Basic

Classes of
Linear
Operators
Israel Gohberg
5eymour Goldberg
Marinus A. Kaashoek

Springer Basel AG
Israel Gohberg Marinus A. Kaashoek
School of Mathematical Sciences Department of Mathematics and Computer Science
Raymond and Beverly Sackler Vrije Universiteit Amsterdam
Faculty of Exact Sciences De Boelelaan 1081a
Tel Aviv University NL-1081 HV Amsterdam
Ramat Aviv 69978 The Netherlands
Israel
e-mail: [email protected] e-mail: [email protected]

Seymour Goldberg
Department of Mathematics
University of Maryland
College Park, MD 20742-4015
USA
e-mail: [email protected]

2000 Mathematics Subject Classification 47-01

A CIP catalogue record for this book is available from the

Ubrary of Congress, Washington D.C., USA

Bibliographic information published by Die Deutsche Bibliothek

Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data is available in the Internet at <http://dnb.ddb.de>.

ISBN 978-3-7643-6930-9 ISBN 978-3-0348-7980-4 (eBook)

DOI 10.1007/978-3-0348-7980-4

This work is subject to copyright. AII rights are reserved, whether the whole or part of the material
is concerned, specifically the rights of translation, reprinting, re-use of iIIustrations, recitation,
broadcasting, reproduction on microfilms or in other ways, and storage in data banks. For any
kind of use permission of the copyright owner must be obtained.

© 2003 Springer Basel AG

Originally published by Birkhăuser Verlag, Basel, Switzerland in 2003
Printed on acid-free paper produced from chlorine-free pulp. TCF ca
Cover design: Micha Lotrovsky, CH-4106 Therwil, Switzerland

ISBN 978-3-7643-6930-9
www.birkhiiuser-science.com
987654321
Dedicated to our grandchildren
Table of Contents

Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . •. . . . . •. . . . . . . . . . . . .. xv
Chapter I Hilbert Spaces. . . . . . . . . . . . . . . . . . . . •. . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Complex n-Space 1
1.2 The Hilbert Space Zj 3
1.3 Definition of Hilbert Space and its Elementary Properties 5
1.4 Distance from a Point to a Finite Dimensional Space 8
1.5 The Gram Determinant 10
1.6 Incompatible Systems of Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.7 Least Square Fit 15
1.8 Distance to a Convex Set and Projections onto Subspaces 16
1.9 Orthonormal Systems 18
1.10 Szego Polynomials 19
1.11 Legendre Polynomials 24
1.12 Orthonormal Bases 26
1.13 Fourier Series 29
1.14 Completeness of the Legendre Polynomials 31
1.15 Bases for the Hilbert Space of Functions on a Square 32
1.16 Stability of Orthonormal Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.17 Separable Spaces 35
1.18 Isometry of Hilbert Spaces 36
1.19 Example of a Non Separable Space 38
Exercises 38

Chapter II Bounded Linear Operators on Hilbert Spaces . . . . . . . . . . . . . 51

2.1 Properties of Bounded Linear Operators 51
2.2 Examples of Bounded Linear Operators with Estimates ofNonns 52
2.3 Continuity of a Linear Operator 56
2.4 Matrix Representations of Bounded Linear Operators 57
2.5 Bounded Linear Functionals 60
2.6 Operators of Finite Rank 63
2.7 Invertible Operators 64
2.8 Inversion of Operators by the Iterative Method . . . . . . . . . . . . . . . . . . . . . 69
2.9 Infinite Systems of Linear Equations 71
2.10 Integral Equations of the Second Kind 73
viii Table ofContents

2.11 Adjoint Operators 76

2.12 Self Adjoint Operators 80
2.13 Orthogonal Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
2.14 Two Fundamental Theorems 82
2.15 Projections and One-Sided Invertibility of Operators 84
2.16 Compact Operators 91
2.17 The Projection Method for Inversion of Linear Operators 96
2.18 The Modified Projection Method 105
2.19 Invariant Subspaces 108
2.20 The Spectrum of an Operator 109
Exercises 118
Chapter III Laurent and Toeplitz Operators on Hilbert Spaces. . . . . . 135

3.1 Laurent Operators , 135

3.2 Toeplitz Operators .. . . .. . . . . . . . .. . .. . . . . . .. . . .. . . . . . . . .. . . . . . . . 141
3.3 Band Toep1itz operators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
3.4 Toeplitz Operators with Continuous Symbols 152
3.5 Finite Section Method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . .. 159
3.6 The Finite Section Method for Laurent Operators 163
Exercises 166
Chapter IV Spectral Theory of Compact Self Adjoint Operators. . . .. 171

4.1 Example of an Infinite Dimensional Generalization .. . . . . . . . . . . . . . . 171

4.2 The Problem of Existence of Eigenvalues and Eigenvectors . . . . . . . . . 172
4.3 Eigenvalues and Eigenvectors of Operators of Finite Rank . . . . . . . . .. 174
4.4 Existence of Eigenvalues . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . .. . . . . . . . . 175
4.5 Spectral Theorem 178
4.6 Basic Systems of Eigenvalues and Eigenvectors 180
4.7 Second Form of the Spectral Theorem 182
4.8 Formula for the Inverse Operator 183
4.9 Minimum-Maximum Properties of Eigenvalues 185
Exercises 188
Chapter V Spectral Theory of Integral Operators. . . . . . . . . . . . . . . . . .. 193

5.1 Hilbert-Schmidt Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

5.2 Preliminaries for Mercer's Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 196
5.3 Mercer's Theorem. . . . .. . .. . .. . .. . ... . . . . . . . . . . . .. . . . . . . . .. . .. . 197
5.4 Trace Formula for Integral Operators 200
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 200
Chapter VI Unbounded Operators on Hilbert Space. . . . . . . . . . . . . . .. 203

6.1 Closed Operators and First Examples 203

6.2 The Second Derivativeas an Operator 204
Table ofContents ix

6.3 The Graph Norm 206

6.4 Adjoint Operators 208
6.5 Sturm-Liouville Operators 211
6.6 Self Adjoint Operators with Compact Inverse . . . . . . . . . . . . . . . . . . . . . 214
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 215

Chapter VII Oscillations of an Elastic String....... 219

7.1 The Displacement Function 219

7.2 Basic Harmonic Oscillations 220
7.3 Harmonic Oscillations with an External Force " 222
Chapter VIII Operational Calculus with Applications. . . . . . . . . . . . . .. 225

8.1 Functions of a Compact Self Adjoint Operator " 225

8.2 Differential Equations in Hilbert Space 230
8.3 Infinite Systems of Differential Equations 232
8.4 Integro-Differential Equations 233
Exercises 234

Chapter IX Solving Linear Equations by Iterative Methods 237

9.1 The Main Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 237
9.2 Preliminaries for the Proof " 238
9.3 Proof of the Main Theorem 240
9.4 Application to Integral Equations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242

Chapter X Further Developments of the Spectral Theorem. . . . . . . . .. 243

10.1 Simultaneous Diagonalization 243

10.2 Compact Normal Operators 244
10.3 Unitary Operators .... .. .. ... .... ... .... ... ... ... .... . 246
10.4 Singular Values 248
10.5 Trace Class and Hilbert Schmidt Operators 253
Exercises 254
Chapter XI Banach Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
11 .1 Definitions and Examples 259
11.2 Finite Dimensional Normed Linear Spaces 262
11 .3 Separable Banach Spaces and Schauder Bases 264
11.4 Conjugate Spaces 265
11.5 Hahn-Banach Theorem 267
Exercises " 272
x Table ofContents

Chapter XII Linear Operators on a Banach Space . . . . . . . . . . . . . . . . . . 277

12.1 Description of Bounded Operators 277

12.2 Closed Linear Operators 279
12.3 Closed Graph Theorem 281
12.4 Applications of the Closed Graph Theorem 283
12.5 Complemented Subspaces and Projections 286
12.6 One-Sided Invertibility Revisited 288
12.7 The Projection Method Revisited . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . 289
12.8 The Spectrum of an Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
12.9 Volterra Integral Operator 293
12.10 Analytic Operator Valued Functions 295
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
Chapter XIII Compact Operators on a Banach Space. . . . . . . . . . . . . .. 299

13.1 Examples of Compact Operators 299

13.2 Decomposition of Operators of Finite Rank 302
13.3 Approximation by Operators of Finite Rank. . . . . . . . . . . . . . . . . . . .. 303
13.4 First Results in Fredholm Theory 305
13.5 Conjugate Operators on a Banach Space 306
13.6 Spectrum of a Compact Operator. . . . . . . . . . . . . . . . . . . . . .. . . . . . . .. 310
13.7 Applications 313
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
Chapter XIV Poincare Operators: Determinant and Trace 317
14.1 Determinant and Trace 317
14.2 Finite Rank Operators, Determinants and Traces 321
14.3 Theorems about the Poincare Determinant 327
14.4 Determinants and Inversion of Operators 330
14.5 Trace and Determinant Formulas for Poincare Operators . . . . . . . . .. 336
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
Chapter XV Fredholm Operators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 347

15.1 Definition and Examples 347

15.2 First Properties 347
15.3 Perturbations Small in Norm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
15.4 Compact Perturbations 355
15.5 Unbounded Fredholm Operators 356
Exercises 358
Table of Contents xi

Chapter XVI Toeplitz and Singular Integral Operators . . . . . . . . . . . . .. 361

16.1 Laurent Operators on i p (Z) 361
16.2 Toeplitz Operators on i p • . . . • . . . . . . • . . . • . . . . . .. . . . . . . . . . . . • . . . . 364
16.3 An Illustrative Example . . .. . . .. .... . . ... ... . . .. . . . . . . . . . . . . . . . 372
16.4 Applications to Pair Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 377
16.5 The Finite Section Method Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
16.6 Singular Integral Operators on the Unit Circle 390
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 395

Chapter XVII Non Linear Operators.. . .. . . 401

17.1 Fixed Point Theorems 401

17.2 Applications of the Contraction Mapping Theorem 402
17.3 Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405

Appendix 1: Countable sets and Separable Hilbert Spaces . . . . . . . . . . . . 409

Appendix 2: The Lebesgue integral and L p Spaces 411

Suggested Reading 415

References 417
List of Symbols 419

Index 421
Preface

The present book is an expanded and enriched version of the text Basic Operator
Theory, written by the first two authors more than twenty years ago. Since then
the three of us have used the basic operator theory text in various courses . This
experience motivated us to update and improve the old text by including a wider
variety of basic classes of operators and their applications. The present book has
also been written in such a way that it can serve as an introduction to our previous
books Classes ofLinear Operators, Volumes I and II. We view the three books as
a unit.
We gratefully acknowledge the support of the mathematical departments of
Tel-Aviv University, the University of Maryland at College Park, and the Vrije
Universiteit at Amsterdam. The generous support ofthe Silver Family Foundation
is highly appreciated.

Amsterdam, November 2002 The authors

Introduction

This elementary text is an introduction to functional analysis, with a strong

emphasis on operator theory and its applications. It is designed for graduate and
senior undergraduate students in mathematics, science, engineering, and other
fields.
The discussion below describes the main aims of the book. This is the same
path we followed in our book Basic Operator Theory which was published more
than twenty years ago. The present book differs substantially from the above men-
tioned one in a number of ways. First, the examples of concrete classes of linear
operators which we study in this book are greatly enriched. For example, it con-
tains the theory of Laurent, Toeplitz and singular integral operators . This book
also presents the theory of traces and determinants in an infinite dimensional set-
ting. Fredholm theory of operators is studied here and the theory of unbounded
operators is expanded. The number of exercises and examples is also increased .
From the beginning of this book to its end, a great deal of attention is paid to
the interplay between abstract methods and concrete problems . Motivation of the
theory and its systematic applications are characteristic of all chapters.
Our aim has been to present the material in a form which the reader can under-
stand and appreciate without too much difficulty. To accomplish this end, we have
not stated the principal results in their most general form. Instead, statements were
chosen which make more transparent the main ideas and the further generalizations
as well.
The book covers only a limited number of topics, but they are sufficient to lay a
firm foundation in operator theory and to demonstrate the power of the theory in
applications. Much of the material in this volume is an essential portion oftraining
in mathematics. It is presented as a natural continuation of the undergraduate
courses in linear algebra and analysis .
In the past, graduate courses in integral equations were usually offered as part
of the mathematics curriculum. Later, some instructors thought to develop the
courses as special cases of results in functional analysis. However, this goal was
not realized and courses in integral equations almost disappeared. One of our aims
is to reestablish a theory ofintegral equations as a significant part ofoperator theory.
A unified approach to some phases of differential equations, approximation theory
and numerical analysis is also provided .
One of the main problems in operator theory is the problem of inverting linear
operators on infinite dimensional spaces. This is achieved by means of projection
methods and the theory of determinants. The projection methods appear in many
chapters of the book.
xvi Introduction

A theme which also appears throughout this book is the use of one sided
invertible operators . For example they appear in the theory of Fredholm, Toeplitz
and singular integral operators.
This book consists, basically, oftwo unequal parts. The major portion ofthe text
is devoted to the theory and applications of linear operators on a Hilbert space.
We begin with a chapter on the geometry of Hilbert space and then proceed to
the theory of linear operators on these spaces . The theory is richly illustrated by
numerous examples.
In the next chapter we study Laurent and Toeplitz operators which provide
further illustrations and applications of the general theory. Here the projection
methods and one sided invertibility play an important role.
Spectral theory of compact self adjoint operators appears in the next chapter
followed by a spectral theory of integral operators . Operational calculus is then
presented as a natural outgrowth of spectral theory. The reader is also shown the
basic theory of unbounded operators on Hilbert space.
The second part of the text concentrates on Banach spaces and linear operators,
both bounded and unbounded, acting on these spaces . It includes, for example, the
three basic principles oflinear analysis and the Riesz-Fredholm theory ofcompact
operators which is based on the theory of one sided invertible operators .
In later chapters we deal with a class of operators, in an infinite dimensional
setting, on which a trace and determinant are defined as extensions of the trace
and determinant of finite matrices. Further developments of Toeplitz operators
and singular integral operators are presented. Projection methods and one sided
invertible operators continue to play an important role. Both parts of the book
contain plenty of applications and examples . All chapters deal exclusively with
linear problems except for the last chapter which is an introduction to the theory
of non linear operators .
In general, in writing this book, the authors were strongly influenced by recent
developments in operator theory which affected the choice of topics, proofs and
exercises .
One ofthe main features ofthis book is the large number ofnew exercises chosen
to expand the reader's comprehension of the material , and to train him or her in
the use of it. In the beginning portion of the book we offer a large selection of
computational exercises; later, the proportion of exercises dealing with theoretical
questions increases . We have, however, omitted exercises after Chapters VII, IX
and XVII due to the specialized nature of the subject matter.
To reach as large an audience as possible, the material is self-contained. Begin-
ning with finite dimensional vector spaces and matrices, the theory is gradually
developed.
Introduction xvii

Since the book contains more material than is needed for a one semester course,
we leave the choice of subject matter to the instructor and the interests of the
students .
It is assumed that the reader is familiar with linear algebra and advanced cal-
culus; some exposure to Lebesgue integration is also helpful. However, any lack
of knowledge of this last topic can be compensated for by referring to Appendix
2 and the references therein.
It has been our experience that exposure to this material stimulated our students
to expand their knowledge ofoperator theory. We hope that this will continue to be
the case. Included in the book is a list of suggested reading material which should
prove helpful.
Chapter I
Hilbert Spaces

In this chapter we review the main properties of the complex n-dimensional space
en and then we study the Hilbert space which is its most natural infinite dimen-
sional generalization. Many applications to classical problems are included (Least
squares, Fourier series and others) .

1.1 Complex n-Space

Let C denote the set of complex numbers and let H = en

be the set of all n-tuples
(~I, . . . , ~n) of complex numbers with addition and scalar multiplication defined
as follows.
For x = (~I , . .. , ~n) and y = (r]I , . . . , r]n),
x +y = (~I + r]1, . .. , ~n + r]n) and a x = (a~l , . . . , a~n), a E C.
The complex valued function <, >, defined on the Cartesian product H x H by
n
(x, y) = L~;ij;
;=1

is called an inner product on H .

It is clear from the definitions that if x , y and z are in H, then

(i) (x, x) > 0 if and only if x =1= 0;

(ii) (x , y) = (y , x) ;
(iii) (ax , y) = a(x , y} ;
(iv) (x + y , z) = (x, z) + (y , z).
It follows from (ii), (iii) and (iv) that

(x, ay + {3z} = &(x , y) + ~(x , z); (x ,O) = (0, x) = O.

Each n-tuple x = (XI, . . . , x n ) is called a vector. The length or norm , [x], of

the vector x is defined by

n ) 1/2
IIxll = (x, x}1 /2 = { ; Ix;l2
(
2 Chapter I. Hilbert Spaces

If x = (XI, X2) and Y = (YI, Y2) are vectors in the plane, then (x , y) = IIx 1111 Y II
cos e, where e is the angle between x andy, 0 S e S Jr. Thus I(x, y)1 S IIxllllyll.
Based only on the properties of the inner product, we shall prove this inequality
in en.
Cauchy-Schwarz Inequality: If x and yare in en, then
I(x, y)1 S IIxllllyll· (1.1)

When y i= 0, equality holds if and only if x = Ay for some A E C.

Proof: For any A E C, it follows from (i}-(iv) that

Os (x - AY,x - Ay) = IIxll 2 - X(x, y) - A(Y, x) + AXllyll2

= IIxll 2 - 2 Re A(Y,X) + IAI 2I1YII 2. (1.2)

If (x, y) = 0, the inequality is trivial. Suppose (x, y) i= O. Substituting

2
A = (yIIxII.x ) in (1.2) gives

from which (1.1) follows.

Suppose y i= O. If (y , x) = 0, then equality holds in (1.1) if and only if x = O.
If (y , x) i= 0, then equality in (1.1) holds if and only if equality in (1.2) holds for
A = (IIyX,xII ) , in which case x = Ay.
2

As a consequence of the Cauchy-Schwarz inequality applied to x = (lXII, . .. ,

IXn I) and y = (IYI J, . . . , IYn I), we have

n ( n 2 ) 1/2 ( n 2 ) 1/2
{ ; IXiYi I S {; IXi 1 { ; IYi 1 (1.3)

The following inequality, when confined to vectors in the plane, states that the
sum of the lengths of two sides of a triangle exceeds the length of the third side.
D
Triangle inequality: For x and Y in en,
IIx + yll S IIxll + IIYII· (1.4)

Proof: By the Cauchy-Schwarz inequality,

2 2
IIx + Y 11 = (x + y, x + y) = IIx 11 + 2 Re (x , y) + II Y 11
2
2 2
S IIxll +211x1l1lYIl + IIYII = (lIxll + II Y11)2 .

D
1.2 The Hilbert Space £2 3

1.2 The Hilbert Space £2

The pas sage from en, the space of n-tuples (~I , . . . , ~n) of com plex numbers, to
the spa ce consisting of certain infinite sequences (~l , ~2 , .. .) of complex numbers
is a very natural one.
If x = (~I , ~2, . ..) and Y = (1/1, 1/2, . ..), define x + Y = (~l + 1/1. ~2 + 1/2, .. .)
and a x = (a~ l, a b . . .) for a E Co The inner product (x , y ) = L~I ~i iji and
the corresponding norm [x ] = (x , x }1 /2 = (L~I 1~i12)1 /2 make sense provided
L~ I 1~i12 < 00 and L~ I 11/i1 2 < 00. Indeed, by ( 1.3) in Section I ,

and by the triangle inequality (1.4 ) in Section I,

00
2) 1/2 ( 00
2) 1/2 ( 00
2) 1/2
( f; I~i + n. 1 S f; I~i 1 + f ; l1/i 1

Thus
I(x , y}1.s IIxli ll YIl and II x + YII s IIx ll + II YII .
Definition of £2: The vector spac e £2 consists of the set of those sequences
(~ I , ~2 , . ..) of complex numbers for which L~I 1~i1 2 < 00 , together with the
operatio ns of addition and scalar multiplication defined abo ve .
The complex valued function ( ,) defined on £2 x £2 by (x , y ) = L~I ~i iji ,
where x = (~d, Y = (1/i ), is called the inner product on £2. Clearly, the inner
product satisfies (i}-(iv) in Section I .
The norm, IIxll , on £2 is defin ed by

00 ) 1/2
II xll=(x, x} 1/2= f;1~i1 2
(

The distance between the vectors x = (~I , ~2 , . ..) and Y = (71 1, 1/2, . . .) is
IIx - yll = (L~II ~i -7Jil2) 1/2.
The extension of e n to £2 is not just an interesting exerc ise. As we shall see ,
£2 arises from the study of Fourier series , integral equations, infinite systems of
linear equations , and many appl ied problems.
In a treatment of advanced calculus it is shown that every Ca uchy sequence in
en converges, i.e., if {Xk} is a sequence ofvectors in en such that IIxm - Xn II -+ 0
as m , n -+ 00, then there exists an x E en such that II x n - x II -+ O. Th is enables
one to prove, for example, the convergence of certain series.
Th e space £2 has the same important prop erty which is called the completeness
property of £2.
4 Chapter I. Hilbert Spaces

Theorem 2.1 (Completeness of £2). If{x n } is a sequence ofvectors in £2 such

that IIx n - X m II -+ 0 as n , m -+ 00, then there exists an x E £2 such that
IIx n - xII -+ O.

Proof: Suppose X n = (~[n) , ~in) , . .. ) E £2 . The idea of the proof is to show that
for each fixed k, the sequence of components ~P ) , ~?) , ...
converges to some
number ~k and that x = (~l , ~2 , ...) is the desired limit of {x n }.
For k fixed,

Thus for each k, {~t)}~l is a Cauchy sequence of complex numbers which there-
fore converges. Take ~k = Iimn ---+ oo ~t) and x = (~l, ~2, . . .). First we show that
x is in £2.
For any positive integer j,

(2.2)

and
j
LI~t)12 ~ IIxn ll 2. (2.3)
k=l
Now sup, IIx n II = M < 00 since IlIxn II - IIx m III ~ IIx n - X m II -+ 0 as m,
n -+ 00. It therefore follows from (2.2) and (2.3) that L:;:l I~k 12 ~ M 2 , i.e., x is
in£2 .
Finally, we prove that IIx n - x II -+ O. Let e > 0 be given . There exists an
integer N such that if m , n ~ N, and p is any positive integer,
p
L I~t) - ~im)12 ~ II x n - x mll 2 ~ c. (2.4)
k=l

Fix n ~ N . From (2.1) and (2.4) we have that for all p,

Thus for n ~ N,
00

II x n - xII 2 = L I~t) - ~d ~ c.
k=l
Sometimes we shall also deal with £2(Z), the two-sided version of £2. The
vector space £2(Z) consists ofall doubly infinite sequences (... , ~-l, ~o , ~l, ...) of
1.3 Definition ofHilbert Space and its Elementary Properties 5

complex numbers that are square summable, that is, L~-oo I~j 12 < 00. As its
one-sided version £2, the space £2G~) is a vector space with addition and scalar
multiplication being defined entry wise . Also, £2(2) has a natural inner product,
namely (x, y) = L~-oo ~j ryj, where x = (~j) jEZ and y = (T/j) j EZ , The
corresponding norm is given by

and Theorem 1.2.1 above remains valid with £2(2) in place of £2. In other words,
the space £2(2) has the completeness property. D

1.3 Definition of Hilbert Space and its Elementary Properties

The vector spaces en

and £2 each have an inner product which satisfy (i)-(iv) in
Section 1. In applications, other vector spaces which possess an inner product
also need to be considered. For this reason we introduce the definition of an inner
product space .
It is obvious from the definitions of addition and scalar multiplication on en
and £2 that if E is either of these spaces, then for x, y and z in E , x + y and ax
are also in E and

(a) x +y = y +x;
(b) (x + y) + z = x + (y + z);
(c) There exists a vector 0 E E such that x +0 = x for all x E E .
(d) For each x E E there exists a vector -x E E such that x + (- x) = o.
For every a and f3 in C ,

(e) a(x + y) = ax + ay ;
(f) (a + f3)x = ax + f3y;
(g) (af3)x = a(f3x);
(h) Ix = x.

These properties are used to define an abstract vector space.

Definition: A vector space over C is a set E, whose elements are called vectors,
together with two rules called addition and scalar multiplication. These rules asso-
ciate with any pair of vectors x and y in E and any a E C, unique vectors in E,
denoted by x + y and ax, such that (a)-(h) hold.
6 Chapter I Hilbert Spaces

A subset M of E is called a subspace of E provided the following two conditions

hold .
(i) Ifu and v are in M, then u + v is in M .
(ii) Ifu is in M and a is in C, then au is in M.
Given a set SeE, the span of S, written sp S, is the subspace of E consisting of
all linear combinations ofvectors in S, i.e., all vectors ofthe form L:7=1 a isr, a , E
e,Si E S.
Vectors VI, . . . , Vn are called linearly independent if L:7=1 a; Vi = 0 holds only
if each ai = O. Otherwise, the vectors are called linearly dependent. Infinitely
many vectors VI, V2, . . . are linearly independent if for each k, VI, .. . , Vk are lin-
early independent.
A subset {VI , ... , Vn} ofa vector space V is a basis for V if V = SP{VI, . . . , vn}
and VI, . . . , V n are linearly independent. In this case we say that V has dimen-
sion n, written dim V = n . V is infinite dimensional if it contains an infinite
linearly independent set of vectors . Thus 12 is infinite dimensional since el
(1,0,0, ...), e2 = (0, 1,0,0, ...), . . . are linearly independent.

Definition: Let E be a vector space over C. An inner product on E is a complex

valued function, ( , ), defined on E x E with the following properties.
(a) (x + y, z) = (x, z) + (y, z) ;
(b) (ax, y) = a(x, y} ;
(c) (x , y) = (y, x) ;
(d) (x , x) ::: 0 and (x , x) > 0 if x ;j; O.
E, together with an inner product, is called an inner product sp ace.
The norm, IIxll, on E is given by [x] = (x ,x}I /2, and the distance between
vectors x and y is IIx - yll .
It follows from (a)-(c) that

(x,ay+{3 z}=a(x,y}+fi(x,z) and II ax II = lalllxll·

For Xo E E and r > 0, the sets {x : IIx - XoII s r}, {x : IIx - XoII = r} are
called the r-ball and r-sphere of E, respectively, with center x o.

Examples: 1. en andl2 are inner product spaces . Givenxo = (171,1/2 , .. .) E 12,

the set of all sequences {~d E 12 such that (L:~I I~k - 17d)I/2 S r is the r-ball
in 12 with center xo.
2. Let L2 ([a , b]) be the vector space ofall complex valued Lebesgue measurable
functions I defined on the interval a S x S b with the property that 1/1 2 is
Lebesgue integrable. By identifying functions which are equal almost everywhere,

(f, g) = l b
I(x)g(x) dx
1.3 Definition ofHilbert Space and its Elementary Properties 7

defines as inner product on L2([a, b]). A discussion of this space appears in

Appendi x 2. Given fo E L2([a , b]), the set of all vectors f E L2([a, b]) for
which (J:
If(x) - fo (x )12d x ) 1/ 2 = r is the r -sphere of L 2([a , b]) with center f o.
3. Let if be the subspace of i2 consisting of all sequences (;d, where s, = 0
for all but at most a finite number of k. With the inner product inherited from
i2, if is an inner product space.
4. The vector space C2([a, b]) of continuous comple x valued functions with
inner product defined in Example 2 is an inner product space .
A number of the geometric propert ies of the plane carry over to inner product
spaces. For example , we have the following basic inequalities.

Theorem 3.1 Let E be an inner product space. For x and y in E,

(i) !(x , Y)I :s IIxlill YIl (Cauchy-Schwarz inequality)

(ii) II x + yll :s IIx II + lIylI (Triangle inequality)
(iii) IIx + Y II 2 + Ilx - Y II 2 = 2( lI x 1l 2 + II YII 2 ) (Parallelogram law).
(iv) II x - Y II 2: IlI xll - II YIII

Proof: The proofs of (i) and (ii) are exactly the same as the proofs of the inequal-
ities (1.1) and (1.4) in Section 1. Also,
2+lIx 2
IIx+Y II -YII
= (x + y , x + y) + (x - Y, x - y )
= II xll + 2 Re (x , y ) + IIYII 2 + II x ll 2 -
2
2Re (x , y) + IIYII 2
2(lIx 1l + II YII 2 ) .
2
=

The proof of (iv) follows from (ii).

Most of the fundamental theorems which are concerned with operators defined
on an inner produ ct space 'H require that 'H be complete. In fact, we rarely consider
inner product spaces which do not have this property. 0

Definition: A sequence {x n } in an inner product space 'H is said to converge to

x E 'H, written X n -+ x, if II x n - x II -+ O.
A sequence {x n } in H is called a Cauchy sequence if II x n - X m II -+ 0 as
n , m -+ 00 .
If every Cauchy sequence in H converges to a vector in H , then H is called
complete. A complete inner product space is called a Hilbert space . It is the most
natural generalization of en .
Examples: 1. en is an n-d imensional Hilbert space.
2. i2 is an infinite dimension al Hilbert space.
8 Chapter 1. Hilbert Spaces

3. L2([a, b]) is a Hilbert space (cf. Appendix 2). This space is infinite
dimensional since the functions 1, x, x 2 , . • . are linearly independent. For
if L:Z=oakxk is the zero function, then ak = 0, 0 :::: k :::: since any n,
polynomial of degree n has at most n zeros .
4. if is not complete. Indeed, let Xn = (!, -b,..., 2~' 0, 0, . . .) E if. Then
forn > m,

IIxn - Xm II 2 < L 00
1 = - 1 ---+ O.
-k
2 2m
k=m+1

Thus {x n } is a Cauchy sequence which converges in l2 to x = ¢ (!, -b, ...)

If. Consequently, {x n } cannot converge to a vector in If since limits of
sequences in l2 are unique .
5. The inner product space P of all polynomials with (f, g) = fol j(x)g(x) dx
is not complete. To see this, let

Then for g(x) = - \ - , 0 :::: x :::: 1, r, ---+ gin L2([0 , 1]). Thus, {Pn }
1- 7x
is a Cauchy sequence in P which does not converge to a vector in P since
g ¢ P C L2([0, 1]).
We shall see later that every finite dimensional inner product space is complete.

1.4 Distance from a Point to a Finite Dimensional Subspace

Throughout this section , E denotes an inner product space

Definition: The distan ce d (v, S) from a point vEE to a set SeE is defined
by
d(v , S) = inf[]» - s] : s E S}.
We shall show that if M is a finite dimensional subspace of E, then for each
VEE there exists a unique W E M such that d(v , M) = IIv - wll. Hence
II v - w II < II v - z II for all Z EM , z f= w.
The following preliminary results are used .

Definition: Vectors u and v in E are called orthogonal, written u ..1 v, if

(u , v) = O.
The vector v is said to be orthogonal to a set M C E if v ..1 m for all m EM.
We write v ..1 M.
1.4 Distancefrom a Point to a Finite Dimensional Subspace 9

A system of vectors {({II, ({I2, . . . } is called orthogonal if ({Ii 1. ({Ii - i i= j. If, in

addition, II ({Ii II = I , I :::; i , the system is called orthonormal.
An orthonormal system {({Ii} is linearly independent since
n
0= Laj({lj
j =1

Every finite dimensional inner product space has an orthonormal basis. This result
is a special case of Theorem 9.1.
The following simple theorem is useful for calculations.

Pythagorean Theorem 4.1lfu 1. v, then

Proof:

lIu + vll 2 = (u + v, u + v) = lIull 2 + (u , v) + (v , u) + IIvll 2

= lIull + IIv1l .
2 2

To return to our problem, we are given a finite dimensional subspace M of E

and a vector VEE . We wish to find aWE M such that d (v , M ) = II v - w ll . If
there exists a wE M such that v - w 1. M, then w is the unique vector in M such
that d (v , M) = II v - w ll. Indeed , if z E M and z i= w , then w - z is in M and by
the Pythagorean theorem appl ied to v - wand w - z,

To find w so that v - w 1. M, let ({II , ..., rpn be an orthonormal basis for M . Then
w = L:J=l a j ({lj and for each k,

0= (v - w , ({Ik) = (v , ({Ik) - (taj({lj, ({Ik) = (v , ({Ik) - ai :

J=I

Thus
n
W = L(v, ({Ij)({Ij. (4.2)
j=1

Conversely, if w is given by (4.2), then v - w 1. M; for if z E M , say z =

L:Z=If3k({lk, then (v - w , z) = L:Z=I
fh(v - w , ({Ik) = O.
Thus we have the following result. 0

Theorem 4.2 Let M be a finite dimensional subspace of E and let {({II , . . . , ({In}
be an orthonormal basisfor M . For each v E E , the vector w = L:J=l (v , ({Ij)({Ij
is the unique vector in M with the property that II v - w II = d (v , M ).
10 Chapter I. Hilbert Spaces

The equivalence of a closest vector and orthogonality is given in the next theorem.

Theorem 4.3 Let M be a subspace of E . Supp ose VEE and W E M . Then

v - w .L M if and only if II v - wll = d (v , M ).

Proof: Ifv- w .L M , then it follows from (4.1) thatd (v , M ) = IIv-wli. Suppose

II v - wll ::: IIv - zll for all z E M . Since M is a subspace, w + AZ is in M for all
Z E M and A E C. Therefore,

II v - w ll 2 ::: II v - (w + )..z)1I 2 = (v - w - AZ, v - w - AZ)

= II v - w ll 2
- 2Re A(Z, v - w } + IA12 11 z 11 2 .

Hence
2 Re A(Z, v - w } ::: IA1 2 11 z 11 2 .
Taking A = r(z , v - w}, where r is a real number, we get

2rl(z , v - w}12 ::: r 2 1(z, v - w}1 2I1zI1 2 .

Since r is arbitrary, it follows that (z, v - w) = O. D

1.5 The Gram Determinant

Throughout this section, E denotes an inner product sp ace . Let YI, ... , Yn be
a basis (not necessarily ortho gon al) for a subspace M of E. We know from
Theorem 4.2 that there exists a unique w = I:7=1cq y, in M such that II Y- w ll =
d (y, M ) or, equivalentl y, y - w .L M . Our aim in this section is to find each a ,
and d (y, M ) in terms of the basis.
Now
n
0= (y - w, Yj) = (y , Yj ) - Lai(Yi , Yj}, 1 ::: j ::: n ,
i= 1

or
al (YI , YI) + a 2(Y2, YI) + ...+ an (Yn, YI) = (y , YI)
(5.1)
al (YI , Yn) + a 2(Y2, Yn) + ...+ an (Yn , Yn) = (Y, Yn).
Since the set of a, is unique , we have by Cramer's rule that

aj = Dj , (5.2)
g(YI, . . . , Yn)
where g(YI, . . . , Yn) is the (non-zero) determinant of the coefficient matri x
1.5 The Gram Determinant 11

and D j is the determinant of the matrix which is obtained by replacing the ph

column of the coefficient matrix by the column «(y , Yi ))7= 1.
The determinant g(Y I, ... , Yn) is called the Gram determinant corre sponding
to the n-tuple (YI, .. . , Yn) .
A neat way to represent w = L: f=1 ctiYi is by the determinant

(Yn, YI) (y, :YI ) ) .

(Yn, Yn) (y , Yn)

Yn 0
(5.3)
This can be verified by expanding the determinant of the last row as minors and
referring to (5.2).
We shall now obtain a formula for the distance d = d (y , M) .
Since Y - w 1- M ,

d 2 = lI y - wll 2 = (y - w, Y - w ) = (y , y ) - (w , y )
n
= (y , y) - L cti (Yi , y )
i=1

or

Attaching (5.4) to the system of equations (5.1), we obtain a non-trivial solution

to the sytem of equations

(YI , Yn)XI + (Y2, Yn)X2 + ...+ (Yn, Yn)x n + (-Y , Yn)Xn+1 = 0

(YI , y )XI + (Y2, Y )X2 + ...+ (Yn, y )x n + (d 2 - (y , y ))Xn+1 = 0

A non-tri vial solution, as seen from (5. 1) and (5.4), is

Xi =cti, 1 s ! S n, Xn+1 = 1.

Hence , by Cramer's rule,

12 Chapter 1. Hilbert Spaces

Let us expand the determinant by using the last column as minors. If C j is

the cofactor corresponding to the number in the i" row of the last column, then
Cn+l = g(Yl , ·· ·, Yn) and
(d 2 - (Y , y })g(Yl , . . . , Yn ) - (Y, Yn} Cn - . .. - (y , yd Cl = O.

Thus
n
2
d g (Yl , ··· , Yn) = (y, y }Cn+l + L(Y, Yk}Ck
k=l
(Yl /l) ...
= det .
( (Yl , Yn)
(Yl, y)
= g(Yl,· · ·, Yn, y ). (5.5)

Now g(Y l) = (Yl , y d > O. Applying (5.5) to M; = Sp{Yl, . .. , ykl and Y = Yk+ l,
and noting that d (Yk+l , Mk) ::::: II Yk+IIl , it follows by induction that

0 < g(Y l, , Yn) ::::: g(Yl , . . . , Yn_l) IIYn Il 2

2 2 211Y2 2 2
::::: g(Y l, , Yn-2) IIYn - i11 II Yn- 2 11 ::::: . . . ::::: IIYl 1l 11 . . . llYn 11 . (5.6)
To summarize, we obtain from (5.3), (5.5) and (5.6) the following theorem.

Theorem 5.1 Let {YI , ... , Yn } be a basis fo r the subspace M C E . For Y E E ,

)
l~
d ey , M ) = g(YI, , Yn, y)
( g(YI , , Yn)
and
0 < g(Y l, . . . , Yn) ::::: II YI1l 2 I1 Y211 2 . .. ll Yn 11 2,
where g( XI, . . " Xj) = det «(xi , Xj }) is the Gram determinant correspo nding to
(XI,X2 , . . . , Xj) .
The vector WE M for which d ey, M ) = lI y - wil is given by

There is a very nice geometric interpretation of the Gram determinant. For

example, let YI, Y2 and Y3 be linearly independent vectors in 3-space. The volume V
of the parallelopiped in Figure 1 is

V = IIY llld(Y2 , Sp{YI })d(Y2, Sp{YI, Y2}).

1.6 Incompatible Systems ofEquations 13

Figure I

Thus by Theorem 5.1,

V2 _ ( )g(Yl' Y2) g(Yl , Y2 , Y3) _ ( )

- YI, YI () ( ) - g YI , Y2, Y3 .
g YI g Yl ,Y2
As an application of Theorem 5.1, we have the follow ing result.

Hadamard's Inequality for a Determinant. Given A = (aU)' an n x n matrix

ofcomplex numbers,
n n
[det AI
2
~ IlL laul 2 .
i=1 j=l
Proof: Let Yi = (ail, . . . , ain) be considered as a vector in the Hilbert space en .
Then
n n
(Yi, Yj) = Laikiijk and IIYill
2
= L laid.
k=1 k=l
Hence if A * is the conjugate transpose of A , then by Theorem 5.1,

IIYI1l
2I1Y211 2 2
. . . llYn 11 ~ g(Yl , . .. , Yn) = det((Yi, Yj)) = det AA*
= det A det A = [det A1 2 ,
which establishes the desired inequality. o
1.6 Incompatible Systems of Equations

Suppose that Y is known to be a linear function of XI, X2, . . . , Xn, i.e., Y =

AIXI + A2X2 + ... + AnXn, but the constants Ai have not been determined. An
experiment is devised in which the inputs XI , . . . , Xn and the corresponding output
yare measured. The experiment is repeated m times so as to provide a system of
equations of the form

YI = aliA/ + al2 A2 + ...+ alnA n

(6.1)
14 Chapter I Hilbert Spaces

where in the i t h experiment, aij is the measurement oix , and y, is the measurement
ofy.
Even if this system of equations could be solved, the solutions would only
approximate the desired Ai since measurements are not exact. Moreover, if m > n,
there might not be a solution, in which case the system of equations is called
incompatible.
As a compromise, we seek to minimize the deviation between the right and left
sides of each equation. To be specific, it is desired to find a I , a2, .. . , an so that
I(AI, , An) = IYI- L:7=1
aIi A;l2+. +!Ym- L:7=,
amiA;l2attains it minimum
at (ai, , an)'
The solution to this minimization problem is obtained directly from Theorem 5.1
as follows.
Let
Ai = (ali, . . . ,ami) E em, 1::: i ::: n, Y = (YI , ... , Ym) E em.
The components of Ai and Y correspond to the columns appearing in (6.1) . Let
M = SP{Ai} C em. Since, in practice, there are usually more experiments than
there are variables, i.e., m > n, we assume that A I, . .. , An are linearly indepen -
dent in em. In this setting we seek (ai , ... , an) so that for w = a, Ai and L:7=,

r
any AI , . . . , An in C,

I(al, . . . , an) = IIY - wll 2 = d 2(y , M) :::lI y - t AiAi

1= 1
= I(AI, ... , An).
Let
m
Cij = (Aj, Ai) = I>kjaki, 1 s i, j ::: n ,
k=1
m m
Yi = (Y, Ai) = ~:::>kali ' 1 ::: i ::: n , Yn+1 = L IYkI 2
.
k=1 k=1
Then by Theorem 5.1, the desired w is given by
CII CI2 c.,
w=-
1
det ·
: ~I)
det(Cij) ( Cnl Cn2 Yn
AI A2 o
and
CII Cln
2 : 1
l(al, . . . ,an)=IIY-wll =d det·
et(Cij) ( Cnl C nn
VI V
1. 7 Least Squares Fit 15

1.7 Least Squares Fit

Given points tl , . .. , tk in the interval [a, b] and given complex numbers YI, .. . , Yk,
let us first consider the problem to find a polynomial P of degree at most k - 1
such that P(tj) = Yj, I S j S k. This problem has a solution which is given by
the Lagrange interpolation formula :

n--.
k k
t - ti
P(t) = LYj Pj (t) , Pj(t) =
j=1 i=1 tj - ti
i#j

In many cases, k is very large and it is reasonable to look for a polynomial P of

degree n < k - 1 which has the properties P (t j) = Yi- I S j S k . It is clear
that, in general, such a polynomial does not exist. As a compromise, we seek a
polynomial P of degree at most n such that

k
2
S(P) = L IYi - P(ti)1
i=1
has the smallest value among all polynomials of degree at most n.
To solve this problem of "least squares," let E be the vector space ofall complex
. k --
valuedfunctions defined onrj . v. : tk. Define (g , h) = Li=l g(ti)h(ti), g, hE E.
Clearly, ( , ) is an inner product on E . Let M be the subspace of E consisting of
all polynomials of degree at most n . The polynomials {I, t, .. . , t n } are linearly
independent in M; for if pet) = LJ=O£ljt j = 0, i.e., P(tj) = 0,1 S j S k,
then P is a polynomial of degree n < k which has at least n + 1 zeros. Thus

°
£lj = 0, I S j S n.
For S i, j S n, let
k k n
Cij = (t l . , t i ) = L i+j
i-:': s, = L Ymtm,
i
Bn+1 = LY;"
m=1 m=l m=1
It follows from Theorem 5.1 that the desired polynomial P is given by

C
COl COn

pet) =
1
det
:
.
c;
~O)
det(Cij) C;o Cnl Bn
t tn
°

C
and
COl COn
1 . Bo )
d 2(P, M) = det :
det( Cij) Sto Cnl c.; Bn
BO BI Bn Bn+1
16 Chapter I Hilbert Spaces

In practice, some of the datapoints YI, ... , Yk are, for various reasons, more
reliable than others. Therefore, certain "weights" 0; are chosen and the corres-
ponding least squares problem is to find the polynomial of degree at most n which
mmmuzes
k
S(P) = L Iy; - P(t;)120 t
;=1
among all polynomials ofdegree at most n . Ifwe apply the above results to E with
the inner product (g, h) = I:1=1 g(t;)h(t;)ot, the desired polynomial is obtained.

1.8 Distance to a Convex Set and Projections onto Subspaces

Throughout this chapter, H denotes a Hilbert space. It was shown in Theorem 4.2
that the distance from a vector to a finite dimensional subspace is always attained.
This result can be extended to closed convex sets . However, a very different proof
is needed.

Definition: A set C cH is convex if for any two vectors x and yin C, the set

{tx + (l - t)y : 0 ::: t ::: 1} is contained in C .

Examples: 1. Any subspace ofH is convex.

2. If x any yare vectors in the plane or 3-space, then {tx + (l - t) Y : 0 ::: t ::: I}
is the line segment joining x to y . Thus a set C in the space is convex if and
only if the line segment joining any two points in C lies in C .
3. The r-ball Sr(XO) = {x : IIx - x o ll ::: r} is convex . For if x and yare in
s, (xo), then for 0 ::: t ::: 1,
IItx + (l - t)y - xoll ::: tllx - xoll + (l - Oily - xoll ::: r.
Thus tx + (l - t)y is in Sr(XO).
4. The set of all functions in L2([a , b]) which are positive almost everywhere
on [a, b] is convex .

Definition: Given a set S C H, the closure of S, written S, is the set of those

vectors in H which are limits of sequences of vectors in S, i.e., XES if X n ~ x
for some sequence {x n } C S.
IfS = S, we call S a closed set.
Every r-ball in H is a closed set.
The remark following Theorem 18.1 shows that every finite dimensional sub-
space ofH is closed.
The subspace of £2 consisting of all sequences (0, ~I , ~2 , .. .) , where ~k =1= 0 for
at most a finite number of k is not closed in £2 .
We are now ready for the generalization of Theorem 4.2.
1.8 Distance to a Convex Set and Projections onto Subspaces 17

Theorem 8.1 Suppose M is a closed convex subset of H. Given Y E H, there

exists a unique wE M such thatd(y, M) = IIY - wll.

Proof: Let d = inf[jy - z] : Z EM} . There exists a sequence {Zn} C M such

that IIY - Zn II --+ d. The idea of the proof is to show that {Zn} converges and that
its limit is the desired w.
By Theorem 3.1 applied to Y - Zn and Y - Zm,

2(lIy-znIl2+lIy-zmIl2)=1I2y-(zn+zm)1I2+lIzn-zmIl2. (8.1)

Since M is convex, i(zn + Zm) is in M and

112y - (Zn +zm)1I = 211 y - ~(Zn +zm)ll2: 2d . (8 .2)

Combining (8.1) and (8.2), we get

2 2)
II zn - zmll 2 ~ 2(IIY - znll + IIY - zmll - 4d --+ 4d - 4d = 0
2 2 2

as n, m --+ 00. Since H is complete and M is closed, there exists awE M such
that Zn --+ w. Thus d = 1imn-;. oo IIY - z; II = lIy - wll. Finally, suppose Z E M
and d = II Y - Z II. Computing the distance between Y and the midpoint of the
segment jointing Z to w, namely i(z + w) E M, we obtain

d
2
s lIy - ~(Z + w)1I 2 = 11~(y - z) + ~(Y - W)r
Hence, by the parallelogram law applied to i(Y - z) and i(Y - w),

d
2
~ II~(Y-Z)+~(Y-W)r
= 2 (11~(y - Z)r +11~(y - W)r) -II~(Z - W)r

Thus Z = wand the theorem is proved. D

Since any r-ball S, in H is closed and convex, we can apply the above theorem
to Sr .

Definition: Given S C H, the orthogonal complement S..L of S is the set

{x E H : x 1- S}.

Theorem 8.2 Let M be a closed subspace of H. Given y E H, there exists a

unique wE M and a unique v E M..L such that y = w + v.
18 Chapter 1. Hilbert Spaces

Proof: By Theorems 8.1 and 4.3, there exists a unique W E M such that v =
Y - WE Ml.. and Y = W + v. Suppose Y = WI + VI, WI EM, VI E Ml... Then
Y - WI E Ml... Hence by the uniqueness of w, W = WI and therefore V = VI . 0

Corollary 8.3 If M is a closed subspace ofH , then (Ml..)l.. = M.

Clearly, M C (Ml..)l... Suppose Y E (Ml..)l... Theorem 8.2 guarantees the

existence of aWE M C (Ml..)l.. and a V E Ml.. such that Y = W + v. Thus
V = Y - wE (Ml..)l.. n Ml.. = (0) . Hence Y = W E M .

1.9 Orthonormal Systems

We recall from linear algebra that given linearly independent vectors YI, Y2, . .. ,yn
in H, there exist an orthonormal set of vectors C{JI, qJ2, . . . , C{Jn in H such that
=
SP{C{Ji If=1 Sp{Yi If=I' 1 ::: k ::: n. The C{Ji are defined indirectively as follows :

YI Yk - Wk-I
C{J = --, C{Jk =
IIYIII IIYk- wk-III'

where,
k-I
Wk-I = L(yk. C{Ji}C{Ji
i=1
and
IIYk - wk-III = d(yk. SP{YI, .. . , Yk-I D·
The following result enables one to find C{Jn without first determining C{JI, C{J2, • • • ,
C{Jn-l·

Theorem 9.1 Let YI, Y2, ••• be linearly independent in inner product space E
and let C{JI , C{J2 , ••• be the corresponding orthonormal set obtained by applying the
Gram-Schmidt orthogonalization procedure to YI, Y2 , .. .. Then

YI
C{JI = lIyIII

... (Yn+;, YI}) ,

n = 1,2, ...
(Yn+I , Yn)
Yn+1
(9.1)
where
(9.2)
1.10 Szego Polynomials 19

Proof: By the Gram-Schmidt procedure and Theorem 5.1,

1/ 2
Yn+1 - w dn
fIln +1 = II Yn+1 - W
II = ( -d)
n+1
(Yn +1 - w),

where w is given in Theorem 5.1 with Y replaced by Yn +l .

Hen ce

Y
I/Il (Y2, YI) .. . (Yn,: YI) (yn+;, YI))

( (YI ,Yn) (Y2 , Yn)

YI Y2
(Yn, Yn) (Yn+I, Yn)
Yn 0
(9.3)
But thi s equality is precisely the right hand side of equality (9.1). Th is follow from
the observation that the cofactors of YI, Y2, ... , Yn are the same in the determinants
appeari ng in (9. 1) and (9.3) wh ile the cofactor of Yn+ I in the determinant in (9. 1)
isdn . 0

It is obvious from the theorem above that every infinite dimensional Hilbert space
contains an infin ite orthonormal set.

