Theory. Advanced
book classics. Addison-Wesley Publishing Company, 1995. (Cited on page 17.)
[62] A. Barut and R. Rączka, Theory of Group Representations and Applications. World
Scientific, 1986. (Cited on pages 32, 40, 46, 47, 55, 57, 58, 61, 64, 70, 71, and 314.)
[63] G. W. Mackey, Induced representations of groups and quantum mechanics. Publicazioni
della Classe di Scienze della Scuola Normale Superiore di Pisa. W. A. Benjamin, 1968.
[64] W. Rudin, Principles of Mathematical Analysis. International series in pure and applied
mathematics. McGraw-Hill, 1976. (Cited on page 32.)
[65] W. Rudin, Real and complex analysis. Mathematics series. McGraw-Hill, 1987. (Cited
on page 32.)
[66] H. Royden and P. Fitzpatrick, Real Analysis. Featured Titles for Real Analysis Series.
Prentice Hall, 2010. (Cited on page 37.)
[67] G. Barnich, F. Brandt, and M. Henneaux, “Local BRST cohomology in gauge theories,”
Phys. Rept. 338 (2000) 439–569, hep-th/0002245. (Cited on page 43.)
Bibliography 389
[68] M. Nakahara, Geometry, Topology and Physics. Graduate student series in physics.
Taylor & Francis, 2003. (Cited on page 44.)
[69] J. R. David, M. R. Gaberdiel, and R. Gopakumar, “The Heat Kernel on AdS(3) and
its Applications,” JHEP 04 (2010) 125, 0911.5085. (Cited on pages 46, 331, 346,
and 364.)
[70] M. R. Gaberdiel, R. Gopakumar, and A. Saha, “Quantum W -symmetry in AdS3 ,”
JHEP 02 (2011) 004, 1009.6087. (Cited on pages 46, 331, 335, 338, 340, and 346.)
[71] R. Camporesi, “Harmonic analysis and propagators on homogeneous spaces,” Phys.
Rept. 196 (1990) 1–134. (Cited on page 46.)
[72] R. Camporesi and A. Higuchi, “On the Eigenfunctions of the Dirac operator on spheres
and real hyperbolic spaces,” J. Geom. Phys. 20 (1996) 1–18, gr-qc/9505009. (Cited
on page 46.)
[73] P. K. Aravind, “The Wigner angle as an anholonomy in rapidity space,” American
Journal of Physics 65 (1997), no. 7, 634–636. (Cited on page 47.)
[74] M. S. Williamson, M. Ericsson, M. Johansson, E. Sjoqvist, A. Sudbery, and V. Vedral,
“Global asymmetry of many-qubit correlations: A lattice gauge theory approach,”
Phys. Rev. A84 (2011) 032302, 1102.5609. (Cited on page 47.)
[75] J.-P. Antoine, Irreversibility and Causality Semigroups and Rigged Hilbert Spaces: A
Selection of Articles Presented at the 21st International Colloquium on Group Theoret-
ical Methods in Physics (ICGTMP) at Goslar, Germany, July 16–21, 1996, ch. Quan-
tum mechanics beyond Hilbert space, pp. 1–33. Springer Berlin Heidelberg, Berlin,
Heidelberg, 1998. (Cited on page 50.)
[76] R. de la Madrid, “The role of the rigged Hilbert space in quantum mechanics,” European
Journal of Physics 26 (Apr., 2005) 287–312, quant-ph/0502053. (Cited on page 50.)
[77] J. Dieudonné and P. Dugac, Abrégé d’histoire des mathématiques: 1700-1900. Her-
mann, 1986. (Cited on page 52.)
[78] T. Gowers, J. Barrow-Green, and I. Leader, The Princeton Companion to Mathematics.
Princeton University Press, 2010. (Cited on page 52.)
[79] S. Sternberg, Group Theory and Physics. Cambridge University Press, 1995. (Cited
[80] H. Joos and R. Schrader, “On the primitive characters of the Poincaré group,” Comm.
Math. Phys. 7 (1968), no. 1, 21–50. (Cited on pages 52 and 78.)
[81] A. A. Kirillov, Elements of the Theory of Representations. Grundlehren der mathema-
tischen Wissenschaften. Springer-Verlag, 1976. (Cited on page 52.)
[82] M. F. Atiyah and R. Bott, “A Lefschetz Fixed Point Formula for Elliptic Complexes: II.
Applications,” Annals of Mathematics 88 (1968), no. 3, 451–491. (Cited on page 54.)
[83] M. F. Atiyah and R. Bott, “A Lefschetz Fixed Point Formula for Elliptic Complexes:
I,” Annals of Mathematics 86 (1967), no. 2, 374–407. (Cited on page 54.)
[84] G. W. Mackey, “On Induced Representations of Groups,” American Journal of Math-
ematics 73 (1951), no. 3, 576–592. (Cited on page 55.)
[85] G. W. Mackey, “Induced Representations of Locally Compact Groups I,” Annals of
Mathematics 55 (1952), no. 1, 101–139. (Cited on page 55.)
[86] G. W. Mackey, “Induced Representations of Locally Compact Groups II. The Frobenius
Reciprocity Theorem,” Annals of Mathematics 58 (1953), no. 2, 193–221. (Cited on
page 55.)
Bibliography 390
[87] E. Kaniuth and K. F. Taylor, Induced Representations of Locally Compact Groups.

Cambridge University Press, 2012. Cambridge Books Online. (Cited on page 55.)
[88] V. Varadarajan, Geometry of Quantum Theory. Springer New York, 2007. (Cited on
page 55.)
[89] E. Galois, “Lettre à Auguste Chevalier,” May, 1832. Available at http://www.galois.
ihp.fr/ressources/vie-et-oeuvre-de-galois/lettres/lettre-testament/.
(Cited on page 57.)
[90] B. Ørsted, “Induced representations and a new proof of the imprimitivity theorem,”
Journal of Functional Analysis 31 (1979), no. 3, 355 – 359. (Cited on page 57.)
[91] J. Cornwell, Group Theory in Physics. Group Theory in Physics. Academic Press,
1984. (Cited on pages 61, 71, and 74.)
[92] X. Bekaert and N. Boulanger, “The Unitary representations of the Poincare group
in any spacetime dimension,” in 2nd Modave Summer School in Theoretical Physics
Modave, Belgium, August 6-12, 2006. 2006. hep-th/0611263. (Cited on pages 61,
71, and 74.)
[93] S. Detournay, T. Hartman, and D. M. Hofman, “Warped Conformal Field Theory,”
Phys. Rev. D86 (2012) 124018, 1210.0539. (Cited on pages 64, 162, 163, and 273.)
[94] V. Bargmann and E. P. Wigner, “Group Theoretical Discussion of Relativistic Wave
Equations,” Proceedings of the National Academy of Sciences 34 (1948), no. 5, 211–223,
(Cited on page 71.)
http://www.pnas.org/content/34/5/211.full.pdf.
[95] M. Henneaux, “Groupes et représentations: I. Groupe des rotations à 3 dimensions,
groupe de Lorentz et groupe de Poincaré.” ULB, 2009. Available at http://www.ulb.
ac.be/sciences/ptm/pmif/membres/notescours.html. (Cited on pages 72 and 87.)
[96] J. Jackson, Classical electrodynamics. Wiley, 1975. (Cited on pages 76 and 83.)
[97] P. Schuster and N. Toro, “On the Theory of Continuous-Spin Particles: Wavefunctions
and Soft-Factor Scattering Amplitudes,” JHEP 09 (2013) 104, 1302.1198. (Cited on
page 78.)
[98] P. Schuster and N. Toro, “On the Theory of Continuous-Spin Particles: Helicity Cor-
respondence in Radiation and Forces,” JHEP 09 (2013) 105, 1302.1577. (Cited on
page 78.)
[99] P. Schuster and N. Toro, “A Gauge Field Theory of Continuous-Spin Particles,” JHEP
10 (2013) 061, 1302.3225. (Cited on page 78.)
[100] H. X. Nghiem, “Harmonic analysis on the Poincaré group. I. Generalized matrix ele-
ments,” Comm. Math. Phys. 12 (1969), no. 4, 331–350. (Cited on page 78.)
[101] A. Garbarz and M. Leston, “Quantization of BMS3 orbits: a perturbative approach,”
Nucl. Phys. B906 (2016) 133–146, 1507.00339. (Cited on page 78.)
[102] A. Peres, P. F. Scudo, and D. R. Terno, “Quantum Entropy and Special Relativity,”
Phys. Rev. Lett. 88 (May, 2002) 230402. (Cited on pages 83 and 84.)
[103] V. Palge, Relativistic entanglement of single and two particle systems. PhD thesis,
[104] P. M. Alsing and G. J. Milburn, “Lorentz Invariance of Entanglement,”
quant-ph/0203051. (Cited on page 83.)
[105] D. Ahn, H.-j. Lee, Y. H. Moon, and S. W. Hwang, “Relativistic entanglement and
Bell’s inequality,” Phys. Rev. A 67 (Jan, 2003) 012103. (Cited on page 83.)
Bibliography 391
[106] L. Thomas, “The motion of the spinning electron,” Nature 117 (April, 1926) 514–514.
(Cited on page 83.)
[107] B. Binegar, “Relativistic Field Theories in Three-dimensions,” J. Math. Phys. 23 (1982)
1511–1517. (Cited on page 85.)
[108] D. R. Grigore, “The Projective unitary irreducible representations of the Poincaré
group in (1+2)-dimensions,” J. Math. Phys. 34 (1993) 4172–4189, hep-th/9304142.
(Cited on page 85.)
[109] B. Oblak, “From the Lorentz Group to the Celestial Sphere,” Notes de la Septième
BSSM, U.L.B. (2015). 1508.00920. (Cited on pages 87 and 260.)
[110] J. M. Leinaas and J. Myrheim, “On the theory of identical particles,” Nuovo Cim. B37
(1977) 1–23. (Cited on page 89.)
[111] F. Wilczek, “Quantum Mechanics of Fractional Spin Particles,” Phys. Rev. Lett. 49
(1982) 957–959. (Cited on page 89.)
[112] J. de Azcárraga and J. Izquierdo, Lie Groups, Lie Algebras, Cohomology and Some
Applications in Physics. Cambridge Monographs on Mathematical Physics. Cambridge
University Press, 1998. (Cited on page 90.)
[113] S. K. Bose, “The galilean group in 2+1 space-times and its central extension,” Com-
munications in Mathematical Physics 169 no. 2, 385–395. (Cited on pages 90 and 91.)
[114] E. Inönü and E. P. Wigner, “Representations of the Galilei group,” Il Nuovo Cimento
(1943-1954) 9 (2007), no. 8, 705–718. (Cited on page 90.)
[115] E. Inönü and E. P. Wigner, “On the contraction of groups and their representations,”
Proc. Nat. Acad. Sci. 39 (1953) 510–524. (Cited on pages 92 and 289.)
[116] N. Woodhouse, Geometric Quantization. Oxford mathematical monographs. Clarendon
Press, 1997. (Cited on pages 97, 111, and 117.)
[117] A. A. Kirillov, Lectures on the Orbit Method. Graduate studies in mathematics. Amer-
ican Mathematical Society, 2004. (Cited on pages 97, 111, and 117.)
[118] S. De Buyl, S. Detournay, and Y. Voglaire, “Symplectic Geometry and Ge-
ometric Quantization,”. Available at http://www.ulb.ac.be/sciences/ptm/pmif/
Rencontres/geosymplgeoquant4.pdf. (Cited on pages 97, 111, and 117.)
[119] V. I. Arnold, Mathematical Methods of Classical Mechanics. Graduate Texts in Math-
ematics. Springer New York, 1997. (Cited on page 107.)
[120] B. Kostant, Lectures in Modern Analysis and Applications III, ch. Quantization and
unitary representations, pp. 87–208. Springer Berlin Heidelberg, Berlin, Heidelberg,
[121] E. Witten, “Coadjoint Orbits of the Virasoro Group,” Commun. Math. Phys. 114
(1988) 1. (Cited on pages 179, 242, and 245.)
[122] A. Alekseev, L. Faddeev, and S. Shatashvili, “Quantization of symplectic orbits of
compact Lie groups by means of the functional integral,” Journal of Geometry and
Physics 5 (1988), no. 3, 391 – 406. (Cited on pages 118 and 124.)
[123] A. Alekseev and S. Shatashvili, “Path integral quantization of the coadjoint orbits of
the Virasoro group and 2-d gravity,” Nuclear Physics B 323 (1989), no. 3, 719–733.
(Cited on page 118.)
[124] H. Aratyn, E. Nissimov, S. Pacheva, and A. H. Zimerman, “Symplectic actions on
coadjoint orbits,” Phys. Lett. B240 (1990) 127. (Cited on pages 118, 120, and 124.)
Bibliography 392
[125] P. Salgado Rebolledo, Symplectic Structure of Constrained Systems: Gribov Ambiguity

