Linear Algebra Done Right: Sheldon Axler

Undergraduate Texts in Mathematics
Sheldon Axler
Linear Algebra
Done Right
Fourth Edition
Series Editors
Pamela Gorkin, Mathematics Department, Bucknell University, Lewisburg, PA, USA
Jessica Sidman, Mathematics and Statistics, Amherst College, Amherst, MA, USA
Advisory Board
Colin Adams, Williams College, Williamstown, MA, USA
Jayadev S. Athreya, University of Washington, Seattle, WA, USA
Nathan Kaplan, University of California, Irvine, CA, USA
Jill Pipher, Brown University, Providence, RI, USA
Jeremy Tyson, University of Illinois at Urbana-Champaign, Urbana, IL, USA
Undergraduate Texts in Mathematics are generally aimed at third- and fourth-

year undergraduate mathematics students at North American universities. These
texts strive to provide students and teachers with new perspectives and novel
approaches. The books include motivation that guides the reader to an appreciation
of interrelations among different aspects of the subject. They feature examples that
illustrate key concepts as well as exercises that strengthen understanding.
Sheldon Axler
Linear Algebra Done Right

Fourth Edition
Sheldon Axler
San Francisco, CA, USA
ISSN 0172-6056 ISSN 2197-5604 (electronic)

ISBN 978-3-031-41025-3 ISBN 978-3-031-41026-0 (eBook)
https://doi.org/10.1007/978-3-031-41026-0
Mathematics Subject Classification (2020): 15-01, 15A03, 15A04, 15A15, 15A18, 15A21
© Sheldon Axler 1996, 1997, 2015, 2024. This book is an open access publication.
Open Access This book is licensed under the terms of the Creative Commons Attribution-NonCommercial
4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncom-
mercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you
give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons
license and indicate if changes were made.
The images or other third party material in this book are included in the book’s Creative Commons
license, unless indicated otherwise in a credit line to the material. If material is not included in the book’s
Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the
permitted use, you will need to obtain permission directly from the copyright holder.
This work is subject to copyright. All commercial rights are reserved by the author(s), whether the whole
or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed. Regarding these commercial rights a non-exclusive
license has been granted to the publisher.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland.
Paper in this product is recyclable.

About the Author
Sheldon Axler received his undergraduate degree from Princeton University,

followed by a PhD in mathematics from the University of California at Berkeley.
As a postdoctoral Moore Instructor at MIT, Axler received a university-wide
teaching award. He was then an assistant professor, associate professor, and
professor at Michigan State University, where he received the first J. Sutherland
Frame Teaching Award and the Distinguished Faculty Award.
Axler received the Lester R. Ford Award for expository writing from the Math-
ematical Association of America in 1996, for a paper that eventually expanded into
this book. In addition to publishing numerous research papers, he is the author of
six mathematics textbooks, ranging from freshman to graduate level. Previous
editions of this book have been adopted as a textbook at over 375 universities and
colleges and have been translated into three languages.
Axler has served as Editor-in-Chief of the Mathematical Intelligencer and
Associate Editor of the American Mathematical Monthly. He has been a member
of the Council of the American Mathematical Society and of the Board of Trustees
of the Mathematical Sciences Research Institute. He has also served on the
editorial board of Springer’s series Undergraduate Texts in Mathematics, Graduate
Texts in Mathematics, Universitext, and Springer Monographs in Mathematics.
Axler is a Fellow of the American Mathematical Society and has been a
recipient of numerous grants from the National Science Foundation.
Axler joined San Francisco State University as chair of the Mathematics
Department in 1997. He served as dean of the College of Science & Engineering
from 2002 to 2015, when he returned to a regular faculty appointment as a
professor in the Mathematics Department.
Carrie Heeter, Bishnu Sarangi
The author and his cat Moon.
Cover equation: Formula for the 𝑛th Fibonacci number. Exercise 21 in Section 5D
uses linear algebra to derive this formula.
v
Contents
About the Author v
Preface for Students xii
Preface for Instructors xiii
Acknowledgments xvii
Chapter 1
Vector Spaces 1
1A 𝐑𝑛 and 𝐂𝑛 2
Complex Numbers 2
Lists 5
𝐅𝑛 6
Digression on Fields 10
Exercises 1A 10
1B Definition of Vector Space 12
Exercises 1B 16
1C Subspaces 18
Sums of Subspaces 19
Direct Sums 21
Exercises 1C 24
Chapter 2
Finite-Dimensional Vector Spaces 27
2A Span and Linear Independence 28
Linear Combinations and Span 28
Linear Independence 31
Exercises 2A 37
vi
Contents vii
2B Bases 39
Exercises 2B 42
2C Dimension 44
Exercises 2C 48
Chapter 3
Linear Maps 51
3A Vector Space of Linear Maps 52
Definition and Examples of Linear Maps 52
Algebraic Operations on ℒ(𝑉, 𝑊) 55
Exercises 3A 57
3B Null Spaces and Ranges 59
Null Space and Injectivity 59
Range and Surjectivity 61
Fundamental Theorem of Linear Maps 62
Exercises 3B 66
3C Matrices 69
Representing a Linear Map by a Matrix 69
Addition and Scalar Multiplication of Matrices 71
Matrix Multiplication 72
Column–Row Factorization and Rank of a Matrix 77
Exercises 3C 79
3D Invertibility and Isomorphisms 82
Invertible Linear Maps 82
Isomorphic Vector Spaces 86
Linear Maps Thought of as Matrix Multiplication 88
Change of Basis 90
Exercises 3D 93
3E Products and Quotients of Vector Spaces 96
Products of Vector Spaces 96
Quotient Spaces 98
Exercises 3E 103
3F Duality 105
Dual Space and Dual Map 105
Null Space and Range of Dual of Linear Map 109
viii Contents
Matrix of Dual of Linear Map 113

Exercises 3F 115
Chapter 4
Polynomials 119
Zeros of Polynomials 122
Division Algorithm for Polynomials 123
Factorization of Polynomials over 𝐂 124
Factorization of Polynomials over 𝐑 127
Exercises 4 129
Chapter 5
Eigenvalues and Eigenvectors 132
5A Invariant Subspaces 133
Eigenvalues 133
Polynomials Applied to Operators 137
Exercises 5A 139
5B The Minimal Polynomial 143
Existence of Eigenvalues on Complex Vector Spaces 143
Eigenvalues and the Minimal Polynomial 144
Eigenvalues on Odd-Dimensional Real Vector Spaces 149
Exercises 5B 150
5C Upper-Triangular Matrices 154
Exercises 5C 160
5D Diagonalizable Operators 163
Diagonal Matrices 163
Conditions for Diagonalizability 165
Gershgorin Disk Theorem 170
Exercises 5D 172
5E Commuting Operators 175
Exercises 5E 179
Chapter 6
Inner Product Spaces 181
6A Inner Products and Norms 182
Inner Products 182
Contents ix
Norms 186
Exercises 6A 191
6B Orthonormal Bases 197
Orthonormal Lists and the Gram–Schmidt Procedure 197
Linear Functionals on Inner Product Spaces 204
Exercises 6B 207
6C Orthogonal Complements and Minimization Problems 211
Orthogonal Complements 211
Minimization Problems 217
Pseudoinverse 220
Exercises 6C 224
Chapter 7
Operators on Inner Product Spaces 227
7A Self-Adjoint and Normal Operators 228
Adjoints 228
Self-Adjoint Operators 233
Normal Operators 235
Exercises 7A 239
7B Spectral Theorem 243
Real Spectral Theorem 243
Complex Spectral Theorem 246
Exercises 7B 247
7C Positive Operators 251
Exercises 7C 255
7D Isometries, Unitary Operators, and Matrix Factorization 258
Isometries 258
Unitary Operators 260
QR Factorization 263
Cholesky Factorization 266
Exercises 7D 268
7E Singular Value Decomposition 270
Singular Values 270
SVD for Linear Maps and for Matrices 273
Exercises 7E 278
x Contents
7F Consequences of Singular Value Decomposition 280

Norms of Linear Maps 280
Approximation by Linear Maps with Lower-Dimensional Range 283
Polar Decomposition 285
Operators Applied to Ellipsoids and Parallelepipeds 287
Volume via Singular Values 291
Properties of an Operator as Determined by Its Eigenvalues 293
Exercises 7F 294
Chapter 8
Operators on Complex Vector Spaces 297
8A Generalized Eigenvectors and Nilpotent Operators 298
Null Spaces of Powers of an Operator 298
Generalized Eigenvectors 300
Nilpotent Operators 303
Exercises 8A 306
8B Generalized Eigenspace Decomposition 308
Generalized Eigenspaces 308
Multiplicity of an Eigenvalue 310
Block Diagonal Matrices 314
Exercises 8B 316
8C Consequences of Generalized Eigenspace Decomposition 319
Square Roots of Operators 319
Jordan Form 321
Exercises 8C 324
8D Trace: A Connection Between Matrices and Operators 326
Exercises 8D 330
Chapter 9
Multilinear Algebra and Determinants 332
9A Bilinear Forms and Quadratic Forms 333
Bilinear Forms 333
Symmetric Bilinear Forms 337
Quadratic Forms 341
Exercises 9A 344
Contents xi
9B Alternating Multilinear Forms 346

Multilinear Forms 346
Alternating Multilinear Forms and Permutations 348
Exercises 9B 352
9C Determinants 354
Defining the Determinant 354
Properties of Determinants 357
Exercises 9C 367
9D Tensor Products 370
Tensor Product of Two Vector Spaces 370
Tensor Product of Inner Product Spaces 376
Tensor Product of Multiple Vector Spaces 378
Exercises 9D 380
Photo Credits 383
Symbol Index 384
Index 385
Colophon: Notes on Typesetting 390

Preface for Students
You are probably about to begin your second exposure to linear algebra. Unlike
your first brush with the subject, which probably emphasized Euclidean spaces
and matrices, this encounter will focus on abstract vector spaces and linear maps.
These terms will be defined later, so don’t worry if you do not know what they
mean. This book starts from the beginning of the subject, assuming no knowledge
of linear algebra. The key point is that you are about to immerse yourself in
serious mathematics, with an emphasis on attaining a deep understanding of the
definitions, theorems, and proofs.
You cannot read mathematics the way you read a novel. If you zip through a
page in less than an hour, you are probably going too fast. When you encounter
the phrase “as you should verify”, you should indeed do the verification, which
will usually require some writing on your part. When steps are left out, you need
to supply the missing pieces. You should ponder and internalize each definition.
For each theorem, you should seek examples to show why each hypothesis is
necessary. Discussions with other students should help.
As a visual aid, definitions are in yellow boxes and theorems are in blue boxes
(in color versions of the book). Each theorem has an infomal descriptive name.
Please check the website below for additional information about the book,
including a link to videos that are freely available to accompany the book.
Your suggestions, comments, and corrections are most welcome.
Best wishes for success and enjoyment in learning linear algebra!
Sheldon Axler
San Francisco State University
website: https://linear.axler.net
e-mail: [email protected]
xii
Preface for Instructors
You are about to teach a course that will probably give students their second
exposure to linear algebra. During their first brush with the subject, your students
probably worked with Euclidean spaces and matrices. In contrast, this course will
emphasize abstract vector spaces and linear maps.
The title of this book deserves an explanation. Most linear algebra textbooks
use determinants to prove that every linear operator on a finite-dimensional com-
plex vector space has an eigenvalue. Determinants are difficult, nonintuitive,
and often defined without motivation. To prove the theorem about existence of
eigenvalues on complex vector spaces, most books must define determinants,
prove that a linear operator is not invertible if and only if its determinant equals 0,
and then define the characteristic polynomial. This tortuous (torturous?) path
gives students little feeling for why eigenvalues exist.
In contrast, the simple determinant-free proofs presented here (for example,
see 5.19) offer more insight. Once determinants have been moved to the end of
the book, a new route opens to the main goal of linear algebra—understanding
the structure of linear operators.
This book starts at the beginning of the subject, with no prerequisites other
than the usual demand for suitable mathematical maturity. A few examples
and exercises involve calculus concepts such as continuity, differentiation, and
integration. You can easily skip those examples and exercises if your students
have not had calculus. If your students have had calculus, then those examples and
exercises can enrich their experience by showing connections between different
parts of mathematics.
Even if your students have already seen some of the material in the first few
chapters, they may be unaccustomed to working exercises of the type presented
here, most of which require an understanding of proofs.
Here is a chapter-by-chapter summary of the highlights of the book:
• Chapter 1: Vector spaces are defined in this chapter, and their basic properties
are developed.
• Chapter 2: Linear independence, span, basis, and dimension are defined in this
chapter, which presents the basic theory of finite-dimensional vector spaces.
• Chapter 3: This chapter introduces linear maps. The key result here is the
fundamental theorem of linear maps: if 𝑇 is a linear map on 𝑉, then dim 𝑉 =
dim null 𝑇 + dim range 𝑇. Quotient spaces and duality are topics in this chapter
at a higher level of abstraction than most of the book; these topics can be
skipped (except that duality is needed for tensor products in Section 9D).
xiii
xiv Preface for Instructors
• Chapter 4: The part of the theory of polynomials that will be needed to un-
derstand linear operators is presented in this chapter. This chapter contains no
linear algebra. It can be covered quickly, especially if your students are already
familiar with these results.
• Chapter 5: The idea of studying a linear operator by restricting it to small sub-
spaces leads to eigenvectors in the early part of this chapter. The highlight of this
chapter is a simple proof that on complex vector spaces, eigenvalues always ex-
ist. This result is then used to show that each linear operator on a complex vector
space has an upper-triangular matrix with respect to some basis. The minimal
polynomial plays an important role here and later in the book. For example, this
chapter gives a characterization of the diagonalizable operators in terms of the
minimal polynomial. Section 5E can be skipped if you want to save some time.
• Chapter 6: Inner product spaces are defined in this chapter, and their basic
properties are developed along with tools such as orthonormal bases and the
Gram–Schmidt procedure. This chapter also shows how orthogonal projections
can be used to solve certain minimization problems. The pseudoinverse is then
introduced as a useful tool when the inverse does not exist. The material on
the pseudoinverse can be skipped if you want to save some time.
• Chapter 7: The spectral theorem, which characterizes the linear operators for
which there exists an orthonormal basis consisting of eigenvectors, is one of
the highlights of this book. The work in earlier chapters pays off here with espe-
cially simple proofs. This chapter also deals with positive operators, isometries,
unitary operators, matrix factorizations, and especially the singular value de-
composition, which leads to the polar decomposition and norms of linear maps.
• Chapter 8: This chapter shows that for each operator on a complex vector space,
there is a basis of the vector space consisting of generalized eigenvectors of the
operator. Then the generalized eigenspace decomposition describes a linear
operator on a complex vector space. The multiplicity of an eigenvalue is defined
as the dimension of the corresponding generalized eigenspace. These tools are
used to prove that every invertible linear operator on a complex vector space
has a square root. Then the chapter gives a proof that every linear operator on
a complex vector space can be put into Jordan form. The chapter concludes
with an investigation of the trace of operators.
• Chapter 9: This chapter begins by looking at bilinear forms and showing that the
vector space of bilinear forms is the direct sum of the subspaces of symmetric
bilinear forms and alternating bilinear forms. Then quadratic forms are diag-
onalized. Moving to multilinear forms, the chapter shows that the subspace of
alternating 𝑛-linear forms on an 𝑛-dimensional vector space has dimension one.
This result leads to a clean basis-free definition of the determinant of an opera-
tor. For complex vector spaces, the determinant turns out to equal the product of
the eigenvalues, with each eigenvalue included in the product as many times as
its multiplicity. The chapter concludes with an introduction to tensor products.
Preface for Instructors xv
This book usually develops linear algebra simultaneously for real and complex
vector spaces by letting 𝐅 denote either the real or the complex numbers. If you and
your students prefer to think of 𝐅 as an arbitrary field, then see the comments at the
end of Section 1A. I prefer avoiding arbitrary fields at this level because they intro-
duce extra abstraction without leading to any new linear algebra. Also, students are
more comfortable thinking of polynomials as functions instead of the more formal
objects needed for polynomials with coefficients in finite fields. Finally, even if the
beginning part of the theory were developed with arbitrary fields, inner product
spaces would push consideration back to just real and complex vector spaces.
You probably cannot cover everything in this book in one semester. Going
through all the material in the first seven or eight chapters during a one-semester
course may require a rapid pace. If you must reach Chapter 9, then consider
skipping the material on quotient spaces in Section 3E, skipping Section 3F
on duality (unless you intend to cover tensor products in Section 9D), covering
Chapter 4 on polynomials in a half hour, skipping Section 5E on commuting
operators, and skipping the subsection in Section 6C on the pseudoinverse.
A goal more important than teaching any particular theorem is to develop in
students the ability to understand and manipulate the objects of linear algebra.
Mathematics can be learned only by doing. Fortunately, linear algebra has many
good homework exercises. When teaching this course, during each class I usually
assign as homework several of the exercises, due the next class. Going over the
homework might take up significant time in a typical class.
Some of the exercises are intended to lead curious students into important
topics beyond what might usually be included in a basic second course in linear
algebra.
The author’s top ten

Listed below are the author’s ten favorite results in the book, in order of their
appearance in the book. Students who leave your course with a good understanding
of these crucial results will have an excellent foundation in linear algebra.
• any two bases of a vector space have the same length (2.34)
• fundamental theorem of linear maps (3.21)
• existence of eigenvalues if 𝐅 = 𝐂 (5.19)
• upper-triangular form always exists if 𝐅 = 𝐂 (5.47)
• Cauchy–Schwarz inequality (6.14)
• Gram–Schmidt procedure (6.32)
• spectral theorem (7.29 and 7.31)
• singular value decomposition (7.70)
• generalized eigenspace decomposition theorem when 𝐅 = 𝐂 (8.22)
• dimension of alternating 𝑛-linear forms on 𝑉 is 1 if dim 𝑉 = 𝑛 (9.37)
xvi Preface for Instructors
Major improvements and additions for the fourth edition
• Over 250 new exercises and over 70 new examples.

• Increasing use of the minimal polynomial to provide cleaner proofs of multiple
results, including necessary and sufficient conditions for an operator to have an
upper-triangular matrix with respect to some basis (see Section 5C), necessary
and sufficient conditions for diagonalizability (see Section 5D), and the real
spectral theorem (see Section 7B).
• New section on commuting operators (see Section 5E).
• New subsection on pseudoinverse (see Section 6C).
• New subsections on QR factorization/Cholesky factorization (see Section 7D).
• Singular value decomposition now done for linear maps from an inner product
space to another (possibly different) inner product space, rather than only deal-
ing with linear operators from an inner product space to itself (see Section 7E).
• Polar decomposition now proved from singular value decomposition, rather than
in the opposite order; this has led to cleaner proofs of both the singular value
decomposition (see Section 7E) and the polar decomposition (see Section 7F).
• New subsection on norms of linear maps on finite-dimensional inner prod-
uct spaces, using the singular value decomposition to avoid even mentioning
supremum in the definition of the norm of a linear map (see Section 7F).
• New subsection on approximation by linear maps with lower-dimensional range
(see Section 7F).
• New elementary proof of the important result that if 𝑇 is an operator on a finite-
dimensional complex vector space 𝑉, then there exists a basis of 𝑉 consisting
of generalized eigenvectors of 𝑇 (see 8.9).
• New Chapter 9 on multilinear algebra, including bilinear forms, quadratic
forms, multilinear forms, and tensor products. Determinants now are defined
using a basis-free approach via alternating multilinear forms.
• New formatting to improve the student-friendly appearance of the book. For
example, the definition and result boxes now have rounded corners instead of
right-angle corners, for a gentler look. The main font size has been reduced
from 11 point to 10.5 point.
Please check the website below for additional links and information about the
book. Your suggestions, comments, and corrections are most welcome.
Best wishes for teaching a successful linear algebra class!
Sheldon Axler Contact the author, or Springer if the
San Francisco State University author is not available, for permission
website: https://linear.axler.net for translations or other commercial
e-mail: [email protected] reuse of the contents of this book.
Acknowledgments
I owe a huge intellectual debt to all the mathematicians who created linear algebra
over the past two centuries. The results in this book belong to the common heritage
of mathematics. A special case of a theorem may first have been proved long ago,
then sharpened and improved by many mathematicians in different time periods.
Bestowing proper credit on all contributors would be a difficult task that I have
not undertaken. In no case should the reader assume that any result presented
here represents my original contribution.
Many people helped make this a better book. The three previous editions of
this book were used as a textbook at over 375 universities and colleges around
the world. I received thousands of suggestions and comments from faculty and
students who used the book. Many of those suggestions led to improvements
in this edition. The manuscript for this fourth edition was class tested at 30
universities. I am extremely grateful for the useful feedback that I received from
faculty and students during this class testing.
The long list of people who should be thanked for their suggestions would
fill up many pages. Lists are boring to read. Thus to represent all contributors
to this edition, I will mention only Noel Hinton, a graduate student at Australian
National University, who sent me more suggestions and corrections for this fourth
edition than anyone else. To everyone who contributed suggestions, let me say
how truly grateful I am to all of you. Many many thanks!
I thank Springer for providing me with help when I needed it and for allowing
me the freedom to make the final decisions about the content and appearance
of this book. Special thanks to the two terrific mathematics editors at Springer
who worked with me on this project—Loretta Bartolini during the first half of
my work on the fourth edition, and Elizabeth Loew during the second half of my
work on the fourth edition. I am deeply indebted to David Kramer, who did a
magnificent job of copyediting and prevented me from making many mistakes.
Extra special thanks to my fantastic partner Carrie Heeter. Her understanding
and encouragement enabled me to work intensely on this new edition. Our won-
derful cat Moon, whose picture appears on the About the Author page, provided
sweet breaks throughout the writing process. Moon died suddenly due to a blood
clot as this book was being finished. We are grateful for five precious years with
him.
Sheldon Axler
xvii
Chapter 1
Vector Spaces
Linear algebra is the study of linear maps on finite-dimensional vector spaces.

Eventually we will learn what all these terms mean. In this chapter we will define
vector spaces and discuss their elementary properties.
In linear algebra, better theorems and more insight emerge if complex numbers
are investigated along with real numbers. Thus we will begin by introducing the
complex numbers and their basic properties.
We will generalize the examples of a plane and of ordinary space to 𝐑𝑛 and
𝐂𝑛, which we then will generalize to the notion of a vector space. As we will see,
a vector space is a set with operations of addition and scalar multiplication that
satisfy natural algebraic properties.
Then our next topic will be subspaces, which play a role for vector spaces
analogous to the role played by subsets for sets. Finally, we will look at sums
of subspaces (analogous to unions of subsets) and direct sums of subspaces
(analogous to unions of disjoint sets). Pierre Louis Dumesnil, Nils Forsberg
René Descartes explaining his work to Queen Christina of Sweden.

Vector spaces are a generalization of the description of a plane
using two coordinates, as published by Descartes in 1637.
© Sheldon Axler 2024
S. Axler, Linear Algebra Done Right, Undergraduate Texts in Mathematics,
1
https://doi.org/10.1007/978-3-031-41026-0_1
2 Chapter 1 Vector Spaces
1A 𝐑𝑛 and 𝐂𝑛
Complex Numbers
You should already be familiar with basic properties of the set 𝐑 of real numbers.
Complex numbers were invented so that we can take square roots of negative
numbers. The idea is to assume we have a square root of −1, denoted by 𝑖, that
obeys the usual rules of arithmetic. Here are the formal definitions.
1.1 definition: complex numbers, 𝐂
• A complex number is an ordered pair (𝑎, 𝑏), where 𝑎, 𝑏 ∈ 𝐑 , but we will

write this as 𝑎 + 𝑏𝑖.
• The set of all complex numbers is denoted by 𝐂 :
𝐂 = {𝑎 + 𝑏𝑖 ∶ 𝑎, 𝑏 ∈ 𝐑}.
• Addition and multiplication on 𝐂 are defined by
(𝑎 + 𝑏𝑖) + (𝑐 + 𝑑𝑖) = (𝑎 + 𝑐) + (𝑏 + 𝑑)𝑖,

(𝑎 + 𝑏𝑖)(𝑐 + 𝑑𝑖) = (𝑎𝑐 − 𝑏𝑑) + (𝑎𝑑 + 𝑏𝑐)𝑖;
here 𝑎, 𝑏, 𝑐, 𝑑 ∈ 𝐑 .
If 𝑎 ∈ 𝐑 , we identify 𝑎 + 0𝑖 with the real number 𝑎. Thus we think of 𝐑 as a

subset of 𝐂 . We usually write 0 + 𝑏𝑖 as just 𝑏𝑖, and we usually write 0 + 1𝑖 as just 𝑖.
To motivate the definition of complex The symbol 𝑖 was first used to denote
multiplication given above, pretend that √−1 by Leonhard Euler in 1777.
we knew that 𝑖2 = −1 and then use the
usual rules of arithmetic to derive the formula above for the product of two
complex numbers. Then use that formula to verify that we indeed have
𝑖2 = −1.
Do not memorize the formula for the product of two complex numbers—you
can always rederive it by recalling that 𝑖2 = −1 and then using the usual rules of
arithmetic (as given by 1.3). The next example illustrates this procedure.
1.2 example: complex arithmetic

The product (2 + 3𝑖)(4 + 5𝑖) can be evaluated by applying the distributive and
commutative properties from 1.3:
(2 + 3𝑖)(4 + 5𝑖) = 2 ⋅ (4 + 5𝑖) + (3𝑖)(4 + 5𝑖)
= 2 ⋅ 4 + 2 ⋅ 5𝑖 + 3𝑖 ⋅ 4 + (3𝑖)(5𝑖)
= 8 + 10𝑖 + 12𝑖 − 15
= −7 + 22𝑖.
Section 1A 𝐑𝑛 and 𝐂𝑛 3
Our first result states that complex addition and complex multiplication have
the familiar properties that we expect.
1.3 properties of complex arithmetic
commutativity
𝛼 + 𝛽 = 𝛽 + 𝛼 and 𝛼𝛽 = 𝛽𝛼 for all 𝛼, 𝛽 ∈ 𝐂 .
associativity
(𝛼 + 𝛽) + 𝜆 = 𝛼 + (𝛽 + 𝜆) and (𝛼𝛽)𝜆 = 𝛼(𝛽𝜆) for all 𝛼, 𝛽, 𝜆 ∈ 𝐂 .
identities
𝜆 + 0 = 𝜆 and 𝜆1 = 𝜆 for all 𝜆 ∈ 𝐂 .
additive inverse
For every 𝛼 ∈ 𝐂 , there exists a unique 𝛽 ∈ 𝐂 such that 𝛼 + 𝛽 = 0.
multiplicative inverse
For every 𝛼 ∈ 𝐂 with 𝛼 ≠ 0, there exists a unique 𝛽 ∈ 𝐂 such that 𝛼𝛽 = 1.
distributive property
𝜆(𝛼 + 𝛽) = 𝜆𝛼 + 𝜆𝛽 for all 𝜆, 𝛼, 𝛽 ∈ 𝐂 .
The properties above are proved using the familiar properties of real numbers
and the definitions of complex addition and multiplication. The next example
shows how commutativity of complex multiplication is proved. Proofs of the
other properties above are left as exercises.
1.4 example: commutativity of complex multiplication

To show that 𝛼𝛽 = 𝛽𝛼 for all 𝛼, 𝛽 ∈ 𝐂 , suppose
𝛼 = 𝑎 + 𝑏𝑖 and 𝛽 = 𝑐 + 𝑑𝑖,
where 𝑎, 𝑏, 𝑐, 𝑑 ∈ 𝐑 . Then the definition of multiplication of complex numbers

shows that
𝛼𝛽 = (𝑎 + 𝑏𝑖)(𝑐 + 𝑑𝑖)
= (𝑎𝑐 − 𝑏𝑑) + (𝑎𝑑 + 𝑏𝑐)𝑖
and
𝛽𝛼 = (𝑐 + 𝑑𝑖)(𝑎 + 𝑏𝑖)
= (𝑐𝑎 − 𝑑𝑏) + (𝑐𝑏 + 𝑑𝑎)𝑖.
The equations above and the commutativity of multiplication and addition of real
numbers show that 𝛼𝛽 = 𝛽𝛼.
Next, we define the additive and multiplicative inverses of complex numbers,

and then use those inverses to define subtraction and division operations with
complex numbers.
1.5 definition: −𝛼, subtraction, 1/𝛼, division
Suppose 𝛼, 𝛽 ∈ 𝐂 .
• Let −𝛼 denote the additive inverse of 𝛼. Thus −𝛼 is the unique complex
number such that
𝛼 + (−𝛼) = 0.
• Subtraction on 𝐂 is defined by
𝛽 − 𝛼 = 𝛽 + (−𝛼).
• For 𝛼 ≠ 0, let 1/𝛼 anddenote the multiplicative inverse of 𝛼. Thus 1/𝛼 is

1
𝛼
the unique complex number such that
𝛼(1/𝛼) = 1.
• For 𝛼 ≠ 0, division by 𝛼 is defined by
𝛽/𝛼 = 𝛽(1/𝛼).
So that we can conveniently make definitions and prove theorems that apply
to both real and complex numbers, we adopt the following notation.
1.6 notation: 𝐅
Throughout this book, 𝐅 stands for either 𝐑 or 𝐂 .
Thus if we prove a theorem involving The letter 𝐅 is used because 𝐑 and 𝐂

𝐅, we will know that it holds when 𝐅 is are examples of what are called fields.
replaced with 𝐑 and when 𝐅 is replaced
with 𝐂 .
Elements of 𝐅 are called scalars. The word “scalar” (which is just a fancy
word for “number”) is often used when we want to emphasize that an object is a
number, as opposed to a vector (vectors will be defined soon).
For 𝛼 ∈ 𝐅 and 𝑚 a positive integer, we define 𝛼𝑚 to denote the product of 𝛼
with itself 𝑚 times:
⏟.
𝛼𝑚 = 𝛼⋯𝛼
𝑚 times
This definition implies that

𝑛
(𝛼𝑚 ) = 𝛼𝑚𝑛 and (𝛼𝛽)𝑚 = 𝛼𝑚 𝛽𝑚
for all 𝛼, 𝛽 ∈ 𝐅 and all positive integers 𝑚, 𝑛.
Lists
Before defining 𝐑𝑛 and 𝐂𝑛, we look at two important examples.
1.7 example: 𝐑2 and 𝐑3

• The set 𝐑2, which you can think of as a plane, is the set of all ordered pairs of
real numbers:
𝐑2 = {(𝑥, 𝑦) ∶ 𝑥, 𝑦 ∈ 𝐑}.
• The set 𝐑3, which you can think of as ordinary space, is the set of all ordered
triples of real numbers:
𝐑3 = {(𝑥, 𝑦, 𝑧) ∶ 𝑥, 𝑦, 𝑧 ∈ 𝐑}.
To generalize 𝐑2 and 𝐑3 to higher dimensions, we first need to discuss the

concept of lists.
1.8 definition: list, length
• Suppose 𝑛 is a nonnegative integer. A list of length 𝑛 is an ordered collec-

tion of 𝑛 elements (which might be numbers, other lists, or more abstract
objects).
• Two lists are equal if and only if they have the same length and the same
elements in the same order.
Lists are often written as elements Many mathematicians call a list of

separated by commas and surrounded by length 𝑛 an 𝑛-tuple.
parentheses. Thus a list of length two is
an ordered pair that might be written as (𝑎, 𝑏). A list of length three is an ordered
triple that might be written as (𝑥, 𝑦, 𝑧). A list of length 𝑛 might look like this:
(𝑧1 , …, 𝑧𝑛 ).
Sometimes we will use the word list without specifying its length. Remember,
however, that by definition each list has a finite length that is a nonnegative integer.
Thus an object that looks like (𝑥1 , 𝑥2 , … ), which might be said to have infinite
length, is not a list.
A list of length 0 looks like this: ( ). We consider such an object to be a list
so that some of our theorems will not have trivial exceptions.
Lists differ from sets in two ways: in lists, order matters and repetitions have
meaning; in sets, order and repetitions are irrelevant.
1.9 example: lists versus sets

• The lists (3, 5) and (5, 3) are not equal, but the sets {3, 5} and {5, 3} are equal.
• The lists (4, 4) and (4, 4, 4) are not equal (they do not have the same length),
although the sets {4, 4} and {4, 4, 4} both equal the set {4}.
𝐅𝑛
To define the higher-dimensional analogues of 𝐑2 and 𝐑3, we will simply replace
𝐑 with 𝐅 (which equals 𝐑 or 𝐂 ) and replace the 2 or 3 with an arbitrary positive
integer.
1.10 notation: 𝑛
Fix a positive integer 𝑛 for the rest of this chapter.
1.11 definition: 𝐅𝑛 , coordinate
𝐅𝑛 is the set of all lists of length 𝑛 of elements of 𝐅 :
𝐅𝑛 = {(𝑥1 , …, 𝑥𝑛 ) ∶ 𝑥𝑘 ∈ 𝐅 for 𝑘 = 1, …, 𝑛}.
For (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛 and 𝑘 ∈ {1, …, 𝑛}, we say that 𝑥𝑘 is the 𝑘 th coordinate of

(𝑥1 , …, 𝑥𝑛 ).
If 𝐅 = 𝐑 and 𝑛 equals 2 or 3, then the definition above of 𝐅𝑛 agrees with our

previous notions of 𝐑2 and 𝐑3.
1.12 example: 𝐂4
𝐂4 is the set of all lists of four complex numbers:
𝐂4 = {(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 ) ∶ 𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 ∈ 𝐂}.
If 𝑛 ≥ 4, we cannot visualize 𝐑𝑛 as Read Flatland: A Romance of Many

a physical object. Similarly, 𝐂1 can be Dimensions, by Edwin A. Abbott, for
thought of as a plane, but for 𝑛 ≥ 2, the an amusing account of how 𝐑3 would
human brain cannot provide a full image be perceived by creatures living in 𝐑2.
of 𝐂𝑛. However, even if 𝑛 is large, we This novel, published in 1884, may
can perform algebraic manipulations in help you imagine a physical space of
𝐅𝑛 as easily as in 𝐑2 or 𝐑3. For example, four or more dimensions.
addition in 𝐅𝑛 is defined as follows.
1.13 definition: addition in 𝐅𝑛

Addition in 𝐅𝑛 is defined by adding corresponding coordinates:
(𝑥1 , …, 𝑥𝑛 ) + (𝑦1 , …, 𝑦𝑛 ) = (𝑥1 + 𝑦1 , …, 𝑥𝑛 + 𝑦𝑛 ).
Often the mathematics of 𝐅𝑛 becomes cleaner if we use a single letter to denote

a list of 𝑛 numbers, without explicitly writing the coordinates. For example, the
next result is stated with 𝑥 and 𝑦 in 𝐅𝑛 even though the proof requires the more
cumbersome notation of (𝑥1 , …, 𝑥𝑛 ) and (𝑦1 , …, 𝑦𝑛 ).
1.14 commutativity of addition in 𝐅𝑛
If 𝑥, 𝑦 ∈ 𝐅𝑛, then 𝑥 + 𝑦 = 𝑦 + 𝑥.
Proof Suppose 𝑥 = (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛 and 𝑦 = (𝑦1 , …, 𝑦𝑛 ) ∈ 𝐅𝑛. Then
𝑥 + 𝑦 = (𝑥1 , …, 𝑥𝑛 ) + (𝑦1 , …, 𝑦𝑛 )
= (𝑥1 + 𝑦1 , …, 𝑥𝑛 + 𝑦𝑛 )
= (𝑦1 + 𝑥1 , …, 𝑦𝑛 + 𝑥𝑛 )
= (𝑦1 , …, 𝑦𝑛 ) + (𝑥1 , …, 𝑥𝑛 )
= 𝑦 + 𝑥,
where the second and fourth equalities above hold because of the definition of
addition in 𝐅𝑛 and the third equality holds because of the usual commutativity of
addition in 𝐅.
If a single letter is used to denote an The symbol means “end of proof ”.

element of 𝐅𝑛, then the same letter with
appropriate subscripts is often used when
coordinates must be displayed. For example, if 𝑥 ∈ 𝐅𝑛, then letting 𝑥 equal
(𝑥1 , …, 𝑥𝑛 ) is good notation, as shown in the proof above. Even better, work with
just 𝑥 and avoid explicit coordinates when possible.
1.15 notation: 0
Let 0 denote the list of length 𝑛 whose coordinates are all 0:
0 = (0, …, 0).
Here we are using the symbol 0 in two different ways—on the left side of the
equation above, the symbol 0 denotes a list of length 𝑛, which is an element of 𝐅𝑛,
whereas on the right side, each 0 denotes a number. This potentially confusing
practice actually causes no problems because the context should always make
clear which 0 is intended.
1.16 example: context determines which 0 is intended

Consider the statement that 0 is an additive identity for 𝐅𝑛 :
𝑥+0=𝑥 for all 𝑥 ∈ 𝐅𝑛.
Here the 0 above is the list defined in 1.15, not the number 0, because we have
not defined the sum of an element of 𝐅𝑛 (namely, 𝑥) and the number 0.
A picture can aid our intuition. We will

draw pictures in 𝐑2 because we can sketch
this space on two-dimensional surfaces
such as paper and computer screens. A
typical element of 𝐑2 is a point 𝑣 = (𝑎, 𝑏).
Sometimes we think of 𝑣 not as a point
but as an arrow starting at the origin and Elements of 𝐑2 can be thought of
ending at (𝑎, 𝑏), as shown here. When we as points or as vectors.
think of an element of 𝐑 as an arrow, we
2
refer to it as a vector.
When we think of vectors in 𝐑2 as arrows, we
can move an arrow parallel to itself (not changing
its length or direction) and still think of it as the
same vector. With that viewpoint, you will often
gain better understanding by dispensing with the
coordinate axes and the explicit coordinates and A vector.
just thinking of the vector, as shown in the figure here. The two arrows shown
here have the same length and same direction, so we think of them as the same
vector.
Whenever we use pictures in 𝐑2 or Mathematical models of the economy
use the somewhat vague language of can have thousands of variables, say
points and vectors, remember that these 𝑥1 , …, 𝑥5000 , which means that we must
are just aids to our understanding, not sub- work in 𝐑5000 . Such a space cannot be
stitutes for the actual mathematics that dealt with geometrically. However, the
we will develop. Although we cannot algebraic approach works well. Thus
draw good pictures in high-dimensional our subject is called linear algebra.
spaces, the elements of these spaces are
as rigorously defined as elements of 𝐑2.
For example, (2, −3, 17, 𝜋, √2) is an element of 𝐑5, and we may casually
refer to it as a point in 𝐑5 or a vector in 𝐑5 without worrying about whether the
geometry of 𝐑5 has any physical meaning.
Recall that we defined the sum of two elements of 𝐅𝑛 to be the element of 𝐅𝑛
obtained by adding corresponding coordinates; see 1.13. As we will now see,
addition has a simple geometric interpretation in the special case of 𝐑2.
Suppose we have two vectors 𝑢 and 𝑣 in 𝐑2
that we want to add. Move the vector 𝑣 parallel
to itself so that its initial point coincides with the
end point of the vector 𝑢, as shown here. The
sum 𝑢 + 𝑣 then equals the vector whose initial
point equals the initial point of 𝑢 and whose end
point equals the end point of the vector 𝑣, as The sum of two vectors.
shown here.
In the next definition, the 0 on the right side of the displayed equation is the
list 0 ∈ 𝐅𝑛.
1.17 definition: additive inverse in 𝐅𝑛 , −𝑥
For 𝑥 ∈ 𝐅𝑛, the additive inverse of 𝑥, denoted by −𝑥, is the vector −𝑥 ∈ 𝐅𝑛

such that
𝑥 + (−𝑥) = 0.
Thus if 𝑥 = (𝑥1 , …, 𝑥𝑛 ), then −𝑥 = (−𝑥1 , …, −𝑥𝑛 ).
The additive inverse of a vector in 𝐑2 is the

vector with the same length but pointing in the
opposite direction. The figure here illustrates
this way of thinking about the additive inverse
in 𝐑2. As you can see, the vector labeled −𝑥 has
the same length as the vector labeled 𝑥 but points A vector and its additive inverse.
in the opposite direction.
Having dealt with addition in 𝐅𝑛, we now turn to multiplication. We could
define a multiplication in 𝐅𝑛 in a similar fashion, starting with two elements of
𝐅𝑛 and getting another element of 𝐅𝑛 by multiplying corresponding coordinates.
Experience shows that this definition is not useful for our purposes. Another
type of multiplication, called scalar multiplication, will be central to our subject.
Specifically, we need to define what it means to multiply an element of 𝐅𝑛 by an
element of 𝐅.
1.18 definition: scalar multiplication in 𝐅𝑛
The product of a number 𝜆 and a vector in 𝐅𝑛 is computed by multiplying

each coordinate of the vector by 𝜆:
𝜆(𝑥1 , …, 𝑥𝑛 ) = (𝜆𝑥1 , …, 𝜆𝑥𝑛 );
here 𝜆 ∈ 𝐅 and (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛.
Scalar multiplication has a nice geo- Scalar multiplication in 𝐅𝑛 multiplies

metric interpretation in 𝐑2. If 𝜆 > 0 and together a scalar and a vector, getting
𝑥 ∈ 𝐑2, then 𝜆𝑥 is the vector that points a vector. In contrast, the dot product in
in the same direction as 𝑥 and whose 𝐑2 or 𝐑3 multiplies together two vec-
length is 𝜆 times the length of 𝑥. In other tors and gets a scalar. Generalizations
words, to get 𝜆𝑥, we shrink or stretch 𝑥 of the dot product will become impor-
by a factor of 𝜆, depending on whether tant in Chapter 6.
𝜆 < 1 or 𝜆 > 1.
If 𝜆 < 0 and 𝑥 ∈ 𝐑2, then 𝜆𝑥 is the
vector that points in the direction opposite
to that of 𝑥 and whose length is |𝜆| times
the length of 𝑥, as shown here.
Scalar multiplication.
Digression on Fields
A field is a set containing at least two distinct elements called 0 and 1, along with
operations of addition and multiplication satisfying all properties listed in 1.3.
Thus 𝐑 and 𝐂 are fields, as is the set of rational numbers along with the usual
operations of addition and multiplication. Another example of a field is the set
{0, 1} with the usual operations of addition and multiplication except that 1 + 1 is
defined to equal 0.
In this book we will not deal with fields other than 𝐑 and 𝐂 . However, many
of the definitions, theorems, and proofs in linear algebra that work for the fields
𝐑 and 𝐂 also work without change for arbitrary fields. If you prefer to do so,
throughout much of this book (except for Chapters 6 and 7, which deal with inner
product spaces) you can think of 𝐅 as denoting an arbitrary field instead of 𝐑
or 𝐂 . For results (except in the inner product chapters) that have as a hypothesis
that 𝐅 is 𝐂 , you can probably replace that hypothesis with the hypothesis that 𝐅
is an algebraically closed field, which means that every nonconstant polynomial
with coefficients in 𝐅 has a zero. A few results, such as Exercise 13 in Section
1C, require the hypothesis on 𝐅 that 1 + 1 ≠ 0.
Exercises 1A
1 Show that 𝛼 + 𝛽 = 𝛽 + 𝛼 for all 𝛼, 𝛽 ∈ 𝐂 .

2 Show that (𝛼 + 𝛽) + 𝜆 = 𝛼 + (𝛽 + 𝜆) for all 𝛼, 𝛽, 𝜆 ∈ 𝐂 .
3 Show that (𝛼𝛽)𝜆 = 𝛼(𝛽𝜆) for all 𝛼, 𝛽, 𝜆 ∈ 𝐂 .
4 Show that 𝜆(𝛼 + 𝛽) = 𝜆𝛼 + 𝜆𝛽 for all 𝜆, 𝛼, 𝛽 ∈ 𝐂 .
5 Show that for every 𝛼 ∈ 𝐂 , there exists a unique 𝛽 ∈ 𝐂 such that 𝛼 + 𝛽 = 0.
6 Show that for every 𝛼 ∈ 𝐂 with 𝛼 ≠ 0, there exists a unique 𝛽 ∈ 𝐂 such
that 𝛼𝛽 = 1.
7 Show that
−1 + √3𝑖
2
is a cube root of 1 (meaning that its cube equals 1).
8 Find two distinct square roots of 𝑖.
9 Find 𝑥 ∈ 𝐑4 such that
(4, −3, 1, 7) + 2𝑥 = (5, 9, −6, 8).
10 Explain why there does not exist 𝜆 ∈ 𝐂 such that

𝜆(2 − 3𝑖, 5 + 4𝑖, −6 + 7𝑖) = (12 − 5𝑖, 7 + 22𝑖, −32 − 9𝑖).
11 Show that (𝑥 + 𝑦) + 𝑧 = 𝑥 + (𝑦 + 𝑧) for all 𝑥, 𝑦, 𝑧 ∈ 𝐅𝑛.

12 Show that (𝑎𝑏)𝑥 = 𝑎(𝑏𝑥) for all 𝑥 ∈ 𝐅𝑛 and all 𝑎, 𝑏 ∈ 𝐅.
13 Show that 1𝑥 = 𝑥 for all 𝑥 ∈ 𝐅𝑛.
14 Show that 𝜆(𝑥 + 𝑦) = 𝜆𝑥 + 𝜆𝑦 for all 𝜆 ∈ 𝐅 and all 𝑥, 𝑦 ∈ 𝐅𝑛.
15 Show that (𝑎 + 𝑏)𝑥 = 𝑎𝑥 + 𝑏𝑥 for all 𝑎, 𝑏 ∈ 𝐅 and all 𝑥 ∈ 𝐅𝑛.
“Can you do addition?” the White Queen asked. “What’s one and one and one
and one and one and one and one and one and one and one?”
“I don’t know,” said Alice. “I lost count.”
—Through the Looking Glass, Lewis Carroll

1B Definition of Vector Space

The motivation for the definition of a vector space comes from properties of
addition and scalar multiplication in 𝐅𝑛 : Addition is commutative, associative,
and has an identity. Every element has an additive inverse. Scalar multiplication
is associative. Scalar multiplication by 1 acts as expected. Addition and scalar
multiplication are connected by distributive properties.
We will define a vector space to be a set 𝑉 with an addition and a scalar
multiplication on 𝑉 that satisfy the properties in the paragraph above.
1.19 definition: addition, scalar multiplication
• An addition on a set 𝑉 is a function that assigns an element 𝑢 + 𝑣 ∈ 𝑉

to each pair of elements 𝑢, 𝑣 ∈ 𝑉.
• A scalar multiplication on a set 𝑉 is a function that assigns an element
𝜆𝑣 ∈ 𝑉 to each 𝜆 ∈ 𝐅 and each 𝑣 ∈ 𝑉.
Now we are ready to give the formal definition of a vector space.

1.20 definition: vector space
A vector space is a set 𝑉 along with an addition on 𝑉 and a scalar multiplication

on 𝑉 such that the following properties hold.
commutativity
𝑢 + 𝑣 = 𝑣 + 𝑢 for all 𝑢, 𝑣 ∈ 𝑉.
associativity
(𝑢 + 𝑣) + 𝑤 = 𝑢 + (𝑣 + 𝑤) and (𝑎𝑏)𝑣 = 𝑎(𝑏𝑣) for all 𝑢, 𝑣, 𝑤 ∈ 𝑉 and for all
𝑎, 𝑏 ∈ 𝐅.
additive identity
There exists an element 0 ∈ 𝑉 such that 𝑣 + 0 = 𝑣 for all 𝑣 ∈ 𝑉.
additive inverse
For every 𝑣 ∈ 𝑉, there exists 𝑤 ∈ 𝑉 such that 𝑣 + 𝑤 = 0.
multiplicative identity
1𝑣 = 𝑣 for all 𝑣 ∈ 𝑉.
distributive properties
𝑎(𝑢 + 𝑣) = 𝑎𝑢 + 𝑎𝑣 and (𝑎 + 𝑏)𝑣 = 𝑎𝑣 + 𝑏𝑣 for all 𝑎, 𝑏 ∈ 𝐅 and all 𝑢, 𝑣 ∈ 𝑉.
The following geometric language sometimes aids our intuition.
1.21 definition: vector, point
Elements of a vector space are called vectors or points.

Section 1B Definition of Vector Space 13
The scalar multiplication in a vector space depends on 𝐅. Thus when we need

to be precise, we will say that 𝑉 is a vector space over 𝐅 instead of saying simply
that 𝑉 is a vector space. For example, 𝐑𝑛 is a vector space over 𝐑 , and 𝐂𝑛 is a
vector space over 𝐂 .
1.22 definition: real vector space, complex vector space
• A vector space over 𝐑 is called a real vector space.

• A vector space over 𝐂 is called a complex vector space.
Usually the choice of 𝐅 is either clear from the context or irrelevant. Thus we
often assume that 𝐅 is lurking in the background without specifically mentioning it.
With the usual operations of addition The simplest vector space is {0}, which
and scalar multiplication, 𝐅𝑛 is a vector contains only one point.
space over 𝐅, as you should verify. The
example of 𝐅𝑛 motivated our definition of vector space.
1.23 example: 𝐅∞
𝐅∞ is defined to be the set of all sequences of elements of 𝐅 :
𝐅∞ = {(𝑥1 , 𝑥2 , … ) ∶ 𝑥𝑘 ∈ 𝐅 for 𝑘 = 1, 2, …}.
Addition and scalar multiplication on 𝐅∞ are defined as expected:
(𝑥1 , 𝑥2 , … ) + (𝑦1 , 𝑦2 , … ) = (𝑥1 + 𝑦1 , 𝑥2 + 𝑦2 , … ),

𝜆(𝑥1 , 𝑥2 , … ) = (𝜆𝑥1 , 𝜆𝑥2 , … ).
With these definitions, 𝐅∞ becomes a vector space over 𝐅, as you should verify.
The additive identity in this vector space is the sequence of all 0’s.
Our next example of a vector space involves a set of functions.
1.24 notation: 𝐅𝑆
• If 𝑆 is a set, then 𝐅𝑆 denotes the set of functions from 𝑆 to 𝐅.

• For 𝑓, 𝑔 ∈ 𝐅𝑆, the sum 𝑓 + 𝑔 ∈ 𝐅𝑆 is the function defined by
( 𝑓 + 𝑔)(𝑥) = 𝑓 (𝑥) + 𝑔(𝑥)
for all 𝑥 ∈ 𝑆.
• For 𝜆 ∈ 𝐅 and 𝑓 ∈ 𝐅𝑆, the product 𝜆 𝑓 ∈ 𝐅𝑆 is the function defined by
(𝜆 𝑓 )(𝑥) = 𝜆 𝑓 (𝑥)
As an example of the notation above, if 𝑆 is the interval [0, 1] and 𝐅 = 𝐑 , then

𝐑 is the set of real-valued functions on the interval [0, 1].
[0, 1]
You should verify all three bullet points in the next example.
1.25 example: 𝐅𝑆 is a vector space
• If 𝑆 is a nonempty set, then 𝐅𝑆 (with the operations of addition and scalar

multiplication as defined above) is a vector space over 𝐅.
• The additive identity of 𝐅𝑆 is the function 0 ∶ 𝑆 → 𝐅 defined by
0(𝑥) = 0
• For 𝑓 ∈ 𝐅𝑆, the additive inverse of 𝑓 is the function − 𝑓 ∶ 𝑆 → 𝐅 defined by
(− 𝑓 )(𝑥) = − 𝑓 (𝑥)
The vector space 𝐅𝑛 is a special case The elements of the vector space 𝐑[0, 1]
of the vector space 𝐅𝑆 because each are real-valued functions on [0, 1], not
(𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛 can be thought of as lists. In general, a vector space is an
a function 𝑥 from the set {1, 2, …, 𝑛} to 𝐅 abstract entity whose elements might
by writing 𝑥(𝑘) instead of 𝑥𝑘 for the 𝑘 th be lists, functions, or weird objects.
coordinate of (𝑥1 , …, 𝑥𝑛 ). In other words,
we can think of 𝐅𝑛 as 𝐅{1, 2, …, 𝑛}. Similarly, we can think of 𝐅∞ as 𝐅{1, 2, … }.
Soon we will see further examples of vector spaces, but first we need to develop
some of the elementary properties of vector spaces.
The definition of a vector space requires it to have an additive identity. The
next result states that this identity is unique.
1.26 unique additive identity
A vector space has a unique additive identity.
Proof Suppose 0 and 0′ are both additive identities for some vector space 𝑉.
Then
0′ = 0′ + 0 = 0 + 0′ = 0,
where the first equality holds because 0 is an additive identity, the second equality
comes from commutativity, and the third equality holds because 0′ is an additive
identity. Thus 0′ = 0, proving that 𝑉 has only one additive identity.
Each element 𝑣 in a vector space has an additive inverse, an element 𝑤 in the

vector space such that 𝑣 + 𝑤 = 0. The next result shows that each element in a
vector space has only one additive inverse.
1.27 unique additive inverse
Every element in a vector space has a unique additive inverse.
Proof Suppose 𝑉 is a vector space. Let 𝑣 ∈ 𝑉. Suppose 𝑤 and 𝑤′ are additive

inverses of 𝑣. Then
𝑤 = 𝑤 + 0 = 𝑤 + (𝑣 + 𝑤′ ) = (𝑤 + 𝑣) + 𝑤′ = 0 + 𝑤′ = 𝑤′.
Thus 𝑤 = 𝑤′, as desired.
Because additive inverses are unique, the following notation now makes sense.
1.28 notation: −𝑣, 𝑤 − 𝑣
Let 𝑣, 𝑤 ∈ 𝑉. Then
• −𝑣 denotes the additive inverse of 𝑣;
• 𝑤 − 𝑣 is defined to be 𝑤 + (−𝑣).
Almost all results in this book involve some vector space. To avoid having to
restate frequently that 𝑉 is a vector space, we now make the necessary declaration
once and for all.
1.29 notation: 𝑉
For the rest of this book, 𝑉 denotes a vector space over 𝐅.
In the next result, 0 denotes a scalar (the number 0 ∈ 𝐅 ) on the left side of the
equation and a vector (the additive identity of 𝑉) on the right side of the equation.
1.30 the number 0 times a vector

0𝑣 = 0 for every 𝑣 ∈ 𝑉.
Proof For 𝑣 ∈ 𝑉, we have The result in 1.30 involves the additive

identity of 𝑉 and scalar multiplication.
0𝑣 = (0 + 0)𝑣 = 0𝑣 + 0𝑣.
The only part of the definition of a vec-
Adding the additive inverse of 0𝑣 to both tor space that connects vector addition
sides of the equation above gives 0 = 0𝑣, and scalar multiplication is the dis-
as desired. tributive property. Thus the distribu-
tive property must be used in the proof
In the next result, 0 denotes the addi- of 1.30.
tive identity of 𝑉. Although their proofs
are similar, 1.30 and 1.31 are not identical. More precisely, 1.30 states that the
product of the scalar 0 and any vector equals the vector 0, whereas 1.31 states that
the product of any scalar and the vector 0 equals the vector 0.
1.31 a number times the vector 0

𝑎0 = 0 for every 𝑎 ∈ 𝐅.
Proof For 𝑎 ∈ 𝐅, we have

𝑎0 = 𝑎(0 + 0) = 𝑎0 + 𝑎0.
Adding the additive inverse of 𝑎0 to both sides of the equation above gives 0 = 𝑎0,
as desired.
Now we show that if an element of 𝑉 is multiplied by the scalar −1, then the
result is the additive inverse of the element of 𝑉.
1.32 the number −1 times a vector

(−1)𝑣 = −𝑣 for every 𝑣 ∈ 𝑉.
Proof For 𝑣 ∈ 𝑉, we have

𝑣 + (−1)𝑣 = 1𝑣 + (−1)𝑣 = (1 + (−1))𝑣 = 0𝑣 = 0.
This equation says that (−1)𝑣, when added to 𝑣, gives 0. Thus (−1)𝑣 is the
additive inverse of 𝑣, as desired.
Exercises 1B
1 Prove that −(−𝑣) = 𝑣 for every 𝑣 ∈ 𝑉.

2 Suppose 𝑎 ∈ 𝐅, 𝑣 ∈ 𝑉, and 𝑎𝑣 = 0. Prove that 𝑎 = 0 or 𝑣 = 0.
3 Suppose 𝑣, 𝑤 ∈ 𝑉. Explain why there exists a unique 𝑥 ∈ 𝑉 such that
𝑣 + 3𝑥 = 𝑤.
4 The empty set is not a vector space. The empty set fails to satisfy only one
of the requirements listed in the definition of a vector space (1.20). Which
one?
5 Show that in the definition of a vector space (1.20), the additive inverse
condition can be replaced with the condition that
0𝑣 = 0 for all 𝑣 ∈ 𝑉.
Here the 0 on the left side is the number 0, and the 0 on the right side is the
additive identity of 𝑉.
The phrase a “condition can be replaced” in a definition means that the
collection of objects satisfying the definition is unchanged if the original
condition is replaced with the new condition.
6 Let ∞ and −∞ denote two distinct objects, neither of which is in 𝐑 . Define

an addition and scalar multiplication on 𝐑 ∪ {∞, −∞} as you could guess
from the notation. Specifically, the sum and product of two real numbers is
as usual, and for 𝑡 ∈ 𝐑 define
⎧−∞ if 𝑡 < 0, ⎧∞ if 𝑡 < 0,
{
{ {
{
𝑡∞ = ⎨0 if 𝑡 = 0, 𝑡(−∞) = ⎨0 if 𝑡 = 0,
{ {
{∞
⎩ if 𝑡 > 0 , {
⎩ −∞ if 𝑡 > 0,
and
𝑡 + ∞ = ∞ + 𝑡 = ∞ + ∞ = ∞,
𝑡 + (−∞) = (−∞) + 𝑡 = (−∞) + (−∞) = −∞,
∞ + (−∞) = (−∞) + ∞ = 0.
With these operations of addition and scalar multiplication, is 𝐑 ∪ {∞, −∞}

a vector space over 𝐑 ? Explain.
7 Suppose 𝑆 is a nonempty set. Let 𝑉 𝑆 denote the set of functions from 𝑆 to 𝑉.
Define a natural addition and scalar multiplication on 𝑉 𝑆, and show that 𝑉 𝑆
is a vector space with these definitions.
8 Suppose 𝑉 is a real vector space.
• The complexification of 𝑉, denoted by 𝑉𝐂 , equals 𝑉× 𝑉. An element of
𝑉𝐂 is an ordered pair (𝑢, 𝑣), where 𝑢, 𝑣 ∈ 𝑉, but we write this as 𝑢 + 𝑖𝑣.
• Addition on 𝑉𝐂 is defined by
(𝑢1 + 𝑖𝑣1 ) + (𝑢2 + 𝑖𝑣2 ) = (𝑢1 + 𝑢2 ) + 𝑖(𝑣1 + 𝑣2 )
for all 𝑢1 , 𝑣1 , 𝑢2 , 𝑣2 ∈ 𝑉.
• Complex scalar multiplication on 𝑉𝐂 is defined by
(𝑎 + 𝑏𝑖)(𝑢 + 𝑖𝑣) = (𝑎𝑢 − 𝑏𝑣) + 𝑖(𝑎𝑣 + 𝑏𝑢)
for all 𝑎, 𝑏 ∈ 𝐑 and all 𝑢, 𝑣 ∈ 𝑉.

Prove that with the definitions of addition and scalar multiplication as above,
𝑉𝐂 is a complex vector space.
Think of 𝑉 as a subset of 𝑉𝐂 by identifying 𝑢 ∈ 𝑉 with 𝑢 + 𝑖0. The construc-
tion of 𝑉𝐂 from 𝑉 can then be thought of as generalizing the construction
of 𝐂𝑛 from 𝐑𝑛.
1C Subspaces
By considering subspaces, we can greatly expand our examples of vector spaces.
1.33 definition: subspace
A subset 𝑈 of 𝑉 is called a subspace of 𝑉 if 𝑈 is also a vector space with the

same additive identity, addition, and scalar multiplication as on 𝑉.
The next result gives the easiest way Some people use the terminology
to check whether a subset of a vector linear subspace, which means the
space is a subspace. same as subspace.
1.34 conditions for a subspace
A subset 𝑈 of 𝑉 is a subspace of 𝑉 if and only if 𝑈 satisfies the following

three conditions.
additive identity
0 ∈ 𝑈.
closed under addition

𝑢, 𝑤 ∈ 𝑈 implies 𝑢 + 𝑤 ∈ 𝑈.
closed under scalar multiplication

𝑎 ∈ 𝐅 and 𝑢 ∈ 𝑈 implies 𝑎𝑢 ∈ 𝑈.
Proof If 𝑈 is a subspace of 𝑉, then 𝑈 The additive identity condition above

satisfies the three conditions above by the could be replaced with the condition
definition of vector space. that 𝑈 is nonempty (because then tak-
Conversely, suppose 𝑈 satisfies the ing 𝑢 ∈ 𝑈 and multiplying it by 0
three conditions above. The first condi- would imply that 0 ∈ 𝑈 ). However,
tion ensures that the additive identity of if a subset 𝑈 of 𝑉 is indeed a sub-
𝑉 is in 𝑈. The second condition ensures space, then usually the quickest way
that addition makes sense on 𝑈. The third to show that 𝑈 is nonempty is to show
condition ensures that scalar multiplica- that 0 ∈ 𝑈.
tion makes sense on 𝑈.
If 𝑢 ∈ 𝑈, then −𝑢 [which equals (−1)𝑢 by 1.32] is also in 𝑈 by the third
condition above. Hence every element of 𝑈 has an additive inverse in 𝑈.
The other parts of the definition of a vector space, such as associativity and
commutativity, are automatically satisfied for 𝑈 because they hold on the larger
space 𝑉. Thus 𝑈 is a vector space and hence is a subspace of 𝑉.
The three conditions in the result above usually enable us to determine quickly
whether a given subset of 𝑉 is a subspace of 𝑉. You should verify all assertions
in the next example.
Section 1C Subspaces 19
1.35 example: subspaces
(a) If 𝑏 ∈ 𝐅, then
{(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) ∈ 𝐅4 ∶ 𝑥3 = 5𝑥4 + 𝑏}
is a subspace of 𝐅4 if and only if 𝑏 = 0.
(b) The set of continuous real-valued functions on the interval [0, 1] is a subspace
of 𝐑[0, 1].
(c) The set of differentiable real-valued functions on 𝐑 is a subspace of 𝐑𝐑.
(d) The set of differentiable real-valued functions 𝑓 on the interval (0, 3) such
that 𝑓 ′(2) = 𝑏 is a subspace of 𝐑(0, 3) if and only if 𝑏 = 0.
(e) The set of all sequences of complex numbers with limit 0 is a subspace of 𝐂∞.
Verifying some of the items above The set {0} is the smallest subspace of
shows the linear structure underlying 𝑉, and 𝑉 itself is the largest subspace
parts of calculus. For example, (b) above of 𝑉. The empty set is not a subspace
requires the result that the sum of two of 𝑉 because a subspace must be a
continuous functions is continuous. As vector space and hence must contain at
another example, (d) above requires the least one element, namely, an additive
result that for a constant 𝑐, the derivative identity.
of 𝑐 𝑓 equals 𝑐 times the derivative of 𝑓.
The subspaces of 𝐑2 are precisely {0}, all lines in 𝐑2 containing the origin,
and 𝐑2. The subspaces of 𝐑3 are precisely {0}, all lines in 𝐑3 containing the origin,
all planes in 𝐑3 containing the origin, and 𝐑3. To prove that all these objects are
indeed subspaces is straightforward—the hard part is to show that they are the
only subspaces of 𝐑2 and 𝐑3. That task will be easier after we introduce some
additional tools in the next chapter.
Sums of Subspaces
When dealing with vector spaces, we are The union of subspaces is rarely a sub-
usually interested only in subspaces, as space (see Exercise 12), which is why
opposed to arbitrary subsets. The notion we usually work with sums rather than
of the sum of subspaces will be useful. unions.
1.36 definition: sum of subspaces
Suppose 𝑉1 , …, 𝑉𝑚 are subspaces of 𝑉. The sum of 𝑉1 , …, 𝑉𝑚 , denoted by

𝑉1 + ⋯ + 𝑉𝑚 , is the set of all possible sums of elements of 𝑉1 , …, 𝑉𝑚 . More
precisely,
𝑉1 + ⋯ + 𝑉𝑚 = {𝑣1 + ⋯ + 𝑣𝑚 ∶ 𝑣1 ∈ 𝑉1 , …, 𝑣𝑚 ∈ 𝑉𝑚 }.
Let’s look at some examples of sums of subspaces.
1.37 example: a sum of subspaces of 𝐅3

Suppose 𝑈 is the set of all elements of 𝐅3 whose second and third coordinates
equal 0, and 𝑊 is the set of all elements of 𝐅3 whose first and third coordinates
equal 0:
𝑈 = {(𝑥, 0, 0) ∈ 𝐅3 ∶ 𝑥 ∈ 𝐅} and 𝑊 = {(0, 𝑦, 0) ∈ 𝐅3 ∶ 𝑦 ∈ 𝐅}.
Then
𝑈 + 𝑊 = {(𝑥, 𝑦, 0) ∈ 𝐅3 ∶ 𝑥, 𝑦 ∈ 𝐅},
as you should verify.
1.38 example: a sum of subspaces of 𝐅4

Suppose
𝑈 = {(𝑥, 𝑥, 𝑦, 𝑦) ∈ 𝐅4 ∶ 𝑥, 𝑦 ∈ 𝐅} and 𝑊 = {(𝑥, 𝑥, 𝑥, 𝑦) ∈ 𝐅4 ∶ 𝑥, 𝑦 ∈ 𝐅}.
Using words rather than symbols, we could say that 𝑈 is the set of elements
of 𝐅4 whose first two coordinates equal each other and whose third and fourth
coordinates equal each other. Similarly, 𝑊 is the set of elements of 𝐅4 whose first
three coordinates equal each other.
To find a description of 𝑈 + 𝑊, consider a typical element (𝑎, 𝑎, 𝑏, 𝑏) of 𝑈 and
a typical element (𝑐, 𝑐, 𝑐, 𝑑) of 𝑊, where 𝑎, 𝑏, 𝑐, 𝑑 ∈ 𝐅. We have
(𝑎, 𝑎, 𝑏, 𝑏) + (𝑐, 𝑐, 𝑐, 𝑑) = (𝑎 + 𝑐, 𝑎 + 𝑐, 𝑏 + 𝑐, 𝑏 + 𝑑),
which shows that every element of 𝑈 + 𝑊 has its first two coordinates equal to
each other. Thus
1.39 𝑈 + 𝑊 ⊆ {(𝑥, 𝑥, 𝑦, 𝑧) ∈ 𝐅4 ∶ 𝑥, 𝑦, 𝑧 ∈ 𝐅}.
To prove the inclusion in the other direction, suppose 𝑥, 𝑦, 𝑧 ∈ 𝐅. Then
(𝑥, 𝑥, 𝑦, 𝑧) = (𝑥, 𝑥, 𝑦, 𝑦) + (0, 0, 0, 𝑧 − 𝑦),
where the first vector on the right is in 𝑈 and the second vector on the right is
in 𝑊. Thus (𝑥, 𝑥, 𝑦, 𝑧) ∈ 𝑈 + 𝑊, showing that the inclusion 1.39 also holds in the
opposite direction. Hence
𝑈 + 𝑊 = {(𝑥, 𝑥, 𝑦, 𝑧) ∈ 𝐅4 ∶ 𝑥, 𝑦, 𝑧 ∈ 𝐅},
which shows that 𝑈 + 𝑊 is the set of elements of 𝐅4 whose first two coordinates
equal each other.
The next result states that the sum of subspaces is a subspace, and is in fact the
smallest subspace containing all the summands (which means that every subspace
containing all the summands also contains the sum).
1.40 sum of subspaces is the smallest containing subspace
Suppose 𝑉1 , …, 𝑉𝑚 are subspaces of 𝑉. Then 𝑉1 + ⋯ + 𝑉𝑚 is the smallest

subspace of 𝑉 containing 𝑉1 , …, 𝑉𝑚 .
Proof The reader can verify that 𝑉1 + ⋯ + 𝑉𝑚 contains the additive identity 0
and is closed under addition and scalar multiplication. Thus 1.34 implies that
𝑉1 + ⋯ + 𝑉𝑚 is a subspace of 𝑉.
The subspaces 𝑉1 , …, 𝑉𝑚 are all con- Sums of subspaces in the theory of vec-
tained in 𝑉1 + ⋯ + 𝑉𝑚 (to see this, consider tor spaces are analogous to unions of
sums 𝑣1 + ⋯ + 𝑣𝑚 where all except one subsets in set theory. Given two sub-
of the 𝑣𝑘 ’s are 0). Conversely, every subspaces of a vector space, the smallest
space of 𝑉 containing 𝑉1 , …, 𝑉𝑚 contains subspace containing them is their sum.
𝑉1 + ⋯ + 𝑉𝑚 (because subspaces must Analogously, given two subsets of a set,
contain all finite sums of their elements). the smallest subset containing them is
Thus 𝑉1 + ⋯ + 𝑉𝑚 is the smallest subspace their union.
of 𝑉 containing 𝑉1 , …, 𝑉𝑚 .
Direct Sums
Suppose 𝑉1 , …, 𝑉𝑚 are subspaces of 𝑉. Every element of 𝑉1 + ⋯ + 𝑉𝑚 can be
written in the form
𝑣1 + ⋯ + 𝑣𝑚 ,
where each 𝑣𝑘 ∈ 𝑉𝑘 . Of special interest are cases in which each vector in
𝑉1 + ⋯ + 𝑉𝑚 can be represented in the form above in only one way. This situation
is so important that it gets a special name (direct sum) and a special symbol (⊕).
1.41 definition: direct sum, ⊕
Suppose 𝑉1 , …, 𝑉𝑚 are subspaces of 𝑉.

• The sum 𝑉1 + ⋯ + 𝑉𝑚 is called a direct sum if each element of 𝑉1 + ⋯ + 𝑉𝑚
can be written in only one way as a sum 𝑣1 + ⋯ + 𝑣𝑚 , where each 𝑣𝑘 ∈ 𝑉𝑘 .
• If 𝑉1 + ⋯ + 𝑉𝑚 is a direct sum, then 𝑉1 ⊕ ⋯ ⊕ 𝑉𝑚 denotes 𝑉1 + ⋯ + 𝑉𝑚 ,
with the ⊕ notation serving as an indication that this is a direct sum.
1.42 example: a direct sum of two subspaces

Suppose 𝑈 is the subspace of 𝐅3 of those vectors whose last coordinate equals 0,
and 𝑊 is the subspace of 𝐅3 of those vectors whose first two coordinates equal 0:
𝑈 = {(𝑥, 𝑦, 0) ∈ 𝐅3 ∶ 𝑥, 𝑦 ∈ 𝐅} and 𝑊 = {(0, 0, 𝑧) ∈ 𝐅3 ∶ 𝑧 ∈ 𝐅}.
Then 𝐅3 = 𝑈 ⊕ 𝑊, as you should verify.
≈ 7𝜋 Chapter 1 Vector Spaces
1.43 example: a direct sum of multiple subspaces

Suppose 𝑉𝑘 is the subspace of 𝐅𝑛 of To produce ⊕ in T X, type \oplus.
E
those vectors whose coordinates are all
0, except possibly in the 𝑘 th slot; for example, 𝑉2 = {(0, 𝑥, 0, …, 0) ∈ 𝐅𝑛 ∶ 𝑥 ∈ 𝐅}.
Then
𝐅𝑛 = 𝑉1 ⊕ ⋯ ⊕ 𝑉𝑛 ,
Sometimes nonexamples add to our understanding as much as examples.
1.44 example: a sum that is not a direct sum

Suppose
𝑉1 = {(𝑥, 𝑦, 0) ∈ 𝐅3 ∶ 𝑥, 𝑦 ∈ 𝐅},
𝑉2 = {(0, 0, 𝑧) ∈ 𝐅3 ∶ 𝑧 ∈ 𝐅},
𝑉3 = {(0, 𝑦, 𝑦) ∈ 𝐅3 ∶ 𝑦 ∈ 𝐅}.
Then 𝐅3 = 𝑉1 + 𝑉2 + 𝑉3 because every vector (𝑥, 𝑦, 𝑧) ∈ 𝐅3 can be written as

(𝑥, 𝑦, 𝑧) = (𝑥, 𝑦, 0) + (0, 0, 𝑧) + (0, 0, 0),
where the first vector on the right side is in 𝑉1 , the second vector is in 𝑉2 , and the
third vector is in 𝑉3 .
However, 𝐅3 does not equal the direct sum of 𝑉1 , 𝑉2 , 𝑉3 , because the vector
(0, 0, 0) can be written in more than one way as a sum 𝑣1 + 𝑣2 + 𝑣3 , with each
𝑣𝑘 ∈ 𝑉𝑘 . Specifically, we have
(0, 0, 0) = (0, 1, 0) + (0, 0, 1) + (0, −1, −1)
and, of course,
(0, 0, 0) = (0, 0, 0) + (0, 0, 0) + (0, 0, 0),
where the first vector on the right side of each equation above is in 𝑉1 , the second
vector is in 𝑉2 , and the third vector is in 𝑉3 . Thus the sum 𝑉1 + 𝑉2 + 𝑉3 is not a
direct sum.
The definition of direct sum requires The symbol ⊕, which is a plus sign
every vector in the sum to have a unique inside a circle, reminds us that we are
representation as an appropriate sum. dealing with a special type of sum of
The next result shows that when deciding subspaces—each element in the direct
whether a sum of subspaces is a direct sum can be represented in only one way
sum, we only need to consider whether 0 as a sum of elements from the specified
can be uniquely written as an appropriate subspaces.
sum.
1.45 condition for a direct sum
Suppose 𝑉1 , …, 𝑉𝑚 are subspaces of 𝑉. Then 𝑉1 + ⋯ + 𝑉𝑚 is a direct sum if

and only if the only way to write 0 as a sum 𝑣1 + ⋯ + 𝑣𝑚 , where each 𝑣𝑘 ∈ 𝑉𝑘 ,
is by taking each 𝑣𝑘 equal to 0.
Proof First suppose 𝑉1 + ⋯ + 𝑉𝑚 is a direct sum. Then the definition of direct

sum implies that the only way to write 0 as a sum 𝑣1 + ⋯ + 𝑣𝑚 , where each 𝑣𝑘 ∈ 𝑉𝑘 ,
is by taking each 𝑣𝑘 equal to 0.
Now suppose that the only way to write 0 as a sum 𝑣1 + ⋯ + 𝑣𝑚 , where each
𝑣𝑘 ∈ 𝑉𝑘 , is by taking each 𝑣𝑘 equal to 0. To show that 𝑉1 + ⋯ + 𝑉𝑚 is a direct
sum, let 𝑣 ∈ 𝑉1 + ⋯ + 𝑉𝑚 . We can write
𝑣 = 𝑣1 + ⋯ + 𝑣𝑚
for some 𝑣1 ∈ 𝑉1 , …, 𝑣𝑚 ∈ 𝑉𝑚 . To show that this representation is unique,
suppose we also have
𝑣 = 𝑢 1 + ⋯ + 𝑢𝑚 ,
where 𝑢1 ∈ 𝑉1 , …, 𝑢𝑚 ∈ 𝑉𝑚 . Subtracting these two equations, we have
0 = (𝑣1 − 𝑢1 ) + ⋯ + (𝑣𝑚 − 𝑢𝑚 ).
Because 𝑣1 − 𝑢1 ∈ 𝑉1 , …, 𝑣𝑚 − 𝑢𝑚 ∈ 𝑉𝑚 , the equation above implies that each
𝑣𝑘 − 𝑢𝑘 equals 0. Thus 𝑣1 = 𝑢1 , …, 𝑣𝑚 = 𝑢𝑚 , as desired.
The next result gives a simple con- The symbol ⟺ used below means
dition for testing whether a sum of two “if and only if ”; this symbol could also
subspaces is a direct sum. be read to mean “is equivalent to”.
1.46 direct sum of two subspaces
Suppose 𝑈 and 𝑊 are subspaces of 𝑉. Then
𝑈 + 𝑊 is a direct sum ⟺ 𝑈 ∩ 𝑊 = {0}.
Proof First suppose that 𝑈 + 𝑊 is a direct sum. If 𝑣 ∈ 𝑈 ∩ 𝑊, then 0 = 𝑣 + (−𝑣),

where 𝑣 ∈ 𝑈 and −𝑣 ∈ 𝑊. By the unique representation of 0 as the sum of a
vector in 𝑈 and a vector in 𝑊, we have 𝑣 = 0. Thus 𝑈 ∩ 𝑊 = {0}, completing
the proof in one direction.
To prove the other direction, now suppose 𝑈 ∩ 𝑊 = {0}. To prove that 𝑈 + 𝑊
is a direct sum, suppose 𝑢 ∈ 𝑈, 𝑤 ∈ 𝑊, and
0 = 𝑢 + 𝑤.
To complete the proof, we only need to show that 𝑢 = 𝑤 = 0 (by 1.45). The
equation above implies that 𝑢 = −𝑤 ∈ 𝑊. Thus 𝑢 ∈ 𝑈 ∩ 𝑊. Hence 𝑢 = 0, which
by the equation above implies that 𝑤 = 0, completing the proof.
The result above deals only with Sums of subspaces are analogous to
the case of two subspaces. When ask- unions of subsets. Similarly, direct
ing about a possible direct sum with sums of subspaces are analogous to
more than two subspaces, it is not disjoint unions of subsets. No two sub-
enough to test that each pair of the spaces of a vector space can be disjoint,
subspaces intersect only at 0. To see because both contain 0. So disjoint-
this, consider Example 1.44. In that ness is replaced, at least in the case
nonexample of a direct sum, we have of two subspaces, with the requirement
𝑉1 ∩ 𝑉2 = 𝑉1 ∩ 𝑉3 = 𝑉2 ∩ 𝑉3 = {0}. that the intersection equal {0}.
Exercises 1C
1 For each of the following subsets of 𝐅3, determine whether it is a subspace

of 𝐅3 .
(a) {(𝑥1 , 𝑥2 , 𝑥3 ) ∈ 𝐅3 ∶ 𝑥1 + 2𝑥2 + 3𝑥3 = 0}
(b) {(𝑥1 , 𝑥2 , 𝑥3 ) ∈ 𝐅3 ∶ 𝑥1 + 2𝑥2 + 3𝑥3 = 4}
(c) {(𝑥1 , 𝑥2 , 𝑥3 ) ∈ 𝐅3 ∶ 𝑥1 𝑥2 𝑥3 = 0}
(d) {(𝑥1 , 𝑥2 , 𝑥3 ) ∈ 𝐅3 ∶ 𝑥1 = 5𝑥3 }
2 Verify all assertions about subspaces in Example 1.35.

3 Show that the set of differentiable real-valued functions 𝑓 on the interval
(−4, 4) such that 𝑓 ′(−1) = 3 𝑓 (2) is a subspace of 𝐑(−4, 4).
4 Suppose 𝑏 ∈ 𝐑 . Show that the set of continuous real-valued functions 𝑓 on

the interval [0, 1] such that ∫01 𝑓 = 𝑏 is a subspace of 𝐑[0, 1] if and only if
𝑏 = 0.
5 Is 𝐑2 a subspace of the complex vector space 𝐂2 ?

6 (a) Is {(𝑎, 𝑏, 𝑐) ∈ 𝐑3 ∶ 𝑎3 = 𝑏3 } a subspace of 𝐑3 ?
(b) Is {(𝑎, 𝑏, 𝑐) ∈ 𝐂3 ∶ 𝑎3 = 𝑏3 } a subspace of 𝐂3 ?
7 Prove or give a counterexample: If 𝑈 is a nonempty subset of 𝐑2 such that
𝑈 is closed under addition and under taking additive inverses (meaning
−𝑢 ∈ 𝑈 whenever 𝑢 ∈ 𝑈), then 𝑈 is a subspace of 𝐑2.
8 Give an example of a nonempty subset 𝑈 of 𝐑2 such that 𝑈 is closed under

scalar multiplication, but 𝑈 is not a subspace of 𝐑2.
9 A function 𝑓 ∶ 𝐑 → 𝐑 is called periodic if there exists a positive number 𝑝
such that 𝑓 (𝑥) = 𝑓 (𝑥 + 𝑝) for all 𝑥 ∈ 𝐑 . Is the set of periodic functions
from 𝐑 to 𝐑 a subspace of 𝐑𝐑 ? Explain.
10 Suppose 𝑉1 and 𝑉2 are subspaces of 𝑉. Prove that the intersection 𝑉1 ∩ 𝑉2
is a subspace of 𝑉.
11 Prove that the intersection of every collection of subspaces of 𝑉 is a subspace

of 𝑉.
12 Prove that the union of two subspaces of 𝑉 is a subspace of 𝑉 if and only if
one of the subspaces is contained in the other.
13 Prove that the union of three subspaces of 𝑉 is a subspace of 𝑉 if and only
if one of the subspaces contains the other two.
This exercise is surprisingly harder than Exercise 12, possibly because this
exercise is not true if we replace 𝐅 with a field containing only two elements.
14 Suppose
𝑈 = {(𝑥, −𝑥, 2𝑥) ∈ 𝐅3 ∶ 𝑥 ∈ 𝐅} and 𝑊 = {(𝑥, 𝑥, 2𝑥) ∈ 𝐅3 ∶ 𝑥 ∈ 𝐅}.
Describe 𝑈 + 𝑊 using symbols, and also give a description of 𝑈 + 𝑊 that
uses no symbols.
15 Suppose 𝑈 is a subspace of 𝑉. What is 𝑈 + 𝑈?
16 Is the operation of addition on the subspaces of 𝑉 commutative? In other
words, if 𝑈 and 𝑊 are subspaces of 𝑉, is 𝑈 + 𝑊 = 𝑊 + 𝑈?
17 Is the operation of addition on the subspaces of 𝑉 associative? In other
words, if 𝑉1 , 𝑉2 , 𝑉3 are subspaces of 𝑉, is
(𝑉1 + 𝑉2 ) + 𝑉3 = 𝑉1 + (𝑉2 + 𝑉3 )?
18 Does the operation of addition on the subspaces of 𝑉 have an additive

identity? Which subspaces have additive inverses?
19 Prove or give a counterexample: If 𝑉1 , 𝑉2 , 𝑈 are subspaces of 𝑉 such that
𝑉1 + 𝑈 = 𝑉2 + 𝑈,
then 𝑉1 = 𝑉2 .
20 Suppose
𝑈 = {(𝑥, 𝑥, 𝑦, 𝑦) ∈ 𝐅4 ∶ 𝑥, 𝑦 ∈ 𝐅}.
Find a subspace 𝑊 of 𝐅4 such that 𝐅4 = 𝑈 ⊕ 𝑊.
21 Suppose
𝑈 = {(𝑥, 𝑦, 𝑥 + 𝑦, 𝑥 − 𝑦, 2𝑥) ∈ 𝐅5 ∶ 𝑥, 𝑦 ∈ 𝐅}.
Find a subspace 𝑊 of 𝐅5 such that 𝐅5 = 𝑈 ⊕ 𝑊.
22 Suppose
𝑈 = {(𝑥, 𝑦, 𝑥 + 𝑦, 𝑥 − 𝑦, 2𝑥) ∈ 𝐅5 ∶ 𝑥, 𝑦 ∈ 𝐅}.
Find three subspaces 𝑊1 , 𝑊2 , 𝑊3 of 𝐅5, none of which equals {0}, such that
𝐅5 = 𝑈 ⊕ 𝑊1 ⊕ 𝑊2 ⊕ 𝑊3 .
23 Prove or give a counterexample: If 𝑉1 , 𝑉2 , 𝑈 are subspaces of 𝑉 such that
𝑉 = 𝑉1 ⊕ 𝑈 and 𝑉 = 𝑉2 ⊕ 𝑈,
then 𝑉1 = 𝑉2 .
Hint: When trying to discover whether a conjecture in linear algebra is true
or false, it is often useful to start by experimenting in 𝐅2.
24 A function 𝑓 ∶ 𝐑 → 𝐑 is called even if
𝑓 (−𝑥) = 𝑓 (𝑥)
for all 𝑥 ∈ 𝐑 . A function 𝑓 ∶ 𝐑 → 𝐑 is called odd if
𝑓 (−𝑥) = − 𝑓 (𝑥)
for all 𝑥 ∈ 𝐑 . Let 𝑉e denote the set of real-valued even functions on 𝐑

and let 𝑉o denote the set of real-valued odd functions on 𝐑 . Show that
𝐑𝐑 = 𝑉e ⊕ 𝑉o .
Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0
International License (https://creativecommons.org/licenses/by-nc/4.0), which permits any noncommercial use,
sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit
to original author and source, provide a link to the Creative Commons license, and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter’s
Creative Commons license, unless indicated otherwise in a credit line to the material. If
material is not included in the chapter’s Creative Commons license and your intended use
is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly
from the copyright holder.
Chapter 2
Finite-Dimensional Vector Spaces
In the last chapter we learned about vector spaces. Linear algebra focuses not
on arbitrary vector spaces, but on finite-dimensional vector spaces, which we
introduce in this chapter.
We begin this chapter by considering linear combinations of lists of vectors.
This leads us to the crucial concept of linear independence. The linear dependence
lemma will become one of our most useful tools.
A list of vectors in a vector space that is small enough to be linearly independent
and big enough so the linear combinations of the list fill up the vector space is
called a basis of the vector space. We will see that every basis of a vector space
has the same length, which will allow us to define the dimension of a vector space.
This chapter ends with a formula for the dimension of the sum of two subspaces.
standing assumptions for this chapter
• 𝐅 denotes 𝐑 or 𝐂 .
• 𝑉 denotes a vector space over 𝐅.
The main building of the Institute for Advanced Study, in Princeton, New Jersey.
Paul Halmos (1916–2006) wrote the first modern linear algebra book in this building.
Halmos’s linear algebra book was published in 1942 (second edition published in 1958).
The title of Halmos’s book was the same as the title of this chapter.
27
https://doi.org/10.1007/978-3-031-41026-0_2
28 Chapter 2 Finite-Dimensional Vector Spaces
2A Span and Linear Independence

We have been writing lists of numbers surrounded by parentheses, and we will
continue to do so for elements of 𝐅𝑛 ; for example, (2, −7, 8) ∈ 𝐅3. However, now
we need to consider lists of vectors (which may be elements of 𝐅𝑛 or of other
vector spaces). To avoid confusion, we will usually write lists of vectors without
surrounding parentheses. For example, (4, 1, 6), (9, 5, 7) is a list of length two of
vectors in 𝐑3.
2.1 notation: list of vectors
We will usually write lists of vectors without surrounding parentheses.
Linear Combinations and Span

A sum of scalar multiples of the vectors in a list is called a linear combination of
the list. Here is the formal definition.
2.2 definition: linear combination

A linear combination of a list 𝑣1 , …, 𝑣𝑚 of vectors in 𝑉 is a vector of the form
𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 ,
where 𝑎1 , …, 𝑎𝑚 ∈ 𝐅.
2.3 example: linear combinations in 𝐑3
• (17, −4, 2) is a linear combination of (2, 1, −3), (1, −2, 4), which is a list of
length two of vectors in 𝐑3, because
(17, −4, 2) = 6(2, 1, −3) + 5(1, −2, 4).
• (17, −4, 5) is not a linear combination of (2, 1, −3), (1, −2, 4), which is a list
of length two of vectors in 𝐑3, because there do not exist numbers 𝑎1 , 𝑎2 ∈ 𝐅
such that
(17, −4, 5) = 𝑎1 (2, 1, −3) + 𝑎2 (1, −2, 4).
In other words, the system of equations
17 = 2𝑎1 + 𝑎2
−4 = 𝑎1 − 2𝑎2
5 = −3𝑎1 + 4𝑎2
has no solutions (as you should verify).

Section 2A Span and Linear Independence 29
2.4 definition: span
The set of all linear combinations of a list of vectors 𝑣1 , …, 𝑣𝑚 in 𝑉 is called

the span of 𝑣1 , …, 𝑣𝑚 , denoted by span(𝑣1 , …, 𝑣𝑚 ). In other words,
span(𝑣1 , …, 𝑣𝑚 ) = {𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 ∶ 𝑎1 , …, 𝑎𝑚 ∈ 𝐅}.
The span of the empty list ( ) is defined to be {0}.
2.5 example: span

The previous example shows that in 𝐅3,
• (17, −4, 2) ∈ span((2, 1, −3), (1, −2, 4));
• (17, −4, 5) ∉ span((2, 1, −3), (1, −2, 4)).
Some mathematicians use the term linear span, which means the same as
span.
2.6 span is the smallest containing subspace
The span of a list of vectors in 𝑉 is the smallest subspace of 𝑉 containing all

vectors in the list.
Proof Suppose 𝑣1 , …, 𝑣𝑚 is a list of vectors in 𝑉.

First we show that span(𝑣1 , …, 𝑣𝑚 ) is a subspace of 𝑉. The additive identity is
in span(𝑣1 , …, 𝑣𝑚 ) because
0 = 0𝑣1 + ⋯ + 0𝑣𝑚 .
Also, span(𝑣1 , …, 𝑣𝑚 ) is closed under addition because
(𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 ) + (𝑐1 𝑣1 + ⋯ + 𝑐𝑚 𝑣𝑚 ) = (𝑎1 + 𝑐1 )𝑣1 + ⋯ + (𝑎𝑚 + 𝑐𝑚 )𝑣𝑚 .
Furthermore, span(𝑣1 , …, 𝑣𝑚 ) is closed under scalar multiplication because
𝜆(𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 ) = 𝜆𝑎1 𝑣1 + ⋯ + 𝜆𝑎𝑚 𝑣𝑚 .
Thus span(𝑣1 , …, 𝑣𝑚 ) is a subspace of 𝑉 (by 1.34).
Each 𝑣𝑘 is a linear combination of 𝑣1 , …, 𝑣𝑚 (to show this, set 𝑎𝑘 = 1 and let
the other 𝑎’s in 2.2 equal 0). Thus span(𝑣1 , …, 𝑣𝑚 ) contains each 𝑣𝑘 . Conversely,
because subspaces are closed under scalar multiplication and addition, every sub-
space of 𝑉 that contains each 𝑣𝑘 contains span(𝑣1 , …, 𝑣𝑚 ). Thus span(𝑣1 , …, 𝑣𝑚 )
is the smallest subspace of 𝑉 containing all the vectors 𝑣1 , …, 𝑣𝑚 .
2.7 definition: spans
If span(𝑣1 , …, 𝑣𝑚 ) equals 𝑉, we say that the list 𝑣1 , …, 𝑣𝑚 spans 𝑉.

2.8 example: a list that spans 𝐅𝑛

Suppose 𝑛 is a positive integer. We want to show that
(1, 0, …, 0), (0, 1, 0, …, 0), …, (0, …, 0, 1)
spans 𝐅𝑛. Here the 𝑘 th vector in the list above has 1 in the 𝑘 th slot and 0 in all other
slots.
Suppose (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛. Then
(𝑥1 , …, 𝑥𝑛 ) = 𝑥1 (1, 0, …, 0) + 𝑥2 (0, 1, 0, …, 0) + ⋯ + 𝑥𝑛 (0, …, 0, 1).
Thus (𝑥1 , …, 𝑥𝑛 ) ∈ span((1, 0, …, 0), (0, 1, 0, …, 0), …, (0, …, 0, 1)), as desired.
Now we can make one of the key definitions in linear algebra.
2.9 definition: finite-dimensional vector space
A vector space is called finite-dimensional if some list of vectors in it spans

the space.
Example 2.8 above shows that 𝐅𝑛 is a Recall that by definition every list has
finite-dimensional vector space for every finite length.
positive integer 𝑛.
The definition of a polynomial is no doubt already familiar to you.
2.10 definition: polynomial, 𝒫(𝐅)
• A function 𝑝 ∶ 𝐅 → 𝐅 is called a polynomial with coefficients in 𝐅 if there

exist 𝑎0 , …, 𝑎𝑚 ∈ 𝐅 such that
𝑝(𝑧) = 𝑎0 + 𝑎1 𝑧 + 𝑎2 𝑧2 + ⋯ + 𝑎𝑚 𝑧𝑚
for all 𝑧 ∈ 𝐅.
• 𝒫(𝐅) is the set of all polynomials with coefficients in 𝐅.
With the usual operations of addition and scalar multiplication, 𝒫(𝐅) is a

vector space over 𝐅, as you should verify. Hence 𝒫(𝐅) is a subspace of 𝐅𝐅, the
vector space of functions from 𝐅 to 𝐅.
If a polynomial (thought of as a function from 𝐅 to 𝐅 ) is represented by two
sets of coefficients, then subtracting one representation of the polynomial from
the other produces a polynomial that is identically zero as a function on 𝐅 and
hence has all zero coefficients (if you are unfamiliar with this fact, just believe
it for now; we will prove it later—see 4.8). Conclusion: the coefficients of a
polynomial are uniquely determined by the polynomial. Thus the next definition
uniquely defines the degree of a polynomial.
2.11 definition: degree of a polynomial, deg 𝑝
• A polynomial 𝑝 ∈ 𝒫(𝐅) is said to have degree 𝑚 if there exist scalars

𝑎0 , 𝑎1 , …, 𝑎𝑚 ∈ 𝐅 with 𝑎𝑚 ≠ 0 such that for every 𝑧 ∈ 𝐅, we have
𝑝(𝑧) = 𝑎0 + 𝑎1 𝑧 + ⋯ + 𝑎𝑚 𝑧𝑚.
• The polynomial that is identically 0 is said to have degree −∞.

• The degree of a polynomial 𝑝 is denoted by deg 𝑝.
In the next definition, we use the convention that −∞ < 𝑚, which means that
the polynomial 0 is in 𝒫𝑚 (𝐅).
2.12 notation: 𝒫𝑚 (𝐅)
For 𝑚 a nonnegative integer, 𝒫𝑚 (𝐅) denotes the set of all polynomials with
coefficients in 𝐅 and degree at most 𝑚.
If 𝑚 is a nonnegative integer, then 𝒫𝑚 (𝐅) = span(1, 𝑧, …, 𝑧𝑚 ) [here we slightly

abuse notation by letting 𝑧𝑘 denote a function]. Thus 𝒫𝑚 (𝐅) is a finite-dimensional
vector space for each nonnegative integer 𝑚.
2.13 definition: infinite-dimensional vector space
A vector space is called infinite-dimensional if it is not finite-dimensional.
2.14 example: 𝒫(𝐅) is infinite-dimensional.

Consider any list of elements of 𝒫(𝐅). Let 𝑚 denote the highest degree of the
polynomials in this list. Then every polynomial in the span of this list has degree
at most 𝑚. Thus 𝑧𝑚 + 1 is not in the span of our list. Hence no list spans 𝒫(𝐅).
Thus 𝒫(𝐅) is infinite-dimensional.
Linear Independence
Suppose 𝑣1 , …, 𝑣𝑚 ∈ 𝑉 and 𝑣 ∈ span(𝑣1 , …, 𝑣𝑚 ). By the definition of span, there
exist 𝑎1 , …, 𝑎𝑚 ∈ 𝐅 such that
𝑣 = 𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 .
Consider the question of whether the choice of scalars in the equation above is
unique. Suppose 𝑐1 , …, 𝑐𝑚 is another set of scalars such that
𝑣 = 𝑐1 𝑣1 + ⋯ + 𝑐𝑚 𝑣𝑚 .
Subtracting the last two equations, we have
0 = (𝑎1 − 𝑐1 )𝑣1 + ⋯ + (𝑎𝑚 − 𝑐𝑚 )𝑣𝑚 .
Thus we have written 0 as a linear combination of (𝑣1 , …, 𝑣𝑚 ). If the only way

to do this is by using 0 for all the scalars in the linear combination, then each
𝑎𝑘 − 𝑐𝑘 equals 0, which means that each 𝑎𝑘 equals 𝑐𝑘 (and thus the choice of
scalars was indeed unique). This situation is so important that we give it a special
name—linear independence—which we now define.
2.15 definition: linearly independent
• A list 𝑣1 , …, 𝑣𝑚 of vectors in 𝑉 is called linearly independent if the only

choice of 𝑎1 , …, 𝑎𝑚 ∈ 𝐅 that makes
𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 = 0
is 𝑎1 = ⋯ = 𝑎𝑚 = 0.
• The empty list ( ) is also declared to be linearly independent.
The reasoning above shows that 𝑣1 , …, 𝑣𝑚 is linearly independent if and only if

each vector in span(𝑣1 , …, 𝑣𝑚 ) has only one representation as a linear combination
of 𝑣1 , …, 𝑣𝑚 .
2.16 example: linearly independent lists

(a) To see that the list (1, 0, 0, 0), (0, 1, 0, 0), (0, 0, 1, 0) is linearly independent in
𝐅4, suppose 𝑎1 , 𝑎2 , 𝑎3 ∈ 𝐅 and
𝑎1 (1, 0, 0, 0) + 𝑎2 (0, 1, 0, 0) + 𝑎3 (0, 0, 1, 0) = (0, 0, 0, 0).
Thus
(𝑎1 , 𝑎2 , 𝑎3 , 0) = (0, 0, 0, 0).
Hence 𝑎1 = 𝑎2 = 𝑎3 = 0. Thus the list (1, 0, 0, 0), (0, 1, 0, 0), (0, 0, 1, 0) is
linearly independent in 𝐅4.
(b) Suppose 𝑚 is a nonnegative integer. To see that the list 1, 𝑧, …, 𝑧𝑚 is linearly
independent in 𝒫(𝐅), suppose 𝑎0 , 𝑎1 , …, 𝑎𝑚 ∈ 𝐅 and
𝑎0 + 𝑎1 𝑧 + ⋯ + 𝑎𝑚 𝑧𝑚 = 0,
where we think of both sides as elements of 𝒫(𝐅). Then
𝑎0 + 𝑎1 𝑧 + ⋯ + 𝑎𝑚 𝑧𝑚 = 0
for all 𝑧 ∈ 𝐅. As discussed earlier (and as follows from 4.8), this implies
that 𝑎0 = 𝑎1 = ⋯ = 𝑎𝑚 = 0. Thus 1, 𝑧, …, 𝑧𝑚 is a linearly independent list in
𝒫(𝐅).
(c) A list of length one in a vector space is linearly independent if and only if the
vector in the list is not 0.
(d) A list of length two in a vector space is linearly independent if and only if
neither of the two vectors in the list is a scalar multiple of the other.
If some vectors are removed from a linearly independent list, the remaining
list is also linearly independent, as you should verify.
2.17 definition: linearly dependent
• A list of vectors in 𝑉 is called linearly dependent if it is not linearly inde-

pendent.
• In other words, a list 𝑣1 , …, 𝑣𝑚 of vectors in 𝑉 is linearly dependent if there
exist 𝑎1 , …, 𝑎𝑚 ∈ 𝐅, not all 0, such that 𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 = 0.
2.18 example: linearly dependent lists

• (2, 3, 1), (1, −1, 2), (7, 3, 8) is linearly dependent in 𝐅3 because
2(2, 3, 1) + 3(1, −1, 2) + (−1)(7, 3, 8) = (0, 0, 0).
• The list (2, 3, 1), (1, −1, 2), (7, 3, 𝑐) is linearly dependent in 𝐅3 if and only if
𝑐 = 8, as you should verify.
• If some vector in a list of vectors in 𝑉 is a linear combination of the other
vectors, then the list is linearly dependent. (Proof: After writing one vector in
the list as equal to a linear combination of the other vectors, move that vector
to the other side of the equation, where it will be multiplied by −1.)
• Every list of vectors in 𝑉 containing the 0 vector is linearly dependent. (This is
a special case of the previous bullet point.)
The next lemma is a terrific tool. It states that given a linearly dependent list
of vectors, one of the vectors is in the span of the previous ones. Furthermore, we
can throw out that vector without changing the span of the original list.
2.19 linear dependence lemma
Suppose 𝑣1 , …, 𝑣𝑚 is a linearly dependent list in 𝑉. Then there exists

𝑘 ∈ {1, 2, …, 𝑚} such that
𝑣𝑘 ∈ span(𝑣1 , …, 𝑣𝑘 − 1 ).
Furthermore, if 𝑘 satisfies the condition above and the 𝑘 th term is removed
from 𝑣1 , …, 𝑣𝑚 , then the span of the remaining list equals span(𝑣1 , …, 𝑣𝑚 ).
Proof Because the list 𝑣1 , …, 𝑣𝑚 is linearly dependent, there exist numbers

𝑎1 , …, 𝑎𝑚 ∈ 𝐅, not all 0, such that
𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 = 0.
Let 𝑘 be the largest element of {1, …, 𝑚} such that 𝑎𝑘 ≠ 0. Then
𝑎1 𝑎𝑘 − 1
𝑣𝑘 = − 𝑣1 − ⋯ − 𝑣 ,
𝑎𝑘 𝑎𝑘 𝑘 − 1
which proves that 𝑣𝑘 ∈ span(𝑣1 , …, 𝑣𝑘 − 1 ), as desired.
Now suppose 𝑘 is any element of {1, …, 𝑚} such that 𝑣𝑘 ∈ span(𝑣1 , …, 𝑣𝑘 − 1 ).

Let 𝑏1 , …, 𝑏𝑘 − 1 ∈ 𝐅 be such that
2.20 𝑣𝑘 = 𝑏1 𝑣1 + ⋯ + 𝑏𝑘 − 1 𝑣𝑘 − 1 .
Suppose 𝑢 ∈ span(𝑣1 , …, 𝑣𝑚 ). Then there exist 𝑐1 , …, 𝑐𝑚 ∈ 𝐅 such that
𝑢 = 𝑐1 𝑣1 + ⋯ + 𝑐𝑚 𝑣𝑚 .
In the equation above, we can replace 𝑣𝑘 with the right side of 2.20, which shows
that 𝑢 is in the span of the list obtained by removing the 𝑘 th term from 𝑣1 , …, 𝑣𝑚 .
Thus removing the 𝑘 th term of the list 𝑣1 , …, 𝑣𝑚 does not change the span of the
list.
If 𝑘 = 1 in the linear dependence lemma, then 𝑣𝑘 ∈ span(𝑣1 , …, 𝑣𝑘 − 1 ) means

that 𝑣1 = 0, because span( ) = {0}. Note also that parts of the proof of the linear
dependence lemma need to be modified if 𝑘 = 1. In general, the proofs in the
rest of the book will not call attention to special cases that must be considered
involving lists of length 0, the subspace {0}, or other trivial cases for which the
result is true but needs a slightly different proof. Be sure to check these special
cases yourself.
2.21 example: smallest 𝑘 in linear dependence lemma

Consider the list
(1, 2, 3), (6, 5, 4), (15, 16, 17), (8, 9, 7)
in 𝐑 . This list of length four is linearly dependent, as we will soon see. Thus the
3
linear dependence lemma implies that there exists 𝑘 ∈ {1, 2, 3, 4} such that the 𝑘 th
vector in this list is a linear combination of the previous vectors in the list. Let’s
see how to find the smallest value of 𝑘 that works.
Taking 𝑘 = 1 in the linear dependence lemma works if and only if the first
vector in the list equals 0. Because (1, 2, 3) is not the 0 vector, we cannot take
𝑘 = 1 for this list.
Taking 𝑘 = 2 in the linear dependence lemma works if and only if the second
vector in the list is a scalar multiple of the first vector. However, there does not
exist 𝑐 ∈ 𝐑 such that (6, 5, 4) = 𝑐(1, 2, 3). Thus we cannot take 𝑘 = 2 for this list.
Taking 𝑘 = 3 in the linear dependence lemma works if and only if the third
vector in the list is a linear combination of the first two vectors. Thus for the list
in this example, we want to know whether there exist 𝑎, 𝑏 ∈ 𝐑 such that
(15, 16, 17) = 𝑎(1, 2, 3) + 𝑏(6, 5, 4).
The equation above is equivalent to a system of three linear equations in the two
unknowns 𝑎, 𝑏. Using Gaussian elimination or appropriate software, we find that
𝑎 = 3, 𝑏 = 2 is a solution of the equation above, as you can verify. Thus for the
list in this example, taking 𝑘 = 3 is the smallest value of 𝑘 that works in the linear
dependence lemma.
Now we come to a key result. It says that no linearly independent list in 𝑉 is

longer than a spanning list in 𝑉.
2.22 length of linearly independent list ≤ length of spanning list
In a finite-dimensional vector space, the length of every linearly independent

list of vectors is less than or equal to the length of every spanning list of
vectors.
Proof Suppose that 𝑢1 , …, 𝑢𝑚 is linearly independent in 𝑉. Suppose also that

𝑤1 , …, 𝑤𝑛 spans 𝑉. We need to prove that 𝑚 ≤ 𝑛. We do so through the process
described below with 𝑚 steps; note that in each step we add one of the 𝑢’s and
remove one of the 𝑤’s.
Step 1
Let 𝐵 be the list 𝑤1 , …, 𝑤𝑛 , which spans 𝑉. Adjoining 𝑢1 at the beginning of
this list produces a linearly dependent list (because 𝑢1 can be written as a linear
combination of 𝑤1 , …, 𝑤𝑚 ). In other words, the list
𝑢1 , 𝑤1 , …, 𝑤𝑛
is linearly dependent.
Thus by the linear dependence lemma (2.19), one of the vectors in the list above
is a linear combination of the previous vectors in the list. We know that 𝑢1 ≠ 0
because the list 𝑢1 , …, 𝑢𝑚 is linearly independent. Thus 𝑢1 is not in the span
of the previous vectors in the list above (because 𝑢1 is not in {0}, which is the
span of the empty list). Hence the linear dependence lemma implies that we
can remove one of the 𝑤’s so that the new list 𝐵 (of length 𝑛) consisting of 𝑢1
and the remaining 𝑤’s spans 𝑉.
Step k, for k = 2, …, m
The list 𝐵 (of length 𝑛) from step 𝑘 − 1 spans 𝑉. In particular, 𝑢𝑘 is in the span of
the list 𝐵. Thus the list of length (𝑛 + 1) obtained by adjoining 𝑢𝑘 to 𝐵, placing
it just after 𝑢1 , …, 𝑢𝑘 − 1 , is linearly dependent. By the linear dependence lemma
(2.19), one of the vectors in this list is in the span of the previous ones, and
because 𝑢1 , …, 𝑢𝑘 is linearly independent, this vector cannot be one of the 𝑢’s.
Hence there still must be at least one remaining 𝑤 at this step. We can remove
from our new list (after adjoining 𝑢𝑘 in the proper place) a 𝑤 that is a linear
combination of the previous vectors in the list, so that the new list 𝐵 (of length
𝑛) consisting of 𝑢1 , …, 𝑢𝑘 and the remaining 𝑤’s spans 𝑉.
After step 𝑚, we have added all the 𝑢’s and the process stops. At each step
as we add a 𝑢 to 𝐵, the linear dependence lemma implies that there is some 𝑤 to
remove. Thus there are at least as many 𝑤’s as 𝑢’s.
The next two examples show how the result above can be used to show, without
any computations, that certain lists are not linearly independent and that certain
lists do not span a given vector space.
2.23 example: no list of length 4 is linearly independent in 𝐑3

The list (1, 0, 0), (0, 1, 0), (0, 0, 1), which has length three, spans 𝐑3. Thus no
list of length larger than three is linearly independent in 𝐑3.
For example, we now know that (1, 2, 3), (4, 5, 8), (9, 6, 7), (−3, 2, 8), which
is a list of length four, is not linearly independent in 𝐑3.
2.24 example: no list of length 3 spans 𝐑4

The list (1, 0, 0, 0), (0, 1, 0, 0), (0, 0, 1, 0), (0, 0, 0, 1), which has length four, is
linearly independent in 𝐑4. Thus no list of length less than four spans 𝐑4.
For example, we now know that (1, 2, 3, −5), (4, 5, 8, 3), (9, 6, 7, −1), which
is a list of length three, does not span 𝐑4.
Our intuition suggests that every subspace of a finite-dimensional vector space

should also be finite-dimensional. We now prove that this intuition is correct.
2.25 finite-dimensional subspaces
Every subspace of a finite-dimensional vector space is finite-dimensional.
Proof Suppose 𝑉 is finite-dimensional and 𝑈 is a subspace of 𝑉. We need to

prove that 𝑈 is finite-dimensional. We do this through the following multistep
construction.
Step 1
If 𝑈 = {0}, then 𝑈 is finite-dimensional and we are done. If 𝑈 ≠ {0}, then
choose a nonzero vector 𝑢1 ∈ 𝑈.
Step k
If 𝑈 = span(𝑢1 , …, 𝑢𝑘 − 1 ), then 𝑈 is finite-dimensional and we are done. If
𝑈 ≠ span(𝑢1 , …, 𝑢𝑘 − 1 ), then choose a vector 𝑢𝑘 ∈ 𝑈 such that
𝑢𝑘 ∉ span(𝑢1 , …, 𝑢𝑘 − 1 ).
After each step, as long as the process continues, we have constructed a list
of vectors such that no vector in this list is in the span of the previous vectors.
Thus after each step we have constructed a linearly independent list, by the linear
dependence lemma (2.19). This linearly independent list cannot be longer than
any spanning list of 𝑉 (by 2.22). Thus the process eventually terminates, which
means that 𝑈 is finite-dimensional.
Exercises 2A
1 Find a list of four distinct vectors in 𝐅3 whose span equals

{(𝑥, 𝑦, 𝑧) ∈ 𝐅3 ∶ 𝑥 + 𝑦 + 𝑧 = 0}.
2 Prove or give a counterexample: If 𝑣1 , 𝑣2 , 𝑣3 , 𝑣4 spans 𝑉, then the list
𝑣1 − 𝑣2 , 𝑣2 − 𝑣3 , 𝑣3 − 𝑣4 , 𝑣4
also spans 𝑉.
3 Suppose 𝑣1 , …, 𝑣𝑚 is a list of vectors in 𝑉. For 𝑘 ∈ {1, …, 𝑚}, let
𝑤𝑘 = 𝑣1 + ⋯ + 𝑣𝑘 .
Show that span(𝑣1 , …, 𝑣𝑚 ) = span(𝑤1 , …, 𝑤𝑚 ).

4 (a) Show that a list of length one in a vector space is linearly independent
if and only if the vector in the list is not 0.
(b) Show that a list of length two in a vector space is linearly independent
if and only if neither of the two vectors in the list is a scalar multiple of
the other.
5 Find a number 𝑡 such that
(3, 1, 4), (2, −3, 5), (5, 9, 𝑡)
is not linearly independent in 𝐑3.

6 Show that the list (2, 3, 1), (1, −1, 2), (7, 3, 𝑐) is linearly dependent in 𝐅3 if
and only if 𝑐 = 8.
7 (a) Show that if we think of 𝐂 as a vector space over 𝐑 , then the list
1 + 𝑖, 1 − 𝑖 is linearly independent.
(b) Show that if we think of 𝐂 as a vector space over 𝐂 , then the list
1 + 𝑖, 1 − 𝑖 is linearly dependent.
8 Suppose 𝑣1 , 𝑣2 , 𝑣3 , 𝑣4 is linearly independent in 𝑉. Prove that the list
𝑣1 − 𝑣2 , 𝑣2 − 𝑣3 , 𝑣3 − 𝑣4 , 𝑣4
is also linearly independent.

9 Prove or give a counterexample: If 𝑣1 , 𝑣2 , …, 𝑣𝑚 is a linearly independent
list of vectors in 𝑉, then
5𝑣1 − 4𝑣2 , 𝑣2 , 𝑣3 , …, 𝑣𝑚
is linearly independent.
10 Prove or give a counterexample: If 𝑣1 , 𝑣2 , …, 𝑣𝑚 is a linearly independent

list of vectors in 𝑉 and 𝜆 ∈ 𝐅 with 𝜆 ≠ 0, then 𝜆𝑣1 , 𝜆𝑣2 , …, 𝜆𝑣𝑚 is linearly
independent.
11 Prove or give a counterexample: If 𝑣1 , …, 𝑣𝑚 and 𝑤1 , …, 𝑤𝑚 are linearly
independent lists of vectors in 𝑉, then the list 𝑣1 + 𝑤1 , …, 𝑣𝑚 + 𝑤𝑚 is linearly
independent.
12 Suppose 𝑣1 , …, 𝑣𝑚 is linearly independent in 𝑉 and 𝑤 ∈ 𝑉. Prove that if
𝑣1 + 𝑤, …, 𝑣𝑚 + 𝑤 is linearly dependent, then 𝑤 ∈ span(𝑣1 , …, 𝑣𝑚 ).
13 Suppose 𝑣1 , …, 𝑣𝑚 is linearly independent in 𝑉 and 𝑤 ∈ 𝑉. Show that
𝑣1 , …, 𝑣𝑚 , 𝑤 is linearly independent ⟺ 𝑤 ∉ span(𝑣1 , …, 𝑣𝑚 ).
𝑤𝑘 = 𝑣1 + ⋯ + 𝑣𝑘 .
Show that the list 𝑣1 , …, 𝑣𝑚 is linearly independent if and only if the list
𝑤1 , …, 𝑤𝑚 is linearly independent.
15 Explain why there does not exist a list of six polynomials that is linearly
independent in 𝒫4 (𝐅).
16 Explain why no list of four polynomials spans 𝒫4 (𝐅).
17 Prove that 𝑉 is infinite-dimensional if and only if there is a sequence 𝑣1 , 𝑣2 , …
of vectors in 𝑉 such that 𝑣1 , …, 𝑣𝑚 is linearly independent for every positive
integer 𝑚.
18 Prove that 𝐅∞ is infinite-dimensional.
19 Prove that the real vector space of all continuous real-valued functions on
the interval [0, 1] is infinite-dimensional.
20 Suppose 𝑝0 , 𝑝1 , …, 𝑝𝑚 are polynomials in 𝒫𝑚 (𝐅) such that 𝑝𝑘 (2) = 0 for each
𝑘 ∈ {0, …, 𝑚}. Prove that 𝑝0 , 𝑝1 , …, 𝑝𝑚 is not linearly independent in 𝒫𝑚 (𝐅).
Section 2B Bases 39
2B Bases
In the previous section, we discussed linearly independent lists and we also
discussed spanning lists. Now we bring these concepts together by considering
lists that have both properties.
2.26 definition: basis

A basis of 𝑉 is a list of vectors in 𝑉 that is linearly independent and spans 𝑉.
2.27 example: bases

(a) The list (1, 0, …, 0), (0, 1, 0, …, 0), …, (0, …, 0, 1) is a basis of 𝐅𝑛, called the
standard basis of 𝐅𝑛.
(b) The list (1, 2), (3, 5) is a basis of 𝐅2. Note that this list has length two, which
is the same as the length of the standard basis of 𝐅2. In the next section, we
will see that this is not a coincidence.
(c) The list (1, 2, −4), (7, −5, 6) is linearly independent in 𝐅3 but is not a basis
of 𝐅3 because it does not span 𝐅3.
(d) The list (1, 2), (3, 5), (4, 13) spans 𝐅2 but is not a basis of 𝐅2 because it is not
linearly independent.
(e) The list (1, 1, 0), (0, 0, 1) is a basis of {(𝑥, 𝑥, 𝑦) ∈ 𝐅3 ∶ 𝑥, 𝑦 ∈ 𝐅}.
(f) The list (1, −1, 0), (1, 0, −1) is a basis of
{(𝑥, 𝑦, 𝑧) ∈ 𝐅3 ∶ 𝑥 + 𝑦 + 𝑧 = 0}.
(g) The list 1, 𝑧, …, 𝑧𝑚 is a basis of 𝒫𝑚 (𝐅), called the standard basis of 𝒫𝑚 (𝐅).
In addition to the standard basis, 𝐅𝑛 has many other bases. For example,
(7, 5), (−4, 9) and (1, 2), (3, 5)
are both bases of 𝐅2.
The next result helps explain why bases are useful. Recall that “uniquely”
means “in only one way”.
2.28 criterion for basis
A list 𝑣1 , …, 𝑣𝑛 of vectors in 𝑉 is a basis of 𝑉 if and only if every 𝑣 ∈ 𝑉 can

be written uniquely in the form
2.29 𝑣 = 𝑎1 𝑣1 + ⋯ + 𝑎𝑛 𝑣𝑛 ,
where 𝑎1 , …, 𝑎𝑛 ∈ 𝐅.
Proof First suppose that 𝑣1 , …, 𝑣𝑛 is a This proof is essentially a repetition of

basis of 𝑉. Let 𝑣 ∈ 𝑉. Because 𝑣1 , …, 𝑣𝑛 the ideas that led us to the definition of
spans 𝑉, there exist 𝑎1 , …, 𝑎𝑛 ∈ 𝐅 such linear independence.
that 2.29 holds. To show that the repre-
sentation in 2.29 is unique, suppose 𝑐1 , …, 𝑐𝑛 are scalars such that we also have
𝑣 = 𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 .
Subtracting the last equation from 2.29, we get
0 = (𝑎1 − 𝑐1 )𝑣1 + ⋯ + (𝑎𝑛 − 𝑐𝑛 )𝑣𝑛 .
This implies that each 𝑎𝑘 − 𝑐𝑘 equals 0 (because 𝑣1 , …, 𝑣𝑛 is linearly independent).
Hence 𝑎1 = 𝑐1 , …, 𝑎𝑛 = 𝑐𝑛 . We have the desired uniqueness, completing the proof
in one direction.
For the other direction, suppose every 𝑣 ∈ 𝑉 can be written uniquely in the
form given by 2.29. This implies that the list 𝑣1 , …, 𝑣𝑛 spans 𝑉. To show that
𝑣1 , …, 𝑣𝑛 is linearly independent, suppose 𝑎1 , …, 𝑎𝑛 ∈ 𝐅 are such that
0 = 𝑎1 𝑣1 + ⋯ + 𝑎𝑛 𝑣𝑛 .
The uniqueness of the representation 2.29 (taking 𝑣 = 0) now implies that
𝑎1 = ⋯ = 𝑎𝑛 = 0. Thus 𝑣1 , …, 𝑣𝑛 is linearly independent and hence is a basis
of 𝑉.
A spanning list in a vector space may not be a basis because it is not linearly
independent. Our next result says that given any spanning list, some (possibly
none) of the vectors in it can be discarded so that the remaining list is linearly
independent and still spans the vector space.
As an example in the vector space 𝐅2, if the procedure in the proof below is
applied to the list (1, 2), (3, 6), (4, 7), (5, 9), then the second and fourth vectors
will be removed. This leaves (1, 2), (4, 7), which is a basis of 𝐅2.
2.30 every spanning list contains a basis
Every spanning list in a vector space can be reduced to a basis of the vector
space.
Proof Suppose 𝑣1 , …, 𝑣𝑛 spans 𝑉. We want to remove some of the vectors from

𝑣1 , …, 𝑣𝑛 so that the remaining vectors form a basis of 𝑉. We do this through the
multistep process described below.
Start with 𝐵 equal to the list 𝑣1 , …, 𝑣𝑛 .
Step 1
If 𝑣1 = 0, then delete 𝑣1 from 𝐵. If 𝑣1 ≠ 0, then leave 𝐵 unchanged.
Step k
If 𝑣𝑘 is in span(𝑣1 , …, 𝑣𝑘 − 1 ), then delete 𝑣𝑘 from the list 𝐵. If 𝑣𝑘 is not in
span(𝑣1 , …, 𝑣𝑘 − 1 ), then leave 𝐵 unchanged.
Section 2B Bases 41
Stop the process after step 𝑛, getting a list 𝐵. This list 𝐵 spans 𝑉 because
our original list spanned 𝑉 and we have discarded only vectors that were already
in the span of the previous vectors. The process ensures that no vector in 𝐵 is
in the span of the previous ones. Thus 𝐵 is linearly independent, by the linear
dependence lemma (2.19). Hence 𝐵 is a basis of 𝑉.
We now come to an important corollary of the previous result.
2.31 basis of finite-dimensional vector space
Every finite-dimensional vector space has a basis.
Proof By definition, a finite-dimensional vector space has a spanning list. The

previous result tells us that each spanning list can be reduced to a basis.
Our next result is in some sense a dual of 2.30, which said that every spanning
list can be reduced to a basis. Now we show that given any linearly independent list,
we can adjoin some additional vectors (this includes the possibility of adjoining
no additional vectors) so that the extended list is still linearly independent but
also spans the space.
2.32 every linearly independent list extends to a basis
Every linearly independent list of vectors in a finite-dimensional vector space

can be extended to a basis of the vector space.
Proof Suppose 𝑢1 , …, 𝑢𝑚 is linearly independent in a finite-dimensional vector

space 𝑉. Let 𝑤1 , …, 𝑤𝑛 be a list of vectors in 𝑉 that spans 𝑉. Thus the list
𝑢1 , …, 𝑢𝑚 , 𝑤1 , …, 𝑤𝑛
spans 𝑉. Applying the procedure of the proof of 2.30 to reduce this list to a
basis of 𝑉 produces a basis consisting of the vectors 𝑢1 , …, 𝑢𝑚 and some of the
𝑤’s (none of the 𝑢’s get deleted in this procedure because 𝑢1 , …, 𝑢𝑚 is linearly
independent).
As an example in 𝐅3, suppose we start with the linearly independent list

(2, 3, 4), (9, 6, 8). If we take 𝑤1 , 𝑤2 , 𝑤3 to be the standard basis of 𝐅3, then applying
the procedure in the proof above produces the list
(2, 3, 4), (9, 6, 8), (0, 1, 0),
which is a basis of 𝐅 .3
As an application of the result above, Using the same ideas but more ad-
we now show that every subspace of a vanced tools, the next result can be
finite-dimensional vector space can be proved without the hypothesis that 𝑉 is
paired with another subspace to form a finite-dimensional.
direct sum of the whole space.
2.33 every subspace of 𝑉 is part of a direct sum equal to 𝑉
Suppose 𝑉 is finite-dimensional and 𝑈 is a subspace of 𝑉. Then there is a

subspace 𝑊 of 𝑉 such that 𝑉 = 𝑈 ⊕ 𝑊.
Proof Because 𝑉 is finite-dimensional, so is 𝑈 (see 2.25). Thus there is a basis

𝑢1 , …, 𝑢𝑚 of 𝑈 (by 2.31). Of course 𝑢1 , …, 𝑢𝑚 is a linearly independent list of
vectors in 𝑉. Hence this list can be extended to a basis 𝑢1 , …, 𝑢𝑚 , 𝑤1 , …, 𝑤𝑛 of 𝑉
(by 2.32). Let 𝑊 = span(𝑤1 , …, 𝑤𝑛 ).
To prove that 𝑉 = 𝑈 ⊕ 𝑊, by 1.46 we only need to show that
𝑉 =𝑈+𝑊 and 𝑈 ∩ 𝑊 = {0}.
To prove the first equation above, suppose 𝑣 ∈ 𝑉. Then, because the list
𝑢1 , …, 𝑢𝑚 , 𝑤1 , …, 𝑤𝑛 spans 𝑉, there exist 𝑎1 , …, 𝑎𝑚 , 𝑏1 , …, 𝑏𝑛 ∈ 𝐅 such that
𝑣 = 𝑎⏟
1𝑢
⏟ +⏟
1⏟ ⋯ +⏟
𝑎⏟
𝑚𝑢
⏟𝑚+⏟
𝑏1 𝑤1+
⏟⏟⏟ ⋯ +⏟𝑏⏟
𝑛𝑤
⏟𝑛 .
𝑢 𝑤
We have 𝑣 = 𝑢 + 𝑤, where 𝑢 ∈ 𝑈 and 𝑤 ∈ 𝑊 are defined as above. Thus

𝑣 ∈ 𝑈 + 𝑊, completing the proof that 𝑉 = 𝑈 + 𝑊.
To show that 𝑈 ∩ 𝑊 = {0}, suppose 𝑣 ∈ 𝑈 ∩ 𝑊. Then there exist scalars
𝑎1 , …, 𝑎𝑚 , 𝑏1 , …, 𝑏𝑛 ∈ 𝐅 such that
𝑣 = 𝑎1 𝑢1 + ⋯ + 𝑎𝑚 𝑢𝑚 = 𝑏1 𝑤1 + ⋯ + 𝑏𝑛 𝑤𝑛 .
Thus
𝑎1 𝑢1 + ⋯ + 𝑎𝑚 𝑢𝑚 − 𝑏1 𝑤1 − ⋯ − 𝑏𝑛 𝑤𝑛 = 0.
Because 𝑢1 , …, 𝑢𝑚 , 𝑤1 , …, 𝑤𝑛 is linearly independent, this implies that
𝑎1 = ⋯ = 𝑎𝑚 = 𝑏1 = ⋯ = 𝑏𝑛 = 0.
Thus 𝑣 = 0, completing the proof that 𝑈 ∩ 𝑊 = {0}.
Exercises 2B
1 Find all vector spaces that have exactly one basis.

2 Verify all assertions in Example 2.27.
3 (a) Let 𝑈 be the subspace of 𝐑5 defined by
𝑈 = {(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 , 𝑥5 ) ∈ 𝐑5 ∶ 𝑥1 = 3𝑥2 and 𝑥3 = 7𝑥4 }.
Find a basis of 𝑈.
(b) Extend the basis in (a) to a basis of 𝐑5.
(c) Find a subspace 𝑊 of 𝐑5 such that 𝐑5 = 𝑈 ⊕ 𝑊.
Section 2B Bases 43
4 (a) Let 𝑈 be the subspace of 𝐂5 defined by
𝑈 = {(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 , 𝑧5 ) ∈ 𝐂5 ∶ 6𝑧1 = 𝑧2 and 𝑧3 + 2𝑧4 + 3𝑧5 = 0}.
Find a basis of 𝑈.
(b) Extend the basis in (a) to a basis of 𝐂5.
(c) Find a subspace 𝑊 of 𝐂5 such that 𝐂5 = 𝑈 ⊕ 𝑊.
5 Suppose 𝑉 is finite-dimensional and 𝑈, 𝑊 are subspaces of 𝑉 such that
𝑉 = 𝑈 + 𝑊. Prove that there exists a basis of 𝑉 consisting of vectors in
𝑈 ∪ 𝑊.
6 Prove or give a counterexample: If 𝑝0 , 𝑝1 , 𝑝2 , 𝑝3 is a list in 𝒫3 (𝐅) such that

none of the polynomials 𝑝0 , 𝑝1 , 𝑝2 , 𝑝3 has degree 2, then 𝑝0 , 𝑝1 , 𝑝2 , 𝑝3 is not
a basis of 𝒫3 (𝐅).
7 Suppose 𝑣1 , 𝑣2 , 𝑣3 , 𝑣4 is a basis of 𝑉. Prove that
𝑣1 + 𝑣2 , 𝑣2 + 𝑣3 , 𝑣3 + 𝑣4 , 𝑣4
is also a basis of 𝑉.
8 Prove or give a counterexample: If 𝑣1 , 𝑣2 , 𝑣3 , 𝑣4 is a basis of 𝑉 and 𝑈 is a
subspace of 𝑉 such that 𝑣1 , 𝑣2 ∈ 𝑈 and 𝑣3 ∉ 𝑈 and 𝑣4 ∉ 𝑈, then 𝑣1 , 𝑣2 is a
basis of 𝑈.
𝑤𝑘 = 𝑣1 + ⋯ + 𝑣𝑘 .
Show that 𝑣1 , …, 𝑣𝑚 is a basis of 𝑉 if and only if 𝑤1 , …, 𝑤𝑚 is a basis of 𝑉.

10 Suppose 𝑈 and 𝑊 are subspaces of 𝑉 such that 𝑉 = 𝑈 ⊕ 𝑊. Suppose also
that 𝑢1 , …, 𝑢𝑚 is a basis of 𝑈 and 𝑤1 , …, 𝑤𝑛 is a basis of 𝑊. Prove that
𝑢1 , …, 𝑢𝑚 , 𝑤1 , …, 𝑤𝑛
is a basis of 𝑉.
11 Suppose 𝑉 is a real vector space. Show that if 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 (as a
real vector space), then 𝑣1 , …, 𝑣𝑛 is also a basis of the complexification 𝑉𝐂
(as a complex vector space).
See Exercise 8 in Section 1B for the definition of the complexification 𝑉𝐂 .
2C Dimension
Although we have been discussing finite-dimensional vector spaces, we have not
yet defined the dimension of such an object. How should dimension be defined?
A reasonable definition should force the dimension of 𝐅𝑛 to equal 𝑛. Notice that
the standard basis
(1, 0, …, 0), (0, 1, 0, …, 0), …, (0, …, 0, 1)
of 𝐅𝑛 has length 𝑛. Thus we are tempted to define the dimension as the length of
a basis. However, a finite-dimensional vector space in general has many different
bases, and our attempted definition makes sense only if all bases in a given vector
space have the same length. Fortunately that turns out to be the case, as we now
show.
2.34 basis length does not depend on basis
Any two bases of a finite-dimensional vector space have the same length.
Proof Suppose 𝑉 is finite-dimensional. Let 𝐵1 and 𝐵2 be two bases of 𝑉. Then

𝐵1 is linearly independent in 𝑉 and 𝐵2 spans 𝑉, so the length of 𝐵1 is at most the
length of 𝐵2 (by 2.22). Interchanging the roles of 𝐵1 and 𝐵2 , we also see that the
length of 𝐵2 is at most the length of 𝐵1 . Thus the length of 𝐵1 equals the length
of 𝐵2 , as desired.
Now that we know that any two bases of a finite-dimensional vector space
have the same length, we can formally define the dimension of such spaces.
2.35 definition: dimension, dim 𝑉
• The dimension of a finite-dimensional vector space is the length of any

basis of the vector space.
• The dimension of a finite-dimensional vector space 𝑉 is denoted by dim 𝑉.
2.36 example: dimensions
• dim 𝐅𝑛 = 𝑛 because the standard basis of 𝐅𝑛 has length 𝑛.

• dim 𝒫𝑚 (𝐅) = 𝑚 + 1 because the standard basis 1, 𝑧, …, 𝑧𝑚 of 𝒫𝑚 (𝐅) has length
𝑚 + 1.
• If 𝑈 = {(𝑥, 𝑥, 𝑦) ∈ 𝐅3 ∶ 𝑥, 𝑦 ∈ 𝐅}, then dim 𝑈 = 2 because (1, 1, 0), (0, 0, 1) is
a basis of 𝑈.
• If 𝑈 = {(𝑥, 𝑦, 𝑧) ∈ 𝐅3 ∶ 𝑥 + 𝑦 + 𝑧 = 0}, then dim 𝑈 = 2 because the list
(1, −1, 0), (1, 0, −1) is a basis of 𝑈.
Section 2C Dimension 45
Every subspace of a finite-dimensional vector space is finite-dimensional

(by 2.25) and so has a dimension. The next result gives the expected inequality
about the dimension of a subspace.
2.37 dimension of a subspace
If 𝑉 is finite-dimensional and 𝑈 is a subspace of 𝑉, then dim 𝑈 ≤ dim 𝑉.
Proof Suppose 𝑉 is finite-dimensional and 𝑈 is a subspace of 𝑉. Think of a basis

of 𝑈 as a linearly independent list in 𝑉, and think of a basis of 𝑉 as a spanning
list in 𝑉. Now use 2.22 to conclude that dim 𝑈 ≤ dim 𝑉.
To check that a list of vectors in 𝑉 The real vector space 𝐑2 has dimen-
is a basis of 𝑉, we must, according to sion two; the complex vector space 𝐂
the definition, show that the list in ques- has dimension one. As sets, 𝐑2 can
tion satisfies two properties: it must be be identified with 𝐂 (and addition is
linearly independent and it must span 𝑉. the same on both spaces, as is scalar
The next two results show that if the list multiplication by real numbers). Thus
in question has the right length, then we when we talk about the dimension of
only need to check that it satisfies one a vector space, the role played by the
of the two required properties. First we choice of 𝐅 cannot be neglected.
prove that every linearly independent list
of the right length is a basis.
2.38 linearly independent list of the right length is a basis
Suppose 𝑉 is finite-dimensional. Then every linearly independent list of

vectors in 𝑉 of length dim 𝑉 is a basis of 𝑉.
Proof Suppose dim 𝑉 = 𝑛 and 𝑣1 , …, 𝑣𝑛 is linearly independent in 𝑉. The list

𝑣1 , …, 𝑣𝑛 can be extended to a basis of 𝑉 (by 2.32). However, every basis of 𝑉 has
length 𝑛, so in this case the extension is the trivial one, meaning that no elements
are adjoined to 𝑣1 , …, 𝑣𝑛 . Thus 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉, as desired.
The next result is a useful consequence of the previous result.
2.39 subspace of full dimension equals the whole space
Suppose that 𝑉 is finite-dimensional and 𝑈 is a subspace of 𝑉 such that

dim 𝑈 = dim 𝑉. Then 𝑈 = 𝑉.
Proof Let 𝑢1 , …, 𝑢𝑛 be a basis of 𝑈. Thus 𝑛 = dim 𝑈, and by hypothesis we

also have 𝑛 = dim 𝑉. Thus 𝑢1 , …, 𝑢𝑛 is a linearly independent list of vectors in 𝑉
(because it is a basis of 𝑈) of length dim 𝑉. From 2.38, we see that 𝑢1 , …, 𝑢𝑛 is
a basis of 𝑉. In particular every vector in 𝑉 is a linear combination of 𝑢1 , …, 𝑢𝑛 .
Thus 𝑈 = 𝑉.
2.40 example: a basis of 𝐅2

Consider the list (5, 7), (4, 3) of vectors in 𝐅2. This list of length two is linearly
independent in 𝐅2 (because neither vector is a scalar multiple of the other). Note
that 𝐅2 has dimension two. Thus 2.38 implies that the linearly independent list
(5, 7), (4, 3) of length two is a basis of 𝐅2 (we do not need to bother checking that
it spans 𝐅2 ).
2.41 example: a basis of a subspace of 𝒫3 (𝐑)

Let 𝑈 be the subspace of 𝒫3 (𝐑) defined by
𝑈 = {𝑝 ∈ 𝒫3 (𝐑) ∶ 𝑝′(5) = 0}.
To find a basis of 𝑈, first note that each of the polynomials 1, (𝑥 − 5)2, and (𝑥 − 5)3
is in 𝑈.
Suppose 𝑎, 𝑏, 𝑐 ∈ 𝐑 and
𝑎 + 𝑏(𝑥 − 5)2 + 𝑐(𝑥 − 5)3 = 0
for every 𝑥 ∈ 𝐑 . Without explicitly expanding the left side of the equation above,
we can see that the left side has a 𝑐𝑥3 term. Because the right side has no 𝑥3
term, this implies that 𝑐 = 0. Because 𝑐 = 0, we see that the left side has a 𝑏𝑥2
term, which implies that 𝑏 = 0. Because 𝑏 = 𝑐 = 0, we can also conclude that
𝑎 = 0. Thus the equation above implies that 𝑎 = 𝑏 = 𝑐 = 0. Hence the list
1, (𝑥 − 5)2, (𝑥 − 5)3 is linearly independent in 𝑈. Thus 3 ≤ dim 𝑈. Hence
3 ≤ dim 𝑈 ≤ dim 𝒫3 (𝐑) = 4,
where we have used 2.37.

The polynomial 𝑥 is not in 𝑈 because its derivative is the constant function 1.
Thus 𝑈 ≠ 𝒫3 (𝐑). Hence dim 𝑈 ≠ 4 (by 2.39). The inequality above now implies
that dim 𝑈 = 3. Thus the linearly independent list 1, (𝑥 − 5)2, (𝑥 − 5)3 in 𝑈 has
length dim 𝑈 and hence is a basis of 𝑈 (by 2.38).
Now we prove that a spanning list of the right length is a basis.
2.42 spanning list of the right length is a basis
Suppose 𝑉 is finite-dimensional. Then every spanning list of vectors in 𝑉 of

length dim 𝑉 is a basis of 𝑉.
Proof Suppose dim 𝑉 = 𝑛 and 𝑣1 , …, 𝑣𝑛 spans 𝑉. The list 𝑣1 , …, 𝑣𝑛 can be

reduced to a basis of 𝑉 (by 2.30). However, every basis of 𝑉 has length 𝑛, so in
this case the reduction is the trivial one, meaning that no elements are deleted
from 𝑣1 , …, 𝑣𝑛 . Thus 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉, as desired.
The next result gives a formula for the dimension of the sum of two subspaces
of a finite-dimensional vector space. This formula is analogous to a familiar
counting formula: the number of elements in the union of two finite sets equals
the number of elements in the first set, plus the number of elements in the second
set, minus the number of elements in the intersection of the two sets.
2.43 dimension of a sum
If 𝑉1 and 𝑉2 are subspaces of a finite-dimensional vector space, then
dim(𝑉1 + 𝑉2 ) = dim 𝑉1 + dim 𝑉2 − dim(𝑉1 ∩ 𝑉2 ).
Proof Let 𝑣1 , …, 𝑣𝑚 be a basis of 𝑉1 ∩ 𝑉2 ; thus dim(𝑉1 ∩ 𝑉2 ) = 𝑚. Because

𝑣1 , …, 𝑣𝑚 is a basis of 𝑉1 ∩ 𝑉2 , it is linearly independent in 𝑉1 . Hence this list can
be extended to a basis 𝑣1 , …, 𝑣𝑚 , 𝑢1 , …, 𝑢𝑗 of 𝑉1 (by 2.32). Thus dim 𝑉1 = 𝑚 + 𝑗.
Also extend 𝑣1 , …, 𝑣𝑚 to a basis 𝑣1 , …, 𝑣𝑚 , 𝑤1 , …, 𝑤𝑘 of 𝑉2 ; thus dim 𝑉2 = 𝑚 + 𝑘.
We will show that
2.44 𝑣1 , …, 𝑣𝑚 , 𝑢1 , …, 𝑢𝑗 , 𝑤1 , …, 𝑤𝑘
is a basis of 𝑉1 + 𝑉2 . This will complete the proof, because then we will have
dim(𝑉1 + 𝑉2 ) = 𝑚 + 𝑗 + 𝑘
= (𝑚 + 𝑗) + (𝑚 + 𝑘) − 𝑚
= dim 𝑉1 + dim 𝑉2 − dim(𝑉1 ∩ 𝑉2 ).
The list 2.44 is contained in 𝑉1 ∪ 𝑉2 and thus is contained in 𝑉1 + 𝑉2 . The
span of this list contains 𝑉1 and contains 𝑉2 and hence is equal to 𝑉1 + 𝑉2 . Thus
to show that 2.44 is a basis of 𝑉1 + 𝑉2 we only need to show that it is linearly
independent.
To prove that 2.44 is linearly independent, suppose
𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 + 𝑏1 𝑢1 + ⋯ + 𝑏𝑗 𝑢𝑗 + 𝑐1 𝑤1 + ⋯ + 𝑐𝑘 𝑤𝑘 = 0,
where all the 𝑎’s, 𝑏’s, and 𝑐’s are scalars. We need to prove that all the 𝑎’s, 𝑏’s,
and 𝑐’s equal 0. The equation above can be rewritten as
2.45 𝑐1 𝑤1 + ⋯ + 𝑐𝑘 𝑤𝑘 = −𝑎1 𝑣1 − ⋯ − 𝑎𝑚 𝑣𝑚 − 𝑏1 𝑢1 − ⋯ − 𝑏𝑗 𝑢𝑗 ,
which shows that 𝑐1 𝑤1 + ⋯ + 𝑐𝑘 𝑤𝑘 ∈ 𝑉1 . All the 𝑤’s are in 𝑉2 , so this implies
that 𝑐1 𝑤1 + ⋯ + 𝑐𝑘 𝑤𝑘 ∈ 𝑉1 ∩ 𝑉2 . Because 𝑣1 , …, 𝑣𝑚 is a basis of 𝑉1 ∩ 𝑉2 , we have
𝑐1 𝑤1 + ⋯ + 𝑐𝑘 𝑤𝑘 = 𝑑1 𝑣1 + ⋯ + 𝑑𝑚 𝑣𝑚
for some scalars 𝑑1 , …, 𝑑𝑚 . But 𝑣1 , …, 𝑣𝑚 , 𝑤1 , …, 𝑤𝑘 is linearly independent, so
the last equation implies that all the 𝑐’s (and 𝑑’s) equal 0. Thus 2.45 becomes the
equation
𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 + 𝑏1 𝑢1 + ⋯ + 𝑏𝑗 𝑢𝑗 = 0.
Because the list 𝑣1 , …, 𝑣𝑚 , 𝑢1 , …, 𝑢𝑗 is linearly independent, this equation implies
that all the 𝑎’s and 𝑏’s are 0, completing the proof.
For 𝑆 a finite set, let #𝑆 denote the number of elements of 𝑆. The table below
compares finite sets with finite-dimensional vector spaces, showing the analogy
between #𝑆 (for sets) and dim 𝑉 (for vector spaces), as well as the analogy between
unions of subsets (in the context of sets) and sums of subspaces (in the context of
vector spaces).
sets vector spaces
𝑆 is a finite set 𝑉 is a finite-dimensional vector space
#𝑆 dim 𝑉
for subsets 𝑆1 , 𝑆2 of 𝑆, the union 𝑆1 ∪ 𝑆2 for subspaces 𝑉1 , 𝑉2 of 𝑉, the sum 𝑉1 + 𝑉2
is the smallest subset of 𝑆 containing 𝑆1 is the smallest subspace of 𝑉 containing
and 𝑆2 𝑉1 and 𝑉2
#(𝑆1 ∪ 𝑆2 ) dim(𝑉1 + 𝑉2 )
= #𝑆1 + #𝑆2 − #(𝑆1 ∩ 𝑆2 ) = dim 𝑉1 + dim 𝑉2 − dim(𝑉1 ∩ 𝑉2 )
#(𝑆1 ∪ 𝑆2 ) = #𝑆1 + #𝑆2 dim(𝑉1 + 𝑉2 ) = dim 𝑉1 + dim 𝑉2
⟺ 𝑆1 ∩ 𝑆2 = ∅ ⟺ 𝑉1 ∩ 𝑉2 = {0}
𝑆1 ∪ ⋯ ∪ 𝑆𝑚 is a disjoint union ⟺ 𝑉1 + ⋯ + 𝑉𝑚 is a direct sum ⟺
#(𝑆1 ∪ ⋯ ∪ 𝑆𝑚 ) = #𝑆1 + ⋯ + #𝑆𝑚 dim(𝑉1 + ⋯ + 𝑉𝑚 )
= dim 𝑉1 + ⋯ + dim 𝑉𝑚
The last row above focuses on the analogy between disjoint unions (for sets)
and direct sums (for vector spaces). The proof of the result in the last box above
will be given in 3.94.
You should be able to find results about sets that correspond, via analogy, to
the results about vector spaces in Exercises 12 through 18.
Exercises 2C
1 Show that the subspaces of 𝐑2 are precisely {0}, all lines in 𝐑2 containing
the origin, and 𝐑2.
2 Show that the subspaces of 𝐑3 are precisely {0}, all lines in 𝐑3 containing
the origin, all planes in 𝐑3 containing the origin, and 𝐑3.
3 (a) Let 𝑈 = {𝑝 ∈ 𝒫4 (𝐅) ∶ 𝑝(6) = 0}. Find a basis of 𝑈.
(b) Extend the basis in (a) to a basis of 𝒫4 (𝐅).
(c) Find a subspace 𝑊 of 𝒫4 (𝐅) such that 𝒫4 (𝐅) = 𝑈 ⊕ 𝑊.
4 (a) Let 𝑈 = {𝑝 ∈ 𝒫4 (𝐑) ∶ 𝑝″(6) = 0}. Find a basis of 𝑈.
(b) Extend the basis in (a) to a basis of 𝒫4 (𝐑).
(c) Find a subspace 𝑊 of 𝒫4 (𝐑) such that 𝒫4 (𝐑) = 𝑈 ⊕ 𝑊.
5 (a) Let 𝑈 = {𝑝 ∈ 𝒫4 (𝐅) ∶ 𝑝(2) = 𝑝(5)}. Find a basis of 𝑈.
6 (a) Let 𝑈 = {𝑝 ∈ 𝒫4 (𝐅) ∶ 𝑝(2) = 𝑝(5) = 𝑝(6)}. Find a basis of 𝑈.