Examples: 1. The standard bas is el = (l , 0, 0, . ..), e2 = (0, 1, 0, . . .), . . . is an

infinite orthonormal system in f.2 .
2. {Jkeinll: _oo is an infinite orthonormal system in L 2([-Jr, Jr]).
3. Jk, c:Ji, s:;¥, .. ., co./f/, si~ht, ...is an infinite orthonorm al system in
L 2([Jr , Jr]) .

1.10 Szego Polynomials

Let w be a strictly pos itive Lebesgu e integrable function defined on the unit circle.
Define on the vector space P of polynomials with complex coeffi cients on inner
product

f
zr

(p , q) w = p (ei()) q (ei()) w (ei ()) dO. (10.1)

- Jr

Applying the Gram-Schmidt process to the polynomials 1, Z, z2, ... , we obtain

pol ynomials fIlo , fill , . .. whic h are orthonormal relative to the inner product (. , .)w.
These polynomials fIlo, fill , . . . are called the Szego p olyn om ials for the weight w .
20 Chapter 1. Hilbert Spaces

If w is the constant function 1/21T, then the corresponding Szego polynomials are
1, z, z2, . . . .
In this section we prove three theorems. The first two describe in two different
ways the Szego polynomials. The third is Szego's theorem which states that all
the zeros of the Szego polynomials lie inside the unit circle.

Theorem 10.1 The Szego polynomials CPo, CPI, ... corresponding to the strictly
positive weight ware given by

CPo (z) = a o-1 /2

CPn(Z) = (dn-I, dn ) - 1/ 2 det

ao
al
:
a-I
ao a-n+1
a_
:
n

1
,n 2: 1, (10.2)
( an-I an - 2 a_I
1 z zn
where

am = !1T

-1T
e-imOw(eiO)dO, mE Z, (10.3)

(10.4)

and n is the Toeplitz matrix

a-I
ao
(10.5)

Proof: Since

( zj _!
,z k) w-
1T

e -i(k- j)e we . -- 0 , 1, 2 , .. . ,
(iO) --ak-j'. k , j (10.6)
-1T

we have

(10.7)

Equality (10.2) now follows directly from (10.6), (10.7) and Theorem 9.1 applied
to the polynomials 1, z, . . . , z", 0
1.10 Szego Polynomials 21

Let Tn be the Toeplitz matrix described in (10.5) with k = n . The matrix Tn is

invertible since det Z; > Oby(1O.7). Therefore there exist numbers Sj , .81 , .. . ,.8n
such that

(10.8)

These numbers can be used to give an alternative description of the n-th Szego
polynomial.

Theorem 10.2 Let .80, .81, ... , .8n be numbers such that (10.8) holds, and put
Pn(Z) =.80 + .8IZ + ... + .8nzn (10.9)
Then the n-th Szego polynomial C{Jn is given by

dn )1 /2
C{Jn(Z) = ( - - Pn(Z), n = 0,1,2, ... , (10.10)
dn-I
where dk is defined by (1004) and d_1 = 1.
Proof: Notice that 1 = To.8o = ao.8o. Therefore .80 = a o and thus
l
,

1/2 1/2 -I -1 /2
do po(z) = a o a o = a o = C{Jo(z),
which proves (10.10) for n = O.
For n ::: 1, we find the phcoefficient .8j of Pn(z) by applying Cramer 's rule to
the system of equations corresponding to the matrix equation (10.8). We get
j+2 /)" .
.8j=(-1)n+ d:' (10.11)

where /)" j is the determinant ofthe matrix obtained by deleting the (k+ l)st column
of the matrix

Now from (10.2) and (10.9) we have that the coefficient of zj of C{J(z) is

(dn_Idn)-1 /2(-1)n+k+2/)"j = (_1)n+ j +2(dn_Idn)-1 /2dn.8j = (d~~J -1 /2 .8j.

Thus formula (10.10) follows . 0
We are now prepared to prove the result about the location of the zeros of the
Szego polynomials. The proof given here is due to H. Landau.
22 Chapter I Hilbert Spaces

Theorem 10.3 (Szegd's Theorem). All the zeros ofthe Szego polynomialsfor the
positive weight ware contained inside the unit circle.

Proof: From equality (10.1 0) we see that it suffices to prove the theorem for Pn (z),
where Pn(Z) is defined by (10.8) and (10.9). Let A be any zero of p«. Then

Pn(Z) = (z - A)q(Z), (10.12)

where q (z) is a polynomial of degree n - 1. Since I{Jn is orthogonal to sp{I{JO, ••• ,

I{Jn- d = sp{1, z. . . . , zn-I }, and Pn is a scalar multiple of I{Jn, if follows that Pn is
orthogonal to q . Now
Pn(Z) + Aq(Z) = zqtz),
and from the definition of the inner product (., .)w we get

Thus
(l - IAI 2 )( q , q)w = (Pn, Pn)w > 0,
which implies IAI < 1. o
Examples: 1) Let WI be the weight given by

-3z
WI (z) = 2z 2 _ 5z +2 (lz] = 1).

From 2z 2 - 5z + 2 = (2z - l)(z - 2) and z= C l

for lz] = 1, we see that

3 3
WI (z) = (2z _ 1)(2cl _ 1) = 12z _ 11 2 ' Izi = 1.

Thus the weight WI is strictly positive on the unit circle . Also, it is easily seen that

WI (z) = -1 + -1- + 1
I
~
= L....- 2- J zJ
1'1 '
(lz] = 1).
1 - lz 1- -C l ,
2 2 J=-OO

Now let us use Theorem 10.2 to find the corresponding n-th orthogonal polynomial
I{Jn. This requires to solve the following system of equations:

f30 + 2/3t + + 2nf3n = 0

2f3o + f31 + + 2n - 1f3n-1 = 0
(10.13)
1.10 Szego Polynomials 23

Solving this system (for instance by elimination) we obtain the following solution:

2 I
fJo = 0, . .. , fJn-2 = 0, fJn-1 = 3' fJn = - 3'
From Theorem 10.2 we conclude that the n-th Szego polynomial for the weig ht
W I is given by

dn
((in(Z) = ( dn-I
)1/2 (23zn-l - 3z
I n) ,
where d_ 1 = I and d n is the determinant of the coefficient matrix of (10.13).
Notice that the zeros of ({in are given by Z = 0 and Z = i
which is in agreement
with Theorem 10.3.
2) In this example we take as a weight the function

W2(Z) = -2z- 1 + 5 - 2z = Iz- ~12 > 0 (lz] = 1).

The Toeplitz matrix that we are interested in now is the following (n + 1) x (n + 1)
tridiagonal diagonal matrix:

5 -2 0 0 0 0
-2 5 -2 0 0 0
0 -2 5 -2 0 0
T; = (10.14)

0 0 0 0 5 2
0 0 0 0 -2 5
Put d n = detTn , n = 0, 1,2, . . . , and set d_1 = 1. From the tridiagonal structure
of the matrix it follows that the determinants dn satisfy the following recurrence
relation:
dk = 5dk-1 - 4dk- 2, k = 1,2, . ...
This together with d-( = 1 and do = 5 yields

k 1
dk = 5 + t ( - I ) j (k+~ -j) (~r ,
where y = [(k + I)/2] denotes the largest integer which does not exceed (k + I) /2 .
The above formula for dk is proved by induction using the above recurrence
relations.
Using Cramer's rule it is straight forward to check that for this example the
solution fJo, . .. , fJn of equation (10.8) is given by

n-k dk-I
fJk = (- 2) - , k=0 ,1, 2, .. . , n.
dn
24 Chapter I. Hilbert Spaces

Thus the n-th Szego polynomial for the weight W2 has the form

(_2)n ~ -k k
((1n(Z) = r;r-:(l L...J(-2) dk-\ Z ,
yUn-\Un k=O

where d_\ = 1 and d n is the determinant of the Toeplitz matrix Tn in (10.14).

1.11 Legendre Polynomials

LetH = L2([-I, 1]) and let un(x) =xn,n = 0,1, . ... Using the Gram-Schmidt
orthogonalization procedure, we shall exhibit an orthonormal system whose span
coincides with the span ofu\, u2, . .. .
Let

Uo 1
({10 = lIuoll ; ({1o(x) = ,J2
V\ = Ul - (U\, ({1o)({10

({1\ = II~: II ; ({1\ (x) = If x

V2 = U2 - (U2, ({1o)({10 - (U2, ({1\ )({1\

V2 (5 1 2
C{J2 = II v 2 11 ; CfJ2(x) = V"2 . "2(3x - 1) .

Continuing in this manner,

n + I 1 dn
~
2 n
({1n ( x ) -- - 2- -2- - - x -1 ) ,
n n! dx" (
n = 0,1, .... (11.1 )

The proof of the formula is as follows .

An inspection of the Gram-Schmidt procedure verifies that each ({1k is a polyno-
mial ofdegree k and the coefficient of x k is positive. Thus, since sp{({1o, ••• , ({1k} =
sp{l, x, . . . , x k }.
k
((1k(X) = L akjx
j,
akk > 0 (11.2)
j=o

and
k
x
k
= Lb kj ({1j (X ) , bu > O. (11.3)
j=O
1.11 Legendre Polynomials 25

We observe that if P; is a polynominal of degree nand Pn 1- xi, j = 0, 1, . . . ,

n - 1, then Pn = Cnf/!n for some number Cn. To see this, we know from (11.3)
that r; =I:7=1 c.« and therefore
n
(P n, f/!k) = L c.i«. f/!k) = c..
;=1
But, by (11.2),
k
(Pn, f/!k) = L iikj(Pn, xi) = 0, 0 ~ k < n.
i=O
Hence Pn = Cnf/!n .
n
Let P (x) = ::n
(x 2 - l) " , which is a polynomial of degree n. To establish
(11.1) , we show Pn 1- x k , k < n, which implies that Pn = Cnf/!n. The constant
Cn is computed and (11.1) is verified.
Since Q~k) (x) = !:k
(x 2 - l)" vanishes at 1 and -1 when 0 ~ k < n, repeated
integration by parts yields for k < n,

(P n, x k) = 1 Q~n)(x)xkdx
1

-I
= (-I)kk! 1 Q~n-k)(x)dx
1

-I

Hence by the observation above,

Pn = Cnf/!n for some Cn E C. (11.4)

Since the leading coefficients of Pn and f/!n are positive, Cn > O. Thus
(11.5)

Integrating by parts , we get

(P n, Pn) = 11 Q~n) (x) Q~n) (x)dx = -11 Q~n-I)Q~n+l) (x)dx

1
-I -I
1
= . .. = (_1)n Qn (x) Q~2n) (x)dx .
-I

Since Q~n (x) = ::;n (x 2 - l )" = (2n)!,

(Pn, Pn) = (-It(2n)! 1 1

(x 2 - It dx

1
-I
1
= (2n)! (1 - xt(1 + x)n dx (11.6)
-I
26 Chapter I. Hilbert Spaces

and

t (1 -
1-1
x)n(1 + x)n dx = _n_
n+1
t (1 -
1-1
x)n-I (1 + x)n+1 ds

= . .. =
(n +
l)(n
n!
+
2) . . . 2n 1-1
(1 t + x)2n dx
(n!)22 2n+ 1
(2n)!(2n + 1)
Thus by (11.4), (11.5) and (11.6),

r, [2i"+T 1 dn 2 n
({In = IIPn ll = V~-2- 2nn! dx n (x -1) .

n
The polynomial -2}, 2
n . ddx n (x - l)n is called the Legendre polynomial of degree
n, We shall refer to ({In as the normalized Legendre polynomial of degree n , This
polynomial ({In has the following interesting property.
Let an E C be chosen so that IPn = an({Jn has a leading coefficient 1, i.e.,
an = J
2;J~W 2n~I ' For any polynomial of degree n with leading coefficient 1,
(11.3) implies that for some ak E C, 0 ::s k ::s n, Q = L~=o ak({Jk. Hence

1.12 Orthonormal Bases

Now that we know that every vector in a Hilbert space has a closest vector w
in a closed subspace M, it remains to find a representation of w. It turns out that
w = Lk(w, ({Jk)({Jk, where {({Jd is a certain orthonormal system in M . It is therefore
necessary to concern ourselves first with the convergence of the series.

Definition: A series L~ I Xk of vectors Xk E 11 converges to x E 11, written

x = L~I Xk, provided Sn -+ x, where Sn = L~=I Xk·
For example, if {ed is the standard basis for i2, and x = (aI, a2, ...) E i2,
then x = L~I akek since

L
00

lakl 2 -+ 0 asn -+ 00 .
k=n+1
1.12 Or thonormal Bases 27

Lemma 12.1 The inner p roduct is continuous on 'H x H , i.e., if X n ---+ x and
Yn ---+ Y in H , then
(Xn , Yn) ---+ (x , y ).

Proof: By Schwarz's inequal ity,

I(x n , Yn ) - (x, Y)I ~ I(xn, Yn - y )I + I(x n - x , Y)I

~ IIxnll llYn - YII + II xn - x IIII YII ---+ 0,

since II x n II ---+ II x II , Yn ---+ Y and X n ---+ x . o

Theorem 12.2 If {!PI , !P2, ... } is an or thonor mal system in 'H, then for every
x EH
(i) I:k (x, !Pk) 2 ~ IIx 1 2 (Bessel's inequality).
1 1

(ii) I:k(x , !Pk)!Pk converges.

(iii) I:ka k!Pk converges if and only if {ad E £2·
(iv) Ify = I:ka k!Pko then a k = (Y, !Pk ).

Proof: (i) 0 ~ (x - I:Z =1(X, !pkl!pk, x - I:Z= I (X, !Pk )!Pk)

= IIxII - 2 I:Z
2
=l I(x , !pkl1 2 + I:Z=l I(x , !Pk )12 .
Thus for all n,

which implies (i).

(ii) For Sn = I:Z=l(x, !Pk)!pk, it follows from (i) that for n > m,

II sn - sm 12 = (I: Z=m+ l (x , !pkl!pk, I:Z=m+l (x , !Pkl!Pk)

= I:Z=m+l I(x , !pkl 12 ---+ 0 as n , m ---+ 00.

Thu s {sn} converge s since 'H is complete, i.e., I:d x , !Pk )!Pk converges.
(iii) Let Sn = I:Z=l ak !Pk, Sn = I:Z=l lad . Then for n > m,
IIsn- sml = (I:Z =m+1 a k!Pk, I:Z=m+l ak!Pk) = I:Z=m+llad
2
= Sn-Sm.
Thus {sn} is a Cauchy sequence if and only if {Sn} is a Cauchy sequence.
Therefore {sn} converges if and only if {Sn} converges, which implies (iii).
(iv) Suppose Y = I:ka k!Pk. Then by Lemma 12.1,
(Y, !Pi) = lim
n -4 00
(I:Z=l ak!Pk, !Pi )= a i'

o
28 Chapter I. Hilbert Spaces

Definition: An orthonormal system {CPI, CP2 , . ..} is called an orthonormal basis

for H if for every v E H, v = L k akCPk for some aI , a2, . . . in Co
By Theorem 12.2 (iv), ak = (y , CPk) . Each (y, CPk) is called a Fourier coefficient
ofy.

Examples: 1. The standard basis {ek} is an orthonormal basis for £2. Also
£2(Z) has a natural orthonormal basis. Indeed, put ek = (8jk)jEZ, where 8jk
is the Kronecker delta. Then ek E £2(Z), and

_ {O if k =F e,
(ek, ee) - 1 if k = £.

The vectors ek, k E Z, have norm one , are mutually orthogonal, and each
x = (~j )jEZ can be written as x = L~-oo ~jej with the series converging
in the norm of £2(Z) . Thus .. . , e_l, eo, el, ... is an orthonormal basis of
£2(Z). We shall refer to this basis as the standard orthonormal basis of
£2(Z).
We shall prove the next assertions 2, 3 and 4 in Sections 13 and 14.
2. beinX, n = 0, ±1, ... is an orthonormal basis for L2([-n, n]) .

.Jiii ' ~
3 . {_I_ -./ii n = I is an orthonormal basis for L 2 ([-n , n] ) .
-./ii ' sinnx}oo

r-:
4. The normalized Legendre polynomials

n
CPn(x) = 1 -d (x 2 - 1)n ,
- - - nn! n = 0,1, .. .
n 2 dx"
is an orthonormal basis for L2([ -1, 1]).

Suppose dimH = nand {CPI , .. . , cpkl are orthonormal in H. Since orthonormal

vectors are linearly independent, {CPI , . . . , CPk} is an orthonormal basis for H if and
only if k = n ,
If we have an infinite orthonormal system, then it need not be an orthonormal
basis. For example, if one of the vectors from an orthonormal basis is deleted,
then the remaining set is no longer an orthonormal basis. The following theorem
provides very useful way to check if an orthonormal system is an orthonormal
basis.

Theorem 12.3 Let {cpI, CP2, . . .} be an orthonormal system in H. The following

statements are equivalent.

(i) {CPI, CP2, . . .} is an orthonormal basis for H.

(ii) If(x , CPk) = Ofor k = 1,2, ... , then x = O.
(iii) Sp{CPk} is dense in H, i.e., every vector in H is a limitofa sequence ofvector
in Sp{CPk}.
1.13 Fourier Series 29

(iv) For every X E H,

IIxll 2 = Lk I(x, qJk)1 2 (Parseval's equality) .

(v) For all x and yin H.

n
(x, y) = lim (sn, Sn) = lim "(x, qJk)(y, qJk) .
n~oo n~OO L...J
k=!

(v) implies (iv). Take x = y in (v).

(iv) implies (iii). Given x E H,

asn ~ 00 .
(iii) implies (ii). If (x, qJk) = 0, k = 1, 2, . . . , then clearly x ..1 sp{qJd.
Therefore, by the continuity of the inner product, x is orthogonal to the closure of
Sp{qJk}, which is H . In particular, x ..1 x and x = O.
(ii) implies (i). For any z E H, W = Lk (z, qJk)qJk converges by Theorem 11.2.
Thus for each j,

(z - w, qJj) = (z, qJj) - n~moo(t(z, qJk)qJk, qJj)

k=!
= (z, qJj) - (z, qJj) = o.
Assumption (ii) assures us that z - W = 0 which establishes (i).
An orthonormal basis for 1-l is also called a complete orthonormal system. 0

1.13 Fourier Series

The proofs of the assertions 2, 3, and 4 preceding Theorem 12.3 rely on the
following two approximation theorems ([R], pp. 174-5).
Weierstrass Approximation Theorem. Iffis a complex valuedfunction which is
continuous on [a, b]. then for every e > 0 there exists a polynominal P such that

I/(x) - P(x)1 < e for all x E [a, b].

30 Chapter 1. Hilbert Spaces

Weierstrass Second Approximation Theorem. Iffis a complex valuedfunction

which is continuous on [-Jr, Jr] and I ( -Jr) = I (zr), then for every 8 > 0 there
exists a trigonometric polynomial
n
Tn(x) = L(aj cosjx + bj sinjx)
j=o
such that
II(x) - Tn(x)1 < 8 for all x E [-Jr , zr].
1. To prove that the orthonormal system

_ { 1 cosnx sinnx .
S- ../2ii' .fiT ' .fiT . n = 1,2, .. . }

is an orthonormal basis for L2([ -Jr, n ]), it suffices, by Theorem 12.3, to show
that the span of S is dense in the space.
Suppose I is a real valued function in L2([ -Jr, Jr]) . Given 8 > 0, there exists
a real valued function g which is continuous on [-Jr, zr] such that

III-gil <8/3 (13.1)

(cf. [12], p. 90). Next we choose a function h which is continuous on [-Jr, Jr] so
that h( -Jr) = h(Jr) and
IIg-hll< 8/3. (13.2)
By the second Weierstrass approximation theorem, there exists a trigonometric
polynomial Tn such that
8
Ih(x) - Tn(x)1 < 10' x E [-Jr,Jr].

Thus

j
Jr 82
IIh - Tn 11 2 = Ih(x) - Tn(x)12 dx < - . (13.3)
-rn 9
Combining (13.1), (13.2) and (13.3) we have

III - T; II < 8. (13.4)

If I is complex valued, it follows from (13.4) applied to mI and ~ I that the

span of S is dense in L2([-Jr, Jr]).
Now that we know that S is an orthonormal basis for L2 ([ -Jr, Jr]), we may
conclude from Theorem 12.3 that given IE L2([ -Jr, Jr]), the series

ao
-+2
L
00

(an cosnx + b; sinnx) ,

n=!
1.14 Completeness ofthe Legendre Polyn omials 31

where

an = -
TC
1 1][
_][
I(x) cos nx dx , b; = -
1
TC _][
1][
I(x) sinn x dx ,

n = 0, 1, . . . , converges in L2([- TC, TC]) to I, i.e.,

2
lim
n-+oo
r
J_][ I/ (X)- ao -
2 L...J
~(ak cos kx + bk sin kX)1 dx = O.
k= 1

The series is called the Fourier series of I ;an, b; are called the Fourier coefficients
of I .
By Parseval's equali ty (Theorem 12.3),

inx
2. Since cos nx = e +/- inx and sin nx = e
inx
2r inx
, it follows from the above

result that sp{e inx : n = 0, ±I , .. .} is dense in L2([-TC, TC]) . Hence { 1.-:=- einx :
v 2][
n = 0, ±I , .. .} is an orthonormal basis for L2( [-TC , TC]) . Therefore , given I E
L2([-TC, TC]) , the Fourier series L~- oo cneinx, where

Cn = -
1 1][ .
l(x)e- In X dx , n=O,±I , . . .
2TC -r tt

converges in L2([-TC,TC]) to I, and by Parseval's equal ity,

1.14 Completeness ofthe Legendre Polynomials

It was shown in Section 11 that the set of normal ized Legendre polynomials

n
~I d
= V-----z--
2 n
({in(x) 2- 2nn ! dx n (x - 1) , n = 0, I , . ..

is orthonormal in L 2([- 1, 1]) and

sp{l , x, ... , x k} = sp{({iO, ({il , ·· ., ({ik}, k = 0,1 , . . . . (14.1)

We now proceed to prove that {({iO, ({iI, . . .}is an orthonormal basis for L2([- I , I])
by showing that (iii) of Theorem 12.3 holds.
32 Chapter I. Hilbert Spaces

Since the set of all complex valued functions which are continuous on [-1, I]
is dense in Lz ([ -1, I]), it follows from the Weierstrass approximation theorem
that the set of all polynomials is also dense in Lz ([ -1 , 1]). Thus from (14.1), the
span of <PI , <Pz, . . . is dense in L2([-n , n]) . This proves that {<po, <PI , . . .} is an
orthonormal basis. Consequently

f I l k:~:~>n<Pn(X) I
Z

lim f(x) - dx = 0,
k-+oo -I n=O

where

By Parseval's equality,

1.15 Bases for the Hilbert Space of Functions on a Square

Let rio = L2([a, b] x [a, b]) be the vector space of all those complex valued
functions f which are Lebesgue measurable on the square [a , b] x [a, b] and for
which

By identifying functions which differ at most on a set of Lebesgue measure zero

on [a, b] x [a, b],

(h, g) = llb b
h(t, s)g(t , s) ds dt

defines the inner product. With this inner product, rio is a Hilbert space .
The reader in referred to [R] for a treatment of rio .
An orthonormal basis for rio can be constructed from an orthonormal basis for
L2([a , b]) as follows .

Theorem 15.1 If<P I, <112 , .. . is an orthonormal basisfor L2([a, b]), then ¢>ij(s , t) =
<Pi(S)<pj(t), 1 ~ i, j,forms an orthonormal basisfor L2([a, b] x [a, b]).
1.15 Basesfor the Hilbert Space ofFunctions on a Square 33

Proof: The set <l> ij is orthonormal since

(<I>jk. <l>mn) = IIb b

rpj(S)rpk(t)rpm(s) rpn(t) ds dt

= I b
rpj(s)rpm(s) ds I b
rpk (t)rpn (t) dt = Ojm Okn ,

where OJ m, called the Kronecker delta , is defined to be 1 if j = m and 0 if j f. m .

To prove that the system {<I>jk} is an orthonormal basis, we show that if g E
L2([a , b] x [a, b]) and (g , <l>jk) = 0, then g = 0 a.e. Now

0= (g , <l>jk) = II b b
g(s , t)rpj(s) rpkCt) ds dt

= I b b
rpj(S){l g(S,t)rpkCt)dt} ds.

I
Thus
b
h(s) = g(s , t)rpk(t) dt

is orthogonal to each rp j which implies h = 0 a.e. Hence there exists a set Z of

Lebesgue measure zero such that for each s If. Z, g(s ,) is orthogonal to rpk. k =
1,2, . .. . Consequently, g(s, t) = 0 for almost every t and each s If. Z. It therefore
follows that

from which we may conclude that g = 0 a.e.

The argument above shows that if {rpl , rp2, ...} and {1f!1 , 1f!2, . . .} are orthonormal
bases for L2 ([a, b]), then .fjk (s , t) = rp j (s) 1/Jk (r), 1 ~ j , k is an orthonormal basis
for L2([a , b] x [a, b]) . We can write {.fjk} as a sequence Fll, F12, F21, F22, . ..
(cf. Appendix 1).
Since { ~eint , n = 0, ±1, . . .} is an orthonormal basis for L2([ -rr, rr]) ,
v2Jr
Theorem 15.1 shows that {~ei(nt+mS) : m, n = 0, ±1 , . . .} is an orthonormal
basis for L2 ([ -rn , rr] x [-rr, n ]). Hence, for f in this space,

I' I'
-Jr -Jr
f (t. s) - i: t a~e;("+·')
n= P m= j
2 ds dt

converges to zero as j, p --+ -00 and k, q --+ 00, where

a nm = -1 jJr jJr .
f(t, s)e-1(nt+ms) ds dt.
2rr - Jr - Jr

o
34 Chapter 1. Hilbert Spaces

1.16 Stability of Orthonormal Bases

The following theorem shows that an orthonormal system which is "close enough"
to an orthonormal basis is also an orthonormal basis .

Lemma 16.1 If M and N are subspaces ofan inner product space and dimM <
dimN, then Ml- n N =1= (0)

Proof: Let CPI, . .. , CPn be a basis for M and let 1/11 , ... , 1/In+1 be linearly inde-
pendent in N . The lemma is proved once it is established that there exist numbers
fh , . .. , f3n+ I, not all zero , such that L7~l f3; 1/1; 1. cPk, k = I, 2, .. . , n, or

n+1
L f3; (1/1;, CPk} = 0, k = 1,2, ... , n.
;=1

But, from linear algebra, this system of n equations in n +I unknowns has a

non-trivial solution. 0

Theorem 16.2 Let CPI , CP2 , .•. be an orthonormal basis for a Hilbert space H. If
1/11,1/12, . .. , is an orthonormal system in H such that L~I IICPn - 1/In 11 2 < 00,
then 1/11, 1/12 , . . . is an orthonormal basis for H.

Proof: Assume {1J!n}~1 is not an orthonormal basis forH . Then by Theorem 12.3,
there exists a 1/10 E H such that 111/10 II = 1 and (1/10 , 1/1 j ) = 0, j = 1, 2, . ...
Choose an integer N so that

00

L 2
IICPn - 1/In 11 < 1. (16.1)
n=N+I

Lemma 16.1 guarantees the existence of a w =1= 0 in sp{1/10, . .. , 1/IN} such that
w 1. tpi> I .:s j .:s N . Since {CPn} ~ I is an orthonormal basis for H and w 1. 1/1n-
n > N,

00 00

0< IIwll 2 = L I(w, CPn}1 2 = L I(w, CPn}!2

n=1 n=N+I
00 00

= L I(w , CPn - 1/In}1 2 .:s IIwll 2 L IICPn - 1/In 11

2
,
n=N+l n=N+I

which contradicts (16.1) . Thus {1/In}~1 is an orthonormal basis for H. 0

1.17 Separable Spaces 35

1.17 Separable Spaces

Now that we have stud ied orthonormal bases in cons iderable detail, it is natural to
determine which Hilbert spaces possess an orthonormal basis .

Definition: A Hilbert space 'H is called separable ifthere exist vectors V I , V2, . ..
wh ich span a subspace dense in 'H.
An equivalent definition appears in Appendix 1.
Every Hilbert space H which has an orthonormal basis is separable since the
span ofthe basis is dense in H by Theorem 12.3(iii). In particular, £2 and L 2([a , b])
are separable.
It turns out that only separable Hilbert spaces have an orthonormal basis .

Theorem 17.1 A Hilbert space contains an orthonormal basis ifand only ifit is
separable.

Proof: Suppose H is a separable Hilbert space and Sp{ WI , W2 , . .. } is dense

in H. By discarding those W k which are linear combinations of W I, ... , Wk-I
it follows that linear independent vectors {VI, V2 , . . . } can be found such that
Sp{ VI , V2 , . .. } = Sp{WI , W2, . . . }. Hence there exists , by Theorem 9.1 , ortho-
normal vectors {qJI, qJ2, . .. } such that sp{qJ;} = sp{v;} which is dense in H . Thus
{qJI , qJ2 , . .. } is an orthonormal basi s for H by Theorem 12.3.
Conversely, if qJI , qJ2, • . . is an orthonormal basis for H , then Sp{qJI, qJ2, . . . } is
dense in H . 0

Theorem 17.2 A closed subspace ofa separable Hilbert space is separable.

Proof: Let sp{VI , V2 , . . . } be dense in the Hilbert space H and let M be a closed
subspace oi'H. For each k there exists a unique vector W k E M such that Vk - Wk ..1
M . We shall show that sp{WI, W2 , . . . } is dense in M which proves that M is
separable.
Given wE M and given e > 0, there exists a vector LJ=I !3jV j such that

n
W - L!3jVj < c. (17.1)
j=1

Since LJ=I !3j(Vj - Wj ) is orthogonal to M , (17 .1) and Theorem 4.3 imply

(17.2)
36 Chapter I. Hilbert Spaces

Thus, from (17.1) and (17.2),

n n n n

W - L{3j Wj S W - L{3j Vj + L{3j Vj - L{3j Wj < 2£.

j=1 j=1 j=1 j=1

Therefore, Sp{WI, W2 , . .. } is dense in M.

We can now give a representation of the projection of a given vector into a
closed , separable subspace of a Hilbert space. 0

Theorem 17.3 Suppose M is a closed, separable subspace of a Hilbert space

H. Then M has an orthonormal basis {cpI , CP2, .. .}. Given y "E 'H, the vector
W = L k(y, CPk)CPk is the proj ection ofy into M , i.e., y - W 1.. M .

Proof: Since M is closed and 1i is complete, it follows that M , with the inner prod-
uct inherited from H , is a separable Hilbert space. Hence M contains an orthonor-
mal basis {cpI , CP2, ·· .}. Now W = L k( y, CPk)CPk converges by Theorem 12.3 and
for each i .

(y - w, CPj ) = (y , CPj ) - n~moo (i)y, CPk )cpk, CPj ) = (y , CPj) - (y , CPj ) = O.

k=1

Thus (y - w ) 1.. sp{CPI , CP2 , . ..} and since sp{CPI, CIl2, • . •} is dense in M , the con-
tinuity of the inner produ ct implies (y - w ) 1.. M. 0

1.18 Isometry of Hilbert Spaces

The existence of orthonormal bases for separable Hilbert spaces enables one to
identify all separable Hilbert spaces as follows .

Definition: Inner product spaces E and F are linearly isometric ifthere exists a
function A which maps E onto F such that for all u , v in E and a , {3 E C,

(i) A (au + {3 v) = aAu + {3Av;

(ii) IIAull = lIuli .
The operator A is called a linear isometry.

Theorem 18.1 Any two infinite dimensional separable Hilbert spaces (over C)
are linearly isometric.

Proof: Suppose 'H I and 1i2 are infinite dimensional separable Hilbert spaces with
orthonormal bases {cpI, CIl2, . • .} and {Vrl , Vr2, . . .}, respectively. For each u E H I,
1.18 Isometry ofHilbert Spaces 37

define Au = Lk(u, ({Jk}1{rk . The series converges by Theorem 12.3. Clearly, A

has property (i). By Parseval's equality

IIAull
2
= L I(u, ({Jk}1 2 = 2
lIull .
k

Finally, A maps 1t1 onto 1t2; for if y E 1t2, then

The same proofshows that every n-dimensional Hilbert space (over C) is linear
isometric to en .
Thus every finite dimensional subspace of a Hilbert space is
closed. 0

Corollary 18.2 Any infinite dimensional separable Hilbert space is linearly

isometric to l2 .

lt is easy to verify that the operator A defined in the proof of Theorem 18.1 has
the property that for all u, v in 'H I .

which means that the spaces 1t I, 1t2 from the point of view of Hilbert spaces,
practically do not differ.
Often one can give an explicit description of the operator that establishes the
linear isometry between two Hilbert spaces. To illustrate this let us consider the
spaces l2 (Z) and L2 ([ -lr, lr]) . Both are infinite dimensional separable Hilbert
spaces, and hence by Theorem 18.1 they are linearly isometric. To make this
connection explicit, let Fbe the map that assigns to each function I in L2 ([ -lr, lr])
the sequence ofthe Fourier coefficients of I with respect to the orthonormal basis
({In(t) = ~eint, n = 0, ±1, ±2, . ... Thus

_1_i
v2Jr

Jr
In = I(t)e- int dt .
-J2ii -Jr

r'
By Parseval's equality, the sequence F I is square summable, that is FIE l2 (Z),
and

Ilfll ~ (,too If,I' ~ IITfll ·

lt follows that the map I 1--+ F I is a linear operator which maps L2 ([ -lr, lr])
in a one to one way onto l2(Z) and preserves the norm. In the sequel we refer
to F as the Fourier transform on L2 ([ -lr, n ]) . Notice that F transforms the
orthonormal basis ((In(t) = Jkeintn = 0, ±l, ±2, . . . of L2([-lr , lr]) into the
standard orthonormal basis of l2(Z) .
38 Chapter 1. Hilbert Spaces

1.19 Example of a Non Separable Space

The Hilbert spaces which are usually encountered in applications are separable.
However, there is a very important non-separable inner product space which has
been a source of active research for more than fifty years. This is the space of
almost periodic functions .

Definition: A complex valued function which is continuous on R = (-00, 00)

is almost periodic if it is the uniform limit on R of a sequence of trigonometric
polynomials of the form L~=l akei)'kt, Ak real.
Let E be the set of almost periodic functions. Under the usual operations of
addition and scalar multiplication, E is a vector space over C.
It can be shown that if f and g are in E, then

(j, g) = lim - liT

r-s-» T 0
f(x)g(x) dx

exists and defines an inner product on E . Since {eiAt : A real} is an uncountable

orthonormal set, E is not separable (cf. Appendix 1).

Exercises I

1. For which a E JR (the set of real numbers) is {n~ }~l in £2?

2. Determine which ofthe following functions are in L2[0, 1]. Then compute
their norms.

(a) .Jt~+l
(b) c!:at' a E JR
(c) rJ" a E JR
(d) t" eat, a E JR, n E N (the set of positive integers).

3. Which of the functions in problem 2 are in L2[I, oo)?

4 . Compute the rj-innerproduct of s = {(~)n}~l with 11 = {(i+i;.J2)n}~I .
5. Check if IIx + yll = [x] + lIyll where x, y in £2 are given by
(a) x = (0, ! 'i , ); y = 0 ,0,0, . . .).
(b) x = 0, 1, ~, );y = 3x.
6. Checkifllx + Yll2 = IIxll 2 + IIYII 2 forthefollowing pairs x, Y in L2[0 , 1].
(a) x(t) = t ; yet) = -t
I . O< t < l 0. os t s !
(b) x(t) =
I ' - - 2
0·' 12 < t -< 1
yet) =
1!
'
1; < t S 1
Exercises I 39

7. Let H be a Hilbert space.

(a) Describe all pairs of vectors x , y for which

IIx + YII = IIxll + IIYII·

(b) Describe all pairs of vectors x , Y for which

8. Let C[a , b] be the vector space of all continuous complex valued functions
on[a, b]. Intoduceanonn 11 ·11 onC[a, b]by 1If11 = maX/E[a ,b] If(t)l . Show
that it is impossible to define an inner product on C[ a, b] such that the norm
it induces is the same as the given norm .
9. Find the intersection of the ball liz - ~oll :5 1 with the line Z = A~I, where
~o = (1, t, ! ,i ,·· .),~I = (1,0, . . .).
10. Find the intersection of the line Z = A({J with the sphere liz - ({Joll = 1 in
L2[0 , 1], where ({Jo(t) = 2t 2, ({J(t) = t.

11. Find the intersection ofthe ball liz - ~o II :5 1 with the plane z = a~1 + (J~2 ,
where

~o= (1,~ ,~, ...), ~I =(1,0,0, ...) and ~2=(0,I,0, .. .).

12. Find the intersection ofthe plane z = a({J1 +{J({J2 with the sphere IIz-({Joll =
1 in L2[0 , 1], where

!
i., 0 -<t<1
- 2
({Jo(t) == y E C, ({JI(t) =
-I ·1<t<I
, 2 -

!
1
2'
. 0 -< t -
< 41 ' 2
1 -
< t -
< 4l
({J2(t) = I. I I 3
-2' 4 < t < 2' 4 < t :5 1

13. Is the intersection ofthe two balls II~I -zil :5 RI and 1I~2 - z] :5 R2 empty?

(a) ~I = (v"(L1)!)nEN, RI = t;~2 = (1,0,0" . .), R2 = t·

(b) ~I = (~y'il)nEN' RI = ! ;~2 = (Jz, 0, 0, . ..), R2 = t·

(c) ~I = (1,0,0,0, 8\ ' 2~3 ' 719' " .), RI = t; ~2 = (0, t,! ,-f.;, 0,
O, . . .) , R2 = 3-v'2
2v'2 '
40 Chapter I Hilbert Spaces

14. (a) Show that the system (Zj}f is linearly independent in L2[0 , 1], where

0; 0 S t < j~1
Zj(t)= . I .
\ 1, H I < t S 1

(b) Let Zj(t) = tje- a t , j = 1,2, . . . . Show that the system is linearly
independent in L2[0, 00).
15. Let W = (WI, W2, . . .), where Wj > O. Define l2(W) to be the set of all
sequences ~ = (~I , ~2 , ... ) of complex numbers with L}:I Wj I~j 12 < 00.
Define an inner product on l2 (w) by
00

(~ , 1)) = LWj~jijj .
j=1

Show that l2(W) is a Hilbert space .

16. (a) Findavectorwsuchthat(l,-b,fr , .. .)¢l2(W) .
(b) Find a vector W such that the set of all ~ = (~I, ~2, . . .) with I~n 1< n"
is in l2(W) .
17. Let g(t) be continuous and strictly positive on [a, b]. Give an inner product
on L2[a , b] such that

b )1/2
1If11 = (1g(t)lf(t)1 2 dt
18. Let H be a Hilbert space. Denote by l2(H) the space of all sequences
(Xj)jEN of vectors in H with Lj=1 IIXj 11 2 < 00. Define the inner product
(x, y) = L}:I (Xj, Yj). Show that l2(H) is Hilbert space .

19. Let z(l) = (zil) , zi 2), . ..), Z(2) = (z? ), zi2), . ..), , z(n) = (zin) , zin) , . . .)
be any n vectors in l2 . Take z(i) (k) = (zii), , zii) , 0, 0, . . .), for i =
1,2, . .. , n.
(a) Iffor some k the vectors (z(l) (k), .. . , z(n)(k)} are linearly independent,
show that (z(l) , . . . , z(n)} are linearly independent too.
(b) Check if the converse statement is true .

20. Check that the following sets are closed subspaces of l2 .

(a) A = {(~I, 2~1, ~3, 4~3 , ~5 , ~6, ~7, .. . )1 L}:I I~j 12 < oo} .
(b) B = {(~I , 0, ~3 , 0, ~5, 0, .. ·)1L}:I 1~2j_112 < oo} .
(c) C = {CO, b 0, ~4, 0, ~6, ... )1 L}:I 1~2j 12 < oo} .

21. Find the pairwise intersections between the subspaces defined problem 20.
Exercises I 41

22. Let rl, ri. rs E C be such that Iril < 1, i = 1,2,3; and rj :j:. ri o Let
Xi = (1 ,ri ,rf, . . .) . Prove that ~ = (~1 ,b~3, .. .) is in the subspace
spanned by XI, X2 and X3 if and only if ~n+3 - (rl + r2 + r3)~n+2 + (rl r: +
rlr3 + r2r3)~n+1 - rlr2r3~n = 0.
23. Let X = (1, !' t, k, /6' ...),
Y = (1, 2.!, 3·t, 4·k, .. .),
z = (2,3 .2.!,
t, k, ...)
4.3. 5.4. be vectors in £2. Prove that ~ = (~I, ~2, .. .) is in the
subspace spanned by x, y and z if and only if
331
~n+3 - 2~n+2 + 4~n+1 - 8~n = 0.

24. Let fl(t) = e', h(t) = e it and f3(t) = e:" , Prove that a vector y E
L2 [a, b] is in the subspace spanned by fl, hand f3 ifand only if y satisfies
the differential equation
y(3) _ /2) + y' _ y = 0.

25. (a) Prove that a vector y E L2[a, b] is in the subspace spanned by {et , te';
e- t } if and only if y satisfies the differential equation /3) - y(2) - y' +
y =0.
(b) What is the dimension of this subspace?
26. (a) Let YI and Y2 be twice continuously differentiable functions in L2[a, b]
with det ( Y) Y;):j:. 0. Prove that Y is an element of the subspace
YI Y2
spanned by YI and Y2 if and only if

det (~;Y;' y~~~ v"

;') == 0.

(b) What is the dimension of that subspace?

27 . Show that the set of all vectors in £2 of the form ~ = (~I , ~I, ~3, ~4, ) is
the orthogonal complement of the subspace spanned by (1, -1, 0, 0, ).
Show that any vector x E £2 can be represented in the form x = XI + X2,
where XI is in sp(1, -1,0,0, . . .) and X2 is in the orthogonal complement.
Show that this representation is unique.
28. Show that the set ofall vectors in £2 ofthe form (~I, ~I, ~I , ~4 , ~5, ) is the
orthogonal complement of the subspace spanned by (1, -1,0,0, ) and
(0, 1, -1 ,0, . . .). Show that any X E £2 can be represented uniquely as X =
XI + X2, where XI is in the span of (1, -1 ,0,0, . ..) and (0, 1, -1 ,0, . . .)
and X2 is in its orthogonal complement.
29. Let xj = (1, -5,6,0, .. .),X2 = (0, 1, -5,6,0, . . .), . . . ,Xj = (0,0,0,0,
'-..-'
j-I times
1, -5 , 6, 0, ...), . .. . Prove that the following three statements are
equivalent:
42 Chapter I. Hilbert Spaces

1. ~ = (~I, ~z, .. .) is a vector in the orthogonal complement (in £z) of the

subspace Sp{XI , Xz , . . .}.
2. For all n E N, 6~n +z - 5~n+1 + ~n = 0.
3. ~ is a vector of the subspace spanned by (1, !-' ~, k, · · ·) and (1, t,!'
-d7 , . ..).
30. Find in £z the orthogonal complement of the subspace

31. Let to = (1, a , a Z, . . .) for a > 1. Denote by £z (w) the Hilbert space
consisting of all sequences ~ = (~I , b . ..) such that L~I a j -ll~jIZ <
00. Prove that s = (1, 1.,
a a
-\, .. .) E £zandfindtheorthogonalcomplement
in £z(w) of sp[s} .
32. Let XI = (1,2,0, ), YI = (1,0,0, );
Xz = (0, 1,2,0, ), YZ = (1, 1,0,0, );
X3 = (0,0, 1,2,0, ), Y3 = (1, 1, 1,0,0, );

be two given systems in £z. Prove that for all j, Yj tt: Sp{XI , Xz , . .. }.
33. (a) Prove that the orthogonal complement in £z of sp{y} , where Y
(1, -t, !' --d7, ...)
is Sp{XI, Xz, .. .)}, where

Xj = (0, . .. , 0, 1, 3, 0, .. .).
'-,--'
j-I times

(b) Decompose each vector with respect to sp{x} and its orthogonal
complement.
34. Prove that ~ E SP{XI , Xz, . . .}, where Xl, xz , . . . are as in problem 32 ifand
only if ~I = - L~I (-!-)j ~J+I . Therefore ~ is of the form (!-~z - ~~3 +
k~4 - ... ,b ~3 , . . .) .
35. Let xj = (a, {3, 0, 0, . . .), Xz = (0, a , {3, 0, . . .), X3 = (0,0, a , {3, . . .), . . . ,
where I~ I > 1. Prove that for all i, the vectors Yj from problem 32 are not
in the subspace Sp{XI, Xz , .. .} of £z.
36. Let XI, Xz, X3, . . . be as in problem 35. Prove that the orthogonal comple-
ment in £z to sp{(1, -~, ~~ , - ~~ , ...)} is SP{XI, Xz, .. .}.
37. Let XI, Xz , X3 , .. . be the vectors in £z as in problem 35. Prove that ~ E
sp{xI ,xz , ...[ if and only if s, = -L~I(-~) j~J+I so that s is of the
form
Exercises I 43

38. Let XI = (a, {3, 0, 0, . ..), X2 = (0, a, {3, 0, ), X3 = (0,0, a , {3, 0, . . .),
. . . , where I~I ~ 1. Check that SP{XI , X2, } = f2 .
39. Let XI = (1,0,0, .. .), X2 = (a , {3 , 0, ), X3 = (0, a, {3, 0, ...), .. . ,
where I~I > 1.
(a) Check that Sp{XI, X2 , . • •} = h
(b) Show that any finite system of these vectors is linearly independent.
(c) Find aI , a2, •.. in C, not all zero , such that L:~I a)x) converges to
zero .

40 . Let XI = (1,0,0, .. .), X2 = (a , {3, 0, . ..), X3 = (0, a, {3, 0, . ..), .. . ,

where I~I s 1.
(a) Check that SP{XI, X2 , ... } = h
(b) Show that each finite system of these vectors is linearly independent.
(c) Show that one cannot find al , a 2, . • . . not all zero, such that L:~I a)x)
converges to zero.

41.
L
et Xr (t) =
{ ° for r < t < 1
1 for 0:;; t < r

(a) Prove that sP{Xr}re[O,lj = L2(0, 1).

(b) Prove that any finite number of these functions with different rare
linearly independent.
(c) Prove that sP{Xr}reQn[o.lj = L2(0, 1), where Q is the set of rational
numbers.

42. Let {Xi} be the vectors in problem 32. For every ~ E f2, find the projection
of ~ into SP{XI , X2 , . • .} and find the distance from ~ to this subspace.
43. Does the line z = AX intersect the ball liz - aoll ~ Ro?
(a) X = (1, ~, ~, .In,...), ao = (0, 1,0, . ..), Ro = ~ .
(b) X = (1, 0, ~,O, i, 0, ~' " .), ao = (0, 1, 0, ~, 0, i, ...), Ro = ~.
44 . Suppose rpl , ... , rpn is an orthonormal system in a Hilbert space 11. Let S
be the ball {z : liz - zo] ~ Ro}, and let M = Sp{rpl, ... , rpn}. The distance
d between Sand M is defined by d = inf'[]« - vII : u E S, v EM} . Prove

t
that

d ~ max { (11Z01l 2 - Ilzo, ~j)l2) 1/2 - Ro, 0 I.

45 Let Sand M be as in problem 44. Prove that the intersection of S with
M is a ball in M with L:J=I (zo, rp)}rp) as the center and (R5 - IIzoll 2 +
L:J=I I(zo, rp)}1 2) 1/2 as the radius.
44 Chapter I Hilbert Spaces

46. Let liz - zi] = R and liz - z211 = R be two spheres with IIzlli = IIz211 .
Show that the intersection of the two spheres is a sphere in the subspace
orthogonal to the vector Z2 - ZI . Find its radius and its center.
47. Prove that Idet(ajk)1 2 = n;=1 L:Z=I lajk l2 if and only if the vectors YI =
(all , a\2, . . . , al n), Y2 = (a21, a22, ... , a2n), . . . , Yn = (ani, an2, . . . , ann)
are orthogonal to each other.
48. Consider the incompatible systems of equations:

(a) 1 = 2Xl + 3X2 (b) 1= Xl - X2

0= Xl + X2 -1 = 2xI + X2
2 = 3xI - X2 -1~ = XI + 2X2
Minimize the deviation between the right and left sides .
49. Let the points (0, 0), (1, 1), (2, 1) be given. Find the polynomial P(t) of
degree 1 with the least squares fit to these three points.
50. Let ZI, . . . , Zn be vectors in a Hilbert space H .
(a) Prove that if ZI , .. . , Zn are linearly independent, then there exists an
e > 0 such that any vectors YI, . . . , Yn in H which satisfy

IIzi - Yi II < e, i = 1, . .. , n,

are also linearly independent. Hint : Consider the Gram determinant.

(b) Let Z I, . .. , Zn be linearly dependent. Determine whether or not there
exists a 8 > 0 such that if YI, .. . , Yn satisfy II zi - Yi II < 8, then the
system YI , .. . , Yn is linearly dependent too.

51. Let A = {ai, . . . , an} be a system of vectors in a Hilbert space H . For any
Y E H, let YA denote the projection of Y into the subspace sp{al , . . . , an}.
Prove that for any e > 0, there exists a 8 > 0 such that for any system of
vectors B = {bl , . .. , bn } with the property lIa j - b j II < 8, 1 .:s j .:s n, the
inequality
IIYA - YBII .:s sHYl1
holds for any Y E H. Hint: Use the formula for YA.
52. Let

Ym = amlAI + ...+ amnAn

where m > n, be an incompatible system. Let aA = (ai, . . . , an) be its
solution as defined in Section 6. Prove that for any e > 0, there exists
a 8 > 0 such that if ZI , •• • , Zn and {bij} ;=I •...•n satisfy IIzk - yd < 8,
J= I •...•m
Exercises I 45

1 .:s k .:s m, lIaij - bijll < 8, 1 .:s i .:s m, 1 .:s j .:s n, then the solution
!JB = (!JI, .. . , !In) of the incompatible system
ZI = bllAI + ... + blnA n

satisfies
liZ
L lai -
n )
z
(
!Jil < s.
1=1

53. Let tl, ... , tk be k points in [0, 1] and let YI, . .. , Yk be in Co Let P(t) be

°
their least square fit polynomial of degree n < k. Prove that for any e > 0,
there exists a 8 > such that if SI, . . . , Sk in [0, 1] and ZI, . .. , Zn E C
satisfy lSi - til < 8, i = 1, . .. , k ; IZi - Yil < 8, i = 1, . . . , k , then their
least square fit polynomial Q(t) (degree < k) satisfies

[1 1
IP(t) - Q(t)IZdtf IZ < e.