and Classical Duals for 3D Gravity. PhD thesis, Concepcion U., 2015-10-28. (Cited
on page 118.)
[126] L. Charles, “Feynman path integral and Toeplitz quantization,” Helvetica Physica Acta
72 (1999), no. 5/6, 341–355. (Cited on page 118.)
[127] J. Wess and B. Zumino, “Consequences of anomalous Ward identities,” Phys. Lett.
B37 (1971) 95–97. (Cited on page 120.)
[128] S. P. Novikov, “The Hamiltonian formalism and a many valued analog of Morse theory,”
Usp. Mat. Nauk 37N5 (1982), no. 5, 3–49. [Russ. Math. Surveys 37, no.5, 1 (1982)].
[129] E. Witten, “Global Aspects of Current Algebra,” Nucl. Phys. B223 (1983) 422–432.
[130] W. Taylor, “Virasoro representations on DiffS 1 /S 1 coadjoint orbits,” hep-th/9204091.
[131] W. Taylor, Coadjoint orbits and conformal field theory. PhD thesis, UC, Berkeley,
1993. hep-th/9310040. (Cited on pages 124 and 245.)
[132] J. H. Rawnsley, “Representations of a semi-direct product by quantization,” Mathemat-
ical Proceedings of the Cambridge Philosophical Society 78 (9, 1975) 345–350. (Cited
[133] P. Baguis, “Semidirect products and the Pukanszky condition,” Journal of Geometry
and Physics 25 (1998), no. 3-4, 245 – 270. (Cited on pages 127, 133, and 135.)
[134] Z. Li, Coadjoint orbits and induced representations. PhD thesis, Massachusetts Insti-
tute of Technology, 1993. (Cited on pages 127, 135, and 136.)
[135] V. Guillemin and S. Sternberg, Symplectic Techniques in Physics, vol. 1. Cambridge
University Press, May, 1990. (Cited on page 127.)
[136] S. T. Ali, J. P. Antoine, and J.-P. Gazeau, Coherent States, Wavelets and Their Gen-
eralizations, vol. 1. Springer, Jan, 2000. (Cited on page 127.)
[137] C. Duval, J. Elhadad, M. J. Gotay, J. Sniatycki, and G. M. Tuynman, “Quantization
and bosonic BRST theory,” Annals Phys. 206 (1991) 1. (Cited on pages 135 and 136.)
[138] R. Cushman and W. van der Kallen, “Adjoint and Coadjoint Orbits of the Poincaré
Group,” Acta Applicandae Mathematica 90 (2006), no. 1-2, 65–89. (Cited on page 138.)
[139] V. Hudon and S. Twareque Ali, “Coadjoint Orbits of the Poincaré Group in 2+1
Dimensions and their Coherent States,” ArXiv e-prints (Nov., 2010) 1011.6620. (Cited
on page 138.)
[140] R. Cushman and W. van der Kallen, “A new interpretation for the mass of a classical
relativistic particle,” Differential Geometry and its Applications 24 (2006), no. 3, 230
– 234. (Cited on page 138.)
[141] A. P. Balachandran, G. Marmo, B. S. Skagerstam, and A. Stern, “Gauge Symmetries
and Fiber Bundles: Applications to Particle Dynamics,” Lect. Notes Phys. 188 (1983)
1–140. (Cited on page 138.)
[142] C. Duval and J. Elhadad, “Geometric quantization and localization of relativistic spin
systems,” in Joint Summer Reasearch Conference on Mathematical Aspects of Classical
Field Theory Seattle, Washington, July 21-25, 1991. 1991. (Cited on page 138.)
[143] J.-M. Souriau, “A mechanistic description of elementary particles,” in Structure of
Dynamical Systems, pp. 173–193. Springer, 1997. (Cited on page 138.)
Bibliography 393
[144] C. Teitelboim, “Quantum Mechanics of the Gravitational Field,” Phys. Rev. D25
(1982) 3159. (Cited on pages 138 and 140.)
[145] M. Henneaux and C. Teitelboim, “Relativistic Quantum Mechanics of Supersymmetric
Particles,” Annals Phys. 143 (1982) 127. (Cited on pages 138 and 140.)
[146] M. Henneaux and C. Teitelboim, “First and second quantized point particles of any
spin,” in 2nd Meeting on Quantum Mechanics of Fundamental Systems (CECS) San-
tiago, Chile, December 17-20, 1987. 1987. (Cited on pages 138 and 140.)
[147] A. Yu. Alekseev and S. L. Shatashvili, “Propagator for the Relativistic Spinning Parti-
cle via Functional Integral Over Trajectories,” Mod. Phys. Lett. A3 (1988) 1551–1559.
[148] R. P. Feynman, “An Operator Calculus Having Applications in Quantum Electrody-
namics,” Phys. Rev. 84 (Oct, 1951) 108–128. (Cited on page 140.)
[149] M. J. Strassler, “Field theory without Feynman diagrams: One loop effective actions,”
Nucl. Phys. B385 (1992) 145–184, hep-ph/9205205. (Cited on page 140.)
[150] M. J. Strassler, The Bern-Kosower rules and their relation to quantum field theory.
PhD thesis, Stanford U., Phys. Dept., 1993. (Cited on page 140.)
[151] M. Henneaux and C. Teitelboim, Quantization of gauge systems. Princeton University
Press, 1992. (Cited on page 141.)
[152] P. H. Ginsparg, “Applied conformal field theory,” in Les Houches Summer School in
Theoretical Physics: Fields, Strings, Critical Phenomena Les Houches, France, June
28-August 5, 1988. 1988. hep-th/9108028. (Cited on pages 145 and 174.)
[153] P. Di Francesco, P. Mathieu, and D. Senechal, Conformal Field Theory. Graduate Texts
in Contemporary Physics. Springer-Verlag, New York, 1997. (Cited on pages 145, 174,
and 247.)
[154] R. Blumenhagen and E. Plauschinn, Introduction to Conformal Field Theory: With
Applications to String Theory. Lecture Notes in Physics. Springer Berlin Heidelberg,
2009. (Cited on pages 145 and 174.)
[155] A. Kriegl and P. W. Michor, The convenient setting of global analysis. No. 53. American
Mathematical Soc., 1997. (Cited on page 146.)
[156] J. Lurie, “Topics in Geometric topology.” MIT, 2009. Available at http://www.math.
harvard.edu/~lurie/937.html. (Cited on page 149.)
[157] G. Barnich, G. Giribet, and M. Leston, “Chern-Simons action for inhomogeneous Vi-
rasoro group as extension of three dimensional flat gravity,” J. Math. Phys. 56 (2015),
no. 7, 071701, 1505.02031. (Cited on page 159.)
[158] R. Bott, “On the characteristic classes of groups of diffeomorphisms,” Enseign. Math
23 (1977), no. 3-4, 209–220. (Cited on page 161.)
[159] L. Goncharova, “Cohomology of Lie algebras of formal vector fields on the line,” Uspekhi
Matematicheskikh Nauk 27 (1972), no. 5, 231–232. (Cited on page 163.)
[160] D. Fuks, “Cohomology of infinite-dimensional Lie algebras,” Sov. Math., Consultants
Bureau, New York (1986). (Cited on page 163.)
[161] V. F. Lazutkin and T. F. Pankratova, “Normal forms and versal deformations for Hill’s
equation,” Functional Analysis and Its Applications 9 (October, 1975) 306–311. (Cited
on page 179.)
Bibliography 394
[162] A. A. Kirillov, “Orbits of the group of diffeomorphisms of a circle and local Lie super-
algebras,” Functional Analysis and Its Applications 15 (April, 1981) 135–137. (Cited
on page 179.)
[163] G. Segal, “Unitary representations of some infinite-dimensional groups,” Comm. Math.
Phys. 80 (1981), no. 3, 301–342. (Cited on pages 179 and 246.)
[164] I. Bakas, “Conformal invariance, the KdV equation and coadjoint orbits of the Virasoro
algebra,” Nuclear Physics B 302 (1988), no. 2, 189 – 203. (Cited on page 179.)
[165] F. Gay-Balmaz, “On the classification of the coadjoint orbits of the Sobolev Bott-
Virasoro group,” Journal of Functional Analysis 256 (2009), no. 9, 2815 – 2841. (Cited
on page 179.)
[166] J. Balog, L. Fehér, and L. Palla, “Coadjoint orbits of the Virasoro algebra and the
global Liouville equation,” Int. J. Mod. Phys. A13 (1998) 315–362, hep-th/9703045.
(Cited on pages 179, 201, 204, 206, 207, 208, 213, and 214.)
[167] G. Barnich, L. Donnay, J. Matulich, and R. Troncoso, “Asymptotic symmetries and
dynamics of three-dimensional flat supergravity,” JHEP 08 (2014) 071, 1407.4275.
(Cited on pages 189, 284, 361, 364, 365, and 366.)
[168] S. Nag and A. Verjovsky, “Diff(S 1 ) and the Teichmüller spaces,” Comm. Math. Phys.
130 (1990), no. 1, 123–138. (Cited on page 196.)
[169] R. Schwartz, “A projectively natural flow for circle diffeomorphisms,” Inventiones math-
ematicae 110 (December, 1992) 627–647. (Cited on page 208.)
[170] R. Schwartz, “On the Integral Curve of a Linear Third Order O.D.E.,” Journal of
Differential Equations 135 (1997), no. 2, 183 – 191. (Cited on page 208.)
[171] S. Tabachnikov, “Variations on R. Schwartz’s inequality for the Schwarzian derivative,”
ArXiv e-prints (June, 2010) 1006.1339. (Cited on page 208.)
[172] R. P. Geroch, “Structure of the gravitational field at spatial infinity,” J. Math. Phys.
13 (1972) 956–968. (Cited on page 217.)
[173] R. Geroch, Asymptotic Structure of Space-Time, ch. Asymptotic Structure of Space-
Time, pp. 1–105. Springer US, Boston, MA, 1977. (Cited on pages 217 and 311.)
[174] A. Ashtekar, “Asymptotic Structure of the Gravitational Field at Spatial Infinity,” in
General Relativity and Gravitation II, A. Held, ed., vol. 2, p. 37. 1980. (Cited on
page 217.)
[175] L. Abbott and S. Deser, “Charge definition in non-abelian gauge theories,” Physics
Letters B 116 (1982), no. 4, 259 – 263. (Cited on page 217.)
[176] G. Barnich and F. Brandt, “Covariant theory of asymptotic symmetries, conservation
laws and central charges,” Nucl. Phys. B633 (2002) 3–82, hep-th/0111246. (Cited
on pages 217, 221, 223, 224, 232, and 264.)
[177] G. Barnich, “Boundary charges in gauge theories: Using Stokes theorem in the bulk,”
Class. Quant. Grav. 20 (2003) 3685–3698, hep-th/0301039. (Cited on page 217.)
[178] G. Barnich and G. Compère, “Surface charge algebra in gauge theories and thermody-
namic integrability,” J. Math. Phys. 49 (2008) 042901, 0708.2378. (Cited on pages 217
and 223.)
[179] M. Bañados, “Three-dimensional quantum geometry and black holes,”
hep-th/9901148. [AIP Conf. Proc. 484, 147 (1999)]. (Cited on pages 217,
233, and 236.)
Bibliography 395
[180] K. Skenderis and S. N. Solodukhin, “Quantum effective action from the AdS / CFT cor-
respondence,” Phys. Lett. B472 (2000) 316–322, hep-th/9910023. (Cited on pages 217
and 233.)
[181] A. Garbarz and M. Leston, “Classification of Boundary Gravitons in AdS3 Gravity,”
JHEP 05 (2014) 141, 1403.3367. (Cited on pages 217, 240, and 241.)
[182] T. Nakatsu, H. Umetsu, and N. Yokoi, “Three-dimensional black holes and Liouville
field theory,” Prog. Theor. Phys. 102 (1999) 867–896, hep-th/9903259. (Cited on
pages 217 and 241.)
[183] J. Navarro-Salas and P. Navarro, “Virasoro orbits, AdS(3) quantum gravity and en-
tropy,” JHEP 05 (1999) 009, hep-th/9903248. (Cited on pages 217 and 241.)
[184] A. Maloney and E. Witten, “Quantum Gravity Partition Functions in Three Dimen-
sions,” JHEP 02 (2010) 029, 0712.0155. (Cited on pages 217, 241, and 252.)
[185] C. Teitelboim, “How commutators of constraints reflect the space-time structure,” An-
nals Phys. 79 (1973) 542–557. (Cited on page 219.)
[186] Virgo, LIGO Scientific Collaboration, B. P. Abbott et al., “Observation of Gravi-
tational Waves from a Binary Black Hole Merger,” Phys. Rev. Lett. 116 (2016), no. 6,
061102, 1602.03837. (Cited on pages 219 and 384.)
[187] S. Deser, R. Jackiw, and S. Templeton, “Topologically Massive Gauge Theories,” An-
nals Phys. 140 (1982) 372–411. [Annals Phys.281,409(2000)]. (Cited on page 219.)
[188] E. A. Bergshoeff, O. Hohm, and P. K. Townsend, “Massive Gravity in Three Dimen-
sions,” Phys. Rev. Lett. 102 (2009) 201301, 0901.1766. (Cited on page 219.)
[189] R. Penrose, Group Theory in Non-Linear Problems: Lectures Presented at the NATO
Advanced Study Institute on Mathematical Physics, held in Istanbul, Turkey, August
7–18, 1972, ch. Relativistic Symmetry Groups, pp. 1–58. Springer Netherlands, Dor-
drecht, 1974. (Cited on page 221.)
[190] J. W. York, “Role of Conformal Three-Geometry in the Dynamics of Gravitation,”
Phys. Rev. Lett. 28 (Apr, 1972) 1082–1085. (Cited on page 221.)
[191] G. W. Gibbons and S. W. Hawking, “Action integrals and partition functions in quan-
tum gravity,” Phys. Rev. D 15 (May, 1977) 2752–2756. (Cited on pages 221 and 232.)
[192] A. Achucarro and P. K. Townsend, “A Chern-Simons Action for Three-Dimensional
anti-De Sitter Supergravity Theories,” Phys. Lett. B180 (1986) 89. (Cited on
page 221.)
[193] E. Witten, “(2+1)-Dimensional Gravity as an Exactly Soluble System,” Nucl. Phys.
B311 (1988) 46. (Cited on page 221.)
[194] M. Blagojevic, Gravitation and Gauge Symmetries. Series in High Energy Physics,
Cosmology and Gravitation. CRC Press, 2001. (Cited on page 221.)
[195] G. Lucena Gómez, “Higher-Spin Theories - Part II : enter dimension three,” PoS Mo-
daveVIII (2012) 003, 1307.3200. (Cited on page 221.)
[196] L. Donnay, “Asymptotic dynamics of three-dimensional gravity,” PoS Modave2015
(2016) 001, 1602.09021. (Cited on page 221.)
[197] O. Coussaert, M. Henneaux, and P. van Driel, “The Asymptotic dynamics of three-
dimensional Einstein gravity with a negative cosmological constant,” Class. Quant.
Grav. 12 (1995) 2961–2966, gr-qc/9506019. (Cited on page 221.)
Bibliography 396
[198] G. Barnich and H. A. González, “Dual dynamics of three dimensional asymptotically