7 (a) Let 𝑈 = {𝑝 ∈ 𝒫4 (𝐑) ∶ ∫−1 1
𝑝 = 0}. Find a basis of 𝑈.
(b) Extend the basis in (a) to a basis of 𝒫4 (𝐑).
(c) Find a subspace 𝑊 of 𝒫4 (𝐑) such that 𝒫4 (𝐑) = 𝑈 ⊕ 𝑊.
8 Suppose 𝑣1 , …, 𝑣𝑚 is linearly independent in 𝑉 and 𝑤 ∈ 𝑉. Prove that
dim span(𝑣1 + 𝑤, …, 𝑣𝑚 + 𝑤) ≥ 𝑚 − 1.
9 Suppose 𝑚 is a positive integer and 𝑝0 , 𝑝1 , …, 𝑝𝑚 ∈ 𝒫(𝐅) are such that each
𝑝𝑘 has degree 𝑘. Prove that 𝑝0 , 𝑝1 , …, 𝑝𝑚 is a basis of 𝒫𝑚 (𝐅).
10 Suppose 𝑚 is a positive integer. For 0 ≤ 𝑘 ≤ 𝑚, let

𝑝𝑘 (𝑥) = 𝑥𝑘 (1 − 𝑥)𝑚 − 𝑘 .
Show that 𝑝0 , …, 𝑝𝑚 is a basis of 𝒫𝑚 (𝐅).
The basis in this exercise leads to what are called Bernstein polynomials.
You can do a web search to learn how Bernstein polynomials are used to
approximate continuous functions on [0, 1].
11 Suppose 𝑈 and 𝑊 are both four-dimensional subspaces of 𝐂6. Prove that

there exist two vectors in 𝑈 ∩ 𝑊 such that neither of these vectors is a scalar
multiple of the other.
12 Suppose that 𝑈 and 𝑊 are subspaces of 𝐑8 such that dim 𝑈 = 3, dim 𝑊 = 5,
and 𝑈 + 𝑊 = 𝐑8. Prove that 𝐑8 = 𝑈 ⊕ 𝑊.
13 Suppose 𝑈 and 𝑊 are both five-dimensional subspaces of 𝐑9. Prove that
𝑈 ∩ 𝑊 ≠ {0}.
14 Suppose 𝑉 is a ten-dimensional vector space and 𝑉1 , 𝑉2 , 𝑉3 are subspaces

of 𝑉 with dim 𝑉1 = dim 𝑉2 = dim 𝑉3 = 7. Prove that 𝑉1 ∩ 𝑉2 ∩ 𝑉3 ≠ {0}.
15 Suppose 𝑉 is finite-dimensional and 𝑉1 , 𝑉2 , 𝑉3 are subspaces of 𝑉 with
dim 𝑉1 + dim 𝑉2 + dim 𝑉3 > 2 dim 𝑉. Prove that 𝑉1 ∩ 𝑉2 ∩ 𝑉3 ≠ {0}.
16 Suppose 𝑉 is finite-dimensional and 𝑈 is a subspace of 𝑉 with 𝑈 ≠ 𝑉. Let
𝑛 = dim 𝑉 and 𝑚 = dim 𝑈. Prove that there exist 𝑛 − 𝑚 subspaces of 𝑉,
each of dimension 𝑛 − 1, whose intersection equals 𝑈.
17 Suppose that 𝑉1 , …, 𝑉𝑚 are finite-dimensional subspaces of 𝑉. Prove that
𝑉1 + ⋯ + 𝑉𝑚 is finite-dimensional and
dim(𝑉1 + ⋯ + 𝑉𝑚 ) ≤ dim 𝑉1 + ⋯ + dim 𝑉𝑚 .
The inequality above is an equality if and only if 𝑉1 + ⋯ + 𝑉𝑚 is a direct

sum, as will be shown in 3.94.
18 Suppose 𝑉 is finite-dimensional, with dim 𝑉 = 𝑛 ≥ 1. Prove that there exist

one-dimensional subspaces 𝑉1 , …, 𝑉𝑛 of 𝑉 such that
𝑉 = 𝑉1 ⊕ ⋯ ⊕ 𝑉𝑛 .
19 Explain why you might guess, motivated by analogy with the formula for
the number of elements in the union of three finite sets, that if 𝑉1 , 𝑉2 , 𝑉3 are
subspaces of a finite-dimensional vector space, then
dim(𝑉1 + 𝑉2 + 𝑉3 )
= dim 𝑉1 + dim 𝑉2 + dim 𝑉3
− dim(𝑉1 ∩ 𝑉2 ) − dim(𝑉1 ∩ 𝑉3 ) − dim(𝑉2 ∩ 𝑉3 )
+ dim(𝑉1 ∩ 𝑉2 ∩ 𝑉3 ).
Then either prove the formula above or give a counterexample.

20 Prove that if 𝑉1 , 𝑉2 , and 𝑉3 are subspaces of a finite-dimensional vector
space, then
dim(𝑉1 + 𝑉2 + 𝑉3 )
= dim 𝑉1 + dim 𝑉2 + dim 𝑉3
dim(𝑉1 ∩ 𝑉2 ) + dim(𝑉1 ∩ 𝑉3 ) + dim(𝑉2 ∩ 𝑉3 )

−
3
dim((𝑉1 + 𝑉2 )∩𝑉3 ) + dim((𝑉1 + 𝑉3 )∩𝑉2 ) + dim((𝑉2 + 𝑉3 )∩𝑉1 )
− .
3
The formula above may seem strange because the right side does not look
like an integer.
I at once gave up my former occupations, set down natural history and all its
progeny as a deformed and abortive creation, and entertained the greatest disdain
for a would-be science which could never even step within the threshold of real
knowledge. In this mood I betook myself to the mathematics and the branches of
study appertaining to that science as being built upon secure foundations, and so
worthy of my consideration.
—Frankenstein, Mary Wollstonecraft Shelley

Chapter 3
Linear Maps
So far our attention has focused on vector spaces. No one gets excited about
vector spaces. The interesting part of linear algebra is the subject to which we
now turn—linear maps.
We will frequently use the powerful fundamental theorem of linear maps,
which states that the dimension of the domain of a linear map equals the dimension
of the subspace that gets sent to 0 plus the dimension of the range. This will imply
the striking result that a linear map from a finite-dimensional vector space to itself
is one-to-one if and only if its range is the whole space.
A major concept that we will introduce in this chapter is the matrix associated
with a linear map and with a basis of the domain space and a basis of the target
space. This correspondence between linear maps and matrices provides much
insight into key aspects of linear algebra.
This chapter concludes by introducing product, quotient, and dual spaces.
In this chapter we will need additional vector spaces, which we call 𝑈 and 𝑊,
in addition to 𝑉. Thus our standing assumptions are now as follows.
• 𝑈, 𝑉, and 𝑊 denote vector spaces over 𝐅.
Stefan Schäfer CC BY-SA
The twelfth-century Dankwarderode Castle in Brunswick (Braunschweig), where Carl

Friedrich Gauss (1777–1855) was born and grew up. In 1809 Gauss published a method
for solving systems of linear equations. This method, now called Gaussian elimination,
was used in a Chinese book written over 1600 years earlier.
51
https://doi.org/10.1007/978-3-031-41026-0_3
52 Chapter 3 Linear Maps
3A Vector Space of Linear Maps

Definition and Examples of Linear Maps
Now we are ready for one of the key definitions in linear algebra.
3.1 definition: linear map
A linear map from 𝑉 to 𝑊 is a function 𝑇 ∶ 𝑉 → 𝑊 with the following

properties.
additivity
𝑇(𝑢 + 𝑣) = 𝑇𝑢 + 𝑇𝑣 for all 𝑢, 𝑣 ∈ 𝑉.
homogeneity
𝑇(𝜆𝑣) = 𝜆(𝑇𝑣) for all 𝜆 ∈ 𝐅 and all 𝑣 ∈ 𝑉.
Note that for linear maps we often Some mathematicians use the phrase
use the notation 𝑇𝑣 as well as the usual linear transformation, which means
function notation 𝑇(𝑣). the same as linear map.
3.2 notation: ℒ(𝑉, 𝑊), ℒ(𝑉)
• The set of linear maps from 𝑉 to 𝑊 is denoted by ℒ(𝑉, 𝑊).

• The set of linear maps from 𝑉 to 𝑉 is denoted by ℒ(𝑉). In other words,
ℒ(𝑉) = ℒ(𝑉, 𝑉).
Let’s look at some examples of linear maps. Make sure you verify that each
of the functions defined in the next example is indeed a linear map:
3.3 example: linear maps

zero
In addition to its other uses, we let the symbol 0 denote the linear map that takes
every element of some vector space to the additive identity of another (or possibly
the same) vector space. To be specific, 0 ∈ ℒ(𝑉, 𝑊) is defined by
0𝑣 = 0.
The 0 on the left side of the equation above is a function from 𝑉 to 𝑊, whereas
the 0 on the right side is the additive identity in 𝑊. As usual, the context should
allow you to distinguish between the many uses of the symbol 0.
identity operator
The identity operator, denoted by 𝐼, is the linear map on some vector space that
takes each element to itself. To be specific, 𝐼 ∈ ℒ(𝑉) is defined by
𝐼𝑣 = 𝑣.
Section 3A Vector Space of Linear Maps 53
differentiation
Define 𝐷 ∈ ℒ(𝒫(𝐑)) by
𝐷𝑝 = 𝑝′.
The assertion that this function is a linear map is another way of stating a basic
result about differentiation: ( 𝑓 + 𝑔)′ = 𝑓 ′ + 𝑔′ and (𝜆 𝑓 )′ = 𝜆 𝑓 ′ whenever 𝑓, 𝑔 are
differentiable and 𝜆 is a constant.
integration
Define 𝑇 ∈ ℒ(𝒫(𝐑), 𝐑) by
1
𝑇𝑝 = ∫ 𝑝.
0
The assertion that this function is linear is another way of stating a basic result
about integration: the integral of the sum of two functions equals the sum of the
integrals, and the integral of a constant times a function equals the constant times
the integral of the function.
multiplication by 𝑥2
Define a linear map 𝑇 ∈ ℒ(𝒫(𝐑)) by
(𝑇𝑝)(𝑥) = 𝑥2 𝑝(𝑥)
for each 𝑥 ∈ 𝐑 .
backward shift
Recall that 𝐅∞ denotes the vector space of all sequences of elements of 𝐅. Define
a linear map 𝑇 ∈ ℒ(𝐅∞ ) by
𝑇(𝑥1 , 𝑥2 , 𝑥3 , … ) = (𝑥2 , 𝑥3 , … ).
from 𝐑3 to 𝐑2
Define a linear map 𝑇 ∈ ℒ(𝐑3, 𝐑2 ) by
𝑇(𝑥, 𝑦, 𝑧) = (2𝑥 − 𝑦 + 3𝑧, 7𝑥 + 5𝑦 − 6𝑧).
from 𝐅𝑛 to 𝐅𝑚
To generalize the previous example, let 𝑚 and 𝑛 be positive integers, let 𝐴𝑗, 𝑘 ∈ 𝐅
for each 𝑗 = 1, …, 𝑚 and each 𝑘 = 1, …, 𝑛, and define a linear map 𝑇 ∈ ℒ(𝐅𝑛, 𝐅𝑚 )
by
𝑇(𝑥1 , …, 𝑥𝑛 ) = (𝐴1, 1 𝑥1 + ⋯ + 𝐴1, 𝑛 𝑥𝑛 , …, 𝐴𝑚, 1 𝑥1 + ⋯ + 𝐴𝑚, 𝑛 𝑥𝑛 ).
Actually every linear map from 𝐅𝑛 to 𝐅𝑚 is of this form.
composition
Fix a polynomial 𝑞 ∈ 𝒫(𝐑). Define a linear map 𝑇 ∈ ℒ(𝒫(𝐑)) by
(𝑇𝑝)(𝑥) = 𝑝(𝑞(𝑥)).
The existence part of the next result means that we can find a linear map that
takes on whatever values we wish on the vectors in a basis. The uniqueness part
of the next result means that a linear map is completely determined by its values
on a basis.
3.4 linear map lemma
Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝑤1 , …, 𝑤𝑛 ∈ 𝑊. Then there exists a

unique linear map 𝑇 ∶ 𝑉 → 𝑊 such that
𝑇𝑣𝑘 = 𝑤𝑘
for each 𝑘 = 1, …, 𝑛.
Proof First we show the existence of a linear map 𝑇 with the desired property.
Define 𝑇 ∶ 𝑉 → 𝑊 by
𝑇(𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 ) = 𝑐1 𝑤1 + ⋯ + 𝑐𝑛 𝑤𝑛 ,
where 𝑐1 , …, 𝑐𝑛 are arbitrary elements of 𝐅. The list 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Thus

the equation above does indeed define a function 𝑇 from 𝑉 to 𝑊 (because each
element of 𝑉 can be uniquely written in the form 𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 ).
For each 𝑘, taking 𝑐𝑘 = 1 and the other 𝑐’s equal to 0 in the equation above
shows that 𝑇𝑣𝑘 = 𝑤𝑘 .
If 𝑢, 𝑣 ∈ 𝑉 with 𝑢 = 𝑎1 𝑣1 + ⋯ + 𝑎𝑛 𝑣𝑛 and 𝑣 = 𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 , then
𝑇(𝑢 + 𝑣) = 𝑇((𝑎1 + 𝑐1 )𝑣1 + ⋯ + (𝑎𝑛 + 𝑐𝑛 )𝑣𝑛 )

= (𝑎1 + 𝑐1 )𝑤1 + ⋯ + (𝑎𝑛 + 𝑐𝑛 )𝑤𝑛
= (𝑎1 𝑤1 + ⋯ + 𝑎𝑛 𝑤𝑛 ) + (𝑐1 𝑤1 + ⋯ + 𝑐𝑛 𝑤𝑛 )
= 𝑇𝑢 + 𝑇𝑣.
Similarly, if 𝜆 ∈ 𝐅 and 𝑣 = 𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 , then
𝑇(𝜆𝑣) = 𝑇(𝜆𝑐1 𝑣1 + ⋯ + 𝜆𝑐𝑛 𝑣𝑛 )

= 𝜆𝑐1 𝑤1 + ⋯ + 𝜆𝑐𝑛 𝑤𝑛
= 𝜆(𝑐1 𝑤1 + ⋯ + 𝑐𝑛 𝑤𝑛 )
= 𝜆𝑇𝑣.
Thus 𝑇 is a linear map from 𝑉 to 𝑊.

To prove uniqueness, now suppose that 𝑇 ∈ ℒ(𝑉, 𝑊) and that 𝑇𝑣𝑘 = 𝑤𝑘 for
each 𝑘 = 1, …, 𝑛. Let 𝑐1 , …, 𝑐𝑛 ∈ 𝐅. Then the homogeneity of 𝑇 implies that
𝑇(𝑐𝑘 𝑣𝑘 ) = 𝑐𝑘 𝑤𝑘 for each 𝑘 = 1, …, 𝑛. The additivity of 𝑇 now implies that
𝑇(𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 ) = 𝑐1 𝑤1 + ⋯ + 𝑐𝑛 𝑤𝑛 .
Thus 𝑇 is uniquely determined on span(𝑣1 , …, 𝑣𝑛 ) by the equation above. Because

𝑣1 , …, 𝑣𝑛 is a basis of 𝑉, this implies that 𝑇 is uniquely determined on 𝑉, as
desired.
Algebraic Operations on ℒ(𝑉, 𝑊)

We begin by defining addition and scalar multiplication on ℒ(𝑉, 𝑊).
3.5 definition: addition and scalar multiplication on ℒ(𝑉, 𝑊)
Suppose 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝜆 ∈ 𝐅. The sum 𝑆 + 𝑇 and the product 𝜆𝑇 are

the linear maps from 𝑉 to 𝑊 defined by
(𝑆 + 𝑇)(𝑣) = 𝑆𝑣 + 𝑇𝑣 and (𝜆𝑇)(𝑣) = 𝜆(𝑇𝑣)
for all 𝑣 ∈ 𝑉.
You should verify that 𝑆 + 𝑇 and 𝜆𝑇 Linear maps are pervasive throughout
as defined above are indeed linear maps. mathematics. However, they are not as
In other words, if 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊) and ubiquitous as imagined by people who
𝜆 ∈ 𝐅, then 𝑆 + 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝜆𝑇 ∈ seem to think cos is a linear map from
ℒ(𝑉, 𝑊). 𝐑 to 𝐑 when they incorrectly write that
Because we took the trouble to de- cos(𝑥 + 𝑦) equals cos 𝑥 +cos 𝑦 and that
fine addition and scalar multiplication on cos 2𝑥 equals 2 cos 𝑥.
ℒ(𝑉, 𝑊), the next result should not be a
surprise.
3.6 ℒ(𝑉, 𝑊) is a vector space
With the operations of addition and scalar multiplication as defined above,

ℒ(𝑉, 𝑊) is a vector space.
The routine proof of the result above is left to the reader. Note that the additive
identity of ℒ(𝑉, 𝑊) is the zero linear map defined in Example 3.3.
Usually it makes no sense to multiply together two elements of a vector space,
but for some pairs of linear maps a useful product exists, as in the next definition.
3.7 definition: product of linear maps
If 𝑇 ∈ ℒ(𝑈, 𝑉) and 𝑆 ∈ ℒ(𝑉, 𝑊), then the product 𝑆𝑇 ∈ ℒ(𝑈, 𝑊) is defined

by
(𝑆𝑇)(𝑢) = 𝑆(𝑇𝑢)
for all 𝑢 ∈ 𝑈.
Thus 𝑆𝑇 is just the usual composition 𝑆 ∘ 𝑇 of two functions, but when both
functions are linear, we usually write 𝑆𝑇 instead of 𝑆 ∘ 𝑇. The product notation
𝑆𝑇 helps make the distributive properties (see next result) seem natural.
Note that 𝑆𝑇 is defined only when 𝑇 maps into the domain of 𝑆. You should
verify that 𝑆𝑇 is indeed a linear map from 𝑈 to 𝑊 whenever 𝑇 ∈ ℒ(𝑈, 𝑉) and
𝑆 ∈ ℒ(𝑉, 𝑊).
3.8 algebraic properties of products of linear maps
associativity
(𝑇1 𝑇2 )𝑇3 = 𝑇1 (𝑇2 𝑇3 ) whenever 𝑇1 , 𝑇2 , and 𝑇3 are linear maps such that
the products make sense (meaning 𝑇3 maps into the domain of 𝑇2 , and 𝑇2
maps into the domain of 𝑇1 ).
identity
𝑇𝐼 = 𝐼𝑇 = 𝑇 whenever 𝑇 ∈ ℒ(𝑉, 𝑊); here the first 𝐼 is the identity operator
on 𝑉, and the second 𝐼 is the identity operator on 𝑊.
distributive properties
(𝑆1 + 𝑆2 )𝑇 = 𝑆1 𝑇 + 𝑆2 𝑇 and 𝑆(𝑇1 + 𝑇2 ) = 𝑆𝑇1 + 𝑆𝑇2 whenever
𝑇, 𝑇1 , 𝑇2 ∈ ℒ(𝑈, 𝑉) and 𝑆, 𝑆1 , 𝑆2 ∈ ℒ(𝑉, 𝑊).
The routine proof of the result above is left to the reader.

Multiplication of linear maps is not commutative. In other words, it is not
necessarily true that 𝑆𝑇 = 𝑇𝑆, even if both sides of the equation make sense.
3.9 example: two noncommuting linear maps from 𝒫(𝐑) to 𝒫(𝐑)

Suppose 𝐷 ∈ ℒ(𝒫(𝐑)) is the differentiation map defined in Example 3.3
and 𝑇 ∈ ℒ(𝒫(𝐑)) is the multiplication by 𝑥2 map defined earlier in this section.
Then
((𝑇𝐷)𝑝)(𝑥) = 𝑥2 𝑝′(𝑥) but ((𝐷𝑇)𝑝)(𝑥) = 𝑥2 𝑝′(𝑥) + 2𝑥𝑝(𝑥).
Thus 𝑇𝐷 ≠ 𝐷𝑇—differentiating and then multiplying by 𝑥2 is not the same as

multiplying by 𝑥2 and then differentiating.
3.10 linear maps take 0 to 0
Suppose 𝑇 is a linear map from 𝑉 to 𝑊. Then 𝑇(0) = 0.
Proof By additivity, we have
𝑇(0) = 𝑇(0 + 0) = 𝑇(0) + 𝑇(0).
Add the additive inverse of 𝑇(0) to each side of the equation above to conclude
that 𝑇(0) = 0.
Suppose 𝑚, 𝑏 ∈ 𝐑 . The function 𝑓 ∶ 𝐑 → 𝐑 defined by
𝑓 (𝑥) = 𝑚𝑥 + 𝑏
is a linear map if and only if 𝑏 = 0 (use 3.10). Thus the linear functions of high
school algebra are not the same as linear maps in the context of linear algebra.
Exercises 3A
1 Suppose 𝑏, 𝑐 ∈ 𝐑 . Define 𝑇 ∶ 𝐑3 → 𝐑2 by
𝑇(𝑥, 𝑦, 𝑧) = (2𝑥 − 4𝑦 + 3𝑧 + 𝑏, 6𝑥 + 𝑐𝑥𝑦𝑧).
Show that 𝑇 is linear if and only if 𝑏 = 𝑐 = 0.
2 Suppose 𝑏, 𝑐 ∈ 𝐑 . Define 𝑇 ∶ 𝒫(𝐑) → 𝐑2 by
2
𝑇𝑝 = (3𝑝(4) + 5𝑝′(6) + 𝑏𝑝(1)𝑝(2), ∫ 𝑥3 𝑝(𝑥) 𝑑𝑥 + 𝑐 sin 𝑝(0)).
−1
Show that 𝑇 is linear if and only if 𝑏 = 𝑐 = 0.
3 Suppose that 𝑇 ∈ ℒ(𝐅𝑛, 𝐅𝑚 ). Show that there exist scalars 𝐴𝑗, 𝑘 ∈ 𝐅 for
𝑗 = 1, …, 𝑚 and 𝑘 = 1, …, 𝑛 such that
𝑇(𝑥1 , …, 𝑥𝑛 ) = (𝐴1, 1 𝑥1 + ⋯ + 𝐴1, 𝑛 𝑥𝑛 , …, 𝐴𝑚, 1 𝑥1 + ⋯ + 𝐴𝑚, 𝑛 𝑥𝑛 )
for every (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛.
This exercise shows that the linear map 𝑇 has the form promised in the
second to last item of Example 3.3.
4 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑣1 , …, 𝑣𝑚 is a list of vectors in 𝑉 such that

𝑇𝑣1 , …, 𝑇𝑣𝑚 is a linearly independent list in 𝑊. Prove that 𝑣1 , …, 𝑣𝑚 is
linearly independent.
5 Prove that ℒ(𝑉, 𝑊) is a vector space, as was asserted in 3.6.
6 Prove that multiplication of linear maps has the associative, identity, and
distributive properties asserted in 3.8.
7 Show that every linear map from a one-dimensional vector space to itself is
multiplication by some scalar. More precisely, prove that if dim 𝑉 = 1 and
𝑇 ∈ ℒ(𝑉), then there exists 𝜆 ∈ 𝐅 such that 𝑇𝑣 = 𝜆𝑣 for all 𝑣 ∈ 𝑉.
8 Give an example of a function 𝜑 ∶ 𝐑2 → 𝐑 such that
𝜑(𝑎𝑣) = 𝑎𝜑(𝑣)
for all 𝑎 ∈ 𝐑 and all 𝑣 ∈ 𝐑2 but 𝜑 is not linear.
This exercise and the next exercise show that neither homogeneity nor
additivity alone is enough to imply that a function is a linear map.
9 Give an example of a function 𝜑 ∶ 𝐂 → 𝐂 such that

𝜑(𝑤 + 𝑧) = 𝜑(𝑤) + 𝜑(𝑧)
for all 𝑤, 𝑧 ∈ 𝐂 but 𝜑 is not linear. (Here 𝐂 is thought of as a complex vector
space.)
There also exists a function 𝜑 ∶ 𝐑 → 𝐑 such that 𝜑 satisfies the additivity
condition above but 𝜑 is not linear. However, showing the existence of such
a function involves considerably more advanced tools.
10 Prove or give a counterexample: If 𝑞 ∈ 𝒫(𝐑) and 𝑇 ∶ 𝒫(𝐑) → 𝒫(𝐑) is

defined by 𝑇𝑝 = 𝑞 ∘ 𝑝, then 𝑇 is a linear map.
The function 𝑇 defined here differs from the function 𝑇 defined in the last
bullet point of 3.3 by the order of the functions in the compositions.
11 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is a scalar

multiple of the identity if and only if 𝑆𝑇 = 𝑇𝑆 for every 𝑆 ∈ ℒ(𝑉).
12 Suppose 𝑈 is a subspace of 𝑉 with 𝑈 ≠ 𝑉. Suppose 𝑆 ∈ ℒ(𝑈, 𝑊) and
𝑆 ≠ 0 (which means that 𝑆𝑢 ≠ 0 for some 𝑢 ∈ 𝑈). Define 𝑇 ∶ 𝑉 → 𝑊 by
⎧
{𝑆𝑣 if 𝑣 ∈ 𝑈,
𝑇𝑣 = ⎨
{0
⎩ if 𝑣 ∈ 𝑉 and 𝑣 ∉ 𝑈.
Prove that 𝑇 is not a linear map on 𝑉.

13 Suppose 𝑉 is finite-dimensional. Prove that every linear map on a subspace
of 𝑉 can be extended to a linear map on 𝑉. In other words, show that if 𝑈 is
a subspace of 𝑉 and 𝑆 ∈ ℒ(𝑈, 𝑊), then there exists 𝑇 ∈ ℒ(𝑉, 𝑊) such that
𝑇𝑢 = 𝑆𝑢 for all 𝑢 ∈ 𝑈.
The result in this exercise is used in the proof of 3.125.
14 Suppose 𝑉 is finite-dimensional with dim 𝑉 > 0, and suppose 𝑊 is infinite-

dimensional. Prove that ℒ(𝑉, 𝑊) is infinite-dimensional.
15 Suppose 𝑣1 , …, 𝑣𝑚 is a linearly dependent list of vectors in 𝑉. Suppose
also that 𝑊 ≠ {0}. Prove that there exist 𝑤1 , …, 𝑤𝑚 ∈ 𝑊 such that no
𝑇 ∈ ℒ(𝑉, 𝑊) satisfies 𝑇𝑣𝑘 = 𝑤𝑘 for each 𝑘 = 1, …, 𝑚.
16 Suppose 𝑉 is finite-dimensional with dim 𝑉 > 1. Prove that there exist

𝑆, 𝑇 ∈ ℒ(𝑉) such that 𝑆𝑇 ≠ 𝑇𝑆.
17 Suppose 𝑉 is finite-dimensional. Show that the only two-sided ideals of

ℒ(𝑉) are {0} and ℒ(𝑉).
A subspace ℰ of ℒ(𝑉) is called a two-sided ideal of ℒ(𝑉) if 𝑇𝐸 ∈ ℰ and
𝐸𝑇 ∈ ℰ for all 𝐸 ∈ ℰ and all 𝑇 ∈ ℒ(𝑉).
Section 3B Null Spaces and Ranges 59
3B Null Spaces and Ranges

Null Space and Injectivity
In this section we will learn about two subspaces that are intimately connected
with each linear map. We begin with the set of vectors that get mapped to 0.
3.11 definition: null space, null 𝑇
For 𝑇 ∈ ℒ(𝑉, 𝑊), the null space of 𝑇, denoted by null 𝑇, is the subset of 𝑉
consisting of those vectors that 𝑇 maps to 0:
null 𝑇 = {𝑣 ∈ 𝑉 ∶ 𝑇𝑣 = 0}.
3.12 example: null space
• If 𝑇 is the zero map from 𝑉 to 𝑊, meaning that 𝑇𝑣 = 0 for every 𝑣 ∈ 𝑉, then

null 𝑇 = 𝑉.
• Suppose 𝜑 ∈ ℒ(𝐂3, 𝐂) is defined by 𝜑(𝑧1 , 𝑧2 , 𝑧3 ) = 𝑧1 + 2𝑧2 + 3𝑧3 . Then
null 𝜑 equals {(𝑧1 , 𝑧2 , 𝑧3 ) ∈ 𝐂3 ∶ 𝑧1 + 2𝑧2 + 3𝑧3 = 0}, which is a subspace of
the domain of 𝜑. We will soon see that the null space of each linear map is a
subspace of its domain.
• Suppose 𝐷 ∈ ℒ(𝒫(𝐑)) is the dif- The word “null” means zero. Thus the
ferentiation map defined by 𝐷𝑝 = 𝑝′. term “null space”should remind you
The only functions whose derivative of the connection to 0. Some mathe-
equals the zero function are the con- maticians use the term kernel instead
stant functions. Thus the null space of of null space.
𝐷 equals the set of constant functions.
• Suppose that 𝑇 ∈ ℒ(𝒫(𝐑)) is the multiplication by 𝑥2 map defined by
(𝑇𝑝)(𝑥) = 𝑥2 𝑝(𝑥). The only polynomial 𝑝 such that 𝑥2 𝑝(𝑥) = 0 for all 𝑥 ∈ 𝐑
is the 0 polynomial. Thus null 𝑇 = {0}.
• Suppose 𝑇 ∈ ℒ(𝐅∞ ) is the backward shift defined by
𝑇(𝑥1 , 𝑥2 , 𝑥3 , … ) = (𝑥2 , 𝑥3 , … ).
Then 𝑇(𝑥1 , 𝑥2 , 𝑥3 , … ) equals 0 if and only if the numbers 𝑥2 , 𝑥3 , … are all 0.

Thus null 𝑇 = {(𝑎, 0, 0, … ) ∶ 𝑎 ∈ 𝐅}.
The next result shows that the null space of each linear map is a subspace of
the domain. In particular, 0 is in the null space of every linear map.
3.13 the null space is a subspace
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Then null 𝑇 is a subspace of 𝑉.

Proof Because 𝑇 is a linear map, 𝑇(0) = 0 (by 3.10). Thus 0 ∈ null 𝑇.

Suppose 𝑢, 𝑣 ∈ null 𝑇. Then
𝑇(𝑢 + 𝑣) = 𝑇𝑢 + 𝑇𝑣 = 0 + 0 = 0.
Hence 𝑢 + 𝑣 ∈ null 𝑇. Thus null 𝑇 is closed under addition.
Suppose 𝑢 ∈ null 𝑇 and 𝜆 ∈ 𝐅. Then
𝑇(𝜆𝑢) = 𝜆𝑇𝑢 = 𝜆0 = 0.
Hence 𝜆𝑢 ∈ null 𝑇. Thus null 𝑇 is closed under scalar multiplication.
We have shown that null 𝑇 contains 0 and is closed under addition and scalar
multiplication. Thus null 𝑇 is a subspace of 𝑉 (by 1.34).
As we will soon see, for a linear map the next definition is closely connected
to the null space.
3.14 definition: injective
A function 𝑇 ∶ 𝑉 → 𝑊 is called injective if 𝑇𝑢 = 𝑇𝑣 implies 𝑢 = 𝑣.
We could rephrase the definition The term one-to-one means the same
above to say that 𝑇 is injective if 𝑢 ≠ 𝑣 as injective.
implies that 𝑇𝑢 ≠ 𝑇𝑣. Thus 𝑇 is injective
if and only if it maps distinct inputs to distinct outputs.
The next result says that we can check whether a linear map is injective
by checking whether 0 is the only vector that gets mapped to 0. As a simple
application of this result, we see that of the linear maps whose null spaces we
computed in 3.12, only multiplication by 𝑥2 is injective (except that the zero map
is injective in the special case 𝑉 = {0}).
3.15 injectivity ⟺ null space equals {0}
Let 𝑇 ∈ ℒ(𝑉, 𝑊). Then 𝑇 is injective if and only if null 𝑇 = {0}.
Proof First suppose 𝑇 is injective. We want to prove that null 𝑇 = {0}. We

already know that {0} ⊆ null 𝑇 (by 3.10). To prove the inclusion in the other
direction, suppose 𝑣 ∈ null 𝑇. Then
𝑇(𝑣) = 0 = 𝑇(0).
Because 𝑇 is injective, the equation above implies that 𝑣 = 0. Thus we can
conclude that null 𝑇 = {0}, as desired.
To prove the implication in the other direction, now suppose null 𝑇 = {0}. We
want to prove that 𝑇 is injective. To do this, suppose 𝑢, 𝑣 ∈ 𝑉 and 𝑇𝑢 = 𝑇𝑣. Then
0 = 𝑇𝑢 − 𝑇𝑣 = 𝑇(𝑢 − 𝑣).
Thus 𝑢 − 𝑣 is in null 𝑇, which equals {0}. Hence 𝑢 − 𝑣 = 0, which implies that
𝑢 = 𝑣. Hence 𝑇 is injective, as desired.
Range and Surjectivity

Now we give a name to the set of outputs of a linear map.
3.16 definition: range
For 𝑇 ∈ ℒ(𝑉, 𝑊), the range of 𝑇 is the subset of 𝑊 consisting of those vectors
that are equal to 𝑇𝑣 for some 𝑣 ∈ 𝑉:
range 𝑇 = {𝑇𝑣 ∶ 𝑣 ∈ 𝑉}.
3.17 example: range
• If 𝑇 is the zero map from 𝑉 to 𝑊, meaning that 𝑇𝑣 = 0 for every 𝑣 ∈ 𝑉, then

range 𝑇 = {0}.
• Suppose 𝑇 ∈ ℒ(𝐑2, 𝐑3 ) is defined by 𝑇(𝑥, 𝑦) = (2𝑥, 5𝑦, 𝑥 + 𝑦). Then
range 𝑇 = {(2𝑥, 5𝑦, 𝑥 + 𝑦) ∶ 𝑥, 𝑦 ∈ 𝐑}.

Note that range 𝑇 is a subspace of 𝐑3 . We will soon see that the range of each
element of ℒ(𝑉, 𝑊) is a subspace of 𝑊.
• Suppose 𝐷 ∈ ℒ(𝒫(𝐑)) is the differentiation map defined by 𝐷𝑝 = 𝑝′. Because
for every polynomial 𝑞 ∈ 𝒫(𝐑) there exists a polynomial 𝑝 ∈ 𝒫(𝐑) such that
𝑝′ = 𝑞, the range of 𝐷 is 𝒫(𝐑).
The next result shows that the range of each linear map is a subspace of the
vector space into which it is being mapped.
3.18 the range is a subspace
If 𝑇 ∈ ℒ(𝑉, 𝑊), then range 𝑇 is a subspace of 𝑊.
Proof Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Then 𝑇(0) = 0 (by 3.10), which implies that
0 ∈ range 𝑇.
If 𝑤1 , 𝑤2 ∈ range 𝑇, then there exist 𝑣1 , 𝑣2 ∈ 𝑉 such that 𝑇𝑣1 = 𝑤1 and
𝑇𝑣2 = 𝑤2 . Thus
𝑇(𝑣1 + 𝑣2 ) = 𝑇𝑣1 + 𝑇𝑣2 = 𝑤1 + 𝑤2 .
Hence 𝑤1 + 𝑤2 ∈ range 𝑇. Thus range 𝑇 is closed under addition.
If 𝑤 ∈ range 𝑇 and 𝜆 ∈ 𝐅, then there exists 𝑣 ∈ 𝑉 such that 𝑇𝑣 = 𝑤. Thus
𝑇(𝜆𝑣) = 𝜆𝑇𝑣 = 𝜆𝑤.
Hence 𝜆𝑤 ∈ range 𝑇. Thus range 𝑇 is closed under scalar multiplication.

We have shown that range 𝑇 contains 0 and is closed under addition and scalar
multiplication. Thus range 𝑇 is a subspace of 𝑊 (by 1.34).
3.19 definition: surjective
A function 𝑇 ∶ 𝑉 → 𝑊 is called surjective if its range equals 𝑊.
To illustrate the definition above, note that of the ranges we computed in 3.17,
only the differentiation map is surjective (except that the zero map is surjective in
the special case 𝑊 = {0}).
Whether a linear map is surjective de- Some people use the term onto, which
pends on what we are thinking of as the means the same as surjective.
vector space into which it maps.
3.20 example: surjectivity depends on the target space

The differentiation map 𝐷 ∈ ℒ(𝒫5 (𝐑)) defined by 𝐷𝑝 = 𝑝′ is not surjective,
because the polynomial 𝑥5 is not in the range of 𝐷. However, the differentiation
map 𝑆 ∈ ℒ(𝒫5 (𝐑), 𝒫4 (𝐑)) defined by 𝑆𝑝 = 𝑝′ is surjective, because its range
equals 𝒫4 (𝐑), which is the vector space into which 𝑆 maps.
Fundamental Theorem of Linear Maps

The next result is so important that it gets a dramatic name.
3.21 fundamental theorem of linear maps
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Then range 𝑇 is finite-

dimensional and
dim 𝑉 = dim null 𝑇 + dim range 𝑇.
Proof Let 𝑢1 , …, 𝑢𝑚 be a basis of null 𝑇; thus dim null 𝑇 = 𝑚. The linearly

independent list 𝑢1 , …, 𝑢𝑚 can be extended to a basis
𝑢1 , …, 𝑢𝑚 , 𝑣1 , …, 𝑣𝑛
of 𝑉 (by 2.32). Thus dim 𝑉 = 𝑚 + 𝑛. To complete the proof, we need to show that
range 𝑇 is finite-dimensional and dim range 𝑇 = 𝑛. We will do this by proving
that 𝑇𝑣1 , …, 𝑇𝑣𝑛 is a basis of range 𝑇.
Let 𝑣 ∈ 𝑉. Because 𝑢1 , …, 𝑢𝑚 , 𝑣1 , …, 𝑣𝑛 spans 𝑉, we can write
𝑣 = 𝑎1 𝑢1 + ⋯ + 𝑎𝑚 𝑢𝑚 + 𝑏1 𝑣1 + ⋯ + 𝑏𝑛 𝑣𝑛 ,
where the 𝑎’s and 𝑏’s are in 𝐅. Applying 𝑇 to both sides of this equation, we get
𝑇𝑣 = 𝑏1 𝑇𝑣1 + ⋯ + 𝑏𝑛 𝑇𝑣𝑛 ,
where the terms of the form 𝑇𝑢𝑘 disappeared because each 𝑢𝑘 is in null 𝑇. The
last equation implies that the list 𝑇𝑣1 , …, 𝑇𝑣𝑛 spans range 𝑇. In particular, range 𝑇
is finite-dimensional.
To show 𝑇𝑣1 , …, 𝑇𝑣𝑛 is linearly independent, suppose 𝑐1 , …, 𝑐𝑛 ∈ 𝐅 and
𝑐1 𝑇𝑣1 + ⋯ + 𝑐𝑛 𝑇𝑣𝑛 = 0.
Then
𝑇(𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 ) = 0.
Hence
𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 ∈ null 𝑇.
Because 𝑢1 , …, 𝑢𝑚 spans null 𝑇, we can write
𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 = 𝑑1 𝑢1 + ⋯ + 𝑑𝑚 𝑢𝑚 ,
where the 𝑑’s are in 𝐅. This equation implies that all the 𝑐’s (and 𝑑’s) are 0
(because 𝑢1 , …, 𝑢𝑚 , 𝑣1 , …, 𝑣𝑛 is linearly independent). Thus 𝑇𝑣1 , …, 𝑇𝑣𝑛 is linearly
independent and hence is a basis of range 𝑇, as desired.
Now we can show that no linear map from a finite-dimensional vector space
to a “smaller” vector space can be injective, where “smaller” is measured by
dimension.
3.22 linear map to a lower-dimensional space is not injective
Suppose 𝑉 and 𝑊 are finite-dimensional vector spaces such that

dim 𝑉 > dim 𝑊. Then no linear map from 𝑉 to 𝑊 is injective.
Proof Let 𝑇 ∈ ℒ(𝑉, 𝑊). Then
dim null 𝑇 = dim 𝑉 − dim range 𝑇

≥ dim 𝑉 − dim 𝑊
> 0,
where the first line above comes from the fundamental theorem of linear maps
(3.21) and the second line follows from 2.37. The inequality above states that
dim null 𝑇 > 0. This means that null 𝑇 contains vectors other than 0. Thus 𝑇 is
not injective (by 3.15).
3.23 example: linear map from 𝐅4 to 𝐅3 is not injective

Define a linear map 𝑇 ∶ 𝐅4 → 𝐅3 by
𝑇(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 ) = (√7𝑧1 + 𝜋𝑧2 + 𝑧4 , 97𝑧1 + 3𝑧2 + 2𝑧3 , 𝑧2 + 6𝑧3 + 7𝑧4 ).
Because dim 𝐅4 > dim 𝐅3, we can use 3.22 to assert that 𝑇 is not injective, without
doing any calculations.
The next result shows that no linear map from a finite-dimensional vector
space to a “bigger” vector space can be surjective, where “bigger” is measured by
dimension.
3.24 linear map to a higher-dimensional space is not surjective
Suppose 𝑉 and 𝑊 are finite-dimensional vector spaces such that

dim 𝑉 < dim 𝑊. Then no linear map from 𝑉 to 𝑊 is surjective.
Proof Let 𝑇 ∈ ℒ(𝑉, 𝑊). Then

dim range 𝑇 = dim 𝑉 − dim null 𝑇
≤ dim 𝑉
< dim 𝑊,
where the equality above comes from the fundamental theorem of linear maps
(3.21). The inequality above states that dim range 𝑇 < dim 𝑊. This means that
range 𝑇 cannot equal 𝑊. Thus 𝑇 is not surjective.
As we will soon see, 3.22 and 3.24 have important consequences in the theory
of linear equations. The idea is to express questions about systems of linear
equations in terms of linear maps. Let’s begin by rephrasing in terms of linear
maps the question of whether a homogeneous system of linear equations has a
nonzero solution.
Fix positive integers 𝑚 and 𝑛, and let Homogeneous, in this context, means
𝐴𝑗, 𝑘 ∈ 𝐅 for 𝑗 = 1, …, 𝑚 and 𝑘 = 1, …, 𝑛. that the constant term on the right side
Consider the homogeneous system of lin- of each equation below is 0.
ear equations
𝑛
∑ 𝐴1, 𝑘 𝑥𝑘 = 0
𝑘=1
⋮
𝑛
∑ 𝐴𝑚, 𝑘 𝑥𝑘 = 0.
𝑘=1
Clearly 𝑥1 = ⋯ = 𝑥𝑛 = 0 is a solution of the system of equations above; the
question here is whether any other solutions exist.
Define 𝑇 ∶ 𝐅𝑛 → 𝐅𝑚 by
𝑛 𝑛
3.25 𝑇(𝑥1 , …, 𝑥𝑛 ) = ( ∑ 𝐴1, 𝑘 𝑥𝑘 , …, ∑ 𝐴𝑚, 𝑘 𝑥𝑘 ).
𝑘=1 𝑘=1
The equation 𝑇(𝑥1 , …, 𝑥𝑛 ) = 0 (the 0 here is the additive identity in 𝐅𝑚, namely,
the list of length 𝑚 of all 0’s) is the same as the homogeneous system of linear
equations above.
Thus we want to know if null 𝑇 is strictly bigger than {0}, which is equivalent
to 𝑇 not being injective (by 3.15). The next result gives an important condition
for ensuring that 𝑇 is not injective.
3.26 homogeneous system of linear equations
A homogeneous system of linear equations with more variables than equations

has nonzero solutions.
Proof Use the notation and result from the discussion above. Thus 𝑇 is a linear
map from 𝐅𝑛 to 𝐅𝑚, and we have a homogeneous system of 𝑚 linear equations
with 𝑛 variables 𝑥1 , …, 𝑥𝑛 . From 3.22 we see that 𝑇 is not injective if 𝑛 > 𝑚.
Example of the result above: a homogeneous system of four linear equations

with five variables has nonzero solutions.
Now we consider the question of Inhomogeneous, as used in this con-
whether an inhomogeneous system of lin- text, means that the constant term on
ear equations has no solutions for some the right side of at least one equation
choice of the constant terms. To rephrase below does not equal 0.
this question in terms of a linear map, fix
positive integers 𝑚 and 𝑛, and let 𝐴𝑗, 𝑘 ∈ 𝐅 for all 𝑗 = 1, …, 𝑚 and all 𝑘 = 1, …, 𝑛.
For 𝑐1 , …, 𝑐𝑚 ∈ 𝐅, consider the system of linear equations
𝑛
∑ 𝐴1, 𝑘 𝑥𝑘 = 𝑐1
𝑘=1
3.27 ⋮
𝑛
∑ 𝐴𝑚, 𝑘 𝑥𝑘 = 𝑐𝑚 .
𝑘=1
The question here is whether there is some choice of 𝑐1 , …, 𝑐𝑚 ∈ 𝐅 such that no

solution exists to the system above.
Define 𝑇 ∶ 𝐅𝑛 → 𝐅𝑚 as in 3.25. The The results 3.26 and 3.28, which com-
equation 𝑇(𝑥1 , …, 𝑥𝑛 ) = (𝑐1 , …, 𝑐𝑚 ) is the pare the number of variables and
same as the system of equations 3.27. the number of equations, can also
Thus we want to know if range 𝑇 ≠ 𝐅𝑚. be proved using Gaussian elimina-
Hence we can rephrase our question tion. The abstract approach taken here
about not having a solution for some seems to provide cleaner proofs.
choice of 𝑐1 , …, 𝑐𝑚 ∈ 𝐅 as follows: What
condition ensures that 𝑇 is not surjective? The next result gives one such condition.
3.28 inhomogeneous system of linear equations
An inhomogeneous system of linear equations with more equations than

variables has no solution for some choice of the constant terms.
Proof Use the notation and result from the example above. Thus 𝑇 is a linear
map from 𝐅𝑛 to 𝐅𝑚, and we have a system of 𝑚 equations with 𝑛 variables 𝑥1 , …, 𝑥𝑛 .
From 3.24 we see that 𝑇 is not surjective if 𝑛 < 𝑚.
Example of the result above: an inhomogeneous system of five linear equations

with four variables has no solution for some choice of the constant terms.
Exercises 3B
1 Give an example of a linear map 𝑇 with dim null 𝑇 = 3 and dim range 𝑇 = 2.
2 Suppose 𝑆, 𝑇 ∈ ℒ(𝑉) are such that range 𝑆 ⊆ null 𝑇. Prove that (𝑆𝑇)2 = 0.
3 Suppose 𝑣1 , …, 𝑣𝑚 is a list of vectors in 𝑉. Define 𝑇 ∈ ℒ(𝐅𝑚, 𝑉) by
𝑇(𝑧1 , …, 𝑧𝑚 ) = 𝑧1 𝑣1 + ⋯ + 𝑧𝑚 𝑣𝑚 .
(a) What property of 𝑇 corresponds to 𝑣1 , …, 𝑣𝑚 spanning 𝑉?

(b) What property of 𝑇 corresponds to the list 𝑣1 , …, 𝑣𝑚 being linearly
independent?
4 Show that {𝑇 ∈ ℒ(𝐑5, 𝐑4 ) ∶ dim null 𝑇 > 2} is not a subspace of ℒ(𝐑5, 𝐑4 ).

5 Give an example of 𝑇 ∈ ℒ(𝐑4 ) such that range 𝑇 = null 𝑇.
6 Prove that there does not exist 𝑇 ∈ ℒ(𝐑5 ) such that range 𝑇 = null 𝑇.
7 Suppose 𝑉 and 𝑊 are finite-dimensional with 2 ≤ dim 𝑉 ≤ dim 𝑊. Show
that {𝑇 ∈ ℒ(𝑉, 𝑊) ∶ 𝑇 is not injective} is not a subspace of ℒ(𝑉, 𝑊).
8 Suppose 𝑉 and 𝑊 are finite-dimensional with dim 𝑉 ≥ dim 𝑊 ≥ 2. Show
that {𝑇 ∈ ℒ(𝑉, 𝑊) ∶ 𝑇 is not surjective} is not a subspace of ℒ(𝑉, 𝑊).
9 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) is injective and 𝑣1 , …, 𝑣𝑛 is linearly independent in 𝑉.
Prove that 𝑇𝑣1 , …, 𝑇𝑣𝑛 is linearly independent in 𝑊.
10 Suppose 𝑣1 , …, 𝑣𝑛 spans 𝑉 and 𝑇 ∈ ℒ(𝑉, 𝑊). Show that 𝑇𝑣1 , …, 𝑇𝑣𝑛 spans
range 𝑇.
11 Suppose that 𝑉 is finite-dimensional and that 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that there
exists a subspace 𝑈 of 𝑉 such that
𝑈 ∩ null 𝑇 = {0} and range 𝑇 = {𝑇𝑢 ∶ 𝑢 ∈ 𝑈}.
12 Suppose 𝑇 is a linear map from 𝐅4 to 𝐅2 such that

null 𝑇 = {(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) ∈ 𝐅4 ∶ 𝑥1 = 5𝑥2 and 𝑥3 = 7𝑥4 }.
Prove that 𝑇 is surjective.
13 Suppose 𝑈 is a three-dimensional subspace of 𝐑8 and that 𝑇 is a linear map
from 𝐑8 to 𝐑5 such that null 𝑇 = 𝑈. Prove that 𝑇 is surjective.
14 Prove that there does not exist a linear map from 𝐅5 to 𝐅2 whose null space
equals {(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 , 𝑥5 ) ∈ 𝐅5 ∶ 𝑥1 = 3𝑥2 and 𝑥3 = 𝑥4 = 𝑥5 }.
15 Suppose there exists a linear map on 𝑉 whose null space and range are both
finite-dimensional. Prove that 𝑉 is finite-dimensional.
16 Suppose 𝑉 and 𝑊 are both finite-dimensional. Prove that there exists an

injective linear map from 𝑉 to 𝑊 if and only if dim 𝑉 ≤ dim 𝑊.
17 Suppose 𝑉 and 𝑊 are both finite-dimensional. Prove that there exists a
surjective linear map from 𝑉 onto 𝑊 if and only if dim 𝑉 ≥ dim 𝑊.
18 Suppose 𝑉 and 𝑊 are finite-dimensional and that 𝑈 is a subspace of 𝑉.
Prove that there exists 𝑇 ∈ ℒ(𝑉, 𝑊) such that null 𝑇 = 𝑈 if and only if
dim 𝑈 ≥ dim 𝑉 − dim 𝑊.
19 Suppose 𝑊 is finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that 𝑇 is injective
if and only if there exists 𝑆 ∈ ℒ(𝑊, 𝑉) such that 𝑆𝑇 is the identity operator
on 𝑉.
20 Suppose 𝑊 is finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that 𝑇 is surjective
if and only if there exists 𝑆 ∈ ℒ(𝑊, 𝑉) such that 𝑇𝑆 is the identity operator
on 𝑊.
21 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉, 𝑊), and 𝑈 is a subspace of 𝑊.
Prove that {𝑣 ∈ 𝑉 ∶ 𝑇𝑣 ∈ 𝑈} is a subspace of 𝑉 and
dim{𝑣 ∈ 𝑉 ∶ 𝑇𝑣 ∈ 𝑈} = dim null 𝑇 + dim(𝑈 ∩ range 𝑇).
22 Suppose 𝑈 and 𝑉 are finite-dimensional vector spaces and 𝑆 ∈ ℒ(𝑉, 𝑊) and

𝑇 ∈ ℒ(𝑈, 𝑉). Prove that
dim null 𝑆𝑇 ≤ dim null 𝑆 + dim null 𝑇.
23 Suppose 𝑈 and 𝑉 are finite-dimensional vector spaces and 𝑆 ∈ ℒ(𝑉, 𝑊) and

𝑇 ∈ ℒ(𝑈, 𝑉). Prove that
dim range 𝑆𝑇 ≤ min{dim range 𝑆, dim range 𝑇}.
24 (a) Suppose dim 𝑉 = 5 and 𝑆, 𝑇 ∈ ℒ(𝑉) are such that 𝑆𝑇 = 0. Prove that
dim range 𝑇𝑆 ≤ 2.
(b) Give an example of 𝑆, 𝑇 ∈ ℒ(𝐅5 ) with 𝑆𝑇 = 0 and dim range 𝑇𝑆 = 2.
25 Suppose that 𝑊 is finite-dimensional and 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that
null 𝑆 ⊆ null 𝑇 if and only if there exists 𝐸 ∈ ℒ(𝑊) such that 𝑇 = 𝐸𝑆.
26 Suppose that 𝑉 is finite-dimensional and 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that
range 𝑆 ⊆ range 𝑇 if and only if there exists 𝐸 ∈ ℒ(𝑉) such that 𝑆 = 𝑇𝐸.
27 Suppose 𝑃 ∈ ℒ(𝑉) and 𝑃2 = 𝑃. Prove that 𝑉 = null 𝑃 ⊕ range 𝑃.
28 Suppose 𝐷 ∈ ℒ(𝒫(𝐑)) is such that deg 𝐷𝑝 = (deg 𝑝) − 1 for every non-
constant polynomial 𝑝 ∈ 𝒫(𝐑). Prove that 𝐷 is surjective.
The notation 𝐷 is used above to remind you of the differentiation map that
sends a polynomial 𝑝 to 𝑝′.
29 Suppose 𝑝 ∈ 𝒫(𝐑). Prove that there exists a polynomial 𝑞 ∈ 𝒫(𝐑) such

that 5𝑞″ + 3𝑞′ = 𝑝.
This exercise can be done without linear algebra, but it’s more fun to do it
using linear algebra.
30 Suppose 𝜑 ∈ ℒ(𝑉, 𝐅) and 𝜑 ≠ 0. Suppose 𝑢 ∈ 𝑉 is not in null 𝜑. Prove

that
𝑉 = null 𝜑 ⊕ {𝑎𝑢 ∶ 𝑎 ∈ 𝐅}.
31 Suppose 𝑉 is finite-dimensional, 𝑋 is a subspace of 𝑉, and 𝑌 is a finite-
dimensional subspace of 𝑊. Prove that there exists 𝑇 ∈ ℒ(𝑉, 𝑊) such that
null 𝑇 = 𝑋 and range 𝑇 = 𝑌 if and only if dim 𝑋 + dim 𝑌 = dim 𝑉.
32 Suppose 𝑉 is finite-dimensional with dim 𝑉 > 1. Show that if 𝜑 ∶ ℒ(𝑉) → 𝐅
is a linear map such that 𝜑(𝑆𝑇) = 𝜑(𝑆)𝜑(𝑇) for all 𝑆, 𝑇 ∈ ℒ(𝑉), then
𝜑 = 0.
Hint: The description of the two-sided ideals of ℒ(𝑉) given by Exercise 17
in Section 3A might be useful.
33 Suppose that 𝑉 and 𝑊 are real vector spaces and 𝑇 ∈ ℒ(𝑉, 𝑊). Define
𝑇𝐂 ∶ 𝑉𝐂 → 𝑊𝐂 by
𝑇𝐂 (𝑢 + 𝑖𝑣) = 𝑇𝑢 + 𝑖𝑇𝑣
for all 𝑢, 𝑣 ∈ 𝑉.
(a) Show that 𝑇𝐂 is a (complex) linear map from 𝑉𝐂 to 𝑊𝐂 .
(b) Show that 𝑇𝐂 is injective if and only if 𝑇 is injective.
(c) Show that range 𝑇𝐂 = 𝑊𝐂 if and only if range 𝑇 = 𝑊.
The linear map 𝑇𝐂 is called the complexification of the linear map 𝑇.
Section 3C Matrices 69
3C Matrices
Representing a Linear Map by a Matrix
We know that if 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝑇 ∶ 𝑉 → 𝑊 is linear, then the values
of 𝑇𝑣1 , …, 𝑇𝑣𝑛 determine the values of 𝑇 on arbitrary vectors in 𝑉—see the linear
map lemma (3.4). As we will soon see, matrices provide an efficient method of
recording the values of the 𝑇𝑣𝑘 ’s in terms of a basis of 𝑊.
3.29 definition: matrix, 𝐴𝑗, 𝑘
Suppose 𝑚 and 𝑛 are nonnegative integers. An 𝑚-by-𝑛 matrix 𝐴 is a rectangular

array of elements of 𝐅 with 𝑚 rows and 𝑛 columns:
𝐴 ⋯ 𝐴 1, 𝑛
⎛ 1, 1 ⎞
𝐴=⎜
⎜
⎜ ⋮ ⋮ ⎟
⎟
⎟.
⎝ 𝐴𝑚, 1 ⋯ 𝐴 𝑚, 𝑛 ⎠
The notation 𝐴𝑗, 𝑘 denotes the entry in row 𝑗, column 𝑘 of 𝐴.
3.30 example: 𝐴𝑗, 𝑘 equals entry in row 𝑗, column 𝑘 of 𝐴

8 4 5 − 3𝑖
Suppose 𝐴 = ( ). When dealing with matrices, the first
1 9 7
Thus 𝐴2, 3 refers to the entry in the sec- index refers to the row number; the sec-
ond row, third column of 𝐴, which means ond index refers to the column number.
that 𝐴2, 3 = 7.
Now we come to the key definition in this section.
3.31 definition: matrix of a linear map, ℳ(𝑇)
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝑤1 , …, 𝑤𝑚 is a basis

of 𝑊. The matrix of 𝑇 with respect to these bases is the 𝑚-by-𝑛 matrix ℳ(𝑇)
whose entries 𝐴𝑗, 𝑘 are defined by
𝑇𝑣𝑘 = 𝐴1, 𝑘 𝑤1 + ⋯ + 𝐴𝑚, 𝑘 𝑤𝑚 .
If the bases 𝑣1 , …, 𝑣𝑛 and 𝑤1 , …, 𝑤𝑚 are not clear from the context, then the
notation ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 ), (𝑤1 , …, 𝑤𝑚 )) is used.
The matrix ℳ(𝑇) of a linear map 𝑇 ∈ ℒ(𝑉, 𝑊) depends on the basis 𝑣1 , …, 𝑣𝑛

of 𝑉 and the basis 𝑤1 , …, 𝑤𝑚 of 𝑊, as well as on 𝑇. However, the bases should be
clear from the context, and thus they are often not included in the notation.
To remember how ℳ(𝑇) is constructed from 𝑇, you might write across the
top of the matrix the basis vectors 𝑣1 , …, 𝑣𝑛 for the domain and along the left the
basis vectors 𝑤1 , …, 𝑤𝑚 for the vector space into which 𝑇 maps, as follows:
𝑣1 ⋯ 𝑣𝑘 ⋯ 𝑣𝑛
𝑤1 𝐴1, 𝑘
⎛
⎜ ⎞
⎟.
ℳ(𝑇) = ⋮ ⎜
⎜ ⋮ ⎟
⎟
𝑤𝑚 ⎝ 𝐴𝑚, 𝑘 ⎠
In the matrix above only the 𝑘 th col-
The 𝑘 th column of ℳ(𝑇) consists of
umn is shown. Thus the second index of the scalars needed to write 𝑇𝑣 as a
each displayed entry of the matrix above linear combination of 𝑤 , …, 𝑤 :
𝑘
is 𝑘. The picture above should remind you

1 𝑚
𝑚
that 𝑇𝑣𝑘 can be computed from ℳ(𝑇) by 𝑇𝑣𝑘 = ∑ 𝐴𝑗, 𝑘 𝑤𝑗 .
multiplying each entry in the 𝑘 th column
𝑗=1
by the corresponding 𝑤𝑗 from the left col-

umn, and then adding up the resulting
vectors.
If 𝑇 is a linear map from 𝐅𝑛 to 𝐅𝑚, If 𝑇 is a linear map from an
then unless stated otherwise, assume the 𝑛-dimensional vector space to an
bases in question are the standard ones 𝑚-dimensional vector space, then
(where the 𝑘 th basis vector is 1 in the 𝑘 th ℳ(𝑇) is an 𝑚-by-𝑛 matrix.
slot and 0 in all other slots). If you think
of elements of 𝐅𝑚 as columns of 𝑚 numbers, then you can think of the 𝑘 th column
of ℳ(𝑇) as 𝑇 applied to the 𝑘 th standard basis vector.
3.32 example: the matrix of a linear map from 𝐅2 to 𝐅3

Suppose 𝑇 ∈ ℒ(𝐅2, 𝐅3 ) is defined by
𝑇(𝑥, 𝑦) = (𝑥 + 3𝑦, 2𝑥 + 5𝑦, 7𝑥 + 9𝑦).
Because 𝑇(1, 0) = (1, 2, 7) and 𝑇(0, 1) = (3, 5, 9), the matrix of 𝑇 with respect
to the standard bases is the 3-by-2 matrix below:
1 3
⎛ ⎞
ℳ(𝑇) = ⎜
⎜ ⎟.
⎜ 2 5 ⎟
⎟
⎝ 7 9 ⎠
When working with 𝒫𝑚 (𝐅), use the standard basis 1, 𝑥, 𝑥2, …, 𝑥𝑚 unless the
context indicates otherwise.
3.33 example: matrix of the differentiation map from 𝒫3 (𝐑) to 𝒫2 (𝐑)

Suppose 𝐷 ∈ ℒ(𝒫3 (𝐑), 𝒫2 (𝐑)) is the differentiation map defined by 𝐷𝑝 = 𝑝′.
Because (𝑥𝑛 )′ = 𝑛𝑥𝑛 − 1, the matrix of 𝐷 with respect to the standard bases is the
3-by-4 matrix below:
0 1 0 0
⎛ ⎞
ℳ(𝐷) = ⎜ ⎜
⎜ 0 0 2 0 ⎟
⎟.
⎟
⎝ 0 0 0 3 ⎠
Addition and Scalar Multiplication of Matrices

For the rest of this section, assume that 𝑈, 𝑉, and 𝑊 are finite-dimensional and
that a basis has been chosen for each of these vector spaces. Thus for each linear
map from 𝑉 to 𝑊, we can talk about its matrix (with respect to the chosen bases).
Is the matrix of the sum of two linear maps equal to the sum of the matrices of
the two maps? Right now this question does not yet make sense because although
we have defined the sum of two linear maps, we have not defined the sum of two
matrices. Fortunately, the natural definition of the sum of two matrices has the
right properties. Specifically, we make the following definition.
3.34 definition: matrix addition

The sum of two matrices of the same size is the matrix obtained by adding
corresponding entries in the matrices:
𝐴1, 1 ⋯ 𝐴1, 𝑛 𝐶1, 1 ⋯ 𝐶 1, 𝑛

⎛
⎜ ⎞
⎟ ⎛
⎜ ⋮ ⎞
⎟
⎜
⎜ ⋮ ⋮ ⎟+⎜
⎟ ⎜ ⋮ ⎟
⎟
⎝ 𝐴𝑚, 1 ⋯ 𝐴 𝑚, 𝑛 ⎠ ⎝ 𝐶𝑚, 1 ⋯ 𝐶 𝑚, 𝑛 ⎠
𝐴1, 1 + 𝐶1, 1 ⋯ 𝐴1, 𝑛 + 𝐶1, 𝑛

⎛
⎜ ⎞
⎟
⎜
=⎜ ⋮ ⋮ ⎟
⎟.
⎝ 𝐴𝑚, 1 + 𝐶𝑚, 1 ⋯ 𝐴𝑚, 𝑛 + 𝐶𝑚, 𝑛 ⎠
In the next result, the assumption is that the same bases are used for all three
linear maps 𝑆 + 𝑇, 𝑆, and 𝑇.
3.35 matrix of the sum of linear maps
Suppose 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊). Then ℳ(𝑆 + 𝑇) = ℳ(𝑆) + ℳ(𝑇).
The verification of the result above follows from the definitions and is left to
the reader.
Still assuming that we have some bases in mind, is the matrix of a scalar times
a linear map equal to the scalar times the matrix of the linear map? Again, the
question does not yet make sense because we have not defined scalar multiplication
on matrices. Fortunately, the natural definition again has the right properties.
3.36 definition: scalar multiplication of a matrix
The product of a scalar and a matrix is the matrix obtained by multiplying

each entry in the matrix by the scalar:
𝐴 ⋯ 𝐴 1, 𝑛 𝜆𝐴1, 1 ⋯ 𝜆𝐴1, 𝑛
⎛ 1, 1 ⎞ ⎛ ⎞
𝜆⎜
⎜
⎜ ⋮ ⋮ ⎟
⎟ =
⎟ ⎜
⎜
⎜ ⋮ ⋮ ⎟
⎟
⎟.
⎝ 𝐴𝑚, 1 ⋯ 𝐴 𝑚, 𝑛 ⎠ ⎝ 𝜆𝐴 𝑚, 1 ⋯ 𝜆𝐴𝑚, 𝑛 ⎠
3.37 example: addition and scalar multiplication of matrices
3 1 4 2 6 2 4 2 10 4
2( )+( )=( )+( )=( )
−1 5 1 6 −2 10 1 6 −1 16
In the next result, the assumption is that the same bases are used for both the
linear maps 𝜆𝑇 and 𝑇.
3.38 the matrix of a scalar times a linear map
Suppose 𝜆 ∈ 𝐅 and 𝑇 ∈ ℒ(𝑉, 𝑊). Then ℳ(𝜆𝑇) = 𝜆ℳ(𝑇).
The verification of the result above is also left to the reader.

Because addition and scalar multiplication have now been defined for matrices,
you should not be surprised that a vector space is about to appear. First we
introduce a bit of notation so that this new vector space has a name, and then we
find the dimension of this new vector space.
3.39 notation: 𝐅𝑚, 𝑛
For 𝑚 and 𝑛 positive integers, the set of all 𝑚-by-𝑛 matrices with entries in 𝐅
is denoted by 𝐅𝑚, 𝑛.
3.40 dim 𝐅𝑚, 𝑛 = 𝑚𝑛

Suppose 𝑚 and 𝑛 are positive integers. With addition and scalar multiplication
defined as above, 𝐅𝑚, 𝑛 is a vector space of dimension 𝑚𝑛.
Proof The verification that 𝐅𝑚, 𝑛 is a vector space is left to the reader. Note that
the additive identity of 𝐅𝑚, 𝑛 is the 𝑚-by-𝑛 matrix all of whose entries equal 0.
The reader should also verify that the list of distinct 𝑚-by-𝑛 matrices that have
0 in all entries except for a 1 in one entry is a basis of 𝐅𝑚, 𝑛. There are 𝑚𝑛 such
matrices, so the dimension of 𝐅𝑚, 𝑛 equals 𝑚𝑛.
Matrix Multiplication
Suppose, as previously, that 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝑤1 , …, 𝑤𝑚 is a basis of 𝑊.
Suppose also that 𝑢1 , …, 𝑢𝑝 is a basis of 𝑈.
Consider linear maps 𝑇 ∶ 𝑈 → 𝑉 and 𝑆 ∶ 𝑉 → 𝑊. The composition 𝑆𝑇 is a
linear map from 𝑈 to 𝑊. Does ℳ(𝑆𝑇) equal ℳ(𝑆)ℳ(𝑇)? This question does
not yet make sense because we have not defined the product of two matrices. We
will choose a definition of matrix multiplication that forces this question to have
a positive answer. Let’s see how to do this.
Suppose ℳ(𝑆) = 𝐴 and ℳ(𝑇) = 𝐵. For 1 ≤ 𝑘 ≤ 𝑝, we have

𝑛
(𝑆𝑇)𝑢𝑘 = 𝑆( ∑ 𝐵𝑟, 𝑘 𝑣𝑟 )
𝑟=1
𝑛
= ∑ 𝐵𝑟, 𝑘 𝑆𝑣𝑟
𝑟=1
𝑛 𝑚
= ∑ 𝐵𝑟, 𝑘 ∑ 𝐴𝑗, 𝑟 𝑤𝑗
𝑟=1 𝑗=1
𝑚 𝑛
= ∑ ( ∑ 𝐴𝑗, 𝑟 𝐵𝑟, 𝑘 )𝑤𝑗 .
𝑗=1 𝑟=1
Thus ℳ(𝑆𝑇) is the 𝑚-by-𝑝 matrix whose entry in row 𝑗, column 𝑘, equals
𝑛
∑ 𝐴𝑗, 𝑟 𝐵𝑟, 𝑘 .
𝑟=1
Now we see how to define matrix multiplication so that the desired equation
ℳ(𝑆𝑇) = ℳ(𝑆)ℳ(𝑇) holds.
3.41 definition: matrix multiplication
Suppose 𝐴 is an 𝑚-by-𝑛 matrix and 𝐵 is an 𝑛-by-𝑝 matrix. Then 𝐴𝐵 is defined

to be the 𝑚-by-𝑝 matrix whose entry in row 𝑗, column 𝑘, is given by the equation
𝑛
(𝐴𝐵)𝑗, 𝑘 = ∑ 𝐴𝑗, 𝑟 𝐵𝑟, 𝑘 .
𝑟=1
Thus the entry in row 𝑗, column 𝑘, of 𝐴𝐵 is computed by taking row 𝑗 of 𝐴 and

column 𝑘 of 𝐵, multiplying together corresponding entries, and then summing.
Note that we define the product of You may have learned this definition
two matrices only when the number of of matrix multiplication in an earlier
columns of the first matrix equals the course, although you may not have
number of rows of the second matrix. seen this motivation for it.
3.42 example: matrix multiplication

Here we multiply together a 3-by-2 matrix and a 2-by-4 matrix, obtaining a
3-by-4 matrix:
1 2 10 7 4 1
⎛
⎜ ⎞
⎟ 6 5 4 3 ⎛
⎜ ⎞
⎟
⎜
⎜ 3 4 ⎟
⎟( 2 ⎜
) = ⎜ 26 19 12 5 ⎟
⎟.
1 0 −1
⎝ 5 6 ⎠ ⎝ 42 31 20 9 ⎠
Matrix multiplication is not commutative—𝐴𝐵 is not necessarily equal to

𝐵𝐴 even if both products are defined (see Exercise 10). Matrix multiplication is
distributive and associative (see Exercises 11 and 12).
In the next result, we assume that the same basis of 𝑉 is used in considering
𝑇 ∈ ℒ(𝑈, 𝑉) and 𝑆 ∈ ℒ(𝑉, 𝑊), the same basis of 𝑊 is used in considering
𝑆 ∈ ℒ(𝑉, 𝑊) and 𝑆𝑇 ∈ ℒ(𝑈, 𝑊), and the same basis of 𝑈 is used in considering
𝑇 ∈ ℒ(𝑈, 𝑉) and 𝑆𝑇 ∈ ℒ(𝑈, 𝑊).
3.43 matrix of product of linear maps
If 𝑇 ∈ ℒ(𝑈, 𝑉) and 𝑆 ∈ ℒ(𝑉, 𝑊), then ℳ(𝑆𝑇) = ℳ(𝑆)ℳ(𝑇).
The proof of the result above is the calculation that was done as motivation
before the definition of matrix multiplication.
In the next piece of notation, note that as usual the first index refers to a row
and the second index refers to a column, with a vertically centered dot used as a
placeholder.
3.44 notation: 𝐴𝑗, ⋅ , 𝐴⋅, 𝑘
Suppose 𝐴 is an 𝑚-by-𝑛 matrix.

• If 1 ≤ 𝑗 ≤ 𝑚, then 𝐴𝑗, ⋅ denotes the 1-by-𝑛 matrix consisting of row 𝑗 of 𝐴.
• If 1 ≤ 𝑘 ≤ 𝑛, then 𝐴⋅, 𝑘 denotes the 𝑚-by-1 matrix consisting of column 𝑘
of 𝐴.
3.45 example: 𝐴𝑗, ⋅ equals 𝑗th row of 𝐴 and 𝐴⋅, 𝑘 equals 𝑘 th column of 𝐴
The notation 𝐴2, ⋅ denotes the second row of 𝐴 and 𝐴⋅, 2 denotes the second
8 4 5
column of 𝐴. Thus if 𝐴 = ( ), then
1 9 7
4
𝐴2, ⋅ = ( 1 9 7 ) and 𝐴⋅, 2 = ( ).
9
The product of a 1-by-𝑛 matrix and an 𝑛-by-1 matrix is a 1-by-1 matrix. How-
ever, we will frequently identify a 1-by-1 matrix with its entry. For example,
6
( 3 4 )( ) = ( 26 )
2
because 3 ⋅ 6 + 4 ⋅ 2 = 26. However, we can identify ( 26 ) with 26, writing

6
( 3 4 )( ) = 26.
2
The next result uses the convention discussed in the paragraph above to give
another way to think of matrix multiplication. For example, the next result and
the calculation in the paragraph above explain why the entry in row 2, column 1,
of the product in Example 3.42 equals 26.
3.46 entry of matrix product equals row times column
Suppose 𝐴 is an 𝑚-by-𝑛 matrix and 𝐵 is an 𝑛-by-𝑝 matrix. Then

(𝐴𝐵)𝑗, 𝑘 = 𝐴𝑗, ⋅ 𝐵⋅, 𝑘
if 1 ≤ 𝑗 ≤ 𝑚 and 1 ≤ 𝑘 ≤ 𝑝. In other words, the entry in row 𝑗, column 𝑘, of
𝐴𝐵 equals (row 𝑗 of 𝐴) times (column 𝑘 of 𝐵).
Proof Suppose 1 ≤ 𝑗 ≤ 𝑚 and 1 ≤ 𝑘 ≤ 𝑝. The definition of matrix multiplication

states that
3.47 (𝐴𝐵)𝑗, 𝑘 = 𝐴𝑗, 1 𝐵1, 𝑘 + ⋯ + 𝐴𝑗, 𝑛 𝐵𝑛, 𝑘 .
The definition of matrix multiplication also implies that the product of the 1-by-𝑛
matrix 𝐴𝑗, ⋅ and the 𝑛-by-1 matrix 𝐵⋅, 𝑘 is the 1-by-1 matrix whose entry is the
number on the right side of the equation above. Thus the entry in row 𝑗, column 𝑘,
of 𝐴𝐵 equals (row 𝑗 of 𝐴) times (column 𝑘 of 𝐵).
The next result gives yet another way to think of matrix multiplication. In the
result below, (𝐴𝐵)⋅, 𝑘 is column 𝑘 of the 𝑚-by-𝑝 matrix 𝐴𝐵. Thus (𝐴𝐵)⋅, 𝑘 is an
𝑚-by-1 matrix. Also, 𝐴𝐵⋅, 𝑘 is an 𝑚-by-1 matrix because it is the product of an
𝑚-by-𝑛 matrix and an 𝑛-by-1 matrix. Thus the two sides of the equation in the
result below have the same size, making it reasonable that they might be equal.
3.48 column of matrix product equals matrix times column
Suppose 𝐴 is an 𝑚-by-𝑛 matrix and 𝐵 is an 𝑛-by-𝑝 matrix. Then

(𝐴𝐵)⋅, 𝑘 = 𝐴𝐵⋅, 𝑘
if 1 ≤ 𝑘 ≤ 𝑝. In other words, column 𝑘 of 𝐴𝐵 equals 𝐴 times column 𝑘 of 𝐵.
Proof As discussed above, (𝐴𝐵)⋅, 𝑘 and 𝐴𝐵⋅, 𝑘 are both 𝑚-by-1 matrices. If 1 ≤
𝑗 ≤ 𝑚, then the entry in row 𝑗 of (𝐴𝐵)⋅, 𝑘 is the left side of 3.47 and the entry in
row 𝑗 of 𝐴𝐵⋅, 𝑘 is the right side of 3.47. Thus (𝐴𝐵)⋅, 𝑘 = 𝐴𝐵⋅, 𝑘 .
Our next result will give another way of thinking about the product of an
𝑚-by-𝑛 matrix and an 𝑛-by-1 matrix, motivated by the next example.
3.49 example: product of a 3-by-2 matrix and a 2-by-1 matrix

Use our definitions and basic arithmetic to verify that
1 2 7 1 2
⎛
⎜ ⎞
⎟ 5 ⎛
⎜ ⎞
⎟ ⎛
⎜ ⎞
⎟ ⎛
⎜ ⎞
⎟.
⎜ 3 4 ⎟( 1 ) = ⎜ 19 ⎟ = 5⎜ 3 ⎟ + 1⎜ 4 ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎝ 5 6 ⎠ ⎝ 31 ⎠ ⎝ 5 ⎠ ⎝ 6 ⎠
Thus in this example, the product of a 3-by-2 matrix and a 2-by-1 matrix is a
linear combination of the columns of the 3-by-2 matrix, with the scalars (5 and 1)
that multiply the columns coming from the 2-by-1 matrix.
The next result generalizes the example above.
3.50 linear combination of columns
𝑏1
⎛ ⎞
Suppose 𝐴 is an 𝑚-by-𝑛 matrix and 𝑏 = ⎜ ⋮ ⎟
⎜
⎜ ⎟ is an 𝑛-by-1 matrix. Then
⎟
⎝ 𝑏𝑛 ⎠
𝐴𝑏 = 𝑏1 𝐴⋅, 1 + ⋯ + 𝑏𝑛 𝐴⋅, 𝑛 .
In other words, 𝐴𝑏 is a linear combination of the columns of 𝐴, with the

scalars that multiply the columns coming from 𝑏.
Proof If 𝑘 ∈ {1, …, 𝑚}, then the definition of matrix multiplication implies that
the entry in row 𝑘 of the 𝑚-by-1 matrix 𝐴𝑏 is
𝐴𝑘, 1 𝑏1 + ⋯ + 𝐴𝑘, 𝑛 𝑏𝑛 .
The entry in row 𝑘 of 𝑏1 𝐴⋅, 1 + ⋯ + 𝑏𝑛 𝐴⋅, 𝑛 also equals the number displayed above.
Because 𝐴𝑏 and 𝑏1 𝐴⋅, 1 + ⋯ + 𝑏𝑛 𝐴⋅, 𝑛 have the same entry in row 𝑘 for each
𝑘 ∈ {1, …, 𝑚}, we conclude that 𝐴𝑏 = 𝑏1 𝐴⋅, 1 + ⋯ + 𝑏𝑛 𝐴⋅, 𝑛 .
Our two previous results focus on the columns of a matrix. Analogous results
hold for the rows of a matrix. Specifically, see Exercises 8 and 9, which can be
proved using appropriate modifications of the proofs of 3.48 and 3.50.
The next result is the main tool used in the next subsection to prove the
column–row factorization (3.56) and to prove that the column rank of a matrix
equals the row rank (3.57). To be consistent with the notation often used with the
column–row factorization, including in the next subsection, the matrices in the
next result are called 𝐶 and 𝑅 instead of 𝐴 and 𝐵.
3.51 matrix multiplication as linear combinations of columns
Suppose 𝐶 is an 𝑚-by-𝑐 matrix and 𝑅 is a 𝑐-by-𝑛 matrix.

(a) If 𝑘 ∈ {1, …, 𝑛}, then column 𝑘 of 𝐶𝑅 is a linear combination of the
columns of 𝐶, with the coefficients of this linear combination coming
from column 𝑘 of 𝑅.
(b) If 𝑗 ∈ {1, …, 𝑚}, then row 𝑗 of 𝐶𝑅 is a linear combination of the rows of 𝑅,
with the coefficients of this linear combination coming from row 𝑗 of 𝐶.
Proof Suppose 𝑘 ∈ {1, …, 𝑛}. Then column 𝑘 of 𝐶𝑅 equals 𝐶𝑅⋅, 𝑘 (by 3.48),
which equals the linear combination of the columns of 𝐶 with coefficients coming
from 𝑅⋅, 𝑘 (by 3.50). Thus (a) holds.
To prove (b), follow the pattern of the proof of (a) but use rows instead of
columns and use Exercises 8 and 9 instead of 3.48 and 3.50.
Column–Row Factorization and Rank of a Matrix

We begin by defining two nonnegative integers associated with each matrix.
3.52 definition: column rank, row rank
Suppose 𝐴 is an 𝑚-by-𝑛 matrix with entries in 𝐅.

• The column rank of 𝐴 is the dimension of the span of the columns of 𝐴
in 𝐅𝑚, 1.
• The row rank of 𝐴 is the dimension of the span of the rows of 𝐴 in 𝐅1, 𝑛.
If 𝐴 is an 𝑚-by-𝑛 matrix, then the column rank of 𝐴 is at most 𝑛 (because 𝐴 has

𝑛 columns) and the column rank of 𝐴 is also at most 𝑚 (because dim 𝐅𝑚, 1 = 𝑚).
Similarly, the row rank of 𝐴 is also at most min{𝑚, 𝑛}.
3.53 example: column rank and row rank of a 2-by-4 matrix

Suppose
4 7 1 8
𝐴=( ).
3 5 2 9
The column rank of 𝐴 is the dimension of
4 7 1 8 ⎞
span ⎛
⎜
⎜( ), ( ), ( ), ( )⎟ ⎟
3 5 2 9
⎝ ⎠
in 𝐅 . Neither of the first two vectors listed above in 𝐅 is a scalar multiple of
2, 1 2, 1
the other. Thus the span of this list of length four has dimension at least two. The
span of this list of vectors in 𝐅2, 1 cannot have dimension larger than two because
dim 𝐅2, 1 = 2. Thus the span of this list has dimension two, which means that the
column rank of 𝐴 is two.
The row rank of 𝐴 is the dimension of
span(( 4 7 1 8 ), ( 3 5 2 9 ))
in 𝐅1, 4. Neither of the two vectors listed above in 𝐅1, 4 is a scalar multiple of the
other. Thus the span of this list of length two has dimension two, which means
that the row rank of 𝐴 is two.
We now define the transpose of a matrix.
3.54 definition: transpose, 𝐴t
The transpose of a matrix 𝐴, denoted by 𝐴t, is the matrix obtained from 𝐴 by

interchanging rows and columns. Specifically, if 𝐴 is an 𝑚-by-𝑛 matrix, then
𝐴t is the 𝑛-by-𝑚 matrix whose entries are given by the equation
(𝐴t )𝑘, 𝑗 = 𝐴𝑗, 𝑘 .

3.55 example: transpose of a matrix

5 −7
⎛ ⎞ t 5 3 −4
If 𝐴 = ⎜
⎜
⎜ 3 8 ⎟ ⎟, then 𝐴 = ( −7 8 2 ).
⎟
⎝ −4 2 ⎠
Note that here 𝐴 is a 3-by-2 matrix and 𝐴t is a 2-by-3 matrix.
The transpose has nice algebraic properties: (𝐴 + 𝐵)t = 𝐴t + 𝐵 t, (𝜆𝐴)t = 𝜆𝐴t,

and (𝐴𝐶)t = 𝐶 t 𝐴t for all 𝑚-by-𝑛 matrices 𝐴, 𝐵, all 𝜆 ∈ 𝐅, and all 𝑛-by-𝑝 matrices
𝐶 (see Exercises 14 and 15).
The next result will be the main tool used to prove that the column rank equals
the row rank (see 3.57).
3.56 column–row factorization
Suppose 𝐴 is an 𝑚-by-𝑛 matrix with entries in 𝐅 and column rank 𝑐 ≥ 1. Then

there exist an 𝑚-by-𝑐 matrix 𝐶 and a 𝑐-by-𝑛 matrix 𝑅, both with entries in 𝐅,
such that 𝐴 = 𝐶𝑅.
Proof Each column of 𝐴 is an 𝑚-by-1 matrix. The list 𝐴⋅, 1 , …, 𝐴⋅, 𝑛 of columns
of 𝐴 can be reduced to a basis of the span of the columns of 𝐴 (by 2.30). This
basis has length 𝑐, by the definition of the column rank. The 𝑐 columns in this
basis can be put together to form an 𝑚-by-𝑐 matrix 𝐶.
If 𝑘 ∈ {1, …, 𝑛}, then column 𝑘 of 𝐴 is a linear combination of the columns
of 𝐶. Make the coefficients of this linear combination into column 𝑘 of a 𝑐-by-𝑛
matrix that we call 𝑅. Then 𝐴 = 𝐶𝑅, as follows from 3.51(a).
In Example 3.53, the column rank and row rank turned out to equal each other.
The next result states that this happens for all matrices.
3.57 column rank equals row rank
Suppose 𝐴 ∈ 𝐅𝑚, 𝑛. Then the column rank of 𝐴 equals the row rank of 𝐴.
Proof Let 𝑐 denote the column rank of 𝐴. Let 𝐴 = 𝐶𝑅 be the column–row

factorization of 𝐴 given by 3.56, where 𝐶 is an 𝑚-by-𝑐 matrix and 𝑅 is a 𝑐-by-𝑛
matrix. Then 3.51(b) tells us that every row of 𝐴 is a linear combination of the
rows of 𝑅. Because 𝑅 has 𝑐 rows, this implies that the row rank of 𝐴 is less than
or equal to the column rank 𝑐 of 𝐴.
To prove the inequality in the other direction, apply the result in the previous
paragraph to 𝐴t, getting
column rank of 𝐴 = row rank of 𝐴t
≤ column rank of 𝐴t
= row rank of 𝐴.
Thus the column rank of 𝐴 equals the row rank of 𝐴.
Because the column rank equals the row rank, the last result allows us to
dispense with the terms “column rank” and “row rank” and just use the simpler
term “rank”.
3.58 definition: rank

The rank of a matrix 𝐴 ∈ 𝐅𝑚, 𝑛 is the column rank of 𝐴.
See 3.133 and Exercise 8 in Section 7A for alternative proofs that the column
rank equals the row rank.
Exercises 3C
1 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Show that with respect to each choice of bases of 𝑉
and 𝑊, the matrix of 𝑇 has at least dim range 𝑇 nonzero entries.
2 Suppose 𝑉 and 𝑊 are finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that
dim range 𝑇 = 1 if and only if there exist a basis of 𝑉 and a basis of 𝑊 such
that with respect to these bases, all entries of ℳ(𝑇) equal 1.
3 Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝑤1 , …, 𝑤𝑚 is a basis of 𝑊.
(a) Show that if 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊), then ℳ(𝑆 + 𝑇) = ℳ(𝑆) + ℳ(𝑇).
(b) Show that if 𝜆 ∈ 𝐅 and 𝑇 ∈ ℒ(𝑉, 𝑊), then ℳ(𝜆𝑇) = 𝜆ℳ(𝑇).
This exercise asks you to verify 3.35 and 3.38.
4 Suppose that 𝐷 ∈ ℒ(𝒫3 (𝐑), 𝒫2 (𝐑)) is the differentiation map defined by

𝐷𝑝 = 𝑝′. Find a basis of 𝒫3 (𝐑) and a basis of 𝒫2 (𝐑) such that the matrix of
𝐷 with respect to these bases is
1 0 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ 0 1 0 0 ⎟
⎟.
⎝ 0 0 1 0 ⎠
Compare with Example 3.33. The next exercise generalizes this exercise.
5 Suppose 𝑉 and 𝑊 are finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that there
exist a basis of 𝑉 and a basis of 𝑊 such that with respect to these bases, all
entries of ℳ(𝑇) are 0 except that the entries in row 𝑘, column 𝑘, equal 1 if
1 ≤ 𝑘 ≤ dim range 𝑇.
6 Suppose 𝑣1 , …, 𝑣𝑚 is a basis of 𝑉 and 𝑊 is finite-dimensional. Suppose

𝑇 ∈ ℒ(𝑉, 𝑊). Prove that there exists a basis 𝑤1 , …, 𝑤𝑛 of 𝑊 such that all
entries in the first column of ℳ(𝑇) [with respect to the bases 𝑣1 , …, 𝑣𝑚 and
𝑤1 , …, 𝑤𝑛 ] are 0 except for possibly a 1 in the first row, first column.
In this exercise, unlike Exercise 5, you are given the basis of 𝑉 instead of
being able to choose a basis of 𝑉.
7 Suppose 𝑤1 , …, 𝑤𝑛 is a basis of 𝑊 and 𝑉 is finite-dimensional. Suppose

𝑇 ∈ ℒ(𝑉, 𝑊). Prove that there exists a basis 𝑣1 , …, 𝑣𝑚 of 𝑉 such that all
entries in the first row of ℳ(𝑇) [with respect to the bases 𝑣1 , …, 𝑣𝑚 and
𝑤1 , …, 𝑤𝑛 ] are 0 except for possibly a 1 in the first row, first column.
In this exercise, unlike Exercise 5, you are given the basis of 𝑊 instead of
being able to choose a basis of 𝑊.
8 Suppose 𝐴 is an 𝑚-by-𝑛 matrix and 𝐵 is an 𝑛-by-𝑝 matrix. Prove that

(𝐴𝐵)𝑗, ⋅ = 𝐴𝑗, ⋅ 𝐵
for each 1 ≤ 𝑗 ≤ 𝑚. In other words, show that row 𝑗 of 𝐴𝐵 equals (row 𝑗 of 𝐴)
times 𝐵.
This exercise gives the row version of 3.48.
9 Suppose 𝑎 = ( 𝑎1 ⋯ 𝑎𝑛 ) is a 1-by-𝑛 matrix and 𝐵 is an 𝑛-by-𝑝 matrix.

Prove that
𝑎𝐵 = 𝑎1 𝐵1, ⋅ + ⋯ + 𝑎𝑛 𝐵𝑛, ⋅ .
In other words, show that 𝑎𝐵 is a linear combination of the rows of 𝐵, with
the scalars that multiply the rows coming from 𝑎.
This exercise gives the row version of 3.50.
10 Give an example of 2-by-2 matrices 𝐴 and 𝐵 such that 𝐴𝐵 ≠ 𝐵𝐴.

11 Prove that the distributive property holds for matrix addition and matrix
multiplication. In other words, suppose 𝐴, 𝐵, 𝐶, 𝐷, 𝐸, and 𝐹 are matrices
whose sizes are such that 𝐴(𝐵 + 𝐶) and (𝐷 + 𝐸)𝐹 make sense. Explain why
𝐴𝐵 + 𝐴𝐶 and 𝐷𝐹 + 𝐸𝐹 both make sense and prove that
𝐴(𝐵 + 𝐶) = 𝐴𝐵 + 𝐴𝐶 and (𝐷 + 𝐸)𝐹 = 𝐷𝐹 + 𝐸𝐹.
12 Prove that matrix multiplication is associative. In other words, suppose 𝐴, 𝐵,

and 𝐶 are matrices whose sizes are such that (𝐴𝐵)𝐶 makes sense. Explain
why 𝐴(𝐵𝐶) makes sense and prove that
(𝐴𝐵)𝐶 = 𝐴(𝐵𝐶).
Try to find a clean proof that illustrates the following quote from Emil Artin:
“It is my experience that proofs involving matrices can be shortened by 50%
if one throws the matrices out.”
13 Suppose 𝐴 is an 𝑛-by-𝑛 matrix and 1 ≤ 𝑗, 𝑘 ≤ 𝑛. Show that the entry in

row 𝑗, column 𝑘, of 𝐴3 (which is defined to mean 𝐴𝐴𝐴) is
𝑛 𝑛
∑ ∑ 𝐴𝑗, 𝑝 𝐴𝑝, 𝑟 𝐴𝑟, 𝑘 .
𝑝=1 𝑟=1
14 Suppose 𝑚 and 𝑛 are positive integers. Prove that the function 𝐴 ↦ 𝐴t is a

linear map from 𝐅𝑚, 𝑛 to 𝐅𝑛, 𝑚.
15 Prove that if 𝐴 is an 𝑚-by-𝑛 matrix and 𝐶 is an 𝑛-by-𝑝 matrix, then
(𝐴𝐶)t = 𝐶 t 𝐴t.
This exercise shows that the transpose of the product of two matrices is the
product of the transposes in the opposite order.
16 Suppose 𝐴 is an 𝑚-by-𝑛 matrix with 𝐴 ≠ 0. Prove that the rank of 𝐴 is 1

if and only if there exist (𝑐1 , …, 𝑐𝑚 ) ∈ 𝐅𝑚 and (𝑑1 , …, 𝑑𝑛 ) ∈ 𝐅𝑛 such that
𝐴𝑗, 𝑘 = 𝑐𝑗 𝑑𝑘 for every 𝑗 = 1, …, 𝑚 and every 𝑘 = 1, …, 𝑛.
17 Suppose 𝑇 ∈ ℒ(𝑉), and 𝑢1 , …, 𝑢𝑛 and 𝑣1 , …, 𝑣𝑛 are bases of 𝑉. Prove that

the following are equivalent.
(a) 𝑇 is injective.
(b) The columns of ℳ(𝑇) are linearly independent in 𝐅𝑛, 1.
(c) The columns of ℳ(𝑇) span 𝐅𝑛, 1.
(d) The rows of ℳ(𝑇) span 𝐅1, 𝑛.
(e) The rows of ℳ(𝑇) are linearly independent in 𝐅1, 𝑛.
Here ℳ(𝑇) means ℳ(𝑇, (𝑢1 , …, 𝑢𝑛 ), (𝑣1 , …, 𝑣𝑛 )).
3D Invertibility and Isomorphisms

Invertible Linear Maps
We begin this section by defining the notions of invertible and inverse in the
context of linear maps.
3.59 definition: invertible, inverse
• A linear map 𝑇 ∈ ℒ(𝑉, 𝑊) is called invertible if there exists a linear map

𝑆 ∈ ℒ(𝑊, 𝑉) such that 𝑆𝑇 equals the identity operator on 𝑉 and 𝑇𝑆 equals
the identity operator on 𝑊.
• A linear map 𝑆 ∈ ℒ(𝑊, 𝑉) satisfying 𝑆𝑇 = 𝐼 and 𝑇𝑆 = 𝐼 is called an
inverse of 𝑇 (note that the first 𝐼 is the identity operator on 𝑉 and the second
𝐼 is the identity operator on 𝑊).
The definition above mentions “an inverse”. However, the next result shows
that we can change this terminology to “the inverse”.
3.60 inverse is unique
An invertible linear map has a unique inverse.
Proof Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) is invertible and 𝑆1 and 𝑆2 are inverses of 𝑇. Then

𝑆1 = 𝑆1 𝐼 = 𝑆1 (𝑇𝑆2 ) = (𝑆1 𝑇)𝑆2 = 𝐼𝑆2 = 𝑆2 .
Thus 𝑆1 = 𝑆2 .
Now that we know that the inverse is unique, we can give it a notation.
3.61 notation: 𝑇 −1
If 𝑇 is invertible, then its inverse is denoted by 𝑇 −1. In other words, if

𝑇 ∈ ℒ(𝑉, 𝑊) is invertible, then 𝑇 −1 is the unique element of ℒ(𝑊, 𝑉) such
that 𝑇 −1 𝑇 = 𝐼 and 𝑇𝑇 −1 = 𝐼.
3.62 example: inverse of a linear map from 𝐑3 to 𝐑3

Suppose 𝑇 ∈ ℒ(𝐑3 ) is defined by 𝑇(𝑥, 𝑦, 𝑧) = (−𝑦, 𝑥, 4𝑧). Thus 𝑇 is a
counterclockwise rotation by 90∘ in the 𝑥𝑦-plane and a stretch by a factor of 4 in
the direction of the 𝑧-axis.
Hence the inverse map 𝑇 −1 ∈ ℒ(𝐑3 ) is the clockwise rotation by 90∘ in the
𝑥𝑦-plane and a stretch by a factor of 14 in the direction of the 𝑧-axis:
𝑇 −1 (𝑥, 𝑦, 𝑧) = (𝑦, −𝑥, 14 𝑧).

Section 3D Invertibility and Isomorphisms 83
The next result shows that a linear map is invertible if and only if it is one-to-
one and onto.
3.63 invertibility ⟺ injectivity and surjectivity
A linear map is invertible if and only if it is injective and surjective.
Proof Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). We need to show that 𝑇 is invertible if and only
if it is injective and surjective.
First suppose 𝑇 is invertible. To show that 𝑇 is injective, suppose 𝑢, 𝑣 ∈ 𝑉
and 𝑇𝑢 = 𝑇𝑣. Then
𝑢 = 𝑇 −1 (𝑇𝑢) = 𝑇 −1 (𝑇𝑣) = 𝑣,
so 𝑢 = 𝑣. Hence 𝑇 is injective.
We are still assuming that 𝑇 is invertible. Now we want to prove that 𝑇 is
surjective. To do this, let 𝑤 ∈ 𝑊. Then 𝑤 = 𝑇(𝑇 −1 𝑤), which shows that 𝑤 is
in the range of 𝑇. Thus range 𝑇 = 𝑊. Hence 𝑇 is surjective, completing this
direction of the proof.
Now suppose 𝑇 is injective and surjective. We want to prove that 𝑇 is invertible.
For each 𝑤 ∈ 𝑊, define 𝑆(𝑤) to be the unique element of 𝑉 such that 𝑇(𝑆(𝑤)) = 𝑤
(the existence and uniqueness of such an element follow from the surjectivity and
injectivity of 𝑇). The definition of 𝑆 implies that 𝑇 ∘ 𝑆 equals the identity operator
on 𝑊.
To prove that 𝑆 ∘ 𝑇 equals the identity operator on 𝑉, let 𝑣 ∈ 𝑉. Then
𝑇((𝑆 ∘ 𝑇)𝑣) = (𝑇 ∘ 𝑆)(𝑇𝑣) = 𝐼(𝑇𝑣) = 𝑇𝑣.
This equation implies that (𝑆 ∘ 𝑇)𝑣 = 𝑣 (because 𝑇 is injective). Thus 𝑆 ∘ 𝑇 equals

the identity operator on 𝑉.
To complete the proof, we need to show that 𝑆 is linear. To do this, suppose
𝑤1 , 𝑤2 ∈ 𝑊. Then
𝑇(𝑆(𝑤1 ) + 𝑆(𝑤2 )) = 𝑇(𝑆(𝑤1 )) + 𝑇(𝑆(𝑤2 )) = 𝑤1 + 𝑤2 .
Thus 𝑆(𝑤1 ) + 𝑆(𝑤2 ) is the unique element of 𝑉 that 𝑇 maps to 𝑤1 + 𝑤2 . By the

definition of 𝑆, this implies that 𝑆(𝑤1 + 𝑤2 ) = 𝑆(𝑤1 ) + 𝑆(𝑤2 ). Hence 𝑆 satisfies
the additive property required for linearity.
The proof of homogeneity is similar. Specifically, if 𝑤 ∈ 𝑊 and 𝜆 ∈ 𝐅, then
𝑇(𝜆𝑆(𝑤)) = 𝜆𝑇(𝑆(𝑤)) = 𝜆𝑤.
Thus 𝜆𝑆(𝑤) is the unique element of 𝑉 that 𝑇 maps to 𝜆𝑤. By the definition of 𝑆,
this implies that 𝑆(𝜆𝑤) = 𝜆𝑆(𝑤). Hence 𝑆 is linear, as desired.
For a linear map from a vector space to itself, you might wonder whether
injectivity alone, or surjectivity alone, is enough to imply invertibility. On infinite-
dimensional vector spaces, neither condition alone implies invertibility, as illus-
trated by the next example, which uses two familiar linear maps from Example 3.3.
3.64 example: neither injectivity nor surjectivity implies invertibility
• The multiplication by 𝑥2 linear map from 𝒫(𝐑) to 𝒫(𝐑) (see 3.3) is injective
but it is not invertible because it is not surjective (the polynomial 1 is not in
the range).
• The backward shift linear map from 𝐅∞ to 𝐅∞ (see 3.3) is surjective but it is
not invertible because it is not injective [the vector (1, 0, 0, 0, … ) is in the null
space].
In view of the example above, the next result is remarkable—it states that for
a linear map from a finite-dimensional vector space to a vector space of the same
dimension, either injectivity or surjectivity alone implies the other condition.
Note that the hypothesis below that dim 𝑉 = dim 𝑊 is automatically satisfied in
the important special case where 𝑉 is finite-dimensional and 𝑊 = 𝑉.
3.65 injectivity is equivalent to surjectivity (if dim 𝑉 = dim 𝑊 < ∞)
Suppose that 𝑉 and 𝑊 are finite-dimensional vector spaces, dim 𝑉 = dim 𝑊,

and 𝑇 ∈ ℒ(𝑉, 𝑊). Then
𝑇 is invertible ⟺ 𝑇 is injective ⟺ 𝑇 is surjective.
Proof The fundamental theorem of linear maps (3.21) states that
3.66 dim 𝑉 = dim null 𝑇 + dim range 𝑇.
If 𝑇 is injective (which by 3.15 is equivalent to the condition dim null 𝑇 = 0),

then the equation above implies that
dim range 𝑇 = dim 𝑉 − dim null 𝑇 = dim 𝑉 = dim 𝑊,
which implies that 𝑇 is surjective (by 2.39).

Conversely, if 𝑇 is surjective, then 3.66 implies that
dim null 𝑇 = dim 𝑉 − dim range 𝑇 = dim 𝑉 − dim 𝑊 = 0,
which implies that 𝑇 is injective.

Thus we have shown that 𝑇 is injective if and only if 𝑇 is surjective. Thus if
𝑇 is either injective or surjective, then 𝑇 is both injective and surjective, which
implies that 𝑇 is invertible. Hence 𝑇 is invertible if and only if 𝑇 is injective if
and only if 𝑇 is surjective.
The next example illustrates the power of the previous result. Although it is
possible to prove the result in the example below without using linear algebra, the
proof using linear algebra is cleaner and easier.
3.67 example: there exists a polynomial 𝑝 such that ((𝑥2 + 5𝑥 + 7)𝑝)″ = 𝑞

The linear map
″
𝑝 ↦ ((𝑥2 + 5𝑥 + 7)𝑝)
from 𝒫(𝐑) to itself is injective, as you can show. Thus we are tempted to use 3.65
to show that this map is surjective. However, Example 3.64 shows that the magic
of 3.65 does not apply to the infinite-dimensional vector space 𝒫(𝐑). We will
get around this problem by restricting attention to the finite-dimensional vector
space 𝒫𝑚 (𝐑).
Suppose 𝑞 ∈ 𝒫(𝐑). There exists a nonnegative integer 𝑚 such that 𝑞 ∈ 𝒫𝑚 (𝐑).
Define 𝑇 ∶ 𝒫𝑚 (𝐑) → 𝒫𝑚 (𝐑) by
″
𝑇𝑝 = ((𝑥2 + 5𝑥 + 7)𝑝) .
Multiplying a nonzero polynomial by (𝑥2 + 5𝑥 + 7) increases the degree by 2, and
then differentiating twice reduces the degree by 2. Thus 𝑇 is indeed a linear map
from 𝒫𝑚 (𝐑) to itself.
Every polynomial whose second derivative equals 0 is of the form 𝑎𝑥 + 𝑏,
where 𝑎, 𝑏 ∈ 𝐑 . Thus null 𝑇 = {0}. Hence 𝑇 is injective.
Thus 𝑇 is surjective (by 3.65), which means that there exists a polynomial
𝑝 ∈ 𝒫𝑚 (𝐑) such that ((𝑥2 + 5𝑥 + 7)𝑝) = 𝑞, as claimed in the title of this example.
″
Exercise 35 in Section 6A gives a similar but more spectacular example of

using 3.65.
The hypothesis in the result below that dim 𝑉 = dim 𝑊 holds in the important
special case in which 𝑉 is finite-dimensional and 𝑊 = 𝑉. Thus in that case, the
equation 𝑆𝑇 = 𝐼 implies that 𝑆𝑇 = 𝑇𝑆, even though we do not have multiplicative
commutativity of arbitrary linear maps from 𝑉 to 𝑉.
3.68 𝑆𝑇 = 𝐼 ⟺ 𝑇𝑆 = 𝐼 (on vector spaces of the same dimension)
Suppose 𝑉 and 𝑊 are finite-dimensional vector spaces of the same dimension,

𝑆 ∈ ℒ(𝑉, 𝑊), and 𝑇 ∈ ℒ(𝑊, 𝑉). Then 𝑆𝑇 = 𝐼 if and only if 𝑇𝑆 = 𝐼.
Proof First suppose 𝑆𝑇 = 𝐼. If 𝑣 ∈ 𝑉 and 𝑇𝑣 = 0, then

𝑣 = 𝐼𝑣 = (𝑆𝑇)𝑣 = 𝑆(𝑇𝑣) = 𝑆(0) = 0.
Thus 𝑇 is injective (by 3.15). Because 𝑉 and 𝑊 have the same dimension, this
implies that 𝑇 is invertible (by 3.65).
Now multiply both sides of the equation 𝑆𝑇 = 𝐼 by 𝑇 −1 on the right, getting
𝑆 = 𝑇 −1.
Thus 𝑇𝑆 = 𝑇𝑇 −1 = 𝐼, as desired.
To prove the implication in the other direction, simply reverse the roles of 𝑆
and 𝑇 (and 𝑉 and 𝑊) in the direction we have already proved, showing that if
𝑇𝑆 = 𝐼, then 𝑆𝑇 = 𝐼.
Isomorphic Vector Spaces

The next definition captures the idea of two vector spaces that are essentially the
same, except for the names of their elements.
3.69 definition: isomorphism, isomorphic
• An isomorphism is an invertible linear map.

• Two vector spaces are called isomorphic if there is an isomorphism from
one vector space onto the other one.
Think of an isomorphism 𝑇 ∶ 𝑉 → 𝑊 as relabeling 𝑣 ∈ 𝑉 as 𝑇𝑣 ∈ 𝑊. This

viewpoint explains why two isomorphic vector spaces have the same vector space
properties. The terms “isomorphism” and “invertible linear map” mean the same
thing. Use “isomorphism” when you want to emphasize that the two spaces are
essentially the same.
It can be difficult to determine whether two mathematical structures (such as
groups or topological spaces) are essentially the same, differing only in the names
of the elements of underlying sets. However, the next result shows that we need
to look at only a single number (the dimension) to determine whether two vector
spaces are isomorphic.
3.70 dimension shows whether vector spaces are isomorphic
Two finite-dimensional vector spaces over 𝐅 are isomorphic if and only if they
have the same dimension.
Proof First suppose 𝑉 and 𝑊 are isomorphic finite-dimensional vector spaces.

Thus there exists an isomorphism 𝑇 from 𝑉 onto 𝑊. Because 𝑇 is invertible, we
have null 𝑇 = {0} and range 𝑇 = 𝑊. Thus
dim null 𝑇 = 0 and dim range 𝑇 = dim 𝑊.
The formula
dim 𝑉 = dim null 𝑇 + dim range 𝑇
(the fundamental theorem of linear maps, which is 3.21) thus becomes the equation
dim 𝑉 = dim 𝑊, completing the proof in one direction.
To prove the other direction, suppose 𝑉 and 𝑊 are finite-dimensional vector
spaces of the same dimension. Let 𝑣1 , …, 𝑣𝑛 be a basis of 𝑉 and 𝑤1 , …, 𝑤𝑛 be a
basis of 𝑊. Let 𝑇 ∈ ℒ(𝑉, 𝑊) be defined by
𝑇(𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 ) = 𝑐1 𝑤1 + ⋯ + 𝑐𝑛 𝑤𝑛 .
Then 𝑇 is a well-defined linear map because 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Also, 𝑇
is surjective because 𝑤1 , …, 𝑤𝑛 spans 𝑊. Furthermore, null 𝑇 = {0} because
𝑤1 , …, 𝑤𝑛 is linearly independent. Thus 𝑇 is injective. Because 𝑇 is injective and
surjective, it is an isomorphism (see 3.63). Hence 𝑉 and 𝑊 are isomorphic.
The previous result implies that each Every finite-dimensional vector space
finite-dimensional vector space 𝑉 is iso- is isomorphic to some 𝐅𝑛. Thus why not
morphic to 𝐅𝑛, where 𝑛 = dim 𝑉. For just study 𝐅𝑛 instead of more general
example, if 𝑚 is a nonnegative integer, vector spaces? To answer this ques-
then 𝒫𝑚 (𝐅) is isomorphic to 𝐅𝑚 + 1. tion, note that an investigation of 𝐅𝑛
Recall that the notation 𝐅𝑚, 𝑛 denotes would soon lead to other vector spaces.
the vector space of 𝑚-by-𝑛 matrices with For example, we would encounter the
entries in 𝐅. If 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 null space and range of linear maps.
and 𝑤1 , …, 𝑤𝑚 is a basis of 𝑊, then for Although each of these vector spaces
each 𝑇 ∈ ℒ(𝑉, 𝑊), we have a matrix is isomorphic to some 𝐅𝑚, thinking of
ℳ(𝑇) ∈ 𝐅𝑚, 𝑛. Thus once bases have them that way often adds complexity
been fixed for 𝑉 and 𝑊, ℳ becomes a but no new insight.
function from ℒ(𝑉, 𝑊) to 𝐅𝑚, 𝑛. Notice
that 3.35 and 3.38 show that ℳ is a lin-
ear map. This linear map is actually an
isomorphism, as we now show.
3.71 ℒ(𝑉, 𝑊) and 𝐅𝑚, 𝑛 are isomorphic
Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝑤1 , …, 𝑤𝑚 is a basis of 𝑊. Then ℳ is

an isomorphism between ℒ(𝑉, 𝑊) and 𝐅𝑚, 𝑛.
Proof We already noted that ℳ is linear. We need to prove that ℳ is injective

and surjective.
We begin with injectivity. If 𝑇 ∈ ℒ(𝑉, 𝑊) and ℳ(𝑇) = 0, then 𝑇𝑣𝑘 = 0 for
each 𝑘 = 1, …, 𝑛. Because 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉, this implies 𝑇 = 0. Thus ℳ
is injective (by 3.15).
To prove that ℳ is surjective, suppose 𝐴 ∈ 𝐅𝑚, 𝑛. By the linear map lemma
(3.4), there exists 𝑇 ∈ ℒ(𝑉, 𝑊) such that
𝑚
𝑇𝑣𝑘 = ∑ 𝐴𝑗, 𝑘 𝑤𝑗
𝑗=1
for each 𝑘 = 1, …, 𝑛. Because ℳ(𝑇) equals 𝐴, the range of ℳ equals 𝐅𝑚, 𝑛, as

desired.
Now we can determine the dimension of the vector space of linear maps from
one finite-dimensional vector space to another.
3.72 dim ℒ(𝑉, 𝑊) = (dim 𝑉)(dim 𝑊)
Suppose 𝑉 and 𝑊 are finite-dimensional. Then ℒ(𝑉, 𝑊) is finite-dimensional

and
dim ℒ(𝑉, 𝑊) = (dim 𝑉)(dim 𝑊).
Proof The desired result follows from 3.71, 3.70, and 3.40.
Linear Maps Thought of as Matrix Multiplication

Previously we defined the matrix of a linear map. Now we define the matrix of a
vector.
3.73 definition: matrix of a vector, ℳ(𝑣)
Suppose 𝑣 ∈ 𝑉 and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. The matrix of 𝑣 with respect to

this basis is the 𝑛-by-1 matrix
𝑏1
⎛ ⎞
ℳ(𝑣) = ⎜ ⋮ ⎟
⎜
⎜ ⎟,
⎟
⎝ 𝑏𝑛 ⎠
where 𝑏1 , …, 𝑏𝑛 are the scalars such that
𝑣 = 𝑏1 𝑣1 + ⋯ + 𝑏𝑛 𝑣𝑛 .
The matrix ℳ(𝑣) of a vector 𝑣 ∈ 𝑉 depends on the basis 𝑣1 , …, 𝑣𝑛 of 𝑉, as

well as on 𝑣. However, the basis should be clear from the context and thus it is
not included in the notation.
3.74 example: matrix of a vector
• The matrix of the polynomial 2 − 7𝑥 + 5𝑥3 + 𝑥4 with respect to the standard

basis of 𝒫4 (𝐑) is
2
⎛
⎜ ⎞
⎟
⎜
⎜ −7 ⎟
⎟
⎜
⎜ ⎟
⎟
⎜
⎜ 0 ⎟.
⎟
⎜
⎜ 5 ⎟
⎟
⎜ ⎟
⎝ 1 ⎠
• The matrix of a vector 𝑥 ∈ 𝐅𝑛 with respect to the standard basis is obtained by

writing the coordinates of 𝑥 as the entries in an 𝑛-by-1 matrix. In other words,
if 𝑥 = (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛, then
𝑥
⎛ 1 ⎞
ℳ(𝑥) = ⎜
⎜
⎜ ⋮ ⎟
⎟.
⎟
⎝ 𝑥𝑛 ⎠
Occasionally we want to think of elements of 𝑉 as relabeled to be 𝑛-by-1

matrices. Once a basis 𝑣1 , …, 𝑣𝑛 is chosen, the function ℳ that takes 𝑣 ∈ 𝑉 to
ℳ(𝑣) is an isomorphism of 𝑉 onto 𝐅𝑛, 1 that implements this relabeling.
Recall that if 𝐴 is an 𝑚-by-𝑛 matrix, then 𝐴⋅, 𝑘 denotes the 𝑘 th column of 𝐴,
thought of as an 𝑚-by-1 matrix. In the next result, ℳ(𝑇𝑣𝑘 ) is computed with
respect to the basis 𝑤1 , …, 𝑤𝑚 of 𝑊.
3.75 ℳ(𝑇)⋅, 𝑘 = ℳ(𝑇𝑣𝑘 ).
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝑤1 , …, 𝑤𝑚 is a basis

of 𝑊. Let 1 ≤ 𝑘 ≤ 𝑛. Then the 𝑘 th column of ℳ(𝑇), which is denoted by
ℳ(𝑇)⋅, 𝑘 , equals ℳ(𝑇𝑣𝑘 ).
Proof The desired result follows immediately from the definitions of ℳ(𝑇) and
ℳ(𝑇𝑣𝑘 ).
The next result shows how the notions of the matrix of a linear map, the matrix
of a vector, and matrix multiplication fit together.
3.76 linear maps act like matrix multiplication
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑣 ∈ 𝑉. Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and

𝑤1 , …, 𝑤𝑚 is a basis of 𝑊. Then
ℳ(𝑇𝑣) = ℳ(𝑇)ℳ(𝑣).
Proof Suppose 𝑣 = 𝑏1 𝑣1 + ⋯ + 𝑏𝑛 𝑣𝑛 , where 𝑏1 , …, 𝑏𝑛 ∈ 𝐅. Thus

3.77 𝑇𝑣 = 𝑏1 𝑇𝑣1 + ⋯ + 𝑏𝑛 𝑇𝑣𝑛 .
Hence
ℳ(𝑇𝑣) = 𝑏1 ℳ(𝑇𝑣1 ) + ⋯ + 𝑏𝑛 ℳ(𝑇𝑣𝑛 )
= 𝑏1 ℳ(𝑇)⋅, 1 + ⋯ + 𝑏𝑛 ℳ(𝑇)⋅, 𝑛
= ℳ(𝑇)ℳ(𝑣),
where the first equality follows from 3.77 and the linearity of ℳ, the second
equality comes from 3.75, and the last equality comes from 3.50.
Each 𝑚-by-𝑛 matrix 𝐴 induces a linear map from 𝐅𝑛, 1 to 𝐅𝑚, 1, namely the
matrix multiplication function that takes 𝑥 ∈ 𝐅𝑛, 1 to 𝐴𝑥 ∈ 𝐅𝑚, 1. The result above
can be used to think of every linear map (from a finite-dimensional vector space
to another finite-dimensional vector space) as a matrix multiplication map after
suitable relabeling via the isomorphisms given by ℳ. Specifically, if 𝑇 ∈ ℒ(𝑉, 𝑊)
and we identify 𝑣 ∈ 𝑉 with ℳ(𝑣) ∈ 𝐅𝑛, 1, then the result above says that we can
identify 𝑇𝑣 with ℳ(𝑇)ℳ(𝑣).
Because the result above allows us to think (via isomorphisms) of each linear
map as multiplication on 𝐅𝑛, 1 by some matrix 𝐴, keep in mind that the specific
matrix 𝐴 depends not only on the linear map but also on the choice of bases. One
of the themes of many of the most important results in later chapters will be the
choice of a basis that makes the matrix 𝐴 as simple as possible.
In this book, we concentrate on linear maps rather than on matrices. However,
sometimes thinking of linear maps as matrices (or thinking of matrices as linear
maps) gives important insights that we will find useful.
Notice that no bases are in sight in the statement of the next result. Although
ℳ(𝑇) in the next result depends on a choice of bases of 𝑉 and 𝑊, the next result
shows that the column rank of ℳ(𝑇) is the same for all such choices (because
range 𝑇 does not depend on a choice of basis).
3.78 dimension of range 𝑇 equals column rank of ℳ(𝑇)
Suppose 𝑉 and 𝑊 are finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Then dim range 𝑇
equals the column rank of ℳ(𝑇).
Proof Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝑤1 , …, 𝑤𝑚 is a basis of 𝑊. The linear

map that takes 𝑤 ∈ 𝑊 to ℳ(𝑤) is an isomorphism from 𝑊 onto the space 𝐅𝑚, 1
of 𝑚-by-1 column vectors. The restriction of this isomorphism to range 𝑇 [which
equals span(𝑇𝑣1 , …, 𝑇𝑣𝑛 ) by Exercise 10 in Section 3B] is an isomorphism from
range 𝑇 onto span(ℳ(𝑇𝑣1 ), …, ℳ(𝑇𝑣𝑛 )). For each 𝑘 ∈ {1, …, 𝑛}, the 𝑚-by-1
matrix ℳ(𝑇𝑣𝑘 ) equals column 𝑘 of ℳ(𝑇). Thus
dim range 𝑇 = the column rank of ℳ(𝑇),
as desired.
Change of Basis
In Section 3C we defined the matrix
ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 ), (𝑤1 , …, 𝑤𝑚 ))
of a linear map 𝑇 from 𝑉 to a possibly different vector space 𝑊, where 𝑣1 , …, 𝑣𝑛
is a basis of 𝑉 and 𝑤1 , …, 𝑤𝑚 is a basis of 𝑊. For linear maps from a vector space
to itself, we usually use the same basis for both the domain vector space and the
target vector space. When using a single basis in both capacities, we often write
the basis only once. In other words, if 𝑇 ∈ ℒ(𝑉) and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉,
then the notation ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )) is defined by the equation
ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )) = ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 ), (𝑣1 , …, 𝑣𝑛 )).
If the basis 𝑣1 , …, 𝑣𝑛 is clear from the context, then we can write just ℳ(𝑇).
3.79 definition: identity matrix, I
Suppose 𝑛 is a positive integer. The 𝑛-by-𝑛 matrix

1 0
⎛
⎜ ⎞
⎟
⎜
⎜ ⋱ ⎟
⎟
⎝ 0 1 ⎠
with 1’s on the diagonal (the entries where the row number equals the column
number) and 0’s elsewhere is called the identity matrix and is denoted by 𝐼.
In the definition above, the 0 in the lower left corner of the matrix indicates that
all entries below the diagonal are 0, and the 0 in the upper right corner indicates
that all entries above the diagonal are 0.
With respect to each basis of 𝑉, the matrix of the identity operator 𝐼 ∈ ℒ(𝑉)
is the identity matrix 𝐼. Note that the symbol 𝐼 is used to denote both the identity
operator and the identity matrix. The context indicates which meaning of 𝐼 is
intended. For example, consider the equation ℳ(𝐼) = 𝐼; on the left side 𝐼 denotes
the identity operator, and on the right side 𝐼 denotes the identity matrix.
If 𝐴 is a square matrix (with entries in 𝐅, as usual) of the same size as 𝐼, then
𝐴𝐼 = 𝐼𝐴 = 𝐴, as you should verify.
3.80 definition: invertible, inverse, 𝐴−1
A square matrix 𝐴 is called invertible if there is a square matrix 𝐵 of the same

size such that 𝐴𝐵 = 𝐵𝐴 = 𝐼; we call 𝐵 the inverse of 𝐴 and denote it by 𝐴−1.
The same proof as used in 3.60 shows Some mathematicians use the terms
that if 𝐴 is an invertible square matrix, nonsingular and singular, which
then there is a unique matrix 𝐵 such that mean the same as invertible and non-
𝐴𝐵 = 𝐵𝐴 = 𝐼 (and thus the notation invertible.
𝐵 = 𝐴−1 is justified).
−1
If 𝐴 is an invertible matrix, then (𝐴−1 ) = 𝐴 because
𝐴−1 𝐴 = 𝐴𝐴−1 = 𝐼.
Also, if 𝐴 and 𝐶 are invertible square matrices of the same size, then 𝐴𝐶 is
invertible and (𝐴𝐶)−1 = 𝐶−1 𝐴−1 because
(𝐴𝐶)(𝐶−1 𝐴−1 ) = 𝐴(𝐶𝐶−1 )𝐴−1
= 𝐴𝐼𝐴−1
= 𝐴𝐴−1
= 𝐼,
and similarly (𝐶−1 𝐴−1 )(𝐴𝐶) = 𝐼.

The next result holds because we defined matrix multiplication to make it
true—see 3.43 and the material preceding it. Now we are just being more explicit
about the bases involved.
3.81 matrix of product of linear maps
Suppose 𝑇 ∈ ℒ(𝑈, 𝑉) and 𝑆 ∈ ℒ(𝑉, 𝑊). If 𝑢1 , …, 𝑢𝑚 is a basis of 𝑈, 𝑣1 , …, 𝑣𝑛

is a basis of 𝑉, and 𝑤1 , …, 𝑤𝑝 is a basis of 𝑊, then
ℳ(𝑆𝑇, (𝑢1 , …, 𝑢𝑚 ), (𝑤1 , …, 𝑤𝑝 )) =

ℳ(𝑆, (𝑣1 , …, 𝑣𝑛 ), (𝑤1 , …, 𝑤𝑝 ))ℳ(𝑇, (𝑢1 , …, 𝑢𝑚 ), (𝑣1 , …, 𝑣𝑛 )).
The next result deals with the matrix of the identity operator 𝐼 with respect
to two different bases. Note that the 𝑘 th column of ℳ(𝐼, (𝑢1 , …, 𝑢𝑛 ), (𝑣1 , …, 𝑣𝑛 ))
consists of the scalars needed to write 𝑢𝑘 as a linear combination of the basis
𝑣1 , …, 𝑣𝑛 .
In the statement of the next result, 𝐼 denotes the identity operator from 𝑉 to 𝑉.
In the proof, 𝐼 also denotes the 𝑛-by-𝑛 identity matrix.
3.82 matrix of identity operator with respect to two bases
Suppose that 𝑢1 , …, 𝑢𝑛 and 𝑣1 , …, 𝑣𝑛 are bases of 𝑉. Then the matrices
ℳ(𝐼, (𝑢1 , …, 𝑢𝑛 ), (𝑣1 , …, 𝑣𝑛 )) and ℳ(𝐼, (𝑣1 , …, 𝑣𝑛 ), (𝑢1 , …, 𝑢𝑛 ))
are invertible, and each is the inverse of the other.
Proof In 3.81, replace 𝑤𝑘 with 𝑢𝑘 , and replace 𝑆 and 𝑇 with 𝐼, getting

𝐼 = ℳ(𝐼, (𝑣1 , …, 𝑣𝑛 ), (𝑢1 , …, 𝑢𝑛 ))ℳ(𝐼, (𝑢1 , …, 𝑢𝑛 ), (𝑣1 , …, 𝑣𝑛 )).
Now interchange the roles of the 𝑢’s and 𝑣’s, getting
𝐼 = ℳ(𝐼, (𝑢1 , …, 𝑢𝑛 ), (𝑣1 , …, 𝑣𝑛 ))ℳ(𝐼, (𝑣1 , …, 𝑣𝑛 ), (𝑢1 , …, 𝑢𝑛 )).
These two equations above give the desired result.
3.83 example: matrix of identity on 𝐅2 with respect to two bases

Consider the bases (4, 2), (5, 3) and (1, 0), (0, 1) of 𝐅2. Because 𝐼(4, 2) =
4(1, 0) + 2(0, 1) and 𝐼(5, 3) = 5(1, 0) + 3(0, 1), we have
4 5
ℳ(𝐼, ((4, 2), (5, 3)), ((1, 0), (0, 1))) = ( ).
2 3
The inverse of the matrix above is
3
⎛ − 25 ⎞
⎜ 2 ⎟,
⎝ −1 2 ⎠
as you should verify. Thus 3.82 implies that
3
− 52 ⎞
ℳ(𝐼, ((1, 0), (0, 1)), ((4, 2), (5, 3))) = ⎛
⎜ 2 ⎟.
⎝ −1 2 ⎠
Our next result shows how the matrix of 𝑇 changes when we change bases. In
the next result, we have two different bases of 𝑉, each of which is used as a basis for
the domain space and as a basis for the target space. Recall our shorthand notation
that allows us to display a basis only once when it is used in both capacities:
ℳ(𝑇, (𝑢1 , …, 𝑢𝑛 )) = ℳ(𝑇, (𝑢1 , …, 𝑢𝑛 ), (𝑢1 , …, 𝑢𝑛 )).
3.84 change-of-basis formula
Suppose 𝑇 ∈ ℒ(𝑉). Suppose 𝑢1 , …, 𝑢𝑛 and 𝑣1 , …, 𝑣𝑛 are bases of 𝑉. Let
𝐴 = ℳ(𝑇, (𝑢1 , …, 𝑢𝑛 )) and 𝐵 = ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 ))
and 𝐶 = ℳ(𝐼, (𝑢1 , …, 𝑢𝑛 ), (𝑣1 , …, 𝑣𝑛 )). Then
𝐴 = 𝐶−1 𝐵𝐶.
Proof In 3.81, replace 𝑤𝑘 with 𝑢𝑘 and replace 𝑆 with 𝐼, getting
3.85 𝐴 = 𝐶−1 ℳ(𝑇, (𝑢1 , …, 𝑢𝑛 ), (𝑣1 , …, 𝑣𝑛 )),
where we have used 3.82.

Again use 3.81, this time replacing 𝑤𝑘 with 𝑣𝑘 . Also replace 𝑇 with 𝐼 and
replace 𝑆 with 𝑇, getting
ℳ(𝑇, (𝑢1 , …, 𝑢𝑛 ), (𝑣1 , …, 𝑣𝑛 )) = 𝐵𝐶.
Substituting the equation above into 3.85 gives the equation 𝐴 = 𝐶−1 𝐵𝐶.
The proof of the next result is left as an exercise.
3.86 matrix of inverse equals inverse of matrix
Suppose that 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝑇 ∈ ℒ(𝑉) is invertible. Then

ℳ(𝑇 −1 ) = (ℳ(𝑇)) , where both matrices are with respect to the basis
−1
𝑣1 , …, 𝑣𝑛 .
Exercises 3D
1 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) is invertible. Show that 𝑇 −1 is invertible and

−1
(𝑇 −1 ) = 𝑇.
2 Suppose 𝑇 ∈ ℒ(𝑈, 𝑉) and 𝑆 ∈ ℒ(𝑉, 𝑊) are both invertible linear maps.

Prove that 𝑆𝑇 ∈ ℒ(𝑈, 𝑊) is invertible and that (𝑆𝑇)−1 = 𝑇 −1 𝑆−1.
3 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that the following
are equivalent.
(a) 𝑇 is invertible.
(b) 𝑇𝑣1 , …, 𝑇𝑣𝑛 is a basis of 𝑉 for every basis 𝑣1 , …, 𝑣𝑛 of 𝑉.
(c) 𝑇𝑣1 , …, 𝑇𝑣𝑛 is a basis of 𝑉 for some basis 𝑣1 , …, 𝑣𝑛 of 𝑉.
4 Suppose 𝑉 is finite-dimensional and dim 𝑉 > 1. Prove that the set of

noninvertible linear maps from 𝑉 to itself is not a subspace of ℒ(𝑉).
5 Suppose 𝑉 is finite-dimensional, 𝑈 is a subspace of 𝑉, and 𝑆 ∈ ℒ(𝑈, 𝑉).

Prove that there exists an invertible linear map 𝑇 from 𝑉 to itself such that
𝑇𝑢 = 𝑆𝑢 for every 𝑢 ∈ 𝑈 if and only if 𝑆 is injective.
6 Suppose that 𝑊 is finite-dimensional and 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that

null 𝑆 = null 𝑇 if and only if there exists an invertible 𝐸 ∈ ℒ(𝑊) such that
𝑆 = 𝐸𝑇.
7 Suppose that 𝑉 is finite-dimensional and 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that

range 𝑆 = range 𝑇 if and only if there exists an invertible 𝐸 ∈ ℒ(𝑉) such
that 𝑆 = 𝑇𝐸.
8 Suppose 𝑉 and 𝑊 are finite-dimensional and 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that
there exist invertible 𝐸1 ∈ ℒ(𝑉) and 𝐸2 ∈ ℒ(𝑊) such that 𝑆 = 𝐸2 𝑇𝐸1 if
and only if dim null 𝑆 = dim null 𝑇.
9 Suppose 𝑉 is finite-dimensional and 𝑇 ∶ 𝑉 → 𝑊 is a surjective linear map
of 𝑉 onto 𝑊. Prove that there is a subspace 𝑈 of 𝑉 such that 𝑇|𝑈 is an
isomorphism of 𝑈 onto 𝑊.
Here 𝑇|𝑈 means the function 𝑇 restricted to 𝑈. Thus 𝑇|𝑈 is the function
whose domain is 𝑈, with 𝑇|𝑈 defined by 𝑇|𝑈 (𝑢) = 𝑇𝑢 for every 𝑢 ∈ 𝑈.
10 Suppose 𝑉 and 𝑊 are finite-dimensional and 𝑈 is a subspace of 𝑉. Let

ℰ = {𝑇 ∈ ℒ(𝑉, 𝑊) ∶ 𝑈 ⊆ null 𝑇}.
(a) Show that ℰ is a subspace of ℒ(𝑉, 𝑊).
(b) Find a formula for dim ℰ in terms of dim 𝑉, dim 𝑊, and dim 𝑈.
Hint: Define Φ ∶ ℒ(𝑉, 𝑊) → ℒ(𝑈, 𝑊) by Φ(𝑇) = 𝑇|𝑈 . What is null Φ?
What is range Φ?
11 Suppose 𝑉 is finite-dimensional and 𝑆, 𝑇 ∈ ℒ(𝑉). Prove that

𝑆𝑇 is invertible ⟺ 𝑆 and 𝑇 are invertible.
12 Suppose 𝑉 is finite-dimensional and 𝑆, 𝑇, 𝑈 ∈ ℒ(𝑉) and 𝑆𝑇𝑈 = 𝐼. Show
that 𝑇 is invertible and that 𝑇 −1 = 𝑈𝑆.
13 Show that the result in Exercise 12 can fail without the hypothesis that 𝑉 is
finite-dimensional.
14 Prove or give a counterexample: If 𝑉 is a finite-dimensional vector space
and 𝑅, 𝑆, 𝑇 ∈ ℒ(𝑉) are such that 𝑅𝑆𝑇 is surjective, then 𝑆 is injective.
15 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑣1 , …, 𝑣𝑚 is a list in 𝑉 such that 𝑇𝑣1 , …, 𝑇𝑣𝑚 spans 𝑉.
Prove that 𝑣1 , …, 𝑣𝑚 spans 𝑉.
16 Prove that every linear map from 𝐅𝑛, 1 to 𝐅𝑚, 1 is given by a matrix multipli-
cation. In other words, prove that if 𝑇 ∈ ℒ(𝐅𝑛, 1, 𝐅𝑚, 1 ), then there exists an
𝑚-by-𝑛 matrix 𝐴 such that 𝑇𝑥 = 𝐴𝑥 for every 𝑥 ∈ 𝐅𝑛, 1.
17 Suppose 𝑉 is finite-dimensional and 𝑆 ∈ ℒ(𝑉). Define 𝒜 ∈ ℒ(ℒ(𝑉)) by

𝒜(𝑇) = 𝑆𝑇
for 𝑇 ∈ ℒ(𝑉).
(a) Show that dim null 𝒜 = (dim 𝑉)(dim null 𝑆).
(b) Show that dim range 𝒜 = (dim 𝑉)(dim range 𝑆).
18 Show that 𝑉 and ℒ(𝐅 , 𝑉) are isomorphic vector spaces.
19 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 has the same
matrix with respect to every basis of 𝑉 if and only if 𝑇 is a scalar multiple
of the identity operator.
20 Suppose 𝑞 ∈ 𝒫(𝐑). Prove that there exists a polynomial 𝑝 ∈ 𝒫(𝐑) such
that
𝑞(𝑥) = (𝑥2 + 𝑥)𝑝″(𝑥) + 2𝑥𝑝′(𝑥) + 𝑝(3)
for all 𝑥 ∈ 𝐑 .
21 Suppose 𝑛 is a positive integer and 𝐴𝑗, 𝑘 ∈ 𝐅 for all 𝑗, 𝑘 = 1, …, 𝑛. Prove that
the following are equivalent (note that in both parts below, the number of
equations equals the number of variables).
(a) The trivial solution 𝑥1 = ⋯ = 𝑥𝑛 = 0 is the only solution to the
homogeneous system of equations
𝑛
∑ 𝐴1, 𝑘 𝑥𝑘 = 0
𝑘=1
⋮
𝑛
∑ 𝐴𝑛, 𝑘 𝑥𝑘 = 0.
𝑘=1
(b) For every 𝑐1 , …, 𝑐𝑛 ∈ 𝐅, there exists a solution to the system of equations

𝑛
∑ 𝐴1, 𝑘 𝑥𝑘 = 𝑐1
𝑘=1
⋮
𝑛
∑ 𝐴𝑛, 𝑘 𝑥𝑘 = 𝑐𝑛 .
𝑘=1
22 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Prove that

ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )) is invertible ⟺ 𝑇 is invertible.
23 Suppose that 𝑢1 , …, 𝑢𝑛 and 𝑣1 , …, 𝑣𝑛 are bases of 𝑉. Let 𝑇 ∈ ℒ(𝑉) be such
that 𝑇𝑣𝑘 = 𝑢𝑘 for each 𝑘 = 1, …, 𝑛. Prove that
ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )) = ℳ(𝐼, (𝑢1 , …, 𝑢𝑛 ), (𝑣1 , …, 𝑣𝑛 )).
24 Suppose 𝐴 and 𝐵 are square matrices of the same size and 𝐴𝐵 = 𝐼. Prove
that 𝐵𝐴 = 𝐼.
3E Products and Quotients of Vector Spaces

Products of Vector Spaces
As usual when dealing with more than one vector space, all vector spaces in use
should be over the same field.
3.87 definition: product of vector spaces
Suppose 𝑉1 , …, 𝑉𝑚 are vector spaces over 𝐅.

• The product 𝑉1 × ⋯ × 𝑉𝑚 is defined by
𝑉1 × ⋯ × 𝑉𝑚 = {(𝑣1 , …, 𝑣𝑚 ) ∶ 𝑣1 ∈ 𝑉1 , …, 𝑣𝑚 ∈ 𝑉𝑚 }.
• Addition on 𝑉1 × ⋯ × 𝑉𝑚 is defined by
(𝑢1 , …, 𝑢𝑚 ) + (𝑣1 , …, 𝑣𝑚 ) = (𝑢1 + 𝑣1 , …, 𝑢𝑚 + 𝑣𝑚 ).
• Scalar multiplication on 𝑉1 × ⋯ × 𝑉𝑚 is defined by
𝜆(𝑣1 , …, 𝑣𝑚 ) = (𝜆𝑣1 , …, 𝜆𝑣𝑚 ).
3.88 example: product of the vector spaces 𝒫5 (𝐑) and 𝐑3

Elements of 𝒫5 (𝐑) × 𝐑3 are lists of length two, with the first item in the list
an element of 𝒫5 (𝐑) and the second item in the list an element of 𝐑3.
For example, (5 − 6𝑥 + 4𝑥2, (3, 8, 7)) and (𝑥 + 9𝑥5, (2, 2, 2)) are elements of
𝒫5 (𝐑) × 𝐑3. Their sum is defined by
(5 − 6𝑥 + 4𝑥2, (3, 8, 7)) + (𝑥 + 9𝑥5, (2, 2, 2))

= (5 − 5𝑥 + 4𝑥2 + 9𝑥5, (5, 10, 9)).
Also, 2(5 − 6𝑥 + 4𝑥2, (3, 8, 7)) = (10 − 12𝑥 + 8𝑥2, (6, 16, 14)).
The next result should be interpreted to mean that the product of vector spaces
is a vector space with the operations of addition and scalar multiplication as
defined by 3.87.
3.89 product of vector spaces is a vector space
Suppose 𝑉1 , …, 𝑉𝑚 are vector spaces over 𝐅. Then 𝑉1 × ⋯ × 𝑉𝑚 is a vector

space over 𝐅.
The proof of the result above is left to the reader. Note that the additive identity
of 𝑉1 × ⋯ × 𝑉𝑚 is (0, …, 0), where the 0 in the 𝑘 th slot is the additive identity of 𝑉𝑘 .
The additive inverse of (𝑣1 , …, 𝑣𝑚 ) ∈ 𝑉1 × ⋯ × 𝑉𝑚 is (−𝑣1 , …, −𝑣𝑚 ).
Section 3E Products and Quotients of Vector Spaces 97
3.90 example: 𝐑2 × 𝐑3 ≠ 𝐑5 but 𝐑2 × 𝐑3 is isomorphic to 𝐑5

Elements of the vector space 𝐑2 × 𝐑3 are lists
((𝑥1 , 𝑥2 ), (𝑥3 , 𝑥4 , 𝑥5 )),
where 𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 , 𝑥5 ∈ 𝐑 . Elements of 𝐑5 are lists
(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 , 𝑥5 ),
where 𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 , 𝑥5 ∈ 𝐑 .
Although elements of 𝐑2 × 𝐑3 and 𝐑5 look similar, they are not the same kind
of object. Elements of 𝐑2 × 𝐑3 are lists of length two (with the first item itself a
list of length two and the second item a list of length three), and elements of 𝐑5
are lists of length five. Thus 𝐑2 × 𝐑3 does not equal 𝐑5.
The linear map This isomorphism is so natural that
we should think of it as a relabel-
((𝑥1 , 𝑥2 ), (𝑥3 , 𝑥4 , 𝑥5 )) ↦ (𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 , 𝑥5 )
ing. Some people informally say that
is an isomorphism of the vector space 𝐑2 ×𝐑3 equals 𝐑5, which is not techni-
𝐑2 × 𝐑3 onto the vector space 𝐑5. Thus cally correct but which captures the
these two vector spaces are isomorphic, al- spirit of identification via relabeling.
though they are not equal.
The next example illustrates the idea that we will use in the proof of 3.92.
3.91 example: a basis of 𝒫2 (𝐑) × 𝐑2

Consider this list of length five of elements of 𝒫2 (𝐑) × 𝐑2 :
(1, (0, 0)), (𝑥, (0, 0)), (𝑥2, (0, 0)), (0, (1, 0)), (0, (0, 1)).
The list above is linearly independent and it spans 𝒫2 (𝐑) × 𝐑2. Thus it is a basis
of 𝒫2 (𝐑) × 𝐑2.
3.92 dimension of a product is the sum of dimensions
Suppose 𝑉1 , …, 𝑉𝑚 are finite-dimensional vector spaces. Then 𝑉1 × ⋯ × 𝑉𝑚 is

finite-dimensional and
dim(𝑉1 × ⋯ × 𝑉𝑚 ) = dim 𝑉1 + ⋯ + dim 𝑉𝑚 .
Proof Choose a basis of each 𝑉𝑘 . For each basis vector of each 𝑉𝑘 , consider the
element of 𝑉1 × ⋯ × 𝑉𝑚 that equals the basis vector in the 𝑘 th slot and 0 in the other
slots. The list of all such vectors is linearly independent and spans 𝑉1 × ⋯ × 𝑉𝑚 .
Thus it is a basis of 𝑉1 × ⋯ × 𝑉𝑚 . The length of this basis is dim 𝑉1 + ⋯ + dim 𝑉𝑚 ,
as desired.
In the next result, the map Γ is surjective by the definition of 𝑉1 + ⋯ + 𝑉𝑚 . Thus

the last word in the result below could be changed from “injective” to “invertible”.
3.93 products and direct sums
Suppose that 𝑉1 , …, 𝑉𝑚 are subspaces of 𝑉. Define a linear map

Γ ∶ 𝑉1 × ⋯ × 𝑉𝑚 → 𝑉1 + ⋯ + 𝑉𝑚 by
Γ(𝑣1 , …, 𝑣𝑚 ) = 𝑣1 + ⋯ + 𝑣𝑚 .
Then 𝑉1 + ⋯ + 𝑉𝑚 is a direct sum if and only if Γ is injective.
Proof By 3.15, Γ is injective if and only if the only way to write 0 as a sum
𝑣1 + ⋯ + 𝑣𝑚 , where each 𝑣𝑘 is in 𝑉𝑘 , is by taking each 𝑣𝑘 equal to 0. Thus 1.45
shows that Γ is injective if and only if 𝑉1 + ⋯ + 𝑉𝑚 is a direct sum, as desired.
3.94 a sum is a direct sum if and only if dimensions add up
Suppose 𝑉 is finite-dimensional and 𝑉1 , …, 𝑉𝑚 are subspaces of 𝑉. Then

𝑉1 + ⋯ + 𝑉𝑚 is a direct sum if and only if
dim(𝑉1 + ⋯ + 𝑉𝑚 ) = dim 𝑉1 + ⋯ + dim 𝑉𝑚 .
Proof The map Γ in 3.93 is surjective. Thus by the fundamental theorem of

linear maps (3.21), Γ is injective if and only if
dim(𝑉1 + ⋯ + 𝑉𝑚 ) = dim(𝑉1 × ⋯ × 𝑉𝑚 ).
Combining 3.93 and 3.92 now shows that 𝑉1 + ⋯ + 𝑉𝑚 is a direct sum if and only
if
dim(𝑉1 + ⋯ + 𝑉𝑚 ) = dim 𝑉1 + ⋯ + dim 𝑉𝑚 ,
as desired.
In the special case 𝑚 = 2, an alternative proof that 𝑉1 + 𝑉2 is a direct sum if

and only if dim(𝑉1 + 𝑉2 ) = dim 𝑉1 + dim 𝑉2 can be obtained by combining 1.46
and 2.43.
Quotient Spaces
We begin our approach to quotient spaces by defining the sum of a vector and a
subset.
3.95 notation: 𝑣 + 𝑈
Suppose 𝑣 ∈ 𝑉 and 𝑈 ⊆ 𝑉. Then 𝑣 + 𝑈 is the subset of 𝑉 defined by
𝑣 + 𝑈 = {𝑣 + 𝑢 ∶ 𝑢 ∈ 𝑈}.
3.96 example: sum of a vector and a one-dimensional subspace of 𝐑2

Suppose
𝑈 = {(𝑥, 2𝑥) ∈ 𝐑2 ∶ 𝑥 ∈ 𝐑}.
Hence 𝑈 is the line in 𝐑2 through the origin with

slope 2. Thus
(17, 20) + 𝑈
is the line in 𝐑2 that contains the point (17, 20)

and has slope 2.
Because (17, 20) + 𝑈 is parallel
to the subspace 𝑈.
(10, 20) ∈ 𝑈 and (17, 20) ∈ (17, 20) + 𝑈,
we see that (17, 20) + 𝑈 is obtained by moving 𝑈

to the right by 7 units.
3.97 definition: translate

For 𝑣 ∈ 𝑉 and 𝑈 a subset of 𝑉, the set 𝑣 + 𝑈 is said to be a translate of 𝑈.
3.98 example: translates
• If 𝑈 is the line in 𝐑2 defined by 𝑈 = {(𝑥, 2𝑥) ∈ 𝐑2 ∶ 𝑥 ∈ 𝐑}, then all lines in

𝐑2 with slope 2 are translates of 𝑈. See Example 3.96 above for a drawing of
𝑈 and one of its translates.
• More generally, if 𝑈 is a line in 𝐑2, then the set of all translates of 𝑈 is the set
of all lines in 𝐑2 that are parallel to 𝑈.
• If 𝑈 = {(𝑥, 𝑦, 0) ∈ 𝐑3 ∶ 𝑥, 𝑦 ∈ 𝐑}, then the translates of 𝑈 are the planes in
𝐑3 that are parallel to the 𝑥𝑦-plane 𝑈.
• More generally, if 𝑈 is a plane in 𝐑3, then the set of all translates of 𝑈 is the
set of all planes in 𝐑3 that are parallel to 𝑈 (see, for example, Exercise 7).
3.99 definition: quotient space, 𝑉/𝑈
Suppose 𝑈 is a subspace of 𝑉. Then the quotient space 𝑉/𝑈 is the set of all
translates of 𝑈. Thus
𝑉/𝑈 = {𝑣 + 𝑈 ∶ 𝑣 ∈ 𝑉}.
3.100 example: quotient spaces
• If 𝑈 = {(𝑥, 2𝑥) ∈ 𝐑2 ∶ 𝑥 ∈ 𝐑}, then 𝐑2/𝑈 is the set of all lines in 𝐑2 that have
slope 2.
• If 𝑈 is a line in 𝐑3 containing the origin, then 𝐑3/𝑈 is the set of all lines in 𝐑3
parallel to 𝑈.
• If 𝑈 is a plane in 𝐑3 containing the origin, then 𝐑3/𝑈 is the set of all planes in
𝐑3 parallel to 𝑈.
Our next goal is to make 𝑉/𝑈 into a vector space. To do this, we will need the
next result.
3.101 two translates of a subspace are equal or disjoint
Suppose 𝑈 is a subspace of 𝑉 and 𝑣, 𝑤 ∈ 𝑉. Then
𝑣 − 𝑤 ∈ 𝑈 ⟺ 𝑣 + 𝑈 = 𝑤 + 𝑈 ⟺ (𝑣 + 𝑈) ∩ (𝑤 + 𝑈) ≠ ∅.
Proof First suppose 𝑣 − 𝑤 ∈ 𝑈. If 𝑢 ∈ 𝑈, then
𝑣 + 𝑢 = 𝑤 + ((𝑣 − 𝑤) + 𝑢) ∈ 𝑤 + 𝑈.
Thus 𝑣 + 𝑈 ⊆ 𝑤 + 𝑈. Similarly, 𝑤 + 𝑈 ⊆ 𝑣 + 𝑈. Thus 𝑣 + 𝑈 = 𝑤 + 𝑈, completing

the proof that 𝑣 − 𝑤 ∈ 𝑈 implies 𝑣 + 𝑈 = 𝑤 + 𝑈.
The equation 𝑣 + 𝑈 = 𝑤 + 𝑈 implies that (𝑣 + 𝑈) ∩ (𝑤 + 𝑈) ≠ ∅.
Now suppose (𝑣 + 𝑈) ∩ (𝑤 + 𝑈) ≠ ∅. Thus there exist 𝑢1 , 𝑢2 ∈ 𝑈 such that
𝑣 + 𝑢1 = 𝑤 + 𝑢2 .
Thus 𝑣 − 𝑤 = 𝑢2 − 𝑢1 . Hence 𝑣 − 𝑤 ∈ 𝑈, showing that (𝑣 + 𝑈) ∩ (𝑤 + 𝑈) ≠ ∅

implies 𝑣 − 𝑤 ∈ 𝑈, which completes the proof.
Now we can define addition and scalar multiplication on 𝑉/𝑈.
3.102 definition: addition and scalar multiplication on 𝑉/𝑈
Suppose 𝑈 is a subspace of 𝑉. Then addition and scalar multiplication are

defined on 𝑉/𝑈 by
(𝑣 + 𝑈) + (𝑤 + 𝑈) = (𝑣 + 𝑤) + 𝑈
𝜆(𝑣 + 𝑈) = (𝜆𝑣) + 𝑈
for all 𝑣, 𝑤 ∈ 𝑉 and all 𝜆 ∈ 𝐅.
As part of the proof of the next result, we will show that the definitions above
make sense.
3.103 quotient space is a vector space
Suppose 𝑈 is a subspace of 𝑉. Then 𝑉/𝑈, with the operations of addition and

scalar multiplication as defined above, is a vector space.
Proof The potential problem with the definitions above of addition and scalar
multiplication on 𝑉/𝑈 is that the representation of a translate of 𝑈 is not unique.
Specifically, suppose 𝑣1 , 𝑣2 , 𝑤1 , 𝑤2 ∈ 𝑉 are such that
𝑣1 + 𝑈 = 𝑣2 + 𝑈 and 𝑤1 + 𝑈 = 𝑤2 + 𝑈.
To show that the definition of addition on 𝑉/𝑈 given above makes sense, we must
show that (𝑣1 + 𝑤1 ) + 𝑈 = (𝑣2 + 𝑤2 ) + 𝑈.
By 3.101, we have
𝑣1 − 𝑣 2 ∈ 𝑈 and 𝑤1 − 𝑤2 ∈ 𝑈.
Because 𝑈 is a subspace of 𝑉 and thus is closed under addition, this implies that
(𝑣1 − 𝑣2 ) + (𝑤1 − 𝑤2 ) ∈ 𝑈. Thus (𝑣1 + 𝑤1 ) − (𝑣2 + 𝑤2 ) ∈ 𝑈. Using 3.101 again,
we see that
(𝑣1 + 𝑤1 ) + 𝑈 = (𝑣2 + 𝑤2 ) + 𝑈,
as desired. Thus the definition of addition on 𝑉/𝑈 makes sense.
Similarly, suppose 𝜆 ∈ 𝐅. We are still assuming that 𝑣1 + 𝑈 = 𝑣2 + 𝑈.
Because 𝑈 is a subspace of 𝑉 and thus is closed under scalar multiplication, we
have 𝜆(𝑣1 − 𝑣2 ) ∈ 𝑈. Thus 𝜆𝑣1 − 𝜆𝑣2 ∈ 𝑈. Hence 3.101 implies that
(𝜆𝑣1 ) + 𝑈 = (𝜆𝑣2 ) + 𝑈.
Thus the definition of scalar multiplication on 𝑉/𝑈 makes sense.