54. Let Lo = {cp E Lz[ -a, a] I cp(t) = -cp( -t) a.e.}

L E = {cp E Lz[-a ,a] Icp(t) = cp(-t) a.e.} ,
(a)Show that both sets are closed infinite dimensional subspaces of Lz [-a , a].
(b)Show that Lo and L E are orthogonal.
(c)Show that L E is the orthogonal complement of Lo.
(d)For 1 E Lz[ -a , a], find its projections into Lo and LE.
(e)Find the distances from l(t) = t Z + t to Lo and to LE. Find the
distances from any 1 E Lz [-a , a] to Lo and L E.
55. Let NI = {(~I, ~I, ~I, ~I , ~z, ~z , ~z, ~z , . ..)}
Nz = {(~I , i~l, -~I, -i~l, ~z , i~z , -~z , -i~z, ·· .)}
N3 = {(~I, -~I, ~I, -~I , ~z , -~z , b -b·· .)}
N4 = {(~I, -i~l , -~I, i~l, b -i~z , -~z , i~z, .. .)}
(a) Prove that they are closed infinite dimensional subspaces of £z .
(a) Check that they are mutually orthogonal.
(c) Show that NI E9 N: E9 N3 E9 N4 = fz.
56. Let {cpI , cpz, . ..} be a set of vectors and let {WI, WZ, • . •} be an orthognal
system. We call {WI , wz, .. .} a backward orthognalization of {CPI , CPZ, . . .}
if sp{cpj, CPHI, .. .} = sp{Wj, WHI, . . .}for j = 1,2, .. .. Prove that there
exists a backward orthogonalization for {CPI , CPZ , .. .} if and only if for every
j we have CPj ¢ Sp{CPHI, CPHZ , · · .},
57. (a) Let CPI = (1,2,0,0, .. .), v: = (-1,2,0,0, . . .); CPj = ej for j =
3,4,5 , . .. , where {ej} is the standard basis for £z. Construct a back-
ward orthongonalization in £z for {cpj heN.
46 Chapter I Hilbert Spaces

(b) Let XI, .. . , Xn be in en. Define (Xi) j to be the ph component of Xi.

Set
~I = «Xlh , (xdz , . · · , (xdn, 0, )
~z = «xzh, (xzh,···, (XZ)n, 0, )

~n = «Xn)l , (xnh , . .. , (xn)n , 0, . ..)

~j = ej for j > n.
(i) Which conditions should be imposed on XI, .. . , Xn, so that there
exists a backward orthogonalization for {~j }jeN?
(ii) Construct, in this case, the backward orthogonalization.

58. Find the Fourier coefficients of the following functions :

(a) f(t) = t
(b) f(t) = t Z
(c) cos at , a E ~\Z (Z is the set of integer)
(d) f(t) = {I ; t2:0
-1; t < 0
(e) f(t) = It I
(t) Use the Parseval equality to prove that L~I dr
(Hint:
Consider t .)
it
(g) Use the Fourier expansion ofcos to prove that 2+4 L~I i=~~n2 = Jr.
59. For f E Lz[ -Jr, Jr], find the projection of f into sp{e- int, .. . , eint} and
find the distance from f to that subspace.
60. (a) Find the Legendre coefficients ao, ai, az (i.e., the Fourier coefficients
with respect to the normalized Legendre polynomials) of the following
functions
(i) f(t) = t Z
(ii) cos~t
(b) Prove that for every m-times differentiable function f , the Legendre
coefficients ak, k S m, are

~ 11
1
ak = k +-
- l - f(k)(t)(1 - tZldt.
k 2kk! -I
61. (a) In Lz[ -1, 1], find the projection of x n into sp{x n- I, x n-Z, ... ,I}.
(b) Express x n = ~(x) + 1/r(x), where
~(x) E sp{x n- I, .. . , I} and 1/r(x) E sp{x n- J , ••• , l}ol.

62. Consider the two vectors cos t and cos t + sin t in Lz[ -Jr, Jrl Change the
inner product on Lz [-Jr, Jr] in such a way that it remains a Hilbert space
and these two vectors become orthogonal.
Exercises I 47

63. Consider the vectors (1,2,0,0, . ..) and (1, 1, 1,0,0, . . .) in h Change
the inner product on £2 such that it remains a Hilbert space and these two
vectors become orthogonal.
64. In general, given linearly independent vectors CPI, ... CPn in a Hilbert space
H, change the inner product on H such that it remains a Hilbert space and
cP' , .. . , CPn became orthogonal.
65. Let L be a closed subspace of a Hilbert space H. Given gEL and f E H,
denote by Pt. f the projection of f into L. Prove that g ..1 PL f if and only
ifg..l f.
66. Prove that for any two subspaces of a Hilbert space H,
(a) (L, + L2).1 = Lt n Lt
(b) (L, n L2).1 = Lt + Lt

67. Generalize problem 66 to the case of n subspaces.

68. Set L, = sp{l, 2, 0, ), (0,1,2,0, .. .), (0,0,1 ,2,0, . .. ), .. .)
L2 = sp{(1, 0, 0, )}.
Prove that LI + L2 is dense in £2.
69. Let L, = sp{l , 2, 0, ), (0,1,2,0, ), (0,0,1,2,0, ), )
L2 = sp{l, 3, 0, ), (0,1 ,3 ,0, ), (0,0,1,3 ,0, ), )
be two subspaces in £2.
(a) Prove that ~ E L, n L2 if and only if ~ is orthogonal to all vectors
(rJ' , rJ2, ...) E £2, where 6rJk+2 + 5rJk+' + rJk = O.
(b) Prove that L, n L2 = sp{(l , 5, 6, 0, . . .), (0,1 ,5,6,0, ...), . . .).

70. Define£2(N x N) to be the set ofall double sequences {~jk} with L:rk=1 I~jd
< 00, and an inner product defined by
00

«, rJ) = L ~jkTjjk '

j ,k=1

(a) Prove that it is a Hilbert space .

(b) Let cP(l), cp(2) , ... be an orthonormal basis in £2, cp(j) = (X?»)~I' Set

(ij) = (x(i) x(j»)oo

cP k p k.p=I '

Prove that {CP(ij)}0=1 is an orthonormal basis for £2(N x N).

71. Determine which of the following systems are orthogonal bases in £2 and
which are not.
(a) (1,2,0,0, . . .), (0,0,1,2,0, . . .), (0,0,0,0,1,2,0, . . .), . . .
48 Chapter I. Hilbert Spaces

(b) (1 , -1, 0, 0, . ..), (1, 1,0,0, . . .), (0,0,1, -1, 0, 0, .. .),

(0,0,1 ,1 ,0,0, . . .), . . . .

72. Given vectors <PI , . . . , <Pn in a Hilbert space H, when is it possible to find
a vector xo E H such that (XO, <PI) = 1 and (XO , <Pj) = 0 for j > I? Ifit
exists, find such a XO .
73. Given vectors <PI, ... , <Pn in a Hilbert space H,
(a) prove that there exist vectors XI, . . . , Xn with (Xj, <Pk) = 8jk, if and
only if <PI, . . . , <Pn are linearly independent. Such a system XI, . . . , Xn
is called a biorthogonal system.
(b) Let rpj be the projection of <Pj into sp{<PI , .. . , <Pj -I , <P j+ I, ... , <Pn}.
Prove that the system {Xj}'j=l' where Xj = lI <PjIl 2~IIq;jIl2(<Pj - rpj) is a
biorthogonal system.
(c) How many biorthogonal systems corresponding to <PI, ... , <Pn exist?

74. Given a system of vectors {<pj }~I in a Hilbert space H,

(a) prove that there exists a system {Xj }~I with (Xj, <Pk) = 8jk if and
only if
<Pj ~ SP{<P;}i;fj .
(b) Prove that the formula in 73b gives a biorthogonal system in this case
too .

75. Let An = (ajk)'j,k=O be a positive definite (n + 1) x (n + 1) matrix with

complex entries . Show that
n
(x , Y}An = L ajkxkYj·
j ,k=O

defines an inner product on C n +l . Denote by CAn the linear space of all

vectors in C n +1 endowed with this inner product. Prove that CAn is a
Hilbert space.
76. Consider the vectors ek = (8kj)'j=o' k = 0, . . . , n, inCAn • Letxn , XI, . . . , Xn
be the vectors in CAn obtained from eo, .. " en by the Gram-Schmidt
orthogonalization procedure in the inner product (, , .) An ; see the previ-
ous exercise. Show that for each k the vector Xk has the form
(k) (k)
Xk = Ck(~O , .. . , ~k , 0, 0, . . . , 0),

where Ck E C and
Exercises I 49

77. Let the matrix An in the previous exercise be replaced by the positive definite
Toeplitz matrix Tn = (tj-kYJ,k=O' Prove that the polynomial

with the coefficients given by (*) in the previous exercise, has all its zeros
inside the unit disc.
78. Let T; be as in the previous exercise, and

Prove that the polynomial

snO + Sn1 Z + ... + Sn« z"

has all its zeros inside the unit circle .
79. Let a E C and laI =1= 1. Prove that the weight
W(A) = a ): + (1 + lal2 ) + CiA- 1
is strictly positive on 1[', the unit circle in the complex plane, and determine
the Szego polynomials for this weight.
80. Consider the matrix
a a2 ...
A, ~ i
(
a ...
a' ) .
a':1
bn b n- I b n- 2

(a) Prove that A n is invertible if and only, if 1 - ab =1= O.

(b) Assume 1 - ab =1= O. Prove that A -I = (fjdJ ,k=O where

100 = (1 - ab)-I, Inn = (1 - ab)-I ,

Ikk = (1 + ab)(1 - ab)-I , k = 2, 3, .. . , n - 1,
Ik-I ,k = -(1 - ab)-I a , k = 1, , n,
Ik ,k-I = -(1 - ab)-Ib , k = 1, , n,

and fjk = 0 for the indices Ij - kl ::: 2. In particular, A;I is a

tridiagonal matrix.

81. Fix a E C, lal < 1. Consider on 1[' the weight

-1 -lal 2
W(A) = _ .
-1 + lal2 - a): - aA- 1
50 Chapter I Hilbert Spaces

(a) Prove that w(A) > 0 for each A E 1[' and that
00 00

w(e it) = I + Laieijt + Liiie-ijt.

i=l i=l

(b) Construct the Szego polynomials for the weight w (hint: use the
previous exercise).

82. Let the weight w be given by

where qn(A) = An + an_lA n- 1 + ... + ao with qn(A) =1= 0 for II..I 2: 1.

Prove that up to a constant the (n + r)-th Szego polynomial for w is equal
to Ar qn(A) .
83. Let lal < I, and consider the weight

2
w()..) = I fah _i1
J=o
Compute up to a constant the Szego polynomials for this weight.
Chapter II
Bounded Linear Operators on Hilbert Spaces

In this chapter, continuous linear functions defined on a Hilbert space are

introduced and studied. These functions are described by infinite matrices in the
same way as linear transformations on en
are represented by finite matrices . In
this way the chapter may be viewed as a beginning of a theory of infinite matrices .
As may be expected , analysis plays a very important role.
Throughout this chapter, H , HI and H2 denote Hilbert spaces over the complex
numbers C.

2.1 Properties of Bounded Linear Operators

Definition: A function A which maps HI into H2 is called a linear operator if

for all x, y in H I and a E C,

(i) A(x + y) = A(x) + A(y) ;

(ii) A(ax) = aA(x).

For convenience, we write Ax instead of A(x) .

Taking a = 0 in (ii) gives AO = O.
For example, with each n x n matrix of complex numbers there is a natural way
to associate a linear operator A mapping into en
namely en ,

where
al) (f3I) L aijaj o
(
n
(aij) : = : i.e., f3i =
an f3n J=I

Definition: The linear operator A : HI -+ H2 is called bounded if

sup [A x] < 00 .
IIx II :::: I

The norm of A, written II A II , is given by

IIAII = sup [Ax]' .

IIxll::::1
52 Chapter II. Bounded Linear Operators on Hilbert Spaces

When there is no cause for confusion, we use the same notation for the norms and
inner products on 'HI and 'H2.

An operator A is bounded if and only ifit takes the l-ball SI with center zero in
'HI into some r-ball in 'H2 . The smallest ball in 'H2 which contains ASI has radius
IIAII·
The identity operator I : 'HI ~ 'HI defined by Ix = x is obviously a bounded
linear operator of norm 1.
The operator A : en ~ en defined above is bounded; for if x = (eq , ... , an),
then by (1.3) in Section 1.1

IIAxll
2~ EIPoI' -s (~I la.I') (~IOjl') = IIxll
2;~l l agl',
from which it follows that II A II S (L~j=1 lay 1/2. e)
It is easy to verify the next three properties of II A II .
(a) IIAII = sUPx;60 II
li
ltll
= sUPllxll=1 [Ar],
(b) IIAII = sUPllxll=lIyll=I!(Ax , Y}I = sUPllxll:ol ,lI yll :oI!(Ax, Y}I· (1.1)
(c) If II Ax II s CJIxll for all x E 'HI, then A is bounded and IIAII s C.
The set of bounded linear operators whick map 'HI into 'H2 is denoted by
£('HI, 'H2)' If'Hl = 'H2, we write £('Hd instead of £('HI, 'HI) .
If A and B are in £('H I, 'H2) , it is a simple matter to verify (i}-(v).
(i) aA + f3B E £('HI , 'H2); a, f3 E C.
lIaAIl = lalliAII; a E C.
(ii)
(iii) IIA + BII s IIAII + IIBII·
(iv) IIA - BII 2: IIiAII- IIBIII ·
(v) If C E £('H2, 'H3), where 'H3 is a Hilbert space, then defining CA by
CAx = C(Ax) , it follows that CA is in £('HI , 'H3) and II CA II S IICJ1I1A II .

2.2 Examples of Bounded Linear Operators

with Estimates of Norms

While it is often easy to verify that a linear operator is bounded, it may be very
difficult to determine its norm. A few examples are discussed in this section.
1. Every linear operator which maps a finite dimensional Hilbert space into a
Hilbert space is bounded. Indeed, suppose A is a linear map from 'HI into 'H2 ,
and dim 'HI < 00 . Let CPI, ... , CPn be an orthonormal basis for 'HI. Then for any
x E 'HI ,
n n
x = L(x , CPk}CPk and Ax = L(x , CPk}Acpk .
i= 1 k=1
2.2 Examples ofBounded Linear Operators with Estimates ofNorms 53

Hence

which implies that

Suppose, in addition , that HI = H2 and A({Jk = Ak({Jk, I :s k :s n. Then

n
21( 2 21lx1l 2
= L IAkI x , ({Jk}1 s M ,
k= I

where M = maxr jAk I. Hence II A II :s M . On the other hand, suppose M = IAj I.

Since lI({Jj II =
1,

Thus
IIAII = max IAkl·
k

2. Let {({JI , ({J2, . . .} be an orthonormal basis for the Hilbert space H and let
{AI , A2, . . .} be a bounded sequence of complex numbers. For each x E H , define

Ax = L Ak(X, ({Jk}({Jk.
k

A is a linear operator on H and by Bessel's inequality,

2 21l 2
IIAxII = L IAdl(x , ({Jk}!2 s m x1I ,
k

where m = sUPk IAkl . Thus IIAII :s m . On the other hand, given 8 > 0, there
exists a Aj such that IA j I > m - 8. Hence

Since 8 was arbitrary, II AII ~ m. Therefore II AII = sup, IAkl .

54 Chapter II. Bounded Linear Operators on Hilbert Spaces

We shall see in Chapter IV that there is a large class of operators A of the form
given above.
3. Certain infinite matrices give rise to bounded linear operators on £2 as follows :
Given an infinite matrix (aij)0=1 ' where

00 00

L L laijl2 < 00, (2.1)

i=1 )=1

A(al, a2, ...) = (f3) , tho ...),

where

L
00

i.e., fJi = aija).

)=1

The operator A is a bounded linear operator on £2 and

00 00

IIAII 2 S L L laijl2
i = 1 )=1

since

implies that for x = (aI, a2, . . .),

00 00 00
2
IIAxll = L IfJil 2
S IIxll 2
L L laijl2.
i= 1 i=1 )=1

Condition (2.1) is not a necessary condition for A to be bounded since the identity
matrix (aij) = (8ij) does not satisfy (2.1), yet A = I .
No conditions on the matrix entries aij have been found which are necessary
and sufficient for A to be bounded, nor has II A II been determined in the general
case .
4. Let H = L2([a, b]) and let a(t) be a complex valued function which is
continuous on [a, b]. Define A : H -+ H by

(Af)(t) = a(t)!(t).
2.2 Examples ofBounded Lin ear Operators with Estimates of Norms 55

A is linear and for M = m ax a :5t::,:b la(t )l,

2
II Afll = l b
2
la(t) f( t) 1 dt ::s M 2 11 f11 2 .
Thus II AII ::s M. To show that II AII = M , suppose M = la(to)l. Define a sequence
{q:>n} in 'H by

Now

lIq:>nll
2
= 1 10+* n
10- 1
n
-dt
2
= 1,

and from the continuity of aCt) at to, it follows that

Hence
IIAII ::: la(to)1 = M.
5. Let k be a complex valued Lebesgue measurable function on [a , b) x [a, b)
such that

ll b b
2
Ik(t , s) 1 dsdt < 00 .

Define the operator K : L 2[a , b) ~ L 2([a , b) ) by

(Kf)(t) = l b
k(t , s )f (s )ds.

The integral exists since for almost every t , k(t , s) f (s) is Lebesgue measurable
on [a, b) and by Schwarz's inequality,

b ( b 2 ) 1/ 2 ( b ) 1/ 2
llk(t, s)f(s)lds::S l l k (t, s )1 ds l lf (S) ,2dS

Thu s

Hence

Clearly, K is linear.
56 Chapter II. Bounded Lin ear Operators on Hilbert Spaces

The operator K is called an integral operator and k(t , s) is called the kernelfunction
corresponding to K.
6. Let Sr: £z -+ £z be defined by

Sr is called theforwardshiftoperator. Obviously, S, is linear and IISrx II = [x], x E

£z· In particular, II s, II =]
7. Let Se: £z -+ £z be defined by

Se is called the backward shift operator. Obviously, Sr is linear and II Se II = 1.

Some authors call a forward shift a right shift and a backward shift a left shift.
8. Let V : £z (Z) -+ £z (Z) be defined by

V( . .. , ~-I , lliJ,~-I , ·· ·) = (. . . , ~-z, 1~-I I, ~o, · · .).

Here the box .. indicates the zero position in the sequence. The operator V is called
the bilateral (forward) shift on £z (Z) . Notice that II V x II = IIx II for each x in
£z(Z), and V maps £z(Z) in a one-one way onto £z(Z) . In fact, the inverse map is
given by
V- I (. • . ,X_I,~,XI, . ..) = (. .. , xo ,~, Xz,·· .),
which is referred to as the bilateral backward shift on £z(Z) .
Linear operators which stem from differential equations are neither defined
on all of Lz([a, b]) nor are bounded on those subspaces of Lz([a , b]) on which
they are defined. For example, the operator Af = 1x
f is not defined for all
f E Lz([-n, n]) . If we want A to map a subspace V(A) of Lz([-n, n]) into
Lz([-n, n]), we could take

However, A is unbounded on V(A); for if 'fin (x) = J..::-e inx , n = 1,2, . . . , then
" Zrr
II 'fin II
= 1 but II A 'fin II = nll'finll = n.
Fortunately, there is a well developed theory of unbounded linear operators
which enables one to treat large classes of differential operators. The reader is
referred to [DS2], [G] and [GGKI] for an extensive treatment of the subject.

2.3 Continuity of a Linear Operator

Whereas a bounded real valued function can be discontinuous at every point, the
situation for bounded linear operators is vastly different as the next theorem shows.
2.4 Matrix Representations ofBounded Linear Operators 57

An operator A mapping HI into H z is said to be continuous at Xo E H I if

Xn -+ Xo implies AXn -+ Axo.
Thus A is continuous at Xo if and only if for every E > 0, there exists a 8 > 0
such that IIx - xoll < 8 implies IIAx - Axoll < c.

Theorem 3.1 Let A be a linear operator whi ch maps HI into H z. The f ollo wing
statements are equivalent.

(i) A is continuous at a p oint.

(ii) A is uniformly continuous on H I.
(iii) A is bounded.

Proof: (i) implies (ii) . Suppose A is continuous at xo. Given e > 0, there exists a
8> 0 such that IIx - xoll < 8 impl ies II Ax -Axoll < c . Thus ifllw - zll < 8, w, z
in HI , then IIxo - (xo + z - w)1I < 8. Therefore,

II Aw - Az il = II Axo - A (xo + z - w) 1I < c.

(ii) imp lies (iii). There exists a 8 > 0 such that

IIAz ll = IIAz - AOII ::: 1, [ z] ::: 8. (3.1)

Suppose [x] = 1, x E HI. Then 118xll = 8 and by (3. 1),

1 ~ IIA(8x)1I = 8IIAxll,

or IIAxll ::: i· Hence II AII ::: i·

(iii) implies (i). Given e > 0 and r ., E H I, if ].r - xoll < I+" AII ' then

II Ax -Axoll ::: II All llx - xoll < c.

o
2.4 Matrix Representations of Bounded Linear Operators

We have seen in Section 2 Example 3 how certain matrices give rise to bounded
linear operators. In this section it is shown how to asso ciate a matrix with a given
bounded linear operator on a separable Hilbert space .
Suppose A E £ (H ), where H is a Hilbert space with orthonormal basis ({II ,
({Iz, . . . . Then for x E H , x = L j (X, ({Ij)({Ij. It follows from the linearity and
continuity of A, appl ied to the sequence of part ial sum s of this serie s, that

Ax = L (x, ({Ij)A({I j. (4.1)

j
58 Chapter II. Bounded Linear Operators on Hilbert Spaces

Now
A({Jj = L(A({Jj, ({Jk}({Jk. (4.2)
k

Combining (4.1) and (4.2) gives

Ax = L L(x , ({Jj}(A({Jj, ({Jk}({Jk = L L(x, ({Jj}(A({Jj, ({Jk}({Jk . (4.3)

j k k j

Thus if Ax = y, then (4.3) implies

This matrix equation leads to the following definition.

Definition: Let ({Jj, ({J2, ... be an orthonormal basis for a Hilbert space 1t and let
A be in £(1t). The matrix (aij) corresponding to A and ({Jj, ({J2, ... is defined by
aij = (A({Jj, ({Jj) .

Examples: 1. Let S, E £(£2) be the forward shift given by

Sr(aj ,a2, ...) = (0,aj ,a2, .. .).

If ej , e2, .. . is the standard orthonormal basis for £2, then

Thus the matrix corresponding to S; and ej , ez , . .. is the lower diagonal matrix

2. Let H = L2([-Jr, Jr]) and let aCt) be a bounded complex valued Lebesgue
measurable function on [-Jr, Jr].
Define A E £(1t) by
(Af)(t) = a(t)!(t) .
The "doubly infinite" matrix {ajdfk=-oo corresponding to A and the orthonormal
basis ({In(t) = b eint, n = 0, ±1, ±2, . . . is obtained as follows. Let

an = - 1
2Jr
l -1T
1T

a(t)e- ln
• t dt.
2.4 Matrix Representations ofBounded Linear Operators 59

Then

Thus the matrix is

which is called a Toeplitz matrix . The entry ~ is located in row zero column
zero .
3. Let K be the integral oprator on L2([a , b]) with kernel funct ion k E
L 2([a , b] x [a, b]) as in Section 2, Example 5. Given an orthonormal basis {lpj}
for L 2( [a , b]) , it follows from the remark at the end ofthe proof ofTheorem L15.1
that <l>ij (s , t ) = lpi(S)lpj (t ) , 1 :'S i , j , forms an orthonormal basi s for L 2([a , b] x
[a , b] ). Now the matrix (aij) corresponding to K and {lpi} is given by

aij = (K lpj , lpi ) = llb b

k (t , s) lpj(s) ;Pi(t) ds dt = (k , <l>j i).

Thus we see that each aij is a Fourier coefficient of k with respe ct to {<I>ij }.
Consequently, Parseval's equality gives

L laijl2 = L j(k, <l>ij) 12 = IIk 1l 2.

i,j i .]

So far we started with a bounded linear operator and considered its matrix repre-
sentation relat ive to an orthonormal basis. In some problems the reverse direction
is important: an infinite matrix is given and one looks for a bounded linear oper-
ator for which the given matrix is the matrix of the operator relative to some
orthonormal basis . In the sequel we shall use the following definition.
Let (aj kr;::k= l be an infinite matri x. We say that this matri x induces a bounded
linear operator A on l2 if (a jk)~k= 1 is the matrix of A with respect to the standard
orthonormal bas is el , e2, ... ofe 2, that is,

]
In this case we simpl y write

l
all a \2 a 13
a 21 a22 a23
..
.. ..
(4.5)
A = ~31 a32 a33
60 Chapter II Bounded Linear Operators on Hilbert Spaces

As a first example consider a diagonal matrix

Here Ojk is the Kronec1an delta and the diagonal entries are WI, W2, W3, . ... This
matrix induces a bounded linear operator Von £2 ifand only ifsuPj ~1 Iw j I < 00,
and in this case
IIVII = sup IWjl· (4.6)
j~ 1

Indeed, ifthe matrix diagtur, )~I induces a bounded linear operator Von £2, then

and hence SUPj~1 IWjl :s IIVII < 00 . Conversely, ifsuPj~1 IWjl < 00, then

is a bounded linear operator on £2 and the matrix of V with respect to the standard
orthonormal basis of £2 is the diagonal matrix diag( W j ) ~ I' Moreover,

and hence (4.6) holds.

Another example appears in parts of Section 11.2. Indeed, if
00 00

L L laijl2 < 00,

i= 1 j=1

then the matrix with respect to the standard orthonormal basis of £2 ofthe operator
A defined in Part 3 of Section IL2 is precisely equal to (aij)'ij=I '

2.5 Bounded Linear Functionals

Definition: A function f which maps a Hilbert space H into C is called a

functional. If f is in £(H , C), we refer to f as a bounded linear functional
on H .

The inner product on H gives rise to bounded linear functionals on H as follows .

For each Y E H, the functional fy defined on H by f y(x) = (x, y) is linear.
Moreover,
Ify(x)1 = I(x, y)l .s IIxIiIlYII·
2.5 Bounded Linear Functionals 61

Thus f y is bounded and II f y II .s II yII. Since

we have II f yll ~ IIYII and therefore

II f yll = lIy l!. YE 11.. (5.1)

In this section we prove the very useful result that every bounded linear
functional on 11. is an f y.
The motivation for the proof is based on the following observations.
Suppose f is a bounded linear functional on 11. and suppose there exists a
Y E 11. such that f (x) = (x , y ) for all x E 11.. In order to find this y, let
Ker f = (x : f(x) = OJ . Then 0 = f(x) = (x , y) for all x E Ker f ,
i.e., y 1- Ker f . It follows readily from the assumption that f is bounded and
linear that Ker f is a closed subspace oi H , The first step in finding y is to choose
v f= 0 which is orthogonal to Ker f (assuming f f= 0). The existence of such a
v is assured by Theorem 1.8.2. But for any a E C, y = av is also orthogonal to
Ker f . In order to determine which a to choose, we note that

af (v ) = f( av) = f (y ) = (y , y ) = aa llvll2 .

Thus we choose a = ffli, and y = a v is our candidate

Lemma 5.1 Iffis a linear fun ctional on 11. and f (xo) f= 0 fo r some Xo E 11., then
every x E 11. has the f orm

x = tho + z. f3 E C, Z E Kerf.

Proof: Take f3 = fg~) and Z = x - f3 xo. o

Riesz Representation Theorem 5.2 If f is a bounded linear functional on a
Hilbert space 11., then there exists a unique y E 11. such that for all x E 11.,

f (x) = (x , y ).

Moreover, IIfll = lI yll·

Proof: If f = 0, take y = O. Suppose f f= O. Then Ker f is a proper closed
subspace of 11.. Hence there exists a v f= 0 in (Kerf) .l . Let y = a v, where
a- w.
- !(v ) Th
en

(i) y 1- Ker f.
(ii) f(y) = (y , y ).
62 Chapter II. Bounded Linear Operators on Hilbert Spaces

Given x E H, we know from Lemma 5.1 that there exists a f3 E C and a Z E Ker
I such that x = f3y + z. From (i) and (ii) we get,
I(x) = 1(f3y) = f31(Y) = f3(y, y} = (f3y + z. y) = {x , y} .
To show that y is unique, suppose there exists awE H such that I (x) = {x , w}
for all x E H. Then
0= I(x) - I(x) = {x, y - w}
for all x E 'H. In particular, {y - w, y - w} = O. Hence y = w. Equation (5.1)
established that II I II = II y II · 0
If ({II , ({I2, ... is an orthonormal basis for 'H, then the vector y corresponding to
the functional I in the Riesz representation theorem is given by

y= L 1«({Ik)({Ik .
k

Indeed, since 1«({Ik) = «({Ik, y),

y = L{y, ({Ik}({Ik = L 1«({Ik)({Ik.

k k
Examples: 1. A functional F on L2 ([a , b]) is bounded and linear if and only if
there exists agE L2 ([a, b]) such that

FU) = l b
I(t)g(t)dt

for all IE L2([a , b]) . In this case, IIFII = IIgli .

2. A functional I on f2 is bounded and linear if and only if there exists a
y = (f3I, f32, .. .) E f2 such that for all x = (aI , a2, . . .) E f2,
00

I(x) = L akfik.
k=1

3. Let {Ad be a bounded sequence of positive numbers. For x = (aI , a2, .. .)

and y = (f3I, f32 , .. .), define a "weighted" inner product {, }w by
00

{x , y} w = L Akakfik.
k=1

It is clear that {, }w is an inner product on f2. Furthermore, f2 is complete with

respect to {, }w ifinf {Ad> O. The proof is essentially the same as the proof of
the completeness of f2. Hence a functional I is bounded and linear on (f2, ( , }w)
if and only if there exists a y = (f3I, f32, ...) E f 2 such that

I(x) = (x , Y}w = LAkakfik

k

for all x = (al ,a2, . . .) E f2 · Also, 1If11 = Lf:,1 Aklf3d .

2.6 Operators ofFinite Rank 63

2.6 Operators of Finite Rank

Definition: The image or range of A E £(11.1,11.2) , written Im A, is the subspace

A11. I = {Ax : x E 11. d.
lfIm A is finite dimensional, A is called an operator offinite rank and dim Im A
is the rank of A .
Let {VI, . . . , vn } and {WI , . .. , wn } be two systems of vectors in HI and 11.2,
respectively. Define K: HI -+ 11.2 by
n
Kx = L(x, Vi}Wi. (6.1)
i=1
K is linear and Im K C sp {WI , .. . , wn }. Thus the rank of K is at most n.
Moreover, IIKII :s I:7=llIvillliwili since
n n
IIKxIl .s L II (x, vi}wdl :s IIxll L II Vi II II Wi II·
n=1 i=1
It turns out that every K E £(11.1,11.2) of finite rank has the form given in (6.1).

Theorem 6.1 SupposeK E £(11.1, H2)isofrankn. Thereexist vectors VI , . . . , Vn

in 11. I and vectors CPI , . .. , CPn in 11.2 such thatfor every x E 11. I,
n
Kx = L(x , Vi}CPi.
i=1
The vectors tpI , . .. , CPn may be chosen to be any orthonormal basis for Im K.

Proof: Let CPI , . . . , CPn be an orthonormal basis for Im K . Then for every x E 11. I,
n
Kx = L(Kx, CPi}CPi . (6.2)
i=1
Now for each i, Ii (x) = (Kx, CPi) is a bounded linear functional on 11. I. There-
fore the Riesz representation theorem guarantees the existence of a Vi E 11. I such
that for all x E HI,

(Kx, CPi) = Ii (x) = (x, Vi), l:s i :s n.

Thus from Equation (6.2) we obtain
n
Kx = L(x, Vi}CPi .
i=1
It is clear that the representation of K is not unique. o
64 Chapter II. Bounded Linear Operators on Hilbert Spaces

Examples: 1. Let'H = L2([a, bD . The general form of an operator K E £('H)

of finite rank is given by

L Wj(t) 1!(S)Vj(s)ds,
n b
(Kf)(t) =
j=1 a

where {VI , . . . , V n } and {w I, . . . , wn } are arbitrary systems of functions in

L2([a, bD .
2. Let K E £(e2) be of rank 1. There exist v = Cfh, fh, .. .) and W =
(ai, a2, ...) in e2 such that for all x = (~I, b ) E e2,
Kx = (x, v)w = (cal, cai , ),

where c = L~ I ~i ~i' The vector Kx, considered as a column matrix, can

be written

2.7 Invertible Operators

One of the fundamental problems in operator theory is to solve, if possible, the

equation Ax = y, where A is a given linear operator and y is some vector. We
shall now consider this problem for some simple, but important, operators.

Definition: An operator A E £('H I, 'H2) is called invertible if there exists an

operator A -I E £('H2, 'Hd such that
A -I Ax = x for every x E 'H I

and
AA-Iy = y for every y E 'H2.
The operator A -I is called the inverse of A.

Clearly, A has at most one inverse. If A is invertible, then for each y E 'H2 there
exists one and only one x E 'H I such that Ax = y, namely x = A -I y. Conversely,
if for each y E 'H2 there exists a unique x E 'HI such that Ax = y, then A is
invertible. The proof that A -I is bounded is given in Theorem XII.4.1.
It is very useful to know that A -I is continuous. For example, suppose that the
equation Ax = y has a unique solution in 'HI for every y E 'H2. It might very
well be that this equation is too difficult to solve whereas Ax = y can be solved
x
rather easily for some y "close" to y. In this case the solution of this equation is
"close" to the solution x of the original equation if A -I is bounded since

IIx - xII = IIA-I y - A-IYII ~ IIA-I li ll y - Yll·

2.7 Invertible Operators 65

Definition: The kernel of A E £(1iI , 1i2), written Ker A , is the closed subspace
{x E HI : Ax = OJ.
A is called inje cti ve or one-one ifKer A = (0) .
A is injective if and only if A x =1= A z whenever x =1= z. Thus , if A is injective ,
then the equation Ax = y has at most one solution.
Suppose A is in £(C m ) and (aij) is the n x n matrix corresponding to A and the
standard basis for en. Then A is invertible if and only if det (aij ) =1= 0, in which
case the matrix corresponding to A - I and the standard basis is

I T
det(aij) (Aij) ,

where Aij is the cofactor corresponding to aij (cf. [S], p. 170).

Theorem 7.1 Suppose K E L(1i) is offinite rank, say

n
Kx = L)x , ({Jj)1/Ij'
j =1

The operator I - K is invertible if and only if

1
In this case. for every y E 'H,

a\2 al n
[all
a21 a22 a2n
(Y•
(y , •({J2)
1)
(I - K )-I y = Y - _ I _ det : (7.1)
det (aij) .
ani a n2 ann (y , ({In)
1/11 1/12 1/In 0

where aij = 8ij - (1/Ij , ({Ji) . Thus I - (I - K)-I is offinite rank.

The kern el of I - K consists of the set of all vectors of the form L::7= I ex j 1/1i-
where

(7.2)

Proof: Given y E 'H, suppose (I - K )x = y . Then

n
X - L(x, ({Jj)1/Ij = y . (7.3)
j=1
66 Chapter II. Bounded Linear Operators on Hilbert Spaces

Taking the inner product of both sides with ({Jk yields

n
(x, ({Jk) - ~)x, ({Jj)('l/Jj, ({Jk) = (y, ({Jk).
j=l

Hence the system of equations

n
ak - L a j('l/!j, ({Jk) = (y , ({Jk), 1 s k S n, (7.4)
j=l

has a solution, namely a j = (x, ((Jj), 1 s j S n.

Conversely, if ai , . . . , an satisfies (7.4), then guided by (7.3), we let
n
x=y+Larl/Jj· (7.5)
j=l

Then
n
(x, ({Jk) = (y, ({Jk) + Laj('l/Jj, ({Jk) = ai:

j=l

Hence
n n
(l- K)x =x - L(x, ({Jk)1/!k =x - L ak1/!k = y.
k= l k= l

To summarize, given Y E H, the equation (l - K)x = y has a solution if and

only if the system of equations (7.4) has a solution. Thus if det(aij) = D f:. 0,
where aij = 8ij - (1/!i - ({Ji ) , then by Cramer's rule, there exists a unique solution to
the system of equations (7.4) given by a j = ~, where D j is the determinant of
the matrix which is obtained by replacing the i" column of (aij) by the column

Once this unique solution ai , . .. , an is obtained, it follows from what we have

shown that (l - K)x = y has a unique solution and this solution is given by (7.5) .
Thus
1 n
X = Y+ D L D j1/!j, (7.6)
j=l

and defining (l - K)-l Y = x, a straightforward computation verifies that

(l - K)-l is bounded and has the representation (7.1) .
2.7 Invertible Operators 67

If det(aij) = 0, then taking y = 0 in (7.4), we have by Cramer's rule, numbers

al , .. . , an, not all zero, such that (7.4) holds and by (7.5), x = L:}= I a j 1/1j is in
Ker(/ - K). Now x :I 0, otherwise
n
0= (x, f{Jk) = Laj(1/Ij , f{Jk) = ai ; I ~ k ~ n.
j=1

Hence I - K is not invertible.

It also follows from the above argument that if x is in Ker(/ - K), then
ak = (x , f{Jk), 1 ~ k ~ n, satisfies the matrix Equation (7.2). D
Examples 1. Given the integral equation
[27r
Jet) - A Jo j(s)sin(s + t)ds = get), (7.7)

we shall
(a) determine those Asuch that for each g E L2 ([0, 2JrD, there exists a solution
in L2([0 , 2JrD to Equation (7.7).
(b) find the solutions .
Define K : L2([0 , 2Jr]) ~ L2([0, 2JrD by

(Kf)(t) = 1 27r
j(s)sin(s + t)ds.
Example 5 in Section 2 shows that K is bounded and linear. From the trigono-
metric identity sin(a + fJ) = sina cosfJ + coso sinfJ, it follows that
AKf = (j, f{Jj)1/I1 + (j, f{J2)1/I2,
where
1/I1(t) = ACOSt , 1/I2(t) = Asint,
f{JI(t) = sin r, f{J2(t) = cost.
Thus AK has rank 2 if A :I O. Using the same notation as in Theorem 7.1, we get
all = 1 - (1/11 , f{JI) = 1, a12 = -(1/12, f{JI) = -AJr ,
a21 = -(1/11, f{J2) = -AJr, a22 = I - (1/12 , f{J2) = 1,
det(aij) = 1 - A2Jr2.

Hence I - AK is invertible if and only if A t=

±~, in which case for each
g E L2 ([0, 2JT]) there exists a unique j ([0, 2JrD (identifying functions
E L2
which are equal almost everywhere) which satisfies (7.7) a.e. The solution is

j(t) = [(l - AK)-I g](t) = get) - I 1 2 det

- A2 Jr
(-~Jr -~Jr. :~)
ACOSt A smt 0

=g(t)+ \ 2 [(bl +AJrb2)cost+(b2+AJrbl)sint] ,

I-AJr
68 Chapter II. Bounded Linear Operators on Hilbert Spaces

where

10r
Z7f
bi = (g, <PI) = g(t) sint dt,

10r
27f
bz = (g, <pz) = g(t) cos t dt .

When A = ±~, Ker (I - AK) consists of vectors of the form (Xp/JI + (Xzo/z,
where

(1 -AJr
-AJr) ((XI) =
1 (Xz
(0)
0
or
(XI - AJr(XZ = 0
-AJr(X1 + (Xz = O.

Thus if A = ~, then (XI = az- If A = -~, then (XI = -(XZ . Hence

Ker (I - ~K) = sp{cost + sint}.

Ker (I + ~K) = sp{cost - sint}.

Therefore, the general solution to the homogeneous equation

f(t) - -1
Jr 0
l Z7f
f(s) sin(s + t) ds = 0

is f(t) = c(cost + sin z), c arbitrary in C, and the general solution to the homo-
geneous equation

f(t) + -I l
Jr 0
Z7f
f(s) sin(s + t) ds = 0

IS
f(t) = c(cos t - sin t), c arbitrary in C.
2. Let a(t) be a continuous complex valued function defined on the closed
interval [a, b]. Define a bounded linear operator A on 1{ = Lz([a, b]) by

(Af)(t) = a(t)f(t).

In Example 4 in Section 2, we saw that A is a bounded linear operator with

IIAII = maXIE[a.bjla(t)l· If aCt) i- 0 for all t E [a, b], then A is invertible,
1
(A-Ig)(t) = -g(t).
a(t)
2.8 Inversion ojOperators by the Iterative Me thod 69

and
1
IIA- 11I = max - - (7.8)
laCt)1
t E[a.b]

Conversely, suppose A is invertible. We show that a Ct ) f= 0 for all t E [a, b]. For
JEH ,
1
IIAjll ~ II A- 11l1lf1l (7.9)

Suppose a Cto ) = 0 for some to E [c, d]. Then given e > 0, there exists a 0 > 0
such that la (t )1 < e if It - rol < o. Define

J (t ) = { 1, It - tol s 0 }.
o It - ro l > 0
Then

f
to+o
II AJ II
2
= la (t )1
2IJ
(t )1
2
:s e 2 11 f11 2 ,
to-o

which contradicts (7.9) if e < IIA~I II'

If A and B are invertible operators in £ (H ), then A B is invertible and (AB) - l =
B-1A - 1.

2.8 Inversion of Operators by the Iterative Method

The next theorem, which is rather easy to prove, is nevertheless very useful. The
formula for (/ - A )-l is suggested by the geometric series

l +a +a 2 + .. . = (l -a) - I ,la l< 1.

Theorem 8.1 Suppose A E £ (H ) and IIAII < 1. Then I - A is invertible, and

for every y E H,
00

(/ - A)-l y = LAky (A o = /).

k=O
Moreover.

i.e.,

00

(/ - A )-l = LA k
k=O
70 Chapter II. Bounded Linear Operators on Hilbert Spaces

and

Proof: Giveny E 1t, the series L~o Akyconverges. Indeed.Iet s., = Lk=O Aky.
Then for n > m,
n n
II sn - smll S L IIA kYIi S IIYII L IIAli k --+ O.
k=m+1 k=m+1

as m, n --+ 00 since L~o IIAli k = I-tAli ' The completeness of1t ensures the
convergence of {sn}, i.e., L~o Aky converges . Define B : H --+ 'H by By =
L~o A k y . B is linear and

Hence IIBII S I-tAli and

00 00

(I - A)By = (I - A) LAky = L(I - A)Aky

k=O k=O
00

= LAk(I - A)y = B(I - A)y

k=O
00 00

= LAky - LAk+l y = y.
k=O k=O

Therefore, I - A is invertible and (I - A) -I = B. Finally,

as n --+ 00. D

Corollary 8.2 Let A E £(1t) be invertible. Suppose B E £(1t) and II A - B II <

IIA~III ' Then B is invertible,

00

B-Iy =L [A-I(A - B)]kA-I y

k=O
2.9 Infinite Systems ofLinear Equations 71

and
1 2I1
II A - I - B - I II< IIA - 1I A - BIl .
- 1 -IIA- I IIIIA - B II
Proof: Since B = A - (A - B ) = A[I - A- I (A - B )] and II A - I( A - B ) II ~
II A -11111 A - B II < 1, it follows from Theorem 8.1 that B is invertible and

L
00

B - 1 = [I - A-I (A - B )rIA- I = [A -I (A - B)]kA - I .

k=O
Hence
1 211A
IIA- I - B-III < IIA-III ~ IIA- I llkIiA _ Bilk =
- 6 IIA- 1I - BII .
1-IIA- IIIIIA - BII

o
Corollary 8.3 Let {An} be a sequence ofoperators in £(1t1, 1t2) whi ch converges
in norm to A E £(1t I, 1t2 ). Then A is invertible ifand only ifthere exists an integer
N such that An is invertiblef or n 2: Nand sUPn >N II A ; III < 00. In this case,
IIA;I - A-III ~ o. -
Proof: Suppose A is invertible, then iffollows from Corollary 8.2 that there exists
an N such that An is invertible for all n 2: Nand IIA; I - A -III ~ O. Thus
sUPn>NIIA;11I < 00.
Conversely, suppose that An is invertible for n 2: Nand M = sUPn>N IIA;I II <
00. Then for n 2: N , -

A = An +A - An = AnU + A;I (A - An)) . (8 .1)

Since IIA;I (A-A n) 1I ~ M IIA-Anll ~ 0, it follows from Theorem 8.1 that there
exists NI 2: N such that I + A;;-I(A - An) is invertible. Since An is invertible for
n 2: NI , we have from (8.1) that A is invertible. Hence II A;;-I - A -III ~ 0 by
Corollary 8.2.
Ifwe define the distance d CA , B ) between the operators A and B to be IIA - BII ,
then Corollary 8.2 shows that the set 0 of invertible operators in £ (1t) is an open
set in the sense that if A is in 0, then there exists an r > 0 such that dCA, B) < r
implies B E O. 0

2.9 Infinite Systems of Linear Equations

In this section we shall solve certain infinite systems of linear equations

00

L aijx j = Yi , i = 1, 2, .. . , (9.1)
j =1

where {Yj} E £2 is given and {x j } is the desired solution in £2.

72 Chapter II. Bounded Linear Operators on Hilbert Spaces

This system is important because any equation Ax = y, where A is in £(H)

and H is a separable Hilbert space, can be written in the form (9.1). Indeed, let
{Cj? j} be an orthonormal basis for H. Then from matrix equation (4.4) in Section
4, we see that (9.1) holds when
Xj = (x, Cj?j), Yj = (y, Cj?j) and aij = (ACj?j , Cj?i) .
A natural approach to the problem of finding a solution to an infinite system
of linear equations is to approximate the solution by solutions to finite sections
of the system. This method does not always work (see Example 2 in Section 17).
However, we do have the following result.

Theorem 9.1 Given the matrix (aij)f.}=I' where L:f.}=1 laije < 1, the system of
equations

L: aijXj = Yi,
00

Xi - i = 1,2,. . . (9.2)
j=1
has a unique solution ~ = (~I, ~2 , . . .) E £2 for every Y = (YI, Y2, .. .) E £2. The
truncated system ofequations
n
Xi - L:aijXj = n. i s i .:s n, (9.3)
j=1
.
has a umque so Iutton
' (n)
xI , . .. , Xn(n) an d Zn = (xI(n) , .. . , Xn(n) , 0 , 0 , ... ) con-
verges in £2 to ~ as n -+ 00.
Proof: Define A E £(£2) by A(ai, a2, ...) = (fh, fh , . . .), where

By Section 2, Example 3, IIAII 2 .:s L:f.}=llaijI2 < 1. Hence we have from

Theorem 8.1 that I - A is invertible, from which it follows that (9.2) has a unique
solution in £2. To approximate this solution, let An E £(£2), n = 1,2, ... , be

1
defined by An (ai, a2, . ..) = (fh, fh, . ..), where

n
a ll al :I
: :I 0
ani . .. ann:
[ -----0----- I

00 00 00 00

IIA - An 11
2
.:s L L laijl2 + L L laij!2 -+ 0 as n -+ 00.
i=n+1 j=1 j=n+1 i=1
2. J0 Integral Equations ofthe Second Kind 73

Hence I - An is invertible and

(9.4)

by Corollary 8.3. Taking (xf n). xi n)• . . .) = (J - An)- I (YI. Y2• . . .), it follows
from the definition of A n that (xf n)• . . . • x~n» is the unique solution to (9.3) and
X]n ) = Yj, for j > n . Since L~ l IYjl2 < 00, (9.4) implies that

l'
n 2.~
«n ) (n) 0 0 )_
x I • . . .• x n • , , . .. - n2.~
u «n ) (n)
Xl • . . . , x n ' Yn+l . Yn+2 • . . .
)

= (J - A)-I(YI . Y2 • .. .) . (9.5)

In fact, (9.4) shows that as (YI. Y2• . . .) ranges over the I-ball of £2, the corres-
ponding sequences in (9.5) converge uniformly. 0

2.10 Integral Equations of the Second Kind

Given k E L 2([a . b] x [a. b]) , we know from Section 2, Example 5 that

(Kf)(t) = l b
k (t , s) J(s) ds

is a bounded linear operator from L 2([a , b]) into L 2([a , b]) with IIK II S IIkll .
Suppose IIkll < 1. By Theorem 8.1, I - K is invertible. Thus for each g E
L 2([a . b]) , there exists a unique J E L 2([a . b]) such that

J(t ) -l b
ktt , s )J (s ) ds = get ). (10.1)

Here" = " means equal almost everywhere .

Equation (10.1) is called an integral efJ!:!ation of!!!e second kind.
We shall show that (J - K) -I = J + K , where K is an integral operator.
By Theorem 8.1,

L
00

(J - K )-l g = K ng. (10.2)

n=O
74 Chapter II. Bounded Linear Operators on Hilbert Spaces

From the definition of K and Fubini's theorem (Appendix 2),

(K 2g)(t) = l b
k(t, x) (Kg) (x)dx

= l {l
b
k(t ,x)
b
k(X,S)g(S)dS}dx

= l {l
b
g(s)
b
k(t ,X)k(X ,S)dX}dS

= l b
k2(t , s)g (s) ds,

where

k2(t, s) = l b
k(t, x)k(x, s)dx.

The function k2 is in L2([a , b] x [a , b]); for, by Schwarz's inequality,

Ik2(t , s)1
2
s {l b
Ik(t, X) 12dX} {l b
Ik(x, S)12dX} .