flat Einstein gravity at null infinity,” JHEP 05 (2013) 016, 1303.1075. (Cited on
pages 221, 265, 284, and 312.)
[199] G. Compère, Symmetries and conservation laws in Lagrangian gauge theories with
applications to the mechanics of black holes and to gravity in three dimensions. PhD
thesis, Brussels U., 2007. 0708.3153. (Cited on pages 223 and 224.)
[200] K. Skenderis, M. Taylor, and B. C. van Rees, “Topologically Massive Gravity and the
AdS/CFT Correspondence,” JHEP 09 (2009) 045, 0906.4926. (Cited on pages 230
and 235.)
[201] C. Fefferman and C. R. Graham, “The ambient metric,” 0710.0919. (Cited on
page 230.)
[202] J. W. York, “Role of Conformal Three-Geometry in the Dynamics of Gravitation,”
Phys. Rev. Lett. 28 (Apr, 1972) 1082–1085. (Cited on page 232.)
[203] J. D. Brown and J. W. York, “Quasilocal energy and conserved charges derived from
the gravitational action,” Phys. Rev. D 47 (Feb, 1993) 1407–1419. (Cited on page 232.)
[204] M. Henningson and K. Skenderis, “Holography and the Weyl anomaly,” Fortsch. Phys.
48 (2000) 125–128, hep-th/9812032. (Cited on page 232.)
[205] V. Balasubramanian and P. Kraus, “A Stress tensor for Anti-de Sitter gravity,” Com-
mun. Math. Phys. 208 (1999) 413–428, hep-th/9902121. (Cited on pages 232
and 238.)
[206] S. de Haro, K. Skenderis, and S. N. Solodukhin, “Gravity in warped compactifica-
tions and the holographic stress tensor,” Class. Quant. Grav. 18 (2001) 3171–3180,
hep-th/0011230. [,495(2000)]. (Cited on page 232.)
[207] C. Troessaert, “Enhanced asymptotic symmetry algebra of AdS 3 ,” JHEP 08 (2013)
[208] G. Compère, W. Song, and A. Strominger, “New Boundary Conditions for AdS3,”
JHEP 05 (2013) 152, 1303.2662. (Cited on page 235.)
[209] M. Henneaux, C. Martínez, and R. Troncoso, “Asymptotically anti-de Sitter spacetimes
in topologically massive gravity,” Phys. Rev. D79 (2009) 081502, 0901.2874. (Cited
on page 235.)
[210] M. Henneaux, C. Martínez, and R. Troncoso, “More on Asymptotically Anti-de Sitter
Spaces in Topologically Massive Gravity,” Phys. Rev. D82 (2010) 064038, 1006.0273.
[211] Y. Liu and Y.-W. Sun, “Consistent Boundary Conditions for New Massive Gravity in
AdS3 ,” JHEP 05 (2009) 039, 0903.2933. (Cited on page 235.)
[212] C. Cunliff, “Non-Fefferman-Graham asymptotics and holographic renormalization in
New Massive Gravity,” JHEP 04 (2013) 141, 1301.1347. (Cited on page 235.)
[213] S. Aminneborg, I. Bengtsson, S. Holst, and P. Peldan, “Making anti-de Sitter black
holes,” Class. Quant. Grav. 13 (1996) 2707–2714, gr-qc/9604005. (Cited on page 238.)
[214] M. Bañados, A. Gomberoff, and C. Martínez, “Anti-de Sitter space and black holes,”
Class. Quant. Grav. 15 (1998) 3575–3598, hep-th/9805087. (Cited on page 238.)
[215] C. Crnković and E. Witten, “Covariant description of canonical formalism in geomet-
rical theories,”. (Cited on page 238.)
[216] E. Witten, “Anti-de Sitter space and holography,” Adv. Theor. Math. Phys. 2 (1998)
253–291, hep-th/9802150. (Cited on pages 238 and 312.)
Bibliography 397
[217] J. Kim and M. Porrati, “On a Canonical Quantization of 3D Anti de Sitter Pure
Gravity,” JHEP 10 (2015) 096, 1508.03638. (Cited on page 239.)
[218] C. Troessaert, “Poisson Structure of the Boundary Gravitons in 3D Gravity with Neg-
ative Λ,” Class. Quant. Grav. 32 (2015), no. 23, 235019, 1507.01580. (Cited on
page 240.)
[219] G. Compère, P.-J. Mao, A. Seraj, and M. M. Sheikh-Jabbari, “Symplectic and Killing
symmetries of AdS3 gravity: holographic vs boundary gravitons,” JHEP 01 (2016)
080, 1511.06079. (Cited on page 240.)
[220] E. J. Martinec, “Conformal field theory, geometry, and entropy,” hep-th/9809021.
[221] E. Witten, “Three-Dimensional Gravity Revisited,” 0706.3359. (Cited on page 241.)
[222] M. M. Sheikh-Jabbari and H. Yavartanoo, “On 3d Bulk Geometry of Virasoro Coad-
joint Orbits: Orbit invariant charges and Virasoro hair on locally AdS3 geometries,”
1603.05272. (Cited on page 241.)
[223] A. Seraj, Conserved charges, surface degrees of freedom, and black hole entropy. PhD
thesis, 2016. 1603.02442. (Cited on page 241.)
[224] R. Schon and S.-T. Yau, “Proof of the positive mass theorem. 2.,” Commun. Math.
Phys. 79 (1981) 231–260. (Cited on page 242.)
[225] E. Witten, “A Simple Proof of the Positive Energy Theorem,” Commun. Math. Phys.
80 (1981) 381. (Cited on page 242.)
[226] R. M. Wald, General Relativity. 1984. (Cited on page 242.)
[227] D. R. Brill and S. Deser, “Variational methods and positive energy in general relativity,”
Annals Phys. 50 (1968) 548–570. (Cited on page 242.)
[228] B. Feigin and D. Fuks, “Casimir operators in modules over virasoro algebra,” Dokl.
Akad. Nauk SSSR 269 (1983), no. 5, 1057–1060. (Cited on pages 245 and 319.)
[229] H.-S. La, P. C. Nelson, and A. S. Schwarz, “Remarks on Virasoro model space,” Conf.
Proc. C9003122 (1990) 259–265. (Cited on page 245.)
[230] H. Airault and P. Malliavin, “Unitarizing probability measures for representations of
Virasoro algebra,” Journal de Mathématiques Pures et Appliquées 80 (2001), no. 6,
627 – 667. (Cited on pages 245 and 301.)
[231] H. Airault, P. Malliavin, and A. Thalmaier, “Support of Virasoro unitarizing measures,”
Comptes Rendus Mathematique 335 (2002), no. 7, 621–626. (Cited on pages 245
and 301.)
[232] B. Feigin and D. Fuks, “Verma modules over the Virasoro algebra,” Funkts. Anal.
Prilozh. 17 (1983), no. 3, 91–92. (Cited on pages 246 and 328.)
[233] B. Feigin and D. Fuchs, “Verma modules over the Virasoro algebra,” in Topology, L. D.
Faddeev and A. A. Malcev, eds., vol. 1060 of Lecture Notes in Mathematics, pp. 230–
245. Springer Berlin Heidelberg, 1984. (Cited on page 246.)
[234] B. Feigin and D. Fuchs, Representations of the Virasoro algebra. Reports: Matematiska
Institutionen. Department, Univ., 1986. (Cited on page 246.)
[235] V. G. Kac, Lie Algebras and Related Topics: Proceedings of a Conference Held at New
Brunswick, New Jersey, May 29–31, 1981, ch. Some problems on infinite dimensional
Lie algebras and their representations, pp. 117–126. Springer Berlin Heidelberg, Berlin,
Heidelberg, 1982. (Cited on page 246.)
Bibliography 398
[236] P. Goddard, A. Kent, and D. Olive, “Unitary representations of the Virasoro and
super-Virasoro algebras,” Comm. Math. Phys. 103 (1986), no. 1, 105–119. (Cited on
page 246.)
[237] H. Salmasian and K.-H. Neeb, “Classification of positive energy representations of the
Virasoro group,” 1402.6572. (Cited on page 248.)
[238] T. Apostol, Introduction to Analytic Number Theory. Undergraduate Texts in Mathe-
matics. Springer New York, 1998. (Cited on page 249.)
[239] A. Wassermann, “Direct proofs of the Feigin-Fuchs character formula for unitary rep-
resentations of the Virasoro algebra,” ArXiv e-prints (Dec., 2010) 1012.6003. (Cited
[240] A. Wassermann, “Kac-Moody and Virasoro algebras,” ArXiv e-prints (Apr., 2010)
1004.1287. (Cited on pages 250 and 328.)
[241] S. Giombi, A. Maloney, and X. Yin, “One-loop Partition Functions of 3D Gravity,”
JHEP 08 (2008) 007, 0804.1773. (Cited on pages 252, 331, 332, 340, 341, and 346.)
[242] C. A. Keller and A. Maloney, “Poincare Series, 3D Gravity and CFT Spectroscopy,”
[243] I. J. R. Aitchison and A. J. G. Hey, Gauge Theories in Particle Physics. Graduate
Student Series in Physics. Taylor & Francis, 2004. (Cited on page 253.)
[244] S. Deser, R. Jackiw, and G. ’t Hooft, “Three-Dimensional Einstein Gravity: Dynamics
of Flat Space,” Annals Phys. 152 (1984) 220. (Cited on page 261.)
[245] S. Deser and R. Jackiw, “String Sources in (2+1)-dimensional Gravity,” Annals Phys.
192 (1989) 352. (Cited on page 261.)
[246] D. Christodoulou and S. Klainerman, “The global nonlinear stability of the Minkowski
space,” Séminaire Équations aux dérivées partielles (Polytechnique) (1993) 1–29.
[247] A. Strominger, “Asymptotic Symmetries of Yang-Mills Theory,” JHEP 07 (2014) 151,
[248] T. He, P. Mitra, A. P. Porfyriadis, and A. Strominger, “New Symmetries of Massless
QED,” JHEP 10 (2014) 112, 1407.3789. (Cited on pages 265 and 312.)
[249] V. Lysov, S. Pasterski, and A. Strominger, “Low’s Subleading Soft Theorem as a
Symmetry of QED,” Phys. Rev. Lett. 113 (2014), no. 11, 111601, 1407.3814. (Cited
[250] A. Strominger and A. Zhiboedov, “Gravitational Memory, BMS Supertranslations and
Soft Theorems,” JHEP 01 (2016) 086, 1411.5745. (Cited on pages 265 and 312.)
[251] D. Kapec, V. Lysov, and A. Strominger, “Asymptotic Symmetries of Massless QED in
Even Dimensions,” 1412.2763. (Cited on pages 265 and 312.)
[252] S. Pasterski, A. Strominger, and A. Zhiboedov, “New Gravitational Memories,”
[253] D. Kapec, V. Lysov, S. Pasterski, and A. Strominger, “Higher-Dimensional Supertrans-
lations and Weinberg’s Soft Graviton Theorem,” 1502.07644. (Cited on pages 265
and 312.)
[254] S. Pasterski, “Asymptotic Symmetries and Electromagnetic Memory,” 1505.00716.
[255] T. He, P. Mitra, and A. Strominger, “2D Kac-Moody Symmetry of 4D Yang-Mills
Theory,” 1503.02663. (Cited on pages 265 and 312.)
Bibliography 399
[256] D. Kapec, M. Pate, and A. Strominger, “New Symmetries of QED,” 1506.02906.