Now that addition and scalar multiplication have been defined on 𝑉/𝑈, the
verification that these operations make 𝑉/𝑈 into a vector space is straightforward
and is left to the reader. Note that the additive identity of 𝑉/𝑈 is 0 + 𝑈 (which
equals 𝑈) and that the additive inverse of 𝑣 + 𝑈 is (−𝑣) + 𝑈.
The next concept will lead to a computation of the dimension of 𝑉/𝑈.
3.104 definition: quotient map, 𝜋
Suppose 𝑈 is a subspace of 𝑉. The quotient map 𝜋 ∶ 𝑉 → 𝑉/𝑈 is the linear

map defined by
𝜋(𝑣) = 𝑣 + 𝑈
for each 𝑣 ∈ 𝑉.
The reader should verify that 𝜋 is indeed a linear map. Although 𝜋 depends
on 𝑈 as well as 𝑉, these spaces are left out of the notation because they should be
clear from the context.
3.105 dimension of quotient space
Suppose 𝑉 is finite-dimensional and 𝑈 is a subspace of 𝑉. Then

dim 𝑉/𝑈 = dim 𝑉 − dim 𝑈.
Proof Let 𝜋 denote the quotient map from 𝑉 to 𝑉/𝑈. If 𝑣 ∈ 𝑉, then 𝑣 + 𝑈 = 0 + 𝑈

if and only if 𝑣 ∈ 𝑈 (by 3.101), which implies that null 𝜋 = 𝑈. The definition of
𝜋 implies range 𝜋 = 𝑉/𝑈. The fundamental theorem of linear maps (3.21) now
implies dim 𝑉 = dim 𝑈 + dim 𝑉/𝑈, which gives the desired result.
Each linear map 𝑇 on 𝑉 induces a linear map 𝑇

̃ on 𝑉/(null 𝑇), which we now
define.
̃
3.106 notation: 𝑇
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Define 𝑇

̃ ∶ 𝑉/(null 𝑇) → 𝑊 by
̃ + null 𝑇) = 𝑇𝑣.
𝑇(𝑣
To show that the definition of 𝑇̃ makes sense, suppose 𝑢, 𝑣 ∈ 𝑉 are such that
𝑢 + null 𝑇 = 𝑣 + null 𝑇. By 3.101, we have 𝑢 − 𝑣 ∈ null 𝑇. Thus 𝑇(𝑢 − 𝑣) = 0.
Hence 𝑇𝑢 = 𝑇𝑣. Thus the definition of 𝑇 ̃ indeed makes sense. The routine
verification that 𝑇 is a linear map from 𝑉/(null 𝑇) to 𝑊 is left to the reader.
̃
The next result shows that we can think of 𝑇
̃ as a modified version of 𝑇, with
a domain that produces a one-to-one map.
3.107 null space and range of 𝑇

̃
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Then

(a) 𝑇
̃ ∘ 𝜋 = 𝑇, where 𝜋 is the quotient map of 𝑉 onto 𝑉/(null 𝑇);
(b) 𝑇
̃ is injective;
(c) range 𝑇̃ = range 𝑇;
(d) 𝑉/(null 𝑇) and range 𝑇 are isomorphic vector spaces.
Proof
(a) If 𝑣 ∈ 𝑉, then (𝑇
̃ ∘ 𝜋)(𝑣) = 𝑇(𝜋(𝑣))
̃ ̃ + null 𝑇) = 𝑇𝑣, as desired.
= 𝑇(𝑣
(b) Suppose 𝑣 ∈ 𝑉 and 𝑇(𝑣 ̃ + null 𝑇) = 0. Then 𝑇𝑣 = 0. Thus 𝑣 ∈ null 𝑇.
Hence 3.101 implies that 𝑣 + null 𝑇 = 0 + null 𝑇. This implies that null 𝑇
̃=
{0 + null 𝑇}. Hence 𝑇 is injective, as desired.
̃
(c) The definition of 𝑇

̃ shows that range 𝑇
̃ = range 𝑇.
(d) Now (b) and (c) imply that if we think of 𝑇

̃ as mapping into range 𝑇, then 𝑇
̃
is an isomorphism from 𝑉/(null 𝑇) onto range 𝑇.
Exercises 3E
1 Suppose 𝑇 is a function from 𝑉 to 𝑊. The graph of 𝑇 is the subset of 𝑉× 𝑊

defined by
graph of 𝑇 = {(𝑣, 𝑇𝑣) ∈ 𝑉 × 𝑊 ∶ 𝑣 ∈ 𝑉}.
Prove that 𝑇 is a linear map if and only if the graph of 𝑇 is a subspace of
𝑉 × 𝑊.
Formally, a function 𝑇 from 𝑉 to 𝑊 is a subset 𝑇 of 𝑉 × 𝑊 such that for
each 𝑣 ∈ 𝑉, there exists exactly one element (𝑣, 𝑤) ∈ 𝑇. In other words,
formally a function is what is called above its graph. We do not usually
think of functions in this formal manner. However, if we do become formal,
then this exercise could be rephrased as follows: Prove that a function 𝑇
from 𝑉 to 𝑊 is a linear map if and only if 𝑇 is a subspace of 𝑉 × 𝑊.
2 Suppose that 𝑉1 , …, 𝑉𝑚 are vector spaces such that 𝑉1 × ⋯ × 𝑉𝑚 is finite-

dimensional. Prove that 𝑉𝑘 is finite-dimensional for each 𝑘 = 1, …, 𝑚.
3 Suppose 𝑉1 , …, 𝑉𝑚 are vector spaces. Prove that ℒ(𝑉1 × ⋯ × 𝑉𝑚 , 𝑊) and
ℒ(𝑉1 , 𝑊) × ⋯ × ℒ(𝑉𝑚 , 𝑊) are isomorphic vector spaces.
4 Suppose 𝑊1 , …, 𝑊𝑚 are vector spaces. Prove that ℒ(𝑉, 𝑊1 × ⋯ × 𝑊𝑚 ) and

ℒ(𝑉, 𝑊1 ) × ⋯ × ℒ(𝑉, 𝑊𝑚 ) are isomorphic vector spaces.
5 For 𝑚 a positive integer, define 𝑉 𝑚 by

⏟
𝑉𝑚 = 𝑉 × ⋯ × 𝑉.
𝑚 times
Prove that 𝑉 and ℒ(𝐅 , 𝑉) are isomorphic vector spaces.

𝑚 𝑚
6 Suppose that 𝑣, 𝑥 are vectors in 𝑉 and that 𝑈, 𝑊 are subspaces of 𝑉 such

that 𝑣 + 𝑈 = 𝑥 + 𝑊. Prove that 𝑈 = 𝑊.
7 Let 𝑈 = {(𝑥, 𝑦, 𝑧) ∈ 𝐑3 ∶ 2𝑥 + 3𝑦 + 5𝑧 = 0}. Suppose 𝐴 ⊆ 𝐑3. Prove that
𝐴 is a translate of 𝑈 if and only if there exists 𝑐 ∈ 𝐑 such that
𝐴 = {(𝑥, 𝑦, 𝑧) ∈ 𝐑3 ∶ 2𝑥 + 3𝑦 + 5𝑧 = 𝑐}.
8 (a) Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑐 ∈ 𝑊. Prove that {𝑥 ∈ 𝑉 ∶ 𝑇𝑥 = 𝑐} is

either the empty set or is a translate of null 𝑇.
(b) Explain why the set of solutions to a system of linear equations such as
3.27 is either the empty set or is a translate of some subspace of 𝐅𝑛.
9 Prove that a nonempty subset 𝐴 of 𝑉 is a translate of some subspace of 𝑉 if
and only if 𝜆𝑣 + (1 − 𝜆)𝑤 ∈ 𝐴 for all 𝑣, 𝑤 ∈ 𝐴 and all 𝜆 ∈ 𝐅.
10 Suppose 𝐴1 = 𝑣 + 𝑈1 and 𝐴2 = 𝑤 + 𝑈2 for some 𝑣, 𝑤 ∈ 𝑉 and some
subspaces 𝑈1 , 𝑈2 of 𝑉. Prove that the intersection 𝐴1 ∩ 𝐴2 is either a
translate of some subspace of 𝑉 or is the empty set.
11 Suppose 𝑈 = {(𝑥1 , 𝑥2 , … ) ∈ 𝐅∞ ∶ 𝑥𝑘 ≠ 0 for only finitely many 𝑘}.

(a) Show that 𝑈 is a subspace of 𝐅∞.
(b) Prove that 𝐅∞/𝑈 is infinite-dimensional.
12 Suppose 𝑣1 , …, 𝑣𝑚 ∈ 𝑉. Let
𝐴 = {𝜆1 𝑣1 + ⋯ + 𝜆𝑚 𝑣𝑚 ∶ 𝜆1 , …, 𝜆𝑚 ∈ 𝐅 and 𝜆1 + ⋯ + 𝜆𝑚 = 1}.

(a) Prove that 𝐴 is a translate of some subspace of 𝑉.
(b) Prove that if 𝐵 is a translate of some subspace of 𝑉 and {𝑣1 , …, 𝑣𝑚 } ⊆ 𝐵,
then 𝐴 ⊆ 𝐵.
(c) Prove that 𝐴 is a translate of some subspace of 𝑉 of dimension less
than 𝑚.
13 Suppose 𝑈 is a subspace of 𝑉 such that 𝑉/𝑈 is finite-dimensional. Prove

that 𝑉 is isomorphic to 𝑈 × (𝑉/𝑈).
14 Suppose 𝑈 and 𝑊 are subspaces of 𝑉 and 𝑉 = 𝑈 ⊕ 𝑊. Suppose 𝑤1 , …, 𝑤𝑚
is a basis of 𝑊. Prove that 𝑤1 + 𝑈, …, 𝑤𝑚 + 𝑈 is a basis of 𝑉/𝑈.
15 Suppose 𝑈 is a subspace of 𝑉 and 𝑣1 + 𝑈, …, 𝑣𝑚 + 𝑈 is a basis of 𝑉/𝑈 and
𝑢1 , …, 𝑢𝑛 is a basis of 𝑈. Prove that 𝑣1 , …, 𝑣𝑚 , 𝑢1 , …, 𝑢𝑛 is a basis of 𝑉.
16 Suppose 𝜑 ∈ ℒ(𝑉, 𝐅) and 𝜑 ≠ 0. Prove that dim 𝑉/(null 𝜑) = 1.

17 Suppose 𝑈 is a subspace of 𝑉 such that dim 𝑉/𝑈 = 1. Prove that there exists
𝜑 ∈ ℒ(𝑉, 𝐅) such that null 𝜑 = 𝑈.
18 Suppose that 𝑈 is a subspace of 𝑉 such that 𝑉/𝑈 is finite-dimensional.

(a) Show that if 𝑊 is a finite-dimensional subspace of 𝑉 and 𝑉 = 𝑈 + 𝑊,
then dim 𝑊 ≥ dim 𝑉/𝑈.
(b) Prove that there exists a finite-dimensional subspace 𝑊 of 𝑉 such that
dim 𝑊 = dim 𝑉/𝑈 and 𝑉 = 𝑈 ⊕ 𝑊.
19 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑈 is a subspace of 𝑉. Let 𝜋 denote the quotient

map from 𝑉 onto 𝑉/𝑈. Prove that there exists 𝑆 ∈ ℒ(𝑉/𝑈, 𝑊) such that
𝑇 = 𝑆 ∘ 𝜋 if and only if 𝑈 ⊆ null 𝑇.
Section 3F Duality 105
3F Duality
Dual Space and Dual Map
Linear maps into the scalar field 𝐅 play a special role in linear algebra, and thus
they get a special name.
3.108 definition: linear functional
A linear functional on 𝑉 is a linear map from 𝑉 to 𝐅. In other words, a linear

functional is an element of ℒ(𝑉, 𝐅).
3.109 example: linear functionals
• Define 𝜑 ∶ 𝐑3 → 𝐑 by 𝜑(𝑥, 𝑦, 𝑧) = 4𝑥 − 5𝑦 + 2𝑧. Then 𝜑 is a linear functional

on 𝐑3.
• Fix (𝑐1 , …, 𝑐𝑛 ) ∈ 𝐅𝑛. Define 𝜑 ∶ 𝐅𝑛 → 𝐅 by 𝜑(𝑥1 , …, 𝑥𝑛 ) = 𝑐1 𝑥1 + ⋯ + 𝑐𝑛 𝑥𝑛 .
Then 𝜑 is a linear functional on 𝐅𝑛.
• Define 𝜑 ∶ 𝒫(𝐑) → 𝐑 by
𝜑(𝑝) = 3𝑝″(5) + 7𝑝(4).
Then 𝜑 is a linear functional on 𝒫(𝐑).
• Define 𝜑 ∶ 𝒫(𝐑) → 𝐑 by
1
𝜑(𝑝) = ∫ 𝑝
0
for each 𝑝 ∈ 𝒫(𝐑). Then 𝜑 is a linear functional on 𝒫(𝐑).
The vector space ℒ(𝑉, 𝐅) also gets a special name and special notation.
3.110 definition: dual space, 𝑉 ′
The dual space of 𝑉, denoted by 𝑉 ′, is the vector space of all linear functionals
on 𝑉. In other words, 𝑉 ′ = ℒ(𝑉, 𝐅).
3.111 dim 𝑉 ′ = dim 𝑉

Suppose 𝑉 is finite-dimensional. Then 𝑉 ′ is also finite-dimensional and
dim 𝑉 ′ = dim 𝑉.
Proof By 3.72 we have

dim 𝑉 ′ = dim ℒ(𝑉, 𝐅) = (dim 𝑉)(dim 𝐅) = dim 𝑉,
as desired.
In the following definition, the linear map lemma (3.4) implies that each 𝜑𝑗 is
well defined.
3.112 definition: dual basis

If 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉, then the dual basis of 𝑣1 , …, 𝑣𝑛 is the list 𝜑1 , …, 𝜑𝑛
of elements of 𝑉 ′, where each 𝜑𝑗 is the linear functional on 𝑉 such that
⎧
{1 if 𝑘 = 𝑗,
𝜑𝑗 (𝑣𝑘 ) = ⎨
{
⎩0 if 𝑘 ≠ 𝑗.
3.113 example: the dual basis of the standard basis of 𝐅𝑛

Suppose 𝑛 is a positive integer. For 1 ≤ 𝑗 ≤ 𝑛, define 𝜑𝑗 to be the linear
functional on 𝐅𝑛 that selects the 𝑗th coordinate of a vector in 𝐅𝑛. Thus
𝜑𝑗 (𝑥1 , …, 𝑥𝑛 ) = 𝑥𝑗
for each (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛.

Let 𝑒1 , …, 𝑒𝑛 be the standard basis of 𝐅𝑛. Then
⎧
{1 if 𝑘 = 𝑗,
𝜑𝑗 (𝑒𝑘 ) = ⎨
{
⎩0 if 𝑘 ≠ 𝑗.
Thus 𝜑1 , …, 𝜑𝑛 is the dual basis of the standard basis 𝑒1 , …, 𝑒𝑛 of 𝐅𝑛.
The next result shows that the dual basis of a basis of 𝑉 consists of the linear
functionals on 𝑉 that give the coefficients for expressing a vector in 𝑉 as a linear
combination of the basis vectors.
3.114 dual basis gives coefficients for linear combination
Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝜑1 , …, 𝜑𝑛 is the dual basis. Then
𝑣 = 𝜑1 (𝑣)𝑣1 + ⋯ + 𝜑𝑛 (𝑣)𝑣𝑛
Proof Suppose 𝑣 ∈ 𝑉. Then there exist 𝑐1 , …, 𝑐𝑛 ∈ 𝐅 such that
3.115 𝑣 = 𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 .
If 𝑗 ∈ {1, …, 𝑛}, then applying 𝜑𝑗 to both sides of the equation above gives
𝜑𝑗 (𝑣) = 𝑐𝑗 .
Substituting the values for 𝑐1 , …, 𝑐𝑛 given by the equation above into 3.115 shows
that 𝑣 = 𝜑1 (𝑣)𝑣1 + ⋯ + 𝜑𝑛 (𝑣)𝑣𝑛 .
The next result shows that the dual basis is indeed a basis of the dual space.
Thus the terminology “dual basis” is justified.
3.116 dual basis is a basis of the dual space
Suppose 𝑉 is finite-dimensional. Then the dual basis of a basis of 𝑉 is a basis

of 𝑉 ′.
Proof Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Let 𝜑1 , …, 𝜑𝑛 denote the dual basis.

To show that 𝜑1 , …, 𝜑𝑛 is a linearly independent list of elements of 𝑉 ′, suppose
𝑎1 , …, 𝑎𝑛 ∈ 𝐅 are such that
3.117 𝑎1 𝜑1 + ⋯ + 𝑎𝑛 𝜑𝑛 = 0.
Now
(𝑎1 𝜑1 + ⋯ + 𝑎𝑛 𝜑𝑛 )(𝑣𝑘 ) = 𝑎𝑘
for each 𝑘 = 1, …, 𝑛. Thus 3.117 shows that 𝑎1 = ⋯ = 𝑎𝑛 = 0. Hence 𝜑1 , …, 𝜑𝑛
Because 𝜑1 , …, 𝜑𝑛 is a linearly independent list in 𝑉 ′ whose length equals
dim 𝑉 ′ (by 3.111), we can conclude that 𝜑1 , …, 𝜑𝑛 is a basis of 𝑉 ′ (see 2.38).
In the definition below, note that if 𝑇 is a linear map from 𝑉 to 𝑊 then 𝑇 ′ is a

linear map from 𝑊 ′ to 𝑉 ′.
3.118 definition: dual map, 𝑇 ′
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). The dual map of 𝑇 is the linear map 𝑇 ′ ∈ ℒ(𝑊 ′, 𝑉 ′ )
defined for each 𝜑 ∈ 𝑊 ′ by
𝑇 ′(𝜑) = 𝜑 ∘ 𝑇.
If 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝜑 ∈ 𝑊 ′, then 𝑇 ′(𝜑) is defined above to be the composition

of the linear maps 𝜑 and 𝑇. Thus 𝑇 ′(𝜑) is indeed a linear map from 𝑉 to 𝐅 ; in
other words, 𝑇 ′(𝜑) ∈ 𝑉 ′.
The following two bullet points show that 𝑇 ′ is a linear map from 𝑊 ′ to 𝑉 ′.
• If 𝜑, 𝜓 ∈ 𝑊 ′, then
𝑇 ′(𝜑 + 𝜓) = (𝜑 + 𝜓) ∘ 𝑇 = 𝜑 ∘ 𝑇 + 𝜓 ∘ 𝑇 = 𝑇 ′(𝜑) + 𝑇 ′(𝜓).
• If 𝜆 ∈ 𝐅 and 𝜑 ∈ 𝑊 ′, then
𝑇 ′(𝜆𝜑) = (𝜆𝜑) ∘ 𝑇 = 𝜆(𝜑 ∘ 𝑇) = 𝜆𝑇 ′(𝜑).
The prime notation appears with two unrelated meanings in the next example:
𝐷′ denotes the dual of the linear map 𝐷, and 𝑝′ denotes the derivative of a
polynomial 𝑝.
3.119 example: dual map of the differentiation linear map

Define 𝐷 ∶ 𝒫(𝐑) → 𝒫(𝐑) by 𝐷𝑝 = 𝑝′.
• Suppose 𝜑 is the linear functional on 𝒫(𝐑) defined by 𝜑(𝑝) = 𝑝(3). Then
𝐷′(𝜑) is the linear functional on 𝒫(𝐑) given by
(𝐷′(𝜑))(𝑝) = (𝜑 ∘ 𝐷)(𝑝) = 𝜑(𝐷𝑝) = 𝜑(𝑝′ ) = 𝑝′(3).
Thus 𝐷′(𝜑) is the linear functional on 𝒫(𝐑) taking 𝑝 to 𝑝′(3).
• Suppose 𝜑 is the linear functional on 𝒫(𝐑) defined by 𝜑(𝑝) = ∫0 𝑝. Then
1
𝐷 (𝜑) is the linear functional on 𝒫(𝐑) given by

′
(𝐷′(𝜑))(𝑝) = (𝜑 ∘ 𝐷)(𝑝)
= 𝜑(𝐷𝑝)
= 𝜑(𝑝′ )
1
= ∫ 𝑝′
0
= 𝑝(1) − 𝑝(0).
Thus 𝐷′(𝜑) is the linear functional on 𝒫(𝐑) taking 𝑝 to 𝑝(1) − 𝑝(0).
In the next result, (a) and (b) imply that the function that takes 𝑇 to 𝑇 ′ is a
linear map from ℒ(𝑉, 𝑊) to ℒ(𝑊 ′, 𝑉 ′ ).
In (c) below, note the reversal of order from 𝑆𝑇 on the left to 𝑇 ′ 𝑆′ on the right.
3.120 algebraic properties of dual maps

(a) (𝑆 + 𝑇)′ = 𝑆′ + 𝑇 ′ for all 𝑆 ∈ ℒ(𝑉, 𝑊);
(b) (𝜆𝑇)′ = 𝜆𝑇 ′ for all 𝜆 ∈ 𝐅 ;
(c) (𝑆𝑇)′ = 𝑇 ′ 𝑆′ for all 𝑆 ∈ ℒ(𝑊, 𝑈).
Proof The proofs of (a) and (b) are left to the reader.
To prove (c), suppose 𝜑 ∈ 𝑈 ′. Then
(𝑆𝑇)′(𝜑) = 𝜑 ∘ (𝑆𝑇) = (𝜑 ∘ 𝑆) ∘ 𝑇 = 𝑇 ′(𝜑 ∘ 𝑆) = 𝑇 ′ (𝑆′(𝜑)) = (𝑇 ′ 𝑆′ )(𝜑),
where the first, third, and fourth equal- Some books use the notation 𝑉 ∗ and
ities above hold because of the defini- 𝑇 ∗ for duality instead of 𝑉 ′ and 𝑇 ′.
tion of the dual map, the second equality However, here we reserve the notation
holds because composition of functions 𝑇 ∗ for the adjoint, which will be intro-
is associative, and the last equality fol- duced when we study linear maps on
lows from the definition of composition. inner product spaces in Chapter 7.
The equation above shows that
(𝑆𝑇)′(𝜑) = (𝑇 ′ 𝑆′ )(𝜑) for all 𝜑 ∈ 𝑈 ′.
Thus (𝑆𝑇)′ = 𝑇 ′ 𝑆′.
Null Space and Range of Dual of Linear Map

Our goal in this subsection is to describe null 𝑇 ′ and range 𝑇 ′ in terms of range 𝑇
and null 𝑇. To do this, we will need the next definition.
3.121 definition: annihilator, 𝑈 0
For 𝑈 ⊆ 𝑉, the annihilator of 𝑈, denoted by 𝑈 0, is defined by
𝑈 0 = {𝜑 ∈ 𝑉 ′ ∶ 𝜑(𝑢) = 0 for all 𝑢 ∈ 𝑈}.
3.122 example: element of an annihilator

Suppose 𝑈 is the subspace of 𝒫(𝐑) consisting of polynomial multiples of 𝑥2.
If 𝜑 is the linear functional on 𝒫(𝐑) defined by 𝜑(𝑝) = 𝑝′(0), then 𝜑 ∈ 𝑈 0.
For 𝑈 ⊆ 𝑉, the annihilator 𝑈 0 is a subset of the dual space 𝑉 ′. Thus 𝑈 0

depends on the vector space containing 𝑈, so a notation such as 𝑈𝑉0 would be
more precise. However, the containing vector space will always be clear from the
context, so we will use the simpler notation 𝑈 0.
3.123 example: the annihilator of a two-dimensional subspace of 𝐑5

Let 𝑒1 , 𝑒2 , 𝑒3 , 𝑒4 , 𝑒5 denote the standard basis of 𝐑5 ; let 𝜑1 , 𝜑2 , 𝜑3 , 𝜑4 , 𝜑5 ∈
5 ′
(𝐑 ) denote the dual basis of 𝑒1 , 𝑒2 , 𝑒3 , 𝑒4 , 𝑒5 . Suppose
𝑈 = span(𝑒1 , 𝑒2 ) = {(𝑥1 , 𝑥2 , 0, 0, 0) ∈ 𝐑5 ∶ 𝑥1 , 𝑥2 ∈ 𝐑}.
We want to show that 𝑈 0 = span(𝜑3 , 𝜑4 , 𝜑5 ).

Recall (see 3.113) that 𝜑𝑗 is the linear functional on 𝐑5 that selects the 𝑗th
coordinate: 𝜑𝑗 (𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 , 𝑥5 ) = 𝑥𝑗 .
First suppose 𝜑 ∈ span(𝜑3 , 𝜑4 , 𝜑5 ). Then there exist 𝑐3 , 𝑐4 , 𝑐5 ∈ 𝐑 such that
𝜑 = 𝑐3 𝜑3 + 𝑐4 𝜑4 + 𝑐5 𝜑5 . If (𝑥1 , 𝑥2 , 0, 0, 0) ∈ 𝑈, then
𝜑(𝑥1 , 𝑥2 , 0, 0, 0) = (𝑐3 𝜑3 + 𝑐4 𝜑4 + 𝑐5 𝜑5 )(𝑥1 , 𝑥2 , 0, 0, 0) = 0.
Thus 𝜑 ∈ 𝑈 0. Hence we have shown that span(𝜑3 , 𝜑4 , 𝜑5 ) ⊆ 𝑈 0.

To show the inclusion in the other direction, suppose that 𝜑 ∈ 𝑈 0. Be-
′
cause the dual basis is a basis of (𝐑5 ) , there exist 𝑐1 , 𝑐2 , 𝑐3 , 𝑐4 , 𝑐5 ∈ 𝐑 such that
𝜑 = 𝑐1 𝜑1 + 𝑐2 𝜑2 + 𝑐3 𝜑3 + 𝑐4 𝜑4 + 𝑐5 𝜑5 . Because 𝑒1 ∈ 𝑈 and 𝜑 ∈ 𝑈 0, we have
0 = 𝜑(𝑒1 ) = (𝑐1 𝜑1 + 𝑐2 𝜑2 + 𝑐3 𝜑3 + 𝑐4 𝜑4 + 𝑐5 𝜑5 )(𝑒1 ) = 𝑐1 .
Similarly, 𝑒2 ∈ 𝑈 and thus 𝑐2 = 0. Hence 𝜑 = 𝑐3 𝜑3 + 𝑐4 𝜑4 + 𝑐5 𝜑5 . Thus

𝜑 ∈ span(𝜑3 , 𝜑4 , 𝜑5 ), which shows that 𝑈 0 ⊆ span(𝜑3 , 𝜑4 , 𝜑5 ).
Thus 𝑈 0 = span(𝜑3 , 𝜑4 , 𝜑5 ).
3.124 the annihilator is a subspace
Suppose 𝑈 ⊆ 𝑉. Then 𝑈 0 is a subspace of 𝑉 ′.
Proof Note that 0 ∈ 𝑈 0 (here 0 is the zero linear functional on 𝑉) because the
zero linear functional applied to every vector in 𝑈 is the zero vector in 𝑉.
Suppose 𝜑, 𝜓 ∈ 𝑈 0. Thus 𝜑, 𝜓 ∈ 𝑉 ′ and 𝜑(𝑢) = 𝜓(𝑢) = 0 for every 𝑢 ∈ 𝑈.
If 𝑢 ∈ 𝑈, then
(𝜑 + 𝜓)(𝑢) = 𝜑(𝑢) + 𝜓(𝑢) = 0 + 0 = 0.
Thus 𝜑 + 𝜓 ∈ 𝑈 0.
Similarly, 𝑈 0 is closed under scalar multiplication. Thus 1.34 implies that 𝑈 0
is a subspace of 𝑉 ′.
The next result shows that dim 𝑈 0 is the difference of dim 𝑉 and dim 𝑈. For
example, this shows that if 𝑈 is a two-dimensional subspace of 𝐑5, then 𝑈 0 is a
′
three-dimensional subspace of (𝐑5 ) , as in Example 3.123.
The next result can be proved following the pattern of Example 3.123: choose
a basis 𝑢1 , …, 𝑢𝑚 of 𝑈, extend to a basis 𝑢1 , …, 𝑢𝑚 , …, 𝑢𝑛 of 𝑉, let 𝜑1 , …, 𝜑𝑚 , …, 𝜑𝑛
be the dual basis of 𝑉 ′, and then show that 𝜑𝑚 + 1 , …, 𝜑𝑛 is a basis of 𝑈 0, which
implies the desired result. You should construct the proof just outlined, even
though a slicker proof is presented here.
3.125 dimension of the annihilator
dim 𝑈 0 = dim 𝑉 − dim 𝑈.
Proof Let 𝑖 ∈ ℒ(𝑈, 𝑉) be the inclusion map defined by 𝑖(𝑢) = 𝑢 for each 𝑢 ∈ 𝑈.
Thus 𝑖′ is a linear map from 𝑉 ′ to 𝑈 ′. The fundamental theorem of linear maps
(3.21) applied to 𝑖′ shows that
dim range 𝑖′ + dim null 𝑖′ = dim 𝑉 ′.
However, null 𝑖′ = 𝑈 0 (as can be seen by thinking about the definitions) and
dim 𝑉 ′ = dim 𝑉 (by 3.111), so we can rewrite the equation above as
3.126 dim range 𝑖′ + dim 𝑈 0 = dim 𝑉.
If 𝜑 ∈ 𝑈 ′, then 𝜑 can be extended to a linear functional 𝜓 on 𝑉 (see, for

example, Exercise 13 in Section 3A). The definition of 𝑖′ shows that 𝑖′(𝜓) = 𝜑.
Thus 𝜑 ∈ range 𝑖′, which implies that range 𝑖′ = 𝑈 ′. Hence
dim range 𝑖′ = dim 𝑈 ′ = dim 𝑈,
and then 3.126 becomes the equation dim 𝑈 + dim 𝑈 0 = dim 𝑉, as desired.
The next result can be a useful tool to show that a subspace is as big as
possible—see (a)—or to show that a subspace is as small as possible—see (b).
3.127 condition for the annihilator to equal {0} or the whole space

(a) 𝑈 0 = {0} ⟺ 𝑈 = 𝑉;
(b) 𝑈 0 = 𝑉 ′ ⟺ 𝑈 = {0}.
Proof To prove (a), we have

𝑈 0 = {0} ⟺ dim 𝑈 0 = 0
⟺ dim 𝑈 = dim 𝑉
⟺ 𝑈 = 𝑉,
where the second equivalence follows from 3.125 and the third equivalence follows
from 2.39.
Similarly, to prove (b) we have
𝑈 0 = 𝑉 ′ ⟺ dim 𝑈 0 = dim 𝑉 ′
⟺ dim 𝑈 0 = dim 𝑉
⟺ dim 𝑈 = 0
⟺ 𝑈 = {0},
where one direction of the first equivalence follows from 2.39, the second equiva-
lence follows from 3.111, and the third equivalence follows from 3.125.
The proof of (a) in the next result does not use the hypothesis that 𝑉 and 𝑊
are finite-dimensional.
3.128 the null space of 𝑇 ′
Suppose 𝑉 and 𝑊 are finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Then

(a) null 𝑇 ′ = (range 𝑇)0 ;
(b) dim null 𝑇 ′ = dim null 𝑇 + dim 𝑊 − dim 𝑉.
Proof
(a) First suppose 𝜑 ∈ null 𝑇 ′. Thus 0 = 𝑇 ′(𝜑) = 𝜑 ∘ 𝑇. Hence
0 = (𝜑 ∘ 𝑇)(𝑣) = 𝜑(𝑇𝑣) for every 𝑣 ∈ 𝑉.
Thus 𝜑 ∈ (range 𝑇)0. This implies that null 𝑇 ′ ⊆ (range 𝑇)0.
To prove the inclusion in the opposite direction, now suppose 𝜑 ∈ (range 𝑇)0.
Thus 𝜑(𝑇𝑣) = 0 for every vector 𝑣 ∈ 𝑉. Hence 0 = 𝜑 ∘ 𝑇 = 𝑇 ′(𝜑). In other
words, 𝜑 ∈ null 𝑇 ′, which shows that (range 𝑇)0 ⊆ null 𝑇 ′, completing the
proof of (a).
(b) We have
dim null 𝑇 ′ = dim(range 𝑇)0
= dim 𝑊 − dim range 𝑇
= dim 𝑊 − (dim 𝑉 − dim null 𝑇)
= dim null 𝑇 + dim 𝑊 − dim 𝑉,
where the first equality comes from (a), the second equality comes from
3.125, and the third equality comes from the fundamental theorem of linear
maps (3.21).
The next result can be useful because sometimes it is easier to verify that 𝑇 ′
is injective than to show directly that 𝑇 is surjective.
3.129 𝑇 surjective is equivalent to 𝑇 ′ injective
𝑇 is surjective ⟺ 𝑇 ′ is injective.
Proof We have
𝑇 ∈ ℒ(𝑉, 𝑊) is surjective ⟺ range 𝑇 = 𝑊
⟺ (range 𝑇)0 = {0}
⟺ null 𝑇 ′ = {0}
⟺ 𝑇 ′ is injective,
where the second equivalence comes from 3.127(a) and the third equivalence
comes from 3.128(a).
3.130 the range of 𝑇 ′

(a) dim range 𝑇 ′ = dim range 𝑇;
(b) range 𝑇 ′ = (null 𝑇)0.
Proof
(a) We have
dim range 𝑇 ′ = dim 𝑊 ′ − dim null 𝑇 ′
= dim 𝑊 − dim(range 𝑇)0
= dim range 𝑇,
where the first equality comes from 3.21, the second equality comes from
3.111 and 3.128(a), and the third equality comes from 3.125.
(b) First suppose 𝜑 ∈ range 𝑇 ′. Thus there exists 𝜓 ∈ 𝑊 ′ such that 𝜑 = 𝑇 ′(𝜓).
If 𝑣 ∈ null 𝑇, then
𝜑(𝑣) = (𝑇 ′(𝜓))𝑣 = (𝜓 ∘ 𝑇)(𝑣) = 𝜓(𝑇𝑣) = 𝜓(0) = 0.
Hence 𝜑 ∈ (null 𝑇)0. This implies that range 𝑇 ′ ⊆ (null 𝑇)0.
We will complete the proof by showing that range 𝑇 ′ and (null 𝑇)0 have the
same dimension. To do this, note that
dim range 𝑇 ′ = dim range 𝑇
= dim 𝑉 − dim null 𝑇
= dim(null 𝑇)0,
where the first equality comes from (a), the second equality comes from 3.21,
and the third equality comes from 3.125.
The next result should be compared to 3.129.
3.131 𝑇 injective is equivalent to 𝑇 ′ surjective
𝑇 is injective ⟺ 𝑇 ′ is surjective.
Proof We have
𝑇 is injective ⟺ null 𝑇 = {0}
⟺ (null 𝑇)0 = 𝑉 ′
⟺ range 𝑇 ′ = 𝑉 ′,
where the second equivalence follows from 3.127(b) and the third equivalence
follows from 3.130(b).
Matrix of Dual of Linear Map

The setting for the next result is the assumption that we have a basis 𝑣1 , …, 𝑣𝑛 of
𝑉, along with its dual basis 𝜑1 , …, 𝜑𝑛 of 𝑉 ′. We also have a basis 𝑤1 , …, 𝑤𝑚 of 𝑊,
along with its dual basis 𝜓1 , …, 𝜓𝑚 of 𝑊 ′. Thus ℳ(𝑇) is computed with respect
to the bases just mentioned of 𝑉 and 𝑊, and ℳ(𝑇 ′ ) is computed with respect to
the dual bases just mentioned of 𝑊 ′ and 𝑉 ′. Using these bases gives the following
pretty result.
3.132 matrix of 𝑇 ′ is transpose of matrix of 𝑇

t
ℳ(𝑇 ′ ) = (ℳ(𝑇)) .
Proof Let 𝐴 = ℳ(𝑇) and 𝐶 = ℳ(𝑇 ′ ). Suppose 1 ≤ 𝑗 ≤ 𝑚 and 1 ≤ 𝑘 ≤ 𝑛.

From the definition of ℳ(𝑇 ′ ) we have
𝑛
𝑇 ′(𝜓𝑗 ) = ∑ 𝐶𝑟, 𝑗 𝜑𝑟 .
𝑟=1
The left side of the equation above equals 𝜓𝑗 ∘ 𝑇. Thus applying both sides of the
equation above to 𝑣𝑘 gives
𝑛
(𝜓𝑗 ∘ 𝑇)(𝑣𝑘 ) = ∑ 𝐶𝑟, 𝑗 𝜑𝑟 (𝑣𝑘 )
𝑟=1
= 𝐶 𝑘, 𝑗 .
We also have
(𝜓𝑗 ∘ 𝑇)(𝑣𝑘 ) = 𝜓𝑗 (𝑇𝑣𝑘 )
𝑚
= 𝜓𝑗 ( ∑ 𝐴𝑟, 𝑘 𝑤𝑟 )
𝑟=1
𝑚
= ∑ 𝐴𝑟, 𝑘 𝜓𝑗 (𝑤𝑟 )
𝑟=1
= 𝐴𝑗, 𝑘 .
Comparing the last line of the last two sets of equations, we have 𝐶𝑘, 𝑗 = 𝐴𝑗, 𝑘 .
Thus 𝐶 = 𝐴t. In other words, ℳ(𝑇 ′ ) = (ℳ(𝑇))t, as desired.
Now we use duality to give an alternative proof that the column rank of a
matrix equals the row rank of the matrix. This result was previously proved using
different tools—see 3.57.
3.133 column rank equals row rank
Suppose 𝐴 ∈ 𝐅𝑚, 𝑛. Then the column rank of 𝐴 equals the row rank of 𝐴.
Proof Define 𝑇 ∶ 𝐅𝑛, 1 → 𝐅𝑚, 1 by 𝑇𝑥 = 𝐴𝑥. Thus ℳ(𝑇) = 𝐴, where ℳ(𝑇) is

computed with respect to the standard bases of 𝐅𝑛, 1 and 𝐅𝑚, 1. Now
column rank of 𝐴 = column rank of ℳ(𝑇)
= dim range 𝑇
= dim range 𝑇 ′
= column rank of ℳ(𝑇 ′ )
= column rank of 𝐴t
= row rank of 𝐴,
where the second equality comes from 3.78, the third equality comes from 3.130(a),
the fourth equality comes from 3.78, the fifth equality comes from 3.132, and the
last equality follows from the definitions of row and column rank.
See Exercise 8 in Section 7A for another alternative proof of the result above.
Exercises 3F
1 Explain why each linear functional is surjective or is the zero map.

2 Give three distinct examples of linear functionals on 𝐑[0, 1].
3 Suppose 𝑉 is finite-dimensional and 𝑣 ∈ 𝑉 with 𝑣 ≠ 0. Prove that there
exists 𝜑 ∈ 𝑉 ′ such that 𝜑(𝑣) = 1.
4 Suppose 𝑉 is finite-dimensional and 𝑈 is a subspace of 𝑉 such that 𝑈 ≠ 𝑉.
Prove that there exists 𝜑 ∈ 𝑉 ′ such that 𝜑(𝑢) = 0 for every 𝑢 ∈ 𝑈 but 𝜑 ≠ 0.
5 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑤1 , …, 𝑤𝑚 is a basis of range 𝑇. Hence for each
𝑣 ∈ 𝑉, there exist unique numbers 𝜑1 (𝑣), …, 𝜑𝑚 (𝑣) such that
𝑇𝑣 = 𝜑1 (𝑣)𝑤1 + ⋯ + 𝜑𝑚 (𝑣)𝑤𝑚 ,
thus defining functions 𝜑1 , …, 𝜑𝑚 from 𝑉 to 𝐅. Show that each of the func-

tions 𝜑1 , …, 𝜑𝑚 is a linear functional on 𝑉.
6 Suppose 𝜑, 𝛽 ∈ 𝑉 ′. Prove that null 𝜑 ⊆ null 𝛽 if and only if there exists
𝑐 ∈ 𝐅 such that 𝛽 = 𝑐𝜑.
7 Suppose that 𝑉1 , …, 𝑉𝑚 are vector spaces. Prove that (𝑉1 × ⋯ × 𝑉𝑚 )′ and

𝑉1 ′ × ⋯ × 𝑉𝑚 ′ are isomorphic vector spaces.
8 Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝜑1 , …, 𝜑𝑛 is the dual basis of 𝑉 ′. Define

Γ ∶ 𝑉 → 𝐅𝑛 and Λ ∶ 𝐅𝑛 → 𝑉 by
Γ(𝑣) = (𝜑1 (𝑣), …, 𝜑𝑛 (𝑣)) and Λ(𝑎1 , …, 𝑎𝑛 ) = 𝑎1 𝑣1 + ⋯ + 𝑎𝑛 𝑣𝑛 .

Explain why Γ and Λ are inverses of each other.
9 Suppose 𝑚 is a positive integer. Show that the dual basis of the basis
1, 𝑥, …, 𝑥𝑚 of 𝒫𝑚 (𝐑) is 𝜑0 , 𝜑1 , …, 𝜑𝑚 , where
𝑝(𝑘) (0)
𝜑𝑘 (𝑝) = .
𝑘!
Here 𝑝(𝑘) denotes the 𝑘 th derivative of 𝑝, with the understanding that the 0th
derivative of 𝑝 is 𝑝.
10 Suppose 𝑚 is a positive integer.

(a) Show that 1, 𝑥 − 5, …, (𝑥 − 5)𝑚 is a basis of 𝒫𝑚 (𝐑).
(b) What is the dual basis of the basis in (a)?
11 Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝜑1 , …, 𝜑𝑛 is the corresponding dual

basis of 𝑉 ′. Suppose 𝜓 ∈ 𝑉 ′. Prove that
𝜓 = 𝜓(𝑣1 )𝜑1 + ⋯ + 𝜓(𝑣𝑛 )𝜑𝑛 .
12 Suppose 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊).

(a) Prove that (𝑆 + 𝑇)′ = 𝑆′ + 𝑇 ′.
(b) Prove that (𝜆𝑇)′ = 𝜆𝑇 ′ for all 𝜆 ∈ 𝐅.
This exercise asks you to verify (a) and (b) in 3.120.
13 Show that the dual map of the identity operator on 𝑉 is the identity operator
on 𝑉 ′.
14 Define 𝑇 ∶ 𝐑3 → 𝐑2 by
𝑇(𝑥, 𝑦, 𝑧) = (4𝑥 + 5𝑦 + 6𝑧, 7𝑥 + 8𝑦 + 9𝑧).
Suppose 𝜑1 , 𝜑2 denotes the dual basis of the standard basis of 𝐑2 and

𝜓1 , 𝜓2 , 𝜓3 denotes the dual basis of the standard basis of 𝐑3.
(a) Describe the linear functionals 𝑇 ′(𝜑1 ) and 𝑇 ′(𝜑2 ).
(b) Write 𝑇 ′(𝜑1 ) and 𝑇 ′(𝜑2 ) as linear combinations of 𝜓1 , 𝜓2 , 𝜓3 .
15 Define 𝑇 ∶ 𝒫(𝐑) → 𝒫(𝐑) by

(𝑇𝑝)(𝑥) = 𝑥2 𝑝(𝑥) + 𝑝″(𝑥)
for each 𝑥 ∈ 𝐑 .
(a) Suppose 𝜑 ∈ 𝒫(𝐑)′ is defined by 𝜑(𝑝) = 𝑝′(4). Describe the linear
functional 𝑇 ′(𝜑) on 𝒫(𝐑).
(b) Suppose 𝜑 ∈ 𝒫(𝐑)′ is defined by 𝜑(𝑝) = ∫01 𝑝. Evaluate (𝑇 ′(𝜑))(𝑥3 ).
16 Suppose 𝑊 is finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that

𝑇 ′ = 0 ⟺ 𝑇 = 0.
17 Suppose 𝑉 and 𝑊 are finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that 𝑇 is

invertible if and only if 𝑇 ′ ∈ ℒ(𝑊 ′, 𝑉 ′ ) is invertible.
18 Suppose 𝑉 and 𝑊 are finite-dimensional. Prove that the map that takes
𝑇 ∈ ℒ(𝑉, 𝑊) to 𝑇 ′ ∈ ℒ(𝑊 ′, 𝑉 ′ ) is an isomorphism of ℒ(𝑉, 𝑊) onto
ℒ(𝑊 ′, 𝑉 ′ ).
19 Suppose 𝑈 ⊆ 𝑉. Explain why

𝑈 0 = {𝜑 ∈ 𝑉 ′ ∶ 𝑈 ⊆ null 𝜑}.
20 Suppose 𝑉 is finite-dimensional and 𝑈 is a subspace of 𝑉. Show that

𝑈 = {𝑣 ∈ 𝑉 ∶ 𝜑(𝑣) = 0 for every 𝜑 ∈ 𝑈 0 }.
21 Suppose 𝑉 is finite-dimensional and 𝑈 and 𝑊 are subspaces of 𝑉.

(a) Prove that 𝑊 0 ⊆ 𝑈 0 if and only if 𝑈 ⊆ 𝑊.
(b) Prove that 𝑊 0 = 𝑈 0 if and only if 𝑈 = 𝑊.
22 Suppose 𝑉 is finite-dimensional and 𝑈 and 𝑊 are subspaces of 𝑉.

(a) Show that (𝑈 + 𝑊)0 = 𝑈 0 ∩ 𝑊 0.
(b) Show that (𝑈 ∩ 𝑊)0 = 𝑈 0 + 𝑊 0.
23 Suppose 𝑉 is finite-dimensional and 𝜑1 , …, 𝜑𝑚 ∈ 𝑉 ′. Prove that the follow-
ing three sets are equal to each other.
(a) span(𝜑1 , …, 𝜑𝑚 )
(b) ((null 𝜑1 ) ∩ ⋯ ∩ (null 𝜑𝑚 ))0
(c) {𝜑 ∈ 𝑉 ′ ∶ (null 𝜑1 ) ∩ ⋯ ∩ (null 𝜑𝑚 ) ⊆ null 𝜑}
24 Suppose 𝑉 is finite-dimensional and 𝑣1 , …, 𝑣𝑚 ∈ 𝑉. Define a linear map
Γ ∶ 𝑉 ′ → 𝐅𝑚 by Γ(𝜑) = (𝜑(𝑣1 ), …, 𝜑(𝑣𝑚 )).
(a) Prove that 𝑣1 , …, 𝑣𝑚 spans 𝑉 if and only if Γ is injective.
(b) Prove that 𝑣1 , …, 𝑣𝑚 is linearly independent if and only if Γ is surjective.
25 Suppose 𝑉 is finite-dimensional and 𝜑1 , …, 𝜑𝑚 ∈ 𝑉 ′. Define a linear map
Γ ∶ 𝑉 → 𝐅𝑚 by Γ(𝑣) = (𝜑1 (𝑣), …, 𝜑𝑚 (𝑣)).
(a) Prove that 𝜑1 , …, 𝜑𝑚 spans 𝑉 ′ if and only if Γ is injective.
(b) Prove that 𝜑1 , …, 𝜑𝑚 is linearly independent if and only if Γ is surjective.
26 Suppose 𝑉 is finite-dimensional and Ω is a subspace of 𝑉 ′. Prove that
Ω = {𝑣 ∈ 𝑉 ∶ 𝜑(𝑣) = 0 for every 𝜑 ∈ Ω} .
0
27 Suppose 𝑇 ∈ ℒ(𝒫5 (𝐑)) and null 𝑇 ′ = span(𝜑), where 𝜑 is the linear

functional on 𝒫5 (𝐑) defined by 𝜑(𝑝) = 𝑝(8). Prove that
range 𝑇 = {𝑝 ∈ 𝒫5 (𝐑) ∶ 𝑝(8) = 0}.
28 Suppose 𝑉 is finite-dimensional and 𝜑1 , …, 𝜑𝑚 is a linearly independent list
in 𝑉 ′. Prove that
dim((null 𝜑1 ) ∩ ⋯ ∩ (null 𝜑𝑚 )) = (dim 𝑉) − 𝑚.
29 Suppose 𝑉 and 𝑊 are finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊).
(a) Prove that if 𝜑 ∈ 𝑊 ′ and null 𝑇 ′ = span(𝜑), then range 𝑇 = null 𝜑.
(b) Prove that if 𝜓 ∈ 𝑉 ′ and range 𝑇 ′ = span(𝜓), then null 𝑇 = null 𝜓.
30 Suppose 𝑉 is finite-dimensional and 𝜑1 , …, 𝜑𝑛 is a basis of 𝑉 ′. Show that
there exists a basis of 𝑉 whose dual basis is 𝜑1 , …, 𝜑𝑛 .
31 Suppose 𝑈 is a subspace of 𝑉. Let 𝑖 ∶ 𝑈 → 𝑉 be the inclusion map defined
by 𝑖(𝑢) = 𝑢. Thus 𝑖′ ∈ ℒ(𝑉 ′, 𝑈 ′ ).
(a) Show that null 𝑖′ = 𝑈 0.
(b) Prove that if 𝑉 is finite-dimensional, then range 𝑖′ = 𝑈 ′.
(c) Prove that if 𝑉 is finite-dimensional, then 𝑖′̃ is an isomorphism from
𝑉 ′/𝑈 0 onto 𝑈 ′.
The isomorphism in (c) is natural in that it does not depend on a choice of
basis in either vector space.
32 The double dual space of 𝑉, denoted by 𝑉 ″, is defined to be the dual space

of 𝑉 ′. In other words, 𝑉 ″ = (𝑉 ′ )′. Define Λ ∶ 𝑉 → 𝑉 ″ by
(Λ𝑣)(𝜑) = 𝜑(𝑣)
for each 𝑣 ∈ 𝑉 and each 𝜑 ∈ 𝑉 ′.

(a) Show that Λ is a linear map from 𝑉 to 𝑉 ″.
(b) Show that if 𝑇 ∈ ℒ(𝑉), then 𝑇 ″ ∘ Λ = Λ ∘ 𝑇, where 𝑇 ″ = (𝑇 ′ )′.
(c) Show that if 𝑉 is finite-dimensional, then Λ is an isomorphism from 𝑉
onto 𝑉 ″.
Suppose 𝑉 is finite-dimensional. Then 𝑉 and 𝑉 ′ are isomorphic, but finding
an isomorphism from 𝑉 onto 𝑉 ′ generally requires choosing a basis of 𝑉.
In contrast, the isomorphism Λ from 𝑉 onto 𝑉 ″ does not require a choice
of basis and thus is considered more natural.
33 Suppose 𝑈 is a subspace of 𝑉. Let 𝜋 ∶ 𝑉 → 𝑉/𝑈 be the usual quotient map.

Thus 𝜋 ′ ∈ ℒ((𝑉/𝑈)′, 𝑉 ′ ).
(a) Show that 𝜋 ′ is injective.
(b) Show that range 𝜋 ′ = 𝑈 0.
(c) Conclude that 𝜋 ′ is an isomorphism from (𝑉/𝑈)′ onto 𝑈 0.
The isomorphism in (c) is natural in that it does not depend on a choice of
basis in either vector space. In fact, there is no assumption here that any of
these vector spaces are finite-dimensional.
Chapter 4
Polynomials
This chapter contains material on polynomials that we will use to investigate

linear maps from a vector space to itself. Many results in this chapter will already
be familiar to you from other courses; they are included here for completeness.
Because this chapter is not about linear algebra, your instructor may go through
it rapidly. You may not be asked to scrutinize all the proofs. Make sure, however,
that you at least read and understand the statements of all results in this chapter—
they will be used in later chapters.
This chapter begins with a brief discussion of algebraic properties of the
complex numbers. Then we prove that a nonconstant polynomial cannot have
more zeros than its degree. We also give a linear-algebra-based proof of the
division algorithm for polynomials, which is worth reading even if you are already
familiar with a proof that does not use linear algebra.
As we will see, the fundamental theorem of algebra leads to a factorization of
every polynomial into degree-one factors if the scalar field is 𝐂 or to factors of
degree at most two if the scalar field is 𝐑 .
standing assumption for this chapter
Alireza Javaheri CC BY
Statue of mathematician and poet Omar Khayyam (1048–1131), whose algebra

book written in 1070 contained the first serious study of cubic polynomials.
119
https://doi.org/10.1007/978-3-031-41026-0_4
120 Chapter 4 Polynomials
Before discussing polynomials with complex or real coefficients, we need to

learn a bit more about the complex numbers.
4.1 definition: real part, Re 𝑧, imaginary part, Im 𝑧
Suppose 𝑧 = 𝑎 + 𝑏𝑖, where 𝑎 and 𝑏 are real numbers.

• The real part of 𝑧, denoted by Re 𝑧, is defined by Re 𝑧 = 𝑎.
• The imaginary part of 𝑧, denoted by Im 𝑧, is defined by Im 𝑧 = 𝑏.
Thus for every complex number 𝑧, we have
𝑧 = Re 𝑧 + (Im 𝑧)𝑖.
4.2 definition: complex conjugate, 𝑧, absolute value, |𝑧|
Suppose 𝑧 ∈ 𝐂 .
• The complex conjugate of 𝑧 ∈ 𝐂 , denoted by 𝑧, is defined by
𝑧 = Re 𝑧 − (Im 𝑧)𝑖.
• The absolute value of a complex number 𝑧, denoted by |𝑧|, is defined by
|𝑧| = √(Re 𝑧)2 + (Im 𝑧)2 .
4.3 example: real and imaginary part, complex conjugate, absolute value
Suppose 𝑧 = 3 + 2𝑖. Then
• Re 𝑧 = 3 and Im 𝑧 = 2;
• 𝑧 = 3 − 2𝑖;
• |𝑧| = √32 + 22 = √13.
Identifying a complex number 𝑧 ∈ 𝐂 with the ordered pair (Re 𝑧, Im 𝑧) ∈ 𝐑2

identifies 𝐂 with 𝐑2. Note that 𝐂 is a one-dimensional complex vector space,
but we can also think of 𝐂 (identified with 𝐑2 ) as a two-dimensional real vector
space.
The absolute value of each complex number is a nonnegative number. Specif-
ically, if 𝑧 ∈ 𝐂 , then |𝑧| equals the distance from the origin in 𝐑2 to the point
(Re 𝑧, Im 𝑧) ∈ 𝐑2.
The real and imaginary parts, com- You should verify that 𝑧 = 𝑧 if and only
plex conjugate, and absolute value have if 𝑧 is a real number.
the properties listed in the following
multipart result.
Chapter 4 Polynomials 121
4.4 properties of complex numbers
Suppose 𝑤, 𝑧 ∈ 𝐂 . Then the following equalities and inequalities hold.

sum of 𝑧 and 𝑧
𝑧 + 𝑧 = 2 Re 𝑧.
difference of 𝑧 and 𝑧
𝑧 − 𝑧 = 2(Im 𝑧)𝑖.
product of 𝑧 and 𝑧
𝑧𝑧 = |𝑧|2.
additivity and multiplicativity of complex conjugate
𝑤 + 𝑧 = 𝑤 + 𝑧 and 𝑤𝑧 = 𝑤 𝑧.
double complex conjugate
𝑧 = 𝑧.
real and imaginary parts are bounded by |𝑧|
| Re 𝑧| ≤ |𝑧| and | Im 𝑧| ≤ |𝑧|.
absolute value of the complex conjugate
∣𝑧∣ = |𝑧|.
multiplicativity of absolute value
|𝑤𝑧| = |𝑤| |𝑧|.
triangle inequality
|𝑤 + 𝑧| ≤ |𝑤| + |𝑧|.
Proof Except for the last item above, Geometric interpretation of triangle in-
the routine verifications of the assertions equality: The length of each side of a
above are left to the reader. To verify the triangle is less than or equal to the sum
triangle inequality, we have of the lengths of the two other sides.
|𝑤 + 𝑧|2 = (𝑤 + 𝑧)(𝑤 + 𝑧)
= 𝑤𝑤 + 𝑧𝑧 + 𝑤𝑧 + 𝑧𝑤
= |𝑤|2 + |𝑧|2 + 𝑤𝑧 + 𝑤𝑧
= |𝑤|2 + |𝑧|2 + 2 Re(𝑤𝑧)
≤ |𝑤|2 + |𝑧|2 + 2∣𝑤𝑧∣
= |𝑤|2 + |𝑧|2 + 2|𝑤| |𝑧|
= (|𝑤| + |𝑧|)2.
Taking square roots now gives the desired See Exercise 2 for the reverse triangle
inequality |𝑤 + 𝑧| ≤ |𝑤| + |𝑧|. inequality.
Zeros of Polynomials
Recall that a function 𝑝 ∶ 𝐅 → 𝐅 is called a polynomial of degree 𝑚 if there exist
𝑎0 , …, 𝑎𝑚 ∈ 𝐅 with 𝑎𝑚 ≠ 0 such that
𝑝(𝑧) = 𝑎0 + 𝑎1 𝑧 + ⋯ + 𝑎𝑚 𝑧𝑚
for all 𝑧 ∈ 𝐅. A polynomial could have more than one degree if the representation
of 𝑝 in the form above were not unique. Our first task is to show that this cannot
happen.
The solutions to the equation 𝑝(𝑧) = 0 play a crucial role in the study of a
polynomial 𝑝 ∈ 𝒫(𝐅). Thus these solutions have a special name.
4.5 definition: zero of a polynomial
A number 𝜆 ∈ 𝐅 is called a zero (or root) of a polynomial 𝑝 ∈ 𝒫(𝐅) if
𝑝(𝜆) = 0.
The next result is the key tool that we will use to show that the degree of a
polynomial is unique.
4.6 each zero of a polynomial corresponds to a degree-one factor
Suppose 𝑚 is a positive integer and 𝑝 ∈ 𝒫(𝐅) is a polynomial of degree 𝑚.

Suppose 𝜆 ∈ 𝐅. Then 𝑝(𝜆) = 0 if and only if there exists a polynomial
𝑞 ∈ 𝒫(𝐅) of degree 𝑚 − 1 such that
𝑝(𝑧) = (𝑧 − 𝜆)𝑞(𝑧)
for every 𝑧 ∈ 𝐅.
Proof First suppose 𝑝(𝜆) = 0. Let 𝑎0 , 𝑎1 , …, 𝑎𝑚 ∈ 𝐅 be such that

𝑝(𝑧) = 𝑎0 + 𝑎1 𝑧 + ⋯ + 𝑎𝑚 𝑧𝑚
for all 𝑧 ∈ 𝐅. Then

4.7 𝑝(𝑧) = 𝑝(𝑧) − 𝑝(𝜆) = 𝑎1 (𝑧 − 𝜆) + ⋯ + 𝑎𝑚 (𝑧𝑚 − 𝜆𝑚 )
for all 𝑧 ∈ 𝐅. For each 𝑘 ∈ {1, …, 𝑚}, the equation

𝑘
𝑧𝑘 − 𝜆𝑘 = (𝑧 − 𝜆) ∑ 𝜆𝑗 − 1 𝑧𝑘 − 𝑗
𝑗=1
shows that 𝑧𝑘 − 𝜆𝑘 equals 𝑧 − 𝜆 times some polynomial of degree 𝑘 − 1. Thus 4.7

shows that 𝑝 equals 𝑧 − 𝜆 times some polynomial of degree 𝑚 − 1, as desired.
To prove the implication in the other direction, now suppose that there is
a polynomial 𝑞 ∈ 𝒫(𝐅) such that 𝑝(𝑧) = (𝑧 − 𝜆)𝑞(𝑧) for every 𝑧 ∈ 𝐅. Then
𝑝(𝜆) = (𝜆 − 𝜆)𝑞(𝜆) = 0, as desired.
Now we can prove that polynomials do not have too many zeros.
4.8 degree 𝑚 implies at most 𝑚 zeros
Suppose 𝑚 is a positive integer and 𝑝 ∈ 𝒫(𝐅) is a polynomial of degree 𝑚.

Then 𝑝 has at most 𝑚 zeros in 𝐅.
Proof We will use induction on 𝑚. The desired result holds if 𝑚 = 1 because

if 𝑎1 ≠ 0 then the polynomial 𝑎0 + 𝑎1 𝑧 has only one zero (which equals −𝑎0 /𝑎1 ).
Thus assume that 𝑚 > 1 and the desired result holds for 𝑚 − 1.
If 𝑝 has no zeros in 𝐅, then the desired result holds and we are done. Thus
suppose 𝑝 has a zero 𝜆 ∈ 𝐅. By 4.6, there is polynomial 𝑞 ∈ 𝒫(𝐅) of degree
𝑚 − 1 such that
𝑝(𝑧) = (𝑧 − 𝜆)𝑞(𝑧)
for every 𝑧 ∈ 𝐅. Our induction hypothesis implies that 𝑞 has at most 𝑚 − 1 zeros
in 𝐅. The equation above shows that the zeros of 𝑝 in 𝐅 are exactly the zeros of 𝑞
in 𝐅 along with 𝜆. Thus 𝑝 has at most 𝑚 zeros in 𝐅.
The result above implies that the coefficients of a polynomial are uniquely
determined (because if a polynomial had two different sets of coefficients, then
subtracting the two representations of the polynomial would give a polynomial
with some nonzero coefficients but infinitely many zeros). In particular, the degree
of a polynomial is uniquely defined.
Recall that the degree of the 0 poly- The 0 polynomial is declared to have
nomial is defined to be −∞. When degree −∞ so that exceptions are not
necessary, use the expected arithmetic needed for various reasonable results
with −∞. For example, −∞ < 𝑚 and such as deg(𝑝𝑞) = deg 𝑝 + deg 𝑞.
−∞ + 𝑚 = −∞ for every integer 𝑚.
Division Algorithm for Polynomials

If 𝑝 and 𝑠 are nonnegative integers, with 𝑠 ≠ 0, then there exist nonnegative
integers 𝑞 and 𝑟 such that
𝑝 = 𝑠𝑞 + 𝑟
and 𝑟 < 𝑠. Think of dividing 𝑝 by 𝑠, getting quotient 𝑞 with remainder 𝑟. Our next
result gives an analogous result for polynomials. Thus the next result is often
called the division algorithm for polynomials, although as stated here it is not
really an algorithm, just a useful result.
The division algorithm for polynomi- Think of the division algorithm for poly-
als could be proved without using any nomials as giving a remainder polyno-
linear algebra. However, as is appropri- mial 𝑟 when the polynomial 𝑝 is divided
ate for a linear algebra textbook, the proof by the polynomial 𝑠.
given here uses linear algebra techniques
and makes nice use of a basis of 𝒫𝑛 (𝐅), which is the (𝑛 + 1)-dimensional vector
space of polynomials with coefficients in 𝐅 and of degree at most 𝑛.
4.9 division algorithm for polynomials
Suppose that 𝑝, 𝑠 ∈ 𝒫(𝐅), with 𝑠 ≠ 0. Then there exist unique polynomials

𝑞, 𝑟 ∈ 𝒫(𝐅) such that
𝑝 = 𝑠𝑞 + 𝑟
and deg 𝑟 < deg 𝑠.
Proof Let 𝑛 = deg 𝑝 and let 𝑚 = deg 𝑠. If 𝑛 < 𝑚, then take 𝑞 = 0 and 𝑟 = 𝑝 to
get the desired equation 𝑝 = 𝑠𝑞 + 𝑟 with deg 𝑟 < deg 𝑠. Thus we now assume that
𝑛 ≥ 𝑚.
The list
4.10 1, 𝑧, …, 𝑧𝑚 − 1 , 𝑠, 𝑧𝑠, …, 𝑧𝑛 − 𝑚 𝑠
is linearly independent in 𝒫𝑛 (𝐅) because each polynomial in this list has a different
degree. Also, the list 4.10 has length 𝑛 + 1, which equals dim 𝒫𝑛 (𝐅). Hence 4.10
is a basis of 𝒫𝑛 (𝐅) [by 2.38].
Because 𝑝 ∈ 𝒫𝑛 (𝐅) and 4.10 is a basis of 𝒫𝑛 (𝐅), there exist unique constants
𝑎0 , 𝑎1 , …, 𝑎𝑚 − 1 ∈ 𝐅 and 𝑏0 , 𝑏1 , …, 𝑏𝑛 − 𝑚 ∈ 𝐅 such that
4.11 𝑝 = 𝑎0 + 𝑎1 𝑧 + ⋯ + 𝑎𝑚 − 1 𝑧𝑚 − 1 + 𝑏0 𝑠 + 𝑏1 𝑧𝑠 + ⋯ + 𝑏𝑛 − 𝑚 𝑧𝑛 − 𝑚 𝑠
0 + 𝑎1 𝑧 + ⋯ + 𝑎𝑚 − 1 𝑧
= 𝑎⏟⏟⏟⏟⏟⏟⏟⏟⏟ 𝑚 − 1 + 𝑠(𝑏 + 𝑏 𝑧 + ⋯ + 𝑏 𝑛 − 𝑚 ).
⏟⏟⏟⏟⏟⏟⏟⏟⏟
0 1 𝑛 − 𝑚𝑧
𝑟 𝑞
With 𝑟 and 𝑞 as defined above, we see that 𝑝 can be written as 𝑝 = 𝑠𝑞 + 𝑟 with

deg 𝑟 < deg 𝑠, as desired.
The uniqueness of 𝑞, 𝑟 ∈ 𝒫(𝐅) satisfying these conditions follows from the
uniqueness of the constants 𝑎0 , 𝑎1 , …, 𝑎𝑚 − 1 ∈ 𝐅 and 𝑏0 , 𝑏1 , …, 𝑏𝑛 − 𝑚 ∈ 𝐅 satisfy-
ing 4.11.
Factorization of Polynomials over 𝐂

W have been handling polynomials with The fundamental theorem of algebra is
complex coefficients and polynomials an existence theorem. Its proof does
with real coefficients simultaneously, let- not lead to a method for finding zeros.
ting 𝐅 denote 𝐑 or 𝐂 . Now we will The quadratic formula gives the zeros
see differences between these two cases. explicitly for polynomials of degree 2.
First we treat polynomials with complex Similar but more complicated formulas
coefficients. Then we will use those re- exist for polynomials of degree 3 and 4.
sults to prove corresponding results for No such formulas exist for polynomials
polynomials with real coefficients. of degree 5 and above.
Our proof of the fundamental theorem
of algebra implicitly uses the result that a continuous real-valued function on a
closed disk in 𝐑2 attains a minimum value. A web search can lead you to several
other proofs of the fundamental theorem of algebra. The proof using Liouville’s
theorem is particularly nice if you are comfortable with analytic functions. All
proofs of the fundamental theorem of algebra need to use some analysis, because
the result is not true if 𝐂 is replaced, for example, with the set of numbers of the
form 𝑐 + 𝑑𝑖 where 𝑐, 𝑑 are rational numbers.
4.12 fundamental theorem of algebra, first version
Every nonconstant polynomial with complex coefficients has a zero in 𝐂 .
Proof De Moivre’s theorem, which you can prove using induction on 𝑘 and the
addition formulas for cosine and sine, states that if 𝑘 is a positive integer and
𝜃 ∈ 𝐑 , then
(cos 𝜃 + 𝑖 sin 𝜃)𝑘 = cos 𝑘𝜃 + 𝑖 sin 𝑘𝜃.
Suppose 𝑤 ∈ 𝐂 and 𝑘 is a positive integer. Using polar coordinates, we know
that there exist 𝑟 ≥ 0 and 𝜃 ∈ 𝐑 such that
𝑟(cos 𝜃 + 𝑖 sin 𝜃) = 𝑤.
De Moivre’s theorem implies that
𝑘
(𝑟1/𝑘 (cos 𝜃𝑘 + 𝑖 sin 𝜃𝑘 )) = 𝑤.
Thus every complex number has a 𝑘 th root, a fact that we will soon use.
Suppose 𝑝 is a nonconstant polynomial with complex coefficients and highest-
order nonzero term 𝑐𝑚 𝑧𝑚. Then |𝑝(𝑧)| → ∞ as |𝑧| → ∞ (because |𝑝(𝑧)|/∣𝑧𝑚 ∣ → |𝑐𝑚 |
as |𝑧| → ∞). Thus the continuous function 𝑧 ↦ |𝑝(𝑧)| has a global minimum at
some point 𝜁 ∈ 𝐂 . To show that 𝑝(𝜁 ) = 0, suppose that 𝑝(𝜁 ) ≠ 0.
Define a new polynomial 𝑞 by
𝑝(𝑧 + 𝜁 )
𝑞(𝑧) = .
𝑝(𝜁 )
The function 𝑧 ↦ |𝑞(𝑧)| has a global minimum value of 1 at 𝑧 = 0. Write
𝑞(𝑧) = 1 + 𝑎𝑘 𝑧𝑘 + ⋯ + 𝑎𝑚 𝑧𝑚,
where 𝑘 is the smallest positive integer such that the coefficient of 𝑧𝑘 is nonzero;
in other words, 𝑎𝑘 ≠ 0.
1
Let 𝛽 ∈ 𝐂 be such that 𝛽𝑘 = − . There is a constant 𝑐 > 1 such that if
𝑎𝑘
𝑡 ∈ (0, 1), then
|𝑞(𝑡𝛽)| ≤ ∣1 + 𝑎𝑘 𝑡𝑘 𝛽𝑘 ∣ + 𝑡𝑘 + 1 𝑐
= 1 − 𝑡𝑘 (1 − 𝑡𝑐).
Thus taking 𝑡 to be 1/(2𝑐) in the inequality above, we have |𝑞(𝑡𝛽)| < 1, which
contradicts the assumption that the global minimum of 𝑧 ↦ |𝑞(𝑧)| is 1. This
contradiction implies that 𝑝(𝜁 ) = 0, showing that 𝑝 has a zero, as desired.
Computers can use clever numerical methods to find good approximations to

the zeros of any polynomial, even when exact zeros cannot be found. For example,
no one will ever give an exact formula for a zero of the polynomial 𝑝 defined by
𝑝(𝑥) = 𝑥5 − 5𝑥4 − 6𝑥3 + 17𝑥2 + 4𝑥 − 7.
However, a computer can find that the zeros of 𝑝 are approximately the five
numbers −1.87, −0.74, 0.62, 1.47, 5.51.
The first version of the fundamental theorem of algebra leads to the following
factorization result for polynomials with complex coefficients. Note that in this
factorization, the zeros of 𝑝 are the numbers 𝜆1 , …, 𝜆𝑚 , which are the only values
of 𝑧 for which the right side of the equation in the next result equals 0.
4.13 fundamental theorem of algebra, second version
If 𝑝 ∈ 𝒫(𝐂) is a nonconstant polynomial, then 𝑝 has a unique factorization

(except for the order of the factors) of the form
𝑝(𝑧) = 𝑐(𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ),
where 𝑐, 𝜆1 , …, 𝜆𝑚 ∈ 𝐂 .
Proof Let 𝑝 ∈ 𝒫(𝐂) and let 𝑚 = deg 𝑝. We will use induction on 𝑚. If 𝑚 = 1,

then the desired factorization exists and is unique. So assume that 𝑚 > 1 and that
the desired factorization exists and is unique for all polynomials of degree 𝑚 − 1.
First we will show that the desired factorization of 𝑝 exists. By the first version
of the fundamental theorem of algebra (4.12), 𝑝 has a zero 𝜆 ∈ 𝐂 . By 4.6, there
is a polynomial 𝑞 of degree 𝑚 − 1 such that
𝑝(𝑧) = (𝑧 − 𝜆)𝑞(𝑧)
for all 𝑧 ∈ 𝐂 . Our induction hypothesis implies that 𝑞 has the desired factorization,
which when plugged into the equation above gives the desired factorization of 𝑝.
Now we turn to the question of uniqueness. The number 𝑐 is uniquely deter-
mined as the coefficient of 𝑧𝑚 in 𝑝. So we only need to show that except for the
order, there is only one way to choose 𝜆1 , …, 𝜆𝑚 . If
(𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ) = (𝑧 − 𝜏1 )⋯(𝑧 − 𝜏𝑚 )
for all 𝑧 ∈ 𝐂 , then because the left side of the equation above equals 0 when
𝑧 = 𝜆1 , one of the 𝜏’s on the right side equals 𝜆1 . Relabeling, we can assume
that 𝜏1 = 𝜆1 . Now if 𝑧 ≠ 𝜆1 , we can divide both sides of the equation above by
𝑧 − 𝜆1 , getting
(𝑧 − 𝜆2 )⋯(𝑧 − 𝜆𝑚 ) = (𝑧 − 𝜏2 )⋯(𝑧 − 𝜏𝑚 )
for all 𝑧 ∈ 𝐂 except possibly 𝑧 = 𝜆1 . Actually the equation above holds for all
𝑧 ∈ 𝐂 , because otherwise by subtracting the right side from the left side we would
get a nonzero polynomial that has infinitely many zeros. The equation above and
our induction hypothesis imply that except for the order, the 𝜆’s are the same as
the 𝜏’s, completing the proof of uniqueness.
Factorization of Polynomials over 𝐑

A polynomial with real coefficients may The failure of the fundamental theorem
have no real zeros. For example, the poly- of algebra for 𝐑 accounts for the differ-
nomial 1 + 𝑥2 has no real zeros. ences between linear algebra on real
To obtain a factorization theorem over and complex vector spaces, as we will
𝐑 , we will use our factorization theorem see in later chapters.
over 𝐂 . We begin with the next result.
4.14 polynomials with real coefficients have nonreal zeros in pairs
Suppose 𝑝 ∈ 𝒫(𝐂) is a polynomial with real coefficients. If 𝜆 ∈ 𝐂 is a zero

of 𝑝, then so is 𝜆.
Proof Let
𝑝(𝑧) = 𝑎0 + 𝑎1 𝑧 + ⋯ + 𝑎𝑚 𝑧𝑚,
where 𝑎0 , …, 𝑎𝑚 are real numbers. Suppose 𝜆 ∈ 𝐂 is a zero of 𝑝. Then
𝑎0 + 𝑎1 𝜆 + ⋯ + 𝑎𝑚 𝜆𝑚 = 0.
Take the complex conjugate of both sides of this equation, obtaining

𝑎0 + 𝑎1 𝜆 + ⋯ + 𝑎𝑚 𝜆𝑚 = 0,
where we have used basic properties of the complex conjugate (see 4.4). The
equation above shows that 𝜆 is a zero of 𝑝.
We want a factorization theorem for Think about the quadratic formula in

polynomials with real coefficients. We connection with the result below.
begin with the following result.
4.15 factorization of a quadratic polynomial
Suppose 𝑏, 𝑐 ∈ 𝐑 . Then there is a polynomial factorization of the form
𝑥2 + 𝑏𝑥 + 𝑐 = (𝑥 − 𝜆1 )(𝑥 − 𝜆2 )
with 𝜆1 , 𝜆2 ∈ 𝐑 if and only if 𝑏2 ≥ 4𝑐.
Proof Notice that

𝑏 2 𝑏2
𝑥2 + 𝑏𝑥 + 𝑐 = (𝑥 + ) + (𝑐 − ).
2 4
First suppose 𝑏2 < 4𝑐. Then the right The equation above is the basis of
side of the equation above is positive for the technique called completing the
every 𝑥 ∈ 𝐑 . Hence the polynomial square.
𝑥2 + 𝑏𝑥 + 𝑐 has no real zeros and thus
cannot be factored in the form (𝑥 − 𝜆1 )(𝑥 − 𝜆2 ) with 𝜆1 , 𝜆2 ∈ 𝐑 .
Conversely, now suppose 𝑏2 ≥ 4𝑐. Then there is a real number 𝑑 such that
2
𝑑 = 𝑏4 − 𝑐. From the displayed equation above, we have
2
𝑏 2
𝑥2 + 𝑏𝑥 + 𝑐 = (𝑥 + ) − 𝑑2
2
𝑏 𝑏
= (𝑥 + + 𝑑)(𝑥 + − 𝑑),
2 2
which gives the desired factorization.
The next result gives a factorization of a polynomial over 𝐑 . The idea of

the proof is to use the second version of the fundamental theorem of algebra
(4.13), which gives a factorization of 𝑝 as a polynomial with complex coefficients.
Complex but nonreal zeros of 𝑝 come in pairs; see 4.14. Thus if the factorization
of 𝑝 as an element of 𝒫(𝐂) includes terms of the form (𝑥 − 𝜆) with 𝜆 a nonreal
complex number, then (𝑥 − 𝜆) is also a term in the factorization. Multiplying
together these two terms, we get
(𝑥2 − 2(Re 𝜆)𝑥 + |𝜆|2 ),
which is a quadratic term of the required form.

The idea sketched in the paragraph above almost provides a proof of the
existence of our desired factorization. However, we need to be careful about
one point. Suppose 𝜆 is a nonreal complex number and (𝑥 − 𝜆) is a term in the
factorization of 𝑝 as an element of 𝒫(𝐂). We are guaranteed by 4.14 that (𝑥 − 𝜆)
also appears as a term in the factorization, but 4.14 does not state that these two
factors appear the same number of times, as needed to make the idea above work.
However, the proof works around this point.
In the next result, either 𝑚 or 𝑀 may equal 0. The numbers 𝜆1 , …, 𝜆𝑚 are
precisely the real zeros of 𝑝, for these are the only real values of 𝑥 for which the
right side of the equation in the next result equals 0.
4.16 factorization of a polynomial over 𝐑
Suppose 𝑝 ∈ 𝒫(𝐑) is a nonconstant polynomial. Then 𝑝 has a unique factor-

ization (except for the order of the factors) of the form
𝑝(𝑥) = 𝑐(𝑥 − 𝜆1 )⋯(𝑥 − 𝜆𝑚 )(𝑥2 + 𝑏1 𝑥 + 𝑐1 )⋯(𝑥2 + 𝑏𝑀 𝑥 + 𝑐𝑀 ),
where 𝑐, 𝜆1 , …, 𝜆𝑚 , 𝑏1 , …, 𝑏𝑀 , 𝑐1 , …, 𝑐𝑀 ∈ 𝐑 , with 𝑏𝑘2 < 4𝑐𝑘 for each 𝑘.
Proof First we will prove that the desired factorization exists, and after that we
will prove the uniqueness.
Think of 𝑝 as an element of 𝒫(𝐂). If all (complex) zeros of 𝑝 are real, then
we have the desired factorization by 4.13. Thus suppose 𝑝 has a zero 𝜆 ∈ 𝐂 with
𝜆 ∉ 𝐑 . By 4.14, 𝜆 is a zero of 𝑝. Thus we can write
𝑝(𝑥) = (𝑥 − 𝜆)(𝑥 − 𝜆)𝑞(𝑥)

= (𝑥2 − 2(Re 𝜆)𝑥 + |𝜆|2 )𝑞(𝑥)
for some polynomial 𝑞 ∈ 𝒫(𝐂) of degree two less than the degree of 𝑝. If we
can prove that 𝑞 has real coefficients, then using induction on the degree of 𝑝
completes the proof of the existence part of this result.
To prove that 𝑞 has real coefficients, we solve the equation above for 𝑞, getting
𝑝(𝑥)
𝑞(𝑥) =
𝑥2 − 2(Re 𝜆)𝑥 + |𝜆|2
for all 𝑥 ∈ 𝐑 . The equation above implies that 𝑞(𝑥) ∈ 𝐑 for all 𝑥 ∈ 𝐑 . Writing
𝑞(𝑥) = 𝑎0 + 𝑎1 𝑥 + ⋯ + 𝑎𝑛 − 2 𝑥𝑛 − 2,
where 𝑛 = deg 𝑝 and 𝑎0 , …, 𝑎𝑛 − 2 ∈ 𝐂 , we thus have
0 = Im 𝑞(𝑥) = (Im 𝑎0 ) + (Im 𝑎1 )𝑥 + ⋯ + (Im 𝑎𝑛 − 2 )𝑥𝑛 − 2
for all 𝑥 ∈ 𝐑 . This implies that Im 𝑎0 , …, Im 𝑎𝑛 − 2 all equal 0 (by 4.8). Thus all
coefficients of 𝑞 are real, as desired. Hence the desired factorization exists.
Now we turn to the question of uniqueness of our factorization. A factor of 𝑝
of the form 𝑥2 + 𝑏𝑘 𝑥 + 𝑐𝑘 with 𝑏𝑘2 < 4𝑐𝑘 can be uniquely written as (𝑥 − 𝜆𝑘 )(𝑥 − 𝜆𝑘 )
with 𝜆𝑘 ∈ 𝐂 . A moment’s thought shows that two different factorizations of 𝑝 as
an element of 𝒫(𝐑) would lead to two different factorizations of 𝑝 as an element
of 𝒫(𝐂), contradicting 4.13.
Exercises 4
1 Suppose 𝑤, 𝑧 ∈ 𝐂 . Verify the following equalities and inequalities.

(a) 𝑧 + 𝑧 = 2 Re 𝑧
(b) 𝑧 − 𝑧 = 2(Im 𝑧)𝑖
(c) 𝑧𝑧 = |𝑧|2
(d) 𝑤 + 𝑧 = 𝑤 + 𝑧 and 𝑤𝑧 = 𝑤 𝑧
(e) 𝑧 = 𝑧
(f) | Re 𝑧| ≤ |𝑧| and | Im 𝑧| ≤ |𝑧|
(g) ∣𝑧∣ = |𝑧|
(h) |𝑤𝑧| = |𝑤| |𝑧|
The results above are the parts of 4.4 that were left to the reader.
2 Prove that if 𝑤, 𝑧 ∈ 𝐂 , then ∣ |𝑤| − |𝑧| ∣ ≤ |𝑤 − 𝑧|.

The inequality above is called the reverse triangle inequality.
3 Suppose 𝑉 is a complex vector space and 𝜑 ∈ 𝑉 ′. Define 𝜎 ∶ 𝑉 → 𝐑 by

𝜎(𝑣) = Re 𝜑(𝑣) for each 𝑣 ∈ 𝑉. Show that
𝜑(𝑣) = 𝜎(𝑣) − 𝑖𝜎(𝑖𝑣)
4 Suppose 𝑚 is a positive integer. Is the set
{0} ∪ {𝑝 ∈ 𝒫(𝐅) ∶ deg 𝑝 = 𝑚}
a subspace of 𝒫(𝐅)?
5 Is the set
{0} ∪ {𝑝 ∈ 𝒫(𝐅) ∶ deg 𝑝 is even}
a subspace of 𝒫(𝐅)?
6 Suppose that 𝑚 and 𝑛 are positive integers with 𝑚 ≤ 𝑛, and suppose
𝜆1 , …, 𝜆𝑚 ∈ 𝐅. Prove that there exists a polynomial 𝑝 ∈ 𝒫(𝐅) with
deg 𝑝 = 𝑛 such that 0 = 𝑝(𝜆1 ) = ⋯ = 𝑝(𝜆𝑚 ) and such that 𝑝 has no
other zeros.
7 Suppose that 𝑚 is a nonnegative integer, 𝑧1 , …, 𝑧𝑚 + 1 are distinct elements
of 𝐅, and 𝑤1 , …, 𝑤𝑚 + 1 ∈ 𝐅. Prove that there exists a unique polynomial
𝑝 ∈ 𝒫𝑚 (𝐅) such that
𝑝(𝑧𝑘 ) = 𝑤𝑘
for each 𝑘 = 1, …, 𝑚 + 1.
This result can be proved without using linear algebra. However, try to find
the clearer, shorter proof that uses some linear algebra.
8 Suppose 𝑝 ∈ 𝒫(𝐂) has degree 𝑚. Prove that 𝑝 has 𝑚 distinct zeros if and
only if 𝑝 and its derivative 𝑝′ have no zeros in common.
9 Prove that every polynomial of odd degree with real coefficients has a real
zero.
10 For 𝑝 ∈ 𝒫(𝐑), define 𝑇𝑝 ∶ 𝐑 → 𝐑 by
⎧ 𝑝(𝑥) − 𝑝(3)
{
{ if 𝑥 ≠ 3,
(𝑇𝑝)(𝑥) = ⎨ 𝑥 − 3
{
{𝑝′(3)
⎩ if 𝑥 = 3
for each 𝑥 ∈ 𝐑 . Show that 𝑇𝑝 ∈ 𝒫(𝐑) for every polynomial 𝑝 ∈ 𝒫(𝐑) and
also show that 𝑇 ∶ 𝒫(𝐑) → 𝒫(𝐑) is a linear map.
11 Suppose 𝑝 ∈ 𝒫(𝐂). Define 𝑞 ∶ 𝐂 → 𝐂 by
𝑞(𝑧) = 𝑝(𝑧) 𝑝(𝑧).
Prove that 𝑞 is a polynomial with real coefficients.
12 Suppose 𝑚 is a nonnegative integer and 𝑝 ∈ 𝒫𝑚 (𝐂) is such that there are

distinct real numbers 𝑥0 , 𝑥1 , …, 𝑥𝑚 with 𝑝(𝑥𝑘 ) ∈ 𝐑 for each 𝑘 = 0, 1, …, 𝑚.
Prove that all coefficients of 𝑝 are real.
13 Suppose 𝑝 ∈ 𝒫(𝐅) with 𝑝 ≠ 0. Let 𝑈 = {𝑝𝑞 ∶ 𝑞 ∈ 𝒫(𝐅)}.
(a) Show that dim 𝒫(𝐅)/𝑈 = deg 𝑝.
(b) Find a basis of 𝒫(𝐅)/𝑈.
14 Suppose 𝑝, 𝑞 ∈ 𝒫(𝐂) are nonconstant polynomials with no zeros in common.

Let 𝑚 = deg 𝑝 and 𝑛 = deg 𝑞. Use linear algebra as outlined below in (a)–(c)
to prove that there exist 𝑟 ∈ 𝒫𝑛 − 1 (𝐂) and 𝑠 ∈ 𝒫𝑚 − 1 (𝐂) such that
𝑟𝑝 + 𝑠𝑞 = 1.
(a) Define 𝑇 ∶ 𝒫𝑛 − 1 (𝐂) × 𝒫𝑚 − 1 (𝐂) → 𝒫𝑚 + 𝑛 − 1 (𝐂) by
𝑇(𝑟, 𝑠) = 𝑟𝑝 + 𝑠𝑞.
Show that the linear map 𝑇 is injective.

(b) Show that the linear map 𝑇 in (a) is surjective.
(c) Use (b) to conclude that there exist 𝑟 ∈ 𝒫𝑛 − 1 (𝐂) and 𝑠 ∈ 𝒫𝑚 − 1 (𝐂)
such that 𝑟𝑝 + 𝑠𝑞 = 1.
Chapter 5
Eigenvalues and Eigenvectors
Linear maps from one vector space to another vector space were the objects of
study in Chapter 3. Now we begin our investigation of operators, which are linear
maps from a vector space to itself. Their study constitutes the most important
part of linear algebra.
To learn about an operator, we might try restricting it to a smaller subspace.
Asking for that restriction to be an operator will lead us to the notion of invariant
subspaces. Each one-dimensional invariant subspace arises from a vector that
the operator maps into a scalar multiple of the vector. This path will lead us to
eigenvectors and eigenvalues.
We will then prove one of the most important results in linear algebra: every
operator on a finite-dimensional nonzero complex vector space has an eigenvalue.
This result will allow us to show that for each operator on a finite-dimensional
complex vector space, there is a basis of the vector space with respect to which
the matrix of the operator has at least almost half its entries equal to 0.
• 𝑉 denotes a vector space over 𝐅.
Hans-Peter Postel CC BY
Statue of Leonardo of Pisa (1170–1250, approximate dates), also known as Fibonacci.

Exercise 21 in Section 5D shows how linear algebra can be used to find
the explicit formula for the Fibonacci sequence shown on the front cover.
132
https://doi.org/10.1007/978-3-031-41026-0_5
Section 5A Invariant Subspaces 133
5A Invariant Subspaces
Eigenvalues
5.1 definition: operator
A linear map from a vector space to itself is called an operator.
Suppose 𝑇 ∈ ℒ(𝑉). If 𝑚 ≥ 2 and Recall that we defined the notation

𝑉 = 𝑉1 ⊕ ⋯ ⊕ 𝑉𝑚 , ℒ(𝑉) to mean ℒ(𝑉, 𝑉).
where each 𝑉𝑘 is a nonzero subspace of 𝑉, then to understand the behavior of
𝑇 we only need to understand the behavior of each 𝑇|𝑉𝑘 ; here 𝑇|𝑉𝑘 denotes the
restriction of 𝑇 to the smaller domain 𝑉𝑘 . Dealing with 𝑇|𝑉𝑘 should be easier than
dealing with 𝑇 because 𝑉𝑘 is a smaller vector space than 𝑉.
However, if we intend to apply tools useful in the study of operators (such
as taking powers), then we have a problem: 𝑇|𝑉𝑘 may not map 𝑉𝑘 into itself; in
other words, 𝑇|𝑉𝑘 may not be an operator on 𝑉𝑘 . Thus we are led to consider only
decompositions of 𝑉 of the form above in which 𝑇 maps each 𝑉𝑘 into itself. Hence
we now give a name to subspaces of 𝑉 that get mapped into themselves by 𝑇.
5.2 definition: invariant subspace
Suppose 𝑇 ∈ ℒ(𝑉). A subspace 𝑈 of 𝑉 is called invariant under 𝑇 if 𝑇𝑢 ∈ 𝑈

for every 𝑢 ∈ 𝑈.
Thus 𝑈 is invariant under 𝑇 if 𝑇|𝑈 is an operator on 𝑈.
5.3 example: subspace invariant under differentiation operator

Suppose that 𝑇 ∈ ℒ(𝒫(𝐑)) is defined by 𝑇𝑝 = 𝑝′. Then 𝒫4 (𝐑), which is a
subspace of 𝒫(𝐑), is invariant under 𝑇 because if 𝑝 ∈ 𝒫(𝐑) has degree at most 4,
then 𝑝′ also has degree at most 4.
5.4 example: four invariant subspaces, not necessarily all different

If 𝑇 ∈ ℒ(𝑉), then the following subspaces of 𝑉 are all invariant under 𝑇.
{0} The subspace {0} is invariant under 𝑇 because if 𝑢 ∈ {0}, then 𝑢 = 0
and hence 𝑇𝑢 = 0 ∈ {0}.
𝑉 The subspace 𝑉 is invariant under 𝑇 because if 𝑢 ∈ 𝑉, then 𝑇𝑢 ∈ 𝑉.
null 𝑇 The subspace null 𝑇 is invariant under 𝑇 because if 𝑢 ∈ null 𝑇, then
𝑇𝑢 = 0, and hence 𝑇𝑢 ∈ null 𝑇.
range 𝑇 The subspace range 𝑇 is invariant under 𝑇 because if 𝑢 ∈ range 𝑇,
then 𝑇𝑢 ∈ range 𝑇.
134 Chapter 5 Eigenvalues and Eigenvectors
Must an operator 𝑇 ∈ ℒ(𝑉) have any invariant subspaces other than {0}
and 𝑉? Later we will see that this question has an affirmative answer if 𝑉 is
finite-dimensional and dim 𝑉 > 1 (for 𝐅 = 𝐂 ) or dim 𝑉 > 2 (for 𝐅 = 𝐑); see
5.19 and Exercise 29 in Section 5B.
The previous example noted that null 𝑇 and range 𝑇 are invariant under 𝑇.
However, these subspaces do not necessarily provide easy answers to the question
above about the existence of invariant subspaces other than {0} and 𝑉, because
null 𝑇 may equal {0} and range 𝑇 may equal 𝑉 (this happens when 𝑇 is invertible).
We will return later to a deeper study of invariant subspaces. Now we turn to
an investigation of the simplest possible nontrivial invariant subspaces—invariant
subspaces of dimension one.
Take any 𝑣 ∈ 𝑉 with 𝑣 ≠ 0 and let 𝑈 equal the set of all scalar multiples of 𝑣:
𝑈 = {𝜆𝑣 ∶ 𝜆 ∈ 𝐅} = span(𝑣).
Then 𝑈 is a one-dimensional subspace of 𝑉 (and every one-dimensional subspace

of 𝑉 is of this form for an appropriate choice of 𝑣). If 𝑈 is invariant under an
operator 𝑇 ∈ ℒ(𝑉), then 𝑇𝑣 ∈ 𝑈, and hence there is a scalar 𝜆 ∈ 𝐅 such that
𝑇𝑣 = 𝜆𝑣.
Conversely, if 𝑇𝑣 = 𝜆𝑣 for some 𝜆 ∈ 𝐅, then span(𝑣) is a one-dimensional

subspace of 𝑉 invariant under 𝑇.
The equation 𝑇𝑣 = 𝜆𝑣, which we have just seen is intimately connected with
one-dimensional invariant subspaces, is important enough that the scalars 𝜆 and
vectors 𝑣 satisfying it are given special names.
5.5 definition: eigenvalue
Suppose 𝑇 ∈ ℒ(𝑉). A number 𝜆 ∈ 𝐅 is called an eigenvalue of 𝑇 if there

exists 𝑣 ∈ 𝑉 such that 𝑣 ≠ 0 and 𝑇𝑣 = 𝜆𝑣.
In the definition above, we require The word eigenvalue is half-German,

that 𝑣 ≠ 0 because every scalar 𝜆 ∈ 𝐅 half-English. The German prefix eigen
satisfies 𝑇0 = 𝜆0. means “own” in the sense of charac-
The comments above show that 𝑉 terizing an intrinsic property.
has a one-dimensional subspace invariant
under 𝑇 if and only if 𝑇 has an eigenvalue.
5.6 example: eigenvalue

Define an operator 𝑇 ∈ ℒ(𝐅3 ) by
𝑇(𝑥, 𝑦, 𝑧) = (7𝑥 + 3𝑧, 3𝑥 + 6𝑦 + 9𝑧, −6𝑦)
for (𝑥, 𝑦, 𝑧) ∈ 𝐅3. Then 𝑇(3, 1, −1) = (18, 6, −6) = 6(3, 1, −1). Thus 6 is an
eigenvalue of 𝑇.
The equivalences in the next result, along with many deep results in linear
algebra, are valid only in the context of finite-dimensional vector spaces.
5.7 equivalent conditions to be an eigenvalue
Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝜆 ∈ 𝐅. Then the following

are equivalent.
(a) 𝜆 is an eigenvalue of 𝑇.
(b) 𝑇 − 𝜆𝐼 is not injective. Reminder: 𝐼 ∈ ℒ(𝑉) is the identity
operator. Thus 𝐼𝑣 = 𝑣 for all 𝑣 ∈ 𝑉.
(c) 𝑇 − 𝜆𝐼 is not surjective.
(d) 𝑇 − 𝜆𝐼 is not invertible.
Proof Conditions (a) and (b) are equivalent because the equation 𝑇𝑣 = 𝜆𝑣
is equivalent to the equation (𝑇 − 𝜆𝐼)𝑣 = 0. Conditions (b), (c), and (d) are
equivalent by 3.65.
5.8 definition: eigenvector
Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆 ∈ 𝐅 is an eigenvalue of 𝑇. A vector 𝑣 ∈ 𝑉 is called

an eigenvector of 𝑇 corresponding to 𝜆 if 𝑣 ≠ 0 and 𝑇𝑣 = 𝜆𝑣.
In other words, a nonzero vector 𝑣 ∈ 𝑉 is an eigenvector of an operator

𝑇 ∈ ℒ(𝑉) if and only if 𝑇𝑣 is a scalar multiple of 𝑣. Because 𝑇𝑣 = 𝜆𝑣 if and only
if (𝑇 − 𝜆𝐼)𝑣 = 0, a vector 𝑣 ∈ 𝑉 with 𝑣 ≠ 0 is an eigenvector of 𝑇 corresponding
to 𝜆 if and only if 𝑣 ∈ null(𝑇 − 𝜆𝐼).
5.9 example: eigenvalues and eigenvectors

Suppose 𝑇 ∈ ℒ(𝐅2 ) is defined by 𝑇(𝑤, 𝑧) = (−𝑧, 𝑤).
(a) First consider the case 𝐅 = 𝐑 . Then 𝑇 is a counterclockwise rotation by 90∘
about the origin in 𝐑2. An operator has an eigenvalue if and only if there
exists a nonzero vector in its domain that gets sent by the operator to a scalar
multiple of itself. A 90∘ counterclockwise rotation of a nonzero vector in 𝐑2
cannot equal a scalar multiple of itself. Conclusion: if 𝐅 = 𝐑 , then 𝑇 has no
eigenvalues (and thus has no eigenvectors).
(b) Now consider the case 𝐅 = 𝐂 . To find eigenvalues of 𝑇, we must find the
scalars 𝜆 such that 𝑇(𝑤, 𝑧) = 𝜆(𝑤, 𝑧) has some solution other than 𝑤 = 𝑧 = 0.
The equation 𝑇(𝑤, 𝑧) = 𝜆(𝑤, 𝑧) is equivalent to the simultaneous equations
5.10 −𝑧 = 𝜆𝑤, 𝑤 = 𝜆𝑧.
Substituting the value for 𝑤 given by the second equation into the first equation
gives
−𝑧 = 𝜆2 𝑧.
Now 𝑧 cannot equal 0 [otherwise 5.10 implies that 𝑤 = 0; we are looking for
solutions to 5.10 such that (𝑤, 𝑧) is not the 0 vector], so the equation above
leads to the equation
−1 = 𝜆2.
The solutions to this equation are 𝜆 = 𝑖 and 𝜆 = −𝑖.
You can verify that 𝑖 and −𝑖 are eigenvalues of 𝑇. Indeed, the eigenvectors
corresponding to the eigenvalue 𝑖 are the vectors of the form (𝑤, −𝑤𝑖), with
𝑤 ∈ 𝐂 and 𝑤 ≠ 0. Furthermore, the eigenvectors corresponding to the
eigenvalue −𝑖 are the vectors of the form (𝑤, 𝑤𝑖), with 𝑤 ∈ 𝐂 and 𝑤 ≠ 0.
In the next proof, we again use the equivalence
𝑇𝑣 = 𝜆𝑣 ⟺ (𝑇 − 𝜆𝐼)𝑣 = 0.
5.11 linearly independent eigenvectors
Suppose 𝑇 ∈ ℒ(𝑉). Then every list of eigenvectors of 𝑇 corresponding to

distinct eigenvalues of 𝑇 is linearly independent.
Proof Suppose the desired result is false. Then there exists a smallest positive
integer 𝑚 such that there exists a linearly dependent list 𝑣1 , …, 𝑣𝑚 of eigenvectors
of 𝑇 corresponding to distinct eigenvalues 𝜆1 , …, 𝜆𝑚 of 𝑇 (note that 𝑚 ≥ 2 because
an eigenvector is, by definition, nonzero). Thus there exist 𝑎1 , …, 𝑎𝑚 ∈ 𝐅, none of
which are 0 (because of the minimality of 𝑚), such that
𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 = 0.
Apply 𝑇 − 𝜆𝑚 𝐼 to both sides of the equation above, getting
𝑎1 (𝜆1 − 𝜆𝑚 )𝑣1 + ⋯ + 𝑎𝑚 − 1 (𝜆𝑚 − 1 − 𝜆𝑚 )𝑣𝑚 − 1 = 0.
Because the eigenvalues 𝜆1 , …, 𝜆𝑚 are distinct, none of the coefficients above
equal 0. Thus 𝑣1 , …, 𝑣𝑚 − 1 is a linearly dependent list of 𝑚 − 1 eigenvectors of 𝑇
corresponding to distinct eigenvalues, contradicting the minimality of 𝑚. This
contradiction completes the proof.
The result above leads to a short proof of the result below, which puts an upper
bound on the number of distinct eigenvalues that an operator can have.
5.12 operator cannot have more eigenvalues than dimension of vector space
Suppose 𝑉 is finite-dimensional. Then each operator on 𝑉 has at most dim 𝑉

distinct eigenvalues.
Proof Let 𝑇 ∈ ℒ(𝑉). Suppose 𝜆1 , …, 𝜆𝑚 are distinct eigenvalues of 𝑇. Let

𝑣1 , …, 𝑣𝑚 be corresponding eigenvectors. Then 5.11 implies that the list 𝑣1 , …, 𝑣𝑚
is linearly independent. Thus 𝑚 ≤ dim 𝑉 (see 2.22), as desired.
Polynomials Applied to Operators

The main reason that a richer theory exists for operators (which map a vector
space into itself) than for more general linear maps is that operators can be raised
to powers. In this subsection we define that notion and the concept of applying a
polynomial to an operator. This concept will be the key tool that we use in the
next section when we prove that every operator on a nonzero finite-dimensional
complex vector space has an eigenvalue.
If 𝑇 is an operator, then 𝑇𝑇 makes sense (see 3.7) and is also an operator on
the same vector space as 𝑇. We usually write 𝑇 2 instead of 𝑇𝑇. More generally,
we have the following definition of 𝑇 𝑚.
5.13 notation: 𝑇 𝑚
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑚 is a positive integer.

• 𝑇 𝑚 ∈ ℒ(𝑉) is defined by 𝑇 𝑚 = 𝑇⋯𝑇
⏟.
𝑚 times
• 𝑇 0 is defined to be the identity operator 𝐼 on 𝑉.
• If 𝑇 is invertible with inverse 𝑇 −1, then 𝑇 −𝑚 ∈ ℒ(𝑉) is defined by

𝑚
𝑇 −𝑚 = (𝑇 −1 ) .
You should verify that if 𝑇 is an operator, then
𝑇𝑚𝑇𝑛 = 𝑇𝑚 + 𝑛 and (𝑇 𝑚 )𝑛 = 𝑇 𝑚𝑛,
where 𝑚 and 𝑛 are arbitrary integers if 𝑇 is invertible and are nonnegative integers
if 𝑇 is not invertible.
Having defined powers of an operator, we can now define what it means to
apply a polynomial to an operator.
5.14 notation: 𝑝(𝑇)
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑝 ∈ 𝒫(𝐅) is a polynomial given by
𝑝(𝑧) = 𝑎0 + 𝑎1 𝑧 + 𝑎2 𝑧2 + ⋯ + 𝑎𝑚 𝑧𝑚
for all 𝑧 ∈ 𝐅. Then 𝑝(𝑇) is the operator on 𝑉 defined by
𝑝(𝑇) = 𝑎0 𝐼 + 𝑎1 𝑇 + 𝑎2 𝑇 2 + ⋯ + 𝑎𝑚 𝑇 𝑚.
This is a new use of the symbol 𝑝 because we are applying 𝑝 to operators, not
just elements of 𝐅. The idea here is that to evaluate 𝑝(𝑇), we simply replace 𝑧 with
𝑇 in the expression defining 𝑝. Note that the constant term 𝑎0 in 𝑝(𝑧) becomes the
operator 𝑎0 𝐼 (which is a reasonable choice because 𝑎0 = 𝑎0 𝑧0 and thus we should
replace 𝑎0 with 𝑎0 𝑇 0, which equals 𝑎0 𝐼).
5.15 example: a polynomial applied to the differentiation operator

Suppose 𝐷 ∈ ℒ(𝒫(𝐑)) is the differentiation operator defined by 𝐷𝑞 = 𝑞′ and
𝑝 is the polynomial defined by 𝑝(𝑥) = 7 − 3𝑥 + 5𝑥2. Then 𝑝(𝐷) = 7𝐼 − 3𝐷 + 5𝐷2.
Thus
(𝑝(𝐷))𝑞 = 7𝑞 − 3𝑞′ + 5𝑞″
for every 𝑞 ∈ 𝒫(𝐑).
If we fix an operator 𝑇 ∈ ℒ(𝑉), then the function from 𝒫(𝐅) to ℒ(𝑉) given
by 𝑝 ↦ 𝑝(𝑇) is linear, as you should verify.
5.16 definition: product of polynomials
If 𝑝, 𝑞 ∈ 𝒫(𝐅), then 𝑝𝑞 ∈ 𝒫(𝐅) is the polynomial defined by
(𝑝𝑞)(𝑧) = 𝑝(𝑧)𝑞(𝑧)
for all 𝑧 ∈ 𝐅.
The order does not matter in taking products of polynomials of a single

operator, as shown by (b) in the next result.
5.17 multiplicative properties
Suppose 𝑝, 𝑞 ∈ 𝒫(𝐅) and 𝑇 ∈ ℒ(𝑉).

Informal proof: When a product of
Then
polynomials is expanded using the dis-
(a) (𝑝𝑞)(𝑇) = 𝑝(𝑇)𝑞(𝑇); tributive property, it does not matter
(b) 𝑝(𝑇)𝑞(𝑇) = 𝑞(𝑇)𝑝(𝑇). whether the symbol is 𝑧 or 𝑇.
Proof
𝑚 𝑛
(a) Suppose 𝑝(𝑧) = ∑ 𝑎𝑗 𝑧𝑗 and 𝑞(𝑧) = ∑ 𝑏𝑘 𝑧𝑘 for all 𝑧 ∈ 𝐅. Then
𝑗=0 𝑘=0
𝑚 𝑛
(𝑝𝑞)(𝑧) = ∑ ∑ 𝑎𝑗 𝑏𝑘 𝑧𝑗 + 𝑘.
𝑗 = 0 𝑘=0
Thus
𝑚 𝑛
(𝑝𝑞)(𝑇) = ∑ ∑ 𝑎𝑗 𝑏𝑘 𝑇 𝑗 + 𝑘
𝑗 = 0 𝑘=0
𝑚 𝑛
= ( ∑ 𝑎𝑗 𝑇 𝑗 )( ∑ 𝑏𝑘 𝑇 𝑘 )
𝑗=0 𝑘=0
= 𝑝(𝑇)𝑞(𝑇).
(b) Using (a) twice, we have 𝑝(𝑇)𝑞(𝑇) = (𝑝𝑞)(𝑇) = (𝑞𝑝)(𝑇) = 𝑞(𝑇)𝑝(𝑇).

We observed earlier that if 𝑇 ∈ ℒ(𝑉), then the subspaces null 𝑇 and range 𝑇
are invariant under 𝑇 (see 5.4). Now we show that the null space and the range of
every polynomial of 𝑇 are also invariant under 𝑇.
5.18 null space and range of 𝑝(𝑇) are invariant under 𝑇
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑝 ∈ 𝒫(𝐅). Then null 𝑝(𝑇) and range 𝑝(𝑇) are
invariant under 𝑇.
Proof Suppose 𝑢 ∈ null 𝑝(𝑇). Then 𝑝(𝑇)𝑢 = 0. Thus

(𝑝(𝑇))(𝑇𝑢) = 𝑇(𝑝(𝑇)𝑢) = 𝑇(0) = 0.
Hence 𝑇𝑢 ∈ null 𝑝(𝑇). Thus null 𝑝(𝑇) is invariant under 𝑇, as desired.

Suppose 𝑢 ∈ range 𝑝(𝑇). Then there exists 𝑣 ∈ 𝑉 such that 𝑢 = 𝑝(𝑇)𝑣. Thus
𝑇𝑢 = 𝑇(𝑝(𝑇)𝑣) = 𝑝(𝑇)(𝑇𝑣).
Hence 𝑇𝑢 ∈ range 𝑝(𝑇). Thus range 𝑝(𝑇) is invariant under 𝑇, as desired.
Exercises 5A
1 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑈 is a subspace of 𝑉.

(a) Prove that if 𝑈 ⊆ null 𝑇, then 𝑈 is invariant under 𝑇.
(b) Prove that if range 𝑇 ⊆ 𝑈, then 𝑈 is invariant under 𝑇.
2 Suppose that 𝑇 ∈ ℒ(𝑉) and 𝑉1 , …, 𝑉𝑚 are subspaces of 𝑉 invariant under 𝑇.

Prove that 𝑉1 + ⋯ + 𝑉𝑚 is invariant under 𝑇.
3 Suppose 𝑇 ∈ ℒ(𝑉). Prove that the intersection of every collection of
subspaces of 𝑉 invariant under 𝑇 is invariant under 𝑇.
4 Prove or give a counterexample: If 𝑉 is finite-dimensional and 𝑈 is a sub-
space of 𝑉 that is invariant under every operator on 𝑉, then 𝑈 = {0} or
𝑈 = 𝑉.
5 Suppose 𝑇 ∈ ℒ(𝐑2 ) is defined by 𝑇(𝑥, 𝑦) = (−3𝑦, 𝑥). Find the eigenvalues

of 𝑇.
6 Define 𝑇 ∈ ℒ(𝐅2 ) by 𝑇(𝑤, 𝑧) = (𝑧, 𝑤). Find all eigenvalues and eigenvec-
tors of 𝑇.
7 Define 𝑇 ∈ ℒ(𝐅3 ) by 𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (2𝑧2 , 0, 5𝑧3 ). Find all eigenvalues and
eigenvectors of 𝑇.
8 Suppose 𝑃 ∈ ℒ(𝑉) is such that 𝑃2 = 𝑃. Prove that if 𝜆 is an eigenvalue of 𝑃,
then 𝜆 = 0 or 𝜆 = 1.
9 Define 𝑇 ∶ 𝒫(𝐑) → 𝒫(𝐑) by 𝑇𝑝 = 𝑝′. Find all eigenvalues and eigenvectors

of 𝑇.
10 Define 𝑇 ∈ ℒ(𝒫4 (𝐑)) by (𝑇𝑝)(𝑥) = 𝑥𝑝′(𝑥) for all 𝑥 ∈ 𝐑 . Find all eigenval-
ues and eigenvectors of 𝑇.
11 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝛼 ∈ 𝐅. Prove that there ex-
ists 𝛿 > 0 such that 𝑇− 𝜆𝐼 is invertible for all 𝜆 ∈ 𝐅 such that 0 < |𝛼 − 𝜆| < 𝛿.
12 Suppose 𝑉 = 𝑈 ⊕ 𝑊, where 𝑈 and 𝑊 are nonzero subspaces of 𝑉. Define
𝑃 ∈ ℒ(𝑉) by 𝑃(𝑢 + 𝑤) = 𝑢 for each 𝑢 ∈ 𝑈 and each 𝑤 ∈ 𝑊. Find all
eigenvalues and eigenvectors of 𝑃.
13 Suppose 𝑇 ∈ ℒ(𝑉). Suppose 𝑆 ∈ ℒ(𝑉) is invertible.
(a) Prove that 𝑇 and 𝑆−1 𝑇𝑆 have the same eigenvalues.
(b) What is the relationship between the eigenvectors of 𝑇 and the eigenvec-
tors of 𝑆−1 𝑇𝑆?
14 Give an example of an operator on 𝐑4 that has no (real) eigenvalues.
15 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝜆 ∈ 𝐅. Show that 𝜆 is
an eigenvalue of 𝑇 if and only if 𝜆 is an eigenvalue of the dual operator
𝑇 ′ ∈ ℒ(𝑉 ′ ).
16 Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝑇 ∈ ℒ(𝑉). Prove that if 𝜆 is an
eigenvalue of 𝑇, then
|𝜆| ≤ 𝑛 max{∣ℳ(𝑇)𝑗, 𝑘 ∣ ∶ 1 ≤ 𝑗, 𝑘 ≤ 𝑛},
where ℳ(𝑇)𝑗, 𝑘 denotes the entry in row 𝑗, column 𝑘 of the matrix of 𝑇 with
respect to the basis 𝑣1 , …, 𝑣𝑛 .
See Exercise 19 in Section 6A for a different bound on |𝜆|.
17 Suppose 𝐅 = 𝐑 , 𝑇 ∈ ℒ(𝑉), and 𝜆 ∈ 𝐑 . Prove that 𝜆 is an eigenvalue of 𝑇

if and only if 𝜆 is an eigenvalue of the complexification 𝑇𝐂 .
See Exercise 33 in Section 3B for the definition of 𝑇𝐂 .
18 Suppose 𝐅 = 𝐑 , 𝑇 ∈ ℒ(𝑉), and 𝜆 ∈ 𝐂 . Prove that 𝜆 is an eigenvalue of

the complexification 𝑇𝐂 if and only if 𝜆 is an eigenvalue of 𝑇𝐂 .
19 Show that the forward shift operator 𝑇 ∈ ℒ(𝐅∞ ) defined by
𝑇(𝑧1 , 𝑧2 , … ) = (0, 𝑧1 , 𝑧2 , … )
has no eigenvalues.
20 Define the backward shift operator 𝑆 ∈ ℒ(𝐅∞ ) by
𝑆(𝑧1 , 𝑧2 , 𝑧3 , … ) = (𝑧2 , 𝑧3 , … ).
(a) Show that every element of 𝐅 is an eigenvalue of 𝑆.
(b) Find all eigenvectors of 𝑆.
Section 5A Invariant Subspaces ≈ 100√2
21 Suppose 𝑇 ∈ ℒ(𝑉) is invertible.

(a) Suppose 𝜆 ∈ 𝐅 with 𝜆 ≠ 0. Prove that 𝜆 is an eigenvalue of 𝑇 if and
only if 1𝜆 is an eigenvalue of 𝑇 −1.
(b) Prove that 𝑇 and 𝑇 −1 have the same eigenvectors.
22 Suppose 𝑇 ∈ ℒ(𝑉) and there exist nonzero vectors 𝑢 and 𝑤 in 𝑉 such that
𝑇𝑢 = 3𝑤 and 𝑇𝑤 = 3𝑢.
Prove that 3 or −3 is an eigenvalue of 𝑇.
23 Suppose 𝑉 is finite-dimensional and 𝑆, 𝑇 ∈ ℒ(𝑉). Prove that 𝑆𝑇 and 𝑇𝑆
have the same eigenvalues.
24 Suppose 𝐴 is an 𝑛-by-𝑛 matrix with entries in 𝐅. Define 𝑇 ∈ ℒ(𝐅𝑛 ) by
𝑇𝑥 = 𝐴𝑥, where elements of 𝐅𝑛 are thought of as 𝑛-by-1 column vectors.
(a) Suppose the sum of the entries in each row of 𝐴 equals 1. Prove that 1
is an eigenvalue of 𝑇.
(b) Suppose the sum of the entries in each column of 𝐴 equals 1. Prove that
1 is an eigenvalue of 𝑇.
25 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑢, 𝑤 are eigenvectors of 𝑇 such that 𝑢 + 𝑤 is also

an eigenvector of 𝑇. Prove that 𝑢 and 𝑤 are eigenvectors of 𝑇 corresponding
to the same eigenvalue.
26 Suppose 𝑇 ∈ ℒ(𝑉) is such that every nonzero vector in 𝑉 is an eigenvector
of 𝑇. Prove that 𝑇 is a scalar multiple of the identity operator.
27 Suppose that 𝑉 is finite-dimensional and 𝑘 ∈ {1, …, dim 𝑉 − 1}. Suppose
𝑇 ∈ ℒ(𝑉) is such that every subspace of 𝑉 of dimension 𝑘 is invariant
under 𝑇. Prove that 𝑇 is a scalar multiple of the identity operator.
28 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 has at most
1 + dim range 𝑇 distinct eigenvalues.
29 Suppose 𝑇 ∈ ℒ(𝐑3 ) and −4, 5, and √7 are eigenvalues of 𝑇. Prove that

there exists 𝑥 ∈ 𝐑3 such that 𝑇𝑥 − 9𝑥 = (−4, 5, √7).
30 Suppose 𝑇 ∈ ℒ(𝑉) and (𝑇 − 2𝐼)(𝑇 − 3𝐼)(𝑇 − 4𝐼) = 0. Suppose 𝜆 is an
eigenvalue of 𝑇. Prove that 𝜆 = 2 or 𝜆 = 3 or 𝜆 = 4.
31 Give an example of 𝑇 ∈ ℒ(𝐑2 ) such that 𝑇 4 = −𝐼.
32 Suppose 𝑇 ∈ ℒ(𝑉) has no eigenvalues and 𝑇 4 = 𝐼. Prove that 𝑇 2 = −𝐼.
33 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑚 is a positive integer.
(a) Prove that 𝑇 is injective if and only if 𝑇 𝑚 is injective.
(b) Prove that 𝑇 is surjective if and only if 𝑇 𝑚 is surjective.
34 Suppose 𝑉 is finite-dimensional and 𝑣1 , …, 𝑣𝑚 ∈ 𝑉. Prove that the list

𝑣1 , …, 𝑣𝑚 is linearly independent if and only if there exists 𝑇 ∈ ℒ(𝑉) such
that 𝑣1 , …, 𝑣𝑚 are eigenvectors of 𝑇 corresponding to distinct eigenvalues.
35 Suppose that 𝜆1 , …, 𝜆𝑛 is a list of distinct real numbers. Prove that the
list 𝑒 𝜆1𝑥 , …, 𝑒 𝜆𝑛𝑥 is linearly independent in the vector space of real-valued
functions on 𝐑 .
Hint: Let 𝑉 = span(𝑒 𝜆1𝑥 , …, 𝑒 𝜆𝑛𝑥 ), and define an operator 𝐷 ∈ ℒ(𝑉) by
𝐷 𝑓 = 𝑓 ′. Find eigenvalues and eigenvectors of 𝐷.
36 Suppose that 𝜆1 , …, 𝜆𝑛 is a list of distinct positive numbers. Prove that

the list cos(𝜆1 𝑥), …, cos(𝜆𝑛 𝑥) is linearly independent in the vector space of
real-valued functions on 𝐑 .
37 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Define 𝒜 ∈ ℒ(ℒ(𝑉)) by
𝒜(𝑆) = 𝑇𝑆
for each 𝑆 ∈ ℒ(𝑉). Prove that the set of eigenvalues of 𝑇 equals the set of
eigenvalues of 𝒜.
38 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝑈 is a subspace of 𝑉
invariant under 𝑇. The quotient operator 𝑇/𝑈 ∈ ℒ(𝑉/𝑈) is defined by
(𝑇/𝑈)(𝑣 + 𝑈) = 𝑇𝑣 + 𝑈
(a) Show that the definition of 𝑇/𝑈 makes sense (which requires using the
condition that 𝑈 is invariant under 𝑇) and show that 𝑇/𝑈 is an operator
on 𝑉/𝑈.
(b) Show that each eigenvalue of 𝑇/𝑈 is an eigenvalue of 𝑇.
39 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 has an eigen-
value if and only if there exists a subspace of 𝑉 of dimension dim 𝑉 − 1 that
is invariant under 𝑇.
40 Suppose 𝑆, 𝑇 ∈ ℒ(𝑉) and 𝑆 is invertible. Suppose 𝑝 ∈ 𝒫(𝐅) is a polynomial.
Prove that
𝑝(𝑆𝑇𝑆−1 ) = 𝑆𝑝(𝑇)𝑆−1.
41 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑈 is a subspace of 𝑉 invariant under 𝑇. Prove that
𝑈 is invariant under 𝑝(𝑇) for every polynomial 𝑝 ∈ 𝒫(𝐅).
42 Define 𝑇 ∈ ℒ(𝐅𝑛 ) by 𝑇(𝑥1 , 𝑥2 , 𝑥3 , …, 𝑥𝑛 ) = (𝑥1 , 2𝑥2 , 3𝑥3 , …, 𝑛𝑥𝑛 ).

(a) Find all eigenvalues and eigenvectors of 𝑇.
(b) Find all subspaces of 𝐅𝑛 that are invariant under 𝑇.
43 Suppose that 𝑉 is finite-dimensional, dim 𝑉 > 1, and 𝑇 ∈ ℒ(𝑉). Prove that
{𝑝(𝑇) ∶ 𝑝 ∈ 𝒫(𝐅)} ≠ ℒ(𝑉).
Section 5B The Minimal Polynomial 143
5B The Minimal Polynomial

Existence of Eigenvalues on Complex Vector Spaces
Now we come to one of the central results about operators on finite-dimensional
complex vector spaces.
5.19 existence of eigenvalues
Every operator on a finite-dimensional nonzero complex vector space has an

eigenvalue.
Proof Suppose 𝑉 is a finite-dimensional complex vector space of dimension

𝑛 > 0 and 𝑇 ∈ ℒ(𝑉). Choose 𝑣 ∈ 𝑉 with 𝑣 ≠ 0. Then
𝑣, 𝑇𝑣, 𝑇 2 𝑣, …, 𝑇 𝑛 𝑣
is not linearly independent, because 𝑉 has dimension 𝑛 and this list has length
𝑛 + 1. Hence some linear combination (with not all the coefficients equal to 0)
of the vectors above equals 0. Thus there exists a nonconstant polynomial 𝑝 of
smallest degree such that
𝑝(𝑇)𝑣 = 0.
By the first version of the fundamental theorem of algebra (see 4.12), there
exists 𝜆 ∈ 𝐂 such that 𝑝(𝜆) = 0. Hence there exists a polynomial 𝑞 ∈ 𝒫(𝐂) such
that
𝑝(𝑧) = (𝑧 − 𝜆)𝑞(𝑧)
for every 𝑧 ∈ 𝐂 (see 4.6). This implies (using 5.17) that
0 = 𝑝(𝑇)𝑣 = (𝑇 − 𝜆𝐼)(𝑞(𝑇)𝑣).
Because 𝑞 has smaller degree than 𝑝, we know that 𝑞(𝑇)𝑣 ≠ 0. Thus the equation
above implies that 𝜆 is an eigenvalue of 𝑇 with eigenvector 𝑞(𝑇)𝑣.
The proof above makes crucial use of the fundamental theorem of algebra.
The comment following Exercise 16 helps explain why the fundamental theorem
of algebra is so tightly connected to the result above.
The hypothesis in the result above that 𝐅 = 𝐂 cannot be replaced with the
hypothesis that 𝐅 = 𝐑 , as shown by Example 5.9. The next example shows that
the finite-dimensional hypothesis in the result above also cannot be deleted.
5.20 example: an operator on a complex vector space with no eigenvalues

Define 𝑇 ∈ ℒ(𝒫(𝐂)) by (𝑇𝑝)(𝑧) = 𝑧𝑝(𝑧). If 𝑝 ∈ 𝒫(𝐂) is a nonzero poly-
nomial, then the degree of 𝑇𝑝 is one more than the degree of 𝑝, and thus 𝑇𝑝 cannot
equal a scalar multiple of 𝑝. Hence 𝑇 has no eigenvalues.
Because 𝒫(𝐂) is infinite-dimensional, this example does not contradict the
result above.
Eigenvalues and the Minimal Polynomial

In this subsection we introduce an important polynomial associated with each
operator. We begin with the following definition.
5.21 definition: monic polynomial
A monic polynomial is a polynomial whose highest-degree coefficient equals 1.
For example, the polynomial 2 + 9𝑧2 + 𝑧7 is a monic polynomial of degree 7.
5.22 existence, uniqueness, and degree of minimal polynomial
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Then there is a unique monic

polynomial 𝑝 ∈ 𝒫(𝐅) of smallest degree such that 𝑝(𝑇) = 0. Furthermore,
deg 𝑝 ≤ dim 𝑉.
Proof If dim 𝑉 = 0, then 𝐼 is the zero operator on 𝑉 and thus we take 𝑝 to be

the constant polynomial 1.
Now use induction on dim 𝑉. Thus assume that dim 𝑉 > 0 and that the desired
result is true for all operators on all complex vector spaces of smaller dimension.
Let 𝑣 ∈ 𝑉 be such that 𝑣 ≠ 0. The list 𝑣, 𝑇𝑣, …, 𝑇 dim 𝑉 𝑣 has length 1 + dim 𝑉
and thus is linearly dependent. By the linear dependence lemma (2.19), there is
a smallest positive integer 𝑚 ≤ dim 𝑉 such that 𝑇 𝑚 𝑣 is a linear combination of
𝑣, 𝑇𝑣, …, 𝑇 𝑚 − 1 𝑣. Thus there exist scalars 𝑐0 , 𝑐1 , 𝑐2 , …, 𝑐𝑚 − 1 ∈ 𝐅 such that
5.23 𝑐0 𝑣 + 𝑐1 𝑇𝑣 + ⋯ + 𝑐𝑚 − 1 𝑇 𝑚 − 1 𝑣 + 𝑇 𝑚 𝑣 = 0.
Define a monic polynomial 𝑞 ∈ 𝒫𝑚 (𝐅) by
𝑞(𝑧) = 𝑐0 + 𝑐1 𝑧 + ⋯ + 𝑐𝑚 − 1 𝑧𝑚 − 1 + 𝑧𝑚.
Then 5.23 implies that 𝑞(𝑇)𝑣 = 0.
If 𝑘 is a nonnegative integer, then
𝑞(𝑇)(𝑇 𝑘 𝑣) = 𝑇 𝑘 (𝑞(𝑇)𝑣) = 𝑇 𝑘 (0) = 0.
The linear dependence lemma (2.19) shows that 𝑣, 𝑇𝑣, …, 𝑇 𝑚 − 1 𝑣 is linearly inde-
pendent. Thus the equation above implies that dim null 𝑞(𝑇) ≥ 𝑚. Hence
dim range 𝑞(𝑇) = dim 𝑉 − dim null 𝑞(𝑇) ≤ dim 𝑉 − 𝑚.
Because range 𝑞(𝑇) is invariant under 𝑇 (by 5.18), we can apply our induction
hypothesis to the operator 𝑇|range 𝑞(𝑇) on the vector space range 𝑞(𝑇). Thus there
is a monic polynomial 𝑠 ∈ 𝒫(𝐅) with
deg 𝑠 ≤ dim 𝑉 − 𝑚 and 𝑠(𝑇|range 𝑞(𝑇) ) = 0.
Hence for all 𝑣 ∈ 𝑉 we have
(𝑠𝑞)(𝑇)(𝑣) = 𝑠(𝑇)(𝑞(𝑇)𝑣) = 0
because 𝑞(𝑇)𝑣 ∈ range 𝑞(𝑇) and 𝑠(𝑇)|range 𝑞(𝑇) = 𝑠(𝑇|range 𝑞(𝑇) ) = 0. Thus 𝑠𝑞 is a
monic polynomial such that deg 𝑠𝑞 ≤ dim 𝑉 and (𝑠𝑞)(𝑇) = 0.
The paragraph above shows that there is a monic polynomial of degree at

most dim 𝑉 that when applied to 𝑇 gives the 0 operator. Thus there is a monic
polynomial of smallest degree with this property, completing the existence part
of this result.
Let 𝑝 ∈ 𝒫(𝐅) be a monic polynomial of smallest degree such that 𝑝(𝑇) = 0.
To prove the uniqueness part of the result, suppose 𝑟 ∈ 𝒫(𝐅) is a monic poly-
nomial of the same degree as 𝑝 and 𝑟(𝑇) = 0. Then (𝑝 − 𝑟)(𝑇) = 0 and also
deg(𝑝 − 𝑟) < deg 𝑝. If 𝑝 − 𝑟 were not equal to 0, then we could divide 𝑝 − 𝑟 by
the coefficient of the highest-order term in 𝑝 − 𝑟 to get a monic polynomial (of
smaller degree than 𝑝) that when applied to 𝑇 gives the 0 operator. Thus 𝑝 − 𝑟 = 0,
as desired.
The previous result justifies the following definition.
5.24 definition: minimal polynomial
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Then the minimal polynomial

of 𝑇 is the unique monic polynomial 𝑝 ∈ 𝒫(𝐅) of smallest degree such that
𝑝(𝑇) = 0.
To compute the minimal polynomial of an operator 𝑇 ∈ ℒ(𝑉), we need to

find the smallest positive integer 𝑚 such that the equation
𝑐0 𝐼 + 𝑐1 𝑇 + ⋯ + 𝑐𝑚 − 1 𝑇 𝑚 − 1 = −𝑇 𝑚
has a solution 𝑐0 , 𝑐1 , …, 𝑐𝑚 − 1 ∈ 𝐅. If we pick a basis of 𝑉 and replace 𝑇 in the
equation above with the matrix of 𝑇, then the equation above can be thought of
as a system of (dim 𝑉)2 linear equations in the 𝑚 unknowns 𝑐0 , 𝑐1 , …, 𝑐𝑚 − 1 ∈ 𝐅.
Gaussian elimination or another fast method of solving systems of linear equations
can tell us whether a solution exists, testing successive values 𝑚 = 1, 2, … until
a solution exists. By 5.22, a solution exists for some smallest positive integer
𝑚 ≤ dim 𝑉. The minimal polynomial of 𝑇 is then 𝑐0 + 𝑐1 𝑧 + ⋯ + 𝑐𝑚 − 1 𝑧𝑚 − 1 + 𝑧𝑚.
Even faster (usually), pick 𝑣 ∈ 𝑉 and consider the equation
5.25 𝑐0 𝑣 + 𝑐1 𝑇𝑣 + ⋯ + 𝑐dim 𝑉 − 1 𝑇 dim 𝑉 − 1 𝑣 = −𝑇 dim 𝑉 𝑣.
Use a basis of 𝑉 to convert the equation above to a system of dim 𝑉 linear equa-
tions in dim 𝑉 unknowns 𝑐0 , 𝑐1 , …, 𝑐dim 𝑉 − 1 . If this system of equations has a
unique solution 𝑐0 , 𝑐1 , …, 𝑐dim 𝑉 − 1 (as happens most of the time), then the scalars
𝑐0 , 𝑐1 , …, 𝑐dim 𝑉 − 1 , 1 are the coefficients of the minimal polynomial of 𝑇 (because
5.22 states that the degree of the minimal polynomial is at most dim 𝑉).
Consider operators on 𝐑4 (thought These estimates are based on testing
of as 4-by-4 matrices with respect to the millions of random matrices.
standard basis), and take 𝑣 = (1, 0, 0, 0)
in the paragraph above. The faster method described above works on over 99.8%
of the 4-by-4 matrices with integer entries in the interval [−10, 10] and on over
99.999% of the 4-by-4 matrices with integer entries in [−100, 100].
The next example illustrates the faster procedure discussed above.
5.26 example: minimal polynomial of an operator on 𝐅5

Suppose 𝑇 ∈ ℒ(𝐅5 ) and
0 0 0 0 −3
⎛
⎜ ⎞
⎟
⎜
⎜ 1 0 0 0 6 ⎟
⎟
⎜ ⎟
ℳ(𝑇) = ⎜
⎜
⎜ 0 1 0 0 0 ⎟
⎟
⎟
⎜
⎜ 0 0 1 0 0 ⎟
⎟
⎜ ⎟
⎝ 0 0 0 1 0 ⎠
with respect to the standard basis 𝑒1 , 𝑒2 , 𝑒3 , 𝑒4 , 𝑒5 . Taking 𝑣 = 𝑒1 for 5.25, we have
𝑇𝑒1 = 𝑒2 , 𝑇 4 𝑒1 = 𝑇(𝑇 3 𝑒1 ) = 𝑇𝑒4 = 𝑒5 ,

𝑇 2 𝑒1 = 𝑇(𝑇𝑒1 ) = 𝑇𝑒2 = 𝑒3 , 𝑇 5 𝑒1 = 𝑇(𝑇 4 𝑒1 ) = 𝑇𝑒5 = −3𝑒1 + 6𝑒2 .
𝑇 3 𝑒1 = 𝑇(𝑇 2 𝑒1 ) = 𝑇𝑒3 = 𝑒4 ,
Thus 3𝑒1 − 6𝑇𝑒1 = −𝑇 5 𝑒1 . The list 𝑒1 , 𝑇𝑒1 , 𝑇 2 𝑒1 , 𝑇 3 𝑒1 , 𝑇 4 𝑒1 , which equals the list
𝑒1 , 𝑒2 , 𝑒3 , 𝑒4 , 𝑒5 , is linearly independent, so no other linear combination of this list
equals −𝑇 5 𝑒1 . Hence the minimal polynomial of 𝑇 is 3 − 6𝑧 + 𝑧5.
Recall that by definition, eigenvalues of operators on 𝑉 and zeros of polyno-

mials in 𝒫(𝐅) must be elements of 𝐅. In particular, if 𝐅 = 𝐑 , then eigenvalues
and zeros must be real numbers.
5.27 eigenvalues are the zeros of the minimal polynomial
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉).

(a) The zeros of the minimal polynomial of 𝑇 are the eigenvalues of 𝑇.
(b) If 𝑉 is a complex vector space, then the minimal polynomial of 𝑇 has the
form
(𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ),
where 𝜆1 , …, 𝜆𝑚 is a list of all eigenvalues of 𝑇, possibly with repetitions.
Proof Let 𝑝 be the minimal polynomial of 𝑇.

(a) First suppose 𝜆 ∈ 𝐅 is a zero of 𝑝. Then 𝑝 can be written in the form
𝑝(𝑧) = (𝑧 − 𝜆)𝑞(𝑧),
where 𝑞 is a monic polynomial with coefficients in 𝐅 (see 4.6). Because

𝑝(𝑇) = 0, we have
0 = (𝑇 − 𝜆𝐼)(𝑞(𝑇)𝑣)
for all 𝑣 ∈ 𝑉. Because the degree of 𝑞 is less than the degree of the minimal
polynomial 𝑝, there exists at least one vector 𝑣 ∈ 𝑉 such that 𝑞(𝑇)𝑣 ≠ 0. The
equation above thus implies that 𝜆 is an eigenvalue of 𝑇, as desired.
To prove that every eigenvalue of 𝑇 is a zero of 𝑝, now suppose 𝜆 ∈ 𝐅 is

an eigenvalue of 𝑇. Thus there exists 𝑣 ∈ 𝑉 with 𝑣 ≠ 0 such that 𝑇𝑣 = 𝜆𝑣.
Repeated applications of 𝑇 to both sides of this equation show that 𝑇 𝑘 𝑣 = 𝜆𝑘 𝑣
for every nonnegative integer 𝑘. Thus
𝑝(𝑇)𝑣 = 𝑝(𝜆)𝑣.
Because 𝑝 is the minimal polynomial of 𝑇, we have 𝑝(𝑇)𝑣 = 0. Hence the

equation above implies that 𝑝(𝜆) = 0. Thus 𝜆 is a zero of 𝑝, as desired.
(b) To get the desired result, use (a) and the second version of the fundamental
theorem of algebra (see 4.13).
A nonzero polynomial has at most as many distinct zeros as its degree (see 4.8).
Thus (a) of the previous result, along with the result that the minimal polynomial
of an operator on 𝑉 has degree at most dim 𝑉, gives an alternative proof of 5.12,
which states that an operator on 𝑉 has at most dim 𝑉 distinct eigenvalues.
Every monic polynomial is the minimal polynomial of some operator, as
shown by Exercise 16, which generalizes Example 5.26. Thus 5.27(a) shows that
finding exact expressions for the eigenvalues of an operator is equivalent to the
problem of finding exact expressions for the zeros of a polynomial (and thus is
not possible for some operators).
5.28 example: An operator whose eigenvalues cannot be found exactly

Let 𝑇 ∈ ℒ(𝐂5 ) be the operator defined by
𝑇(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 , 𝑧5 ) = (−3𝑧5 , 𝑧1 + 6𝑧5 , 𝑧2 , 𝑧3 , 𝑧4 ).
The matrix of 𝑇 with respect to the standard basis of 𝐂5 is the 5-by-5 matrix in
Example 5.26. As we showed in that example, the minimal polynomial of 𝑇 is
the polynomial
3 − 6𝑧 + 𝑧5.
No zero of the polynomial above can be expressed using rational numbers,
roots of rational numbers, and the usual rules of arithmetic (a proof of this would
take us considerably beyond linear algebra). Because the zeros of the polynomial
above are the eigenvalues of 𝑇 [by 5.27(a)], we cannot find an exact expression
for any eigenvalue of 𝑇 in any familiar form.
Numeric techniques, which we will not discuss here, show that the zeros of the
polynomial above, and thus the eigenvalues of 𝑇, are approximately the following
five complex numbers:
−1.67, 0.51, 1.40, −0.12 + 1.59𝑖, −0.12 − 1.59𝑖.
Note that the two nonreal zeros of this polynomial are complex conjugates of
each other, as we expect for a polynomial with real coefficients (see 4.14).
The next result completely characterizes the polynomials that when applied to
an operator give the 0 operator.
5.29 𝑞(𝑇) = 0 ⟺ 𝑞 is a polynomial multiple of the minimal polynomial
Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝑞 ∈ 𝒫(𝐅). Then 𝑞(𝑇) = 0

if and only if 𝑞 is a polynomial multiple of the minimal polynomial of 𝑇.
Proof Let 𝑝 denote the minimal polynomial of 𝑇.

First suppose 𝑞(𝑇) = 0. By the division algorithm for polynomials (4.9), there
exist polynomials 𝑠, 𝑟 ∈ 𝒫(𝐅) such that
5.30 𝑞 = 𝑝𝑠 + 𝑟
and deg 𝑟 < deg 𝑝. We have
0 = 𝑞(𝑇) = 𝑝(𝑇)𝑠(𝑇) + 𝑟(𝑇) = 𝑟(𝑇).
The equation above implies that 𝑟 = 0 (otherwise, dividing 𝑟 by its highest-degree
coefficient would produce a monic polynomial that when applied to 𝑇 gives 0;
this polynomial would have a smaller degree than the minimal polynomial, which
would be a contradiction). Thus 5.30 becomes the equation 𝑞 = 𝑝𝑠. Hence 𝑞 is a
polynomial multiple of 𝑝, as desired.
To prove the other direction, now suppose 𝑞 is a polynomial multiple of 𝑝.
Thus there exists a polynomial 𝑠 ∈ 𝒫(𝐅) such that 𝑞 = 𝑝𝑠. We have
𝑞(𝑇) = 𝑝(𝑇)𝑠(𝑇) = 0 𝑠(𝑇) = 0,
as desired.
The next result is a nice consequence of the result above.
5.31 minimal polynomial of a restriction operator
Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝑈 is a subspace of 𝑉 that is

invariant under 𝑇. Then the minimal polynomial of 𝑇 is a polynomial multiple
of the minimal polynomial of 𝑇|𝑈 .
Proof Suppose 𝑝 is the minimal polynomial of 𝑇. Thus 𝑝(𝑇)𝑣 = 0 for all 𝑣 ∈ 𝑉.

In particular,
𝑝(𝑇)𝑢 = 0 for all 𝑢 ∈ 𝑈.
Thus 𝑝(𝑇|𝑈 ) = 0. Now 5.29, applied to the operator 𝑇|𝑈 in place of 𝑇, implies
that 𝑝 is a polynomial multiple of the minimal polynomial of 𝑇|𝑈 .
See Exercise 25 for a result about quotient operators that is analogous to the
result above.
The next result shows that the constant term of the minimal polynomial of an
operator determines whether the operator is invertible.
5.32 𝑇 not invertible ⟺ constant term of minimal polynomial of 𝑇 is 0
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Then 𝑇 is not invertible if

and only if the constant term of the minimal polynomial of 𝑇 is 0.
Proof Suppose 𝑇 ∈ ℒ(𝑉) and 𝑝 is the minimal polynomial of 𝑇. Then

𝑇 is not invertible ⟺ 0 is an eigenvalue of 𝑇
⟺ 0 is a zero of 𝑝
⟺ the constant term of 𝑝 is 0,
where the first equivalence holds by 5.7, the second equivalence holds by 5.27(a),
and the last equivalence holds because the constant term of 𝑝 equals 𝑝(0).
Eigenvalues on Odd-Dimensional Real Vector Spaces

The next result will be the key tool that we use to show that every operator on an
odd-dimensional real vector space has an eigenvalue.
5.33 even-dimensional null space
Suppose 𝐅 = 𝐑 and 𝑉 is finite-dimensional. Suppose also that 𝑇 ∈ ℒ(𝑉)

and 𝑏, 𝑐 ∈ 𝐑 with 𝑏2 < 4𝑐. Then dim null(𝑇 2 + 𝑏𝑇 + 𝑐𝐼) is an even number.
Proof Recall that null(𝑇 2 + 𝑏𝑇 + 𝑐𝐼) is invariant under 𝑇 (by 5.18). By replacing
𝑉 with null(𝑇 2 + 𝑏𝑇 + 𝑐𝐼) and replacing 𝑇 with 𝑇 restricted to null(𝑇 2 + 𝑏𝑇 + 𝑐𝐼),
we can assume that 𝑇 2 + 𝑏𝑇 + 𝑐𝐼 = 0; we now need to prove that dim 𝑉 is even.
Suppose 𝜆 ∈ 𝐑 and 𝑣 ∈ 𝑉 are such that 𝑇𝑣 = 𝜆𝑣. Then
2 𝑏2
0 = (𝑇 2 + 𝑏𝑇 + 𝑐𝐼)𝑣 = (𝜆2 + 𝑏𝜆 + 𝑐)𝑣 = ((𝜆 + 2𝑏 ) + 𝑐 − 4
)𝑣.
The term in large parentheses above is a positive number. Thus the equation above
implies that 𝑣 = 0. Hence we have shown that 𝑇 has no eigenvectors.
Let 𝑈 be a subspace of 𝑉 that is invariant under 𝑇 and has the largest dimension
among all subspaces of 𝑉 that are invariant under 𝑇 and have even dimension. If
𝑈 = 𝑉, then we are done; otherwise assume there exists 𝑤 ∈ 𝑉 such that 𝑤 ∉ 𝑈.
Let 𝑊 = span(𝑤, 𝑇𝑤). Then 𝑊 is invariant under 𝑇 because 𝑇(𝑇𝑤) =
−𝑏𝑇𝑤 − 𝑐𝑤. Furthermore, dim 𝑊 = 2 because otherwise 𝑤 would be an eigen-
vector of 𝑇. Now
dim(𝑈 + 𝑊) = dim 𝑈 + dim 𝑊 − dim(𝑈 ∩ 𝑊) = dim 𝑈 + 2,
where 𝑈 ∩ 𝑊 = {0} because otherwise 𝑈 ∩ 𝑊 would be a one-dimensional
subspace of 𝑉 that is invariant under 𝑇 (impossible because 𝑇 has no eigenvectors).
Because 𝑈 + 𝑊 is invariant under 𝑇, the equation above shows that there exists
a subspace of 𝑉 invariant under 𝑇 of even dimension larger than dim 𝑈. Thus the
assumption that 𝑈 ≠ 𝑉 was incorrect. Hence 𝑉 has even dimension.
The next result states that on odd-dimensional vector spaces, every operator
has an eigenvalue. We already know this result for finite-dimensional complex
vectors spaces (without the odd hypothesis). Thus in the proof below, we will
assume that 𝐅 = 𝐑 .
5.34 operators on odd-dimensional vector spaces have eigenvalues
Every operator on an odd-dimensional vector space has an eigenvalue.
Proof Suppose 𝐅 = 𝐑 and 𝑉 is finite-dimensional. Let 𝑛 = dim 𝑉, and suppose

𝑛 is an odd number. Let 𝑇 ∈ ℒ(𝑉). We will use induction on 𝑛 in steps of size
two to show that 𝑇 has an eigenvalue. To get started, note that the desired result
holds if dim 𝑉 = 1 because then every nonzero vector in 𝑉 is an eigenvector of 𝑇.
Now suppose that 𝑛 ≥ 3 and the desired result holds for all operators on all
odd-dimensional vector spaces of dimension less than 𝑛. Let 𝑝 denote the minimal
polynomial of 𝑇. If 𝑝 is a polynomial multiple of 𝑥 − 𝜆 for some 𝜆 ∈ 𝐑 , then 𝜆 is
an eigenvalue of 𝑇 [by 5.27(a)] and we are done. Thus we can assume that there
exist 𝑏, 𝑐 ∈ 𝐑 such that 𝑏2 < 4𝑐 and 𝑝 is a polynomial multiple of 𝑥2 + 𝑏𝑥 + 𝑐 (see
4.16).
There exists a monic polynomial 𝑞 ∈ 𝒫(𝐑) such that 𝑝(𝑥) = 𝑞(𝑥)(𝑥2 + 𝑏𝑥 + 𝑐)
for all 𝑥 ∈ 𝐑 . Now
0 = 𝑝(𝑇) = (𝑞(𝑇))(𝑇 2 + 𝑏𝑇 + 𝑐𝐼),
which means that 𝑞(𝑇) equals 0 on range(𝑇 2 + 𝑏𝑇 + 𝑐𝐼). Because deg 𝑞 < deg 𝑝
and 𝑝 is the minimal polynomial of 𝑇, this implies that range(𝑇 2 + 𝑏𝑇 + 𝑐𝐼) ≠ 𝑉.
The fundamental theorem of linear maps (3.21) tells us that
dim 𝑉 = dim null(𝑇 2 + 𝑏𝑇 + 𝑐𝐼) + dim range(𝑇 2 + 𝑏𝑇 + 𝑐𝐼).
Because dim 𝑉 is odd (by hypothesis) and dim null(𝑇 2 + 𝑏𝑇 + 𝑐𝐼) is even (by 5.33),
the equation above shows that dim range(𝑇 2 + 𝑏𝑇 + 𝑐𝐼) is odd.
Hence range(𝑇 2 + 𝑏𝑇 + 𝑐𝐼) is a subspace of 𝑉 that is invariant under 𝑇 (by
5.18) and that has odd dimension less than dim 𝑉. Our induction hypothesis now
implies that 𝑇 restricted to range(𝑇 2 + 𝑏𝑇 + 𝑐𝐼) has an eigenvalue, which means
that 𝑇 has an eigenvalue.
See Exercise 23 in Section 8B and Exercise 10 in Section 9C for alternative

proofs of the result above.
Exercises 5B
1 Suppose 𝑇 ∈ ℒ(𝑉). Prove that 9 is an eigenvalue of 𝑇 2 if and only if 3 or

−3 is an eigenvalue of 𝑇.
2 Suppose 𝑉 is a complex vector space and 𝑇 ∈ ℒ(𝑉) has no eigenvalues.
Prove that every subspace of 𝑉 invariant under 𝑇 is either {0} or infinite-
dimensional.
3 Suppose 𝑛 is a positive integer and 𝑇 ∈ ℒ(𝐅𝑛 ) is defined by

𝑇(𝑥1 , …, 𝑥𝑛 ) = (𝑥1 + ⋯ + 𝑥𝑛 , …, 𝑥1 + ⋯ + 𝑥𝑛 ).
(a) Find all eigenvalues and eigenvectors of 𝑇.
(b) Find the minimal polynomial of 𝑇.
The matrix of 𝑇 with respect to the standard basis of 𝐅𝑛 consists of all 1’s.
4 Suppose 𝐅 = 𝐂 , 𝑇 ∈ ℒ(𝑉), 𝑝 ∈ 𝒫(𝐂), and 𝛼 ∈ 𝐂 . Prove that 𝛼 is an

eigenvalue of 𝑝(𝑇) if and only if 𝛼 = 𝑝(𝜆) for some eigenvalue 𝜆 of 𝑇.
5 Give an example of an operator on 𝐑2 that shows the result in Exercise 4
does not hold if 𝐂 is replaced with 𝐑 .
6 Suppose 𝑇 ∈ ℒ(𝐅2 ) is defined by 𝑇(𝑤, 𝑧) = (−𝑧, 𝑤). Find the minimal
polynomial of 𝑇.
7 (a) Give an example of 𝑆, 𝑇 ∈ ℒ(𝐅2 ) such that the minimal polynomial of
𝑆𝑇 does not equal the minimal polynomial of 𝑇𝑆.
(b) Suppose 𝑉 is finite-dimensional and 𝑆, 𝑇 ∈ ℒ(𝑉). Prove that if at least
one of 𝑆, 𝑇 is invertible, then the minimal polynomial of 𝑆𝑇 equals the
minimal polynomial of 𝑇𝑆.
Hint: Show that if 𝑆 is invertible and 𝑝 ∈ 𝒫(𝐅), then 𝑝(𝑇𝑆) = 𝑆−1 𝑝(𝑆𝑇)𝑆.
8 Suppose 𝑇 ∈ ℒ(𝐑2 ) is the operator of counterclockwise rotation by 1∘ . Find

the minimal polynomial of 𝑇.
Because dim 𝐑2 = 2, the degree of the minimal polynomial of 𝑇 is at most 2.
Thus the minimal polynomial of 𝑇 is not the tempting polynomial 𝑥180 + 1,
even though 𝑇 180 = −𝐼.
9 Suppose 𝑇 ∈ ℒ(𝑉) is such that with respect to some basis of 𝑉, all entries
of the matrix of 𝑇 are rational numbers. Explain why all coefficients of the
minimal polynomial of 𝑇 are rational numbers.
10 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝑣 ∈ 𝑉. Prove that
span(𝑣, 𝑇𝑣, …, 𝑇 𝑚 𝑣) = span(𝑣, 𝑇𝑣, …, 𝑇 dim 𝑉 − 1 𝑣)
for all integers 𝑚 ≥ dim 𝑉 − 1.
11 Suppose 𝑉 is a two-dimensional vector space, 𝑇 ∈ ℒ(𝑉), and the matrix of
𝑎 𝑐
𝑇 with respect to some basis of 𝑉 is ( ).
𝑏 𝑑
(a) Show that 𝑇 2 − (𝑎 + 𝑑)𝑇 + (𝑎𝑑 − 𝑏𝑐)𝐼 = 0.
(b) Show that the minimal polynomial of 𝑇 equals
⎧
{𝑧 − 𝑎 if 𝑏 = 𝑐 = 0 and 𝑎 = 𝑑,
⎨𝑧2 − (𝑎 + 𝑑)𝑧 + (𝑎𝑑 − 𝑏𝑐)
{ otherwise.
⎩
12 Define 𝑇 ∈ ℒ(𝐅𝑛 ) by 𝑇(𝑥1 , 𝑥2 , 𝑥3 , …, 𝑥𝑛 ) = (𝑥1 , 2𝑥2 , 3𝑥3 , …, 𝑛𝑥𝑛 ). Find the

minimal polynomial of 𝑇.
13 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑝 ∈ 𝒫(𝐅). Prove that there exists a unique 𝑟 ∈ 𝒫(𝐅)
such that 𝑝(𝑇) = 𝑟(𝑇) and deg 𝑟 is less than the degree of the minimal
polynomial of 𝑇.
14 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉) has minimal polynomial
4 + 5𝑧 − 6𝑧2 − 7𝑧3 + 2𝑧4 + 𝑧5. Find the minimal polynomial of 𝑇 −1.
15 Suppose 𝑉 is a finite-dimensional complex vector space with dim 𝑉 > 0

and 𝑇 ∈ ℒ(𝑉). Define 𝑓 ∶ 𝐂 → 𝐑 by
𝑓 (𝜆) = dim range(𝑇 − 𝜆𝐼).
Prove that 𝑓 is not a continuous function.