Therefore,

and

ll b b
Ik2(t, s)1
2ds
dt

~ {l l {l l
b b b b
2dx 2dx
Ik(t, x)1 dt} Ik(x , s)1 dS} '

Proceeding in this manner, an induction argument establishes that

(Kng)(t) = l b
kn(t ,s)g(s)ds, n= 1,2, . .. , (10.3)

and IIkn II .s Ilkll n , where

k!(t ,s) =k(t,s) andkn(t ,s) = l b

k(t,x)kn_!(x, s) dx,n > 1, (lOA)

is in L2([a, b] x [a, b]) . Since

00 00

L IIk nil s L IIkil n

< 00 ,
n=! n=!
2.10 Integral Equations ofthe Second Kind 75

L~I kn = k for so~e k E L2([a, b] x [a, b]) . Let K be the integral operator
with kernel function k. It follows from (10.3) and Schwarz's inequality that

as p ~ 00. Hence
00

(1- K)-I g = L Kng = (1 + K)g,

n=O
which shows that the solution (a. e.) to (10 .1) is given by
b
Jet) = get) +l k(t, s)g(s) ds. (10.5)

l
If, in addition,
b
c = sup Ik(t, s)12ds < 00,
t a
then
00

«(1- K)-Ig)(t) = L(Kng)(t) a.e.

n=O
The series converges absolutely and uniformly on [a, b]. Indeed, for any h E
L2([a , b]), we get from Schwarz's inequality,

b )1/2
I(Kh)(t) I S IIhil ( llk(t , s)1 2 ds S -JCllhil .

Replacing h by K n - I g gives

II(Kng)(t)1 S -JCIIKn- ' gil S -JCIIKll n- ' IIgll. n = 1,2, .. . .

Hence
00 00

L I(Kng)(t)1 S -JCIIgil L IIKll n- ' < 00, t E [a, b] .

n=1 n=l

Example: To solve the integral equation

Jet) - A 1 1
e(t-s) j(s) ds = get) E L2([0, 1]), (10.6)

let k1 (t , s) = Aet - s. Then from (lOA) in the discussion above,

k2(t,S) = A211 et-xex-s dx = A2et- s.

76 Chapter II. Bounded Linear Operators on Hilbert Spaces

In general, kn(t, s) = Anet - s, and from (10.5),

I(t) = get) + I>n

00

n=1
1 I
et-Sg(s) ds

1
0
1
= get) + -A- et- s g(s) ds (10.7)
1 -A 0

is the solution to the integral equation for IAI < 1.

Even though the series converges only for IAI < 1, a straighforward computation
verifies that for all A f= 1, the expression after the second equality in (10.7) is still
a solution to (10.6).
Another way to find the solution to an integral equation of the second kind is
to make use of the results in Section 9. Namely, suppose IIkll < 1. Let {<pj} be
an orthonormal basis for Lz([a, b]). Then for aij = iKo], <p;), we have from
Example 3, Section 5, that L0=1 laijlz = IIkll z < 1. Now by Theorem 9.1, the
finite system of equations
n
Xi - L:aijXj = (g, <Pi), l:s i :s n ,
j=1
has a unique solution

where B n is the inverse of the matrix (Ojk - ajk)'J ,k=l.

The sequence of vectors (x}n) , . . . , x~n), 0, 0, ...) converges in £z to some
vector, say, (XI, X2, . . .) . The solution to the integral equation is I = L1==1 Xj<Pj
(convergence in Lz([a, b]).

2.11 Adjoint Operators

In this section we discuss the infinite dimensional counterpart to the adjoint of a

matrix , namely the adjoint of a bounded linear operator.
Frequently, knowledge about an operator can be gained from information about
its adjoint, which may be simpler to deal with .
Suppose A E £(11.1, Hz) . For each y E Hz , the functional Iy(x) = (Ax, y) is
a bounded linear functional. Hence the Riesz representation theorem guarantees
the existence of a unique y* E 11.I such that for all X E 11.I,

(Ax, y) = I y(x) = (x, y*) .

2.11 Adjoint Operators 77

This gives rise to an operator A * : H2 ~ H I defined by A *Y = y* . Thus for all

x E HI,
(Ax, y) = (x, y*) = (x, A*y) .
The operator A * is called the adjoint of A .

A* is in £(H2, Hd and IIAII = IIA*II.

Indeed, given u and v E H2 and a, fJ E C,

(Ax, au + fJV) = a(Ax, u} + ~(Ax , v)

= (x, aA*u + fJA*v)

for all x E HI . Thus, by definition, A*(au + fJv) = aA*u + fJA*v. Also, we

have from (Ll) in Section 1, that

IIAII = sup I(Ax, y}1 = sup I(x, A*Y}I = IIA*II·

IIxll=lIylI=1 IIxll=lIylI=1

Examples 11.1 1. We have 0* = 0 and 1* = 1.

2. Let S; be the forward shift onz-. Giveny = (YI, Y2, . . .)andx = (XI,X2 , . . .)
in £2,

is,x, y) = ((0, XI, X2 , ·· .), (YI, Y2,· · .)}

= XI Y2 + X2Y3 + ... = (x , y*) .

where y* = (Y2, Y3, . . .) . Thus S: is the backward shift operator.

3. Let Sf be the backward shift operator on £2. For x = (XI , X2, ...) and
Y = (YI, Y2, . . .) in £2,

(SIX, y) = X2YI + X3Y2 + ... = (x , y*),

where y* = (0, YI, Y2, . ..) . Thus S; is the forward shift operator.
In a similar way one shows that the adjoint of the bilateral forward shift
Von £2(Z) is equal to the bilateral backward shift on £2(Z) . Thus in this
case V* = V-I.
4. Let K E £(H I, H2) be the operator of finite rank defined by Kx = L:J=I
(x, u j }V j, U j E HI , Vj E H2 . Then for all x E HI and Y E H2,

(Kx , Y) ~ t,(X, Uj)(Vj, y} ~ (x , t,IY, Vj)u j ).

Hence
n
K*y = L(Y' Vj}Uj .
j=1
78 Chapter II Bounded Linear Operators on Hilbert Spaces

5. Let a(t) be a bounded complex valued Lebesgue measurable function on

[a, b], and let A: L2([a , b]) ---+ L2([a, b]) be the bounded linear operator
defined by
(Af)(t) = a(t)!(t).
For all!, gin L2([a, b]),

(Aj, g) = I b
a(t)!(t)g(t)dt = (j, iig) .

Thus (A*g)(t) = ii(t)g(t).

6. Let 1{ be a Hilbert space with an orthonormal basis !PI, !P2, .. .. For A E
,C(1{), the matrix (aij) corresponding to A and !PI, !P2, . . . is (aij) = «(A!pl :
!Pi}) (cf. Section 4). Since

(A*!pj, !Pi) = (!pj' A!pi) = iiji'

The matrix corresponding to A * and !PI, !P2 , . .. is the conj ugate transpose
(iip) of the matrix (aij) '

The next example is important enough to be designated as a theorem.

Theorem 11.2 Let K : L2([a, b]) ---+ L2([a, b]) be the bounded linear operator

I
defined by
b
(Kf)(t) = k(t, s)!(s)ds,

where k is in L2([a, b] x [a, b]). For all g E L2([a , b]),

(K*g)(t) = I b
k(s, t)g(s)ds.

Proof: By Fubini 's theorem,

(Kj, g) = I (l
b b
k(t , S)!(S)dS) g(t)dt

= I (l
b
!(s)
b
k(t, s)g(t)dt ) ds = (j, g*),
where

g*(s) = I b
k(t,s)g(t) dt.

o
2.11 Adjoint Operators 79

Theorem 11.3 If A and B are in .c(11I , 11z), then

(i) (A + B)* = A* + B*;

(ii) (aA)*=aA*,aEC;
(iii) A ** = A;
(iv) If D is in .c(11z , 113), where 113 is a Hilbert space, then (DA)* = A* D*.

Proof: (i) Given x E 111 and y E 11z,

(A + B)x, y) = (Ax, y) + (Bx, y) = (x, A*y + B*y).

Therefore, by definition, (A + B)*y = A*y + B*y .

(ii) (aAx,y) =a(x ,A*y) = (x,aA*y).
Hence (aA)*y = aA*y.
(iii) (A*y, x) = (x, A*y) = (y, Ax) .
Thus, by the definition of the adjoint of A*, A**x = Ax .
(iv) For v E 113,

(DAx , v) = (Ax, D*v) = (x, A*D*v) .

Hence (DA)*v = A*D*v.

o
The next theorem gives some useful relationships between the kernels and ranges
of an operator and its adjoint.

Theorem 11.4 For A E £(11 I, 11z), the following dual properties hold.
(i) KerA = ~A*-1 .
(ii) KerA* = ~A-1 .
(iii) ~A = KerA*-1.
(iv) ~A* = KerA-1.

Proof: (i) Given x E Ker A and y E 11z,

0= (Ax, y) = (x , A*y).
Thus Ker A C ~A *-1. On the other hand, if v E ~A *-1,then for all y E 11z,

0= (v, A*y) = (Av, y).

In particular, 0 = (Av, Av) . Hence v E KerA and therefore ImA*-1 C

KerA.
80 Chapter II. Bounded Linear Operators on Hilbert Spaces

(ii) Applying (i) to A * and A ** = A, we get

KerA* = ImA**l- = ImAl- .

(iii) Corollary 1.8.3, the continuity of the inner product and (ii) imply

(iv) Apply (iii) to A * and A ** = A.

o

2.12 Self Adjoint Operators

The requirement that an operator be self adjoint is very restrictive. Nevertheless,

a wide variety of problems in mathematics and physics give rise to self adjoint
operators. Some of these operators are studied in subsequent chapters.

Definition: An operator A E £(1i) is called selfadjoint if A * = A.

As a trivial consequence of Theorem 11.4(iv) we have the following result.

Theorem 12.1 If A E £(1i) is self adjoint, then Ker Al- = :SA. Thus H =
Ker A EB :SA.

Examples 12.2 1. Let K E £(1i) be the operator of finite rank given by Kx =

I:j=l (x , Uj}Vj . By 11.1, Example 4, K is selfadjoint if and only iffor all
x E 'H,
n n
L(x, U j}Vj = L(x , Vj}Uj'
j=l j=l

2. Let (Af)(t) = a(t)f(t) be the operator defined in 11.1, Example 5. Then

A is self adjoint if and only if aCt) = att).
3. Let K be the integral operator with kernel function k defined in Theorem 11.2.
Ifk(t, s) = k(s , t) a.e., then A is selfadjoint. The converse also holds.
4. If A E £(1i), then A *A is self adjoint since

(A* A)* = A* A** = A* A .

Theorem 12.3 An operator A E £(1i) is selfadjoint if and only if (Ax , x) is real

for all x E 1i.
2.13 Orthogonal Projections 81

Proof: If A = A *, then for x E H,

(Ax, x) = (x , Ax) = (Ax , x).

Thus (Ax, x) is real

Suppose (Au, u) is real for all u E H . Then for all x, y E H and A E C,

(A(x + Ay), (x + Ay)) = (x + Ay, A(x + Ay))

which, together with the assumption, implies that

I(Ax, y) + A(Ay, x) = I(x , Ay) + A(y, Ax) .

Hence
~A(Ay, x) = ~A(Y, Ax) .
Taking A = 1 and A = i, it follows that (Ay, x) = (y, Ax) . Thus A = A*. 0

2.13 Orthogonal Projections

There is a special class of self adjoint operators, called orthogonal projections,

which are building blocks for the theory of self adjoint operators, in the sense that
every bounded linear self adjoint operator is the limit, in norm, of a sequence of
linear combinations of orthogonal projections. This result is proved in Chapter IV,
but only for a certain class of self adjoint operators.

Definition: Given a closed subspace M of H, an operator P defined on H is

called the orthogonal projection onto M if

P(m + n) = m , for all m E Mand n E M.1. .

It is easy to see that P is linear on H , ~ P = M, Ker P = M.1. and Pm = m for

all m EM . Clearly, I - P is an orthogonal projection onto M.1. with kernel M .
The projection P is called orthogonal since Ker P is orthogonal to ~ P .
If M :F 0, then IIPII = 1; for given x = u + v, where u EM and v E M.1., we
have by the Pythagorean theorem,

IIPxll
2
= lIull 2 :s lIull 2 + IIvII 2 = lIu + vII 2 = IIx1l 2 .
Thus IIPII :s 1. But for m :F 0 in M, IIPlilimll ~ IIPmll = IImli. Hence IIPII ~ 1.
From Theorems 1.4.3 and 1.8.1 we know that if M is a closed subspace of H,
then for each y E H, there exists a unique W E M such that (y - w)l.M or,
equivalently, II Y- wll = d(y, M). Ifwe define Py = w, then P is the orthogonal
projection of y onto M since y = w + y - w, W E M , y - W E M.1. .

Theorem 13.1 An operator P E £(H) is an orthogonal projection if and only if

p 2 = P and P is selfadjoint.
82 Chapter II. Bounded Linear Operators on Hilbert Spaces

Proof: Suppose P is the orthogonal projection onto a closed subspace M of H .

Given x = m + n, m E M and n E Ml..,

P(Px) = Pm = m = Px .

Thus p 2 = P . Now for y = ml + nl, ml EM, nl E Ml.., (Px , y) = (m, ml) =

(m + n , ml) = (x, Py). Hence P = P* .
Assume r?
= P and P* = P. Let M = ::5P. Since M = Ker(I - P) . Mis
closed and by Theorem 1104,

Ml.. = ::5Pl.. = KerP .

Therefore, given U E M and v E M l..,

P(U + v) = Pu + Pv = Pu = U

since U E M = ::5 P . Hence P is the orthogonal projection onto M. D

2.14 Two Fundamental Theorems

The following two theorems, which play a major role in functional analysis,
will be used often throughout the remaining chapters of the text. The proof of
Theorem 14.1 which is given in the more general setting ofBanach spaces, appears
in Theorem XIIAA.

Theorem 14.1 Suppose A E L(HI, H2) has the properties that KerA = {O} and
ImA = H2. Then its inverse A-I is bounded.

Corollary 14.2 An operator A E L(HI, H2) is one-one and has a closed range
if and only if there exists an m
> 0 such that

IIAxll ~ mllxll for all x E HI. (14.1 )

Proof: Assume inequality (14.1). Clearly, A is one-one. Suppose AXn --+ y .

Then the sequence {Axnl is Cauchy and therefore the sequence {xnl is Cauchy by
(14.1) . Hencex = limn.-+ooxn exists since HI is complete. Thus Ax, --+ Ax = y,
which shows that ::5A is closed.
Conversely, suppose that A is one-one and::5A is closed. Let Al be the operator
A considered as a map from H I onto the Hilbert space ::5A. Then All is bounded
by Theorem 14.1. Hence

IIxll = IIAllAxll s IIAlllIlIAxll

and inequality (14.1) follows. D

The proofofthe next theorem is elementary. It differs from the usual proofs which
depend on the closed graph theorem or the Baire category theorem.
2.14 Two Fundamental Theorems 83

Theorem 14.3 Suppose {An} is a sequence in £(HI, Hz) with the property that
sUPn II Anx II < 00 for each x E H I . Then sUPn II An II < 00.

Proof: Assume sUPn IIAn II = 00 . Let {ad be a sequence of positive numbers

which we shall choose later. There exists operators Ani , Anz, . . . and vectors
XI , Xz, .. . in HI such that

(14.2)

and
(14.3)

Let x = L~I x i - Note that the series converges since HI is complete and
L~I II x j ll = L~14- j < 00 . Since

it follows from (14.2) and (14.3) that for k > 1

where

Now choose a k > 3 · 4k ( C k_ 1 + k) , k = 1,2, . . . . Then sUP n IIA nx ll = 00, which

is a contradiction. 0

CoroUary 14.4 Let {An} and {B n} be sequences in £(HI , Hz) and £(Hz, H3) ,
respectively, where H i is a Hilbert space, i = 1, 2, 3. Suppose Ax = limn-.+ ooAnx
and By = lim Bny existforeach x E HI and y E Hz. Then A and B are bounded
n-.+ oo
linear operators and

BnA nx ~ BAx for each x E HI .

Proof: Since the sequence {Bny} converges for each y E H z, sup, II Bnyll < 00 .
Hence M = sUPn IIBn II < 00 by Theorem 14.3. Thus IIByll = limn -.+ oollBny ll S
84 Chapter II Bounded Linear Operators on Hilbert Spaces

MlIylI . Clearly B is linear. Hence B E L(H2, H3) . Similarly A E £(Hl, H2).

Therefore

IIBnAnx - BAxll ~ IIBnAnx - BnAxll + II BnAx - BAxll

~ MIiAnx - Axil + IIBnAx - BAxll --+ O.

o
2.15 Projections and One-Sided Invertibility of Operators

Definition: An operator P acting on a Hilbert space H is called a projection if

P is a bounded linear operator on Hand p 2 = P .
If P is a projection on H, then I - P is also a projection on H since (I _ p)2 =
1- 2P + p 2 = I - P .

Examples: Each operator P defined on £2 in a) - c) is a projection.

(a) P(al,a2, )=(0 ,0, .. . 0,a n,an+I, ...)

(b) P(al,a2 , )=(al,0,a3,0,as, . .. )
(c) P(al , a2 , ) = (aI, aI, . .. , aI, an+l, a n+2, . ..) .
--.
n
.-
(d) Let H = L2([0, 1]) and let E be a measurable subset of [0,1]. Define P
onHby
(Pf)(t) = !(t)CE(t) ,
where Cdt) is the characteristic function of E. Clearly, P is a projection
on H.

Theorem 15.1 Let P be a projection on a Hilbert space H . Then

(i) 5:5 P = Ker (I - P) .

(ii) 5:5 P is closed.
(iii) Every vector v E H can be uniquely written in the form v = x + y, where
Px = 0 and P y = y.

Proof: (i) Since (I - P)Px = Px - Px = 0, we have 5:5P ~ Ker (I - P). Ify E

Ker (I - P), then y - Py = O. Thus y = Py E 5:5P . Equality (i) follows
(ii) Statement ii) is an immediate consequence of (i)
(iii) Take x = (I - P)v and y = Pv . Then x and y have the properties stated
above . To prove the uniqueness ofthe representation, suppose v = Xl + Yl,
where PXl = Oand Pv, = Yl. Then r -cc, = YI-yandO = P(X-Xl) =
P(YI - y) = Yl - y .

Hence Y = Yl and x = Xl. o

2.15 Projections and One-Sided Invertibility ofOperators 85

Recall that we have already met projections in Section 13.

Definition: A Hilbert space 11. is said to be the direct sum of subspaces M and
N, written 11. = M EB N, if every vector v E 11. has a unique representation of the
form v = x + y, where x E M and yEN.
The subspace M is said to be complemented in 11. ifthere exists a closed subspace
N of 11. such that 11. = M EB N . Note that the representation of v = x + y, x E
M, yEN is unique if and only if M n N = {OJ .
We have seen in Theorem 15.1 that if P is a projection on 11., then

H = Ker P EB:SP. (15.1)

Examples: (i) Let P be the projection in a). Then

:sP = {(ak) E £z!ak °

= 0, .:s k < n},
Ker P = {(,Bd,,Bk = 0, k >: n)

(ii) Let P be the projection in b). Then

:sP = {(ak) E £zlazk = 0, k = 1,2, },

Ker P = {(,Bk) E £zl,Bzk+1 = 0, k = 0,1, }

(iii) Let P be the projection in c). Then

:sP = {(ak)lak = aI, 1 .:s k.:s n),

Ker P = {(,Bdl,Bk = 0, k = 1 and k > n}

(iv) Let P be the projection in (d). Then

:sP = {f E Lz([O, 1])lf(t) = 0, t ¢ E},

Ker P = {g E Lz([O, 1])lg(t) = 0, tEE}

The projections P in examples (i), (ii) and (iv) are orthogonal projections since
:sP is orthogonal to Ker P . This is not the case for the projection in example
(iii) since (ak), (,Bk») = Lk=1 ak~k need not be zero for (ad E :sP and (,Bd E
KerP.
Let HI and Hz be Hilbert spaces. An operator A E £(11.1 , Hz) is said to be
left invertible if there exists an operator L E E (Hz , 11.» such that LA = 1. The
operator L is called a left inverse of A.
Examples: 1. Let A be the forward shift operator on £z, i.e., A(al , az , . . .) =
(0, aI , a2, ...) . Then A is left invertible and every left inverse L has the form
86 Chapter II. Bounded Linear Operators on Hilbert Spaces

whe re {,Bd is an arbitrary vector in f,2. Indeed, Let L ei = {.Bd k::I' Since Le; =
LAek-1 = ek- I, k > 1, equali ty (15.2) follows .
2. Let AI be the bounded linear operator defined on f,2 by

AI (ai, a 2, . . .) = (0, a i , a 2 - ~al ' a 3 - ~a2 ' .. .) .

Then A I is left invertible and every left inverse L I of A I is of the form

1
+L
k
LI «ak» = (rk), where rk = al.Bk +ak+1 2k+I - ja j , (15 .3)
j= 1

and (.Bd is an arbitrary vector in f,2. This can be shown as follows: First we note
that AI = S - !S2 = S(I - is), where S is a forward shift operator on f,2. If LI
is a left inverse of A , then

(15.4)

Since S is a forward shift , we have from (15.4) and the above example that L I x
(I - i s) = L , where L is given in (15.2). Now II !SII = ! . Hence I - ! Sis
invertible, LI = L(I - ! S) -I , and a simple computation shows that

( 1-
1
"2S
)-1 « ad) = (11k) , where

k-I 1
111 =a l ,11k=ak + L 2
k- j a j , k > 1. (15.5)
j= 1

Thus from (15.2) and (15.5) we get

3. Let 1t = L 2([0 , 1]). For 0 < a < 1, define the operators Ra and Rt / a on H
by
(Raf)(t) = fea t), 0 ~ t ~ 1, (15.6)
and
(R I/af)(t) = {f ( ~t) , 0 s t .s a , (15.7)
0, a < t ~ 1.
2. J5 Projections and One-Sided Invertibility ofOperators 87

Clearly R a is linear and

(15.8)

Thus II s; II ~ 1j.ja. On the other hand, define

rp(t) = {.la'
0,
o ~ t ~ a,
a<t~1.

Then IIrpll = 1, and from (15.8) we have

Hence II s; II 2: 1j.ja. We have shown that

1
IIR a ll = .ja' (15.9)

Similarly

Hence
(15.10)

Now
(RaRlrp)(t) = (Rlrp)(at) = rp(t), 0 ~ t ~ 1, (15.11)
a a

which shows that R 1 is a right inverse of R a .

Iffollows from «(5.6) and (15.11) that

Ker R; = {rp E Hlrp = 0 on [0, a]} .

and
:sRa = H.

Thus R a has a right inverse but is not invertible . Also, from (15.7) and (15.11) we
have this
Ker Rl = {O}
a

and
:sRl = {g E Hig = 0 on [a, I]}. (15.12)
a

Indeed, given such a g, define rp(t) = g(at) , 0 ~ t ~ 1. Then Rl /arp = g .

88 Chapter II. Bounded Linear Operators on Hilbert Spaces

Finally, the adjoint R~ = ~ Rl ja. Indeed

(Raf, g) = 11 feat) get) dt = ~ 1 a

f(s) g (~s) ds
= (f' ~Rljag),
where the last equality follows from (15.7).

Theorem 15.2 Let operator A E £('HI , 'H2) have a left inverse L. Then ~A is
closed and complemented in 'H by Ker I. The operator AL is a projection onto
~A and Ker AL = Ker I.

Proof: Suppose A has a left inverse I. Clearly, Ker A S; Ker LA = Ker I = {O}.
The bounded operator AL is a projection since (AL)2 = A(LA)L = AI. From
~AL S; ~A and ~A = ~(AL)A S; ~AL, we have ~AL = ~A . Also, KerAL S;
Ker LAL = Ker Land Ker L S; Ker AI. Thus Ker AL = Ker L and

'H = ~A EB Ker I. (15.13)

Definition: Let M and N be subspaces (not necessarily closed) and let 'H =
M EB N. The codimension of M, written codim M , is defined to be dim N (finite
or infinite).
The codim M is independent of the subspace N . Indeed, suppose 'H = M EB Z .
Define <p : N ~ Z as follows :
For each v EN, there exists a unique u E M and a unique Z E Z such that
v = u + z. Let <p(v) = z. The linear operator <p is a I - I map from N onto Z .
Hence dim Z = dim N .

Theorem 15.3 Lettheoperator A E £(HI, H2) have a left inverse L E £('H2, HI) .
If B is an operator in £('H I, 'H2) and II A - B II < II L II-I, then B has a left inverse
LI given by

LI = L(I - (A - B)L)-I = L (fk=O

[(A - B)L]k) . (15.14)

Moreover,
codim ~B = codim ~A. (15.15)

Proof: Since IIA - BII < liLli-I, the operator I - (A - B)L is invertible and
(I - (A - B)L)-I = L~o [(A - B)L]k by Theorem 8.1. Hence

LIB = L(I - (A - B)L)-I B = L(I - (A - B)L)-I[I - (A - B)L]A

= LA = I.
2.16 Compact Operators 89

From the identity B = [l - (A - B )L]A and the invertibility of 1 - (A - B )L ,

equality ( 15.15) is valid. Indeed, for C = (l - (A - B )L ) it follows from the
invert ibility of C and equality (15 .13) that

H 2 = CH2 = :sCA EB CKer L = :sB EB C Ker L .

Hence
codim 'JB = dim C Ker L = dim Ker L = codim :sA
o
Ifan operator A E L (H\ , H 2) has a left inverse L but A is not invertible (e.g., A
is a forward shift operator on £2), then A has infinitely many left inverses ; namely,
for each A E C , L + A(1 - AL) is a left inverse of A . Indeed,

[L + A(l - AL )]A = LA + AA - AAL A = LA = 1.

Definition: An operator A E .c(H i . H 2) is said to be right invertible if there
exists an operator R E L (H2, H d such that A R = l. Th e operator R is called a
right inverse of A .

Example: Let A be the backward shift operator on £2, i.e., A(I1\, 112 , .. .)
(112, 113 , ... ) Then A is right invertible and every right inverse R of A has the form

(15.16)

where (Yk) is in £2 and {ed is the standard basis in £2. To see this, let R ek =
(fh , fh, .. .) now
ek = ARek = (fh , fh , . . .).

Thus fh is arbitrary, fh+\ = 1, and f3 j = 0 when j =f. I and j =f. k + 1. Henc e

R e; = Yke\ + ek+l for k = 1,2, . . . , and (15.16) follows. To see that (Yk) is in
£2 let f be the bounded linear functional on £2 defined by f « 11k)) = 11 \ . Since
f R is a bounded linear functional on £2, there exists, by the Riesz repre sentation
theorem, a unique vector (zr) in £2 such that for all (Ck) E £2 we have

00 00

L q Yk = f R « cd ) = ( q ) , (Zk) ) = L CkZk .
k= \ k=l

Hence (Yk) = (zd E h

If the operator A has a right inverse R, then the equation A x = y has a solution
x = Ry.
If R is a right inverse of A, then R* is a left inverse of A *.
90 Chapter II Bounded Linear Operators on Hilbert Spaces

Theorem 15.4 Let A E 'c(HI , H2) have a right inverse R. Then ~A = H2 and
the operator I - RA is a projectionjrom HI onto Ker A with kernel ~ R . Thus ~R
is closed and
(15.17)

Proof: Clearly H2 = ~ AR S; ~A S; H2 . The bounded operator RA is a pro-

jection since (RA)2 = R(AR)A = RA . From ImRA S; 1m R and 1m R =
1m RAR S; 1m RA, we have 1m RA = 1m R . Also, Ker A S; Ker RA and
Ker RA S; Ker ARA = Ker A . Thus 1m (/ - RA) = Ker RA = Ker A and
Ker (/ - RA) = ~ RA = ~ R. 0

Theorem 15.5 Let the operator A E 'c(HI, H2) have a right inverse R E
'c(H2 , Hd . If B is an operator in 'c(Hl, H2) and

IIA - BII < IIRII- 1 ,

then B has a right inverse RI E 'c(H2, HI) given by

00

R, = (/ - R(A - B))-I R = ~)R(A - B)]k R .

k=O

Moreover,
dimKer A = dimKer B .

Proof: Since IIR(A - B)II < 1, the operator I - R(A - B) is invertible and
(/ - R(A - B))-l = L~o[R(A - B)]k by Theorem 8.1. Hence

BRI = B(/ - R(A - B))-l R = A[I - R(A - B)] [(/ - R(A - B))-I R]
=AR = I.

Let C = (/ - R(A - B)). From the identity B = A[I - R(A - B)] = AC and
the invertibilty of C, we have

dim Ker B = dim Ker AC = dim C-1Ker A = dim Ker A.

o
If an operator A E 'c(HI, H2) has a right inverse R but A is not invertible (e.g.,
A is a backward shift operator on £2), then A has infinitely many right inverses,
namely for each A E C, R + A(/ - RA) is a right inverse of A. Indeed,

A[R + A(/ - RA)] = AR + AA - AARA = AR = I .

2.16 Compact Operators 91

2.16 Compact Operators

A very important class of bounded linear operators which arise in the study
of integral equations is the class of compact operators . This section presents
some properties and examples of compact operators which are used in subsequent
chapters .

Definition: An operator K E £(HI, H2) is compact if for each sequence {xn } in

HI , II x n II = 1, the sequence {K x n } has a subsequence which converges in H2.

(a) If K E £(HI, H2) is of finite rank, then K is compact. For suppose

{x n } C HI , IIxn II = 1. Then {Kx n } is a bounded sequence in the finite
dimensional space ~K . Since ~ K is linearly isometric to C k for some k, it
follows that {Kx n } has a convergent subsequence.
(b) If dim HI < 00, then any linear map K from HI into H2 is bounded by
Example 1 in Section 2. Since ~ K = KHI is finite dimensional, K is
compact by (a).
(c) The identity operator on an infinite dimensional Hilbert space is not com-
pact; for if <PI, <P2 , • • . is an infinite orthonormal set in H, then

which implies that {I <Pn} has no convergent subsequence even though

lI<Pnll=1.
If K E £(HI, H2) is compact and {Zn} is a bounded sequence in HI,
then {Kzn } has a convergent subsequence .

Theorem 16.1 Suppose K and L are compact operators in J:(HI, H2) . Then

(i) K + L is compact.
(ii) If AE £(H3 , HI) and B E £(H2 , H3), where H3 is a Hilbert space, then
KA and BK are compact.

Proof: (i) Given {x n } C HI , IIxn II = 1, the sequence {Kx n } has a convergent

subsequence {Kxn ,, }. Since L is compact, {Lx n , } has a convergent subse-
quence {Lx n" }. Thus {(K + L)xn ,, } converges.
(ii) Given {Zn} C H3 , IIznll = 1, the sequence {Azn } is bounded. Therefore it
follows from the compactness of K that {KAzn } has a convergent subse-
quence . Hence KA is compact.

If {x n } CHI, Ilxnll = 1, there is a subsequence {Kx n ' } of {Kx n } which con-

verges. Therefore , by the continuity ofB , {BKxn , } converges. Thus BK is compact.
o
92 Chapter II. Bounded Linear Operators on Hilbert Spaces

Theorem 16.2 An operator A E £(HI, Hz) is compact ifand only ifits adjoint
A * is compact.

Proof: Suppose A is compact. Let {xn} C Hz , IIx n II = 1. Since AA * is compact

by Theorem 16.1, there exists a subsequence {xnl} of {xn} such that {AA*xn,}
converges. Thus

IIA*xn, - A*xm1ll z = (AA*(xn , - Xml), Xn' - xml}

::; IIAA*(x n, -xm')lIl1xn, -xm'll
::; 2I1AA*xn, - AA*xm'll -+ 0

as n' and m' -+ 00 .

The convergence of {A *xnl} is ensured by the completeness of H I . Therefore
A * is compact.
If A * is compact, then by the result we just proved, A ** = A is compact. 0

In order to exhibit some important examples of compact operators, we use the

following result.

Theorem 16.3 Suppose {K n} is a sequence ofcomp act operators in £(H I , Hz)

and II K n - K II -+ 0, where K is in £(H I , Hz) . Then K is compact.

Proof: The proof employs a diagonalization procedure as follows.

Let {xn} be a sequence in HI, II x n II = 1. Since K 1 is compact, there exists
a subsequence {Xln} of {xn} such that {KIXln} converges. Since Kz is compact,
there exists a subsequence {xzn} of {Xln} such that {Kzxzn} converges. Continuing
in this manner, we obtain for each integer j ::: 2, a subsequence {Xjn}~1 of
{X(j-I)n}~1 such that {Kjxjn}~1 converges. We now show that the "diagonal"
sequence {Kxnn} converges, which proves that K is compact.
Given e > 0, there exists, by hypothesis, an integer p such that

11K - Kpll < £/2. (16.1)

Now {K pXnn} converges since n ::: p implies that {K pXnn} is a subsequence of

the convergent sequence {Kpx pn}. By (16.1) ,

IIKx nn - Kx mm II ::; + II K pXnn - K pXmm II

IIKx nn - Kpx nn II
+ II p - KXmm II
K xmm
::; 211 K p - KII + II K pxnn - Kpx mm II
< £ + II K pxnn - Kpxmmll -+ £

as n, m -+ 00 . Thus it follows that {Kx nn} is a Cauchy sequence which must

converge since Hz is complete. 0
2.16 Compact Operators 93

Examples: 1. Let {A.k} be a sequence in C which converges to zero . Define

K E £(£z) by
K(al , az , · . ·) = (Alai , Azaz, . . .) .
For each positive integer n, let K n E £(£z) be the operator defined by

Kn(al, az , .. .) = (Alai, . . . , Anan, 0, 0, ...).

Kn is of finite rank and therefore is compact. Since

IIKn - KII ::::: sup IAkl ---+ 0, as n ---+ 00,
k?:.n

K is also compact by Theorem 16.3.

2. Given the infinite matrix (aij) 7.'j=l' where L 0=1 laijlz < 00, let A E E (£z)
be the operator corresponding to this matrix as defined in Section 2, Example 3.
To show that A is compact, let An E £(£z) , n = 1,2, . . . , be the operator corres-
pon dimg to th e matrix
. «(n)
a r ioo h l ' . d (n)
,j=I' were aij = aij , ::::: I, ) ::::: n, an aij =
(n) 0
otherwise. Since An is offinite rank, it is compact. Moreover,

IIA - Anll---+ O.

Hence A is compact.
3. Let K: Lz ([a , b]) ---+ Lz ([a, b]) be the integral operator

(Kf)(t) = l b
ktt , s)f(s)ds,

where k is in Lz([a , b] x [a , b]) . By Section 2, Example 5, K is a bounded linear

operator and

IIKII ::::: (l l
b b
Ik(t, S)IZdsdt) I /Z = IIkll . (16.2)

We construct a sequence of operators of finite rank which converges, in norm, to

K as follows :
Let CPI , cpz , .. . be an orthonormal basis for Lz ([a , b]) . Then <I> ij (t , s) = CPi (t) x
i/Jj(s), i, j 1,2, ... is an orthonormal basis for Lz([a, b] x [a, b])
by Theorem 1.14.1.
Thus k = L 0=1 (k , <l>ij)<I>ij . Define
n
kn(t , s) = L (k, <l>ij)<I>ij(t, s).
i ,j=1

Then
Ilk - kn II ---+ O. (16.3)
Let K n be the integral operator defined on Lz([a, b]) by

(Knf)(t) = l b
kn(t, s)f(s)ds.
94 Chapter II. Bounded Linear Operators on Hilbert Spaces

Now K n isaboundedlinearoperatoroffiniteranksince~Kn C Sp{gll, . .. , gln} .

By (16.3) and (16.2) applied to K - K n .
II K - s; II S Ilk - kn II -+ O.
Hence K is compact.
4. Given a complex valued function aCt) which is continuous on [a, b], let
A: Lz([a, b]) -+ Lz([a, b]) be the bounded linear operator given by
(Af)(t) = a(t)f(t).
If a (to) f= 0 for some to E [a, b], then A is not compact. Indeed, it follows from
the continuity of aCt) that la(t)1 ::: la~o)1 > 0 for all t in some compact interval J
containing to. Let {(,on} be an orthonormal basis for Lz(1). Define gln to be 'iin on
J and zero elsewhere on [a, b]. Then II gln II = 1 and for n f= m,

IIAgl n - Aglm liZ = l b

la(t)!z!gln(t) - glm(t)IZdt

::: la(~)IZ i l'iin(t) _ 'iim(t)IZdt = la(~)IZ.

Hence {Agln} does not have a convergent subsequence.
We conclude this section with a compactness result for band operators. This
requires some preparations.
An infinite matrix (ajdfk=O is called a band matrix if all its nonzero entries are
in a finite number of diagonals parallel to the main diagonal, that is, there exists
an integer N such that
ajk = 0 if U - kl > N. (16.4)
If this number N is equal to 1, then the matrix is called tridiagonal.
An operator A on £z induced by a band matrix is called a band operator. In
other words, an operator A on £z is a band operator if and only if the matrix of A
with respect to the standard orthogonal basis eo, el , ez , .. . oflz is a band matrix .

Theorem 16.4 In order that a band matrix (ajk)J:k=O induce a bounded linear
operator A on £z it is necessary and sufficient that
sup lajkl < 00, (16.5)
j.k=I.Z....
and in this case
N
IIAII S L sup lajkl,
v=_N j-k=v
(16.6)

where N is the integer in (16.4) . Furthermore. the band operator A is compact if

and only if
lim ajk = O. (16.7)
j.k-*oo
2.16 Compact Operators 95

Proof: If the operator A induced by the matrix (ajdJ:k=O is bounded, then

for each j and k. Hence (16.5) holds. To prove the reverse implication, assume
(16.5) holds. For v ~ 0 let D; and D- v be the operators on £z defined by

D; (XO, Xl, Xz, . ..) = (0, . . . 0, avoxo, avlxl, .. .),

'-v-'
v

Condition (16.5) implies that the operator D n is bounded for each integer nand

IIDn II = sup lajkl. (16.8)

j-k=n

Let A = L~=-N D n . Then A is a bounded linear operator on £z, and the matrix
of A with respect to the standard basis of £z is precisely the band matrix (ajk)J,k=O'
Thus (16.5) implies that the matrix (ajk)J,k=O induces a bounded linear operator on
lz. Futhermore, IIA II :s L~=-N IIDn II . By combining this with (16.8) we obtain
(16.6). Notice that (16.6) implies that

IIAII :s (2N + 1) sup lajkl . (16.9)

j.k=1.2....

Now, let condition (16.7) hold. Let Pn be the orthogonal projection on £z

defined by
Pn(XO , xl , Xz, . ..) = (Xo , . .. , Xn , 0, 0, .. .),

and set Qn = I - Pn . Relative to the standard orthogonal basis of £2 the matrix

of the operator Qn A Qn has the following block form

where An is the band matrix (ajk)J,k=n+I' Notice that

and hence A - QnAQn is finite rank for each n . From (16.9) it follows that

II QnAQn II :s (2N + 1) sup lajkl,

j.k=n+l.n+Z....

and hence (16.7) implies that II QnAQn II --+ 0 for n --+ 00. Thus in the operator
norm, A is the limit ofa sequence of finite rank operators, and hence A is compact.
96 Chapter Il. Bounded Linear Operators on Hilbert Spaces

Conversely, let A be compact. Then from Lemma 17.8 (which is independent

of the results in this section) it follows that II QnAQn II -+ 0 for n -+ 00. For
j, k ~ n + 1 we have

lajkl = I(Aeb ej}1 = I(QnAQneb ej}1

~ IIQn AQnll .

Since II QnAQn II -+ 0 for n -+ 00 we see that (16.7) holds. o

2.17 The Projection Method for Inversion of Linear Operators

A scheme for approximating the solution to an infinite system of linear equations

was given in Theorem 9.1. The proofofthe theorem relied on the fact that the oper-
ator corresponding to the coefficient matrix was the limit, in norm, of a sequence
of operators of finite rank.
Ifthe restrictions on the coefficient matrix are weakened so that the sequence of
these operators does not necessarily converge in norm, then other arguments may
be used.
Let us begin with the infinite system

CllXl + Cl2X2 + = YI
C2l Xl + C22 X2 + = Y2 (i)

where (Yl, Y2, ... ) is in £2. As in Theorem 9.1, the natural approach to approxi-
mating a solution to (i) is the following :
For all n sufficiently large, find a solution (x~I), ... , x~n» to the finite system

ClIXI + + ClnXn = Yl
C21Xl + + C2n Xn = Y2
(ii)

CnlXl + ... + CnnXn = Yn·

Then, hope full y, Vn = (xl(n) , .•• , X n(n) , 0 , 0 ) converges III

, ... . L20
to a so Iunon
.
of (i).
To put this scheme within the framework of operator theory, define the operator
C onl2 by

where
00

{3j = I>jkab j = 1,2, . .. .

k=l
2.17 The Projection Method for Inversion ofLinear Operators 97

We shall assume that E]:I E~I ICjd < 00 so that C is in .c(lz) (Example 3 in
Section 2). Define P; E .c(lz) by

we see that (ii) is equivalent to

Note that
P; = Pn, IIPn ll = 1 and Pnx --+ x for every x E lz.

The scheme prescribes that Un = (xi n), .. . , x~n), 0, 0, .. .) be found so that

and {un} converges in lz to a solution of (i).

We now consider a more general approximation scheme.

Definition: Let {P n } be a bounded sequence of projections in .cOol) with the

property that Pnx --+ x for all x E 11.. Given an invertible operator A E .c(11.), the
projection method for A seeks to approximate a solution to the equation Ax = y
by a sequence {x n } of solutions to the equations

PnAPnx = Pny, n = 1,2, ... (17.1)

This projection method for A is said to converge if there exists an integer N such
that for each y E 11. and n ~ N, there exists a unique solution X n to equation (17.1)
and, in addition, the sequence {x n } converges to A -I y .
We denote by Il(Pn) the set of invertible operators A for which the above
projection method for A converges .
Unless a specific sequence of projections is used, we shall always assume that
we have the general sequence {P n } with the properties above .
Since Pnx --+ x for each x E 11., sUPn II Pnx II < 00 for each x E 11.. Hence

sup IIPn ll < 00 (17.2)

n

by Theorem 14.3. Also, since Pn is a projection, we know from Theorem 15.1

that ::SPn is closed and therefore is a Hilbert space.

Theorem 17.1 Let A E .c(H) be invertible. Then A is in Il(Pn) if and only if

there exists an integer N such thatfor n ~ N , the restriction ofthe operatro PnAPn
to ::SPn has a bounded inverse, denoted by (PnAPn) -I, and

sup II (PnAPn)-1 II < 00 . (17.3)

n?:.N
98 Chapter II. Bounded Linear Operators on Hilbert Spaces

Proof: Suppose there exists an integer N such that for n ~ Nand Y E 71., the
equation PnAPnx = Pny has a unique solution X n E ;:S Pn. In addition, assume
(17.3) holds . By (17.2) and (17.3),

M = sup {II (PnAPn)-1 II + IIPnlD < 00 .

n~N

Then for x = A-1y,

IIx n - Pnxil = II (PnAPn)-1 PnAPn(x n - Pnx)1I ::s MIlPnAPnx n - PnAPnxll

= MIIPny - PnAPnxll ::s M 11 Y- APnxll -+ M 2 11 Y- Axil = o.
2

Hence
X n= X n - Pnx + Pnx -+ X = A -I Y,
which shows that A E I1(Pn). Conversely, assume A E 11 (Pn) . Then there
exists an integer N such that for n ~ N, the operator PnAPn restricted to ;:SPn
is a one-one map from the Hilbert space ;:SPn onto ;:SPn. Hence (PnAPn)-1 is a
bounded linear operator on ;:S Pn by Theorem 14.1. Since X n = (PnAPn)-1 PnY is
the unique solution to equation (17.1) for n ~ N and since A E 11 (P n ), we have
that

Hence
m = sup II (PnAPn)-1 Pnll < 00
n~N

by Theorem 14.3. Thus

II (PnAPn)-1 Pnyll = II (PnAPn)-1 r, PnYII s mllPnyll

which shows that
sup II (PnApn)-1 II s m.
n~ N

o
Corollary 17.2 Let {Pn} be a sequence ojprojections on a Hilbert space H. Sup-
pose II Pn II = 1, n = 1, 2, ... and Pnx -+ x for every x E H . If A E £(71.) with
III - All < 1, then A E I1(Pn).

Proof: For any x E 71.,

(17.4)

Let B = I-A . Since IIPnBPnll ::s IIBII < 1, both A = I - B and 1- PnBPn
are invertible . It is easy to see that (I - PnBPn)-1 maps ;:SPn onto ;:S Pn. Thus it
follows from (17.4) that PnAPn is invertible on ;:s Pn and

(PnAPn)-1 PnY = (1- Pn(I - A)Pn)-1 Pny. (17.5)

2.17 The Projection Methodfor Inversion ofLinear Operators 99

By Theorem 8.1,

I 1
11(1 - Pn (1 - A)Pn)- II :s -l---II-Pn-(-I---A-)-Pn-1I
:s (l - III - AID-I . (17.6)

The corollary now follows from (17.5), (17.6) and Theorem 17.1.
Let Pn be the orthogonal projection of £2 onto the first n coordinates, i.e.,

Pn(Xo, XI, X2, . . .) = (xo, ... , Xn-I,O, 0, . ..).

For any X = (xQ, XI, X2 , .. .) E £2 we have that

IIx - PnxII2 = L IXjl2 ---+ 0 (n ---+ 00) ,

j?:.n

and hence Pnx ---+ X for each X E £2. Now, let A be invertible on £2. We say that
the finite section method converges for A if the projections method for A relative
to the projections {Pn}~o converges, that is, A E TI{Pn}.
In this setting identify ':SPn with en, and we refer to PnAPn as the nth section
of A.
Similarly, for an invertible operator A on £2 (Z) we say that the finite section
method converges for A if A E TI{Pn}, where now Pn is the orthogonal projection
on £2 GZ) given by

Pn(Xj)jEZ = (... ,0,0, x- n, . . . , x n, 0, 0, . ..) ,

with n = 0, 1, 2, .. . notice that

IIPnx _x1I 2 = L IXjl2 ---+ 0 (n ---+ 00),

Ijl >n

and thus Pnx ---+ X for each X E £2(Z) . o

Examples: 1. Let A be the operator on £2(Z) given by the following two diagonal
doubly infinite matrix

o 0
A= -c OJ 0
o -c 1

Here lei < 1, and, as usual, the box indicates the entry in the zero-zero position.
We shall show that the finite section method converges for A .
100 Chapter II Bounded Linear Operators on Hilbert Spaces

Notice that A = I -cV, where V is the forward shift on z-fz). Since IIVII = 1,
we have II c V II < 1, and hence A is invertible with

00

A-Ix = L::>JvJ x E l2(Z).

J=o

It follows that the matrix of A -I relative to the standard basis of l2 (Z) is given by

0 0 0 0

C 0 0 0
A- I = c2 c ITJ 0 0
c3 c2 c 1 0

c4 c3 c2 C

Next, consider the truncation An = PnAPnI~Pn» where Pn is as in the paragraph

preceding this example. Identifying ~Pn with C 2n+1 we see that the matrix of An
is given by

-c
1 01 00 0)
0
An = 0 -c 1 O.
[ .. ..
o 0 0 1
A straight forward calculation shows that An is invertible and

o 0
1 0 ... OJ
... 0
c 1 o .
1

By comparing the matrices for A -I and A; I it is clear that

and hence for the operator A the finite section method converges.
The above conclusion can also be derived directly from Corollory 17.2. Indeed,
I - A = c V, and thus II I - A II < 1.
2.17 The Projection Method for Inversion ofLinear Operators 101

2. In this example we consider the operator B on £z(Z) given by

-c 0 0
B = 1 I-c I 0
o 1 -c

Again the matrix is two diagonal and [c] < 1. Notice that B differs from A in the
previous example only in the fact that the order of the two diagonals with nonzero
entries is interchanged.
Recall that Y is invertible, and that y-I is the backward shift on £z(Z) . Thus
B = -c I +Y= (/ - C y-I) Y.
Since IIcy- 11i = [c] < 1, we conclude that B is invertible, and
00
B-Ix = y-I(/ -cy-I)-I x = Lcjy-j-I x,
j=O
for each x E £z(Z) . Thus

0 1 c cZ c3
0 0 1 c cZ
B- 1 = 0 0 @] 1 c
0 0 0 0 1
0 0 0 0 0

Let us now consider the truncated operator B n = PnBPn, where Pn is as in

previous example.
Identifying ::SPn with C Zn+ 1 we see that

B, (I
~ ;c ;c
!J
A straightforward calculation shows that

!)
0 0
l/c 1 0
B-1 = _~ ( Iz
l/c l/c 1
n c .

1/c2n+ 1 l/c Zn l/c Zn - 1

102 Chapter II Bounded Linear Operators on Hilbert Spaces

o
By comparing the matrix of B -I with those for B I, B ,I, B2"1 ... we see that for
B the finite section method does not converge . This conclusion also follows from
Theorem 17.1 because in this case

(n ~ 00) .

Let us remark that in this example the operator B is invertible and the equation

has a unique solution x, E ';SPn for each y E f2(Z), but the sequence xn, XI, X2, ...
need not converge. For instance, if y = (0jO) jEZ, then

Xn = -~ (1, ~, ... , c2 :+ 1) , n = 0, 1,2, . .. ,

and this sequence does not converge .

A further analysis of the different behavior of the finite section approximation
in these two examples and the reason behind it will be presented in the last section
of Chapter III.

Lemma 17.3 Let H be a Hilbert space. Given A E £(H), suppose there exists a
number c > 0 such that
I(Ax, x)] ::: cllxll 2 (17.7)
for all X E H. Then A is invertible and II A -I II .:::: ~.
Proof: Since
IIAxllllxll::: I(Ax,x)l::: cllxll 2
we have
IIAxll ::: cllxll . (17.8)
Thus A is injective and its range ';SA is closed by Corollary 14.2. We show that
';SA = H. Let P be the orthogonal projection onto ';SA. Given y E H, we know
that y - Py E KerP = ';Spl.. = ';SAl... Hence

0= I(A(y - Py), y - Py)l ::: clly _ pY1l2 .

Therefore, y = Py E ';SA . The inequality IIA -III .:::: ~ follows from inequality
(17.8). 0

Theorem 17.4 Let H be a Hilbert space. Given A E £(H), suppose there exists
a number c > 0 such that
I(Ax,x)1 ::: cllxll2
for all X E H. Let {Pn} be a sequence oforthogonal projections on H with the
property that Pnx ~ x for all x E H . Then A E Il (Pn).
2.17 The Projection Methodfor Inversion ofLin ear Operators 103

Proof: Since Pn is a self adjoint projection,

/(PnAPnx, Pnx)1 = /(AP nx, Pnx)1 :::: c ll Pnx ll 2 , x E H.
lt follows from Lemma 17.3, that PnAPn, restricted to (:sPn), is invertible and
II (PnAPn)-1 II S ~, m = 1,2, . . .
Hence A E IT (Pn ) by Theorem 17.1. 0

Coronary 17.5 Given A E £ (H ), suppose there is a number c > 0 such that

2
Re (Ax , x) :::: cllxll .

Let {Pn} be a sequence of orthogonal projections on H with the property that

Pnx -+ x f or all x E H . Then A E IT(Pn ).