[257] A. Strominger, “Magnetic Corrections to the Soft Photon Theorem,” Phys. Rev. Lett.
116 (2016), no. 3, 031602, 1509.00543. (Cited on pages 265 and 312.)
[258] T. T. Dumitrescu, T. He, P. Mitra, and A. Strominger, “Infinite-Dimensional Fermionic
Symmetry in Supersymmetric Gauge Theories,” 1511.07429. (Cited on pages 265
and 312.)
[259] M. Henneaux, “Asymptotically flat spaces in D spacetime dimensions: a review of the
Hamiltonian approach,” April, 2016. Talk given at the workshop Flat Holography at
the Simons Center for Geometry and Physics. (Cited on page 265.)
[260] J. Hartong, “Holographic Reconstruction of 3D Flat Space-Time,” 1511.01387. (Cited
[261] A. Bagchi, S. Detournay, and D. Grumiller, “Flat-Space Chiral Gravity,” Phys. Rev.
Lett. 109 (2012) 151301, 1208.1658. (Cited on pages 268 and 312.)
[262] D. Grumiller and W. Merbis, “Free energy of topologically massive gravity and flat
space holography,” in 2nd Karl Schwarzschild Meeting on Gravitational Physics (KSM
2015) Frankfurt am Main, Germany, July 20-24, 2015. 2015. 1509.08505. (Cited on
page 268.)
[263] H. A. González and M. Pino, “Asymptotically flat spacetimes in 3D bigravity,” Phys.
Rev. D86 (2012) 084045, 1207.0882. (Cited on pages 268 and 312.)
[264] H. R. Afshar, “Flat/AdS boundary conditions in three dimensional conformal gravity,”
JHEP 10 (2013) 027, 1307.4855. (Cited on pages 268 and 312.)
[265] G. Barnich, C. Troessaert, D. Tempo, and R. Troncoso, “Asymptotically locally flat
spacetimes and dynamical nonspherically-symmetric black holes in three dimensions,”
Phys. Rev. D93 (2016), no. 8, 084001, 1512.05410. (Cited on page 268.)
[266] C. Troessaert, D. Tempo, and R. Troncoso, “Asymptotically flat black holes and grav-
itational waves in three-dimensional massive gravity,” in 8th Aegean Summer School:
Gravitational Waves: From Theory to Observations Rethymno, Crete, Greece, June
29-July 4, 2015. 2015. 1512.09046. (Cited on pages 268 and 312.)
[267] L. Cornalba and M. S. Costa, “A New cosmological scenario in string theory,” Phys.
Rev. D66 (2002) 066001, hep-th/0203031. (Cited on page 269.)
[268] L. Cornalba and M. S. Costa, “Time dependent orbifolds and string cosmology,”
Fortsch. Phys. 52 (2004) 145–199, hep-th/0310099. (Cited on page 269.)
[269] P. J. McCarthy, “The Bondi-Metzner-Sachs Group in the Nuclear Topol-
ogy,” Proceedings of the Royal Society of London A: Mathematical,
Physical and Engineering Sciences 343 (1975), no. 1635, 489–523,
[270] A. Ashtekar and M. Streubel, “Symplectic Geometry of Radiative Modes and Con-
served Quantities at Null Infinity,” Proc. Roy. Soc. Lond. A376 (1981) 585–607. (Cited
[271] O. M. Moreschi, “Supercenter of Mass System at Future Null Infinity,” Class. Quant.
Grav. 5 (1988) 423–435. (Cited on page 280.)
[272] B. R. Iyer, A. Kembhavi, J. V. Narlikar, and C. V. Vishveshwara, eds., Highlights in
gravitation and cosmology. Proceedings, International Conference, Goa, India, Decem-
ber 14-19, 1987. 1988. (Cited on page 280.)
Bibliography 400
[273] A. D. Helfer, “A phase space for gravitational radiation,” Comm. Math. Phys. 170
(1995), no. 3, 483–502. (Cited on page 280.)
[274] E. E. Flanagan and D. A. Nichols, “Conserved charges of the extended Bondi-Metzner-
Sachs algebra,” 1510.03386. (Cited on page 280.)
[275] J. Garecki, “Canonical angular supermomentum tensors in general relativity,” Journal
of Mathematical Physics 40 (1999), no. 8, 4035–4055. (Cited on page 280.)
[276] J. Garecki, “Superenergy and angular supermomentum tensors in general relativity,”
Reports on Mathematical Physics 44 (1999), no. 1, 95–100. (Cited on page 280.)
[277] G. Barnich, A. Gomberoff, and H. A. González, “A 2D field theory equivalent to 3D
gravity with no cosmological constant,” Springer Proc. Math. Stat. 60 (2014) 135–138,
1303.3568. (Cited on page 284.)
[278] G. Barnich, A. Gomberoff, and H. A. González, “Three-dimensional Bondi-Metzner-
Sachs invariant two-dimensional field theories as the flat limit of Liouville theory,”
Phys. Rev. D87 (2013), no. 12, 124032, 1210.0731. (Cited on pages 284 and 312.)
[279] H. A. González and M. Pino, “Boundary dynamics of asymptotically flat 3D gravity
coupled to higher spin fields,” JHEP 05 (2014) 127, 1403.4898. (Cited on page 284.)
[280] G. Barnich, L. Donnay, J. Matulich, and R. Troncoso, “Super-BMS3 invariant boundary
theory from three-dimensional flat supergravity,” 1510.08824. (Cited on pages 284,
361, and 364.)
[281] G. Barnich, A. Gomberoff, and H. A. González, “The Flat limit of three dimensional
asymptotically anti-de Sitter spacetimes,” Phys. Rev. D86 (2012) 024020, 1204.3288.
[282] L. Susskind, “Holography in the flat space limit,” hep-th/9901079. [AIP Conf. Proc.
493, 98 (1999)]. (Cited on page 290.)
[283] J. Polchinski, “S matrices from AdS space-time,” hep-th/9901076. (Cited on
page 290.)
[284] M. Gary, S. B. Giddings, and J. Penedones, “Local bulk S-matrix elements and CFT
singularities,” Phys. Rev. D80 (2009) 085005, 0903.4437. (Cited on page 290.)
[285] R. N. Caldeira Costa, “Aspects of the zero Λ limit in the AdS/CFT correspondence,”
[286] C. Krishnan, A. Raju, and S. Roy, “A Grassmann path from AdS3 to flat space,” JHEP
03 (2014) 036, 1312.2941. (Cited on pages 290 and 312.)
[287] A. Bagchi, R. Gopakumar, I. Mandal, and A. Miwa, “GCA in 2d,” JHEP 08 (2010)
004, 0912.1090. (Cited on pages 291, 317, 321, 322, and 357.)
[288] A. Bagchi, “The BMS/GCA correspondence,” 1006.3354. (Cited on pages 291
and 312.)
[289] A. Bagchi and I. Mandal, “Supersymmetric Extension of Galilean Conformal Algebras,”
Phys. Rev. D80 (2009) 086011, 0905.0580. (Cited on page 291.)
[290] A. Bagchi and I. Mandal, “On Representations and Correlation Functions of Galilean
Conformal Algebras,” Phys. Lett. B675 (2009) 393–397, 0903.4524. (Cited on
page 291.)
[291] A. Bagchi and R. Gopakumar, “Galilean Conformal Algebras and AdS/CFT,” JHEP
07 (2009) 037, 0902.1385. (Cited on page 291.)
Bibliography 401
[292] A. Bagchi, “Correspondence between Asymptotically Flat Spacetimes and Nonrela-

tivistic Conformal Field Theories,” Phys. Rev. Lett. 105 (2010) 171601. (Cited on
pages 291 and 312.)
[293] R. Fareghbal and A. Naseh, “Flat-Space Energy-Momentum Tensor from BMS/GCA
Correspondence,” JHEP 03 (2014) 005, 1312.2109. (Cited on pages 291 and 312.)
[294] N. Banerjee, D. P. Jatkar, S. Mukhi, and T. Neogi, “Free-field realisations of the BMS3
algebra and its extensions,” 1512.06240. (Cited on pages 291 and 312.)
[295] R. Fareghbal and Y. Izadi, “Flat-Space Holography and Stress Tensor of Kerr Black
Hole,” 1603.04137. (Cited on pages 291 and 312.)
[296] N. A. of Sciences, Biographical Memoirs vol. 61. The National Academies Press, Wash-
ington, DC, 1992. (Cited on page 301.)
[297] A. Alekseev and S. Shatashvili, “From geometric quantization to conformal field the-
ory,” Comm. Math. Phys. 128 (1990), no. 1, 197–212. (Cited on page 301.)
[298] H. Airault, “Mesure unitarisante : algèbre de Heisenberg, algèbre de Virasoro,”
Comptes Rendus Mathématiques 334 (2002) 787–792. (Cited on page 301.)
[299] H. Airault, “Affine coordinates and Virasoro unitarizing measures,” Journal de Math-
ématiques Pures et Appliquées 82 (2003), no. 4, 425 – 455. (Cited on page 301.)
[300] J. Dai and D. Pickrell, “The orbit method and the Virasoro extension of Diff+ (S 1 ). I.
Orbital integrals,” Journal of Geometry and Physics 44 (1, 2003) 623–653. (Cited on
pages 301 and 363.)
[301] E. T. Shavgulidze, “A measure that is quasi-invariant with respect to the action of a
group of diffeomorphisms of a finite-dimensional manifold,” Dokl. Akad. Nauk SSSR
303 (1988), no. 4, 811–814. (Cited on page 301.)
[302] E. T. Shavgulidze, “An example of a measure quasi-invariant under the action of the
diffeomorphism group of the circle,” Funkts. Anal. Prilozh. 12 (1978), no. 3, 55–60.
[303] E. T. Shavgulidze, “Mesures quasi-invariantes sur les groupes de difféomorphismes
des variétés riemaniennes,” C. R. Acad. Sci. Paris 321 (1995) 229–232. (Cited on
page 301.)
[304] E. T. Shavgulidze, “Quasiinvariant measures on groups of diffeomorphisms.,” in Loop
spaces and groups of diffeomorphisms. Collected papers, pp. 181–202 (1997); transla-
tion from tr. mat. inst. steklova 217, 189–208. Moscow: MAIK Nauka/Interperiodica
Publishing, 1997. (Cited on page 301.)
[305] V. Bogachev, Gaussian Measures. Mathematical surveys and monographs. American
Mathematical Society, 1998. (Cited on page 301.)
[306] H. Shimomura, “Quasi-invariant Measures on the Group of Diffeomorphisms and
Smooth Vectors of Unitary Representations,” Journal of Functional Analysis 187
(2001), no. 2, 406 – 441. (Cited on page 301.)
[307] R. Penrose, “Asymptotic properties of fields and space-times,” Phys. Rev. Lett. 10
(1963) 66–68. (Cited on page 311.)
[308] A. Komar, “Quantized Gravitational Theory and Internal Symmetries,” Phys. Rev.
Lett. 15 (1965) 76–78. (Cited on page 311.)
[309] P. J. McCarthy and M. Crampin, “Representations of the Bondi-
Metzner-Sachs Group. III. Poincaré Spin Multiplicities and Irreducibil-
ity,” Proceedings of the Royal Society of London A: Mathematical,
Physical and Engineering Sciences 335 (1973), no. 1602, 301–311,
Bibliography 402
[310] M. Crampin and P. J. McCarthy, “Representations of the Bondi-Metzner-Sachs Group.
IV. Cantoni Representations are Induced,” Proceedings of the Royal Society of Lon-
don A: Mathematical, Physical and Engineering Sciences 351 (1976), no. 1664, 55–
70, http://rspa.royalsocietypublishing.org/content/351/1664/55.full.pdf.
[311] V. Cantoni, “A Class of Representations of the Generalized Bondi-Metzner Group,”
Journal of Mathematical Physics 7 (1966), no. 8, 1361–1364. (Cited on page 311.)
[312] V. Cantoni, “Reduction of Some Representations of the Generalized Bondi-Metzner
Group,” Journal of Mathematical Physics 8 (1967), no. 8, 1700–1706. (Cited on
page 311.)
[313] L. Girardello and G. Parravicini, “Continuous spins in the Bondi-Metzner-Sachs group
of asymptotic symmetry in general relativity,” Phys. Rev. Lett. 32 (1974) 565–568.
[314] M. Crampin, “Physical significance of the topology of the Bondi-Metzner-Sachs group,”
Phys. Rev. Lett. 33 (1974) 547–550. (Cited on page 311.)
[315] A. Piard, “Unitary representations of semi-direct product groups with infinite dimen-
sional Abelian normal subgroup,” Reports on Mathematical Physics 11 (1977), no. 2,
259 – 278. (Cited on page 311.)
[316] A. Piard, “Representations of the Bondi-Metzner-Sachs Group with the Hilbert Topol-
ogy,” Rept. Math. Phys. 11 (1977) 279–283. (Cited on page 311.)
[317] U. Cattaneo, “Borel multipliers for the Bondi-Metzner-Sachs group,” J. Math. Phys.
20 (1979) 2257–2263. (Cited on page 311.)
[318] M. A. Awada, G. W. Gibbons, and W. T. Shaw, “Conformal supergravity, Twistors
and the Super BMS group,” Annals Phys. 171 (1986) 52. (Cited on page 311.)
[319] R. P. Geroch, “Null infinity is not a good initial data surface,” J. Math. Phys. 19 (1978)
1300–1303. (Cited on page 311.)
[320] A. Ashtekar and R. O. Hansen, “A unified treatment of null and spatial infinity in gen-
eral relativity. I - Universal structure, asymptotic symmetries, and conserved quantities
at spatial infinity,” J. Math. Phys. 19 (1978) 1542–1566. (Cited on page 311.)
[321] A. Ashtekar, “Asymptotic Quantization of the Gravitational Field,” Phys. Rev. Lett.
46 (1981) 573–576. (Cited on page 311.)
[322] A. Ashtekar, Asymptotic quantization: Based on 1984 Naples lectures. 1987. (Cited
on page 311.)
[323] S. S. Gubser, I. R. Klebanov, and A. M. Polyakov, “Gauge theory correlators from
noncritical string theory,” Phys. Lett. B428 (1998) 105–114, hep-th/9802109. (Cited
on page 312.)
[324] R. Bousso, “Holography in general space-times,” JHEP 06 (1999) 028,
hep-th/9906022. (Cited on page 312.)
[325] R. Bousso, “The Holographic principle for general backgrounds,” Class. Quant. Grav.
17 (2000) 997–1005, hep-th/9911002. (Cited on page 312.)
[326] G. Arcioni and C. Dappiaggi, “Exploring the holographic principle in asymptot-
ically flat space-times via the BMS group,” Nucl. Phys. B674 (2003) 553–592,
hep-th/0306142. (Cited on page 312.)
Bibliography 403
[327] G. Arcioni and C. Dappiaggi, “Holography in asymptotically flat space-times and