16 Suppose 𝑎0 , …, 𝑎𝑛 − 1 ∈ 𝐅. Let 𝑇 be the operator on 𝐅𝑛 whose matrix (with
respect to the standard basis) is
0 −𝑎0
⎛
⎜ ⎞
⎟
⎜
⎜ 1 0 −𝑎1 ⎟
⎟
⎜
⎜ ⎟
⎟
⎜ 1 ⋱ −𝑎2 ⎟
⎜
⎜ ⎟.
⎟
⎜
⎜ ⋱ ⋮ ⎟
⎟
⎜
⎜ ⎟
⎟
⎜ 0 −𝑎𝑛 − 2 ⎟
⎝ 1 −𝑎𝑛 − 1 ⎠
Here all entries of the matrix are 0 except for all 1’s on the line under the
diagonal and the entries in the last column (some of which might also be 0).
Show that the minimal polynomial of 𝑇 is the polynomial
𝑎0 + 𝑎1 𝑧 + ⋯ + 𝑎𝑛 − 1 𝑧𝑛 − 1 + 𝑧𝑛.
The matrix above is called the companion matrix of the polynomial above.
This exercise shows that every monic polynomial is the minimal polynomial
of some operator. Hence a formula or an algorithm that could produce
exact eigenvalues for each operator on each 𝐅𝑛 could then produce exact
zeros for each polynomial [by 5.27(a)]. Thus there is no such formula or
algorithm. However, efficient numeric methods exist for obtaining very good
approximations for the eigenvalues of an operator.
17 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝑝 is the minimal polynomial

of 𝑇. Suppose 𝜆 ∈ 𝐅. Show that the minimal polynomial of 𝑇 − 𝜆𝐼 is the
polynomial 𝑞 defined by 𝑞(𝑧) = 𝑝(𝑧 + 𝜆).
18 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝑝 is the minimal polynomial
of 𝑇. Suppose 𝜆 ∈ 𝐅\{0}. Show that the minimal polynomial of 𝜆𝑇 is the
𝑧
polynomial 𝑞 defined by 𝑞(𝑧) = 𝜆deg 𝑝 𝑝( ).
𝜆
19 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Let ℰ be the subspace of

ℒ(𝑉) defined by
ℰ = {𝑞(𝑇) ∶ 𝑞 ∈ 𝒫(𝐅)}.
Prove that dim ℰ equals the degree of the minimal polynomial of 𝑇.
20 Suppose 𝑇 ∈ ℒ(𝐅4 ) is such that the eigenvalues of 𝑇 are 3, 5, 8. Prove that
(𝑇 − 3𝐼)2 (𝑇 − 5𝐼)2 (𝑇 − 8𝐼)2 = 0.
21 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that the minimal
polynomial of 𝑇 has degree at most 1 + dim range 𝑇.
If dim range 𝑇 < dim 𝑉 − 1, then this exercise gives a better upper bound
than 5.22 for the degree of the minimal polynomial of 𝑇.
22 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is invertible if
and only if 𝐼 ∈ span(𝑇, 𝑇 2, …, 𝑇 dim 𝑉 ).
23 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Let 𝑛 = dim 𝑉. Prove that
if 𝑣 ∈ 𝑉, then span(𝑣, 𝑇𝑣, …, 𝑇 𝑛 − 1 𝑣) is invariant under 𝑇.
24 Suppose 𝑉 is a finite-dimensional complex vector space. Suppose 𝑇 ∈ ℒ(𝑉)
is such that 5 and 6 are eigenvalues of 𝑇 and that 𝑇 has no other eigenvalues.
Prove that (𝑇 − 5𝐼)dim 𝑉 − 1 (𝑇 − 6𝐼)dim 𝑉 − 1 = 0.
25 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝑈 is a subspace of 𝑉 that
is invariant under 𝑇.
(a) Prove that the minimal polynomial of 𝑇 is a polynomial multiple of the
minimal polynomial of the quotient operator 𝑇/𝑈.
(b) Prove that
(minimal polynomial of 𝑇|𝑈 ) × (minimal polynomial of 𝑇/𝑈)
is a polynomial multiple of the minimal polynomial of 𝑇.
The quotient operator 𝑇/𝑈 was defined in Exercise 38 in Section 5A.
26 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝑈 is a subspace of 𝑉 that
is invariant under 𝑇. Prove that the set of eigenvalues of 𝑇 equals the union
of the set of eigenvalues of 𝑇|𝑈 and the set of eigenvalues of 𝑇/𝑈.
27 Suppose 𝐅 = 𝐑 , 𝑉 is finite-dimensional, and 𝑇 ∈ ℒ(𝑉). Prove that the
minimal polynomial of 𝑇𝐂 equals the minimal polynomial of 𝑇.
The complexification 𝑇𝐂 was defined in Exercise 33 of Section 3B.
28 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that the minimal
polynomial of 𝑇 ′ ∈ ℒ(𝑉 ′ ) equals the minimal polynomial of 𝑇.
The dual map 𝑇 ′ was defined in Section 3F.
29 Show that every operator on a finite-dimensional vector space of dimension
at least two has an invariant subspace of dimension two.
Exercise 6 in Section 5C will give an improvement of this result when 𝐅 = 𝐂 .
5C Upper-Triangular Matrices
In Chapter 3 we defined the matrix of a linear map from a finite-dimensional vector
space to another finite-dimensional vector space. That matrix depends on a choice
of basis of each of the two vector spaces. Now that we are studying operators,
which map a vector space to itself, the emphasis is on using only one basis.
5.35 definition: matrix of an operator, ℳ(𝑇)
Suppose 𝑇 ∈ ℒ(𝑉). The matrix of 𝑇 with respect to a basis 𝑣1 , …, 𝑣𝑛 of 𝑉 is

the 𝑛-by-𝑛 matrix
𝐴 ⋯ 𝐴 1, 𝑛
⎛ 1, 1 ⎞
ℳ(𝑇) = ⎜
⎜
⎜ ⋮ ⋮ ⎟
⎟
⎟
⎝ 𝐴𝑛, 1 ⋯ 𝐴𝑛, 𝑛 ⎠
whose entries 𝐴𝑗, 𝑘 are defined by
𝑇𝑣𝑘 = 𝐴1, 𝑘 𝑣1 + ⋯ + 𝐴𝑛, 𝑘 𝑣𝑛 .
The notation ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )) is used if the basis is not clear from the context.
Operators have square matrices (meaning that the number of rows equals the
number of columns), rather than the more general rectangular matrices that we
considered earlier for linear maps.
If 𝑇 is an operator on 𝐅𝑛 and no ba-
The 𝑘 th column of the matrix ℳ(𝑇) is
sis is specified, assume that the basis in formed from the coefficients used to
question is the standard one (where the write 𝑇𝑣 as a linear combination of
𝑘 th basis vector is 1 in the 𝑘 th slot and 0
𝑘
the basis 𝑣1 , …, 𝑣𝑛 .
in all other slots). You can then think of
the 𝑘 th column of ℳ(𝑇) as 𝑇 applied to the 𝑘 th basis vector, where we identify
𝑛-by-1 column vectors with elements of 𝐅𝑛.
5.36 example: matrix of an operator with respect to standard basis

Define 𝑇 ∈ ℒ(𝐅3 ) by 𝑇(𝑥, 𝑦, 𝑧) = (2𝑥 + 𝑦, 5𝑦 + 3𝑧, 8𝑧). Then the matrix of 𝑇
with respect to the standard basis of 𝐅3 is
2 1 0
⎛ ⎞
ℳ(𝑇) = ⎜
⎜ ⎟,
⎜ 0 5 3 ⎟
⎟
⎝ 0 0 8 ⎠
A central goal of linear algebra is to show that given an operator 𝑇 on a finite-

dimensional vector space 𝑉, there exists a basis of 𝑉 with respect to which 𝑇 has
a reasonably simple matrix. To make this vague formulation a bit more precise,
we might try to choose a basis of 𝑉 such that ℳ(𝑇) has many 0’s.
Section 5C Upper-Triangular Matrices 155
If 𝑉 is a finite-dimensional complex vector space, then we already know

enough to show that there is a basis of 𝑉 with respect to which the matrix of 𝑇
has 0’s everywhere in the first column, except possibly the first entry. In other
words, there is a basis of 𝑉 with respect to which the matrix of 𝑇 looks like
𝜆
⎛
⎜ ⎞
⎟
⎜
⎜ 0 ∗ ⎟
⎟;
⎜
⎜ ⋮ ⎟
⎟
⎜ ⎟
⎝ 0 ⎠
here ∗ denotes the entries in all columns other than the first column. To prove
this, let 𝜆 be an eigenvalue of 𝑇 (one exists by 5.19) and let 𝑣 be a corresponding
eigenvector. Extend 𝑣 to a basis of 𝑉. Then the matrix of 𝑇 with respect to this
basis has the form above. Soon we will see that we can choose a basis of 𝑉 with
respect to which the matrix of 𝑇 has even more 0’s.
5.37 definition: diagonal of a matrix
The diagonal of a square matrix consists of the entries on the line from the
upper left corner to the bottom right corner.
For example, the diagonal of the matrix

2 1 0
⎛ ⎞
ℳ(𝑇) = ⎜
⎜ ⎟
⎜ 0 5 3 ⎟
⎟
⎝ 0 0 8 ⎠
from Example 5.36 consists of the entries 2, 5, 8, which are shown in red in the
matrix above.
5.38 definition: upper-triangular matrix
A square matrix is called upper triangular if all entries below the diagonal
are 0.
For example, the 3-by-3 matrix above is upper triangular.

Typically we represent an upper-triangular matrix in the form
⎛
𝜆1 ∗ ⎞
⎜ ⎟
⎜
⎜ ⋱ ⎟;
⎟
⎝ 0 𝜆𝑛 ⎠
the 0 in the matrix above indicates that We often use ∗ to denote matrix entries
all entries below the diagonal in this that we do not know or that are irrele-
𝑛-by-𝑛 matrix equal 0. Upper-triangular vant to the questions being discussed.
matrices can be considered reasonably
simple—if 𝑛 is large, then at least almost half the entries in an 𝑛-by-𝑛 upper-
triangular matrix are 0.
The next result provides a useful connection between upper-triangular matrices

and invariant subspaces.
5.39 conditions for upper-triangular matrix
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Then the following are

equivalent.
(a) The matrix of 𝑇 with respect to 𝑣1 , …, 𝑣𝑛 is upper triangular.
(b) span(𝑣1 , …, 𝑣𝑘 ) is invariant under 𝑇 for each 𝑘 = 1, …, 𝑛.
(c) 𝑇𝑣𝑘 ∈ span(𝑣1 , …, 𝑣𝑘 ) for each 𝑘 = 1, …, 𝑛.
Proof First suppose (a) holds. To prove that (b) holds, suppose 𝑘 ∈ {1, …, 𝑛}. If
𝑗 ∈ {1, …, 𝑛}, then
𝑇𝑣𝑗 ∈ span(𝑣1 , …, 𝑣𝑗 )
because the matrix of 𝑇 with respect to 𝑣1 , …, 𝑣𝑛 is upper triangular. Because
span(𝑣1 , …, 𝑣𝑗 ) ⊆ span(𝑣1 , …, 𝑣𝑘 ) if 𝑗 ≤ 𝑘, we see that
𝑇𝑣𝑗 ∈ span(𝑣1 , …, 𝑣𝑘 )
for each 𝑗 ∈ {1, …, 𝑘}. Thus span(𝑣1 , …, 𝑣𝑘 ) is invariant under 𝑇, completing the
proof that (a) implies (b).
Now suppose (b) holds, so span(𝑣1 , …, 𝑣𝑘 ) is invariant under 𝑇 for each
𝑘 = 1, …, 𝑛. In particular, 𝑇𝑣𝑘 ∈ span(𝑣1 , …, 𝑣𝑘 ) for each 𝑘 = 1, …, 𝑛. Thus
(b) implies (c).
Now suppose (c) holds, so 𝑇𝑣𝑘 ∈ span(𝑣1 , …, 𝑣𝑘 ) for each 𝑘 = 1, …, 𝑛. This
means that when writing each 𝑇𝑣𝑘 as a linear combination of the basis vectors
𝑣1 , …, 𝑣𝑛 , we need to use only the vectors 𝑣1 , …, 𝑣𝑘 . Hence all entries under the
diagonal of ℳ(𝑇) are 0. Thus ℳ(𝑇) is an upper-triangular matrix, completing
the proof that (c) implies (a).
We have shown that (a) ⟹ (b) ⟹ (c) ⟹ (a), which shows that (a), (b),
and (c) are equivalent.
The next result tells us that if 𝑇 ∈ ℒ(𝑉) and with respect to some basis of 𝑉
we have
𝜆
⎛ 1
∗ ⎞
ℳ(𝑇) = ⎜ ⎜
⎜ ⋱ ⎟
⎟,
⎟
⎝ 0 𝜆𝑛 ⎠
then 𝑇 satisfies a simple equation depending on 𝜆1 , …, 𝜆𝑛 .
5.40 equation satisfied by operator with upper-triangular matrix
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑉 has a basis with respect to which 𝑇 has an upper-
triangular matrix with diagonal entries 𝜆1 , …, 𝜆𝑛 . Then
(𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑛 𝐼) = 0.
Proof Let 𝑣1 , …, 𝑣𝑛 denote a basis of 𝑉 with respect to which 𝑇 has an upper-

triangular matrix with diagonal entries 𝜆1 , …, 𝜆𝑛 . Then 𝑇𝑣1 = 𝜆1 𝑣1 , which
means that (𝑇 − 𝜆1 𝐼)𝑣1 = 0, which implies that (𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑚 𝐼)𝑣1 = 0 for
𝑚 = 1, …, 𝑛 (using the commutativity of each 𝑇 − 𝜆𝑗 𝐼 with each 𝑇 − 𝜆𝑘 𝐼).
Note that (𝑇 − 𝜆2 𝐼)𝑣2 ∈ span(𝑣1 ). Thus (𝑇 − 𝜆1 𝐼)(𝑇 − 𝜆2 𝐼)𝑣2 = 0 (by
the previous paragraph), which implies that (𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑚 𝐼)𝑣1 = 0 for
𝑚 = 2, …, 𝑛 (using the commutativity of each 𝑇 − 𝜆𝑗 𝐼 with each 𝑇 − 𝜆𝑘 𝐼).
Note that (𝑇 − 𝜆3 𝐼)𝑣3 ∈ span(𝑣1 , 𝑣2 ). Thus by the previous paragraph,
(𝑇−𝜆1 𝐼)(𝑇−𝜆2 𝐼)(𝑇−𝜆3 𝐼)𝑣3 = 0, which implies that (𝑇−𝜆1 𝐼)⋯(𝑇−𝜆𝑚 𝐼)𝑣1 = 0
for 𝑚 = 3, …, 𝑛 (using the commutativity of each 𝑇 − 𝜆𝑗 𝐼 with each 𝑇 − 𝜆𝑘 𝐼).
Continuing this pattern, we see that (𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑛 𝐼)𝑣𝑘 = 0 for each
𝑘 = 1, …, 𝑛. Thus (𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑛 𝐼) is the 0 operator because it is 0 on each
vector in a basis of 𝑉.
Unfortunately no method exists for exactly computing the eigenvalues of an

operator from its matrix. However, if we are fortunate enough to find a basis with
respect to which the matrix of the operator is upper triangular, then the problem
of computing the eigenvalues becomes trivial, as the next result shows.
5.41 determination of eigenvalues from upper-triangular matrix
Suppose 𝑇 ∈ ℒ(𝑉) has an upper-triangular matrix with respect to some basis

of 𝑉. Then the eigenvalues of 𝑇 are precisely the entries on the diagonal of
that upper-triangular matrix.
Proof Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 with respect to which 𝑇 has an upper-

triangular matrix
⎛ 1
𝜆 ∗ ⎞
ℳ(𝑇) = ⎜ ⎜
⎜ ⋱ ⎟
⎟
⎟.
⎝ 0 𝜆𝑛 ⎠
Because 𝑇𝑣1 = 𝜆1 𝑣1 , we see that 𝜆1 is an eigenvalue of 𝑇.
Suppose 𝑘 ∈ {2, …, 𝑛}. Then (𝑇 − 𝜆𝑘 𝐼)𝑣𝑘 ∈ span(𝑣1 , …, 𝑣𝑘 − 1 ). Thus 𝑇 − 𝜆𝑘 𝐼
maps span(𝑣1 , …, 𝑣𝑘 ) into span(𝑣1 , …, 𝑣𝑘 − 1 ). Because
dim span(𝑣1 , …, 𝑣𝑘 ) = 𝑘 and dim span(𝑣1 , …, 𝑣𝑘 − 1 ) = 𝑘 − 1,
this implies that 𝑇 − 𝜆𝑘 𝐼 restricted to span(𝑣1 , …, 𝑣𝑘 ) is not injective (by 3.22).
Thus there exists 𝑣 ∈ span(𝑣1 , …, 𝑣𝑘 ) such that 𝑣 ≠ 0 and (𝑇 − 𝜆𝑘 𝐼)𝑣 = 0. Thus
𝜆𝑘 is an eigenvalue of 𝑇. Hence we have shown that every entry on the diagonal
of ℳ(𝑇) is an eigenvalue of 𝑇.
To prove 𝑇 has no other eigenvalues, let 𝑞 be the polynomial defined by
𝑞(𝑧) = (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑛 ). Then 𝑞(𝑇) = 0 (by 5.40). Hence 𝑞 is a polynomial
multiple of the minimal polynomial of 𝑇 (by 5.29). Thus every zero of the minimal
polynomial of 𝑇 is a zero of 𝑞. Because the zeros of the minimal polynomial of
𝑇 are the eigenvalues of 𝑇 (by 5.27), this implies that every eigenvalue of 𝑇 is a
zero of 𝑞. Hence the eigenvalues of 𝑇 are all contained in the list 𝜆1 , …, 𝜆𝑛 .
5.42 example: eigenvalues via an upper-triangular matrix

Define 𝑇 ∈ ℒ(𝐅3 ) by 𝑇(𝑥, 𝑦, 𝑧) = (2𝑥 + 𝑦, 5𝑦 + 3𝑧, 8𝑧). The matrix of 𝑇 with
respect to the standard basis is
2 1 0
⎛ ⎞
ℳ(𝑇) = ⎜ ⎜
⎜ 0 5 3 ⎟
⎟.
⎟
⎝ 0 0 8 ⎠
Now 5.41 implies that the eigenvalues of 𝑇 are 2, 5, and 8.
The next example illustrates 5.44: an operator has an upper-triangular matrix

with respect to some basis if and only if the minimal polynomial of the operator
is the product of polynomials of degree 1.
5.43 example: whether 𝑇 has an upper-triangular matrix can depend on 𝐅

Define 𝑇 ∈ ℒ(𝐅4 ) by
𝑇(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 ) = (−𝑧2 , 𝑧1 , 2𝑧1 + 3𝑧3 , 𝑧3 + 3𝑧4 ).
Thus with respect to the standard basis of 𝐅4 , the matrix of 𝑇 is
0 −1 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ 1 0 0 0 ⎟
⎟
⎜
⎜ ⎟
⎟ .
⎜ 2 0 3 0 ⎟
⎝ 0 0 1 3 ⎠
You can ask a computer to verify that the minimal polynomial of 𝑇 is the polyno-
mial 𝑝 defined by
𝑝(𝑧) = 9 − 6𝑧 + 10𝑧2 − 6𝑧3 + 𝑧4.
First consider the case 𝐅 = 𝐑 . Then the polynomial 𝑝 factors as
𝑝(𝑧) = (𝑧2 + 1)(𝑧 − 3)(𝑧 − 3),
with no further factorization of 𝑧2 + 1 as the product of two polynomials of degree
1 with real coefficients. Thus 5.44 states that there does not exist a basis of 𝐑4
with respect to which 𝑇 has an upper-triangular matrix.
Now consider the case 𝐅 = 𝐂 . Then the polynomial 𝑝 factors as
𝑝(𝑧) = (𝑧 − 𝑖)(𝑧 + 𝑖)(𝑧 − 3)(𝑧 − 3),
where all factors above have the form 𝑧− 𝜆𝑘 . Thus 5.44 states that there is a basis of
𝐂4 with respect to which 𝑇 has an upper-triangular matrix. Indeed, you can verify
that with respect to the basis (4 − 3𝑖, −3 − 4𝑖, −3 + 𝑖, 1), (4 + 3𝑖, −3 + 4𝑖, −3 − 𝑖, 1),
(0, 0, 0, 1), (0, 0, 1, 0) of 𝐂4 , the operator 𝑇 has the upper-triangular matrix
𝑖 0 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ 0 −𝑖 0 0 ⎟
⎟
⎜
⎜ ⎟
⎟ .
⎜ 0 0 3 1 ⎟
⎝ 0 0 0 3 ⎠
5.44 necessary and sufficient condition to have an upper-triangular matrix
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Then 𝑇 has an upper-

triangular matrix with respect to some basis of 𝑉 if and only if the minimal
polynomial of 𝑇 equals (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ) for some 𝜆1 , …, 𝜆𝑚 ∈ 𝐅.
Proof First suppose 𝑇 has an upper-triangular matrix with respect to some basis
of 𝑉. Let 𝛼1 , …, 𝛼𝑛 denote the diagonal entries of that matrix. Define a polynomial
𝑞 ∈ 𝒫(𝐅) by
𝑞(𝑧) = (𝑧 − 𝛼1 )⋯(𝑧 − 𝛼𝑛 ).
Then 𝑞(𝑇) = 0, by 5.40. Hence 𝑞 is a polynomial multiple of the minimal polyno-
mial of 𝑇, by 5.29. Thus the minimal polynomial of 𝑇 equals (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 )
for some 𝜆1 , …, 𝜆𝑚 ∈ 𝐅 with {𝜆1 , …, 𝜆𝑚 } ⊆ {𝛼1 , …, 𝛼𝑛 }.
To prove the implication in the other direction, now suppose the minimal
polynomial of 𝑇 equals (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ) for some 𝜆1 , …, 𝜆𝑚 ∈ 𝐅. We will use
induction on 𝑚. To get started, if 𝑚 = 1 then 𝑧 − 𝜆1 is the minimal polynomial of
𝑇, which implies that 𝑇 = 𝜆1 𝐼, which implies that the matrix of 𝑇 (with respect
to any basis of 𝑉) is upper triangular.
Now suppose 𝑚 > 1 and the desired result holds for all smaller positive
integers. Let
𝑈 = range(𝑇 − 𝜆𝑚 𝐼).
Then 𝑈 is invariant under 𝑇 [this is a special case of 5.18 with 𝑝(𝑧) = 𝑧 − 𝜆𝑚 ].
Thus 𝑇|𝑈 is an operator on 𝑈.
If 𝑢 ∈ 𝑈, then 𝑢 = (𝑇 − 𝜆𝑚 𝐼)𝑣 for some 𝑣 ∈ 𝑉 and
(𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑚 − 1 𝐼)𝑢 = (𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑚 𝐼)𝑣 = 0.
Hence (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 − 1 ) is a polynomial multiple of the minimal polynomial

of 𝑇|𝑈 , by 5.29. Thus the minimal polynomial of 𝑇|𝑈 is the product of at most
𝑚 − 1 terms of the form 𝑧 − 𝜆𝑘 .
By our induction hypothesis, there is a basis 𝑢1 , …, 𝑢𝑀 of 𝑈 with respect to
which 𝑇|𝑈 has an upper-triangular matrix. Thus for each 𝑘 ∈ {1, …, 𝑀}, we have
(using 5.39)
5.45 𝑇𝑢𝑘 = (𝑇|𝑈 )(𝑢𝑘 ) ∈ span(𝑢1 , …, 𝑢𝑘 ).
Extend 𝑢1 , …, 𝑢𝑀 to a basis 𝑢1 , …, 𝑢𝑀 , 𝑣1 , …, 𝑣𝑁 of 𝑉. For each 𝑘 ∈ {1, …, 𝑁},

we have
𝑇𝑣𝑘 = (𝑇 − 𝜆𝑚 𝐼)𝑣𝑘 + 𝜆𝑚 𝑣𝑘 .
The definition of 𝑈 shows that (𝑇 − 𝜆𝑚 𝐼)𝑣𝑘 ∈ 𝑈 = span(𝑢1 , …, 𝑢𝑀 ). Thus the
equation above shows that
5.46 𝑇𝑣𝑘 ∈ span(𝑢1 , …, 𝑢𝑀 , 𝑣1 , …, 𝑣𝑘 ).
From 5.45 and 5.46, we conclude (using 5.39) that 𝑇 has an upper-triangular
matrix with respect to the basis 𝑢1 , …, 𝑢𝑀 , 𝑣1 , …, 𝑣𝑁 of 𝑉, as desired.
The set of numbers {𝜆1 , …, 𝜆𝑚 } from the previous result equals the set of
eigenvalues of 𝑇 (because the set of zeros of the minimal polynomial of 𝑇 equals
the set of eigenvalues of 𝑇, by 5.27), although the list 𝜆1 , …, 𝜆𝑚 in the previous
result may contain repetitions.
In Chapter 8 we will improve even the wonderful result below; see 8.37 and
8.46.
5.47 if 𝐅 = 𝐂 , then every operator on 𝑉 has an upper-triangular matrix
Suppose 𝑉 is a finite-dimensional complex vector space and 𝑇 ∈ ℒ(𝑉). Then

𝑇 has an upper-triangular matrix with respect to some basis of 𝑉.
Proof The desired result follows immediately from 5.44 and the second version
of the fundamental theorem of algebra (see 4.13).
For an extension of the result above to two operators 𝑆 and 𝑇 such that
𝑆𝑇 = 𝑇𝑆,
see 5.80. Also, for an extension to more than two operators, see Exercise 9(b) in
Section 5E.
Caution: If an operator 𝑇 ∈ ℒ(𝑉) has a upper-triangular matrix with respect
to some basis 𝑣1 , …, 𝑣𝑛 of 𝑉, then the eigenvalues of 𝑇 are exactly the entries on
the diagonal of ℳ(𝑇), as shown by 5.41, and furthermore 𝑣1 is an eigenvector of
𝑇. However, 𝑣2 , …, 𝑣𝑛 need not be eigenvectors of 𝑇. Indeed, a basis vector 𝑣𝑘 is
an eigenvector of 𝑇 if and only if all entries in the 𝑘 th column of the matrix of 𝑇
are 0, except possibly the 𝑘 th entry.
You may recall from a previous The row echelon form of the matrix
course that every matrix of numbers can of an operator does not give us a list
be changed to a matrix in what is called of the eigenvalues of the operator. In
row echelon form. If one begins with a contrast, an upper-triangular matrix
square matrix, the matrix in row echelon with respect to some basis gives us a
form will be an upper-triangular matrix. list of all the eigenvalues of the op-
Do not confuse this upper-triangular ma- erator. However, there is no method
trix with the upper-triangular matrix of for computing exactly such an upper-
an operator with respect to some basis triangular matrix, even though 5.47
whose existence is proclaimed by 5.47 (if guarantees its existence if 𝐅 = 𝐂 .
𝐅 = 𝐂 )—there is no connection between
these upper-triangular matrices.
Exercises 5C
1 Prove or give a counterexample: If 𝑇 ∈ ℒ(𝑉) and 𝑇 2 has an upper-triangular

matrix with respect to some basis of 𝑉, then 𝑇 has an upper-triangular matrix
with respect to some basis of 𝑉.
2 Suppose 𝐴 and 𝐵 are upper-triangular matrices of the same size, with

𝛼1 , …, 𝛼𝑛 on the diagonal of 𝐴 and 𝛽1 , …, 𝛽𝑛 on the diagonal of 𝐵.
(a) Show that 𝐴 + 𝐵 is an upper-triangular matrix with 𝛼1 + 𝛽1 , …, 𝛼𝑛 + 𝛽𝑛
on the diagonal.
(b) Show that 𝐴𝐵 is an upper-triangular matrix with 𝛼1 𝛽1 , …, 𝛼𝑛 𝛽𝑛 on the
diagonal.
The results in this exercise are used in the proof of 5.81.
3 Suppose 𝑇 ∈ ℒ(𝑉) is invertible and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 with respect

to which the matrix of 𝑇 is upper triangular, with 𝜆1 , …, 𝜆𝑛 on the diagonal.
Show that the matrix of 𝑇 −1 is also upper triangular with respect to the basis
𝑣1 , …, 𝑣𝑛 , with
1 1
, …,
𝜆1 𝜆𝑛
on the diagonal.
4 Give an example of an operator whose matrix with respect to some basis
contains only 0’s on the diagonal, but the operator is invertible.
This exercise and the exercise below show that 5.41 fails without the hypoth-
esis that an upper-triangular matrix is under consideration.
5 Give an example of an operator whose matrix with respect to some basis

contains only nonzero numbers on the diagonal, but the operator is not
invertible.
6 Suppose 𝐅 = 𝐂 , 𝑉 is finite-dimensional, and 𝑇 ∈ ℒ(𝑉). Prove that if
𝑘 ∈ {1, …, dim 𝑉}, then 𝑉 has a 𝑘-dimensional subspace invariant under 𝑇.
7 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝑣 ∈ 𝑉.

(a) Prove that there exists a unique monic polynomial 𝑝𝑣 of smallest degree
such that 𝑝𝑣 (𝑇)𝑣 = 0.
(b) Prove that the minimal polynomial of 𝑇 is a polynomial multiple of 𝑝𝑣 .
8 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and there exists a nonzero

vector 𝑣 ∈ 𝑉 such that 𝑇 2 𝑣 + 2𝑇𝑣 = −2𝑣.
(a) Prove that if 𝐅 = 𝐑 , then there does not exist a basis of 𝑉 with respect
to which 𝑇 has an upper-triangular matrix.
(b) Prove that if 𝐅 = 𝐂 and 𝐴 is an upper-triangular matrix that equals
the matrix of 𝑇 with respect to some basis of 𝑉, then −1 + 𝑖 or −1 − 𝑖
appears on the diagonal of 𝐴.
9 Suppose 𝐵 is a square matrix with complex entries. Prove that there exists
an invertible square matrix 𝐴 with complex entries such that 𝐴−1 𝐵𝐴 is an
upper-triangular matrix.
10 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Show that the following

are equivalent.
(a) The matrix of 𝑇 with respect to 𝑣1 , …, 𝑣𝑛 is lower triangular.
(b) span(𝑣𝑘 , …, 𝑣𝑛 ) is invariant under 𝑇 for each 𝑘 = 1, …, 𝑛.
(c) 𝑇𝑣𝑘 ∈ span(𝑣𝑘 , …, 𝑣𝑛 ) for each 𝑘 = 1, …, 𝑛.
A square matrix is called lower triangular if all entries above the diagonal
are 0.
11 Suppose 𝐅 = 𝐂 and 𝑉 is finite-dimensional. Prove that if 𝑇 ∈ ℒ(𝑉), then

there exists a basis of 𝑉 with respect to which 𝑇 has a lower-triangular matrix.
12 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉) has an upper-triangular matrix
with respect to some basis of 𝑉, and 𝑈 is a subspace of 𝑉 that is invariant
under 𝑇.
(a) Prove that 𝑇|𝑈 has an upper-triangular matrix with respect to some basis
of 𝑈.
(b) Prove that the quotient operator 𝑇/𝑈 has an upper-triangular matrix with
respect to some basis of 𝑉/𝑈.
13 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Suppose there exists

a subspace 𝑈 of 𝑉 that is invariant under 𝑇 such that 𝑇|𝑈 has an upper-
triangular matrix with respect to some basis of 𝑈 and also 𝑇/𝑈 has an
upper-triangular matrix with respect to some basis of 𝑉/𝑈. Prove that 𝑇 has
an upper-triangular matrix with respect to some basis of 𝑉.
14 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 has an upper-
triangular matrix with respect to some basis of 𝑉 if and only if the dual
operator 𝑇 ′ has an upper-triangular matrix with respect to some basis of the
dual space 𝑉 ′.
Section 5D Diagonalizable Operators 163
5D Diagonalizable Operators
Diagonal Matrices
5.48 definition: diagonal matrix
A diagonal matrix is a square matrix that is 0 everywhere except possibly on

the diagonal.
5.49 example: diagonal matrix

8 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ 0 5 0 ⎟
⎟
⎝ 0 0 5 ⎠
is a diagonal matrix.
If an operator has a diagonal matrix Every diagonal matrix is upper tri-

with respect to some basis, then the en- angular. Diagonal matrices typically
tries on the diagonal are precisely the have many more 0’s than most upper-
eigenvalues of the operator; this follows triangular matrices of the same size.
from 5.41 (or find an easier direct proof
for diagonal matrices).
5.50 definition: diagonalizable
An operator on 𝑉 is called diagonalizable if the operator has a diagonal matrix

with respect to some basis of 𝑉.
5.51 example: diagonalization may require a different basis

Define 𝑇 ∈ ℒ(𝐑2 ) by
𝑇(𝑥, 𝑦) = (41𝑥 + 7𝑦, −20𝑥 + 74𝑦).
The matrix of 𝑇 with respect to the standard basis of 𝐑2 is
41 7
( ),
−20 74
which is not a diagonal matrix. However, 𝑇 is diagonalizable. Specifically, the
matrix of 𝑇 with respect to the basis (1, 4), (7, 5) is
69 0
( )
0 46
because 𝑇(1, 4) = (69, 276) = 69(1, 4) and 𝑇(7, 5) = (322, 230) = 46(7, 5).
For 𝜆 ∈ 𝐅, we will find it convenient to have a name and a notation for the set
of vectors that an operator 𝑇 maps to 𝜆 times the vector.
5.52 definition: eigenspace, 𝐸(𝜆, 𝑇)
Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆 ∈ 𝐅. The eigenspace of 𝑇 corresponding to 𝜆 is

the subspace 𝐸(𝜆, 𝑇) of 𝑉 defined by
𝐸(𝜆, 𝑇) = null(𝑇 − 𝜆𝐼) = {𝑣 ∈ 𝑉 ∶ 𝑇𝑣 = 𝜆𝑣}.
Hence 𝐸(𝜆, 𝑇) is the set of all eigenvectors of 𝑇 corresponding to 𝜆, along

with the 0 vector.
For 𝑇 ∈ ℒ(𝑉) and 𝜆 ∈ 𝐅, the set 𝐸(𝜆, 𝑇) is a subspace of 𝑉 because the null
space of each linear map on 𝑉 is a subspace of 𝑉. The definitions imply that 𝜆 is
an eigenvalue of 𝑇 if and only if 𝐸(𝜆, 𝑇) ≠ {0}.
5.53 example: eigenspaces of an operator

Suppose the matrix of an operator 𝑇 ∈ ℒ(𝑉) with respect to a basis 𝑣1 , 𝑣2 , 𝑣3
of 𝑉 is the matrix in Example 5.49. Then
𝐸(8, 𝑇) = span(𝑣1 ), 𝐸(5, 𝑇) = span(𝑣2 , 𝑣3 ).
If 𝜆 is an eigenvalue of an operator 𝑇 ∈ ℒ(𝑉), then 𝑇 restricted to 𝐸(𝜆, 𝑇) is

just the operator of multiplication by 𝜆.
5.54 sum of eigenspaces is a direct sum
Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆1 , …, 𝜆𝑚 are distinct eigenvalues of 𝑇. Then
𝐸(𝜆1 , 𝑇) + ⋯ + 𝐸(𝜆𝑚 , 𝑇)
is a direct sum. Furthermore, if 𝑉 is finite-dimensional, then
dim 𝐸(𝜆1 , 𝑇) + ⋯ + dim 𝐸(𝜆𝑚 , 𝑇) ≤ dim 𝑉.
Proof To show that 𝐸(𝜆1 , 𝑇) + ⋯ + 𝐸(𝜆𝑚 , 𝑇) is a direct sum, suppose

𝑣1 + ⋯ + 𝑣𝑚 = 0,
where each 𝑣𝑘 is in 𝐸(𝜆𝑘 , 𝑇). Because eigenvectors corresponding to distinct
eigenvalues are linearly independent (by 5.11), this implies that each 𝑣𝑘 equals 0.
Thus 𝐸(𝜆1 , 𝑇) + ⋯ + 𝐸(𝜆𝑚 , 𝑇) is a direct sum (by 1.45), as desired.
Now suppose 𝑉 is finite-dimensional. Then
dim 𝐸(𝜆1 , 𝑇) + ⋯ + dim 𝐸(𝜆𝑚 , 𝑇) = dim(𝐸(𝜆1 , 𝑇) ⊕ ⋯ ⊕ 𝐸(𝜆𝑚 , 𝑇))
≤ dim 𝑉,
where the first line follows from 3.94 and the second line follows from 2.37.
Conditions for Diagonalizability

The following characterizations of diagonalizable operators will be useful.
5.55 conditions equivalent to diagonalizability
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Let 𝜆1 , …, 𝜆𝑚 denote the

distinct eigenvalues of 𝑇. Then the following are equivalent.
(a) 𝑇 is diagonalizable.
(b) 𝑉 has a basis consisting of eigenvectors of 𝑇.
(c) 𝑉 = 𝐸(𝜆1 , 𝑇) ⊕ ⋯ ⊕ 𝐸(𝜆𝑚 , 𝑇).
(d) dim 𝑉 = dim 𝐸(𝜆1 , 𝑇) + ⋯ + dim 𝐸(𝜆𝑚 , 𝑇).
Proof An operator 𝑇 ∈ ℒ(𝑉) has a diagonal matrix

𝜆1 0
⎛
⎜ ⎞
⎟
⎜
⎜ ⋱ ⎟
⎟
⎝ 0 𝜆𝑛 ⎠
with respect to a basis 𝑣1 , …, 𝑣𝑛 of 𝑉 if and only if 𝑇𝑣𝑘 = 𝜆𝑘 𝑣𝑘 for each 𝑘. Thus
(a) and (b) are equivalent.
Suppose (b) holds; thus 𝑉 has a basis consisting of eigenvectors of 𝑇. Hence
every vector in 𝑉 is a linear combination of eigenvectors of 𝑇, which implies that
𝑉 = 𝐸(𝜆1 , 𝑇) + ⋯ + 𝐸(𝜆𝑚 , 𝑇).
Now 5.54 shows that (c) holds, proving that (b) implies (c).
That (c) implies (d) follows immediately from 3.94.
Finally, suppose (d) holds; thus
5.56 dim 𝑉 = dim 𝐸(𝜆1 , 𝑇) + ⋯ + dim 𝐸(𝜆𝑚 , 𝑇).
Choose a basis of each 𝐸(𝜆𝑘 , 𝑇); put all these bases together to form a list 𝑣1 , …, 𝑣𝑛
of eigenvectors of 𝑇, where 𝑛 = dim 𝑉 (by 5.56). To show that this list is linearly
independent, suppose
𝑎1 𝑣1 + ⋯ + 𝑎𝑛 𝑣𝑛 = 0,
where 𝑎1 , …, 𝑎𝑛 ∈ 𝐅. For each 𝑘 = 1, …, 𝑚, let 𝑢𝑘 denote the sum of all the terms
𝑎𝑗 𝑣𝑗 such that 𝑣𝑗 ∈ 𝐸(𝜆𝑘 , 𝑇). Thus each 𝑢𝑘 is in 𝐸(𝜆𝑘 , 𝑇), and
𝑢1 + ⋯ + 𝑢𝑚 = 0.
Because eigenvectors corresponding to distinct eigenvalues are linearly indepen-
dent (see 5.11), this implies that each 𝑢𝑘 equals 0. Because each 𝑢𝑘 is a sum of
terms 𝑎𝑗 𝑣𝑗 , where the 𝑣𝑗 ’s were chosen to be a basis of 𝐸(𝜆𝑘 , 𝑇), this implies that
all 𝑎𝑗 ’s equal 0. Thus 𝑣1 , …, 𝑣𝑛 is linearly independent and hence is a basis of 𝑉
(by 2.38). Thus (d) implies (b), completing the proof.
For additional conditions equivalent to diagonalizability, see 5.62, Exercises 5

and 15 in this section, Exercise 24 in Section 7B, and Exercise 15 in Section 8A.
As we know, every operator on a finite-dimensional complex vector space

has an eigenvalue. However, not every operator on a finite-dimensional complex
vector space has enough eigenvectors to be diagonalizable, as shown by the next
example.
5.57 example: an operator that is not diagonalizable

Define an operator 𝑇 ∈ ℒ(𝐅3 ) by 𝑇(𝑎, 𝑏, 𝑐) = (𝑏, 𝑐, 0). The matrix of 𝑇 with
respect to the standard basis of 𝐅3 is
0 1 0
⎛
⎜ ⎞
⎟
⎜
⎜ 0 0 1 ⎟,
⎟
⎝ 0 0 0 ⎠
which is an upper-triangular matrix but is not a diagonal matrix.
As you should verify, 0 is the only eigenvalue of 𝑇 and furthermore
𝐸(0, 𝑇) = {(𝑎, 0, 0) ∈ 𝐅3 ∶ 𝑎 ∈ 𝐅}.
Hence conditions (b), (c), and (d) of 5.55 fail (of course, because these conditions
are equivalent, it is sufficient to check that only one of them fails). Thus condition
(a) of 5.55 also fails. Hence 𝑇 is not diagonalizable, regardless of whether 𝐅 = 𝐑
or 𝐅 = 𝐂 .
The next result shows that if an operator has as many distinct eigenvalues as
the dimension of its domain, then the operator is diagonalizable.
5.58 enough eigenvalues implies diagonalizability
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉) has dim 𝑉 distinct eigenvalues.

Then 𝑇 is diagonalizable.
Proof Suppose 𝑇 has distinct eigenvalues 𝜆1 , …, 𝜆dim 𝑉 . For each 𝑘, let 𝑣𝑘 ∈ 𝑉

be an eigenvector corresponding to the eigenvalue 𝜆𝑘 . Because eigenvectors corre-
sponding to distinct eigenvalues are linearly independent (see 5.11), 𝑣1 , …, 𝑣dim 𝑉
A linearly independent list of dim 𝑉 vectors in 𝑉 is a basis of 𝑉 (see 2.38); thus
𝑣1 , …, 𝑣dim 𝑉 is a basis of 𝑉. With respect to this basis consisting of eigenvectors,
𝑇 has a diagonal matrix.
In later chapters we will find additional conditions that imply that certain
operators are diagonalizable. For example, see the real spectral theorem (7.29)
and the complex spectral theorem (7.31).
The result above gives a sufficient condition for an operator to be diagonal-
izable. However, this condition is not necessary. For example, the operator 𝑇
on 𝐅3 defined by 𝑇(𝑥, 𝑦, 𝑧) = (6𝑥, 6𝑦, 7𝑧) has only two eigenvalues (6 and 7) and
dim 𝐅3 = 3, but 𝑇 is diagonalizable (by the standard basis of 𝐅3 ).
The next example illustrates the im- For a spectacular application of these
portance of diagonalization, which can techniques, see Exercise 21, which
be used to compute high powers of an shows how to use diagonalization to
operator, taking advantage of the equa- find an exact formula for the 𝑛th term
tion 𝑇 𝑘 𝑣 = 𝜆𝑘 𝑣 if 𝑣 is an eigenvector of of the Fibonacci sequence.
𝑇 with eigenvalue 𝜆.
5.59 example: using diagonalization to compute 𝑇 100

Define 𝑇 ∈ ℒ(𝐅3 ) by 𝑇(𝑥, 𝑦, 𝑧) = (2𝑥 + 𝑦, 5𝑦 + 3𝑧, 8𝑧). With respect to the
standard basis, the matrix of 𝑇 is
2 1 0
⎛
⎜ ⎞
⎟
⎜
⎜ 0 5 3 ⎟
⎟.
⎝ 0 0 8 ⎠
The matrix above is an upper-triangular matrix but it is not a diagonal matrix. By
5.41, the eigenvalues of 𝑇 are 2, 5, and 8. Because 𝑇 is an operator on a vector
space of dimension three and 𝑇 has three distinct eigenvalues, 5.58 assures us that
there exists a basis of 𝐅3 with respect to which 𝑇 has a diagonal matrix.
To find this basis, we only have to find an eigenvector for each eigenvalue. In
other words, we have to find a nonzero solution to the equation
𝑇(𝑥, 𝑦, 𝑧) = 𝜆(𝑥, 𝑦, 𝑧)
for 𝜆 = 2, then for 𝜆 = 5, and then for 𝜆 = 8. Solving these simple equations
shows that for 𝜆 = 2 we have an eigenvector (1, 0, 0), for 𝜆 = 5 we have an
eigenvector (1, 3, 0), and for 𝜆 = 8 we have an eigenvector (1, 6, 6).
Thus (1, 0, 0), (1, 3, 0), (1, 6, 6) is a basis of 𝐅3 consisting of eigenvectors of 𝑇,
and with respect to this basis the matrix of 𝑇 is the diagonal matrix
2 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ 0 5 0 ⎟
⎟.
⎝ 0 0 8 ⎠
To compute 𝑇 100 (0, 0, 1), for example, write (0, 0, 1) as a linear combination
of our basis of eigenvectors:
(0, 0, 1) = 16 (1, 0, 0) − 31 (1, 3, 0) + 16 (1, 6, 6).
Now apply 𝑇 100 to both sides of the equation above, getting
𝑇 100 (0, 0, 1) = 61 (𝑇 100 (1, 0, 0)) − 13 (𝑇 100 (1, 3, 0)) + 61 (𝑇 100 (1, 6, 6))
= 16 (2100 (1, 0, 0) − 2 ⋅ 5100 (1, 3, 0) + 8100 (1, 6, 6))
= 16 (2100 − 2 ⋅ 5100 + 8100 , 6 ⋅ 8100 − 6 ⋅ 5100 , 6 ⋅ 8100 ).

We saw earlier that an operator 𝑇 on a finite-dimensional vector space 𝑉 has an

upper-triangular matrix with respect to some basis of 𝑉 if and only if the minimal
polynomial of 𝑇 equals (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ) for some 𝜆1 , …, 𝜆𝑚 ∈ 𝐅 (see 5.44).
As we previously noted (see 5.47), this condition is always satisfied if 𝐅 = 𝐂 .
Our next result 5.62 states that an operator 𝑇 ∈ ℒ(𝑉) has a diagonal matrix
with respect to some basis of 𝑉 if and only if the minimal polynomial of 𝑇 equals
(𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ) for some distinct 𝜆1 , …, 𝜆𝑚 ∈ 𝐅. Before formally stating this
result, we give two examples of using it.
5.60 example: diagonalizable, but with no known exact eigenvalues

Define 𝑇 ∈ ℒ(𝐂5 ) by
𝑇(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 , 𝑧5 ) = (−3𝑧5 , 𝑧1 + 6𝑧5 , 𝑧2 , 𝑧3 , 𝑧4 ).
The matrix of 𝑇 is shown in Example 5.26, where we showed that the minimal
polynomial of 𝑇 is 3 − 6𝑧 + 𝑧5.
As mentioned in Example 5.28, no exact expression is known for any of the
zeros of this polynomial, but numeric techniques show that the zeros of this
polynomial are approximately −1.67, 0.51, 1.40, −0.12 + 1.59𝑖, −0.12 − 1.59𝑖.
The software that produces these approximations is accurate to more than
three digits. Thus these approximations are good enough to show that the five
numbers above are distinct. The minimal polynomial of 𝑇 equals the fifth degree
monic polynomial with these zeros. Now 5.62 shows that 𝑇 is diagonalizable.
5.61 example: showing that an operator is not diagonalizable

Define 𝑇 ∈ ℒ(𝐅3 ) by
𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (6𝑧1 + 3𝑧2 + 4𝑧3 , 6𝑧2 + 2𝑧3 , 7𝑧3 ).
The matrix of 𝑇 with respect to the standard basis of 𝐅3 is
6 3 4
⎛
⎜ ⎞
⎟
⎜
⎜ 0 6 2 ⎟
⎟.
⎝ 0 0 7 ⎠
The matrix above is an upper-triangular matrix but is not a diagonal matrix. Might
𝑇 have a diagonal matrix with respect to some other basis of 𝐅3 ?
To answer this question, we will find the minimal polynomial of 𝑇. First note
that the eigenvalues of 𝑇 are the diagonal entries of the matrix above (by 5.41).
Thus the zeros of the minimal polynomial of 𝑇 are 6, 7 [by 5.27(a)]. The diagonal
of the matrix above tells us that (𝑇 − 6𝐼)2 (𝑇 − 7𝐼) = 0 (by 5.40). The minimal
polynomial of 𝑇 has degree at most 3 (by 5.22). Putting all this together, we see
that the minimal polynomial of 𝑇 is either (𝑧 − 6)(𝑧 − 7) or (𝑧 − 6)2 (𝑧 − 7).
A simple computation shows that (𝑇 − 6𝐼)(𝑇 − 7𝐼) ≠ 0. Thus the minimal
polynomial of 𝑇 is (𝑧 − 6)2 (𝑧 − 7).
Now 5.62 shows that 𝑇 is not diagonalizable.
5.62 necessary and sufficient condition for diagonalizability
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Then 𝑇 is diagonalizable if

and only if the minimal polynomial of 𝑇 equals (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ) for some
list of distinct numbers 𝜆1 , …, 𝜆𝑚 ∈ 𝐅.
Proof First suppose 𝑇 is diagonalizable. Thus there is a basis 𝑣1 , …, 𝑣𝑛 of 𝑉

consisting of eigenvectors of 𝑇. Let 𝜆1 , …, 𝜆𝑚 be the distinct eigenvalues of 𝑇.
Then for each 𝑣𝑗 , there exists 𝜆𝑘 with (𝑇 − 𝜆𝑘 𝐼)𝑣𝑗 = 0. Thus
(𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑚 𝐼)𝑣𝑗 = 0,
which implies that the minimal polynomial of 𝑇 equals (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ).
To prove the implication in the other direction, now suppose the minimal
polynomial of 𝑇 equals (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ) for some list of distinct numbers
𝜆1 , …, 𝜆𝑚 ∈ 𝐅. Thus
5.63 (𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑚 𝐼) = 0.
We will prove that 𝑇 is diagonalizable by induction on 𝑚. To get started,
suppose 𝑚 = 1. Then 𝑇 − 𝜆1 𝐼 = 0, which means that 𝑇 is a scalar multiple of the
identity operator, which implies that 𝑇 is diagonalizable.
Now suppose that 𝑚 > 1 and the desired result holds for all smaller values of
𝑚. The subspace range(𝑇 − 𝜆𝑚 𝐼) is invariant under 𝑇 [this is a special case of
5.18 with 𝑝(𝑧) = 𝑧 − 𝜆𝑚 ]. Thus 𝑇 restricted to range(𝑇 − 𝜆𝑚 𝐼) is an operator on
range(𝑇 − 𝜆𝑚 𝐼).
If 𝑢 ∈ range(𝑇 − 𝜆𝑚 𝐼), then 𝑢 = (𝑇 − 𝜆𝑚 𝐼)𝑣 for some 𝑣 ∈ 𝑉, and 5.63 implies
5.64 (𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑚 − 1 𝐼)𝑢 = (𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑚 𝐼)𝑣 = 0.
Hence (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 − 1 ) is a polynomial multiple of the minimal polynomial
of 𝑇 restricted to range(𝑇 − 𝜆𝑚 𝐼) [by 5.29]. Thus by our induction hypothesis,
there is a basis of range(𝑇 − 𝜆𝑚 𝐼) consisting of eigenvectors of 𝑇.
Suppose that 𝑢 ∈ range(𝑇 − 𝜆𝑚 𝐼) ∩ null(𝑇 − 𝜆𝑚 𝐼). Then 𝑇𝑢 = 𝜆𝑚 𝑢. Now
5.64 implies that
0 = (𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑚 − 1 𝐼)𝑢
= (𝜆𝑚 − 𝜆1 )⋯(𝜆𝑚 − 𝜆𝑚 − 1 )𝑢.
Because 𝜆1 , …, 𝜆𝑚 are distinct, the equation above implies that 𝑢 = 0. Hence
range(𝑇 − 𝜆𝑚 𝐼) ∩ null(𝑇 − 𝜆𝑚 𝐼) = {0}.
Thus range(𝑇− 𝜆𝑚 𝐼)+null(𝑇− 𝜆𝑚 𝐼) is a direct sum (by 1.46) whose dimension
is dim 𝑉 (by 3.94 and 3.21). Hence range(𝑇 − 𝜆𝑚 𝐼) ⊕ null(𝑇 − 𝜆𝑚 𝐼) = 𝑉. Every
vector in null(𝑇 − 𝜆𝑚 𝐼) is an eigenvector of 𝑇 with eigenvalue 𝜆𝑚 . Earlier in this
proof we saw that there is a basis of range(𝑇 − 𝜆𝑚 𝐼) consisting of eigenvectors
of 𝑇. Adjoining to that basis a basis of null(𝑇 − 𝜆𝑚 𝐼) gives a basis of 𝑉 consisting
of eigenvectors of 𝑇. The matrix of 𝑇 with respect to this basis is a diagonal
matrix, as desired.
No formula exists for the zeros of polynomials of degree 5 or greater. However,

the previous result can be used to determine whether an operator on a complex
vector space is diagonalizable without even finding approximations of the zeros
of the minimal polynomial—see Exercise 15.
The next result will be a key tool when we prove a result about the simul-
taneous diagonalization of two operators; see 5.76. Note how the use of the
characterization of diagonalizable operators in terms of the minimal polynomial
(see 5.62) leads to a short proof of the next result.
5.65 restriction of diagonalizable operator to invariant subspace
Suppose 𝑇 ∈ ℒ(𝑉) is diagonalizable and 𝑈 is a subspace of 𝑉 that is invariant

under 𝑇. Then 𝑇|𝑈 is a diagonalizable operator on 𝑈.
Proof Because the operator 𝑇 is diagonalizable, the minimal polynomial of 𝑇

equals (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ) for some list of distinct numbers 𝜆1 , …, 𝜆𝑚 ∈ 𝐅 (by
5.62). The minimal polynomial of 𝑇 is a polynomial multiple of the minimal
polynomial of 𝑇|𝑈 (by 5.31). Hence the minimal polynomial of 𝑇|𝑈 has the form
required by 5.62, which shows that 𝑇|𝑈 is diagonalizable.
Gershgorin Disk Theorem
5.66 definition: Gershgorin disks
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Let 𝐴 denote the matrix of

𝑇 with respect to this basis. A Gershgorin disk of 𝑇 with respect to the basis
𝑣1 , …, 𝑣𝑛 is a set of the form
𝑛
{𝑧 ∈ 𝐅 ∶ |𝑧 − 𝐴𝑗, 𝑗 | ≤ ∑ |𝐴𝑗, 𝑘 |},
𝑘=1
𝑘≠𝑗
where 𝑗 ∈ {1, …, 𝑛}.
Because there are 𝑛 choices for 𝑗 in the definition above, 𝑇 has 𝑛 Gershgorin
disks. If 𝐅 = 𝐂 , then for each 𝑗 ∈ {1, …, 𝑛}, the corresponding Gershgorin disk
is a closed disk in 𝐂 centered at 𝐴𝑗, 𝑗 , which is the 𝑗th entry on the diagonal of 𝐴.
The radius of this closed disk is the sum of the absolute values of the entries in
row 𝑗 of 𝐴, excluding the diagonal entry. If 𝐅 = 𝐑 , then the Gershgorin disks are
closed intervals in 𝐑 .
In the special case that the square matrix 𝐴 above is a diagonal matrix, each
Gershgorin disk consists of a single point that is a diagonal entry of 𝐴 (and
each eigenvalue of 𝑇 is one of those points, as required by the next result). One
consequence of our next result is that if the nondiagonal entries of 𝐴 are small,
then each eigenvalue of 𝑇 is near a diagonal entry of 𝐴.
5.67 Gershgorin disk theorem
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Then each eigenvalue of 𝑇

is contained in some Gershgorin disk of 𝑇 with respect to the basis 𝑣1 , …, 𝑣𝑛 .
Proof Suppose 𝜆 ∈ 𝐅 is an eigenvalue of 𝑇. Let 𝑤 ∈ 𝑉 be a corresponding

eigenvector. There exist 𝑐1 , …, 𝑐𝑛 ∈ 𝐅 such that
5.68 𝑤 = 𝑐1 𝑣1 + ⋯ + 𝑐𝑛 𝑣𝑛 .
Let 𝐴 denote the matrix of 𝑇 with respect to the basis 𝑣1 , …, 𝑣𝑛 . Applying 𝑇

to both sides of the equation above gives
𝑛
5.69 𝜆𝑤 = ∑ 𝑐𝑘 𝑇𝑣𝑘
𝑘=1
𝑛 𝑛
= ∑ 𝑐𝑘 ∑ 𝐴𝑗, 𝑘 𝑣𝑗
𝑘=1 𝑗=1
𝑛 𝑛
5.70 = ∑ ( ∑ 𝐴𝑗, 𝑘 𝑐𝑘 )𝑣𝑗 .
𝑗 = 1 𝑘=1
Let 𝑗 ∈ {1, …, 𝑛} be such that

|𝑐𝑗 | = max{|𝑐1 |, …, |𝑐𝑛 |}.
Using 5.68, we see that the coefficient of 𝑣𝑗 on the left side of 5.69 equals 𝜆𝑐𝑗 ,
which must equal the coefficient of 𝑣𝑗 on the right side of 5.70. In other words,
𝑛
𝜆𝑐𝑗 = ∑ 𝐴𝑗, 𝑘 𝑐𝑘 .
𝑘=1
Subtract 𝐴𝑗, 𝑗 𝑐𝑗 from each side of the equation above and then divide both sides
by 𝑐𝑗 to get
𝑛 𝑐𝑘
|𝜆 − 𝐴𝑗, 𝑗 | = ∣ ∑ 𝐴𝑗, 𝑘 ∣
𝑘=1 𝑐𝑗
𝑘≠𝑗
𝑛
≤ ∑ |𝐴𝑗, 𝑘 |.
𝑘=1
𝑘≠𝑗
Thus 𝜆 is in the 𝑗th Gershgorin disk with respect to the basis 𝑣1 , …, 𝑣𝑛 .
Exercise 22 gives a nice application The Gershgorin disk theorem is named

of the Gershgorin disk theorem. for Semyon Aronovich Gershgorin,
Exercise 23 states that the radius of who published this result in 1931.
each Gershgorin disk could be changed
to the sum of the absolute values of corresponding column entries (instead of row
entries), excluding the diagonal entry, and the theorem above would still hold.
Exercises 5D
1 Suppose 𝑉 is a finite-dimensional complex vector space and 𝑇 ∈ ℒ(𝑉).

(a) Prove that if 𝑇 4 = 𝐼, then 𝑇 is diagonalizable.
(b) Prove that if 𝑇 4 = 𝑇, then 𝑇 is diagonalizable.
(c) Give an example of an operator 𝑇 ∈ ℒ(𝐂2 ) such that 𝑇 4 = 𝑇 2 and 𝑇 is
not diagonalizable.
2 Suppose 𝑇 ∈ ℒ(𝑉) has a diagonal matrix 𝐴 with respect to some basis

of 𝑉. Prove that if 𝜆 ∈ 𝐅, then 𝜆 appears on the diagonal of 𝐴 precisely
dim 𝐸(𝜆, 𝑇) times.
3 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that if the operator 𝑇
is diagonalizable, then 𝑉 = null 𝑇 ⊕ range 𝑇.
4 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that the following
are equivalent.
(a) 𝑉 = null 𝑇 ⊕ range 𝑇.
(b) 𝑉 = null 𝑇 + range 𝑇.
(c) null 𝑇 ∩ range 𝑇 = {0}.
5 Suppose 𝑉 is a finite-dimensional complex vector space and 𝑇 ∈ ℒ(𝑉).

Prove that 𝑇 is diagonalizable if and only if
𝑉 = null(𝑇 − 𝜆𝐼) ⊕ range(𝑇 − 𝜆𝐼)
for every 𝜆 ∈ 𝐂 .
6 Suppose 𝑇 ∈ ℒ(𝐅5 ) and dim 𝐸(8, 𝑇) = 4. Prove that 𝑇 − 2𝐼 or 𝑇 − 6𝐼 is
invertible.
7 Suppose 𝑇 ∈ ℒ(𝑉) is invertible. Prove that
𝐸(𝜆, 𝑇) = 𝐸( 1𝜆 , 𝑇 −1 )
for every 𝜆 ∈ 𝐅 with 𝜆 ≠ 0.

8 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Let 𝜆1 , …, 𝜆𝑚 denote the
distinct nonzero eigenvalues of 𝑇. Prove that
dim 𝐸(𝜆1 , 𝑇) + ⋯ + dim 𝐸(𝜆𝑚 , 𝑇) ≤ dim range 𝑇.
9 Suppose 𝑅, 𝑇 ∈ ℒ(𝐅3 ) each have 2, 6, 7 as eigenvalues. Prove that there

exists an invertible operator 𝑆 ∈ ℒ(𝐅3 ) such that 𝑅 = 𝑆−1 𝑇𝑆.
10 Find 𝑅, 𝑇 ∈ ℒ(𝐅4 ) such that 𝑅 and 𝑇 each have 2, 6, 7 as eigenvalues, 𝑅 and
𝑇 have no other eigenvalues, and there does not exist an invertible operator
𝑆 ∈ ℒ(𝐅4 ) such that 𝑅 = 𝑆−1 𝑇𝑆.
11 Find 𝑇 ∈ ℒ(𝐂3 ) such that 6 and 7 are eigenvalues of 𝑇 and such that 𝑇 does
not have a diagonal matrix with respect to any basis of 𝐂3.
12 Suppose 𝑇 ∈ ℒ(𝐂3 ) is such that 6 and 7 are eigenvalues of 𝑇. Furthermore,
suppose 𝑇 does not have a diagonal matrix with respect to any basis of 𝐂3.
Prove that there exists (𝑧1 , 𝑧2 , 𝑧3 ) ∈ 𝐂3 such that
𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (6 + 8𝑧1 , 7 + 8𝑧2 , 13 + 8𝑧3 ).
13 Suppose 𝐴 is a diagonal matrix with distinct entries on the diagonal and 𝐵

is a matrix of the same size as 𝐴. Show that 𝐴𝐵 = 𝐵𝐴 if and only if 𝐵 is a
diagonal matrix.
14 (a) Give an example of a finite-dimensional complex vector space and an
operator 𝑇 on that vector space such that 𝑇 2 is diagonalizable but 𝑇 is
not diagonalizable.
(b) Suppose 𝐅 = 𝐂 , 𝑘 is a positive integer, and 𝑇 ∈ ℒ(𝑉) is invertible.
Prove that 𝑇 is diagonalizable if and only if 𝑇 𝑘 is diagonalizable.
15 Suppose 𝑉 is a finite-dimensional complex vector space, 𝑇 ∈ ℒ(𝑉), and 𝑝
is the minimal polynomial of 𝑇. Prove that the following are equivalent.
(a) 𝑇 is diagonalizable.
(b) There does not exist 𝜆 ∈ 𝐂 such that 𝑝 is a polynomial multiple of
(𝑧 − 𝜆)2.
(c) 𝑝 and its derivative 𝑝′ have no zeros in common.
(d) The greatest common divisor of 𝑝 and 𝑝′ is the constant polynomial 1.
The greatest common divisor of 𝑝 and 𝑝′ is the monic polynomial 𝑞 of
largest degree such that 𝑝 and 𝑝′ are both polynomial multiples of 𝑞. The
Euclidean algorithm for polynomials (look it up) can quickly determine
the greatest common divisor of two polynomials, without requiring any
information about the zeros of the polynomials. Thus the equivalence of (a)
and (d) above shows that we can determine whether 𝑇 is diagonalizable
without knowing anything about the zeros of 𝑝.
16 Suppose that 𝑇 ∈ ℒ(𝑉) is diagonalizable. Let 𝜆1 , …, 𝜆𝑚 denote the distinct

eigenvalues of 𝑇. Prove that a subspace 𝑈 of 𝑉 is invariant under 𝑇 if and
only if there exist subspaces 𝑈1 , …, 𝑈𝑚 of 𝑉 such that 𝑈𝑘 ⊆ 𝐸(𝜆𝑘 , 𝑇) for
each 𝑘 and 𝑈 = 𝑈1 ⊕ ⋯ ⊕ 𝑈𝑚 .
17 Suppose 𝑉 is finite-dimensional. Prove that ℒ(𝑉) has a basis consisting of
diagonalizable operators.
18 Suppose that 𝑇 ∈ ℒ(𝑉) is diagonalizable and 𝑈 is a subspace of 𝑉 that is
invariant under 𝑇. Prove that the quotient operator 𝑇/𝑈 is a diagonalizable
operator on 𝑉/𝑈.
19 Prove or give a counterexample: If 𝑇 ∈ ℒ(𝑉) and there exists a subspace 𝑈

of 𝑉 that is invariant under 𝑇 such that 𝑇|𝑈 and 𝑇/𝑈 are both diagonalizable,
then 𝑇 is diagonalizable.
See Exercise 13 in Section 5C for an analogous statement about upper-
triangular matrices.
20 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is diagonaliz-

able if and only if the dual operator 𝑇 ′ is diagonalizable.
21 The Fibonacci sequence 𝐹0 , 𝐹1 , 𝐹2 , … is defined by
𝐹0 = 0, 𝐹1 = 1, and 𝐹𝑛 = 𝐹𝑛 − 2 + 𝐹𝑛 − 1 for 𝑛 ≥ 2.
Define 𝑇 ∈ ℒ(𝐑2 ) by 𝑇(𝑥, 𝑦) = (𝑦, 𝑥 + 𝑦).
(a) Show that 𝑇 𝑛 (0, 1) = (𝐹𝑛 , 𝐹𝑛 + 1 ) for each nonnegative integer 𝑛.
(b) Find the eigenvalues of 𝑇.
(c) Find a basis of 𝐑2 consisting of eigenvectors of 𝑇.
(d) Use the solution to (c) to compute 𝑇 𝑛 (0, 1). Conclude that
𝑛 𝑛
1 1 + √5 1 − √5
𝐹𝑛 = [( ) −( )]
√5 2 2
for each nonnegative integer 𝑛.
(e) Use (d) to conclude that if 𝑛 is a nonnegative integer, then the Fibonacci
number 𝐹𝑛 is the integer that is closest to
𝑛
1 + √5
1
( ).
√5 2
Each 𝐹𝑛 is a nonnegative integer, even though the right side of the formula
in (d) does not look like an integer. The number
1 + √5
2
is called the golden ratio.
22 Suppose 𝑇 ∈ ℒ(𝑉) and 𝐴 is an 𝑛-by-𝑛 matrix that is the matrix of 𝑇 with

respect to some basis of 𝑉. Prove that if
𝑛
|𝐴𝑗, 𝑗 | > ∑ |𝐴𝑗, 𝑘 |
𝑘=1
𝑘≠𝑗
for each 𝑗 ∈ {1, …, 𝑛}, then 𝑇 is invertible.

This exercise states that if the diagonal entries of the matrix of 𝑇 are large
compared to the nondiagonal entries, then 𝑇 is invertible.
23 Suppose the definition of the Gershgorin disks is changed so that the radius of
the 𝑘 th disk is the sum of the absolute values of the entries in column (instead
of row) 𝑘 of 𝐴, excluding the diagonal entry. Show that the Gershgorin disk
theorem (5.67) still holds with this changed definition.
Section 5E Commuting Operators 175
5E Commuting Operators
5.71 definition: commute
• Two operators 𝑆 and 𝑇 on the same vector space commute if 𝑆𝑇 = 𝑇𝑆.

• Two square matrices 𝐴 and 𝐵 of the same size commute if 𝐴𝐵 = 𝐵𝐴.
For example, if 𝑇 is an operator and 𝑝, 𝑞 ∈ 𝒫(𝐅), then 𝑝(𝑇) and 𝑞(𝑇) commute
[see 5.17(b)].
As another example, if 𝐼 is the identity operator on 𝑉, then 𝐼 commutes with
every operator on 𝑉.
5.72 example: partial differentiation operators commute

Suppose 𝑚 is a nonnegative integer. Let 𝒫𝑚(𝐑2 ) denote the real vector space of
polynomials (with real coefficients) in two real variables and of degree at most 𝑚,
with the usual operations of addition and scalar multiplication of real-valued
functions. Thus the elements of 𝒫𝑚(𝐑2 ) are functions 𝑝 on 𝐑2 of the form
5.73 𝑝= ∑ 𝑎𝑗, 𝑘 𝑥𝑗 𝑦𝑘,

𝑗+𝑘 ≤ 𝑚
where the indices 𝑗 and 𝑘 take on all nonnegative integer values such that 𝑗 + 𝑘 ≤ 𝑚,
each 𝑎𝑗, 𝑘 is in 𝐑 , and 𝑥𝑗 𝑦𝑘 denotes the function on 𝐑2 defined by (𝑥, 𝑦) ↦ 𝑥𝑗 𝑦𝑘.
Define operators 𝐷𝑥 , 𝐷𝑦 ∈ ℒ(𝒫𝑚(𝐑2 )) by
𝜕𝑝 𝜕𝑝
𝐷𝑥 𝑝 = = ∑ 𝑗𝑎 𝑥𝑗 − 1 𝑦𝑘 and 𝐷𝑦 𝑝 = = ∑ 𝑘𝑎 𝑥𝑗 𝑦𝑘 − 1,
𝜕𝑥 𝑗 + 𝑘 ≤ 𝑚 𝑗, 𝑘 𝜕𝑦 𝑗 + 𝑘 ≤ 𝑚 𝑗, 𝑘
where 𝑝 is as in 5.73. The operators 𝐷𝑥 and 𝐷𝑦 are called partial differentiation

operators because each of these operators differentiates with respect to one of the
variables while pretending that the other variable is a constant.
The operators 𝐷𝑥 and 𝐷𝑦 commute because if 𝑝 is as in 5.73, then
(𝐷𝑥 𝐷𝑦 )𝑝 = ∑ 𝑗𝑘𝑎𝑗, 𝑘 𝑥𝑗 − 1 𝑦𝑘 − 1 = (𝐷𝑦 𝐷𝑥 )𝑝.

𝑗+𝑘 ≤ 𝑚
The equation 𝐷𝑥 𝐷𝑦 = 𝐷𝑦 𝐷𝑥 on 𝒫𝑚(𝐑2 ) illustrates a more general result that

the order of partial differentiation does not matter for nice functions.
Commuting matrices are unusual. All 214,358,881 (which equals 118 )

For example, there are 214,358,881 pairs pairs of the 2-by-2 matrices under con-
of 2-by-2 matrices all of whose entries sideration were checked by a computer
are integers in the interval [−5, 5]. Only to discover that only 674,609 of these
about 0.3% of these pairs of matrices pairs of matrices commute.
commute.
The next result shows that two operators commute if and only if their matrices
(with respect to the same basis) commute.
5.74 commuting operators correspond to commuting matrices
Suppose 𝑆, 𝑇 ∈ ℒ(𝑉) and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Then 𝑆 and 𝑇 commute if

and only if ℳ(𝑆, (𝑣1 , …, 𝑣𝑛 )) and ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )) commute.
Proof We have
𝑆 and 𝑇 commute ⟺ 𝑆𝑇 = 𝑇𝑆
⟺ ℳ(𝑆𝑇) = ℳ(𝑇𝑆)
⟺ ℳ(𝑆)ℳ(𝑇) = ℳ(𝑇)ℳ(𝑆)
⟺ ℳ(𝑆) and ℳ(𝑇) commute,
as desired.
The next result shows that if two operators commute, then every eigenspace
for one operator is invariant under the other operator. This result, which we will
use several times, is one of the main reasons why a pair of commuting operators
behaves better than a pair of operators that does not commute.
5.75 eigenspace is invariant under commuting operator
Suppose 𝑆, 𝑇 ∈ ℒ(𝑉) commute and 𝜆 ∈ 𝐅. Then 𝐸(𝜆, 𝑆) is invariant under 𝑇.
Proof Suppose 𝑣 ∈ 𝐸(𝜆, 𝑆). Then

𝑆(𝑇𝑣) = (𝑆𝑇)𝑣 = (𝑇𝑆)𝑣 = 𝑇(𝑆𝑣) = 𝑇(𝜆𝑣) = 𝜆𝑇𝑣.
The equation above shows that 𝑇𝑣 ∈ 𝐸(𝜆, 𝑆). Thus 𝐸(𝜆, 𝑆) is invariant under 𝑇.
Suppose we have two operators, each of which is diagonalizable. If we want

to do computations involving both operators (for example, involving their sum),
then we want the two operators to be diagonalizable by the same basis, which
according to the next result is possible when the two operators commute.
5.76 simultaneous diagonalizablity ⟺ commutativity
Two diagonalizable operators on the same vector space have diagonal matrices
with respect to the same basis if and only if the two operators commute.
Proof First suppose 𝑆, 𝑇 ∈ ℒ(𝑉) have diagonal matrices with respect to the
same basis. The product of two diagonal matrices of the same size is the diagonal
matrix obtained by multiplying the corresponding elements of the two diagonals.
Thus any two diagonal matrices of the same size commute. Thus 𝑆 and 𝑇 commute,
by 5.74.
To prove the implication in the other direction, now suppose that 𝑆, 𝑇 ∈ ℒ(𝑉)
are diagonalizable operators that commute. Let 𝜆1 , …, 𝜆𝑚 denote the distinct
eigenvalues of 𝑆. Because 𝑆 is diagonalizable, 5.55(c) shows that
5.77 𝑉 = 𝐸(𝜆1 , 𝑆) ⊕ ⋯ ⊕ 𝐸(𝜆𝑚 , 𝑆).
For each 𝑘 = 1, …, 𝑚, the subspace 𝐸(𝜆𝑘 , 𝑆) is invariant under 𝑇 (by 5.75).
Because 𝑇 is diagonalizable, 5.65 implies that 𝑇|𝐸( 𝜆𝑘, 𝑆) is diagonalizable for
each 𝑘. Hence for each 𝑘 = 1, …, 𝑚, there is a basis of 𝐸(𝜆𝑘 , 𝑆) consisting of
eigenvectors of 𝑇. Putting these bases together gives a basis of 𝑉 (because of
5.77), with each vector in this basis being an eigenvector of both 𝑆 and 𝑇. Thus 𝑆
and 𝑇 both have diagonal matrices with respect to this basis, as desired.
See Exercise 2 for an extension of the result above to more than two operators.
Suppose 𝑉 is a finite-dimensional nonzero complex vector space. Then every
operator on 𝑉 has an eigenvector (see 5.19). The next result shows that if two
operators on 𝑉 commute, then there is a vector in 𝑉 that is an eigenvector for both
operators (but the two commuting operators might not have a common eigenvalue).
For an extension of the next result to more than two operators, see Exercise 9(a).
5.78 common eigenvector for commuting operators
Every pair of commuting operators on a finite-dimensional nonzero complex

vector space has a common eigenvector.
Proof Suppose 𝑉 is a finite-dimensional nonzero complex vector space and

𝑆, 𝑇 ∈ ℒ(𝑉) commute. Let 𝜆 be an eigenvalue of 𝑆 (5.19 tells us that 𝑆 does
indeed have an eigenvalue). Thus 𝐸(𝜆, 𝑆) ≠ {0}. Also, 𝐸(𝜆, 𝑆) is invariant
under 𝑇 (by 5.75).
Thus 𝑇|𝐸( 𝜆, 𝑆) has an eigenvector (again using 5.19), which is an eigenvector
for both 𝑆 and 𝑇, completing the proof.
5.79 example: common eigenvector for partial differentiation operators

Let 𝒫𝑚(𝐑2 ) be as in Example 5.72 and let 𝐷𝑥 , 𝐷𝑦 ∈ ℒ(𝒫𝑚(𝐑2 )) be the
commuting partial differentiation operators in that example. As you can verify, 0
is the only eigenvalue of each of these operators. Also
𝑚
𝐸(0, 𝐷𝑥 ) = { ∑ 𝑎𝑘 𝑦𝑘 ∶ 𝑎0 , …, 𝑎𝑚 ∈ 𝐑},
𝑘=0
𝑚
𝐸(0, 𝐷𝑦 ) = { ∑ 𝑐𝑗 𝑥𝑗 ∶ 𝑐0 , …, 𝑐𝑚 ∈ 𝐑}.
𝑗=0
The intersection of these two eigenspaces is the set of common eigenvectors of

the two operators. Because 𝐸(0, 𝐷𝑥 ) ∩ 𝐸(0, 𝐷𝑦 ) is the set of constant functions,
we see that 𝐷𝑥 and 𝐷𝑦 indeed have a common eigenvector, as promised by 5.78.
The next result extends 5.47 (the existence of a basis that gives an upper-
triangular matrix) to two commuting operators.
5.80 commuting operators are simultaneously upper triangularizable
Suppose 𝑉 is a finite-dimensional complex vector space and 𝑆, 𝑇 are

commuting operators on 𝑉. Then there is a basis of 𝑉 with respect to which
both 𝑆 and 𝑇 have upper-triangular matrices.
Proof Let 𝑛 = dim 𝑉. We will use induction on 𝑛. The desired result holds if
𝑛 = 1 because all 1-by-1 matrices are upper triangular. Now suppose 𝑛 > 1 and
the desired result holds for all complex vector spaces whose dimension is 𝑛 − 1.
Let 𝑣1 be any common eigenvector of 𝑆 and 𝑇 (using 5.78). Hence 𝑆𝑣1 ∈
span(𝑣1 ) and 𝑇𝑣1 ∈ span(𝑣1 ). Let 𝑊 be a subspace of 𝑉 such that
𝑉 = span(𝑣1 ) ⊕ 𝑊;
see 2.33 for the existence of 𝑊. Define a linear map 𝑃 ∶ 𝑉 → 𝑊 by
𝑃(𝑎𝑣1 + 𝑤) = 𝑤
for each 𝑎 ∈ 𝐂 and each 𝑤 ∈ 𝑊. Define 𝑆̂, 𝑇̂ ∈ ℒ(𝑊) by
̂ = 𝑃(𝑆𝑤) and 𝑇𝑤
𝑆𝑤 ̂ = 𝑃(𝑇𝑤)
for each 𝑤 ∈ 𝑊. To apply our induction hypothesis to 𝑆̂ and 𝑇̂ , we must first show
that these two operators on 𝑊 commute. To do this, suppose 𝑤 ∈ 𝑊. Then there
exists 𝑎 ∈ 𝐂 such that
(𝑆̂𝑇)𝑤
̂ ̂
= 𝑆(𝑃(𝑇𝑤)) ̂
= 𝑆(𝑇𝑤 − 𝑎𝑣 ) = 𝑃(𝑆(𝑇𝑤 − 𝑎𝑣 )) = 𝑃((𝑆𝑇)𝑤),
1 1
where the last equality holds because 𝑣1 is an eigenvector of 𝑆 and 𝑃𝑣1 = 0.

Similarly,
(𝑇̂ 𝑆)𝑤
̂ = 𝑃((𝑇𝑆)𝑤).
Because the operators 𝑆 and 𝑇 commute, the last two displayed equations show
that (𝑆̂𝑇)𝑤
̂ ̂ . Hence 𝑆̂ and 𝑇̂ commute.
= (𝑇̂ 𝑆)𝑤
Thus we can use our induction hypothesis to state that there exists a basis
𝑣2 , …, 𝑣𝑛 of 𝑊 such that 𝑆̂ and 𝑇̂ both have upper-triangular matrices with respect
to this basis. The list 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉.
If 𝑘 ∈ {2, …, 𝑛}, then there exist 𝑎𝑘 , 𝑏𝑘 ∈ 𝐂 such that
𝑆𝑣 = 𝑎 𝑣 + 𝑆𝑣
𝑘 𝑘 1
̂
𝑘 and 𝑇𝑣 = 𝑏 𝑣 + 𝑇𝑣
𝑘 𝑘 1
̂ .
𝑘
Because 𝑆̂ and 𝑇̂ have upper-triangular matrices with respect to 𝑣2 , …, 𝑣𝑛 , we

know that 𝑆𝑣
̂ ∈ span(𝑣 , …, 𝑣 ) and 𝑇𝑣
𝑘 2 𝑘
̂ ∈ span(𝑣 , …, 𝑣 ). Hence the equations
𝑘 2 𝑘
above imply that
𝑆𝑣𝑘 ∈ span(𝑣1 , …, 𝑣𝑘 ) and 𝑇𝑣𝑘 ∈ span(𝑣1 , …, 𝑣𝑘 ).
Thus 𝑆 and 𝑇 have upper-triangular matrices with respect to 𝑣1 , …, 𝑣𝑛 , as desired.
Exercise 9(b) extends the result above to more than two operators.
In general, it is not possible to determine the eigenvalues of the sum or product

of two operators from the eigenvalues of the two operators. However, the next
result shows that something nice happens when the two operators commute.
5.81 eigenvalues of sum and product of commuting operators
Suppose 𝑉 is a finite-dimensional complex vector space and 𝑆, 𝑇 are commut-

ing operators on 𝑉. Then
• every eigenvalue of 𝑆 + 𝑇 is an eigenvalue of 𝑆 plus an eigenvalue of 𝑇,
• every eigenvalue of 𝑆𝑇 is an eigenvalue of 𝑆 times an eigenvalue of 𝑇.
Proof There is a basis of 𝑉 with respect to which both 𝑆 and 𝑇 have upper-
triangular matrices (by 5.80). With respect to that basis,
ℳ(𝑆 + 𝑇) = ℳ(𝑆) + ℳ(𝑇) and ℳ(𝑆𝑇) = ℳ(𝑆)ℳ(𝑇),
as stated in 3.35 and 3.43.
The definition of matrix addition shows that each entry on the diagonal of
ℳ(𝑆 + 𝑇) equals the sum of the corresponding entries on the diagonals of ℳ(𝑆)
and ℳ(𝑇). Similarly, because ℳ(𝑆) and ℳ(𝑇) are upper-triangular matrices,
the definition of matrix multiplication shows that each entry on the diagonal of
ℳ(𝑆𝑇) equals the product of the corresponding entries on the diagonals of ℳ(𝑆)
and ℳ(𝑇). Furthermore, ℳ(𝑆 + 𝑇) and ℳ(𝑆𝑇) are upper-triangular matrices
(see Exercise 2 in Section 5B).
Every entry on the diagonal of ℳ(𝑆) is an eigenvalue of 𝑆, and every entry
on the diagonal of ℳ(𝑇) is an eigenvalue of 𝑇 (by 5.41). Every eigenvalue
of 𝑆 + 𝑇 is on the diagonal of ℳ(𝑆 + 𝑇), and every eigenvalue of 𝑆𝑇 is on
the diagonal of ℳ(𝑆𝑇) (these assertions follow from 5.41). Putting all this
together, we conclude that every eigenvalue of 𝑆 + 𝑇 is an eigenvalue of 𝑆 plus
an eigenvalue of 𝑇, and every eigenvalue of 𝑆𝑇 is an eigenvalue of 𝑆 times an
eigenvalue of 𝑇.
Exercises 5E
1 Give an example of two commuting operators 𝑆, 𝑇 on 𝐅4 such that there

is a subspace of 𝐅4 that is invariant under 𝑆 but not under 𝑇 and there is a
subspace of 𝐅4 that is invariant under 𝑇 but not under 𝑆.
2 Suppose ℰ is a subset of ℒ(𝑉) and every element of ℰ is diagonalizable.
Prove that there exists a basis of 𝑉 with respect to which every element of ℰ
has a diagonal matrix if and only if every pair of elements of ℰ commutes.
This exercise extends 5.76, which considers the case in which ℰ contains
only two elements. For this exercise, ℰ may contain any number of elements,
and ℰ may even be an infinite set.
3 Suppose 𝑆, 𝑇 ∈ ℒ(𝑉) are such that 𝑆𝑇 = 𝑇𝑆. Suppose 𝑝 ∈ 𝒫(𝐅).

(a) Prove that null 𝑝(𝑆) is invariant under 𝑇.
(b) Prove that range 𝑝(𝑆) is invariant under 𝑇.
See 5.18 for the special case 𝑆 = 𝑇.
4 Prove or give a counterexample: If 𝐴 is a diagonal matrix and 𝐵 is an

upper-triangular matrix of the same size as 𝐴, then 𝐴 and 𝐵 commute.
5 Prove that a pair of operators on a finite-dimensional vector space commute
if and only if their dual operators commute.
See 3.118 for the definition of the dual of an operator.
6 Suppose 𝑉 is a finite-dimensional complex vector space and 𝑆, 𝑇 ∈ ℒ(𝑉)

commute. Prove that there exist 𝛼, 𝜆 ∈ 𝐂 such that
range(𝑆 − 𝛼𝐼) + range(𝑇 − 𝜆𝐼) ≠ 𝑉.
7 Suppose 𝑉 is a complex vector space, 𝑆 ∈ ℒ(𝑉) is diagonalizable, and

𝑇 ∈ ℒ(𝑉) commutes with 𝑆. Prove that there is a basis of 𝑉 such that 𝑆 has
a diagonal matrix with respect to this basis and 𝑇 has an upper-triangular
matrix with respect to this basis.
8 Suppose 𝑚 = 3 in Example 5.72 and 𝐷𝑥 , 𝐷𝑦 are the commuting partial
differentiation operators on 𝒫3(𝐑2 ) from that example. Find a basis of
𝒫3(𝐑2 ) with respect to which 𝐷𝑥 and 𝐷𝑦 each have an upper-triangular
matrix.
9 Suppose 𝑉 is a finite-dimensional nonzero complex vector space. Suppose
that ℰ ⊆ ℒ(𝑉) is such that 𝑆 and 𝑇 commute for all 𝑆, 𝑇 ∈ ℰ.
(a) Prove that there is a vector in 𝑉 that is an eigenvector for every element
of ℰ.
(b) Prove that there is a basis of 𝑉 with respect to which every element of
ℰ has an upper-triangular matrix.
This exercise extends 5.78 and 5.80, which consider the case in which ℰ
contains only two elements. For this exercise, ℰ may contain any number of
elements, and ℰ may even be an infinite set.
10 Give an example of two commuting operators 𝑆, 𝑇 on a finite-dimensional

real vector space such that 𝑆 + 𝑇 has a eigenvalue that does not equal an
eigenvalue of 𝑆 plus an eigenvalue of 𝑇 and 𝑆𝑇 has a eigenvalue that does
not equal an eigenvalue of 𝑆 times an eigenvalue of 𝑇.
This exercise shows that 5.81 does not hold on real vector spaces.
Chapter 6
Inner Product Spaces
In making the definition of a vector space, we generalized the linear structure

(addition and scalar multiplication) of 𝐑2 and 𝐑3. We ignored geometric features
such as the notions of length and angle. These ideas are embedded in the concept
of inner products, which we will investigate in this chapter.
Every inner product induces a norm, which you can think of as a length.
This norm satisfies key properties such as the Pythagorean theorem, the triangle
inequality, the parallelogram equality, and the Cauchy–Schwarz inequality.
The notion of perpendicular vectors in Euclidean geometry gets renamed to
orthogonal vectors in the context of an inner product space. We will see that
orthonormal bases are tremendously useful in inner product spaces. The Gram–
Schmidt procedure constructs such bases. This chapter will conclude by putting
together these tools to solve minimization problems.
• 𝑉 and 𝑊 denote vector spaces over 𝐅.
Matthew Petroff CC BY-SA
The George Peabody Library, now part of Johns Hopkins University, opened while
James Sylvester (1814–1897) was the university’s first mathematics professor. Sylvester’s
publications include the first use of the word matrix in mathematics.
181
https://doi.org/10.1007/978-3-031-41026-0_6
182 Chapter 6 Inner Product Spaces
6A Inner Products and Norms

Inner Products
To motivate the concept of inner product,
think of vectors in 𝐑2 and 𝐑3 as arrows
with initial point at the origin. The length
of a vector 𝑣 in 𝐑2 or 𝐑3 is called the
norm of 𝑣 and is denoted by ‖𝑣‖. Thus
for 𝑣 = (𝑎, 𝑏) ∈ 𝐑2, we have
This vector 𝑣 has norm √𝑎2 + 𝑏2 .
‖𝑣‖ = √𝑎2 + 𝑏2 .
Similarly, if 𝑣 = (𝑎, 𝑏, 𝑐) ∈ 𝐑3, then ‖𝑣‖ = √𝑎2 + 𝑏2 + 𝑐2 .

Even though we cannot draw pictures in higher dimensions, the generalization
to 𝐑𝑛 is easy: we define the norm of 𝑥 = (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐑𝑛 by
‖𝑥‖ = √𝑥12 + ⋯ + 𝑥𝑛2 .
The norm is not linear on 𝐑𝑛. To inject linearity into the discussion, we
introduce the dot product.
6.1 definition: dot product
For 𝑥, 𝑦 ∈ 𝐑𝑛, the dot product of 𝑥 and 𝑦, denoted by 𝑥 ⋅ 𝑦, is defined by
𝑥 ⋅ 𝑦 = 𝑥1 𝑦1 + ⋯ + 𝑥𝑛 𝑦𝑛 ,
where 𝑥 = (𝑥1 , …, 𝑥𝑛 ) and 𝑦 = (𝑦1 , …, 𝑦𝑛 ).
The dot product of two vectors in 𝐑𝑛 If we think of a vector as a point instead

is a number, not a vector. Notice that of as an arrow, then ‖𝑥‖ should be
𝑥 ⋅ 𝑥 = ‖𝑥‖2 for all 𝑥 ∈ 𝐑𝑛. Furthermore, interpreted to mean the distance from
the dot product on 𝐑𝑛 has the following the origin to the point 𝑥.
properties.
• 𝑥 ⋅ 𝑥 ≥ 0 for all 𝑥 ∈ 𝐑𝑛.
• 𝑥 ⋅ 𝑥 = 0 if and only if 𝑥 = 0.
• For 𝑦 ∈ 𝐑𝑛 fixed, the map from 𝐑𝑛 to 𝐑 that sends 𝑥 ∈ 𝐑𝑛 to 𝑥 ⋅ 𝑦 is linear.
• 𝑥 ⋅ 𝑦 = 𝑦 ⋅ 𝑥 for all 𝑥, 𝑦 ∈ 𝐑𝑛.
An inner product is a generalization of the dot product. At this point you may
be tempted to guess that an inner product is defined by abstracting the properties
of the dot product discussed in the last paragraph. For real vector spaces, that
guess is correct. However, so that we can make a definition that will be useful
for both real and complex vector spaces, we need to examine the complex case
before making the definition.
Section 6A Inner Products and Norms 183
Recall that if 𝜆 = 𝑎 + 𝑏𝑖, where 𝑎, 𝑏 ∈ 𝐑 , then

• the absolute value of 𝜆, denoted by |𝜆|, is defined by |𝜆| = √𝑎2 + 𝑏2 ;
• the complex conjugate of 𝜆, denoted by 𝜆, is defined by 𝜆 = 𝑎 − 𝑏𝑖;
• |𝜆|2 = 𝜆𝜆.
See Chapter 4 for the definitions and the basic properties of the absolute value
and complex conjugate.
For 𝑧 = (𝑧1 , …, 𝑧𝑛 ) ∈ 𝐂𝑛, we define the norm of 𝑧 by
‖𝑧‖ = √|𝑧1 |2 + ⋯ + |𝑧𝑛 |2 .

The absolute values are needed because we want ‖𝑧‖ to be a nonnegative number.
Note that
‖𝑧‖2 = 𝑧1 𝑧1 + ⋯ + 𝑧𝑛 𝑧𝑛 .
We want to think of ‖𝑧‖2 as the inner product of 𝑧 with itself, as we did
in 𝐑𝑛. The equation above thus suggests that the inner product of the vector
𝑤 = (𝑤1 , …, 𝑤𝑛 ) ∈ 𝐂𝑛 with 𝑧 should equal
𝑤1 𝑧1 + ⋯ + 𝑤𝑛 𝑧𝑛 .
If the roles of the 𝑤 and 𝑧 were interchanged, the expression above would be
replaced with its complex conjugate. Thus we should expect that the inner product
of 𝑤 with 𝑧 equals the complex conjugate of the inner product of 𝑧 with 𝑤. With
that motivation, we are now ready to define an inner product on 𝑉, which may be
a real or a complex vector space.
One comment about the notation used in the next definition:
• For 𝜆 ∈ 𝐂 , the notation 𝜆 ≥ 0 means 𝜆 is real and nonnegative.
6.2 definition: inner product
An inner product on 𝑉 is a function that takes each ordered pair (𝑢, 𝑣) of

elements of 𝑉 to a number ⟨𝑢, 𝑣⟩ ∈ 𝐅 and has the following properties.
positivity
⟨𝑣, 𝑣⟩ ≥ 0 for all 𝑣 ∈ 𝑉.
definiteness
⟨𝑣, 𝑣⟩ = 0 if and only if 𝑣 = 0.
additivity in first slot

⟨𝑢 + 𝑣, 𝑤⟩ = ⟨𝑢, 𝑤⟩ + ⟨𝑣, 𝑤⟩ for all 𝑢, 𝑣, 𝑤 ∈ 𝑉.
homogeneity in first slot

⟨𝜆𝑢, 𝑣⟩ = 𝜆⟨𝑢, 𝑣⟩ for all 𝜆 ∈ 𝐅 and all 𝑢, 𝑣 ∈ 𝑉.
conjugate symmetry
⟨𝑢, 𝑣⟩ = ⟨𝑣, 𝑢⟩ for all 𝑢, 𝑣 ∈ 𝑉.
Every real number equals its complex Most mathematicians define inner
conjugate. Thus if we are dealing with products as above, but many physicists
a real vector space, then in the last con- use a definition that requires homo-
dition above we can dispense with the geneity in the second slot instead of
complex conjugate and simply state that the first slot.
⟨𝑢, 𝑣⟩ = ⟨𝑣, 𝑢⟩ for all 𝑢, 𝑣 ∈ 𝑉.
6.3 example: inner products
(a) The Euclidean inner product on 𝐅𝑛 is defined by

⟨(𝑤1 , …, 𝑤𝑛 ), (𝑧1 , …, 𝑧𝑛 )⟩ = 𝑤1 𝑧1 + ⋯ + 𝑤𝑛 𝑧𝑛
for all (𝑤1 , …, 𝑤𝑛 ), (𝑧1 , …, 𝑧𝑛 ) ∈ 𝐅𝑛.

(b) If 𝑐1 , …, 𝑐𝑛 are positive numbers, then an inner product can be defined on 𝐅𝑛
by
⟨(𝑤1 , …, 𝑤𝑛 ), (𝑧1 , …, 𝑧𝑛 )⟩ = 𝑐1 𝑤1 𝑧1 + ⋯ + 𝑐𝑛 𝑤𝑛 𝑧𝑛
for all (𝑤1 , …, 𝑤𝑛 ), (𝑧1 , …, 𝑧𝑛 ) ∈ 𝐅𝑛.
(c) An inner product can be defined on the vector space of continuous real-valued
functions on the interval [−1, 1] by
1
⟨ 𝑓, 𝑔⟩ = ∫ 𝑓𝑔
−1
for all 𝑓, 𝑔 continuous real-valued functions on [−1, 1].

(d) An inner product can be defined on 𝒫(𝐑) by
1
⟨𝑝, 𝑞⟩ = 𝑝(0)𝑞(0) + ∫ 𝑝′ 𝑞′
−1
for all 𝑝, 𝑞 ∈ 𝒫(𝐑).

(e) An inner product can be defined on 𝒫(𝐑) by
∞
⟨𝑝, 𝑞⟩ = ∫ 𝑝(𝑥)𝑞(𝑥)𝑒−𝑥 𝑑𝑥
0
for all 𝑝, 𝑞 ∈ 𝒫(𝐑).
6.4 definition: inner product space
An inner product space is a vector space 𝑉 along with an inner product on 𝑉.
The most important example of an inner product space is 𝐅𝑛 with the Euclidean
inner product given by (a) in the example above. When 𝐅𝑛 is referred to as an
inner product space, you should assume that the inner product is the Euclidean
inner product unless explicitly told otherwise.
So that we do not have to keep repeating the hypothesis that 𝑉 and 𝑊 are inner
product spaces, we make the following assumption.
6.5 notation: 𝑉, 𝑊
For the rest of this chapter and the next chapter, 𝑉 and 𝑊 denote inner product
spaces over 𝐅.
Note the slight abuse of language here. An inner product space is a vector
space along with an inner product on that vector space. When we say that a vector
space 𝑉 is an inner product space, we are also thinking that an inner product on
𝑉 is lurking nearby or is clear from the context (or is the Euclidean inner product
if the vector space is 𝐅𝑛 ).
6.6 basic properties of an inner product
(a) For each fixed 𝑣 ∈ 𝑉, the function that takes 𝑢 ∈ 𝑉 to ⟨𝑢, 𝑣⟩ is a linear
map from 𝑉 to 𝐅.
(b) ⟨0, 𝑣⟩ = 0 for every 𝑣 ∈ 𝑉.
(c) ⟨𝑣, 0⟩ = 0 for every 𝑣 ∈ 𝑉.
(d) ⟨𝑢, 𝑣 + 𝑤⟩ = ⟨𝑢, 𝑣⟩ + ⟨𝑢, 𝑤⟩ for all 𝑢, 𝑣, 𝑤 ∈ 𝑉.
(e) ⟨𝑢, 𝜆𝑣⟩ = 𝜆⟨𝑢, 𝑣⟩ for all 𝜆 ∈ 𝐅 and all 𝑢, 𝑣 ∈ 𝑉.
Proof
(a) For 𝑣 ∈ 𝑉, the linearity of 𝑢 ↦ ⟨𝑢, 𝑣⟩ follows from the conditions of additivity
and homogeneity in the first slot in the definition of an inner product.
(b) Every linear map takes 0 to 0. Thus (b) follows from (a).
(c) If 𝑣 ∈ 𝑉, then the conjugate symmetry property in the definition of an inner
product and (b) show that ⟨𝑣, 0⟩ = ⟨0, 𝑣⟩ = 0 = 0.
(d) Suppose 𝑢, 𝑣, 𝑤 ∈ 𝑉. Then
⟨𝑢, 𝑣 + 𝑤⟩ = ⟨𝑣 + 𝑤, 𝑢⟩
= ⟨𝑣, 𝑢⟩ + ⟨𝑤, 𝑢⟩
= ⟨𝑣, 𝑢⟩ + ⟨𝑤, 𝑢⟩
= ⟨𝑢, 𝑣⟩ + ⟨𝑢, 𝑤⟩.
(e) Suppose 𝜆 ∈ 𝐅 and 𝑢, 𝑣 ∈ 𝑉. Then

⟨𝑢, 𝜆𝑣⟩ = ⟨𝜆𝑣, 𝑢⟩
= 𝜆⟨𝑣, 𝑢⟩
= 𝜆 ⟨𝑣, 𝑢⟩
= 𝜆⟨𝑢, 𝑣⟩.
Norms
Our motivation for defining inner products came initially from the norms of
vectors on 𝐑2 and 𝐑3. Now we see that each inner product determines a norm.
6.7 definition: norm, ‖𝑣‖
For 𝑣 ∈ 𝑉, the norm of 𝑣, denoted by ‖𝑣‖, is defined by
‖𝑣‖ = √⟨𝑣, 𝑣⟩.
6.8 example: norms
(a) If (𝑧1 , …, 𝑧𝑛 ) ∈ 𝐅𝑛 (with the Euclidean inner product), then
‖(𝑧1 , …, 𝑧𝑛 )‖ = √|𝑧1 |2 + ⋯ + |𝑧𝑛 |2 .
(b) For 𝑓 in the vector space of continuous real-valued functions on [−1, 1] and
with inner product given as in 6.3(c), we have
1
‖ 𝑓 ‖ = √∫ 𝑓 2.
−1
6.9 basic properties of the norm
Suppose 𝑣 ∈ 𝑉.
(a) ‖𝑣‖ = 0 if and only if 𝑣 = 0.
(b) ‖𝜆𝑣‖ = |𝜆| ‖𝑣‖ for all 𝜆 ∈ 𝐅.
Proof
(a) The desired result holds because ⟨𝑣, 𝑣⟩ = 0 if and only if 𝑣 = 0.
(b) Suppose 𝜆 ∈ 𝐅. Then
‖𝜆𝑣‖2 = ⟨𝜆𝑣, 𝜆𝑣⟩

= 𝜆⟨𝑣, 𝜆𝑣⟩
= 𝜆𝜆⟨𝑣, 𝑣⟩
= |𝜆|2 ‖𝑣‖2.
Taking square roots now gives the desired equality.
The proof of (b) in the result above illustrates a general principle: working
with norms squared is usually easier than working directly with norms.
Now we come to a crucial definition.
6.10 definition: orthogonal
Two vectors 𝑢, 𝑣 ∈ 𝑉 are called orthogonal if ⟨𝑢, 𝑣⟩ = 0.
In the definition above, the order of The word orthogonal comes from the
the two vectors does not matter, because Greek word orthogonios, which means
⟨𝑢, 𝑣⟩ = 0 if and only if ⟨𝑣, 𝑢⟩ = 0. In- right-angled.
stead of saying 𝑢 and 𝑣 are orthogonal,
sometimes we say 𝑢 is orthogonal to 𝑣.
Exercise 15 asks you to prove that if 𝑢, 𝑣 are nonzero vectors in 𝐑2, then
⟨𝑢, 𝑣⟩ = ‖𝑢‖ ‖𝑣‖ cos 𝜃,
where 𝜃 is the angle between 𝑢 and 𝑣 (thinking of 𝑢 and 𝑣 as arrows with initial
point at the origin). Thus two nonzero vectors in 𝐑2 are orthogonal (with respect
to the Euclidean inner product) if and only if the cosine of the angle between
them is 0, which happens if and only if the vectors are perpendicular in the usual
sense of plane geometry. Thus you can think of the word orthogonal as a fancy
word meaning perpendicular.
We begin our study of orthogonality with an easy result.
6.11 orthogonality and 0
(a) 0 is orthogonal to every vector in 𝑉.

(b) 0 is the only vector in 𝑉 that is orthogonal to itself.
Proof
(a) Recall that 6.6(b) states that ⟨0, 𝑣⟩ = 0 for every 𝑣 ∈ 𝑉.
(b) If 𝑣 ∈ 𝑉 and ⟨𝑣, 𝑣⟩ = 0, then 𝑣 = 0 (by definition of inner product).
For the special case 𝑉 = 𝐑2, the next theorem was known over 3,500 years ago
in Babylonia and then rediscovered and proved over 2,500 years ago in Greece.
Of course, the proof below is not the original proof.
6.12 Pythagorean theorem
Suppose 𝑢, 𝑣 ∈ 𝑉. If 𝑢 and 𝑣 are orthogonal, then
‖𝑢 + 𝑣‖2 = ‖𝑢‖2 + ‖𝑣‖2.
Proof Suppose ⟨𝑢, 𝑣⟩ = 0. Then

‖𝑢 + 𝑣‖2 = ⟨𝑢 + 𝑣, 𝑢 + 𝑣⟩
= ⟨𝑢, 𝑢⟩ + ⟨𝑢, 𝑣⟩ + ⟨𝑣, 𝑢⟩ + ⟨𝑣, 𝑣⟩
= ‖𝑢‖2 + ‖𝑣‖2.
Suppose 𝑢, 𝑣 ∈ 𝑉, with 𝑣 ≠ 0. We would like to write 𝑢 as a scalar multiple

of 𝑣 plus a vector 𝑤 orthogonal to 𝑣, as suggested in the picture here.
An orthogonal decomposition:
𝑢 expressed as a scalar multiple of 𝑣 plus a vector orthogonal to 𝑣.
To discover how to write 𝑢 as a scalar multiple of 𝑣 plus a vector orthogonal

to 𝑣, let 𝑐 ∈ 𝐅 denote a scalar. Then
𝑢 = 𝑐𝑣 + (𝑢 − 𝑐𝑣).
Thus we need to choose 𝑐 so that 𝑣 is orthogonal to (𝑢 − 𝑐𝑣). Hence we want
0 = ⟨𝑢 − 𝑐𝑣, 𝑣⟩ = ⟨𝑢, 𝑣⟩ − 𝑐‖𝑣‖2.
The equation above shows that we should choose 𝑐 to be ⟨𝑢, 𝑣⟩/‖𝑣‖2. Making this
choice of 𝑐, we can write
⟨𝑢, 𝑣⟩ ⟨𝑢, 𝑣⟩
𝑢= 2
𝑣 + (𝑢 − 𝑣).
‖𝑣‖ ‖𝑣‖2
As you should verify, the equation displayed above explicitly writes 𝑢 as a scalar
multiple of 𝑣 plus a vector orthogonal to 𝑣. Thus we have proved the following
key result.
6.13 an orthogonal decomposition
⟨𝑢, 𝑣⟩ ⟨𝑢, 𝑣⟩
Suppose 𝑢, 𝑣 ∈ 𝑉, with 𝑣 ≠ 0. Set 𝑐 = and 𝑤 = 𝑢 − 𝑣. Then
‖𝑣‖2 ‖𝑣‖2
𝑢 = 𝑐𝑣 + 𝑤 and ⟨𝑤, 𝑣⟩ = 0.
The orthogonal decomposition 6.13 will be used in the proof of the Cauchy–
Schwarz inequality, which is our next result and is one of the most important
inequalities in mathematics.
6.14 Cauchy–Schwarz inequality
Suppose 𝑢, 𝑣 ∈ 𝑉. Then
|⟨𝑢, 𝑣⟩| ≤ ‖𝑢‖ ‖𝑣‖.
This inequality is an equality if and only if one of 𝑢, 𝑣 is a scalar multiple of
the other.
Proof If 𝑣 = 0, then both sides of the desired inequality equal 0. Thus we can
assume that 𝑣 ≠ 0. Consider the orthogonal decomposition
⟨𝑢, 𝑣⟩
𝑢= 𝑣+𝑤
‖𝑣‖2
given by 6.13, where 𝑤 is orthogonal to 𝑣. By the Pythagorean theorem,
⟨𝑢, 𝑣⟩ 2
‖𝑢‖2 = ∥ 𝑣∥ + ‖𝑤‖2
‖𝑣‖2
2
∣⟨𝑢, 𝑣⟩∣
= + ‖𝑤‖2
‖𝑣‖2
2
∣⟨𝑢, 𝑣⟩∣
6.15 ≥ .
‖𝑣‖2
Multiplying both sides of this inequality by ‖𝑣‖2 and then taking square roots
gives the desired inequality.
The proof in the paragraph above Augustin-Louis Cauchy (1789–1857)
shows that the Cauchy–Schwarz inequal- proved 6.16(a) in 1821. In 1859,
ity is an equality if and only if 6.15 is Cauchy’s student Viktor Bunyakovsky
an equality. This happens if and only (1804–1889) proved integral inequal-
if 𝑤 = 0. But 𝑤 = 0 if and only if 𝑢 ities like the one in 6.16(b). A few
is a multiple of 𝑣 (see 6.13). Thus the decades later, similar discoveries by
Cauchy–Schwarz inequality is an equal- Hermann Schwarz (1843–1921) at-
ity if and only if 𝑢 is a scalar multiple of 𝑣 tracted more attention and led to the
or 𝑣 is a scalar multiple of 𝑢 (or both; the name of this inequality.
phrasing has been chosen to cover cases
in which either 𝑢 or 𝑣 equals 0).
6.16 example: Cauchy–Schwarz inequality
(a) If 𝑥1 , …, 𝑥𝑛 , 𝑦1 , …, 𝑦𝑛 ∈ 𝐑 , then
(𝑥1 𝑦1 + ⋯ + 𝑥𝑛 𝑦𝑛 )2 ≤ (𝑥12 + ⋯ + 𝑥𝑛2 )(𝑦12 + ⋯ + 𝑦𝑛2 ),
as follows from applying the Cauchy–Schwarz inequality to the vectors

(𝑥1 , …, 𝑥𝑛 ), (𝑦1 , …, 𝑦𝑛 ) ∈ 𝐑𝑛, using the usual Euclidean inner product.
(b) If 𝑓, 𝑔 are continuous real-valued functions on [−1, 1], then

1 2 1 1
∣∫ 𝑓 𝑔 ∣ ≤ (∫ 𝑓 2 )(∫ 𝑔2 ),
−1 −1 −1
as follows from applying the Cauchy–Schwarz inequality to Example 6.3(c).
The next result, called the triangle inequality,
has the geometric interpretation that the length
of each side of a triangle is less than the sum of
the lengths of the other two sides.
Note that the triangle inequality implies that
the shortest polygonal path between two points is In this triangle, the length of
a single line segment (a polygonal path consists 𝑢 + 𝑣 is less than the length
of line segments). of 𝑢 plus the length of 𝑣.
6.17 triangle inequality
‖𝑢 + 𝑣‖ ≤ ‖𝑢‖ + ‖𝑣‖.
This inequality is an equality if and only if one of 𝑢, 𝑣 is a nonnegative real

multiple of the other.
Proof We have
‖𝑢 + 𝑣‖2 = ⟨𝑢 + 𝑣, 𝑢 + 𝑣⟩
= ⟨𝑢, 𝑢⟩ + ⟨𝑣, 𝑣⟩ + ⟨𝑢, 𝑣⟩ + ⟨𝑣, 𝑢⟩
= ⟨𝑢, 𝑢⟩ + ⟨𝑣, 𝑣⟩ + ⟨𝑢, 𝑣⟩ + ⟨𝑢, 𝑣⟩
= ‖𝑢‖2 + ‖𝑣‖2 + 2 Re⟨𝑢, 𝑣⟩
6.18 ≤ ‖𝑢‖2 + ‖𝑣‖2 + 2∣⟨𝑢, 𝑣⟩∣
6.19 ≤ ‖𝑢‖2 + ‖𝑣‖2 + 2‖𝑢‖ ‖𝑣‖
2
= (‖𝑢‖ + ‖𝑣‖) ,
where 6.19 follows from the Cauchy–Schwarz inequality (6.14). Taking square
roots of both sides of the inequality above gives the desired inequality.
The proof above shows that the triangle inequality is an equality if and only if
we have equality in 6.18 and 6.19. Thus we have equality in the triangle inequality
if and only if
6.20 ⟨𝑢, 𝑣⟩ = ‖𝑢‖ ‖𝑣‖.
If one of 𝑢, 𝑣 is a nonnegative real multiple of the other, then 6.20 holds. Con-
versely, suppose 6.20 holds. Then the condition for equality in the Cauchy–
Schwarz inequality (6.14) implies that one of 𝑢, 𝑣 is a scalar multiple of the other.
This scalar must be a nonnegative real number, by 6.20, completing the proof.
For the reverse triangle inequality, see Exercise 20.

The next result is called the parallel-

ogram equality because of its geometric
interpretation: in every parallelogram, the
sum of the squares of the lengths of the
diagonals equals the sum of the squares of
the lengths of the four sides. Note that the
proof here is more straightforward than The diagonals of this parallelogram
the usual proof in Euclidean geometry. are 𝑢 + 𝑣 and 𝑢 − 𝑣.
6.21 parallelogram equality
‖𝑢 + 𝑣‖2 + ‖𝑢 − 𝑣‖2 = 2(‖𝑢‖2 + ‖𝑣‖2 ).
Proof We have
‖𝑢 + 𝑣‖2 + ‖𝑢 − 𝑣‖2 = ⟨𝑢 + 𝑣, 𝑢 + 𝑣⟩ + ⟨𝑢 − 𝑣, 𝑢 − 𝑣⟩
= ‖𝑢‖2 + ‖𝑣‖2 + ⟨𝑢, 𝑣⟩ + ⟨𝑣, 𝑢⟩

+ ‖𝑢‖2 + ‖𝑣‖2 − ⟨𝑢, 𝑣⟩ − ⟨𝑣, 𝑢⟩
= 2(‖𝑢‖2 + ‖𝑣‖2 ),
as desired.
Exercises 6A
1 Prove or give a counterexample: If 𝑣1 , …, 𝑣𝑚 ∈ 𝑉, then

𝑚 𝑚
∑ ∑ ⟨𝑣𝑗 , 𝑣𝑘 ⟩ ≥ 0.
𝑗 = 1 𝑘=1
2 Suppose 𝑆 ∈ ℒ(𝑉). Define ⟨⋅, ⋅⟩1 by

⟨𝑢, 𝑣⟩1 = ⟨𝑆𝑢, 𝑆𝑣⟩
for all 𝑢, 𝑣 ∈ 𝑉. Show that ⟨⋅, ⋅⟩1 is an inner product on 𝑉 if and only if 𝑆 is
injective.
3 (a) Show that the function taking an ordered pair ((𝑥1 , 𝑥2 ), (𝑦1 , 𝑦2 )) of
elements of 𝐑2 to |𝑥1 𝑦1 | + |𝑥2 𝑦2 | is not an inner product on 𝐑2.
(b) Show that the function taking an ordered pair ((𝑥1 , 𝑥2 , 𝑥3 ), (𝑦1 , 𝑦2 , 𝑦3 ))
of elements of 𝐑3 to 𝑥1 𝑦1 + 𝑥3 𝑦3 is not an inner product on 𝐑3.
4 Suppose 𝑇 ∈ ℒ(𝑉) is such that ‖𝑇𝑣‖ ≤ ‖𝑣‖ for every 𝑣 ∈ 𝑉. Prove that
𝑇 − √2 𝐼 is injective.
5 Suppose 𝑉 is a real inner product space.

(a) Show that ⟨𝑢 + 𝑣, 𝑢 − 𝑣⟩ = ‖𝑢‖2 − ‖𝑣‖2 for every 𝑢, 𝑣 ∈ 𝑉.
(b) Show that if 𝑢, 𝑣 ∈ 𝑉 have the same norm, then 𝑢 + 𝑣 is orthogonal to
𝑢 − 𝑣.
(c) Use (b) to show that the diagonals of a rhombus are perpendicular to
each other.
6 Suppose 𝑢, 𝑣 ∈ 𝑉. Prove that ⟨𝑢, 𝑣⟩ = 0 ⟺ ‖𝑢‖ ≤ ‖𝑢 + 𝑎𝑣‖ for all 𝑎 ∈ 𝐅.

7 Suppose 𝑢, 𝑣 ∈ 𝑉. Prove that ‖𝑎𝑢 + 𝑏𝑣‖ = ‖𝑏𝑢 + 𝑎𝑣‖ for all 𝑎, 𝑏 ∈ 𝐑 if and
only if ‖𝑢‖ = ‖𝑣‖.
8 Suppose 𝑎, 𝑏, 𝑐, 𝑥, 𝑦 ∈ 𝐑 and 𝑎2 + 𝑏2 + 𝑐2 + 𝑥2 + 𝑦2 ≤ 1. Prove that
𝑎 + 𝑏 + 𝑐 + 4𝑥 + 9𝑦 ≤ 10.
9 Suppose 𝑢, 𝑣 ∈ 𝑉 and ‖𝑢‖ = ‖𝑣‖ = 1 and ⟨𝑢, 𝑣⟩ = 1. Prove that 𝑢 = 𝑣.

10 Suppose 𝑢, 𝑣 ∈ 𝑉 and ‖𝑢‖ ≤ 1 and ‖𝑣‖ ≤ 1. Prove that
√1 − ‖𝑢‖2 √1 − ‖𝑣‖2 ≤ 1 − ∣⟨𝑢, 𝑣⟩∣.
11 Find vectors 𝑢, 𝑣 ∈ 𝐑2 such that 𝑢 is a scalar multiple of (1, 3), 𝑣 is orthog-

onal to (1, 3), and (1, 2) = 𝑢 + 𝑣.
12 Suppose 𝑎, 𝑏, 𝑐, 𝑑 are positive numbers.
1 1 1 1
(a) Prove that (𝑎 + 𝑏 + 𝑐 + 𝑑)( + + + ) ≥ 16.
𝑎 𝑏 𝑐 𝑑
(b) For which positive numbers 𝑎, 𝑏, 𝑐, 𝑑 is the inequality above an equality?
13 Show that the square of an average is less than or equal to the average of the
squares. More precisely, show that if 𝑎1 , …, 𝑎𝑛 ∈ 𝐑 , then the square of the
average of 𝑎1 , …, 𝑎𝑛 is less than or equal to the average of 𝑎12, …, 𝑎𝑛2.
14 Suppose 𝑣 ∈ 𝑉 and 𝑣 ≠ 0. Prove that 𝑣/‖𝑣‖ is the unique closest element on
the unit sphere of 𝑉 to 𝑣. More precisely, prove that if 𝑢 ∈ 𝑉 and ‖𝑢‖ = 1,
then
𝑣
∥𝑣 − ∥ ≤ ‖𝑣 − 𝑢‖,
‖𝑣‖
with equality only if 𝑢 = 𝑣/‖𝑣‖.
15 Suppose 𝑢, 𝑣 are nonzero vectors in 𝐑2. Prove that
⟨𝑢, 𝑣⟩ = ‖𝑢‖ ‖𝑣‖ cos 𝜃,
where 𝜃 is the angle between 𝑢 and 𝑣 (thinking of 𝑢 and 𝑣 as arrows with

initial point at the origin).
Hint: Use the law of cosines on the triangle formed by 𝑢, 𝑣, and 𝑢 − 𝑣.
16 The angle between two vectors (thought of as arrows with initial point at
the origin) in 𝐑2 or 𝐑3 can be defined geometrically. However, geometry is
not as clear in 𝐑𝑛 for 𝑛 > 3. Thus the angle between two nonzero vectors
𝑥, 𝑦 ∈ 𝐑𝑛 is defined to be
⟨𝑥, 𝑦⟩
arccos ,
‖𝑥‖ ‖𝑦‖
where the motivation for this definition comes from Exercise 15. Explain
why the Cauchy–Schwarz inequality is needed to show that this definition
makes sense.
17 Prove that
𝑛 2 𝑛 𝑛 𝑏𝑘2
( ∑ 𝑎𝑘 𝑏𝑘 ) ≤ ( ∑ 𝑘𝑎𝑘2 )( ∑ )
𝑘=1 𝑘=1 𝑘=1 𝑘
for all real numbers 𝑎1 , …, 𝑎𝑛 and 𝑏1 , …, 𝑏𝑛 .
18 (a) Suppose 𝑓 ∶ [1, ∞) → [0, ∞) is continuous. Show that
∞ 2 ∞
2
(∫ 𝑓 ) ≤ ∫ 𝑥2 ( 𝑓 (𝑥)) 𝑑𝑥.
1 1
(b) For which continuous functions 𝑓 ∶ [1, ∞) → [0, ∞) is the inequality in

(a) an equality with both sides finite?
19 Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 and 𝑇 ∈ ℒ(𝑉). Prove that if 𝜆 is an
eigenvalue of 𝑇, then
𝑛 𝑛
|𝜆|2 ≤ ∑ ∑ |ℳ(𝑇)𝑗, 𝑘 |2,
𝑗 = 1 𝑘=1
where ℳ(𝑇)𝑗, 𝑘 denotes the entry in row 𝑗, column 𝑘 of the matrix of 𝑇 with
respect to the basis 𝑣1 , …, 𝑣𝑛 .
20 Prove that if 𝑢, 𝑣 ∈ 𝑉, then ∣ ‖𝑢‖ − ‖𝑣‖ ∣ ≤ ‖𝑢 − 𝑣‖.
The inequality above is called the reverse triangle inequality. For the
reverse triangle inequality when 𝑉 = 𝐂 , see Exercise 2 in Chapter 4.
21 Suppose 𝑢, 𝑣 ∈ 𝑉 are such that

‖𝑢‖ = 3, ‖𝑢 + 𝑣‖ = 4, ‖𝑢 − 𝑣‖ = 6.
What number does ‖𝑣‖ equal?
22 Show that if 𝑢, 𝑣 ∈ 𝑉, then
‖𝑢 + 𝑣‖ ‖𝑢 − 𝑣‖ ≤ ‖𝑢‖2 + ‖𝑣‖2.
23 Suppose 𝑣1 , …, 𝑣𝑚 ∈ 𝑉 are such that ‖𝑣𝑘 ‖ ≤ 1 for each 𝑘 = 1, …, 𝑚. Show

that there exist 𝑎1 , …, 𝑎𝑚 ∈ {1, −1} such that
√
‖𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 ‖ ≤ 𝑚.
24 Prove or give a counterexample: If ‖ ⋅ ‖ is the norm associated with an inner

product on 𝐑2, then there exists (𝑥, 𝑦) ∈ 𝐑2 such that ‖(𝑥, 𝑦)‖ ≠ max{|𝑥|, |𝑦|}.
25 Suppose 𝑝 > 0. Prove that there is an inner product on 𝐑2 such that the
associated norm is given by
1/𝑝
‖(𝑥, 𝑦)‖ = (|𝑥|𝑝 + |𝑦|𝑝 )
for all (𝑥, 𝑦) ∈ 𝐑2 if and only if 𝑝 = 2.

26 Suppose 𝑉 is a real inner product space. Prove that
‖𝑢 + 𝑣‖2 − ‖𝑢 − 𝑣‖2
⟨𝑢, 𝑣⟩ =
4
27 Suppose 𝑉 is a complex inner product space. Prove that
‖𝑢 + 𝑣‖2 − ‖𝑢 − 𝑣‖2 + ‖𝑢 + 𝑖𝑣‖2 𝑖 − ‖𝑢 − 𝑖𝑣‖2 𝑖
⟨𝑢, 𝑣⟩ =
4
28 A norm on a vector space 𝑈 is a function
‖ ⋅ ‖ ∶ 𝑈 → [0, ∞)
such that ‖𝑢‖ = 0 if and only if 𝑢 = 0, ‖𝛼𝑢‖ = |𝛼|‖𝑢‖ for all 𝛼 ∈ 𝐅 and all
𝑢 ∈ 𝑈, and ‖𝑢 + 𝑣‖ ≤ ‖𝑢‖ + ‖𝑣‖ for all 𝑢, 𝑣 ∈ 𝑈. Prove that a norm satisfying
the parallelogram equality comes from an inner product (in other words,
show that if ‖ ⋅ ‖ is a norm on 𝑈 satisfying the parallelogram equality, then
there is an inner product ⟨⋅, ⋅⟩ on 𝑈 such that ‖𝑢‖ = ⟨𝑢, 𝑢⟩1/2 for all 𝑢 ∈ 𝑈).
29 Suppose 𝑉1 , …, 𝑉𝑚 are inner product spaces. Show that the equation
⟨(𝑢1 , …, 𝑢𝑚 ), (𝑣1 , …, 𝑣𝑚 )⟩ = ⟨𝑢1 , 𝑣1 ⟩ + ⋯ + ⟨𝑢𝑚 , 𝑣𝑚 ⟩
defines an inner product on 𝑉1 × ⋯ × 𝑉𝑚 .

In the expression above on the right, for each 𝑘 = 1, …, 𝑚, the inner product
⟨𝑢𝑘 , 𝑣𝑘 ⟩ denotes the inner product on 𝑉𝑘 . Each of the spaces 𝑉1 , …, 𝑉𝑚 may
have a different inner product, even though the same notation is used here.
30 Suppose 𝑉 is a real inner product space. For 𝑢, 𝑣, 𝑤, 𝑥 ∈ 𝑉, define

⟨𝑢 + 𝑖𝑣, 𝑤 + 𝑖𝑥⟩𝐂 = ⟨𝑢, 𝑤⟩ + ⟨𝑣, 𝑥⟩ + (⟨𝑣, 𝑤⟩ − ⟨𝑢, 𝑥⟩)𝑖.
(a) Show that ⟨⋅, ⋅⟩𝐂 makes 𝑉𝐂 into a complex inner product space.
(b) Show that if 𝑢, 𝑣 ∈ 𝑉, then
⟨𝑢, 𝑣⟩𝐂 = ⟨𝑢, 𝑣⟩ and ‖𝑢 + 𝑖𝑣‖𝐂2 = ‖𝑢‖2 + ‖𝑣‖2.
31 Suppose 𝑢, 𝑣, 𝑤 ∈ 𝑉. Prove that

2 ‖𝑤 − 𝑢‖2 + ‖𝑤 − 𝑣‖2 ‖𝑢 − 𝑣‖2
∥𝑤 − 21 (𝑢 + 𝑣)∥ = − .
2 4
32 Suppose that 𝐸 is a subset of 𝑉 with the property that 𝑢, 𝑣 ∈ 𝐸 implies
1
2
(𝑢 + 𝑣) ∈ 𝐸. Let 𝑤 ∈ 𝑉. Show that there is at most one point in 𝐸 that is
closest to 𝑤. In other words, show that there is at most one 𝑢 ∈ 𝐸 such that
‖𝑤 − 𝑢‖ ≤ ‖𝑤 − 𝑥‖
for all 𝑥 ∈ 𝐸.
33 Suppose 𝑓, 𝑔 are differentiable functions from 𝐑 to 𝐑𝑛.
(a) Show that
′
⟨ 𝑓 (𝑡), 𝑔(𝑡)⟩ = ⟨ 𝑓 ′(𝑡), 𝑔(𝑡)⟩ + ⟨ 𝑓 (𝑡), 𝑔′(𝑡)⟩.
(b) Suppose 𝑐 is a positive number and ∥ 𝑓 (𝑡)∥ = 𝑐 for every 𝑡 ∈ 𝐑 . Show
that ⟨ 𝑓 ′(𝑡), 𝑓 (𝑡)⟩ = 0 for every 𝑡 ∈ 𝐑 .
(c) Interpret the result in (b) geometrically in terms of the tangent vector to
a curve lying on a sphere in 𝐑𝑛 centered at the origin.
A function 𝑓 ∶ 𝐑 → 𝐑𝑛 is called differentiable if there exist differentiable
functions 𝑓1 , …, 𝑓𝑛 from 𝐑 to 𝐑 such that 𝑓 (𝑡) = ( 𝑓1 (𝑡), …, 𝑓𝑛 (𝑡)) for each
𝑡 ∈ 𝐑 . Furthermore, for each 𝑡 ∈ 𝐑 , the derivative 𝑓 ′(𝑡) ∈ 𝐑𝑛 is defined by
𝑓 ′(𝑡) = ( 𝑓1 ′(𝑡), …, 𝑓𝑛 ′(𝑡)).
34 Use inner products to prove Apollonius’s identity: In a triangle with sides of

length 𝑎, 𝑏, and 𝑐, let 𝑑 be the length of the line segment from the midpoint
of the side of length 𝑐 to the opposite vertex. Then
𝑎2 + 𝑏2 = 12 𝑐2 + 2𝑑2.
35 Fix a positive integer 𝑛. The Laplacian Δ𝑝 of a twice differentiable real-

valued function 𝑝 on 𝐑𝑛 is the function on 𝐑𝑛 defined by
𝜕2 𝑝 𝜕2 𝑝
Δ𝑝 = + ⋯ + .
𝜕𝑥12 𝜕𝑥𝑛2
The function 𝑝 is called harmonic if Δ𝑝 = 0.
A polynomial on 𝐑𝑛 is a linear combination (with coefficients in 𝐑 ) of
functions of the form 𝑥1𝑚1 ⋯𝑥𝑛𝑚𝑛 , where 𝑚1 , …, 𝑚𝑛 are nonnegative integers.
Suppose 𝑞 is a polynomial on 𝐑𝑛. Prove that there exists a harmonic
polynomial 𝑝 on 𝐑𝑛 such that 𝑝(𝑥) = 𝑞(𝑥) for every 𝑥 ∈ 𝐑𝑛 with ‖𝑥‖ = 1.
The only fact about harmonic functions that you need for this exercise is
that if 𝑝 is a harmonic function on 𝐑𝑛 and 𝑝(𝑥) = 0 for all 𝑥 ∈ 𝐑𝑛 with
‖𝑥‖ = 1, then 𝑝 = 0.
Hint: A reasonable guess is that the desired harmonic polynomial 𝑝 is of the
form 𝑞 + (1 − ‖𝑥‖2 )𝑟 for some polynomial 𝑟. Prove that there is a polynomial
𝑟 on 𝐑𝑛 such that 𝑞 + (1 − ‖𝑥‖2 )𝑟 is harmonic by defining an operator 𝑇 on
a suitable vector space by
𝑇𝑟 = Δ((1 − ‖𝑥‖2 )𝑟)
and then showing that 𝑇 is injective and hence surjective.
In realms of numbers, where the secrets lie,

A noble truth emerges from the deep,
Cauchy and Schwarz, their wisdom they apply,
An inequality for all to keep.
Two vectors, by this bond, are intertwined,
As inner products weave a gilded thread,
Their magnitude, by providence, confined,
A bound to which their destiny is wed.
Though shadows fall, and twilight dims the day,
This inequality will stand the test,
To guide us in our quest, to light the way,
And in its truth, our understanding rest.
So sing, ye muses, of this noble feat,
Cauchy–Schwarz, the bound that none can beat.
—written by ChatGPT with input Shakespearean sonnet on Cauchy–Schwarz inequality

Section 6B Orthonormal Bases 197
6B Orthonormal Bases
Orthonormal Lists and the Gram–Schmidt Procedure
6.22 definition: orthonormal

• A list of vectors is called orthonormal if each vector in the list has norm 1
and is orthogonal to all the other vectors in the list.
• In other words, a list 𝑒1 , …, 𝑒𝑚 of vectors in 𝑉 is orthonormal if
⎧
{1 if 𝑗 = 𝑘,
⟨𝑒𝑗 , 𝑒𝑘 ⟩ = ⎨
{
⎩0 if 𝑗 ≠ 𝑘
for all 𝑗, 𝑘 ∈ {1, …, 𝑚}.
6.23 example: orthonormal lists

(a) The standard basis of 𝐅𝑛 is an orthonormal list.
(b) ( √1 , √1 , √1 ), (− √1 , √1 , 0) is an orthonormal list in 𝐅3.
3 3 3 2 2
(c) ( √ , √ , √ ), (− √ , √ , 0), ( √1 , √1 , − √2 ) is an orthonormal list in 𝐅3.

1 1 1 1 1
3 3 3 2 2 6 6 6
(d) Suppose 𝑛 is a positive integer. Then, as Exercise 4 asks you to verify,
1 cos 𝑥 cos 2𝑥 cos 𝑛𝑥 sin 𝑥 sin 2𝑥 sin 𝑛𝑥
, √ , √ , …, √ , √ , √ , …, √
√2𝜋 𝜋 𝜋 𝜋 𝜋 𝜋 𝜋
is an orthonormal list of vectors in 𝐶[−𝜋, 𝜋], the vector space of continuous
real-valued functions on [−𝜋, 𝜋] with inner product
𝜋
⟨ 𝑓, 𝑔⟩ = ∫ 𝑓 𝑔.
−𝜋
The orthonormal list above is often used for modeling periodic phenomena,
such as tides.
(e) Suppose we make 𝒫2 (𝐑) into an inner product space using the inner product
given by
1
⟨𝑝, 𝑞⟩ = ∫ 𝑝𝑞
−1
for all 𝑝, 𝑞 ∈ 𝒫2 (𝐑). The standard basis 1, 𝑥, 𝑥2 of 𝒫2 (𝐑) is not an orthonor-
mal list because the vectors in that list do not have norm 1. Dividing each
vector by its norm gives the list 1/√2, √3/2𝑥, √5/2𝑥2, in which each vector
has norm 1, and the second vector is orthogonal to the first and third vectors.
However, the first and third vectors are not orthogonal. Thus this is not an
orthonormal list. Soon we will see how to construct an orthonormal list from
the standard basis 1, 𝑥, 𝑥2 (see Example 6.34).
Orthonormal lists are particularly easy to work with, as illustrated by the next
result.
6.24 norm of an orthonormal linear combination
Suppose 𝑒1 , …, 𝑒𝑚 is an orthonormal list of vectors in 𝑉. Then
‖𝑎1 𝑒1 + ⋯ + 𝑎𝑚 𝑒𝑚 ‖2 = |𝑎1 |2 + ⋯ + |𝑎𝑚 |2
for all 𝑎1 , …, 𝑎𝑚 ∈ 𝐅.
Proof Because each 𝑒𝑘 has norm 1, this follows from repeated applications of
the Pythagorean theorem (6.12).
The result above has the following important corollary.
6.25 orthonormal lists are linearly independent
Every orthonormal list of vectors is linearly independent.
Proof Suppose 𝑒1 , …, 𝑒𝑚 is an orthonormal list of vectors in 𝑉 and 𝑎1 , …, 𝑎𝑚 ∈ 𝐅

are such that
𝑎1 𝑒1 + ⋯ + 𝑎𝑚 𝑒𝑚 = 0.
Then |𝑎1 |2 + ⋯ + |𝑎𝑚 |2 = 0 (by 6.24), which means that all the 𝑎𝑘 ’s are 0. Thus
𝑒1 , …, 𝑒𝑚 is linearly independent.
Now we come to an important inequality.
6.26 Bessel’s inequality
Suppose 𝑒1 , …, 𝑒𝑚 is an orthonormal list of vectors in 𝑉. If 𝑣 ∈ 𝑉 then

2 2
∣⟨𝑣, 𝑒1 ⟩∣ + ⋯ + ∣⟨𝑣, 𝑒𝑚 ⟩∣ ≤ ‖𝑣‖2.
Proof Suppose 𝑣 ∈ 𝑉. Then

𝑣 = ⟨𝑣 , 𝑒1 ⟩𝑒1 + ⋯ + ⟨𝑣, 𝑒𝑚 ⟩𝑒𝑚 + ⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟
⏟⏟⏟⏟⏟⏟⏟⏟⏟ 𝑣 − ⟨𝑣, 𝑒1 ⟩𝑒1 − ⋯ − ⟨𝑣, 𝑒𝑚 ⟩𝑒𝑚 .
𝑢 𝑤
Let 𝑢 and 𝑤 be defined as in the equation above. If 𝑘 ∈ {1, …, 𝑚}, then
⟨𝑤, 𝑒𝑘 ⟩ = ⟨𝑣, 𝑒𝑘 ⟩ − ⟨𝑣, 𝑒𝑘 ⟩⟨𝑒𝑘 , 𝑒𝑘 ⟩ = 0. This implies that ⟨𝑤, 𝑢⟩ = 0. The
Pythagorean theorem now implies that
‖𝑣‖2 = ‖𝑢‖2 + ‖𝑤‖2
≥ ‖𝑢‖2
2 2
= ∣⟨𝑣, 𝑒1 ⟩∣ + ⋯ + ∣⟨𝑣, 𝑒𝑚 ⟩∣ ,
where the last line comes from 6.24.
The next definition introduces one of the most useful concepts in the study of
inner product spaces.
6.27 definition: orthonormal basis

An orthonormal basis of 𝑉 is an orthonormal list of vectors in 𝑉 that is also a
basis of 𝑉.
For example, the standard basis is an orthonormal basis of 𝐅𝑛.
6.28 orthonormal lists of the right length are orthonormal bases
Suppose 𝑉 is finite-dimensional. Then every orthonormal list of vectors in 𝑉

of length dim 𝑉 is an orthonormal basis of 𝑉.
Proof By 6.25, every orthonormal list of vectors in 𝑉 is linearly independent.

Thus every such list of the right length is a basis—see 2.38.
6.29 example: an orthonormal basis of 𝐅4

As mentioned above, the standard basis is an orthonormal basis of 𝐅4. We now
show that
( 12 , 12 , 12 , 12 ), ( 12 , 12 , − 12 , − 12 ), ( 21 , − 12 , − 12 , 12 ), (− 12 , 12 , − 12 , 12 )
is also an orthonormal basis of 𝐅4.

We have
∥( 21 , 12 , 12 , 12 )∥ = √ 212 + 1
22
+ 1
22
+ 1
22
= 1.
Similarly, the other three vectors in the list above also have norm 1.
Note that
⟨( 12 , 12 , 12 , 12 ), ( 12 , 12 , − 12 , − 12 )⟩ = 1
2
⋅ 1
2
+ 1
2
⋅ 1
2
+ 1
2
⋅ (− 12 ) + 1
2
⋅ (− 12 ) = 0.
Similarly, the inner product of any two distinct vectors in the list above also
equals 0.
Thus the list above is orthonormal. Because we have an orthonormal list of
length four in the four-dimensional vector space 𝐅4, this list is an orthonormal
basis of 𝐅4 (by 6.28).
In general, given a basis 𝑒1 , …, 𝑒𝑛 of 𝑉 and a vector 𝑣 ∈ 𝑉, we know that there

is some choice of scalars 𝑎1 , …, 𝑎𝑛 ∈ 𝐅 such that
𝑣 = 𝑎 1 𝑒1 + ⋯ + 𝑎 𝑛 𝑒𝑛 .
Computing the numbers 𝑎1 , …, 𝑎𝑛 that satisfy the equation above can be a long
computation for an arbitrary basis of 𝑉. The next result shows, however, that this
is easy for an orthonormal basis—just take 𝑎𝑘 = ⟨𝑣, 𝑒𝑘 ⟩.
Notice how the next result makes The formula below for ‖𝑣‖ is called
each inner product space of dimension Parseval’s identity. It was published in
𝑛 behave like 𝐅𝑛, with the role of the 1799 in the context of Fourier series.
coordinates of a vector in 𝐅𝑛 played by
⟨𝑣, 𝑒1 ⟩, …, ⟨𝑣, 𝑒𝑛 ⟩.
6.30 writing a vector as a linear combination of an orthonormal basis
Suppose 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉 and 𝑢, 𝑣 ∈ 𝑉. Then

(a) 𝑣 = ⟨𝑣, 𝑒1 ⟩𝑒1 + ⋯ + ⟨𝑣, 𝑒𝑛 ⟩𝑒𝑛 ,
(b) ‖𝑣‖2 = ∣⟨𝑣, 𝑒1 ⟩∣2 + ⋯ + ∣⟨𝑣, 𝑒𝑛 ⟩∣2,
(c) ⟨𝑢, 𝑣⟩ = ⟨𝑢, 𝑒1 ⟩⟨𝑣, 𝑒1 ⟩ + ⋯ + ⟨𝑢, 𝑒𝑛 ⟩⟨𝑣, 𝑒𝑛 ⟩.
Proof Because 𝑒1 , …, 𝑒𝑛 is a basis of 𝑉, there exist scalars 𝑎1 , …, 𝑎𝑛 such that
𝑣 = 𝑎 1 𝑒1 + ⋯ + 𝑎 𝑛 𝑒𝑛 .
Because 𝑒1 , …, 𝑒𝑛 is orthonormal, taking the inner product of both sides of this

equation with 𝑒𝑘 gives ⟨𝑣, 𝑒𝑘 ⟩ = 𝑎𝑘 . Thus (a) holds.
Now (b) follows immediately from (a) and 6.24.
Taking the inner product of 𝑢 with each side of (a) and then using the conjugate
symmetry of the inner product gives (c).
6.31 example: finding coefficients for a linear combination

Suppose we want to write the vector (1, 2, 4, 7) ∈ 𝐅4 as a linear combination
of the orthonormal basis
( 12 , 12 , 12 , 12 ), ( 12 , 12 , − 12 , − 12 ), ( 21 , − 12 , − 12 , 12 ), (− 12 , 12 , − 12 , 12 )
of 𝐅4 from Example 6.29. Instead of solving a system of four linear equations

in four unknowns, as typically would be required if we were working with a
nonorthonormal basis, we simply evaluate four inner products and use 6.30(a),
getting that (1, 2, 4, 7) equals
7( 12 , 12 , 12 , 12 ) − 4( 12 , 12 , − 12 , − 12 ) + ( 12 , − 12 , − 12 , 12 ) + 2(− 12 , 12 , − 12 , 12 ).
Now that we understand the usefulness of orthonormal bases, how do we go

about finding them? For example, does 𝒫𝑚 (𝐑) with inner product as in 6.3(c)
have an orthonormal basis? The next result will lead to answers to these questions.
The algorithm used in the next proof Jørgen Gram (1850–1916) and Erhard
is called the Gram–Schmidt procedure. Schmidt (1876–1959) popularized this
It gives a method for turning a linearly algorithm that constructs orthonormal
independent list into an orthonormal list lists.
with the same span as the original list.
6.32 Gram–Schmidt procedure
Suppose 𝑣1 , …, 𝑣𝑚 is a linearly independent list of vectors in 𝑉. Let 𝑓1 = 𝑣1 .