Next we show that the projection method is stable under certa in perturbations.

Theorem 17.6 Suppose A E IT (Pn) . There exists a y > 0 such that if B E L (H )

and liB II < y , then A + B E IT ( Pn) .
Proof: There exists an integer N such that for n :::: N , the operator PnAPn,
restricted to :sPn , is invertible and
M = sup (II ( PnAPn)- 1II + II PnlD < 00 (17.9)
n?:.N
now for n > N ,
(17.10)
and
II (PnAPn)- IPnBIl < ~ifllBIl < 2~2 ' « : », (17.11)

Hence I + ( PnAPn)-1 PnB is invertible and

1
11 (1 + ( PnAPn)-1 Pn B ) - I II -< 1 - II( PnA P n) Pn BIl
_< 2, n >
_ N. ( 17.12)

by Theorem 8.1 and inequality (17 .11).

Since (I + ( PnAPn) - 1PnB )-1 maps :SPn onto :sPn, it follows from (17.9 ),
(17.10) and (17.12) that for n :::: N ,
( Pn( A + B )Pn )-1 = (I + (PnAPn )-1 PnB )-1 (PnAPn)- 1 (17.13)

and
sup II (Pn (A + B )Pn)-11l S 2M. (17.14)
n?:.N

Hence for li B II < n1m(z;&r, I A~ I I ) '

A + B = A (I + A -I B )
is invertible and sup II (Pn (A + B )Pn)-11l < 00. The theorem now follows from
n?:.N
Theorem 17.1. o
104 Chapter II. Bounded Linear Operators on Hilbert Spaces

Theorem 17.7 Suppose A E Il(Pn) and suppose K is a compact operator in

£(H) with the property that A + K is invertible. Then A + K E Il(Pn).

First we prove the following Lemma.

Lemma 17.8 Suppose {Tn} is a sequence ofoperators in £(H2) such that Tny ~
Tyfor all Y E H2 . If K is a compact operator in £(HI, H2) , then

IITnK - TKII ~ o. (17.15)

Proof: It follows from Theorem 14.3 that M = sup II Tn II < 00 . Suppose (17.15)
n
does not hold. Then there exists an B > 0, and a subsequence in'} of {n} , vectors
{xnl} such that IIxn 'll = 1 and

(17.16)

Since K is compact, there exists a subsequence {xnll} of {xn l} such that Kx;» ~
Y E H2 . Hence

IITnllKxnll - TKxnll1l ::s II TnllKxnll - Tnllyll + II TnllY - Tyll

+ IITy- TKxnll1l
::s MIIKxnll - YII + II TnllY - TYII + 1IT11I1Y -Kxnllil ~ 0

which contradicts (17.16).

Proof of Theorem 17.7 Since the projection method for A converges there exists
an interger N such that for n ~ N, the operator PnAPn, restricted to <;SPn, is
invertible and
sup (II(PnAPn)-1 II + IIPnll) < 00 (17.17)
n?:.N
now form ~ N,

(17.18)

moreover
(17.19)
Indeed, from the definition of convergence of the projection method for A, we
know that
(PnAPn)-1 Ps Kx ~ A-I Kx , x E H.
Hence (17.19) follows from Lemma 17.8. Since I + A -I K = A -I (A + K) is
invertible, we have from (17.19) and Corollary 1.8.3 that I + (PnAPn)-1 PnK is
invertible for n ~ NI ~ Nand

sup 11(/ + (PnAPn)-1 PnK)-11i < 00 . (17.20)

m?:.Nl
2. J8 The Modified Projection Method 105

Since (/ + PnAPn)-1 maps <;J Pn onto <;J Pn, it follows from (17.17), (17.18) and
(17.20) that

( Pn( A + K )Pn )-1 = (/ + ( PnAPn) - 1PnK )-I (PnAPn )-1

and
sup II (P n(A + K )Pn)-1 1l < 00.
n?NI

The theorem is now a consequence of Theorem 17.1.

In our theorem on the convergence of the projection method we always assume
that the operator A is invertible. However, if we only assume

sup II ( PnAPn)- 1II < 00 ,

n

then A is one-one and has a closed range. Indeed suppose M = sup II (PnAPn)-IIi .
Then for all x E 11,
1
IIPnAPnxll 2: M IIPnxll ·

If follows from Corollary 14.4 that PnAPnx -+ Ax .

Hence

·
IIAxll = Iim II PnAP nx ll 2: -
I I'im IIP x ll = -lIxll·
1
n - HXl M n-->oo n M
Thus
1
IIAx ll 2: M IIxll for all x E 'H.

By Corollary 14.2, A is one-one and has a closed range. o

2.18 The Modified Projection Method

Let {Pn } be a sequence of projections in £(11) with the property that Pi;» -+ x
for all x E 11. In the ordinary projection method for {P n }, the solution x of an
equation Ax = Y, when A is an invertible operator on 11, in approximated by
solutions X n E <;J Pn of the equations

PnAx n = Pn Y, n = 1,2, . . . . (18.1)

In some cases it is more convenient to approximate the original equation by the

equations
Fnx n = PnY, n = 1,2, .. . , (18.2)
where F; E £ (<;JPn) for each nand

lim II Fn Pnx - Axil = 0, x E 11. (18.3)

n--> oo
106 Chapter Il. Bounded Linear Operators on Hilbert Spaces

We refer to this way of inverting A as the modified projection method. One of the
reasons for replacing (18.1) by (18.2) is that it is sometimes easier to invert the
operators F n than the operators PnAI ~Pn .
The next result is the analogue of Theorem 17.1 for the modified projection
method

Theorem 18.1 Let {Pn} be a sequence ofprojections in £(H) such that Pnx -+ x
for each x E H, and let A be an invertible operator on H . Also, let Fn E
£(~ Pn ) , n = 1, 2, . .., be a sequence ofoperators satisfying (18.3), and assume
that for n ~ N the operator Fn is invertible. Then

(18.4)

if and only if
sup IlFn-11i < 00. (18.5)
n?:.N

Proof: The necessity of (18.4) is an immediate corollary of Theorem 14.3. To

prove the sufficiency, take n ~ N. Then Fn is invertible, and for each y E H we
have

The second term in the right hand side converges to zero. The first can be evaluated
as follows:

IlFn-IPny-PnA-IYII:S IlFn- IIIIIPny-FnPnA-lyll

:s IlFn-III(I!Pny - yll + II(FnPn - A)A-lylI) .

Recall that PnY -+ y. Using (18.3) with x = A-Iy and (18.4) we see that the
preceding inequality implies that the first term in the right hand side of (18.6)
converges to zero when n -+ 00. Thus (18.4) is proved. 0

Example: Let the operator A on £2 be defined by the infinite tridiagonal matrix

6 -37 -30 0"

-2 0 ,)
A = 0 -2 7 -3 .
( o.. 0 -2 7
..

We shall use the modified projection method to show that A is invertible and to
find its inverse.
For each n we let Pn be the usual orthogonal projection onto the first n + 1

0,0,...).
coordinates :
Pn(XO, XI, X2, ...) = (xo , . .. , Xn ,
Here we identify ~ Pn with C n + l . Notice that Fn differs from PnAPn only in the
entry in the right lower comer. It follows that

lim IlFnPnx - PnAPnxll = 0, x E 'H,

n--+oo

1
and hence (18.3) holds. A direct computation shows that
2- 1 2- 2 n
1 2- 1 .. .. .. 2-
2-n + 1
3- 1 1 . .. 2- n +2

By letting n go to infinity it is natural to expect that the inverse of A is equal to the

operator R given by

..
.... ) (18.7)

Notice that R is a bounded linear operator on f2 . In fact (cf., the proof of

Theorem II.16.4) we have

1(( "21)-1 + 31( 31)-1) "21

IIRII ::: "5 1- 1- = (18.8)

By a direct computation one sees that RA = AR = I, and hence A is invertible

and A-I = R .
Since Fn-I = PnRI ~Pn , we have IlFn-11i ::: IIRII, and hence

sup IlFn-11i < 00.

n:,:O

Thus we can apply Theorem 8.1 to show that for this example the modified pro-
jection method converges, that is, for each y E 'H.
108 Chapter II. Bounded Linear Operators on Hilbert Spaces

2.19 Invariant Subspaces

Sometimes properties of an operator A E £(H) can be determined rather easily

by considering simpler operators which are restrictions of A to certain subspaces.
We now present some elementary facts about invariant subspaces.

Definition: Let A be in £(H). A subspace M of H is called A-invariant if

AMcM.

It follows readily from the continuity of A that if M is A-invariant, then so is M.

Examples: 1. Let {qJI, qJ2, . .. } be an orthonormal basis for H . Suppose that

the matrix corresponding to A E £(H) and {qJn} (cf. Section 4) is upper
triangular, i.e.,
aij = (AqJj, qJd = 0, i > j.

2. If, in Example 1, (aij) is lower triangular, i.e., aij = 0 if i < j, then for each
n , SP{qJn+l, qJn+2, ... } is A-invariant.
3. Given k E L2([a, b] x [a, b]), define K on L2([a, b]) by

(Kf)(t) = it k(t, s)!(s) ds.

For each t E [a, b], the subspace

M t = (f E L2([a , b]) : ! = 0 a.e. on [a, t]}

is invariant under K .

Theorem 19.1 If a subspace M is A-invariant, then M1.. is A*-invariant.

In particular, if A is selfadjoint, then M 1.. is A-invariant.

Proof: Suppose v E M1.. . For any U EM, Au is also in M . Hence

0= (Au, v) = (u, A*v).

Therefore, A*v E M1... o

The following theorem gives a connection between invariant subspaces and orthog-
onal projections.

Theorem 19.2 A closed subspace M cHis A-invariant ifand only ifAP = PAP,
where P is the orthogonal projection onto M
2.20 The Spectrum ofan Operator 109

Proof: If AM eM, then for each u E 'H, APu E AM eM. Therefore, PAPu =
APu. Conversely, if PAP = AP, then for v E M,

Av =APv =PAPv E M.

Given A E £(1t) and given a closed subspace M C 'H, let P be the orthogonal
projection onto M . Then Q = I - P is the orthogonal projection onto Ml.. and
x = Px + Qx. If we identify with each x E n the column vector (~: ), then
the operator A can be represented as a block matrix

(
A ll
A21
AI2) .
A22

To be specific, suppose Ax = y . Since

A =PAP+QAP+PAQ+QAQ

and x = Px + Qx, Y = Py + 0', we get

PAP PAQ ) (Px) = (py)
( QAP QAQ Qx 0"
It is easy to see that M is A-invariant if and only if A21 = QAP = 0,
i.e., the block matrix is upper triangular. Similarly, M l.. is A-invariant if and
only if A 12 = PAQ = 0, i.e., the block matrix is lower triangular. These remarks
give alternate proofs of Theorems 19.1 and 19.2.
A problem which is still unsolved after years of effort by numerous mathemati-
cians is whether every bounded linear operator on a Hilbert space has a nontrivial
closed invariant subspace . It is known, for example, that every compact linear
operator on a Hilbert space (more generally on a Banach space) has a nontrivial
closed invariant subsapce. For an exposition of the invariant subspace problem
see [PS].

2.20 The Spectrum of an Operator

For a linear operator on a finite dimensional Hilbert space 'H. it is well known from
linear algebra that the equation Ax - Ax = y has a unique solution for every y E 'H
if and only if det().oij - aij) t- 0, where oij is the Kronecker delta and (aij) is the
matrix corresponding to A and some orthonormal basis of 'H. Therefore, in this
case AI - A is invertible for all but a finite number of )..
If1t is infinite dimensional, then the set a(A) of those ). for which AI - A is
not invertible in more complex. In this section we touch upon some properties
of a(A ).
110 Chapter II Bounded Linear Operators on Hilbert Spaces

Definition: Given A E .cO-i), a point A E C is called a regular point of A if

AI - A is invertible. The set peA) ofregular points is called the resolvent set of A .
The spectrum a(A) of A is the complement of peA) .

Theorem 20.1 The resolvent set ofA E .c(1i) is an open set containing {A I IAI >
IIAII}. Hence a(A) is a closed bounded set contained in {A I IAI .:s IIAII}. Fur-
thermore.for A E oa(A), the boundary ofa(A), the operator AI - A is neither
left nor right invertible.

Proof: If IAI > IIAII, then AI - A = A(I - fA) is invertible since II fAil < 1.
Thus A E peA) . Suppose AO E peA) . Since AoI - A is invertible, we have by
Corollary 8.2 that AI - A is invertible if IA- Aol < II (AO - A)-III-I. Hence peA)
is an open set.
Next, take A E oa(A), and assume AI - A is left invertible. Since a(A) is
closed. A belongs to a(A), and hence AI - A is not two-sided invertible. It
follows that codim ;s(AI - A) =I- O. By Theorem 15.3 we have

codim;s(A'I - A) = codim;s(AI - A)

whenever IA- A'I is sufficiently small. But each open neighborhood of A contains
points of resolvent set of A. In particular, each open neighborhood of A contains
points A' such that codim ;s(A'I - A) = O. We reached a contradiction. Thus
AI - A is not left invertible. In a similar way, using Theorem 15.5, one proves
that AI - A cannot be right invertible. 0

Notice that an operator A E L(1i) is invertible if and only if its adjoint A * is

invertible, and in that case

Indeed, if AB = BA = I, then by Theorem 11.3(iv) we have B* A * = A * B* =

I . Hence A* is invertible, and the above identity holds . Since (A*)* = A,
the invertibility of A * also implies that of A .
From the above remark it follows that AI - A is invertible if and only if XI - A *
is invertible, and hence A E a(A) if and only if XE a(A*) . We conclude that

a(A*) = {XI A E a (A)}

Examples: 1. Let K be the finite rank operator on the Hilbert space 'H given by
n
Kx = ~)x, C{Jj} Vrj '
j=1
2.20 The Spectrum ofan Operator 111

If 'H is infinite dimensional, then the spectrum

a (K ) = {A =I- 0 I det (A8ij - (ljI j , ffJi»)7, j=1 = O} U {OJ.

This follows directly from Theorem 7.1.
2. Let ffJ I, ffJx , . . . be an orthonormal basis of the Hilbert space 'H, and let
{AI , A2 , . . .} be a bounded sequence of complex numbers. For x E 'H define
00

Ax = LAK(X,ffJk)ffJk.
k=1

From Example 2 in Section 2 we know that A is a bounded linear operator

on 'H. Notice that the matrix of A corresponding to the basis ffJI , ffJ2, . .. is an

J
infinite diagonal matrix with diagonal entries AI , A2, . .., i.e.,

= diag(ll . 12. '3 • . . .J.

We claim that the spectrum of A is equal to the closure of the set

{Ajij = 1,2, ...}. In other word s,

a (A ) = {A E q i~flA - Aj l = O} .

To prove this result , it is sufficient to show that A E p (A ) if and only if

inf'[), - Aj l > O. (20.1)

J

Assume condition (20.1) holds . Then , according to Example 2 in Section 2, the

formula
00 I
Bx = L A _ A . (x , ffJ j ) ffJj (x E 1t)
j=1 J

defines a bounded linear opeator on 'H. In fact,

II BII = sup
j
1__1
A - Aj
1
= {inf lA - Ajl} -I <
J
00.

It is straightforward to check that AB = BA = I , and thus B = (AI - A ) -I, and

A E p (A ) .
Conversely, let A E p ( A ), and assume (20.1) is not satisfied. Then there is a
subsequence {Ak'} of {Ad such that Ak' --+ A. But then
112 Chapter II Bounded Linear Operators on Hilbert Spaces

But this is impossible, because

1= IIqJk'iI = II(U - A)-IUI.! - A)qJk'iI ~ II(U - A)-III(U - A)qJd,

and thus II(U - A)qJk' II ::=: II(U - A)-III-I . Thus AE peA) implies (20.1).
3. Let A be the backward shift on £2, i.e.,

A(al, a2, . ..) = (a2, a3,·· .).

We shall show that a(A) = {A E C] IAI ~ I}. If IAI > 1, then U - A =

A(I - A-I A) is invertible since IIA-I A II = lA-II < 1. If IAI < 1, then A E a(A)
because x = (1 , A, A2 , .. .) E Ker (U - A) . Finally, since a(A) is a closed set,
it follows that a(A) is the closed unit disc in C.
The spectrum of the forward shift S on £2 is also equal to the closed unit disc.
To see this note that S = A *, and use that

a(A*) = (XI A E a(A)} .

Since IAI ~ 1 if and only if IXI ~ 1, we get that a (S) = a (A *) is the closed unit
disc .
4. Let a(·) be a continuous complex valued function defined on the closed
bounded interval [c, d] . Define a bounded linear operator A on H = L2 ([ c, d]) by

(Af)(t) = a(t)f(t) , c ~ t ~ d.

Since «U - A)f)(t) = (A - a(t»f(t),

We have from Example 2 in Section 7 that U - A is invertible if and only if
A - aCt) =1= 0 for all t E [c, d],
1
«U - A)-I g)(t) = get),
A - aCt)
and
II(U-A)-III= max IA-a(t)I-I .
IE [c,d]

Hence a(A) = {a(t)lt E [c, d]} .

5. Let H = L2([0,I]) . For 0 < a < 1 let R a and R1 be the operators
introduced in Section 15, Example 3. As we have proved ther"e, the operator R1
Ja and
a
is a right inverse of R a and R a is not left invertible. Also, II R a II =
II R 111 = "ja. We show that
a

a(R a ) = {A I IAI s .Ja} . (20.2)

If IAI < ~, then IIAR11I < 1 and therefore! - AR1 is invertible . Since R; is
V U
a a
not invertible and
U - R a = Ra(AR1 - I),
a
2.20 The Spectrum ofan Operator 113

it follows thatAI-Ra is not invertible, i.e., A E a(Ra ) · IflAI > Ja = IIRall,then

A E p(Ra ) by Theorem 19.1. Sincea(Ra ) is closed and {A IIAI < Jal S; a(R
a ),
equality (20.2) follows.
We shall also determine the spectrum of the operator Ra on L2([0, 00]); see
Example 6 below. For this purpose we need the following results .

Theorem 20.2 Given A E L(H), where H is a Hilbert space, let PI (A) be an

open, connected subset of p(A). Let Ho be a closed A-invariant subspace ofH .
Define Ao = AjHo, the restriction of A to Ho. Then either

(i) PI (A) c p(Ao) or (ii)PI (A) C a(Ao) .

/f PI (A) c a(Ao), then for all A E PI (A), the operator AI - Ao is strictly left
invertible and 0 < codim(AI - Ao) is constant on PI (A).

Proof: Obviously,

PI (A) = (PI (A) n p(Ao» U (PI (A) n a(Ao» . (20.3)

The set PI (A) n p(Ao) is open . Also , PI (A) n a(Ao) is open . To see this , suppose
A E PI(A) n a(Ao) . Let P be the orthogonal projection onto Ho. Then (AI-
A)-ljHo its a left inverse for AI - Ao on Ho since

P(A - A)-I(AI - Ao)xo = P(A - A)-I(A - A)xo = Pxo = xo.

Since A E a(Ao) , AI - Ao is strictly left invertible . Hence it follows from

Theorem 15.3 than is an interior point of PI (A) na(A) . Therefore PI (A) na(A)
is an open set. Since PI (A) is an open connected set, it cannot be expressed as the
union of two non-empty mutually disjoint open sets . Thus from (20.3) we have
that either PI (A) n p(Ao) = 0 or PI (A) n a(Ao) = 0. Hence either (i) or (ii)
holds .
If PI (A) S; a(Ao), then by Theorem 15.3, thefunctioncp(A) = codim(AI -Ao)
is continuous on PI (A) . Since PI (A) is connected, cp(A) is constant (and greater
than zero). 0

Corollary 20.3 /f p(A) is connected, a(Ao) C a(A).

Proof: The open set p(A) is not contained in a(A) since p(A) is unbounded.
Hence p(A) C p(Ao) by Theorem 19.2. Thus a(Ao) C a(A). 0

Corollary 20.4 If): is in the boundary ofa(Ao), i.e.,

A E a(Ao) n p(Ao) , then A E a(A).

114 Chapter II. Bounded Linear Operators on Hilbert Spaces

Proof: Suppose A E p(A) . Then there exists an open disc D such that A E D C
p(A) . Since A E p(Ao), D n p(Ao) i= 0 . Also, A E D n a(Ao). This contradicts
Theorem 19.2 as D is an open connected subset of p(A) .
An operator U E £(H) is called an isometry if IIUxll = [x] for each x E H .
If, in addition, ;sU = H, the operator U is said to be unitary. Notice that a unitary
operator U is invertible, and both U and U- I have norm 1. This implies that for
a unitary operator U the spectrum belongs to the unit circle, i.e.,

a(U) C {A E q IAI = I}. (20.4)

Indeed, for IAI > 1 we have A E p(U) by Theorem 19.1, and for IAI < 1 we have
IIAU- III < 1, and thus AI - U = -U(I - AU-I) is again invertible. In general,
the inclusion in (20.4) can be strict. For instance, if U = I, where I is the identity
operator on H, then a(U) = {l}. D

Examples: 6. Let H = L2([0, 00]) . For 0 < ex < 1, define Rex on H by

(Rexcp)(t) = cp(ext) .

The operator y'ciRex is an isometry since

(20.5)

also ;s(y'ciRexcp) = H, for given 1/1 E H, take cp(t) = 1/I(~t), 0:::; t < 00. Then
(Rexcp)(t) = Rex1/l(~t) = 1/I(t). Hence y'ciRex is unitary and by (20.4) we have
a(y'ciR ex) C {AlAI = I} which implies

(20.6)

By making use of Corollary 20.4 we shall show that

(20.7)

Let Ho = {cp E HI cp(t) = 0, 0:::; t :::; I}. Clearly Ho is a closed Rex-invariant

subspace ofH. Define the linear operators andR2 R1 on Ho as follows:
a

and
o { 0, 0 :::; t :::; 1,
(R bcp)(t) = cp(~t), t > 1.
2.20 The Spectrum ofan Operator 115

Then

Hence II R~ II s .ja. Infact,

a
(20.9)

Indeed, define
I,
cp(t) =
{ 0,
Then cp E 1-lo and from (20.8) we have

and
1 1
IIcpll2 = f a Idt = _.
ha ex
Hence
1
IIR~II.ja = IIR~III1CPIl2: IIR~cpli = 1

or II R~ II 2: .ja, Thus (20.9) holds . Also

a

Ro °
1 Racp = sp, cp E 1-lo, (20.10)
a

since
0 0 )
( R ~ Racp (t)
{ 0, 0 s t ~ 1
= R2(cp(~t» = cp(t), t > 1.
The operator R2 is not invertible since 1m R2 i= 1-lo. To see this, define

1ft(t) = {01" t E [1, ~],

t rt [1, ~] .

Clearly 1ft E 1-lo and for all cp E 1-lo,

1
(R~cp)(t) = cp(ext) = 0, 1 < t < -.
-ex

Thus 1ft rt :JR2. Next we show that

(20.11)
116 Chapter II Bounded Linear Operators on Hilbert Spaces

From (20.5) we have II R211 = Jet. Hence if 1.1..1 > Jet, then A E p(R2) . For
1.1..1 < Jet, IIAR~.;a II < Jet.va = 1 by (20.9). Thus I - AR1 a
is invertible .
Equality (20.10) shows that

AI - RaO = (AR o1 - I)RaO (20.12)

a

Since (AR1 - /) is invertible and R2 is not, it follows from (20.12) that AI - R2

is not inve~ible, i.e., A E a(R2) , 1.1..1 < Jet . To summarize, we have shown that
{AIIAI < Ja} C a(R~) C {AIIAI ~ ~} .
Since a(R2) is closed, equality (20.11) follows. Suppose 1.1..1 = Jet. Then by
(20.11), Ais in the boundary of a (R2), hence we may conclude from Corollary 20.4
that A E a(R a ) . Equality (20.7) is now a consequence of(20.6).
7. Let'H = Lz ([ -Jr , it ]), and let

where a-I , ao and al are real numbers. On 'H we consider the operator A of
multiplications by a, that is

(Af)(t) = a(t)f(t), -Jr ~ t ~ Jr .

Example 4 shows that the spectrum a(A) consists of those A E <C for which

(20.13)

for some t . Taking the real and imaginary parts in (20.13) and using eit = cost
+ i sin t gives
ReA = ao + (a_1 +al)cost, (20.14)

~A = (al - a_I> sint. (20.15)

To describe a(A) a bit better we consider two cases .

Case 1: First assume that lall =1= la-II. In this case

ao - Re A ~A
- - - = cost, = sint. (20.16)
a-I + al al - a-I

Hence

ao - ReA)Z ( ~A )Z (20.17)
( a-I +al + al -a_1 = 1.
2.20 The Spectrum ofan Operator 117

Thus A lies on the ellipse given by (20.17). Conversely, if A lies on the above
ellipse , then there exists atE [-n . zr] such that (20.16) holds . Hence for this
choice of t , equation (20.13) is valid. So, if jaI l =1= la-II, then a (A) is precisely
the ellipse given by (20.17).

Case 2: Assume a l = a- I. Then ~A = 0, and from (20.14) and (20. 15) we get
A = ao + 2aI cos t . Thus a (A ) is a real line segment, namely

a (A ) = {r E IR I ao - 21aII ::: r ::: ao + 21a II} .

Case 3: Finally, assume at = -a_I . Then we see from (20.14) and (20.15) that
ffi A = ao, ~A = 2al sint .

Thus
a(A) = {A = ao + irl - 21all ::: r ::: 21all}.
We conclude with a remark about the operator A. Recall (see Section 1.13), that
the functions
1 .
'Pn(t) = ~ emf , nE Z , (20.18)
v2n
form an orthonormal basis of7i = L2([- n. n) .

The matrix of A with respect to this basis in the doubl y infinite tridiagonal matrix

ao al 0 0 0
a_I ao al 0 0
0 a-I §J al 0
0 0 a-I ao al
0 0 0 a-I ao

Here the box indicates the entry in the (0 , O)-position. Operators defined by matri-
ces of the above type are examples of Laurent operators. The latter operators will
be studied in the first section of the next chapter.
From Example 7 above we derive the following result.

Corollary 20.5 The spectrum ofthe forward shift V on £2 (Z) is equal to the unit
circle.
Proof: Let F be the map which assigns to each f in L2([-n. n )) the sequence
of Fourier coefficients of f relative to the orthonormal basis (20. 18). Then F is
an invertible operator from L2([-ot , n)) onto £2(Z) , Put A = F - I V :F. Then
a (V ) = a (A ). But A is the operator considered in Example 7 with al = 1 and
ao = a-I = O. The spectrum of A is given by (20.17), which in this case is the
unit circle. 0
118 Chapter II. Bounded Linear Operators on Hilbert Spaces

Exercises II

1. Let (Wj)~1 be a sequence of complex numbers. Define an operator D w

on £z by
Dw~ = (WI~I, wz~z, W3~3 , . . .).

(a) Prove that D w is bounded if and only if (Wj)~1 is bounded. In this

case IIDwll = SUPj I Wj I.
(b) Prove that inf', IWjlll~1I ~ IIDw~lI·
(c) Compute D~ for any kEN.
(d) Prove that D w is invertible ifand only if'inf', IWj I > O. Give a formula
for D;;;I.

2. LetDw be as in problem 1 and let inf', IWjl > oand suplui, I < 00. Which
of the following equalities or inequalities hold for any w?

(a) IIDwll = IID~'II

(b) IIDwll ~ IID~'II
(c) IIDwll s IID~'II
(d) IIDwll < IID~'II
(e) IIDwll > IID~'II'
3. Let W be such that SUPj IWj I< 00, find '25D", and KerD", .
4. Let A be an operator on £z given by the matrix (ajdrk=1 with
00

L sup lajj-kl < 00 .

k=-oo J

Prove that A is bounded and that

00

IIAII s L
k=-oo J
s~p laii-kl .

5. Let HI and Hz be Hilbert spaces. Define H = HI EB Hz to be the Hilbert

space consisting of all pairs {(UI, uz) : UI E HI, Uz E Hz} with

and an inner product

Exercises II 119

H is called the direct sum of HI, H2. Given Aj E £(HdandA2 E £(H2),

define A on H by the matrix

A = (~I 12)
i.e., A(UI, U2) = (AIUI, A2U2). Prove that A is in £(H) and IIAII =
max(IIAIII,II A211) ·
6. Foreachn E N, let U; be an operator on £2given by Un~ = (~n, ~n-I, . . . ,
~I, 0, 0, . . .). Find ;s u-. Ker u., II u, II and a matrix representation for
U« with respect to the standard basis.
7. Let aCt) be a continuous complex valued function on [a, b] . Define A :
L2[a , b] ~ L2[a , b] by (Af)(t) = a(t)!(t) . Find KerA and ;SA.
8. Given the operator A in problem 7, prove that ;s A is closed if and only if
aCt) i- 0 for t E [a , b] ora(t) is identically zero on [a , b].
9. Does the statement of problem 8 remain true if aCt) is allowed to be
discontinuous?
10. Let D w be an operator on £2 as defined in problem 1. Prove that ;sD w is
closed if and only if

11. Let K be an operator of finite rank on a Hilbert space H. For C{J E H,

n
K C{J = 'I)C{J, C{J; )1/1;.
;= 1

Suppose 1/1; E span{C{JI , . . . , C{Jn}.l for i = 1, . . . , n . Prove that I +aK is

invertible for any ex and find its inverse.
12. Which of the following operators

have finite rank and which do not?

(a) (Kf)(t) = L:J=I C{Jj(t) f: 1/Ij(s)!(s)ds.
(b) (Kf)(t) = L:J=I C{Jj(t) f~ 1/Ij(s)!(s)ds.

13. For the following operators K of finite rank, find the orthonormal basis
for which K has the diagonal matrix representation
120 Chapter II Bounded Linear Operators on Hilbert Spaces

(a) K~ = (3~1 + ~2, ~I + 3~2, 0, 0, . ..), K : £2 ---+ h

(b) K = (L:~I ~iC.fi)j-1 +~I, .fi~I ' ~~I ' .. .), K : £2 ---+ £2.

(c) (K({l)(t) = t f!:.rr ((l(X) COSX dx + cost f!:.rr X({l(X) dx, K : L2[-rr, zr]
---+ L2[-rr,rr] .

14. For the operators from problem 13 find K 1980 .

15. Given vectors ({ll, ,({In, 1f!1, . . . , 1f!n in a Hilbert space H, let N = span
{({ll, . .. , ({In, 1f!1, , 1f!n}. Define K E £(1i) by Kv = L:J=I (v , ({lj)x
1f!j. Prove that KN C N and that KNJ.. = O.

16. Given a separable Hilbert space H, Let K E £(1i) be an operator of

finite rank . Prove that there exists an orthonormal basis in 'H such that
with respect to this basis, K has the matrix representation of the form
(aij)0=1' where aij = 0 if; > nor j > n.

17. Let 1i be a Hilbert space and let K E L (1i) be an operator of finite rank
given by

n
Kv = I)v, ({lj)1f!j, ({lj ,1f!j E 'H, j = 1, . . . , n ,
j=1

Which of the following statements are true and which are not?

(a) rank K = dim Sp{({ll , , ({In} .

(b) rank K = dim Sp{1f!1 , , 1f!n}.
(c) rank K = minjdim sp [ej , , ({In), dim Sp{1f!1 , , 1f!n}].
(d) rank K = max [dim Sp{({ll , , ({In}, dim Sp{1f!1 , , 1f!n}].
(e) rank K ::: min [dim Sp{({ll , , ({In), dim sp{ 1f!1 , , 1f!n}].

18. Let H be a Hilbert space and let K I, K2 E £(1i) be two operators of finite
rank . Prove that rank (KI + K2) ::: rank KI + rankK2

19. Let 'H be a Hilbert space and let B, C. D E £(1i). On 1i(3) = 'H EB 'H. EB H
define A by the matrix
Exercises II 121

Prove
(a) A E £(1i(3»
(b) A 3=0

(c) I - (X A is invertible for any (X E C and I + (X A + (XZ A is its inverse.

20. LetIL = (ILk>k=1 be a given sequence ofcomplex numbers with sUPk IILkl <
1. Prove that the following two systems have a unique solution for any
right side in £z . Find the solution for 1Jk = Olk and ILk = Zk~l .
(a) ~k - ILk~k+I = 1Jk. k = 1,2, .
(b) ~k - ILk~k-1 = 1Jk. k = 2, 3, , ~I = 1JI·
21. Let IL = (ILk)~1 and v = (Vk)~1 be two sequences of complex numbers
with sUPk IILk I + sup, IVkl < 1. Prove that the following system of
equations has a unique solution in £z for any {1Jk} in £z.

~I + ILI~Z = 1JI

Vk-I~k-I + ~k + ILk~k+1 = 1Jk. k = 2, 3, ... .

22. Find the inverse of 1- K, where K is of the form

(a) K~ = (HI + b !-~I, k~I , ...), K: £z -"" ez.
(b) (Kf)(t) = [j~~~z f(t) cos t dt]~ cos t + [j~~~z f(t)t dt] sin t
K: Lz[-f, f] -"" Lz[-f , flo

i:
23. Let K : LZ[-JT, JT] -"" LZ[-JT, JT] be given by

(Krp)(t) = k(t - s)rp(s) ds.

Find the matrix of K with respect to the basis {einl}nE Z in each of the
following cases:
(a) k(t) = It I
(a) k(t) = sin t

24. Solve the integral equation

rp(t) = sin t + 1 1

rp(s) ds.

25. Let K : Lz[O, 00) -"" Lz[O , 00) be given by (K rp )(t) = k(t +s)rp(s)ds 10
00

where k(t) is continuous and

00
10
Ik(t)1 dt < 00 . Show that K is compact,
26. Find the adjoint of the operator K given in problem 22.
27. When is the operator D w given in problem 1 self adjoint?
122 Chapter II. Bounded Linear Operators on Hilbert Spaces

28. Which of the following operators are self adjoint and which are not?
(a) The operator K: L2[-rr,rr] ~ L2[-rr,rr] defined by (Kcp)(t) =
J~:rr ei(t-s)cp(s) ds.
(b) The operator K: L2[-rr,rr] ~ L2[-rr ,rr] defined by (Kcp)(t) =
J~:rr cos(t - s)cp(s) ds.
(c) The operator K2: L2[O, 1] ~ L2[O, 1] defined by (Kcp)(t) =
J~ cp(s) ds.

29. Let A be in £(H), where H is a Hilbert space . Define on the direct sum
H(2) = H E9 H (cf. exercise 7) the operator B by

B= (° °.
-iA*
iA)

Prove that II A II = II B II and that B is self adjoint.

30. Suppose A E £(H) and dim Im A = 1. When is A self adjoint?
31. Let X and tp be given vectors in a Hilbert space H . When does there exist a
self adjoint operator A in £(H) such that AX = cp? When is A of rank I?
32. Give a necessary and sufficient condition for the product oftwo selfadjoint
operators to be self adjoint.
33. Let A j k E £(H) , j, k = 1, . . . , n. Define on the direct sum H(n) =
?i E9 H ~ . . . E91f, an operator A by
n

Give necessary and sufficient conditions for A to be self adjoint.

34. Let A : £2 ~ £2 be given by the matrix

ao

A=
al
a2
ao
al ao °
a2 al

with L~o la j I < 00.

(a) Prove that A is bounded and that IIAII ::: L~o lajl.
(b) Find the matrix representation of A*.
Exercises II 123

35. Let w = ()I.I, A2, . . .) , where A j > 0, j = 1,2, ... be such that
SUPjAj < 00 and inf', A j > 0. Define £2(W) as in Exerci se 1-15. For
; = (;1, b ;3 , · · ·) E £2(W), define
S; (; I , ;2 , . . .) = (0, ~I , ;2 , . . .).
Prove that (S;) * is given by

A * ('71 , 712, . .. ) = (A2

(Sr) A3
-7/2 ,-713, · · ·) .
Al A2

36. Describe all selfadjoint operators of finite rank for which K I 980 = 0.
37. Define an operator U on £2 by U ; = (;n, ;n-l, . .. , ; 1, ;n+d . Prove that
(a) IIU;II = 11;11 for all s E h
(b) U- I = U = U* and U 2 = I.
(c) if a =/= ±1 , a E C , then I - all is invertible and (I - aU)-1
= 1~Q'2 (I + aU) .
(d) Give a matrix representation for U.

38. Define the operators S;r ) and Sir ) on L 2[0 , 1] by

(S(r ) )(t) = { rp(t - r )

° for
for
1~ t ~ r
os t < r
°
r ip

()
(S/ rp)(t) = {
L
rp (t
°+ r) for
for
_ 1- r
_< t <
1 - r < t ::s 1.
Prove that
(a) [S;r )] * = Si r )
(b) S? Il S;r2 ) = S; r l +Q )
(c) si i
rIls r2) = SY I+Q)

(d) S(r2)
e e
S (ril = !p Q Sc(rl- r2)
P S (r2- ril
>c:
lor rl _ rz

°::s
c.
r2 r lor r2 ~ rl
where PQ is the projection onto {rp E L 2[a , b] I rp(t) = 0, t ::s
r2}.
(e) li S;') II = IISy )II = 11 for r E [0, 1].
39. Let 'H be a Hilbert space . Prove that for any A I , A 2 E £ (1i),
(a) {~AI + ~A2 }-l = Ker At n Ker A ~
(b) ~ A I n ~A2 = (Ker At + Ker A ~ )-l.
40. Give a formula for the orthogonal proje ction Ponto sp {rpl , tp2 , rp3}, where
(a) rpl = (1, 0, 0, 1, 0, . . .),rp2 = (1,0,1 ,0, . . .) ,rp3 = (1, 1, 0, . ..)
are in £2.
(b) rpl (t) = cos t, tp2(t) = e', rp3(t) = t are in L 2[ - Jr , zr],
124 Chapter J[ Bounded Linear Operators on Hilbert Spaces

41. Let Lo be the subspace of all odd functions in L2[-rr, rr] and let L E be
the subspace of all even functions . Denote by Po and PE the orthogonal
projections onto Lo and L E, respectively. Give a formula for Po and PE.
42. Let Nl = {(~l, ~l, ~l , ~2, ~2 , b, ...)} ,

N2 = {(~l , 8~1 , 82~1 , ~2, 8~2 , 82~2, " .) },

N3 = {(~l, 82~1, 8~1, ~2, 82~2, 8b, . . .)},

be subspaces in £2, where 8 = -!

+ !J3i. Given an expression for each
of the orthogonal projectors Pj onto the subspace N], j = 1,2,3 .
43. Find the orthogonal projections onto the following subspaces of £2:
(a) sp{(1, 2, 0, ...), (0,1 ,2,0, . . .), (0,0,1,2,0, . . .) , .. .}
(b) sp(l, -5 ,6,0, . . .) , (0,1, -5, 6, 0, ...), ...}.

44. Find the orthogonal projection onto the intersection of the following pair
of subspaces in £2:

sp{(1, 2, 0, ), (0,1 ,2,0, ), },

sp{(1,4,0, ), (0,1 ,4,0, ), }.

45. Given Ajk E £(1-{), j , k = 1,2, define on 1-{(2) = 1-{ EB 1-{ an operator A
by
A= (All Al2)
Al2 .
A2l

Prove that A is compact if and only if each Ajk is compact.

46. Suppose A, B E £(1-{) and AB is compact. Which statements must be
true?
(a) Both A and B are compact.
(b) At least A or B is compact.

47. Which of the following statements about compact operators on a Hilbert

space are true?
(a) There exists a compact operator with a closed image.
(b) The image of any compact operator is closed.
(c) The image of any compact operator is not closed.
(d) There exists a compact operator with a nonclosed image.
(e) There exist a compact operator with a finite dimensional kernel.
(f) The kernel of any compact operator is finite dimensional.
Exercises II 125

48. Let (aj)~1 be a sequence of complex numbers with L~llajl < 00.
Define an operator on £2 by the matrix

C )
a2 a3
A = a: a3
a3

Prove that A is compact.

49. Let A E £(£2) be given by the matrix

A=

C
W2
W3 0)
0

i.e., A = D w as in exercise 1. Show that for any kl :s k2 :s ... :s k m in

N, the subspace SP{ekl' ek2 ' . . . , ek m } is A-invariant.
50. Let A be the operator in problem 49. Suppose Wi E JR, i = 1,2, .. . and
Wi =F Wj for i =F j . Prove that all the closed A-invariant subspaces of A
have the form described in problem 49.
51. Let A E £(H) be invertible and let M be an A-invariant subspace. Prove
or disprove the following statements.
(a) Ml.. is A-I invariant.
(b) M is A -I invariant if M is finite dimensional.
(c) M is A -I invariant. Hint: Define A on £2(Z) by Aej = e2j-l, 1 :s i,
Ae-2j = e2j , 0 :s i , Ae-2}+1 = e_j, 1 :s j .

52. Given A E £(H), let L be a closed A invariant subspace. Denote by PL the

orthogonal projection onto L. Prove or disprove the following statements.
(a) PLA = APL
(b) (/ - PdA = A(/ - Pt.)
(c) PLAPL = Ah
(d) PLAPL = PLA
(e) (/ - h)A(/ - Pd = A(/ - Pd
(f) (/ - PdA(/ - Pd = (/ - PdA .

53. What can one say about the A-invariance of L and Ll.. for each of the
properties (a)--(f) in problem 52?
54. Given A E £(H) , let M be a closed A-invariant subspace. Denote by AM
the restriction of A to M . Show that (AM)* = PMA*IM , where A*IM is
the restriction of A * to M and PM is the orthogonal projection onto M.
126 Chapter II. Bounded Linear Operators on Hilbert Spaces

55. Find an invertible operator A and an A-invariant subspace M such that

M.1 is not A-invariant but is A - I-invariant. Hint: exercise 51.
56. Let M be a closed subspace of a Hilbert space 'H. With respect to the
decomposition 1t = M ED M.1, let AI , A2 , A 3 E £ (1t ) be given by the
matrices

Al -
- (All0 o)
B22 ,
A =
3
(Cll
0

List all obvious invariant subspaces for these operators .

57. Let 1t1, 1t2 and 1t3 be mutually orthogonal subspaces ofa Hilbert space

lr
'H such that 'H = H I ED 1t2 ED 1t3. Let A , B , Cand D E £(1t ) be given by

ell
the matrices

A= 0
Al2
A22
All )
A23 , B = B21
0
B22 o),

ell
0 0 A33 B31 B32 B33

cD
CI2
c= 0 C22 D =
(D II
0 D22 o).
0 0 0 D32 D33
List all the obvious invariant subspaces for A, B , C and D .
58. List some A-invariant subspaces different from sp{ek, ek+I , . . .},
k = 1,2, . .. , where the operator A : £2 --+ £2 is given by the matrix

0
2 1
A =
2

2
0

59. List some A-invariant subspaces different from sp{el , .. . , ed, k = 1, 2, .. . ,

where the operator A: £2 --+ £2 is given by the matrix

60. Given A E £(1t ), suppose L is A-invariant. Prove that

(a) L is A -I -invariant if and only if the restriction of A to L is invertible
on L (assuming A is invertible).
(b) The closed subspace L is A *-invariant ifandon1y if L.1 is A-invariant.
Exercises II 127

61. Let H be a Hilbert space. Construct a bounded linear operator A on H

such that
(a) for a given pair of vectors x and yin H we have Ax = y.
(b) for two given systems of vectors XI, . •. , X n and YI , • .• , Yn in H we
have

62. Let M and N be closed subspaces of a separable Hilbert space H . Under

what conditions on M and N does there exist an operator A E £(H) such
that AM = N?
63. Do the two previous exercises under the additional constraint that A is self
adjoint.
64. Do Exercises 61 and 62 under the additional constraint that A is invertible .
65. Do Exercises 61 and 62 under the additional constraint that A is right (left)
invertible but not two-sided invertible .
66. Let N be a given closed subspace of a separable Hilbert space H . When
does there exist a right (left) invertible operator A E £(H) such that its
kernel (image) is N?
67. Let M and N be closed subspaces of a separable Hilbert space H. Does
there exist an operator A E £(H) such that

M = Irn A , N = Ker A?

When can A be chosen to be a projection?

68. Let M and N be as in the previous exercise. Does there exist an operator
A E £(H) such that

M = 1mA* , N = Ker A*?

69. For i = 1,2 let M; and N ; be closed subspaces of the separable Hilbert
space H ; such that H ; = M ; E9 N; . Find (if possible) an operator A E
£(H I, Hz) with the property that

NI = KerA , N2 = Ker A* ,
MI = 1mA*, Mz = ImA .

70. Let n be a positive integer. If for A E £(H) the operator I - An is one-

sided invertible, then I - A is invertible from the same side. Prove or
disprove this statement.
128 Chapter II Bounded Linear Operators on Hilbert Spaces

71. Let 'H = 'HI ffi 'H2, where 'HI and 'H2 are closed subspaces of the
Hilbert space 'H, and let T E £('H) have the following operator matrix
representation

T = (~ ~)
relative to the given decomposition of 'H. Assume T is invertible. Deter-
mine which of the following statements are true :
(a) A is invertible,
(b) A is left invertible,
(c) A is right invertible,
(d) B is invertible,
(e) B is right invertible,
(f) B is left invertible.

72. As in the previous exercise, let'H = 'HI ffi 'H2 and T E £('H) be given by

Let A and B both be left (right) invertible. Is the operator T left (right)
invertible?
73. Do Exercise 71 with the operator matrix representation of T being replaced
by

74. Do Exercise 72 with the operator matrix representation ofT being replaced
by

75. (a) Given an example ofa product oftwo orthogonal projections that is not
an orthogonal projection.
(b) When is the product oftwo orthogonal projections again an orthogonal
projection?

76. Let ta = (1, a, a 2 , . ..), where 0 < a E ~, and consider the corresponding
Hilbert spaces l2 (cv); see Exercise 1-15. Define Sf to be the forward shift
on l2(CV) and Sb the backward shift.
(a) Prove that Sf and Sb are bounded linear operators on l2(CV) and com-
pute their norms.
(b) Find the spectra of Sf and Sb.
Exercises II 129

77. Let W be as in the previous exercise, and fix 0 i= q E C . Let Sf ,q and Sb,q
be the weighted shift operators on £z(w) given by

Sf ,q(xo , XI , Xz, ) = (0, qx«, qZxI , ),

Sb,q(XO, XI , xz, ) = (XI , qxi , qZx3, ).

Solve the problems (a) and (b) in the previous exercise with Sf ,q and Sb,q
in place of Sf and Sb, respectively.
78 . Do Exercise 76 with the sequence W = (1 , a , a Z , • • •) being replaced by
W = (wo, WI , Wz .. .), where Wj > 0 for j = 0, 1,2, .. ..
79 . Do Exercise 77withw = (wo, WI, Wz, )andwj > ofor j = 0,1,2, . . . .
80. Do Exercise 77 with W = (wo, WI, WZ, ) as in the previous exercise and
with Sf,Q and Si,Q being given by

Sf,Q(xO, XI , xz, ) = (0, qIXO, qZXI, . . .) ,

Sb,Q(XO, XI , xz, ) = (qOxI, qIXZ, qZx3, ·· .) .

Here 0 i= qj E C for j = 0,1 ,2, ....

81. Let Sj ,Q and Sb,Q be the operators on £z(w) defined in the previous exercise.
Find the matrix representations of these operators relative to the standard
basis of £z (r»).
82. Let W = (wo, WI, W2, . ..) with W j > 0 for each j , and let U be the operator
from £z(w) into £z defined by

U(xo , XI , XZ , . . .) = (.JWOxo, ..jW\XI, -JWi,xz , .. .).

(a) Show that U is invertible, and determine U- I and U* .

(b) Let Sj ,Q and Sb,Q be the operators on £z(w) defined in Exercise 80.
What kind of operators are the operators U Sj ,QU - 1 and U Sb,QU- I ?

83. Let'H = 'HI ®'Hz ff/H3 be the direct sum ofthe Hilbert spaces 'HI , 'Hz and
'H3 (cf., Exercise 5). Relative to this decompostion of'H let the operator
A E £('H) be given by the lower triangular operator matrix

What can we conclude about the operators A II , Azz, A33 when the operator
A is invert ible? Illustrate the answers with examples.
84. Do the previous exercise for the case when the operator A is given by

All
A= 0
(
o
130 Chapter II Bounded Linear Operators on Hilbert Spaces

85. Fix t > O. Let W be the operator on L2(0 , 00) defined by

f (x - t) for x > t,
(Wf)(x) =
Io for 0 < x :::: t.
,

(a) Show that W is an isometry and W is not unitary.

(b) Prove that its adjoint is given by

(W* f)(x) = f(x + t) , x > O.

(c) Show that the spectrum of W is equal to the closed unit disc

86. Fix t > O. Let V be the operator on L2(-00 , 00) defined by

(Vf)(x) = f(x + t) , x E lR

(a) Prove that V is unitary and determine V-I .

(b) Show that a(V) = 11'.

87. Let A E £(£2), and let Pn be the orthogonal projection onto span {eo, .. . ,
en }, where eo, el , . .. is the standard basis of £2. Show that the operator
A is a band operator if and only if for some nonnegative integer m the
following equalities hold
(i) PkA(I - Pm+k+d = 0, k = 0, 1,2, ,
(ii) (I - Pm+k+dAPk = 0, k = 0,1,2, .

Give an interpretation of these equalities in terms of the spaces Im Pn and

Ker Pn , n = 0, 1,2, . .. .
88. Let Lo be the linear span of the vectors eo, el , . .. in the standard basis of
£2. If A on £2 is a band operator, show that

ALo C Lo, A* Lo C Lo .

Is the converse implication also true?