the BMS group,” Class. Quant. Grav. 21 (2004) 5655, hep-th/0312186. (Cited on
page 312.)
[328] C. Dappiaggi, “Elementary particles, holography and the BMS group,” Phys. Lett.
B615 (2005) 291–296, hep-th/0412142. (Cited on page 312.)
[329] C. Dappiaggi, “BMS field theory and holography in asymptotically flat space-times,”
JHEP 11 (2004) 011, hep-th/0410026. (Cited on page 312.)
[330] G. Arcioni and C. Dappiaggi, “Holography and BMS field theory,” AIP Conf. Proc.
751 (2005) 176–178, hep-th/0409313. [,176(2004)]. (Cited on page 312.)
[331] A. Virmani, “Supertranslations and Holographic Stress Tensor,” JHEP 02 (2012) 024,
1112.2146. (Cited on page 312.)
[332] G. Barnich and C. Troessaert, “Comments on holographic current algebras and asymp-
totically flat four dimensional spacetimes at null infinity,” JHEP 11 (2013) 003,
1309.0794. (Cited on page 312.)
[333] G. Barnich and P.-H. Lambert, “Einstein-Yang-Mills theory: Asymptotic symmetries,”
Phys. Rev. D88 (2013) 103006, 1310.2698. (Cited on page 312.)
[334] P.-H. Lambert, Conformal symmetries of gravity from asymptotic methods: further
developments. PhD thesis, Brussels U., 2014. 1409.4693. (Cited on page 312.)
[335] I. Fujisawa and R. Nakayama, “Bondi-Metzner-Sachs surface-charge algebra via a
Hamiltonian framework,” Phys. Rev. D91 (2015), no. 12, 126005, 1503.03225. (Cited
on page 312.)
[336] G. Barnich and C. Troessaert, “Supertranslations call for superrotations,” PoS (2010)
010, 1102.4632. [Ann. U. Craiova Phys.21,S11(2011)]. (Cited on pages 312 and 313.)
[337] G. Barnich, “Entropy of three-dimensional asymptotically flat cosmological solutions,”
JHEP 10 (2012) 095, 1208.4371. (Cited on pages 312 and 326.)
[338] W. Schulgin and J. Troost, “Asymptotic symmetry groups and operator algebras,”
[339] V. Bonzom and B. Dittrich, “3D holography: from discretum to continuum,” JHEP
03 (2016) 208, 1511.05441. (Cited on pages 312 and 341.)
[340] A. Bagchi and R. Fareghbal, “BMS/GCA Redux: Towards Flatspace Holography from
Non-Relativistic Symmetries,” JHEP 10 (2012) 092, 1203.5795. (Cited on page 312.)
[341] A. Bagchi, S. Detournay, R. Fareghbal, and J. Simón, “Holography of 3D Flat Cosmo-
logical Horizons,” Phys. Rev. Lett. 110 (2013), no. 14, 141302, 1208.4372. (Cited on
pages 312 and 326.)
[342] A. Bagchi, D. Grumiller, J. Salzer, S. Sarkar, and F. SchÃűller, “Flat space cosmologies
in two dimensions - Phase transitions and asymptotic mass-domination,” Phys. Rev.
D90 (2014), no. 8, 084041, 1408.5337. (Cited on page 312.)
[343] R. Fareghbal and A. Naseh, “Aspects of Flat/CCFT Correspondence,” Class. Quant.
Grav. 32 (2015) 135013, 1408.6932. (Cited on page 312.)
[344] R. Fareghbal and S. M. Hosseini, “Holography of 3D Asymptotically Flat Black Holes,”
[345] A. Bagchi, D. Grumiller, and W. Merbis, “Stress tensor correlators in three-dimensional
gravity,” Phys. Rev. D93 (2016), no. 6, 061502, 1507.05620. (Cited on page 312.)
Bibliography 404
[346] S. M. Hosseini and A. Véliz-Osorio, “Gravitational anomalies, entanglement entropy,

and flat-space holography,” Phys. Rev. D93 (2016), no. 4, 046005, 1507.06625. (Cited
on page 312.)
[347] R. Fareghbal, A. Naseh, and S. Rouhani, “Scale vs Conformal Invariance in Ultra-
Relativistic Field Theory,” 1511.01774. (Cited on page 312.)
[348] M. Campiglia and A. Laddha, “Asymptotic symmetries of QED and Weinberg’s soft
photon theorem,” JHEP 07 (2015) 115, 1505.05346. (Cited on page 312.)
[349] M. Campiglia and A. Laddha, “Asymptotic symmetries of gravity and soft theorems
for massive particles,” JHEP 12 (2015) 094, 1509.01406. (Cited on page 312.)
[350] S. W. Hawking, M. J. Perry, and A. Strominger, “Soft Hair on Black Holes,”
1601.00921. (Cited on page 312.)
[351] G. Barnich and C. Troessaert, “Finite BMS transformations,” JHEP 03 (2016) 167,
1601.04090. (Cited on page 313.)
[352] M. Campiglia and A. Laddha, “Asymptotic symmetries and subleading soft graviton
theorem,” Phys. Rev. D90 (2014), no. 12, 124028, 1408.2228. (Cited on page 313.)
[353] M. Campiglia and A. Laddha, “New symmetries for the Gravitational S-matrix,” JHEP
04 (2015) 076, 1502.02318. (Cited on page 313.)
[354] G. Compère and J. Long, “Vacua of the gravitational field,” 1601.04958. (Cited on
page 313.)
[355] M. Schottenloher, A Mathematical Introduction to Conformal Field Theory. Lecture
Notes in Physics. Springer Berlin Heidelberg, 2008. (Cited on page 315.)
[356] D. Grumiller, M. Riegler, and J. Rosseel, “Unitarity in three-dimensional flat space
higher spin theories,” JHEP 07 (2014) 015, 1403.5297. (Cited on pages 317, 331,
346, 351, 353, 354, 355, and 357.)
[357] A. Schild, “Classical Null Strings,” Phys. Rev. D16 (1977) 1722. (Cited on page 323.)
[358] F. Lizzi, B. Rai, G. Sparano, and A. Srivastava, “Quantization of the Null String
and Absence of Critical Dimensions,” Phys. Lett. B182 (1986) 326–330. (Cited on
page 323.)
[359] J. Gamboa, C. Ramirez, and M. Ruiz-Altaba, “Null spinning strings,” Nucl. Phys.
B338 (1990) 143–187. (Cited on page 323.)
[360] J. Gamboa, C. Ramirez, and M. Ruiz-Altaba, “Quantum null (super)strings,” Phys.
Lett. B225 (1989) 335–339. (Cited on page 323.)
[361] A. Bagchi, “Tensionless Strings and Galilean Conformal Algebra,” JHEP 05 (2013)
141, 1303.0291. (Cited on page 324.)
[362] A. Bagchi, S. Chakrabortty, and P. Parekh, “Tensionless Strings from Worldsheet Sym-
metries,” JHEP 01 (2016) 158, 1507.04361. (Cited on page 324.)
[363] E. Casali and P. Tourkine, “On the null origin of the ambitwistor string,” 1606.05636.
[364] D. V. Vassilevich, “Heat kernel expansion: User’s manual,” Phys. Rept. 388 (2003)
279–360, hep-th/0306138. (Cited on page 332.)
[365] L. P. S. Singh and C. R. Hagen, “Lagrangian formulation for arbitrary spin. 1. The
boson case,” Phys. Rev. D9 (1974) 898–909. (Cited on page 334.)
[366] Yu. M. Zinoviev, “On massive high spin particles in AdS,” hep-th/0108192. (Cited
on pages 334, 335, and 340.)
Bibliography 405
[367] C. Fronsdal, “Massless Fields with Integer Spin,” Phys. Rev. D18 (1978) 3624. (Cited
on page 334.)
[368] R. Rahman, “Higher Spin Theory - Part I,” PoS ModaveVIII (2012) 004, 1307.3199.
[369] G. W. Gibbons, M. J. Perry, and C. N. Pope, “Partition functions, the Bekenstein
bound and temperature inversion in anti-de Sitter space and its conformal boundary,”
Phys. Rev. D74 (2006) 084009, hep-th/0606186. (Cited on pages 338 and 342.)
[370] R. Gopakumar, R. K. Gupta, and S. Lal, “The Heat Kernel on AdS,” JHEP 11 (2011)
[371] M. Beccaria and A. A. Tseytlin, “On higher spin partition functions,” J. Phys. A48
(2015), no. 27, 275401, 1503.08143. (Cited on page 338.)
[372] A. Bagchi, S. Detournay, D. Grumiller, and J. Simon, “Cosmic Evolution from Phase
Transition of Three-Dimensional Flat Space,” Phys. Rev. Lett. 111 (2013), no. 18,
181301, 1305.2919. (Cited on page 340.)
[373] F. A. Dolan, “Character formulae and partition functions in higher dimensional con-
formal field theory,” J. Math. Phys. 47 (2006) 062303, hep-th/0508031. (Cited on
page 342.)
[374] J. Balog, L. Fehér, L. O’Raifeartaigh, P. Forgacs, and A. Wipf, “Toda Theory and
W Algebra From a Gauged WZNW Point of View,” Annals Phys. 203 (1990) 76–136.
[375] B. A. Khesin and B. Z. Shapiro, “Nondegenerate curves on S**2 and orbit classification
of the Zamolodchikov algebra,” Commun. Math. Phys. 145 (1992) 357–362. (Cited
[376] Z. Bajnok and D. Nogradi, “Geometry of W algebras from the affine Lie algebra point
of view,” J. Phys. A34 (2001) 4811–4830, hep-th/0012190. (Cited on pages 343, 344,
and 348.)
[377] M. Henneaux and S.-J. Rey, “Nonlinear Winf inity as Asymptotic Symmetry of Three-
Dimensional Higher Spin Anti-de Sitter Gravity,” JHEP 12 (2010) 007, 1008.4579.
[378] A. Campoleoni, S. Fredenhagen, S. Pfenninger, and S. Theisen, “Asymptotic symme-
tries of three-dimensional gravity coupled to higher-spin fields,” JHEP 11 (2010) 007,
[379] M. R. Gaberdiel and T. Hartman, “Symmetries of Holographic Minimal Models,” JHEP
05 (2011) 031, 1101.2910. (Cited on page 343.)
[380] A. Campoleoni, S. Fredenhagen, and S. Pfenninger, “Asymptotic W-symmetries in
three-dimensional higher-spin gauge theories,” JHEP 09 (2011) 113, 1107.0290.
[381] H. Afshar, A. Bagchi, R. Fareghbal, D. Grumiller, and J. Rosseel, “Spin-3 Grav-
ity in Three-Dimensional Flat Space,” Phys. Rev. Lett. 111 (2013), no. 12, 121603,
[382] H. A. González, J. Matulich, M. Pino, and R. Troncoso, “Asymptotically flat space-
times in three-dimensional higher spin gravity,” JHEP 09 (2013) 016, 1307.5651.
[383] A. Campoleoni, “Higher Spins in D = 2 + 1,” Subnucl. Ser. 49 (2013) 385–396,
1110.5841. (Cited on page 351.)
Bibliography 406
[384] L. P. S. Singh and C. R. Hagen, “Lagrangian formulation for arbitrary spin. 2. The
fermion case,” Phys. Rev. D9 (1974) 910–920. (Cited on page 358.)
[385] R. R. Metsaev, “Gauge invariant formulation of massive totally symmetric fermionic
fields in (A)dS space,” Phys. Lett. B643 (2006) 205–212, hep-th/0609029. (Cited on
page 358.)
[386] J. Fang and C. Fronsdal, “Massless Fields with Half Integral Spin,” Phys. Rev. D18
(1978) 3630. (Cited on page 358.)
[387] T. Creutzig, Y. Hikida, and P. B. Ronne, “Higher spin AdS3 supergravity and its dual
CFT,” JHEP 02 (2012) 109, 1111.2139. (Cited on pages 358 and 361.)
[388] O. Fuentealba, J. Matulich, and R. Troncoso, “Extension of the Poincaré group
with half-integer spin generators: hypergravity and beyond,” JHEP 09 (2015) 003,
[389] O. Fuentealba, J. Matulich, and R. Troncoso, “Asymptotically flat structure of hyper-
gravity in three spacetime dimensions,” JHEP 10 (2015) 009, 1508.04663. (Cited on
pages 361 and 368.)
[390] I. Mandal, “Supersymmetric Extension of GCA in 2d,” JHEP 11 (2010) 018,
1003.0209. (Cited on page 361.)
[391] P. Deligne and I. Study, Quantum Fields and Strings: A Course for Mathematicians,
vol. 2 of Quantum Fields and Strings: A Course for Mathematicians. American Math-
ematical Soc., 1999. (Cited on page 362.)
[392] C. Carmeli, G. Cassinelli, A. Toigo, and V. S. Varadarajan, “Unitary representations of
super Lie groups and applications to the classification and multiplet structure of super
particles,” Commun. Math. Phys. 263 (2006) 217–258, hep-th/0501061. [Erratum:
Commun. Math. Phys. 307, 565 (2011)]. (Cited on page 362.)
[393] C. Carmeli, Super Lie Groups: Structure and Representations. PhD thesis, 2006. Avail-
able on http://www.ge.infn.it/~carmeli/publication/publication.html. (Cited
on page 362.)
[394] M. Bañados, K. Bautier, O. Coussaert, M. Henneaux, and M. Ortiz, “Anti-de Sitter
/ CFT correspondence in three-dimensional supergravity,” Phys. Rev. D58 (1998)
085020, hep-th/9805165. (Cited on page 364.)
[395] M. Henneaux, L. Maoz, and A. Schwimmer, “Asymptotic dynamics and asymptotic
symmetries of three-dimensional extended AdS supergravity,” Annals Phys. 282 (2000)
31–66, hep-th/9910013. (Cited on page 364.)
[396] O. Coussaert and M. Henneaux, “Supersymmetry of the (2+1) black holes,” Phys. Rev.
Lett. 72 (1994) 183–186, hep-th/9310194. (Cited on pages 364 and 366.)
[397] I. MacDonald, Symmetric Functions and Hall Polynomials. Second edition in Oxford
Mathematical Monographs. Oxford Science Publications, 1995. (Cited on page 370.)
[398] W. Fulton and J. Harris, Representation Theory: A First Course. Graduate Texts in
Mathematics. Springer, New York, 1991. (Cited on pages 372, 373, 374, and 378.)
[399] D. E. Littlewood, The Theory of Group Characters and Matrix Representations of
Groups. American Mathematical Soc., 1950. (Cited on pages 378 and 379.)
[400] G. Barnich, A. Gomberoff, H. González, B. Oblak, and P. Salgado-Rebolledo, in prepa-
ration. (Cited on page 383.)
Index
Symbols allowed gauge transformation, 223