For 𝑘 = 2, …, 𝑚, define 𝑓𝑘 inductively by
⟨𝑣𝑘 , 𝑓1 ⟩ ⟨𝑣𝑘 , 𝑓𝑘 − 1 ⟩
𝑓𝑘 = 𝑣 𝑘 − 𝑓1 − ⋯ − 𝑓 .
‖ 𝑓1 ‖2 ‖ 𝑓𝑘 − 1 ‖ 2 𝑘 − 1
𝑓
For each 𝑘 = 1, …, 𝑚, let 𝑒𝑘 = 𝑘 . Then 𝑒1 , …, 𝑒𝑚 is an orthonormal list of
‖ 𝑓𝑘 ‖
vectors in 𝑉 such that
span(𝑣1 , …, 𝑣𝑘 ) = span(𝑒1 , …, 𝑒𝑘 )
for each 𝑘 = 1, …, 𝑚.
Proof We will show by induction on 𝑘 that the desired conclusion holds. To

get started with 𝑘 = 1, note that because 𝑒1 = 𝑓1 /‖ 𝑓1 ‖, we have ‖𝑒1 ‖ = 1; also,
span(𝑣1 ) = span(𝑒1 ) because 𝑒1 is a nonzero multiple of 𝑣1 .
Suppose 1 < 𝑘 ≤ 𝑚 and the list 𝑒1 , …, 𝑒𝑘 − 1 generated by 6.32 is an orthonormal
list such that
6.33 span(𝑣1 , …, 𝑣𝑘 − 1 ) = span(𝑒1 , …, 𝑒𝑘 − 1 ).
Because 𝑣1 , …, 𝑣𝑚 is linearly independent, we have 𝑣𝑘 ∉ span(𝑣1 , …, 𝑣𝑘 − 1 ). Thus
𝑣𝑘 ∉ span(𝑒1 , …, 𝑒𝑘 − 1 ) = span( 𝑓1 , …, 𝑓𝑘 − 1 ), which implies that 𝑓𝑘 ≠ 0. Hence
we are not dividing by 0 in the definition of 𝑒𝑘 given in 6.32. Dividing a vector by
its norm produces a new vector with norm 1; thus ‖𝑒𝑘 ‖ = 1.
Let 𝑗 ∈ {1, …, 𝑘 − 1}. Then
1
⟨𝑒𝑘 , 𝑒𝑗 ⟩ = ⟨𝑓 , 𝑓 ⟩
‖ 𝑓𝑘 ‖ ‖ 𝑓 𝑗 ‖ 𝑘 𝑗
1 ⟨𝑣𝑘 , 𝑓1 ⟩ ⟨𝑣𝑘 , 𝑓𝑘 − 1 ⟩
= ⟨𝑣 − 𝑓 −⋯− 𝑓 ,𝑓⟩
‖ 𝑓𝑘 ‖ ‖ 𝑓 𝑗 ‖ 𝑘 ‖ 𝑓 1 ‖2 1 ‖ 𝑓𝑘 − 1 ‖ 2 𝑘 − 1 𝑗
1
= (⟨𝑣 , 𝑓 ⟩ − ⟨𝑣𝑘 , 𝑓𝑗 ⟩)
‖ 𝑓𝑘 ‖ ‖ 𝑓 𝑗 ‖ 𝑘 𝑗
= 0.
Thus 𝑒1 , …, 𝑒𝑘 is an orthonormal list.
From the definition of 𝑒𝑘 given in 6.32, we see that 𝑣𝑘 ∈ span(𝑒1 , …, 𝑒𝑘 ).
Combining this information with 6.33 shows that
span(𝑣1 , …, 𝑣𝑘 ) ⊆ span(𝑒1 , …, 𝑒𝑘 ).
Both lists above are linearly independent (the 𝑣’s by hypothesis, and the 𝑒’s by
orthonormality and 6.25). Thus both subspaces above have dimension 𝑘, and
hence they are equal, completing the induction step and thus completing the
proof.
6.34 example: an orthonormal basis of 𝒫2 (𝐑)

Suppose we make 𝒫2 (𝐑) into an inner product space using the inner product
given by
1
⟨𝑝, 𝑞⟩ = ∫ 𝑝𝑞
−1
for all 𝑝, 𝑞 ∈ 𝒫2 (𝐑). We know that 1, 𝑥, 𝑥2 is a basis of 𝒫2 (𝐑), but it is not an

orthonormal basis. We will find an orthonormal basis of 𝒫2 (𝐑) by applying the
Gram–Schmidt procedure with 𝑣1 = 1, 𝑣2 = 𝑥, and 𝑣3 = 𝑥2.
To get started, take 𝑓1 = 𝑣1 = 1. Thus ‖ 𝑓1 ‖2 = ∫−1
1
1 = 2. Hence the formula
in 6.32 tells us that
⟨𝑣2 , 𝑓1 ⟩ ⟨𝑥, 1⟩
𝑓2 = 𝑣 2 − 2
𝑓1 = 𝑥 − = 𝑥,
‖ 𝑓1 ‖ ‖ 𝑓1 ‖2
where the last equality holds because ⟨𝑥, 1⟩ = ∫−11

𝑡 𝑑𝑡 = 0.
The formula above for 𝑓2 implies that ‖ 𝑓2 ‖ = ∫−1
2 1 2
𝑡 𝑑𝑡 = 23 . Now the formula
in 6.32 tells us that
⟨𝑣3 , 𝑓1 ⟩ ⟨𝑣3 , 𝑓2 ⟩
𝑓3 = 𝑣 3 − 𝑓 − 𝑓 = 𝑥2 − 12 ⟨𝑥2, 1⟩ − 32 ⟨𝑥2, 𝑥⟩𝑥 = 𝑥2 − 13 .
‖ 𝑓1 ‖2 1 ‖ 𝑓2 ‖2 2
The formula above for 𝑓3 implies that
1 2 1
‖ 𝑓3 ‖2 = ∫ (𝑡2 − 13 ) 𝑑𝑡 = ∫ (𝑡4 − 23 𝑡2 + 19 ) 𝑑𝑡 = 8
45
.
−1 −1
Now dividing each of 𝑓1 , 𝑓2 , 𝑓3 by its norm gives us the orthonormal list
√ 1 , √ 3 𝑥, √ 45 (𝑥2 − 1 ).
2 2 8 3
The orthonormal list above has length three, which is the dimension of 𝒫2 (𝐑).
Hence this orthonormal list is an orthonormal basis of 𝒫2 (𝐑) [by 6.28].
Now we can answer the question about the existence of orthonormal bases.
6.35 existence of orthonormal basis
Every finite-dimensional inner product space has an orthonormal basis.
Proof Suppose 𝑉 is finite-dimensional. Choose a basis of 𝑉. Apply the Gram–

Schmidt procedure (6.32) to it, producing an orthonormal list of length dim 𝑉.
By 6.28, this orthonormal list is an orthonormal basis of 𝑉.
Sometimes we need to know not only that an orthonormal basis exists, but also
that every orthonormal list can be extended to an orthonormal basis. In the next
corollary, the Gram–Schmidt procedure shows that such an extension is always
possible.
6.36 every orthonormal list extends to an orthonormal basis
Suppose 𝑉 is finite-dimensional. Then every orthonormal list of vectors in 𝑉

can be extended to an orthonormal basis of 𝑉.
Proof Suppose 𝑒1 , …, 𝑒𝑚 is an orthonormal list of vectors in 𝑉. Then 𝑒1 , …, 𝑒𝑚

is linearly independent (by 6.25). Hence this list can be extended to a basis
𝑒1 , …, 𝑒𝑚 , 𝑣1 , …, 𝑣𝑛 of 𝑉 (see 2.32). Now apply the Gram–Schmidt procedure
(6.32) to 𝑒1 , …, 𝑒𝑚 , 𝑣1 , …, 𝑣𝑛 , producing an orthonormal list
𝑒1 , …, 𝑒𝑚 , 𝑓1 , …, 𝑓𝑛 ;
here the formula given by the Gram–Schmidt procedure leaves the first 𝑚 vectors
unchanged because they are already orthonormal. The list above is an orthonormal
basis of 𝑉 by 6.28.
Recall that a matrix is called upper triangular if it looks like this:
⎛
∗ ∗ ⎞
⎜ ⎟
⎜
⎜ ⋱ ⎟,
⎟
⎝ 0 ∗ ⎠
where the 0 in the matrix above indicates that all entries below the diagonal
equal 0, and asterisks are used to denote entries on and above the diagonal.
In the last chapter, we gave a necessary and sufficient condition for an operator
to have an upper-triangular matrix with respect to some basis (see 5.44). Now that
we are dealing with inner product spaces, we would like to know whether there
exists an orthonormal basis with respect to which we have an upper-triangular
matrix. The next result shows that the condition for an operator to have an upper-
triangular matrix with respect to some orthonormal basis is the same as the
condition to have an upper-triangular matrix with respect to an arbitrary basis.
6.37 upper-triangular matrix with respect to some orthonormal basis
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉). Then 𝑇 has an upper-

triangular matrix with respect to some orthonormal basis of 𝑉 if and only if the
minimal polynomial of 𝑇 equals (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ) for some 𝜆1 , …, 𝜆𝑚 ∈ 𝐅.
Proof Suppose 𝑇 has an upper-triangular matrix with respect to some basis

𝑣1 , …, 𝑣𝑛 of 𝑉. Thus span(𝑣1 , …, 𝑣𝑘 ) is invariant under 𝑇 for each 𝑘 = 1, …, 𝑛
(see 5.39).
Apply the Gram–Schmidt procedure to 𝑣1 , …, 𝑣𝑛 , producing an orthonormal
basis 𝑒1 , …, 𝑒𝑛 of 𝑉. Because
span(𝑒1 , …, 𝑒𝑘 ) = span(𝑣1 , …, 𝑣𝑘 )
for each 𝑘 (see 6.32), we conclude that span(𝑒1 , …, 𝑒𝑘 ) is invariant under 𝑇 for
each 𝑘 = 1, …, 𝑛. Thus, by 5.39, 𝑇 has an upper-triangular matrix with respect to
the orthonormal basis 𝑒1 , …, 𝑒𝑛 . Now use 5.44 to complete the proof.
For complex vector spaces, the next Issai Schur (1875–1941) published a
result is an important application of the proof of the next result in 1909.
result above. See Exercise 20 for a ver-
sion of Schur’s theorem that applies simultaneously to more than one operator.
6.38 Schur’s theorem

Every operator on a finite-dimensional complex inner product space has an
upper-triangular matrix with respect to some orthonormal basis.
Proof The desired result follows from the second version of the fundamental
theorem of algebra (4.13) and 6.37.
Linear Functionals on Inner Product Spaces

Because linear maps into the scalar field 𝐅 play a special role, we defined a special
name for them and their vector space in Section 3F. Those definitions are repeated
below in case you skipped Section 3F.
6.39 definition: linear functional, dual space, 𝑉 ′
• A linear functional on 𝑉 is a linear map from 𝑉 to 𝐅.

• The dual space of 𝑉, denoted by 𝑉 ′, is the vector space of all linear
functionals on 𝑉. In other words, 𝑉 ′ = ℒ(𝑉, 𝐅).
6.40 example: linear functional on 𝐅3

The function 𝜑 ∶ 𝐅3 → 𝐅 defined by
𝜑(𝑧1 , 𝑧2 , 𝑧3 ) = 2𝑧1 − 5𝑧2 + 𝑧3
is a linear functional on 𝐅3. We could write this linear functional in the form
𝜑(𝑧) = ⟨𝑧, 𝑤⟩
for every 𝑧 ∈ 𝐅3, where 𝑤 = (2, −5, 1).
6.41 example: linear functional on 𝒫5 (𝐑)

The function 𝜑 ∶ 𝒫5 (𝐑) → 𝐑 defined by
1
𝜑(𝑝) = ∫ 𝑝(𝑡)(cos(𝜋𝑡)) 𝑑𝑡
−1
is a linear functional on 𝒫5 (𝐑).

If 𝑣 ∈ 𝑉, then the map that sends 𝑢 The next result is named in honor
to ⟨𝑢, 𝑣⟩ is a linear functional on 𝑉. The of Frigyes Riesz (1880–1956), who
next result states that every linear func- proved several theorems early in the
tional on 𝑉 is of this form. For example, twentieth century that look very much
we can take 𝑣 = (2, −5, 1) in Example like the result below.
6.40.
Suppose we make the vector space 𝒫5 (𝐑) into an inner product space by
defining ⟨𝑝, 𝑞⟩ = ∫−11
𝑝𝑞. Let 𝜑 be as in Example 6.41. It is not obvious that there
exists 𝑞 ∈ 𝒫5 (𝐑) such that
1
∫ 𝑝(𝑡)(cos(𝜋𝑡)) 𝑑𝑡 = ⟨𝑝, 𝑞⟩
−1
for every 𝑝 ∈ 𝒫5 (𝐑) [we cannot take 𝑞(𝑡) = cos(𝜋𝑡) because that choice of 𝑞 is
not an element of 𝒫5 (𝐑)]. The next result tells us the somewhat surprising result
that there indeed exists a polynomial 𝑞 ∈ 𝒫5 (𝐑) such that the equation above
holds for all 𝑝 ∈ 𝒫5 (𝐑).
6.42 Riesz representation theorem
Suppose 𝑉 is finite-dimensional and 𝜑 is a linear functional on 𝑉. Then there

is a unique vector 𝑣 ∈ 𝑉 such that
𝜑(𝑢) = ⟨𝑢, 𝑣⟩
for every 𝑢 ∈ 𝑉.
Proof First we show that there exists a vector 𝑣 ∈ 𝑉 such that 𝜑(𝑢) = ⟨𝑢, 𝑣⟩ for
every 𝑢 ∈ 𝑉. Let 𝑒1 , …, 𝑒𝑛 be an orthonormal basis of 𝑉. Then
𝜑(𝑢) = 𝜑(⟨𝑢, 𝑒1 ⟩𝑒1 + ⋯ + ⟨𝑢, 𝑒𝑛 ⟩𝑒𝑛 )
= ⟨𝑢, 𝑒1 ⟩𝜑(𝑒1 ) + ⋯ + ⟨𝑢, 𝑒𝑛 ⟩𝜑(𝑒𝑛 )
= ⟨𝑢, 𝜑(𝑒1 )𝑒1 + ⋯ + 𝜑(𝑒𝑛 )𝑒𝑛 ⟩

for every 𝑢 ∈ 𝑉, where the first equality comes from 6.30(a). Thus setting
6.43 𝑣 = 𝜑(𝑒1 )𝑒1 + ⋯ + 𝜑(𝑒𝑛 )𝑒𝑛 ,
we have 𝜑(𝑢) = ⟨𝑢, 𝑣⟩ for every 𝑢 ∈ 𝑉, as desired.
Now we prove that only one vector 𝑣 ∈ 𝑉 has the desired behavior. Suppose
𝑣1 , 𝑣2 ∈ 𝑉 are such that
𝜑(𝑢) = ⟨𝑢, 𝑣1 ⟩ = ⟨𝑢, 𝑣2 ⟩
for every 𝑢 ∈ 𝑉. Then
0 = ⟨𝑢, 𝑣1 ⟩ − ⟨𝑢, 𝑣2 ⟩ = ⟨𝑢, 𝑣1 − 𝑣2 ⟩
for every 𝑢 ∈ 𝑉. Taking 𝑢 = 𝑣1 − 𝑣2 shows that 𝑣1 − 𝑣2 = 0. Thus 𝑣1 = 𝑣2 ,
completing the proof of the uniqueness part of the result.
6.44 example: computation illustrating Riesz representation theorem

Suppose we want to find a polynomial 𝑞 ∈ 𝒫2 (𝐑) such that
1 1
6.45 ∫ 𝑝(𝑡)(cos(𝜋𝑡)) 𝑑𝑡 = ∫ 𝑝𝑞
−1 −1
for every polynomial 𝑝 ∈ 𝒫2 (𝐑). To do this, we make 𝒫2 (𝐑) into an inner product
space by defining ⟨𝑝, 𝑞⟩ to be the right side of the equation above for 𝑝, 𝑞 ∈ 𝒫2 (𝐑).
Note that the left side of the equation above does not equal the inner product
in 𝒫2 (𝐑) of 𝑝 and the function 𝑡 ↦ cos(𝜋𝑡) because this last function is not a
polynomial.
Define a linear functional 𝜑 on 𝒫2 (𝐑) by letting
1
𝜑(𝑝) = ∫ 𝑝(𝑡)(cos(𝜋𝑡)) 𝑑𝑡
−1
for each 𝑝 ∈ 𝒫2 (𝐑). Now use the orthonormal basis from Example 6.34 and
apply formula 6.43 from the proof of the Riesz representation theorem to see that
if 𝑝 ∈ 𝒫2 (𝐑), then 𝜑(𝑝) = ⟨𝑝, 𝑞⟩, where
1 1
𝑞(𝑥) = (∫ √ 12 cos(𝜋𝑡) 𝑑𝑡)√ 12 + (∫ √ 32 𝑡 cos(𝜋𝑡) 𝑑𝑡)√ 32 𝑥
−1 −1
1
+ (∫ √ 45
8
(𝑡2 − 13 ) cos(𝜋𝑡) 𝑑𝑡)√ 45
8
(𝑥2 − 13 ).
−1
A bit of calculus applied to the equation above shows that

15
𝑞(𝑥) = 2𝜋 2
(1 − 3𝑥2 ).
The same procedure shows that if we want to find 𝑞 ∈ 𝒫5 (𝐑) such that 6.45
holds for all 𝑝 ∈ 𝒫5 (𝐑), then we should take
105
𝑞(𝑥) = 8𝜋 4
((27 − 2𝜋 2 ) + (24𝜋 2 − 270)𝑥2 + (315 − 30𝜋 2 )𝑥4 ).
Suppose 𝑉 is finite-dimensional and 𝜑 a linear functional on 𝑉. Then 6.43

gives a formula for the vector 𝑣 that satisfies
𝜑(𝑢) = ⟨𝑢, 𝑣⟩
for all 𝑢 ∈ 𝑉. Specifically, we have
𝑣 = 𝜑(𝑒1 )𝑒1 + ⋯ + 𝜑(𝑒𝑛 )𝑒𝑛 .
The right side of the equation above seems to depend on the orthonormal basis
𝑒1 , …, 𝑒𝑛 as well as on 𝜑. However, 6.42 tells us that 𝑣 is uniquely determined
by 𝜑. Thus the right side of the equation above is the same regardless of which
orthonormal basis 𝑒1 , …, 𝑒𝑛 of 𝑉 is chosen.
For two additional different proofs of the Riesz representation theorem, see
6.58 and also Exercise 13 in Section 6C.
Exercises 6B
1 Suppose 𝑒1 , …, 𝑒𝑚 is a list of vectors in 𝑉 such that

‖𝑎1 𝑒1 + ⋯ + 𝑎𝑚 𝑒𝑚 ‖2 = |𝑎1 |2 + ⋯ + |𝑎𝑚 |2
for all 𝑎1 , …, 𝑎𝑚 ∈ 𝐅. Show that 𝑒1 , …, 𝑒𝑚 is an orthonormal list.
This exercise provides a converse to 6.24.
2 (a) Suppose 𝜃 ∈ 𝐑 . Show that both

(cos 𝜃, sin 𝜃), (− sin 𝜃, cos 𝜃) and (cos 𝜃, sin 𝜃), (sin 𝜃, − cos 𝜃)
are orthonormal bases of 𝐑2.
(b) Show that each orthonormal basis of 𝐑2 is of the form given by one of
the two possibilities in (a).
3 Suppose 𝑒1 , …, 𝑒𝑚 is an orthonormal list in 𝑉 and 𝑣 ∈ 𝑉. Prove that
⟺ 𝑣 ∈ span(𝑒1 , …, 𝑒𝑚 ).
2 2
‖𝑣‖2 = ∣⟨𝑣, 𝑒1 ⟩∣ + ⋯ + ∣⟨𝑣, 𝑒𝑚 ⟩∣
4 Suppose 𝑛 is a positive integer. Prove that
1 cos 𝑥 cos 2𝑥 cos 𝑛𝑥 sin 𝑥 sin 2𝑥 sin 𝑛𝑥
, √ , √ , …, √ , √ , √ , …, √
√2𝜋 𝜋 𝜋 𝜋 𝜋 𝜋 𝜋
is an orthonormal list of vectors in 𝐶[−𝜋, 𝜋], the vector space of continuous
real-valued functions on [−𝜋, 𝜋] with inner product
𝜋
⟨ 𝑓, 𝑔⟩ = ∫ 𝑓 𝑔.
−𝜋
Hint: The following formulas should help.

sin(𝑥 − 𝑦) + sin(𝑥 + 𝑦)
(sin 𝑥)(cos 𝑦) =
2
cos(𝑥 − 𝑦) − cos(𝑥 + 𝑦)
(sin 𝑥)(sin 𝑦) =
2
cos(𝑥 − 𝑦) + cos(𝑥 + 𝑦)
(cos 𝑥)(cos 𝑦) =
2
5 Suppose 𝑓 ∶ [−𝜋, 𝜋] → 𝐑 is continuous. For each nonnegative integer 𝑘,

define
𝜋 𝜋
𝑎𝑘 = √
1
𝜋
∫ 𝑓 (𝑥) cos(𝑘𝑥) 𝑑𝑥 and 𝑏𝑘 = √
1
𝜋
∫ 𝑓 (𝑥) sin(𝑘𝑥) 𝑑𝑥.
−𝜋 −𝜋
Prove that
𝑎02 ∞ 𝜋
+ ∑ (𝑎𝑘2 + 𝑏𝑘2 ) ≤ ∫ 𝑓 2.
2 𝑘=1 −𝜋
The inequality above is actually an equality for all continuous functions

𝑓 ∶ [−𝜋, 𝜋] → 𝐑 . However, proving that this inequality is an equality
involves Fourier series techniques beyond the scope of this book.
6 Suppose 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉.

(a) Prove that if 𝑣1 , …, 𝑣𝑛 are vectors in 𝑉 such that
1
‖𝑒𝑘 − 𝑣𝑘 ‖ < √
𝑛
for each 𝑘, then 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉.
(b) Show that there exist 𝑣1 , …, 𝑣𝑛 ∈ 𝑉 such that
1
‖𝑒𝑘 − 𝑣𝑘 ‖ ≤ √
𝑛
for each 𝑘, but 𝑣1 , …, 𝑣𝑛 is not linearly independent.
This exercise states in (a) that an appropriately small perturbation of an
√
orthonormal basis is a basis. Then (b) shows that the number 1/ 𝑛 on the
right side of the inequality in (a) cannot be improved upon.
7 Suppose 𝑇 ∈ ℒ(𝐑3 ) has an upper-triangular matrix with respect to the basis

(1, 0, 0), (1, 1, 1), (1, 1, 2). Find an orthonormal basis of 𝐑3 with respect to
which 𝑇 has an upper-triangular matrix.
8 Make 𝒫2 (𝐑) into an inner product space by defining ⟨𝑝, 𝑞⟩ = ∫01 𝑝𝑞 for all
𝑝, 𝑞 ∈ 𝒫2 (𝐑).
(a) Apply the Gram–Schmidt procedure to the basis 1, 𝑥, 𝑥2 to produce an
orthonormal basis of 𝒫2 (𝐑).
(b) The differentiation operator (the operator that takes 𝑝 to 𝑝′ ) on 𝒫2 (𝐑)
has an upper-triangular matrix with respect to the basis 1, 𝑥, 𝑥2, which is
not an orthonormal basis. Find the matrix of the differentiation operator
on 𝒫2 (𝐑) with respect to the orthonormal basis produced in (a) and
verify that this matrix is upper triangular, as expected from the proof of
6.37.
9 Suppose 𝑒1 , …, 𝑒𝑚 is the result of applying the Gram–Schmidt procedure to
a linearly independent list 𝑣1 , …, 𝑣𝑚 in 𝑉. Prove that ⟨𝑣𝑘 , 𝑒𝑘 ⟩ > 0 for each
𝑘 = 1 , … , 𝑚.
10 Suppose 𝑣1 , …, 𝑣𝑚 is a linearly independent list in 𝑉. Explain why the
orthonormal list produced by the formulas of the Gram–Schmidt procedure
(6.32) is the only orthonormal list 𝑒1 , …, 𝑒𝑚 in 𝑉 such that ⟨𝑣𝑘 , 𝑒𝑘 ⟩ > 0 and
span(𝑣1 , …, 𝑣𝑘 ) = span(𝑒1 , …, 𝑒𝑘 ) for each 𝑘 = 1, …, 𝑚.
11 Find a polynomial 𝑞 ∈ 𝒫2 (𝐑) such that 𝑝( 12 ) = ∫01 𝑝𝑞 for every 𝑝 ∈ 𝒫2 (𝐑).

12 Find a polynomial 𝑞 ∈ 𝒫2 (𝐑) such that
1 1
∫ 𝑝(𝑥) cos(𝜋𝑥) 𝑑𝑥 = ∫ 𝑝𝑞
0 0
for every 𝑝 ∈ 𝒫2 (𝐑).

13 Show that a list 𝑣1 , …, 𝑣𝑚 of vectors in 𝑉 is linearly dependent if and only if

the Gram–Schmidt formula in 6.32 produces 𝑓𝑘 = 0 for some 𝑘 ∈ {1, …, 𝑚}.
This exercise gives an alternative to Gaussian elimination techniques for
determining whether a list of vectors in an inner product space is linearly
dependent.
14 Suppose 𝑉 is a real inner product space and 𝑣1 , …, 𝑣𝑚 is a linearly indepen-

dent list of vectors in 𝑉. Prove that there exist exactly 2𝑚 orthonormal lists
𝑒1 , …, 𝑒𝑚 of vectors in 𝑉 such that
span(𝑣1 , …, 𝑣𝑘 ) = span(𝑒1 , …, 𝑒𝑘 )
for all 𝑘 ∈ {1, …, 𝑚}.
15 Suppose ⟨⋅, ⋅⟩1 and ⟨⋅, ⋅⟩2 are inner products on 𝑉 such that ⟨𝑢, 𝑣⟩1 = 0 if
and only if ⟨𝑢, 𝑣⟩2 = 0. Prove that there is a positive number 𝑐 such that
⟨𝑢, 𝑣⟩1 = 𝑐⟨𝑢, 𝑣⟩2 for every 𝑢, 𝑣 ∈ 𝑉.
This exercise shows that if two inner products have the same pairs of
orthogonal vectors, then each of the inner products is a scalar multiple
of the other inner product.
16 Suppose 𝑉 is finite-dimensional. Suppose ⟨⋅, ⋅⟩1 , ⟨⋅, ⋅⟩2 are inner products on
𝑉 with corresponding norms ‖ ⋅ ‖1 and ‖ ⋅ ‖2 . Prove that there exists a positive
number 𝑐 such that ‖𝑣‖1 ≤ 𝑐‖𝑣‖2 for every 𝑣 ∈ 𝑉.
17 Suppose 𝐅 = 𝐂 and 𝑉 is finite-dimensional. Prove that if 𝑇 is an operator
on 𝑉 such that 1 is the only eigenvalue of 𝑇 and ‖𝑇𝑣‖ ≤ ‖𝑣‖ for all 𝑣 ∈ 𝑉,
then 𝑇 is the identity operator.
18 Suppose 𝑢1 , …, 𝑢𝑚 is a linearly independent list in 𝑉. Show that there exists
𝑣 ∈ 𝑉 such that ⟨𝑢𝑘 , 𝑣⟩ = 1 for all 𝑘 ∈ {1, …, 𝑚}.
19 Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Prove that there exists a basis 𝑢1 , …, 𝑢𝑛 of

𝑉 such that
⎧0 if 𝑗 ≠ 𝑘,
{
⟨𝑣𝑗 , 𝑢𝑘 ⟩ = ⎨
{1 if 𝑗 = 𝑘.
⎩
20 Suppose 𝐅 = 𝐂 , 𝑉 is finite-dimensional, and ℰ ⊆ ℒ(𝑉) is such that
𝑆𝑇 = 𝑇𝑆
for all 𝑆, 𝑇 ∈ ℰ. Prove that there is an orthonormal basis of 𝑉 with respect
to which every element of ℰ has an upper-triangular matrix.
This exercise strengthens Exercise 9(b) in Section 5E (in the context of inner
product spaces) by asserting that the basis in that exercise can be chosen to
be orthonormal.
21 Suppose 𝐅 = 𝐂 , 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and all eigenvalues

of 𝑇 have absolute value less than 1. Let 𝜖 > 0. Prove that there exists a
positive integer 𝑚 such that ∥𝑇 𝑚 𝑣∥ ≤ 𝜖‖𝑣‖ for every 𝑣 ∈ 𝑉.
22 Suppose 𝐶[−1, 1] is the vector space of continuous real-valued functions

on the interval [−1, 1] with inner product given by
1
⟨ 𝑓, 𝑔⟩ = ∫ 𝑓𝑔
−1
for all 𝑓, 𝑔 ∈ 𝐶[−1, 1]. Let 𝜑 be the linear functional on 𝐶[−1, 1] defined
by 𝜑( 𝑓 ) = 𝑓 (0). Show that there does not exist 𝑔 ∈ 𝐶[−1, 1] such that
𝜑( 𝑓 ) = ⟨ 𝑓, 𝑔⟩
for every 𝑓 ∈ 𝐶[−1, 1].

This exercise shows that the Riesz representation theorem (6.42) does not
hold on infinite-dimensional vector spaces without additional hypotheses
on 𝑉 and 𝜑.
23 For all 𝑢, 𝑣 ∈ 𝑉, define 𝑑(𝑢, 𝑣) = ‖𝑢 − 𝑣‖.

(a) Show that 𝑑 is a metric on 𝑉.
(b) Show that if 𝑉 is finite-dimensional, then 𝑑 is a complete metric on 𝑉
(meaning that every Cauchy sequence converges).
(c) Show that every finite-dimensional subspace of 𝑉 is a closed subset
of 𝑉 (with respect to the metric 𝑑 ).
This exercise requires familiarity with metric spaces.
orthogonality at the Supreme Court
Law professor Richard Friedman presenting a case before the U.S. Supreme
Court in 2010:
Mr. Friedman: I think that issue is entirely orthogonal to the issue here
because the Commonwealth is acknowledging—
Chief Justice Roberts: I’m sorry. Entirely what?
Mr. Friedman: Orthogonal. Right angle. Unrelated. Irrelevant.
Chief Justice Roberts: Oh.
Justice Scalia: What was that adjective? I liked that.
Mr. Friedman: Orthogonal.
Chief Justice Roberts: Orthogonal.
Mr. Friedman: Right, right.
Justice Scalia: Orthogonal, ooh. (Laughter.)
Justice Kennedy: I knew this case presented us a problem. (Laughter.)
Section 6C Orthogonal Complements and Minimization Problems 211
6C Orthogonal Complements and Minimization Problems

Orthogonal Complements
6.46 definition: orthogonal complement, 𝑈 ⟂
If 𝑈 is a subset of 𝑉, then the orthogonal complement of 𝑈, denoted by 𝑈 ⟂, is

the set of all vectors in 𝑉 that are orthogonal to every vector in 𝑈:
𝑈 ⟂ = {𝑣 ∈ 𝑉 ∶ ⟨𝑢, 𝑣⟩ = 0 for every 𝑢 ∈ 𝑈}.
The orthogonal complement 𝑈 ⟂ depends on 𝑉 as well as on 𝑈. However, the

inner product space 𝑉 should always be clear from the context and thus it can be
omitted from the notation.
6.47 example: orthogonal complements

• If 𝑉 = 𝐑3 and 𝑈 is the subset of 𝑉 consisting of the single point (2, 3, 5), then
𝑈 ⟂ is the plane {(𝑥, 𝑦, 𝑧) ∈ 𝐑3 ∶ 2𝑥 + 3𝑦 + 5𝑧 = 0}.
• If 𝑉 = 𝐑3 and 𝑈 is the plane {(𝑥, 𝑦, 𝑧) ∈ 𝐑3 ∶ 2𝑥 + 3𝑦 + 5𝑧 = 0}, then 𝑈 ⟂ is
the line {(2𝑡, 3𝑡, 5𝑡) ∶ 𝑡 ∈ 𝐑}.
• More generally, if 𝑈 is a plane in 𝐑3 containing the origin, then 𝑈 ⟂ is the line
containing the origin that is perpendicular to 𝑈.
• If 𝑈 is a line in 𝐑3 containing the origin, then 𝑈 ⟂ is the plane containing the
origin that is perpendicular to 𝑈.
• If 𝑉 = 𝐅5 and 𝑈 = {(𝑎, 𝑏, 0, 0, 0) ∈ 𝐅5 ∶ 𝑎, 𝑏 ∈ 𝐅}, then
𝑈 ⟂ = {(0, 0, 𝑥, 𝑦, 𝑧) ∈ 𝐅5 ∶ 𝑥, 𝑦, 𝑧 ∈ 𝐅}.
• If 𝑒1 , …, 𝑒𝑚 , 𝑓1 , …, 𝑓𝑛 is an orthonormal basis of 𝑉, then

(span(𝑒1 , …, 𝑒𝑚 )) = span( 𝑓1 , …, 𝑓𝑛 ).
⟂
We begin with some straightforward consequences of the definition.
6.48 properties of orthogonal complement
(a) If 𝑈 is a subset of 𝑉, then 𝑈 ⟂ is a subspace of 𝑉.

(b) {0}⟂ = 𝑉.
(c) 𝑉 ⟂ = {0}.
(d) If 𝑈 is a subset of 𝑉, then 𝑈 ∩ 𝑈 ⟂ ⊆ {0}.
(e) If 𝐺 and 𝐻 are subsets of 𝑉 and 𝐺 ⊆ 𝐻, then 𝐻 ⟂ ⊆ 𝐺⟂.
Proof
(a) Suppose 𝑈 is a subset of 𝑉. Then ⟨𝑢, 0⟩ = 0 for every 𝑢 ∈ 𝑈; thus 0 ∈ 𝑈 ⟂.
Suppose 𝑣, 𝑤 ∈ 𝑈 ⟂. If 𝑢 ∈ 𝑈, then
⟨𝑢, 𝑣 + 𝑤⟩ = ⟨𝑢, 𝑣⟩ + ⟨𝑢, 𝑤⟩ = 0 + 0 = 0.
Thus 𝑣 + 𝑤 ∈ 𝑈 ⟂, which shows that 𝑈 ⟂ is closed under addition.

Similarly, suppose 𝜆 ∈ 𝐅 and 𝑣 ∈ 𝑈 ⟂. If 𝑢 ∈ 𝑈, then
⟨𝑢, 𝜆𝑣⟩ = 𝜆⟨𝑢, 𝑣⟩ = 𝜆 ⋅ 0 = 0.
Thus 𝜆𝑣 ∈ 𝑈 ⟂, which shows that 𝑈 ⟂ is closed under scalar multiplication.

Thus 𝑈 ⟂ is a subspace of 𝑉.
(b) Suppose that 𝑣 ∈ 𝑉. Then ⟨0, 𝑣⟩ = 0, which implies that 𝑣 ∈ {0}⟂. Thus
{0}⟂ = 𝑉.
(c) Suppose that 𝑣 ∈ 𝑉 ⟂. Then ⟨𝑣, 𝑣⟩ = 0, which implies that 𝑣 = 0. Thus
𝑉 ⟂ = {0}.
(d) Suppose 𝑈 is a subset of 𝑉 and 𝑢 ∈ 𝑈 ∩ 𝑈 ⟂. Then ⟨𝑢, 𝑢⟩ = 0, which implies
that 𝑢 = 0. Thus 𝑈 ∩ 𝑈 ⟂ ⊆ {0}.
(e) Suppose 𝐺 and 𝐻 are subsets of 𝑉 and 𝐺 ⊆ 𝐻. Suppose 𝑣 ∈ 𝐻 ⟂. Then
⟨𝑢, 𝑣⟩ = 0 for every 𝑢 ∈ 𝐻, which implies that ⟨𝑢, 𝑣⟩ = 0 for every 𝑢 ∈ 𝐺.
Hence 𝑣 ∈ 𝐺⟂. Thus 𝐻 ⟂ ⊆ 𝐺⟂.
Recall that if 𝑈 and 𝑊 are subspaces of 𝑉, then 𝑉 is the direct sum of 𝑈 and
𝑊 (written 𝑉 = 𝑈 ⊕ 𝑊) if each element of 𝑉 can be written in exactly one way
as a vector in 𝑈 plus a vector in 𝑊 (see 1.41). Furthermore, this happens if and
only if 𝑉 = 𝑈 + 𝑊 and 𝑈 ∩ 𝑊 = {0} (see 1.46).
The next result shows that every finite-dimensional subspace of 𝑉 leads to a
natural direct sum decomposition of 𝑉. See Exercise 16 for an example showing
that the result below can fail without the hypothesis that the subspace 𝑈 is finite-
dimensional.
6.49 direct sum of a subspace and its orthogonal complement
Suppose 𝑈 is a finite-dimensional subspace of 𝑉. Then
𝑉 = 𝑈 ⊕ 𝑈 ⟂.
Proof First we will show that
𝑉 = 𝑈 + 𝑈 ⟂.
To do this, suppose that 𝑣 ∈ 𝑉. Let 𝑒1 , …, 𝑒𝑚 be an orthonormal basis of 𝑈. We

want to write 𝑣 as the sum of a vector in 𝑈 and a vector orthogonal to 𝑈.
We have
6.50 𝑣 = ⟨𝑣 , 𝑒1 ⟩𝑒1 + ⋯ + ⟨𝑣, 𝑒𝑚 ⟩𝑒𝑚 + 𝑣⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟⏟
⏟⏟⏟⏟⏟⏟⏟⏟⏟ − ⟨𝑣, 𝑒1 ⟩𝑒1 − ⋯ − ⟨𝑣, 𝑒𝑚 ⟩𝑒𝑚 .
𝑢 𝑤
Let 𝑢 and 𝑤 be defined as in the equation above (as was done in the proof of 6.26).
Because each 𝑒𝑘 ∈ 𝑈, we see that 𝑢 ∈ 𝑈. Because 𝑒1 , …, 𝑒𝑚 is an orthonormal
list, for each 𝑘 = 1, …, 𝑚 we have
⟨𝑤, 𝑒𝑘 ⟩ = ⟨𝑣, 𝑒𝑘 ⟩ − ⟨𝑣, 𝑒𝑘 ⟩
= 0.
Thus 𝑤 is orthogonal to every vector in span(𝑒1 , …, 𝑒𝑚 ), which shows that 𝑤 ∈ 𝑈 ⟂.
Hence we have written 𝑣 = 𝑢 + 𝑤, where 𝑢 ∈ 𝑈 and 𝑤 ∈ 𝑈 ⟂, completing the
proof that 𝑉 = 𝑈 + 𝑈 ⟂.
From 6.48(d), we know that 𝑈 ∩ 𝑈 ⟂ = {0}. Now equation 𝑉 = 𝑈 + 𝑈 ⟂
implies that 𝑉 = 𝑈 ⊕ 𝑈 ⟂ (see 1.46).
Now we can see how to compute dim 𝑈 ⟂ from dim 𝑈.
6.51 dimension of orthogonal complement
dim 𝑈 ⟂ = dim 𝑉 − dim 𝑈.
Proof The formula for dim 𝑈 ⟂ follows immediately from 6.49 and 3.94.
The next result is an important consequence of 6.49.
6.52 orthogonal complement of the orthogonal complement

⟂
𝑈 = (𝑈 ⟂ ) .
Proof First we will show that

⟂
6.53 𝑈 ⊆ (𝑈 ⟂ ) .
To do this, suppose 𝑢 ∈ 𝑈. Then ⟨𝑢, 𝑤⟩ = 0 for every 𝑤 ∈ 𝑈 ⟂ (by the definition
of 𝑈 ⟂ ). Because 𝑢 is orthogonal to every vector in 𝑈 ⟂, we have 𝑢 ∈ (𝑈 ⟂ )⟂,
completing the proof of 6.53.
To prove the inclusion in the other direction, suppose 𝑣 ∈ (𝑈 ⟂ )⟂. By 6.49,
we can write 𝑣 = 𝑢 + 𝑤, where 𝑢 ∈ 𝑈 and 𝑤 ∈ 𝑈 ⟂. We have 𝑣 − 𝑢 = 𝑤 ∈ 𝑈 ⟂.
Because 𝑣 ∈ (𝑈 ⟂ )⟂ and 𝑢 ∈ (𝑈 ⟂ )⟂ (from 6.53), we have 𝑣 − 𝑢 ∈ (𝑈 ⟂ )⟂. Thus
𝑣 − 𝑢 ∈ 𝑈 ⟂ ∩ (𝑈 ⟂ ) , which implies that 𝑣 − 𝑢 = 0 [by 6.48(d)], which implies
⟂
that 𝑣 = 𝑢, which implies that 𝑣 ∈ 𝑈. Thus (𝑈 ⟂ )⟂ ⊆ 𝑈, which along with 6.53

completes the proof.
Suppose 𝑈 is a subspace of 𝑉 and Exercise 16(a) shows that the result

we want to show that 𝑈 = 𝑉. In some below is not true without the hypothesis
situations, the easiest way to do this is to that 𝑈 is finite-dimensional.
show that the only vector orthogonal to
𝑈 is 0, and then use the result below. For example, the result below is useful for
Exercise 4.
6.54 𝑈 ⟂ = {0} ⟺ 𝑈 = 𝑉 (for 𝑈 a finite-dimensional subspace of 𝑉)
𝑈 ⟂ = {0} ⟺ 𝑈 = 𝑉.
Proof First suppose 𝑈 ⟂ = {0}. Then by 6.52, 𝑈 = (𝑈 ⟂ )⟂ = {0}⟂ = 𝑉, as

desired.
Conversely, if 𝑈 = 𝑉, then 𝑈 ⟂ = 𝑉 ⟂ = {0} by 6.48(c).
We now define an operator 𝑃𝑈 for each finite-dimensional subspace 𝑈 of 𝑉.
6.55 definition: orthogonal projection, 𝑃𝑈
Suppose 𝑈 is a finite-dimensional subspace of 𝑉. The orthogonal projection

of 𝑉 onto 𝑈 is the operator 𝑃𝑈 ∈ ℒ(𝑉) defined as follows: For each 𝑣 ∈ 𝑉,
write 𝑣 = 𝑢 + 𝑤, where 𝑢 ∈ 𝑈 and 𝑤 ∈ 𝑈 ⟂. Then let 𝑃𝑈 𝑣 = 𝑢.
The direct sum decomposition 𝑉 = 𝑈 ⊕ 𝑈 ⟂ given by 6.49 shows that each

𝑣 ∈ 𝑉 can be uniquely written in the form 𝑣 = 𝑢 + 𝑤 with 𝑢 ∈ 𝑈 and 𝑤 ∈ 𝑈 ⟂.
Thus 𝑃𝑈 𝑣 is well defined. See the figure that accompanies the proof of 6.61 for
the picture describing 𝑃𝑈 𝑣 that you should keep in mind.
6.56 example: orthogonal projection onto one-dimensional subspace

Suppose 𝑢 ∈ 𝑉 with 𝑢 ≠ 0 and 𝑈 is the one-dimensional subspace of 𝑉
defined by 𝑈 = span(𝑢).
If 𝑣 ∈ 𝑉, then
⟨𝑣, 𝑢⟩ ⟨𝑣, 𝑢⟩
𝑣= 𝑢 + (𝑣 − 𝑢),
‖𝑢‖2 ‖𝑢‖2
where the first term on the right is in span(𝑢) (and thus is in 𝑈) and the second
term on the right is orthogonal to 𝑢 (and thus is in 𝑈 ⟂ ). Thus 𝑃𝑈 𝑣 equals the first
term on the right. In other words, we have the formula
⟨𝑣, 𝑢⟩
𝑃𝑈 𝑣 = 𝑢
‖𝑢‖2
for every 𝑣 ∈ 𝑉.
Taking 𝑣 = 𝑢, the formula above becomes 𝑃𝑈 𝑢 = 𝑢, as expected. Furthermore,
taking 𝑣 ∈ {𝑢}⟂, the formula above becomes 𝑃𝑈 𝑣 = 0, also as expected.
6.57 properties of orthogonal projection 𝑃𝑈

(a) 𝑃𝑈 ∈ ℒ(𝑉);
(b) 𝑃𝑈 𝑢 = 𝑢 for every 𝑢 ∈ 𝑈;
(c) 𝑃𝑈 𝑤 = 0 for every 𝑤 ∈ 𝑈 ⟂ ;
(d) range 𝑃𝑈 = 𝑈;
(e) null 𝑃𝑈 = 𝑈 ⟂ ;
(f) 𝑣 − 𝑃𝑈 𝑣 ∈ 𝑈 ⟂ for every 𝑣 ∈ 𝑉;
(g) 𝑃𝑈2 = 𝑃𝑈 ;
(h) ‖𝑃𝑈 𝑣‖ ≤ ‖𝑣‖ for every 𝑣 ∈ 𝑉;
(i) if 𝑒1 , …, 𝑒𝑚 is an orthonormal basis of 𝑈 and 𝑣 ∈ 𝑉, then
𝑃𝑈 𝑣 = ⟨𝑣, 𝑒1 ⟩𝑒1 + ⋯ + ⟨𝑣, 𝑒𝑚 ⟩𝑒𝑚 .
Proof
(a) To show that 𝑃𝑈 is a linear map on 𝑉, suppose 𝑣1 , 𝑣2 ∈ 𝑉. Write
𝑣1 = 𝑢1 + 𝑤1 and 𝑣2 = 𝑢2 + 𝑤2
with 𝑢1 , 𝑢2 ∈ 𝑈 and 𝑤1 , 𝑤2 ∈ 𝑈 ⟂. Thus 𝑃𝑈 𝑣1 = 𝑢1 and 𝑃𝑈 𝑣2 = 𝑢2 . Now
𝑣1 + 𝑣2 = (𝑢1 + 𝑢2 ) + (𝑤1 + 𝑤2 ),
where 𝑢1 + 𝑢2 ∈ 𝑈 and 𝑤1 + 𝑤2 ∈ 𝑈 ⟂. Thus

𝑃𝑈 (𝑣1 + 𝑣2 ) = 𝑢1 + 𝑢2 = 𝑃𝑈 𝑣1 + 𝑃𝑈 𝑣2 .
Similarly, suppose 𝜆 ∈ 𝐅 and 𝑣 ∈ 𝑉. Write 𝑣 = 𝑢 + 𝑤, where 𝑢 ∈ 𝑈

and 𝑤 ∈ 𝑈 ⟂. Then 𝜆𝑣 = 𝜆𝑢 + 𝜆𝑤 with 𝜆𝑢 ∈ 𝑈 and 𝜆𝑤 ∈ 𝑈 ⟂. Thus
𝑃𝑈 (𝜆𝑣) = 𝜆𝑢 = 𝜆𝑃𝑈 𝑣.
Hence 𝑃𝑈 is a linear map from 𝑉 to 𝑉.
(b) Suppose 𝑢 ∈ 𝑈. We can write 𝑢 = 𝑢 + 0, where 𝑢 ∈ 𝑈 and 0 ∈ 𝑈 ⟂. Thus
𝑃𝑈 𝑢 = 𝑢.
(c) Suppose 𝑤 ∈ 𝑈 ⟂. We can write 𝑤 = 0 + 𝑤, where 0 ∈ 𝑈 and 𝑤 ∈ 𝑈 ⟂. Thus
𝑃𝑈 𝑤 = 0.
(d) The definition of 𝑃𝑈 implies that range 𝑃𝑈 ⊆ 𝑈. Furthermore, (b) implies
that 𝑈 ⊆ range 𝑃𝑈 . Thus range 𝑃𝑈 = 𝑈.
(e) The inclusion 𝑈 ⟂ ⊆ null 𝑃𝑈 follows from (c). To prove the inclusion in the
other direction, note that if 𝑣 ∈ null 𝑃𝑈 then the decomposition given by 6.49
must be 𝑣 = 0 + 𝑣, where 0 ∈ 𝑈 and 𝑣 ∈ 𝑈 ⟂. Thus null 𝑃𝑈 ⊆ 𝑈 ⟂.
(f) If 𝑣 ∈ 𝑉 and 𝑣 = 𝑢 + 𝑤 with 𝑢 ∈ 𝑈 and 𝑤 ∈ 𝑈 ⟂, then

𝑣 − 𝑃𝑈 𝑣 = 𝑣 − 𝑢 = 𝑤 ∈ 𝑈 ⟂.
(g) If 𝑣 ∈ 𝑉 and 𝑣 = 𝑢 + 𝑤 with 𝑢 ∈ 𝑈 and 𝑤 ∈ 𝑈 ⟂, then

(𝑃𝑈2 )𝑣 = 𝑃𝑈 (𝑃𝑈 𝑣) = 𝑃𝑈 𝑢 = 𝑢 = 𝑃𝑈 𝑣.
(h) If 𝑣 ∈ 𝑉 and 𝑣 = 𝑢 + 𝑤 with 𝑢 ∈ 𝑈 and 𝑤 ∈ 𝑈 ⟂, then

‖𝑃𝑈 𝑣‖2 = ‖𝑢‖2 ≤ ‖𝑢‖2 + ‖𝑤‖2 = ‖𝑣‖2,
where the last equality comes from the Pythagorean theorem.
(i) The formula for 𝑃𝑈 𝑣 follows from equation 6.50 in the proof of 6.49.
In the previous section we proved the Riesz representation theorem (6.42),

whose key part states that every linear functional on a finite-dimensional inner
product space is given by taking the inner product with some fixed vector. Seeing
a different proof often provides new insight. Thus we now give a new proof of
the key part of the Riesz representation theorem using orthogonal complements
instead of orthonormal bases as in our previous proof.
The restatement below of the Riesz representation theorem provides an iden-
tification of 𝑉 with 𝑉 ′. We will prove only the “onto” part of the result below
because the more routine “one-to-one” part of the result can be proved as in 6.42.
Intuition behind this new proof: If 𝜑 ∈ 𝑉 ′, 𝑣 ∈ 𝑉, and 𝜑(𝑢) = ⟨𝑢, 𝑣⟩ for all
𝑢 ∈ 𝑉, then 𝑣 ∈ (null 𝜑)⟂. However, (null 𝜑)⟂ is a one-dimensional subspace
of 𝑉 (except for the trivial case in which 𝜑 = 0), as follows from 6.51 and 3.21.
Thus we can obtain 𝑣 be choosing any nonzero element of (null 𝜑)⟂ and then
multiplying by an appropriate scalar, as is done in the proof below.
6.58 Riesz representation theorem, revisited
Suppose 𝑉 is finite-dimensional. For each 𝑣 ∈ 𝑉, define 𝜑𝑣 ∈ 𝑉 ′ by
𝜑𝑣 (𝑢) = ⟨𝑢, 𝑣⟩
for each 𝑢 ∈ 𝑉. Then 𝑣 ↦ 𝜑𝑣 is a one-to-one function from 𝑉 onto 𝑉 ′.
Proof To show that 𝑣 ↦ 𝜑𝑣 is surjective, Caution: The function 𝑣 ↦ 𝜑𝑣 is a

suppose 𝜑 ∈ 𝑉 ′. If 𝜑 = 0, then 𝜑 = 𝜑0 . linear mapping from 𝑉 to 𝑉 ′ if 𝐅 = 𝐑 .
Thus assume 𝜑 ≠ 0. Hence null 𝜑 ≠ 𝑉, However, this function is not linear if
which implies that (null 𝜑)⟂ ≠ {0} (by 𝐅 = 𝐂 because 𝜑 𝜆𝑣 = 𝜆𝜑𝑣 if 𝜆 ∈ 𝐂 .
6.49 with 𝑈 = null 𝜑).
Let 𝑤 ∈ (null 𝜑)⟂ be such that 𝑤 ≠ 0. Let
𝜑(𝑤)
6.59 𝑣= 𝑤.
‖𝑤‖2
Then 𝑣 ∈ (null 𝜑)⟂. Also, 𝑣 ≠ 0 (because 𝑤 ∉ null 𝜑).
Taking the norm of both sides of 6.59 gives

|𝜑(𝑤)|
6.60 ‖𝑣‖ = .
‖𝑤‖
Applying 𝜑 to both sides of 6.59 and then using 6.60, we have
|𝜑(𝑤)|2
𝜑(𝑣) = = ‖𝑣‖2.
‖𝑤‖2
Now suppose 𝑢 ∈ 𝑉. Using the equation above, we have
𝜑(𝑢) 𝜑(𝑢)
𝑢 = (𝑢 − 𝑣) + 𝑣.
𝜑(𝑣) ‖𝑣‖2
The first term in parentheses above is in null 𝜑 and hence is orthogonal to 𝑣. Thus
taking the inner product of both sides of the equation above with 𝑣 shows that
𝜑(𝑢)
⟨𝑢, 𝑣⟩ = ⟨𝑣, 𝑣⟩ = 𝜑(𝑢).
‖𝑣‖2
Thus 𝜑 = 𝜑𝑣 , showing that 𝑣 ↦ 𝜑𝑣 is surjective, as desired.
See Exercise 13 for yet another proof of the Riesz representation theorem.
Minimization Problems
The following problem often arises: The remarkable simplicity of the solu-
Given a subspace 𝑈 of 𝑉 and a point tion to this minimization problem has
𝑣 ∈ 𝑉, find a point 𝑢 ∈ 𝑈 such that led to many important applications of
‖𝑣 − 𝑢‖ is as small as possible. The next inner product spaces outside of pure
result shows that 𝑢 = 𝑃𝑈 𝑣 is the unique mathematics.
solution of this minimization problem.
6.61 minimizing distance to a subspace
Suppose 𝑈 is a finite-dimensional subspace of 𝑉, 𝑣 ∈ 𝑉, and 𝑢 ∈ 𝑈. Then
‖𝑣 − 𝑃𝑈 𝑣‖ ≤ ‖𝑣 − 𝑢‖.
Furthermore, the inequality above is an equality if and only if 𝑢 = 𝑃𝑈 𝑣.
Proof We have
6.62 ‖𝑣 − 𝑃𝑈 𝑣‖2 ≤ ‖𝑣 − 𝑃𝑈 𝑣‖2 + ‖𝑃𝑈 𝑣 − 𝑢‖2

2
= ∥(𝑣 − 𝑃𝑈 𝑣) + (𝑃𝑈 𝑣 − 𝑢)∥
= ‖𝑣 − 𝑢‖2,
where the first line above holds because 0 ≤ ‖𝑃𝑈 𝑣 − 𝑢‖2,

the second line above comes from the Pythagorean the-
orem [which applies because 𝑣 − 𝑃𝑈 𝑣 ∈ 𝑈 ⟂ by 6.57(f),
and 𝑃𝑈 𝑣 − 𝑢 ∈ 𝑈], and the third line above holds by
simple algebra. Taking square roots gives the desired
inequality.
The inequality proved above is an equality if and
only if 6.62 is an equality, which happens if and only if
‖𝑃𝑈 𝑣 − 𝑢‖ = 0, which happens if and only if 𝑢 = 𝑃𝑈 𝑣.
𝑃𝑈 𝑣 is the closest
The last result is often combined with the formula point in 𝑈 to 𝑣.
6.57(i) to compute explicit solutions to minimization
problems, as in the following example.
6.63 example: using linear algebra to approximate the sine function

Suppose we want to find a polynomial 𝑢 with real coefficients and of degree
at most 5 that approximates the sine function as well as possible on the interval
[−𝜋, 𝜋], in the sense that
𝜋
∣sin 𝑥 − 𝑢(𝑥)∣ 𝑑𝑥
2
∫
−𝜋
is as small as possible.
Let 𝐶[−𝜋, 𝜋] denote the real inner product space of continuous real-valued
functions on [−𝜋, 𝜋] with inner product
𝜋
6.64 ⟨ 𝑓, 𝑔⟩ = ∫ 𝑓 𝑔.
−𝜋
Let 𝑣 ∈ 𝐶[−𝜋, 𝜋] be the function defined by 𝑣(𝑥) = sin 𝑥. Let 𝑈 denote the
subspace of 𝐶[−𝜋, 𝜋] consisting of the polynomials with real coefficients and of
degree at most 5. Our problem can now be reformulated as follows:
Find 𝑢 ∈ 𝑈 such that ‖𝑣 − 𝑢‖ is as small as possible.
To compute the solution to our ap- A computer that can integrate is useful
proximation problem, first apply the here.
Gram–Schmidt procedure (using the in-
ner product given by 6.64) to the basis 1, 𝑥, 𝑥2, 𝑥3, 𝑥4, 𝑥5 of 𝑈, producing an ortho-
normal basis 𝑒1 , 𝑒2 , 𝑒3 , 𝑒4 , 𝑒5 , 𝑒6 of 𝑈. Then, again using the inner product given
by 6.64, compute 𝑃𝑈 𝑣 using 6.57(i) (with 𝑚 = 6). Doing this computation shows
that 𝑃𝑈 𝑣 is the function 𝑢 defined by
6.65 𝑢(𝑥) = 0.987862𝑥 − 0.155271𝑥3 + 0.00564312𝑥5,
where the 𝜋’s that appear in the exact answer have been replaced with a good
decimal approximation. By 6.61, the polynomial 𝑢 above is the best approximation
to the sine function on [−𝜋, 𝜋] using polynomials of degree at most 5 (here “best
approximation” means in the sense of minimizing ∫−𝜋 𝜋
| sin 𝑥 − 𝑢(𝑥)|2 𝑑𝑥).
To see how good this approximation is, the next figure shows the graphs of
both the sine function and our approximation 𝑢 given by 6.65 over the interval
[−𝜋, 𝜋].
Graphs on [−𝜋, 𝜋] of the sine function (red) and its best

fifth degree polynomial approximation 𝑢 (blue) from 6.65.
Our approximation 6.65 is so accurate that the two graphs are almost identical—
our eyes may see only one graph! Here the red graph is placed almost exactly
over the blue graph. If you are viewing this on an electronic device, enlarge the
picture above by 400% near 𝜋 or −𝜋 to see a small gap between the two graphs.
Another well-known approximation to the sine function by a polynomial of
degree 5 is given by the Taylor polynomial 𝑝 defined by
𝑥3 𝑥5
6.66 𝑝(𝑥) = 𝑥 − + .
3! 5!
To see how good this approximation is, the next picture shows the graphs of both
the sine function and the Taylor polynomial 𝑝 over the interval [−𝜋, 𝜋].
Graphs on [−𝜋, 𝜋] of the sine function (red)

and the Taylor polynomial (blue) from 6.66.
The Taylor polynomial is an excellent approximation to sin 𝑥 for 𝑥 near 0. But

the picture above shows that for |𝑥| > 2, the Taylor polynomial is not so accurate,
especially compared to 6.65. For example, taking 𝑥 = 3, our approximation 6.65
estimates sin 3 with an error of approximately 0.001, but the Taylor series 6.66
estimates sin 3 with an error of approximately 0.4. Thus at 𝑥 = 3, the error in
the Taylor series is hundreds of times larger than the error given by 6.65. Linear
algebra has helped us discover an approximation to the sine function that improves
upon what we learned in calculus!
Pseudoinverse
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑏 ∈ 𝑊. Consider the problem of finding 𝑥 ∈ 𝑉 such
that
𝑇𝑥 = 𝑏.
For example, if 𝑉 = 𝐅𝑛 and 𝑊 = 𝐅𝑚, then the equation above could represent a
system of 𝑚 linear equations in 𝑛 unknowns.
If 𝑇 is invertible, then the unique solution to the equation above is 𝑥 = 𝑇 −1 𝑏.
However, if 𝑇 is not invertible, then for some 𝑏 ∈ 𝑊 there may not exist any
solutions of the equation above, and for some 𝑏 ∈ 𝑊 there may exist infinitely
many solutions of the equation above.
If 𝑇 is not invertible, then we can still try to do as well as possible with the
equation above. For example, if the equation above has no solutions, then instead
of solving the equation 𝑇𝑥 − 𝑏 = 0, we can try to find 𝑥 ∈ 𝑉 such that ‖𝑇𝑥 − 𝑏‖
is as small as possible. As another example, if the equation above has infinitely
many solutions 𝑥 ∈ 𝑉, then among all those solutions we can try to find one such
that ‖𝑥‖ is as small as possible.
The pseudoinverse will provide the tool to solve the equation above as well
as possible, even when 𝑇 is not invertible. We need the next result to define the
pseudoinverse.
In the next two proofs, we will use without further comment the result that if
𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊), then null 𝑇, (null 𝑇)⟂, and range 𝑇 are
all finite-dimensional.
6.67 restriction of a linear map to obtain a one-to-one and onto map
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Then 𝑇|(null 𝑇)⟂ is a one-

to-one map of (null 𝑇)⟂ onto range 𝑇.
Proof Suppose that 𝑣 ∈ (null 𝑇)⟂ and 𝑇|(null 𝑇)⟂ 𝑣 = 0. Hence 𝑇𝑣 = 0 and
thus 𝑣 ∈ (null 𝑇) ∩ (null 𝑇)⟂, which implies that 𝑣 = 0 [by 6.48(d)]. Hence
null 𝑇|(null 𝑇)⟂ = {0}, which implies that 𝑇|(null 𝑇)⟂ is injective, as desired.
Clearly range 𝑇|(null 𝑇)⟂ ⊆ range 𝑇. To prove the inclusion in the other direction,
suppose 𝑤 ∈ range 𝑇. Hence there exists 𝑣 ∈ 𝑉 such that 𝑤 = 𝑇𝑣. There exist
𝑢 ∈ null 𝑇 and 𝑥 ∈ (null 𝑇)⟂ such that 𝑣 = 𝑢 + 𝑥 (by 6.49). Now
𝑇|(null 𝑇)⟂ 𝑥 = 𝑇𝑥 = 𝑇𝑣 − 𝑇𝑢 = 𝑤 − 0 = 𝑤,
which shows that 𝑤 ∈ range 𝑇|(null 𝑇)⟂ . Hence range 𝑇 ⊆ range 𝑇|(null 𝑇)⟂ , complet-
ing the proof that range 𝑇|(null 𝑇)⟂ = range 𝑇.
Now we can define the pseudoinverse To produce the pseudoinverse notation

𝑇† (pronounced “𝑇 dagger”) of a linear 𝑇 † in T X, type T ̂\dagger.
E
map 𝑇. In the next definition (and from
now on), think of 𝑇|(null 𝑇)⟂ as an invertible linear map from (null 𝑇)⟂ onto range 𝑇,
as is justified by the result above.
6.68 definition: pseudoinverse, 𝑇 †
Suppose that 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). The pseudoinverse

𝑇 † ∈ ℒ(𝑊, 𝑉) of 𝑇 is the linear map from 𝑊 to 𝑉 defined by
𝑇 † 𝑤 = (𝑇|(null 𝑇)⟂ )−1 𝑃range 𝑇 𝑤
for each 𝑤 ∈ 𝑊.
Recall that 𝑃range 𝑇 𝑤 = 0 if 𝑤 ∈ (range 𝑇)⟂ and 𝑃range 𝑇 𝑤 = 𝑤 if 𝑤 ∈ range 𝑇.

Thus if 𝑤 ∈ (range 𝑇)⟂, then 𝑇 † 𝑤 = 0, and if 𝑤 ∈ range 𝑇, then 𝑇 † 𝑤 is the
unique element of (null 𝑇)⟂ such that 𝑇(𝑇 † 𝑤) = 𝑤.
The pseudoinverse behaves much like an inverse, as we will see.
6.69 algebraic properties of the pseudoinverse
Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊).

(a) If 𝑇 is invertible, then 𝑇 † = 𝑇 −1.
(b) 𝑇𝑇 † = 𝑃range 𝑇 = the orthogonal projection of 𝑊 onto range 𝑇.
(c) 𝑇 † 𝑇 = 𝑃(null 𝑇)⟂ = the orthogonal projection of 𝑉 onto (null 𝑇)⟂.
Proof
(a) Suppose 𝑇 is invertible. Then (null 𝑇)⟂ = 𝑉 and range 𝑇 = 𝑊. Thus
𝑇|(null 𝑇)⟂ = 𝑇 and 𝑃range 𝑇 is the identity operator on 𝑊. Hence 𝑇 † = 𝑇 −1.
(b) Suppose 𝑤 ∈ range 𝑇. Thus
𝑇𝑇 † 𝑤 = 𝑇(𝑇|(null 𝑇)⟂ )−1 𝑤 = 𝑤 = 𝑃range 𝑇 𝑤.
If 𝑤 ∈ (range 𝑇)⟂, then 𝑇 † 𝑤 = 0. Hence 𝑇𝑇 † 𝑤 = 0 = 𝑃range 𝑇 𝑤. Thus 𝑇𝑇 †

and 𝑃range 𝑇 agree on range 𝑇 and on (range 𝑇)⟂. Hence these two linear maps
are equal (by 6.49).
(c) Suppose 𝑣 ∈ (null 𝑇)⟂. Because 𝑇𝑣 ∈ range 𝑇, the definition of 𝑇 † shows
that
𝑇 † (𝑇𝑣) = (𝑇|(null 𝑇)⟂ )−1 (𝑇𝑣) = 𝑣 = 𝑃(null 𝑇)⟂ 𝑣.
If 𝑣 ∈ null 𝑇, then 𝑇 † 𝑇𝑣 = 0 = 𝑃(null 𝑇)⟂ 𝑣. Thus 𝑇 † 𝑇 and 𝑃(null 𝑇)⟂ agree on

(null 𝑇)⟂ and on null 𝑇. Hence these two linear maps are equal (by 6.49).
Suppose that 𝑇 ∈ ℒ(𝑉, 𝑊). If 𝑇 is The pseudoinverse is also called the

surjective, then 𝑇𝑇 † is the identity opera- Moore–Penrose inverse.
tor on 𝑊, as follows from (b) in the result
above. If 𝑇 is injective, then 𝑇 † 𝑇 is the identity operator on 𝑉, as follows from (c)
in the result above. For additional algebraic properties of the pseudoinverse, see
Exercises 19–23.
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊), 𝑏 ∈ 𝑊, and we want to find 𝑥 ∈ 𝑉 that solves the

equation
𝑇𝑥 = 𝑏.
If 𝑇 is invertible, then 𝑥 = 𝑇 −1 𝑏 is the unique solution. If 𝑇 is not invertible, then
𝑇 −1 is not defined. However, the pseudoinverse 𝑇 † is defined. Taking 𝑥 = 𝑇 † 𝑏
makes 𝑇𝑥 as close to 𝑏 as possible, as shown by (a) of the next result. Thus the
pseudoinverse provides what is called a best fit to the equation above.
Among all vectors 𝑥 ∈ 𝑉 that make 𝑇𝑥 as close as possible to 𝑏, the vector
𝑇 † 𝑏 has the smallest norm, as shown by combining (b) in the next result with the
condition for equality in (a).
6.70 pseudoinverse provides best approximate solution or best solution
Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉, 𝑊), and 𝑏 ∈ 𝑊.

(a) If 𝑥 ∈ 𝑉, then
∥𝑇(𝑇 † 𝑏) − 𝑏∥ ≤ ‖𝑇𝑥 − 𝑏‖,
with equality if and only if 𝑥 ∈ 𝑇 † 𝑏 + null 𝑇.
(b) If 𝑥 ∈ 𝑇 † 𝑏 + null 𝑇, then
∥𝑇 † 𝑏∥ ≤ ‖𝑥‖,
with equality if and only if 𝑥 = 𝑇 † 𝑏.
Proof
(a) Suppose 𝑥 ∈ 𝑉. Then

𝑇𝑥 − 𝑏 = (𝑇𝑥 − 𝑇𝑇 † 𝑏) + (𝑇𝑇 † 𝑏 − 𝑏).
The first term in parentheses above is in range 𝑇. Because the operator 𝑇𝑇 †

is the orthogonal projection of 𝑊 onto range 𝑇 [by 6.69(b)], the second term
in parentheses above is in (range 𝑇)⟂ [see 6.57(f)].
Thus the Pythagorean theorem implies the desired inequality that the norm of
the second term in parentheses above is less than or equal to ‖𝑇𝑥 − 𝑏‖, with
equality if and only if the first term in parentheses above equals 0. Hence
we have equality if and only if 𝑥 − 𝑇 † 𝑏 ∈ null 𝑇, which is equivalent to the
statement that 𝑥 ∈ 𝑇 † 𝑏 + null 𝑇, completing the proof of (a).
(b) Suppose 𝑥 ∈ 𝑇 † 𝑏 + null 𝑇. Hence 𝑥 − 𝑇 † 𝑏 ∈ null 𝑇. Now
𝑥 = (𝑥 − 𝑇 † 𝑏) + 𝑇 † 𝑏.
The definition of 𝑇 † implies that 𝑇 † 𝑏 ∈ (null 𝑇)⟂. Thus the Pythagorean

theorem implies that ∥𝑇 † 𝑏∥ ≤ ‖𝑥‖, with equality if and only if 𝑥 = 𝑇 † 𝑏.
A formula for 𝑇 † will be given in the next chapter (see 7.78).

6.71 example: pseudoinverse of a linear map from 𝐅4 to 𝐅3

Suppose 𝑇 ∈ ℒ(𝐅4 , 𝐅3 ) is defined by
𝑇(𝑎, 𝑏, 𝑐, 𝑑) = (𝑎 + 𝑏 + 𝑐, 2𝑐 + 𝑑, 0).
This linear map is neither injective nor surjective, but we can compute its pseudo-
inverse. To do this, first note that range 𝑇 = {(𝑥, 𝑦, 0) ∶ 𝑥, 𝑦 ∈ 𝐅}. Thus
𝑃range 𝑇 (𝑥, 𝑦, 𝑧) = (𝑥, 𝑦, 0)
for each (𝑥, 𝑦, 𝑧) ∈ 𝐅3. Also,

null 𝑇 = {(𝑎, 𝑏, 𝑐, 𝑑) ∈ 𝐅4 ∶ 𝑎 + 𝑏 + 𝑐 = 0 and 2𝑐 + 𝑑 = 0}.
The list (−1, 1, 0, 0), (−1, 0, 1, −2) of two vectors in null 𝑇 spans null 𝑇 because
if (𝑎, 𝑏, 𝑐, 𝑑) ∈ null 𝑇 then
(𝑎, 𝑏, 𝑐, 𝑑) = 𝑏(−1, 1, 0, 0) + 𝑐(−1, 0, 1, −2).
Because the list (−1, 1, 0, 0), (−1, 0, 1, −2) is linearly independent, this list is a
basis of null 𝑇.
Now suppose (𝑥, 𝑦, 𝑧) ∈ 𝐅3. Then
6.72 𝑇 † (𝑥, 𝑦, 𝑧) = (𝑇|(null 𝑇)⟂ )−1 𝑃range 𝑇 (𝑥, 𝑦, 𝑧) = (𝑇|(null 𝑇)⟂ )−1 (𝑥, 𝑦, 0).
The right side of the equation above is the vector (𝑎, 𝑏, 𝑐, 𝑑) ∈ 𝐅4 such that
𝑇(𝑎, 𝑏, 𝑐, 𝑑) = (𝑥, 𝑦, 0) and (𝑎, 𝑏, 𝑐, 𝑑) ∈ (null 𝑇)⟂. In other words, 𝑎, 𝑏, 𝑐, 𝑑 must
satisfy the following equations:
𝑎+𝑏+𝑐=𝑥
2𝑐 + 𝑑 = 𝑦
−𝑎 + 𝑏 = 0
−𝑎 + 𝑐 − 2𝑑 = 0,
where the first two equations are equivalent to the equation 𝑇(𝑎, 𝑏, 𝑐, 𝑑) = (𝑥, 𝑦, 0)
and the last two equations come from the condition for (𝑎, 𝑏, 𝑐, 𝑑) to be orthogo-
nal to each of the basis vectors (−1, 1, 0, 0), (−1, 0, 1, −2) in this basis of null 𝑇.
Thinking of 𝑥 and 𝑦 as constants and 𝑎, 𝑏, 𝑐, 𝑑 as unknowns, we can solve the
system above of four equations in four unknowns, getting
1 1 1 1
𝑎= 11
(5𝑥 − 2𝑦), 𝑏 = 11
(5𝑥 − 2𝑦), 𝑐 = 11
(𝑥 + 4𝑦), 𝑑 = 11
(−2𝑥 + 3𝑦).
Hence 6.72 tells us that
1
𝑇 † (𝑥, 𝑦, 𝑧) = 11
(5𝑥 − 2𝑦, 5𝑥 − 2𝑦, 𝑥 + 4𝑦, −2𝑥 + 3𝑦).
The formula above for 𝑇 † shows that 𝑇𝑇 † (𝑥, 𝑦, 𝑧) = (𝑥, 𝑦, 0) for all (𝑥, 𝑦, 𝑧) ∈ 𝐅3,
which illustrates the equation 𝑇𝑇 † = 𝑃range 𝑇 from 6.69(b).
Exercises 6C
1 Suppose 𝑣1 , …, 𝑣𝑚 ∈ 𝑉. Prove that

{𝑣1 , …, 𝑣𝑚 }⟂ = (span(𝑣1 , …, 𝑣𝑚 )) .
⟂
2 Suppose 𝑈 is a subspace of 𝑉 with basis 𝑢1 , …, 𝑢𝑚 and

𝑢1 , …, 𝑢𝑚 , 𝑣1 , …, 𝑣𝑛
is a basis of 𝑉. Prove that if the Gram–Schmidt procedure is applied to the

basis of 𝑉 above, producing a list 𝑒1 , …, 𝑒𝑚 , 𝑓1 , …, 𝑓𝑛 , then 𝑒1 , …, 𝑒𝑚 is an
orthonormal basis of 𝑈 and 𝑓1 , …, 𝑓𝑛 is an orthonormal basis of 𝑈 ⟂.
3 Suppose 𝑈 is the subspace of 𝐑4 defined by
𝑈 = span((1, 2, 3, −4), (−5, 4, 3, 2)).
Find an orthonormal basis of 𝑈 and an orthonormal basis of 𝑈 ⟂.

4 Suppose 𝑒1 , …, 𝑒𝑛 is a list of vectors in 𝑉 with ‖𝑒𝑘 ‖ = 1 for each 𝑘 = 1, …, 𝑛
and
2 2
‖𝑣‖2 = ∣⟨𝑣, 𝑒1 ⟩∣ + ⋯ + ∣⟨𝑣, 𝑒𝑛 ⟩∣
for all 𝑣 ∈ 𝑉. Prove that 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉.
This exercise provides a converse to 6.30(b).
5 Suppose that 𝑉 is finite-dimensional and 𝑈 is a subspace of 𝑉. Show that

𝑃𝑈⟂ = 𝐼 − 𝑃𝑈 , where 𝐼 is the identity operator on 𝑉.
6 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Show that

𝑇 = 𝑇𝑃(null 𝑇)⟂ = 𝑃range 𝑇 𝑇.
7 Suppose that 𝑋 and 𝑌 are finite-dimensional subspaces of 𝑉. Prove that

𝑃𝑋 𝑃𝑌 = 0 if and only if ⟨𝑥, 𝑦⟩ = 0 for all 𝑥 ∈ 𝑋 and all 𝑦 ∈ 𝑌.
8 Suppose 𝑈 is a finite-dimensional subspace of 𝑉 and 𝑣 ∈ 𝑉. Define a linear

functional 𝜑 ∶ 𝑈 → 𝐅 by
𝜑(𝑢) = ⟨𝑢, 𝑣⟩
for all 𝑢 ∈ 𝑈. By the Riesz representation theorem (6.42) as applied to the
inner product space 𝑈, there exists a unique vector 𝑤 ∈ 𝑈 such that
𝜑(𝑢) = ⟨𝑢, 𝑤⟩
for all 𝑢 ∈ 𝑈. Show that 𝑤 = 𝑃𝑈 𝑣.

9 Suppose 𝑉 is finite-dimensional. Suppose 𝑃 ∈ ℒ(𝑉) is such that 𝑃2 = 𝑃
and every vector in null 𝑃 is orthogonal to every vector in range 𝑃. Prove
that there exists a subspace 𝑈 of 𝑉 such that 𝑃 = 𝑃𝑈 .
10 Suppose 𝑉 is finite-dimensional and 𝑃 ∈ ℒ(𝑉) is such that 𝑃2 = 𝑃 and
‖𝑃𝑣‖ ≤ ‖𝑣‖
for every 𝑣 ∈ 𝑉. Prove that there exists a subspace 𝑈 of 𝑉 such that 𝑃 = 𝑃𝑈 .

11 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑈 is a finite-dimensional subspace of 𝑉. Prove that
𝑈 is invariant under 𝑇 ⟺ 𝑃𝑈 𝑇𝑃𝑈 = 𝑇𝑃𝑈 .
12 Suppose 𝑉 is finite-dimensional, 𝑇 ∈ ℒ(𝑉), and 𝑈 is a subspace of 𝑉. Prove

that
𝑈 and 𝑈 ⟂ are both invariant under 𝑇 ⟺ 𝑃𝑈 𝑇 = 𝑇𝑃𝑈 .
13 Suppose 𝐅 = 𝐑 and 𝑉 is finite-dimensional. For each 𝑣 ∈ 𝑉, let 𝜑𝑣 denote
the linear functional on 𝑉 defined by
𝜑𝑣 (𝑢) = ⟨𝑢, 𝑣⟩
for all 𝑢 ∈ 𝑉.
(a) Show that 𝑣 ↦ 𝜑𝑣 is an injective linear map from 𝑉 to 𝑉 ′.
(b) Use (a) and a dimension-counting argument to show that 𝑣 ↦ 𝜑𝑣 is an
isomorphism from 𝑉 onto 𝑉 ′.
The purpose of this exercise is to give an alternative proof of the Riesz
representation theorem (6.42 and 6.58) when 𝐅 = 𝐑 . Thus you should not
use the Riesz representation theorem as a tool in your solution.
14 Suppose that 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉. Explain why the dual

basis (see 3.112) of 𝑒1 , …, 𝑒𝑛 is 𝑒1 , …, 𝑒𝑛 under the identification of 𝑉 ′ with
𝑉 provided by the Riesz representation theorem (6.58).
15 In 𝐑4 , let
𝑈 = span((1, 1, 0, 0), (1, 1, 1, 2)).
Find 𝑢 ∈ 𝑈 such that ‖𝑢 − (1, 2, 3, 4)‖ is as small as possible.
16 Suppose 𝐶[−1, 1] is the vector space of continuous real-valued functions
on the interval [−1, 1] with inner product given by
1
⟨ 𝑓, 𝑔⟩ = ∫ 𝑓𝑔
−1
for all 𝑓, 𝑔 ∈ 𝐶[−1, 1]. Let 𝑈 be the subspace of 𝐶[−1, 1] defined by
𝑈 = { 𝑓 ∈ 𝐶[−1, 1] ∶ 𝑓 (0) = 0}.
(a) Show that 𝑈 ⟂ = {0}.

(b) Show that 6.49 and 6.52 do not hold without the finite-dimensional
hypothesis.
1
17 Find 𝑝 ∈ 𝒫3 (𝐑) such that 𝑝(0) = 0, 𝑝′(0) = 0, and ∫ ∣2 + 3𝑥 − 𝑝(𝑥)∣2 𝑑𝑥 is
0
as small as possible.
𝜋
18 Find 𝑝 ∈ 𝒫5 (𝐑) that makes ∫ ∣sin 𝑥 − 𝑝(𝑥)∣ 𝑑𝑥 as small as possible.
2
−𝜋
The polynomial 6.65 is an excellent approximation to the answer to this
exercise, but here you are asked to find the exact solution, which involves
powers of 𝜋. A computer that can perform symbolic integration should
help.
19 Suppose 𝑉 is finite-dimensional and 𝑃 ∈ ℒ(𝑉) is an orthogonal projection

of 𝑉 onto some subspace of 𝑉. Prove that 𝑃† = 𝑃.
20 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Show that
null 𝑇 † = (range 𝑇)⟂ and range 𝑇 † = (null 𝑇)⟂.
21 Suppose 𝑇 ∈ ℒ(𝐅3 , 𝐅2 ) is defined by
𝑇(𝑎, 𝑏, 𝑐) = (𝑎 + 𝑏 + 𝑐, 2𝑏 + 3𝑐).
(a) For (𝑥, 𝑦) ∈ 𝐅2, find a formula for 𝑇 † (𝑥, 𝑦).

(b) Verify that the equation 𝑇𝑇 † = 𝑃range 𝑇 from 6.69(b) holds with the
formula for 𝑇 † obtained in (a).
(c) Verify that the equation 𝑇 † 𝑇 = 𝑃(null 𝑇)⟂ from 6.69(c) holds with the
formula for 𝑇 † obtained in (a).
22 Suppose 𝑉 is finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that
𝑇𝑇 † 𝑇 = 𝑇 and 𝑇 † 𝑇𝑇 † = 𝑇 †.
Both formulas above clearly hold if 𝑇 is invertible because in that case we
can replace 𝑇 † with 𝑇 −1.
23 Suppose 𝑉 and 𝑊 are finite-dimensional and 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that

†
(𝑇 † ) = 𝑇.
The equation above is analogous to the equation (𝑇 −1 )−1 = 𝑇 that holds if
𝑇 is invertible.
Chapter 7
Operators on Inner Product Spaces
The deepest results related to inner product spaces deal with the subject to which
we now turn—linear maps and operators on inner product spaces. As we will see,
good theorems can be proved by exploiting properties of the adjoint.
The hugely important spectral theorem will provide a complete description
of self-adjoint operators on real inner product spaces and of normal operators
on complex inner product spaces. We will then use the spectral theorem to help
understand positive operators and unitary operators, which will lead to unitary
matrices and matrix factorizations. The spectral theorem will also lead to the
popular singular value decomposition, which will lead to the polar decomposition.
The most important results in the rest of this book are valid only in finite
dimensions. Thus from now on we assume that 𝑉 and 𝑊 are finite-dimensional.
• 𝑉 and 𝑊 are nonzero finite-dimensional inner product spaces over 𝐅.
Petar Milošević CC BY-SA
Market square in Lviv, a city that has had several names and has been in several
countries because of changing international borders. From 1772 until 1918, the city was
in Austria and was called Lemberg. Between World War I and World War II, the city was
in Poland and was called Lwów. During this time, mathematicians in Lwów, particularly
Stefan Banach (1892–1945) and his colleagues, developed the basic results of modern
functional analysis, using tools of analysis to study infinite-dimensional vector spaces.
Since the end of World War II, Lviv has been in Ukraine, which was part of the
Soviet Union until Ukraine became an independent country in 1991.
227
https://doi.org/10.1007/978-3-031-41026-0_7
228 Chapter 7 Operators on Inner Product Spaces
7A Self-Adjoint and Normal Operators

Adjoints
7.1 definition: adjoint, 𝑇 ∗
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). The adjoint of 𝑇 is the function 𝑇 ∗ ∶ 𝑊 → 𝑉 such

that
⟨𝑇𝑣, 𝑤⟩ = ⟨𝑣, 𝑇 ∗ 𝑤⟩
for every 𝑣 ∈ 𝑉 and every 𝑤 ∈ 𝑊.
To see why the definition above The word adjoint has another meaning
makes sense, suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Fix in linear algebra. In case you en-
𝑤 ∈ 𝑊. Consider the linear functional counter the second meaning elsewhere,
be warned that the two meanings for
𝑣 ↦ ⟨𝑇𝑣, 𝑤⟩
adjoint are unrelated to each other.
on 𝑉 that maps 𝑣 ∈ 𝑉 to ⟨𝑇𝑣, 𝑤⟩; this
linear functional depends on 𝑇 and 𝑤. By the Riesz representation theorem (6.42),
there exists a unique vector in 𝑉 such that this linear functional is given by taking
the inner product with it. We call this unique vector 𝑇 ∗ 𝑤. In other words, 𝑇 ∗ 𝑤 is
the unique vector in 𝑉 such that
⟨𝑇𝑣, 𝑤⟩ = ⟨𝑣, 𝑇 ∗ 𝑤⟩
In the equation above, the inner product on the left takes place in 𝑊 and the
inner product on the right takes place in 𝑉. However, we use the same notation
⟨⋅, ⋅⟩ for both inner products.
7.2 example: adjoint of a linear map from 𝐑3 to 𝐑2

Define 𝑇 ∶ 𝐑3 → 𝐑2 by
𝑇(𝑥1 , 𝑥2 , 𝑥3 ) = (𝑥2 + 3𝑥3 , 2𝑥1 ).
To compute 𝑇 ∗, suppose (𝑥1 , 𝑥2 , 𝑥3 ) ∈ 𝐑3 and (𝑦1 , 𝑦2 ) ∈ 𝐑2. Then
⟨𝑇(𝑥1 , 𝑥2 , 𝑥3 ), (𝑦1 , 𝑦2 )⟩ = ⟨(𝑥2 + 3𝑥3 , 2𝑥1 ), (𝑦1 , 𝑦2 )⟩

= 𝑥2 𝑦1 + 3𝑥3 𝑦1 + 2𝑥1 𝑦2
= ⟨(𝑥1 , 𝑥2 , 𝑥3 ), (2𝑦2 , 𝑦1 , 3𝑦1 )⟩.
The equation above and the definition of the adjoint imply that
𝑇 ∗ (𝑦1 , 𝑦2 ) = (2𝑦2 , 𝑦1 , 3𝑦1 ).

Section 7A Self-Adjoint and Normal Operators 229
7.3 example: adjoint of a linear map with range of dimension at most 1

Fix 𝑢 ∈ 𝑉 and 𝑥 ∈ 𝑊. Define 𝑇 ∈ ℒ(𝑉, 𝑊) by
𝑇𝑣 = ⟨𝑣, 𝑢⟩𝑥
for each 𝑣 ∈ 𝑉. To compute 𝑇 ∗, suppose 𝑣 ∈ 𝑉 and 𝑤 ∈ 𝑊. Then
⟨𝑇𝑣, 𝑤⟩ = ⟨⟨𝑣, 𝑢⟩𝑥, 𝑤⟩
= ⟨𝑣, 𝑢⟩⟨𝑥, 𝑤⟩
= ⟨𝑣, ⟨𝑤, 𝑥⟩𝑢⟩.
Thus
𝑇 ∗ 𝑤 = ⟨𝑤, 𝑥⟩𝑢.
In the two examples above, 𝑇 ∗ turned The two examples above and the proof
out to be not just a function from 𝑉 to below use a common technique for
𝑊 but a linear map from 𝑉 to 𝑊. This computing 𝑇 ∗ : start with a formula
behavior is true in general, as shown by for ⟨𝑇𝑣, 𝑤⟩ then manipulate it to get
the next result. just 𝑣 in the first slot; the entry in the
second slot will then be 𝑇 ∗ 𝑤.
7.4 adjoint of a linear map is a linear map
If 𝑇 ∈ ℒ(𝑉, 𝑊), then 𝑇 ∗ ∈ ℒ(𝑊, 𝑉).
Proof Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). If 𝑣 ∈ 𝑉 and 𝑤1 , 𝑤2 ∈ 𝑊, then

⟨𝑇𝑣, 𝑤1 + 𝑤2 ⟩ = ⟨𝑇𝑣, 𝑤1 ⟩ + ⟨𝑇𝑣, 𝑤2 ⟩
= ⟨𝑣, 𝑇 ∗ 𝑤1 ⟩ + ⟨𝑣, 𝑇 ∗ 𝑤2 ⟩
= ⟨𝑣, 𝑇 ∗ 𝑤1 + 𝑇 ∗ 𝑤2 ⟩.
𝑇 ∗ (𝑤1 + 𝑤2 ) = 𝑇 ∗ 𝑤1 + 𝑇 ∗ 𝑤2 .
If 𝑣 ∈ 𝑉, 𝜆 ∈ 𝐅, and 𝑤 ∈ 𝑊, then
⟨𝑇𝑣, 𝜆𝑤⟩ = 𝜆⟨𝑇𝑣, 𝑤⟩
= 𝜆⟨𝑣, 𝑇 ∗ 𝑤⟩
= ⟨𝑣, 𝜆𝑇 ∗ 𝑤⟩.
𝑇 ∗ (𝜆𝑤) = 𝜆𝑇 ∗ 𝑤.
Thus 𝑇 ∗ is a linear map, as desired.
7.5 properties of the adjoint

(a) (𝑆 + 𝑇)∗ = 𝑆∗ + 𝑇 ∗ for all 𝑆 ∈ ℒ(𝑉, 𝑊);
(b) (𝜆𝑇)∗ = 𝜆𝑇 ∗ for all 𝜆 ∈ 𝐅 ;
(c) (𝑇 ∗ )∗ = 𝑇;
(d) (𝑆𝑇)∗ = 𝑇 ∗ 𝑆∗ for all 𝑆 ∈ ℒ(𝑊, 𝑈) (here 𝑈 is a finite-dimensional inner
product space over 𝐅 );
(e) 𝐼 ∗ = 𝐼, where 𝐼 is the identity operator on 𝑉;
(f) if 𝑇 is invertible, then 𝑇 ∗ is invertible and (𝑇 ∗ )−1 = (𝑇 −1 )∗.
Proof Suppose 𝑣 ∈ 𝑉 and 𝑤 ∈ 𝑊.

(a) If 𝑆 ∈ ℒ(𝑉, 𝑊), then
⟨(𝑆 + 𝑇)𝑣, 𝑤⟩ = ⟨𝑆𝑣, 𝑤⟩ + ⟨𝑇𝑣, 𝑤⟩
= ⟨𝑣, 𝑆∗ 𝑤⟩ + ⟨𝑣, 𝑇 ∗ 𝑤⟩
= ⟨𝑣, 𝑆∗ 𝑤 + 𝑇 ∗ 𝑤⟩.
Thus (𝑆 + 𝑇)∗ 𝑤 = 𝑆∗ 𝑤 + 𝑇 ∗ 𝑤, as desired.
(b) If 𝜆 ∈ 𝐅, then
⟨(𝜆𝑇)𝑣, 𝑤⟩ = 𝜆⟨𝑇𝑣, 𝑤⟩ = 𝜆⟨𝑣, 𝑇 ∗ 𝑤⟩ = ⟨𝑣, 𝜆𝑇 ∗ 𝑤⟩.
Thus (𝜆𝑇)∗ 𝑤 = 𝜆𝑇 ∗ 𝑤, as desired.
(c) We have
⟨𝑇 ∗ 𝑤, 𝑣⟩ = ⟨𝑣, 𝑇 ∗ 𝑤⟩ = ⟨𝑇𝑣, 𝑤⟩ = ⟨𝑤, 𝑇𝑣⟩.
Thus (𝑇 ∗ )∗ 𝑣 = 𝑇𝑣, as desired.
(d) Suppose 𝑆 ∈ ℒ(𝑊, 𝑈) and 𝑢 ∈ 𝑈. Then
⟨(𝑆𝑇)𝑣, 𝑢⟩ = ⟨𝑆(𝑇𝑣), 𝑢⟩ = ⟨𝑇𝑣, 𝑆∗ 𝑢⟩ = ⟨𝑣, 𝑇 ∗ (𝑆∗ 𝑢)⟩.
Thus (𝑆𝑇)∗ 𝑢 = 𝑇 ∗ (𝑆∗ 𝑢), as desired.
(e) Suppose 𝑢 ∈ 𝑉. Then
⟨𝐼𝑢, 𝑣⟩ = ⟨𝑢, 𝑣⟩.
Thus 𝐼 ∗ 𝑣 = 𝑣, as desired.
(f) Suppose 𝑇 is invertible. Take adjoints of both sides of the equation 𝑇 −1 𝑇 = 𝐼,
then use (d) and (e) to show that 𝑇 ∗ (𝑇 −1 )∗ = 𝐼. Similarly, the equation
∗ ∗
𝑇𝑇 −1 = 𝐼 implies (𝑇 −1 ) 𝑇 ∗ = 𝐼. Thus (𝑇 −1 ) is the inverse of 𝑇 ∗, as
desired.
If 𝐅 = 𝐑 , then the map 𝑇 ↦ 𝑇 ∗ is a linear map from ℒ(𝑉, 𝑊) to ℒ(𝑊, 𝑉),

as follows from (a) and (b) of the result above. However, if 𝐅 = 𝐂 , then this map
is not linear because of the complex conjugate that appears in (b).
The next result shows the relationship between the null space and the range of
a linear map and its adjoint.
7.6 null space and range of 𝑇 ∗

(a) null 𝑇 ∗ = (range 𝑇)⟂ ;
(b) range 𝑇 ∗ = (null 𝑇)⟂ ;
(c) null 𝑇 = (range 𝑇 ∗ )⟂ ;
(d) range 𝑇 = (null 𝑇 ∗ )⟂ .
Proof We begin by proving (a). Let 𝑤 ∈ 𝑊. Then

𝑤 ∈ null 𝑇 ∗ ⟺ 𝑇 ∗ 𝑤 = 0
⟺ ⟨𝑣, 𝑇 ∗ 𝑤⟩ = 0 for all 𝑣 ∈ 𝑉
⟺ ⟨𝑇𝑣, 𝑤⟩ = 0 for all 𝑣 ∈ 𝑉
⟺ 𝑤 ∈ (range 𝑇)⟂ .
Thus null 𝑇 ∗ = (range 𝑇)⟂ , proving (a).
If we take the orthogonal complement of both sides of (a), we get (d), where
we have used 6.52. Replacing 𝑇 with 𝑇 ∗ in (a) gives (c), where we have used
7.5(c). Finally, replacing 𝑇 with 𝑇 ∗ in (d) gives (b).
As we will soon see, the next definition is intimately connected to the matrix
of the adjoint of a linear map.
7.7 definition: conjugate transpose, 𝐴∗
The conjugate transpose of an 𝑚-by-𝑛 matrix 𝐴 is the 𝑛-by-𝑚 matrix 𝐴∗

obtained by interchanging the rows and columns and then taking the complex
conjugate of each entry. In other words, if 𝑗 ∈ {1, …, 𝑛} and 𝑘 ∈ {1, …, 𝑚},
then
(𝐴∗ )𝑗, 𝑘 = 𝐴𝑘, 𝑗 .
7.8 example: conjugate transpose of a 2-by-3 matrix

The conjugate transpose of the 2-by-3 If a matrix 𝐴 has only real entries,
2 3 + 4𝑖 7 then 𝐴∗ = 𝐴t, where 𝐴t denotes the
matrix ( ) is the 3-by-2
6 5 8𝑖 transpose of 𝐴 (the matrix obtained
matrix by interchanging the rows and the
⎛
2 6
⎞ columns).
⎜
⎜ 3 − 4𝑖 5 ⎟⎟
⎜ ⎟.
⎝ 7 −8𝑖 ⎠
The next result shows how to compute The adjoint of a linear map does not
the matrix of 𝑇 ∗ from the matrix of 𝑇. depend on a choice of basis. Thus
Caution: With respect to nonorthonor- we frequently emphasize adjoints of
mal bases, the matrix of 𝑇 ∗ does not nec- linear maps instead of transposes or
essarily equal the conjugate transpose of conjugate transposes of matrices.
the matrix of 𝑇.
7.9 matrix of 𝑇 ∗ equals conjugate transpose of matrix of 𝑇
Let 𝑇 ∈ ℒ(𝑉, 𝑊). Suppose 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉 and

𝑓1 , …, 𝑓𝑚 is an orthonormal basis of 𝑊. Then ℳ(𝑇 ∗, ( 𝑓1 , …, 𝑓𝑚 ), (𝑒1 , …, 𝑒𝑛 ))
is the conjugate transpose of ℳ(𝑇, (𝑒1 , …, 𝑒𝑛 ), ( 𝑓1 , …, 𝑓𝑚 )). In other words,
∗
ℳ(𝑇 ∗ ) = (ℳ(𝑇)) .
Proof In this proof, we will write ℳ(𝑇) and ℳ(𝑇 ∗ ) instead of the longer
expressions ℳ(𝑇, (𝑒1 , …, 𝑒𝑛 ), ( 𝑓1 , …, 𝑓𝑚 )) and ℳ(𝑇 ∗, ( 𝑓1 , …, 𝑓𝑚 ), (𝑒1 , …, 𝑒𝑛 )).
Recall that we obtain the 𝑘 th column of ℳ(𝑇) by writing 𝑇𝑒𝑘 as a linear
combination of the 𝑓𝑗 ’s; the scalars used in this linear combination then become
the 𝑘 th column of ℳ(𝑇). Because 𝑓1 , …, 𝑓𝑚 is an orthonormal basis of 𝑊, we
know how to write 𝑇𝑒𝑘 as a linear combination of the 𝑓𝑗 ’s [see 6.30(a)]:
𝑇𝑒𝑘 = ⟨𝑇𝑒𝑘 , 𝑓1 ⟩ 𝑓1 + ⋯ + ⟨𝑇𝑒𝑘 , 𝑓𝑚 ⟩ 𝑓𝑚 .
Thus
the entry in row 𝑗, column 𝑘, of ℳ(𝑇) is ⟨𝑇𝑒𝑘 , 𝑓𝑗 ⟩.
In the statement above, replace 𝑇 with 𝑇 ∗ and interchange 𝑒1 , …, 𝑒𝑛 and
𝑓1 , …, 𝑓𝑚 . This shows that the entry in row 𝑗, column 𝑘, of ℳ(𝑇 ∗ ) is ⟨𝑇 ∗ 𝑓𝑘 , 𝑒𝑗 ⟩,
which equals ⟨ 𝑓𝑘 , 𝑇𝑒𝑗 ⟩, which equals ⟨𝑇𝑒𝑗 , 𝑓𝑘 ⟩, which equals the complex conjugate
of the entry in row 𝑘, column 𝑗, of ℳ(𝑇). Thus ℳ(𝑇 ∗ ) = (ℳ(𝑇))∗.
The Riesz representation theorem as stated in 6.58 provides an identification of

𝑉 with its dual space 𝑉 ′ defined in 3.110. Under this identification, the orthogonal
complement 𝑈 ⟂ of a subset 𝑈 ⊆ 𝑉 corresponds to the annihilator 𝑈 0 of 𝑈. If 𝑈
is a subspace of 𝑉, then the formulas for the dimensions of 𝑈 ⟂ and 𝑈 0 become
identical under this identification—see 3.125 and 6.51.
Suppose 𝑇 ∶ 𝑉 → 𝑊 is a linear map. Because orthogonal complements and
Under the identification of 𝑉 with 𝑉 ′ and adjoints are easier to deal with than
the identification of 𝑊 with 𝑊 ′, the ad- annihilators and dual maps, there is
joint map 𝑇 ∗ ∶ 𝑊 → 𝑉 corresponds to no need to work with annihilators
the dual map 𝑇 ′ ∶ 𝑊 ′ → 𝑉 ′ defined in and dual maps in the context of inner
3.118, as Exercise 32 asks you to verify. product spaces.
Under this identification, the formulas for
null 𝑇 ∗ and range 𝑇 ∗ [7.6(a) and (b)] then become identical to the formulas for
null 𝑇 ′ and range 𝑇 ′ [3.128(a) and 3.130(b)]. Furthermore, the theorem about the
matrix of 𝑇 ∗ (7.9) is analogous to the theorem about the matrix of 𝑇 ′ (3.132).
Self-Adjoint Operators
Now we switch our attention to operators on inner product spaces. Instead of
considering linear maps from 𝑉 to 𝑊, we will focus on linear maps from 𝑉 to 𝑉;
recall that such linear maps are called operators.
7.10 definition: self-adjoint
An operator 𝑇 ∈ ℒ(𝑉) is called self-adjoint if 𝑇 = 𝑇 ∗.
If 𝑇 ∈ ℒ(𝑉) and 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉, then 𝑇 is self-adjoint

if and only if ℳ(𝑇, (𝑒1 , …, 𝑒𝑛 )) = ℳ(𝑇, (𝑒1 , …, 𝑒𝑛 ))∗, as follows from 7.9.
7.11 example: determining whether 𝑇 is self-adjoint from its matrix

Suppose 𝑐 ∈ 𝐅 and 𝑇 is the operator on 𝐅2 whose matrix (with respect to the
standard basis) is
2 𝑐
ℳ(𝑇) = ( ).
3 7
The matrix of 𝑇 ∗ (with respect to the standard basis) is
2 3
ℳ(𝑇 ∗ ) = ( ).
𝑐 7
Thus ℳ(𝑇) = ℳ(𝑇 ∗ ) if and only if 𝑐 = 3. Hence the operator 𝑇 is self-adjoint
if and only if 𝑐 = 3.
A good analogy to keep in mind is that the adjoint on ℒ(𝑉) plays a role similar
to that of the complex conjugate on 𝐂 . A complex number 𝑧 is real if and only if
𝑧 = 𝑧; thus a self-adjoint operator (𝑇 = 𝑇 ∗ ) is analogous to a real number.
We will see that the analogy discussed An operator 𝑇 ∈ ℒ(𝑉) is self-adjoint
above is reflected in some important prop- if and only if
erties of self-adjoint operators, beginning
with eigenvalues in the next result. ⟨𝑇𝑣, 𝑤⟩ = ⟨𝑣, 𝑇𝑤⟩
If 𝐅 = 𝐑 , then by definition every for all 𝑣, 𝑤 ∈ 𝑉.
eigenvalue is real, so the next result is
interesting only when 𝐅 = 𝐂 .
7.12 eigenvalues of self-adjoint operators
Every eigenvalue of a self-adjoint operator is real.
Proof Suppose 𝑇 is a self-adjoint operator on 𝑉. Let 𝜆 be an eigenvalue of 𝑇,

and let 𝑣 be a nonzero vector in 𝑉 such that 𝑇𝑣 = 𝜆𝑣. Then
𝜆‖𝑣‖2 = ⟨𝜆𝑣, 𝑣⟩ = ⟨𝑇𝑣, 𝑣⟩ = ⟨𝑣, 𝑇𝑣⟩ = ⟨𝑣, 𝜆𝑣⟩ = 𝜆‖𝑣‖2.
Thus 𝜆 = 𝜆, which means that 𝜆 is real, as desired.
The next result is false for real inner product spaces. As an example, consider
the operator 𝑇 ∈ ℒ(𝐑2 ) that is a counterclockwise rotation of 90∘ around the
origin; thus 𝑇(𝑥, 𝑦) = (−𝑦, 𝑥). Notice that 𝑇𝑣 is orthogonal to 𝑣 for every 𝑣 ∈ 𝐑2,
even though 𝑇 ≠ 0.
7.13 𝑇𝑣 is orthogonal to 𝑣 for all 𝑣 ⟺ 𝑇 = 0 (assuming 𝐅 = 𝐂 )
Suppose 𝑉 is a complex inner product space and 𝑇 ∈ ℒ(𝑉). Then
⟨𝑇𝑣, 𝑣⟩ = 0 for every 𝑣 ∈ 𝑉 ⟺ 𝑇 = 0.
Proof If 𝑢, 𝑤 ∈ 𝑉, then
⟨𝑇(𝑢 + 𝑤), 𝑢 + 𝑤⟩ − ⟨𝑇(𝑢 − 𝑤), 𝑢 − 𝑤⟩
⟨𝑇𝑢, 𝑤⟩ =
4
⟨𝑇(𝑢 + 𝑖𝑤), 𝑢 + 𝑖𝑤⟩ − ⟨𝑇(𝑢 − 𝑖𝑤), 𝑢 − 𝑖𝑤⟩
+ 𝑖,
4
as can be verified by computing the right side. Note that each term on the right
side is of the form ⟨𝑇𝑣, 𝑣⟩ for appropriate 𝑣 ∈ 𝑉.
Now suppose ⟨𝑇𝑣, 𝑣⟩ = 0 for every 𝑣 ∈ 𝑉. Then the equation above implies
that ⟨𝑇𝑢, 𝑤⟩ = 0 for all 𝑢, 𝑤 ∈ 𝑉, which then implies that 𝑇𝑢 = 0 for every 𝑢 ∈ 𝑈
(take 𝑤 = 𝑇𝑢). Hence 𝑇 = 0, as desired.
The next result is false for real inner The next result provides another good
product spaces, as shown by considering example of how self-adjoint operators
any operator on a real inner product space behave like real numbers.
that is not self-adjoint.
7.14 ⟨𝑇𝑣, 𝑣⟩ is real for all 𝑣 ⟺ 𝑇 is self-adjoint (assuming 𝐅 = 𝐂 )
Suppose 𝑉 is a complex inner product space and 𝑇 ∈ ℒ(𝑉). Then
𝑇 is self-adjoint ⟺ ⟨𝑇𝑣, 𝑣⟩ ∈ 𝐑 for every 𝑣 ∈ 𝑉.
Proof If 𝑣 ∈ 𝑉, then
7.15 ⟨𝑇 ∗ 𝑣, 𝑣⟩ = ⟨𝑣, 𝑇 ∗ 𝑣⟩ = ⟨𝑇𝑣, 𝑣⟩.
Now
𝑇 is self-adjoint ⟺ 𝑇 − 𝑇 ∗ = 0
⟺ ⟨(𝑇 − 𝑇 ∗ )𝑣, 𝑣⟩ = 0 for every 𝑣 ∈ 𝑉
⟺ ⟨𝑇𝑣, 𝑣⟩ − ⟨𝑇𝑣, 𝑣⟩ = 0 for every 𝑣 ∈ 𝑉
⟺ ⟨𝑇𝑣, 𝑣⟩ ∈ 𝐑 for every 𝑣 ∈ 𝑉,
where the second equivalence follows from 7.13 as applied to 𝑇 − 𝑇 ∗ and the
third equivalence follows from 7.15.
On a real inner product space 𝑉, a nonzero operator 𝑇 might satisfy ⟨𝑇𝑣, 𝑣⟩ = 0

for all 𝑣 ∈ 𝑉. However, the next result shows that this cannot happen for a self-
adjoint operator.
7.16 𝑇 self-adjoint and ⟨𝑇𝑣, 𝑣⟩ = 0 for all 𝑣 ⟺ 𝑇 = 0
Suppose 𝑇 is a self-adjoint operator on 𝑉. Then
⟨𝑇𝑣, 𝑣⟩ = 0 for every 𝑣 ∈ 𝑉 ⟺ 𝑇 = 0.
Proof We have already proved this (without the hypothesis that 𝑇 is self-adjoint)
when 𝑉 is a complex inner product space (see 7.13). Thus we can assume that 𝑉
is a real inner product space. If 𝑢, 𝑤 ∈ 𝑉, then
⟨𝑇(𝑢 + 𝑤), 𝑢 + 𝑤⟩ − ⟨𝑇(𝑢 − 𝑤), 𝑢 − 𝑤⟩
7.17 ⟨𝑇𝑢, 𝑤⟩ = ,
4
as can be proved by computing the right side using the equation
⟨𝑇𝑤, 𝑢⟩ = ⟨𝑤, 𝑇𝑢⟩ = ⟨𝑇𝑢, 𝑤⟩,
where the first equality holds because 𝑇 is self-adjoint and the second equality
holds because we are working in a real inner product space.
Now suppose ⟨𝑇𝑣, 𝑣⟩ = 0 for every 𝑣 ∈ 𝑉. Because each term on the right
side of 7.17 is of the form ⟨𝑇𝑣, 𝑣⟩ for appropriate 𝑣, this implies that ⟨𝑇𝑢, 𝑤⟩ = 0
for all 𝑢, 𝑤 ∈ 𝑉. This implies that 𝑇𝑢 = 0 for every 𝑢 ∈ 𝑉 (take 𝑤 = 𝑇𝑢). Hence
𝑇 = 0, as desired.
Normal Operators
7.18 definition: normal

• An operator on an inner product space is called normal if it commutes with
its adjoint.
• In other words, 𝑇 ∈ ℒ(𝑉) is normal if 𝑇𝑇 ∗ = 𝑇 ∗ 𝑇.
Every self-adjoint operator is normal, because if 𝑇 is self-adjoint then 𝑇 ∗ = 𝑇

and hence 𝑇 commutes with 𝑇 ∗.
7.19 example: an operator that is normal but not self-adjoint

Let 𝑇 be the operator on 𝐅2 whose matrix (with respect to the standard basis)
is
2 −3
( ).
3 2
Thus 𝑇(𝑤, 𝑧) = (2𝑤 − 3𝑧, 3𝑤 + 2𝑧).
This operator 𝑇 is not self-adjoint because the entry in row 2, column 1 (which
equals 3) does not equal the complex conjugate of the entry in row 1, column 2
(which equals −3).
The matrix of 𝑇𝑇 ∗ equals
2 −3 2 3 13 0
( )( ), which equals ( ).
3 2 −3 2 0 13
Similarly, the matrix of 𝑇 ∗ 𝑇 equals

2 3 2 −3 13 0
( )( ), which equals ( ).
−3 2 3 2 0 13
Because 𝑇𝑇 ∗ and 𝑇 ∗ 𝑇 have the same matrix, we see that 𝑇𝑇 ∗ = 𝑇 ∗ 𝑇. Thus 𝑇 is

normal.
In the next section we will see why normal operators are worthy of special
attention. The next result provides a useful characterization of normal operators.
7.20 𝑇 is normal if and only if 𝑇𝑣 and 𝑇 ∗ 𝑣 have the same norm
Suppose 𝑇 ∈ ℒ(𝑉). Then
𝑇 is normal ⟺ ‖𝑇𝑣‖ = ‖𝑇 ∗ 𝑣‖ for every 𝑣 ∈ 𝑉.
Proof We have
𝑇 is normal ⟺ 𝑇 ∗ 𝑇 − 𝑇𝑇 ∗ = 0
⟺ ⟨(𝑇 ∗ 𝑇 − 𝑇𝑇 ∗ )𝑣, 𝑣⟩ = 0 for every 𝑣 ∈ 𝑉
⟺ ⟨𝑇 ∗ 𝑇𝑣, 𝑣⟩ = ⟨𝑇𝑇 ∗ 𝑣, 𝑣⟩ for every 𝑣 ∈ 𝑉
⟺ ⟨𝑇𝑣, 𝑇𝑣⟩ = ⟨𝑇 ∗ 𝑣, 𝑇 ∗ 𝑣⟩ for every 𝑣 ∈ 𝑉
⟺ ‖𝑇𝑣‖2 = ∥𝑇 ∗ 𝑣∥ for every 𝑣 ∈ 𝑉

2
⟺ ‖𝑇𝑣‖ = ∥𝑇 ∗ 𝑣∥ for every 𝑣 ∈ 𝑉,
where we used 7.16 to establish the second equivalence (note that the operator
𝑇 ∗ 𝑇 − 𝑇𝑇 ∗ is self-adjoint).
The next result presents several consequences of the result above. Compare
(e) of the next result to Exercise 3. That exercise states that the eigenvalues of
the adjoint of each operator are equal (as a set) to the complex conjugates of
the eigenvalues of the operator. The exercise says nothing about eigenvectors,
because an operator and its adjoint may have different eigenvectors. However,
(e) of the next result implies that a normal operator and its adjoint have the same
eigenvectors.
7.21 range, null space, and eigenvectors of a normal operator
Suppose 𝑇 ∈ ℒ(𝑉) is normal. Then

(a) null 𝑇 = null 𝑇 ∗ ;
(b) range 𝑇 = range 𝑇 ∗ ;
(c) 𝑉 = null 𝑇 ⊕ range 𝑇;
(d) 𝑇 − 𝜆𝐼 is normal for every 𝜆 ∈ 𝐅 ;
(e) if 𝑣 ∈ 𝑉 and 𝜆 ∈ 𝐅 , then 𝑇𝑣 = 𝜆𝑣 if and only if 𝑇 ∗ 𝑣 = 𝜆𝑣.
Proof
(a) Suppose 𝑣 ∈ 𝑉. Then
𝑣 ∈ null 𝑇 ⟺ ‖𝑇𝑣‖ = 0 ⟺ ∥𝑇 ∗ 𝑣∥ = 0 ⟺ 𝑣 ∈ null 𝑇 ∗,
where the middle equivalence above follows from 7.20. Thus null 𝑇 = null 𝑇 ∗.
(b) We have
range 𝑇 = (null 𝑇 ∗ )⟂ = (null 𝑇)⟂ = range 𝑇 ∗,
where the first equality comes from 7.6(d), the second equality comes from
(a) in this result, and the third equality comes from 7.6(b).
(c) We have
𝑉 = (null 𝑇) ⊕ (null 𝑇)⟂ = null 𝑇 ⊕ range 𝑇 ∗ = null 𝑇 ⊕ range 𝑇,
where the first equality comes from 6.49, the second equality comes from
7.6(b), and the third equality comes from (b) in this result.
(d) Suppose 𝜆 ∈ 𝐅. Then
(𝑇 − 𝜆𝐼)(𝑇 − 𝜆𝐼)∗ = (𝑇 − 𝜆𝐼)(𝑇 ∗ − 𝜆𝐼)
= 𝑇𝑇 ∗ − 𝜆𝑇 − 𝜆𝑇 ∗ + |𝜆|2 𝐼
= 𝑇 ∗ 𝑇 − 𝜆𝑇 − 𝜆𝑇 ∗ + |𝜆|2 𝐼
= (𝑇 ∗ − 𝜆𝐼)(𝑇 − 𝜆𝐼)
= (𝑇 − 𝜆𝐼)∗ (𝑇 − 𝜆𝐼).
Thus 𝑇 − 𝜆𝐼 commutes with its adjoint. Hence 𝑇 − 𝜆𝐼 is normal.

(e) Suppose 𝑣 ∈ 𝑉 and 𝜆 ∈ 𝐅. Then (d) and 7.20 imply that
‖(𝑇 − 𝜆𝐼)𝑣‖ = ∥(𝑇 − 𝜆𝐼)∗ 𝑣∥ = ∥(𝑇 ∗ − 𝜆𝐼)𝑣∥.
Thus ‖(𝑇 − 𝜆𝐼)𝑣‖ = 0 if and only if ∥(𝑇 ∗ − 𝜆𝐼)𝑣∥ = 0. Hence 𝑇𝑣 = 𝜆𝑣 if

and only if 𝑇 ∗ 𝑣 = 𝜆𝑣.
Because every self-adjoint operator is normal, the next result applies in partic-
ular to self-adjoint operators.
7.22 orthogonal eigenvectors for normal operators
Suppose 𝑇 ∈ ℒ(𝑉) is normal. Then eigenvectors of 𝑇 corresponding to

distinct eigenvalues are orthogonal.
Proof Suppose 𝛼, 𝛽 are distinct eigenvalues of 𝑇, with corresponding eigen-

vectors 𝑢, 𝑣. Thus 𝑇𝑢 = 𝛼𝑢 and 𝑇𝑣 = 𝛽𝑣. From 7.21(e) we have 𝑇 ∗ 𝑣 = 𝛽𝑣.
Thus
(𝛼 − 𝛽)⟨𝑢, 𝑣⟩ = ⟨𝛼𝑢, 𝑣⟩ − ⟨𝑢, 𝛽𝑣⟩
= ⟨𝑇𝑢, 𝑣⟩ − ⟨𝑢, 𝑇 ∗ 𝑣⟩
= 0.
Because 𝛼 ≠ 𝛽, the equation above implies that ⟨𝑢, 𝑣⟩ = 0. Thus 𝑢 and 𝑣 are
orthogonal, as desired.
As stated here, the next result makes sense only when 𝐅 = 𝐂 . However, see
Exercise 12 for a version that makes sense when 𝐅 = 𝐂 and when 𝐅 = 𝐑 .
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Under the analogy between ℒ(𝑉) and 𝐂 ,
with the adjoint on ℒ(𝑉) playing a similar role to that of the complex conjugate on
𝐂 , the operators 𝐴 and 𝐵 as defined by 7.24 correspond to the real and imaginary
parts of 𝑇. Thus the informal title of the result below should make sense.
7.23 𝑇 is normal ⟺ the real and imaginary parts of 𝑇 commute
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Then 𝑇 is normal if and only if there exist

commuting self-adjoint operators 𝐴 and 𝐵 such that 𝑇 = 𝐴 + 𝑖𝐵.
Proof First suppose 𝑇 is normal. Let

𝑇 + 𝑇∗ 𝑇 − 𝑇∗
7.24 𝐴= and 𝐵 = .
2 2𝑖
Then 𝐴 and 𝐵 are self-adjoint and 𝑇 = 𝐴 + 𝑖𝐵. A quick computation shows that
𝑇 ∗ 𝑇 − 𝑇𝑇 ∗
7.25 𝐴𝐵 − 𝐵𝐴 = .
2𝑖
Because 𝑇 is normal, the right side of the equation above equals 0. Thus the
operators 𝐴 and 𝐵 commute, as desired.
To prove the implication in the other direction, now suppose there exist com-
muting self-adjoint operators 𝐴 and 𝐵 such that 𝑇 = 𝐴 + 𝑖𝐵. Then 𝑇 ∗ = 𝐴 − 𝑖𝐵.
Adding the last two equations and then dividing by 2 produces the equation for 𝐴
in 7.24. Subtracting the last two equations and then dividing by 2𝑖 produces the
equation for 𝐵 in 7.24. Now 7.24 implies 7.25. Because 𝐵 and 𝐴 commute, 7.25
implies that 𝑇 is normal, as desired.
Exercises 7A
1 Suppose 𝑛 is a positive integer. Define 𝑇 ∈ ℒ(𝐅𝑛 ) by

𝑇(𝑧1 , …, 𝑧𝑛 ) = (0, 𝑧1 , …, 𝑧𝑛 − 1 ).
Find a formula for 𝑇 ∗ (𝑧1 , …, 𝑧𝑛 ).

2 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that
𝑇 = 0 ⟺ 𝑇 ∗ = 0 ⟺ 𝑇 ∗ 𝑇 = 0 ⟺ 𝑇𝑇 ∗ = 0.
3 Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆 ∈ 𝐅. Prove that

𝜆 is an eigenvalue of 𝑇 ⟺ 𝜆 is an eigenvalue of 𝑇 ∗.
4 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑈 is a subspace of 𝑉. Prove that

𝑈 is invariant under 𝑇 ⟺ 𝑈 ⟂ is invariant under 𝑇 ∗.
5 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Suppose 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉 and

𝑓1 , …, 𝑓𝑚 is an orthonormal basis of 𝑊. Prove that
‖𝑇𝑒1 ‖2 + ⋯ + ‖𝑇𝑒𝑛 ‖2 = ∥𝑇 ∗ 𝑓1 ∥ + ⋯ + ∥𝑇 ∗ 𝑓𝑚 ∥ .
2 2
The numbers ‖𝑇𝑒1 ‖2, …, ‖𝑇𝑒𝑛 ‖2 in the equation above depend on the ortho-
normal basis 𝑒1 , …, 𝑒𝑛 , but the right side of the equation does not depend on
𝑒1 , …, 𝑒𝑛 . Thus the equation above shows that the sum on the left side does
not depend on which orthonormal basis 𝑒1 , …, 𝑒𝑛 is used.
6 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that

(a) 𝑇 is injective ⟺ 𝑇 ∗ is surjective;
(b) 𝑇 is surjective ⟺ 𝑇 ∗ is injective.
7 Prove that if 𝑇 ∈ ℒ(𝑉, 𝑊), then

(a) dim null 𝑇 ∗ = dim null 𝑇 + dim 𝑊 − dim 𝑉;
(b) dim range 𝑇 ∗ = dim range 𝑇.
8 Suppose 𝐴 is an 𝑚-by-𝑛 matrix with entries in 𝐅. Use (b) in Exercise 7 to

prove that the row rank of 𝐴 equals the column rank of 𝐴.
This exercise asks for yet another alternative proof of a result that was
previously proved in 3.57 and 3.133.
9 Prove that the product of two self-adjoint operators on 𝑉 is self-adjoint if

and only if the two operators commute.
10 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is self-adjoint if and only if
⟨𝑇𝑣, 𝑣⟩ = ⟨𝑇 ∗ 𝑣, 𝑣⟩
11 Define an operator 𝑆 ∶ 𝐅2 → 𝐅2 by 𝑆(𝑤, 𝑧) = (−𝑧, 𝑤).

(a) Find a formula for 𝑆∗.
(b) Show that 𝑆 is normal but not self-adjoint.
(c) Find all eigenvalues of 𝑆.
If 𝐅 = 𝐑 , then 𝑆 is the operator on 𝐑2 of counterclockwise rotation by 90∘ .
12 An operator 𝐵 ∈ ℒ(𝑉) is called skew if
𝐵∗ = −𝐵.
Suppose that 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is normal if and only if there exist
commuting operators 𝐴 and 𝐵 such that 𝐴 is self-adjoint, 𝐵 is a skew operator,
and 𝑇 = 𝐴 + 𝐵.
13 Suppose 𝐅 = 𝐑 . Define 𝒜 ∈ ℒ(ℒ(𝑉)) by 𝒜𝑇 = 𝑇 ∗ for all 𝑇 ∈ ℒ(𝑉).
(a) Find all eigenvalues of 𝒜.
(b) Find the minimal polynomial of 𝒜.
14 Define an inner product on 𝒫2 (𝐑) by ⟨𝑝, 𝑞⟩ = ∫01 𝑝𝑞. Define an operator

𝑇 ∈ ℒ(𝒫2 (𝐑)) by
𝑇(𝑎𝑥2 + 𝑏𝑥 + 𝑐) = 𝑏𝑥.
(a) Show that with this inner product, the operator 𝑇 is not self-adjoint.
(b) The matrix of 𝑇 with respect to the basis 1, 𝑥, 𝑥2 is
0 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ 0 1 0 ⎟.
⎟
⎝ 0 0 0 ⎠
This matrix equals its conjugate transpose, even though 𝑇 is not self-
adjoint. Explain why this is not a contradiction.

(a) 𝑇 is self-adjoint ⟺ 𝑇 −1 is self-adjoint;
(b) 𝑇 is normal ⟺ 𝑇 −1 is normal.
16 Suppose 𝐅 = 𝐑 .
(a) Show that the set of self-adjoint operators on 𝑉 is a subspace of ℒ(𝑉).
(b) What is the dimension of the subspace of ℒ(𝑉) in (a) [in terms of
dim 𝑉]?
17 Suppose 𝐅 = 𝐂 . Show that the set of self-adjoint operators on 𝑉 is not a

subspace of ℒ(𝑉).
18 Suppose dim 𝑉 ≥ 2. Show that the set of normal operators on 𝑉 is not a
subspace of ℒ(𝑉).
19 Suppose 𝑇 ∈ ℒ(𝑉) and ∥𝑇 ∗ 𝑣∥ ≤ ‖𝑇𝑣‖ for every 𝑣 ∈ 𝑉. Prove that 𝑇 is

normal.
This exercise fails on infinite-dimensional inner product spaces, leading to
what are called hyponormal operators, which have a well-developed theory.
20 Suppose 𝑃 ∈ ℒ(𝑉) is such that 𝑃2 = 𝑃. Prove that the following are

equivalent.
(a) 𝑃 is self-adjoint.
(b) 𝑃 is normal.
(c) There is a subspace 𝑈 of 𝑉 such that 𝑃 = 𝑃𝑈 .
21 Suppose 𝐷 ∶ 𝒫8 (𝐑) → 𝒫8 (𝐑) is the differentiation operator defined by

𝐷𝑝 = 𝑝′. Prove that there does not exist an inner product on 𝒫8 (𝐑) that
makes 𝐷 a normal operator.
22 Give an example of an operator 𝑇 ∈ ℒ(𝐑3 ) such that 𝑇 is normal but not
self-adjoint.
23 Suppose 𝑇 is a normal operator on 𝑉. Suppose also that 𝑣, 𝑤 ∈ 𝑉 satisfy the
equations
‖𝑣‖ = ‖𝑤‖ = 2, 𝑇𝑣 = 3𝑣, 𝑇𝑤 = 4𝑤.
Show that ‖𝑇(𝑣 + 𝑤)‖ = 10.
24 Suppose 𝑇 ∈ ℒ(𝑉) and
𝑎0 + 𝑎1 𝑧 + 𝑎2 𝑧2 + ⋯ + 𝑎𝑚 − 1 𝑧𝑚 − 1 + 𝑧𝑚
is the minimal polynomial of 𝑇. Prove that the minimal polynomial of 𝑇 ∗ is

𝑎0 + 𝑎1 𝑧 + 𝑎2 𝑧2 + ⋯ + 𝑎𝑚 − 1 𝑧𝑚 − 1 + 𝑧𝑚.
This exercise shows that the minimal polynomial of 𝑇 ∗ equals the minimal
polynomial of 𝑇 if 𝐅 = 𝐑 .
25 Suppose 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is diagonalizable if and only if 𝑇 ∗ is

diagonalizable.
26 Fix 𝑢, 𝑥 ∈ 𝑉. Define 𝑇 ∈ ℒ(𝑉) by 𝑇𝑣 = ⟨𝑣, 𝑢⟩𝑥 for every 𝑣 ∈ 𝑉.
(a) Prove that if 𝑉 is a real vector space, then 𝑇 is self-adjoint if and only if
the list 𝑢, 𝑥 is linearly dependent.
(b) Prove that 𝑇 is normal if and only if the list 𝑢, 𝑥 is linearly dependent.
27 Suppose 𝑇 ∈ ℒ(𝑉) is normal. Prove that

null 𝑇 𝑘 = null 𝑇 and range 𝑇 𝑘 = range 𝑇
for every positive integer 𝑘.
28 Suppose 𝑇 ∈ ℒ(𝑉) is normal. Prove that if 𝜆 ∈ 𝐅, then the minimal
polynomial of 𝑇 is not a polynomial multiple of (𝑥 − 𝜆)2.
29 Prove or give a counterexample: If 𝑇 ∈ ℒ(𝑉) and there is an orthonormal

basis 𝑒1 , …, 𝑒𝑛 of 𝑉 such that ‖𝑇𝑒𝑘 ‖ = ∥𝑇 ∗ 𝑒𝑘 ∥ for each 𝑘 = 1, …, 𝑛, then 𝑇 is
normal.
30 Suppose that 𝑇 ∈ ℒ(𝐅3 ) is normal and 𝑇(1, 1, 1) = (2, 2, 2). Suppose
(𝑧1 , 𝑧2 , 𝑧3 ) ∈ null 𝑇. Prove that 𝑧1 + 𝑧2 + 𝑧3 = 0.
31 Fix a positive integer 𝑛. In the inner product space of continuous real-valued

functions on [−𝜋, 𝜋] with inner product ⟨ 𝑓, 𝑔⟩ = ∫−𝜋 𝜋
𝑓 𝑔, let
𝑉 = span(1, cos 𝑥, cos 2𝑥, …, cos 𝑛𝑥, sin 𝑥, sin 2𝑥, …, sin 𝑛𝑥).
(a) Define 𝐷 ∈ ℒ(𝑉) by 𝐷 𝑓 = 𝑓 ′. Show that 𝐷∗ = −𝐷. Conclude that 𝐷

is normal but not self-adjoint.
(b) Define 𝑇 ∈ ℒ(𝑉) by 𝑇 𝑓 = 𝑓 ″. Show that 𝑇 is self-adjoint.
32 Suppose 𝑇 ∶ 𝑉 → 𝑊 is a linear map. Show that under the standard identifica-

tion of 𝑉 with 𝑉 ′ (see 6.58) and the corresponding identification of 𝑊 with
𝑊 ′, the adjoint map 𝑇 ∗ ∶ 𝑊 → 𝑉 corresponds to the dual map 𝑇 ′ ∶ 𝑊 ′ → 𝑉 ′ .
More precisely, show that
𝑇 ′(𝜑𝑤 ) = 𝜑𝑇 ∗𝑤
for all 𝑤 ∈ 𝑊, where 𝜑𝑤 and 𝜑𝑇∗𝑤 are defined as in 6.58.
Section 7B Spectral Theorem 243
7B Spectral Theorem
Recall that a diagonal matrix is a square matrix that is 0 everywhere except
possibly on the diagonal. Recall that an operator on 𝑉 is called diagonalizable if
the operator has a diagonal matrix with respect to some basis of 𝑉. Recall also
that this happens if and only if there is a basis of 𝑉 consisting of eigenvectors of
the operator (see 5.55).
The nicest operators on 𝑉 are those for which there is an orthonormal basis
of 𝑉 with respect to which the operator has a diagonal matrix. These are precisely
the operators 𝑇 ∈ ℒ(𝑉) such that there is an orthonormal basis of 𝑉 consisting
of eigenvectors of 𝑇. Our goal in this section is to prove the spectral theorem,
which characterizes these operators as the self-adjoint operators when 𝐅 = 𝐑 and
as the normal operators when 𝐅 = 𝐂 .
The spectral theorem is probably the most useful tool in the study of operators
on inner product spaces. Its extension to certain infinite-dimensional inner product
spaces (see, for example, Section 10D of the author’s book Measure, Integration
& Real Analysis) plays a key role in functional analysis.
Because the conclusion of the spectral theorem depends on 𝐅, we will break
the spectral theorem into two pieces, called the real spectral theorem and the
complex spectral theorem.
Real Spectral Theorem

To prove the real spectral theorem, we will need two preliminary results. These
preliminary results hold on both real and complex inner product spaces, but they
are not needed for the proof of the complex spectral theorem.
You could guess that the next result is This completing-the-square technique
true and even discover its proof by think- can be used to derive the quadratic
ing about quadratic polynomials with formula.
real coefficients. Specifically, suppose
𝑏, 𝑐 ∈ 𝐑 and 𝑏2 < 4𝑐. Let 𝑥 be a real number. Then
𝑏 2 𝑏2
𝑥2 + 𝑏𝑥 + 𝑐 = (𝑥 + ) + (𝑐 − ) > 0.
2 4
In particular, 𝑥2 + 𝑏𝑥 + 𝑐 is an invertible real number (a convoluted way of saying
that it is not 0). Replacing the real number 𝑥 with a self-adjoint operator (recall the
analogy between real numbers and self-adjoint operators) leads to the next result.
7.26 invertible quadratic expressions
Suppose 𝑇 ∈ ℒ(𝑉) is self-adjoint and 𝑏, 𝑐 ∈ 𝐑 are such that 𝑏2 < 4𝑐. Then
𝑇 2 + 𝑏𝑇 + 𝑐𝐼
is an invertible operator.
Proof Let 𝑣 be a nonzero vector in 𝑉. Then

⟨(𝑇 2 + 𝑏𝑇 + 𝑐𝐼)𝑣, 𝑣⟩ = ⟨𝑇 2 𝑣, 𝑣⟩ + 𝑏⟨𝑇𝑣, 𝑣⟩ + 𝑐⟨𝑣, 𝑣⟩
= ⟨𝑇𝑣, 𝑇𝑣⟩ + 𝑏⟨𝑇𝑣, 𝑣⟩ + 𝑐‖𝑣‖2
≥ ‖𝑇𝑣‖2 − |𝑏| ‖𝑇𝑣‖ ‖𝑣‖ + 𝑐‖𝑣‖2
|𝑏| ‖𝑣‖ 2 𝑏2
= (‖𝑇𝑣‖ − ) + (𝑐 − )‖𝑣‖2
2 4
> 0,
where the third line above holds by the Cauchy–Schwarz inequality (6.14). The
last inequality implies that (𝑇 2 + 𝑏𝑇 + 𝑐𝐼)𝑣 ≠ 0. Thus 𝑇 2 + 𝑏𝑇 + 𝑐𝐼 is injective,
which implies that it is invertible (see 3.65).
The next result will be a key tool in our proof of the real spectral theorem.
7.27 minimal polynomial of self-adjoint operator
Suppose 𝑇 ∈ ℒ(𝑉) is self-adjoint. Then the minimal polynomial of 𝑇 equals

(𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ) for some 𝜆1 , …, 𝜆𝑚 ∈ 𝐑 .
Proof First suppose 𝐅 = 𝐂 . The zeros of the minimal polynomial of 𝑇 are the
eigenvalues of 𝑇 [by 5.27(a)]. All eigenvalues of 𝑇 are real (by 7.12). Thus the
second version of the fundamental theorem of algebra (see 6.69) tells us that the
minimal polynomial of 𝑇 has the desired form.
Now suppose 𝐅 = 𝐑 . By the factorization of a polynomial over 𝐑 (see 4.16)
there exist 𝜆1 , …, 𝜆𝑚 ∈ 𝐑 and 𝑏1 , …, 𝑏𝑁 , 𝑐1 , …, 𝑐𝑁 ∈ 𝐑 with 𝑏𝑘2 < 4𝑐𝑘 for each 𝑘
such that the minimal polynomial of 𝑇 equals
7.28 (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 )(𝑧2 + 𝑏1 𝑧 + 𝑐1 )⋯(𝑧2 + 𝑏𝑁 𝑧 + 𝑐𝑁 );
here either 𝑚 or 𝑁 might equal 0, meaning that there are no terms of the corre-
sponding form. Now
(𝑇 − 𝜆1 𝐼)⋯(𝑇 − 𝜆𝑚 𝐼)(𝑇 2 + 𝑏1 𝑇 + 𝑐1 𝐼)⋯(𝑇 2 + 𝑏𝑁 𝑇 + 𝑐𝑁 𝐼) = 0.
If 𝑁 > 0, then we could multiply both sides of the equation above on the right by
the inverse of 𝑇 2 + 𝑏𝑁 𝑇 + 𝑐𝑁 𝐼 (which is an invertible operator by 7.26) to obtain a
polynomial expression of 𝑇 that equals 0. The corresponding polynomial would
have degree two less than the degree of 7.28, violating the minimality of the
degree of the polynomial with this property. Thus we must have 𝑁 = 0, which
means that the minimal polynomial in 7.28 has the form (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ), as
desired.
The result above along with 5.27(a) implies that every self-adjoint operator
has an eigenvalue. In fact, as we will see in the next result, self-adjoint operators
have enough eigenvectors to form a basis.
The next result, which gives a complete description of the self-adjoint operators
on a real inner product space, is one of the major theorems in linear algebra.
7.29 real spectral theorem
Suppose 𝐅 = 𝐑 and 𝑇 ∈ ℒ(𝑉). Then the following are equivalent.

(a) 𝑇 is self-adjoint.
(b) 𝑇 has a diagonal matrix with respect to some orthonormal basis of 𝑉.
(c) 𝑉 has an orthonormal basis consisting of eigenvectors of 𝑇.
Proof First suppose (a) holds, so 𝑇 is self-adjoint. Our results on minimal poly-
nomials, specifically 6.37 and 7.27, imply that 𝑇 has an upper-triangular matrix
with respect to some orthonormal basis of 𝑉. With respect to this orthonormal
basis, the matrix of 𝑇 ∗ is the transpose of the matrix of 𝑇. However, 𝑇 ∗ = 𝑇.
Thus the transpose of the matrix of 𝑇 equals the matrix of 𝑇. Because the matrix
of 𝑇 is upper-triangular, this means that all entries of the matrix above and below
the diagonal are 0. Hence the matrix of 𝑇 is a diagonal matrix with respect to the
orthonormal basis. Thus (a) implies (b).
Conversely, now suppose (b) holds, so 𝑇 has a diagonal matrix with respect to
some orthonormal basis of 𝑉. That diagonal matrix equals its transpose. Thus
with respect to that basis, the matrix of 𝑇 ∗ equals the matrix of 𝑇. Hence 𝑇 ∗ = 𝑇,
proving that (b) implies (a).
The equivalence of (b) and (c) follows from the definitions [or see the proof
that (a) and (b) are equivalent in 5.55].
7.30 example: an orthonormal basis of eigenvectors for an operator

Consider the operator 𝑇 on 𝐑3 whose matrix (with respect to the standard
basis) is
14 −13 8
⎛
⎜ ⎞
⎜
⎜ −13 14 8 ⎟⎟
⎟.
⎝ 8 8 −7 ⎠
This matrix with real entries equals its transpose; thus 𝑇 is self-adjoint. As you
can verify,
(1, −1, 0) (1, 1, 1) (1, 1, −2)
, ,
√2 √3 √6
is an orthonormal basis of 𝐑 consisting of eigenvectors of 𝑇. With respect to
3
this basis, the matrix of 𝑇 is the diagonal matrix

27 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ 0 9 0 ⎟
⎟.
⎝ 0 0 −15 ⎠
See Exercise 17 for a version of the real spectral theorem that applies simulta-
neously to more than one operator.
Complex Spectral Theorem

The next result gives a complete description of the normal operators on a complex
inner product space.
7.31 complex spectral theorem
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Then the following are equivalent.

(a) 𝑇 is normal.
(b) 𝑇 has a diagonal matrix with respect to some orthonormal basis of 𝑉.
(c) 𝑉 has an orthonormal basis consisting of eigenvectors of 𝑇.
Proof First suppose (a) holds, so 𝑇 is normal. By Schur’s theorem (6.38), there is
an orthonormal basis 𝑒1 , …, 𝑒𝑛 of 𝑉 with respect to which 𝑇 has an upper-triangular
matrix. Thus we can write
𝑎1, 1 ⋯ 𝑎 1, 𝑛
⎛ ⎞
7.32 ℳ(𝑇, (𝑒1 , …, 𝑒𝑛 )) = ⎜
⎜
⎜ ⋱ ⋮ ⎟
⎟
⎟.
⎝ 0 𝑎 𝑛, 𝑛 ⎠
We will show that this matrix is actually a diagonal matrix.
We see from the matrix above that
‖𝑇𝑒1 ‖2 = |𝑎1, 1 |2,
∥𝑇 ∗ 𝑒1 ∥ = |𝑎1, 1 |2 + |𝑎1, 2 |2 + ⋯ + |𝑎1, 𝑛 |2.
2
Because 𝑇 is normal, ‖𝑇𝑒1 ‖ = ∥𝑇 ∗ 𝑒1 ∥ (see 7.20). Thus the two equations above
imply that all entries in the first row of the matrix in 7.32, except possibly the first
entry 𝑎1, 1 , equal 0.
Now 7.32 implies
‖𝑇𝑒2 ‖2 = |𝑎2, 2 |2
(because 𝑎1, 2 = 0, as we showed in the paragraph above) and
∥𝑇 ∗ 𝑒2 ∥ = |𝑎2, 2 |2 + |𝑎2, 3 |2 + ⋯ + |𝑎2, 𝑛 |2.
2
Because 𝑇 is normal, ‖𝑇𝑒2 ‖ = ∥𝑇 ∗ 𝑒2 ∥. Thus the two equations above imply that
all entries in the second row of the matrix in 7.32, except possibly the diagonal
entry 𝑎2, 2 , equal 0.
Continuing in this fashion, we see that all nondiagonal entries in the matrix
7.32 equal 0. Thus (b) holds, completing the proof that (a) implies (b).
Now suppose (b) holds, so 𝑇 has a diagonal matrix with respect to some
orthonormal basis of 𝑉. The matrix of 𝑇 ∗ (with respect to the same basis) is
obtained by taking the conjugate transpose of the matrix of 𝑇; hence 𝑇 ∗ also has a
diagonal matrix. Any two diagonal matrices commute; thus 𝑇 commutes with 𝑇 ∗,
which means that 𝑇 is normal. In other words, (a) holds, completing the proof
that (b) implies (a).
The equivalence of (b) and (c) follows from the definitions (also see 5.55).
See Exercises 13 and 20 for alternative proofs that (a) implies (b) in the
previous result.
Exercises 14 and 15 interpret the real spectral theorem and the complex
spectral theorem by expressing the domain space as an orthogonal direct sum of
eigenspaces.
See Exercise 16 for a version of the complex spectral theorem that applies
simultaneously to more than one operator.
The main conclusion of the complex spectral theorem is that every normal
operator on a complex finite-dimensional inner product space is diagonalizable
by an orthonormal basis, as illustrated by the next example.
7.33 example: an orthonormal basis of eigenvectors for an operator

Consider the operator 𝑇 ∈ ℒ(𝐂2 ) defined by 𝑇(𝑤, 𝑧) = (2𝑤 − 3𝑧, 3𝑤 + 2𝑧).
The matrix of 𝑇 (with respect to the standard basis) is
2 −3
( ).
3 2
As we saw in Example 7.19, 𝑇 is a normal operator.

As you can verify,
1
√
(𝑖, 1), √1 (−𝑖, 1)
2 2
is an orthonormal basis of 𝐂 consisting of eigenvectors of 𝑇, and with respect to

2
this basis the matrix of 𝑇 is the diagonal matrix

2 + 3𝑖 0
( ).
0 2 − 3𝑖
Exercises 7B
1 Prove that a normal operator on a complex inner product space is self-adjoint

if and only if all its eigenvalues are real.
This exercise strengthens the analogy (for normal operators) between self-
adjoint operators and real numbers.
2 Suppose 𝐅 = 𝐂 . Suppose 𝑇 ∈ ℒ(𝑉) is normal and has only one eigenvalue.

Prove that 𝑇 is a scalar multiple of the identity operator.
3 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉) is normal. Prove that the set of eigenvalues
of 𝑇 is contained in {0, 1} if and only if there is a subspace 𝑈 of 𝑉 such that
𝑇 = 𝑃𝑈 .
4 Prove that a normal operator on a complex inner product space is skew

(meaning it equals the negative of its adjoint) if and only if all its eigenvalues
are purely imaginary (meaning that they have real part equal to 0).
5 Prove or give a counterexample: If 𝑇 ∈ ℒ(𝐂3 ) is a diagonalizable operator,

then 𝑇 is normal (with respect to the usual inner product).
6 Suppose 𝑉 is a complex inner product space and 𝑇 ∈ ℒ(𝑉) is a normal
operator such that 𝑇 9 = 𝑇 8. Prove that 𝑇 is self-adjoint and 𝑇 2 = 𝑇.
7 Give an example of an operator 𝑇 on a complex vector space such that
𝑇 9 = 𝑇 8 but 𝑇 2 ≠ 𝑇.
8 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is normal if and only if every

eigenvector of 𝑇 is also an eigenvector of 𝑇 ∗.
9 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is normal if and only if there
exists a polynomial 𝑝 ∈ 𝒫(𝐂) such that 𝑇 ∗ = 𝑝(𝑇).
10 Suppose 𝑉 is a complex inner product space. Prove that every normal
operator on 𝑉 has a square root.
An operator 𝑆 ∈ ℒ(𝑉) is called a square root of 𝑇 ∈ ℒ(𝑉) if 𝑆2 = 𝑇. We
will discuss more about square roots of operators in Sections 7C and 8C.
11 Prove that every self-adjoint operator on 𝑉 has a cube root.

An operator 𝑆 ∈ ℒ(𝑉) is called a cube root of 𝑇 ∈ ℒ(𝑉) if 𝑆3 = 𝑇.
12 Suppose 𝑉 is a complex vector space and 𝑇 ∈ ℒ(𝑉) is normal. Prove that

if 𝑆 is an operator on 𝑉 that commutes with 𝑇, then 𝑆 commutes with 𝑇 ∗.
The result in this exercise is called Fuglede’s theorem.
13 Without using the complex spectral theorem, use the version of Schur’s
theorem that applies to two commuting operators (take ℰ = {𝑇, 𝑇 ∗ } in
Exercise 20 in Section 6B) to give a different proof that if 𝐅 = 𝐂 and
𝑇 ∈ ℒ(𝑉) is normal, then 𝑇 has a diagonal matrix with respect to some
orthonormal basis of 𝑉.
14 Suppose 𝐅 = 𝐑 and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is self-adjoint if and only
if all pairs of eigenvectors corresponding to distinct eigenvalues of 𝑇 are
orthogonal and 𝑉 = 𝐸(𝜆1 , 𝑇) ⊕ ⋯ ⊕ 𝐸(𝜆𝑚 , 𝑇), where 𝜆1 , …, 𝜆𝑚 denote the
distinct eigenvalues of 𝑇.
15 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is normal if and only if all pairs
of eigenvectors corresponding to distinct eigenvalues of 𝑇 are orthogonal
and 𝑉 = 𝐸(𝜆1 , 𝑇) ⊕ ⋯ ⊕ 𝐸(𝜆𝑚 , 𝑇), where 𝜆1 , …, 𝜆𝑚 denote the distinct
eigenvalues of 𝑇.
16 Suppose 𝐅 = 𝐂 and ℰ ⊆ ℒ(𝑉). Prove that there is an orthonormal basis
of 𝑉 with respect to which every element of ℰ has a diagonal matrix if and
only if 𝑆 and 𝑇 are commuting normal operators for all 𝑆, 𝑇 ∈ ℰ.
This exercise extends the complex spectral theorem to the context of a
collection of commuting normal operators.
17 Suppose 𝐅 = 𝐑 and ℰ ⊆ ℒ(𝑉). Prove that there is an orthonormal basis

of 𝑉 with respect to which every element of ℰ has a diagonal matrix if and
only if 𝑆 and 𝑇 are commuting self-adjoint operators for all 𝑆, 𝑇 ∈ ℰ.
This exercise extends the real spectral theorem to the context of a collection
of commuting self-adjoint operators.
18 Give an example of a real inner product space 𝑉, an operator 𝑇 ∈ ℒ(𝑉),

and real numbers 𝑏, 𝑐 with 𝑏2 < 4𝑐 such that
𝑇 2 + 𝑏𝑇 + 𝑐𝐼
is not invertible.
This exercise shows that the hypothesis that 𝑇 is self-adjoint cannot be
deleted in 7.26, even for real vector spaces.
19 Suppose 𝑇 ∈ ℒ(𝑉) is self-adjoint and 𝑈 is a subspace of 𝑉 that is invariant

under 𝑇.
(a) Prove that 𝑈 ⟂ is invariant under 𝑇.
(b) Prove that 𝑇|𝑈 ∈ ℒ(𝑈) is self-adjoint.
(c) Prove that 𝑇|𝑈⟂ ∈ ℒ(𝑈 ⟂ ) is self-adjoint.
20 Suppose 𝑇 ∈ ℒ(𝑉) is normal and 𝑈 is a subspace of 𝑉 that is invariant

under 𝑇.
(a) Prove that 𝑈 ⟂ is invariant under 𝑇.
(b) Prove that 𝑈 is invariant under 𝑇 ∗.
(c) Prove that (𝑇|𝑈 )∗ = (𝑇 ∗ )|𝑈 .
(d) Prove that 𝑇|𝑈 ∈ ℒ(𝑈) and 𝑇|𝑈⟂ ∈ ℒ(𝑈 ⟂ ) are normal operators.
This exercise can be used to give yet another proof of the complex spectral
theorem (use induction on dim 𝑉 and the result that 𝑇 has an eigenvector).
21 Suppose that 𝑇 is a self-adjoint operator on a finite-dimensional inner product

space and that 2 and 3 are the only eigenvalues of 𝑇. Prove that
𝑇 2 − 5𝑇 + 6𝐼 = 0.
22 Give an example of an operator 𝑇 ∈ ℒ(𝐂3 ) such that 2 and 3 are the only
eigenvalues of 𝑇 and 𝑇 2 − 5𝑇 + 6𝐼 ≠ 0.
23 Suppose 𝑇 ∈ ℒ(𝑉) is self-adjoint, 𝜆 ∈ 𝐅, and 𝜖 > 0. Suppose there exists
𝑣 ∈ 𝑉 such that ‖𝑣‖ = 1 and
‖𝑇𝑣 − 𝜆𝑣‖ < 𝜖.
Prove that 𝑇 has an eigenvalue 𝜆′ such that ∣𝜆 − 𝜆′ ∣ < 𝜖.