89. LetA be an operator on £2 satisfying (*) in the previous exercise. Describe
the form of the matrix of A relative to the standard basis of £2.
90. Let Lo be as in Exercise 88, and let A E £(£2) . Describe in terms of the
matrix of A relative to the standard basis of £2 the property that

ALo C u;
91. Do the previous exercise with A * in place of A .
92. Is the inverse of a lower triangular invertible band operator always given
by a lower triangular matrix?
Exercises II 131

)
93. Let T be the operator on £z given by the matrix

...
.. .

where a and b are complex number with lal < 1 and Ibl < l.
(a) Prove that indeed T is a bounded operator and find an evaluation of
the norm of T.
(b) Use the finite section method to find the conditions of invertibility of
T and to invert T .

94. Do the previous exercise with £z(Z) in place of £z and with

I a aZ a 3 a4
b 1 a aZ a3
T= bZ b IT] a aZ
b 3 bZ b I a
b4 b 3 bZ b

95. Consider on £z(Z) the operator T given by the two-diagonal lower trian-
gular matrix

bOO
T= a [B 0
o a b

Use the finite section method to find that T is invertible and to determine
its inverse for the following cases:
(i) a = 1 and b = -4;
(ii) a = -4 and b = 1;
(iii) a , bEe and
lal < Ibl;
(iv) a, bEe and Ibl < [zz] .

96. Solve the problems from the previous exercise in an alternative way by
writing
T=bI+aV .
where V is the forward shift on £z(Z) .
132 Chapter II. Bounded Linear Operators on Hilbert Spaces

97. Do Exercise 95 with the matrix of T replaced by its transpose .

98. Use a modified finite section method to invert the operator T on £20'~)
given by the following tridiagonal matrix

7 -3 0
T = -2 [2J -3
o -2 7

99. Use the finite section method to find the conditions of invertibility of
T E £2 and to construct its inverse when T is given by the following
tridiagonal matrix

(1 +dal' ii
1 + lal 2
0
ii
0
0 . ..
T= 0 a 1 + lal 2 ii OJ
0 0 a 1 + lal 2

100. Do the previous exercise with £2 replaced by £2 (Z) and with the matrix
replaced by the corresponding doubly infinite tridiagonal analog .
101. Let A E £(1i\ , 1i2) and B E £(1i2, 1iJ). Assume that the product AB is
invertible . Which of the following statements is true?
(a) A is invertible .
(b) B is invertible.
(c) A and B are both invertible.
(d) A right invertible and B is left invertible .
(e) A is left invertible and B is right invertible .

102. If AB is invertible, does it follow that BA is invertible?

103. Find the spectrum a(A) and the inverse (A./ - A)-l for A in the resolvent
set of A, when A E £(1i) is one of the following operators:
(a) Ax = (x , cp)lfr where (cp , lfr) = 0;
(b) Ax = LJ=1 (x, CPj)lfrj where (cpj, lfrk) = 0 for i .k = 1, . . . , n ;
(c) A is given by
1
L
00

Ax = 2j (x, CPj)lfrj,
j=l

where (cpj, lfrk) = 0 and IIcpjll = IIlfrkli = 1 for i .k = 1,2,3 , . . . .

2.20 The Spectrum ofan Operator 133

104. Find the connection between the spectra of the operators A 1 and Az when

(a) AI = ( ~z ~I), Az = BI Bz;

(b) A = (~ ~z ~I ) , Az = B I BZB3.
B3 0 0
(c) What can you say additionally if the operators B j are compact?

105. Find X E £ (1i) such that X is a solution of the operator equation

X-AXB=Y,

where A and B are bounded linear operators on H of norm strictly less

than one .
106. Let A and B be as in the previous exercise, and let F be the linear operator
on £ (1i) defined by
F (X ) = X - AXB.

(a) Prove that F is bounded and give an evaluation of its norm.

(b) Show that F is invertible and determine its inverse.
(c) Determine the spectrum of F.
Chapter III
Laurent and Toeplitz Operators

This chapter deals with operators on £2 (Z) and £2 with the property that the matrix
relative to the standard basis in these spaces has a special structure, namely the
elements on diagonals parallel to the main diagonal are the same, i.e., the matrix
entries ajk depend on the difference j - k only. On £2 (Z) these operators are
called Laurent operators (and in that case the matrix is doubly infinite); on £2 they
are called Toeplitz operators. These operators form important classes of operators
and they appear in many applications. They also have remarkable properties.
For instance, there are different methods to invert explicitly these operators, and to
compute their spectra . This chapter reviews these results starting from the simplest
class.

3.1 Laurent Operators

A Laurent operator A is a bounded linear operator on £2 (Z) with the property that
the matrix of A with respect to the standard orthonormal basis {ej }~-oo of £2 (Z)
is of the form

ao a -I a-2
al ~ a-I (Ll)

a2 al ao

Here ~ denotes the entry ao located in the zero row zero column position. In
other words, a bounded linear operator A on £2 (Z) is a Laurent operator if and
only if (Aek, ej ) depends on the difference j - k only.

Proposition 1.1 A bounded linear operator A on £2(Z) is a Laurent operator if

and only if A commutes with the bilateral shift on £2(Z),

Proof: Let {ej }{~-oo be the standard orthonormal basis of £2 (Z), and put ajk =
(Aek, ej) . Reca I that the bilateral shift V on £2(Z) is given by

V (.. . , ~-I , lliJ,~I , . . .) = (. . . , ~-2 , [G], ~o , .. .).

136 Chapter III. Laurent and Toeplitz Operators

Thus Vej = ej+1 for each j E Z. Using V* = V-I, we have

Also, (AVek, ej) = (Aek+ I, ej) = a j .k+ I . Next, observe that VA = AV if and
only if

Thus A and V commute if and only if aj-I ,k = a j ,k+1 for each j , k E Z, that is,
if and only if A is a Laurent operator. 0
In a somewhat different form we have already met Laurent operators in
Example 2 of Section 11.4. Indeed, let a be a bounded complex valued Lebesgue
measurable function on [-lr, zr], and let M be the corresponding operator ofmul-
tiplication by a on L2 ([-lr, lr]), that is,

(Mf)(t) = a(t)f(t), f E L2([-lr, lr]).

Example 4 in Section 11.2 shows that M is a bounded linear operator on L2 ([-lr, lr])
and the matrix of M with respect to the orthonormal basis ({in (t) = Jkeint
,
n = 0, ±l, ±2, ... , is given by (1.1), where

an = - -
1 1]( .
a(t)e- tn t dt, n E Z. (1.2)
.J2ii -](

It follows that the operator A = F M F- I , where F is the Fourier transform (see

the last paragraph of Section 1.18) on L2 ([-lr, n ]), is a bounded linear operator
on f2(Z) and the matrix of A with respect standard basis of f2(Z) is also given
by (1.1). Thus A is a Laurent operator. In this case we say that A is the Laurent
operator defined by the function a, and we refer to a as the defining function of
A . Often the defining function a will be of the form aCt) = w(e it ) , where w is
defined on the unit circle. In that case we refer to w as the symbol of A.
The procedure described in the previous section yields all Laurent operators
on f2(Z) . In other words, given a Laurent operator A on f2(Z), we then find a
bounded complex valued Lebesgue measurable function on [-lr, lr] such that A
is the Laurent operator defined by a. More precisely, if A is the Laurent operator
given by the infinite matrix (1.1), then one can find a bounded complex valued
Lebesgue measurable function a on [-lr, zr] such that (1.2) holds. For this result,
which we will not need in this chapter, we refer the reader to Section XXIIl.2 of
[GGK2].
3.1 Laurent Operators 137

Let A be the Laurent operator on £2(1£) defined by the continuous function a ,

and let M be the corresponding operator of multiplication on L2([ - 7(, 7(]) . Since
the Fourier transform pre serves the norm, II A II = II M il , and we can use Example
4 in Section 11.2 to show that

II AII = max la(t) l· (1.3 )

-If ::99r

The relation A = F M F - I also yie lds the following invers ion theorem.

Theorem 1.2 Let A be the Laurent operator defined by the continuous fun ction
a. Then A is invertible if and only if a(t ) i= 0 for each -7( S t S 7(, and in this
case, A - I is the Laurent operator defined by the fun ction I I a, that is

bo b_1 b-2
bl bo b_1
b2 bl bo

where
bn = - -
I j If I .
__ e- mt dt , n E Z.
,J2ii - If a Ct )
Proof: Let M be the ope rator of mult iplication by a on L 2([ -7( , rr]) . From
Example 2 in Section 11.7 we kno w that M is invertible if and only if a Ct ) i= 0 for
- 7( S t S 7(] , and in that case M - I is the operator of multiplicat ion by b = Y]«.
Since A = F M F - I , the operator A is invertible if and only if M is invertible ,
and then A -I = F M - 1F- 1. By combining these results the theorem follows . 0

Corollary 1.3 The spectrum ofthe Laurent operator A defin ed by the continuous
f unctions consists ofall p oints on the curve param etrized by a, that is

a (A ) = {aCt ) I- 7( S t S zr}, (1.4)

Proof: Take A E C. Notice that AI - A is again a Laurent operator. In fact,

AI - A is the Laurent operator defined by the continuous function t t-+ A - aCt).
But then we can use the prev ious theorem to show that AI - A is not invertible if
and only if A - aCt) = 0 for some t E [- rr,7(]. It follows that A E a (A ) if and
only if A = a Ct ) for som e t E [-7(, 7(] , which proves (1.4 ). 0

The set A of all Laurent operators with a cont inuous defining function is closed
under addition and multiplication . Moreover multiplication in A is comm utative.
The se facts follow immediately from the fact that F AF- 1 consists ofall operators
ofmultiplication by a cont inuous function. Thus A has the same algebraic structure
as the ring of all complex valued continuous funct ions on [-7(, 7(].
138 Chapter III. Laurent and Toeplitz Operators

Examples: 1. Let A be the Laurent operator on f2(Z) given by the following

tridiagonal matrix representation:

7/5 -3/5 0 0 0
-2/5 7/5 -3/5 0 0
A= 0 -2/5 17/51 -3/5 0
o 0 -2/5 7/5 -3/5
o 0 0 -2/5 7/5

The function
2 it 7 3_ it
aCt) = -Se + S- Se

is the defining function of A. It follows that aCt) = w(e if ) , where

W(A) = -~A + 2- ~A -I
5 5 5
= ~ (~ -
5 A
2) (A - 3) .

Herea(t) =1= 0 foreacht , and we can apply Theorem 1.2 to show that A is invertible.
To compute A -I, notice that
1 lA-I
2 1 + --1- ' IAI = 1,
W(A) I - 2A-I 1 - 3A
and hence
bet) = - 1 =
aCt)
Loo ----,.e-
1
2J
lJ
"f
+L
oo I
3J
"•
----,.e lJf
)= 1 )=0

Thus the inverse operator A -I has the form

1 1/2 1/22 1/23 1/24

1/3 I 1/2 1/22 1/23
A- 1 = 2
1/3 1/3
3 2
OJ
1/2 1/22
1/3 1/3 1/3 1 1/2
1/34 1/33 1/32 1/3

Since ei f = cos t + i sin t, we see that the defining function a is also given by
7
aCt) = ( S- cost
) 1 smt
+ Si . . (1.5)
3.1 Laurent Operators 139

Thus cos t = ~ - ffla(t) and sin t = 5;Ja(t) . Using Corollary 1.3 we see that the
spectrum of A is precisely given by the ellipse

2. Next we consider the Laurent operator

1 1/2 1/2 2
A= 1
2" OJ 1/2
2
1/2 1/2

The defining function is given by

+L
OO 1 .. oo 1 ..
a(t) =
L
j=1
----ce-ljf
2J
j=O
----celjf .
2J

Thus a(t) = w(e it ) with co being given by

W(A) =
1 -I
2"A ( 1 - i1 ·-I
)-1 + ( 1-
1
2"A )-1 =
(53 - 3A
2 -I -
)-1
2A
3

Since w (A) =I- 0 for IA 1 = 1, we see that a (t) =I- 0 for each t , and hence the operator
A is invertible by Theorem 1.2, and the defining function b of A -I is given by

b(t) = -
1 2 '
= __ e- 1f
z ;.
+ -5 - _e '
a(t) 3 3 3

We conclude that A -I has the following tridiagonal matrix representation:

5/3 -2/3 0 0 0
-2/3 5/3 -2/3 0 0
A- 1 = 0 -2/3 15/31 -2/3 0
0 0 -2/3 5/3 -2/3
0 0 0 -2/3 5/3
140 Chapter III Laurent and Toeplitz Operators

Notice that fRb(t) = -1cost + i and ':3b(t ) = O. Thus, by Corollary 1.3,

a(A- I ) = {r E fR 11/3 ~ r ~ 3}.

Since o'{A'<" ) = pl.-I I).. E a (A )} , we obtaina(A ) = {r E lR 1 1/ 3 ~ r ~ 3}.

The fact that a(A) is real also follows from the fact that A is selfadjoint.
3. The final example is the tridiagonal Laurent operator

8/9 1/3 1/3 0 0

-1/3 8/9 1/3 0 0
10
A=- 0 -1/3 18/91 1/3 0
9
0 0 -1/3 8/9 1/3
0 0 0-1/3 8/9

of which the defining function a is given by a(t) = w(e it ) , where

+ -8 + -)..
w ().. ) = -10 (--)..
1
9393
1 -I) = -10 ( 1 + <);
9
1
3
-I) (1 - 1) .
<):
3

We see that a (t) i- 0 for each t , and

it
b(t ) = - 1 = 1 + 13 e = 00 (
__1) j ..
+ 00 (1)
_ j ..
~ ~
e -Ijt eljt.
a(t ) 1+ l e- it 1 - le it 3 3
3 3 1=0 1= 1

By Theorem 1.2 the operator A is invertible, and its inverse is given by

1 -1/3 1/3 2 -1/3 3 1/3 4

1/3 1 -1/3 1/3 2 -1/3 3
A-I = 1/3 2 1/3 IT]
-1/3 1/3 2
3 2 1 -1/3
1/3 1/3 1/3

1/3 4 1/3 3 1/3 2 1/3

Notice that fRa(t ) = ~~ and ':3a (t) = - ~~ sint. Thus by Corollary 1.3, we have

a (A ) = { z = -80
81
+ ir I- -20 < r < -20} .
27 - - 27
3.2 Toeplitz Operators 141

3.2 Toeplitz Operators

Let T be a bounded linear operator on £2. In the sequel the formula

a ll a I2 an ]
a21 a22 a23 .
T = a31 a32 a33 (2.1)
[ .. .. ..
. . .
means that the matrix in the right hand side of (2.1) is the matrix corresponding to
T and the standard orthonormal basis e l , e2, .. . in £2. We say that T is a Toeplitz
operator if

(2.2)

In this case we refer to the right hand side of (2.2) as the standard matrix repre-
sentation ofT. Recall that ej = (0, . . . , 0, 1, 0, . . .), where the one appears in the
j -th entry. Thus T is a Toeplitz operator if and only if (Tek , ej ) depends only on
the difference j - k. This remark yields the follow ing theorem.

Theorem 2.1 An operato r T on £2 is a Toeplitz operator ifand only ifT = S*TS,

where S is the f orward shift on £2.

Proof: Assume T is given by (2.1). By definition , Sen = en+ I for n = 1, 2 . ...

Thus

Therefore, T = S*TS if and only if akj = ak+I.j +I for all j , k = 1, 2, . ... Thus
T = S*TS if and only if T is Toeplitz. 0
With each continuous complex valued function on the unit circle T in C we
can associate a Toeplitz operator on £2. Indeed, let w be such a function, and put
a(t ) = w (e i t ) . Notice that a is continuous on [-Jr, Jr] and a (-Jr ) = a(Jr) . Now
let A be the Laurent operator on £2(Z) defined by a, i.e.,

ao a-I a-2
A= al [§]
a- I (2.3)
142 Chapter III. Laurent and Toeplitz Operators

where
an = - -
1 j:rr w(e1t)e-
. . ln t
dt, n E Z. (2.4)
.j2ii -:rr
In the sequel we identify the space f.2 with the closed subspce of f.2 (Z) consisting
of all sequences (~j) jeZ such that ~j = 0 for j :s -1. Put
(2.5)

where Pez is the orthogonal projection of f.2(Z) onto f.2. Then T is a bounded
linear operator on f.2, and from (2.3) we see that (2.2) holds with an given by (2.4).
Thus T is a Toeplitz operator. We shall refer to T as the Toeplitz operator with
continuous symbol co.

Theorem 2.2 The norm of a Toeplitz operator T with continuous symbol co is

given by
IITII = max Iw().)I ·
1).1=1
Proof: We continue to use the notation introduced in the paragraph preceding the
present theorem. From (2.5) and formula (1.3) in the previous section we see that

1IT11 :s IIAII = max

-:rr9~:rr
la(t)1 = max Iw().)I .
1).1=1
So it suffices to prove that II T II ~ II A II . To do this we first make some preparations.
For N = 0, 1,2, . .. , let 1iN be the closed subspace of f.2(Z) consisting of
all sequences (~j) j eZ such that ~ j = 0 for j :s - N - 1. Thus 1io = f.2 and
1io C 1il C 1i2 C . " . Denote by PH.N the orthogonal projection of f.2(Z) onto
1iN . Notice that for each x = (~j)jez in f.2(Z) we have
-N-l
IIx - P1-{NxIl2 = L l~jl2 --+ 0 (N --+ 00) . (2.6)
j=-oo

Also, if ... ,e-l,eo,el, ' " is the standard basis of f.2(Z), then e-tc.e-u-i«,
e-N+2,'" is an orthonormal basis of 1iN, and from (2.5) if follows that with
respect this basis the matrix of P1-{NAI1iN is given by

:::]
[
:~ aa~1 :=~
a2 al ao .
··· ..
.
It follows that II PH.N A I1iN II = II T II for each N.
Now, fix e > 0, and choose x E f.2(Z) such that [x] = I and II Ax II > IIAII- e.
By (2.6), we have P1-{NX --+ x and AP1-{NX --+ Ax for N --+ 00. Thus we can find
a positive integer N' such that x' = P1-{N'X has the following properties:

o t= IIx'lI < 1 + e, IIAx'lI > IIAII- e. (2.7)

3.3 Band Toeplitz Operators 143

According to (2.6), we also have P?tN Ax' ~ Ax' for N ~ 00. Thus we can
choose N il ::: N ' such that II P?t N" Ax' II > II A 11- 8. Since x ' E 1tN' and N il ::: N ',
we have x' E 1tN", and thus

P IIAII- 8
li P?t N" AI1t N "II >
-
II ?t N" Ax ' ll >
IIx' ll 1+ 8

Therefore , II T II > (1 + 8) - 1( II A II - 8). This holds for each 8 > O. Hence

1IT11 ::: IIAII· o
The construction given in the paragraph preceding Theorem 2.2 can be carried
out for larger classes of functions than continuous ones. Also Theorem 2.2 holds
in a more general setting. See, for instance, Section XXIIl.3 in [GGK2].

3.3 Band Toeplitz Operators

In this section we study the invertibility ofa Toeplitz operator that has the additional
property that in its standard matrix representation all diagonals are zero with the
exception of a finite number. In that case all the non-zero entries in the standard
matrix representation are located in a band, and for that reason we refer to such
an operator as a band Toeplitz operator (cf., the last part of Section II.I6). Notice
that a band Toeplitz operator has a continuous symbol of the form
q
W(A) = L Akak. (3.1)
k= -p

Thus the symbol of a band Toeplitz operator is a trigonometric polynomial.

The function W in (3.1) can be represented in the form

W(A) = c); -r CA - ad . . . (A - a e)(A - f3[) . •. (A - f3m ), (3.2)

where c is a non-zero constant, r is a nonnegative integer, ai , . .. ,ae are complex

numbers inside the open unit disk, and f31, . . . , f3m are complex numbers on the
unit circle or outside the closed unit disk. The representation (3.2) of W is not
unique, but the number K = l - r is. In fact, K is just equal to the number of zeros
(multiplicities taken into account) of W in the open unit disk minus the order of the
pole of W at zero. If W(A) =1= 0 for each IAI = 1, i.e., if lf3 j I > 1 for each j, then K
is equal to the number of times the oriented curve t t-* w(e it ) circles around zero
when t runs from -rr to tt , and in that case we call K the winding number of W
relative to zero. In other words

I it tt
K = 2rr [arg wee )]t=-rr '

We shall prove the following theorems .

144 Chapter III Laurent and Toeplitz Operators

Theorem 3.1 Let T be the band Toeplitz operator with symbol w given by (3.2).
Then T is two-sided invertible if and only if w(e it ) f= 0 for -Jr :s t :s Jr and
K = I - r = O. In that case

1 m
c J=I (S-,BJ'I)-Irr 1=
T-1=-rr ~ 1 (I-a 1'S*)-1 , (3.3)

where S is the forward shift on £2.

Theorem 3.2 Let T be the band Toeplitz operator with symbol w given by (3.1),
and put K = £ - r. Then T is left or rightinvertible if and only if w(e it ) f= Ofor
-Jr :s t :s Jr . Furthermore, if this condition is satisfied, we have

(i) T is left invertible if and only if K 2: 0, and in that case codim 1m T = K,

and a left inverse ofT is given by

T(-I)=~rr~
c J=I (S-,BJ·I)-\s*)Krr~1=1 (I-a ·S*)-I . 1 ,
(3.4)

(ii) T is right invertible if and only if K :s 0, and in that case dim Ker T = -K
and a right inverse ofT is given by
1
m i
T(-I)=-rr
c J=I (S-,B J'I)-IS-Krr 1=1 (I-a 1'S*)-1 • (3.5)

Here S is the forward shift on £2.

The proofs of Theorems 3.1 and 3.2 are considerably more involved than the
inversion theorem for Laurent operators. There are two reasons for this . The
first is the fact that in general, in contrast to the Laurent operators in Section 3.1,
Toeplitz operators do not commute. Secondly, also unlike Laurent operators, the
product of two Toeplitz operators does not have to be a Toeplitz operator. These
phenomena can already be illustrated in the forward shift S and the backward shift
S*. Both Sand S* are Toeplitz operators, S*S is the identity operator on £2, and
o0 0 0
o 1 0 0
SS* = 0 0 1 0
000 1

Thus SS* f= S*S (and thus the Toeplitz operators S and S* do not commute), and
the product SS* is not Toeplitz (but a projection).
Notice that for n 2: 0 and m 2: 0 we have
s *)m- n m > n
(S*)mS n = {( ' -, (3.6)
s":", n 2: m
These identities will turn out to be very useful in the proofs.
In the proofs of Theorems 3.1 and 3.2 we will also need the following lemma.
3.3 Band Toeplitz Operators 145

Lemma 3.3 For Jal < 1 and 1,81 > 1 the operators 1 - a S" and S - ,81 are
invertible, and the respective inverses are given by

1 a a2 a3
o 1 a a2
o0 1 a
o0 0

1 0 o 0
,8-1 1 o 0
(S - ,81)-1 = -,8 ,8-2 ,8-1 1 0 (3.7)
,8-3 ,8-2 ,8-1 1

For lal = 1,81 = 1 the operators 1 - a S" and S - ,81 are neither left nor right
invertible.

Proof: For laJ < 1 we have lIaS*1I < 1, and hence, by Theorem II.8.!, the
operator 1 - a S* is invertible and its inverse is given by the first part of(3.7). Since
S -,81 = -,8(/ - ,8-1 S), a similar argument shows that S -,81 is invertible, and
that its inverse is given by the second part of(3 .7). From the results in Section II.20
we know that the spectra of Sand S* are both equal to the closed unit disc. Thus
the perturbation results Theorems II.15.3 and 11.15.5 imply the operators 1 - a S"
and S - ,81 are neither left nor right invertible when lal = 1,81 = 1. 0

Now, let R be a band Toeplitz operator, and let its symbol to be given by (3.1).
Since we allow a: p and aq to be zero, we may assume that p and q are positive
integers, and in that case (3.1) means that

R = a_pS*P + ...+ a-1S* + aol + alS + ... + aqSq,

where S is the forward shift on £2. The fact that S*S = 1 implies that for each pair
of complex a and b the operator R(a + bS) is again a band Toeplitz operator. In
fact, the symbol of R(a + bS) is given by w(A)(a + Ab). Similarly, for each pair
of complex numbers c and d the operator (c + dS*)R is a Toeplitz operator and
its symbol is the function (c + d): - 1)W(A). As we remarked above (see (3.6)) in
these product formulas the order of the factors is important. In general, (a + bS)R
and R(c + dS*) are not Toeplitz operators. We are now ready to prove Theorems
3.1 and 3.2.
Proof of Theorem 3.2 We split the proof in two parts. In the first part we assume
that w(e it ) =1= 0 for -7f :::: t :::: n and prove statements (i) and (ii) . In the second
part we prove that w(e it ) =1= 0 for -7f :::: t :::: n is necessary for left or right
invertibility.
146 Chapter III. Laurent and Toeplitz Operators

Part 1. Assume w(e it ) i= 0 for -T{ ::s t ::s T{ . Here co is given by (3.2). It
follows that in (3.2) the numbers f31, . .. , f3m are outside the closed unit disk, i.e.,
1f3 j I > 1 for j = 1, .. . , m.
Assume K = f - r ::: O. Then from the results proved in the paragraph preceding
the present proof we see that

(3.8)

Indeed, s':' is the band Toeplitz operator with symbol A':', and thus by repeatedly
applying the above mentioned multiplication rules, we see that the right hand side
of(3 .8) is the band Toeplitz operator with symbol

W(A) = cn7=1 ( l - cq): - 1)Al- r n f=1(A - f3i) .

Since W = co, it follows that (3.8) holds.
The fact that lad < 1 and lf3jl > 1 for each i and j implies (see Lemma 3.3)
that the factors I - a, S* and S - f3 j I appearing in the right hand side of (3.8) are
invertible operators, and hence the operators

(3.9)

are invertible operators. Also , the operator T(-I) in (3.4) is well-defined, and ,
using (S*)KSK = I, we see that T(-I)T = I . Thus T is left invertible and T(-I)
is a left inverse ofT. From (3.8) and the invertibility of the operators in (3.9) we
obtain that
codim Im T = codim Im SK = K ,
which completes the proof of statement (i).
Next, assume K ::s 0, and hence f ::s r . Then, by repeated application of the
product rules proved in the paragraph preceding the present proof, we have

(3.10)

Since the operators I - a, S* and S - f3 j I are invertible for each i and j, the
operator T(-I ) in (3.5) is well-defined. From (S*)-K S-K = I we conclude that
TT(-I) = I, and hence T(-I) is a right inverse of T. Furthermore, using the
invertibility of the operators in (3.9) we see that

codim Ker T = dimKer (S*)-K = -K,

which completes the proof of (ii).

Part 2. In this part T is left or right invertible. We want to show that w(e it ) i= 0
for -T{ ::s t ::s T{ . The proof is by contradiction. So we assume that w has a zero
on the unit circle IAI = 1, that is, in the representation (3.2) we have lf3kl = 1 for
at least one k. Choose such a f3k and put
W(A) W(A)
WI(A)=-- , W2(A) = A - - .
A - f3k A - f3k
3.3 Band Toeplitz Operators 147

Then WI and wz are functions of the same type as co. Let TI and T2 be the band
Toeplitz operators with symbol s WI and W2 , respe ctively. Since
w(}.) = WI ( }.)(). - fh ), w (}. ) = (1 - fh). - I )WZ (}.) ,

the product rules from the paragraph preceding the present proo f yield

T = T I (S - fh I) , T = (/ - fhS*)T2. (3.11)
By Lemma 3.3 the operator S - Ih I is not left invertible, because Ifh I = I. Hence
the first equality in (3.11) shows that T cannot be left invertible. It follows that T
must be right invertible, but then the second equality in (3.11) implies that I - Ih S*
is right invertible. Thus I - 71k S is left invertible. However the latter is impossible,
again by Lemma 3.3 and the fact that I13k I = 1.
Proof of Theorem 3.1 By Theorem 3.2 (i) and (ii) the operator T is both left and
right invertible if and only if w (ei !) =1= 0 for -rn S t S tt and K = £ - r = O.
Moreover in that case, formula (3.4) with K = 0 yields the formula for the inverse
in (3.3). 0
Corollary 3.4 Let T be the band Toeplitz operator with symbol ta . Then the
spectrum a (T) of T is given by

a (T ) = C\Q , (3.12)
where Q is the set ofall points }. in C such that}. =1= w(e it ) fo r all - Jr S t S tt
and the winding number ofco (.) - }.relative to zero is equal to zero. In particular,
the spectral radius of T is equal to the norm ofT.
Proof: Notice that for each }.in C the operator T -).J is the band Toeplitz operator
with symbol w(·) <): Thus Theorem 3.1 shows that}. E a CT ) ifandonly if}. E Q ,
which proves (3.12).
Since w(e it ) E a CT ) for each -Jr S t .:::: n , the spectral radius r eT ) =
sup{I}.1 I }. E a (T)} is larger than or equal to max-JT ~t ~JT Iw (eit)l. On the other
hand, the latter quantity is equal to II TIl , by Theorem 2.2. Since 1IT11 ~ r eT) , we
conclude that r eT) = II T11 . 0
Example : In this example we consider the Toeplitz operator T on £2 defined by
the tridiagonal matrix :

7/5 - 3/ 5 0
-2/5 7/5 -3/5
. .. J
T = 0 - 2/ 5 7/5 .
[
··· ..
.
The corresponding Laurent operator has been considered in Example I of the first
section . The operator T is bounded, and its symbol co is given by

w(}.)
2 +S
= - S}. 7 - S}.
3 - I = -S}.
2 - I ( }. - "2I) (). - 3) . (3.13)
148 Chapter III. Laurent and Toeplitz Operators

The aim is to study the invertibility properties of zI - T for each z in <C, and to
compute an inverse (one-sided or two-sided) whenever it exists. We split the text
into five parts.
Part 1. Notice thatw(A) i= 0 for each IAI = 1. Moreover, for this W the numbers
land r in (3.2) are given by l = 1 and r = 1. Thus K = O. By Theorem 3.1 the
operator T is invertible, and

(3.14)

By formula (3.7) in Lemma 3.3 we have

( 1-
1
2S*
)-1 = L
00 1 .
2}(S*)J, (3.15)
J=o
and
1
-(S - 3/)- = (31 - S)- =
1
3"1 ~
~
1 .
3} SJ, (3.16)
}=o
and the formula for T- 1 can be rewritten in matrix form as follows

1/3
1
~ [~ 1~2 l:t: •••]
::]

1/22
7/12
43/36
:: ]
Let us examine the matrix entries)~ of T- 1 in more detail. From (3.14)-(3.16)
we see that

)~ = ~ miI: k
) (~r-V (~y-v
v=o
An easy calculation shows that

)kx -_ 3J+ 112k+ 1 (6J+I _ 1)"lor j ~ k ,

and
)kx = 3J+ 112k+ 1 (6k+1 -
1)"lor j ~ k .
3.3 Band Toeplitz Operators 149

From these formulas it follows that)~ can be represented in the form

j ,k=O,I, . . . ,

where
3 k - ' , for k S i .
si-«> { zi:", fork~j,

and
1
lik = 3j+ 12k+ 1 j,k=O ,I, . . . .

Notice that the operator

G = (gj-dfk=O
is a Toeplitz operator and the operator

is an operator of rank one.

Now let us compare this with Example I in the first section. There we considered
the Laurent operator A with the same symbol co as for T , and we showed that A is
invertible with its inverse being given by
A -I ( x)oo
= a j-k j, k=O'
where
«: = {3~j, for j ~ 0,
J 2J , for j S O.
This means that G = PA- 1 P I ImP, where P is the projection defined by the
equality

We proved that
T - 1 =PA-1PIImP+F.
The latter equality can be derived for other Toeplitz operators T too and can be
used for the computation of T- 1 •
Part 2. Let us determine the spectrum a (T) of the Toeplitz operator T with
symbol w given by (3.13). As we have seen in Example I of the first section the
curve t t-+ w(e it ) is precisely equal to the ellipse

( !JtA _ ~) 2 + 25(~A)2 = 1. (3.17)

Thus we can apply Corollary 3.4 to show that a (T) consists of all points A E C
that lie on or are inside the ellipse (3.17), that is,

a(T) = I
A Eel ( !JtA - ~y + 25(~A)2 S IJ .
150 Chapter III Laurent and Toeplitz Operators

On the other hand, as we have seen in Example 1 of the first section, the spectrum
of the Laurent operator A with symbol co given by (3.13) is precisely the ellipse
(3.17). Thus in this case a(A) is just the boundary iJa(T) of the spectrum ofT .
Part 3. We continue with an analysis of zol - T for the case when zo is a point
strictly inside the ellipse (3.17). From the result in the previous part we know
that zol - T is not two-sided invertible. We claim that zol - T is right invertible
with dim Ker (zol - T) = 1. To see this, notice that WO(A) = W(A) - zo is the
symbol of the Toeplitz operator To = T - zcl , and A E a(To) if and only if
A - zo E a(T) . Now, apply Theorem 3.2 to To. Since zo is inside the ellipse
(3.17) , we have W(eit ) - zo i- 0 for each t, and hence Wo (e") i- 0 for each t.
Furthermore, the winding number of the curve t f-+ wo(e it ) with respect to zero is
precisely equal to the winding number of the curve t f-+ w(e it ) with respect to ZOo
Since (cf., formula (5) in the first section)

7 1
w(e')
1
= 5" - cost + 5"i sint ,
the orientation on the curve t f-+ w(e it ) is clockwise, and hence the winding
number is -I . Thus we can apply Theorem 3.2 to show that To = T - zo I is right
invertible and dim Ker To = I.
Let us specify further the above for the case when zo = 7/5 (the center of the
ellipse (3.17)) . Then the symbol WO(A) is given by

Notice that the roots ±i.../3 /2 are outside the unit circle . Thus in this case the
number K = e- r corresponding to (3.2) is equal to -1 (which we already know
from the previous paragraph). Furthermore, by Theorem 3.2 (ii), the operator
T - ~ I is right invertible and a right inverse is given by

7 )(-1) =-2:5 ( S-i (3)-1

2:
(S+i (3)-1
2: S.
( T-5"1 Y Y
More generally, let zo be an arbitrary point strictly inside the ellipse (3.17).
Then

WO(A) = --A
5
2-I [2 5(7
2 5
zo) + -3]
A - -
2 '
- -

and the roots ai, a2 of the quadratic polynomial A2 - ~ (~ - zo) + ~ are outside
the unit circle. To see this, recall that wo(e it ) i- 0 for each t and each choice of
zo strictly inside the ellipse (3.17). Next, notice that the roots al and a2 depend
continuously on the point zo, and for Zo = 7/5 they are outside the unit circle.
Thus, if for i = 1 or i = 2 the root a ; would not be outside the unit circle, then
3.3 Band Toeplitz Operators 151

the root a; has to cross the unit circle, which contradicts the fact that wo(e it ) =1= 0
for each t. Thus

with laii > 1, la21 > 1, and by Theorem 3.2 (ii) the operator T - zoI is right
invertible and a right inverse is given by

5
(T - zoI)(-I) = -2"(S - aII)-\S - a2I)-IS.

By using the second part of formula (3.7) in Lemma 3.3 we obtain that

(zoI - T)(-I) = _~_1 f (L (~)j (~)k) Sn+l .

2 ala2 n=O j+k=n al a2

Thus (zoI - T)(-I) is a strictly lower triangular Toeplitz operator, namely

o 0 0 0
». 0 0 0
I) 5 1 bi bl 0 0
(zoI - T) ( - = ---
2 a la2 b3 b3 bl 0

where

Part4. Next let zo be a point strictly outside the ellipse (3.17) . We already know
from Part 2 that in that case zo rj. a(T) . Thus zoI - T is invertible. To compute
its inverse, we consider again the corresponding symbol

In this case one root, al say, of the quadratic polynomial

is inside the open unit disc , and one root, a2 say, is outside the unit circle. For
zo = 0 we know this from Example 1. Indeed, for zo = 0 we have al = 1/2 and
a2 = 3. For an arbitrary zo it follows by using a continuity argument similar to
152 Chapter III. Laurent and Toeplitz Operators

the one used in the third paragraph of the previous part. Thus Theorem 3.1 yields
that

Using the formulas appearing in (3.7) we see that

.'.J [1 ar '.'J
1 0
-I 5
-I
a2
1 o 0 al1 al
(zoI - T) = -- -2 -I 1 0 0 1 .
2a2 a a2
[ 2 .. : ·0
" .

This product we can compute in the same way as we compute T- 1 at the end of
Part 1. We have

where
. _ [a2". j
gJ -
for j = -1 , -2, .
af for j=0,1 ,2, .
and

i .k = 0,1 ,2, . . . .

Part 5. Finally, let zo be a point on the ellipse (3.17) . As we have already seen in
Part 2, this implies that zo E oa(T) . We claim that in this case zoI - T is neither
left nor right invertible. Indeed, let zoI - T be left or right invertible . Since zo
belongs to the spectrum of T, the operator zoI - T is not two-sided invertible.
Thus by the stability results of Section 11.15 there is an open neighborhood of zo
such that for each point z~ in this neighborhood the operator z~I - T is one-sided
invertible but not two-sided invertible . Since zo E Ba (T) , this is impossible.

3.4 Toeplitz Operators with Continuous Symbols

In this section we study invertibility, one-sided or two-sided, of Toeplitz operators

with a continuous symbol to . To state the main theorem we need the notion of a
winding number.
Let co be a continuous function on the unit circle T, and assume that w(e it ) i= 0
for - n :s t :s n . Then the closed curve parametrized by t r+ w (e") , where
-n :s t :s n , does not pass through zero . Again we define the winding number K
of w relative to zero to be the number oftimes the oriented curve t r+ w(e it ) circles
3.4 Toeplitz Operators with Continuous Symbols 153

around 0 when t runs from - l r ~ t ~ n . In other words, the winding number is

equal to 1/2lr times the total variation of the argument function arg(w(e it)) when
the variable t varies over - l r ~ t ~ n . In the sequel we shall need the following
fact (which is known as Rouche's theorem) : if co and iiJ are both continuous
functions on the unit circle 1I' and

(4.1)

then the two curves t t--+ w(e it) and t t--+ iiJ(e it ) , with - l r ~ t ~ tt , do not pass
through zero and the corresponding winding numbers are equal.
The following theorem is the main result of this section.

Theorem 4.1 Let T be the Toeplitz operator with continuous symbol w. Then T
is left or right invertible ifand only ifw(e it) i= Ofor - l r ~ t ~ tt . Furthermore,
ifthis condition is satisfied, then
(i) T is left invertible ifand only if K 2: 0, and in that case

codim 1m T = K, (4.2)

(ii) T is right invertible ifand only if K ~ 0, and in that case

dim Ker T = -K. (4.3)

Here K is the winding number ofw relative to zero. In particular, T is two-sided

invertible ifand only ifw(e it ) i= Ofor - l r ~ t ~ tt and the winding number ofw
relative to zero is equal to zero.

Proof: We split the proof into two parts. In the first part we assume that co has
no zero on the unit circle and prove (i) and (ii). The second part concerns the
necessity of the above condition.
Part I. Assume w(e it ) i= 0 for - l r ~ t ~ n . We prove (i) and (ii) by an
approximation argument using Theorem 3.2. By the second Weierstrass approxi-
mation theorem (see Section 1.13) there exists a sequence WI (e") , w2(e it), .. . of
trigonometric polynomials such that

max Iw(e it ) - wn(eit)1 ~ 0 (n ~ 00). (4.4)

-7TS:t~rr

From (4.4) it follows that for n sufficiently large wn(A) i= 0 for IAI I and
wn(A) -lw(A) ~ I uniformly on IAI = I when n ~ 00. Thus we can find a
trigonometric polynomial iiJ such that iiJ(e it ) i= 0 for - l r ~ t ~ n and

(4.5)

The function iiJ can be written in the form

(4.6)
154 Chapter III Laurent and Toeplitz Operators

where c is a non-zero constant, r is a non-negative integer, ai, ... , ae are complex

numbers inside the open unit disk, and fh, . . . , fJm are complex numbers outside
the closed unit disk . Using (4.5) and the remark related to the inequality (4.1)
we see that the winding numbers of wand iiJ relative to zero are equal. Thus the
number l - r in (4.6) is equal to the winding number K of w relative to zero .
Let C be the Toeplitz operator with continuous symbolequalto YeA) = iiJ(A)-1 x
w(A) - 1. From (4.5) and Theorem 2.2 we see that 1IC11 < 1, and the Toeplitz
operator I + C is invertible. Next put

These operators are also invertible. This follows from Lemma 3.3 and the fact that
la; I < I and IfJj I > I for each i and j. We claim that

(4.8)

or
(4.9)

Indeed, if K 2: 0, then K = l - r and the product rules appearing in the paragraph

before the proof of Theorem 3.2 imply that the right hand side of (4.8) is the
Toeplitz operator of which the symbol w# is given by

Since iiJ is given by (4.6) and all scalar functions commute, we see that w# = w,
and hence (4.8) holds. The identity (4.9) is proved in a similar way.
From (4.8) and the invertibility ofthe operators T_, T+ and I + C we conclude
that T is left invertible if K 2: 0. In fact, in that case a left inverse T(-l) of Tis
given by

(4.10)

Moreover,
codim Im T = codim Im SK = K.

Similarly, (4.9) yields that T is right invertible for K :::: 0, a right inverse of T(-l)
of T is given by

(4.11)

and
dimKer T = dimKer (S*)-K = -K.

This completes the proof of (i) and (ii).

3.4 Toeplitz Operators with Continuous Symbols 155

Part 2. In this part T is left or right invertible, and we prove that w(e it ) i= 0 for
- ] f ::::: t ::::: it . The proof is by contradiction. We assume that w(Ao) = 0 for some

lAO I = 1. Since T is left or right invertible, the perturbation results of Section I.15
show that there exists e > 0 such that the operator T on £z is left or right invertible
whenever
liT-TIl <c. (4.12)

By the second Weierstrass approximation theorem (see Section I.13) there exists
a trigonometric polynomial w(e it ) such that

(4.13)

Now let T be the Toeplitz operator with symbol WCA) = w(A) - w(Ao) . Since
w(Ao) = 0, we see from (4.13) that IW(Ao)1 < te. Hence, again using (4.13), we
have
IWCe it ) - w(eit)1 < e, -]f::::: t ::::: tt .

But then we can use Theorem 2.2 to show that T satisfies (4.12), and hence T is left
or right invertible. However, w(Ao) = 0 by construction, and thus Theorem 3.2
tells us that T cannot be left or right invertible . So we reached a contradiction.
We conclude that w(e it) i= 0 for - ] f ::::: t ::::: it . 0

Corollary 4.2 Let T be the Toeplitz operator with continuous symbol to. Then the
sp ectrum ofT is given by
a(T)=C\Q , (4.14)

where Q is the set ofall points A in C such that A i= w(e it) for all - ] f ::::: t ::::: x
and the winding number ofco (.) - A relative to zero is equal to zero. In particular,
the spectral radius ofT is equal to the norm ofT.

The proof ofthe above corollary is the same as that of Corollary 3.4, except that
the reference to Theorem 3.1 has to be replaced by a reference to the final part of
Theorem 4.1. We omit further details. 0
We conclude this section with the case when the symbol ofto ofthe Toeplitz T is a
rational function. In that case the process of inverting T can be carried out without
the approximation step described in the first part of the proof of Theorem 4.1 .
Indeed, let
w(A) = PI (A) , IAI = 1,
pZ(A)

where PI and PZ are polynomials which have no common zeros . Furthermore, we

assume that PI and PZ have no zeros on T , which means that we assume that co is
continuous on T and to (eit ) i= 0 for each t.
156 Chapter III. Laurent and Toeplitz Operators

Because of our assumptions on PI and P2 we can write

n n
k+ r:
PI (>..) = CJ (>.. - tt) (>.. - ti),
j=1 j=1

n rt) n(>.. - ri),

e+ t:
p2(>") = C2 (>.. -
j=1 j=1

where tt and rt are inside the unit circle and the points ti and ri are outside 11'.
It follows that co can be represented in the form

>.. E 11', (4.15)

where K = k+ - l+, and the factors co.: and w+ are rational functions which have
their zeros and poles inside 11' and outside 11', respectively. In fact,

Notice that K = k+ - l+ is precisely the winding number of o» relative to zero.

Thus T will be invertible if and only if co = w_w+ .
Let T_ and T+ be the Toeplitz operators with symbols co.: and w+, respectively.
Using the multiplication rules appearing in the paragraphs before the proof of
Theorem 3.2 we see that

T_ = CI
C2
nU- 7- *) IT u -
k+
t S
J
e+
r + S *)- I,
J
j=1 j=1
r:
n
k-

T+ IT (S - ti I) (S - ri I)-I ,
j =1 j=1

and
(4.16)

where
Sen) _ {sn for n = 0, 1, 2, ,
- (s*)-n for n=-I ,-2, .
Notice that because of commutativity the order of the factors in the formulas for
Land T+ is not important. However in (4.16) the order is essential.
3.4 Toeplitz Operators with Continuous Symbols 157

From Lemma 3.3 and the location of the points tt

and 1:; it follows that the
Toeplitz operators T_ and T+ are invertible and their inverses are given by

T~I = C2
CI
nU-
f+
1:+S*)
J
nU- t7-
k+

J
S*)-I,
j=I j=1

T;I = nk:

j =1
(S - 1: -
j
I) n
r:

j=I
(S - t; I) -I .
In other words, T~I and T;I are the Toeplitz operators with symbols l/uL and
l / w+, respectively. Furthermore, we see that T~I and T;I are of the form

:J
- -
Y-I Y- 2
[ Yo- Yo-
T~I = 0 Yo~
Y-=-1
(4.17)

J
0 0
[ y~+ Yo+ 0
-I YI
T+ = Y( Yo+ yt (4.18)

where

0
co.: (A) = L
j =-oo
j
Yj- A , IAI ~ 1, (4.19)

00

W+(A) = L yf Aj , IAI .s 1. (4.20)

j =O
The first conclusion is that the Toeplitz operator with the (rational) symbol (4.15)
is invertible if and only if K = 0, and in that case its inverse is given by
I
T- = (YjdJ:k=O'

where
min(j . k )

Yjk = L yf-r Yr-=-k

r=O

with Yj- and yf being given by (4.19) and (4.20), respectively.

158 Chapter III. Laurent and Toeplitz Operators

Furthermore, if K < 0, then the operator T is right invertible (but not left) and
a right inverse is given by

(4.21)

If K > 0, then the operator T is left invertible (but not right), and a left inverse is
given by
T(-I) = T;Ics*tT.=-I. (4.22)

The formulas (4.21) and (4.22) are the analogs of formulas (3.4) and (3.5) in
Theorem 3.2. As in the case K = 0, formulas (4.17) and (4.18) can be used to
obtain the matrix entries in the matrix representations of the operators T(-l) in
(4.21) and (4.22).
Example: Consider the Toeplitz operator on £2 given by

1 1/2 1/4 .. . ]
1/3 1 1/2
2
T = 1/:3 1/3 1 ". .
[

We met the corresponding Laurent operator as the inverse of the operator A in

Example 1 of Section 1. The symbol W of T is easy to compute:

W(A) -
_ 00
~
(~)j
2 A
r i
+L
00 (~)k
3
k_
A -
1
1 _ A-1 /2
A/3
+ 1- A/3
)=0 k=1

= <;5 ( 1 - 1
2A -I
)-1 ( 1- 3I A)-1
Thus the symbol is a rational function , and we have the representation (4.15) with
K = 0 and

5 ( 1- 1 A- I
W_(A)=<;
2
)-1 ,

We conclude that T is invertible and

Hence T- I is the tridiagonal operator

6/5 -3/5 0
-I -2/5 7/5 -3/5
T = 0 -2/5 7/5

r
3.5 Finit e Section Method 159

Notice that T - I = G - F , where

7/5 -3/5 0

G=
[
-2/5 7/5 -3/5
0 -2/5 7/5
"'J
..
.
' F=
[~~~" 'J
000
.'
.
..
'
.

Thus T- I is a rank one perturbation of a tridiagonal Toeplitz operator G . More-

over the Laurent operator corresponding to G is the operator A in Example 1 of
Section 1.

3.5 Finite Section Method

Let T be a Toeplitz operator on £2 with continuous symbol co. Thus

ao a _I a- z
al ao a-I .
J
T = oz al ao ' (5.1)
[ ·· ..
· .
where

an = -
1 j :n: w (e't. )e- rnt
. dt , n E Z. (5.2)
2rr -:n:
Given a non-negative integer N , the N -th section TN of T is the linear operator
on eNgiven by

[ aaDl
a_ I
ao
... G- N
: : . a_~+ 1
J . (5.3)
TN = a~ . .
aN - I . . . ao

Here we identify a linear operator on en

with its matrix with respect to the standard
basis of en.Assume T is invertible . In this section we study the problem to get
the solution x of the equation Tx = y , where y = (Yo, YI, Y2 , .. .) in ez, as a limit
of solutions of the truncated equation

(5.4)

More precisely, we say that the finite sec tion method for T con verges if TN is
invertible for N sufficiently large, N ~ No say, and for each y = (Yo, YI, Yz , . . .) in
160 Chapter III Laurent and Toeplitz Operators

l2 the vector x(N) = (xo(N), . .. , xN(N), 0, 0, ...), where (xo(N), ... , xN(N))
is the solution of (5.4) for N ~ No, converges in the norm of l2 to the solution x
ofTx = y .
In the sequel we identify the space eN with the subspace HN of l2 consisting of
all x = (x j )'/=,0 in l2 such that x j = 0 for j > N, and we let PN be the orthogonal
projection of l2 onto HN = eN. It then follows that the finite section method for
T converges ifand only ifthe projection method relative to {PN }'f!=o is applicable
to T, that is, using the notation of Section II.17, the operator T belongs to n {PN }.
The following theorem is the main result of this section.

Theorem 5.1 Let T be a Toeplitz operator with a continuous symbol co. If T is

invertible, then the finite section methodfor T converges.

Before we prove the theorem it will be convenient to make some preparations. The
first is a general proposition and the second a lemma.