2πZ-equivariance, 148 almost everywhere, 34
3D gravity, 218 ambient construction, 230
dimensional reduction, 221 angular momentum, 128, 141, 234
no local degrees of freedom, 342 in AdS3 , 234
partition function, 252, 340 in flat space, 266
orbital versus spin, 129
A angular momentum aspect, 267
A∗ (space of momenta), 63 angular potential, 141, 332
absolute time, 90 angular supermomentum, 280
action functional angular velocity, 142
differentiable, 220 annihilator, 131
for gravity, 218, 232 anomaly, 40, 43
for higher spin field, 334, 358 antilinear, 17
on symplectic manifold, 119 antiperiodic, 358
action of a group, 38 antiunitary, 17
adjoint representation, 99 anyon, 21, 89, 309
of Diff(S 1 ), 150, 275 asymptotic flatness, 4, see also BMS3
of centrally extended group, 171 fall-offs
of semi-direct product, 128, 272 asymptotic Killing spinor, 364
adjoint vector, 100, 173 asymptotic Killing vector, 224
for BMS3 , see bms3 algebra for AdS3 , 231
ADM mass, 234 for flat space, 263
admissible momentum, 362, 366 asymptotic quantization, 311
AdS radius, 226 asymptotic state, 49
as infrared regulator, 317 asymptotic symmetry, 1, 224
AdS3 , 226, see also Brown-Henneaux for AdS3 , 232
boundary, 228 for flat space, 264
conical deficit, 229 for higher spins, 343, 346
Killing vectors, 227 for supergravity, 364
null infinity, 286 in the bulk, 232
AdS/CFT, 7, 238 Atiyah-Bott theorem, 54, 80
and coadjoint representation, 239 average lemma, 208
flat limit, 290
non-relativistic limit, 291 B
advanced Bondi coordinates, 4, 260 Banach space, 35
affine module, 28, 173 Bañados metric, 233, see also Brown-
and quasi-regular reps, 43 Henneaux metric
407
Index 408
Bargmann character, 95, 142 energy spectrum, 321

Bargmann group, 91 mass, 297
coadjoint orbits, 140 scalar, 302
momentum, 92 spin, 309
Bessel function, 336 wavefunction, 302
bigravity, 268 with definite supermomentum, 304
black hole, 6, 236 with positive energy, 296
BMS gauge, 262 BMS3 stress tensor, 280, 284
BMS group, 2, 5, 311 as flat limit, 290
extended, 6 BMS3 surface charge, 266
in 3D, see BMS3 group BMS3 tachyon, 295
BMS particle, 9 BMS3 vacuum, 307
in 4D, 311 with spin, 322
BMS world line, 383 bms3 algebra, 264, 278
BMS3 central charge, 265, 299 and gca2 algebra, 291
as mass scale, 268 as flat limit, 288
BMS3 character, 326 Casimir operators, 318
for vacuum representation, 329 centrally extended, 267, 278
universal, 328 cohomology, 281
BMS3 fall-offs, 262 induced module, 319
angular momentum, 266 perfect, 281
asymptotic Killing vector, 263 supersymmetric version, 364
asymptotic symmetry algebra, 264 bms3 module, 319
Hamiltonian, 266 as flat limit, 321
on-shell metric, 265 vacuum, 320
phase space, 283 Bohr-Sommerfeld quantization, 119
surface charge, 266 Bondi coordinates, 2, 4, 258
surface charge algebra, 267 globality, 264
BMS3 group, 8, 276 Bondi mass (aspect), 266
as flat limit, 287 as Virasoro energy functional, 207
at null infinity, 261 boost, 62, 76, 83, see also standard boost
central extension, 277 acting on supermomentum, 298
coadjoint vector, 280 as projective transformation, 212
connected, 276 Borel measure, 33
Kirillov-Kostant bracket, 281 bosonic partition function, 335, 338
Poincaré subgroups, 279 in 3D, 340
universal cover, 277 massless limit, 339
BMS3 metric, 265 Bott-Thurston cocycle, 161, 164
as stress tensor, 280 and Gelfand-Fuks, 162
orbit under BMS3 , 285 and Schwarzian derivative, 166
positive energy, 286 boundary, 220
zero-mode, 268 boundary conditions, see fall-off cond.
BMS3 momentum, see supermomentum boundary degree of freedom, 221
BMS3 particle, 294 boundary graviton, 240, 285, 308
as dressed particle, 306 dressing, 252
as projective representation, 303 partition function, 252, 340
as quantized metric, 304, 331 quantization, 251, 304
Index 409
boundary term, 220, 232 adjoint representation, 171

bra, 15 classical, 107, 225
broken symmetry, 253 coadjoint representation, 172
Brown-Henneaux central charge, 7, 233 Lie bracket, 172
Brown-Henneaux fall-offs, 229 of bms3 algebra, 278
angular momentum, 234 of BMS3 group, 277
asymptotic Killing vector, 231 of Diff(S 1 ), 173
asymptotic symmetry group, 232 of group, 19, 29
at null infinity, 286 of Lie algebra, 25
Hamiltonian, 234 of PSL(2, R), 171
on-shell metric, 233 of Vect(S 1 ), 174
phase space, 238 topological, 20
surface charge, 234 trivial, 19, 26, 29
symplectic leaves, 240 universal, 26, 30, 174
Brown-Henneaux metric, 233 centre of mass, 129
as CFT stress tensor, 233, 238 CFT vacuum, see Virasoro vacuum
for BTZ black hole, 236 character, 52, see also Frobenius for-
for conical deficit, 236 mula
orbit under Virasoro, 240 and fixed points, 54
positive energy, 241 for BMS3 , 326, 329
zero-mode, 235 for flat WN , 349, 352
BRST, 43 for hypersymmetric BMS3 , 369
BTZ black hole, 236 for induced representation, 53
as quotient of AdS3 , 238 for Poincaré, 80–82, 89
bundle of little group orbits, 133 for semi-direct product, 69
for SL(2, R), 127
C for SO(n), 337, 373, 378, 379
c2 (BMS3 central charge), 265 for supersymmetric BMS3 , 368
as mass scale, 268, 299 for Virasoro, 249, 251, 252
not a Virasoro central charge, 268 for WN , 346
canonical symplectic form, 102, 134 within partition function, 341
canonical system of imprimitivity, 57 Chern-Simons theory, 221
canonical transformation, 103 Chevalley-Eilenberg differential, 23, 27
Cartan subalgebra, 31, 337, 372, 373 circle, see S 1
Casimir operator, 314 class function, 52
for bms3 , 318 classical central extension, 107, 225
for sl(2, R), 244 in AdS3 , 233
for Virasoro, 245 in flat space, 265, 267
Čech cohomology, 116 classical flat W3 algebra, 347
celestial circle, 260 classical observable, 103
celestial sphere, 3, 230 classical spin, 132
central charge, 172, 175 classical W3 algebra, 345
Brown-Henneaux, 233 classification
flat limit, 290 of BMS3 particles, 295
for Bargmann group, 92 of coadjoint orbits, 134, 181
for BMS3 , 265, 299 of particles, 70
central extension, 19 of projective reps, 22
Index 410
of Virasoro reps, 248 of Vect(S 1 ), 157, 159, 163

Clifford algebra, 362, 367 commutator, 30, 113
closed time-like curve, 226, 229, 237 complete homogeneous symmetric poly-
coadjoint orbit, 100 nomial, 369
as cotangent bundle, 131 complete metric space, 15
as symplectic leaf, 105 completeness relation, 50
classification, 134 complex rotation, 325
integrality condition, 117 composition, 147
momentum map, 109 conformal compactification, 220, 230, 311
of Diff(S 1 ), 180 conformal field theory, 6, see also AdS/CFT
of Poincaré group, 138 conformal Killing vector, 5
of semi-direct product, 130, 133 conformal transformation, 232, see also
of SL(2, R), 125 Möbius transformation
of Virasoro group, 182 conformal weight, 152
scalar, 130 conical deficit, 229, 236, 261, 269
spinning, 132 conjugacy class, 52, 187
coadjoint representation, 100 elliptic/hyperbolic/parabolic, 191
in gravity, 239, 283 conjugate representation, 47
of Diff(S 1 ), 155 connected BMS3 group, 276
of BMS3 , 280 connected Lorentz group, 72
of centrally extended group, 172 connected Poincaré group, 73
of semi-direct product/sum, 129 connection, 112
of Virasoro group/algebra, 175 conserved superpotential, 223
of W3 algebra, 344 constraint, 139
coadjoint vector, 100 continuous representation, 19
for BMS3 , 280 continuous spin, 78
for semi-direct product, 128, 273 for BMS particles, 311
for Virasoro, 175 contragredient representation, 65
cochain coordinates
Lie derivative, 157 Bondi, 2, 4, 258
local, 158, 162 cylindrical, 227
on group, 27 light-cone, 228
on Lie algebra, 22 on orbit, 328
cocycle correspondence principle, 113
differentiation, 28 coset space, see quotient space
integration, 29 cosmic censorship, 237, 270
on group, 27 cosmological constant, 227
on Lie algebra, 23 cotangent bundle, 102
trivial, 23, 27 momentum map, 109
cocycle condition, 19, 161 of momentum orbit, 131
codimension parity, 181 tangent space, 134
cohomology cotangent space, 131
interpretation, 28 counting degrees of freedom, 219
non-local, 159 covariant derivative, 112
of bms3 , 282 covariant phase space, 238
of groups, 27 as coadjoint representation, 239
of Lie algebras, 23 cross product, 129, 273, 274
Index 411
cross ratio, 169 disintegration of a measure, 308

curvature of a connection, 112 distribution, 154
cylindrical coordinates, 227, 258 double zero, 180
dressed particle, 252
D dual representation, 65
Darboux theorem, 103 dual space, 63, 153
de Rham cohomology, 23 dual stress tensor, 284
Dedekind eta function, 250
deformation of Lie algebra, 24 E
degenerate conical deficit, 234, 266 Einstein-Hilbert action, 218
degenerate parabolic matrix, 193 variation, 220
degrees of freedom, 218, 342 electric charge, 223
delta function, 48, 304, 326 elliptic matrix, 191
density matrix, 84 elliptic Virasoro orbit, 193
density on the circle, 152 energy functional, 207
derivation, 101 on elliptic Virasoro orbit, 211
of Lie algebra, 24 on parabolic Virasoro orbit, 213
descendant state, 246 on vacuum Virasoro orbit, 211
for sl(2, R), 243 unbounded, 214
diffeomorphism, 147 energy spectrum, 321
infinitesimal, 150 energy-momentum, 73
differentiable action, 220 entanglement, 36
for AdS3 gravity, 232 due to Wigner rotations, 84
differential, 98 equations of motion
for group, 27 for gravitational field, 218
for Lie algebra, 23 on Poisson manifold, 101
differentiation of cocycles, 28 equivalent measures, 36
Diff(S 1 ), 147 and representations, 42
action on one-forms, 152 equivalent representations, 58
adjoint representation, 150, 275 equivariant function, 148
central extension, see Virasoro group Erlangen programme, 8
coadjoint orbit, 180 Euclidean group, 63
coadjoint representation, 155 Euclidean time, 333
cohomology, 163 Euler-Poisson equation, 104
connected components, 148 exact representation, 18
exponential map, 156 exceptional constant, 183, 196
fundamental group, 148 deformation, 200, 203
orientation-preserving, 147 exceptional semi-direct product/sum, 271,
superrotation, 275 347
universal cover, 148 exhaustivity theorem, 70, 248
Diff(S 1 )/S 1 , 180 existence and uniqueness, 146
dimensional analysis, 92, 328 exponential map, 99
dimensional reduction, 221 for Diff(S 1 ), 156
dimensionful central charge, 92, 265, 299 extended BMS group, 6, 313
Dirac distribution, see delta function in 3D, see BMS3
direct integral, 64 extreme conical deficit, 229
discrete spin, 78 extrinsic curvature, 232
Index 412
F Frobenius formula, 53, 324, see also char-