This exercise shows that for a self-adjoint operator, a number that is close
to satisfying an equation that would make it an eigenvalue is close to an
eigenvalue.
24 Suppose 𝑈 is a finite-dimensional vector space and 𝑇 ∈ ℒ(𝑈).

(a) Suppose 𝐅 = 𝐑 . Prove that 𝑇 is diagonalizable if and only if there is a
basis of 𝑈 such that the matrix of 𝑇 with respect to this basis equals its
transpose.
(b) Suppose 𝐅 = 𝐂 . Prove that 𝑇 is diagonalizable if and only if there is a
basis of 𝑈 such that the matrix of 𝑇 with respect to this basis commutes
with its conjugate transpose.
This exercise adds another equivalence to the list of conditions equivalent
to diagonalizability in 5.55.
25 Suppose that 𝑇 ∈ ℒ(𝑉) and there is an orthonormal basis 𝑒1 , …, 𝑒𝑛 of 𝑉

consisting of eigenvectors of 𝑇, with corresponding eigenvalues 𝜆1 , …, 𝜆𝑛 .
Show that if 𝑘 ∈ {1, …, 𝑛}, then the pseudoinverse 𝑇 † satisfies the equation
⎧1𝑒
{𝜆 𝑘 if 𝜆𝑘 ≠ 0,
𝑇†𝑒 𝑘 =⎨ 𝑘
{0
⎩ if 𝜆𝑘 = 0.
Section 7C Positive Operators 251
7C Positive Operators
7.34 definition: positive operator
An operator 𝑇 ∈ ℒ(𝑉) is called positive if 𝑇 is self-adjoint and

⟨𝑇𝑣, 𝑣⟩ ≥ 0
If 𝑉 is a complex vector space, then the requirement that 𝑇 be self-adjoint can

be dropped from the definition above (by 7.14).
7.35 example: positive operators

(a) Let 𝑇 ∈ ℒ(𝐅2 ) be the operator whose matrix (using the standard basis) is
2 −1 ). Then 𝑇 is self-adjoint and ⟨𝑇(𝑤, 𝑧), (𝑤, 𝑧)⟩ = 2|𝑤|2 −2 Re(𝑤𝑧) + |𝑧|2
( −1 1
= |𝑤 − 𝑧|2 + |𝑤|2 ≥ 0 for all (𝑤, 𝑧) ∈ 𝐅2 . Thus 𝑇 is a positive operator.
(b) If 𝑈 is a subspace of 𝑉, then the orthogonal projection 𝑃𝑈 is a positive operator,
(c) If 𝑇 ∈ ℒ(𝑉) is self-adjoint and 𝑏, 𝑐 ∈ 𝐑 are such that 𝑏2 < 4𝑐, then 𝑇 2 +𝑏𝑇+𝑐𝐼
is a positive operator, as shown by the proof of 7.26.
7.36 definition: square root
An operator 𝑅 is called a square root of an operator 𝑇 if 𝑅2 = 𝑇.
7.37 example: square root of an operator

If 𝑇 ∈ ℒ(𝐅3 ) is defined by 𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (𝑧3 , 0, 0), then the operator
𝑅 ∈ ℒ(𝐅3 ) defined by 𝑅(𝑧1 , 𝑧2 , 𝑧3 ) = (𝑧2 , 𝑧3 , 0) is a square root of 𝑇 because
𝑅2 = 𝑇, as you can verify.
The characterizations of the positive Because positive operators correspond

operators in the next result correspond to nonnegative numbers, better termi-
to characterizations of the nonnegative nology would use the term nonnegative
numbers among 𝐂 . Specifically, a num- operators. However, operator theorists
ber 𝑧 ∈ 𝐂 is nonnegative if and only consistently call these positive opera-
if it has a nonnegative square root, cor- tors, so we follow that custom. Some
responding to condition (d). Also, 𝑧 is mathematicians use the term positive
nonnegative if and only if it has a real semidefinite operator, which means
square root, corresponding to condition the same as positive operator.
(e). Finally, 𝑧 is nonnegative if and only
if there exists 𝑤 ∈ 𝐂 such that 𝑧 = 𝑤𝑤, corresponding to condition (f). See
Exercise 20 for another condition that is equivalent to being a positive operator.
7.38 characterization of positive operators
Let 𝑇 ∈ ℒ(𝑉). Then the following are equivalent.

(a) 𝑇 is a positive operator.
(b) 𝑇 is self-adjoint and all eigenvalues of 𝑇 are nonnegative.
(c) With respect to some orthonormal basis of 𝑉, the matrix of 𝑇 is a diagonal
matrix with only nonnegative numbers on the diagonal.
(d) 𝑇 has a positive square root.
(e) 𝑇 has a self-adjoint square root.
(f) 𝑇 = 𝑅∗ 𝑅 for some 𝑅 ∈ ℒ(𝑉).
Proof We will prove that (a) ⇒ (b) ⇒ (c) ⇒ (d) ⇒ (e) ⇒ (f) ⇒ (a).
First suppose (a) holds, so that 𝑇 is positive, which implies that 𝑇 is self-adjoint
(by definition of positive operator). To prove the other condition in (b), suppose
𝜆 is an eigenvalue of 𝑇. Let 𝑣 be an eigenvector of 𝑇 corresponding to 𝜆. Then
0 ≤ ⟨𝑇𝑣, 𝑣⟩ = ⟨𝜆𝑣, 𝑣⟩ = 𝜆⟨𝑣, 𝑣⟩.
Thus 𝜆 is a nonnegative number. Hence (b) holds, showing that (a) implies (b).
Now suppose (b) holds, so that 𝑇 is self-adjoint and all eigenvalues of 𝑇 are
nonnegative. By the spectral theorem (7.29 and 7.31), there is an orthonormal
basis 𝑒1 , …, 𝑒𝑛 of 𝑉 consisting of eigenvectors of 𝑇. Let 𝜆1 , …, 𝜆𝑛 be the eigenval-
ues of 𝑇 corresponding to 𝑒1 , …, 𝑒𝑛 ; thus each 𝜆𝑘 is a nonnegative number. The
matrix of 𝑇 with respect to 𝑒1 , …, 𝑒𝑛 is the diagonal matrix with 𝜆1 , …, 𝜆𝑛 on the
diagonal, which shows that (b) implies (c).
Now suppose (c) holds. Suppose 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉 such
that the matrix of 𝑇 with respect to this basis is a diagonal matrix with nonnegative
numbers 𝜆1 , …, 𝜆𝑛 on the diagonal. The linear map lemma (3.4) implies that
there exists 𝑅 ∈ ℒ(𝑉) such that
𝑅𝑒𝑘 = √ 𝜆𝑘 𝑒𝑘
for each 𝑘 = 1, …, 𝑛. As you should verify, 𝑅 is a positive operator. Furthermore,
𝑅2 𝑒𝑘 = 𝜆𝑘 𝑒𝑘 = 𝑇𝑒𝑘 for each 𝑘, which implies that 𝑅2 = 𝑇. Thus 𝑅 is a positive
square root of 𝑇. Hence (d) holds, which shows that (c) implies (d).
Every positive operator is self-adjoint (by definition of positive operator).
Thus (d) implies (e).
Now suppose (e) holds, meaning that there exists a self-adjoint operator 𝑅 on
𝑉 such that 𝑇 = 𝑅2. Then 𝑇 = 𝑅∗ 𝑅 (because 𝑅∗ = 𝑅). Hence (e) implies (f).
Finally, suppose (f) holds. Let 𝑅 ∈ ℒ(𝑉) be such that 𝑇 = 𝑅∗ 𝑅. Then
∗ ∗
𝑇 = (𝑅∗ 𝑅) = 𝑅∗ (𝑅∗ ) = 𝑅∗ 𝑅 = 𝑇. Hence 𝑇 is self-adjoint. To complete the
∗
proof that (a) holds, note that
⟨𝑇𝑣, 𝑣⟩ = ⟨𝑅∗ 𝑅𝑣, 𝑣⟩ = ⟨𝑅𝑣, 𝑅𝑣⟩ ≥ 0
for every 𝑣 ∈ 𝑉. Thus 𝑇 is positive, showing that (f) implies (a).
Every nonnegative number has a unique nonnegative square root. The next
result shows that positive operators enjoy a similar property.
7.39 each positive operator has only one positive square root
Every positive operator on 𝑉 has a unique positive square root.
Proof Suppose 𝑇 ∈ ℒ(𝑉) is positive. A positive operator can have infinitely

Suppose 𝑣 ∈ 𝑉 is an eigenvector of 𝑇. many square roots (although only one
Hence there exists a real number 𝜆 ≥ 0 of them can be positive). For example,
such that 𝑇𝑣 = 𝜆𝑣. the identity operator on 𝑉 has infinitely
Let 𝑅 be a positive square root of 𝑇. many square roots if dim 𝑉 > 1.
We will prove that 𝑅𝑣 = √ 𝜆𝑣. This will
imply that the behavior of 𝑅 on the eigenvectors of 𝑇 is uniquely determined.
Because there is a basis of 𝑉 consisting of eigenvectors of 𝑇 (by the spectral
theorem), this will imply that 𝑅 is uniquely determined.
To prove that 𝑅𝑣 = √ 𝜆𝑣, note that the spectral theorem asserts that there is an
orthonormal basis 𝑒1 , …, 𝑒𝑛 of 𝑉 consisting of eigenvectors of 𝑅. Because 𝑅 is a
positive operator, all its eigenvalues are nonnegative. Thus there exist nonnegative
numbers 𝜆1 , …, 𝜆𝑛 such that 𝑅𝑒𝑘 = √ 𝜆𝑘 𝑒𝑘 for each 𝑘 = 1, …, 𝑛.
Because 𝑒1 , …, 𝑒𝑛 is a basis of 𝑉, we can write
𝑣 = 𝑎1 𝑒1 + ⋯ + 𝑎𝑛 𝑒𝑛
for some numbers 𝑎1 , …, 𝑎𝑛 ∈ 𝐅. Thus
𝑅𝑣 = 𝑎1 √ 𝜆1 𝑒1 + ⋯ + 𝑎𝑛 √ 𝜆𝑛 𝑒𝑛 .
Hence
𝜆𝑣 = 𝑇𝑣 = 𝑅2 𝑣 = 𝑎1 𝜆1 𝑒1 + ⋯ + 𝑎𝑛 𝜆𝑛 𝑒𝑛 .
The equation above implies that
𝑎1 𝜆𝑒1 + ⋯ + 𝑎𝑛 𝜆𝑒𝑛 = 𝑎1 𝜆1 𝑒1 + ⋯ + 𝑎𝑛 𝜆𝑛 𝑒𝑛 .
Thus 𝑎𝑘 (𝜆 − 𝜆𝑘 ) = 0 for each 𝑘 = 1, …, 𝑛. Hence
𝑣= ∑ 𝑎𝑘 𝑒𝑘 .
{𝑘 ∶ 𝜆𝑘 = 𝜆}
Thus
𝑅𝑣 = ∑ 𝑎𝑘 √ 𝜆𝑒𝑘 = √ 𝜆𝑣,
{𝑘 ∶ 𝜆𝑘 = 𝜆}
as desired.
The notation defined below makes sense thanks to the result above.
7.40 notation: √𝑇
For 𝑇 a positive operator, √𝑇 denotes the unique positive square root of 𝑇.

7.41 example: square root of positive operators

Define operators 𝑆, 𝑇 on 𝐑2 (with the usual Euclidean inner product) by
𝑆(𝑥, 𝑦) = (𝑥, 2𝑦) and 𝑇(𝑥, 𝑦) = (𝑥 + 𝑦, 𝑥 + 𝑦).
Then with respect to the standard basis of 𝐑2 we have
1 0 1 1
7.42 ℳ(𝑆) = ( ) and ℳ(𝑇) = ( ).
0 2 1 1
Each of these matrices equals its transpose; thus 𝑆 and 𝑇 are self-adjoint.
If (𝑥, 𝑦) ∈ 𝐑2 , then
⟨𝑆(𝑥, 𝑦), (𝑥, 𝑦)⟩ = 𝑥2 + 2𝑦2 ≥ 0
and
⟨𝑇(𝑥, 𝑦), (𝑥, 𝑦)⟩ = 𝑥2 + 2𝑥𝑦 + 𝑦2 = (𝑥 + 𝑦)2 ≥ 0.
Thus 𝑆 and 𝑇 are positive operators.
The standard basis of 𝐑2 is an orthonormal basis consisting of eigenvectors of
𝑆. Note that
( √1 , √1 ), ( √1 , − √1 )
2 2 2 2
is an orthonormal basis of eigenvectors of 𝑇, with eigenvalue 2 for the first
eigenvector and eigenvalue 0 for the second eigenvector. Thus √𝑇 has the same
eigenvectors, with eigenvalues √2 and 0.
You can verify that
1 1
⎛ ⎞
⎛ 1
ℳ(√𝑆 ) = ⎜
0 ⎞
⎟ and ℳ(√𝑇 ) = ⎜
⎜
⎜
√2 √2 ⎟
⎟
⎟
√2 ⎜ ⎟
⎝ 0 ⎠ 1 1
⎝ √2 √2 ⎠
with respect to the standard basis by showing that the squares of the matrices
above are the matrices in 7.42 and that each matrix above is the matrix of a positive
operator.
The statement of the next result does not involve a square root, but the clean
proof makes nice use of the square root of a positive operator.
7.43 𝑇 positive and ⟨𝑇𝑣, 𝑣⟩ = 0 ⟹ 𝑇𝑣 = 0
Suppose 𝑇 is a positive operator on 𝑉 and 𝑣 ∈ 𝑉 is such that ⟨𝑇𝑣, 𝑣⟩ = 0.

Then 𝑇𝑣 = 0.
Proof We have
2
0 = ⟨𝑇𝑣, 𝑣⟩ = ⟨√𝑇√𝑇𝑣, 𝑣⟩ = ⟨√𝑇𝑣, √𝑇𝑣⟩ = ∥√𝑇𝑣∥ .
Hence √𝑇𝑣 = 0. Thus 𝑇𝑣 = √𝑇(√𝑇𝑣) = 0, as desired.

Exercises 7C
1 Suppose 𝑇 ∈ ℒ(𝑉). Prove that if both 𝑇 and −𝑇 are positive operators, then
𝑇 = 0.
2 Suppose 𝑇 ∈ ℒ(𝐅4 ) is the operator whose matrix (with respect to the

standard basis) is
2 −1 0 0
⎛
⎜ ⎞
⎜
⎜ −1 2 −1 0 ⎟⎟
⎟
⎜
⎜ ⎟
⎟ .
⎜ 0 −1 2 −1 ⎟
⎝ 0 0 −1 2 ⎠
Show that 𝑇 is an invertible positive operator.
3 Suppose 𝑛 is a positive integer and 𝑇 ∈ ℒ(𝐅𝑛 ) is the operator whose matrix
(with respect to the standard basis) consists of all 1’s. Show that 𝑇 is a
positive operator.
4 Suppose 𝑛 is an integer with 𝑛 > 1. Show that there exists an 𝑛-by-𝑛 matrix
𝐴 such that all of the entries of 𝐴 are positive numbers and 𝐴 = 𝐴∗, but the
operator on 𝐅𝑛 whose matrix (with respect to the standard basis) equals 𝐴 is
not a positive operator.
5 Suppose 𝑇 ∈ ℒ(𝑉) is self-adjoint. Prove that 𝑇 is a positive operator if and
only if for every orthonormal basis 𝑒1 , …, 𝑒𝑛 of 𝑉, all entries on the diagonal
of ℳ(𝑇, (𝑒1 , …, 𝑒𝑛 )) are nonnegative numbers.
6 Prove that the sum of two positive operators on 𝑉 is a positive operator.
7 Suppose 𝑆 ∈ ℒ(𝑉) is an invertible positive operator and 𝑇 ∈ ℒ(𝑉) is a
positive operator. Prove that 𝑆 + 𝑇 is invertible.
8 Suppose 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is a positive operator if and only if the
pseudoinverse 𝑇 † is a positive operator.
9 Suppose 𝑇 ∈ ℒ(𝑉) is a positive operator and 𝑆 ∈ ℒ(𝑊, 𝑉). Prove that
𝑆∗ 𝑇𝑆 is a positive operator on 𝑊.
10 Suppose 𝑇 is a positive operator on 𝑉. Suppose 𝑣, 𝑤 ∈ 𝑉 are such that
𝑇𝑣 = 𝑤 and 𝑇𝑤 = 𝑣.
Prove that 𝑣 = 𝑤.
11 Suppose 𝑇 is a positive operator on 𝑉 and 𝑈 is a subspace of 𝑉 invariant
under 𝑇. Prove that 𝑇|𝑈 ∈ ℒ(𝑈) is a positive operator on 𝑈.
12 Suppose 𝑇 ∈ ℒ(𝑉) is a positive operator. Prove that 𝑇 𝑘 is a positive operator
13 Suppose 𝑇 ∈ ℒ(𝑉) is self-adjoint and 𝛼 ∈ 𝐑 .

(a) Prove that 𝑇 − 𝛼𝐼 is a positive operator if and only if 𝛼 is less than or
equal to every eigenvalue of 𝑇.
(b) Prove that 𝛼𝐼 − 𝑇 is a positive operator if and only if 𝛼 is greater than or
equal to every eigenvalue of 𝑇.
14 Suppose 𝑇 is a positive operator on 𝑉 and 𝑣1 , …, 𝑣𝑚 ∈ 𝑉. Prove that

𝑚 𝑚
∑ ∑ ⟨𝑇𝑣𝑘 , 𝑣𝑗 ⟩ ≥ 0.
𝑗 = 1 𝑘=1
15 Suppose 𝑇 ∈ ℒ(𝑉) is self-adjoint. Prove that there exist positive operators

𝐴, 𝐵 ∈ ℒ(𝑉) such that
𝑇 =𝐴−𝐵 and √𝑇 ∗ 𝑇 = 𝐴 + 𝐵 and 𝐴𝐵 = 𝐵𝐴 = 0.
16 Suppose 𝑇 is a positive operator on 𝑉. Prove that
null √𝑇 = null 𝑇 and range √𝑇 = range 𝑇.

17 Suppose that 𝑇 ∈ ℒ(𝑉) is a positive operator. Prove that there exists a
polynomial 𝑝 with real coefficients such that √𝑇 = 𝑝(𝑇).
18 Suppose 𝑆 and 𝑇 are positive operators on 𝑉. Prove that 𝑆𝑇 is a positive
operator if and only if 𝑆 and 𝑇 commute.
19 Show that the identity operator on 𝐅2 has infinitely many self-adjoint square
roots.
20 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉. Prove that 𝑇
is a positive operator if and only if there exist 𝑣1 , …, 𝑣𝑛 ∈ 𝑉 such that
⟨𝑇𝑒𝑘 , 𝑒𝑗 ⟩ = ⟨𝑣𝑘 , 𝑣𝑗 ⟩
for all 𝑗, 𝑘 = 1, …, 𝑛.
The numbers {⟨𝑇𝑒𝑘 , 𝑒𝑗 ⟩}𝑗, 𝑘 = 1, …, 𝑛 are the entries in the matrix of 𝑇 with
respect to the orthonormal basis 𝑒1 , …, 𝑒𝑛 .
21 Suppose 𝑛 is a positive integer. The 𝑛-by-𝑛 Hilbert matrix is the 𝑛-by-𝑛

matrix whose entry in row 𝑗, column 𝑘 is 𝑗 + 𝑘1− 1 . Suppose 𝑇 ∈ ℒ(𝑉) is an
operator whose matrix with respect to some orthonormal basis of 𝑉 is the
𝑛-by-𝑛 Hilbert matrix. Prove that 𝑇 is a positive invertible operator.
Example: The 4-by-4 Hilbert matrix is
1 1 1
1
⎛
⎜ 2 3 4 ⎞
⎟
⎜
⎜ 1 1 1 1 ⎟
⎟
⎜
⎜ ⎟
⎟
⎜
⎜
2 3 4 5 ⎟
⎟.
⎜
⎜ 1 1 1 1 ⎟
⎟
⎜
⎜ 3 4 5 6
⎟
⎟
⎜
⎜ ⎟
⎟
1 1 1 1
⎝ 4 5 6 7 ⎠
22 Suppose 𝑇 ∈ ℒ(𝑉) is a positive operator and 𝑢 ∈ 𝑉 is such that ‖𝑢‖ = 1

and ‖𝑇𝑢‖ ≥ ‖𝑇𝑣‖ for all 𝑣 ∈ 𝑉 with ‖𝑣‖ = 1. Show that 𝑢 is an eigenvector
of 𝑇 corresponding to the largest eigenvalue of 𝑇.
23 For 𝑇 ∈ ℒ(𝑉) and 𝑢, 𝑣 ∈ 𝑉, define ⟨𝑢, 𝑣⟩𝑇 by ⟨𝑢, 𝑣⟩𝑇 = ⟨𝑇𝑢, 𝑣⟩.
(a) Suppose 𝑇 ∈ ℒ(𝑉). Prove that ⟨⋅, ⋅⟩𝑇 is an inner product on 𝑉 if and
only if 𝑇 is an invertible positive operator (with respect to the original
inner product ⟨⋅, ⋅⟩).
(b) Prove that every inner product on 𝑉 is of the form ⟨⋅, ⋅⟩𝑇 for some positive
invertible operator 𝑇 ∈ ℒ(𝑉).
24 Suppose 𝑆 and 𝑇 are positive operators on 𝑉. Prove that
null(𝑆 + 𝑇) = null 𝑆 ∩ null 𝑇.
25 Let 𝑇 be the second derivative operator in Exercise 31(b) in Section 7A.

Show that −𝑇 is a positive operator.
7D Isometries, Unitary Operators, and Matrix Factorization

Isometries
Linear maps that preserve norms are sufficiently important to deserve a name.
7.44 definition: isometry
A linear map 𝑆 ∈ ℒ(𝑉, 𝑊) is called an isometry if
‖𝑆𝑣‖ = ‖𝑣‖
for every 𝑣 ∈ 𝑉. In other words, a linear map is an isometry if it preserves

norms.
If 𝑆 ∈ ℒ(𝑉, 𝑊) is an isometry and The Greek word isos means equal; the
𝑣 ∈ 𝑉 is such that 𝑆𝑣 = 0, then Greek word metron means measure.
‖𝑣‖ = ‖𝑆𝑣‖ = ‖0‖ = 0, Thus isometry literally means equal
measure.
which implies that 𝑣 = 0. Thus every
isometry is injective.
7.45 example: orthonormal basis maps to orthonormal list ⟹ isometry

Suppose 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉 and 𝑔1 , …, 𝑔𝑛 is an orthonormal
list in 𝑊. Let 𝑆 ∈ ℒ(𝑉, 𝑊) be the linear map such that 𝑆𝑒𝑘 = 𝑔𝑘 for each
𝑘 = 1, …, 𝑛. To show that 𝑆 is an isometry, suppose 𝑣 ∈ 𝑉. Then
7.46 𝑣 = ⟨𝑣, 𝑒1 ⟩𝑒1 + ⋯ + ⟨𝑣, 𝑒𝑛 ⟩𝑒𝑛
and
2 2
7.47 ‖𝑣‖2 = ∣⟨𝑣, 𝑒1 ⟩∣ + ⋯ + ∣⟨𝑣, 𝑒𝑛 ⟩∣ ,
where we have used 6.30(b). Applying 𝑆 to both sides of 7.46 gives
𝑆𝑣 = ⟨𝑣, 𝑒1 ⟩𝑆𝑒1 + ⋯ + ⟨𝑣, 𝑒𝑛 ⟩𝑆𝑒𝑛 = ⟨𝑣, 𝑒1 ⟩𝑔1 + ⋯ + ⟨𝑣, 𝑒𝑛 ⟩𝑔𝑛 .
Thus
2
7.48 ‖𝑆𝑣‖2 = ∣⟨𝑣, 𝑒1 ⟩∣ + ⋯ + |⟨𝑣, 𝑒𝑛 ⟩|2.
Comparing 7.47 and 7.48 shows that ‖𝑣‖ = ‖𝑆𝑣‖. Thus 𝑆 is an isometry.
The next result gives conditions equivalent to being an isometry. The equiv-
alence of (a) and (c) shows that a linear map is an isometry if and only if it
preserves inner products. The equivalence of (a) and (d) shows that a linear map
is an isometry if and only if it maps some orthonormal basis to an orthonormal list.
Thus the isometries given by Example 7.45 include all isometries. Furthermore,
a linear map is an isometry if and only if it maps every orthonormal basis to an
orthonormal list [because whether or not (a) holds does not depend on the basis
𝑒1 , …, 𝑒𝑛 ].
Section 7D Isometries, Unitary Operators, and Matrix Factorization 259
The equivalence of (a) and (e) in the next result shows that a linear map is an
isometry if and only if the columns of its matrix (with respect to any orthonormal
bases) form an orthonormal list. Here we are identifying the columns of an 𝑚-by-𝑛
matrix with elements of 𝐅𝑚 and then using the Euclidean inner product on 𝐅𝑚 .
7.49 characterization of isometries
Suppose 𝑆 ∈ ℒ(𝑉, 𝑊). Suppose 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉 and

𝑓1 , …, 𝑓𝑚 is an orthonormal basis of 𝑊. Then the following are equivalent.
(a) 𝑆 is an isometry.
(b) 𝑆∗ 𝑆 = 𝐼.
(c) ⟨𝑆𝑢, 𝑆𝑣⟩ = ⟨𝑢, 𝑣⟩ for all 𝑢, 𝑣 ∈ 𝑉.
(d) 𝑆𝑒1 , …, 𝑆𝑒𝑛 is an orthonormal list in 𝑊.
(e) The columns of ℳ(𝑆, (𝑒1 , …, 𝑒𝑛 ), ( 𝑓1 , …, 𝑓𝑚 )) form an orthonormal list
in 𝐅𝑚 with respect to the Euclidean inner product.
Proof First suppose (a) holds, so 𝑆 is an isometry. If 𝑣 ∈ 𝑉 then

⟨(𝐼 − 𝑆∗ 𝑆)𝑣, 𝑣⟩ = ⟨𝑣, 𝑣⟩ − ⟨𝑆∗ 𝑆𝑣, 𝑣⟩ = ‖𝑣‖2 − ⟨𝑆𝑣, 𝑆𝑣⟩ = ‖𝑣‖2 − ‖𝑆𝑣‖2 = 0.
Hence the self-adjoint operator 𝐼 − 𝑆∗ 𝑆 equals 0 (by 7.16). Thus 𝑆∗ 𝑆 = 𝐼, proving
that (a) implies (b).
Now suppose (b) holds, so 𝑆∗ 𝑆 = 𝐼. If 𝑢, 𝑣 ∈ 𝑉 then
⟨𝑆𝑢, 𝑆𝑣⟩ = ⟨𝑆∗ 𝑆𝑢, 𝑣⟩ = ⟨𝐼𝑢, 𝑣⟩ = ⟨𝑢, 𝑣⟩,
proving that (b) implies (c).
Now suppose that (c) holds, so ⟨𝑆𝑢, 𝑆𝑣⟩ = ⟨𝑢, 𝑣⟩ for all 𝑢, 𝑣 ∈ 𝑉. Thus if
𝑗, 𝑘 ∈ {1, …, 𝑛}, then
⟨𝑆𝑒𝑗 , 𝑆𝑒𝑘 ⟩ = ⟨𝑒𝑗 , 𝑒𝑘 ⟩.
Hence 𝑆𝑒1 , …, 𝑆𝑒𝑛 is an orthonormal list in 𝑊, proving that (c) implies (d).
Now suppose that (d) holds, so 𝑆𝑒1 , …, 𝑆𝑒𝑛 is an orthonormal list in 𝑊. Let
𝐴 = ℳ(𝑆, (𝑒1 , …, 𝑒𝑛 ), ( 𝑓1 , …, 𝑓𝑚 )). If 𝑘, 𝑟 ∈ {1, …, 𝑛}, then
𝑚 𝑚 𝑚 ⎧
{1 if 𝑘 = 𝑟,
7.50 ∑ 𝐴𝑗, 𝑘 𝐴𝑗, 𝑟 = ⟨ ∑ 𝐴𝑗, 𝑘 𝑓𝑗 , ∑ 𝐴𝑗, 𝑟 𝑓𝑗 ⟩ = ⟨𝑆𝑒𝑘 , 𝑆𝑒𝑟 ⟩ = ⎨
𝑗=1 𝑗=1 𝑗=1 {
⎩0 if 𝑘 ≠ 𝑟.
The left side of 7.50 is the inner product in 𝐅𝑚 of columns 𝑘 and 𝑟 of 𝐴. Thus the
columns of 𝐴 form an orthonormal list in 𝐅𝑚 , proving that (d) implies (e).
Now suppose (e) holds, so the columns of the matrix 𝐴 defined in the paragraph
above form an orthonormal list in 𝐅𝑛 . Then 7.50 shows that 𝑆𝑒1 , …, 𝑆𝑒𝑛 is an
orthonormal list in 𝑊. Thus Example 7.45, with 𝑆𝑒1 , …, 𝑆𝑒𝑛 playing the role of
𝑔1 , …, 𝑔𝑛 , shows that 𝑆 is an isometry, proving that (e) implies (a).
See Exercises 1 and 11 for additional conditions that are equivalent to being
an isometry.
Unitary Operators
In this subsection, we confine our attention to linear maps from a vector space to
itself. In other words, we will be working with operators.
7.51 definition: unitary operator
An operator 𝑆 ∈ ℒ(𝑉) is called unitary if 𝑆 is an invertible isometry.
As previously noted, every isometry Although the words “unitary” and

is injective. Every injective operator on “isometry” mean the same thing for
a finite-dimensional vector space is in- operators on finite-dimensional inner
vertible (see 3.65). A standing assump- product spaces, remember that a uni-
tion for this chapter is that 𝑉 is a finite- tary operator maps a vector space to
dimensional inner product space. Thus itself, while an isometry maps a vector
we could delete the word “invertible” space to another (possibly different)
from the definition above without chang- vector space.
ing the meaning. The unnecessary word
“invertible” has been retained in the definition above for consistency with the
definition readers may encounter when learning about inner product spaces that
are not necessarily finite-dimensional.
7.52 example: rotation of 𝐑2

Suppose 𝜃 ∈ 𝐑 and 𝑆 is the operator on 𝐅2 whose matrix with respect to the
standard basis of 𝐅2 is
cos 𝜃 − sin 𝜃
( ).
sin 𝜃 cos 𝜃
The two columns of this matrix form an orthonormal list in 𝐅2 ; hence 𝑆 is an
isometry [by the equivalence of (a) and (e) in 7.49]. Thus 𝑆 is a unitary operator.
If 𝐅 = 𝐑 , then 𝑆 is the operator of counterclockwise rotation by 𝜃 radians
around the origin of 𝐑2 . This observation gives us another way to think about why
𝑆 is an isometry, because each rotation around the origin of 𝐑2 preserves norms.
The next result (7.53) lists several conditions that are equivalent to being a
unitary operator. All the conditions equivalent to being an isometry in 7.49 should
be added to this list. The extra conditions in 7.53 arise because of limiting the
context to linear maps from a vector space to itself. For example, 7.49 shows that
a linear map 𝑆 ∈ ℒ(𝑉, 𝑊) is an isometry if and only if 𝑆∗ 𝑆 = 𝐼, while 7.53 shows
that an operator 𝑆 ∈ ℒ(𝑉) is a unitary operator if and only if 𝑆∗ 𝑆 = 𝑆𝑆∗ = 𝐼.
Another difference is that 7.49(d) mentions an orthonormal list, while 7.53(d)
mentions an orthonormal basis. Also, 7.49(e) mentions the columns of ℳ(𝑇),
while 7.53(e) mentions the rows of ℳ(𝑇). Furthermore, ℳ(𝑇) in 7.49(e) is with
respect to an orthonormal basis of 𝑉 and an orthonormal basis of 𝑊, while ℳ(𝑇)
in 7.53(e) is with respect to a single basis of 𝑉 doing double duty.
7.53 characterization of unitary operators
Suppose 𝑆 ∈ ℒ(𝑉). Suppose 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉. Then the

following are equivalent.
(a) 𝑆 is a unitary operator.
(b) 𝑆∗ 𝑆 = 𝑆𝑆∗ = 𝐼.
(c) 𝑆 is invertible and 𝑆−1 = 𝑆∗.
(d) 𝑆𝑒1 , …, 𝑆𝑒𝑛 is an orthonormal basis of 𝑉.
(e) The rows of ℳ(𝑆, (𝑒1 , …, 𝑒𝑛 )) form an orthonormal basis of 𝐅𝑛 with
respect to the Euclidean inner product.
(f) 𝑆∗ is a unitary operator.
Proof First suppose (a) holds, so 𝑆 is a unitary operator. Hence
𝑆∗ 𝑆 = 𝐼
by the equivalence of (a) and (b) in 7.49. Multiply both sides of this equation by
𝑆−1 on the right, getting 𝑆∗ = 𝑆−1 . Thus 𝑆𝑆∗ = 𝑆𝑆−1 = 𝐼, as desired, proving
The definitions of invertible and inverse show that (b) implies (c).
Now suppose (c) holds, so 𝑆 is invertible and 𝑆−1 = 𝑆∗. Thus 𝑆∗ 𝑆 = 𝐼. Hence
𝑆𝑒1 , …, 𝑆𝑒𝑛 is an orthonormal list in 𝑉, by the equivalence of (b) and (d) in 7.49.
The length of this list equals dim 𝑉. Thus 𝑆𝑒1 , …, 𝑆𝑒𝑛 is an orthonormal basis of 𝑉,
proving that (c) implies (d).
Now suppose (d) holds, so 𝑆𝑒1 , …, 𝑆𝑒𝑛 is an orthonormal basis of 𝑉. The
equivalence of (a) and (d) in 7.49 shows that 𝑆 is a unitary operator. Thus
∗
(𝑆∗ ) 𝑆∗ = 𝑆𝑆∗ = 𝐼,
where the last equation holds because we have already shown that (a) implies (b) in
this result. The equation above and the equivalence of (a) and (b) in 7.49 show that
𝑆∗ is an isometry. Thus the columns of ℳ(𝑆∗, (𝑒1 , …, 𝑒𝑛 )) form an orthonormal ba-
sis of 𝐅𝑛 [by the equivalence of (a) and (e) of 7.49]. The rows of ℳ(𝑆, (𝑒1 , …, 𝑒𝑛 ))
are the complex conjugates of the columns of ℳ(𝑆∗, (𝑒1 , …, 𝑒𝑛 )). Thus the rows
of ℳ(𝑆, (𝑒1 , …, 𝑒𝑛 )) form an orthonormal basis of 𝐅𝑛 , proving that (d) implies (e).
Now suppose (e) holds. Thus the columns of ℳ(𝑆∗, (𝑒1 , …, 𝑒𝑛 )) form an
orthonormal basis of 𝐅𝑛 . The equivalence of (a) and (e) in 7.49 shows that 𝑆∗ is
an isometry, proving that (e) implies (f).
Now suppose (f) holds, so 𝑆∗ is a unitary operator. The chain of implications
we have already proved in this result shows that (a) implies (f). Applying this
result to 𝑆∗ shows that (𝑆∗ )∗ is a unitary operator, proving that (f) implies (a).
We have shown that (a) ⇒ (b) ⇒ (c) ⇒ (d) ⇒ (e) ⇒ (f) ⇒ (a), completing the
proof.
Recall our analogy between 𝐂 and ℒ(𝑉). Under this analogy, a complex
number 𝑧 corresponds to an operator 𝑆 ∈ ℒ(𝑉), and 𝑧 corresponds to 𝑆∗. The
real numbers (𝑧 = 𝑧) correspond to the self-adjoint operators (𝑆 = 𝑆∗ ), and the
nonnegative numbers correspond to the (badly named) positive operators.
Another distinguished subset of 𝐂 is the unit circle, which consists of the
complex numbers 𝑧 such that |𝑧| = 1. The condition |𝑧| = 1 is equivalent to the
condition 𝑧𝑧 = 1. Under our analogy, this corresponds to the condition 𝑆∗ 𝑆 = 𝐼,
which is equivalent to 𝑆 being a unitary operator. Hence the analogy shows that
the unit circle in 𝐂 corresponds to the set of unitary operators. In the next two
results, this analogy appears in the eigenvalues of unitary operators. Also see
Exercise 15 for another example of this analogy.
7.54 eigenvalues of unitary operators have absolute value 1
Suppose 𝜆 is an eigenvalue of a unitary operator. Then |𝜆| = 1.
Proof Suppose 𝑆 ∈ ℒ(𝑉) is a unitary operator and 𝜆 is an eigenvalue of 𝑆. Let

𝑣 ∈ 𝑉 be such that 𝑣 ≠ 0 and 𝑆𝑣 = 𝜆𝑣. Then
|𝜆| ‖𝑣‖ = ‖𝜆𝑣‖ = ‖𝑆𝑣‖ = ‖𝑣‖.
Thus |𝜆| = 1, as desired.
The next result characterizes unitary operators on finite-dimensional complex

inner product spaces, using the complex spectral theorem as the main tool.
7.55 description of unitary operators on complex inner product spaces
Suppose 𝐅 = 𝐂 and 𝑆 ∈ ℒ(𝑉). Then the following are equivalent.

(a) 𝑆 is a unitary operator.
(b) There is an orthonormal basis of 𝑉 consisting of eigenvectors of 𝑆 whose
corresponding eigenvalues all have absolute value 1.
Proof Suppose (a) holds, so 𝑆 is a unitary operator. The equivalence of (a) and
(b) in 7.53 shows that 𝑆 is normal. Thus the complex spectral theorem (7.31)
shows that there is an orthonormal basis 𝑒1 , …, 𝑒𝑛 of 𝑉 consisting of eigenvectors
of 𝑆. Every eigenvalue of 𝑆 has absolute value 1 (by 7.54), completing the proof
Now suppose (b) holds. Let 𝑒1 , …, 𝑒𝑛 be an orthonormal basis of 𝑉 consisting
of eigenvectors of 𝑆 whose corresponding eigenvalues 𝜆1 , …, 𝜆𝑛 all have absolute
value 1. Then 𝑆𝑒1 , …, 𝑆𝑒𝑛 is also an orthonormal basis of 𝑉 because
⎧
{0 if 𝑗 ≠ 𝑘,
⟨𝑆𝑒𝑗 , 𝑆𝑒𝑘 ⟩ = ⟨𝜆𝑗 𝑒𝑗 , 𝜆𝑘 𝑒𝑘 ⟩ = 𝜆𝑗 𝜆𝑘 ⟨𝑒𝑗 , 𝑒𝑘 ⟩ = ⎨
{
⎩1 if 𝑗 = 𝑘
for all 𝑗, 𝑘 = 1, …, 𝑛. Thus the equivalence of (a) and (d) in 7.53 shows that 𝑆 is
unitary, proving that (b) implies (a).
QR Factorization
In this subsection, we shift our attention from operators to matrices. This switch
should give you good practice in identifying an operator with a square matrix
(after picking a basis of the vector space on which the operator is defined). You
should also become more comfortable with translating concepts and results back
and forth between the context of operators and the context of square matrices.
When starting with 𝑛-by-𝑛 matrices instead of operators, unless otherwise
specified assume that the associated operators live on 𝐅𝑛 (with the Euclidean inner
product) and that their matrices are computed with respect to the standard basis
of 𝐅𝑛 .
We begin by making the following definition, transferring the notion of a
unitary operator to a unitary matrix.
7.56 definition: unitary matrix
An 𝑛-by-𝑛 matrix is called unitary if its columns form an orthonormal list

in 𝐅𝑛 .
In the definition above, we could have replaced “orthonormal list in 𝐅𝑛 ” with

“orthonormal basis of 𝐅𝑛 ” because every orthonormal list of length 𝑛 in an 𝑛-
dimensional inner product space is an orthonormal basis. If 𝑆 ∈ ℒ(𝑉) and
𝑒1 , …, 𝑒𝑛 and 𝑓1 , …, 𝑓𝑛 are orthonormal bases of 𝑉, then 𝑆 is a unitary operator
if and only if ℳ(𝑆, (𝑒1 , …, 𝑒𝑛 ), ( 𝑓1 , …, 𝑓𝑛 )) is a unitary matrix, as shown by the
equivalence of (a) and (e) in 7.49. Also note that we could also have replaced
“columns” in the definition above with “rows” by using the equivalence between
conditions (a) and (e) in 7.53.
The next result, whose proof will be left as an exercise for the reader, gives
some equivalent conditions for a square matrix to be unitary. In (c), 𝑄𝑣 denotes
the matrix product of 𝑄 and 𝑣, identifying elements of 𝐅𝑛 with 𝑛-by-1 matrices
(sometimes called column vectors). The norm in (c) below is the usual Euclidean
norm on 𝐅𝑛 that comes from the Euclidean inner product. In (d), 𝑄∗ denotes
the conjugate transpose of the matrix 𝑄, which corresponds to the adjoint of the
associated operator.
7.57 characterizations of unitary matrices
Suppose 𝑄 is an 𝑛-by-𝑛 matrix. Then the following are equivalent.

(a) 𝑄 is a unitary matrix.
(b) The rows of 𝑄 form an orthonormal list in 𝐅𝑛.
(c) ‖𝑄𝑣‖ = ‖𝑣‖ for every 𝑣 ∈ 𝐅𝑛.
(d) 𝑄∗ 𝑄 = 𝑄𝑄∗ = 𝐼, the 𝑛-by-𝑛 matrix with 1’s on the diagonal and 0’s
elsewhere.
The QR factorization stated and proved below is the main tool in the widely
used QR algorithm (not discussed here) for finding good approximations to
eigenvalues and eigenvectors of square matrices. In the result below, if the matrix
𝐴 is in 𝐅𝑛, 𝑛, then the matrices 𝑄 and 𝑅 are also in 𝐅𝑛, 𝑛.
7.58 QR factorization
Suppose 𝐴 is a square matrix with linearly independent columns. Then there

exist unique matrices 𝑄 and 𝑅 such that 𝑄 is unitary, 𝑅 is upper triangular
with only positive numbers on its diagonal, and
𝐴 = 𝑄𝑅.
Proof Let 𝑣1 , …, 𝑣𝑛 denote the columns of 𝐴, thought of as elements of 𝐅𝑛 . Apply

the Gram–Schmidt procedure (6.32) to the list 𝑣1 , …, 𝑣𝑛 , getting an orthonormal
basis 𝑒1 , …, 𝑒𝑛 of 𝐅𝑛 such that
7.59 span(𝑣1 , …, 𝑣𝑘 ) = span(𝑒1 , …, 𝑒𝑘 )
for each 𝑘 = 1, …, 𝑛. Let 𝑅 be the 𝑛-by-𝑛 matrix defined by
𝑅𝑗, 𝑘 = ⟨𝑣𝑘 , 𝑒𝑗 ⟩,
where 𝑅𝑗, 𝑘 denotes the entry in row 𝑗, column 𝑘 of 𝑅. If 𝑗 > 𝑘, then 𝑒𝑗 is orthogonal
to span(𝑒1 , …, 𝑒𝑘 ) and hence 𝑒𝑗 is orthogonal to 𝑣𝑘 (by 7.59). In other words, if
𝑗 > 𝑘 then ⟨𝑣𝑘 , 𝑒𝑗 ⟩ = 0. Thus 𝑅 is an upper-triangular matrix.
Let 𝑄 be the unitary matrix whose columns are 𝑒1 , …, 𝑒𝑛 . If 𝑘 ∈ {1, …, 𝑛},
then the 𝑘 th column of 𝑄𝑅 equals a linear combination of the columns of 𝑄, with
the coefficients for the linear combination coming from the 𝑘 th column of 𝑅—see
3.51(a). Hence the 𝑘 th column of 𝑄𝑅 equals
⟨𝑣𝑘 , 𝑒1 ⟩𝑒1 + ⋯ + ⟨𝑣𝑘 , 𝑒𝑘 ⟩𝑒𝑘 ,
which equals 𝑣𝑘 [by 6.30(a)], the 𝑘 th column of 𝐴. Thus 𝐴 = 𝑄𝑅, as desired.

The equations defining the Gram–Schmidt procedure (see 6.32) show that
each 𝑣𝑘 equals a positive multiple of 𝑒𝑘 plus a linear combination of 𝑒1 , …, 𝑒𝑘 − 1 .
Thus each ⟨𝑣𝑘 , 𝑒𝑘 ⟩ is a positive number. Hence all entries on the diagonal of 𝑅 are
positive numbers, as desired.
Finally, to show that 𝑄 and 𝑅 are unique, suppose we also have 𝐴 = 𝑄 ̂𝑅̂, where
̂ is unitary and 𝑅
𝑄 ̂ is upper triangular with only positive numbers on its diagonal.
Let 𝑞1 , …, 𝑞𝑛 denote the columns of 𝑄 ̂. Thinking of matrix multiplication as above,
we see that each 𝑣𝑘 is a linear combination of 𝑞1 , …, 𝑞𝑘 , with the coefficients coming
from the 𝑘 th column of 𝑅 ̂. This implies that span(𝑣 , …, 𝑣 ) = span(𝑞 , …, 𝑞 ) and
1 𝑘 1 𝑘
⟨𝑣𝑘 , 𝑞𝑘 ⟩ > 0. The uniqueness of the orthonormal lists satisfying these conditions
(see Exercise 10 in Section 6B) now shows that 𝑞𝑘 = 𝑒𝑘 for each 𝑘 = 1, …, 𝑛. Hence
̂ = 𝑄, which then implies that 𝑅
𝑄 ̂ = 𝑅, completing the proof of uniqueness.
The proof of the QR factorization shows that the columns of the unitary matrix
can be computed by applying the Gram–Schmidt procedure to the columns of the
matrix to be factored. The next example illustrates the computation of the QR
factorization based on the proof that we just completed.
7.60 example: QR factorization of a 3-by-3 matrix

To find the QR factorization of the matrix
1 2 1
⎛ ⎞
𝐴=⎜ ⎜
⎜ 0 1 −4 ⎟
⎟,
⎟
⎝ 0 3 2 ⎠
follow the proof of 7.58. Thus set 𝑣1 , 𝑣2 , 𝑣3 equal to the columns of 𝐴:
𝑣1 = (1, 0, 0), 𝑣2 = (2, 1, 3), 𝑣3 = (1, −4, 2).
Apply the Gram–Schmidt procedure to 𝑣1 , 𝑣2 , 𝑣3 , producing the orthonormal list
𝑒1 = (1, 0, 0), 𝑒2 = (0, √1 , √3 ), 𝑒3 = (0, − √3 , √1 ).
10 10 10 10
Still following the proof of 7.58, let 𝑄 be the unitary matrix whose columns are
𝑒1 , 𝑒2 , 𝑒3 :
1 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ 0 1
− 3 ⎟
⎟
𝑄=⎜ ⎜ √10 √10 ⎟⎟ .
⎜
⎜ ⎟
⎟
⎜ 3 1 ⎟
⎝ 0 √ 10 √ ⎠10
As in the proof of 7.58, let 𝑅 be the 3-by-3 matrix whose entry in row 𝑗, column 𝑘
is ⟨𝑣𝑘 , 𝑒𝑗 ⟩, which gives
1 2 1
⎛
⎜ ⎞
⎟
⎜
⎜ 0 √10 ⎟
⎟
√10
𝑅=⎜
⎜
⎜ 5
⎟
⎟
⎟.
⎜
⎜ ⎟
⎟
7√10
⎝ 0 0 5 ⎠
Note that 𝑅 is indeed an upper-triangular matrix with only positive numbers on
the diagonal, as required by the QR factorization.
Now matrix multiplication can verify that 𝐴 = 𝑄𝑅 is the desired factorization
of 𝐴:
1 0 0 1 2 1
⎛
⎜ ⎞
⎟ ⎛
⎜ ⎞
⎟ 1 2 1
⎜
⎜ 0 1
− 3 ⎟
⎟ ⎜
⎜ √ 10 ⎟
⎟ ⎛ ⎞
𝑄𝑅 = ⎜⎜ ⎟
√10 ⎟⎜ ⎜ 0 √ 10 ⎟ ⎜
= ⎜ 0 1 −4 ⎟
⎜ ⎟ = 𝐴.
⎜ √10
⎟ ⎜ 5 ⎟
⎟ ⎟
⎜
⎜ ⎟⎜
⎟ ⎜ ⎟
⎟ 0 3 2
3 1 √ ⎝ ⎠
⎝ 0 √10
7 10
√10 ⎠⎝ 0 0 5 ⎠
Thus 𝐴 = 𝑄𝑅, as expected.
The QR factorization will be the major tool used in the proof of the Cholesky
factorization (7.63) in the next subsection. For another nice application of the QR
factorization, see the proof of Hadamard’s inequality (9.66).
If a QR factorization is available, then it can be used to solve a corresponding

system of linear equations without using Gaussian elimination. Specifically,
suppose 𝐴 is an 𝑛-by-𝑛 square matrix with linearly independent columns. Suppose
that 𝑏 ∈ 𝐅𝑛 and we want to solve the equation 𝐴𝑥 = 𝑏 for 𝑥 = (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛
(as usual, we are identifying elements of 𝐅𝑛 with 𝑛-by-1 column vectors).
Suppose 𝐴 = 𝑄𝑅, where 𝑄 is unitary and 𝑅 is upper triangular with only
positive numbers on its diagonal (𝑄 and 𝑅 are computable from 𝐴 using just the
Gram–Schmidt procedure, as shown in the proof of 7.58). The equation 𝐴𝑥 = 𝑏 is
equivalent to the equation 𝑄𝑅𝑥 = 𝑏. Multiplying both sides of this last equation
by 𝑄∗ on the left and using 7.57(d) gives the equation
𝑅𝑥 = 𝑄∗ 𝑏.
The matrix 𝑄∗ is the conjugate transpose of the matrix 𝑄. Thus computing

𝑄 𝑏 is straightforward. Because 𝑅 is an upper-triangular matrix with positive
∗
numbers on its diagonal, the system of linear equations represented by the equation
above can quickly be solved by first solving for 𝑥𝑛 , then for 𝑥𝑛 − 1 , and so on.
Cholesky Factorization
We begin this subsection with a characterization of positive invertible operators
in terms of inner products.
7.61 positive invertible operator
A self-adjoint operator 𝑇 ∈ ℒ(𝑉) is a positive invertible operator if and only

if ⟨𝑇𝑣, 𝑣⟩ > 0 for every nonzero 𝑣 ∈ 𝑉.
Proof First suppose 𝑇 is a positive invertible operator. If 𝑣 ∈ 𝑉 and 𝑣 ≠ 0, then

because 𝑇 is invertible we have 𝑇𝑣 ≠ 0. This implies that ⟨𝑇𝑣, 𝑣⟩ ≠ 0 (by 7.43).
Hence ⟨𝑇𝑣, 𝑣⟩ > 0.
To prove the implication in the other direction, suppose now that ⟨𝑇𝑣, 𝑣⟩ > 0
for every nonzero 𝑣 ∈ 𝑉. Thus 𝑇𝑣 ≠ 0 for every nonzero 𝑣 ∈ 𝑉. Hence 𝑇 is
injective. Thus 𝑇 is invertible, as desired.
The next definition transfers the result above to the language of matrices. Here
we are using the usual Euclidean inner product on 𝐅𝑛 and identifying elements of
𝐅𝑛 with 𝑛-by-1 column vectors.
7.62 definition: positive definite
A matrix 𝐵 ∈ 𝐅𝑛, 𝑛 is called positive definite if 𝐵∗ = 𝐵 and
⟨𝐵𝑥, 𝑥⟩ > 0
for every nonzero 𝑥 ∈ 𝐅𝑛.

A matrix is upper triangular if and only if its conjugate transpose is lower

triangular (meaning that all entries above the diagonal are 0). The factorization
below, which has important consequences in computational linear algebra, writes
a positive definite matrix as the product of a lower triangular matrix and its
conjugate transpose.
Our next result is solely about matrices, although the proof makes use of the
identification of results about operators with results about square matrices. In the
result below, if the matrix 𝐵 is in 𝐅𝑛, 𝑛, then the matrix 𝑅 is also in 𝐅𝑛, 𝑛.
7.63 Cholesky factorization
Suppose 𝐵 is a positive definite matrix. Then there exists a unique upper-

triangular matrix 𝑅 with only positive numbers on its diagonal such that
𝐵 = 𝑅∗ 𝑅.
Proof Because 𝐵 is positive definite, there exists an invertible square matrix 𝐴

of the same size as 𝐵 such that 𝐵 = 𝐴∗ 𝐴 [by the equivalence of (a) and (f) in
7.38].
Let 𝐴 = 𝑄𝑅 be the QR factorization of 𝐴 (see 7.58), where 𝑄 is unitary and 𝑅
is upper triangular with only positive numbers on its diagonal. Then 𝐴∗ = 𝑅∗ 𝑄∗.
Thus André-Louis Cholesky (1875–1918)
discovered this factorization, which
𝐵 = 𝐴∗ 𝐴 = 𝑅∗ 𝑄∗ 𝑄𝑅 = 𝑅∗ 𝑅,
was published posthumously in 1924.
as desired.
To prove the uniqueness part of this result, suppose 𝑆 is an upper-triangular
matrix with only positive numbers on its diagonal and 𝐵 = 𝑆∗ 𝑆. The matrix 𝑆 is
invertible because 𝐵 is invertible (see Exercise 11 in Section 3D). Multiplying both
sides of the equation 𝐵 = 𝑆∗ 𝑆 by 𝑆−1 on the left gives the equation 𝐵𝑆−1 = 𝑆∗.
Let 𝐴 be the matrix from the first paragraph of this proof. Then
∗
(𝐴𝑆−1 ) (𝐴𝑆−1 ) = (𝑆∗ )−1 𝐴∗ 𝐴𝑆−1
= (𝑆∗ )−1 𝐵𝑆−1
= (𝑆∗ )−1 𝑆∗
= 𝐼.
Thus 𝐴𝑆−1 is unitary.

Hence 𝐴 = (𝐴𝑆−1 )𝑆 is a factorization of 𝐴 as the product of a unitary matrix
and an upper-triangular matrix with only positive numbers on its diagonal. The
uniqueness of the QR factorization, as stated in 7.58, now implies that 𝑆 = 𝑅.
In the first paragraph of the proof above, we could have chosen 𝐴 to be the
unique positive definite matrix that is a square root of 𝐵 (see 7.39). However,
the proof was presented with the more general choice of 𝐴 because for specific
positive definite matrices 𝐵, it may be easier to find a different choice of 𝐴.
Exercises 7D
1 Suppose dim 𝑉 ≥ 2 and 𝑆 ∈ ℒ(𝑉, 𝑊). Prove that 𝑆 is an isometry if and

only if 𝑆𝑒1 , 𝑆𝑒2 is an orthonormal list in 𝑊 for every orthonormal list 𝑒1 , 𝑒2
of length two in 𝑉.
2 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that 𝑇 is a scalar multiple of an isometry if
and only if 𝑇 preserves orthogonality.
The phrase “𝑇 preserves orthogonality” means that ⟨𝑇𝑢, 𝑇𝑣⟩ = 0 for all
𝑢, 𝑣 ∈ 𝑉 such that ⟨𝑢, 𝑣⟩ = 0.
3 (a) Show that the product of two unitary operators on 𝑉 is a unitary operator.
(b) Show that the inverse of a unitary operator on 𝑉 is a unitary operator.
This exercise shows that the set of unitary operators on 𝑉 is a group, where
the group operation is the usual product of two operators.
4 Suppose 𝐅 = 𝐂 and 𝐴, 𝐵 ∈ ℒ(𝑉) are self-adjoint. Show that 𝐴 + 𝑖𝐵 is

unitary if and only if 𝐴𝐵 = 𝐵𝐴 and 𝐴2 + 𝐵2 = 𝐼.
5 Suppose 𝑆 ∈ ℒ(𝑉). Prove that the following are equivalent.
(a) 𝑆 is a self-adjoint unitary operator.
(b) 𝑆 = 2𝑃 − 𝐼 for some orthogonal projection 𝑃 on 𝑉.
(c) There exists a subspace 𝑈 of 𝑉 such that 𝑆𝑢 = 𝑢 for every 𝑢 ∈ 𝑈 and
𝑆𝑤 = −𝑤 for every 𝑤 ∈ 𝑈 ⟂.
6 Suppose 𝑇1 , 𝑇2 are both normal operators on 𝐅3 with 2, 5, 7 as eigenvalues.
Prove that there exists a unitary operator 𝑆 ∈ ℒ(𝐅3 ) such that 𝑇1 = 𝑆∗ 𝑇2 𝑆.
7 Give an example of two self-adjoint operators 𝑇1 , 𝑇2 ∈ ℒ(𝐅4 ) such that the
eigenvalues of both operators are 2, 5, 7 but there does not exist a unitary
operator 𝑆 ∈ ℒ(𝐅4 ) such that 𝑇1 = 𝑆∗ 𝑇2 𝑆. Be sure to explain why there is
no unitary operator with the required property.
8 Prove or give a counterexample: If 𝑆 ∈ ℒ(𝑉) and there exists an orthonormal
basis 𝑒1 , …, 𝑒𝑛 of 𝑉 such that ‖𝑆𝑒𝑘 ‖ = 1 for each 𝑒𝑘 , then 𝑆 is a unitary operator.
9 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Suppose every eigenvalue of 𝑇 has absolute
value 1 and ‖𝑇𝑣‖ ≤ ‖𝑣‖ for every 𝑣 ∈ 𝑉. Prove that 𝑇 is a unitary operator.
10 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉) is a self-adjoint operator such that ‖𝑇𝑣‖ ≤ ‖𝑣‖
(a) Show that 𝐼 − 𝑇 2 is a positive operator.
(b) Show that 𝑇 + 𝑖√𝐼 − 𝑇 2 is a unitary operator.
11 Suppose 𝑆 ∈ ℒ(𝑉). Prove that 𝑆 is a unitary operator if and only if
{𝑆𝑣 ∶ 𝑣 ∈ 𝑉 and ‖𝑣‖ ≤ 1} = {𝑣 ∈ 𝑉 ∶ ‖𝑣‖ ≤ 1}.
12 Prove or give a counterexample: If 𝑆 ∈ ℒ(𝑉) is invertible and ∥𝑆−1 𝑣∥ = ‖𝑆𝑣‖
for every 𝑣 ∈ 𝑉, then 𝑆 is unitary.
13 Explain why the columns of a square matrix of complex numbers form

an orthonormal list in 𝐂𝑛 if and only if the rows of the matrix form an
orthonormal list in 𝐂𝑛 .
14 Suppose 𝑣 ∈ 𝑉 with ‖𝑣‖ = 1 and 𝑏 ∈ 𝐅. Also suppose dim 𝑉 ≥ 2. Prove
that there exists a unitary operator 𝑆 ∈ ℒ(𝑉) such that ⟨𝑆𝑣, 𝑣⟩ = 𝑏 if and
only if |𝑏| ≤ 1.
15 Suppose 𝑇 is a unitary operator on 𝑉 such that 𝑇 − 𝐼 is invertible.
(a) Prove that (𝑇 + 𝐼)(𝑇 − 𝐼)−1 is a skew operator (meaning that it equals
the negative of its adjoint).
(b) Prove that if 𝐅 = 𝐂 , then 𝑖(𝑇 + 𝐼)(𝑇 − 𝐼)−1 is a self-adjoint operator.
The function 𝑧 ↦ 𝑖(𝑧 + 1)(𝑧 − 1)−1 maps the unit circle in 𝐂 (except for the
point 1) to 𝐑 . Thus (b) illustrates the analogy between the unitary operators
and the unit circle in 𝐂 , along with the analogy between the self-adjoint
operators and 𝐑 .
16 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉) is self-adjoint. Prove that (𝑇 + 𝑖𝐼)(𝑇 − 𝑖𝐼)−1

is a unitary operator and 1 is not an eigenvalue of this operator.
17 Explain why the characterization of unitary matrices given by 7.57 holds.
18 A square matrix 𝐴 is called symmetric if it equals its transpose. Prove that if
𝐴 is a symmetric matrix with real entries, then there exists a unitary matrix
𝑄 with real entries such that 𝑄∗ 𝐴𝑄 is a diagonal matrix.
19 Suppose 𝑛 is a positive integer. For this exercise, we adopt the notation that
a typical element 𝑧 of 𝐂𝑛 is denoted by 𝑧 = (𝑧0 , 𝑧1 , …, 𝑧𝑛 − 1 ). Define linear
functionals 𝜔0 , 𝜔1 , …, 𝜔𝑛 − 1 on 𝐂𝑛 by
1 𝑛−1
𝜔𝑗 (𝑧0 , 𝑧1 , …, 𝑧𝑛 − 1 ) = √ ∑ 𝑧𝑚 𝑒−2𝜋𝑖𝑗𝑚/𝑛 .
𝑛 𝑚=0
The discrete Fourier transform is the operator ℱ ∶ 𝐂𝑛 → 𝐂𝑛 defined by
ℱ𝑧 = (𝜔0 (𝑧), 𝜔1 (𝑧), …, 𝜔𝑛 − 1 (𝑧)).
(a) Show that ℱ is a unitary operator on 𝐂𝑛.
(b) Show that if (𝑧0 , …, 𝑧𝑛 − 1 ) ∈ 𝐂𝑛 and 𝑧𝑛 is defined to equal 𝑧0 , then
ℱ−1 (𝑧0 , 𝑧1 , …, 𝑧𝑛 − 1 ) = ℱ(𝑧𝑛 , 𝑧𝑛 − 1 , …, 𝑧1 ).
(c) Show that ℱ4 = 𝐼.
The discrete Fourier transform has many important applications in data
analysis. The usual Fourier transform involves expressions of the form
∫−∞ 𝑓 (𝑥)𝑒−2𝜋𝑖𝑡𝑥 𝑑𝑥 for complex-valued integrable functions 𝑓 defined on 𝐑 .
∞
20 Suppose 𝐴 is a square matrix with linearly independent columns. Prove that

there exist unique matrices 𝑅 and 𝑄 such that 𝑅 is lower triangular with only
positive numbers on its diagonal, 𝑄 is unitary, and 𝐴 = 𝑅𝑄.
7E Singular Value Decomposition

Singular Values
We will need the following result in this section.
7.64 properties of 𝑇 ∗ 𝑇

(a) 𝑇 ∗ 𝑇 is a positive operator on 𝑉;
(b) null 𝑇 ∗ 𝑇 = null 𝑇;
(c) range 𝑇 ∗ 𝑇 = range 𝑇 ∗ ;
(d) dim range 𝑇 = dim range 𝑇 ∗ = dim range 𝑇 ∗ 𝑇.
Proof
(a) We have
∗ ∗
(𝑇 ∗ 𝑇) = 𝑇 ∗ (𝑇 ∗ ) = 𝑇 ∗ 𝑇.
Thus 𝑇 ∗ 𝑇 is self-adjoint.
⟨(𝑇 ∗ 𝑇)𝑣, 𝑣⟩ = ⟨𝑇 ∗ (𝑇𝑣), 𝑣⟩ = ⟨𝑇𝑣, 𝑇𝑣⟩ = ‖𝑇𝑣‖2 ≥ 0.
Thus 𝑇 ∗ 𝑇 is a positive operator.

(b) First suppose 𝑣 ∈ null 𝑇 ∗ 𝑇. Then
‖𝑇𝑣‖2 = ⟨𝑇𝑣, 𝑇𝑣⟩ = ⟨𝑇 ∗ 𝑇𝑣, 𝑣⟩ = ⟨0, 𝑣⟩ = 0.
Thus 𝑇𝑣 = 0, proving that null 𝑇 ∗ 𝑇 ⊆ null 𝑇.
The inclusion in the other direction is clear, because if 𝑣 ∈ 𝑉 and 𝑇𝑣 = 0,
then 𝑇 ∗ 𝑇𝑣 = 0.
Thus null 𝑇 ∗ 𝑇 = null 𝑇, completing the proof of (b).
(c) We already know from (a) that 𝑇 ∗ 𝑇 is self-adjoint. Thus
range 𝑇 ∗ 𝑇 = (null 𝑇 ∗ 𝑇)⟂ = (null 𝑇)⟂ = range 𝑇 ∗,
where the first and last equalities come from 7.6 and the second equality
comes from (b).
(d) To verify the first equation in (d), note that
dim range 𝑇 = dim(null 𝑇 ∗ )⟂ = dim 𝑊 − dim null 𝑇 ∗ = dim range 𝑇 ∗,
where the first equality comes from 7.6(d), the second equality comes from
6.51, and the last equality comes from the fundamental theorem of linear
maps (3.21).
The equality dim range 𝑇 ∗ = dim range 𝑇 ∗ 𝑇 follows from (c).
Section 7E Singular Value Decomposition 271
The eigenvalues of an operator tell us something about the behavior of the

operator. Another collection of numbers, called the singular values, is also useful.
Eigenspaces and the notation 𝐸 (used in the examples) were defined in 5.52.
7.65 definition: singular values
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). The singular values of 𝑇 are the nonnegative square
roots of the eigenvalues of 𝑇 ∗ 𝑇, listed in decreasing order, each included as
many times as the dimension of the corresponding eigenspace of 𝑇 ∗ 𝑇.
7.66 example: singular values of an operator on 𝐅4

Define 𝑇 ∈ ℒ(𝐅4 ) by 𝑇(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 ) = (0, 3𝑧1 , 2𝑧2 , −3𝑧4 ). A calculation
shows that
𝑇 ∗ 𝑇(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 ) = (9𝑧1 , 4𝑧2 , 0, 9𝑧4 ),
as you should verify. Thus the standard basis of 𝐅4 diagonalizes 𝑇 ∗ 𝑇, and we
see that the eigenvalues of 𝑇 ∗ 𝑇 are 9, 4, and 0. Also, the dimensions of the
eigenspaces corresponding to the eigenvalues are
dim 𝐸(9, 𝑇 ∗ 𝑇) = 2 and dim 𝐸(4, 𝑇 ∗ 𝑇) = 1 and dim 𝐸(0, 𝑇 ∗ 𝑇) = 1.
Taking nonnegative square roots of these eigenvalues of 𝑇 ∗ 𝑇 and using dimension
information from above, we conclude that the singular values of 𝑇 are 3, 3, 2, 0.
The only eigenvalues of 𝑇 are −3 and 0. Thus in this case, the collection of
eigenvalues did not pick up the number 2 that appears in the definition (and hence
the behavior) of 𝑇, but the list of singular values does include 2.
7.67 example: singular values of a linear map from 𝐅4 to 𝐅3

Suppose 𝑇 ∈ ℒ(𝐅4 , 𝐅3 ) has matrix (with respect to the standard bases)
0 0 0 −5
⎛
⎜ ⎞
⎟
⎜
⎜ 0 0 0 0 ⎟
⎟.
⎝ 1 1 0 0 ⎠
You can verify that the matrix of 𝑇 ∗ 𝑇 is
1 1 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ 1 1 0 0 ⎟
⎟
⎜
⎜ ⎟
⎟
⎜ 0 0 0 0 ⎟
⎝ 0 0 0 25 ⎠
and that the eigenvalues of the operator 𝑇 ∗ 𝑇 are 25, 2, 0, with dim 𝐸(25, 𝑇 ∗ 𝑇) = 1,
dim 𝐸(2, 𝑇 ∗ 𝑇) = 1, and dim 𝐸(0, 𝑇 ∗ 𝑇) = 2. Thus the singular values of 𝑇 are
5, √2, 0, 0.
See Exercise 2 for a characterization of the positive singular values.

≈ 100 𝑒 Chapter 7 Operators on Inner Product Spaces
7.68 role of positive singular values
Suppose that 𝑇 ∈ ℒ(𝑉, 𝑊). Then

(a) 𝑇 is injective ⟺ 0 is not a singular value of 𝑇;
(b) the number of positive singular values of 𝑇 equals dim range 𝑇;
(c) 𝑇 is surjective ⟺ number of positive singular values of 𝑇 equals dim 𝑊.
Proof The linear map 𝑇 is injective if and only if null 𝑇 = {0}, which happens
if and only if null 𝑇 ∗ 𝑇 = {0} [by 7.64(b)], which happens if and only if 0 is not
an eigenvalue of 𝑇 ∗ 𝑇, which happens if and only if 0 is not a singular value of 𝑇,
completing the proof of (a).
The spectral theorem applied to 𝑇 ∗ 𝑇 shows that dim range 𝑇 ∗ 𝑇 equals the num-
ber of positive eigenvalues of 𝑇 ∗ 𝑇 (counting repetitions). Thus 7.64(c) implies
that dim range 𝑇 equals the number of positive singular values of 𝑇, proving (b).
Use (b) and 2.39 to show that (c) holds.
The table below compares eigenvalues with singular values.

list of eigenvalues list of singular values
context: vector spaces context: inner product spaces
defined only for linear maps from a vector defined for linear maps from an inner
space to itself product space to a possibly different inner
product space
can be arbitrary real numbers (if 𝐅 = 𝐑 ) are nonnegative numbers
or complex numbers (if 𝐅 = 𝐂 )
can be the empty list if 𝐅 = 𝐑 length of list equals dimension of domain
includes 0 ⟺ operator is not invertible includes 0 ⟺ linear map is not injective
no standard order, especially if 𝐅 = 𝐂 always listed in decreasing order
The next result nicely characterizes isometries in terms of singular values.
7.69 isometries characterized by having all singular values equal 1
Suppose that 𝑆 ∈ ℒ(𝑉, 𝑊). Then
𝑆 is an isometry ⟺ all singular values of 𝑆 equal 1.
Proof We have
𝑆 is an isometry ⟺ 𝑆∗ 𝑆 = 𝐼
⟺ all eigenvalues of 𝑆∗ 𝑆 equal 1
⟺ all singular values of 𝑆 equal 1,
where the first equivalence comes from 7.49 and the second equivalence comes
from the spectral theorem (7.29 or 7.31) applied to the self-adjoint operator 𝑆∗ 𝑆.
SVD for Linear Maps and for Matrices

The next result shows that every linear The singular value decomposition is
map from 𝑉 to 𝑊 has a remarkably clean useful in computational linear alge-
description in terms of its singular val- bra because good techniques exist for
ues and orthonormal lists in 𝑉 and 𝑊. approximating eigenvalues and eigen-
In the next section we will see several vectors of positive operators such as
important applications of the singular 𝑇 ∗ 𝑇, whose eigenvalues and eigenvec-
value decomposition (often called the tors lead to the singular value decom-
SVD). position.
7.70 singular value decomposition
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and the positive singular values of 𝑇 are 𝑠1 , …, 𝑠𝑚 .

Then there exist orthonormal lists 𝑒1 , …, 𝑒𝑚 in 𝑉 and 𝑓1 , …, 𝑓𝑚 in 𝑊 such that
7.71 𝑇𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩ 𝑓𝑚
Proof Let 𝑠1 , …, 𝑠𝑛 denote the singular values of 𝑇 (thus 𝑛 = dim 𝑉). Because
𝑇 ∗ 𝑇 is a positive operator [see 7.64(a)], the spectral theorem implies that there
exists an orthonormal basis 𝑒1 , …, 𝑒𝑛 of 𝑉 with
7.72 𝑇 ∗ 𝑇𝑒𝑘 = 𝑠𝑘2 𝑒𝑘
for each 𝑘 = 1, …, 𝑛.
For each 𝑘 = 1, …, 𝑚, let
𝑇𝑒𝑘
7.73 𝑓𝑘 = .
𝑠𝑘
If 𝑗, 𝑘 ∈ {1, …, 𝑚}, then
1 1 𝑠𝑘 ⎧
{0 if 𝑗 ≠ 𝑘,
⟨ 𝑓𝑗 , 𝑓𝑘 ⟩ = ⟨𝑇𝑒𝑗 , 𝑇𝑒𝑘 ⟩ = ⟨𝑒𝑗 , 𝑇 ∗ 𝑇𝑒𝑘 ⟩ = ⟨𝑒𝑗 , 𝑒𝑘 ⟩ = ⎨
𝑠𝑗 𝑠𝑘 𝑠𝑗 𝑠𝑘 𝑠𝑗 {
⎩1 if 𝑗 = 𝑘.
Thus 𝑓1 , …, 𝑓𝑚 is an orthonormal list in 𝑊.
If 𝑘 ∈ {1, …, 𝑛} and 𝑘 > 𝑚, then 𝑠𝑘 = 0 and hence 𝑇 ∗ 𝑇𝑒𝑘 = 0 (by 7.72), which
implies that 𝑇𝑒𝑘 = 0 [by 7.64(b)].
Suppose 𝑣 ∈ 𝑉. Then
𝑇𝑣 = 𝑇(⟨𝑣, 𝑒1 ⟩𝑒1 + ⋯ + ⟨𝑣, 𝑒𝑛 ⟩𝑒𝑛 )
= ⟨𝑣, 𝑒1 ⟩𝑇𝑒1 + ⋯ + ⟨𝑣, 𝑒𝑚 ⟩𝑇𝑒𝑚
= 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩ 𝑓𝑚 ,
where the last index in the first line switched from 𝑛 to 𝑚 in the second line
because 𝑇𝑒𝑘 = 0 if 𝑘 > 𝑚 (as noted in the paragraph above) and the third line
follows from 7.73. The equation above is our desired result.
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊), the positive singular values of 𝑇 are 𝑠1 , …, 𝑠𝑚 , and

𝑒1 , …, 𝑒𝑚 and 𝑓1 , …, 𝑓𝑚 are as in the singular value decomposition 7.70. The
orthonormal list 𝑒1 , …, 𝑒𝑚 can be extended to an orthonormal basis 𝑒1 , …, 𝑒dim 𝑉
of 𝑉 and the orthonormal list 𝑓1 , …, 𝑓𝑚 can be extended to an orthonormal basis
𝑓1 , …, 𝑓dim 𝑊 of 𝑊. The formula 7.71 shows that
⎧𝑠 𝑓
{ if 1 ≤ 𝑘 ≤ 𝑚,
𝑇𝑒𝑘 = ⎨ 𝑘 𝑘
{0
⎩ if 𝑚 < 𝑘 ≤ dim 𝑉.
Thus the matrix of 𝑇 with respect to the orthonormal bases (𝑒1 , …, 𝑒dim 𝑉 ) and
( 𝑓1 , …, 𝑓dim 𝑊 ) has the simple form
⎧
{𝑠 if 1 ≤ 𝑗 = 𝑘 ≤ 𝑚,
ℳ(𝑇, (𝑒1 , …, 𝑒dim 𝑉 ), ( 𝑓1 , …, 𝑓dim 𝑊 ))𝑗, 𝑘 = ⎨ 𝑘
{
⎩0 otherwise.
If dim 𝑉 = dim 𝑊 (as happens, for example, if 𝑊 = 𝑉), then the matrix
described in the paragraph above is a diagonal matrix. If we extend the definition
of diagonal matrix as follows to apply to matrices that are not necessarily square,
then we have proved the wonderful result that every linear map from 𝑉 to 𝑊 has
a diagonal matrix with respect to appropriate orthonormal bases.
7.74 definition: diagonal matrix
An 𝑀-by-𝑁 matrix 𝐴 is called a diagonal matrix if all entries of the matrix

are 0 except possibly 𝐴𝑘, 𝑘 for 𝑘 = 1, …, min{𝑀, 𝑁}.
The table below compares the spectral theorem (7.29 and 7.31) with the
singular value decomposition (7.70).
spectral theorem singular value decomposition
describes only self-adjoint operators describes arbitrary linear maps from an
(when 𝐅 = 𝐑 ) or normal operators (when inner product space to a possibly different
𝐅 = 𝐂) inner product space
produces a single orthonormal basis produces two orthonormal lists, one for
domain space and one for range space,
that are not necessarily the same even
when range space equals domain space
different proofs depending on whether same proof works regardless of whether
𝐅 = 𝐑 or 𝐅 = 𝐂 𝐅 = 𝐑 or 𝐅 = 𝐂
The singular value decomposition gives us a new way to understand the adjoint
and the inverse of a linear map. Specifically, the next result shows that given a
singular value decomposition of a linear map 𝑇 ∈ ℒ(𝑉, 𝑊), we can obtain the
adjoint of 𝑇 simply by interchanging the roles of the 𝑒’s and the 𝑓 ’s (see 7.77).
Similarly, we can obtain the pseudoinverse 𝑇 † (see 6.68) of 𝑇 by interchanging
the roles of the 𝑒’s and the 𝑓 ’s and replacing each positive singular value 𝑠𝑘 of 𝑇
with 1/𝑠𝑘 (see 7.78).
Recall that the pseudoinverse 𝑇 † in 7.78 below equals the inverse 𝑇 −1 if 𝑇 is

invertible [see 6.69(a)].
7.75 singular value decomposition of adjoint and pseudoinverse
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and the positive singular values of 𝑇 are 𝑠1 , …, 𝑠𝑚 .

Suppose 𝑒1 , …, 𝑒𝑚 and 𝑓1 , …, 𝑓𝑚 are orthonormal lists in 𝑉 and 𝑊 such that
7.76 𝑇𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩ 𝑓𝑚
for every 𝑣 ∈ 𝑉. Then
7.77 𝑇 ∗ 𝑤 = 𝑠1 ⟨𝑤, 𝑓1 ⟩𝑒1 + ⋯ + 𝑠𝑚 ⟨𝑤, 𝑓𝑚 ⟩𝑒𝑚
and
⟨𝑤, 𝑓1 ⟩ ⟨𝑤, 𝑓𝑚 ⟩
7.78 𝑇†𝑤 = 𝑒1 + ⋯ + 𝑒𝑚
𝑠1 𝑠𝑚
for every 𝑤 ∈ 𝑊.
Proof If 𝑣 ∈ 𝑉 and 𝑤 ∈ 𝑊 then

⟨𝑇𝑣, 𝑤⟩ = ⟨𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩ 𝑓𝑚 , 𝑤⟩
= 𝑠1 ⟨𝑣, 𝑒1 ⟩⟨ 𝑓1 , 𝑤⟩ + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩⟨ 𝑓𝑚 , 𝑤⟩
= ⟨𝑣, 𝑠1 ⟨𝑤, 𝑓1 ⟩𝑒1 + ⋯ + 𝑠𝑚 ⟨𝑤, 𝑓𝑚 ⟩𝑒𝑚 ⟩.

This implies that
𝑇 ∗ 𝑤 = 𝑠1 ⟨𝑤, 𝑓1 ⟩𝑒1 + ⋯ + 𝑠𝑚 ⟨𝑤, 𝑓𝑚 ⟩𝑒𝑚 ,
proving 7.77.
To prove 7.78, suppose 𝑤 ∈ 𝑊. Let
⟨𝑤, 𝑓1 ⟩ ⟨𝑤, 𝑓𝑚 ⟩
𝑣= 𝑒1 + ⋯ + 𝑒𝑚 .
𝑠1 𝑠𝑚
Apply 𝑇 to both sides of the equation above, getting
⟨𝑤, 𝑓1 ⟩ ⟨𝑤, 𝑓𝑚 ⟩
𝑇𝑣 = 𝑇𝑒1 + ⋯ + 𝑇𝑒𝑚
𝑠1 𝑠𝑚
= ⟨𝑤, 𝑓1 ⟩ 𝑓1 + ⋯ + ⟨𝑤, 𝑓𝑚 ⟩ 𝑓𝑚
= 𝑃range 𝑇 𝑤,
where the second line holds because 7.76 implies that 𝑇𝑒𝑘 = 𝑠𝑘 𝑓𝑘 if 𝑘 = 1, …, 𝑚,
and the last line above holds because 7.76 implies that 𝑓1 , …, 𝑓𝑚 spans range 𝑇 and
thus is an orthonormal basis of range 𝑇 [and hence 6.57(i) applies]. The equation
above, the observation that 𝑣 ∈ (null 𝑇)⟂ [see Exercise 8(b)], and the definition
of 𝑇 † 𝑤 (see 6.68) show that 𝑣 = 𝑇 † 𝑤, proving 7.78.
7.79 example: finding a singular value decomposition

Define 𝑇 ∈ ℒ(𝐅4 , 𝐅3 ) by 𝑇(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ) = (−5𝑥4 , 0, 𝑥1 + 𝑥2 ). We want to
find a singular value decomposition of 𝑇. The matrix of 𝑇 (with respect to the
standard bases) is
0 0 0 −5
⎛
⎜ ⎞
⎟.
⎜
⎜ 0 0 0 0 ⎟ ⎟
⎝ 1 1 0 0 ⎠
Thus, as discussed in Example 7.67, the matrix of 𝑇 ∗ 𝑇 is
1 1 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ 1 1 0 0 ⎟
⎟
⎜
⎜ ⎟
⎟ ,
⎜ 0 0 0 0 ⎟
⎝ 0 0 0 25 ⎠
and the positive eigenvalues of 𝑇 ∗ 𝑇 are 25, 2, with dim 𝐸(25, 𝑇 ∗ 𝑇) = 1 and
dim 𝐸(2, 𝑇 ∗ 𝑇) = 1. Hence the positive singular values of 𝑇 are 5, √2.
Thus to find a singular value decomposition of 𝑇, we must find an orthonormal
list 𝑒1 , 𝑒2 in 𝐅4 and an orthonormal list 𝑓1 , 𝑓2 in 𝐅3 such that
𝑇𝑣 = 5⟨𝑣, 𝑒1 ⟩ 𝑓1 + √2⟨𝑣, 𝑒2 ⟩ 𝑓2
for all 𝑣 ∈ 𝐅4 .
An orthonormal basis of 𝐸(25, 𝑇 ∗ 𝑇) is the vector (0, 0, 0, 1); an orthonormal
basis of 𝐸(2, 𝑇 ∗ 𝑇) is the vector ( √1 , √1 , 0, 0). Thus, following the proof of 7.70,
2 2
we take
1 1
𝑒1 = (0, 0, 0, 1) and 𝑒2 = ( , , 0, 0)
√2 √2
and
𝑇𝑒1 𝑇𝑒2
𝑓1 = = (−1, 0, 0) and 𝑓2 = = (0, 0, 1).
5 √2
Then, as expected, we see that 𝑒1 , 𝑒2 is an orthonormal list in 𝐅4 and 𝑓1 , 𝑓2 is an

orthonormal list in 𝐅3 and
𝑇𝑣 = 5⟨𝑣, 𝑒1 ⟩ 𝑓1 + √2⟨𝑣, 𝑒2 ⟩ 𝑓2
for all 𝑣 ∈ 𝐅4 . Thus we have found a singular value decomposition of 𝑇.
The next result translates the singular value decomposition from the context
of linear maps to the context of matrices. Specifically, the following result gives
a factorization of an arbitrary matrix as the product of three nice matrices. The
proof gives an explicit construction of these three matrices in terms of the singular
value decomposition.
In the next result, the phrase “orthogonal columns” should be interpreted to
mean that the columns are orthogonal with respect to the standard Euclidean inner
product.
7.80 matrix version of SVD
Suppose 𝐴 is an 𝑀-by-𝑛 matrix of rank 𝑚 ≥ 1. Then there exist an 𝑀-by-𝑚

matrix 𝐵 with orthonormal columns, an 𝑚-by-𝑚 diagonal matrix 𝐷 with
positive numbers on the diagonal, and an 𝑛-by-𝑚 matrix 𝐶 with orthonormal
columns such that
𝐴 = 𝐵𝐷𝐶∗.
Proof Let 𝑇 ∶ 𝐅𝑛 → 𝐅𝑀 be the linear map whose matrix with respect to the
standard bases equals 𝐴. Then dim range 𝑇 = 𝑚 (by 3.78). Let
7.81 𝑇𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩ 𝑓𝑚
be a singular value decomposition of 𝑇. Let

𝐵 = the 𝑀-by-𝑚 matrix whose columns are 𝑓1 , …, 𝑓𝑚 ,
𝐷 = the 𝑚-by-𝑚 diagonal matrix whose diagonal entries are 𝑠1 , …, 𝑠𝑚 ,
𝐶 = the 𝑛-by-𝑚 matrix whose columns are 𝑒1 , …, 𝑒𝑚 .
Let 𝑢1 , …, 𝑢𝑚 denote the standard basis of 𝐅𝑚 . If 𝑘 ∈ {1, …, 𝑚} then

(𝐴𝐶 − 𝐵𝐷)𝑢𝑘 = 𝐴𝑒𝑘 − 𝐵(𝑠𝑘 𝑢𝑘 ) = 𝑠𝑘 𝑓𝑘 − 𝑠𝑘 𝑓𝑘 = 0.
Thus 𝐴𝐶 = 𝐵𝐷.
Multiply both sides of this last equation by 𝐶∗ (the conjugate transpose of 𝐶)
on the right to get
𝐴𝐶𝐶∗ = 𝐵𝐷𝐶∗.
Note that the rows of 𝐶∗ are the complex conjugates of 𝑒1 , …, 𝑒𝑚 . Thus if
𝑘 ∈ {1, …, 𝑚}, then the definition of matrix multiplication shows that 𝐶∗ 𝑒𝑘 = 𝑢𝑘 ;
hence 𝐶𝐶∗ 𝑒𝑘 = 𝑒𝑘 . Thus 𝐴𝐶𝐶∗ 𝑣 = 𝐴𝑣 for all 𝑣 ∈ span(𝑒1 , …, 𝑒𝑚 ).
If 𝑣 ∈ (span(𝑒1 , …, 𝑒𝑚 ))⟂ , then 𝐴𝑣 = 0 (as follows from 7.81) and 𝐶∗ 𝑣 = 0
(as follows from the definition of matrix multiplication). Hence 𝐴𝐶𝐶∗ 𝑣 = 𝐴𝑣 for
all 𝑣 ∈ (span(𝑒1 , …, 𝑒𝑚 ))⟂ .
Because 𝐴𝐶𝐶∗ and 𝐴 agree on span(𝑒1 , …, 𝑒𝑚 ) and on (span(𝑒1 , …, 𝑒𝑚 ))⟂ , we
conclude that 𝐴𝐶𝐶∗ = 𝐴. Thus the displayed equation above becomes
𝐴 = 𝐵𝐷𝐶∗,
as desired.
Note that the matrix 𝐴 in the result above has 𝑀𝑛 entries. In comparison, the
matrices 𝐵, 𝐷, and 𝐶 above have a total of
𝑚(𝑀 + 𝑚 + 𝑛)
entries. Thus if 𝑀 and 𝑛 are large numbers and the rank 𝑚 is considerably less
than 𝑀 and 𝑛, then the number of entries that must be stored on a computer to
represent 𝐴 is considerably less than 𝑀𝑛.
Exercises 7E
1 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Show that 𝑇 = 0 if and only if all singular values of
𝑇 are 0.
2 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑠 > 0. Prove that 𝑠 is a singular value of 𝑇 if and

only if there exist nonzero vectors 𝑣 ∈ 𝑉 and 𝑤 ∈ 𝑊 such that
𝑇𝑣 = 𝑠𝑤 and 𝑇 ∗ 𝑤 = 𝑠𝑣.
The vectors 𝑣, 𝑤 satisfying both equations above are called a Schmidt pair.
Erhard Schmidt introduced the concept of singular values in 1907.
3 Give an example of 𝑇 ∈ ℒ(𝐂2 ) such that 0 is the only eigenvalue of 𝑇 and

the singular values of 𝑇 are 5, 0.
4 Suppose that 𝑇 ∈ ℒ(𝑉, 𝑊), 𝑠1 is the largest singular value of 𝑇, and 𝑠𝑛 is
the smallest singular value of 𝑇. Prove that
{‖𝑇𝑣‖ ∶ 𝑣 ∈ 𝑉 and ‖𝑣‖ = 1} = [𝑠𝑛 , 𝑠1 ].
5 Suppose 𝑇 ∈ ℒ(𝐂2 ) is defined by 𝑇(𝑥, 𝑦) = (−4𝑦, 𝑥). Find the singular

values of 𝑇.
6 Find the singular values of the differentiation operator 𝐷 ∈ ℒ(𝒫2 (𝐑))
defined by 𝐷𝑝 = 𝑝′, where the inner product on 𝒫2 (𝐑) is as in Example 6.34.
7 Suppose that 𝑇 ∈ ℒ(𝑉) is self-adjoint or that 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉) is
normal. Let 𝜆1 , …, 𝜆𝑛 be the eigenvalues of 𝑇, each included in this list
as many times as the dimension of the corresponding eigenspace. Show
that the singular values of 𝑇 are |𝜆1 |, …, |𝜆𝑛 |, after these numbers have been
sorted into decreasing order.
8 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Suppose 𝑠1 ≥ 𝑠2 ≥ ⋯ ≥ 𝑠𝑚 > 0 and 𝑒1 , …, 𝑒𝑚 is an
orthonormal list in 𝑉 and 𝑓1 , …, 𝑓𝑚 is an orthonormal list in 𝑊 such that
𝑇𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩ 𝑓𝑚
(a) Prove that 𝑓1 , …, 𝑓𝑚 is an orthonormal basis of range 𝑇.
(b) Prove that 𝑒1 , …, 𝑒𝑚 is an orthonormal basis of (null 𝑇)⟂ .
(c) Prove that 𝑠1 , …, 𝑠𝑚 are the positive singular values of 𝑇.
(d) Prove that if 𝑘 ∈ {1, …, 𝑚}, then 𝑒𝑘 is an eigenvector of 𝑇 ∗ 𝑇 with corre-
sponding eigenvalue 𝑠𝑘2.
(e) Prove that
𝑇𝑇 ∗ 𝑤 = 𝑠12 ⟨𝑤, 𝑓1 ⟩ 𝑓1 + ⋯ + 𝑠𝑚2 ⟨𝑤, 𝑓𝑚 ⟩ 𝑓𝑚
for all 𝑤 ∈ 𝑊.
9 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Show that 𝑇 and 𝑇 ∗ have the same positive singular
values.
10 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) has singular values 𝑠1 , …, 𝑠𝑛 . Prove that if 𝑇 is an
invertible linear map, then 𝑇 −1 has singular values
1 1
, …, .
𝑠𝑛 𝑠1
11 Suppose that 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑣1 , …, 𝑣𝑛 is an orthonormal basis of 𝑉. Let
𝑠1 , …, 𝑠𝑛 denote the singular values of 𝑇.
(a) Prove that ‖𝑇𝑣1 ‖2 + ⋯ + ‖𝑇𝑣𝑛 ‖2 = 𝑠12 + ⋯ + 𝑠𝑛2.
(b) Prove that if 𝑊 = 𝑉 and 𝑇 is a positive operator, then
⟨𝑇𝑣1 , 𝑣1 ⟩ + ⋯ + ⟨𝑇𝑣𝑛 , 𝑣𝑛 ⟩ = 𝑠1 + ⋯ + 𝑠𝑛 .
See the comment after Exercise 5 in Section 7A.
12 (a) Give an example of a finite-dimensional vector space and an operator 𝑇

on it such that the singular values of 𝑇 2 do not equal the squares of the
singular values of 𝑇.
(b) Suppose 𝑇 ∈ ℒ(𝑉) is normal. Prove that the singular values of 𝑇 2
equal the squares of the singular values of 𝑇.
13 Suppose 𝑇1 , 𝑇2 ∈ ℒ(𝑉). Prove that 𝑇1 and 𝑇2 have the same singular
values if and only if there exist unitary operators 𝑆1 , 𝑆2 ∈ ℒ(𝑉) such that
𝑇1 = 𝑆1 𝑇2 𝑆2 .
14 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Let 𝑠𝑛 denote the smallest singular value of 𝑇. Prove
that 𝑠𝑛 ‖𝑣‖ ≤ ‖𝑇𝑣‖ for every 𝑣 ∈ 𝑉.
15 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑠1 ≥ ⋯ ≥ 𝑠𝑛 are the singular values of 𝑇. Prove
that if 𝜆 is an eigenvalue of 𝑇, then 𝑠1 ≥ |𝜆| ≥ 𝑠𝑛 .
16 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that (𝑇 ∗ )† = (𝑇 † )∗.

Compare the result in this exercise to the analogous result for invertible
linear maps [see 7.5( f )].
17 Suppose 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is self-adjoint if and only if 𝑇 † is self-

adjoint.
Matrices unfold
Singular values gleam like stars
Order in chaos shines
—written by ChatGPT with input haiku about SVD

7F Consequences of Singular Value Decomposition

Norms of Linear Maps
The singular value decomposition leads to the following upper bound for ‖𝑇𝑣‖.
7.82 upper bound for ‖𝑇𝑣‖
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Let 𝑠1 be the largest singular value of 𝑇. Then
‖𝑇𝑣‖ ≤ 𝑠1 ‖𝑣‖
Proof Let 𝑠1 , …, 𝑠𝑚 denote the positive For a lower bound on ‖𝑇𝑣‖, look at
singular values of 𝑇, and let 𝑒1 , …, 𝑒𝑚 be Exercise 14 in Section 7E.
an orthonormal list in 𝑉 and 𝑓1 , …, 𝑓𝑚 be
an orthonormal list in 𝑊 that provide a singular value decomposition of 𝑇. Thus
7.83 𝑇𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩ 𝑓𝑚
for all 𝑣 ∈ 𝑉. Hence if 𝑣 ∈ 𝑉 then
2 2
‖𝑇𝑣‖2 = 𝑠12 ∣⟨𝑣, 𝑒1 ⟩∣ + ⋯ + 𝑠𝑚2 ∣⟨𝑣, 𝑒𝑚 ⟩∣
2 2
≤ 𝑠12 (∣⟨𝑣, 𝑒1 ⟩∣ + ⋯ + ∣⟨𝑣, 𝑒𝑚 ⟩∣ )
≤ 𝑠12 ‖𝑣‖2,
where the last inequality follows from Bessel’s inequality (6.26). Taking square
roots of both sides of the inequality above shows that ‖𝑇𝑣‖ ≤ 𝑠1 ‖𝑣‖, as desired.
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑠1 is the largest singular value of 𝑇. The result

above shows that
7.84 ‖𝑇𝑣‖ ≤ 𝑠1 for all 𝑣 ∈ 𝑉 with ‖𝑣‖ ≤ 1.
Taking 𝑣 = 𝑒1 in 7.83 shows that 𝑇𝑒1 = 𝑠1 𝑓1 . Because ‖ 𝑓1 ‖ = 1, this implies that
‖𝑇𝑒1 ‖ = 𝑠1 . Thus because ‖𝑒1 ‖ = 1, the inequality in 7.84 leads to the equation
7.85 max{‖𝑇𝑣‖ ∶ 𝑣 ∈ 𝑉 and ‖𝑣‖ ≤ 1} = 𝑠1 .
The equation above is the motivation for the following definition, which defines
the norm of 𝑇 to be the left side of the equation above without needing to refer to
singular values or the singular value decomposition.
7.86 definition: norm of a linear map, ‖ ⋅ ‖
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Then the norm of 𝑇, denoted by ‖𝑇‖, is defined by
‖𝑇‖ = max{‖𝑇𝑣‖ ∶ 𝑣 ∈ 𝑉 and ‖𝑣‖ ≤ 1}.

Section 7F Consequences of Singular Value Decomposition 281
In general, the maximum of an infinite set of nonnegative numbers need

not exist. However, the discussion before 7.86 shows that the maximum in the
definition of the norm of a linear map 𝑇 from 𝑉 to 𝑊 does indeed exist (and equals
the largest singular value of 𝑇).
We now have two different uses of the word norm and the notation ‖ ⋅ ‖. Our
first use of this notation was in connection with an inner product on 𝑉, when we
defined ‖𝑣‖ = √⟨𝑣, 𝑣⟩ for each 𝑣 ∈ 𝑉. Our second use of the norm notation and
terminology is with the definition we just made of ‖𝑇‖ for 𝑇 ∈ ℒ(𝑉, 𝑊). The
norm ‖𝑇‖ for 𝑇 ∈ ℒ(𝑉, 𝑊) does not usually come from taking an inner product
of 𝑇 with itself (see Exercise 21). You should be able to tell from the context and
from the symbols used which meaning of the norm is intended.
The properties of the norm on ℒ(𝑉, 𝑊) listed below look identical to properties
of the norm on an inner product space (see 6.9 and 6.17). The inequality in (d) is
called the triangle inequality, thus using the same terminology that we used for
the norm on 𝑉. For the reverse triangle inequality, see Exercise 1.
7.87 basic properties of norms of linear maps

(a) ‖𝑇‖ ≥ 0;
(b) ‖𝑇‖ = 0 ⟺ 𝑇 = 0;
(c) ‖𝜆𝑇‖ = |𝜆| ‖𝑇‖ for all 𝜆 ∈ 𝐅 ;
(d) ‖𝑆 + 𝑇‖ ≤ ‖𝑆‖ + ‖𝑇‖ for all 𝑆 ∈ ℒ(𝑉, 𝑊).
Proof
(a) Because ‖𝑇𝑣‖ ≥ 0 for every 𝑣 ∈ 𝑉, the definition of ‖𝑇‖ implies that ‖𝑇‖ ≥ 0.
(b) Suppose ‖𝑇‖ = 0. Thus 𝑇𝑣 = 0 for all 𝑣 ∈ 𝑉 with ‖𝑣‖ ≤ 1. If 𝑢 ∈ 𝑉 with
𝑢 ≠ 0, then 𝑢
𝑇𝑢 = ‖𝑢‖ 𝑇( ) = 0,
‖𝑢‖
where the last equality holds because 𝑢/‖𝑢‖ has norm 1. Because 𝑇𝑢 = 0 for
all 𝑢 ∈ 𝑉, we have 𝑇 = 0.
Conversely, if 𝑇 = 0 then 𝑇𝑣 = 0 for all 𝑣 ∈ 𝑉 and hence ‖𝑇‖ = 0.
(c) Suppose 𝜆 ∈ 𝐅. Then
‖𝜆𝑇‖ = max{‖𝜆𝑇𝑣‖ ∶ 𝑣 ∈ 𝑉 and ‖𝑣‖ ≤ 1}
= |𝜆| max{‖𝑇𝑣‖ ∶ 𝑣 ∈ 𝑉 and ‖𝑣‖ ≤ 1}
= |𝜆| ‖𝑇‖.
(d) Suppose 𝑆 ∈ ℒ(𝑉, 𝑊). The definition of ‖𝑆 + 𝑇‖ implies that there exists
𝑣 ∈ 𝑉 such that ‖𝑣‖ ≤ 1 and ‖𝑆 + 𝑇‖ = ∥(𝑆 + 𝑇)𝑣∥. Now
‖𝑆 + 𝑇‖ = ∥(𝑆 + 𝑇)𝑣∥ = ‖𝑆𝑣 + 𝑇𝑣‖ ≤ ‖𝑆𝑣‖ + ‖𝑇𝑣‖ ≤ ‖𝑆‖ + ‖𝑇‖,
completing the proof of (d).
For 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊), the quantity ‖𝑆 − 𝑇‖ is often called the distance between
𝑆 and 𝑇. Informally, think of the condition that ‖𝑆 − 𝑇‖ is a small number as
meaning that 𝑆 and 𝑇 are close together. For example, Exercise 9 asserts that for
every 𝑇 ∈ ℒ(𝑉), there is an invertible operator as close to 𝑇 as we wish.
7.88 alternative formulas for ‖𝑇‖

(a) ‖𝑇‖ = the largest singular value of 𝑇;
(b) ‖𝑇‖ = max{‖𝑇𝑣‖ ∶ 𝑣 ∈ 𝑉 and ‖𝑣‖ = 1};
(c) ‖𝑇‖ = the smallest number 𝑐 such that ‖𝑇𝑣‖ ≤ 𝑐‖𝑣‖ for all 𝑣 ∈ 𝑉.
Proof
(a) See 7.85.
(b) Let 𝑣 ∈ 𝑉 be such that 0 < ‖𝑣‖ ≤ 1. Let 𝑢 = 𝑣/‖𝑣‖. Then
𝑣 𝑣 ‖𝑇𝑣‖
‖𝑢‖ = ∥ ∥=1 and ‖𝑇𝑢‖ = ∥𝑇( )∥ = ≥ ‖𝑇𝑣‖.
‖𝑣‖ ‖𝑣‖ ‖𝑣‖
Thus when finding the maximum of ‖𝑇𝑣‖ with ‖𝑣‖ ≤ 1, we can restrict
attention to vectors in 𝑉 with norm 1, proving (b).
(c) Suppose 𝑣 ∈ 𝑉 and 𝑣 ≠ 0. Then the definition of ‖𝑇‖ implies that
𝑣
∥𝑇( )∥ ≤ ‖𝑇‖,
‖𝑣‖
which implies that
7.89 ‖𝑇𝑣‖ ≤ ‖𝑇‖ ‖𝑣‖.
Now suppose 𝑐 ≥ 0 and ‖𝑇𝑣‖ ≤ 𝑐‖𝑣‖ for all 𝑣 ∈ 𝑉. This implies that
‖𝑇𝑣‖ ≤ 𝑐
for all 𝑣 ∈ 𝑉 with ‖𝑣‖ ≤ 1. Taking the maximum of the left side of the
inequality above over all 𝑣 ∈ 𝑉 with ‖𝑣‖ ≤ 1 shows that ‖𝑇‖ ≤ 𝑐. Thus ‖𝑇‖ is
the smallest number 𝑐 such that ‖𝑇𝑣‖ ≤ 𝑐‖𝑣‖ for all 𝑣 ∈ 𝑉.
When working with norms of linear maps, you will probably frequently use
the inequality 7.89.
For computing an approximation of the norm of a linear map 𝑇 given the
matrix of 𝑇 with respect to some orthonormal bases, 7.88(a) is likely to be most
useful. The matrix of 𝑇 ∗ 𝑇 is quickly computable from matrix multiplication.
Then a computer can be asked to find an approximation for the largest eigenvalue
of 𝑇 ∗ 𝑇 (excellent numeric algorithms exist for this purpose). Then taking the
square root and using 7.88(a) gives an approximation for the norm of 𝑇 (which
usually cannot be computed exactly).
You should verify all assertions in the example below.
7.90 example: norms

• If 𝐼 denotes the usual identity operator on 𝑉, then ‖𝐼‖ = 1.
• If 𝑇 ∈ ℒ(𝐅𝑛 ) and the matrix of 𝑇 with respect to the standard basis of 𝐅𝑛
consists of all 1’s, then ‖𝑇‖ = 𝑛.
• If 𝑇 ∈ ℒ(𝑉) and 𝑉 has an orthonormal basis consisting of eigenvectors of
𝑇 with corresponding eigenvalues 𝜆1 , …, 𝜆𝑛 , then ‖𝑇‖ is the maximum of the
numbers |𝜆1 |, …, |𝜆𝑛 |.
• Suppose 𝑇 ∈ ℒ(𝐑5 ) is the operator whose matrix (with respect to the stan-
dard basis) is the 5-by-5 matrix whose entry in row 𝑗, column 𝑘 is 1/(𝑗2 + 𝑘).
Standard mathematical software shows that the largest singular value of 𝑇 is
approximately 0.8 and the smallest singular value of 𝑇 is approximately 10−6.
Thus ‖𝑇‖ ≈ 0.8 and (using Exercise 10 in Section 7E) ∥𝑇 −1 ∥ ≈ 106. It is not
possible to find exact formulas for these norms.
A linear map and its adjoint have the same norm, as shown by the next result.
7.91 norm of the adjoint
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Then ∥𝑇 ∗ ∥ = ‖𝑇‖.
Proof Suppose 𝑤 ∈ 𝑊. Then

∥𝑇 ∗ 𝑤∥ = ⟨𝑇 ∗ 𝑤, 𝑇 ∗ 𝑤⟩ = ⟨𝑇𝑇 ∗ 𝑤, 𝑤⟩ ≤ ∥𝑇𝑇 ∗ 𝑤∥ ‖𝑤‖ ≤ ‖𝑇‖ ∥𝑇 ∗ 𝑤∥ ‖𝑤‖.
2
The inequality above implies that

∥𝑇 ∗ 𝑤∥ ≤ ‖𝑇‖ ‖𝑤‖,
which along with 7.88(𝑐) implies that ∥𝑇 ∗ ∥ ≤ ‖𝑇‖.
Replacing 𝑇 with 𝑇 ∗ in the inequality ∥𝑇 ∗ ∥ ≤ ‖𝑇‖ and then using the equation
∗
(𝑇 ) = 𝑇 shows that ‖𝑇‖ ≤ ∥𝑇 ∗ ∥. Thus ∥𝑇 ∗ ∥ = ‖𝑇‖, as desired.
∗
You may want to construct an alternative proof of the result above using
Exercise 9 in Section 7E, which asserts that a linear map and its adjoint have the
same positive singular values.
Approximation by Linear Maps with Lower-Dimensional Range

The next result is a spectacular application of the singular value decomposition.
It says that to best approximate a linear map by a linear map whose range has
dimension at most 𝑘, chop off the singular value decomposition after the first
𝑘 terms. Specifically, the linear map 𝑇𝑘 in the next result has the property that
dim range 𝑇𝑘 = 𝑘 and 𝑇𝑘 minimizes the distance to 𝑇 among all linear maps with
range of dimension at most 𝑘. This result leads to algorithms for compressing
huge matrices while preserving their most important information.
7.92 best approximation by linear map whose range has dimension ≤ 𝑘
Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑠1 ≥ ⋯ ≥ 𝑠𝑚 are the positive singular values of 𝑇.

Suppose 1 ≤ 𝑘 < 𝑚. Then
min{‖𝑇 − 𝑆‖ ∶ 𝑆 ∈ ℒ(𝑉, 𝑊) and dim range 𝑆 ≤ 𝑘} = 𝑠𝑘 + 1 .
Furthermore, if
𝑇𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩ 𝑓𝑚
is a singular value decomposition of 𝑇 and 𝑇𝑘 ∈ ℒ(𝑉, 𝑊) is defined by
𝑇𝑘 𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑘 ⟨𝑣, 𝑒𝑘 ⟩ 𝑓𝑘
for each 𝑣 ∈ 𝑉, then dim range 𝑇𝑘 = 𝑘 and ‖𝑇 − 𝑇𝑘 ‖ = 𝑠𝑘 + 1 .
Proof If 𝑣 ∈ 𝑉 then
2 2
∥(𝑇 − 𝑇𝑘 )𝑣∥ = ∥𝑠𝑘 + 1 ⟨𝑣, 𝑒𝑘 + 1 ⟩ 𝑓𝑘 + 1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩ 𝑓𝑚 ∥
2 2
= 𝑠𝑘 + 12 ∣⟨𝑣, 𝑒𝑘 + 1 ⟩∣ + ⋯ + 𝑠𝑚2 ∣⟨𝑣, 𝑒𝑚 ⟩∣
2 2
≤ 𝑠𝑘 + 12 (∣⟨𝑣, 𝑒𝑘 + 1 ⟩∣ + ⋯ + ∣⟨𝑣, 𝑒𝑚 ⟩∣ )
≤ 𝑠𝑘 + 12 ‖𝑣‖2.
Thus ‖𝑇 − 𝑇𝑘 ‖ ≤ 𝑠𝑘 + 1 . The equation (𝑇 − 𝑇𝑘 )𝑒𝑘 + 1 = 𝑠𝑘 + 1 𝑓𝑘 + 1 now shows that
‖𝑇 − 𝑇𝑘 ‖ = 𝑠𝑘 + 1 .
Suppose 𝑆 ∈ ℒ(𝑉, 𝑊) and dim range 𝑆 ≤ 𝑘. Thus 𝑆𝑒1 , …, 𝑆𝑒𝑘 + 1 , which is a
list of length 𝑘 + 1, is linearly dependent. Hence there exist 𝑎1 , …, 𝑎𝑘 + 1 ∈ 𝐅, not
all 0, such that
𝑎1 𝑆𝑒1 + ⋯ + 𝑎𝑘 + 1 𝑆𝑒𝑘 + 1 = 0.
Now 𝑎1 𝑒1 + ⋯ + 𝑎𝑘 + 1 𝑒𝑘 + 1 ≠ 0 because 𝑎1 , …, 𝑎𝑘 + 1 are not all 0. We have
2 2
∥(𝑇 − 𝑆)(𝑎1 𝑒1 + ⋯ + 𝑎𝑘 + 1 𝑒𝑘 + 1 )∥ = ∥𝑇(𝑎1 𝑒1 + ⋯ + 𝑎𝑘 + 1 𝑒𝑘 + 1 )∥
= ‖𝑠1 𝑎1 𝑓1 + ⋯ + 𝑠𝑘 + 1 𝑎𝑘 + 1 𝑓𝑘 + 1 ‖2
= 𝑠12 |𝑎1 |2 + ⋯ + 𝑠𝑘 + 12 |𝑎𝑘 + 1 |2
≥ 𝑠𝑘 + 12 (|𝑎1 |2 + ⋯ + |𝑎𝑘 + 1 |2 )
= 𝑠𝑘 + 12 ‖𝑎1 𝑒1 + ⋯ + 𝑎𝑘 + 1 𝑒𝑘 + 1 ‖2.
Because 𝑎1 𝑒1 + ⋯ + 𝑎𝑘 + 1 𝑒𝑘 + 1 ≠ 0, the inequality above implies that
‖𝑇 − 𝑆‖ ≥ 𝑠𝑘 + 1 .
Thus 𝑆 = 𝑇𝑘 minimizes ‖𝑇 − 𝑆‖ among 𝑆 ∈ ℒ(𝑉, 𝑊) with dim range 𝑆 ≤ 𝑘.
For other examples of the use of the singular value decomposition in best
approximation, see Exercise 22, which finds a subspace of given dimension on
which the restriction of a linear map is as small as possible, and Exercise 27,
which finds a unitary operator that is as close as possible to a given operator.
Polar Decomposition
Recall our discussion before 7.54 of the analogy between complex numbers 𝑧
with |𝑧| = 1 and unitary operators. Continuing with this analogy, note that every
complex number 𝑧 except 0 can be written in the form
𝑧
𝑧=( )|𝑧|
|𝑧|
𝑧
=( )√𝑧𝑧,
|𝑧|
where the first factor, namely, 𝑧/|𝑧|, has absolute value 1.
Our analogy leads us to guess that every operator 𝑇 ∈ ℒ(𝑉) can be written as
a unitary operator times √𝑇 ∗ 𝑇. That guess is indeed correct. The corresponding
result is called the polar decomposition, which gives a beautiful description of an
arbitrary operator on 𝑉.
Note that if 𝑇 ∈ ℒ(𝑉), then 𝑇 ∗ 𝑇 is a positive operator [as was shown in
7.64(a)]. Thus the operator √𝑇 ∗ 𝑇 makes sense and is well defined as a positive
operator on 𝑉.
The polar decomposition that we are about to state and prove says that every
operator on 𝑉 is the product of a unitary operator and a positive operator. Thus
we can write an arbitrary operator on 𝑉 as the product of two nice operators,
each of which comes from a class that we can completely describe and that we
understand reasonably well. The unitary operators are described by 7.55 if 𝐅 = 𝐂 ;
the positive operators are described by the real and complex spectral theorems
(7.29 and 7.31).
Specifically, consider the case 𝐅 = 𝐂 , and suppose
𝑇 = 𝑆√𝑇 ∗ 𝑇
is a polar decomposition of an operator 𝑇 ∈ ℒ(𝑉), where 𝑆 is a unitary operator.

Then there is an orthonormal basis of 𝑉 with respect to which 𝑆 has a diagonal
matrix, and there is an orthonormal basis of 𝑉 with respect to which √𝑇 ∗ 𝑇 has
a diagonal matrix. Warning: There may not exist an orthonormal basis that
simultaneously puts the matrices of both 𝑆 and √𝑇 ∗ 𝑇 into these nice diagonal
forms—𝑆 may require one orthonormal basis and √𝑇 ∗ 𝑇 may require a different
orthonormal basis.
However (still assuming that 𝐅 = 𝐂 ), if 𝑇 is normal, then an orthonormal
basis of 𝑉 can be chosen such that both 𝑆 and √𝑇 ∗ 𝑇 have diagonal matrices with
respect to this basis—see Exercise 31. The converse is also true: If 𝑇 ∈ ℒ(𝑉)
and 𝑇 = 𝑆√𝑇 ∗ 𝑇 for some unitary operator 𝑆 ∈ ℒ(𝑉) such that 𝑆 and √𝑇 ∗ 𝑇 both
have diagonal matrices with respect to the same orthonormal basis of 𝑉, then 𝑇
is normal. This holds because 𝑇 then has a diagonal matrix with respect to this
same orthonormal basis, which implies that 𝑇 is normal [by the equivalence of
(c) and (a) in 7.31].
The polar decomposition below is valid on both real and complex inner product
spaces and for all operators on those spaces.
7.93 polar decomposition
Suppose 𝑇 ∈ ℒ(𝑉). Then there exists a unitary operator 𝑆 ∈ ℒ(𝑉) such that
𝑇 = 𝑆√𝑇 ∗ 𝑇.
Proof Let 𝑠1 , …, 𝑠𝑚 be the positive singular values of 𝑇, and let 𝑒1 , …, 𝑒𝑚 and

𝑓1 , …, 𝑓𝑚 be orthonormal lists in 𝑉 such that
7.94 𝑇𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩ 𝑓𝑚
for every 𝑣 ∈ 𝑉. Extend 𝑒1 , …, 𝑒𝑚 and 𝑓1 , …, 𝑓𝑚 to orthonormal bases 𝑒1 , …, 𝑒𝑛
and 𝑓1 , …, 𝑓𝑛 of 𝑉.
Define 𝑆 ∈ ℒ(𝑉) by
𝑆𝑣 = ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + ⟨𝑣, 𝑒𝑛 ⟩ 𝑓𝑛
for each 𝑣 ∈ 𝑉. Then
2
‖𝑆𝑣‖2 = ∥⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + ⟨𝑣, 𝑒𝑛 ⟩ 𝑓𝑛 ∥
2 2
= ∣⟨𝑣, 𝑒1 ⟩∣ + ⋯ + ∣⟨𝑣, 𝑒𝑛 ⟩∣
= ‖𝑣‖2.
Thus 𝑆 is a unitary operator.
Applying 𝑇 ∗ to both sides of 7.94 and then using the formula for 𝑇 ∗ given by
7.77 shows that
𝑇 ∗ 𝑇𝑣 = 𝑠12 ⟨𝑣, 𝑒1 ⟩𝑒1 + ⋯ + 𝑠𝑚2 ⟨𝑣, 𝑒𝑚 ⟩𝑒𝑚
for every 𝑣 ∈ 𝑉. Thus if 𝑣 ∈ 𝑉, then
√𝑇 ∗ 𝑇𝑣 = 𝑠 ⟨𝑣, 𝑒 ⟩𝑒 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩𝑒𝑚
1 1 1
because the operator that sends 𝑣 to the right side of the equation above is a
positive operator whose square equals 𝑇 ∗ 𝑇. Now
𝑆√𝑇 ∗ 𝑇𝑣 = 𝑆(𝑠1 ⟨𝑣, 𝑒1 ⟩𝑒1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩𝑒𝑚 )
= 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑚 ⟨𝑣, 𝑒𝑚 ⟩ 𝑓𝑚
= 𝑇𝑣,
where the last equation follows from 7.94.
Exercise 27 shows that the unitary operator 𝑆 produced in the proof above is
as close as a unitary operator can be to 𝑇.
Alternative proofs of the polar decomposition directly use the spectral theorem,
avoiding the singular value decomposition. However, the proof above seems
cleaner than those alternative proofs.
Operators Applied to Ellipsoids and Parallelepipeds
7.95 definition: ball, 𝐵
The ball in 𝑉 of radius 1 centered at 0, denoted by 𝐵, is defined by
𝐵 = {𝑣 ∈ 𝑉 ∶ ‖𝑣‖ < 1}.
If dim 𝑉 = 2, the word disk is sometimes used instead of

ball. However, using ball in all dimensions is less confusing.
Similarly, if dim 𝑉 = 2, then the word ellipse is sometimes
used instead of the word ellipsoid that we are about to define.
Again, using ellipsoid in all dimensions is less confusing.
You can think of the ellipsoid defined below as obtained
by starting with the ball 𝐵 and then stretching by a factor of
𝑠𝑘 along each 𝑓𝑘 axis. The ball 𝐵 in 𝐑2 .
7.96 definition: ellipsoid, 𝐸(𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 ), principal axes
Suppose that 𝑓1 , …, 𝑓𝑛 is an orthonormal basis of 𝑉 and 𝑠1 , …, 𝑠𝑛 are positive

numbers. The ellipsoid 𝐸(𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 ) with principal axes 𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 is
defined by
|⟨𝑣, 𝑓1 ⟩|2 |⟨𝑣, 𝑓𝑛 ⟩|2
𝐸(𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 ) = {𝑣 ∈ 𝑉 ∶ + ⋯ + < 1}.
𝑠12 𝑠𝑛2
The ellipsoid notation 𝐸(𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 ) does not explicitly include the inner
product space 𝑉, even though the definition above depends on 𝑉. However, the in-
ner product space 𝑉 should be clear from the context and also from the requirement
that 𝑓1 , …, 𝑓𝑛 be an orthonormal basis of 𝑉.
7.97 example: ellipsoids
The ellipsoid 𝐸(2 𝑓1 , 𝑓2 ) in 𝐑2 , where The ellipsoid 𝐸(2 𝑓1 , 𝑓2 ) in 𝐑2 , where

𝑓1 , 𝑓2 is the standard basis of 𝐑2 . 𝑓1 = ( √1 , √1 ) and 𝑓2 = (− √1 , √1 ).
2 2 2 2
The ellipsoid
𝐸(4 𝑓1 , 3 𝑓2 , 2 𝑓3 ) in 𝐑3 ,
where 𝑓1 , 𝑓2 , 𝑓3 is the
standard basis of 𝐑3 .
The ellipsoid 𝐸( 𝑓1 , …, 𝑓𝑛 ) equals the ball 𝐵 in 𝑉 for every orthonormal basis

𝑓1 , …, 𝑓𝑛 of 𝑉 [by Parseval’s identity 6.30(b)].
7.98 notation: 𝑇(Ω)
For 𝑇 a function defined on 𝑉 and Ω ⊆ 𝑉, define 𝑇(Ω) by
𝑇(Ω) = {𝑇𝑣 ∶ 𝑣 ∈ Ω}.
Thus if 𝑇 is a function defined on 𝑉, then 𝑇(𝑉) = range 𝑇.

The next result states that every invertible operator 𝑇 ∈ ℒ(𝑉) maps the ball
𝐵 in 𝑉 onto an ellipsoid in 𝑉. The proof shows that the principal axes of this
ellipsoid come from the singular value decomposition of 𝑇.
7.99 invertible operator takes ball to ellipsoid
Suppose 𝑇 ∈ ℒ(𝑉) is invertible. Then 𝑇 maps the ball 𝐵 in 𝑉 onto an ellipsoid

in 𝑉.
Proof Suppose 𝑇 has singular value decomposition

7.100 𝑇𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑛 ⟨𝑣, 𝑒𝑛 ⟩ 𝑓𝑛
for all 𝑣 ∈ 𝑉, where 𝑠1 , …, 𝑠𝑛 are the singular values of 𝑇 and 𝑒1 , …, 𝑒𝑛 and 𝑓1 , …, 𝑓𝑛
are both orthonormal bases of 𝑉. We will show that 𝑇(𝐵) = 𝐸(𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 ).
First suppose 𝑣 ∈ 𝐵. Because 𝑇 is invertible, none of the singular values
𝑠1 , …, 𝑠𝑛 equals 0 (see 7.68). Thus 7.100 implies that
2 2
∣⟨𝑇𝑣, 𝑓1 ⟩∣ ∣⟨𝑇𝑣, 𝑓𝑛 ⟩∣
2
+⋯+ = |⟨𝑣, 𝑒1 ⟩|2 + ⋯ + |⟨𝑣, 𝑒𝑛 ⟩|2 < 1.
𝑠1 𝑠𝑛2
Thus 𝑇𝑣 ∈ 𝐸(𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 ). Hence 𝑇(𝐵) ⊆ 𝐸(𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 ).
To prove inclusion in the other direction, now suppose 𝑤 ∈ 𝐸(𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 ).
Let ⟨𝑤, 𝑓1 ⟩ ⟨𝑤, 𝑓𝑛 ⟩
𝑣= 𝑒1 + ⋯ + 𝑒𝑛 .
𝑠1 𝑠𝑛
Then ‖𝑣‖ < 1 and 7.100 implies that 𝑇𝑣 = ⟨𝑤, 𝑓1 ⟩ 𝑓1 + ⋯ + ⟨𝑤, 𝑓𝑛 ⟩ 𝑓𝑛 = 𝑤. Thus
𝑇(𝐵) ⊇ 𝐸(𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 ).
We now use the previous result to show that invertible operators take all
ellipsoids, not just the ball of radius 1, to ellipsoids.
7.101 invertible operator takes ellipsoids to ellipsoids
Suppose 𝑇 ∈ ℒ(𝑉) is invertible and 𝐸 is an ellipsoid in 𝑉. Then 𝑇(𝐸) is an

ellipsoid in 𝑉.
Proof There exist orthonormal basis 𝑓1 , …, 𝑓𝑛 of 𝑉 and positive numbers 𝑠1 , …, 𝑠𝑛

such that 𝐸 = 𝐸(𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 ). Define 𝑆 ∈ ℒ(𝑉) by
𝑆(𝑎1 𝑓1 + ⋯ + 𝑎𝑛 𝑓𝑛 ) = 𝑎1 𝑠1 𝑓1 + ⋯ + 𝑎𝑛 𝑠𝑛 𝑓𝑛 .
Then 𝑆 maps the ball 𝐵 of 𝑉 onto 𝐸, as you can verify. Thus
𝑇(𝐸) = 𝑇(𝑆(𝐵)) = (𝑇𝑆)(𝐵).
The equation above and 7.99, applied to 𝑇𝑆, show that 𝑇(𝐸) is an ellipsoid in 𝑉.
Recall (see 3.95) that if 𝑢 ∈ 𝑉 and Ω ⊆ 𝑉 then 𝑢 + Ω is defined by

𝑢 + Ω = {𝑢 + 𝑤 ∶ 𝑤 ∈ Ω}.
Geometrically, the sets Ω and 𝑢 + Ω look the same, but they are in different
locations.
In the following definition, if dim 𝑉 = 2 then the word parallelogram is often
used instead of parallelepiped.
7.102 definition: 𝑃(𝑣1 , …, 𝑣𝑛 ), parallelepiped
Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Let
𝑃(𝑣1 , …, 𝑣𝑛 ) = {𝑎1 𝑣1 + ⋯ + 𝑎𝑛 𝑣𝑛 ∶ 𝑎1 , …, 𝑎𝑛 ∈ (0, 1)}.
A parallelepiped is a set of the form 𝑢 + 𝑃(𝑣1 , …, 𝑣𝑛 ) for some 𝑢 ∈ 𝑉. The

vectors 𝑣1 , …, 𝑣𝑛 are called the edges of this parallelepiped.
7.103 example: parallelepipeds
The parallelepiped A parallelepiped in 𝐑3 .

(0.3, 0.5) + 𝑃((1, 0), (1, 1)) in 𝐑2 .
7.104 invertible operator takes parallelepipeds to parallelepipeds
Suppose 𝑢 ∈ 𝑉 and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. Suppose 𝑇 ∈ ℒ(𝑉) is invertible.

Then
𝑇(𝑢 + 𝑃(𝑣1 , …, 𝑣𝑛 )) = 𝑇𝑢 + 𝑃(𝑇𝑣1 , …, 𝑇𝑣𝑛 ).
Proof Because 𝑇 is invertible, the list 𝑇𝑣1 , …, 𝑇𝑣𝑛 is a basis of 𝑉. The linearity
of 𝑇 implies that
𝑇(𝑢 + 𝑎1 𝑣1 + ⋯ + 𝑎𝑛 𝑣𝑛 ) = 𝑇𝑢 + 𝑎1 𝑇𝑣1 + ⋯ + 𝑎𝑛 𝑇𝑣𝑛
for all 𝑎1 , …, 𝑎𝑛 ∈ (0, 1). Thus 𝑇(𝑢 + 𝑃(𝑣1 , …, 𝑣𝑛 )) = 𝑇𝑢 + 𝑃(𝑇𝑣1 , …, 𝑇𝑣𝑛 ).
Just as the rectangles are distinguished among the parallelograms in 𝐑2 , we

give a special name to the parallelepipeds in 𝑉 whose defining edges are orthogonal
to each other.
7.105 definition: box

A box in 𝑉 is a set of the form
𝑢 + 𝑃(𝑟1 𝑒1 , …, 𝑟𝑛 𝑒𝑛 ),
where 𝑢 ∈ 𝑉 and 𝑟1 , …, 𝑟𝑛 are positive numbers and 𝑒1 , …, 𝑒𝑛 is an orthonormal
basis of 𝑉.
Note that in the special case of 𝐑2 each box is a rectangle, but the terminology
box can be used in all dimensions.
7.106 example: boxes
The box (1, 0) + 𝑃(√2 𝑒1 , √2 𝑒2 ), where The box 𝑃(𝑒1 , 2𝑒2 , 𝑒3 ), where 𝑒1 , 𝑒2 , 𝑒3
1
𝑒1 = ( √ , 1
) and 𝑒2 = (− √ ,
1 1
). is the standard basis of 𝐑3 .
2 √2 2 √2
Suppose 𝑇 ∈ ℒ(𝑉) is invertible. Then 𝑇 maps every parallelepiped in 𝑉

to a parallelepiped in 𝑉 (by 7.104). In particular, 𝑇 maps every box in 𝑉 to a
parallelepiped in 𝑉. This raises the question of whether 𝑇 maps some boxes in
𝑉 to boxes in 𝑉. The following result answers this question, with the help of the
singular value decomposition.
7.107 every invertible operator takes some boxes to boxes
Suppose 𝑇 ∈ ℒ(𝑉) is invertible. Suppose 𝑇 has singular value decomposition
𝑇𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑛 ⟨𝑣, 𝑒𝑛 ⟩ 𝑓𝑛 ,
where 𝑠1 , …, 𝑠𝑛 are the singular values of 𝑇 and 𝑒1 , …, 𝑒𝑛 and 𝑓1 , …, 𝑓𝑛 are

orthonormal bases of 𝑉 and the equation above holds for all 𝑣 ∈ 𝑉. Then 𝑇
maps the box 𝑢 + 𝑃(𝑟1 𝑒1 , …, 𝑟𝑛 𝑒𝑛 ) onto the box 𝑇𝑢 + 𝑃(𝑟1 𝑠1 𝑓1 , …, 𝑟𝑛 𝑠𝑛 𝑓𝑛 ) for
all positive numbers 𝑟1 , …, 𝑟𝑛 and all 𝑢 ∈ 𝑉.
Proof If 𝑎1 , …, 𝑎𝑛 ∈ (0, 1) and 𝑟1 , …, 𝑟𝑛 are positive numbers and 𝑢 ∈ 𝑉, then

𝑇(𝑢 + 𝑎1 𝑟1 𝑒1 + ⋯ + 𝑎𝑛 𝑟𝑛 𝑒𝑛 ) = 𝑇𝑢 + 𝑎1 𝑟1 𝑠1 𝑓1 + ⋯ + 𝑎𝑛 𝑟𝑛 𝑠𝑛 𝑓𝑛 .
Thus 𝑇(𝑢 + 𝑃(𝑟1 𝑒1 , …, 𝑟𝑛 𝑒𝑛 )) = 𝑇𝑢 + 𝑃(𝑟1 𝑠1 𝑓1 , …, 𝑟𝑛 𝑠𝑛 𝑓𝑛 ).
Volume via Singular Values

Our goal in this subsection is to understand how an operator changes the volume
of subsets of its domain. Because notions of volume belong to analysis rather
than to linear algebra, we will work only with an intuitive notion of volume. Our
intuitive approach to volume can be converted into appropriate correct definitions,
correct statements, and correct proofs using the machinery of analysis.
Our intuition about volume works best in real inner product spaces. Thus the
assumption that 𝐅 = 𝐑 will appear frequently in the rest of this subsection.
If dim 𝑉 = 𝑛, then by volume we will mean 𝑛-dimensional volume. You
should be familiar with this concept in 𝐑3 . When 𝑛 = 2, this is usually called area
instead of volume, but for consistency we use the word volume in all dimensions.
The most fundamental intuition about volume is that the volume of a box (whose
defining edges are by definition orthogonal to each other) is the product of the
lengths of the defining edges. Thus we make the following definition.
7.108 definition: volume of a box
Suppose 𝐅 = 𝐑 . If 𝑢 ∈ 𝑉 and 𝑟1 , …, 𝑟𝑛 are positive numbers and 𝑒1 , …, 𝑒𝑛 is

an orthonormal basis of 𝑉, then
volume(𝑢 + 𝑃(𝑟1 𝑒1 , …, 𝑟𝑛 𝑒𝑛 )) = 𝑟1 × ⋯ × 𝑟𝑛 .
The definition above agrees with the familiar formulas for the area (which we
are calling the volume) of a rectangle in 𝐑2 and for the volume of a box in 𝐑3 . For
example, the first box in Example 7.106 has two-dimensional volume (or area) 2
because the defining edges of that box have length √2 and √2. The second box
in Example 7.106 has three-dimensional volume 2 because the defining edges of
that box have length 1, 2, and 1.
To define the volume of a subset of 𝑉, approximate the

subset by a finite collection of disjoint boxes, and then add up
the volumes of the approximating collection of boxes. As we
approximate a subset of 𝑉 more accurately by disjoint unions
of more boxes, we get a better approximation to the volume.
These ideas should remind you of how the Riemann integral Volume of this
is defined by approximating the area under a curve by a disjoint ball ≈ sum of the
collection of rectangles. This discussion leads to the following volumes of the
nonrigorous but intuitive definition. five boxes.
7.109 definition: volume

Suppose 𝐅 = 𝐑 and Ω ⊆ 𝑉. Then the volume of Ω, denoted by volume Ω, is
approximately the sum of the volumes of a collection of disjoint boxes that
approximate Ω.
We are ignoring many reasonable questions by taking an intuitive approach to

volume. For example, if we approximate Ω by boxes with respect to one basis,
do we get the same volume if we approximate Ω by boxes with respect to a
different basis? If Ω1 and Ω2 are disjoint subsets of 𝑉, is volume(Ω1 ∪ Ω1 ) =
volume Ω1 + volume Ω2 ? Provided that we consider only reasonably nice subsets
of 𝑉, techniques of analysis show that both these questions have affirmative
answers that agree with our intuition about volume.
7.110 example: volume change by a linear map

Suppose that 𝑇 ∈ ℒ(𝐑2 ) is defined by
𝑇𝑣 = 2⟨𝑣, 𝑒1 ⟩𝑒1 + ⟨𝑣, 𝑒2 ⟩𝑒2 , where 𝑒1 , 𝑒2 is the
standard basis of 𝐑2 . This linear map stretches
by a factor of 2 along the 𝑒1 axis. The ball
approximated by five boxes above gets mapped
by 𝑇 to the ellipsoid shown here. Each of the
five boxes in the original figure gets mapped to Each box here has twice the width
a box of twice the width and the same height and the same height as the boxes in
as in the original figure. Hence each box gets the previous figure.
mapped to a box of twice the volume (area) as in the original figure. The sum
of the volumes of the five new boxes approximates the volume of the ellipsoid.
Thus 𝑇 changes the volume of the ball by a factor of 2.
In the example above, 𝑇 maps boxes with respect to the basis 𝑒1 , 𝑒2 to boxes
with respect to the same basis; thus we can see how 𝑇 changes volume. In general,
an operator maps boxes to parallelepipeds that are not boxes. However, if we
choose the right basis (coming from the singular value decomposition!), then
boxes with respect to that basis get mapped to boxes with respect to a possibly
different basis, as shown in 7.107. This observation leads to a natural proof of
the following result.
7.111 volume changes by a factor of the product of the singular values
Suppose 𝐅 = 𝐑 , 𝑇 ∈ ℒ(𝑉) is invertible, and Ω ⊆ 𝑉. Then
volume 𝑇(Ω) = (product of singular values of 𝑇)(volume Ω).
Proof Suppose 𝑇 has singular value decomposition

𝑇𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑛 ⟨𝑣, 𝑒𝑛 ⟩ 𝑓𝑛
for all 𝑣 ∈ 𝑉, where 𝑒1 , …, 𝑒𝑛 and 𝑓1 , …, 𝑓𝑛 are orthonormal bases of 𝑉.
Approximate Ω by boxes of the form 𝑢 + 𝑃(𝑟1 𝑒1 , …, 𝑟𝑛 𝑒𝑛 ), which have volume
𝑟1 × ⋯ × 𝑟𝑛 . The operator 𝑇 maps each box 𝑢 + 𝑃(𝑟1 𝑒1 , …, 𝑟𝑛 𝑒𝑛 ) onto the box
𝑇𝑢 + 𝒫(𝑟1 𝑠1 𝑓1 , …, 𝑟𝑛 𝑠𝑛 𝑓𝑛 ), which has volume (𝑠1 × ⋯ × 𝑠𝑛 )(𝑟1 × ⋯ × 𝑟𝑛 ).
The operator 𝑇 maps a collection of boxes that approximate Ω onto a collection
of boxes that approximate 𝑇(Ω). Because 𝑇 changes the volume of each box in a
collection that approximates Ω by a factor of 𝑠1 × ⋯ × 𝑠𝑛 , the linear map 𝑇 changes
the volume of Ω by the same factor.
Suppose 𝑇 ∈ ℒ(𝑉). As we will see when we get to determinants, the product

of the singular values of 𝑇 equals |det 𝑇|; see 9.60 and 9.61.
Properties of an Operator as Determined by Its Eigenvalues

We conclude this chapter by presenting the table below. The context of this
table is a finite-dimensional complex inner product space. The first column of
the table shows a property that a normal operator on such a space might have.
The second column of the table shows a subset of 𝐂 such that the operator has
the corresponding property if and only if all eigenvalues of the operator lie in
the specified subset. For example, the first row of the table states that a normal
operator is invertible if and only if all its eigenvalues are nonzero (this first row
is the only one in the table that does not need the hypothesis that the operator is
normal).
Make sure you can explain why all results in the table hold. For example,
the last row of the table holds because the norm of an operator equals its largest
singular value (by 7.85) and the singular values of a normal operator, assuming
𝐅 = 𝐂 , equal the absolute values of the eigenvalues (by Exercise 7 in Section 7E).
properties of a normal operator eigenvalues are contained in

invertible 𝐂\{0}
self-adjoint 𝐑
skew {𝜆 ∈ 𝐂 ∶ Re 𝜆 = 0}
orthogonal projection {0, 1}
positive [0, ∞)
unitary {𝜆 ∈ 𝐂 ∶ |𝜆| = 1}
norm is less than 1 {𝜆 ∈ 𝐂 ∶ |𝜆| < 1}
Exercises 7F
1 Prove that if 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊), then ∣ ‖𝑆‖ − ‖𝑇‖ ∣ ≤ ‖𝑆 − 𝑇‖.

The inequality above is called the reverse triangle inequality.
2 Suppose that 𝑇 ∈ ℒ(𝑉) is self-adjoint or that 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉) is

normal. Prove that
‖𝑇‖ = max{|𝜆| ∶ 𝜆 is an eigenvalue of 𝑇}.
3 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑣 ∈ 𝑉. Prove that
‖𝑇𝑣‖ = ‖𝑇‖ ‖𝑣‖ ⟺ 𝑇 ∗ 𝑇𝑣 = ‖𝑇‖2 𝑣.
4 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊), 𝑣 ∈ 𝑉, and ‖𝑇𝑣‖ = ‖𝑇‖ ‖𝑣‖. Prove that if 𝑢 ∈ 𝑉 and
⟨𝑢, 𝑣⟩ = 0, then ⟨𝑇𝑢, 𝑇𝑣⟩ = 0.
5 Suppose 𝑈 is a finite-dimensional inner product space, 𝑇 ∈ ℒ(𝑉, 𝑈), and
𝑆 ∈ ℒ(𝑈, 𝑊). Prove that
‖𝑆𝑇‖ ≤ ‖𝑆‖ ‖𝑇‖.
6 Prove or give a counterexample: If 𝑆, 𝑇 ∈ ℒ(𝑉), then ‖𝑆𝑇‖ = ‖𝑇𝑆‖.
7 Show that defining 𝑑(𝑆, 𝑇) = ‖𝑆 − 𝑇‖ for 𝑆, 𝑇 ∈ ℒ(𝑉, 𝑊) makes 𝑑 a metric
on ℒ(𝑉, 𝑊).
This exercise is intended for readers who are familiar with metric spaces.
8 (a) Prove that if 𝑇 ∈ ℒ(𝑉) and ‖𝐼 − 𝑇‖ < 1, then 𝑇 is invertible.

(b) Suppose that 𝑆 ∈ ℒ(𝑉) is invertible. Prove that if 𝑇 ∈ ℒ(𝑉) and
‖𝑆 − 𝑇‖ < 1/∥𝑆−1 ∥, then 𝑇 is invertible.
This exercise shows that the set of invertible operators in ℒ(𝑉) is an open
subset of ℒ(𝑉), using the metric defined in Exercise 7.
9 Suppose 𝑇 ∈ ℒ(𝑉). Prove that for every 𝜖 > 0, there exists an invertible
operator 𝑆 ∈ ℒ(𝑉) such that 0 < ‖𝑇 − 𝑆‖ < 𝜖.
10 Suppose dim 𝑉 > 1 and 𝑇 ∈ ℒ(𝑉) is not invertible. Prove that for every
𝜖 > 0, there exists 𝑆 ∈ ℒ(𝑉) such that 0 < ‖𝑇 − 𝑆‖ < 𝜖 and 𝑆 is not
invertible.
11 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Prove that for every 𝜖 > 0 there exists a
diagonalizable operator 𝑆 ∈ ℒ(𝑉) such that 0 < ‖𝑇 − 𝑆‖ < 𝜖.
12 Suppose 𝑇 ∈ ℒ(𝑉) is a positive operator. Show that ∥√𝑇 ∥ = √‖𝑇‖.

13 Suppose 𝑆, 𝑇 ∈ ℒ(𝑉) are positive operators. Show that
‖𝑆 − 𝑇‖ ≤ max{‖𝑆‖, ‖𝑇‖} ≤ ‖𝑆 + 𝑇‖.
14 Suppose 𝑈 and 𝑊 are subspaces of 𝑉 such that ‖𝑃𝑈 − 𝑃𝑊 ‖ < 1. Prove that
dim 𝑈 = dim 𝑊.
15 Define 𝑇 ∈ ℒ(𝐅3 ) by
𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (𝑧3 , 2𝑧1 , 3𝑧2 ).
Find (explicitly) a unitary operator 𝑆 ∈ ℒ(𝐅3 ) such that 𝑇 = 𝑆√𝑇 ∗ 𝑇.

16 Suppose 𝑆 ∈ ℒ(𝑉) is a positive invertible operator. Prove that there exists
𝛿 > 0 such that 𝑇 is a positive operator for every self-adjoint operator
𝑇 ∈ ℒ(𝑉) with ‖𝑆 − 𝑇‖ < 𝛿.
17 Prove that if 𝑢 ∈ 𝑉 and 𝜑𝑢 is the linear functional on 𝑉 defined by the
equation 𝜑𝑢 (𝑣) = ⟨𝑣, 𝑢⟩, then ‖𝜑𝑢 ‖ = ‖𝑢‖.
Here we are thinking of the scalar field 𝐅 as an inner product space with
⟨𝛼, 𝛽⟩ = 𝛼𝛽 for all 𝛼, 𝛽 ∈ 𝐅. Thus ‖𝜑𝑢 ‖ means the norm of 𝜑𝑢 as a linear
map from 𝑉 to 𝐅.
18 Suppose 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉 and 𝑇 ∈ ℒ(𝑉, 𝑊).

(a) Prove that max{‖𝑇𝑒1 ‖, …, ‖𝑇𝑒𝑛 ‖} ≤ ‖𝑇‖ ≤ (‖𝑇𝑒1 ‖2 + ⋯ + ‖𝑇𝑒𝑛 ‖2 )1/2.
(b) Prove that ‖𝑇‖ = (‖𝑇𝑒1 ‖2 + ⋯ + ‖𝑇𝑒𝑛 ‖2 )1/2 if and only if dim range 𝑇 ≤ 1.
Here 𝑒1 , …, 𝑒𝑛 is an arbitrary orthonormal basis of 𝑉, not necessarily con-
nected with a singular value decomposition of 𝑇. If 𝑠1 , …, 𝑠𝑛 is the list
of singular values of 𝑇, then the right side of the inequality above equals
(𝑠12 + ⋯ + 𝑠𝑛2 ) , as was shown in Exercise 11(a) in Section 7E.
1/2
19 Prove that if 𝑇 ∈ ℒ(𝑉, 𝑊), then ∥𝑇 ∗ 𝑇∥ = ‖𝑇‖2.

This formula for ∥𝑇 ∗ 𝑇∥ leads to the important subject of 𝐶∗ -algebras.
20 Suppose 𝑇 ∈ ℒ(𝑉) is normal. Prove that ∥𝑇 𝑘 ∥ = ‖𝑇‖𝑘 for every positive

integer 𝑘.
21 Suppose dim 𝑉 > 1 and dim 𝑊 > 1. Prove that the norm on ℒ(𝑉, 𝑊) does
not come from an inner product. In other words, prove that there does not
exist an inner product on ℒ(𝑉, 𝑊) such that
max{‖𝑇𝑣‖ ∶ 𝑣 ∈ 𝑉 and ‖𝑣‖ ≤ 1} = √⟨𝑇, 𝑇⟩

for all 𝑇 ∈ ℒ(𝑉, 𝑊).
22 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Let 𝑛 = dim 𝑉 and let 𝑠1 ≥ ⋯ ≥ 𝑠𝑛 denote the
singular values of 𝑇. Prove that if 1 ≤ 𝑘 ≤ 𝑛, then
min{‖𝑇|𝑈 ‖ ∶ 𝑈 is a subspace of 𝑉 with dim 𝑈 = 𝑘} = 𝑠𝑛 − 𝑘 + 1 .
23 Suppose 𝑇 ∈ ℒ(𝑉, 𝑊). Show that 𝑇 is uniformly continuous with respect
to the metrics on 𝑉 and 𝑊 that arise from the norms on those spaces (see
Exercise 23 in Section 6B).
𝑇
∥𝑇 −1 ∥ = ‖𝑇‖−1 ⟺ is a unitary operator.
‖𝑇‖
25 Fix 𝑢, 𝑥 ∈ 𝑉 with 𝑢 ≠ 0. Define 𝑇 ∈ ℒ(𝑉) by 𝑇𝑣 = ⟨𝑣, 𝑢⟩𝑥 for every 𝑣 ∈ 𝑉.