Proposition 5.2 Let A be an invertible operator on l2 , and assume that with

respect to the standard orthonormal basis of l2 the matrices of A and A -I are
both lower triangular or both upper triangular. Then for each N the N -th section
AN of A is invertible and

(AN )-1 PN = PNA -I (lower triangular case) (5.5)

(AN)-l pN =A-1PN (upper triangular case) (5.6)

where PN is the orthogonal projection of l2 onto the space spanned by the first
N + 1 vectors in the standard basis. Furthermore

(5.7)

Proof: First consider the case when the standard matrices of A and A -I are both
lower triangular. So

with zero elements in the strictly upper triangular part. Then the N -th sections of
A and A -I are given by

:~~. aoo :~~. bll

..
AN= . (A-1)N = .
[
.; [
b~o b~1
3.5 Finite Section Method 161

Since AA -I = A -I A is the identity on f2, it follows that AN is invertible and

(AN)-I = (A-1)N . Now, take y E f2 . Then

(AN)-I PNY = (A-1)NPNy = PNA-1y,

which proves (5.5). Moreover, (5.7) follows by using that IIA-I y- PN A -I yll ---+ 0
if N ---+ 00. This completes the proof for the case when the standard matrices of
A and A -I are both lower triangular. The upper triangular case is proved in a
similar way. 0

Lemma 5.3 Let T be an arbitrary Toeplitz operator, and let R = nm=1 (S - f3 j l),
where S is the forward shift on f2 and f3, .. . ,f3m are complex numbers. Then

rank (RT - TR) < 00. (5.8)

Proof: First notice that ST - TS has rank at most one . Indeed, let eo, el, e2, . ..
be the standard orthonormal basis of f2, and let

::: ]
[
:~ aa~1 :=~
T = a2 al ao .
·· ..
· .
Then for each n = 0, 1,2, .. . we have

Thus rank (ST - TS) :::s 1. Next observe that

L sr:' (ST -
P
SPT - T se = TS)Sj-l,
j=1
and thus
rank (SPT - TSP) :::s p, p = 1,2, . . .. (5.9)
Now notice that R is a polynomial in S of degree m, and hence (5.9) yields that
the rank of RT - TR is at most mp. 0
Proof of Theorem 5.1: Since T is assumed to be invertible, we know from
Theorem 4.1 that w(e it ) =1= 0 for -T( ::::; t :::s T( and the winding number K of w
relative to zero is equal to zero . But then we can use the first part of the proof of
Theorem 4.1 to conclude that T can be represented in the form

(5.10)
162 Chapter III Laurent and Toeplitz Operators

Here c is a non-zero constant, C is a Toeplitz operator with continuous symbol y

satisfying ly(eit)1 < 1 for -Jr S t S tt , and

(5.11)

where S is the forward shift on £2, and ICti 1< 1, 1,8j 1> 1 for each i and j . As we
have seen before, the fact that ICti 1< 1 and 1,8j I > 1 for each i and j implies that
the operators 1'_ and 1'+ are invertible (by Lemma 3.3) . In fact,

1'-I=IT
- ~,=1 (I-Ct ''S*)-1 , 1'-I=IT~
+ )=1 (S-,8 )·I)-1 . (5.12)

From (5.11) and (5.12)we also see that with respect to the standard basis 0[£2 the
matrices of 1'_ and 1'~ 1 are upper triangular, and those of 1'+ and 1'; 1 are lower
triangular.
Consider the operator
(5.13)
By comparing (5.10) and (5.13) we see that F is obtained from T by interchanging
the factors 1'_ and 1'+. According to Lemma 5.3 the operators 1'+ C - C 1'+ and
1'+1'_ - 1'_ 1'+ are of finite rank. It follows that T - «i:+1'_ (l + C) is an operator
of finite rank . Also , by taking adjoints in Lemma 5.3 we see that 1'_C - C1'_ is
of finite rank . So we have proved that the operator T - F is of finite rank.
Notice that the three factors in the right hand side of(5.13) are invertible, and
hence F is invertible . But then, by Theorem 11.17.7, the result of the previous
paragraph shows that T E IT {PN} if and only if F E IT {PN }. Moreover, since
IICII < 1 we also know (see Theorem 11.17.6) that I + C E IT{PN}.
Let 1'+,N and 1'_,N be the N-th sections of 1'+ and 1'_, respectively, and let FN
and C N be those of F and C, respectively. Since 1'+ is upper triangular and 1'_
lower triangular with respect to the standard basis of £2, we have

and thus
(5.14)
From the triangular properties ofthe operators in (5.11) and (5.13) it follows that
we can apply Proposition 5.2 to show that 1'+.N and 1'_,N are invertible . Also,
IICNII S 1IC11 < 1, and therefore I + CN is invertible . Using (5.14) we see that
for each N the N -th section F N is invertible , and

FN1 = -1(-L ,N)-1 (l + CN)- 1(T+,N

- )-1 . (5.15)
C

Next, take y E h Formula (5.5) applied to A = 1'+ yields

3.6 The Finite Section Methodfor Laurent Operators 163

Now use that 1+ C E IT{PN} . It follows that

XN := (l + CN)-I(1'+,N)-1 PNY = (l + CN)-I PN1';l y ~ (l + C)-I1';l y

for N ~ 00 . Put x = (l + C)-I 1';1 y. Since 1'_ and 1'~1 are upper triangular,
Proposition 5.2 applied to A = 1'_ yields

We conclude that

and hence FE IT{PN}. o

3.6 The Finite Section Method for Laurent Operators

In this section we consider the finite section method for Laurent operators. Let
L be an invertible Laurent operator on f2 (Z) . Recall (see Section 11.17) that the
finite section method is said to converge for L if L E IT{Qn}, where Qn is the
orthogonal projection OU2(Z) onto span {e- n, . .. , en}. Here . . . , e_l, eo, el, . ..
is the standard orthonormal basis of f2(Z).
In general, in contrast to Toeplitz operators , from the invertibility of L it does
not follow that the finite section method converges for 1. For instance, assume
that L = V , where V is the bilateral forward shift on f2 (Z). Then L is invertible,
but

000 o0
100 o 0
010 o0

o0 0 1 0

Hence there is no n for which L n is invertible, and therefore L = V ¢ IT{ Qn}. To

get the convergence of the finite section method for an invertible Laurent operator
an additional condition is required (see the next theorem) which in the Toeplitz
case follows from invertibility.

Theorem 6.1 Let L be an invertible Laurent operator on f2(Z) defined by the

fun ction aCt) = w(e it ), where to is continuous on the unit circle 11' (satisfying
w(e it ) i= Ofor each -1r :s t :s zr). In order thatthefinite section method converge
for L it is necessary and sufficient that the winding number K ofco relative to zero
is equal to zero.
164 Chapter III Laurent and Toeplitz Operators

Proof: The sufficiency follows from Theorems 4.1 and 5.1. Indeed, assume the
winding number K of co relative to zero is equal to zero. Let T be the Toeplitz
operator on £2 with symbol co. Since K = 0, Theorem 4.1 implies that T is
invertible. But then we can use Theorem 5.1 to show that the finite section method
converges for T. Thus for n sufficiently large, n ~ no say, the operator PnTPn is
invertible on Im P, and

sup IICPnTPn)-IIIm Pnll < 00.

n ~n o

Here Pn is the orthogonal projection onto the space spanned by the first n
vectors in the standard orthonormal basis of £2. Notice that we may identify
both Im P2n+1 and Im Qn with C 2n+ l , and in that case we have P2n+1 TP2n+1 =
QnLQn . We conclude that QnLQn is invertible on Im Qn for n ~ no and
sUPn >n
_ 0 IICQnLQn)-IIm Qnll < 00. By Theorem ILl7.! this implies that the
finite section method converges for L.
For the proof of the necessity of the winding number condition we refer to
the proof of Theorem XVI.5 .2 where the necessity is established in a somewhat
different setting. The proof given there carries over to the case considered here.
o
To understand Theorem 6.1 better, consider the Laurent operator

ao a-I a-2
L = al §] a-I

Let co be the symbol of L which we take to be continuous on 11'. Assume L is

invertible, i.e., wCe it ) =I- 0 for each -Jr ~ t ~ Jr . Let K be the corresponding
winding number. Notice LI = V- KL is again a Laurent operator. In fact,

aK aK-I aK-2
LI = aK+I ~ aK-I

The difference between the matrices for L and L I is only in the location of the
entry in the zero-zero position .
3.6 The Finite Section Methodfor Laurent Operators 165

The symbol of L I is equal to CVI (A) = A-K cv (A). We have CVI (e") :F 0 for each t
and the corresponding winding number is equal to zero. Thus L I is invertible , and
Theorem 6.1 tells us that the finite section method converges for L I. This means
that the matrices
QnLI QnlIm Qn = (aK+J-r)'}.r=-n
are invertible for n large enough and

lim IIWnLIQn)-IQny-Ljlyll =0.

n~oo

Since LI = V- KL, the latter limit can be rewritten in the form

Notice that
V KQn V- K = Qn,K'
where Qn,K is the orthogonal projection of f2(Z) onto span {e-n+K" ' " en+K}
with .. . , e_l, eo, el , ... being the standard orthonormal basis Off2(Z) . We arrive
at the conclusion that for the operator L the projection method with respect to two
sequences of projections (Qn , K' Qn }nEZ is convergent. In other words to have the
finite section method converging for an invertible Laurent operator L one has to
take the sections around the x-th diagonal, where K is the winding number of the
symbol of L.
As we have seen in Section II. 18 it is sometimes convenient to a consider a
modified finite section method. This is also true for Laurent operators.

Example: Consider the Laurent operator L with symbol CV(A) = -2A+ 7 -3A -I .
Since

we see that cv(eit ) :F 0 for each t and relative to zero the winding number of
the corresponding curve is equal to zero . It follows that the finite section method
converges for L.
To find the inverse of L it is convenient to use a modified finite section method.
Notice that the n-th section of L is the following (2n + I) x (2n + I) matrix

7 -3 0 0 0 0
-2 7 -3 0 0 0
0 -2 7 -3 0 0
Ln = 0 0 -2 7 0 0

0 0 0 0 7 -3
0 0 0 0 -2 7
166 Chapter III. Laurent and Toeplitz Operators

Now replace the entries in the left upper comer and in the right lower comer by 6,
and let Fn be the resulting matrix . As we have seen in Section II.18 the inverse of
Fn is given by

I 2- 1 2- 2 2- 2n
3- 1 I 2- 1 2- 2n + 1
F- I = ~ r 2
r l
I 2- 2n +2
n 5=

3- 2n 3-2n+ 1 3- 2n +2

Hence one expects the inverse of L to be given by

I 2- 1 2- 2 2- 3 2- 4
3- 1 I 2- 1 2- 2 2- 3
L- I = ~
5
= 3- 2 3- 1 r l 2-2OJ
r 3 r 2
r l
I r l
3- 4 3- 3 3- 2 3- 1 I

It is straightforward to check that indeed this is the correct inverse of L.

The sources for the material covered in this chapter can be found in [Kre],
Section 3.13, [GF], Chapters 1-111; see also [BS] and [GGK2], Part VI. In these
monographs one can also find further developments ofthe contents ofthis chapter.
For a partial extension of the theory to Banach spaces, see Chapter XVI. Finally
we would like to note that there is some difference in terminology between the
above mentioned sources and Chapters 111 and XVI of the present book.

Exercises III

1. Let A be the Laurent operator on l2Gl) defined by

aCt) = _4e- it + 17 - 4e it •

(a) Prove that the operator A is strictly positive, that is, for some s > 0

(b) Show that A is invertible, and find the matrix of A -I with respect to the
standard basis of l2Gl) .
(c) Find the spectra of the operators A and A -I .

2. Let lal =1= 1. Do the previous exercise for the Laurent operator A on e2G~)
defined by
Exercises III 167

3. Consider on l2 (Z) the Laurent operator

1 a a 2 a 3 a4
b 1 a a2 a3
A= b2 b ITJa a2
b3 b2 b 1 a
b4 b3 b2 b

where lal < 1 and Ibl < 1.

(a) Compute the symbol of A .
(b) Find the inverse of A when it exists.
(c) Determine the spectrum of A .

4. Let [zz I i= 1, and let A be the Laurent operator on l2 (Z) with symbol

1- lal 2
W(A) = 2 I' A E T.
1 + lal - a): - aA-

(a) Prove that A is invertible, and determine its inverse.

(b) Find the spectra of A and A -I.

5. Can it happen that the inverse ofa lower triangular Laurent operator is upper
triangular?
6. Prove or disprove the statement: a left invertible Laurent operator is always
two-sided invertible.
7. Fix t E JR, and let U be the operator on L2 (-00 , (0) given by

(Uf)(x) = f(x + t) , x ER

Recall (see Exercise 86 to Chapter II) that U is unitary and a(U) = T .

(a) Prove that the operator

A=-2U+71-3U- I .

is invertible and determine its inverse .

(b) Given g E L2(-00 , (0) solve in L2(-00 , (0) the functional equation

-2f(x + t) + 7 f(x) - 3f(x - t) = g(x), x ER

(c) Determine the spectrum of the operator A defined in (a).

168 Chapter III. Laurent and Toeplitz Operators

8. Fix a > O. Let Ra be the operator on L2(0, 00) defined by

(Raf)(t) = feat), t ::: o.
(a) Given g E L2(0, 00) solve in L2(0, 00) the functional equation
00 00

Lb j f(a-jt) + f(t) + Laj f(ajt) = get) , t::: O.

j=1 j=1

Here lal <..;a

and Ibl <~.
(b) Determine the spectrum of the operator A given by
00 00

(Af)(t) = Lb j f(a -jt) + La j f(ajt) , i e: 0,

j=1 j=1

where lal < -JCi and Ibl < ~.

Hint : Use the results of Example 6 in Section H.20.
9. Let T be the Toeplitz operator on £2 given by the following tridiagonal
matrix representation:
17 -4 0 0
-4 17 -4 0
T= o -4 17 -4
o 0 -4 17

(a) Prove that the operator T is strictly positive, that is, for some e > 0
(x E £2) .

(b) Show that T is invertible, and find the matrix of T- 1 with respect to
the standard basis of £2.
(c) Find the spectra of T and T- 1•
10. Let lal =1= 1. Do the previous exercise for the Toeplitz operator T on £2
with symbol
a(A) = a): -I + 1 + lal2 + CiA, A E T.

11. Consider on £2 the Toeplitz operator T given by

2
1 a a ... ]

T=
[
:, ~ 7 ". !
where lal < 1 and Ibl < 1.
Exercises III 169

(a) Compute the symbol ofT.

(b) Find the inverse of T when it exists.
(c) Determine the spectrum of T .

12. Let lal =1= 1, and let T be the Toeplitz operator on £2 with symbol

1 -lal 2
(A) - ---;:-------:- A E T.
- 1 + lal 2 - a): - aA -I '
(J)

(a) Find the matrix of T with respect to the standard basis of h

(b) Prove that T is invertible, and determine its inverse.
(c) Find the spectra of T and T- 1•

13. Let T be the Toeplitz operator on £2 with the following tridiagonal matrix
representation:
oa 0 0
bOa 0
T=
o bOa
o0 b 0

(a) For which valuesofa andb in C is the operator T left, right or two-sided
invertible? For these values determine the one- or two-sided inverse of
T.
(b) Solve the equation Tx = y, where y = (l, q, q2, . . .) with Iql < 1.

14. Fix t > O. Let W be the operator on L2(0 , 00) defined by

(Wj)(x) = {/(X - t), for x> t ,

0, for 0 < x ~ t.

Recall (see Exercise 85 to Chapter II) that W is an isometry and a(W) is

the closed unit disc .
(a) Determine Ker W* and 1m W.
(b) Prove that the operator

T = -2 W + 7I - 3W*

is invertible and determine the inverse .

(c) Given g E L2(0 , 00) solve in L2(0, 00) the functional equation

-2f(x + t) + 7 f(x) - 3f(x - t) = g(x), x> O.

Here we follow the convention that f(x - t) = 0 for 0 < x ~ t.

(d) Find the spectrum of the operator T in (b).
170 Chapter III. Laurent and Toeplitz Operators

15. Fix 0 < a < 1. Let Ma be the operator Lz([O, 1]) defined by

I
Ja!(~t), 0 ~ t ~ a,
(Maf)(t) =
0, a < t ~ 1.

(a) Prove that M a is an isometry, and show that

(M~f)(t) = .Ja!(at),

(b) Determine Im u; and Ker M~.

(c) Find the spectra of the operators M a and M;.

16. Fix 0 < a < 1. On la ([0, 1]) consider the operator

T = -4Ma + 17/ -4M~,

where M a is defined in the previous exercise . For the operator T solve the
problems posed in (a), (b) and (c) of Exercise 9.
17. Fix 0 < a < 1. On Lz ([0, 1]) consider the operator

where Ma is defined in Exercise 15.

(a) For which values of a and bin <C is the operator T left, right or two-sided
invertible ? For these values determine the one- or two-sided inverse of
T.
(b) Find the spectrum of T .

18. Fix 0 < a < 1. Find the spectrum of T, where T is the operator on
Lz ([0, 1]) given by
00 00

T = 'Laj M a + / + 'Lbj(M~)j .
j=! j=!

Here Ma is as in Exercise 15, and lal < 1, Ibl < 1.

19. Let U be an isometry on a Hilbert space H, and U is not unitary. Generalize
the results of Section 3 to operators of the form
p q
T = LCjUj + Ldj(U*)j.
j=O j=!

In particular, for the operator T prove the analogs of Theorems 3.1 and 3.2,
and determine its spectrum.
Chapter IV
Spectral Theory of Compact Self
Adjoint Operators

One ofthe fundamental results in linear algebra is the spectral theorem which states
that if H is a finite dimensional Hilbert space and A E £(H) is self adjoint, then
there exists an orthonormal basis cpI, . .. , CPn for H and real numbers AI, ... , An
such that
ACPi = Aicpi, l::s i ::s n.
The matrix (aij ) «Acp j , cpi) ) corresponding to A and cpI , . . . , CPn is the
diagonal matrix

A natural question is whether this spectral theorem can be generalized to the

case where A is self adjoint and H is infinite dimensional. That is to say, is there
an orthonormal basis cpI , CPZ, . . . for H and numbers AI, AZ, . . . such that

ACPi = AiCPi, i s i?

This means thatthe matrix corresponding to A and cpI , <fJ2 , . •. is an infinite diagonal
matrix.
In this chapter it is shown that the spectral theorem admits an important gener-
alization to compact self adjoint operators .
Let us first consider an example which indicates the possibility of a
generalization.

4.1 Example of an Infinite Dimensional Generalization

Let h be a continuous complex valued function of period 2JT . The operator K

i:
defined on L z([-JT , JT]) by

(Kf)(t) = h(t - s)j(s ) ds

172 Chapter Iv. Spectral Theory ofCompact SelfAdjoint Operators

is a bounded linear operator with range in L2([-n, n]). Taking CPn (I) = ~eint,
v2TC
n = 0, ±l, . . . as the orthonormal basis for L2([-n,n]), it follows from the

t:
periodicity of hand CPn that

f
TC
(Kcpn)(l)= _TCh(t-S)CPn(s)ds=

1 . fTC
t-TC h(s)CPn(t-s)ds

.
= _ _ e mt h(s)e- tn S ds = AnCPn(t),
,.j2ii -TC
where
An = fTC h(s)e- ins ds,n = 0, ±l , ... .
-TC
The matrix corresponding to the operator K and (CPn}~-oo' is the doubly
infinite diagonal matrix

4.2 The Problem of Existence of Eigenvalues and Eigenvectors

The examples described above show that the spectral representation theory starts
with the problem of the existence of eigenvalues and eigenvectors.

Definition: A complex number A is called an eigenvalue of A E £(H) if there

exists a cP f= 0 in H such that Acp = Acp. The vector cP is called an eigenvector of
A corresponding to the eigenvalue A.

We shall see later the significance ofeigenvalues and eigenvectors which appear
in various problems in mathematical analysis and mechanics.

Example: Let K be the integral operator defined on H = L2([0.1]) with kernel

function ik, where
k(t ,s) = {I,-1, s::::t
s » r.
Example 5 in 11.2 shows that K is in £(H). To find the eigenvalues and eigenvectors
of K, let us suppose that cP f= 0 and

(Kcp)(t) =i 1t
cp(s)ds-i [1 cp(S)dS=Acp(t)a.e. (2.1)
4.2 The Problem ofExistence ofEigenvalues and Eigenvectors 173

Differentiating each side of (2.1) yields

2icp (I) = ACP' (t) a.e. (2.2)

Thus A = 0 is not an eigenvalue of K . If A =1= 0, it follows from (2.2) that

o.
2;
cp (I) = ceT! a.e ., c =1= (2.3)

By identifying functions which are equal a.e., we may assume (2.3) holds for
all t . Now (2.1) implies that
2;
O=cp(O) + cp(1)=c(l+e T ), c =1= O.

Hence
2
A = (2k + l)rr, k 0, ±1, . . . .

By reversing our steps, it follows that

2
Ak = , k = 0, ± 1, .. .
(2k + l)rr

are the eigenvalues of K and e i (Zk+ l )1T! are eigenvectors corresponding to Ak.

Every linear operator on a finite dimensional Hilbert space over C has an eigen-
value. However, even a self adjoint operator on an infinite dimensional Hilbert
space need not have an eigenvalue.
For example, let A : Lz([a, b]) -+ Lz([a, b]) be the operator defined by

(Ai) (t) = if (t).

Now A is a bounded linear self adjoint operator by 11.12.2 - Example 2. How-

ever, A has no eigenvalue; for if Acp = Acp, then

(r - A)cp (t) = 0 a.e

Thus cp (I) = 0 a.e., which means that cp = 0 when considered as a vector in

Lz([a, b]).
The following results are used throughout this chapter.

2.1 (a) Any eigenvalue ofa selfadjoint operator is real. For if A is self adjoint and
Ax = Ax, x =1= 0, then

whence A = X.
174 Chapter TV. Spectral Theory ofCompact SelfAdjoint Operators

(b) Eigenvectors corresponding to distinct eigenvalues of a self adjoint operator

are orthogonalIndeed, ifA is self adjoint and

Ax = AX, Ay = p.y; Y i- 0,

then I-t is real and

A(X, y} = (Ax, y) = (x, Ay) = ut», y}.

Hence A i- I-t implies (x , y) = O.

(c) If): is an eigenvalue ofA E £(H), then IAI S IIAII. For if Ax = Ax, then

IIAllllxll ~ IIAxll = IAlllxll ·

4.3 Eigenvalues and Eigenvectors of Operators of Finite Rank

Let K E £(H) be an operator of finite rank, say

n
Kx = L(x, ({Jj}1/!j'
j=l

The eigenvalues and eigenvectors of K are determined as follows .

Suppose A i- 0 is an eigenvalue of K with eigenvector x. Then
n
Ax - L(x, ({Jj}1/!j = 0 (3.1)
j=l

and
n
A(X, ({Jk) - L(x , ({Jj}(1/!j, ({Jk) = 0, 1 S k S n. (3.2)
j=l
Now (x, ({Jj) i- 0 for some j, otherwise x = 0 by (3.1). Thus {(x , ({Jj}}~=l is a
non trivial solution to the system of equations J

n
Aq - L(1/!j, ({Jk}Cj = 0, < k S n. (3.3)
j=l

Therefore, by Cramer's rule,

(3.4)

Conversely, if (3.4) holds, then there exist ci , ... , Cn , not all zero, which satisfy
t
(3.3) . Guided by (3.1), we take x = L:J=1 Cj1/!j and getKx = AX.
4.4 Existence ofEigenvalues 175

To summarize, we have shown that A i= ° is an eigenvalue of K if and only if

and x is an eigenvector corresponding to A if and only if it is a non -zero vector of

the form. i L:
J= 1 CX j 0/j » where
(0/1, <fJd - A (0/2, <fJI)
(0/1 , <fJ2) (0/2, <fJ2) - A (o/n, <fJI)
(o/n , <fJ2) ) (CXI)
CX2 (0)°
··· .
..
(
(0/1 , <fJn) (0/2, <fJn) (o/n:<fJn)-A ~n - b
If {o/I, 0/2, .. . , 0/n} is linearly independent, then it is clear from the definition
of K that zero is an eigenvalue of K with eigenvector x if and only if x i= and
xJ..<fJj , 1::::; j::::; n,
°
The above results can also be obtained by applying Theorem II.7.1 to I - K. i
4.4 Existence of Eigenvalues

In this section it is shown that every compact self adjoint operator has an
eigenvalue.
Let us start with the following theorem.

Theorem 4.1 IfA E I: (1-£) is selfadjoint, then

IIAII = sup I(Ax, x)l.
IIxll=1

Proof: Let m = sUPllxll=1 I(Ax , x}l. Then for IIxll = 1,

I(Ax ,x)l::::; IIAxll::::; IIAII·

Hence m::::; II A II. To prove that m ~ II A II , let x and y be arbitrary vectors in 'H.
Then
(A(x ± y), x ± y) = (Ax, x) ± 2 lR(Ax , y) + (Ay , y) .

Therefore,

4 lR(Ax, y) = (A(x + y) , x + y) - (A (x - y) , x - y).

Combining this with the definition of m and the parallelogram law (Theorem I.3.1),
we get
176 Chapter IV. Spectral Theory ofCompact SelfAdjoint Operators

Now (Ax, y) = I(Ax , y) I e i O for some real number (). Substituting e- i Ox for x
in (4.1) yields
I(Ax, Y)I .s ~ (lIxll 2 + IIYII 2 ) . (4.2)

Suppose Ax =1= O. Then taking Y = II~;III Ax in (4.2), we get

Hence IIAxll ~ mllxll for all x E Hand IIAII ~ m. o

The following corollary is an immediate consequence of Theorems Il.I2.3
and 4.1.

Corollary 4.2 IfA E L (H) and (Ax, x) = Ofor all x E H , then A = O.

Now that we know that the least upper bound of the set {1(Ax,x)1 : IIxll = I}
is II A II (A self adjoint), the next problem is to determine if II A II is attained, i.e.,
IIAII = I(Axo , xo)1 for some xo, IIxoll = 1. The next theorem shows that if this
is the case, then at least one of the numbers IIAII or -IIAII is an eigenvalue of A .
Thus if A is a self adjoint operator which does not have an eigenvalue, as in the
example in Section 2, then

IIAII > I(Ax,x)1 for all x, IIxll = 1.

Theorem 4.3 Suppose A is selfadjoint. Let

A= inf (Ax, x).

IIxll=1

If there exists an Xo E H such that IIxoll = I and

A = (Axo , xo) ,

then A is an eigenvalue ofA with corresponding eigenvector xo.

Let
/1 = sup (Ax , x) .
IIxll=1
If there exists an Xl E H such that IIxlli = I and

then /1 is an eigenvalue ofA with corresponding eigenvector Xl.

Proof: For every a E <C and every v E H, it follows from the definition of A that

(A(xo + av), Xo + av) ::=: A(XO + av, Xo + av) .

4.4 Existence ofEigenvalues 177

Expanding the inner products and setting A = (Axo, xo), we obtain the inequality

2 ffi a(v, (A - A)XO) + lal2 «A - A)V, v) ~ O. (4.3)

Taking a = r(v, (A - A) xo}, where r is an arbitrary real number, it is easy to

see that the inequality can only hold for all a if

(v , (A - A)xo) = O.

Since v is arbitrary, Axo - AXo = O.

The second conclusion of the theorem follows from the above result applied to
the self adjoint operator -A. 0
When are we guaranteed that (Ax, x) has at least a largest or a smallest value
as x ranges over the l-sphere ofH? The next corollary supplies an answer to this
question.

Theorem 4.4 IfA E £(H) is compact and selfadjoint, H =1= {O}, then at least one
ofthe numbers II A II or - II A II is an eigenvalue ofA.

Proof: The theorem is trivial if A = O. Assume A =1= o. It follows from Theorem 4.1
that there exists a sequence {xn} in H, IIxn II = 1, and a real number A such that
IAI = II A II =1= 0 and (Axn ,xn) -+ A. To prove that A is an eigenvalue of A , we
first note that

0:::: IIAx n - Axn ll 2 = IIAx n f - 2A(Axn, x n} + A2

:::: 2A2 - 2A(Ax n,xn} -+ O.

Thus
(4.4)
Since A is compact, there exists a subsequence {Axn'} of {Axn} which converges
t
to some Y E H. Consequently, (4.3) implies that X n ' -+ Y , and by the continuity
of A,
Y = lim Ax n , =
n->OO
~Ay
A
.
Hence Ay = Ay and y =1= 0 since II YII = lim n->oollAxn ' II = IAI = IIAII . Thus A is
an eigenvalue of A. 0

Corollary 4.5 If A E £(H) is compact and selfadjoint, then maxj, 11 =1 I(Ax, x) I

exists and equals II A II.
Proof: By Theorem 4.4, there exists a A which is an eigenvalue of A with II..I =
IIAII . LetrpbeaneigenvectorcorrespondingtoAwithllrpil = 1. Then for llx] = 1,

I(Arp, rp} 1 = IAI = IIA II ~ I(Ax , x}l · o

A compact operator need not have an eigenvalue if it is not self adjoint.
178 Chapter IV. Spectral Theory ofCompact SelfAdjoint Operators

Examples: Let K: Lz([O, 1]) ~ Lz([O, 1]) be defined by

(Kf)(t) = l' !(s) ds.

Since

(Kf)(t) = 1 1
ktt , s)! (s)ds,

Where ktt , s) = 1 if 0 S S S t and zero otherwise, it follows from I1.16,

Example 3, that K is compact. However, K has no eigenvalue . For if K q; = Aq;,

l'
then
q;(s) ds = Aq;(t) a.e . (4.5)

Since Lz[O, 1] consists ofequivalence classes ofsquare integrable functions which

are equal almost everywhere, we may redefine q; so that (4.1) holds for all t . Hence

q;(t) = Aq;' (t), q;(0) = 0,

from which it follows that tp = O.

4.5 Spectral Theorem

We shall now prove the spectral theorem for compact self adjoint operators. The
proof depends on successive applications of Theorem 4.4.

Theorem 5.1 Suppose A is a compact self adjoint operator on H. There exist

an orthonormal system q;1, q;z , .. . ofeigenvectors ofA and corresponding eigen-
values }1oJ , Az , .. . such that for all x E H,

Ax = L:).t{x , q;k)q;k ·
k

If {Ak} is an infinite sequence, then it converges to zero.

It is clear from the theorem that if {q;d is an orthonormal basis for H, then the
matrix corresponding to A and {q;k} is the diagonal matrix

Proof: Let HI = H and Al = A. By Theorem 4.4, there exists an eigen-

value Al of Al and a corresponding eigenvector q;1 such that 1Iq;11I = 1 and
IAJI = IIAIIi. Now Hz = {q;d.l is a closed subspace of HI and AHz C Hz by
4.5 Spectral Theorem 179

Theorem IU8.I. Let A2 be the restriction of A to H 2. Then A2 is a compact self

adjoint operator in £ (H2). If A2 =1= 0, there exists an eigenvalue A2 of A 2 and a
corresponding eigenvector CfJ2 such that II CfJ2 11 = 1 and

Clearly, {({II , CfJ2} is orthonormal. Now H 3 = {({II , CfJ2}.l is a closed subspace of

H , H3 C H 2 and AH3 C H3 . Letting A3 be the restriction of A to H3, we have
that A 3 is a compact selfadjoint operator in £ (H3)' Continuing in this manner, the
process either stops when An = 0 or else we get a sequence {An} of eigenvalues
of A and a corresponding orthonormal set {({II , CfJ2, . .. } or eigenvectors such that

If {An} is an infinite sequence, then An -+ O.

Assume this is not the case . Since IAn I 2: IAn+ll, there exists an 6 > 0 such
that IAn I 2: 6 for all n . Hence for n =1= m ,
2 2
IIA({In - A({1m 11 = IIA n ({In - Am ({1m 11 = A~ + A; >
2
6 .

But this is impossible since {A({In} has a convergent sub-sequence due to the com-
pactness of A.
We are now ready to prove the representation of A as asserted in the theorem.
Let x E H be given.

Case 1: An = 0 for some n .

Since Xn = x - L~= I (x , ({Ik)({Ik is orthogonal to ({I I , lSi S n, the vector Xn

is in Hn . Hence
n
0= An Xn = Ax - L Adx , ({Ik)({Ik.
k= 1

Case 2: An =1= 0 for all n .

From what we have seen in case 1,

IIAX - ~ Ak (x, ({Ik)({Ikll = II Anx n II S IIAnllllxnll

= IAnl IIxn II S IAnlllxlI -+ 0,

which means that

L
00

Ax = Ak(X, ({Ik)({Ik .
k= 1

o
180 Chapter IV. Spectral Theory ofCompact SelfAdjoint Operators

4.6 Basic Systems of Eigenvalues and Eigenvectors

For convenience we introduce the following definition.

Definition: An orthonormal system rpl , rp2 , . . . of eigenvectors of A E £(H)

with corresponding non-zero eigenvalues AI, A2 , . . . is called a basic system of
eigenvectors and eigenvalues of A if for all x E H.

Ax = I:>dx,rpk}rpk. (6.1)
k

The existence of a basic system of eigenvectors and eigenvalues of any non-zero

compact self adjoint operator is ensured by the spectral theorem.
The following observations (a)-(d) pertain to a compact self adjoint operator
A which has a basic system of eigenvectors {rpn} and eigenvalues {An}.
6.1. (a) The orthonormal system {rpn} is an orthonormal basisfor 1m A = Ker A.1,
and for each x E H ,
x = Pox + L (x, rpn) rpn, (6.2)
n

where Po is the orthogonal projection onto Ker A.

Indeed, since rpn = 1 Arpn E 1m A, formula (6.1) in the definition impl ies that
sp{rpn} is dense in 1m A~ It now follows from Theorem 11.11.4 that

- -- .1
sp {rpn} = ImA = Ker A .

Thus {rpn} is an orthonormal basis for Ker A.1. Hence, given x E H, there exists a
v E Ker A.1 such that

x = Pox +v= Pox + L(v, rpk}rpk ,

k

which implies (6.2) since (v, rpd = (x, rpk).

If Ao is the restriction of A to 1m A , then the matrix corresponding to Ao and
{rpn} is the diagonal matrix (aij Ai).
(b) H has an orthonormal basis consisting ofa basic system ofeigenvectors of
A ifand only ifKer A = (0).
The statement follows from (6.2).
IfKer A =I- (0) is separable, let {1frd be an orthonormal basis for Ker A. Then
H has an orthonormal basis consisting of eigenvectors of A, namely {1frd U {rpn} .
(c) If A =I- 0 is an eigenvalue ofA, then A = Ak for some k.
4.6 Basic Systems ofEigenvalues and Eigenvectors 181

Otherwise any eigenvector v with eigenvalue Awould be orthogonal to each C{Jk

by 2.1 (b). Hence
AV = Av =L Ak(V, C{JdC{Jk = 0,
k

which is impossible since A =1= 0 and v =1= o.

It is shown in Theorem 8.1 that AI - A is invertible if A =1= 0 and A =1= Ak for all
k. In addition, a formula for (AI - A)-l is given.
(d) Each Aj is repeated in the sequence {An} exactly Pj times, where Pj =
dim Ker(AjI - A) .
To see this, let us recall from the spectral theorem that An --+ 0 if {An} is an
infinite sequence. Consequently, Aj appears only a finite number of times in the
sequence {An}. Suppose Aj = Ani ' i = 1,2, . . . , p . Then

Ker(Aj - A) = SP{C{Jni };=l '

For if this is not the case, then there exists a v =1= 0 in Ker(Aj - A) such that
v ..1 C{Jni ' 1 .:s i .:s p. But for k =1= n., 1 .:s i .:s p, v is also orthogonal to C{Jk since
Aj =1= Ak . Hence
AjV = Av = L Ak(V, C{Jk)C{Jk = 0
k

which is impossible since Aj =1= 0 and v =1= O.

The dimension of Ker(Aj - A) is called the multiplicity of Aj. In view of the
above results, every basic system of eigenvectors and eigenvalues of a compact
self adjoint operator A can be obtained as follows .
Suppose {JLI, JL2, . . .} is the set of non-zero eigenvalues of A . Choose any
orthonormal basis B, for Ker(JLj I - A) . Then B, is a finite set by (d). Let {C{Jn} be
a sequence consisting ofall the vectors in Uj B i - Since eigenvectors corresponding
to distinct eigenvalues of A are orthogonal, it follows that {C{Jn} is an orthonormal
set of eigenvectors of A . Let An be the eigenvalue of A corresponding to C{Jn . Then
{C{Jn} , {An} is a basic system of eigenvectors and eigenvalues of A .
The converse of the spectral theorem also holds .

Theorem 6.2 Suppose {C{Jd is an orthonormal system in H and {Ad is a sequence

ofreal numbers which is either a finite sequence or converges to zero. The linear
operator A defined on H by

Ax = L Ak(X, C{Jk)C{Jk
k

is compact and selfadjoint.

182 Chapter rv. Spectral Theory ofCompact SelfAdjoint Operators
Proof: A is in .cOt) by 11.2 Example 2. If {Ad is a finite sequence, then A is of
finite rank and is therefore compact. Assume An --+ O. Let An E .c(H) be the
compact operator defined by

n
Anx = L Ak(X, f{Jk}f{Jk.
k=!

Since
2
00
2
IIA-A n Il = sup L Ak(X,f{Jk}f{Jk :s sup lAd --+ 0
IIxll=! k=n+! k >n

as n --+ 00 , A is compact. For x and y in H,

(Ax, y) = LAk(X , f{Jk}(f{Jk, y} = (x , Ay) .

k

Thus A is self adjoint. o

4.7 Second Form of the Spectral Theorem

The construction above of a basic system of eigenvectors and eigenvalues sheds

light on how to approximate a compact selfadjoint operator by linear combinations
of orthogonal projections. This is done in the following theorem.

Theorem 7.1 Given a compact selfadjoint operator A E .c(H), let {J-L j} be the
set of non-zero eigenvalues of A and let Pj be the orthogonal projection onto
Ker(J-L j I - A) . Then

(i) PjPk = 0, j i= k ;
(ii) A = LjJ-LjPj,
where convergence ofthe series is with respect to the norm on .c(H).

(iii) For each x E H ,

x = POX + LPjx,
j

where Po is the orthogonal projection onto Ker A .

4.8 Formulafor the Inverse Operator 183

Proof: Let {~n}, {,A.n} be a basic system of eigenvectors and eigenvalues of A. For
each k, a subset of {~n} is an orthonormal basis for Ker(JLk - A) , say ~ni ' 1 :5
i :5 p. Then PkX = L;=l (x , ~ni )~n i' and since each An is a JLk. it follows from
(6.2) that

x = Pox +L PkX and Ax = L JLk PkX.

k k

Furthermore, Pj Pk = 0, j =1= k, since Ker (JL j - A) ..1 Ker(JLk - A) . Finally,

if {A j} is an infinite sequence, then

i:
2

I
A- k=l JLkPk 11 = su~
IIxll-l
IIAX- 'tJLkPkXI12
k=l

:5 sup L A}I(x, ~j)12:5 ~up A} -+ 0

IIxll=1 ie» ie»

as n -+ 00. D

4.8 Formula for the Inverse Operator

The following theorem gives sufficient conditions for the equation Ax - K x = Y

to have a unique solution for every y E 'H..

Theorem 8.1 Let K be a compact self adjoint operator in £('H.) with a basic
system ofeigenvectors {~n} and eigenvalues {An}. If A =1= 0 and A =1= An, n 2: 1,
then AI - K is invertible and

(A1- K)-ly =!A [y + Lk ~(y,

A - Ak
~k)~k].
Proof: Suppose (AI - K)x = y . Then
AX = Y + Kx = Y + LAkC x , ~k)~k. (8.1)
k

which implies that

or
(A - Aj)(X , ~j) = (y, ~j), 1:5 j. (8.2)
Thus, from (8.1) and (8.2) ,

x = ~ [y + ~ A ~\k (y, ~k)~k] . (8.3)

184 Chapter IV. Spectral Theory ofCompact SelfAdjoint Operators

Conversely, if x is defined by (8.3), (the series converges since the sequence

{ A~\k (Y, CPk)} is in 12) then a straightforward computation verifies that
(AI - K)x = y. We have shown that for each y E 'H, there exists a unique
x E 1i such that (AI - K)x = y and this x is given by (8.3) . Hence

(8.4)

where KIY = Lk A~\k (Y, CPk}CPk.

Now K, is in £(1i) . In fact, Theorem 6.2 shows that K, is compact. Hence
(8.4) implies that AI - K is invertible . n
We shall show in Section VIII.l, equation (1.14) that

II(AI - K)-lll = sup { _ l _

IA - J.LI
: J.L an eigenvalue of K} .
If1i is infinite dimensional and K E £(1i) is compact, then K is not invertible,
otherwise, I = K- 1K would be compact.
The following theorem characterizes 1m K when K is compact and self adjoint.

Theorem 8.2 Let K E £(1i) be a compact self adjoint operator with a basic
system of eigenvectors {CPn} and eigenvalues {An} . Given y E H, the equation
K x = y has a solution if and only if

(i) y..l Ker K

and
(ii) Ln ;'21(Y,
n
CPn}l2 < 00.

Every solution is ofthe form

x = u+ Ln -(Y
1
An
, CPn}CPn, u E KerK.

1
Proof: Suppose (i) and (ii) hold . Then XQ = Ln n (Y, CPn) is in 'H and it follows
from (2) in Section 6 that

If also Kx = Y, then u = x - XQ E Ker K and

4.9 Minimum-Maximum Properties ofEigenvalues 185

Conversely, suppose Kx = y. Then y E 1mK C Ker K.L and

L(Y , C{Jn}C{Jn = Y = Kx = LAn(X, C{Jn}C{Jn .

n n

Hence (y, C{Jn) = An(x, C{Jn) and

1 z z
L Az1(y, C{Jn}l = L I(x, C{Jn}l :5 IIxIi Z•
n n n
o
Suppose that in Theorem 8.1, A = An =1= 0 for some n. By an argument similar
to the one given in the proof of Theorem 8.2, the following result can be shown.
The equation Ax - Kx = y has a solution ifand only ify E Ker(An - K).L . In
this case the general solution to the equation is

where Z E Ker (An - K).

4.9 Minimum-Maximum Properties of Eigenvalues

Definition: An operator A E £(H) is called non-negative or positive if (Ax, x)

2: 0 for all x E H,
A compact self adjoint operator A is non-negative if and only if its eigenvalues
are non-negative. For suppose {C{Jn}, {An} is a basic system of eigenvectors and
eigenvalues of A . If A is non-negative, then Ak = (AC{Jk, C{Jk) 2: O. On the other
hand, if each Ak is non-negative, then it follows from 6.1(a) that

(Ax, x) = (L; Ak(X, C{Jk}C{Jk, Pox + L; (x , C{Jk}C{Jk)

= LAkl(x, C{Jk}lZ 2: o.
k
Let A be compact and non-negative. In the proof of the spectral theorem, a basic
system {C{Jn}, {An} of eigenvectors and eigenvalues of A was obtained by taking
AI = max (Ax,x) = (AC{JI,C{JI)
IIxll=I
AZ = max (Ax, x) = (AC{Jz , CPZ)
IIxll=1
x.J..SP('Pl}

max (Ax, x) = (AC{Jn, C{Jn) (9.1)

IIxll=1
x.J..SP{'I'1,···,'I'n - l )
186 Chapter I V. Spectral Theory ofCompact SelfAdjoint Operators

The following theorem shows that it is unnecessary to find the eigenvectors

gil, . . . , gin-I in order to determine An as in (9.1) . The result has numerous
applications-especially to numerical methods for estimating eigenvalues.

Theorem 9.1 Let A E £ (H ) be compact and non-negative and let Al 2: A2 2: ...

be the basic system ofeigenvalues of A. Thenfor each positive integer n,

min max (Ax , x).

M Dx U= 1
dim Meen-e l x L'M

Proof: We note that max{(Ax ,x) : IIxll = 1, x ..1 M} is attained. This can
be seen from Corollary 4.5 applied to the restriction of PA to M.l , where P is
the orthogonal projection onto M.l . For n = 1, the only subspace of dimension
zero is (0). Thus the formula for Al reduces to Al = maxllxll=1 (Ax, x), which we
already know.
Let {gin} be a basic system of eigenvectors of A corresponding to {An} . Given
any subspace M of dimension n - 1, Lemma 1.16.1 implies that there exists an
Xo E Sp{gil , ... , gin} such that Xo ..1 M and II xo II = 1. Suppose Xo = L:Z=I a kgik.
Since Ak 2: An , 1 ::; k ::; n,

max (Ax , x ) 2: (A xo , x o) =
Ilx n= l
xl-M
(t k= 1
Akakgik, i:
k=1
akgik)

n
2
= LAk lad 2: An llxoll = An. (9.2)
k= 1

But

An = max (A x , x ). (9.3)
DxD = 1
x L sp{I"I •...•l"n - I }

Since M is an arbitrary subspace of dimn - 1, the theorem follows from (9.2)

and (9.3). 0

Now for some simple applications of the min-max theorem. Suppose A and B
are compact non-negative operators in £(H) . Let gil, gi2, . . . and Al (A) 2: A2 (A)
2: . . . be a basic system of eigenvectors and eigenvalues of A. Let Vrl , Vr2 , ... and
Al (B) 2: A2 (B) 2: ... be a basic system of eigenvectors and eigenvalues of B.

(a) If(Ax,x)::; (Bx , x)for all x E H, then

nun max(Ax , x)::; min max (Bx, X) = An(B).

M Bx D= 1 M UxD= 1
dim M=n -I x l- M dim Meen-cl x l- M

(b) IAn(A ) - An(B )1 ::; II A - BII·

4.9 Min imum-Maxim um Prop erties ofEige nvalues 187

To see this, we note that if IIxII = 1, then

I(Ax , x} - (Bx , x}1 = I(A - B )x, x }1 :::: IIA - BII.

Hence

(Ax , x) :::: (B x , x) + II A - BII (9.4)

(Bx, x) :::: (A x , x ) + II A - B II. (9.5)

It follows from (9.4), (9.5) and Theorem 9.1 that

An( A) :::: An( B) + IIA - BII

An(B ) :::: An ( A) + IIA - BII
or, equ ivalently,

Thus if IIKj - Koll ~ 0, where each K ] , j > 0, is a compact non-negative

operator in £ ('H), then Ko is also compact and non-negative. LeUI ( Kj) :::: A2(Kj)
:::: ... be the basic system of eigenvalues of K j , j = 0 , 1, . . .. Then from the
inequality (b),An(Kj ) ~ An(Ko )asj ~ 00 .

(c) Sin ce A andB are comp ac t and non-negative, so is A + B. Let Al (A +B ) ::::

A2(A + B) :::: ... be the basic sys tem ofeigenvalues of A + B. Then

(9 .6)

Indeed, from Equations (9.1),

max (Ax, x ) (9.7)

IIx D= 1
x 10(<PI •... •<Pn - il

Am(B ) = ma x (B x , x) . (9.8)
IIxll=1
.r L {h .....'i' m- I )

Take M = Sp{<pl, . .. , <Pn-l , 1f!1 , .. . , 1f!m-d . Then dim M = n + m - j for some

j :::: 2. Since Mol = {<PI , . .. , <Pn -dol n {1f!I , ... , 1f!m-dol, it follows from (9.3),
(9.4) and Theorem 9.1 that

An(A ) +A.m(B ):::: max (A x , x }

UxD= 1
+ max(Bx
IIxll= 1
,x}
xl-M x l- M

:::: max (A
IIxU= 1
+ B )x , x) :::: An+ m- j+1 ( A + B)
x l- M
188 Chapter IV. Spectral Theory ofCompact SelfAdjoint Operators

Exercises IV

1. Find all the eigenvectors and eigenvalues of the following operators K.

(a) K~ = (L~l ~j(~)j-l +~l, ~~l' ~~I' ...); K : £2 ~ £2
(b) (Kq;)(t) =tJ~rrq;(x) cosxdx +costJ~rrxq;(x)dx;K : L2[-Jr ,Jr]
~ L2[-Jr , zr],

2. LetK : L2[-Jr, Jr] ~ L2[-Jr, zrjbe given by (Kq;)(t) = J~rr k(t-s)q;(s)

ds . Find all the eigenvectors and eigenvalues of K when K is given by
(a) K(t) = It I
(b) K(t) = sin t.

3. What are the eigenvectors and eigenvalues of an orthogonal projection?

4. If A E £(1i) and A 2 = 0, find the eigenvectors and eigenvalues of A .
5. Let P be an orthogonal projection on a Hilbert space 'H. Find all the eigen-
vectors and eigenvalues of 1- 2P .
6. Given an orthogonal projection P on a Hilbert space, prove that S = 1- 2P
if and only if S = S* and S2 = I .
7. Let K be the operator in exercises I and 2. Define A on 'H E91i by

Find the eigenvectors and eigenvalues of A .

8. Let K be a compact self adjoint operator on a Hilbert space H, If one
knows the eigenvalues and eigenvectors of K, what are the eigenvalues and
eigenvectors of A, defined on 'H E9 H , if

9. Let K be a compact operator on a Hilbert space . Let {A j } be the eigenvalues

of K* K . Find the eigenvalues of

10. Show that every IAI < 1 is an eigenvalue of the backward shift operator Se.
11. Given that (aI - A)-l exists , find x, where
(a) (aI - A)x = el and A : £2 ~ £2 is given by A~
3~2, 0, 0, . ..).
Exercises IV 189

(b) (a I - A)x cos2t, and A : L2[-Jr,Jr] --+ L2[-Jr,Jr] is given by

(AqJ)(t) = f::'n(t - s)2qJ(s) ds.

12. Let A , B be two operators on a Hilbert space and let qJ be an eigenvector of

AB corresponding to an eigenvalue A :j:. O. Prove that Brp is an eigenvector
of BA corresponding to A.
13. Prove that all non zero eigenvalues A or AB and BA are the same and that

B[Ker(AB - A1)] = Ker(BA - A1).

Is it always correct that dim Ker(AB - A1) = dim Ker(BA - A1) if A :j:. 0
is an eigenvalue of AB?
14. What conclusions can one draw in problems 12 and 13 if A = O?
15. Given A :j:. 0, prove that ifone ofthe two operators (AB-A1) and (BA -A1)
is invertible, then so is the other and

(AB - A1)-1 = ~[-I + A(BA - AI)-I B] .

A

16. Let A be defined on H EB H by

A =
B
0
00 C0)
( 000

Where B, C E £(H), B compact and self adjoint. Prove that A 2 is compact

and self adjoint.
17. Given A E £(H) , Suppose A 2 is compact and self adjoint. Show that A
has an eigenvalue and that the eigenvectors of A corresponding to )" and p.,
where A2 :j:. u.2 , are orthogonal.
18. Let A be a self adjoint operator on a Hilbert space.
(a) Prove that if A 2 is compact, then A is compact.
(b) Generalize this result to the case when A n is compact for some n EN.