Fabri-Picasso theorem, 253 acter
fall-off conditions, 219 and fixed points, 54
BMS3 , 262 for semi-direct products, 69
Brown-Henneaux, 230 Frobenius’ theorem, 101
Fang-Fronsdal action, 358 Fronsdal action, 334
Farey tail, 252 functional determinant, 332, 336, 358,
Fefferman-Graham expansion, 230, 233 360
fermion, 21 fundamental group
heat kernel, 359 of Diff(S 1 ), 148
higher spin, 358 of Lie group, 20
partition function, 358, 360 fundamental representation, 362
Fh (S 1 ) (space of densities), 152 fundamental vector field, see infinitesi-
dual space, 154 mal generator
fibre bundle, 111
finite measure, 32 G
fixed point, 54, 80 G n A, see semi-direct product
flat connection, 112 Galilean conformal algebra, see gca2
flat limit, 287 Galilean flat W3 algebra, 355
as ultrarelativistic limit, 317 Galilean highest-weight reps, 321
of AdS/CFT, 290 Galilean limit, see non-relativistic limit
of highest-weight reps, 316 Galilean particle, 90, 93, 140
of normal ordering, 354 Galilean vacuum, 322
of sl(2, R), 316 Galilei group, 90, see also Bargmann
of Virasoro character, 327 group
of Virasoro group, 287 projective representation, 95
of Virasoro reps, 321 γ-tracelessness, 359
of W3 algebra, 353 gauge transformation, 223, 334, 338
of Witt algebra, 288 gauged Sigma model, 123
flat space cosmology, 269 gca2 , 291, 321
flat space holography, 7, 290 higher spin version, 355
modular transformations, 326 Gelfand-Fuks cocycle, 157
flat W3 algebra, 347, 354 and Bott-Thurston, 162
character, 349 generic constant, 183, 194
Galilean version, 355 geometric action functional, 119
induced representation, 347, 356 and gauge invariance, 124
little groups, 348 and Maurer-Cartan form, 123
vacuum module, 356 as non-linear Sigma model, 123
flat WN algebra, 350 for semi-direct product, 137
vacuum character, 352 on group, 122
flow, 104, 155 geometric quantization, 51
foliation, 101 and boundary gravitons, 251, 304
forbidden gauge transformation, 223 and line bundles, 116
Fourier series, 176, 325 and unitary reps, 117, 137, 245
Fréchet space, 146 ghost, 338, 342, 358
free boundary conditions, 235 Gibbons-Hawking term, 221, 232
free particle, 141 global BMS group, 5, 311
Index 413
in 3D, 312 character, 249, 250

Godbillon-Vey cocycle, 159 exhaustivity, 248
Goldstone boson, 253 flat limit, 316
Gram matrix, 247 of gca2 , 321
gravitational dressing, 298, 306 of sl(2, R), 243
gravitational wave, 219, 384 of Virasoro, 246
gravitino, 365 unitarity, 248
gravity in 3D, 9, 218 vacuum, 248
group, 17 Hilbert space, 15
central extension, 19, 29 ray, 16
perfect, 30 rigged, 50
simply connected, 20 Hill’s equation, 184
group action, 38 inequivalent solutions, 188
Hamiltonian, 108 monodromy, 186
symplectic, 107 holography, 6
and topological field theories, 221
H in AdS3 , 238
Haar measure, 39 in flat space, 7, 290
Hamiltonian, 51, 101 holonomy, 47
in AdS3 , 234 homogeneous space, 38
in flat space, 266 as coset space, 44
Hamiltonian group action, 108 homotopy, 148
Hamiltonian vector field, 101, 103 hot flat space, 333
harmonic analysis, 46 hyperbolic matrix, 191
harmonic oscillator, 126, 142 hyperbolic Virasoro orbit, 198, 199
heat kernel, 332, 359 hypergravity, 368
traceless, 335
Heisenberg algebra, 26 I
Heisenberg group, 30, 91 I + (future null infinity), 260
Hermitian operator, 16 i regularization, 325, 326, 329, 336, 340,
Hermitian structure, 112 361
Hermiticity conditions imprimitivity, 56
for BMS3 , 318 improved action, 232
for flat W3 , 354 Inönü-Wigner contraction, 92, 289
for Poincaré, 314 induced module, 313, 315, 319
for sl(2, R), 243 flat limit, 317
for Virasoro, 246 for bms3 algebra, 319
for W3 , 345 for flat W3 algebra, 356
higher spin, 334, 358 for Poincaré algebra, 315
asymptotic symmetries, 343, 346 for Virasoro algebra, 247
partition function, 340, 346, 360 induced representation, 45, 310
supermomentum, 351 and imprimitivity, 56
surface charge, 343, 346 character, 53
vacuum configuration, 351 equivalence, 58
highest-weight representation, 320, 337 exhaustive, 70
at c ≤ 1, 248 in terms of fcts on a group, 46
by quantization, 245 in terms of plane waves, 50, 69
Index 414
irreducible, 58 for Virasoro, 177

notation, 45 Kirillov-Kostant symplectic form, 106
of Bargmann group, 94 and Maurer-Cartan, 121
of BMS3 , see BMS3 particle for semi-direct product, 135
of flat W3 algebra, 347 for SL(2, R), 126
of Lie algebra, see induced module for Virasoro, 245
of Poincaré group, 77 Korteweg-de Vries equation, 105
of semi-direct product, 67, 69
of super semi-direct product, 362 L
induction in stages, 48 L2 space, 35
induction-reduction theorem, 306 Lagrange multiplier, 139
inertia operator, 105 large central charge, 245
inertial coordinates, 258 large ` limit, 287
infinite spin, 78 leaking wavefunction, 305
infinite-dimensional group, 146, 277 Lebesgue integral, 33
infinitesimal cocycle, 29 Lefschetz number, 54, 80
infinitesimal generator, 106 left-invariant vector field, 98
infinitesimal Schwarzian derivative, 163 Legendre transformation, 139
infinitesimal Souriau cocycle, 172 Leibniz rule, 100, 112
infinitesimal transformation, 153 level (of descendant state), 243, 247
infinity, 219, 228 Lie algebra, 98
inhomogeneous Lorentz group, see Poincaré deformation, 24
group of BMS3 , see bms3
integral curve, 99, 155 of Poincaré in 3D, 259
integral symplectic form, 116 perfect, 24
integration of cocycles, 29 representation, 22
intertwiner, 42, 47, 58 Lie bracket, 98
invariant measure, 39 for centrally extended algebra, 172
invariant polarization, 117 for semi-direct sum, 128, 272
invariant vector, 23, 28 of vector fields, 151
irreducible representation, 58, 356 Lie derivative, 157, 229
isometry algebra, 227, 259 Lie group, 98
Iwasawa decomposition, 86 Lie-Fréchet group, 146, 277
lift (of diffeomorphism), 148
J light-cone coordinates, 228
Jacobi identity, 100 Minkoswki metric, 231
Jacobian, 40 light-like infinity, see null infinity
and delta functions, 49 light-like vector, 71
line bundle, 68, 111
K Liouville one-form, 102, 134
Kac-Moody algebra, 162 Liouville theory, 221
Kac-Moody level, 120 Liouville volume form, 102
ket, 15 little group, 65, see also stabilizer
Killing vector, 227, 258 for BMS3 particle, 295
Kirillov-Kostant bracket, 104 for flat W3 , 348
applied to AdS3 , 239 for higher spin vacuum, 351
for BMS3 , 281 for non-relativistic particle, 93
for SL(2, R), 125
Index 415
for relativistic particle, 74, 88 Lorentz-invariant, 33

fundamental representation, 362 on Virasoro orbit, 300, 326, 329
local cochain, 158, 162 projection-valued, 55
localization, 54, 324, 328 quasi-invariant, 40
Lorentz group, 2, 5, 72 method of images, 333, 335, 359
as projective group, 168 metric, see AdS3 , Minkowski space, on-
at null infinity, 261 shell metric
in 3D, 86 Milne universe, 269
topology, 72 Minkowski space, 2, 71, 258
Lorentz-invariant delta function, 49 as homogeneous space, 258
Lorentz-invariant measure, 33 light-cone coordinates, 231
Möbius transformation, 5, 169
M Modave, 97
map of Virasoro orbits, 205, 215 modified Lie bracket, 232
mass, 9 modular parameter, 252
as central charge, 92 modular transformation, 326
in AdS3 , 234 momentum, 64, 128
in flat space, 266 admissible, 362
of BMS3 particle, 297 eigenstate, see plane wave
mass gap, 237 for Bargmann group, 92
mass operator, 297, 318 for BMS3 , see supermomentum
mass shell, 139 for Poincaré group, 73
mass squared, 74, 314 map, 107
for BMS3 , 297, 318 orbit, 65, 74, 131
massive BMS3 particle, 295, see also el- monodromy, 186
liptic Virasoro orbit as bms3 Casimir, 318
character, 326 as Wilson loop, 187
factorization, 308 projective, 189
massive bms3 module, 319 multiplet, 363
massive particle, 74, 77, 94 multiply connected, 20
in 3D, 89
partition function, 81 N
massless BMS3 particle, 295, see also Neveu-Schwarz sector, 363
parabolic Virasoro orbit new massive gravity, 219, 235, 268
massless particle, 75, 78, 93 Newton constant, 218
in 3D, 89 Newton identities, 370
partition function, 82 nilpotent, 23
Maurer-Cartan form, 120 no gravitons in 3D, 218, 342
and geometric action, 123 Hamiltonian counting, 219
and Kirillov-Kostant, 121 Noether’s theorem, 108, 222
for semi-direct product, 137 non-linear Sigma model, 123
maximal torus, 31 non-local cohomology, 159
maximally symmetric, 218, 227, 258 non-relativistic limit, 291, 323, 355
measure, 32 non-relativistic mass, 92
disintegration, 308 non-relativistic particle, 90, 93, 140
equivalence, 36 non-trivial gauge transformation, 224
invariant, 39 non-unitary representation, 322
Index 416
normal ordering, 322, 345, 357 non-degenerate, 202

flat limit, 354 parity, 147, 258
normal subgroup, 62 particle, 8
null infinity, 3, 260 as induced representation, 68
for AdS3 , 286 by geometric quantization, 136
Poincaré transformations, 261 classification, 70
null orbifold, 266, 269 dressed with gravitons, 252
null state, 34, 248 in AdS3 , 251
null string, 323 scalar, 67
null vector, 71 spinning, 69
with definite momentum, 69
O partition function, 51, 141, 332, 341
O(2, 2), 226 for 3D gravity, 252, 340
observable, 103, 114 for fermions, 358, 360
on-shell metric, 233, 265 for higher spins, 346, 350, 352, 360
one-cocycle, 28 for hypergravity, 369
anomaly, 43 for massive particle, 81, 90
Radon-Nikodym derivative, 43 for supergravity, 368
Schwarzian derivative, 164 partition of integers, 247
Souriau, 165 path integral, 141, 332
one-form, 152 Penrose diagram, 3, 228
one-loop determinant, 336, 360 perfect group, 30
one-loop exactness, 252, 341 perfect Lie algebra, 24, 281
one-loop partition function, see parti- universal central extension, 26
tion function PGL(2, R), see PSL(2, R)
one-particle state, see particle phase space, 100
open set, 33, 115, 119, 181 as coadjoint representation, 239, 283
orbifold, 266, 340 covariant, 238
orbit, 38, 65 foliation, 240
coordinates, 325, 328 Planck mass, 218, 265, 299
for BMS3 , 295 plane wave, 49, 69, 304
for Poincaré, 74 p(N ) (partition of integers), 247
representative, 70 Poincaré algebra, 259, 314
orbit method, 51 Poincaré character, 80
orbital angular momentum, 129 in 3D, 89
orientation-preserving, 147 in a partition function, 341
orthochronous Lorentz group, 72 Poincaré group, 2, 63, 72, 258
P at null infinity, 261
p0 (constant supermomentum), 182 coadjoint orbits, 138
exceptional, 196 embedded in BMS3 , 279
generic, 194 in 3D, 87
parabolic matrix, 191 topology, 73
parabolic Virasoro orbit, 202 Poincaré momentum, 73
and exceptional constants, 196 Poincaré vacuum, 77
by deformation of a constant, 203 point at infinity, 167
degenerate, 196 point mass, 229
energy functional, 213 point particle, 237
Index 417
point transformation, 110 quotient of AdS3 , 238