Prove that
‖𝑥‖
√𝑇 ∗ 𝑇𝑣 = ⟨𝑣, 𝑢⟩𝑢
‖𝑢‖
26 Suppose 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is invertible if and only if there exists a
unique unitary operator 𝑆 ∈ ℒ(𝑉) such that 𝑇 = 𝑆√𝑇 ∗ 𝑇.
27 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑠1 , …, 𝑠𝑛 are the singular values of 𝑇. Let 𝑒1 , …, 𝑒𝑛
and 𝑓1 , …, 𝑓𝑛 be orthonormal bases of 𝑉 such that
𝑇𝑣 = 𝑠1 ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + 𝑠𝑛 ⟨𝑣, 𝑒𝑛 ⟩ 𝑓𝑛
for all 𝑣 ∈ 𝑉. Define 𝑆 ∈ ℒ(𝑉) by
𝑆𝑣 = ⟨𝑣, 𝑒1 ⟩ 𝑓1 + ⋯ + ⟨𝑣, 𝑒𝑛 ⟩ 𝑓𝑛 .
(a) Show that 𝑆 is unitary and ‖𝑇 − 𝑆‖ = max{|𝑠1 − 1|, …, |𝑠𝑛 − 1|}.
(b) Show that if 𝐸 ∈ ℒ(𝑉) is unitary, then ‖𝑇 − 𝐸‖ ≥ ‖𝑇 − 𝑆‖.
This exercise finds a unitary operator 𝑆 that is as close as possible (among
the unitary operators) to a given operator 𝑇.
28 Suppose 𝑇 ∈ ℒ(𝑉). Prove that there exists a unitary operator 𝑆 ∈ ℒ(𝑉)

such that 𝑇 = √𝑇𝑇 ∗ 𝑆.
29 Suppose 𝑇 ∈ ℒ(𝑉).
(a) Use the polar decomposition to show that there exists a unitary operator
𝑆 ∈ ℒ(𝑉) such that 𝑇𝑇 ∗ = 𝑆𝑇 ∗ 𝑇𝑆∗.
(b) Show how (a) implies that 𝑇 and 𝑇 ∗ have the same singular values.
30 Suppose 𝑇 ∈ ℒ(𝑉), 𝑆 ∈ ℒ(𝑉) is a unitary operator, and 𝑅 ∈ ℒ(𝑉) is a

positive operator such that 𝑇 = 𝑆𝑅. Prove that 𝑅 = √𝑇 ∗ 𝑇.
This exercise shows that if we write 𝑇 as the product of a unitary operator
and a positive operator (as in the polar decomposition 7.93), then the
positive operator equals √𝑇 ∗ 𝑇.
31 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉) is normal. Prove that there exists a unitary

operator 𝑆 ∈ ℒ(𝑉) such that 𝑇 = 𝑆√𝑇 ∗ 𝑇 and such that 𝑆 and √𝑇 ∗ 𝑇 both
have diagonal matrices with respect to the same orthonormal basis of 𝑉.
32 Suppose that 𝑇 ∈ ℒ(𝑉, 𝑊) and 𝑇 ≠ 0. Let 𝑠1 , …, 𝑠𝑚 denote the positive
singular values of 𝑇. Show that there exists an orthonormal basis 𝑒1 , …, 𝑒𝑚
of (null 𝑇)⟂ such that
𝑒1 𝑒𝑚
𝑇(𝐸( , …, ))
𝑠1 𝑠𝑚
equals the ball in range 𝑇 of radius 1 centered at 0.
Chapter 8
Operators on Complex Vector Spaces
In this chapter we delve deeper into the structure of operators, with most of the
attention on complex vector spaces. Some of the results in this chapter apply to
both real and complex vector spaces; thus we do not make a standing assumption
that 𝐅 = 𝐂 . Also, an inner product does not help with this material, so we return
to the general setting of a finite-dimensional vector space.
Even on a finite-dimensional complex vector space, an operator may not have
enough eigenvectors to form a basis of the vector space. Thus we will consider the
closely related objects called generalized eigenvectors. We will see that for each
operator on a finite-dimensional complex vector space, there is a basis of the vector
space consisting of generalized eigenvectors of the operator. The generalized
eigenspace decomposition then provides a good description of arbitrary operators
on a finite-dimensional complex vector space.
Nilpotent operators, which are operators that when raised to some power
equal 0, have an important role in these investigations. Nilpotent operators provide
a key tool in our proof that every invertible operator on a finite-dimensional
complex vector space has a square root and in our approach to Jordan form.
This chapter concludes by defining the trace and proving its key properties.
• 𝑉 denotes a finite-dimensional nonzero vector space over 𝐅.
David Iliff CC BY-SA
The Long Room of the Old Library at the University of Dublin, where William Hamilton
(1805–1865) was a student and then a faculty member. Hamilton proved a special case
of what we now call the Cayley–Hamilton theorem in 1853.
297
https://doi.org/10.1007/978-3-031-41026-0_8
298 Chapter 8 Operators on Complex Vector Spaces
8A Generalized Eigenvectors and Nilpotent Operators

Null Spaces of Powers of an Operator
We begin this chapter with a study of null spaces of powers of an operator.
8.1 sequence of increasing null spaces
{0} = null 𝑇 0 ⊆ null 𝑇 1 ⊆ ⋯ ⊆ null 𝑇 𝑘 ⊆ null 𝑇 𝑘 + 1 ⊆ ⋯ .
Proof Suppose 𝑘 is a nonnegative integer and 𝑣 ∈ null 𝑇 𝑘. Then 𝑇 𝑘 𝑣 = 0,

which implies that 𝑇 𝑘 + 1 𝑣 = 𝑇(𝑇 𝑘 𝑣) = 𝑇(0) = 0. Thus 𝑣 ∈ null 𝑇 𝑘 + 1. Hence
null 𝑇 𝑘 ⊆ null 𝑇 𝑘 + 1, as desired.
The following result states that if two For similar results about decreasing
consecutive terms in the sequence of sub- sequences of ranges, see Exercises 6,
spaces above are equal, then all later 7, and 8.
terms in the sequence are equal.
8.2 equality in the sequence of null spaces
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑚 is a nonnegative integer such that
null 𝑇 𝑚 = null 𝑇 𝑚 + 1.
Then
null 𝑇 𝑚 = null 𝑇 𝑚 + 1 = null 𝑇 𝑚 + 2 = null 𝑇 𝑚 + 3 = ⋯ .
Proof Let 𝑘 be a positive integer. We want to prove that
null 𝑇 𝑚 + 𝑘 = null 𝑇 𝑚 + 𝑘 + 1.
We already know from 8.1 that null 𝑇 𝑚 + 𝑘 ⊆ null 𝑇 𝑚 + 𝑘 + 1.

To prove the inclusion in the other direction, suppose 𝑣 ∈ null 𝑇 𝑚 + 𝑘 + 1. Then
𝑇 𝑚 + 1 (𝑇 𝑘 𝑣) = 𝑇 𝑚 + 𝑘 + 1 𝑣 = 0.
Hence
𝑇 𝑘 𝑣 ∈ null 𝑇 𝑚 + 1 = null 𝑇 𝑚.
Thus 𝑇 𝑚 + 𝑘 𝑣 = 𝑇 𝑚 (𝑇 𝑘 𝑣) = 0, which means that 𝑣 ∈ null 𝑇 𝑚 + 𝑘. This implies that
null 𝑇 𝑚 + 𝑘 + 1 ⊆ null 𝑇 𝑚 + 𝑘, completing the proof.
The result above raises the question of whether there exists a nonnegative
integer 𝑚 such that null 𝑇 𝑚 = null 𝑇 𝑚 + 1. The next result shows that this equality
holds at least when 𝑚 equals the dimension of the vector space on which 𝑇
operates.
Section 8A Generalized Eigenvectors and Nilpotent Operators 299
8.3 null spaces stop growing
null 𝑇 dim 𝑉 = null 𝑇 dim 𝑉 + 1 = null 𝑇 dim 𝑉 + 2 = ⋯ .
Proof We only need to prove that null 𝑇 dim 𝑉 = null 𝑇 dim 𝑉 + 1 (by 8.2). Suppose
this is not true. Then, by 8.1 and 8.2, we have
{0} = null 𝑇 0 ⊊ null 𝑇 1 ⊊ ⋯ ⊊ null 𝑇 dim 𝑉 ⊊ null 𝑇 dim 𝑉 + 1,
where the symbol ⊊ means “contained in but not equal to”. At each of the
strict inclusions in the chain above, the dimension increases by at least 1. Thus
dim null 𝑇 dim 𝑉 + 1 ≥ dim 𝑉 + 1, a contradiction because a subspace of 𝑉 cannot
have a larger dimension than dim 𝑉.
It is not true that 𝑉 = null 𝑇 ⊕ range 𝑇 for every 𝑇 ∈ ℒ(𝑉). However, the
next result can be a useful substitute.
8.4 𝑉 is the direct sum of null 𝑇 dim 𝑉 and range 𝑇 dim 𝑉
𝑉 = null 𝑇 dim 𝑉 ⊕ range 𝑇 dim 𝑉.
Proof Let 𝑛 = dim 𝑉. First we show that
8.5 (null 𝑇 𝑛 ) ∩ (range 𝑇 𝑛 ) = {0}.
Suppose 𝑣 ∈ (null 𝑇 𝑛 ) ∩ (range 𝑇 𝑛 ). Then 𝑇 𝑛 𝑣 = 0, and there exists 𝑢 ∈ 𝑉

such that 𝑣 = 𝑇 𝑛 𝑢. Applying 𝑇 𝑛 to both sides of the last equation shows that
𝑇 𝑛 𝑣 = 𝑇 2𝑛 𝑢. Hence 𝑇 2𝑛 𝑢 = 0, which implies that 𝑇 𝑛 𝑢 = 0 (by 8.3). Thus
𝑣 = 𝑇 𝑛 𝑢 = 0, completing the proof of 8.5.
Now 8.5 implies that null 𝑇 𝑛 + range 𝑇 𝑛 is a direct sum (by 1.46). Also,
dim(null 𝑇 𝑛 ⊕ range 𝑇 𝑛 ) = dim null 𝑇 𝑛 + dim range 𝑇 𝑛 = dim 𝑉,
where the first equality above comes from 3.94 and the second equality comes
from the fundamental theorem of linear maps (3.21). The equation above implies
that null 𝑇 𝑛 ⊕ range 𝑇 𝑛 = 𝑉 (see 2.39), as desired.
For an improvement of the result above, see Exercise 19.
8.6 example: 𝐅3 = null 𝑇 3 ⊕ range 𝑇 3 for 𝑇 ∈ ℒ(𝐅3 )

Suppose 𝑇 ∈ ℒ(𝐅3 ) is defined by
𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (4𝑧2 , 0, 5𝑧3 ).

Then null 𝑇 = {(𝑧1 , 0, 0) ∶ 𝑧1 ∈ 𝐅} and range 𝑇 = {(𝑧1 , 0, 𝑧3 ) ∶ 𝑧1 , 𝑧3 ∈ 𝐅}. Thus

null 𝑇 ∩ range 𝑇 ≠ {0}. Hence null 𝑇 + range 𝑇 is not a direct sum. Also note that
null 𝑇 + range 𝑇 ≠ 𝐅3. However, we have 𝑇 3 (𝑧1 , 𝑧2 , 𝑧3 ) = (0, 0, 125𝑧3 ). Thus we
see that
null 𝑇 3 = {(𝑧1 , 𝑧2 , 0) ∶ 𝑧1 , 𝑧2 ∈ 𝐅} and range 𝑇 3 = {(0, 0, 𝑧3 ) ∶ 𝑧3 ∈ 𝐅}.
Hence 𝐅3 = null 𝑇 3 ⊕ range 𝑇 3, as expected by 8.4.
Generalized Eigenvectors
Some operators do not have enough eigenvectors to lead to good descriptions of
their behavior. Thus in this subsection we introduce the concept of generalized
eigenvectors, which will play a major role in our description of the structure of an
operator.
To understand why we need more than eigenvectors, let’s examine the question
of describing an operator by decomposing its domain into invariant subspaces. Fix
𝑇 ∈ ℒ(𝑉). We seek to describe 𝑇 by finding a “nice” direct sum decomposition
𝑉 = 𝑉1 ⊕ ⋯ ⊕ 𝑉𝑚 ,
where each 𝑉𝑘 is a subspace of 𝑉 invariant under 𝑇. The simplest possible nonzero
invariant subspaces are one-dimensional. A decomposition as above in which
each 𝑉𝑘 is a one-dimensional subspace of 𝑉 invariant under 𝑇 is possible if and
only if 𝑉 has a basis consisting of eigenvectors of 𝑇 (see 5.55). This happens if
and only if 𝑉 has an eigenspace decomposition
8.7 𝑉 = 𝐸(𝜆1 , 𝑇) ⊕ ⋯ ⊕ 𝐸(𝜆𝑚 , 𝑇),
where 𝜆1 , …, 𝜆𝑚 are the distinct eigenvalues of 𝑇 (see 5.55).
The spectral theorem in the previous chapter shows that if 𝑉 is an inner product
space, then a decomposition of the form 8.7 holds for every self-adjoint operator
if 𝐅 = 𝐑 and for every normal operator if 𝐅 = 𝐂 because operators of those types
have enough eigenvectors to form a basis of 𝑉 (see 7.29 and 7.31).
However, a decomposition of the form 8.7 may not hold for more general
operators, even on a complex vector space. An example was given by the operator
in 5.57, which does not have enough eigenvectors for 8.7 to hold. Generalized
eigenvectors and generalized eigenspaces, which we now introduce, will remedy
this situation.
8.8 definition: generalized eigenvector
Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆 is an eigenvalue of 𝑇. A vector 𝑣 ∈ 𝑉 is called a

generalized eigenvector of 𝑇 corresponding to 𝜆 if 𝑣 ≠ 0 and
(𝑇 − 𝜆𝐼)𝑘 𝑣 = 0
for some positive integer 𝑘.

A nonzero vector 𝑣 ∈ 𝑉 is a general- Generalized eigenvalues are not de-

ized eigenvector of 𝑇 corresponding to 𝜆 fined because doing so would not lead
if and only if to anything new. Reason: if (𝑇 − 𝜆𝐼)𝑘
is not injective for some positive inte-
(𝑇 − 𝜆𝐼)dim 𝑉 𝑣 = 0,
ger 𝑘, then 𝑇 − 𝜆𝐼 is not injective, and
as follows from applying 8.1 and 8.3 to hence 𝜆 is an eigenvalue of 𝑇.
the operator 𝑇 − 𝜆𝐼.
As we know, an operator on a complex vector space may not have enough
eigenvectors to form a basis of the domain. The next result shows that on a
complex vector space there are enough generalized eigenvectors to do this.
8.9 a basis of generalized eigenvectors
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Then there is a basis of 𝑉 consisting of

generalized eigenvectors of 𝑇.
Proof Let 𝑛 = dim 𝑉. We will use induction on 𝑛. To get started, note that
the desired result holds if 𝑛 = 1 because then every nonzero vector in 𝑉 is an
eigenvector of 𝑇.
Now suppose 𝑛 > 1 and the de- This step is where we use the hypothesis
sired result holds for all smaller values that 𝐅 = 𝐂 , because if 𝐅 = 𝐑 then 𝑇
of dim 𝑉. Let 𝜆 be an eigenvalue of 𝑇. may not have any eigenvalues.
Applying 8.4 to 𝑇 − 𝜆𝐼 shows that
𝑉 = null(𝑇 − 𝜆𝐼)𝑛 ⊕ range(𝑇 − 𝜆𝐼)𝑛.
If null(𝑇 − 𝜆𝐼)𝑛 = 𝑉, then every nonzero vector in 𝑉 is a generalized eigen-

vector of 𝑇, and thus in this case there is a basis of 𝑉 consisting of generalized
eigenvectors of 𝑇. Hence we can assume that null(𝑇 − 𝜆𝐼)𝑛 ≠ 𝑉, which implies
that range(𝑇 − 𝜆𝐼)𝑛 ≠ {0}.
Also, null(𝑇 − 𝜆𝐼)𝑛 ≠ {0}, because 𝜆 is an eigenvalue of 𝑇. Thus we have
0 < dim range(𝑇 − 𝜆𝐼)𝑛 < 𝑛.
Furthermore, range(𝑇 − 𝜆𝐼)𝑛 is invariant under 𝑇 [by 5.18 with 𝑝(𝑧) = (𝑧 − 𝜆)𝑛 ].
Let 𝑆 ∈ ℒ(range(𝑇 − 𝜆𝐼)𝑛 ) equal 𝑇 restricted to range(𝑇 − 𝜆𝐼)𝑛. Our induction
hypothesis applied to the operator 𝑆 implies that there is a basis of range(𝑇 − 𝜆𝐼)𝑛
consisting of generalized eigenvectors of 𝑆, which of course are generalized
eigenvectors of 𝑇. Adjoining that basis of range(𝑇−𝜆𝐼)𝑛 to a basis of null(𝑇−𝜆𝐼)𝑛
gives a basis of 𝑉 consisting of generalized eigenvectors of 𝑇.
If 𝐅 = 𝐑 and dim 𝑉 > 1, then some operators on 𝑉 have the property that
there exists a basis of 𝑉 consisting of generalized eigenvectors of the operator,
and (unlike what happens when 𝐅 = 𝐂 ) other operators do not have this property.
See Exercise 11 for a necessary and sufficient condition that determines whether
an operator has this property.
8.10 example: generalized eigenvectors of an operator on 𝐂3

𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (4𝑧2 , 0, 5𝑧3 )
for each (𝑧1 , 𝑧2 , 𝑧3 ) ∈ 𝐂3. A routine use of the definition of eigenvalue shows that
the eigenvalues of 𝑇 are 0 and 5. Furthermore, the eigenvectors corresponding to
the eigenvalue 0 are the nonzero vectors of the form (𝑧1 , 0, 0), and the eigenvectors
corresponding to the eigenvalue 5 are the nonzero vectors of the form (0, 0, 𝑧3 ).
Hence this operator does not have enough eigenvectors to span its domain 𝐂3.
We compute that 𝑇 3 (𝑧1 , 𝑧2 , 𝑧3 ) = (0, 0, 125𝑧3 ). Thus 8.1 and 8.3 imply that the
generalized eigenvectors of 𝑇 corresponding to the eigenvalue 0 are the nonzero
vectors of the form (𝑧1 , 𝑧2 , 0).
We also have (𝑇 − 5𝐼)3 (𝑧1 , 𝑧2 , 𝑧3 ) = (−125𝑧1 + 300𝑧2 , −125𝑧2 , 0). Thus the
generalized eigenvectors of 𝑇 corresponding to the eigenvalue 5 are the nonzero
vectors of the form (0, 0, 𝑧3 ).
The paragraphs above show that each of the standard basis vectors of 𝐂3 is a
generalized eigenvector of 𝑇. Thus 𝐂3 indeed has a basis consisting of generalized
eigenvectors of 𝑇, as promised by 8.9.
If 𝑣 is an eigenvector of 𝑇 ∈ ℒ(𝑉), then the corresponding eigenvalue 𝜆 is

uniquely determined by the equation 𝑇𝑣 = 𝜆𝑣, which can be satisfied by only one
𝜆 ∈ 𝐅 (because 𝑣 ≠ 0). However, if 𝑣 is a generalized eigenvector of 𝑇, then it
is not obvious that the equation (𝑇 − 𝜆𝐼)dim 𝑉 𝑣 = 0 can be satisfied by only one
𝜆 ∈ 𝐅. Fortunately, the next result tells us that all is well on this issue.
8.11 generalized eigenvector corresponds to a unique eigenvalue
Suppose 𝑇 ∈ ℒ(𝑉). Then each generalized eigenvector of 𝑇 corresponds to

only one eigenvalue of 𝑇.
Proof Suppose 𝑣 ∈ 𝑉 is a generalized eigenvector of 𝑇 corresponding to eigen-

values 𝛼 and 𝜆 of 𝑇. Let 𝑚 be the smallest positive integer such that (𝑇−𝛼𝐼)𝑚 𝑣 = 0.
Let 𝑛 = dim 𝑉. Then
0 = (𝑇 − 𝜆𝐼)𝑛 𝑣
𝑛
= ((𝑇 − 𝛼𝐼) + (𝛼 − 𝜆)𝐼) 𝑣
𝑛
= ∑ 𝑏𝑘 (𝛼 − 𝜆)𝑛 − 𝑘 (𝑇 − 𝛼𝐼)𝑘 𝑣,
𝑘=0
where 𝑏0 = 1 and the values of the other binomial coefficients 𝑏𝑘 do not matter.
Apply the operator (𝑇 − 𝛼𝐼)𝑚 − 1 to both sides of the equation above, getting
0 = (𝛼 − 𝜆)𝑛 (𝑇 − 𝛼𝐼)𝑚 − 1 𝑣.
Because (𝑇 − 𝛼𝐼)𝑚 − 1 𝑣 ≠ 0, the equation above implies that 𝛼 = 𝜆, as desired.
We saw earlier (5.11) that eigenvectors corresponding to distinct eigenvalues

are linearly independent. Now we prove a similar result for generalized eigen-
vectors, with a proof that roughly follows the pattern of the proof of that earlier
result.
8.12 linearly independent generalized eigenvectors
Suppose that 𝑇 ∈ ℒ(𝑉). Then every list of generalized eigenvectors of 𝑇

corresponding to distinct eigenvalues of 𝑇 is linearly independent.
Proof Suppose the desired result is false. Then there exists a smallest positive
integer 𝑚 such that there exists a linearly dependent list 𝑣1 , …, 𝑣𝑚 of generalized
eigenvectors of 𝑇 corresponding to distinct eigenvalues 𝜆1 , …, 𝜆𝑚 of 𝑇 (note that
𝑚 ≥ 2 because a generalized eigenvector is, by definition, nonzero). Thus there
exist 𝑎1 , …, 𝑎𝑚 ∈ 𝐅, none of which are 0 (because of the minimality of 𝑚), such
that
𝑎1 𝑣1 + ⋯ + 𝑎𝑚 𝑣𝑚 = 0.
Let 𝑛 = dim 𝑉. Apply (𝑇 − 𝜆𝑚 𝐼)𝑛 to both sides of the equation above, getting
8.13 𝑎1 (𝑇 − 𝜆𝑚 𝐼)𝑛 𝑣1 + ⋯ + 𝑎𝑚 − 1 (𝑇 − 𝜆𝑚 𝐼)𝑛 𝑣𝑚 − 1 = 0.
Suppose 𝑘 ∈ {1, …, 𝑚 − 1}. Then
(𝑇 − 𝜆𝑚 𝐼)𝑛 𝑣𝑘 ≠ 0
because otherwise 𝑣𝑘 would be a generalized eigenvector of 𝑇 corresponding to
the distinct eigenvalues 𝜆𝑘 and 𝜆𝑚 , which would contradict 8.11. However,
(𝑇 − 𝜆𝑘 𝐼)𝑛 ((𝑇 − 𝜆𝑚 𝐼)𝑛 𝑣𝑘 ) = (𝑇 − 𝜆𝑚 𝐼)𝑛 ((𝑇 − 𝜆𝑘 𝐼)𝑛 𝑣𝑘 ) = 0.
Thus the last two displayed equations show that (𝑇 − 𝜆𝑚 𝐼)𝑛 𝑣𝑘 is a generalized
eigenvector of 𝑇 corresponding to the eigenvalue 𝜆𝑘 . Hence
(𝑇 − 𝜆𝑚 𝐼)𝑛 𝑣1 , …, (𝑇 − 𝜆𝑚 𝐼)𝑛 𝑣𝑚 − 1
is a linearly dependent list (by 8.13) of 𝑚 − 1 generalized eigenvectors correspond-
ing to distinct eigenvalues, contradicting the minimality of 𝑚. This contradiction
completes the proof.
Nilpotent Operators
8.14 definition: nilpotent
An operator is called nilpotent if some power of it equals 0.
Thus an operator on 𝑉 is nilpotent if every nonzero vector in 𝑉 is a generalized

eigenvector of 𝑇 corresponding to the eigenvalue 0.
8.15 example: nilpotent operators
(a) The operator 𝑇 ∈ ℒ(𝐅4 ) defined by
𝑇(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 ) = (0, 0, 𝑧1 , 𝑧2 )
is nilpotent because 𝑇 2 = 0.
(b) The operator on 𝐅3 whose matrix (with respect to the standard basis) is
−3 9 0
⎛
⎜ ⎞
⎟
⎜
⎜ −7 9 6 ⎟
⎟
⎝ 4 0 −6 ⎠
is nilpotent, as can be shown by cubing the matrix above to get the zero matrix.
(c) The operator of differentiation on 𝒫𝑚 (𝐑) is nilpotent because the (𝑚 + 1)th
derivative of every polynomial of degree at most 𝑚 equals 0. Note that on
this space of dimension 𝑚 + 1, we need to raise the nilpotent operator to the
power 𝑚 + 1 to get the 0 operator.
The next result shows that when rais- The Latin word nil means nothing or
ing a nilpotent operator to a power, we zero; the Latin word potens means
never need to use a power higher than the having power. Thus nilpotent literally
dimension of the space. For a slightly means having a power that is zero.
stronger result, see Exercise 18.
8.16 nilpotent operator raised to dimension of domain is 0
Suppose 𝑇 ∈ ℒ(𝑉) is nilpotent. Then 𝑇 dim 𝑉 = 0.
Proof Because 𝑇 is nilpotent, there exists a positive integer 𝑘 such that 𝑇 𝑘 = 0.

Thus null 𝑇 𝑘 = 𝑉. Now 8.1 and 8.3 imply that null 𝑇 dim 𝑉 = 𝑉. Thus 𝑇 dim 𝑉 = 0.
8.17 eigenvalues of nilpotent operator
Suppose 𝑇 ∈ ℒ(𝑉).
(a) If 𝑇 is nilpotent, then 0 is an eigenvalue of 𝑇 and 𝑇 has no other
eigenvalues.
(b) If 𝐅 = 𝐂 and 0 is the only eigenvalue of 𝑇, then 𝑇 is nilpotent.
Proof
(a) To prove (a), suppose 𝑇 is nilpotent. Hence there is a positive integer 𝑚 such
that 𝑇 𝑚 = 0. This implies that 𝑇 is not injective. Thus 0 is an eigenvalue
of 𝑇.
To show that 𝑇 has no other eigenvalues, suppose 𝜆 is an eigenvalue of 𝑇.

Then there exists a nonzero vector 𝑣 ∈ 𝑉 such that
𝜆𝑣 = 𝑇𝑣.
Repeatedly applying 𝑇 to both sides of this equation shows that
𝜆𝑚 𝑣 = 𝑇 𝑚 𝑣 = 0.
Thus 𝜆 = 0, as desired.
(b) Suppose 𝐅 = 𝐂 and 0 is the only eigenvalue of 𝑇. By 5.27(b), the minimal
polynomial of 𝑇 equals 𝑧𝑚 for some positive integer 𝑚. Thus 𝑇 𝑚 = 0. Hence
𝑇 is nilpotent.
Exercise 23 shows that the hypothesis that 𝐅 = 𝐂 cannot be deleted in (b) of

the result above.
Given an operator on 𝑉, we want to find a basis of 𝑉 such that the matrix of
the operator with respect to this basis is as simple as possible, meaning that the
matrix contains many 0’s. The next result shows that if 𝑇 is nilpotent, then we can
choose a basis of 𝑉 such that the matrix of 𝑇 with respect to this basis has more
than half of its entries equal to 0. Later in this chapter we will do even better.
8.18 minimal polynomial and upper-triangular matrix of nilpotent operator
Suppose 𝑇 ∈ ℒ(𝑉). Then the following are equivalent.

(a) 𝑇 is nilpotent.
(b) The minimal polynomial of 𝑇 is 𝑧𝑚 for some positive integer 𝑚.
(c) There is a basis of 𝑉 with respect to which the matrix of 𝑇 has the form
⎛
0 ∗ ⎞
⎜ ⎟
⎜
⎜ ⋱ ⎟,
⎟
⎝ 0 0 ⎠
where all entries on and below the diagonal equal 0.
Proof Suppose (a) holds, so 𝑇 is nilpotent. Thus there exists a positive integer
𝑛 such that 𝑇 𝑛 = 0. Now 5.29 implies that 𝑧𝑛 is a polynomial multiple of the
minimal polynomial of 𝑇. Thus the minimal polynomial of 𝑇 is 𝑧𝑚 for some
positive integer 𝑚, proving that (a) implies (b).
Now suppose (b) holds, so the minimal polynomial of 𝑇 is 𝑧𝑚 for some positive
integer 𝑚. This implies, by 5.27(a), that 0 (which is the only zero of 𝑧𝑚 ) is the
only eigenvalue of 𝑇. This further implies, by 5.44, that there is a basis of 𝑉 with
respect to which the matrix of 𝑇 is upper triangular. This also implies, by 5.41,
that all entries on the diagonal of this matrix are 0, proving that (b) implies (c).
Now suppose (c) holds. Then 5.40 implies that 𝑇 dim 𝑉 = 0. Thus 𝑇 is nilpotent,
proving that (c) implies (a).
Exercises 8A
1 Suppose 𝑇 ∈ ℒ(𝑉). Prove that if dim null 𝑇 4 = 8 and dim null 𝑇 6 = 9, then
dim null 𝑇 𝑚 = 9 for all integers 𝑚 ≥ 5.
2 Suppose 𝑇 ∈ ℒ(𝑉), 𝑚 is a positive integer, 𝑣 ∈ 𝑉, and 𝑇 𝑚 − 1 𝑣 ≠ 0 but
𝑇 𝑚 𝑣 = 0. Prove that 𝑣, 𝑇𝑣, 𝑇 2 𝑣, …, 𝑇 𝑚 − 1 𝑣 is linearly independent.
3 Suppose 𝑇 ∈ ℒ(𝑉). Prove that

𝑉 = null 𝑇 ⊕ range 𝑇 ⟺ null 𝑇 2 = null 𝑇.
4 Suppose 𝑇 ∈ ℒ(𝑉), 𝜆 ∈ 𝐅, and 𝑚 is a positive integer such that the minimal
polynomial of 𝑇 is a polynomial multiple of (𝑧 − 𝜆)𝑚. Prove that
dim null(𝑇 − 𝜆𝐼)𝑚 ≥ 𝑚.
5 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑚 is a positive integer. Prove that
dim null 𝑇 𝑚 ≤ 𝑚 dim null 𝑇.
Hint: Exercise 21 in Section 3B may be useful.
6 Suppose 𝑇 ∈ ℒ(𝑉). Show that

𝑉 = range 𝑇 0 ⊇ range 𝑇 1 ⊇ ⋯ ⊇ range 𝑇 𝑘 ⊇ range 𝑇 𝑘 + 1 ⊇ ⋯ .
7 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑚 is a nonnegative integer such that
range 𝑇 𝑚 = range 𝑇 𝑚 + 1.
Prove that range 𝑇 𝑘 = range 𝑇 𝑚 for all 𝑘 > 𝑚.
8 Suppose 𝑇 ∈ ℒ(𝑉). Prove that
range 𝑇 dim 𝑉 = range 𝑇 dim 𝑉 + 1 = range 𝑇 dim 𝑉 + 2 = ⋯ .
9 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑚 is a nonnegative integer. Prove that
null 𝑇 𝑚 = null 𝑇 𝑚 + 1 ⟺ range 𝑇 𝑚 = range 𝑇 𝑚 + 1.
10 Define 𝑇 ∈ ℒ(𝐂2 ) by 𝑇(𝑤, 𝑧) = (𝑧, 0). Find all generalized eigenvectors
of 𝑇.
11 Suppose that 𝑇 ∈ ℒ(𝑉). Prove that there is a basis of 𝑉 consisting of
generalized eigenvectors of 𝑇 if and only if the minimal polynomial of 𝑇
equals (𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑚 ) for some 𝜆1 , …, 𝜆𝑚 ∈ 𝐅.
Assume 𝐅 = 𝐑 because the case 𝐅 = 𝐂 follows from 5.27(b) and 8.9.
This exercise states that the condition for there to be a basis of 𝑉 consisting
of generalized eigenvectors of 𝑇 is the same as the condition for there to be
a basis with respect to which 𝑇 has an upper-triangular matrix (see 5.44).
Caution: If 𝑇 has an upper-triangular matrix with respect to a basis
𝑣1 , …, 𝑣𝑛 of 𝑉, then 𝑣1 is an eigenvector of 𝑇 but it is not necessarily true
that 𝑣2 , …, 𝑣𝑛 are generalized eigenvectors of 𝑇.
12 Suppose 𝑇 ∈ ℒ(𝑉) is such that every vector in 𝑉 is a generalized eigenvector

of 𝑇. Prove that there exists 𝜆 ∈ 𝐅 such that 𝑇 − 𝜆𝐼 is nilpotent.
13 Suppose 𝑆, 𝑇 ∈ ℒ(𝑉) and 𝑆𝑇 is nilpotent. Prove that 𝑇𝑆 is nilpotent.
14 Suppose 𝑇 ∈ ℒ(𝑉) is nilpotent and 𝑇 ≠ 0. Prove 𝑇 is not diagonalizable.
15 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Prove that 𝑇 is diagonalizable if and only if
every generalized eigenvector of 𝑇 is an eigenvector of 𝑇.
For 𝐅 = 𝐂 , this exercise adds another equivalence to the list of conditions
for diagonalizability in 5.55.
16 (a) Give an example of nilpotent operators 𝑆, 𝑇 on the same vector space

such that neither 𝑆 + 𝑇 nor 𝑆𝑇 is nilpotent.
(b) Suppose 𝑆, 𝑇 ∈ ℒ(𝑉) are nilpotent and 𝑆𝑇 = 𝑇𝑆. Prove that 𝑆 + 𝑇 and
𝑆𝑇 are nilpotent.
17 Suppose 𝑇 ∈ ℒ(𝑉) is nilpotent and 𝑚 is a positive integer such that 𝑇 𝑚 = 0.
(a) Prove that 𝐼 − 𝑇 is invertible and that (𝐼 − 𝑇)−1 = 𝐼 + 𝑇 + ⋯ + 𝑇 𝑚 − 1.
(b) Explain how you would guess the formula above.
18 Suppose 𝑇 ∈ ℒ(𝑉) is nilpotent. Prove that 𝑇 1 + dim range 𝑇 = 0.
If dim range 𝑇 < dim 𝑉 − 1, then this exercise improves 8.16.
19 Suppose 𝑇 ∈ ℒ(𝑉) is not nilpotent. Show that

𝑉 = null 𝑇 dim 𝑉 − 1 ⊕ range 𝑇 dim 𝑉 − 1.
For operators that are not nilpotent, this exercise improves 8.4.
20 Suppose 𝑉 is an inner product space and 𝑇 ∈ ℒ(𝑉) is normal and nilpotent.

Prove that 𝑇 = 0.
21 Suppose 𝑇 ∈ ℒ(𝑉) is such that null 𝑇 dim 𝑉 − 1 ≠ null 𝑇 dim 𝑉. Prove that 𝑇 is
nilpotent and that dim null 𝑇 𝑘 = 𝑘 for every integer 𝑘 with 0 ≤ 𝑘 ≤ dim 𝑉.
22 Suppose 𝑇 ∈ ℒ(𝐂5 ) is such that range 𝑇 4 ≠ range 𝑇 5. Prove that 𝑇 is
nilpotent.
23 Give an example of an operator 𝑇 on a finite-dimensional real vector space
such that 0 is the only eigenvalue of 𝑇 but 𝑇 is not nilpotent.
This exercise shows that the implication (b) ⟹ (a) in 8.17 does not hold
without the hypothesis that 𝐅 = 𝐂 .
24 For each item in Example 8.15, find a basis of the domain vector space such
that the matrix of the nilpotent operator with respect to that basis has the
upper-triangular form promised by 8.18(c).
25 Suppose that 𝑉 is an inner product space and 𝑇 ∈ ℒ(𝑉) is nilpotent. Show
that there is an orthonormal basis of 𝑉 with respect to which the matrix of 𝑇
has the upper-triangular form promised by 8.18(c).
8B Generalized Eigenspace Decomposition

Generalized Eigenspaces
8.19 definition: generalized eigenspace, 𝐺(𝜆, 𝑇)
Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆 ∈ 𝐅. The generalized eigenspace of 𝑇 correspond-

ing to 𝜆, denoted by 𝐺(𝜆, 𝑇), is defined by
𝐺(𝜆, 𝑇) = {𝑣 ∈ 𝑉 ∶ (𝑇 − 𝜆𝐼)𝑘 𝑣 = 0 for some positive integer 𝑘}.
Thus 𝐺(𝜆, 𝑇) is the set of generalized eigenvectors of 𝑇 corresponding to 𝜆,

along with the 0 vector.
Because every eigenvector of 𝑇 is a generalized eigenvector of 𝑇 (take 𝑘 = 1

in the definition of generalized eigenvector), each eigenspace is contained in the
corresponding generalized eigenspace. In other words, if 𝑇 ∈ ℒ(𝑉) and 𝜆 ∈ 𝐅,
then 𝐸(𝜆, 𝑇) ⊆ 𝐺(𝜆, 𝑇).
The next result implies that if 𝑇 ∈ ℒ(𝑉) and 𝜆 ∈ 𝐅, then the generalized
eigenspace 𝐺(𝜆, 𝑇) is a subspace of 𝑉 (because the null space of each linear map
on 𝑉 is a subspace of 𝑉).
8.20 description of generalized eigenspaces
Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆 ∈ 𝐅. Then 𝐺(𝜆, 𝑇) = null(𝑇 − 𝜆𝐼)dim 𝑉.
Proof Suppose 𝑣 ∈ null(𝑇 − 𝜆𝐼)dim 𝑉. The definitions imply 𝑣 ∈ 𝐺(𝜆, 𝑇). Thus
𝐺(𝜆, 𝑇) ⊇ null(𝑇 − 𝜆𝐼)dim 𝑉.
Conversely, suppose 𝑣 ∈ 𝐺(𝜆, 𝑇). Thus there is a positive integer 𝑘 such
that 𝑣 ∈ null(𝑇 − 𝜆𝐼)𝑘. From 8.1 and 8.3 (with 𝑇 − 𝜆𝐼 replacing 𝑇), we get
𝑣 ∈ null(𝑇 − 𝜆𝐼)dim 𝑉. Thus 𝐺(𝜆, 𝑇) ⊆ null(𝑇 − 𝜆𝐼)dim 𝑉, completing the proof.
8.21 example: generalized eigenspaces of an operator on 𝐂3

𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (4𝑧2 , 0, 5𝑧3 ).
In Example 8.10, we saw that the eigenvalues of 𝑇 are 0 and 5, and we found
the corresponding sets of generalized eigenvectors. Taking the union of those sets
with {0}, we have
𝐺(0, 𝑇) = {(𝑧1 , 𝑧2 , 0) ∶ 𝑧1 , 𝑧2 ∈ 𝐂} and 𝐺(5, 𝑇) = {(0, 0, 𝑧3 ) ∶ 𝑧3 ∈ 𝐂}.
Note that 𝐂3 = 𝐺(0, 𝑇) ⊕ 𝐺(5, 𝑇).

Section 8B Generalized Eigenspace Decomposition 309
In Example 8.21, the domain space 𝐂3 is the direct sum of the generalized
eigenspaces of the operator 𝑇 in that example. Our next result shows that this
behavior holds in general. Specifically, the following major result shows that if
𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉), then 𝑉 is the direct sum of the generalized eigenspaces
of 𝑇, each of which is invariant under 𝑇 and on which 𝑇 is a nilpotent operator
plus a scalar multiple of the identity. Thus the next result achieves our goal of
decomposing 𝑉 into invariant subspaces on which 𝑇 has a known behavior.
As we will see, the proof follows from putting together what we have learned
about generalized eigenspaces and then using our result that for each operator
𝑇 ∈ ℒ(𝑉), there exists a basis of 𝑉 consisting of generalized eigenvectors of 𝑇.
8.22 generalized eigenspace decomposition
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Let 𝜆1 , …, 𝜆𝑚 be the distinct eigenvalues

of 𝑇. Then
(a) 𝐺(𝜆𝑘 , 𝑇) is invariant under 𝑇 for each 𝑘 = 1, …, 𝑚;
(b) (𝑇 − 𝜆𝑘 𝐼)|𝐺( 𝜆𝑘, 𝑇) is nilpotent for each 𝑘 = 1, …, 𝑚;
(c) 𝑉 = 𝐺(𝜆1 , 𝑇) ⊕ ⋯ ⊕ 𝐺(𝜆𝑚 , 𝑇).
Proof
(a) Suppose 𝑘 ∈ {1, …, 𝑚}. Then 8.20 shows that
𝐺(𝜆𝑘 , 𝑇) = null(𝑇 − 𝜆𝑘 𝐼)dim 𝑉.
Thus 5.18, with 𝑝(𝑧) = (𝑧− 𝜆𝑘 )dim 𝑉, implies that 𝐺(𝜆𝑘 , 𝑇) is invariant under 𝑇,
proving (a).
(b) Suppose 𝑘 ∈ {1, …, 𝑚}. If 𝑣 ∈ 𝐺(𝜆𝑘 , 𝑇), then (𝑇 − 𝜆𝑘 𝐼)dim 𝑉 𝑣 = 0 (by 8.20).
Thus ((𝑇 − 𝜆𝑘 𝐼)|𝐺( 𝜆𝑘, 𝑇) )dim 𝑉 = 0. Hence (𝑇 − 𝜆𝑘 𝐼)|𝐺( 𝜆𝑘, 𝑇) is nilpotent,
proving (b).
(c) To show that 𝐺(𝜆1 , 𝑇) + ⋯ + 𝐺(𝜆𝑚 , 𝑇) is a direct sum, suppose
𝑣1 + ⋯ + 𝑣𝑚 = 0,
where each 𝑣𝑘 is in 𝐺(𝜆𝑘 , 𝑇). Because generalized eigenvectors of 𝑇 cor-
responding to distinct eigenvalues are linearly independent (by 8.12), this
implies that each 𝑣𝑘 equals 0. Thus 𝐺(𝜆1 , 𝑇) + ⋯ + 𝐺(𝜆𝑚 , 𝑇) is a direct sum
(by 1.45).
Finally, each vector in 𝑉 can be written as a finite sum of generalized eigen-
vectors of 𝑇 (by 8.9). Thus
𝑉 = 𝐺(𝜆1 , 𝑇) ⊕ ⋯ ⊕ 𝐺(𝜆𝑚 , 𝑇),
proving (c).
For the analogous result when 𝐅 = 𝐑 , see Exercise 8.

Multiplicity of an Eigenvalue
If 𝑉 is a complex vector space and 𝑇 ∈ ℒ(𝑉), then the decomposition of 𝑉 pro-
vided by the generalized eigenspace decomposition (8.22) can be a powerful tool.
The dimensions of the subspaces involved in this decomposition are sufficiently
important to get a name, which is given in the next definition.
8.23 definition: multiplicity
• Suppose 𝑇 ∈ ℒ(𝑉). The multiplicity of an eigenvalue 𝜆 of 𝑇 is defined to

be the dimension of the corresponding generalized eigenspace 𝐺(𝜆, 𝑇).
• In other words, the multiplicity of an eigenvalue 𝜆 of 𝑇 equals
dim null(𝑇 − 𝜆𝐼)dim 𝑉.
The second bullet point above holds because 𝐺(𝜆, 𝑇) = null(𝑇 − 𝜆𝐼)dim 𝑉 (see
8.20).
8.24 example: multiplicity of each eigenvalue of an operator

Suppose 𝑇 ∈ ℒ(𝐂3 ) is defined by
𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (6𝑧1 + 3𝑧2 + 4𝑧3 , 6𝑧2 + 2𝑧3 , 7𝑧3 ).

6 3 4
⎛
⎜ ⎞
⎟
⎜
⎜ 0 6 2 ⎟
⎟.
⎝ 0 0 7 ⎠
The eigenvalues of 𝑇 are the diagonal entries 6 and 7, as follows from 5.41. You
can verify that the generalized eigenspaces of 𝑇 are as follows:
𝐺(6, 𝑇) = span((1, 0, 0), (0, 1, 0)) and 𝐺(7, 𝑇) = span((10, 2, 1)).
Thus the eigenvalue 6 has multiplicity 2 In this example, the multiplicity of each
and the eigenvalue 7 has multiplicity 1. eigenvalue equals the number of times
The direct sum 𝐂3 = 𝐺(6, 𝑇) ⊕ 𝐺(7, 𝑇) that eigenvalue appears on the diago-
is the generalized eigenspace decom- nal of an upper-triangular matrix rep-
position promised by 8.22. A basis resenting the operator. This behavior
of 𝐂3 consisting of generalized eigen- always happens, as we will see in 8.31.
vectors of 𝑇, as promised by 8.9, is
(1, 0, 0), (0, 1, 0), (10, 2, 1). There does not exist a basis of 𝐂3 consisting of eigen-
vectors of this operator.
In the example above, the sum of the multiplicities of the eigenvalues of 𝑇

equals 3, which is the dimension of the domain of 𝑇. The next result shows that
this holds for all operators on finite-dimensional complex vector spaces.
8.25 sum of the multiplicities equals dim 𝑉
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Then the sum of the multiplicities of all

eigenvalues of 𝑇 equals dim 𝑉.
Proof The desired result follows from the generalized eigenspace decomposition
(8.22) and the formula for the dimension of a direct sum (see 3.94).
The terms algebraic multiplicity and geometric multiplicity are used in some
books. In case you encounter this terminology, be aware that the algebraic multi-
plicity is the same as the multiplicity defined here and the geometric multiplicity
is the dimension of the corresponding eigenspace. In other words, if 𝑇 ∈ ℒ(𝑉)
and 𝜆 is an eigenvalue of 𝑇, then
algebraic multiplicity of 𝜆 = dim null(𝑇 − 𝜆𝐼)dim 𝑉 = dim 𝐺(𝜆, 𝑇),
geometric multiplicity of 𝜆 = dim null(𝑇 − 𝜆𝐼) = dim 𝐸(𝜆, 𝑇).
Note that as defined above, the algebraic multiplicity also has a geometric meaning
as the dimension of a certain null space. The definition of multiplicity given here
is cleaner than the traditional definition that involves determinants; 9.62 implies
that these definitions are equivalent.
If 𝑉 is an inner product space, 𝑇 ∈ ℒ(𝑉) is normal, and 𝜆 is an eigenvalue
of 𝑇, then the algebraic multiplicity of 𝜆 equals the geometric multiplicity of 𝜆,
as can be seen from applying Exercise 27 in Section 7A to the normal operator
𝑇 − 𝜆𝐼. As a special case, the singular values of 𝑆 ∈ ℒ(𝑉, 𝑊) (here 𝑉 and 𝑊 are
both finite-dimensional inner product spaces) depend on the multiplicities (either
algebraic or geometric) of the eigenvalues of the self-adjoint operator 𝑆∗ 𝑆.
The next definition associates a monic polynomial with each operator on a
finite-dimensional complex vector space.
8.26 definition: characteristic polynomial
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Let 𝜆1 , …, 𝜆𝑚 denote the distinct eigenvalues

of 𝑇, with multiplicities 𝑑1 , …, 𝑑𝑚 . The polynomial
(𝑧 − 𝜆1 )𝑑1 ⋯(𝑧 − 𝜆𝑚 )𝑑𝑚
is called the characteristic polynomial of 𝑇.
8.27 example: the characteristic polynomial of an operator

Suppose 𝑇 ∈ ℒ(𝐂3 ) is defined as in Example 8.24. Because the eigenvalues of
𝑇 are 6, with multiplicity 2, and 7, with multiplicity 1, we see that the characteristic
polynomial of 𝑇 is (𝑧 − 6)2 (𝑧 − 7).
8.28 degree and zeros of characteristic polynomial
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Then

(a) the characteristic polynomial of 𝑇 has degree dim 𝑉;
(b) the zeros of the characteristic polynomial of 𝑇 are the eigenvalues of 𝑇.
Proof Our result about the sum of the multiplicities (8.25) implies (a). The
definition of the characteristic polynomial implies (b).
Most texts define the characteristic polynomial using determinants (the two
definitions are equivalent by 9.62). The approach taken here, which is considerably
simpler, leads to the following nice proof of the Cayley–Hamilton theorem.
8.29 Cayley–Hamilton theorem
Suppose 𝐅 = 𝐂 , 𝑇 ∈ ℒ(𝑉), and 𝑞 is the characteristic polynomial of 𝑇. Then

𝑞(𝑇) = 0.
Proof Let 𝜆1 , …, 𝜆𝑚 be the distinct eigenvalues of 𝑇, and let 𝑑𝑘 = dim 𝐺(𝜆𝑘 , 𝑇).
For each 𝑘 ∈ {1, …, 𝑚}, we know that (𝑇 − 𝜆𝑘 𝐼)|𝐺( 𝜆𝑘, 𝑇) is nilpotent. Thus we have
(𝑇 − 𝜆𝑘 𝐼)𝑑𝑘 |𝐺( 𝜆𝑘, 𝑇) = 0 Arthur Cayley (1821–1895) published
three mathematics papers before com-
(by 8.16) for each 𝑘 ∈ {1, …, 𝑚}. pleting his undergraduate degree.
The generalized eigenspace decom-
position (8.22) states that every vector in 𝑉 is a sum of vectors in
𝐺(𝜆1 , 𝑇), …, 𝐺(𝜆𝑚 , 𝑇). Thus to prove that 𝑞(𝑇) = 0, we only need to show
that 𝑞(𝑇)|𝐺( 𝜆𝑘, 𝑇) = 0 for each 𝑘.
Fix 𝑘 ∈ {1, …, 𝑚}. We have
𝑞(𝑇) = (𝑇 − 𝜆1 𝐼)𝑑1 ⋯(𝑇 − 𝜆𝑚 𝐼)𝑑𝑚.
The operators on the right side of the equation above all commute, so we can
move the factor (𝑇 − 𝜆𝑘 𝐼)𝑑𝑘 to be the last term in the expression on the right.
Because (𝑇 − 𝜆𝑘 𝐼)𝑑𝑘 |𝐺( 𝜆𝑘, 𝑇) = 0, we have 𝑞(𝑇)|𝐺( 𝜆𝑘, 𝑇) = 0, as desired.
The next result implies that if the minimal polynomial of an operator 𝑇 ∈ ℒ(𝑉)
has degree dim 𝑉 (as happens almost always—see the paragraphs following 5.24),
then the characteristic polynomial of 𝑇 equals the minimal polynomial of 𝑇.
8.30 characteristic polynomial is a multiple of minimal polynomial
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Then the characteristic polynomial of 𝑇 is a

polynomial multiple of the minimal polynomial of 𝑇.
Proof The desired result follows immediately from the Cayley–Hamilton theo-
rem (8.29) and 5.29.
Now we can prove that the result suggested by Example 8.24 holds for all
operators on finite-dimensional complex vector spaces.
8.31 multiplicity of an eigenvalue equals number of times on diagonal
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Suppose 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 such that

ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )) is upper triangular. Then the number of times that each
eigenvalue 𝜆 of 𝑇 appears on the diagonal of ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )) equals the
multiplicity of 𝜆 as an eigenvalue of 𝑇.
Proof Let 𝐴 = ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )). Thus 𝐴 is an upper-triangular matrix. Let

𝜆1 , …, 𝜆𝑛 denote the entries on the diagonal of 𝐴. Thus for each 𝑘 ∈ {1, …, 𝑛},
we have
𝑇𝑣𝑘 = 𝑢𝑘 + 𝜆𝑘 𝑣𝑘
for some 𝑢𝑘 ∈ span(𝑣1 , …, 𝑣𝑘 − 1 ). Hence if 𝑘 ∈ {1, …, 𝑛} and 𝜆𝑘 ≠ 0, then 𝑇𝑣𝑘 is
not a linear combination of 𝑇𝑣1 , …, 𝑇𝑣𝑘 − 1 . The linear dependence lemma (2.19)
now implies that the list of those 𝑇𝑣𝑘 such that 𝜆𝑘 ≠ 0 is linearly independent.
Let 𝑑 denote the number of indices 𝑘 ∈ {1, …, 𝑛} such that 𝜆𝑘 = 0. The
conclusion of the previous paragraph implies that
dim range 𝑇 ≥ 𝑛 − 𝑑.
Because 𝑛 = dim 𝑉 = dim null 𝑇 + dim range 𝑇, the inequality above implies that
8.32 dim null 𝑇 ≤ 𝑑.
The matrix of the operator 𝑇 𝑛 with respect to the basis 𝑣1 , …, 𝑣𝑛 is the upper-
triangular matrix 𝐴𝑛, which has diagonal entries 𝜆1𝑛 , …, 𝜆𝑛𝑛 [see Exercise 2(b) in
Section 5C]. Because 𝜆𝑘𝑛 = 0 if and only if 𝜆𝑘 = 0, the number of times that 0
appears on the diagonal of 𝐴𝑛 equals 𝑑. Thus applying 8.32 with 𝑇 replaced with
𝑇 𝑛, we have
8.33 dim null 𝑇 𝑛 ≤ 𝑑.

For 𝜆 an eigenvalue of 𝑇, let 𝑚 𝜆 denote the multiplicity of 𝜆 as an eigenvalue
of 𝑇 and let 𝑑 𝜆 denote the number of times that 𝜆 appears on the diagonal of 𝐴.
Replacing 𝑇 in 8.33 with 𝑇 − 𝜆𝐼, we see that
8.34 𝑚𝜆 ≤ 𝑑𝜆
for each eigenvalue 𝜆 of 𝑇. The sum of the multiplicities 𝑚 𝜆 over all eigenvalues
𝜆 of 𝑇 equals 𝑛, the dimension of 𝑉 (by 8.25). The sum of the numbers 𝑑 𝜆 over
all eigenvalues 𝜆 of 𝑇 also equals 𝑛, because the diagonal of 𝐴 has length 𝑛.
Thus summing both sides of 8.34 over all eigenvalues 𝜆 of 𝑇 produces an
equality. Hence 8.34 must actually be an equality for each eigenvalue 𝜆 of 𝑇.
Thus the multiplicity of 𝜆 as an eigenvalue of 𝑇 equals the number of times that
𝜆 appears on the diagonal of 𝐴, as desired.
≈ 100𝜋 Chapter 8 Operators on Complex Vector Spaces
Block Diagonal Matrices

To interpret our results in matrix form, Often we can understand a matrix
we make the following definition, gener- better by thinking of it as composed
alizing the notion of a diagonal matrix. of smaller matrices.
If each matrix 𝐴𝑘 in the definition below
is a 1-by-1 matrix, then we actually have a diagonal matrix.
8.35 definition: block diagonal matrix
A block diagonal matrix is a square matrix of the form

𝐴1 0
⎛
⎜ ⎞
⎟
⎜
⎜ ⋱ ⎟,
⎟
⎝ 0 𝐴𝑚 ⎠
where 𝐴1 , …, 𝐴𝑚 are square matrices lying along the diagonal and all other
entries of the matrix equal 0.
8.36 example: a block diagonal matrix

The 5-by-5 matrix
( 4 ) 0 0 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ ⎟
⎟
⎜
⎜ 0 ⎛ 2 −3 ⎞ 0 0 ⎟
⎟
⎜ ⎜ ⎟ ⎟
𝐴=⎜
⎜
⎜ 0 ⎝ 0 2 ⎠ 0 0
⎟
⎟
⎟
⎜
⎜ ⎟
⎟
⎜
⎜ 0 0 0 1 7 ⎞ ⎟
⎟
⎜
⎜ ⎛ ⎟
⎜ ⎟ ⎟
⎝ 0 0 0 ⎝ 0 1 ⎠ ⎠
is a block diagonal matrix with
𝐴1 0
⎛
⎜ ⎞
⎟
𝐴=⎜
⎜
⎜ 𝐴2 ⎟
⎟
⎟,
⎜ ⎟
⎝ 0 𝐴3 ⎠
where
2 −3 ⎞ 1 7 ⎞
𝐴1 = ( 4 ), 𝐴2 = ⎛
⎜ ⎟, 𝐴3 = ⎛ ⎜ ⎟.
⎝ 0 2 ⎠ ⎝ 0 1 ⎠
Here the inner matrices in the 5-by-5 matrix above are blocked off to show how
we can think of it as a block diagonal matrix.
Note that in the example above, each of 𝐴1 , 𝐴2 , 𝐴3 is an upper-triangular

matrix whose diagonal entries are all equal. The next result shows that with
respect to an appropriate basis, every operator on a finite-dimensional complex
vector space has a matrix of this form. Note that this result gives us many more
zeros in the matrix than are needed to make it upper triangular.
8.37 block diagonal matrix with upper-triangular blocks
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Let 𝜆1 , …, 𝜆𝑚 be the distinct eigenvalues

of 𝑇, with multiplicities 𝑑1 , …, 𝑑𝑚 . Then there is a basis of 𝑉 with respect to
which 𝑇 has a block diagonal matrix of the form
𝐴1 0
⎛
⎜ ⎞
⎟
⎜
⎜ ⋱ ⎟,
⎟
⎝ 0 𝐴𝑚 ⎠
where each 𝐴𝑘 is a 𝑑𝑘 -by-𝑑𝑘 upper-triangular matrix of the form
𝜆
⎛ 𝑘
∗ ⎞
𝐴𝑘 = ⎜
⎜
⎜ ⋱ ⎟
⎟
⎟.
⎝ 0 𝜆𝑘 ⎠
Proof Each (𝑇 − 𝜆𝑘 𝐼)|𝐺( 𝜆𝑘 , 𝑇) is nilpotent (see 8.22). For each 𝑘, choose a basis
of 𝐺(𝜆𝑘 , 𝑇), which is a vector space of dimension 𝑑𝑘 , such that the matrix of
(𝑇 − 𝜆𝑘 𝐼)|𝐺( 𝜆𝑘, 𝑇) with respect to this basis is as in 8.18(c). Thus with respect to
this basis, the matrix of 𝑇|𝐺( 𝜆𝑘, 𝑇) , which equals (𝑇 − 𝜆𝑘 𝐼)|𝐺( 𝜆𝑘, 𝑇) + 𝜆𝑘 𝐼|𝐺( 𝜆𝑘, 𝑇) ,
looks like the desired form shown above for 𝐴𝑘 .
The generalized eigenspace decomposition (8.22) shows that putting together
the bases of the 𝐺(𝜆𝑘 , 𝑇)’s chosen above gives a basis of 𝑉. The matrix of 𝑇 with
respect to this basis has the desired form.
8.38 example: block diagonal matrix via generalized eigenvectors

Let 𝑇 ∈ ℒ(𝐂3 ) be defined by 𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (6𝑧1 + 3𝑧2 + 4𝑧3 , 6𝑧2 + 2𝑧3 , 7𝑧3 ).
6 3 4
⎛
⎜ ⎞
⎟
⎜
⎜ 0 6 2 ⎟,
⎟
⎝ 0 0 7 ⎠
which is an upper-triangular matrix but is not of the form promised by 8.37.
As we saw in Example 8.24, the eigenvalues of 𝑇 are 6 and 7, and
𝐺(6, 𝑇) = span((1, 0, 0), (0, 1, 0)) and 𝐺(7, 𝑇) = span((10, 2, 1)).
We also saw that a basis of 𝐂3 consisting of generalized eigenvectors of 𝑇 is
(1, 0, 0), (0, 1, 0), (10, 2, 1).
The matrix of 𝑇 with respect to this basis is
6 3 0
⎛
⎜ ( ) ⎞
⎜
⎜ 0 6 0 ⎟ ⎟,
⎟
⎝ 0 0 ( 7 ) ⎠
which is a matrix of the block diagonal form promised by 8.37.
Exercises 8B
1 Define 𝑇 ∈ ℒ(𝐂2 ) by 𝑇(𝑤, 𝑧) = (−𝑧, 𝑤). Find the generalized eigenspaces

corresponding to the distinct eigenvalues of 𝑇.
2 Suppose 𝑇 ∈ ℒ(𝑉) is invertible. Prove that 𝐺(𝜆, 𝑇) = 𝐺( 1𝜆 , 𝑇 −1 ) for every
𝜆 ∈ 𝐅 with 𝜆 ≠ 0.
3 Suppose 𝑇 ∈ ℒ(𝑉). Suppose 𝑆 ∈ ℒ(𝑉) is invertible. Prove that 𝑇 and

𝑆−1 𝑇𝑆 have the same eigenvalues with the same multiplicities.
4 Suppose dim 𝑉 ≥ 2 and 𝑇 ∈ ℒ(𝑉) is such that null 𝑇 dim 𝑉 − 2 ≠ null 𝑇 dim 𝑉 − 1.
Prove that 𝑇 has at most two distinct eigenvalues.
5 Suppose 𝑇 ∈ ℒ(𝑉) and 3 and 8 are eigenvalues of 𝑇. Let 𝑛 = dim 𝑉. Prove
that 𝑉 = (null 𝑇 𝑛 − 2 ) ⊕ (range 𝑇 𝑛 − 2 ).
6 Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆 is an eigenvalue of 𝑇. Explain why the exponent
of 𝑧 − 𝜆 in the factorization of the minimal polynomial of 𝑇 is the smallest
positive integer 𝑚 such that (𝑇 − 𝜆𝐼)𝑚 |𝐺( 𝜆, 𝑇) = 0.
7 Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆 is an eigenvalue of 𝑇 with multiplicity 𝑑. Prove
that 𝐺(𝜆, 𝑇) = null(𝑇 − 𝜆𝐼)𝑑.
If 𝑑 < dim 𝑉, then this exercise improves 8.20.
8 Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆1 , …, 𝜆𝑚 are the distinct eigenvalues of 𝑇. Prove

that
𝑉 = 𝐺(𝜆1 , 𝑇) ⊕ ⋯ ⊕ 𝐺(𝜆𝑚 , 𝑇)
if and only if the minimal polynomial of 𝑇 equals (𝑧 − 𝜆1 )𝑘1 ⋯(𝑧 − 𝜆𝑚 )𝑘𝑚
for some positive integers 𝑘1 , …, 𝑘𝑚 .
The case 𝐅 = 𝐂 follows immediately from 5.27(b) and the generalized
eigenspace decomposition (8.22); thus this exercise is interesting only when
𝐅 = 𝐑.
9 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Prove that there exist 𝐷, 𝑁 ∈ ℒ(𝑉)

such that 𝑇 = 𝐷 + 𝑁, the operator 𝐷 is diagonalizable, 𝑁 is nilpotent, and
𝐷𝑁 = 𝑁𝐷.
10 Suppose 𝑉 is a complex inner product space, 𝑒1 , …, 𝑒𝑛 is an orthonormal

basis of 𝑇, and 𝑇 ∈ ℒ(𝑉). Let 𝜆1 , …, 𝜆𝑛 be the eigenvalues of 𝑇, each
included as many times as its multiplicity. Prove that
|𝜆1 |2 + ⋯ + |𝜆𝑛 |2 ≤ ‖𝑇𝑒1 ‖2 + ⋯ + ‖𝑇𝑒𝑛 ‖2.
See the comment after Exercise 5 in Section 7A.
11 Give an example of an operator on 𝐂4 whose characteristic polynomial

equals (𝑧 − 7)2 (𝑧 − 8)2.
12 Give an example of an operator on 𝐂4 whose characteristic polynomial

equals (𝑧 − 1)(𝑧 − 5)3 and whose minimal polynomial equals (𝑧 − 1)(𝑧 − 5)2.
13 Give an example of an operator on 𝐂4 whose characteristic and minimal
polynomials both equal 𝑧(𝑧 − 1)2 (𝑧 − 3).
14 Give an example of an operator on 𝐂4 whose characteristic polynomial equals
𝑧(𝑧 − 1)2 (𝑧 − 3) and whose minimal polynomial equals 𝑧(𝑧 − 1)(𝑧 − 3).
15 Let 𝑇 be the operator on 𝐂4 defined by 𝑇(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 ) = (0, 𝑧1 , 𝑧2 , 𝑧3 ). Find

the characteristic polynomial and the minimal polynomial of 𝑇.
16 Let 𝑇 be the operator on 𝐂6 defined by
𝑇(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 , 𝑧5 , 𝑧6 ) = (0, 𝑧1 , 𝑧2 , 0, 𝑧4 , 0).
Find the characteristic polynomial and the minimal polynomial of 𝑇.

17 Suppose 𝐅 = 𝐂 and 𝑃 ∈ ℒ(𝑉) is such that 𝑃2 = 𝑃. Prove that the characteris-
tic polynomial of 𝑃 is 𝑧𝑚 (𝑧−1)𝑛, where 𝑚 = dim null 𝑃 and 𝑛 = dim range 𝑃.
18 Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆 is an eigenvalue of 𝑇. Explain why the following
four numbers equal each other.
(a) The exponent of 𝑧 − 𝜆 in the factorization of the minimal polynomial
of 𝑇.
(b) The smallest positive integer 𝑚 such that (𝑇 − 𝜆𝐼)𝑚 |𝐺( 𝜆, 𝑇) = 0.
(c) The smallest positive integer 𝑚 such that
null(𝑇 − 𝜆𝐼)𝑚 = null(𝑇 − 𝜆𝐼)𝑚 + 1.

(d) The smallest positive integer 𝑚 such that
range(𝑇 − 𝜆𝐼)𝑚 = range(𝑇 − 𝜆𝐼)𝑚 + 1.
19 Suppose 𝐅 = 𝐂 and 𝑆 ∈ ℒ(𝑉) is a unitary operator. Prove that the constant

term in the characteristic polynomial of 𝑆 has absolute value 1.
20 Suppose that 𝐅 = 𝐂 and 𝑉1 , …, 𝑉𝑚 are nonzero subspaces of 𝑉 such that
𝑉 = 𝑉1 ⊕ ⋯ ⊕ 𝑉𝑚 .
Suppose 𝑇 ∈ ℒ(𝑉) and each 𝑉𝑘 is invariant under 𝑇. For each 𝑘, let 𝑝𝑘

denote the characteristic polynomial of 𝑇|𝑉𝑘 . Prove that the characteristic
polynomial of 𝑇 equals 𝑝1 ⋯𝑝𝑚 .
21 Suppose 𝑝, 𝑞 ∈ 𝒫(𝐂) are monic polynomials with the same zeros and 𝑞 is a
polynomial multiple of 𝑝. Prove that there exists 𝑇 ∈ ℒ(𝐂deg 𝑞 ) such that
the characteristic polynomial of 𝑇 is 𝑞 and the minimal polynomial of 𝑇 is 𝑝.
This exercise implies that every monic polynomial is the characteristic
polynomial of some operator.
22 Suppose 𝐴 and 𝐵 are block diagonal matrices of the form

𝐴 0 𝐵 0
⎛ 1 ⎞ ⎛ 1 ⎞
𝐴=⎜
⎜
⎜ ⋱ ⎟
⎟,
⎟ 𝐵=⎜
⎜
⎜ ⋱ ⎟
⎟,
⎟
⎝ 0 𝐴𝑚 ⎠ ⎝ 0 𝐵𝑚 ⎠
where 𝐴𝑘 and 𝐵𝑘 are square matrices of the same size for each 𝑘 = 1, …, 𝑚.
Show that 𝐴𝐵 is a block diagonal matrix of the form
𝐴1 𝐵1 0
⎛
⎜ ⎞
⎟
⎜
𝐴𝐵 = ⎜ ⋱ ⎟
⎟.
⎝ 0 𝐴𝑚 𝐵𝑚 ⎠
23 Suppose 𝐅 = 𝐑 , 𝑇 ∈ ℒ(𝑉), and 𝜆 ∈ 𝐂 .

(a) Show that 𝑢 + 𝑖𝑣 ∈ 𝐺(𝜆, 𝑇𝐂 ) if and only if 𝑢 − 𝑖𝑣 ∈ 𝐺(𝜆, 𝑇𝐂 ).
(b) Show that the multiplicity of 𝜆 as an eigenvalue of 𝑇𝐂 equals the
multiplicity of 𝜆 as an eigenvalue of 𝑇𝐂 .
(c) Use (b) and the result about the sum of the multiplicities (8.25) to show
that if dim 𝑉 is an odd number, then 𝑇𝐂 has a real eigenvalue.
(d) Use (c) and the result about real eigenvalues of 𝑇𝐂 (Exercise 17 in
Section 5A) to show that if dim 𝑉 is an odd number, then 𝑇 has an
eigenvalue (thus giving an alternative proof of 5.34).
See Exercise 33 in Section 3B for the definition of the complexification 𝑇𝐂 .
Section 8C Consequences of Generalized Eigenspace Decomposition 319
8C Consequences of Generalized Eigenspace Decomposition

Square Roots of Operators
Recall that a square root of an operator 𝑇 ∈ ℒ(𝑉) is an operator 𝑅 ∈ ℒ(𝑉) such
that 𝑅2 = 𝑇 (see 7.36). Every complex number has a square root, but not every
operator on a complex vector space has a square root. For example, the operator
on 𝐂3 defined by 𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (𝑧2 , 𝑧3 , 0) does not have a square root, as you are
asked to show in Exercise 1. The noninvertibility of that operator is no accident,
as we will soon see. We begin by showing that the identity plus any nilpotent
operator has a square root.
8.39 identity plus nilpotent has a square root
Suppose 𝑇 ∈ ℒ(𝑉) is nilpotent. Then 𝐼 + 𝑇 has a square root.
Proof Consider the Taylor series for the function √1 + 𝑥:
8.40 √1 + 𝑥 = 1 + 𝑎 𝑥 + 𝑎 𝑥2
1 2 + ⋯.
We do not find an explicit formula for

Because 𝑎1 = 12 , the formula above
the coefficients or worry about whether
implies that 1 + 𝑥2 is a good estimate
the infinite sum converges because we
for √1 + 𝑥 when 𝑥 is small.
use this equation only as motivation.
Because 𝑇 is nilpotent, 𝑇 𝑚 = 0 for
some positive integer 𝑚. In 8.40, suppose we replace 𝑥 with 𝑇 and 1 with 𝐼. Then
the infinite sum on the right side becomes a finite sum (because 𝑇 𝑘 = 0 for all
𝑘 ≥ 𝑚). Thus we guess that there is a square root of 𝐼 + 𝑇 of the form
𝐼 + 𝑎1 𝑇 + 𝑎2 𝑇 2 + ⋯ + 𝑎𝑚 − 1 𝑇 𝑚 − 1.
Having made this guess, we can try to choose 𝑎1 , 𝑎2 , …, 𝑎𝑚 − 1 such that the operator
above has its square equal to 𝐼 + 𝑇. Now
2
(𝐼+𝑎1 𝑇 + 𝑎2 𝑇 2 + 𝑎3 𝑇 3 + ⋯ + 𝑎𝑚 − 1 𝑇 𝑚 − 1 )
= 𝐼 + 2𝑎1 𝑇 + (2𝑎2 + 𝑎12 )𝑇 2 + (2𝑎3 + 2𝑎1 𝑎2 )𝑇 3 + ⋯
+ (2𝑎𝑚 − 1 + terms involving 𝑎1 , …, 𝑎𝑚 − 2 )𝑇 𝑚 − 1.
We want the right side of the equation above to equal 𝐼 + 𝑇. Hence choose 𝑎1
such that 2𝑎1 = 1 (thus 𝑎1 = 1/2). Next, choose 𝑎2 such that 2𝑎2 + 𝑎12 = 0 (thus
𝑎2 = −1/8). Then choose 𝑎3 such that the coefficient of 𝑇 3 on the right side of
the equation above equals 0 (thus 𝑎3 = 1/16). Continue in this fashion for each
𝑘 = 4, …, 𝑚 − 1, at each step solving for 𝑎𝑘 so that the coefficient of 𝑇 𝑘 on the right
side of the equation above equals 0. Actually we do not care about the explicit
formula for the 𝑎𝑘 ’s. We only need to know that some choice of the 𝑎𝑘 ’s gives a
square root of 𝐼 + 𝑇.
The previous lemma is valid on real and complex vector spaces. However, the
result below holds only on complex vector spaces. For example, the operator of
multiplication by −1 on the one-dimensional real vector space 𝐑 has no square
root.
For the proof below, we need to know that every 𝑧 ∈ 𝐂 has
a square root in 𝐂 . To show this, write
𝑧 = 𝑟(cos 𝜃 + 𝑖 sin 𝜃),
where 𝑟 is the length of the line segment in the complex plane

from the origin to 𝑧 and 𝜃 is the angle of that line segment with
the positive horizontal axis. Then
√
Representation
𝑟(cos 2𝜃 + 𝑖 sin 2𝜃 ) of a complex
number with
is a square root of 𝑧, as you can verify by showing that the square polar
of the complex number above equals 𝑧. coordinates.
8.41 over 𝐂 , invertible operators have square roots
Suppose 𝑉 is a complex vector space and 𝑇 ∈ ℒ(𝑉) is invertible. Then 𝑇 has

a square root.
Proof Let 𝜆1 , …, 𝜆𝑚 be the distinct eigenvalues of 𝑇. For each 𝑘, there exists a

nilpotent operator 𝑇𝑘 ∈ ℒ(𝐺(𝜆𝑘 , 𝑇)) such that 𝑇|𝐺( 𝜆𝑘, 𝑇) = 𝜆𝑘 𝐼 + 𝑇𝑘 [see 8.22(c)].
Because 𝑇 is invertible, none of the 𝜆𝑘 ’s equals 0, so we can write
𝑇𝑘
𝑇|𝐺( 𝜆𝑘, 𝑇) = 𝜆𝑘 (𝐼 + )
𝜆𝑘
for each 𝑘. Because 𝑇𝑘 /𝜆𝑘 is nilpotent, 𝐼 + 𝑇𝑘 /𝜆𝑘 has a square root (by 8.39).
Multiplying a square root of the complex number 𝜆𝑘 by a square root of 𝐼 + 𝑇𝑘 /𝜆𝑘 ,
we obtain a square root 𝑅𝑘 of 𝑇|𝐺( 𝜆𝑘, 𝑇) .
By the generalized eigenspace decomposition (8.22), a typical vector 𝑣 ∈ 𝑉
can be written uniquely in the form
𝑣 = 𝑢 1 + ⋯ + 𝑢𝑚 ,
where each 𝑢𝑘 is in 𝐺(𝜆𝑘 , 𝑇). Using this decomposition, define an operator

𝑅 ∈ ℒ(𝑉) by
𝑅𝑣 = 𝑅1 𝑢1 + ⋯ + 𝑅𝑚 𝑢𝑚 .
You should verify that this operator 𝑅 is a square root of 𝑇, completing the proof.
By imitating the techniques in this subsection, you should be able to prove that
if 𝑉 is a complex vector space and 𝑇 ∈ ℒ(𝑉) is invertible, then 𝑇 has a 𝑘 th root
Jordan Form
We know that if 𝑉 is a complex vector space, then for every 𝑇 ∈ ℒ(𝑉) there is a
basis of 𝑉 with respect to which 𝑇 has a nice upper-triangular matrix (see 8.37).
In this subsection we will see that we can do even better—there is a basis of 𝑉
with respect to which the matrix of 𝑇 contains 0’s everywhere except possibly on
the diagonal and the line directly above the diagonal.
We begin by looking at two examples of nilpotent operators.
8.42 example: nilpotent operator with nice matrix

Let 𝑇 be the operator on 𝐂4 defined by
𝑇(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 ) = (0, 𝑧1 , 𝑧2 , 𝑧3 ).
Then 𝑇 4 = 0; thus 𝑇 is nilpotent. If 𝑣 = (1, 0, 0, 0), then 𝑇 3 𝑣, 𝑇 2 𝑣, 𝑇𝑣, 𝑣 is a basis
of 𝐂4. The matrix of 𝑇 with respect to this basis is
0 1 0 0
⎛
⎜ ⎞
⎟
⎜
⎜ 0 0 1 0 ⎟
⎟
⎜
⎜ ⎟
⎟.
⎜ 0 0 0 1 ⎟
⎝ 0 0 0 0 ⎠
The next example of a nilpotent operator has more complicated behavior than
the example above.
8.43 example: nilpotent operator with slightly more complicated matrix

Let 𝑇 be the operator on 𝐂6 defined by
𝑇(𝑧1 , 𝑧2 , 𝑧3 , 𝑧4 , 𝑧5 , 𝑧6 ) = (0, 𝑧1 , 𝑧2 , 0, 𝑧4 , 0).
Then 𝑇 3 = 0; thus 𝑇 is nilpotent. In contraast to the nice behavior of the nilpotent
operator of the previous example, for this nilpotent operator there does not exist
a vector 𝑣 ∈ 𝐂6 such that 𝑇 5 𝑣, 𝑇 4 𝑣, 𝑇 3 𝑣, 𝑇 2 𝑣, 𝑇𝑣, 𝑣 is a basis of 𝐂6. However, if
we take 𝑣1 = (1, 0, 0, 0, 0, 0), 𝑣2 = (0, 0, 0, 1, 0, 0), and 𝑣3 = (0, 0, 0, 0, 0, 1), then
𝑇 2 𝑣1 , 𝑇𝑣1 , 𝑣1 , 𝑇𝑣2 , 𝑣2 , 𝑣3 is a basis of 𝐂6. The matrix of 𝑇 with respect to this
basis is
0 1 0 0 0 0
⎛
⎜ ⎛
⎜ ⎞
⎟ ⎞
⎜
⎜ ⎜
⎜ 0 0 1 ⎟ ⎟ 0 0 0 ⎟ ⎟
⎟
⎜
⎜ ⎟
⎜ ⎝ 0 0 0 ⎠ 0 0 0 ⎟ ⎟
⎜
⎜ ⎟
⎟ .
⎜
⎜ 0 0 0 0 1 0 ⎟
⎟
⎜ ( )
⎜
⎜ 0 0 0 0 0 0 ⎟ ⎟
⎟
⎝ 0 0 0 0 0 ( 0 ) ⎠
Here the inner matrices are blocked off to show that we can think of the 6-by-6
matrix above as a block diagonal matrix consisting of a 3-by-3 block with 1’s on
the line above the diagonal and 0’s elsewhere, a 2-by-2 block with 1 above the
diagonal and 0’s elsewhere, and a 1-by-1 block containing 0.
Our next goal is to show that every nilpotent operator 𝑇 ∈ ℒ(𝑉) behaves
similarly to the operator in the previous example. Specifically, there is a finite
collection of vectors 𝑣1 , …, 𝑣𝑛 ∈ 𝑉 such that there is a basis of 𝑉 consisting of
the vectors of the form 𝑇 𝑗 𝑣𝑘 , as 𝑘 varies from 1 to 𝑛 and 𝑗 varies (in reverse order)
from 0 to the largest nonnegative integer 𝑚𝑘 such that 𝑇 𝑚𝑘 𝑣𝑘 ≠ 0. With respect to
this basis, the matrix of 𝑇 looks like the matrix in the previous example. More
specifically, 𝑇 has a block diagonal matrix with respect to this basis, with each
block a square matrix that is 0 everywhere except on the line above the diagonal.
In the next definition, the diagonal of each 𝐴𝑘 is filled with some eigenvalue
𝜆𝑘 of 𝑇, the line directly above the diagonal of 𝐴𝑘 is filled with 1’s, and all other
entries in 𝐴𝑘 are 0 (to understand why each 𝜆𝑘 is an eigenvalue of 𝑇, see 5.41).
The 𝜆𝑘 ’s need not be distinct. Also, 𝐴𝑘 may be a 1-by-1 matrix (𝜆𝑘 ) containing
just an eigenvalue of 𝑇. If each 𝜆𝑘 is 0, then the next definition captures the
behavior described in the paragraph above (recall that if 𝑇 is nilpotent, then 0 is
the only eigenvalue of 𝑇).
8.44 definition: Jordan basis

Suppose 𝑇 ∈ ℒ(𝑉). A basis of 𝑉 is called a Jordan basis for 𝑇 if with respect
to this basis 𝑇 has a block diagonal matrix
𝐴1 0
⎛
⎜ ⎞
⎟
⎜
⎜ ⋱ ⎟
⎟
⎝ 0 𝐴𝑝 ⎠
in which each 𝐴𝑘 is an upper-triangular matrix of the form

𝜆𝑘 1 0
⎛
⎜ ⎞
⎟
⎜ ⋱ ⋱ ⎟
𝐴𝑘 = ⎜
⎜
⎜
⎟
⎟.
⎜ ⋱ 1 ⎟⎟
⎝ 0 𝜆𝑘 ⎠
Most of the work in proving that every operator on a finite-dimensional com-

plex vector space has a Jordan basis occurs in proving the special case below
of nilpotent operators. This special case holds on real vector spaces as well as
complex vector spaces.
8.45 every nilpotent operator has a Jordan basis
Suppose 𝑇 ∈ ℒ(𝑉) is nilpotent. Then there is a basis of 𝑉 that is a Jordan

basis for 𝑇.
Proof We will prove this result by induction on dim 𝑉. To get started, note that
the desired result holds if dim 𝑉 = 1 (because in that case, the only nilpotent
operator is the 0 operator). Now assume that dim 𝑉 > 1 and that the desired result
holds on all vector spaces of smaller dimension.
Let 𝑚 be the smallest positive integer such that 𝑇 𝑚 = 0. Thus there exists
𝑢 ∈ 𝑉 such that 𝑇 𝑚 − 1 𝑢 ≠ 0. Let
𝑈 = span(𝑢, 𝑇𝑢, …, 𝑇 𝑚 − 1 𝑢).
The list 𝑢, 𝑇𝑢, …, 𝑇 𝑚 − 1 𝑢 is linearly independent (see Exercise 2 in Section 8A).
If 𝑈 = 𝑉, then writing this list in reverse order gives a Jordan basis for 𝑇 and we
are done. Thus we can assume that 𝑈 ≠ 𝑉.
Note that 𝑈 is invariant under 𝑇. By our induction hypothesis, there is a basis
of 𝑈 that is a Jordan basis for 𝑇|𝑈 . The strategy of our proof is that we will find a
subspace 𝑊 of 𝑉 such that 𝑊 is also invariant under 𝑇 and 𝑉 = 𝑈 ⊕ 𝑊. Again
by our induction hypothesis, there will be a basis of 𝑊 that is a Jordan basis for
𝑇|𝑊 . Putting together the Jordan bases for 𝑇|𝑈 and 𝑇|𝑊 , we will have a Jordan
basis for 𝑇.
Let 𝜑 ∈ 𝑉 ′ be such that 𝜑(𝑇 𝑚 − 1 𝑢) ≠ 0. Let
𝑊 = {𝑣 ∈ 𝑉 ∶ 𝜑(𝑇 𝑘 𝑣) = 0 for each 𝑘 = 0, …, 𝑚 − 1}.
Then 𝑊 is a subspace of 𝑉 that is invariant under 𝑇 (the invariance holds because

if 𝑣 ∈ 𝑊 then 𝜑(𝑇 𝑘 (𝑇𝑣)) = 0 for 𝑘 = 0, …, 𝑚 − 1, where the case 𝑘 = 𝑚 − 1
holds because 𝑇 𝑚 = 0). We will show that 𝑉 = 𝑈 ⊕ 𝑊, which by the previous
paragraph will complete the proof.
To show that 𝑈 + 𝑊 is a direct sum, suppose 𝑣 ∈ 𝑈 ∩ 𝑊 with 𝑣 ≠ 0. Because
𝑣 ∈ 𝑈, there exist 𝑐0 , …, 𝑐𝑚 − 1 ∈ 𝐅 such that
𝑣 = 𝑐0 𝑢 + 𝑐1 𝑇𝑢 + ⋯ + 𝑐𝑚 − 1 𝑇 𝑚 − 1 𝑢.
Let 𝑗 be the smallest index such that 𝑐𝑗 ≠ 0. Apply 𝑇 𝑚 − 𝑗 − 1 to both sides of the
equation above, getting
𝑇 𝑚 − 𝑗 − 1 𝑣 = 𝑐𝑗 𝑇 𝑚 − 1 𝑢,
where we have used the equation 𝑇 𝑚 = 0. Now apply 𝜑 to both sides of the
equation above, getting
𝜑(𝑇 𝑚 − 𝑗 − 1 𝑣) = 𝑐𝑗 𝜑(𝑇 𝑚 − 1 𝑢) ≠ 0.
The equation above shows that 𝑣 ∉ 𝑊. Hence we have proved that 𝑈 ∩ 𝑊 = {0},
which implies that 𝑈 + 𝑊 is a direct sum (see 1.46).
To show that 𝑈 ⊕ 𝑊 = 𝑉, define 𝑆 ∶ 𝑉 → 𝐅𝑚 by
𝑆𝑣 = (𝜑(𝑣), 𝜑(𝑇𝑣), …, 𝜑(𝑇 𝑚 − 1 𝑣)).
Thus null 𝑆 = 𝑊. Hence
dim 𝑊 = dim null 𝑆 = dim 𝑉 − dim range 𝑆 ≥ dim 𝑉 − 𝑚,
where the second equality comes from the fundamental theorem of linear maps
(3.21). Using the inequality above, we have
dim(𝑈 ⊕ 𝑊) = dim 𝑈 + dim 𝑊 ≥ 𝑚 + (dim 𝑉 − 𝑚) = dim 𝑉.
Thus 𝑈 ⊕ 𝑊 = 𝑉 (by 2.39), completing the proof.
Now the generalized eigenspace de- Camille Jordan (1838–1922) pub-

composition allows us to extend the pre- lished a proof of 8.46 in 1870.
vious result to operators that may not be
nilpotent. Doing this requires that we deal with complex vector spaces.
8.46 Jordan form
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Then there is a basis of 𝑉 that is a Jordan

basis for 𝑇.
Proof Let 𝜆1 , …, 𝜆𝑚 be the distinct eigenvalues of 𝑇. The generalized eigenspace

decomposition states that
𝑉 = 𝐺(𝜆1 , 𝑇) ⊕ ⋯ ⊕ 𝐺(𝜆𝑚 , 𝑇),
where each (𝑇 − 𝜆𝑘 𝐼)|𝐺( 𝜆𝑘, 𝑇) is nilpotent (see 8.22). Thus 8.45 implies that some
basis of each 𝐺(𝜆𝑘 , 𝑇) is a Jordan basis for (𝑇 − 𝜆𝑘 𝐼)|𝐺( 𝜆𝑘, 𝑇) . Put these bases
together to get a basis of 𝑉 that is a Jordan basis for 𝑇.
Exercises 8C
1 Suppose 𝑇 ∈ ℒ(𝐂3 ) is the operator defined by 𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (𝑧2 , 𝑧3 , 0).

Prove that 𝑇 does not have a square root.
2 Define 𝑇 ∈ ℒ(𝐅5 ) by 𝑇(𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 , 𝑥5 ) = (2𝑥2 , 3𝑥3 , −𝑥4 , 4𝑥5 , 0).
(a) Show that 𝑇 is nilpotent.
(b) Find a square root of 𝐼 + 𝑇.
3 Suppose 𝑉 is a complex vector space. Prove that every invertible operator

on 𝑉 has a cube root.
4 Suppose 𝑉 is a real vector space. Prove that the operator −𝐼 on 𝑉 has a
square root if and only if dim 𝑉 is an even number.
5 Suppose 𝑇 ∈ ℒ(𝐂2 ) is the operator defined by 𝑇(𝑤, 𝑧) = (−𝑤 − 𝑧, 9𝑤 + 5𝑧).
Find a Jordan basis for 𝑇.
6 Find a basis of 𝒫4 (𝐑) that is a Jordan basis for the differentiation operator
𝐷 on 𝒫4 (𝐑) defined by 𝐷𝑝 = 𝑝′.
7 Suppose 𝑇 ∈ ℒ(𝑉) is nilpotent and 𝑣1 , …, 𝑣𝑛 is a Jordan basis for 𝑇. Prove

that the minimal polynomial of 𝑇 is 𝑧𝑚 + 1 , where 𝑚 is the length of the
longest consecutive string of 1’s that appears on the line directly above the
diagonal in the matrix of 𝑇 with respect to 𝑣1 , …, 𝑣𝑛 .
8 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 that is a Jordan basis for 𝑇.
Describe the matrix of 𝑇 2 with respect to this basis.
9 Suppose 𝑇 ∈ ℒ(𝑉) is nilpotent. Explain why there exist 𝑣1 , …, 𝑣𝑛 ∈ 𝑉 and

nonnegative integers 𝑚1 , …, 𝑚𝑛 such that (a) and (b) below both hold.
(a) 𝑇 𝑚1 𝑣1 , …, 𝑇𝑣1 , 𝑣1 , …, 𝑇 𝑚𝑛 𝑣𝑛 , …, 𝑇𝑣𝑛 , 𝑣𝑛 is a basis of 𝑉.
(b) 𝑇 𝑚1 + 1 𝑣1 = ⋯ = 𝑇 𝑚𝑛 + 1 𝑣𝑛 = 0.
10 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉 that is a Jordan basis for 𝑇.

Describe the matrix of 𝑇 with respect to the basis 𝑣𝑛 , …, 𝑣1 obtained by
reversing the order of the 𝑣’s.
11 Suppose 𝑇 ∈ ℒ(𝑉). Explain why every vector in each Jordan basis for 𝑇 is
a generalized eigenvector of 𝑇.
12 Suppose 𝑇 ∈ ℒ(𝑉) is diagonalizable. Show that ℳ(𝑇) is a diagonal matrix
with respect to every Jordan basis for 𝑇.
13 Suppose 𝑇 ∈ ℒ(𝑉) is nilpotent. Prove that if 𝑣1 , …, 𝑣𝑛 are vectors in 𝑉 and
𝑚1 , …, 𝑚𝑛 are nonnegative integers such that
𝑇 𝑚1 𝑣1 , …, 𝑇𝑣1 , 𝑣1 , …, 𝑇 𝑚𝑛 𝑣𝑛 , …, 𝑇𝑣𝑛 , 𝑣𝑛 is a basis of 𝑉
and
𝑇 𝑚1 + 1 𝑣1 = ⋯ = 𝑇 𝑚𝑛 + 1 𝑣𝑛 = 0,
then 𝑇 𝑚1 𝑣1 , …, 𝑇 𝑚𝑛 𝑣𝑛 is a basis of null 𝑇.
This exercise shows that 𝑛 = dim null 𝑇. Thus the positive integer 𝑛 that
appears above depends only on 𝑇 and not on the specific Jordan basis
chosen for 𝑇.
14 Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Prove that there does not exist a direct sum
decomposition of 𝑉 into two nonzero subspaces invariant under 𝑇 if and
only if the minimal polynomial of 𝑇 is of the form (𝑧 − 𝜆)dim 𝑉 for some
𝜆 ∈ 𝐂.
8D Trace: A Connection Between Matrices and Operators

We begin this section by defining the trace of a square matrix. After developing
some properties of the trace of a square matrix, we will use this concept to define
the trace of an operator.
8.47 definition: trace of a matrix
Suppose 𝐴 is a square matrix with entries in 𝐅. The trace of 𝐴, denoted by

tr 𝐴, is defined to be the sum of the diagonal entries of 𝐴.
8.48 example: trace of a 3-by-3 matrix

Suppose
3 −1 −2
⎛ ⎞
𝐴=⎜
⎜
⎜ 3 2 −3 ⎟
⎟
⎟.
⎝ 1 2 0 ⎠
The diagonal entries of 𝐴, which are shown in red above, are 3, 2, and 0. Thus
tr 𝐴 = 3 + 2 + 0 = 5.
Matrix multiplication is not commutative, but the next result shows that the
order of matrix multiplication does not matter to the trace.
8.49 trace of 𝐴𝐵 equals trace of 𝐵𝐴
Suppose 𝐴 is an 𝑚-by-𝑛 matrix and 𝐵 is an 𝑛-by-𝑚 matrix. Then
tr(𝐴𝐵) = tr(𝐵𝐴).
Proof Suppose
𝐴1, 1 ⋯ 𝐴1, 𝑛 𝐵1, 1 ⋯ 𝐵1, 𝑚
⎛
⎜ ⎞
⎟ ⎛
⎜ ⎞
⎟
⎜ ⎟ ⎜ ⎟
⎜ ⋮
𝐴=⎜ ⋮ ⎟,
⎟ ⎜ ⋮
𝐵=⎜ ⋮ ⎟.
⎟
⎝ 𝐴𝑚, 1 ⋯ 𝐴 𝑚, 𝑛 ⎠ ⎝ 𝐵𝑛, 1 ⋯ 𝐵 𝑛, 𝑚 ⎠
th
The 𝑗 term on the diagonal of the 𝑚-by-𝑚 matrix 𝐴𝐵 equals ∑𝑛𝑘 = 1 𝐴𝑗, 𝑘 𝐵𝑘, 𝑗 . Thus
𝑚 𝑛
tr(𝐴𝐵) = ∑ ∑ 𝐴𝑗, 𝑘 𝐵𝑘, 𝑗
𝑗 = 1 𝑘=1
𝑛 𝑚
= ∑ ∑ 𝐵𝑘, 𝑗 𝐴𝑗, 𝑘
𝑘=1 𝑗 = 1
𝑛
= ∑ (𝑘 th term on diagonal of the 𝑛-by-𝑛 matrix 𝐵𝐴)
𝑘=1
= tr(𝐵𝐴),
as desired.
Section 8D Trace: A Connection Between Matrices and Operators 327
We want to define the trace of an operator 𝑇 ∈ ℒ(𝑉) to be the trace of the

matrix of 𝑇 with respect to some basis of 𝑉. However, this definition should not
depend on the choice of basis. The following result will make this possible.
8.50 trace of matrix of operator does not depend on basis
Suppose 𝑇 ∈ ℒ(𝑉). Suppose 𝑢1 , …, 𝑢𝑛 and 𝑣1 , …, 𝑣𝑛 are bases of 𝑉. Then
tr ℳ(𝑇, (𝑢1 , …, 𝑢𝑛 )) = tr ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )).
Proof Let 𝐴 = ℳ(𝑇, (𝑢1 , …, 𝑢𝑛 )) and 𝐵 = ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )). The change-of-
basis formula tells us that there exists an invertible 𝑛-by-𝑛 matrix 𝐶 such that
𝐴 = 𝐶−1 𝐵𝐶 (see 3.84). Thus
tr 𝐴 = tr((𝐶−1 𝐵)𝐶)
= tr(𝐶(𝐶−1 𝐵))
= tr((𝐶𝐶−1 )𝐵)
= tr 𝐵,
where the second line comes from 8.49.
Because of 8.50, the following definition now makes sense.
8.51 definition: trace of an operator
Suppose 𝑇 ∈ ℒ(𝑉). The trace of 𝑇, denote tr 𝑇, is defined by
tr 𝑇 = tr ℳ(𝑇, (𝑣1 , …, 𝑣𝑛 )),
where 𝑣1 , …, 𝑣𝑛 is any basis of 𝑉.
Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆 is an eigenvalue of 𝑇. Recall that we defined the

multiplicity of 𝜆 to be the dimension of the generalized eigenspace 𝐺(𝜆, 𝑇) (see
8.23); we proved that this multiplicity equals dim null(𝑇 − 𝜆𝐼)dim 𝑉 (see 8.20).
Recall also that if 𝑉 is a complex vector space, then the sum of the multiplicities
of all eigenvalues of 𝑇 equals dim 𝑉 (see 8.25).
In the definition below, the sum of the eigenvalues “with each eigenvalue
included as many times as its multiplicity” means that if 𝜆1 , …, 𝜆𝑚 are the distinct
eigenvalues of 𝑇 with multiplicities 𝑑1 , …, 𝑑𝑚 , then the sum is
𝑑1 𝜆1 + ⋯ + 𝑑𝑚 𝜆𝑚 .
Or if you prefer to work with a list of not-necessarily-distinct eigenvalues, with
each eigenvalue included as many times as its multiplicity, then the eigenvalues
could be denoted by 𝜆1 , …, 𝜆𝑛 (where 𝑛 equals dim 𝑉) and the sum is
𝜆1 + ⋯ + 𝜆𝑛 .
8.52 on complex vector spaces, trace equals sum of eigenvalues
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Then tr 𝑇 equals the sum of the eigenvalues

of 𝑇, with each eigenvalue included as many times as its multiplicity.
Proof There is a basis of 𝑉 with respect to which 𝑇 has an upper-triangular

matrix with the diagonal entries of the matrix consisting of the eigenvalues of 𝑇,
with each eigenvalue included as many times as its multiplicity—see 8.37. Thus
the definition of the trace of an operator along with 8.50, which allows us to use a
basis of our choice, implies that tr 𝑇 equals the sum of the eigenvalues of 𝑇, with
each eigenvalue included as many times as its multiplicity.
8.53 example: trace of an operator on 𝐂3

Suppose 𝑇 ∈ ℒ(𝐂3 ) is defined by
𝑇(𝑧1 , 𝑧2 , 𝑧3 ) = (3𝑧1 − 𝑧2 − 2𝑧3 , 3𝑧1 + 2𝑧2 − 3𝑧3 , 𝑧1 + 2𝑧2 ).
Then the matrix of 𝑇 with respect to the standard basis of 𝐂3 is
3 −1 −2
⎛
⎜ ⎞
⎟
⎜
⎜ 3 2 −3 ⎟
⎟.
⎝ 1 2 0 ⎠
Adding up the diagonal entries of this matrix, we see that tr 𝑇 = 5.
The eigenvalues of 𝑇 are 1, 2 + 3𝑖, and 2 − 3𝑖, each with multiplicity 1, as
you can verify. The sum of these eigenvalues, each included as many times as its
multiplicity, is 1 + (2 + 3𝑖) + (2 − 3𝑖), which equals 5, as expected by 8.52.
The trace has a close connection with the characteristic polynomial. Suppose
𝐅 = 𝐂 , 𝑇 ∈ ℒ(𝑉), and 𝜆1 , …, 𝜆𝑛 are the eigenvalues of 𝑇, with each eigenvalue
included as many times as its multiplicity. Then by definition (see 8.26), the
characteristic polynomial of 𝑇 equals
(𝑧 − 𝜆1 )⋯(𝑧 − 𝜆𝑛 ).
Expanding the polynomial above, we can write the characteristic polynomial of 𝑇
in the form
𝑧𝑛 − (𝜆1 + ⋯ + 𝜆𝑛 )𝑧𝑛 − 1 + ⋯ + (−1)𝑛 (𝜆1 ⋯𝜆𝑛 ).
The expression above immediately leads to the next result. Also see 9.65,
which does not require the hypothesis that 𝐅 = 𝐂 .
8.54 trace and characteristic polynomial
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Let 𝑛 = dim 𝑉. Then tr 𝑇 equals the negative

of the coefficient of 𝑧𝑛 − 1 in the characteristic polynomial of 𝑇.
The next result gives a nice formula for the trace of an operator on an inner
product space.
8.55 trace on an inner product space
Suppose 𝑉 is an inner product space, 𝑇 ∈ ℒ(𝑉), and 𝑒1 , …, 𝑒𝑛 is an orthonor-

mal basis of 𝑉. Then
tr 𝑇 = ⟨𝑇𝑒1 , 𝑒1 ⟩ + ⋯ + ⟨𝑇𝑒𝑛 , 𝑒𝑛 ⟩.
Proof The desired formula follows from the observation that the entry in row 𝑘,
column 𝑘 of ℳ(𝑇, (𝑒1 , …, 𝑒𝑛 )) equals ⟨𝑇𝑒𝑘 , 𝑒𝑘 ⟩ [use 6.30(a) with 𝑣 = 𝑇𝑒𝑘 ].
The algebraic properties of the trace as defined on square matrices translate

to algebraic properties of the trace as defined on operators, as shown in the next
result.
8.56 trace is linear

The function tr ∶ ℒ(𝑉) → 𝐅 is a linear functional on ℒ(𝑉) such that
tr(𝑆𝑇) = tr(𝑇𝑆)
for all 𝑆, 𝑇 ∈ ℒ(𝑉).
Proof Choose a basis of 𝑉. All matrices of operators in this proof will be with
respect to that basis. Suppose 𝑆, 𝑇 ∈ ℒ(𝑉).
If 𝜆 ∈ 𝐅, then
tr(𝜆𝑇) = tr ℳ(𝜆𝑇) = tr(𝜆ℳ(𝑇)) = 𝜆 tr ℳ(𝑇) = 𝜆 tr 𝑇,
where the first and last equalities come from the definition of the trace of an
operator, the second equality comes from 3.38, and the third equality follows
from the definition of the trace of a square matrix.
Also,
tr(𝑆 + 𝑇) = tr ℳ(𝑆 + 𝑇) = tr(ℳ(𝑆) + ℳ(𝑇)) = tr ℳ(𝑆) + tr ℳ(𝑇) = tr 𝑆 + tr 𝑇,
where the first and last equalities come from the definition of the trace of an
operator, the second equality comes from 3.35, and the third equality follows
from the definition of the trace of a square matrix. The two paragraphs above
show that tr ∶ ℒ(𝑉) → 𝐅 is a linear functional on ℒ(𝑉).
Furthermore,
tr(𝑆𝑇) = tr ℳ(𝑆𝑇) = tr(ℳ(𝑆)ℳ(𝑇)) = tr(ℳ(𝑇)ℳ(𝑆)) = tr ℳ(𝑇𝑆) = tr(𝑇𝑆),
where the second and fourth equalities come from 3.43 and the crucial third
equality comes from 8.49.
The equations tr(𝑆𝑇) = tr(𝑇𝑆) and tr 𝐼 = dim 𝑉 uniquely characterize the

trace among the linear functionals on ℒ(𝑉)—see Exercise 10.
The equation tr(𝑆𝑇) = tr(𝑇𝑆) leads The statement of the next result does
to our next result, which does not hold on not involve traces, but the short proof
infinite-dimensional vector spaces (see uses traces. When something like this
Exercise 13). However, additional hy- happens in mathematics, then usually
potheses on 𝑆, 𝑇, and 𝑉 lead to an infinite- a good definition lurks in the back-
dimensional generalization of the result ground.
below, with important applications to
quantum theory.
8.57 identity operator is not the difference of 𝑆𝑇 and 𝑇𝑆
There do not exist operators 𝑆, 𝑇 ∈ ℒ(𝑉) such that 𝑆𝑇 − 𝑇𝑆 = 𝐼.
Proof Suppose 𝑆, 𝑇 ∈ ℒ(𝑉). Then

tr(𝑆𝑇 − 𝑇𝑆) = tr(𝑆𝑇) − tr(𝑇𝑆) = 0,
where both equalities come from 8.56. The trace of 𝐼 equals dim 𝑉, which is not 0.
Because 𝑆𝑇 − 𝑇𝑆 and 𝐼 have different traces, they cannot be equal.
Exercises 8D
1 Suppose 𝑉 is an inner product space and 𝑣, 𝑤 ∈ 𝑉. Define an operator

𝑇 ∈ ℒ(𝑉) by 𝑇𝑢 = ⟨𝑢, 𝑣⟩𝑤. Find a formula for tr 𝑇.
2 Suppose 𝑃 ∈ ℒ(𝑉) satisfies 𝑃2 = 𝑃. Prove that
tr 𝑃 = dim range 𝑃.
3 Suppose 𝑇 ∈ ℒ(𝑉) and 𝑇 = 𝑇. Prove that the real and imaginary parts of
5
tr 𝑇 are both integers.

4 Suppose 𝑉 is an inner product space and 𝑇 ∈ ℒ(𝑉). Prove that
tr 𝑇 ∗ = tr 𝑇.
5 Suppose 𝑉 is an inner product space. Suppose 𝑇 ∈ ℒ(𝑉) is a positive
operator and tr 𝑇 = 0. Prove that 𝑇 = 0.
6 Suppose 𝑉 is an inner product space and 𝑃, 𝑄 ∈ ℒ(𝑉) are orthogonal
projections. Prove that tr(𝑃𝑄) ≥ 0.
7 Suppose 𝑇 ∈ ℒ(𝐂3 ) is the operator whose matrix is
51 −12 −21
⎛
⎜ ⎞
⎟
⎜
⎜ 60 −40 −28 ⎟
⎟.
⎝ 57 −68 1 ⎠
Someone tells you (accurately) that −48 and 24 are eigenvalues of 𝑇. Without
using a computer or writing anything down, find the third eigenvalue of 𝑇.
8 Prove or give a counterexample: If 𝑆, 𝑇 ∈ ℒ(𝑉), then tr(𝑆𝑇) = (tr 𝑆)(tr 𝑇).

9 Suppose 𝑇 ∈ ℒ(𝑉) is such that tr(𝑆𝑇) = 0 for all 𝑆 ∈ ℒ(𝑉). Prove that
𝑇 = 0.
10 Prove that the trace is the only linear functional 𝜏 ∶ ℒ(𝑉) → 𝐅 such that
𝜏(𝑆𝑇) = 𝜏(𝑇𝑆)
for all 𝑆, 𝑇 ∈ ℒ(𝑉) and 𝜏(𝐼) = dim 𝑉.
Hint: Suppose that 𝑣1 , …, 𝑣𝑛 is a basis of 𝑉. For 𝑗, 𝑘 ∈ {1, …, 𝑛}, define
𝑃𝑗, 𝑘 ∈ ℒ(𝑉) by 𝑃𝑗, 𝑘 (𝑎1 𝑣1 + ⋯ + 𝑎𝑛 𝑣𝑛 ) = 𝑎𝑘 𝑣𝑗 . Prove that
⎧
{1 if 𝑗 = 𝑘,
𝜏(𝑃𝑗, 𝑘 ) = ⎨
{0
⎩ if 𝑗 ≠ 𝑘.
Then for 𝑇 ∈ ℒ(𝑉), use the equation 𝑇 = ∑𝑘𝑛 = 1 ∑𝑛𝑗= 1 ℳ(𝑇)𝑗, 𝑘 𝑃𝑗, 𝑘 to
show that 𝜏(𝑇) = tr 𝑇.
11 Suppose 𝑉 and 𝑊 are inner product spaces and 𝑇 ∈ ℒ(𝑉, 𝑊). Prove that if
𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉 and 𝑓1 , …, 𝑓𝑚 is an orthonormal basis
of 𝑊, then
𝑛 𝑚
tr(𝑇 ∗ 𝑇) = ∑ ∑ |⟨𝑇𝑒𝑘 , 𝑓𝑗 ⟩|2.
𝑘=1 𝑗 = 1
The numbers ⟨𝑇𝑒𝑘 , 𝑓𝑗 ⟩ are the entries of the matrix of 𝑇 with respect to the
orthonormal bases 𝑒1 , …, 𝑒𝑛 and 𝑓1 , …, 𝑓𝑚 . These numbers depend on the
bases, but tr(𝑇 ∗ 𝑇) does not depend on a choice of bases. Thus this exercise
shows that the sum of the squares of the absolute values of the matrix entries
does not depend on which orthonormal bases are used.
12 Suppose 𝑉 and 𝑊 are finite-dimensional inner product spaces.

(a) Prove that ⟨𝑆, 𝑇⟩ = tr(𝑇 ∗ 𝑆) defines an inner product on ℒ(𝑉, 𝑊).
(b) Suppose 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉 and 𝑓1 , …, 𝑓𝑚 is an or-
thonormal basis of 𝑊. Show that the inner product on ℒ(𝑉, 𝑊) from
(a) is the same as the standard inner product on 𝐅𝑚𝑛, where we identify
each element of ℒ(𝑉, 𝑊) with its matrix (with respect to the bases just
mentioned) and then with an element of 𝐅𝑚𝑛.
Caution: The norm of a linear map 𝑇 ∈ ℒ(𝑉, 𝑊) as defined by 7.86 is not
the same as the norm that comes from the inner product in (a) above. Unless
explicitly stated otherwise, always assume that ‖𝑇‖ refers to the norm as
defined by 7.86. The norm that comes from the inner product in (a) is called
the Frobenius norm or the Hilbert–Schmidt norm.
13 Find 𝑆, 𝑇 ∈ ℒ(𝒫(𝐅)) such that 𝑆𝑇 − 𝑇𝑆 = 𝐼.

Hint: Make an appropriate modification of the operators in Example 3.9.
This exercise shows that additional hypotheses are needed on 𝑆 and 𝑇 to
extend 8.57 to the setting of infinite-dimensional vector spaces.
Chapter 9
Multilinear Algebra and Determinants
We begin this chapter by investigating bilinear forms and quadratic forms on a

vector space. Then we will move on to multilinear forms. We will show that the
vector space of alternating 𝑛-linear forms has dimension one on a vector space of
dimension 𝑛. This result will allow us to give a clean basis-free definition of the
determinant of an operator.
This approach to the determinant via alternating multilinear forms leads to
straightforward proofs of key properties of determinants. For example, we will see
that the determinant is multiplicative, meaning that det(𝑆𝑇) = (det 𝑆)(det 𝑇) for
all operators 𝑆 and 𝑇 on the same vector space. We will also see that 𝑇 is invertible
if and only if det 𝑇 ≠ 0. Another important result states that the determinant of
an operator on a complex vector space equals the product of the eigenvalues of
the operator, with each eigenvalue included as many times as its multiplicity.
The chapter concludes with an introduction to tensor products.
• 𝑉 and 𝑊 denote finite-dimensional nonzero vector spaces over 𝐅.
Matthew Petroff CC BY-SA
The Mathematical Institute at the University of Göttingen. This building opened in 1930,
when Emmy Noether (1882–1935) had already been a research mathematician and
faculty member at the university for 15 years (the first eight years without salary).
Noether was fired by the Nazi government in 1933. By then Noether and her
collaborators had created many of the foundations of modern algebra, including an
abstract algebra viewpoint that contributed to the development of linear algebra.
332
https://doi.org/10.1007/978-3-031-41026-0_9
Section 9A Bilinear Forms and Quadratic Forms 333
9A Bilinear Forms and Quadratic Forms

Bilinear Forms
A bilinear form on 𝑉 is a function from 𝑉 × 𝑉 to 𝐅 that is linear in each slot
separately, meaning that if we hold either slot fixed then we have a linear function
in the other slot. Here is the formal definition.
9.1 definition: bilinear form
A bilinear form on 𝑉 is a function 𝛽 ∶ 𝑉 × 𝑉 → 𝐅 such that
𝑣 ↦ 𝛽(𝑣, 𝑢) and 𝑣 ↦ 𝛽(𝑢, 𝑣)
are both linear functionals on 𝑉 for every 𝑢 ∈ 𝑉.
For example, if 𝑉 is a real inner prod- Recall that the term linear functional,
uct space, then the function that takes an used in the definition above, means
ordered pair (𝑢, 𝑣) ∈ 𝑉 × 𝑉 to ⟨𝑢, 𝑣⟩ is a linear function that maps into the
a bilinear form on 𝑉. If 𝑉 is a nonzero scalar field 𝐅. Thus the term bilinear
complex inner product space, then this functional would be more consistent
function is not a bilinear form because terminology than bilinear form, which
the inner product is not linear in the sec- unfortunately has become standard.
ond slot (complex scalars come out of the
second slot as their complex conjugates).
If 𝐅 = 𝐑 , then a bilinear form differs from an inner product in that an inner
product requires symmetry [meaning that 𝛽(𝑣, 𝑤) = 𝛽(𝑤, 𝑣) for all 𝑣, 𝑤 ∈ 𝑉]
and positive definiteness [meaning that 𝛽(𝑣, 𝑣) > 0 for all 𝑣 ∈ 𝑉\{0}], but these
properties are not required for a bilinear form.
9.2 example: bilinear forms
• The function 𝛽 ∶ 𝐅3 × 𝐅3 → 𝐅 defined by
𝛽((𝑥1 , 𝑥2 , 𝑥3 ), (𝑦1 , 𝑦2 , 𝑦3 )) = 𝑥1 𝑦2 − 5𝑥2 𝑦3 + 2𝑥3 𝑦1
is a bilinear form on 𝐅3 .
• Suppose 𝐴 is an 𝑛-by-𝑛 matrix with 𝐴𝑗, 𝑘 ∈ 𝐅 in row 𝑗, column 𝑘. Define a
bilinear form 𝛽𝐴 on 𝐅𝑛 by
𝑛 𝑛
𝛽𝐴 ((𝑥1 , …, 𝑥𝑛 ), (𝑦1 , …, 𝑦𝑛 )) = ∑ ∑ 𝐴𝑗, 𝑘 𝑥𝑗 𝑦𝑘 .
𝑘=1 𝑗 = 1
The first bullet point is a special case of this bullet point with 𝑛 = 3 and
0 1 0
⎛ ⎞
𝐴=⎜
⎜
⎜ 0 0 −5 ⎟
⎟
⎟.
⎝ 2 0 0 ⎠
334 Chapter 9 Multilinear Algebra and Determinants
• Suppose 𝑉 is a real inner product space and 𝑇 ∈ ℒ(𝑉). Then the function
𝛽 ∶ 𝑉 × 𝑉 → 𝐑 defined by
𝛽(𝑢, 𝑣) = ⟨𝑢, 𝑇𝑣⟩
is a bilinear form on 𝑉.
• If 𝑛 is a positive integer, then the function 𝛽 ∶ 𝒫𝑛 (𝐑) × 𝒫𝑛 (𝐑) → 𝐑 defined by
𝛽(𝑝, 𝑞) = 𝑝(2) ⋅ 𝑞′ (3)
is a bilinear form on 𝒫𝑛 (𝐑).
• Suppose 𝜑, 𝜏 ∈ 𝑉 ′. Then the function 𝛽 ∶ 𝑉 × 𝑉 → 𝐅 defined by
𝛽(𝑢, 𝑣) = 𝜑(𝑢) ⋅ 𝜏(𝑣)
• More generally, suppose that 𝜑1 , …, 𝜑𝑛 , 𝜏1 , …, 𝜏𝑛 ∈ 𝑉 ′. Then the function
𝛽 ∶ 𝑉 × 𝑉 → 𝐅 defined by
𝛽(𝑢, 𝑣) = 𝜑1 (𝑢) ⋅ 𝜏1 (𝑣) + ⋯ + 𝜑𝑛 (𝑢) ⋅ 𝜏𝑛 (𝑣)
A bilinear form on 𝑉 is a function from 𝑉 × 𝑉 to 𝐅. Because 𝑉 × 𝑉 is a vector
space, this raises the question of whether a bilinear form can also be a linear map
from 𝑉× 𝑉 to 𝐅. Note that none of the bilinear forms in 9.2 are linear maps except
in some special cases in which the bilinear form is the zero map. Exercise 3 shows
that a bilinear form 𝛽 on 𝑉 is a linear map on 𝑉 × 𝑉 only if 𝛽 = 0.
9.3 definition: 𝑉 (2)
The set of bilinear forms on 𝑉 is denoted by 𝑉 (2).
With the usual operations of addition and scalar multiplication of functions,

𝑉 is a vector space.
(2)
For 𝑇 an operator on an 𝑛-dimensional vector space 𝑉 and a basis 𝑒1 , …, 𝑒𝑛

of 𝑉, we used an 𝑛-by-𝑛 matrix to provide information about 𝑇. We now do the
same thing for bilinear forms on 𝑉.
9.4 definition: matrix of a bilinear form, ℳ(𝛽)
Suppose 𝛽 is a bilinear form on 𝑉 and 𝑒1 , …, 𝑒𝑛 is a basis of 𝑉. The matrix of

𝛽 with respect to this basis is the 𝑛-by-𝑛 matrix ℳ(𝛽) whose entry ℳ(𝛽)𝑗, 𝑘
in row 𝑗, column 𝑘 is given by
ℳ(𝛽)𝑗, 𝑘 = 𝛽(𝑒𝑗 , 𝑒𝑘 ).
If the basis 𝑒1 , …, 𝑒𝑛 is not clear from the context, then the notation
ℳ(𝛽, (𝑒1 , …, 𝑒𝑛 )) is used.
Recall that 𝐅𝑛, 𝑛 denotes the vector space of 𝑛-by-𝑛 matrices with entries in 𝐅
and that dim 𝐅𝑛, 𝑛 = 𝑛2 (see 3.39 and 3.40).
9.5 dim 𝑉 (2) = (dim 𝑉)2
Suppose 𝑒1 , …, 𝑒𝑛 is a basis of 𝑉. Then the map 𝛽 ↦ ℳ(𝛽) is an isomorphism

of 𝑉 (2) onto 𝐅𝑛, 𝑛. Furthermore, dim 𝑉 (2) = (dim 𝑉)2.
Proof The map 𝛽 ↦ ℳ(𝛽) is clearly a linear map of 𝑉 (2) into 𝐅𝑛, 𝑛.
For 𝐴 ∈ 𝐅𝑛, 𝑛, define a bilinear form 𝛽𝐴 on 𝑉 by
𝑛 𝑛
𝛽𝐴 (𝑥1 𝑒1 + ⋯ + 𝑥𝑛 𝑒𝑛 , 𝑦1 𝑒1 + ⋯ + 𝑦𝑛 𝑒𝑛 ) = ∑ ∑ 𝐴𝑗, 𝑘 𝑥𝑗 𝑦𝑘
𝑘=1 𝑗 = 1
for 𝑥1 , …, 𝑥𝑛 , 𝑦1 , …, 𝑦𝑛 ∈ 𝐅 (if 𝑉 = 𝐅 and 𝑒1 , …, 𝑒𝑛 is the standard basis of 𝐅𝑛, this

𝑛
𝛽𝐴 is the same as the bilinear form 𝛽𝐴 in the second bullet point of Example 9.2).
The linear map 𝛽 ↦ ℳ(𝛽) from 𝑉 (2) to 𝐅𝑛, 𝑛 and the linear map 𝐴 ↦ 𝛽𝐴 from
𝐅 to 𝑉 (2) are inverses of each other because 𝛽ℳ(𝛽) = 𝛽 for all 𝛽 ∈ 𝑉 (2) and
𝑛, 𝑛
ℳ(𝛽𝐴 ) = 𝐴 for all 𝐴 ∈ 𝐅𝑛, 𝑛, as you should verify.

Thus both maps are isomorphisms and the two spaces that they connect have
the same dimension. Hence dim 𝑉 (2) = dim 𝐅𝑛, 𝑛 = 𝑛2 = (dim 𝑉)2.
Recall that 𝐶 t denotes the transpose of a matrix 𝐶. The matrix 𝐶 t is obtained

by interchanging the rows and the columns of 𝐶.
9.6 composition of a bilinear form and an operator
Suppose 𝛽 is a bilinear form on 𝑉 and 𝑇 ∈ ℒ(𝑉). Define bilinear forms 𝛼

and 𝜌 on 𝑉 by
𝛼(𝑢, 𝑣) = 𝛽(𝑢, 𝑇𝑣) and 𝜌(𝑢, 𝑣) = 𝛽(𝑇𝑢, 𝑣).
Let 𝑒1 , …, 𝑒𝑛 be a basis of 𝑉. Then
ℳ(𝛼) = ℳ(𝛽)ℳ(𝑇) and ℳ(𝜌) = ℳ(𝑇)t ℳ(𝛽).
Proof If 𝑗, 𝑘 ∈ {1, …, 𝑛}, then

ℳ(𝛼)𝑗, 𝑘 = 𝛼(𝑒𝑗 , 𝑒𝑘 )
= 𝛽(𝑒𝑗 , 𝑇𝑒𝑘 )
𝑛
= 𝛽(𝑒𝑗 , ∑ ℳ(𝑇)𝑚, 𝑘 𝑒𝑚 )
𝑚=1
𝑛
= ∑ 𝛽(𝑒𝑗 , 𝑒𝑚 )ℳ(𝑇)𝑚, 𝑘
𝑚=1
= (ℳ(𝛽)ℳ(𝑇))𝑗, 𝑘 .
Thus ℳ(𝛼) = ℳ(𝛽)ℳ(𝑇). The proof that ℳ(𝜌) = ℳ(𝑇)t ℳ(𝛽) is similar.
The result below shows how the matrix of a bilinear form changes if we change
the basis. The formula in the result below should be compared to the change-
of-basis formula for the matrix of an operator (see 3.84). The two formulas are
similar, except that the transpose 𝐶 t appears in the formula below and the inverse
𝐶−1 appears in the change-of-basis formula for the matrix of an operator.
9.7 change-of-basis formula
Suppose 𝛽 ∈ 𝑉 (2). Suppose 𝑒1 , …, 𝑒𝑛 and 𝑓1 , …, 𝑓𝑛 are bases of 𝑉. Let
𝐴 = ℳ(𝛽, (𝑒1 , …, 𝑒𝑛 )) and 𝐵 = ℳ(𝛽, ( 𝑓1 , …, 𝑓𝑛 ))
and 𝐶 = ℳ(𝐼, (𝑒1 , …, 𝑒𝑛 ), ( 𝑓1 , …, 𝑓𝑛 )). Then
𝐴 = 𝐶 t 𝐵𝐶.
Proof The linear map lemma (3.4) tells us that there exists an operator 𝑇 ∈ ℒ(𝑉)
such that 𝑇 𝑓𝑘 = 𝑒𝑘 for each 𝑘 = 1, …, 𝑛. The definition of the matrix of an operator
with respect to a basis implies that
ℳ(𝑇, ( 𝑓1 , …, 𝑓𝑛 )) = 𝐶.
Define bilinear forms 𝛼, 𝜌 on 𝑉 by

𝛼(𝑢, 𝑣) = 𝛽(𝑢, 𝑇𝑣) and 𝜌(𝑢, 𝑣) = 𝛼(𝑇𝑢, 𝑣) = 𝛽(𝑇𝑢, 𝑇𝑣).
Then 𝛽(𝑒𝑗 , 𝑒𝑘 ) = 𝛽(𝑇 𝑓𝑗 , 𝑇 𝑓𝑘 ) = 𝜌( 𝑓𝑗 , 𝑓𝑘 ) for all 𝑗, 𝑘 ∈ {1, …, 𝑛}. Thus
𝐴 = ℳ(𝜌, ( 𝑓1 , …, 𝑓𝑛 ))
= 𝐶 t ℳ(𝛼, 𝑓1 , …, 𝑓𝑛 ))
= 𝐶 t 𝐵𝐶,
where the second and third lines each follow from 9.6.
9.8 example: the matrix of a bilinear form on 𝒫2 (𝐑)

Define a bilinear form 𝛽 on 𝒫2 (𝐑) by 𝛽(𝑝, 𝑞) = 𝑝(2) ⋅ 𝑞′ (3). Let
𝐴 = ℳ(𝛽, (1, 𝑥 − 2, (𝑥 − 3)2 )) and 𝐵 = ℳ(𝛽, (1, 𝑥, 𝑥2 ))
and 𝐶 = ℳ(𝐼, (1, 𝑥 − 2, (𝑥 − 3)2 ), (1, 𝑥, 𝑥2 )). Then
0 1 0 0 1 6 1 −2 9
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
𝐴=⎜
⎜
⎜ 0 0 0 ⎟
⎟
⎟ and 𝐵 = ⎜
⎜
⎜ 0 2 12 ⎟
⎟
⎟ and 𝐶 = ⎜
⎜ ⎟.
⎜ 0 1 −6 ⎟
⎟
⎝ 0 1 0 ⎠ ⎝ 0 4 24 ⎠ ⎝ 0 0 1 ⎠
Now the change-of-basis formula 9.7 asserts that 𝐴 = 𝐶 t 𝐵𝐶, which you can verify
with matrix multiplication using the matrices above.
Symmetric Bilinear Forms
9.9 definition: symmetric bilinear form, 𝑉sym

(2)
A bilinear form 𝜌 ∈ 𝑉 (2) is called symmetric if
𝜌(𝑢, 𝑤) = 𝜌(𝑤, 𝑢)
for all 𝑢, 𝑤 ∈ 𝑉. The set of symmetric bilinear forms on 𝑉 is denoted by 𝑉sym

(2)
.
9.10 example: symmetric bilinear forms

• If 𝑉 is a real inner product space and 𝜌 ∈ 𝑉 (2) is defined by
𝜌(𝑢, 𝑤) = ⟨𝑢, 𝑤⟩,
then 𝜌 is a symmetric bilinear form on 𝑉.
• Suppose 𝑉 is a real inner product space and 𝑇 ∈ ℒ(𝑉). Define 𝜌 ∈ 𝑉 (2) by
𝜌(𝑢, 𝑤) = ⟨𝑢, 𝑇𝑤⟩.
Then 𝜌 is a symmetric bilinear form on 𝑉 if and only if 𝑇 is a self-adjoint
operator (the previous bullet point is the special case 𝑇 = 𝐼).
• Suppose 𝜌 ∶ ℒ(𝑉) × ℒ(𝑉) → 𝐅 is defined by
𝜌(𝑆, 𝑇) = tr(𝑆𝑇).
Then 𝜌 is a symmetric bilinear form on ℒ(𝑉) because trace is a linear functional
on ℒ(𝑉) and tr(𝑆𝑇) = tr(𝑇𝑆) for all 𝑆, 𝑇 ∈ ℒ(𝑉); see 8.56.
9.11 definition: symmetric matrix
A square matrix 𝐴 is called symmetric if it equals its transpose.
An operator on 𝑉 may have a symmetric matrix with respect to some but not all
bases of 𝑉. In contrast, the next result shows that a bilinear form on 𝑉 has a sym-
metric matrix with respect to either all bases of 𝑉 or with respect to no bases of 𝑉.
9.12 symmetric bilinear forms are diagonalizable
Suppose 𝜌 ∈ 𝑉 (2). Then the following are equivalent.

(a) 𝜌 is a symmetric bilinear form on 𝑉.
(b) ℳ(𝜌, (𝑒1 , …, 𝑒𝑛 )) is a symmetric matrix for every basis 𝑒1 , …, 𝑒𝑛 of 𝑉.
(c) ℳ(𝜌, (𝑒1 , …, 𝑒𝑛 )) is a symmetric matrix for some basis 𝑒1 , …, 𝑒𝑛 of 𝑉.
(d) ℳ(𝜌, (𝑒1 , …, 𝑒𝑛 )) is a diagonal matrix for some basis 𝑒1 , …, 𝑒𝑛 of 𝑉.
Proof First suppose (a) holds, so 𝜌 is a symmetric bilinear form. Suppose

𝑒1 , …, 𝑒𝑛 is a basis of 𝑉 and 𝑗, 𝑘 ∈ {1, …, 𝑛}. Then 𝜌(𝑒𝑗 , 𝑒𝑘 ) = 𝜌(𝑒𝑘 , 𝑒𝑗 ) because 𝜌
is symmetric. Thus ℳ(𝜌, (𝑒1 , …, 𝑒𝑛 )) is a symmetric matrix, showing that (a)
implies (b).
Clearly (b) implies (c).
Now suppose (c) holds and 𝑒1 , …, 𝑒𝑛 is a basis of 𝑉 such that ℳ(𝜌, (𝑒1 , …, 𝑒𝑛 ))
is a symmetric matrix. Suppose 𝑢, 𝑤 ∈ 𝑉. There exist 𝑎1 , …, 𝑎𝑛 , 𝑏1 , …, 𝑏𝑛 ∈ 𝐅
such that 𝑢 = 𝑎1 𝑒1 + ⋯ + 𝑎𝑛 𝑒𝑛 and 𝑤 = 𝑏1 𝑒1 + ⋯ + 𝑏𝑛 𝑒𝑛 . Now
𝑛 𝑛
𝜌(𝑢, 𝑤) = 𝜌( ∑ 𝑎𝑗 𝑒𝑗 , ∑ 𝑏𝑘 𝑒𝑘 )
𝑗=1 𝑘=1
𝑛 𝑛
= ∑ ∑ 𝑎𝑗 𝑏𝑘 𝜌(𝑒𝑗 , 𝑒𝑘 )
𝑗 = 1 𝑘=1
𝑛 𝑛
= ∑ ∑ 𝑎𝑗 𝑏𝑘 𝜌(𝑒𝑘 , 𝑒𝑗 )
𝑗 = 1 𝑘=1
𝑛 𝑛
= 𝜌( ∑ 𝑏𝑘 𝑒𝑘 , ∑ 𝑎𝑗 𝑒𝑗 )
𝑘=1 𝑗=1
= 𝜌(𝑤, 𝑢),
where the third line holds because ℳ(𝜌) is a symmetric matrix. The equation
above shows that 𝜌 is a symmetric bilinear form, proving that (c) implies (a).
At this point, we have proved that (a), (b), (c) are equivalent. Because every
diagonal matrix is symmetric, (d) implies (c). To complete the proof, we will
show that (a) implies (d) by induction on 𝑛 = dim 𝑉.
If 𝑛 = 1, then (a) implies (d) because every 1-by-1 matrix is diagonal. Now
suppose 𝑛 > 1 and the implication (a) ⟹ (d) holds for one less dimension.
Suppose (a) holds, so 𝜌 is a symmetric bilinear form. If 𝜌 = 0, then the matrix of
𝜌 with respect to every basis of 𝑉 is the zero matrix, which is a diagonal matrix.
Hence we can assume that 𝜌 ≠ 0, which means there exist 𝑢, 𝑤 ∈ 𝑉 such that
𝜌(𝑢, 𝑤) ≠ 0. Now
2𝜌(𝑢, 𝑤) = 𝜌(𝑢 + 𝑤, 𝑢 + 𝑤) − 𝜌(𝑢, 𝑢) − 𝜌(𝑤, 𝑤).
Because the left side of the equation above is nonzero, the three terms on the right
cannot all equal 0. Hence there exists 𝑣 ∈ 𝑉 such that 𝜌(𝑣, 𝑣) ≠ 0.
Let 𝑈 = {𝑢 ∈ 𝑉 ∶ 𝜌(𝑢, 𝑣) = 0}. Thus 𝑈 is the null space of the linear
functional 𝑢 ↦ 𝜌(𝑢, 𝑣) on 𝑉. This linear functional is not the zero linear functional
because 𝑣 ∉ 𝑈. Thus dim 𝑈 = 𝑛 − 1. By our induction hypothesis, there is a
basis 𝑒1 , …, 𝑒𝑛 − 1 of 𝑈 such that the symmetric bilinear form 𝜌|𝑈 × 𝑈 has a diagonal
matrix with respect to this basis.
Because 𝑣 ∉ 𝑈, the list 𝑒1 , …, 𝑒𝑛 − 1 , 𝑣 is a basis of 𝑉. Suppose 𝑘 ∈ {1, …, 𝑛−1}.
Then 𝜌(𝑒𝑘 , 𝑣) = 0 by the construction of 𝑈. Because 𝜌 is symmetric, we also
have 𝜌(𝑣, 𝑒𝑘 ) = 0. Thus the matrix of 𝜌 with respect to 𝑒1 , …, 𝑒𝑛 − 1 , 𝑣 is a diagonal
matrix, completing the proof that (a) implies (d).
The previous result states that every symmetric bilinear form has a diagonal
matrix with respect to some basis. If our vector space happens to be a real inner
product space, then the next result shows that every symmetric bilinear form has
a diagonal matrix with respect to some orthonormal basis. Note that the inner
product here is unrelated to the bilinear form.
9.13 diagonalization of a symmetric bilinear form by an orthonormal basis
Suppose 𝑉 is a real inner product space and 𝜌 is a symmetric bilinear form on

𝑉. Then 𝜌 has a diagonal matrix with respect to some orthonormal basis of 𝑉.
Proof Let 𝑓1 , …, 𝑓𝑛 be an orthonormal basis of 𝑉. Let 𝐵 = ℳ(𝜌, ( 𝑓1 , …, 𝑓𝑛 )).

Then 𝐵 is a symmetric matrix (by 9.12). Let 𝑇 ∈ ℒ(𝑉) be the operator such that
ℳ(𝑇, ( 𝑓1 , …, 𝑓𝑛 )) = 𝐵. Thus 𝑇 is self-adjoint.
The real spectral theorem (7.29) states that 𝑇 has a diagonal matrix with respect
to some orthonormal basis 𝑒1 , …, 𝑒𝑛 of 𝑉. Let 𝐶 = ℳ(𝐼, (𝑒1 , …, 𝑒𝑛 ), ( 𝑓1 , …, 𝑓𝑛 )).
Thus 𝐶−1 𝑇𝐶 is the matrix of 𝑇 with respect to the basis 𝑒1 , …, 𝑒𝑛 (by 3.84). Hence
𝐶−1 𝑇𝐶 is a diagonal matrix. Now
𝑀(𝜌, (𝑒1 , …, 𝑒𝑛 )) = 𝐶 t 𝑇𝐶 = 𝐶−1 𝑇𝐶,
where the first equality holds by 9.7 and the second equality holds because 𝐶 is a
unitary matrix with real entries (which implies that 𝐶−1 = 𝐶 t ; see 7.57).
Now we turn our attention to alternating bilinear forms. Alternating multilinear

forms will play a major role in our approach to determinants later in this chapter.
9.14 definition: alternating bilinear form, 𝑉alt

(2)
A bilinear form 𝛼 ∈ 𝑉 (2) is called alternating if
𝛼(𝑣, 𝑣) = 0
for all 𝑣 ∈ 𝑉. The set of alternating bilinear forms on 𝑉 is denoted by 𝑉alt

(2)
.
9.15 example: alternating bilinear forms

• Suppose 𝑛 ≥ 3 and 𝛼 ∶ 𝐅𝑛 × 𝐅𝑛 → 𝐅 is defined by
𝛼((𝑥1 , …, 𝑥𝑛 ), (𝑦1 , …, 𝑦𝑛 )) = 𝑥1 𝑦2 − 𝑥2 𝑦1 + 𝑥1 𝑦3 − 𝑥3 𝑦1 .
Then 𝛼 is an alternating bilinear form on 𝐅𝑛.
• Suppose 𝜑, 𝜏 ∈ 𝑉 ′. Then the bilinear form 𝛼 on 𝑉 defined by
𝛼(𝑢, 𝑤) = 𝜑(𝑢)𝜏(𝑤) − 𝜑(𝑤)𝜏(𝑢)
is alternating.
The next result shows that a bilinear form is alternating if and only if switching
the order of the two inputs multiplies the output by −1.
9.16 characterization of alternating bilinear forms
A bilinear form 𝛼 on 𝑉 is alternating if and only if
𝛼(𝑢, 𝑤) = −𝛼(𝑤, 𝑢)
for all 𝑢, 𝑤 ∈ 𝑉.
Proof First suppose that 𝛼 is alternating. If 𝑢, 𝑤 ∈ 𝑉, then

0 = 𝛼(𝑢 + 𝑤, 𝑢 + 𝑤)
= 𝛼(𝑢, 𝑢) + 𝛼(𝑢, 𝑤) + 𝛼(𝑤, 𝑢) + 𝛼(𝑤, 𝑤)
= 𝛼(𝑢, 𝑤) + 𝛼(𝑤, 𝑢).
Thus 𝛼(𝑢, 𝑤) = −𝛼(𝑤, 𝑢), as desired.
To prove the implication in the other direction, suppose 𝛼(𝑢, 𝑤) = −𝛼(𝑤, 𝑢)
for all 𝑢, 𝑤 ∈ 𝑉. Then 𝛼(𝑣, 𝑣) = −𝛼(𝑣, 𝑣) for all 𝑣 ∈ 𝑉, which implies that
𝛼(𝑣, 𝑣) = 0 for all 𝑣 ∈ 𝑉. Thus 𝛼 is alternating.
Now we show that the vector space of bilinear forms on 𝑉 is the direct sum of
the symmetric bilinear forms on 𝑉 and the alternating bilinear forms on 𝑉.
9.17
(2) (2)
𝑉 (2) = 𝑉sym ⊕ 𝑉alt
The sets 𝑉sym

(2)
and 𝑉alt
(2)
are subspaces of 𝑉 (2) . Furthermore,
(2) (2)
𝑉 (2) = 𝑉sym ⊕ 𝑉alt .
Proof The definition of symmetric bilinear form implies that the sum of any two
symmetric bilinear forms on 𝑉 is a bilinear form on 𝑉, and any scalar multiple of
any bilinear form on 𝑉 is a bilinear form on 𝑉. Thus 𝑉sym (2)
is a subspace of 𝑉 (2).
Similarly, the verification that 𝑉alt is a subspace of 𝑉 is straightforward.
(2) (2)
Next, we want to show that 𝑉 (2) = 𝑉sym (2) (2)

+ 𝑉alt . To do this, suppose 𝛽 ∈ 𝑉 (2).
Define 𝜌, 𝛼 ∈ 𝑉 (2) by
𝛽(𝑢, 𝑤) + 𝛽(𝑤, 𝑢) 𝛽(𝑢, 𝑤) − 𝛽(𝑤, 𝑢)
𝜌(𝑢, 𝑤) = and 𝛼(𝑢, 𝑤) = .
2 2
Then 𝜌 ∈ 𝑉sym(2)
and 𝛼 ∈ 𝑉alt
(2)
, and 𝛽 = 𝜌 + 𝛼. Thus 𝑉 (2) = 𝑉sym(2) (2)
+ 𝑉alt .
Finally, to show that the intersection of the two subspaces under consideration
equals {0}, suppose 𝛽 ∈ 𝑉sym(2) (2)
∩ 𝑉alt . Then 9.16 implies that
𝛽(𝑢, 𝑤) = −𝛽(𝑤, 𝑢) = −𝛽(𝑢, 𝑤)
for all 𝑢, 𝑤 ∈ 𝑉, which implies that 𝛽 = 0. Thus 𝑉 (2) = 𝑉sym
(2) (2)
⊕ 𝑉alt , as implied
by 1.46.
Quadratic Forms
9.18 definition: quadratic form associated with a bilinear form, 𝑞𝛽
For 𝛽 a bilinear form on 𝑉, define a function 𝑞𝛽 ∶ 𝑉 → 𝐅 by 𝑞𝛽 (𝑣) = 𝛽(𝑣, 𝑣).

A function 𝑞 ∶ 𝑉 → 𝐅 is called a quadratic form on 𝑉 if there exists a bilinear
form 𝛽 on 𝑉 such that 𝑞 = 𝑞𝛽 .
Note that if 𝛽 is a bilinear form, then 𝑞𝛽 = 0 if and only if 𝛽 is alternating.
9.19 example: quadratic form

Suppose 𝛽 is the bilinear form on 𝐑3 defined by
𝛽((𝑥1 , 𝑥2 , 𝑥3 ), (𝑦1 , 𝑦2 , 𝑦3 )) = 𝑥1 𝑦1 − 4𝑥1 𝑦2 + 8𝑥1 𝑦3 − 3𝑥3 𝑦3 .
Then 𝑞𝛽 is the quadratic form on 𝐑3 given by the formula
𝑞𝛽 (𝑥1 , 𝑥2 , 𝑥3 ) = 𝑥12 − 4𝑥1 𝑥2 + 8𝑥1 𝑥3 − 3𝑥32.
The quadratic form in the example above is typical of quadratic forms on 𝐅𝑛,
as shown in the next result.
9.20 quadratic forms on 𝐅𝑛
Suppose 𝑛 is a positive integer and 𝑞 is a function from 𝐅𝑛 to 𝐅. Then 𝑞

is a quadratic form on 𝐅𝑛 if and only if there exist numbers 𝐴𝑗, 𝑘 ∈ 𝐅 for
𝑗, 𝑘 ∈ {1, …, 𝑛} such that
𝑛 𝑛
𝑞(𝑥1 , …, 𝑥𝑛 ) = ∑ ∑ 𝐴𝑗, 𝑘 𝑥𝑗 𝑥𝑘
𝑘=1 𝑗 = 1
for all (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛.
Proof First suppose 𝑞 is a quadratic form on 𝐅𝑛. Thus there exists a bilinear form
𝛽 on 𝐅𝑛 such that 𝑞 = 𝑞𝛽 . Let 𝐴 be the matrix of 𝛽 with respect to the standard
basis of 𝐅𝑛. Then for all (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛, we have the desired equation
𝑛 𝑛
𝑞(𝑥1 , …, 𝑥𝑛 ) = 𝛽((𝑥1 , …, 𝑥𝑛 ), (𝑥1 , …, 𝑥𝑛 )) = ∑ ∑ 𝐴𝑗, 𝑘 𝑥𝑗 𝑥𝑘 .
𝑘=1 𝑗 = 1
Conversely, suppose there exist numbers 𝐴𝑗, 𝑘 ∈ 𝐅 for 𝑗, 𝑘 ∈ {1, …, 𝑛} such that
𝑛 𝑘
𝑞(𝑥1 , …, 𝑥𝑛 ) = ∑ ∑ 𝐴𝑗, 𝑘 𝑥𝑗 𝑥𝑘
𝑘=1 𝑗 = 1
for all (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅 . Define a bilinear form 𝛽 on 𝐅𝑛 by

𝑛
𝑛 𝑘
𝛽((𝑥1 , …, 𝑥𝑛 ), (𝑦1 , …, 𝑦𝑛 )) = ∑ ∑ 𝐴𝑗, 𝑘 𝑥𝑗 𝑦𝑘 .
𝑘=1 𝑗 = 1
Then 𝑞 = 𝑞𝛽 , as desired.
Although quadratic forms are defined in terms of an arbitrary bilinear form,

the equivalence of (a) and (b) in the result below shows that a symmetric bilinear
form can always be used.
9.21 characterization of quadratic forms
Suppose 𝑞 ∶ 𝑉 → 𝐅 is a function. The following are equivalent.

(a) 𝑞 is a quadratic form.
(b) There exists a unique symmetric bilinear form 𝜌 on 𝑉 such that 𝑞 = 𝑞𝜌 .
(c) 𝑞(𝜆𝑣) = 𝜆2 𝑞(𝑣) for all 𝜆 ∈ 𝐅 and all 𝑣 ∈ 𝑉, and the function
(𝑢, 𝑤) ↦ 𝑞(𝑢 + 𝑤) − 𝑞(𝑢) − 𝑞(𝑤)
is a symmetric bilinear form on 𝑉.

(d) 𝑞(2𝑣) = 4𝑞(𝑣) for all 𝑣 ∈ 𝑉, and the function
(𝑢, 𝑤) ↦ 𝑞(𝑢 + 𝑤) − 𝑞(𝑢) − 𝑞(𝑤)
is a symmetric bilinear form on 𝑉.
Proof First suppose (a) holds, so 𝑞 is a quadratic form. Hence there exists a
bilinear form 𝛽 such that 𝑞 = 𝑞𝛽 . By 9.17, there exist a symmetric bilinear form 𝜌
on 𝑉 and an alternating bilinear form 𝛼 on 𝑉 such that 𝛽 = 𝜌 + 𝛼. Now
𝑞 = 𝑞𝛽 = 𝑞𝜌 + 𝑞𝛼 = 𝑞𝜌 .
If 𝜌′ ∈ 𝑉sym
(2)
also satisfies 𝑞𝜌′ = 𝑞, then 𝑞𝜌′ − 𝜌 = 0; thus 𝜌′ − 𝜌 ∈ 𝑉sym
(2) (2)
∩ 𝑉alt ,
which implies that 𝜌 = 𝜌 (by 9.17). This completes the proof that (a) implies (b).
′
Now suppose (b) holds, so there exists a symmetric bilinear form 𝜌 on 𝑉 such
that 𝑞 = 𝑞𝜌 . If 𝜆 ∈ 𝐅 and 𝑣 ∈ 𝑉 then
𝑞(𝜆𝑣) = 𝜌(𝜆𝑣, 𝜆𝑣) = 𝜆𝜌(𝑣, 𝜆𝑣) = 𝜆2 𝜌(𝑣, 𝑣) = 𝜆2 𝑞(𝑣),
showing that the first part of (c) holds.
If 𝑢, 𝑤 ∈ 𝑉, then
𝑞(𝑢 + 𝑤) − 𝑞(𝑢) − 𝑞(𝑤) = 𝜌(𝑢 + 𝑤, 𝑢 + 𝑤) − 𝜌(𝑢, 𝑢) − 𝜌(𝑤, 𝑤) = 2𝜌(𝑢, 𝑤).
Thus the function (𝑢, 𝑤) ↦ 𝑞(𝑢 + 𝑤)−𝑞(𝑢)−𝑞(𝑤) equals 2𝜌, which is a symmetric
bilinear form on 𝑉, completing the proof that (b) implies (c).
Clearly (c) implies (d).
Now suppose (d) holds. Let 𝜌 be the symmetric bilinear form on 𝑉 defined by
𝑞(𝑢 + 𝑤) − 𝑞(𝑢) − 𝑞(𝑤)
𝜌(𝑢, 𝑤) = .
2
𝑞(2𝑣) − 𝑞(𝑣) − 𝑞(𝑣) 4𝑞(𝑣) − 2𝑞(𝑣)
𝜌(𝑣, 𝑣) = = = 𝑞(𝑣).
2 2
Thus 𝑞 = 𝑞𝜌 , completing the proof that (d) implies (a).
9.22 example: symmetric bilinear form associated with a quadratic form

Suppose 𝑞 is the quadratic form on 𝐑3 given by the formula
𝑞(𝑥1 , 𝑥2 , 𝑥3 ) = 𝑥12 − 4𝑥1 𝑥2 + 8𝑥1 𝑥3 − 3𝑥32.
A bilinear form 𝛽 on 𝐑3 such that 𝑞 = 𝑞𝛽 is given by Example 9.19, but this
bilinear form is not symmetric, as promised by 9.21(b). However, the bilinear
form 𝜌 on 𝐑3 defined by
𝜌((𝑥1 , 𝑥2 , 𝑥3 ), (𝑦1 , 𝑦2 , 𝑦3 )) = 𝑥1 𝑦1 − 2𝑥1 𝑦2 − 2𝑥2 𝑦1 + 4𝑥1 𝑦3 + 4𝑥3 𝑦1 − 3𝑥3 𝑦3
is symmetric and satisfies 𝑞 = 𝑞𝜌 .
The next result states that for each quadratic form we can choose a basis such
that the quadratic form looks like a weighted sum of squares of the coordinates,
meaning that there are no cross terms of the form 𝑥𝑗 𝑥𝑘 with 𝑗 ≠ 𝑘.
9.23 diagonalization of quadratic form
Suppose 𝑞 is a quadratic form on 𝑉.

(a) There exist a basis 𝑒1 , …, 𝑒𝑛 of 𝑉 and 𝜆1 , …, 𝜆𝑛 ∈ 𝐅 such that
𝑞(𝑥1 𝑒1 + ⋯ + 𝑥𝑛 𝑒𝑛 ) = 𝜆1 𝑥12 + ⋯ + 𝜆𝑛 𝑥𝑛2
for all 𝑥1 , …, 𝑥𝑛 ∈ 𝐅.
(b) If 𝐅 = 𝐑 and 𝑉 is an inner product space, then the basis in (a) can be
chosen to be an orthonormal basis of 𝑉.
Proof
(a) There exists a symmetric bilinear form 𝜌 on 𝑉 such that 𝑞 = 𝑞𝜌 (by 9.21). Now
there exists a basis 𝑒1 , …, 𝑒𝑛 of 𝑉 such that ℳ(𝜌, (𝑒1 , …, 𝑒𝑛 )) is a diagonal
matrix (by 9.12). Let 𝜆1 , …, 𝜆𝑛 denote the entries on the diagonal of this
matrix. Thus
{ 𝜆 if 𝑗 = 𝑘,
⎧
𝜌(𝑒𝑗 , 𝑒𝑘 ) = ⎨ 𝑗
⎩0 if 𝑗 ≠ 𝑘
{
for all 𝑗, 𝑘 ∈ {1, …, 𝑛}. If 𝑥1 , …, 𝑥𝑛 ∈ 𝐅, then
𝑞(𝑥1 𝑒1 + ⋯ + 𝑥𝑛 𝑒𝑛 ) = 𝜌(𝑥1 𝑒1 + ⋯ + 𝑥𝑛 𝑒𝑛 , 𝑥1 𝑒1 + ⋯ + 𝑥𝑛 𝑒𝑛 )
𝑛 𝑛
= ∑ ∑ 𝑥𝑗 𝑥𝑘 𝜌(𝑒𝑗 , 𝑒𝑘 )
𝑘=1 𝑗 = 1
= 𝜆1 𝑥12 + ⋯ + 𝜆𝑛 𝑥𝑛2,
as desired.
(b) Suppose 𝐅 = 𝐑 and 𝑉 is an inner product space. Then 9.13 tells us that the
basis in (a) can be chosen to be an orthonormal basis of 𝑉.
Exercises 9A
1 Prove that if 𝛽 is a bilinear form on 𝐅, then there exists 𝑐 ∈ 𝐅 such that

𝛽(𝑥, 𝑦) = 𝑐𝑥𝑦
for all 𝑥, 𝑦 ∈ 𝐅.
2 Let 𝑛 = dim 𝑉. Suppose 𝛽 is a bilinear form on 𝑉. Prove that there exist
𝜑1 , …, 𝜑𝑛 , 𝜏1 , …, 𝜏𝑛 ∈ 𝑉 ′ such that
𝛽(𝑢, 𝑣) = 𝜑1 (𝑢) ⋅ 𝜏1 (𝑣) + ⋯ + 𝜑𝑛 (𝑢) ⋅ 𝜏𝑛 (𝑣)
This exercise shows that if 𝑛 = dim 𝑉, then every bilinear form on 𝑉 is of
the form given by the last bullet point of Example 9.2.
3 Suppose 𝛽 ∶ 𝑉× 𝑉 → 𝐅 is a bilinear form on 𝑉 and also is a linear functional

on 𝑉 × 𝑉. Prove that 𝛽 = 0.
4 Suppose 𝑉 is a real inner product space and 𝛽 is a bilinear form on 𝑉. Show
that there exists a unique operator 𝑇 ∈ ℒ(𝑉) such that
𝛽(𝑢, 𝑣) = ⟨𝑢, 𝑇𝑣⟩
This exercise states that if 𝑉 is a real inner product space, then every
bilinear form on 𝑉 is of the form given by the third bullet point in 9.2.
5 Suppose 𝛽 is a bilinear form on a real inner product space 𝑉 and 𝑇 is the

unique operator on 𝑉 such that 𝛽(𝑢, 𝑣) = ⟨𝑢, 𝑇𝑣⟩ for all 𝑢, 𝑣 ∈ 𝑉 (see
Exercise 4). Show that 𝛽 is an inner product on 𝑉 if and only if 𝑇 is an
invertible positive operator on 𝑉.
6 Prove or give a counterexample: If 𝜌 is a symmetric bilinear form on 𝑉, then
{𝑣 ∈ 𝑉 ∶ 𝜌(𝑣, 𝑣) = 0}
is a subspace of 𝑉.
7 Explain why the proof of 9.13 (diagonalization of a symmetric bilinear form
by an orthonormal basis on a real inner product space) fails if the hypothesis
that 𝐅 = 𝐑 is dropped.
8 Find formulas for dim 𝑉sym
(2)
and dim 𝑉alt
(2)
in terms of dim 𝑉.
9 Suppose that 𝑛 is a positive integer and 𝑉 = {𝑝 ∈ 𝒫𝑛 (𝐑) ∶ 𝑝(0) = 𝑝(1)}.
Define 𝛼 ∶ 𝑉 × 𝑉 → 𝐑 by
1
𝛼(𝑝, 𝑞) = ∫ 𝑝𝑞′.
0
Show that 𝛼 is an alternating bilinear form on 𝑉.

10 Suppose that 𝑛 is a positive integer and
𝑉 = {𝑝 ∈ 𝒫𝑛 (𝐑) ∶ 𝑝(0) = 𝑝(1) and 𝑝′ (0) = 𝑝′ (1)}.
Define 𝜌 ∶ 𝑉 × 𝑉 → 𝐑 by
1
𝜌(𝑝, 𝑞) = ∫ 𝑝𝑞″.
0
Show that 𝜌 is a symmetric bilinear form on 𝑉.
9B Alternating Multilinear Forms

Multilinear Forms
9.24 definition: 𝑉 𝑚
For 𝑚 a positive integer, define 𝑉 𝑚 by
⏟
𝑉𝑚 = 𝑉 × ⋯ × 𝑉.
𝑚 times
Now we can define 𝑚-linear forms as a generalization of the bilinear forms

that we discussed in the previous section.
9.25 definition: 𝑚-linear form, 𝑉 (𝑚), multilinear form
• For 𝑚 a positive integer, an 𝑚-linear form on 𝑉 is a function 𝛽 ∶ 𝑉 𝑚 → 𝐅

that is linear in each slot when the other slots are held fixed. This means
that for each 𝑘 ∈ {1, …, 𝑚} and all 𝑢1 , …, 𝑢𝑚 ∈ 𝑉, the function
𝑣 ↦ 𝛽(𝑢1 , …, 𝑢𝑘 − 1 , 𝑣, 𝑢𝑘 + 1 , …, 𝑢𝑚 )
is a linear map from 𝑉 to 𝐅.

• The set of 𝑚-linear forms on 𝑉 is denoted by 𝑉 (𝑚).
• A function 𝛽 is called a multilinear form on 𝑉 if it is an 𝑚-linear form on 𝑉
for some positive integer 𝑚.
In the definition above, the expression 𝛽(𝑢1 , …, 𝑢𝑘 − 1 , 𝑣, 𝑢𝑘 + 1 , …, 𝑢𝑚 ) means

𝛽(𝑣, 𝑢2 , …, 𝑢𝑚 ) if 𝑘 = 1 and means 𝛽(𝑢1 , …, 𝑢𝑚 − 1 , 𝑣) if 𝑘 = 𝑚.
A 1-linear form on 𝑉 is a linear functional on 𝑉. A 2-linear form on 𝑉 is
a bilinear form on 𝑉. You can verify that with the usual addition and scalar
multiplication of functions, 𝑉 (𝑚) is a vector space.
9.26 example: 𝑚-linear forms
• Suppose 𝛼, 𝜌 ∈ 𝑉 (2). Define a function 𝛽 ∶ 𝑉 4 → 𝐅 by
𝛽(𝑣1 , 𝑣2 , 𝑣3 , 𝑣4 ) = 𝛼(𝑣1 , 𝑣2 ) 𝜌(𝑣3 , 𝑣4 ).
Then 𝛽 ∈ 𝑉 (4).
• Define 𝛽 ∶ (ℒ(𝑉)) → 𝐅 by
𝑚
𝛽(𝑇1 , …, 𝑇𝑚 ) = tr(𝑇1 ⋯𝑇𝑚 ).
Then 𝛽 is an 𝑚-linear form on ℒ(𝑉).

Section 9B Alternating Multilinear Forms 347
Alternating multilinear forms, which we now define, play an important role as

we head toward defining determinants.
9.27 definition: alternating forms, 𝑉alt

(𝑚)
Suppose 𝑚 is a positive integer.

• An 𝑚-linear form 𝛼 on 𝑉 is called alternating if 𝛼(𝑣1 , …, 𝑣𝑚 ) = 0 whenever
𝑣1 , …, 𝑣𝑚 is a list of vectors in 𝑉 with 𝑣𝑗 = 𝑣𝑘 for some two distinct values
of 𝑗 and 𝑘 in {1, …, 𝑚}.
(𝑚)
• 𝑉alt = {𝛼 ∈ 𝑉 (𝑚) ∶ 𝛼 is an alternating 𝑚-linear form on 𝑉}.
You should verify that 𝑉alt (𝑚)

is a subspace of 𝑉 (𝑚). See Example 9.15 for
examples of alternating 2-linear forms. See Exercise 2 for an example of an
alternating 3-linear form.
The next result tells us that if a linearly dependent list is input to an alternating
multilinear form, then the output equals 0.
9.28 alternating multilinear forms and linear dependence
Suppose 𝑚 is a positive integer and 𝛼 is an alternating 𝑚-linear form on 𝑉. If

𝑣1 , …, 𝑣𝑚 is a linearly dependent list in 𝑉, then
𝛼(𝑣1 , …, 𝑣𝑚 ) = 0.
Proof Suppose 𝑣1 , …, 𝑣𝑚 is a linearly dependent list in 𝑉. By the linear depen-

dence lemma (2.19), some 𝑣𝑘 is a linear combination of 𝑣1 , …, 𝑣𝑘 − 1 . Thus there
exist 𝑏1 , …, 𝑏𝑘 − 1 such that 𝑣𝑘 = 𝑏1 𝑣1 + ⋯ + 𝑏𝑘 − 1 𝑣𝑘 − 1 . Now
𝑘−1
𝛼(𝑣1 , …, 𝑣𝑚 ) = 𝛼(𝑣1 , …, 𝑣𝑘 − 1 , ∑ 𝑏𝑗 𝑣𝑗 , 𝑣𝑘 + 1 , …, 𝑣𝑚 )
𝑗=1
𝑘−1
= ∑ 𝑏𝑗 𝛼(𝑣1 , …, 𝑣𝑘 − 1 , 𝑣𝑗 , 𝑣𝑘 + 1 , …, 𝑣𝑚 )
𝑗=1
= 0.
The next result states that if 𝑚 > dim 𝑉, then there are no alternating 𝑚-linear
forms on 𝑉 other than the function on 𝑉 𝑚 that is identically 0.
9.29 no nonzero alternating 𝑚-linear forms for 𝑚 > dim 𝑉
Suppose 𝑚 > dim 𝑉. Then 0 is the only alternating 𝑚-linear form on 𝑉.
Proof Suppose that 𝛼 is an alternating 𝑚-linear form on 𝑉 and 𝑣1 , …, 𝑣𝑚 ∈ 𝑉.

Because 𝑚 > dim 𝑉, this list is not linearly independent (by 2.22). Thus 9.28
implies that 𝛼(𝑣1 , …, 𝑣𝑚 ) = 0. Hence 𝛼 is the zero function from 𝑉 𝑚 to 𝐅.
Alternating Multilinear Forms and Permutations
9.30 swapping input vectors in an alternating multilinear form
Suppose 𝑚 is a positive integer, 𝛼 is an alternating 𝑚-linear form on 𝑉, and

𝑣1 , …, 𝑣𝑚 is a list of vectors in 𝑉. Then swapping the vectors in any two slots
of 𝛼(𝑣1 , …, 𝑣𝑚 ) changes the value of 𝛼 by a factor of −1.
Proof Put 𝑣1 + 𝑣2 in both the first two slots, getting

0 = 𝛼(𝑣1 + 𝑣2 , 𝑣1 + 𝑣2 , 𝑣3 , …, 𝑣𝑚 ).
Use the multilinear properties of 𝛼 to expand the right side of the equation above
(as in the proof of 9.16) to get
𝛼(𝑣2 , 𝑣1 , 𝑣3 , …, 𝑣𝑚 ) = −𝛼(𝑣1 , 𝑣2 , 𝑣3 , …, 𝑣𝑚 ).
Similarly, swapping the vectors in any two slots of 𝛼(𝑣1 , …, 𝑣𝑚 ) changes the
value of 𝛼 by a factor of −1.
To see what can happen with multiple swaps, suppose 𝛼 is an alternating

3-linear form on 𝑉 and 𝑣1 , 𝑣2 , 𝑣3 ∈ 𝑉. To evaluate 𝛼(𝑣3 , 𝑣1 , 𝑣2 ) in terms of
𝛼(𝑣1 , 𝑣2 , 𝑣3 ), start with 𝛼(𝑣3 , 𝑣1 , 𝑣2 ) and swap the entries in the first and third
slots, getting 𝛼(𝑣3 , 𝑣1 , 𝑣2 ) = −𝛼(𝑣2 , 𝑣1 , 𝑣3 ). Now in the last expression, swap the
entries in the first and second slots, getting
𝛼(𝑣3 , 𝑣1 , 𝑣2 ) = −𝛼(𝑣2 , 𝑣1 , 𝑣3 ) = 𝛼(𝑣1 , 𝑣2 , 𝑣3 ).
More generally, we see that if we do an odd number of swaps, then the value of 𝛼
changes by a factor of −1, and if we do an even number of swaps, then the value
of 𝛼 does not change.
To deal with arbitrary multiple swaps, we need a bit of information about
permutations.
9.31 definition: permutation, perm 𝑚
Suppose 𝑚 is a positive integer.

• A permutation of (1, …, 𝑚) is a list (𝑗1 , …, 𝑗𝑚 ) that contains each of the
numbers 1, …, 𝑚 exactly once.
• The set of all permutations of (1, …, 𝑚) is denoted by perm 𝑚.
For example, (2, 3, 4, 5, 1) ∈ perm 5. You should think of an element of

perm 𝑚 as a rearrangement of the first 𝑚 positive integers.
The number of swaps used to change a permutation (𝑗1 , …, 𝑗𝑚 ) to the stan-
dard order (1, …, 𝑚) can depend on the specific swaps selected. The following
definition has the advantage of assigning a well-defined sign to every permutation.
9.32 definition: sign of a permutation
The sign of a permutation (𝑗1 , …, 𝑗𝑚 ) is defined by
sign(𝑗1 , …, 𝑗𝑚 ) = (−1)𝑁,
where 𝑁 is the number of pairs of integers (𝑘, ℓ) with 1 ≤ 𝑘 < ℓ ≤ 𝑚 such

that 𝑘 appears after ℓ in the list (𝑗1 , …, 𝑗𝑚 ).
Hence the sign of a permutation equals 1 if the natural order has been changed
an even number of times and equals −1 if the natural order has been changed an
odd number of times.
9.33 example: signs

• The permutation (1, …, 𝑚) [no changes in the natural order] has sign 1.
• The only pair of integers (𝑘, ℓ) with 𝑘 < ℓ such that 𝑘 appears after ℓ in the list
(2, 1, 3, 4) is (1, 2). Thus the permutation (2, 1, 3, 4) has sign −1.
• In the permutation (2, 3, …, 𝑚, 1), the only pairs (𝑘, ℓ) with 𝑘 < ℓ that appear
with changed order are (1, 2), (1, 3), …, (1, 𝑚). Because we have 𝑚 − 1 such
pairs, the sign of this permutation equals (−1)𝑚 − 1 .
9.34 swapping two entries in a permutation
Swapping two entries in a permutation multiplies the sign of the permutation

by −1.
Proof Suppose we have two permutations, where the second permutation is

obtained from the first by swapping two entries. The two swapped entries were
in their natural order in the first permutation if and only if they are not in their
natural order in the second permutation. Thus we have a net change (so far) of 1
or −1 (both odd numbers) in the number of pairs not in their natural order.
Consider each entry between the two swapped entries. If an intermediate entry
was originally in the natural order with respect to both swapped entries, then it
is now in the natural order with respect to neither swapped entry. Similarly, if
an intermediate entry was originally in the natural order with respect to neither
of the swapped entries, then it is now in the natural order with respect to both
swapped entries. If an intermediate entry was originally in the natural order with
respect to exactly one of the swapped entries, then that is still true. Thus the net
change (for each pair containing an entry between the two swapped entries) in the
number of pairs not in their natural order is 2, −2, or 0 (all even numbers).
For all other pairs of entries, there is no change in whether or not they are in
their natural order. Thus the total net change in the number of pairs not in their
natural order is an odd number. Hence the sign of the second permutation equals
−1 times the sign of the first permutation.
9.35 permutations and alternating multilinear forms
Suppose 𝑚 is a positive integer and 𝛼 ∈ 𝑉alt

(𝑚)
. Then
𝛼(𝑣𝑗1 , …, 𝑣𝑗𝑚 ) = (sign(𝑗1 , …, 𝑗𝑚 ))𝛼(𝑣1 , …, 𝑣𝑚 )
for every list 𝑣1 , …, 𝑣𝑚 of vectors in 𝑉 and all (𝑗1 , …, 𝑗𝑚 ) ∈ perm 𝑚.
Proof Suppose 𝑣1 , …, 𝑣𝑚 ∈ 𝑉 and (𝑗1 , …, 𝑗𝑚 ) ∈ perm 𝑚. We can get from

(𝑗1 , …, 𝑗𝑚 ) to (1, …, 𝑚) by a series of swaps of entries in different slots. Each such
swap changes the value of 𝛼 by a factor of −1 (by 9.30) and also changes the sign
of the remaining permutation by a factor of −1 (by 9.34). After an appropriate
number of swaps, we reach the permutation 1, …, 𝑚, which has sign 1. Thus the
value of 𝛼 changed signs an even number of times if sign(𝑗1 , …, 𝑗𝑚 ) = 1 and an
odd number of times if sign(𝑗1 , …, 𝑗𝑚 ) = −1, which gives the desired result.
Our use of permutations now leads in a natural way to the following beautiful
formula for alternating 𝑛-linear forms on an 𝑛-dimensional vector space.
9.36 formula for (dim 𝑉)-linear alternating forms on 𝑉
Let 𝑛 = dim 𝑉. Suppose 𝑒1 , …, 𝑒𝑛 is a basis of 𝑉 and 𝑣1 , …, 𝑣𝑛 ∈ 𝑉. For each

𝑘 ∈ {1, …, 𝑛}, let 𝑏1, 𝑘 , …, 𝑏𝑛, 𝑘 ∈ 𝐅 be such that
𝑛
𝑣𝑘 = ∑ 𝑏𝑗, 𝑘 𝑒𝑗 .
𝑗=1
Then
𝛼(𝑣1 , …, 𝑣𝑛 ) = 𝛼(𝑒1 , …, 𝑒𝑛 ) ∑ (sign(𝑗1 , …, 𝑗𝑛 ))𝑏𝑗1, 1 ⋯𝑏𝑗𝑛, 𝑛

(𝑗1 , …, 𝑗𝑛 ) ∈ perm 𝑛
for every alternating 𝑛-linear form 𝛼 on 𝑉.
Proof Suppose 𝛼 is an alternating 𝑛-linear form 𝛼 on 𝑉. Then

𝑛 𝑛
𝛼(𝑣1 , …, 𝑣𝑛 ) = 𝛼( ∑ 𝑏𝑗1, 1 𝑒𝑗1 , …, ∑ 𝑏𝑗𝑛, 𝑛 𝑒𝑗𝑛 )
𝑗1 = 1 𝑗𝑛 = 1
𝑛 𝑛
= ∑ ⋯ ∑ 𝑏𝑗1, 1 ⋯𝑏𝑗𝑛, 𝑛 𝛼(𝑒𝑗1 , …, 𝑒𝑗𝑛 )
𝑗1 = 1 𝑗𝑛 = 1
= ∑ 𝑏𝑗1, 1 ⋯𝑏𝑗𝑛, 𝑛 𝛼(𝑒𝑗1 , …, 𝑒𝑗𝑛 )

(𝑗1 , …, 𝑗𝑛 ) ∈ perm 𝑛
= 𝛼(𝑒1 , …, 𝑒𝑛 ) ∑ (sign(𝑗1 , …, 𝑗𝑛 ))𝑏𝑗1, 1 ⋯𝑏𝑗𝑛, 𝑛 ,

(𝑗1 , …, 𝑗𝑛 ) ∈ perm 𝑛
where the third line holds because 𝛼(𝑒𝑗1 , …, 𝑒𝑗𝑛 ) = 0 if 𝑗1 , …, 𝑗𝑛 are not distinct
integers, and the last line holds by 9.35.
The following result will be the key to our definition of the determinant in the
next section.
(dim 𝑉)
9.37 dim 𝑉alt =1
(dim 𝑉)
The vector space 𝑉alt has dimension one.
Proof Let 𝑛 = dim 𝑉. Suppose 𝛼 and 𝛼′ are alternating 𝑛-linear forms on 𝑉 with
𝛼 ≠ 0. Let 𝑒1 , …, 𝑒𝑛 be such that 𝛼(𝑒1 , …, 𝑒𝑛 ) ≠ 0. There exists 𝑐 ∈ 𝐅 such that
𝛼′ (𝑒1 , …, 𝑒𝑛 ) = 𝑐𝛼(𝑒1 , …, 𝑒𝑛 ).
Furthermore, 9.28 implies that 𝑒1 , …, 𝑒𝑛 is linearly independent and thus is a basis
of 𝑉.
Suppose 𝑣1 , …, 𝑣𝑛 ∈ 𝑉. Let 𝑏𝑗, 𝑘 be as in 9.36 for 𝑗, 𝑘 = 1, …, 𝑛. Then
𝛼′ (𝑣1 , …, 𝑣𝑛 ) = 𝛼′ (𝑒1 , …, 𝑒𝑛 ) ∑ (sign(𝑗1 , …, 𝑗𝑛 ))𝑏𝑗1, 1 ⋯𝑏𝑗𝑛, 𝑛
(𝑗1 , …, 𝑗𝑛 ) ∈ perm 𝑛
= 𝑐𝛼(𝑒1 , …, 𝑒𝑛 ) ∑ (sign(𝑗1 , …, 𝑗𝑛 ))𝑏𝑗1, 1 ⋯𝑏𝑗𝑛, 𝑛

(𝑗1 , …, 𝑗𝑛 ) ∈ perm 𝑛
= 𝑐𝛼(𝑣1 , …, 𝑣𝑛 ),
where the first and last lines above come from 9.36. The equation above implies
that 𝛼′ = 𝑐𝛼. Thus 𝛼′, 𝛼 is not a linearly independent list, which implies that
dim 𝑉alt
(𝑛)
≤ 1.
To complete the proof, we only need to show that there exists a nonzero
alternating 𝑛-linear form 𝛼 on 𝑉 (thus eliminating the possibility that dim 𝑉alt (𝑛)
equals 0). To do this, let 𝑒1 , …, 𝑒𝑛 be any basis of 𝑉, and let 𝜑1 , …, 𝜑𝑛 ∈ 𝑉 be

′
the linear functionals on 𝑉 that allow us to express each element of 𝑉 as a linear

combination of 𝑒1 , …, 𝑒𝑛 . In other words,
𝑛
𝑣 = ∑ 𝜑𝑗 (𝑣)𝑒𝑗
𝑗=1
for every 𝑣 ∈ 𝑉 (see 3.114). Now for 𝑣1 , …, 𝑣𝑛 ∈ 𝑉, define

9.38 𝛼(𝑣1 , …, 𝑣𝑛 ) = ∑ (sign(𝑗1 , …, 𝑗𝑛 ))𝜑𝑗1 (𝑣1 )⋯𝜑𝑗𝑛 (𝑣𝑛 ).
(𝑗1 , …, 𝑗𝑛 ) ∈ perm 𝑛
The verification that 𝛼 is an 𝑛-linear form on 𝑉 is straightforward.

To see that 𝛼 is alternating, suppose 𝑣1 , …, 𝑣𝑛 ∈ 𝑉 with 𝑣1 = 𝑣2 . For each
(𝑗1 , …, 𝑗𝑛 ) ∈ perm 𝑛, the permutation (𝑗2 , 𝑗1 , 𝑗3 , …, 𝑗𝑛 ) has the opposite sign. Be-
cause 𝑣1 = 𝑣2 , the contributions from these two permutations to the sum in 9.38
cancel either other. Hence 𝛼(𝑣1 , 𝑣1 , 𝑣3 , …, 𝑣𝑛 ) = 0. Similarly, 𝛼(𝑣1 , …, 𝑣𝑛 ) = 0 if
any two vectors in the list 𝑣1 , …, 𝑣𝑛 are equal. Thus 𝛼 is alternating.
Finally, consider 9.38 with each 𝑣𝑘 = 𝑒𝑘 . Because 𝜑𝑗 (𝑒𝑘 ) equals 0 if 𝑗 ≠ 𝑘 and
equals 1 if 𝑗 = 𝑘, only the permutation (1, …, 𝑛) makes a nonzero contribution to
the right side of 9.38 in this case, giving the equation 𝛼(𝑒1 , …, 𝑒𝑛 ) = 1. Thus we
have produced a nonzero alternating 𝑛-linear form 𝛼 on 𝑉, as desired.
Earlier we showed that the value of The formula 9.38 used in the last proof
an alternating multilinear form applied to construct a nonzero alternating 𝑛-
to a linearly dependent list is 0; see 9.28. linear form came from the formula in
The next result provides a converse of 9.36, and that formula arose naturally
9.28 for 𝑛-linear multilinear forms when from the properties of an alternating
𝑛 = dim 𝑉. In the following result, the multilinear form.
statement that 𝛼 is nonzero means (as
usual for a function) that 𝛼 is not the function on 𝑉 𝑛 that is identically 0.
9.39 alternating (dim 𝑉)-linear forms and linear independence
Let 𝑛 = dim 𝑉. Suppose 𝛼 is a nonzero alternating 𝑛-linear form on 𝑉 and

𝑒1 , …, 𝑒𝑛 is a list of vectors in 𝑉. Then
𝛼(𝑒1 , …, 𝑒𝑛 ) ≠ 0
if and only if 𝑒1 , …, 𝑒𝑛 is linearly independent.
Proof First suppose 𝛼(𝑒1 , …, 𝑒𝑛 ) ≠ 0. Then 9.28 implies that 𝑒1 , …, 𝑒𝑛 is linearly

independent.
To prove the implication in the other direction, now suppose 𝑒1 , …, 𝑒𝑛 is linearly
independent. Because 𝑛 = dim 𝑉, this implies that 𝑒1 , …, 𝑒𝑛 is a basis of 𝑉 (see
2.38).
Because 𝛼 is not the zero 𝑛-linear form, there exist 𝑣1 , …, 𝑣𝑛 ∈ 𝑉 such that
𝛼(𝑣1 , …, 𝑣𝑛 ) ≠ 0. Now 9.36 implies that 𝛼(𝑒1 , …, 𝑒𝑛 ) ≠ 0.
Exercises 9B
1 Suppose 𝑚 is a positive integer. Show that dim 𝑉 (𝑚) = (dim 𝑉)𝑚.

2 Suppose 𝑛 ≥ 3 and 𝛼 ∶ 𝐅𝑛 × 𝐅𝑛 × 𝐅𝑛 → 𝐅 is defined by
𝛼((𝑥1 , …, 𝑥𝑛 ), (𝑦1 , …, 𝑦𝑛 ), (𝑧1 , …, 𝑧𝑛 ))

= 𝑥1 𝑦2 𝑧3 − 𝑥2 𝑦1 𝑧3 − 𝑥3 𝑦2 𝑧1 − 𝑥1 𝑦3 𝑧2 + 𝑥3 𝑦1 𝑧2 + 𝑥2 𝑦3 𝑧1 .
Show that 𝛼 is an alternating 3-linear form on 𝐅𝑛.

3 Suppose 𝑚 is a positive integer and 𝛼 is an 𝑚-linear form on 𝑉 such that
𝛼(𝑣1 , …, 𝑣𝑚 ) = 0 whenever 𝑣1 , …, 𝑣𝑚 is a list of vectors in 𝑉 with 𝑣𝑗 = 𝑣𝑗 + 1
for some 𝑗 ∈ {1, …, 𝑚 − 1}. Prove that 𝛼 is an alternating 𝑚-linear form on 𝑉.
4 Prove or give a counterexample: If 𝛼 ∈ 𝑉alt
(4)
, then
{(𝑣1 , 𝑣2 , 𝑣3 , 𝑣4 ) ∈ 𝑉 4 ∶ 𝛼(𝑣1 , 𝑣2 , 𝑣3 , 𝑣4 ) = 0}
is a subspace of 𝑉 4.
5 Suppose 𝑚 is a positive integer and 𝛽 is an 𝑚-linear form on 𝑉. Define an

𝑚-linear form 𝛼 on 𝑉 by
𝛼(𝑣1 , …, 𝑣𝑚 ) = ∑ (sign(𝑗1 , …, 𝑗𝑚 ))𝛽(𝑣𝑗1 , …, 𝑣𝑗𝑚 )

(𝑗1 , …, 𝑗𝑚 ) ∈ perm 𝑚
for 𝑣1 , …, 𝑣𝑚 ∈ 𝑉. Explain why 𝛼 ∈ 𝑉alt

(𝑚)
.
6 Suppose 𝑚 is a positive integer and 𝛽 is an 𝑚-linear form on 𝑉. Define an
𝑚-linear form 𝛼 on 𝑉 by
𝛼(𝑣1 , …, 𝑣𝑚 ) = ∑ 𝛽(𝑣𝑗1 , …, 𝑣𝑗𝑚 )

(𝑗1 , …, 𝑗𝑚 ) ∈ perm 𝑚
for 𝑣1 , …, 𝑣𝑚 ∈ 𝑉. Explain why
𝛼(𝑣𝑘1 , …, 𝑣𝑘𝑚 ) = 𝛼(𝑣1 , …, 𝑣𝑚 )
for all 𝑣1 , …, 𝑣𝑚 ∈ 𝑉 and all (𝑘1 , …, 𝑘𝑚 ) ∈ perm 𝑚.

7 Give an example of a nonzero alternating 2-linear form 𝛼 on 𝐑3 and a linearly
independent list 𝑣1 , 𝑣2 in 𝐑3 such that 𝛼(𝑣1 , 𝑣2 ) = 0.
This exercise shows that 9.39 can fail if the hypothesis that 𝑛 = dim 𝑉 is
deleted.
9C Determinants
Defining the Determinant
The next definition will lead us to a clean, beautiful, basis-free definition of the
determinant of an operator.
9.40 definition: 𝛼𝑇
Suppose that 𝑚 is a positive integer and 𝑇 ∈ ℒ(𝑉). For 𝛼 ∈ 𝑉alt

(𝑚)
, define
𝛼𝑇 ∈ 𝑉alt by
(𝑚)
𝛼𝑇 (𝑣1 , …, 𝑣𝑚 ) = 𝛼(𝑇𝑣1 , …, 𝑇𝑣𝑚 )

for each list 𝑣1 , …, 𝑣𝑚 of vectors in 𝑉.
Suppose 𝑇 ∈ ℒ(𝑉). If 𝛼 ∈ 𝑉alt (𝑚)

and 𝑣1 , …, 𝑣𝑚 is a list of vectors in 𝑉 with
𝑣𝑗 = 𝑣𝑘 for some 𝑗 ≠ 𝑘, then 𝑇𝑣𝑗 = 𝑇𝑣𝑘 , which implies that 𝛼𝑇 (𝑣1 , …, 𝑣𝑚 ) =
𝛼(𝑇𝑣1 , …, 𝑇𝑣𝑚 ) = 0. Thus the function 𝛼 ↦ 𝛼𝑇 is a linear map of 𝑉alt (𝑚)
to itself.
(dim 𝑉)
We know that dim 𝑉alt = 1 (see 9.37). Every linear map from a one-
dimensional vector space to itself is multiplication by some unique scalar. For
the linear map 𝛼 ↦ 𝛼𝑇 , we now define det 𝑇 to be that scalar.
9.41 definition: determinant of an operator, det 𝑇
Suppose 𝑇 ∈ ℒ(𝑉). The determinant of 𝑇, denoted by det 𝑇, is defined to be

the unique number in 𝐅 such that
𝛼𝑇 = (det 𝑇) 𝛼
(dim 𝑉)
for all 𝛼 ∈ 𝑉alt .
9.42 example: determinants of operators

Let 𝑛 = dim 𝑉.
• If 𝐼 is the identity operator on 𝑉, then 𝛼𝐼 = 𝛼 for all 𝛼 ∈ 𝑉alt
(𝑛)
. Thus det 𝐼 = 1.
• More generally, if 𝜆 ∈ 𝐅, then 𝛼 𝜆𝐼 = 𝜆𝑛 𝛼 for all 𝛼 ∈ 𝑉alt
(𝑛)
. Thus det(𝜆𝐼) = 𝜆𝑛.
• Still more generally, if 𝑇 ∈ ℒ(𝑉) and 𝜆 ∈ 𝐅, then 𝛼 𝜆𝑇 = 𝜆𝑛 𝛼𝑇 = 𝜆𝑛 (det 𝑇)𝛼
for all 𝛼 ∈ 𝑉alt
(𝑛)
. Thus det(𝜆𝑇) = 𝜆𝑛 det 𝑇.
• Suppose 𝑇 ∈ ℒ(𝑉) and there is a basis 𝑒1 , …, 𝑒𝑛 of 𝑉 consisting of eigenvectors
of 𝑇, with corresponding eigenvalues 𝜆1 , …, 𝜆𝑛 . If 𝛼 ∈ 𝑉alt
(𝑛)
, then
𝛼𝑇 (𝑒1 , …, 𝑒𝑛 ) = 𝛼(𝜆1 𝑒1 , …, 𝜆𝑛 𝑒𝑛 ) = (𝜆1 ⋯𝜆𝑛 )𝛼(𝑒1 , …, 𝑒𝑛 ).
If 𝛼 ≠ 0, then 9.39 implies 𝛼(𝑒1 , …, 𝑒𝑛 ) ≠ 0. Thus the equation above implies
det 𝑇 = 𝜆1 ⋯𝜆𝑛 .
Section 9C Determinants 355
Our next task is to define and give a formula for the determinant of a square
matrix. To do this, we associate with each square matrix an operator and then
define the determinant of the matrix to be the determinant of the associated
operator.
9.43 definition: determinant of a matrix, det 𝐴
Suppose that 𝑛 is a positive integer and 𝐴 is an 𝑛-by-𝑛 square matrix with

entries in 𝐅. Let 𝑇 ∈ ℒ(𝐅𝑛 ) be the operator whose matrix with respect to
the standard basis of 𝐅𝑛 equals 𝐴. The determinant of 𝐴, denoted by det 𝐴, is
defined by det 𝐴 = det 𝑇.
9.44 example: determinants of matrices
• If 𝐼 is the 𝑛-by-𝑛 identity matrix, then the corresponding operator on 𝐅𝑛 is the

identity operator 𝐼 on 𝐅𝑛. Thus the first bullet point of 9.42 implies that the
determinant of the identity matrix is 1.
• Suppose 𝐴 is a diagonal matrix with 𝜆1 , …, 𝜆𝑛 on the diagonal. Then the
corresponding operator on 𝐅𝑛 has the standard basis of 𝐅𝑛 as eigenvectors,
with eigenvalues 𝜆1 , …, 𝜆𝑛 . Thus the last bullet point of 9.42 implies that
det 𝐴 = 𝜆1 ⋯𝜆𝑛 .
For the next result, think of each list 𝑣1 , …, 𝑣𝑛 of 𝑛 vectors in 𝐅𝑛 as a list of
𝑛-by-1 column vectors. The notation ( 𝑣1 ⋯ 𝑣𝑛 ) then denotes the 𝑛-by-𝑛
square matrix whose 𝑘 th column is 𝑣𝑘 for each 𝑘 = 1, …, 𝑛.
9.45 determinant is an alternating multilinear form
Suppose that 𝑛 is a positive integer. The map that takes a list 𝑣1 , …, 𝑣𝑛 of

vectors in 𝐅𝑛 to det ( 𝑣1 ⋯ 𝑣𝑛 ) is an alternating 𝑛-linear form on 𝐅𝑛.
Proof Let 𝑒1 , …, 𝑒𝑛 be the standard basis of 𝐅𝑛 and suppose 𝑣1 , …, 𝑣𝑛 is a list of

vectors in 𝐅𝑛. Let 𝑇 ∈ ℒ(𝐅𝑛 ) be the operator such that 𝑇𝑒𝑘 = 𝑣𝑘 for 𝑘 = 1, …, 𝑛.
Thus 𝑇 is the operator whose matrix with respect to 𝑒1 , …, 𝑒𝑛 is ( 𝑣1 ⋯ 𝑣𝑛 ).
Hence det ( 𝑣1 ⋯ 𝑣𝑛 ) = det 𝑇, by definition of the determinant of a matrix.
Let 𝛼 be an alternating 𝑛-linear form on 𝐅𝑛 such that 𝛼(𝑒1 , …, 𝑒𝑛 ) = 1. Then
det ( 𝑣1 ⋯ 𝑣𝑛 ) = det 𝑇
= (det 𝑇) 𝛼(𝑒1 , …, 𝑒𝑛 )
= 𝛼(𝑇𝑒1 , …, 𝑇𝑒𝑛 )
= 𝛼(𝑣1 , …, 𝑣𝑛 ),
where the third line follows from the definition of the determinant of an operator.
The equation above shows that the map that takes a list of vectors 𝑣1 , …, 𝑣𝑛 in 𝐅𝑛
to det ( 𝑣1 ⋯ 𝑣𝑛 ) is the alternating 𝑛-linear form 𝛼 on 𝐅𝑛.
The previous result has several important consequences. For example, it

immediately implies that a matrix with two identical columns has determinant 0.
We will come back to other consequences later, but for now we want to give a
formula for the determinant of a square matrix. Recall that if 𝐴 is a matrix, then
𝐴𝑗, 𝑘 denotes the entry in row 𝑗, column 𝑘 of 𝐴.
9.46 formula for determinant of a matrix
Suppose that 𝑛 is a positive integer and 𝐴 is an 𝑛-by-𝑛 square matrix. Then
det 𝐴 = ∑ (sign(𝑗1 , …, 𝑗𝑛 ))𝐴𝑗1, 1 ⋯𝐴𝑗𝑛, 𝑛 .

(𝑗1 , …, 𝑗𝑛 ) ∈ perm 𝑛
Proof Apply 9.36 with 𝑉 = 𝐅𝑛 and 𝑒1 , …, 𝑒𝑛 the standard basis of 𝐅𝑛 and 𝛼 the
alternating 𝑛-linear form on 𝐅𝑛 that takes 𝑣1 , …, 𝑣𝑛 to det ( 𝑣1 ⋯ 𝑣𝑛 ) [see
9.45]. If each 𝑣𝑘 is the 𝑘 th column of 𝐴, then each 𝑏𝑗, 𝑘 in 9.36 equals 𝐴𝑗, 𝑘 . Finally,
𝛼(𝑒1 , …, 𝑒𝑛 ) = det ( 𝑒1 ⋯ 𝑒𝑛 ) = det 𝐼 = 1.
Thus the formula in 9.36 becomes the formula stated in this result.
9.47 example: explicit formula for determinant

• If 𝐴 is a 2-by-2 matrix, then the formula in 9.46 becomes
det 𝐴 = 𝐴1, 1 𝐴2, 2 − 𝐴2, 1 𝐴1, 2 .
• If 𝐴 is a 3-by-3 matrix, then the formula in 9.46 becomes

det 𝐴 =𝐴1, 1 𝐴2, 2 𝐴3, 3 − 𝐴2, 1 𝐴1, 2 𝐴3, 3 − 𝐴3, 1 𝐴2, 2 𝐴1, 3
− 𝐴1, 1 𝐴3, 2 𝐴2, 3 + 𝐴3, 1 𝐴1, 2 𝐴2, 3 + 𝐴2, 1 𝐴3, 2 𝐴1, 3 .
The sum in the formula in 9.46 contains 𝑛! terms. Because 𝑛! grows rapidly as
𝑛 increases, the formula in 9.46 is not a viable method to evaluate determinants
even for moderately sized 𝑛. For example, 10! is over three million, and 100! is
approximately 10158 , leading to a sum that the fastest computer cannot evaluate.
We will soon see some results that lead to faster evaluations of determinants than
direct use of the sum in 9.46.
9.48 determinant of upper-triangular matrix
Suppose that 𝐴 is an upper-triangular matrix with 𝜆1 , …, 𝜆𝑛 on the diagonal.

Then det 𝐴 = 𝜆1 ⋯𝜆𝑛 .
Proof If (𝑗1 , …, 𝑗𝑛 ) ∈ perm 𝑛 with (𝑗1 , …, 𝑗𝑛 ) ≠ (1, …, 𝑛), then 𝑗𝑘 > 𝑘 for some
𝑘 ∈ {1, …, 𝑛}, which implies that 𝐴𝑗𝑘, 𝑘 = 0. Thus the only permutation that
can make a nonzero contribution to the sum in 9.46 is the permutation (1, …, 𝑛).
Because 𝐴𝑘, 𝑘 = 𝜆𝑘 for each 𝑘 = 1, …, 𝑛, this implies that det 𝐴 = 𝜆1 ⋯𝜆𝑛 .
Properties of Determinants
Our definition of the determinant leads to the following magical proof that the
determinant is multiplicative.
9.49 determinant is multiplicative
(a) Suppose 𝑆, 𝑇 ∈ ℒ(𝑉). Then det(𝑆𝑇) = (det 𝑆)(det 𝑇).

(b) Suppose 𝐴 and 𝐵 are square matrices of the same size. Then
det(𝐴𝐵) = (det 𝐴)(det 𝐵)
Proof
(a) Let 𝑛 = dim 𝑉. Suppose 𝛼 ∈ 𝑉alt
(𝑛)
and 𝑣1 , …, 𝑣𝑛 ∈ 𝑉. Then
𝛼𝑆𝑇 (𝑣1 , …, 𝑣𝑛 ) = 𝛼(𝑆𝑇𝑣1 , …, 𝑆𝑇𝑣𝑛 )
= (det 𝑆)𝛼(𝑇𝑣1 , …, 𝑇𝑣𝑛 )
= (det 𝑆)(det 𝑇)𝛼(𝑣1 , …, 𝑣𝑛 ),
where the first equation follows from the definition of 𝛼𝑆𝑇 , the second equation
follows from the definition of det 𝑆, and the third equation follows from the
definition of det 𝑇. The equation above implies that det(𝑆𝑇) = (det 𝑆)(det 𝑇).
(b) Let 𝑆, 𝑇 ∈ ℒ(𝐅𝑛 ) be such that ℳ(𝑆) = 𝐴 and ℳ(𝑇) = 𝐵, where all matrices
of operators in this proof are with respect to the standard basis of 𝐅𝑛. Then
ℳ(𝑆𝑇) = ℳ(𝑆)ℳ(𝑇) = 𝐴𝐵 (see 3.43). Thus
det(𝐴𝐵) = det(𝑆𝑇) = (det 𝑆)(det 𝑇) = (det 𝐴)(det 𝐵),
where the second equality comes from the result in (a).
The determinant of an operator determines whether the operator is invertible.
9.50 invertible ⟺ nonzero determinant

An operator 𝑇 ∈ ℒ(𝑉) is invertible if and only if det 𝑇 ≠ 0. Furthermore, if
𝑇 is invertible, then det(𝑇 −1 ) = det1 𝑇 .
Proof First suppose 𝑇 is invertible. Thus 𝑇𝑇 −1 = 𝐼. Now 9.49 implies that

1 = det 𝐼 = det(𝑇𝑇 −1 ) = (det 𝑇)(det(𝑇 −1 )).
Hence det 𝑇 ≠ 0 and det(𝑇 −1 ) is the multiplicative inverse of det 𝑇.
To prove the other direction, now suppose det 𝑇 ≠ 0. Suppose 𝑣 ∈ 𝑉 and
𝑣 ≠ 0. Let 𝑣, 𝑒2 , …, 𝑒𝑛 be a basis of 𝑉 and let 𝛼 ∈ 𝑉alt
(𝑛)
be such that 𝛼 ≠ 0. Then
𝛼(𝑣, 𝑒2 , …, 𝑒𝑛 ) ≠ 0 (by 9.39). Now
𝛼(𝑇𝑣, 𝑇𝑒2 , …, 𝑇𝑒𝑛 ) = (det 𝑇)𝛼(𝑣, 𝑒2 , …, 𝑒𝑛 ) ≠ 0,
Thus 𝑇𝑣 ≠ 0. Hence 𝑇 is invertible.
An 𝑛-by-𝑛 matrix 𝐴 is invertible (see 3.80 for the definition of an invertible

matrix) if and only if the operator on 𝐅𝑛 associated with 𝐴 (via the standard basis
of 𝐅𝑛 ) is invertible. Thus the previous result shows that a square matrix 𝐴 is
invertible if and only if det 𝐴 ≠ 0.
9.51 eigenvalues and determinants
Suppose 𝑇 ∈ ℒ(𝑉) and 𝜆 ∈ 𝐅. Then 𝜆 is an eigenvalue of 𝑇 if and only if

det(𝜆𝐼 − 𝑇) = 0.
Proof The number 𝜆 is an eigenvalue of 𝑇 if and only if 𝑇 − 𝜆𝐼 is not invertible

(see 5.7), which happens if and only if 𝜆𝐼 − 𝑇 is not invertible, which happens if
and only if det(𝜆𝐼 − 𝑇) = 0 (by 9.50).
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑆 ∶ 𝑊 → 𝑉 is an invertible linear map. To prove that

det(𝑆−1 𝑇𝑆) = det 𝑇, we could try to use 9.49 and 9.50, writing
det(𝑆−1 𝑇𝑆) = (det 𝑆−1 )(det 𝑇)(det 𝑆)

= det 𝑇.
That proof works if 𝑊 = 𝑉, but if 𝑊 ≠ 𝑉 then it makes no sense because the

determinant is defined only for linear maps from a vector space to itself, and 𝑆
maps 𝑊 to 𝑉, making det 𝑆 undefined. The proof given below works around this
issue and is valid when 𝑊 ≠ 𝑉.
9.52 determinant is a similarity invariant
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑆 ∶ 𝑊 → 𝑉 is an invertible linear map. Then
det(𝑆−1 𝑇𝑆) = det 𝑇.
Proof Let 𝑛 = dim 𝑊 = dim 𝑉. Suppose 𝜏 ∈ 𝑊alt

(𝑛)
. Define 𝛼 ∈ 𝑉alt
(𝑛)
by
𝛼(𝑣1 , …, 𝑣𝑛 ) = 𝜏(𝑆−1 𝑣1 , …, 𝑆−1 𝑣𝑛 )
for 𝑣1 , …, 𝑣𝑛 ∈ 𝑉. Suppose 𝑤1 , …, 𝑤𝑛 ∈ 𝑊. Then
𝜏𝑆−1𝑇𝑆 (𝑤1 , …, 𝑤𝑛 ) = 𝜏(𝑆−1 𝑇𝑆𝑤1 , …, 𝑆−1 𝑇𝑆𝑤𝑛 )
= 𝛼(𝑇𝑆𝑤1 , …, 𝑇𝑆𝑤𝑛 )
= 𝛼𝑇 (𝑆𝑤1 , …, 𝑆𝑤𝑛 )
= (det 𝑇)𝛼(𝑆𝑤1 , …, 𝑆𝑤𝑛 )
= (det 𝑇)𝜏(𝑤1 , …, 𝑤𝑛 ).
The equation above and the definition of the determinant of the operator 𝑆−1 𝑇𝑆
imply that det(𝑆−1 𝑇𝑆) = det 𝑇.
For the special case in which 𝑉 = 𝐅𝑛 and 𝑒1 , …, 𝑒𝑛 is the standard basis of 𝐅𝑛,
the next result is true by the definition of the determinant of a matrix. The left
side of the equation in the next result does not depend on a choice of basis, which
means that the right side is independent of the choice of basis.
9.53 determinant of operator equals determinant of its matrix
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑒1 , …, 𝑒𝑛 is a basis of 𝑉. Then
det 𝑇 = det ℳ(𝑇, (𝑒1 , …, 𝑒𝑛 )).
Proof Let 𝑓1 , …, 𝑓𝑛 be the standard basis of 𝐅𝑛 . Let 𝑆 ∶ 𝐅𝑛 → 𝑉 be the linear

map such that 𝑆 𝑓𝑘 = 𝑒𝑘 for each 𝑘 = 1, …, 𝑛. Thus ℳ(𝑆, ( 𝑓1 , …, 𝑓𝑛 ), (𝑒1 , …, 𝑒𝑛 ))
and ℳ(𝑆−1 , (𝑒1 , …, 𝑒𝑛 ), ( 𝑓1 , …, 𝑓𝑛 )) both equal the 𝑛-by-𝑛 identity matrix. Hence
9.54 ℳ(𝑆−1 𝑇𝑆, ( 𝑓1 , …, 𝑓𝑛 )) = ℳ(𝑇, (𝑒1 , …, 𝑒𝑛 )),
as follows from two applications of 3.43. Thus

det 𝑇 = det(𝑆−1 𝑇𝑆)
= det ℳ(𝑆−1 𝑇𝑆, ( 𝑓1 , …, 𝑓𝑛 ))
= det ℳ(𝑇, (𝑒1 , …, 𝑒𝑛 )),
where the first line comes from 9.52, the second line comes from the definition of
the determinant of a matrix, and the third line follows from 9.54.
The next result gives a more intuitive way to think about determinants than the
definition or the formula in 9.46. We could make the characterization in the result
below the definition of the determinant of an operator on a finite-dimensional
complex vector space, with the current definition then becoming a consequence
of that definition.
9.55 if 𝐅 = 𝐂 , then determinant equals product of eigenvalues
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Then det 𝑇 equals the product of the eigen-
values of 𝑇, with each eigenvalue included as many times as its multiplicity.
Proof There is a basis of 𝑉 with respect to which 𝑇 has an upper-triangular

matrix with the diagonal entries of the matrix consisting of the eigenvalues of 𝑇,
with each eigenvalue included as many times as its multiplicity—see 8.37. Thus
9.53 and 9.48 imply that det 𝑇 equals the product of the eigenvalues of 𝑇, with
each eigenvalue included as many times as its multiplicity.
As the next result shows, the determinant interacts nicely with the transpose of
a square matrix, with the dual of an operator, and with the adjoint of an operator
on an inner product space.
9.56 determinant of transpose, dual, or adjoint
(a) Suppose 𝐴 is a square matrix. Then det 𝐴t = det 𝐴.

(b) Suppose 𝑇 ∈ ℒ(𝑉). Then det 𝑇 ′ = det 𝑇.
(c) Suppose 𝑉 is an inner product space and 𝑇 ∈ ℒ(𝑉). Then
det(𝑇 ∗ ) = det 𝑇.
Proof
(a) Let 𝑛 be a positive integer. Define 𝛼 ∶ (𝐅𝑛 )𝑛 → 𝐅 by
t
𝛼(( 𝑣1 ⋯ 𝑣𝑛 )) = det(( 𝑣1 ⋯ 𝑣𝑛 ) )
for all 𝑣1 , …, 𝑣𝑛 ∈ 𝐅𝑛. The formula in 9.46 for the determinant of a matrix
shows that 𝛼 is an 𝑛-linear form on 𝐅𝑛.
Suppose 𝑣1 , …, 𝑣𝑛 ∈ 𝐅𝑛 and 𝑣𝑗 = 𝑣𝑘 for some 𝑗 ≠ 𝑘. If 𝐵 is an 𝑛-by-𝑛 matrix,
t
then ( 𝑣1 ⋯ 𝑣𝑛 ) 𝐵 cannot equal the identity matrix because row 𝑗 and
t t
row 𝑘 of ( 𝑣1 ⋯ 𝑣𝑛 ) 𝐵 are equal. Thus ( 𝑣1 ⋯ 𝑣𝑛 ) is not invertible,
which implies that 𝛼(( 𝑣1 ⋯ 𝑣𝑛 )) = 0. Hence 𝛼 is an alternating 𝑛-
linear form on 𝐅𝑛.
Note that 𝛼 applied to the standard basis of 𝐅𝑛 equals 1. Because the vector
space of alternating 𝑛-linear forms on 𝐅𝑛 has dimension one (by 9.37), this
implies that 𝛼 is the determinant function. Thus (a) holds.
(b) The equation det 𝑇 ′ = det 𝑇 follows from (a) and 9.53 and 3.132.
(c) Pick an orthonormal basis of 𝑉. The matrix of 𝑇 ∗ with respect to that basis is
the conjugate transpose of the matrix of 𝑇 with respect to that basis (by 7.9).
Thus 9.53, 9.46, and (a) imply that det(𝑇 ∗ ) = det 𝑇.
9.57 helpful results in evaluating determinants
(a) If either two columns or two rows of a square matrix are equal, then the
determinant of the matrix equals 0.
(b) Suppose 𝐴 is a square matrix and 𝐵 is the matrix obtained from 𝐴 by
swapping either two columns or two rows. Then det 𝐴 = − det 𝐵.
(c) If one column or one row of a square matrix is multiplied by a scalar, then
the value of the determinant is multiplied by the same scalar.
(d) If a scalar multiple of one column of a square matrix to added to another
column, then the value of the determinant is unchanged.
(e) If a scalar multiple of one row of a square matrix to added to another row,
then the value of the determinant is unchanged.
Proof All the assertions in this result follow from the result that the maps
t
𝑣1 , …, 𝑣𝑛 ↦ det ( 𝑣1 ⋯ 𝑣𝑛 ) and 𝑣1 , …, 𝑣𝑛 ↦ det ( 𝑣1 ⋯ 𝑣𝑛 ) are both
alternating 𝑛-linear forms on 𝐅𝑛 [see 9.45 and 9.56(a)].
For example, to prove (d) suppose 𝑣1 , …, 𝑣𝑛 ∈ 𝐅𝑛 and 𝑐 ∈ 𝐅. Then
det( 𝑣1 + 𝑐𝑣2 𝑣2 ⋯ 𝑣𝑛 )
= det ( 𝑣1 𝑣2 ⋯ 𝑣𝑛 ) + 𝑐 det ( 𝑣2 𝑣2 𝑣3 ⋯ 𝑣𝑛 )
= det ( 𝑣1 𝑣2 ⋯ 𝑣 𝑛 ),
where the first equation follows from the multilinearity property and the second
equation follows from the alternating property. The equation above shows that
adding a multiple of the second column to the first column does not change the
value of the determinant. The same conclusion holds for any two columns. Thus
(d) holds.
The proof of (e) follows from (d) and from 9.56(a). The proofs of (a), (b), and
(c) use similar tools and are left to the reader.
For matrices whose entries are concrete numbers, the result above leads to a
much faster way to evaluate the determinant than direct application of the formula
in 9.46. Specifically, apply the Gaussian elimination procedure of swapping
rows [by 9.48(b), this changes the determinant by a factor of −1], multiplying
a row by a nonzero constant [by 9.48(c), this changes the determinant by the
same constant], and adding a multiple of one row to another row [by 9.48(e), this
does not change the determinant] to produce an upper-triangular matrix, whose
determinant is the product of the diagonal entries (by 9.48). If your software keeps
track of the number of row swaps and of the constants used when multiplying a
row by a constant, then the determinant of the original matrix can be computed.
Because a number 𝜆 ∈ 𝐅 is an eigenvalue of an operator 𝑇 ∈ ℒ(𝑉) if and
only if det(𝜆𝐼 − 𝑇) = 0 (by 9.51), you may be tempted to think that one way
to find eigenvalues quickly is to choose a basis of 𝑉, let 𝐴 = ℳ(𝑇), evaluate
det(𝜆𝐼 − 𝐴), and then solve the equation det(𝜆𝐼 − 𝐴) = 0 for 𝜆. However, that
procedure is rarely efficient, except when dim 𝑉 = 2 (or when dim 𝑉 equals 3 or
4 if you are willing to use the cubic or quartic formulas). One problem is that the
procedure described in the paragraph above for evaluating a determinant does not
work when the matrix includes a symbol (such as the 𝜆 in 𝜆𝐼 − 𝐴). This problem
arises because decisions need to be made in the Gaussian elimination procedure
about whether certain quantities equal 0, and those decisions become complicated
in expressions involving a symbol 𝜆.
Recall that an operator on a finite-dimensional inner product space is unitary
if it preserves norms (see 7.51 and the paragraph following it). Every eigenvalue
of a unitary operator has absolute value 1 (by 7.54). Thus the product of the
eigenvalues of a unitary operator has absolute value 1. Hence (at least in the case
𝐅 = 𝐂 ) the determinant of a unitary operator has absolute value 1 (by 9.55). The
next result gives a proof that works without the assumption that 𝐅 = 𝐂 .
9.58 every unitary operator has determinant with absolute value 1
Suppose 𝑉 is an inner product space and 𝑆 ∈ ℒ(𝑉) is a unitary operator.

Then |det 𝑆| = 1.
Proof Because 𝑆 is unitary, 𝐼 = 𝑆∗ 𝑆 (see 7.53). Thus

1 = det(𝑆∗ 𝑆) = (det 𝑆∗ )(det 𝑆) = (det 𝑆)(det 𝑆) = |det 𝑆|2,
where the second equality comes from 9.49(a) and the third equality comes from
9.56(c). The equation above implies that |det 𝑆| = 1.
The determinant of a positive operator on an inner product space meshes well

with the analogy that such operators correspond to the nonnegative real numbers.
9.59 every positive operator has nonnegative determinant
Suppose 𝑉 is an inner product space and 𝑇 ∈ ℒ(𝑉) is a positive operator.

Then det 𝑇 ≥ 0.
Proof By the spectral theorem (7.29 or 7.31), 𝑉 has an orthonormal basis con-
sisting of eigenvectors of 𝑇. Thus by the last bullet point of 9.42, det 𝑇 equals a
product of the eigenvalues of 𝑇, possibly with repetitions. Each eigenvalue of 𝑇 is
a nonnegative number (by 7.38). Thus we conclude that det 𝑇 ≥ 0.
Suppose 𝑉 is an inner product space and 𝑇 ∈ ℒ(𝑉). Recall that the list of
nonnegative square roots of the eigenvalues of 𝑇 ∗ 𝑇 (each included as many times
as its multiplicity) is called the list of singular values of 𝑇 (see Section 7E).
9.60 |det 𝑇| = product of singular values of 𝑇
Suppose 𝑉 is an inner product space and 𝑇 ∈ ℒ(𝑉). Then
|det 𝑇| = √det(𝑇 ∗ 𝑇) = product of singular values of 𝑇.
Proof We have
|det 𝑇|2 = (det 𝑇)(det 𝑇) = (det(𝑇 ∗ ))(det 𝑇) = det(𝑇 ∗ 𝑇),
where the middle equality comes from 9.56(c) and the last equality comes from
9.49(a). Taking square roots of both sides of the equation above shows that
|det 𝑇| = √det(𝑇 ∗ 𝑇).
Let 𝑠1 , …, 𝑠𝑛 denote the list of singular values of 𝑇. Thus 𝑠12, …, 𝑠𝑛2 is the
list of eigenvalues of 𝑇 ∗ 𝑇 (with appropriate repetitions), corresponding to an
orthonormal basis of 𝑉 consisting of eigenvectors of 𝑇 ∗ 𝑇. Hence the last bullet
point of 9.42 implies that
det(𝑇 ∗ 𝑇) = 𝑠12 ⋯𝑠𝑛2.
Thus |det 𝑇| = 𝑠1 ⋯𝑠𝑛 , as desired.
An operator 𝑇 on a real inner product space changes volume by a factor of the

product of the singular values (by 7.111). Thus the next result follows immediately
from 7.111 and 9.60. This result explains why the absolute value of a determinant
appears in the change of variables formula in multivariable calculus.
9.61 𝑇 changes volume by factor of |det 𝑇|
Suppose 𝑇 ∈ ℒ(𝐑𝑛 ) and Ω ⊆ 𝐑𝑛. Then
volume 𝑇(Ω) = |det 𝑇|(volume Ω).
For operators on finite-dimensional complex vector spaces, we now connect

the determinant to a polynomial that we have previously seen.
9.62 if 𝐅 = 𝐂 , then characteristic polynomial of 𝑇 equals det(𝑧𝐼 − 𝑇)
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). Let 𝜆1 , …, 𝜆𝑚 denote the distinct eigenvalues

of 𝑇, and let 𝑑1 , …, 𝑑𝑚 denote their multiplicities. Then
det(𝑧𝐼 − 𝑇) = (𝑧 − 𝜆1 )𝑑1 ⋯(𝑧 − 𝜆𝑚 )𝑑𝑚.
Proof There exists a basis of 𝑉 with respect to which 𝑇 has an upper-triangular

matrix with each 𝜆𝑘 appearing on the diagonal exactly 𝑑𝑘 times (by 8.37). With
respect to this basis, 𝑧𝐼 − 𝑇 has an upper-triangular matrix with 𝑧 − 𝜆𝑘 appearing
on the diagonal exactly 𝑑𝑘 times for each 𝑘. Thus 9.48 gives the desired equation.
Suppose 𝐅 = 𝐂 and 𝑇 ∈ ℒ(𝑉). The characteristic polynomial of 𝑇 was

defined in 8.26 as the polynomial on the right side of the equation in 9.62. We
did not previously define the characteristic polynomial of an operator on a finite-
dimensional real vector space because such operators may have no eigenvalues,
making a definition using the right side of the equation in 9.62 inappropriate.
We now present a new definition of the characteristic polynomial, motivated
by 9.62. This new definition is valid for both real and complex vector spaces.
The equation in 9.62 shows that this new definition is equivalent to our previous
definition when 𝐅 = 𝐂 (8.26).
9.63 definition: characteristic polynomial
Suppose 𝑇 ∈ ℒ(𝑉). The polynomial defined by
𝑧 ↦ det(𝑧𝐼 − 𝑇)
is called the characteristic polynomial of 𝑇.
The formula in 9.46 shows that the characteristic polynomial of an opera-

tor 𝑇 ∈ ℒ(𝑉) is a monic polynomial of degree dim 𝑉. The zeros in 𝐅 of the
characteristic polynomial of 𝑇 are exactly the eigenvalues of 𝑇 (by 9.51).
Previously we proved the Cayley–Hamilton theorem (8.29) in the complex

case. Now we can extend that result to operators on real vector spaces.
9.64 Cayley–Hamilton theorem
Suppose 𝑇 ∈ ℒ(𝑉) and 𝑞 is the characteristic polynomial of 𝑇. Then 𝑞(𝑇) = 0.
Proof If 𝐅 = 𝐂 , then the equation 𝑞(𝑇) = 0 follows from 9.62 and 8.29.
Now suppose 𝐅 = 𝐑 . Fix a basis of 𝑉, and let 𝐴 be the matrix of 𝑇 with
respect to this basis. Let 𝑆 be the operator on 𝐂dim 𝑉 such that the matrix of 𝑆
(with respect to the standard basis of 𝐂dim 𝑉 ) is 𝐴. For all 𝑧 ∈ 𝐑 we have
𝑞(𝑧) = det(𝑧𝐼 − 𝑇) = det(𝑧𝐼 − 𝐴) = det(𝑧𝐼 − 𝑆).
Thus 𝑞 is the characteristic polynomial of 𝑆. The case 𝐅 = 𝐂 (first sentence of
this proof) now implies that 0 = 𝑞(𝑆) = 𝑞(𝐴) = 𝑞(𝑇).
The Cayley–Hamilton theorem (9.64) implies that the characteristic polyno-

mial of an operator 𝑇 ∈ ℒ(𝑉) is a polynomial multiple of the minimal polynomial
of 𝑇 (by 5.29). Thus if the degree of the minimal polynomial of 𝑇 equals dim 𝑉,
then the characteristic polynomial of 𝑇 equals the minimal polynomial of 𝑇. This
happens for a very large percentage of operators, including over 99.999% of
4-by-4 matrices with integer entries in [−100, 100] (see the paragraph following
5.25).
The last sentence in our next result was previously proved in the complex case
(see 8.54). Now we can give a proof that works on both real and complex vector
spaces.
9.65 characteristic polynomial, trace, and determinant
Suppose 𝑇 ∈ ℒ(𝑉). Let 𝑛 = dim 𝑉. Then the characteristic polynomial of 𝑇

can be written as
𝑧𝑛 − (tr 𝑇)𝑧𝑛 − 1 + ⋯ + (−1)𝑛 (det 𝑇).
Proof The constant term of a polynomial function of 𝑧 is the value of the poly-
nomial when 𝑧 = 0. Thus the constant term of the characteristic polynomial of 𝑇
equals det(−𝑇), which equals (−1)𝑛 det 𝑇 (by the third bullet point of 9.42).
Fix a basis of 𝑉, and let 𝐴 be the matrix of 𝑇 with respect to this basis. The
matrix of 𝑧𝐼 − 𝑇 with respect to this basis is 𝑧𝐼 − 𝐴. The term coming from the
identity permutation {1, …, 𝑛} in the formula 9.46 for det(𝑧𝐼 − 𝐴) is
(𝑧 − 𝐴1, 1 )⋯(𝑧 − 𝐴𝑛, 𝑛 ).
The coefficient of 𝑧𝑛 − 1 in the expression above is −(𝐴1, 1 + ⋯ + 𝐴𝑛, 𝑛 ), which equals
− tr 𝑇. The terms in the formula for det(𝑧𝐼 − 𝐴) coming from other elements of
perm 𝑛 contain at most 𝑛 − 2 factors of the form 𝑧 − 𝐴𝑘, 𝑘 and thus do not contribute
to the coefficient of 𝑧𝑛 − 1 in the characteristic polynomial of 𝑇.
In the result below, think of the The next result was proved by Jacques
columns of the 𝑛-by-𝑛 matrix 𝐴 as ele- Hadamard (1865–1963) in 1893.
ments of 𝐅𝑛. The norms appearing below
then arise from the standard inner product on 𝐅𝑛 . Recall that the notation 𝑅⋅, 𝑘 in
the proof below means the 𝑘 th column of the matrix 𝑅 (as was defined in 3.44).
9.66 Hadamard’s inequality
Suppose 𝐴 is an 𝑛-by-𝑛 matrix. Let 𝑣1 , …, 𝑣𝑛 denote the columns of 𝐴. Then

𝑛
|det 𝐴| ≤ ∏ ‖𝑣𝑘 ‖.
𝑘=1
Proof If 𝐴 is not invertible, then det 𝐴 = 0 and hence the desired inequality
holds in this case.
Thus assume that 𝐴 is invertible. The QR factorization (7.58) tells us that
there exist a unitary matrix 𝑄 and an upper-triangular matrix 𝑅 whose diagonal
contains only positive numbers such that 𝐴 = 𝑄𝑅. We have
|det 𝐴| = |det 𝑄| |det 𝑅|
= |det 𝑅|
𝑛
= ∏ 𝑅 𝑘, 𝑘
𝑘=1
𝑛
≤ ∏ ‖𝑅⋅, 𝑘 ‖
𝑘=1
𝑛
= ∏ ‖𝑄𝑅⋅, 𝑘 ‖
𝑘=1
𝑛
= ∏ ‖𝑣𝑘 ‖,
𝑘=1
where the first line comes from 9.49(b), the second line comes from 9.58, the
third line comes from 9.48, and the fifth line holds because 𝑄 is an isometry.
To give a geometric interpretation to Hadamard’s inequality, suppose 𝐅 = 𝐑 .

Let 𝑇 ∈ ℒ(𝐑𝑛 ) be the operator such that 𝑇𝑒𝑘 = 𝑣𝑘 for each 𝑘 = 1, …, 𝑛, where
𝑒1 , …, 𝑒𝑛 is the standard basis of 𝐑𝑛. Then 𝑇 maps the box 𝑃(𝑒1 , …, 𝑒𝑛 ) onto the
parallelepiped 𝑃(𝑣1 , …, 𝑣𝑛 ) [see 7.102 and 7.105 for a review of this notation
and terminology]. Because the box 𝑃(𝑒1 , …, 𝑒𝑛 ) has volume 1, this implies (by
9.61) that the parallelepiped 𝑃(𝑣1 , …, 𝑣𝑛 ) has volume |det 𝑇|, which equals |det 𝐴|.
Thus Hadamard’s inequality above can be interpreted to say that among all paral-
lelepipeds whose edges have lengths ‖𝑣1 ‖, …, ‖𝑣𝑛 ‖, the ones with largest volume
have orthogonal edges (and thus have volume ∏𝑛𝑘 = 1 ‖𝑣𝑘 ‖).
For a necessary and sufficient condition for Hadamard’s inequality to be an
equality, see Exercise 18.
The matrix in the next result is called the Vandermonde matrix. Vandermonde
matrices have important applications in polynomial interpolation, the discrete
Fourier transform, and other areas of mathematics. The proof of the next result is
a nice illustration of the power of switching between matrices and linear maps.
9.67 determinant of Vandermonde matrix
Suppose 𝑛 > 1 and 𝛽1 , …, 𝛽𝑛 ∈ 𝐅. Then
1 𝛽1 𝛽12 ⋯ 𝛽1𝑛 − 1
⎛
⎜ ⎞
⎟
⎜ ⎟
⎜
⎜
⎜ 1 𝛽2 𝛽22 ⋯ 𝛽2𝑛 − 1 ⎟
⎟
⎟
⎜
⎜ ⎟
⎟
det ⎜
⎜ ⎟
⎟ = ∏ (𝛽𝑘 − 𝛽𝑗 ).
⎜
⎜ ⋱ ⎟
⎟ 1 ≤ 𝑗 <𝑘≤𝑛
⎜
⎜ ⎟
⎟
⎜
⎜ ⎟
⎟
⎝ 1 𝛽𝑛 𝛽𝑛2 ⋯ 𝛽𝑛𝑛 − 1 ⎠
Proof Let 1, 𝑧, …, 𝑧𝑛 − 1 be the standard basis of 𝒫𝑛 − 1 (𝐅) and let 𝑒1 , …, 𝑒𝑛 denote

the standard basis of 𝐅𝑛. Define a linear map 𝑆 ∶ 𝒫𝑛 − 1 (𝐅) → 𝐅𝑛 by
𝑆𝑝 = (𝑝(𝛽1 ), …, 𝑝(𝛽𝑛 )).
Let 𝐴 denote the Vandermonde matrix shown in the statement of this result.
Note that
𝐴 = ℳ(𝑆, (1, 𝑧, …, 𝑧𝑛 − 1 ), (𝑒1 , …, 𝑒𝑛 )).
Let 𝑇 ∶ 𝒫𝑛 − 1 (𝐅) → 𝒫𝑛 − 1 (𝐅) be the operator on 𝒫𝑛 − 1 (𝐅) such that 𝑇1 = 1
and
𝑇𝑧𝑘 = (𝑧 − 𝛽1 )(𝑧 − 𝛽2 )⋯(𝑧 − 𝛽𝑘 )
for 𝑘 = 1, …, 𝑛 − 1. Let 𝐵 = ℳ(𝑇, (1, 𝑧, …, 𝑧𝑛 − 1 ), (1, 𝑧, …, 𝑧𝑛 − 1 )). Then 𝐵 is an
upper-triangular matrix all of whose diagonal entries equal 1. Thus det 𝐵 = 1 (by
9.48).
Let 𝐶 = ℳ(𝑆𝑇, (1, 𝑧, …, 𝑧𝑛 − 1 ), (𝑒1 , …, 𝑒𝑛 )). Thus 𝐶 = 𝐴𝐵 (by 3.81), which
implies that
det 𝐴 = (det 𝐴)(det 𝐵) = det 𝐶.
The definitions of 𝐶, 𝑆, and 𝑇 show that 𝐶 equals
1 0 0 ⋯ 0
⎛
⎜ ⎞
⎟
⎜
⎜ 1 𝛽2 − 𝛽 1 0 ⋯ 0 ⎟
⎟
⎜
⎜ ⎟
⎟
⎜
⎜ ⎟
⎟
⎜
⎜ 1 𝛽2 − 𝛽 1 (𝛽3 − 𝛽1 )(𝛽3 − 𝛽2 ) ⋯ 0 ⎟
⎟
⎜
⎜ ⎟
⎟ .
⎜
⎜ ⎟
⎟
⎜
⎜ ⎟
⎟
⎜
⎜ ⋱ ⎟
⎟
⎜
⎜ ⎟
⎟
⎜ ⎟
⎝ 1 𝛽2 − 𝛽 1 (𝛽3 − 𝛽1 )(𝛽3 − 𝛽2 ) ⋯ (𝛽𝑛 − 𝛽1 )(𝛽𝑛 − 𝛽2 )⋯(𝛽𝑛 − 𝛽𝑛 − 1 ) ⎠
Now det 𝐴 = det 𝐶 = ∏ (𝛽𝑘 − 𝛽𝑗 ), where we have used 9.56(a) and 9.48.
1≤𝑗<𝑘≤𝑛
Exercises 9C
1 Prove or give a counterexample: 𝑆, 𝑇 ∈ ℒ(𝑉) ⟹ det(𝑆 + 𝑇) = det 𝑆 +det 𝑇.

2 Suppose the first column of a square matrix 𝐴 consists of all zeros except
possibly the first entry 𝐴1, 1 . Let 𝐵 be the matrix obtained from 𝐴 by deleting
the first row and the first column of 𝐴. Show that det 𝐴 = 𝐴1, 1 det 𝐵.
3 Suppose 𝑇 ∈ ℒ(𝑉) is nilpotent. Prove that det(𝐼 + 𝑇) = 1.
4 Suppose 𝑆 ∈ ℒ(𝑉). Prove that 𝑆 is unitary if and only if |det 𝑆| = ‖𝑆‖ = 1.
5 Suppose 𝐴 is a block upper-triangular matrix
𝐴
⎛ 1
∗ ⎞
𝐴=⎜
⎜
⎜ ⋱ ⎟
⎟,
⎟
⎝ 0 𝐴𝑚 ⎠
where each 𝐴𝑘 along the diagonal is a square matrix. Prove that
det 𝐴 = (det 𝐴1 )⋯(det 𝐴𝑚 ).
6 Suppose 𝐴 = ( 𝑣1 ⋯ 𝑣𝑛 ) is an 𝑛-by-𝑛 matrix, with 𝑣𝑘 denoting the 𝑘 th

column of 𝐴. Show that if (𝑚1 , …, 𝑚𝑛 ) ∈ perm 𝑛, then
det ( 𝑣𝑚1 ⋯ 𝑣𝑚𝑛 ) = (sign(𝑚1 , …, 𝑚𝑛 )) det 𝐴.
7 Suppose 𝑇 ∈ ℒ(𝑉) is invertible. Let 𝑝 denote the characteristic polynomial

of 𝑇 and let 𝑞 denote the characteristic polynomial of 𝑇 −1 . Prove that
1 dim 𝑉 1
𝑞(𝑧) = 𝑧 𝑝( )
𝑝(0) 𝑧
for all nonzero 𝑧 ∈ 𝐅.
8 Suppose 𝑇 ∈ ℒ(𝑉) is an operator with no eigenvalues (which implies that
𝐅 = 𝐑 ). Prove that det 𝑇 > 0.
9 Suppose that 𝑉 is a real vector space of even dimension, 𝑇 ∈ ℒ(𝑉), and

det 𝑇 < 0. Prove that 𝑇 has at least two distinct eigenvalues.
10 Suppose 𝑉 is a real vector space of odd dimension and 𝑇 ∈ ℒ(𝑉). Without
using the minimal polynomial, prove that 𝑇 has an eigenvalue.
This result was previously proved without using determinants or the charac-
teristic polynomial—see 5.34.
11 Prove or give a counterexample: If 𝐅 = 𝐑 , 𝑇 ∈ ℒ(𝑉), and det 𝑇 > 0, then

𝑇 has a square root.
If 𝐅 = 𝐂 , 𝑇 ∈ ℒ(𝑉), and det 𝑉 ≠ 0, then 𝑉 has a square root (see 8.41).
12 Suppose 𝑆, 𝑇 ∈ ℒ(𝑉) and 𝑆 is invertible. Define 𝑝 ∶ 𝐅 → 𝐅 by

𝑝(𝑧) = det(𝑧𝑆 − 𝑇).
Prove that 𝑝 is a polynomial of degree dim 𝑉 and that the coefficient of 𝑧dim 𝑉
in this polynomial is det 𝑆.
13 Suppose 𝐅 = 𝐂 , 𝑇 ∈ ℒ(𝑉), and 𝑛 = dim 𝑉 > 2. Let 𝜆1 , …, 𝜆𝑛 denote
the eigenvalues of 𝑇, with each eigenvalue included as many times as its
multiplicity.
(a) Find a formula for the coefficient of 𝑧𝑛 − 2 in the characteristic polynomial
of 𝑇 in terms of 𝜆1 , …, 𝜆𝑛 .
(b) Find a formula for the coefficient of 𝑧 in the characteristic polynomial
of 𝑇 in terms of 𝜆1 , …, 𝜆𝑛 .
14 Suppose 𝑉 is an inner product space and 𝑇 is a positive operator on 𝑉. Prove

that
det √𝑇 = √det 𝑇.
15 Suppose 𝑉 is an inner product space and 𝑇 ∈ ℒ(𝑉). Use the polar decom-
position to give a proof that
|det 𝑇| = √det(𝑇 ∗ 𝑇)
that is different from the proof given earlier (see 9.60).

16 Suppose 𝑇 ∈ ℒ(𝑉). Define 𝑔 ∶ 𝐅 → 𝐅 by 𝑔(𝑥) = det(𝐼 + 𝑥𝑇). Show that
𝑔’(0) = tr 𝑇.
Look for a clean solution to this exercise, without using the explicit but
complicated formula for the determinant of a matrix.
17 Suppose 𝑎, 𝑏, 𝑐 are positive numbers. Find the volume of the ellipsoid

𝑥2 𝑦2 𝑧2
{(𝑥, 𝑦, 𝑧) ∈ 𝐑3 ∶ + 2 + 2 < 1}
𝑎2 𝑏 𝑐
by finding a set Ω ⊆ 𝐑3 whose volume you know and an operator 𝑇 on 𝐑3
such that 𝑇(Ω) equals the ellipsoid above.
18 Suppose that 𝐴 is an invertible square matrix. Prove that Hadamard’s
inequality (9.66) is an equality if and only if each column of 𝐴 is orthogonal
to the other columns.
19 Suppose 𝑉 is an inner product space, 𝑒1 , …, 𝑒𝑛 is an orthonormal basis of 𝑉,
and 𝑇 ∈ ℒ(𝑉) is a positive operator.
(a) Prove that det 𝑇 ≤ ∏𝑛𝑘 = 1 ⟨𝑇𝑒𝑘 , 𝑒𝑘 ⟩.
(b) Prove that if 𝑇 is invertible, then the inequality in (a) is an equality if
and only if 𝑒𝑘 is an eigenvector of 𝑇 for each 𝑘 = 1, …, 𝑛.
20 Suppose 𝐴 is an 𝑛-by-𝑛 matrix, and suppose 𝑐 is such that |𝐴𝑗, 𝑘 | ≤ 𝑐 for all
𝑗, 𝑘 ∈ {1, …, 𝑛}. Prove that
|det 𝐴| ≤ 𝑐𝑛 𝑛𝑛/2.
The formula for the determinant of a matrix (9.46) shows that |det 𝐴| ≤ 𝑐𝑛 𝑛!.
However, the estimate given by this exercise is much better. For example, if
𝑐 = 1 and 𝑛 = 100, then 𝑐𝑛 𝑛! ≈ 10158, but the estimate given by this exercise
is the much smaller number 10100. If 𝑛 is an integer power of 2, then the
inequality above is sharp and cannot be improved.
21 Suppose 𝑛 is a positive integer and 𝛿 ∶ 𝐂𝑛, 𝑛 → 𝐂 is a function such that
𝛿(𝐴𝐵) = 𝛿(𝐴) ⋅ 𝛿(𝐵)
for all 𝐴, 𝐵 ∈ 𝐂𝑛, 𝑛 and 𝛿(𝐴) equals the product of the diagonal entries of 𝐴
for each diagonal matrix 𝐴 ∈ 𝐂𝑛, 𝑛 . Prove that
𝛿(𝐴) = det 𝐴
for all 𝐴 ∈ 𝐂𝑛, 𝑛 .

Recall that 𝐂𝑛, 𝑛 denotes set of 𝑛-by-𝑛 matrices with entries in 𝐂 . This
exercise shows that the determinant is the unique function defined on square
matrices that is multiplicative and has the desired behavior on diagonal
matrices. This result is analogous to Exercise 10 in Section 8D, which
shows that the trace is uniquely determined by its algebraic properties.
I find that in my own elementary lectures, I have, for pedagogical reasons, pushed
determinants more and more into the background. Too often I have had the expe-
rience that, while the students acquired facility with the formulas, which are so
useful in abbreviating long expressions, they often failed to gain familiarity with
their meaning, and skill in manipulation prevented the student from going into all
the details of the subject and so gaining a mastery.
—Elementary Mathematics from an Advanced Standpoint: Geometry, Felix Klein

9D Tensor Products
Tensor Product of Two Vector Spaces
The motivation for our next topic comes from wanting to form the product of
a vector 𝑣 ∈ 𝑉 and a vector 𝑤 ∈ 𝑊. This product will be denoted by 𝑣 ⊗ 𝑤,
pronounced “𝑣 tensor 𝑤”, and will be an element of some new vector space called
𝑉 ⊗ 𝑊 (also pronounced “𝑉 tensor 𝑊 ”).
We already have a vector space 𝑉 × 𝑊 (see Section 3E), called the product of
𝑉 and 𝑊. However, 𝑉 × 𝑊 will not serve our purposes here because it does not
provide a natural way to multiply an element of 𝑉 by an element of 𝑊. We would
like our tensor product to satisfy some of the usual properties of multiplication.
For example, we would like the distributive property to be satisfied, meaning that
if 𝑣1 , 𝑣2 , 𝑣 ∈ 𝑉 and 𝑤1 , 𝑤2 , 𝑤 ∈ 𝑊, then
(𝑣1 + 𝑣2 ) ⊗ 𝑤 = 𝑣1 ⊗ 𝑤 + 𝑣2 ⊗ 𝑤 and 𝑣 ⊗ (𝑤1 + 𝑤2 ) = 𝑣 ⊗ 𝑤1 + 𝑣 ⊗ 𝑤2 .
We would also like scalar multiplica- To produce ⊗ in TEX, type \otimes.
tion to interact well with this new multi-
plication, meaning that
𝜆(𝑣 ⊗ 𝑤) = (𝜆𝑣) ⊗ 𝑤 = 𝑣 ⊗ (𝜆𝑤)
for all 𝜆 ∈ 𝐅, 𝑣 ∈ 𝑉, and 𝑤 ∈ 𝑊.
Furthermore, it would be nice if each basis of 𝑉 when combined with each
basis of 𝑊 produced a basis of 𝑉 ⊗ 𝑊. Specifically, if 𝑒1 , …, 𝑒𝑚 is a basis of 𝑉
and 𝑓1 , …, 𝑓𝑛 is a basis of 𝑊, then we would like a list (in any order) consisting
of 𝑒𝑗 ⊗ 𝑓𝑘 , as 𝑗 ranges from 1 to 𝑚 and 𝑘 ranges from 1 to 𝑛, to be a basis of
𝑉⊗ 𝑊. This implies that dim(𝑉⊗ 𝑊) should equal (dim 𝑉)(dim 𝑊). Recall that
dim(𝑉 × 𝑊) = dim 𝑉 + dim 𝑊 (see 3.92), which shows that the product 𝑉 × 𝑊
will not serve our purposes here.
To produce a vector space whose dimension is (dim 𝑉)(dim 𝑊) in a natural
fashion from 𝑉 and 𝑊, we look at the vector space of bilinear functionals, as
defined below.
9.68 definition: bilinear functional on 𝑉 × 𝑊, the vector space ℬ(𝑉, 𝑊)
• A bilinear functional on 𝑉 × 𝑊 is a function 𝛽 ∶ 𝑉 × 𝑊 → 𝐅 such that

𝑣 ↦ 𝛽(𝑣, 𝑤) is a linear functional on 𝑉 for each 𝑤 ∈ 𝑊 and 𝑤 ↦ 𝛽(𝑣, 𝑤)
is a linear functional on 𝑊 for each 𝑣 ∈ 𝑉.
• The vector space of bilinear functionals on 𝑉 × 𝑊 is denoted by ℬ(𝑉, 𝑊).
If 𝑊 = 𝑉, then a bilinear functional on 𝑉 × 𝑊 is a bilinear form; see 9.1.

The operations of addition and scalar multiplication on ℬ(𝑉, 𝑊) are defined
to be the usual operations of addition and scalar multiplication of functions. As
you can verify, these operations make ℬ(𝑉, 𝑊) into a vector space whose additive
identity is the zero function from 𝑉 × 𝑊 to 𝐅.
Section 9D Tensor Products 371
9.69 example: bilinear functionals
• Suppose 𝜑 ∈ 𝑉 ′ and 𝜏 ∈ 𝑊 ′. Define 𝛽 ∶ 𝑉 × 𝑊 → 𝐅 by 𝛽(𝑣, 𝑤) = 𝜑(𝑣)𝜏(𝑤).

Then 𝛽 is a bilinear functional on 𝑉 × 𝑊.
• Suppose 𝑣 ∈ 𝑉 and 𝑤 ∈ 𝑊. Define 𝛽 ∶ 𝑉 ′ × 𝑊 ′ → 𝐅 by 𝛽(𝜑, 𝜏) = 𝜑(𝑣)𝜏(𝑤).
Then 𝛽 is a bilinear functional on 𝑉 ′ × 𝑊 ′.
• Define 𝛽 ∶ 𝑉 × 𝑉 ′ → 𝐅 by 𝛽(𝑣, 𝜑) = 𝜑(𝑣). Then 𝛽 is a bilinear functional on
𝑉 × 𝑉 ′.
• Suppose 𝜑 ∈ 𝑉 ′. Define 𝛽 ∶ 𝑉 × ℒ(𝑉) → 𝐅 by 𝛽(𝑣, 𝑇) = 𝜑(𝑇𝑣). Then 𝛽 is a
bilinear functional on 𝑉 × ℒ(𝑉).
• Suppose 𝑚 and 𝑛 are positive integers. Define 𝛽 ∶ 𝐅𝑚, 𝑛 ×𝐅𝑛, 𝑚 → 𝐅 by 𝛽(𝐴, 𝐵) =
tr(𝐴𝐵). Then 𝛽 is a bilinear functional on 𝐅𝑚, 𝑛 × 𝐅𝑛, 𝑚.
9.70 dimension of the vector space of bilinear functionals
dim ℬ(𝑉, 𝑊) = (dim 𝑉)(dim 𝑊).
Proof Let 𝑒1 , …, 𝑒𝑚 be a basis of 𝑉 and 𝑓1 , …, 𝑓𝑛 be a basis of 𝑊. For a bilinear

functional 𝛽 ∈ ℬ(𝑉, 𝑊), let ℳ(𝛽) be the 𝑚-by-𝑛 matrix whose entry in row 𝑗,
column 𝑘 is 𝛽(𝑒𝑗 , 𝑓𝑘 ). The map 𝛽 ↦ ℳ(𝛽) is a linear map of ℬ(𝑉, 𝑊) into 𝐅𝑚, 𝑛.
For a matrix 𝐶 ∈ 𝐅𝑚, 𝑛, define a bilinear functional 𝛽𝐶 on 𝑉 × 𝑊 by
𝑛 𝑚
𝛽𝐶 (𝑎1 𝑒1 + ⋯ + 𝑎𝑚 𝑒𝑚 , 𝑏1 𝑓1 + ⋯ + 𝑏𝑛 𝑓𝑛 ) = ∑ ∑ 𝐶𝑗, 𝑘 𝑎𝑗 𝑏𝑘
𝑘=1 𝑗 = 1
for 𝑎1 , …, 𝑎𝑚 , 𝑏1 , …, 𝑏𝑛 ∈ 𝐅.
The linear map 𝛽 ↦ ℳ(𝛽) from ℬ(𝑉, 𝑊) to 𝐅𝑚, 𝑛 and the linear map 𝐶 ↦ 𝛽𝐶
from 𝐅𝑚, 𝑛 to ℬ(𝑉, 𝑊) are inverses of each other because 𝛽ℳ(𝛽) = 𝛽 for all
𝛽 ∈ ℬ(𝑉, 𝑊) and ℳ(𝛽𝐶 ) = 𝐶 for all 𝐶 ∈ 𝐅𝑚, 𝑛, as you should verify.
Thus both maps are isomorphisms and the two spaces that they connect have the
same dimension. Hence dim ℬ(𝑉, 𝑊) = dim 𝐅𝑚, 𝑛 = 𝑚𝑛 = (dim 𝑉)(dim 𝑊).
Several different definitions of 𝑉 ⊗ 𝑊 appear in the mathematical literature.

These definitions are equivalent to each other, at least in the finite-dimensional
context, because any two vector spaces of the same dimension are isomorphic.
The result above states that ℬ(𝑉, 𝑊) has the dimension that we seek, as do
ℒ(𝑉, 𝑊) and 𝐅dim 𝑉,dim 𝑊. Thus it may be tempting to define 𝑉⊗ 𝑊 to be ℬ(𝑉, 𝑊)
or ℒ(𝑉, 𝑊) or 𝐅dim 𝑉,dim 𝑊. However, none of those definitions would lead to a
basis-free definition of 𝑣 ⊗ 𝑤 for 𝑣 ∈ 𝑉 and 𝑤 ∈ 𝑊.
The following definition, while it may seem a bit strange and abstract at first,
has the huge advantage that it defines 𝑣 ⊗ 𝑤 in a basis-free fashion. We define
𝑉 ⊗ 𝑊 to be the vector space of bilinear functionals on 𝑉 ′ × 𝑊 ′ instead of the
more tempting choice of the vector space of bilinear functionals on 𝑉 × 𝑊.
9.71 definition: tensor product, 𝑉 ⊗ 𝑊, 𝑣 ⊗ 𝑤
• The tensor product 𝑉 ⊗ 𝑊 is defined to be ℬ(𝑉 ′ , 𝑊 ′ ).

• For 𝑣 ∈ 𝑉 and 𝑤 ∈ 𝑊, the tensor product 𝑣 ⊗ 𝑤 is the element of 𝑉 ⊗ 𝑊
defined by
(𝑣 ⊗ 𝑤)(𝜑, 𝜏) = 𝜑(𝑣)𝜏(𝑤)
for all (𝜑, 𝜏) ∈ 𝑉 ′ × 𝑊 ′.
We can quickly prove that the definition of 𝑉⊗𝑊 gives it the desired dimension.
9.72 dimension of the tensor product of two vector spaces
dim(𝑉 ⊗ 𝑊) = (dim 𝑉)(dim 𝑊).
Proof Because a vector space and its dual have the same dimension (by 3.111),
we have dim 𝑉 ′ = dim 𝑉 and dim 𝑊 ′ = dim 𝑊. Thus 9.70 tells us that the
dimension of ℬ(𝑉 ′ , 𝑊 ′ ) equals (dim 𝑉)(dim 𝑊).
To understand the definition of the tensor product 𝑣 ⊗ 𝑤 of two vectors 𝑣 ∈ 𝑉

and 𝑤 ∈ 𝑊, focus on the kind of object it is. An element of 𝑉 ⊗ 𝑊 is a bilinear
functional on 𝑉 ′ × 𝑊 ′, and in particular it is a function from 𝑉 ′ × 𝑊 ′ to 𝐅. Thus for
each element of 𝑉 ′ × 𝑊 ′, it should produce an element of 𝐅. The definition above
has this behavior, because 𝑣 ⊗ 𝑤 applied to a typical element (𝜑, 𝜏) of 𝑉 ′ × 𝑊 ′
produces the number 𝜑(𝑣)𝜏(𝑤).
The somewhat abstract nature of 𝑣 ⊗ 𝑤 should not matter. The important point
is the behavior of these objects. The next result shows that tensor products of
vectors have the desired bilinearity properties.
9.73 bilinearity of tensor product
Suppose 𝑣, 𝑣1 , 𝑣2 ∈ 𝑉 and 𝑤, 𝑤1 , 𝑤2 ∈ 𝑊 and 𝜆 ∈ 𝐅. Then
(𝑣1 + 𝑣2 ) ⊗ 𝑤 = 𝑣1 ⊗ 𝑤 + 𝑣2 ⊗ 𝑤 and 𝑣 ⊗ (𝑤1 + 𝑤2 ) = 𝑣 ⊗ 𝑤1 + 𝑣 ⊗ 𝑤2
and
𝜆(𝑣 ⊗ 𝑤) = (𝜆𝑣) ⊗ 𝑤 = 𝑣 ⊗ (𝜆𝑤).
Proof Suppose (𝜑, 𝜏) ∈ 𝑉 ′ × 𝑊 ′. Then

((𝑣1 + 𝑣2 ) ⊗ 𝑤)(𝜑, 𝜏) = 𝜑(𝑣1 + 𝑣2 )𝜏(𝑤)
= 𝜑(𝑣1 )𝜏(𝑤) + 𝜑(𝑣2 )𝜏(𝑤)
= (𝑣1 ⊗ 𝑤)(𝜑, 𝜏) + (𝑣2 ⊗ 𝑤)(𝜑, 𝜏)
= (𝑣1 ⊗ 𝑤 + 𝑣2 ⊗ 𝑤)(𝜑, 𝜏).
Thus (𝑣1 + 𝑣2 ) ⊗ 𝑤 = 𝑣1 ⊗ 𝑤 + 𝑣2 ⊗ 𝑤.
The other two equalities are proved similarly.
Lists are, by definition, ordered. The order matters when, for example, we
form the matrix of an operator with respect to a basis. For lists in this section
with two indices, such as {𝑒𝑗 ⊗ 𝑓𝑘 }𝑗 = 1, …, 𝑚; 𝑘 = 1, …, 𝑛 in the next result, the ordering
does not matter and we do not specify it—just choose any convenient ordering.
The linear independence of elements of 𝑉 ⊗ 𝑊 in (a) of the result below
captures the idea that there are no relationships among vectors in 𝑉 ⊗ 𝑊 other
than the relationships that come from bilinearity of the tensor product (see 9.73)
and the relationships that may be present due to linear dependence of a list of
vectors in 𝑉 or a list of vectors in 𝑊.
9.74 basis of 𝑉 ⊗ 𝑊
Suppose 𝑒1 , …, 𝑒𝑚 is a list of vectors in 𝑉 and 𝑓1 , …, 𝑓𝑛 is a list of vectors in 𝑊.

(a) If 𝑒1 , …, 𝑒𝑚 and 𝑓1 , …, 𝑓𝑛 are both linearly independent lists, then
{𝑒𝑗 ⊗ 𝑓𝑘 }𝑗 = 1, …, 𝑚; 𝑘 = 1, …, 𝑛
is a linearly independent list in 𝑉 ⊗ 𝑊.

(b) If 𝑒1 , …, 𝑒𝑚 is a basis of 𝑉 and 𝑓1 , …, 𝑓𝑛 is a basis of 𝑊, then the list
{𝑒𝑗 ⊗ 𝑓𝑘 }𝑗 = 1, …, 𝑚; 𝑘 = 1, …, 𝑛 is a basis of 𝑉 ⊗ 𝑊.
Proof To prove (a), suppose 𝑒1 , …, 𝑒𝑚 and 𝑓1 , …, 𝑓𝑛 are both linearly independent

lists. This linear independence and the linear map lemma (3.4) imply that there
exist 𝜑1 , …, 𝜑𝑚 ∈ 𝑉 ′ and 𝜏1 , …, 𝜏𝑛 ∈ 𝑊 ′ such that
⎧
{1 if 𝑗 = 𝑘, {1 if 𝑗 = 𝑘,
⎧
𝜑𝑗 (𝑒𝑘 ) = ⎨ and
𝜏𝑗 ( 𝑓𝑘 ) = ⎨
{
⎩0 if 𝑗 ≠ 𝑘 ⎩0 if 𝑗 ≠ 𝑘,
{
where 𝑗, 𝑘 ∈ {1, …, 𝑚} in the first equation and 𝑗, 𝑘 ∈ {1, …, 𝑛} in the second
equation.
Suppose {𝑎𝑗, 𝑘 }𝑗 = 1, …, 𝑚; 𝑘 = 1, …, 𝑛 is a list of scalars such that
𝑛 𝑚
9.75 ∑ ∑ 𝑎𝑗, 𝑘 (𝑒𝑗 ⊗ 𝑓𝑘 ) = 0.
𝑘=1 𝑗 = 1
Note that (𝑒𝑗 ⊗ 𝑓𝑘 )(𝜑𝑀 , 𝜏𝑁 ) equals 1 if 𝑗 = 𝑀 and 𝑘 = 𝑁, and equals 0 otherwise.

Thus applying both sides of 9.75 to (𝜑𝑀 , 𝜏𝑁 ) shows that 𝑎𝑀, 𝑁 = 0, proving that
{𝑒𝑗 ⊗ 𝑓𝑘 }𝑗 = 1, …, 𝑚; 𝑘 = 1, …, 𝑛 is linearly independent.
Now (b) follows from (a), the equation dim 𝑉 ⊗ 𝑊 = (dim 𝑉)(dim 𝑊) [see
9.72], and the result that a linearly independent list of the right length is a basis
(see 2.38).
Every element of 𝑉 × 𝑊 is a finite sum of elements of the form 𝑣 ⊗ 𝑤, where

𝑣 ∈ 𝑉 and 𝑤 ∈ 𝑊, as implied by (b) in the result above. However, if dim 𝑉 > 1
and dim 𝑊 > 1, then Exercise 4 shows that
{𝑣 ⊗ 𝑤 ∶ (𝑣, 𝑤) ∈ 𝑉 × 𝑊} ≠ 𝑉 ⊗ 𝑊.
9.76 example: tensor product of element of 𝐅𝑚 with element of 𝐅𝑛

Suppose 𝑚 and 𝑛 are positive integers. Let 𝑒1 , …, 𝑒𝑚 denote the standard basis
of 𝐅𝑚 and let 𝑓1 , …, 𝑓𝑛 denote the standard basis of 𝐅𝑛. Suppose
𝑣 = (𝑣1 , …, 𝑣𝑚 ) ∈ 𝐅𝑚 and 𝑤 = (𝑤1 , …, 𝑤𝑛 ) ∈ 𝐅𝑛 .
Then
𝑚 𝑛
𝑣 ⊗ 𝑤 = ( ∑ 𝑣𝑗 𝑒𝑗 ) ⊗ ( ∑ 𝑤𝑘 𝑓𝑘 )
𝑗=1 𝑘=1
𝑛 𝑚
= ∑ ∑ (𝑣𝑗 𝑤𝑘 )(𝑒𝑗 ⊗ 𝑓𝑘 ).
𝑘=1 𝑗 = 1
Thus with respect to the basis {𝑒𝑗 ⊗ 𝑓𝑘 }𝑗 = 1, …, 𝑚; 𝑘 = 1, …, 𝑛 of 𝐅𝑚 ⊗ 𝐅𝑛 provided

by 9.74(b), the coefficients of 𝑣 ⊗ 𝑤 are the numbers {𝑣𝑗 𝑤𝑘 }𝑗 = 1, …, 𝑚; 𝑘 = 1, …, 𝑛 . If
instead of writing these numbers in a list, we write them in an 𝑚-by-𝑛 matrix with
𝑣𝑗 𝑤𝑘 in row 𝑗, column 𝑘, then we can identify 𝑣 ⊗ 𝑤 with the 𝑚-by-𝑛 matrix
𝑣1 𝑤1 ⋯ 𝑣1 𝑤𝑛
⎛
⎜ ⎞
⎟
⎜
⎜ ⎟
⎟
⎜
⎜ ⋱ ⎟
⎟ .
⎜ ⎟
⎝ 𝑣𝑚 𝑤1 ⋯ 𝑣𝑚 𝑤𝑛 ⎠
See Exercises 5 and 6 for practice in using the identification from the example
above.
We now define bilinear maps, which differ from bilinear functionals in that
the target space can be an arbitrary vector space rather than just the scalar field.
9.77 definition: bilinear map
A bilinear map from 𝑉 × 𝑊 to a vector space 𝑈 is a function Γ ∶ 𝑉 × 𝑊 → 𝑈

such that 𝑣 ↦ Γ(𝑣, 𝑤) is a linear map from 𝑉 to 𝑈 for each 𝑤 ∈ 𝑊 and
𝑤 ↦ Γ(𝑣, 𝑤) is a linear map from 𝑊 to 𝑈 for each 𝑣 ∈ 𝑉.
9.78 example: bilinear maps
• Every bilinear functional on 𝑉 × 𝑊 is a bilinear map from 𝑉 × 𝑊 to 𝐅.

• The function Γ ∶ 𝑉× 𝑊 → 𝑉⊗ 𝑊 defined by Γ(𝑣, 𝑤) = 𝑣 ⊗ 𝑤 is a bilinear map
from 𝑉 × 𝑊 to 𝑉 ⊗ 𝑊 (by 9.73).
• The function Γ ∶ ℒ(𝑉) × ℒ(𝑉) → ℒ(𝑉) defined by Γ(𝑆, 𝑇) = 𝑆𝑇 is a bilinear
map from ℒ(𝑉) × ℒ(𝑉) to ℒ(𝑉).
• The function Γ ∶ 𝑉 × ℒ(𝑉, 𝑊) → 𝑊 defined by Γ(𝑣, 𝑇) = 𝑇𝑣 is a bilinear map
from 𝑉 × ℒ(𝑉, 𝑊) to 𝑊.
Tensor products allow us to convert bilinear maps on 𝑉× 𝑊 into linear maps on

𝑉⊗ 𝑊 (and vice versa), as shown by the next result. In the mathematical literature,
(a) of the result below is called the “universal property” of tensor products.
9.79 converting bilinear maps to linear maps
Suppose 𝑈 is a vector space.

(a) Suppose Γ ∶ 𝑉 × 𝑊 → 𝑈 is a bilinear map. Then there exists a unique
linear map Γ̂ ∶ 𝑉 ⊗ 𝑊 → 𝑈 such that
̂ ⊗ 𝑤) = Γ(𝑣, 𝑤)
Γ(𝑣
for all (𝑣, 𝑤) ∈ 𝑉 × 𝑊.

(b) Conversely, suppose 𝑇 ∶ 𝑉 ⊗ 𝑊 → 𝑈 is a linear map. There there exists a
unique bilinear map 𝑇 # ∶ 𝑉 × 𝑊 → 𝑈 such that
𝑇 # (𝑣, 𝑤) = 𝑇(𝑣 ⊗ 𝑤)
for all (𝑣, 𝑤) ∈ 𝑉 × 𝑊.
Proof Let 𝑒1 , …, 𝑒𝑚 be a basis of 𝑉 and let 𝑓1 , …, 𝑓𝑛 be a basis of 𝑊. By the linear

map lemma (3.4) and 9.74(b), there exists a unique linear map Γ̂ ∶ 𝑉 ⊗ 𝑊 → 𝑈
such that
̂ ⊗ 𝑓 ) = Γ(𝑒 , 𝑓 )
Γ(𝑒 𝑗 𝑘 𝑗 𝑘
for all 𝑗 ∈ {1, …, 𝑚} and 𝑘 ∈ {1, …, 𝑛}.

Now suppose (𝑣, 𝑤) ∈ 𝑉 × 𝑊. There exist 𝑎1 , …, 𝑎𝑚 , 𝑏1 , …, 𝑏𝑛 ∈ 𝐅 such that
𝑣 = 𝑎1 𝑒1 + ⋯ + 𝑎𝑚 𝑒𝑚 and 𝑤 = 𝑏1 𝑓1 + ⋯ + 𝑏𝑛 𝑓𝑛 . Thus
𝑛 𝑚
̂ ⊗ 𝑤) = Γ(
Γ(𝑣 ̂ ∑ ∑ (𝑎 𝑏 )(𝑒 ⊗ 𝑓 ))
𝑗 𝑘 𝑗 𝑘
𝑘=1 𝑗 = 1
𝑛 𝑚
̂ ⊗ 𝑓)
= ∑ ∑ 𝑎𝑗 𝑏𝑘 Γ(𝑒𝑗 𝑘
𝑘=1 𝑗 = 1
𝑛 𝑚
= ∑ ∑ 𝑎𝑗 𝑏𝑘 Γ(𝑒𝑗 , 𝑓𝑘 )
𝑘=1 𝑗 = 1
= Γ(𝑣, 𝑤),
as desired, where the second line holds because Γ̂ is linear, the third line holds by
the definition of Γ̂, and the fourth line holds because Γ is bilinear.
The uniqueness of the linear map Γ̂ satisfying Γ(𝑣 ̂ ⊗ 𝑤) = Γ(𝑣, 𝑤) follows
from 9.74(b), completing the proof of (a).
To prove (b), define a function 𝑇 # ∶ 𝑉× 𝑊 → 𝑈 by 𝑇 # (𝑣, 𝑤) = 𝑇(𝑣 ⊗ 𝑤) for all
(𝑣, 𝑤) ∈ 𝑉 × 𝑊. The bilinearity of the tensor product (see 9.73) and the linearity
of 𝑇 imply that 𝑇 # is bilinear.
Clearly the choice of 𝑇 # that satisfies the conditions is unique.
To prove 9.79(a), we could not just define Γ(𝑣̂ ⊗ 𝑤) = Γ(𝑣, 𝑤) for all 𝑣 ∈ 𝑉
and 𝑤 ∈ 𝑊 (and then extend Γ linearly to all of 𝑉 ⊗ 𝑊) because elements of
̂
𝑉 ⊗ 𝑊 do not have unique representations as finite sums of elements of the form
𝑣 ⊗ 𝑤. Our proof used a basis of 𝑉 and a basis of 𝑊 to get around this problem.
Although our construction of Γ̂ in the proof of 9.79(a) depended on a basis of
̂ ⊗ 𝑤) = Γ(𝑣, 𝑤) that holds for all 𝑣 ∈ 𝑉 and
𝑉 and a basis of 𝑊, the equation Γ(𝑣
𝑤 ∈ 𝑊 shows that Γ̂ does not depend on the choice of bases for 𝑉 and 𝑊.
Tensor Product of Inner Product Spaces

The result below features three inner products—one on 𝑉⊗ 𝑊, one on 𝑉, and one
on 𝑊, although we use the same symbol ⟨⋅, ⋅⟩ for all three inner products.
9.80 inner product on tensor product of two inner product spaces
Suppose 𝑉 and 𝑊 are inner product spaces. Then there is a unique inner
product on 𝑉 ⊗ 𝑊 such that
⟨𝑣 ⊗ 𝑤, 𝑢 ⊗ 𝑥⟩ = ⟨𝑣, 𝑢⟩⟨𝑤, 𝑥⟩
for all 𝑣, 𝑢 ∈ 𝑉 and 𝑤, 𝑥 ∈ 𝑊.
Proof Suppose 𝑒1 , …, 𝑒𝑚 is an orthonormal basis of 𝑉 and 𝑓1 , …, 𝑓𝑛 is an ortho-

normal basis of 𝑊. Define an inner product on 𝑉 ⊗ 𝑊 by
𝑛 𝑚 𝑛 𝑚 𝑛 𝑚
9.81 ⟨ ∑ ∑ 𝑏𝑗, 𝑘 𝑒𝑗 ⊗ 𝑓𝑘 , ∑ ∑ 𝑐𝑗, 𝑘 𝑒𝑗 ⊗ 𝑓𝑘 ⟩ = ∑ ∑ 𝑏𝑗, 𝑘 𝑐𝑗, 𝑘 .
𝑘=1 𝑗 = 1 𝑘=1 𝑗 = 1 𝑘=1 𝑗 = 1
The straightforward verification that 9.81 defines an inner product on 𝑉 × 𝑊

is left to the reader [use 9.74(b)].
Suppose that 𝑣, 𝑢 ∈ 𝑉 and 𝑤, 𝑥 ∈ 𝑊. Let 𝑣1 , …, 𝑣𝑚 ∈ 𝐅 be such that
𝑣 = 𝑣1 𝑒1 + ⋯ + 𝑣𝑚 𝑒𝑚 , with similar expressions for 𝑢, 𝑤, and 𝑥. Then
𝑚 𝑛 𝑚 𝑛
⟨𝑣 ⊗ 𝑤, 𝑢 ⊗ 𝑥⟩ = ⟨ ∑ 𝑣𝑗 𝑒𝑗 ⊗ ∑ 𝑤𝑘 𝑓𝑘 , ∑ 𝑢𝑗 𝑒𝑗 ⊗ ∑ 𝑥𝑘 𝑓𝑘 ⟩
𝑗=1 𝑘=1 𝑗=1 𝑘=1
𝑛 𝑚 𝑛 𝑚
= ⟨ ∑ ∑ 𝑣𝑗 𝑤𝑘 𝑒𝑗 ⊗ 𝑓𝑘 , ∑ ∑ 𝑢𝑗 𝑥𝑘 𝑒𝑗 ⊗ 𝑓𝑘 ⟩
𝑘=1 𝑗 = 1 𝑘=1 𝑗 = 1
𝑛 𝑚
= ∑ ∑ 𝑣𝑗 𝑢𝑗 𝑤𝑘 𝑥𝑘
𝑘=1 𝑗 = 1
𝑚 𝑛
= ( ∑ 𝑣𝑗 𝑢𝑗 )( ∑ 𝑤𝑘 𝑥𝑘 )
𝑗=1 𝑘=1
= ⟨𝑣, 𝑢⟩⟨𝑤, 𝑥⟩.

There is only one inner product on 𝑉⊗𝑊 such that ⟨𝑣⊗𝑤, 𝑢⊗𝑥⟩ = ⟨𝑣, 𝑢⟩⟨𝑤, 𝑥⟩
for all 𝑣, 𝑢 ∈ 𝑉 and 𝑤, 𝑥 ∈ 𝑊 because every element of 𝑉 ⊗ 𝑊 can be written as
a linear combination of elements of the form 𝑣 ⊗ 𝑤 [by 9.74(b)].
The definition below of a natural inner product on 𝑉 ⊗ 𝑊 is now justified by

9.80. We could not have simply defined ⟨𝑣 ⊗ 𝑤, 𝑢 ⊗ 𝑥⟩ to be ⟨𝑣, 𝑢⟩⟨𝑤, 𝑥⟩ (and then
used additivity in each slot separately to extend the definition to 𝑉 ⊗ 𝑊) without
some proof because elements of 𝑉 ⊗ 𝑊 do not have unique representations as
finite sums of elements of the form 𝑣 ⊗ 𝑤.
9.82 definition: inner product on tensor product of two inner product spaces
Suppose 𝑉 and 𝑊 are inner product spaces. The inner product on 𝑉 ⊗ 𝑊 is

the unique function ⟨⋅, ⋅⟩ from (𝑉 ⊗ 𝑊) × (𝑉 ⊗ 𝑊) to 𝐅 such that
⟨𝑣 ⊗ 𝑤, 𝑢 ⊗ 𝑥⟩ = ⟨𝑣, 𝑢⟩⟨𝑤, 𝑥⟩
for all 𝑣, 𝑢 ∈ 𝑉 and 𝑤, 𝑥 ∈ 𝑊.
Take 𝑢 = 𝑣 and 𝑥 = 𝑤 in the equation above and then take square roots to
show that
‖𝑣 ⊗ 𝑤‖ = ‖𝑣‖ ‖𝑤‖
for all 𝑣 ∈ 𝑉 and all 𝑤 ∈ 𝑊.
The construction of the inner product in the proof of 9.80 depended on an
orthonormal basis 𝑒1 , …, 𝑒𝑚 of 𝑉 and an orthonormal basis 𝑓1 , …, 𝑓𝑛 of 𝑊. Formula
9.81 implies that {𝑒𝑗 ⊗ 𝑓𝑘 }𝑗 = 1, …, 𝑚; 𝑘 = 1, …, 𝑛 is a doubly indexed orthonormal list in
𝑉⊗ 𝑊 and hence is an orthonormal basis of 𝑉⊗ 𝑊 [by 9.74(b)]. The importance
of the next result arises because the orthonormal bases used there can be different
from the orthonormal bases used to define the inner product in 9.80. Although
the notation for the bases is the same in the proof of 9.80 and in the result below,
think of them as two different sets of orthonormal bases.
9.83 orthonormal basis of 𝑉 ⊗ 𝑊
Suppose 𝑉 and 𝑊 are inner product spaces, and 𝑒1 , …, 𝑒𝑚 is an orthonormal

basis of 𝑉 and 𝑓1 , …, 𝑓𝑛 is an orthonormal basis of 𝑊. Then
{𝑒𝑗 ⊗ 𝑓𝑘 }𝑗 = 1, …, 𝑚; 𝑘 = 1, …, 𝑛
is an orthonormal basis of 𝑉 ⊗ 𝑊.
Proof We know that {𝑒𝑗 ⊗ 𝑓𝑘 }𝑗 = 1, …, 𝑚; 𝑘 = 1, …, 𝑛 is a basis of 𝑉 ⊗ 𝑊 [by 9.74(b)].

Thus we only need to verify orthonormality. To do this, suppose 𝑗, 𝑀 ∈ {1, …, 𝑚}
and 𝑘, 𝑁 ∈ {1, …, 𝑛}. Then
{1 if 𝑗 = 𝑁 and 𝑘 = 𝑀,
⎧
⟨𝑒𝑗 ⊗ 𝑓𝑘 , 𝑒𝑁 ⊗ 𝑓𝑀 ⟩ = ⟨𝑒𝑗 , 𝑒𝑁 ⟩⟨ 𝑓𝑘 , 𝑓𝑀 ⟩ = ⎨
⎩0 otherwise.
{
Hence the doubly indexed list {𝑒𝑗 ⊗ 𝑓𝑘 }𝑗 = 1, …, 𝑚; 𝑘 = 1, …, 𝑛 is indeed an orthonormal
basis of 𝑉 ⊗ 𝑊.
See Exercise 11 for an example of how the inner product structure on 𝑉 ⊗ 𝑊

interacts with operators on 𝑉 and 𝑊.
Tensor Product of Multiple Vector Spaces

We have been discussing properties of the tensor product of two finite-dimensional
vector spaces. Now we turn our attention to the tensor product of multiple finite-
dimensional vector spaces. This generalization requires no new ideas, only some
slightly more complicated notation. Readers with a good understanding of the
tensor product of two vector spaces should be able to make the extension to the
tensor product of more than two vector spaces.
Thus in this subsection, no proofs will be provided. The definitions and the
statements of results that will be provided should be enough information to enable
readers to fill in the details, using what has already been learned about the tensor
product of two vector spaces.
We begin with the following notational assumption.
9.84 notation: 𝑉1 , …, 𝑉𝑚
For the rest of this subsection, 𝑚 denotes an integer greater than 1 and
𝑉1 , …, 𝑉𝑚 denote finite-dimensional vector spaces.
The notion of an 𝑚-linear functional, which we are about to define, generalizes

the notion of a bilinear functional (see 9.68). Recall that the use of the word
“functional” indicates that we are mapping into the scalar field 𝐅. Recall also that
the terminology “𝑚-linear form” is used in the special case 𝑉1 = ⋯ = 𝑉𝑚 (see
9.25). The notation ℬ(𝑉1 , …, 𝑉𝑚 ) generalizes our previous notation ℬ(𝑉, 𝑊).
9.85 definition: 𝑚-linear functional, the vector space ℬ(𝑉1 , …, 𝑉𝑚 )
• An 𝑚-linear functional on 𝑉1 × ⋯ × 𝑉𝑚 is a function 𝛽 ∶ 𝑉1 × ⋯ × 𝑉𝑚 → 𝐅

that is a linear functional in each slot when the other slots are held fixed.
• The vector space of 𝑚-linear functionals on 𝑉1 × ⋯ × 𝑉𝑚 is denoted by
ℬ(𝑉1 , …, 𝑉𝑚 ).
9.86 example: 𝑚-linear functional

Suppose 𝜑𝑘 ∈ (𝑉𝑘 )′ for each 𝑘 ∈ {1, …, 𝑚}. Define 𝛽 ∶ 𝑉1 × ⋯ × 𝑉𝑚 → 𝐅 by
𝛽(𝑣1 , …, 𝑣𝑚 ) = 𝜑1 (𝑣1 ) × ⋯ × 𝜑𝑚 (𝑣𝑚 ).
Then 𝛽 is an 𝑚-linear functional on 𝑉1 × ⋯ × 𝑉𝑚 .
The next result can be proved by imitating the proof of 9.70.
9.87 dimension of the vector space of 𝑚-linear functionals
dim ℬ(𝑉1 , …, 𝑉𝑚 ) = (dim 𝑉1 ) × ⋯ × (dim 𝑉𝑚 ).

Now we can define the tensor product of multiple vector spaces and the tensor
product of elements of those vector spaces. The following definition is completely
analogous to our previous definition (9.71) in the case 𝑚 = 2.
9.88 definition: tensor product, 𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 , 𝑣1 ⊗ ⋯ ⊗ 𝑣𝑚
• The tensor product 𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 is defined to be ℬ(𝑉1′ , …, 𝑉𝑚′ ).

• For 𝑣1 ∈ 𝑉1 , …, 𝑣𝑚 ∈ 𝑉𝑚 , the tensor product 𝑣1 ⊗ ⋯ ⊗ 𝑣𝑚 is the element
of 𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 defined by
(𝑣1 ⊗ ⋯ ⊗ 𝑣𝑚 )(𝜑1 , …, 𝜑𝑚 ) = 𝜑1 (𝑣1 )⋯𝜑𝑚 (𝑣𝑚 )
for all (𝜑1 …, 𝜑𝑚 ) ∈ 𝑉1′ × ⋯ × 𝑉𝑚′ .
The next result can be proved by following the pattern of the proof of the
analogous result when 𝑚 = 2 (see 9.72).
9.89 dimension of the tensor product
dim(𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 ) = (dim 𝑉1 )⋯(dim 𝑉𝑚 ).
Our next result generalizes 9.74.
9.90 basis of 𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚
Suppose dim 𝑉𝑘 = 𝑛𝑘 and 𝑒1𝑘 , …, 𝑒𝑛𝑘𝑘 is a basis of 𝑉𝑘 for 𝑘 = 1, …, 𝑚. Then
{𝑒𝑗11 ⊗ ⋯ ⊗ 𝑒𝑗𝑚𝑚 }𝑗1 = 1, …, 𝑛1; ⋯; 𝑗𝑚 = 1, …, 𝑛𝑚
is a basis of 𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 .
Suppose 𝑚 = 2 and 𝑒11 , …, 𝑒𝑛11 is a basis of 𝑉1 and 𝑒12 , …, 𝑒𝑛22 is a basis of 𝑉2 .

Then with respect to the basis {𝑒𝑗11 ⊗ 𝑒𝑗22 }𝑗1 = 1, …, 𝑛1; 𝑗2 = 1, …, 𝑛2 in the result above, the
coefficients of an element of 𝑉1 ⊗𝑉2 can be represented by an 𝑛1 -by-𝑛2 matrix that
contains the coefficient of 𝑒𝑗11 ⊗ 𝑒𝑗22 in row 𝑗1 , column 𝑗2 . Thus we need a matrix,
which is an array specified by two indices, to represent an element of 𝑉1 ⊗ 𝑉2 .
If 𝑚 > 2, then the result above shows that we need an array specified by 𝑚
indices to represent an arbitrary element of 𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 . Thus tensor products
may appear when we deal with objects specified by arrays with multiple indices.
The next definition generalizes the notion of a bilinear map (see 9.77). As
with bilinear maps, the target space can be an arbitrary vector space.
9.91 definition: 𝑚-linear map
An 𝑚-linear map from 𝑉1 × ⋯ × 𝑉𝑚 to a vector space 𝑈 is a function

Γ ∶ 𝑉1 × ⋯ × 𝑉𝑚 → 𝑈 that is a linear map in each slot when the other slots are
held fixed.
The next result can be proved by following the pattern of the proof of 9.79.
9.92 converting 𝑚-linear maps to linear maps
Suppose 𝑈 is a vector space.

(a) Suppose that Γ ∶ 𝑉1 × ⋯ × 𝑉𝑚 → 𝑈 is an 𝑚-linear map. Then there exists
a unique linear map Γ̂ ∶ 𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 → 𝑈 such that
̂
Γ(𝑣1 ⊗ ⋯ ⊗ 𝑣𝑚 ) = Γ(𝑣1 , …, 𝑣𝑚 )
for all (𝑣1 , …, 𝑣𝑚 ) ∈ 𝑉1 × ⋯ × 𝑉𝑚 .

(b) Conversely, suppose 𝑇 ∶ 𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 → 𝑈 is a linear map. There there
exists a unique 𝑚-linear map 𝑇 # ∶ 𝑉1 × ⋯ × 𝑉𝑚 → 𝑈 such that
𝑇 # (𝑣1 , …, 𝑣𝑚 ) = 𝑇(𝑣1 ⊗ ⋯ ⊗ 𝑣𝑚 )
for all (𝑣1 , …, 𝑣𝑚 ) ∈ 𝑉1 × ⋯ × 𝑉𝑚 .
See Exercises 12 and 13 for tensor products of multiple inner product spaces.
Exercises 9D
1 Suppose 𝑣 ∈ 𝑉 and 𝑤 ∈ 𝑊. Prove that 𝑣 ⊗ 𝑤 = 0 if and only if 𝑣 = 0 or

𝑤 = 0.
2 Give an example of six distinct vectors 𝑣1 , 𝑣2 , 𝑣3 , 𝑤1 , 𝑤2 , 𝑤3 in 𝐑3 such that
𝑣1 ⊗ 𝑤1 + 𝑣2 ⊗ 𝑤2 + 𝑣3 ⊗ 𝑤3 = 0
but none of 𝑣1 ⊗ 𝑤1 , 𝑣2 ⊗ 𝑤2 , 𝑣3 ⊗ 𝑤3 is a scalar multiple of another element

of this list.
3 Suppose that 𝑣1 , …, 𝑣𝑚 is a linearly independent list in 𝑉. Suppose also that
𝑤1 , …, 𝑤𝑚 is a list in 𝑊 such that
𝑣1 ⊗ 𝑤1 + ⋯ + 𝑣𝑚 ⊗ 𝑤𝑚 = 0.
Prove that 𝑤1 = ⋯ = 𝑤𝑚 = 0.
4 Suppose dim 𝑉 > 1 and dim 𝑊 > 1. Prove that
{𝑣 ⊗ 𝑤 ∶ (𝑣, 𝑤) ∈ 𝑉 × 𝑊}
is not a subspace of 𝑉 ⊗ 𝑊.
This exercise implies that if dim 𝑉 > 1 and dim 𝑊 > 1, then
{𝑣 ⊗ 𝑤 ∶ (𝑣, 𝑤) ∈ 𝑉 × 𝑊} ≠ 𝑉 ⊗ 𝑊.
5 Suppose 𝑚 and 𝑛 are positive integers. For 𝑣 ∈ 𝐅𝑚 and 𝑤 ∈ 𝐅𝑛 , identify

𝑣 ⊗ 𝑤 with an 𝑚-by-𝑛 matrix as in Example 9.76. With that identification,
show that the set
{𝑣 ⊗ 𝑤 ∶ 𝑣 ∈ 𝐅𝑚 and 𝑤 ∈ 𝐅𝑛 }
is the set of 𝑚-by-𝑛 matrices (with entries in 𝐅 ) that have rank at most one.
6 Suppose 𝑚 and 𝑛 are positive integers. Give a description, analogous to
Exercise 5, of the set of 𝑚-by-𝑛 matrices (with entries in 𝐅 ) that have rank
at most two.
7 Suppose dim 𝑉 > 2 and dim 𝑊 > 2. Prove that
{𝑣1 ⊗ 𝑤1 + 𝑣2 ⊗ 𝑤2 ∶ 𝑣1 , 𝑣2 ∈ 𝑉 and 𝑤1 , 𝑤2 ∈ 𝑊} ≠ 𝑉 ⊗ 𝑊.
8 Suppose 𝑣1 , …, 𝑣𝑚 ∈ 𝑉 and 𝑤1 , …, 𝑤𝑚 ∈ 𝑊 are such that
𝑣1 ⊗ 𝑤1 + ⋯ + 𝑣𝑚 ⊗ 𝑤𝑚 = 0.
Suppose that 𝑈 is a vector space and Γ ∶ 𝑉× 𝑊 → 𝑈 is a bilinear map. Show

that
Γ(𝑣1 , 𝑤1 ) + ⋯ + Γ(𝑣𝑚 , 𝑤𝑚 ) = 0.
9 Suppose 𝑆 ∈ ℒ(𝑉) and 𝑇 ∈ ℒ(𝑊). Prove that there exists a unique operator
on 𝑉 ⊗ 𝑊 that takes 𝑣 ⊗ 𝑤 to 𝑆𝑣 ⊗ 𝑇𝑤 for all 𝑣 ∈ 𝑉 and 𝑤 ∈ 𝑊.
In an abuse of notation, the operator on 𝑉 ⊗ 𝑊 given by this exercise is
often called 𝑆 ⊗ 𝑇.
10 Suppose 𝑆 ∈ ℒ(𝑉) and 𝑇 ∈ ℒ(𝑊). Prove that 𝑆⊗𝑇 is an invertible operator

on 𝑉 ⊗ 𝑊 if and only if both 𝑆 and 𝑇 are invertible operators. Also, prove
that if both 𝑆 and 𝑇 are invertible operators, then (𝑆 ⊗ 𝑇)−1 = 𝑆−1 ⊗ 𝑇 −1 ,
where we are using the notation from the comment after Exercise 9.
11 Suppose 𝑉 and 𝑊 are inner product spaces. Prove that if 𝑆 ∈ ℒ(𝑉) and
𝑇 ∈ ℒ(𝑊), then (𝑆 ⊗ 𝑇)∗ = 𝑆∗ ⊗ 𝑇 ∗, where we are using the notation from
the comment after Exercise 9.
12 Suppose that 𝑉1 , …, 𝑉𝑚 are finite-dimensional inner product spaces. Prove
that there is a unique inner product on 𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 such that
⟨𝑣1 ⊗ ⋯ ⊗ 𝑣𝑚 , 𝑢1 ⊗ ⋯ ⊗ 𝑢𝑚 ⟩ = ⟨𝑣1 , 𝑢1 ⟩⋯⟨𝑣𝑚 , 𝑢𝑚 ⟩
for all (𝑣1 , …, 𝑣𝑚 ) and (𝑢1 , …, 𝑢𝑚 ) in 𝑉1 × ⋯ × 𝑉𝑚 .

Note that the equation above implies that
‖𝑣1 ⊗ ⋯ ⊗ 𝑣𝑚 ‖ = ‖𝑣1 ‖ × ⋯ × ‖𝑣𝑚 ‖
for all (𝑣1 , …, 𝑣𝑚 ) ∈ 𝑉1 × ⋯ × 𝑉𝑚 .

13 Suppose that 𝑉1 , …, 𝑉𝑚 are finite-dimensional inner product spaces and

𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 is made into an inner product space using the inner product
from Exercise 12. Suppose 𝑒1𝑘 , …, 𝑒𝑛𝑘𝑘 is an orthonormal basis of 𝑉𝑘 for each
𝑘 = 1, …, 𝑚. Show that the list
{𝑒𝑗11 ⊗ ⋯ ⊗ 𝑒𝑗𝑚𝑚 }𝑗1 = 1, …, 𝑛1; ⋯; 𝑗𝑚 = 1, …, 𝑛𝑚
is an orthonormal basis of 𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 .
Photo Credits
• page v: Photos by Carrie Heeter and Bishnu Sarangi. Public domain image.
• page 1: Original painting by Pierre Louis Dumesnil; 1884 copy by Nils
Forsberg. Public domain image downloaded on 29 March 2022 from
https://commons.wikimedia.org/wiki/File:René_Descartes_i_samtal_med_Sveriges_drottning,_Kristina.jpg.
• page 27: Public domain image downloaded on 4 February 2022 from

https://commons.wikimedia.org/wiki/File:IAS_Princeton.jpg.
• page 51: Photo by Stefan Schäfer; Creative Commons Attribution Share Alike
license. Downloaded on 28 March 2022 from
https://commons.wikimedia.org/wiki/File:BurgDankwarderode2016.jpg.
• page 119: Photo by Alireza Javaheri. Creative Commons Attribution license.
Downloaded on 12 March 2023 from
https://commons.wikimedia.org/wiki/File:Hakim_Omar_Khayam_-_panoramio.jpg.
• page 132: Statue completed by Giovanni Paganucci in 1863. Photo by
Hans-Peter Postel; Creative Commons Attribution license. Downloaded on 14
March 2022 from
https://commons.wikimedia.org/wiki/File:Leonardo_da_Pisa.jpg.
• page 181: Photo by Matthew Petroff; Creative Commons Attribution Share
Alike license. Downloaded on 31 March 2022 from
https://commons.wikimedia.org/wiki/File:George-peabody-library.jpg.
• page 227: Photo by Petar Milošević; Creative Commons Attribution Share
Alike license. Downloaded on 30 March 2022 from
https://en.wikipedia.org/wiki/Lviv.
• page 297: Photo by David Iliff; Creative Commons Attribution Share Alike
license. Downloaded on 30 March 2022 from
https://en.wikipedia.org/wiki/File:Long_Room_Interior,_Trinity_College_Dublin,_Ireland_-_Diliff.jpg.
• page 332: Photo by Daniel Schwen; Creative Commons Attribution Share

Alike license. Downloaded on 9 July 2019 from
https://commons.wikimedia.org/wiki/File:Mathematik_Göttingen.jpg.

383
https://doi.org/10.1007/978-3-031-41026-0
Symbol Index
𝐴−1, 91 𝐼, 52, 90 𝑇 †, 221

𝐴𝑗, ⋅ , 74 ⟺ , 23 𝑇 𝑚, 137
𝐴𝑗, 𝑘 , 69 Im, 120 ‖𝑇‖, 280
𝐴⋅, 𝑘 , 74 −∞, 31 𝑇 #, 375, 380
𝛼𝑇 , 354 tr 𝐴, 326
𝐴∗, 231
ℒ(𝑉), 52 tr 𝑇, 327
ℒ(𝑉, 𝑊), 52 𝑇|𝑈 , 133
𝐴t , 77
𝑇/𝑈, 142
ℳ(𝛽), 334
Γ̂, 375, 380
ℳ(𝑇), 69, 154 𝑈 ⟂ , 211
𝐵, 287
ℳ(𝑣), 88 𝑈 0 , 109
ℬ(𝑉1 , …, 𝑉𝑚 ), 378
⟨𝑢, 𝑣⟩, 184
ℬ(𝑉, 𝑊), 370 perm, 348
𝒫(𝐅), 30 𝑉, 15
𝐂, 2 𝑉 ′, 105, 204
𝜋, 101
∘, 55 𝑉/𝑈, 99
𝒫𝑚 (𝐅), 31
𝑝(𝑇), 137 −𝑣, 15
deg, 31
𝑃𝑈 , 214 𝑉1 ⊗ ⋯ ⊗ 𝑉𝑚 , 379
Δ, 196
𝑣1 ⊗ ⋯ ⊗ 𝑣𝑚 , 379
det 𝐴, 355 𝑞𝛽 , 341 𝑉 (2), 334
det 𝑇, 354 ,7 (2)
𝑉alt , 339
dim, 44
𝑉sym , 337
(2)
⊕, 21 𝐑, 2
𝑉𝐂 , 17
Re, 120
𝑉 𝑚, 103, 346
𝐸(𝑠1 𝑓1 , …, 𝑠𝑛 𝑓𝑛 ), 287
𝑉 (𝑚), 346
𝐸(𝜆, 𝑇), 164 𝑆 ⊗ 𝑇, 381 (𝑚)
𝑉alt , 347
⊊, 299
𝐅, 4 𝑉 ⊗ 𝑊, 372
𝐅∞, 13 √𝑇, 253 𝑣 ⊗ 𝑤, 372
𝐅𝑚, 𝑛, 72
̃, 102
𝑇 𝑣 + 𝑈, 98
𝐅𝑛, 6 𝑇 ′, 107
||𝑣||, 186
𝐅𝑆, 13 𝑇 ∗, 228
𝑇 −1, 82 𝑧, 120
𝐺(𝜆, 𝑇), 308 𝑇(Ω), 288 |𝑧|, 120
384
https://doi.org/10.1007/978-3-031-41026-0
Index
Abbott, Edwin A., 6 bilinear form, 333

absolute value, 120 bilinear functional, 370
addition bilinear map, 374
in quotient space, 100 block diagonal matrix, 314
of complex numbers, 2 Bunyakovsky, Viktor, 189
of functions, 13
of linear maps, 55 𝐶∗ -algebras, 295
of matrices, 71 Carroll, Lewis, 11
of subspaces, 19 Cauchy, Augustin-Louis, 189
of vectors, 12 Cauchy–Schwarz inequality, 189
of vectors in 𝐅𝑛 , 6 Cayley, Arthur, 312
additive inverse Cayley–Hamilton theorem, 364
in 𝐂 , 3, 4 on complex vector space, 312
in 𝐅𝑛 , 9 change-of-basis formula
in vector space, 12, 15 for bilinear forms, 336
additivity, 52 for operators, 93
adjoint of a linear map, 228 characteristic polynomial, 363
algebraic multiplicity, 311 on complex vector space, 311
alternating bilinear form, 339 ChatGPT, 196, 279
alternating 𝑚-linear form, 347 Cholesky factorization, 267
annihilator, 109 Cholesky, André-Louis, 267
Apollonius’s identity, 195 Christina, Queen of Sweden, 1
Artin, Emil, 80 closed under addition, 18
associativity, 3, 12, 56 closed under scalar multiplication, 18
column rank of a matrix, 77, 114, 239
backward shift, 53, 59, 84, 140 column–row factorization, 78
ball, 287 commutativity, 3, 7, 12, 25, 56, 73, 80
Banach, Stefan, 227 commuting operators, 138, 175–180,
basis, 39 209, 235, 248–249, 256
of eigenvectors, 165, 245, 246, companion matrix, 152
250 complex conjugate, 120
of generalized eigenvectors, 301 complex number, 2
Bernstein polynomials, 49 complex spectral theorem, 246
Bessel’s inequality, 198 complex vector space, 13
385
https://doi.org/10.1007/978-3-031-41026-0
386 Index
complexification double dual space, 118

eigenvalues of, 140 dual
generalized eigenvectors of, 318 of a basis, 106
minimal polynomial of, 153 of a linear map, 107, 153, 162,
multiplicity of eigenvalues, 318 174
of a linear map, 68 of a linear operator, 140
of a vector space, 17, 43 of a vector space, 105, 204
of an inner product space, 194
conjugate symmetry, 183 eigenspace, 164
conjugate transpose of a matrix, 231 eigenvalue
coordinate, 6 of adjoint, 239
cube root of an operator, 248 of dual of linear operator, 140
of operator, 134
De Moivre’s theorem, 125 of positive operator, 252
degree of a polynomial, 31 of self-adjoint operator, 233
Descartes, René, 1 of unitary operator, 262
determinant on odd-dimensional space, 150,
of matrix, 355 318, 367
of operator, 354 eigenvector, 135
of positive operator, 362 ellipsoid, 287
of unitary operator, 362 Euclidean inner product, 184
diagonal matrix, 163, 274
diagonal of a square matrix, 155 Fibonacci, 132
diagonalizable, 163, 172, 176, 245, Fibonacci sequence, 174
246, 294, 307, 316 field, 10
differentiation linear map, 53, 56, 59, finite-dimensional vector space, 30
61, 62, 67, 70, 79, 138, 208, Flatland, 6
304 forward shift, 140
dimension, 44 Frankenstein, 50
of a sum of subspaces, 47 Frobenius norm, 331
direct sum, 21, 42, 98 Fuglede’s theorem, 248
of a subspace and its orthogonal fundamental theorem of algebra, 125
complement, 212 fundamental theorem of linear maps,
of null 𝑇 dim 𝑉 and range 𝑇 dim 𝑉 , 62
299
discrete Fourier transform, 269 Gauss, Carl Friedrich, 51
distributive property, 3, 12, 15, 56, 80 Gaussian elimination, 51, 65, 361
division algorithm for polynomials, generalized eigenspace, 308
124 generalized eigenvector, 300
division of complex numbers, 4 geometric multiplicity, 311
dot product, 182 Gershgorin disk, 170
Index 387
Gershgorin disk theorem, 171 Khayyam, Omar, 119

Gershgorin, Semyon Aronovich, 171
Gram, Jørgen, 200 Laplacian, 196
Gram–Schmidt procedure, 200 length of list, 5
graph of a linear map, 103 Leonardo of Pisa, 132
linear combination, 28
Hadamard’s inequality, 365 linear dependence lemma, 33
Halmos, Paul, 27 linear equations, 64–65, 95
Hamilton, William, 297 linear functional, 105, 204
harmonic function, 196 linear map, 52
Hilbert matrix, 256 linear map lemma, 54
Hilbert–Schmidt norm, 331 linear span, 29
homogeneity, 52 linear subspace, 18
homogeneous system of linear linear transformation, 52
equations, 65, 95 linearly dependent, 33
hyponormal operator, 241 linearly independent, 32
list, 5
identity matrix, 90
of vectors, 28
identity operator, 52, 56
lower-triangular matrix, 162, 267
imaginary part, 120
Lviv, 227
infinite-dimensional vector space, 31
Lwów, 227
inhomogeneous system of linear
equations, 65, 95 matrix, 69
injective, 60 multiplication, 73
inner product, 183 of bilinear form, 334
inner product space, 184 of linear map, 69
Institute for Advanced Study, 27 of nilpotent operator, 305
invariant subspace, 133 of operator, 154
inverse of product of linear maps, 74, 91
of a linear map, 82 of 𝑇 ′ , 113
of a matrix, 91 of 𝑇 ∗, 232
invertible linear map, 82 of vector, 88
invertible matrix, 91 minimal polynomial
isometry, 258 and basis of generalized
isomorphic vector spaces, 86 eigenvectors, 306
isomorphism, 86 and characteristic polynomial, 312
Jordan basis, 322 and diagonalizability, 169
Jordan form, 324 and generalized eigenspace
Jordan, Camille, 324 decomposition, 316
and generalized eigenspaces, 317
kernel, 59 and invertibility, 149
388 Index
and upper-triangular matrices, orthogonal

159, 203 complement, 211
computing, 145 projection, 214
definition of, 145 vectors, 187
gcd with its derivative, 173 orthonormal
no direct sum decomposition, 325 basis, 199
of adjoint, 241 list, 197
of companion matrix, 152
of complexification, 153 parallelogram equality, 191
of dual map, 153 Parseval’s identity, 200
of nilpotent operator, 305, 324 partial differentiation operator, 175
of normal operator, 241 Peabody Library, 181
of quotient operator, 153 permutation, 348
of restriction operator, 148 photo credits, 383
of self-adjoint operator, 244 point, 12
polynomial multiple of, 148 polar decomposition, 286
zeros of, 146 polynomial, 30
minimizing distance, 217 positive definite, 266
𝑚-linear form, 346 positive operator, 251
𝑚-linear functional, 378 positive semidefinite operator, 251
𝑚-linear map, 379 principal axes, 287
monic polynomial, 144 product
Moon, v, xvii of complex numbers, 2
Moore–Penrose inverse, 221 of linear maps, 55
multilinear form, 346 of matrices, 73
multiplication, see product of polynomials, 138
multiplicity of an eigenvalue, 310 of scalar and linear map, 55
of scalar and vector, 12
nilpotent operator, 303, 322 of scalar and vector in 𝐅𝑛 , 9
Noether, Emmy, 332 of vector spaces, 96
nonsingular matrix, 91 pseudoinverse, 221, 250, 255, 275, 279
norm, 182, 186 Pythagorean theorem, 187
normal operator, 235
null space, 59 QR factorization, 264, 365
of powers of an operator, 298 quadratic form, 341
of 𝑇 ′, 111 quotient
of 𝑇 ∗, 231 map, 101
operator, 142, 153, 162, 173
one-to-one, 60 space, 99
onto, 62
operator, 133 range, 61
Index 389
of powers of an operator, 306 square root of an operator, 248, 251,

of 𝑇 ′, 112 253, 320
of 𝑇 ∗, 231 standard basis
rank of a matrix, 79, 114, 239 of 𝐅𝑛 , 39
real part, 120 of 𝒫𝑚 (𝐅), 39
real spectral theorem, 245 subspace, 18
real vector space, 13 subtraction of complex numbers, 4
reverse triangle inequality, 129, 193, sum, see addition
294 sum of subspaces, 19
Riesz representation theorem, 205, 210, Supreme Court, 210
216, 224, 225 surjective, 62
Riesz, Frigyes, 205 SVD, see singular value decomposition
row rank of a matrix, 77, 114, 239 Sylvester, James, 181
symmetric bilinear form, 337
scalar, 4 symmetric matrix, 269, 337
scalar multiplication, 9, 12
tensor product, 372, 379
in quotient space, 100
Through the Looking Glass, 11
of linear maps, 55
trace
of matrices, 71
of a matrix, 326
Schmidt pair, 278
of an operator, 327
Schmidt, Erhard, 200, 278
translate, 99
Schur’s theorem, 204
transpose of a matrix, 77, 231
Schur, Issai, 204
triangle inequality, 121, 190, 281
Schwarz, Hermann, 189 tuple, 5
self-adjoint operator, 233 two-sided ideal, 58
Shelley, Mary Wollstonecraft, 50
sign of a permutation, 349 unit circle in 𝐂 , 262, 269
simultaneous diagonalization, 176 unitary matrix, 263
simultaneously upper triangularizable, unitary operator, 260
178 University of Dublin, 297
singular matrix, 91 University of Göttingen, 332
singular value decomposition upper-triangular matrix, 155–160, 264,
of adjoint, 275 267, 314
of linear map, 273
Vandermonde matrix, 366
of pseudoinverse, 275
vector, 8, 12
singular values, 271, 362
vector space, 12
skew operator, 240, 247, 269
volume, 292, 363
span, 29
of a box, 291
spans, 29
spectral theorem, 245, 246 zero of a polynomial, 122
Colophon: Notes on Typesetting
• This book was typeset in LuaLATEX by the author, who wrote the LATEX code to
implement the book’s design.
• The LATEX software used for this book was written by Leslie Lamport. The TEX
software, which forms the base for LATEX, was written by Donald Knuth.
• The main text font in this book is the Open Type Format version of TEX Gyre
Termes, a font based on Times, which was designed by Stanley Morison and
Victor Lardent for the British newspaper The Times in 1931.
• The main math font in this book is the Open Type Format version of TEX Gyre
Pagella Math, a font based on Palatino, which was designed by Hermann Zapf.
• The sans serif font used for page headings and some other design elements is
the Open Type Format version of TEX Gyre Heros, a font based on Helvetica,
which was designed by Max Miedinger and Eduard Hoffmann.
• The LuaLATEX packages fontspec and unicode-math, both written by Will
Robertson, were used to manage fonts.
• The LATEX package fontsize, written by Ivan Valbusa, was used to gracefully
change the main fonts to 10.5 point size.
• The figures in the book were produced by Mathematica, using Mathematica
code written by the author. Mathematica was created by Stephen Wolfram.
The Mathematica package MaTeX, written by Szabolcs Horvát, was used to
place LATEX-generated labels in the Mathematica figures.
• The LATEX package graphicx, written by David Carlisle and Sebastian Rahtz,
was used to include photos and figures.
• The LATEX package multicol, written by Frank Mittelbach, was used to get
around LATEX’s limitation that two-column format must start on a new page
(needed for the Symbol Index and the Index).
• The LATEX packages TikZ, written by Till Tantau, and tcolorbox, written by
Thomas Sturm, were used to produce the definition boxes and result boxes.
• The LATEX package color, written by David Carlisle, was used to add appropriate
color to various design elements.
• The LATEX package wrapfig, written by Donald Arseneau, was used to wrap
text around the comment boxes.
• The LATEX package microtype, written by Robert Schlicht, was used to reduce
hyphenation and produce more pleasing right justification.
390
https://doi.org/10.1007/978-3-031-41026-0

Linear Algebra Done Right: Sheldon Axler

Uploaded by

Copyright:

Available Formats

Linear Algebra Done Right: Sheldon Axler

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Linear Algebra Done Right: Sheldon Axler

Uploaded by

Copyright:

Available Formats

Undergraduate Texts in Mathematics

Undergraduate Texts in Mathematics are generally aimed at third- and fourth-

Linear Algebra Done Right

ISSN 0172-6056 ISSN 2197-5604 (electronic)

Paper in this product is recyclable.

Sheldon Axler received his undergraduate degree from Princeton University,

The author and his cat Moon.

About the Author v

Preface for Students xii

Preface for Instructors xiii

Matrix of Dual of Linear Map 113

7F Consequences of Singular Value Decomposition 280

9B Alternating Multilinear Forms 346

Photo Credits 383

Symbol Index 384

Colophon: Notes on Typesetting 390

The author’s top ten

Major improvements and additions for the fourth edition

• Over 250 new exercises and over 70 new examples.

Linear algebra is the study of linear maps on finite-dimensional vector spaces.

René Descartes explaining his work to Queen Christina of Sweden.

1.1 definition: complex numbers, 𝐂

• A complex number is an ordered pair (𝑎, 𝑏), where 𝑎, 𝑏 ∈ 𝐑 , but we will

(𝑎 + 𝑏𝑖) + (𝑐 + 𝑑𝑖) = (𝑎 + 𝑐) + (𝑏 + 𝑑)𝑖,

If 𝑎 ∈ 𝐑 , we identify 𝑎 + 0𝑖 with the real number 𝑎. Thus we think of 𝐑 as a

1.2 example: complex arithmetic

1.3 properties of complex arithmetic

1.4 example: commutativity of complex multiplication

where 𝑎, 𝑏, 𝑐, 𝑑 ∈ 𝐑 . Then the definition of multiplication of complex numbers

Next, we define the additive and multiplicative inverses of complex numbers,

1.5 definition: −𝛼, subtraction, 1/𝛼, division

• For 𝛼 ≠ 0, let 1/𝛼 anddenote the multiplicative inverse of 𝛼. Thus 1/𝛼 is

• For 𝛼 ≠ 0, division by 𝛼 is defined by

Throughout this book, 𝐅 stands for either 𝐑 or 𝐂 .

Thus if we prove a theorem involving The letter 𝐅 is used because 𝐑 and 𝐂

This definition implies that

1.7 example: 𝐑2 and 𝐑3

To generalize 𝐑2 and 𝐑3 to higher dimensions, we first need to discuss the

1.8 definition: list, length

• Suppose 𝑛 is a nonnegative integer. A list of length 𝑛 is an ordered collec-

Lists are often written as elements Many mathematicians call a list of

1.9 example: lists versus sets

Fix a positive integer 𝑛 for the rest of this chapter.

1.11 definition: 𝐅𝑛 , coordinate

𝐅𝑛 is the set of all lists of length 𝑛 of elements of 𝐅 :

𝐅𝑛 = {(𝑥1 , …, 𝑥𝑛 ) ∶ 𝑥𝑘 ∈ 𝐅 for 𝑘 = 1, …, 𝑛}.

For (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛 and 𝑘 ∈ {1, …, 𝑛}, we say that 𝑥𝑘 is the 𝑘 th coordinate of

If 𝐅 = 𝐑 and 𝑛 equals 2 or 3, then the definition above of 𝐅𝑛 agrees with our

𝐂4 is the set of all lists of four complex numbers:

If 𝑛 ≥ 4, we cannot visualize 𝐑𝑛 as Read Flatland: A Romance of Many

1.13 definition: addition in 𝐅𝑛

(𝑥1 , …, 𝑥𝑛 ) + (𝑦1 , …, 𝑦𝑛 ) = (𝑥1 + 𝑦1 , …, 𝑥𝑛 + 𝑦𝑛 ).

Often the mathematics of 𝐅𝑛 becomes cleaner if we use a single letter to denote

1.14 commutativity of addition in 𝐅𝑛

Proof Suppose 𝑥 = (𝑥1 , …, 𝑥𝑛 ) ∈ 𝐅𝑛 and 𝑦 = (𝑦1 , …, 𝑦𝑛 ) ∈ 𝐅𝑛. Then

If a single letter is used to denote an The symbol means “end of proof ”.

Let 0 denote the list of length 𝑛 whose coordinates are all 0:

1.16 example: context determines which 0 is intended

𝑥+0=𝑥 for all 𝑥 ∈ 𝐅𝑛.