Poisson algebra, 101 quotient space, 38, 44, 333
Poisson bracket, 100, 103 of AdS3 , 238
and commutator, 113 thermal, 332
polarization, 114, 116, 117
polarized observable, 115 R
positive energy theorem, 242, 285 Radon-Nikodym derivative, 37, 40
for BMS3 particles, 296 as anomaly, 40
prequantization, 114, 116 as cocycle, 43
primary field, 152 Radon-Nikodym theorem, 36
primary state, 246, 320 Ramond sector, 363
principal bundle, 44 ray, 16
projection-valued measure, 55 reduced density matrix, 84
projective group, 18, 168 Regge calculus, 341
projective line, 167 regular dual, 154
projective monodromy, 189 regular representation, 42
projective representation, 18 regular semi-direct product, 71, 301
classification, 22 relative cocycle, 170
of BMS3 , 303 relativistic momentum, 73
of Galilei group, 95 relativistic particle, see particle
of SL(2, R), 244 representation, 18
of U(1), 21 adjoint, 99
projective space, 16 by quantization, 137
projective transformation, 168 conjugate, 47
boost, 212 continuous, 19
projector, 55 dual, 65
propagator, 96, 142 geometrization, 51
proper Lorentz group, 72 induced, 45
PSL(2, R), 86, 168 irreducible, 58, 356
n-fold cover, 170, 197 non-unitary, 322
two-cocycle, 171 of BMS3 , see BMS3 particle
pullback, 103 of Lie algebra, 22
pure boost, 76, 83, 212 of Poincaré, see particle
of semi-direct product, 67, 70
Q of sl(2, R), see highest-weight rep-
quantization conditions, 116 resentation
from path integral, 119 of super semi-direct product, 362
quantum flat W3 algebra, 354 of vector group, 63
quantum gravity, 382 of Virasoro, see highest-weight rep-
and BMS3 particles, 304 resentation
quantum state, 16 projective, 18
quantum W3 algebra, 345 quasi-regular, 41
quasi-invariant measure, 40 regular, 42
on Virasoro orbit, 300 unitary, 18
quasi-primary field, 176 rest frame, 74, 93, 315, 319
quasi-regular representation, 41 for flat W3 algebra, 356
as affine module, 43
Index 418
retarded Bondi coordinates, see Bondi unitary representations, 70

coordinates semi-direct sum, 127
retarded time, 3, 258 coadjoint representation, 129
Riemann tensor, 218 exceptional, 271
rigged Hilbert space, 50 Lie bracket, 128
rigid rotation, 147 separable Hilbert space, 16
Rn , 32, 37 σ-additivity, 32, 56
robustness of induced reps, 46 σ-finite measure, 32
rotation, 62, 79 sigma model, 123, 137
action on momenta, 65 simple zero, 180
action on supermomenta, 325 simply connected, 20
1
RP (projective line), 167 SL(2, C), 5
SL(2, R), 85
S and harmonic oscillators, 126
S 1 (circle), 146 and SO(2, 1)↑ , 86
as projective line, 167 character, 127
projective transformation, 168 coadjoint orbits, 125
stereographic coordinate, 167 conjugacy classes, 191
universal cover, 21, 146 Iwasawa decomposition, 86
saddle point approximation, 332 Kirillov-Kostant bracket, 125
scalar particle, 67, 302 Kirillov-Kostant form, 126
scalar product, 35 projective representation, 244
Schwarzian derivative, 164 topology, 85
and Bott-Thurston, 166 sl(2, R), 86, 124, 227, 242
as one-cocycle, 164 Casimir operator, 244
as Souriau cocycle, 166 commutation relations, 124
average lemma, 208 descendant state, 243
cocycle identity, 165 flat limit, 316
infinitesimal, 163 Hermiticity conditions, 243
PSL(2, R)-invariant, 169 highest-weight representation, 243
section, 111 ultrarelativistic limit, 317
self-adjoint operator, 16 smooth dual, 154
semi-classical limit, 245 SNAG theorem, 70
in symplectic geometry, 117 SO(2, 1)↑ , 72, 86
semi-direct product, 2, 62 SO(2, 2), 226
adjoint representation, 128 so(2, 2), 227
and space-time, 138 soft graviton, 9, 307, 312
character, 69 solution vector, 186
coadjoint orbit, 130, 133 so(n)
coadjoint representation, 129 Cartan subalgebra, 372, 373
coadjoint vector, 128 character, 337, 373, 378, 379
exceptional, 271 highest weight, 337
geometric action functional, 137 Souriau cocycle, 165
Kirillov-Kostant form, 135 and Schwarzian derivative, 166
Maurer-Cartan form, 137 infinitesimal, 172
quantization conditions, 135 space-like vector, 71
regular, 71, 301 spatial infinity, 228
Index 419
spherical harmonic, 5 for higher spins, 347

spin, 9, 77, 94 supermultiplet, 363, 367
classical, 132 superpotential, 223
for BMS3 particle, 309 superrotation, 6, 264, 275, 276
spin angular momentum, 129 supertranslation, 2, 5, 264, 276, 277
spin group, 72 surface charge, 223, 224
spin-statistics, 21 algebra, 234, 267
spinning particle, 69 as dual Noether charge, 234
spinning vacuum, 77 as symmetry generator, 225
for BMS3 , 322 for 3D supergravity, 365
square integrability, 34 for higher spins, 343, 346
stabilizer, 38, 65, see also little group for superrotation, 267
for Virasoro coadjoint vector, 182 for supertranslation, 266
standard boost, 44, 47, 194 in AdS3 , 234
for BMS3 , 310 in flat space, 266
for Poincaré, 76 symmetrization, 334
standard decomposition theorem, 72 symmetry, 1, 16
stereographic coordinate, 3, 167 and unitarity, 8
Stokes’ theorem, 220 asymptotic, see asymptotic symme-
strength tensor, 223 try
stress tensor, 175 symmetry enhancement, 6
dual, 284 symmetry generator, 225
for BMS3 -invariant theories, 280 symmetry representation thm, 8, 17
transformation law, 176 symplectic form, 102
structure constant, 98 integral, 116
Sturm-Liouville operator, 185 symplectic group action, 107
super bms3 algebra, 364 symplectic induction, 135
super BMS3 character, 368 symplectic leaf, 101, 347
super BMS3 multiplet, 367 in AdS3 , 240
super centre of mass, 280 in flat space, 285
super Lie group, 361 symplectic manifold, 102
super Virasoro algebra, 363 codimension parity, 181
super Witt algebra, 363 symplectomorphism, 103
superboost, 276, see also superrotation system of imprimitivity, 56
supergravity, 364
supermomentum, 280, 294 T
admissible, 366 tachyon, 75, 78
angular, 280 tachyonic BMS3 particle, 200, 295, see
for higher spins, 351 also hyperbolic Virasoro orbit
for moving particle, 297 tensor algebra, 362
gravitational dressing, 298 tensor product, 36, 308
Lorentz-inequivalent, 305 thermal quotient, 332
massless, 203 Thomas precession, 84, see also Wigner
tachyonic, 200 rotation
transformation under boosts, 298 three-dimensional gravity, see 3D grav-
supermomentum orbit, 295 ity
coordinates, 325, 328 time slicing, 141
Index 420
time-like vector, 71 for flat W3 algebra, 356

topological central extension, 20 for Virasoro algebra, 248
topological field theory, 219 vacuum orbit
and holography, 221 energy, 211
topologically massive gravity, 219, 235, for BMS3 , 307
268 for Virasoro, 196
traceless heat kernel, 335 vacuum with spin, 77
transitive action, see homogeneous space Van Est theorem, 29
translation, 1, 62, 260 variation of action functional, 220
trivial bundle, 44, 111 vector bundle, 111
trivial central extension, 19, 26, 29 vector field
trivial cocycle, 23, 27 generating asymptotic symmetries,
trivial gauge transformation, 224 see asymptotic Killing vector
twisted derivative, 159, 160, 163 Hamiltonian, 101, 103
two-cocycle, see central extension left-invariant, 98
vector group, 62
U unitary representation, 63
U(1), 21 Vect(S 1 ), 151
ubiquitousness of induced reps, 31 cohomology, 157, 159, 163
ultrarelativistic limit, 317, see also flat supersymmetric version, 363
limit Verma module, 247
uncertainty principle, 111 Virasoro algebra, 174
unitary representation, 18, see also in- Casimir operators, 245
duced representation character, 249
by quantization, 117, 137 coadjoint representation, 176
of BMS3 , see BMS3 particle commutation relations, 174
of Poincaré, see particle descendant state, 246
of vector group, 63 Hermiticity conditions, 246
universal central extension, 26, 30, 174 highest-weight rep, 246, 320
universal cover, 20 non-relativistic limit, 323
of BMS3 group, 277 supersymmetric version, 363
of Diff(S 1 ), 148 Virasoro character, 250
of Lorentz group, 72 Virasoro cocycle, see Gelfand-Fuks
of Poincaré group, 73 Virasoro energy, 207
of S 1 , 21, 146 on parabolic orbits, 213
universal Teichmüller space, 196 on vacuum orbit, 211
universality, 328 unbounded from above, 208
V Virasoro group, 173
vacuum character coadjoint orbit, 181
for BMS3 , 329 coadjoint representation, 175
for flat W3 , 350 flat limit, 287
for flat WN , 352 Kirillov-Kostant bracket, 177
for Virasoro, 251 Virasoro measure, 300
for WN , 346 Virasoro orbits, 182
vacuum coadjoint vector, 183 classification, 191
vacuum module Kirillov-Kostant form, 245
for bms3 algebra, 320 map, 205, 215
Index 421
wedge algebra, 346, 356

massless, 204 weight, 152
measure, 300 Wess-Zumino-Witten model, 120
quantization, 245 Whitehead’s lemma, 25
stabilizer, 206 Wigner rotation, 46, 69, 83, 310
with constant representative, 204 as holonomy, 47
with positive energy, 211, 214 producing entanglement, 84
Virasoro stabilizer, 182, 188 Wilson loop, 187
Virasoro vacuum, 183, 248 winding number, 190
character, 251, 252 for elliptic Virasoro orbit, 195
descendant state, 248 Witt algebra, 152, 232
flat limit, 288
W
non-relativistic limit, 291
W3 algebra, 345
supersymmetric version, 363
flat limit, 353
world line, 137, 139
non-relativistic limit, 355
for BMS3 particle, 383
WN algebra
non-relativistic, 140
character, 346
on coadjoint orbit, 123
coadjoint representation, 344
on symplectic manifold, 119
warped CFT, 64, 163, 164, 273
world line formalism, 140
wavefunction, 34, 302
Wronskian, 186
in momentum space, 66
on group manifold, 46 Z
scalar product, 35 Z-equivariance, 148
wavefunctional, 302 zero-mode metric, 235, 268

BMS Particlesin Three Dimensions

Uploaded by

Copyright:

Available Formats

BMS Particlesin Three Dimensions

Uploaded by

Document Information

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

BMS Particlesin Three Dimensions

Uploaded by

Copyright:

Available Formats

BMS Particles

This thesis is devoted to the group-theoretic aspects of three-dimensional

faculté des sciences

service de physique théorique et mathématique

Thèse présentée en vue de l’obtention du titre de Docteur en Sciences

Réalisée sous la direction du Prof. Glenn Barnich

Année académique 2015-2016

Prof. Guillaume Bossard,

Typeset in LATEX with the memoir class.

Note: sections marked with an asterisk may be skipped on a first reading.

3.1.2 Hilbert spaces of wavefunctions.......................................................... 34

5 Coadjoint orbits and geometric quantization 97

5.1.3 Poisson structures ............................................................................. 100

II Virasoro symmetry and AdS3 gravity 143

6.4.2 Virasoro group.................................................................................. 173

7 Virasoro coadjoint orbits 179

8 Symmetries of gravity in AdS3 217

III BMS3 symmetry and gravity in flat space 255

10 Quantum BMS3 symmetry 293

11 Partition functions and characters 331

11.1.2 Bosonic higher spins ........................................................................ 334

How to read this thesis

• H. Afshar, S. Detournay, D. Grumiller, and B. Oblak, “Near-Horizon Geometry

• A. Campoleoni, H. A. González, B. Oblak, and M. Riegler, “BMS Modules

The friendly environment at the physics department of the Université Libre de

Finally, I am grateful to Guillaume Bossard, Axel Kleinschmidt and Petr Tinyakov

Family and friends

Physics and art

Financial and logistic support

The quantization of gravity is one of the long-standing puzzles of theoretical physics.

1.1 Asymptotic BMS symmetry

In this thesis we shall be concerned with symmetries of gravitational systems, that

BMS = Lorentz n Supertranslations. (1.1)

Poincaré = Lorentz n Translations. (1.2)

The latter is a subgroup of BMS: the group of space-time translations is contained in

1.2 Global BMS and extended BMS

Here r is a space-like radial coordinate, z is a stereographic coordinate on the sphere

In terms of Bondi coordinates, the region reached by massless particles emitted at

Figure 1.2: A representation of celestial spheres on the Penrose diagram of Minkowski

There exists a parallel construction of Bondi coordinates which is convenient for

Asymptotic flatness and the BMS group

that there are two families of such diffeomorphisms:

u 7→ u + α(z, z̄), (1.7)

as asymptotic symmetries, at least infinitesimally.

The extension of the BMS group obtained by replacing Lorentz transformations by

Extended BMS = Superrotations n Supertranslations (1.9)

where now both factors of the semi-direct product are infinite-dimensional.

has come to be known as the Anti-de Sitter/Conformal Field Theory (AdS/CFT)

The case of a three-dimensional bulk space-time (D = 3) is especially important

Flat space holography

The most interesting incarnation of BMS symmetry is the four-dimensional one,

Holography as an Erlangen programme

1.4 BMS particles and soft gravitons

A BMS particle is an irreducible, unitary representation of the BMS group.

Recent developments in the study of BMS symmetry provide a simple interpreta-

understanding of BMS particles in four dimensions is still an open problem; it suggests

1.5 Plan of the thesis

As an application, in chapter 4 we describe the irreducible unitary representations

Remark. The tools used in chapters 2 to 4 rely on elementary group theory; we

Virasoro symmetry and AdS3 gravity