19. Let A be a positive compact operator on a Hilbert space. Let {qJj } ~ I be

a basic system of eigenvectors of A corresponding to eigenvalues {Aj }~I
(AI::: A2 ::: .. .). Let L j-I = SP{qJI, .. . , qJj-d, j ::: 1, where L« = {O}.
Verify that
. I
mm (x ,x) = - .
(Ax.x) = 1 A'
x .l L j _ 1 J

20 . Let A be a compact positive operator on a Hilbert space. Let x I ::: A2 :::

::: 0 be the eigenvalues of A . Show that
190 Chapter IV. Spectral Theory ofCompact SelfAdjoint Operators

(a) max
L
min (x, x) =
(Ax .x)= 1
1.
J
dim Le j-e l X.LL

(b) min
L
max (x, x) =
(Ax .x) = !
1..
J
dim Leej XEL

21. Let P be an orthogonal projection on a Hilbert space 'H. Prove that for any
A E £(1i) , A* PA ~ A* A , i.e., A* A - A*PA is positive.
22. Define K E £(1i) byKh = LJ=1 Aj(h, CPj}CPj, where {CPj}j=1 is an ortho-
normal system and Aj ~ O. Let P be an orthogonal projection on 'H. Define
Kph = LJ=1 Aj(h , PCPj}PCPj . Prove thatAj(K p) ~ Aj(K) .
23. Let P be an orthogonal projection on a Hilbert space 'H. Prove that if
A E £(1i) is compact, then

24. Prove or disprove the statement that if P is an orthogonal projection on a

Hilbert space 'H and A E £(1i) , then PA*AP ~ A * A .
25. Let A be a compact positive operator on a Hilbert space 'H. Let the matrix
representation of A with respect to some orthonormal basis be (aij) r,o.= l ' Let
An be the operator corresponding to the orthonormal basis and th~ matrix
(bij)7.}=l' where bij = aij, 1 ~ i, j ~ n, and zero otherwise. Prove that
Aj(A n) ~ Aj(A), j = 1,2, ... .
26. Let A be a compact positive operator on a Hilbert space 'H. Find an operator
Z such that Z is compact and Z2 = A . Also, find an operator Z such that
Z is compact and Z" = A .
27. Prove that if A is a compact positive operator on a Hilbert space 'H and P
is an orthogonal projection on 'H, then

28. Does the conclusion in exercise 27 remain true if A is not assumed to be

positive?
29. Let A be a compact operator or a Hilbert space 1i and let CPl, . .. , CPn be an
orthonormal system of vectors in 'H. Prove that

g(Acpl, .. . , Acpn) ~ Al (A * A) .. . An (A * A) ,

where g denotes the Gram determinant.

30 . Prove that for a compact operator A on a Hilbert space 'H,

max g(Acpl, ... , Acpn) =Al(A*A) .. . An(A*A).

{'PI,. ·.,'Pn}
orthonormal
Exercises IV 191

31. Let A be a compact positive definite operator on a Hilbert space H. Prove

that for any orthonormal system of vectors {<Pi 17=1 in H,
n n
L(A<Pi, <PI) ::::: LAi(A) .
i=1 i= 1

32. Prove that under the conditions in Exercise 31,

Chapter V
Spectral Theory of Integral Operators

Using the theory developed in Chapter IV, we now present some fundamental
theorems concerning the spectral theory of compact self adjo int integral operators.
In general, the spectral series representations of these operators converge in the
L 2-norm which is not strong enough for many applications. Therefore we prove
the Hilbert-Schmidt theorem and Mercer's theorem since each of these theorems
gives conditions for a uniform convergence of the spectral decomposition of the
integra l operators. As a corollary of Mercer's theorem we obtain the trace formula
for positive integral operators with cont inuous kernel function .

5.1 Hilbert-Schmidt Theorem

We recall that if k is in L 2([a , b] x [a, b]) and k(t , s) = k (s , t ) a.e., then the
integral operator K defined by

(Kf)(t) = l b
k(t , s )j (s )ds

is a compact selfadjoint operator on L 2([a, b]) . Con sequently, there exists a basic
system of eigenvectors {cpkl and eigenvalues {A.k} of K (K i= 0) such that

(1.1)

Convergence of the series means convergence with respect to the norm on

L 2([a , b]), i.e.,

For many purposes this type of convergence is too weak, whereas uniform conver-
gence in (1.1) is most desirable. The problem is to give sufficient, but reasonable,
conditions so that for each jEL2([a , b]) , the series in ( 1.1) converges uniformly
on [a, b]. The Hilbert-Schmidt theorem which follows now is very useful in this
regard.
194 Chapter V. Spectral Theory ofIntegral Operators

Theorem 1.1 Let k be a Lebesgue measurable function on [a, b) x [a, b] such

that
k(t , s) = k(s, t) a,e.
and
suplb Ik(t, s)12ds < 00 .
t a
Let {({In), {An} be a basic system ofeigenvectors and eigenvalues ofK , where K is
the integral operator with kernel function k. Thenfor all j E L2([a, b)) ,

I b
k(t, s)j(s)ds = ~ Ak (l b
j(S)({Ik(S)dS) ({IkCt) a.e.

The series converges absolutely and uniformly on [a, b].

Proof: By Schwarz's inequality,

n
L IAj(f, ({Ij)({Ij(t)1
j=m

(1.2)

Now

Aj({lj(t) = (K({Ij)(t) = I b
ktt , s)({Ij(s) ds = (k t , iPj),

where k t (s) = k(t, s) . Therefore, since k, is in L2([a, b)), it follows from Bessel's
inequality and the hypotheses that

L IAj({lj(t)1 2
= L I(kr, iPj)1 2 ~ IIk ll 2 t

I
j j
b
2
= Ik(t,s)1 ds s suplb Ik(t,s)1 2 ds
a t a
= C2 < 00. (1.3)

Let e > 0 be given . Since L j l(f, ({Ij)1 2 ~ 1If11 2 , there exists an integer N such
that ifn > m ~ N, then
n
L l(f, ({Ij)1
2
s s2. (1.4)
j=m
Thus from (1.2) , (1.3), and (1.4)
n
L IAj(f, ({Ij)({I j(t)1 s c« n > m ~ N, t E [a, b).
j=m
5.1 Hilbert-Schmidt Theorem 195

Hence L j Aj(f, ({Jj }({Jj(t) converges absolutely and uniformly on [a , b] by the

Cauchy criterion. Since this series also converges to (Kf)(t) with respect to the
norm on L 2 ( [a , b D, it follows that (Kf) (t) is the limit of the series for almost
every t . D

Disregarding the validity of termwise integration of the series in the above

theorem, we formally write

la
b
k (t , s) f (s ) ds = L
k
la
b
Ak({Jk(t )i/Jk (s ) f (s ) ds

= l a
b
L Ak({Jk(t)i/Jk (s ) f (s ) ds.
k

Since f is arbitrary in L 2([a, bD , it is not unreasonable to expect that under the

appropriate conditions,

k (t , s ) = L Ak({Jk (t)qik (S) a.e. (1.5)

k

The result we shall prove in Section 3 is Mercer 's theorem which states that
if k is continuous and each Ak is non-negative, then the series (1.5) converges
uniformly and absolutely on [a , b] x [a, b] to k (t , s) .
Now let us prove another fact. We know from IY. 6.l (b) that L 2([a , bD has an
orthonorm al basis ({J\, qJ2, . .. consisting of eigenvectors of K . Since cf> ij (t , s) =
({J; (t ) ({J j (s) form an orthonormal basis for L 2([a, b] x [a , b D,

00

k = L (k , cf> ij}cf>ij (1.6)

; ,j= \

and

(k , cf> ij) = ll b b
k(t , s)qi; (t) ({Jj (s) ds dt

= l [l b
qi;(t)
b
k(t, s)({Jj (s ) dsJ dt

= (K ({Jj , ({J;) = Aj(({Jj , ({J;) = AjDji . (1.7)

Thus ( 1.1) and (1.2) imply

We have therefore obtained the following result.

196 Chapter V Spectral Theory ofIntegral Operators

Theorem 1.2 Suppose k E Lz([a , b] x [a, b]) and k(t, s) = k(s, t) a.e. .If{Akl is
the basic system ofeigenvalue ofK , where K is the integral operator with kernel
function k, then

The formula is the continuous analogue of the fact that if (aij) is an n x n self
adjoint matrix with eigenvalues AI, .. . , An, counted according to multiplicity, then
n n
L laijlZ = L A7·
;,j= 1 ;=1

Here ktt , s) is replaced by k(i, j) = aij and the intergral is replaced by a sum.

5.2 Preliminaries for Mercer's Theorem

Lemma 2.1 If k is continuous on [a, b] x [a, b] and

ll b b
ktt , s)f(s)f(t) ds dt 2: 0

for all f E Lz([a, b]), then

k(t ,t) 2: Oforallt E [a,b].

Proof: The function k(t, t) is real valued . Indeed, the integral operator with
kernel function k is positive and therefore self adjoint. Hence k(t , t) = k(t , t) .
Suppose keto, to) < 0 for some to E [a, b]. It follows from the continuity of k that
ffi kit ; s) < 0 for all (r, s) in some square [c, d] x [c, d] containing (to, to). But
then for g(s) = 1 if s E [c, d] and zero otherwise,

o~ llb b
kit , s )g (s )g (t ) ds dt = ffi II d d
k(t, s) ds dt < 0

which is absurd . Hence ktt ; t) 2: 0 for all t E [a, b] . o

Lemma 2.2 .lfk is a continuous complex valuedfunction on [a, b] x [a, b], then
for any tp E
Lz([a,b]),

h(t) = l b
k(t ,s)cp(s) ds

is continuous on [a, b].

5.3 Mercer's Theorem 197

Proof: By Schwarz's inequality,

Ih(t) - h (to)1 ~ l b
Ik( t , s) - k (to, s)llcp(s)1 ds

(l
I
b
~ IIcpli Ik(t , s) -
z
k tt«, s)l dS) 1

This last inequality, together with the uniform continuity of k on [a, b] x [a , b] ,

imply the theorem. 0

Dini's Theorem 2.3 Let Un} be a sequence ofreal valued continuousfun ctions on
[a, b]. Suppo se /J (t ) ~ h (t ) ~ . . . for all t E [a, b] and f (t) = limn->oo f n(t )
is continuous on [a, b]. Then U n} converges uniformly to f on [a, b].

Proof: Given e > 0, let

Fn = {t : f (t ) - fn(t) ~ s}, n = 1, 2, .. . .

Then Fn ::> Fn + l and it follows from the continuity of f - [« that F; is a closed

set. Furthermore, the pointwise convergence of Un } impl ies that n~l Fn = 0 .
Hence [a , b] C U~l F~ , where F~ is the comple ment of r; Since [a, b] is closed
and bounded and each F~ is an open set, there exists an integer N such that
[a , b] C U~= l F~. Therefore

n r, =
N
FN = 0.
n= l

Th is implies that for all n ::: N and all t E [a , b] ,

If (t ) - f n(t )! = f( t) - fn (t) ~ f(t) - f N(t ) < e.

Hence U n} converges uniforml y to f on [a, b]. o

5.3 Mercer's Theorem

We are now ready to prove the series expansion for k which was discussed in
Section 1.

Theorem 3.1 Let k be continuous on [a, b] x [a, b]. Supp ose that f or all f E
L z([a , b]),

ll b b
k (t , s ) f (s ) j(t) ds dt ~ o.
198 Chapter V. Spectral Theory ofIntegral Operators

if {C{ln},
{An} is a basic system of eigenvectors and eigenvalues of the integral
operator with kernelfunction k, thenfor all t and s in [a, b],

k(t,s) = LAjC{lj(t)(jJj(s).
j

The series converges absolutely and uniformly on [a, b] x [a, b].

Proof: Let K be the integral operator with kernel function k. It follows from
the hypotheses that K is compact, positive and Aj = tKip], C{lj) ~ O. Schwarz's
inequality applied to the sequences {AC{lj(t)} and {AC{lj(s)} yields

(3.1)

First we establish that for all t

LAjIC{lj(t)1
2
s m:xk(s,s) . (3.2)
j

Let
n
kn(t, s) = k(t, s) - LAjC{lj(t)(jJj(s).
j=!

Since each <P j is an eigenvector of K, it follows from Lemma 2.2 that <P j is
continuous which implies that k n is continuous. A straightforward computation
verifies that

inr inr kn (t, s)f(s)f(t) ds dt =

b b
n 2
(Kf, f) - LA j l(f, C{lj }1
a a j=!

= LAjl(f, C{lj}1 2 ~ o.
j>n

Hence we have from Lemma 2.1 that for each t,

n
Os kn(t, t) = k(t, t) - LAjIC{lj(t)1 2 •
j=!

Since n was arbitrary, inequality (3.2) follows . For fixed t and e > 0, (3.1) and
(3.2) imply the existence of an integer N such that for n > m ~ N
n
L Aj IC{lj (t)C{lj (s)1 s ec. s E [a, b] ,
j=m
5.3 Mercer 's Theorem 199

where C 2 = maXxE[a,b] k (x , x). Therefore we have from the uniform Cauchy

criterion that L j Aj ({Jj (t)i{Jj (s) converges absolutely and uniforml y in s for each t .
The next step in the proof is to show that the series converges to k (t , s). Once
this is done, an applic ation of Dini's theorem will conclude the proof.
Let

k (t , s) = LAj ({Jj (t )i/Jj (s ).

j

For / E L 2([a , b]) and t fixed, the uniform convergence of the series in s and the
cont inuity of each i{J j imply that k(t , s) is continuous as a funct ion of s and

I b
[k(t , s) - k (t, s )]/ (s) ds = (Kf)( t ) - "2;. Aj (f, ({J j}({Jj (t ).
J
(3.3)

1.
If / E Ker K = 1m K.L , then since ({Jj = K({Jj Elm K , we have that Kf = 0
J
and (f, ({Jj) = O. Thus the right side of(3.3) is zero . If T = ({J0.. then the right side
of (3.3) is A;({J; (t) - A;({J; (t ) = O. Thus for each t , k (t , ) - k (t , ) is orthogonal
to L 2([a , b]) by IY.6.1(a). Hence, k (t , s) = k (t , s) for each t and almost every s.
But then k (t , s) = k (t , s) for every t and s since k (t , ) and k (t , ) are continuous.
We have shown that

k (t , s) = k (t , s) = L Aj ({Jj (t) i{J j (s) .

j

In part icular,
2
k (t , t ) = L Aj l({Jj (t ) 1 .
j

The partial sums of this series form an increasing sequen ce of continuous func-
tions which converges pointwise to the continuous function k (t , t ) . Dini's theorem
asserts that this series converges uniforml y to k (t , t). Thus given e > 0, there
exists an integer N such that for n > m 2: N ,
n
L Ajl({Jj(t)1 2
< e
2
for all t E [a, b] .
j=m

This observation, together with (3.1) and (3.2), imply that for all n > m 2: Nand
all (t, s) E [a, b] x [a, b],
n
L Ajl({Jj(t)i{Jj(s) 1 ::::: «:
j =m

Hence L j Aj({Jj (t) i{Jj (s) converges absolutely and uniformly on [a , b] x [a , b].
200 Chapter V. Spectral Theory ofIntegral Operators

Mercer's theorem is not true if we remove the assumption that K is positive.

To see this, let us consider the example in IV.I where k(t, s) = h (t - s) and h is
continuous. If the conclusion in Mercer's theorem were to hold, then h would be
the uniform limit of its Fourier series. However, there are well known examples
of continuous functions where this is not the case. 0

5.4 Trace Formula for Integral Operators

The trace formula for finite matrices states that if (aij) is an n x n matrix
with eigenvalues AI, . . . , An, counted according to multiplicity, then LJ=I ajj =
LJ=I Al : The following theorem is the continuous analogue of the trace formula.
Theorem 4.1 Let k be continuous on [a, b] x [a, b]. Suppose that for all

ll
f E L2([a, b]) ,
b b
k(t, s)f(s)j(t) ds dt 2: O.

If K is the integral operator with kernel function k and {A j} is the basic system of
eigenvalues of K, then

L>j =
J
l b
k(t, t) dt.

Proof: Let {q; j } be a basic system of eigenvectors of K corresponding to {A j }. By

Mercer's theorem, the series

ktt , t) = L>j Iq;j (t)12

j

converges uniformly on [a, b]. Hence

l b
k(t,t) dt = ~:~:>jllq;jIl2 = ~Aj.
J J
o
If, more generally, k is in L2([a, b] x [a , b]) and k(t , s) = k(s, t) a.e., then
L j Aj may diverge whereas L j A] < 00 by Theorem 1.2.

Exercises V

1. Which of the integral operators on L2[-7T,7T] with the following kernel

functions are positive and which are not?
(a) k(t, s) = cos(t - s)
(b) k(t, s) = (t - s)3
(c) ktt ; s) = (t 2 - 2s 2)e t +s .
Exercises V 201

2. Prove that if a bounded integral operator with kernel function k is positive

on Lz[a, b], then it is positive on Lz[c, d], where [c, d] C [a , b]. Is the
converse true?
3. For c E (a , b), suppose an integral operator K with kernel function k is
positive on Lz[a, c] and Lz[c, b]. Is it true that k is positive on Lz[a, b]?
4. Let k be a compact positive integral operator on Lz[O, 1] with a kernel
function defined on [0, 1] x [0, 1]. Let K' be the integral operator with the
same kernel function restricted to [1},~] x [1},~] C [0, 1] x [0, 1]. Prove
that An(K') S An(K), where An(K) denotes the nth eigenvalue of K (in the
decreasing sequence of eigenvalues).
5. Under what conditions does a kernel h (t - s) defined on [-1 , 1] x [-1 , 1]
define a positive bounded integral operator on L z[0, I]? Hint: Example IV.I.
6. Using Mercer's theorem, what can you say about the convergence of the
Fourier series of a continuous periodic function with period 2n?
7. Let k(Z)(t,s) = f: k(t ,x)k(x ,s)dx, where k is a self adjoint Hilbert-
Schmidt kernel, i.e. k(t, s) = k(s, t) a.e. and k E Lz([a , b] x [a , b]).
Prove that if {A j } ~ I is the basic system of eigenvalues of the operator
(K cp)(t) = f:k(t, s)cp(s) ds, then L:~I Aj = f: f:
Ik(Z) (t , s) lz ds dt.
8. Letg be a continuous complex valued function on [a, b] x [a, b] . Prove that

k(t,s) = l b
g(x,t)g(x ,s)dx

satisfies the hypotheses of Mercer's theorem.

9. Let k be defined as in exercise 8 and let K be the corresponding integral
operator. Show that

I:>j(K) =
J
ll b b
z
Ig( s, t)l ds dt.

10. Generalize the results in exercises 8 and 9 to the case where g(t, s) is a

°
Hilbert-Schmidt kernel which is Lz-continuous. This means that for any
e > 0, there exists a 0 > such that

It I - tzl < 0 implies l b

z
Ig(tl , s) - g(tz , s)l ds < sand

ISI-szl < 0 implies l b

Ig(t ,sl) -g(t,sz) IZdt < E,

II . Prove that the conclusions of Mercer's theorems remain valid as long as

all but afinite number of the eigenvalues of the self adjoint operator (with
continuous kernel) are positive.
202 Chapter V. Spectral Theory ofIntegral Operators

12. Let K be an integral operator on Ljtc , b) withkernelk(t, s) = L:j=1 a j (t ) x

bj(s), aj , bj E L2([a , b]) , 1 :s j :s n. Prove that K can be written in the

I
form
b
(Kq;)(t) = A(t)B(s)q;(s) ds,

where A and B are matrix valued functions.

13. Let K be the operator in exercise 12. Using the integral representation of K
obtained in the exercise, determine those A for which AI - K is invertible
and find (AI - K)-I.
14. Let K be the integral operator on L2[0 , 1] with kernel function k(t, s) =
min(t , s), 0 :s t, s :s 1. Show that K is positive and find its eigenvalues and
eigenvectors.
15. Let K be the integral operator on L2 ([0,00]) with kernel function

~
k(x , y) = ~
1
2 [cos (k + ..
l)x smky - sm (k + ]
l)x cosky .
k=o(k+l)

Show that K has no eigenvalues.

Chapter VI
Unbounded Operators on Hilbert Space

The theory developed thus far concentrated on bounded linear operators on a

Hilbert space which had applications to integral equations . However, differential
equations give rise to an important class of unbounded linear operators which are
not defined on all of L2([a, b]). In this chapter an introduction to unbounded
operators is presented which includes the spectral theorem for the Sturm-Liouville
operator. Simple examples of the spectral theory of unbounded self adjoint oper-
ators are also given. For a more detailed theory the reader is referred to [G],
[GGKl], [K] and [DS2].

6.1 Closed Operators and First Examples

It is indeed fortunate that essentially all the important differential operators form a
class ofoperators for which an extensive theory is developed. These are the closed
operators which we now define.

Definition: Let H I and H 2be Hilbert spaces . An operator A with domain V (A) ~
H I and range in H 2 is called a closed operator if it has the property that whenever
{xn} is a sequence in V (A ) satisfying Xn -+ x in HI and AXn -+ y in H 2, then
x E V (A ) and Ax = y .
Clearly, a bounded linear operator on a Hilbert space is closed.

We now give an example of an unbounded closed operator.

Let H = L2([0, 1]). Define the differential operator A by V (A ) = (f E Hlf
is absolutely continuous, I' E H , f(O) = OJ, and Af = f",
The operator A is unbounded since fn (t) = t", n = 1,2, .. . is in V(A),
2
IIfn 11 = fo l 2ndt
t = 2n~1 ~ 1 and

The operator A is closed. To see this, we first note that Ker A = {OJ and
1m A = H. Indeed, given g E H, take f (t ) = f~ g(s)ds. (Recall that
L2([0, 1]) ~ L I([O , 1])). Then f E V (A)andAf = g. DefineA - Ig = f , g E H.
204 Chapter VI Unbounded Operators on Hilbert Space

The operator A-I is a bounded linear operator on H with range V(A) since the
Schwarz inequality gives

Therefore

Hence IIA - I II :s 1. Suppose

!n ----+ ! ,!n E V(A) and A!n ----+ h E H.

Then jj, =A-IA!n ----+ A-Ih. Hence! = A-Ih E V(A) and A! = hwhich
shows that A is closed.

Definition: Let HI and H2 be Hilbert spaces and let A be a linear operator with
domain V(A) ~ HI and range 1m A ~ H2 . Suppose there exists a bounded linear
operator A - I mapping H2 into HI with the properties AA -I Y = Y for all y E H2
and A -I Ax = x for all x E V(A) . In this case we say that A is invertible and A -I
is the inverse of A . Note that if A is invertible, then Ker A = {OJ and 1m A = H2.
Clearly, A cannot have more than one inverse.

The differential operator in the above example is invertible and its inverse is
(A-Ig)(t) = f~ g(s)ds,g E L2([0, 1]).
The notation A (HI ----+ H2) signifies that A is a linear operator with domain in
Hilbert space HI and range in Hilbert space H2 .

Theorem 1.1 Let A(HI ----+ H2) be invertible. Then A is a closed operator.

Proof: Suppose Xn ----+ x , Xn E V(A) and AXn ----+ y. Since A - I is bounded

onH2,xn = A-IAxn ----+ A-Iy. Therefore X = A-Iy E V(A) and Ax =
AA-Iy = y. 0

We shall show in Theorem XII.4.2 that the converse theorem also holds, i.e., if
the operator A is closed with Ker A = {OJ and 1m A = H2 , then A is invertible.

6.2 The Second Derivative as an Operator

One of the main motivations for the development of a theory of integral operators
is that certain differential equations with prescribed boundary conditions can be
6.2 The Second Derivative as an Operator 205

transformed into equivalent integral equations. As an illustration, let us begin with

the simple boundary value problem .

- y"(x ) = f(x) (2.1)

y (O) = y(1) = 0, (2.2)

where f is a function in L2([0, 1]). To find a solution, we integrate both sides

of (2.1) twice and obtain

y(x) = - l
x
1 t
f(s)dsdt + C1X + C2, (2.3)

where

C2 = y(O) = 0 and C1 = . : f(s)dsdt . (2.4)

Interchanging the order of integration in (2.3) and (2.4) yields

y(X)=-l
x
1x
f(s)dtds+x 1 11 1
f(s)dtds

= 1 x
(s - x )f(s) ds +1
1
x(1 - s)f(s) ds. (2.5)

Hence

y(x) = 1 1
g (x , s)f(s ) ds, (2.6)

where
s(l - x ) , 0:::: s :::: x
g(x , s ) = { x(1 _ s) , x:::: s :::: 1. (2.7)

Conversely, if y is given by (2.6), a straightforward computation verifies that y

satisfies (2.1) and (2.2) a.e. The function g is called the Green's function corres-
ponding to the boundary value problem.
Let us consider the above result from the point of view of operator theory.
We want to express the differential expression - y" with the boundary conditions
(2.2) as a linear operator. The action of the operator is clear. However, we must
define its domain. To do this, we note that (2.3) implies that the derivative y' is
an indefinite integral or, equivalently, y' is absolutely continuous (cf. [RD. An
important property of absolutely continuous functions, which we shall use later,
is that the usual "integration by parts" formula holds for the integral of f g, where
f is absolutely continuous and g is Lebesgue integrable.
Let the domain V (L ) of L be the subspace of L2([0 , 1]) consisting of those
complex valued functions y which satisfy (2.2), have first order derivatives which
206 Chapter VI. Unbounded Operators on Hilbert Space

are absolutely continuous on [0, 1] and have second order derivatives which are
in L2 ([0 , 1]). Note that v" (x) exists for almost every x since y' is absolutely
continuous. Define Ly = -y" . It is clear from (2.2) that L is injective.
Take G to be the integral operator with kernel function g defined in (2.7). Since
g is continuous on [0, 1] x [0, 1] and g (x , s) = g (s, x), G is a compact selfadjoint
operator on L2 ([0 , 1]). From the discussion above, y = Gf satisfies (2.1) a.e. for
every f E L2([O, 1]). Thus
LGj= f. (2.8)
Since L is also injective, we have that L is invertible with L -1 = G . Thus L is
a closed linear operator by Theorem 1.1. If follows that rp is an eigenvector of L
with eigenvalue Aif and only if rp is an eigenvector of G with eigenvalue t.
Thus,
since G is compact self adjoint and Ker G = (0), L2 ([0, 1]) has an orthonormal
basis consisting of eigenvectors of L. The eigenvalues of L are those real A for
which
y" +Ay = 0 (2.9)
and
y(O) = y(l) = 0 (2.10)
has a non-trivial solution. Since the general solution to (2.9) is

y = ax +b, A = 0
Y = acos~x +bsin~x, A > 0
Y = ae A x + be-Ax, A < 0

it follows from the boundary conditions (2.10) that the eigenvalues are )..
n 2 rr 2 , n = 1, 2, ... , with bn sin nitx, bn I- 0 the corresponding eigenvectors.
The eigenvectors ..fi sin ntt x , n = 1, 2, . . . , therefore form an orthonormal basis
for L2([0, 1]).
Similarly, if we change the domain of L by replacing the boundary conditions
(2.10) by y(-rr) = y(rr), y'(-rr) = y'(rr), then the eigenvalues of L are those A
for which the boundary value problem

v" +Ay = 0

y(-rr) = y(rr), y'(-rr) = y'(rr)

has a non trivial solution. It follows that A = n 2 , n = 0, 1, ... are the eigenvalues
of L with an cos nt + bn sin nt (Ian 12 + Ibn 12 I- 0) the corresponding eigenvectors.
Thus , {_1_
.Jiii ' cost nt sin nt }OO is an orthonormal system ofeigenvectors of L which
.j1i , .j1i n = 1
we know forms an orthonormal basis for L2([ -rr, rr]) .

6.3 The Graph Norm

A closed linear operator in a Hilbert space can be considered as a bounded operator

acting between two Hilbert spaces. This is achieved by introducing an appropriate
6.3 The Graph Norm 207

norm on its domain. We need the following preliminary results. Let 1t1 and 1t2
be Hilbert spaces. The product sp ace H I x 1t2 is the set of all ordered pairs (x , y)
with x E 'H I and y E 1t2. The operations of addi tion + and scalar multiplication
. are defined in the usual way by

(XI, Yd + (X2, Y2) = (XI + X2 , YI + Y2)

a · (x , y) = (ax , a y )

Under these operations , H I x 1t2 is a vector space. It is easy to see that 'H I x 1t2
is a Hilbert space with respect to the inner product

(3.1)

and corresponding norm

II (x , Y) II = ( x , y ), (x, y) }1/2 = (lIxll 2 + II YIl 2)1/ 2. (3.2)

Given operator A (1t1 --+ 1t2), the grap h G (A ) of A is the subspace of1t1 x 1t2
consisting of the ordered pairs (x , Ax ) , x E D (A ) . It follow s readil y from the
defin ition of the norm on 'H 1 X 1t2 that A is a closed operator if and only if its
graph G ( A ) is a closed subspace of1t1 x 1t2.
On the domain D (A ) we define the inne r product ( , }A as follow s:

(u , V}A = (u, v) + (A u, Av ), u , v E D (A ).

The corresponding norm II II A is

wh ich is exactl y the norm of the pair (u, Au ) E G (A ) . We call the norm III1 A
the gr aph norm on D ( A ) . If A is closed, then the inner product space D (A ) with
norm II II A is complete. Indeed, suppose IIxn - Xm II A --+ 0 as n , m --+ 00 . Then
II x n- x mll--+ OandllAx n-Axmll --+ Oasn ,m --+ 00 . Hence there exists x E 1t1
and y E 1t2 such that Xn --+ x and AXn --+ y. Since A is closed, x E D(A) and
A x = y . Therefore

The operator A mapping (D A : II II A) into 1t2 is bounded since

Th is graph norm allows one to reduce theorems and problems for closed operators
to corresponding results for bounded linear operators. For exampl e, let the closed
operator A have the property that it is invert ible when con sidered as a (boun ded)
208 Chapter VI. Unbounded Operators on Hilbert Space

map from the Hilbert space (V(A) : III1A) onto 1h Suppose B(HI -+ H2) has
domain V(B) ;2 V(A) and satisfies

IIBxll ~ C(lIxll 2 + IIAxIl 2) 1/ 2 = CIIxIlA, X E V(A).

Then for IAI sufficiently small, A + AB is invertible . This follows from

Corollary II.8.2 applied to the bounded operators A and B on the Hilbert space
(V(A) : III1A).
Let L be the second order differential operator described in Section 2. Then , as
we have seen, L is a closed operator. With respect to its graph norm,

IIf11L =
(
I
1If(t)12dt
) 1/2
+
(
1
I
Ifl/(t)1 2dt
) 1/2
, f E V(L).

The operator L is a bounded linear operator on the Hilbert space (V(L) : II lid
with range L2([0, 1]).

6.4 Adjoint Operators

In this section we extend the concept of the adjoint of a bounded operator to the
adjoint of an unbounded operator.

Definition: Let A(HI -+ H 2) have domain V(A) dense in HI, i.e., the closure
V(A) = HI . We say that A is densely defined. The adjoint A*(H2 -+ HI) is
defined as follows :
V(A*) = {y E H21 there exists z E HI such that (Ax , y) = (x, z)
for all x E V(A)}
Obviously 0 E V(A*). The vector z with the above property is unique since
(x, z) = (x, u) for all x E V(A) implies (v - z) 1.. V(A) = HI . Hence
v - z = O. We define A*y = z. Thus
(Ax, y) = (x , z) = (x, A *y) , x E V(A), y E V (A*).

Note that if A is bounded on HI, then A * is the adjoint defined in Section II.11.
It is easy to see that A * is a linear operator.

Examples: Let H = L2([0 , 1]). Define A(H -+ H) by

V(A) = {J E HI fis absolutely continuous on [0,1], r E H,

f(O) = f(l) = O}, Af = 1'. (4.1)

The space Cgo ([0 , 1]) of infinitely differentiable functions which vanish outside a
closed subinterval ofthe open interval (0, 1) is well known to be dense in L2([0 , 1])
with respect to the L2-norm. Since Cgo([O, 1]) ~ V(A), we have that A is densely
defined. We shall show that A* = S, where
6.4 Adjoint Operators 209

V (S ) = {g E 'H [g is absolutely continuous on [0, 1], g' E 7t}

Sg = - s' (4.2)

Suppose g E V (A *) and A *g = h . Then for f E V (A ),

(Af, g) = 1 1
!,(t)g(t) dt = (f,A*g) = 1 1
f (t )h (t )dt. (4.3)

Since f EV (A ), f(O) = f(1 ) = O. Integration by parts yields

1 1
f (t ) h (t ) dt = -1 1
(H(t) + C )f'(t )dt , (4.4)

where C is an arbitray constant and H(t ) = J~ h(s)ds. Then by (4.3) and (4.4),

0= 1 1
(g (t ) + H (t ) + C)f'(t)d~ f E V (A ). (4.5)

it
Let
f o(t ) = (g(s) + H (s ) + Co) ds , (4.6 )

where Co is chosen so that f o(1) = O. Then fo E V (A ) and iffollows from (4.5)

and (4.6), with f replaced by f o, that

0= 1 1
Ig(t) + H (t ) + Col2dt

Hence get) = -H (t) - Co = - J~ h(s)ds - Co, which shows that g is abso-

lutely continuous on [0, 1] and g' = - h E 'H, Hence g E V (S ) and (Af, g ) =
(f, - s'; = (f, Sg) by (4.3). Therefore V (A *) S; V (S ) and A*g = Sg, g E
V (A*).
It remai ns to prove that V (S ) S; V(A*) . Given v E V (S ) and u E V (A ),

(Au , v) = 1 1
u' (t )v (t )dt = -1 1
u (t )v' (t )dt = (u ,Sv)
Therefore v E V(A *) and A*v = Sv = -v'.

Theorem 4.1 Let A (7t1 ~ 7t 2) be densely defined. Then its adjoint A* is a

closed operator.

Proof: Suppose Yn ~ Y, Yn E V (A *) and A *Yn ~ v. Then for all x E V (A ),

(Ax, y) = lim
n
(Ax , Yn) = lim (x , A*Yn) = (x , v)
n

Hence Y E V (A *) and A*y = v . o

210 Chapter VI. Unbounded Operato rs on Hilbert Space

Since the operator S in the above example is A *, Theorem 4.1 shows that S is a
closed operator.

Theorem 4.2 Let A('HI -+ 'H2) be a densely defined invertible operator. Then
A * is invertible and
(A*)-I = (A- I)*

Proof: Given v E V(A *) and y E 'H2, it follows from the boundedness of A-I
that
(A-I)*A*v,y) = (A*v,A-Iy) = (v,AA - Iy) = (v,y)
Since y is arbitrary in 'H2,
(A- I)* A*v = v, v E V(A*) . (4.7)

Now for x E V(A) and w E 'HI, it follows from the boundedness of (A- I)* on
'HI that
(Ax , (A -I)*w) = (A -I Ax, w) = (x , w) .
Thus , by definition of A* , (A-I) *w E V(A*) and
A*(A-I)*w = w . (4.8)

Since (A -1)* E L('HI , 'H2), we have from (4.7) and (4.8) that A* is invertible and
(A*)-I = (A- I)* E L('H I, 'H2)' 0

Definition : A densely defined linear operator A ('H -+ 'H) is called selfadjoint

if A* = A.
The following result is an immediate consequence of Theorem 4.1.

Corollary 4.3 A selfadjoint operator is closed.

Theorem 4.4 A densely defined invertible operator A ('H -+ 'H) is selfadjoint if

and only if its inverse A -I is selfadjoint.

Proof: Suppose A is self adjoint. Then by Theorem 4.2 (A- I)* = (A*)-I =
A -I , i.e., A -I is self adjoint. Conversely, if A -I is self adjoint, then V(A *) =
Im(A*)-1 = Im(A- 1)* = ImA- 1 = V(A) . Given y E V(A*) = V(A),

(A- 1)*A*y = (A*)-I A*y = y .

and
(A- 1)*Ay = (A-1)Ay = y .
Since (A- 1)* is 1 - 1, Ay = A*y. o
It was shown that the second order differential operator L discussed in Section 2
is invertible with inverse L -I, a compact self adjoint operator.
Hence L is self adjoint by Theorem 4.4.
6.5 Sturm-Liouville Operators 211

6.5 Sturm-Liouville Operators

As an application of Theorem IV.5.1, we consider the following Sturm-Liouville

system.
A Sturm-Liouville system is a differential equation of the form

dx dY)
d ( p(x) dx + q(x)y = f(x) (i)

together with boundary conditions

al y(a) +a2y'(a) = 0
bly(b) +b2y'(b) = 0, (ii)
where a., b, are real numbers with ar +ai i- 0, br +bi i- O. The system is used to
describe, for example, motions of vibrating strings, elastic bars, and membranes.
For more details see [CH] Chapter V.
Suppose p , p' and q are continuous real valued functions and p(x) i- 0 for
each x E [a, b] . We now give some important properties of the eigenvalues and
eigenvectors of the corresponding Sturm-Liouville operator L. In order to do so,
we use some facts from differential equations and the theory ofcompact selfadjoint
operators.
Define the linear differential operator L as follows: The domain V(L) of L
consists of those functions y which satisfy (ii), have first order derivatives which
are absolutely continuous on [a, b] and have second order derivatives in L2 ([a , bD.
Let

d ( p(x) ddxY)
Ly = dx + q(x)y.
Assume that zero is not an eigenvalue of L, i.e., L is injective. Now for f = 0,
there exist real valued functions YI ¢ 0 and Y2 ¢ 0 such that YI , Y2 satisfy (i),
Y;'and Y~ are continuous, YI satisfies the first condition in (ii), and Y2 satisfies the
second condition in (ii) ([DS2]), XIII.2.32). Let

W(t) = det (YI(t) Y2(t»)

Y; (t) Y~ (I) ,
which is called the Wronskian of (YI, Y2). A straightforward computation verifies
that (p W) ' = O. Thus p W is a real valued non-zero constant function. Let
(p W) -I = c. We shall show that the function g given by

(x s) = {cY2(S) YI(X) a:::: x:::: s (5.1)

g, CYI(S)Y2(X) s:::: x :::: b
is the Green's function corresponding to L , i.e., for each f E L2 ([a , b]) ,

y(x) = I b
g(x , s ) f (s )ds

lies in V(L) andLy = f.

212 Chapter VI. Unbounded Operators on Hilbert Space

Let G be the integral operator with kernel function g . Since g is a continuous

real valued function on [a, b] x [a, b] and g(x , s) = g(s, x) , G is compact and
self adjoint. We now prove that

LGf = f, f E L2([a , b]) . (5.2)

First we show that Y = Gf is in Vel) . It is clear from the definition of g that

(5.3)

where

Yl (x) i X

CYI (s)f(s) ds,

Y2(X) = l b
cY2(s)f(s) ds.

Differentiation of both sides of (5.3) yields

y' = yi Y2 - YICY2f + Y2cyI! + Y~Yl

= yi Y2 + Y~Yl a.e. (5.4)

Actually, (5.4) holds for all x . To see this, let h = yi Y2 + Y~Yl and let

y(x) = yea) + i X

h(t) dt .

Now y' = v' a.e. and yand yare absolutely continuous. This follows from the
absolute continuity of Yi and Yi , i = 1, 2. Hence

(y- y )(x ) = i (y-

X

y)'(t) dt = 0,

or y = y' , Thus (5.4) holds for all x which implies that y' is absolutely continuous.
Moreover,

Finally, since Yl (a) = 0 and Yl satisfies the first equation in (ii),

al yea) + a2Y' (a) = (alY2(a) + a2 y~(a»Yl (a) + (a lYl (a) + a2yi (a» Y2(a)
=0.

Similarly,
6.5 Sturm-Liouville Operators 213

We have therefore shown that y is in Vel) . It remains to prove that Ly = f . It

follows from (5.4), the definition of c and properties of YI and Y2 that
(py')' = (py; Y2 + PY;YI)' = (Py;)'Y2 + (py;)'YI + pC(YIY~ - Y2Y;)f
= -qY 1Y2 - qY2YI + f
= -qy+ f a.e.
Thus Ly = f. Hence we have established (5.2). Since it is assumed that L is
injective, we have from (5.2) that L is invertible with L -I = G which is compact
and self adjoint. Each eigenvalue X of G is simple, i.e., dim Ker (H - G) = 1.
For ifu and v are linearly independent in Ker(H - G) = Ker(t - L), then by a
basic result from differential equations, any solution w to the differential equation

!..-
dx
[P(X) d
Y]
dx
+ (q(X) - ~)
)...
Y= 0 (5.5)

is a linear combination of u and v. Therefore w satisfies (ii). But there exists a

solution to (5.5) with boundary conditions yea) = ex , y'(a) = f3, which can be
chosen so that (ii) fails to hold. Hence x is simple.
The following theorem summarizes these results.

Theorem 5.1 Let L be the Sturm-Liouville operator with Ker L = {OJ. Then L is
a selfadjoint invertible operator with compact inverse L -I = G, where G is the
integral operator on L2 ([a, b]) with kernelfunction g defined in (5 .1) . In addition,
dim Ker(H - L) ::::: 1,)... E C and L has infinitely many eigenvalues.

Corollary 5.2 The eigenvalues ofthe Sturm-Liouville operator L form an infinite

sequence {ILk} ofreal numbers which can be ordered so that

IILII < IIL21 < . .. -+ 00 .

If ipi- II cP j II =
1, is an eigenvector ofL corresponding to ILi- then cp] , CP2 , . . . is an
orthonormal basis for L2([a, b]).
A vector y E L2 ([a, b]) is in the domain ofL ifand only if

L IILjI21(y, cpj}1
00
2
< 00 .
j=1
For such vectors,
00

(Ly)(t) = L ILj(y, CPj}CPj(t) ;

j=1
convergence is with respect to the norm on L2([a, b]) . The series
00

yet) = L(y, CPj}CPj(t) , y E Vel) ,

j=1
converges uniformly and absolutely on [a, b].
214 Chapter VI Unbounded Operators on Hilbert Space

Proof: If KerL = (0), then the corollary follows from the theorem above and
Theorems V.l.l, IV.8.2.
Suppose zero is an eigenvalue of L. Choose a real number r such that r is not
an eigenvalue of L. The existence of such an r may be seen as follows.
Integrating by parts and making use of(ii), a straight forward computation yields

(Lu , v) = (u, Lv), u, V E D(L) .

Hence eigenvectors corresponding to distinct eigenvalues of L are orthogonal.

Thus the set ofeigenvalues of L is countable, otherwise L2 ([a, b]) would contain an
uncountable orthonormal set of eigenvectors of L, which contradicts the corollary
in Appendix 1. Since the set of real numbers is uncountable, the existence of r is
assured.
The corollary now follows from the above results applied to L - r I in place
of L.
Since r I - L is self-adjoint for r real and not an eigenvalue of L, it follows
from Corollary 4.3 that L is a closed operator. 0

6.6 Self Adjoint Operators with Compact Inverse

Theorem 6.1 LetHbeaninfinitedimensionalHilbertspace. Suppose At'H ~ H)

is selfadjoint and has a compact inverse. Then

(a) there exists an orthonormal basis {CPI , CP2 , . . .} of'H consisting ofeigenvec-
tors of A . If J-t I , J-t2, .. . are the corresponding eigenvalues, then each J-t j
is real and IJ-t j I ~ 00. The numbers ofrepetitions of J-t j in the sequence
J-tl, J-t2, . .. is finite and equals dim Kertu j - A).
(b) D(A) = {x E HI Lj 1J-t1 21(x , cp j}l2 < oo}
(c) Ax = Lj J-tj(x , CPj}CP j , x E D(A) .

Proof: Let K = A -I . Then K is compact and self adjoint by Theorem 4.4.

Hence by Theorem IV.5.1 and observations (a )-(d) in IV.6.l, there exists an ortho-
normal basis {CPI , CP2 , . .. } ofH consisting ofeigenvectors of K with corresponding
non-zero real eigenvalues AI, A2, which converge to zero. The number of repeti-
tions of Aj in the sequence AI, A2 , . .. is finite and equals dim Ker (Aj - K). The
operators A and K have the same eigenvectors cP since A -I cP = K cP = ACP if and
only if A i= 0 and Acp = tcp.
Statement (a) follows with J-tj = 1..
J
Suppose that for some x E H,L j IJ-t j 121 (x , cpg}1 2 < 00.
Exercises VI 215

Then the series Lj J.L j (x , ({Jj }({Jj converges. Now A -I is bounded and A -1({J j =
1.
)
({J j = J.L j ({Ji- Hence

x = I>x, ({Jj}({Jj = L J.Lj (x, ({Jj}A-1({Jj

j j

~ A-I ( ~ I'j (x, ~j)~j) E VIA).

On the other hand, if x E D(A), then since K = A -I is self adjoint and K ({Jj =
Aj({Jj, we have

Ax = L(Ax, ({Jj}({Jj = L J.Lj(Ax, K({Jj}({Jj

j j

= LJ.Lj(KAx, ({Jj}({Jj = L J.Lj (x, ({Jj}({Jj

j j

which proves (c) as well as (b) since

2 2
IIAxll = LJ.LJI(x, ({Jj}1 .
j

o
Exercises VI

1. An operator A(H --+ H) is called symmetric if (Au, v) = tu, Av) for

all u, v in D(A) (A need not be densely defined). Prove the following
statements .
(a) IIA + ixll 2 = IIxll 2 + IIAxll 2
(b) A is a closed operator if and only if 1m (A + if) is closed .
2. Prove that if A (H --+ H) is densely defined, symmetric and invertible, then
A is self adjoint.
3. Let A(H --+ H) be densely defined and let G(A) denote the graph of A
with the graph norm. Define V on the Hilbert space H x H by V (x, y) =
(-y ,x) . Prove that G(A*) = (VG(A».L , the orthogonal complement of
VG(A) in H x H . If, in addition, A is closed, show that H x H =
VG(A) ~ G(A*) .
4. Given A(HI --+ H2) and B(HI --+ H2), the operator A +B is defined as
follows: its domain
D(A + B) = D(A) n D(B) and
(A + B)x = Ax + Bx , x E D(A + B).
216 Chapter VI. Unbounded Operators on Hilbert Space

Suppose A + B is densely defined. Prove that

(a) (A + B)* is an extension of A * + B*, i.e., V(A * + B*) C
V(A + B)*, (A* + B*)y = (A + B)*y, y E V(A* + B*) .
(b) If, in addition, A is bounded on H, then (A + B)* = A * + B* .
(c) If B is an extension of A, then A * is an extension of B* .

5. Given S(H -+ H) and T(H -+ H) , the operater ST is defined as follows :

VeST) = {x E V(T)ITx E V(S)}

STx = S(Tx), x E VeST) .

Suppose S, T and ST are densely defined. Prove that (ST)* is an extension

of T* S*. If, in addition, S is bounded on H , prove that (ST)* = T* S*.
6. Set H = £2 and let T be the operator defined as follows : VeT) = sp{ek}f,
where {ek}f, is the standard basis in 12,

Prove that V(T*) = {CO, th, th· · ·)ICBj) E £2}

T*y = y, Y E V(T*).

7. Let H = L2([0, 1]). Define operators Ti(H -+ H) , i = 1,2,3 as follows:

V(Tj) = U E Hl/is absolutely continuous on [0,1], I' E L2([a, b]) .}

Ts f = if'
V(T2) = V(Tj) n UI/(O) = 1(1)}
T21 = if'
V(T3) = V(Tj) n UI/(O) = 1(1) = OJ.
T31 = if'
Prove that
a) Tt = T3
b) T2* = h

8. Suppose A(H -+ H) is densely defined. Prove

a) KerA* = ImA.L n V(A*)
b) If A is closed, then KerA = ImA*-L n V(A)

9. Let H = L2(ffi.). Suppose A(H -+ H) is the operator defined by V(A) =

Cgo(ffi.) and (Ax)(t) = tx(t) . Is A a closed operator?
Exercises VI 217

10. Find the eigenvalues, eigenvectors and the Green's function for the
following Sturm-Liouville operators.
= -y" ; y'(O) = 0, y(l) = °
°
(a) Ly
= -v" . yeO) = 0 , y(l) + y'(I) =
°
(b) Ly
(c) Ly = <y"; yeO) = 0, y/(l) =
(d) Ly = -s" - y ; y'(O) = 0, y(l) = 0.
11. Given the Sturm-Liouville operator

Ly = -y"- y ; yeO) = y (7t)

2' = 0,
express the inverse of L as an integral operator on L2 ([0, n /2]).
12. Consider the boundary value problem

cy = y(n) + ao(x)yn-l + ... + an-l (x)y = f(x)

y'(O) = leo) = ... = yn-l(O) = 0,
where each ak has a continuous nth order derivative and f is in L2[0, 1].
Define the domain V(L) of L to be those y such that y<n-l) is absolutely
continuous on [0,1], y(n) is in L2[0 , 1], and y satisfies (*) . Let Ly = ry.
Show that the equation Ly = f has a solution if and only if the integral

l
equation
x
q.>(x) = f(x) + k(x, s)q.>(s)ds

has a solution in V(L), where

n-l
' " aj(x)
k(x, s) = - L...J - .,-(x - s)J..
j=O J.

Hint : Let q.> = y(n) and integrate by parts.

13. In exercise 12, let n = 2 and verify that the equation Ly = f has a solution

1
given by
t
yet) = g(t,s)f(s)ds,

where
Yl (S)Y2(t) - Yl (t)Y2(S)
g (t, s ) = -=-----'---''--:-----=-..:....:...:..:....:-:.....:...
Yl (s)y~(s) - yi
(S)Y2(S)
with Yl , Y2 any basis for Ker L.
Chapter VII
Oscillations of an Elastic String

The aim of this chapter is to describe the motion of a vibrating string in terms of
the eigenvalues and eigenvectors of an integral operator.

7.1 The Displacement Function

Let us consider an elastic string whose ends are fastened at 0 and e. We assume that
the segment Oi is much greater than the natural length of the string. This means
that the string is always under large tension.
When a vertical force F is applied to the string, each point x on the string has
a displacement u (x ) from the segment. We shall determine u (x ), first when F is
concentrated at a point, then when F is continuously distributed. Finally, we study
the motion of the string by regarding acceleration as an inertial force.
A vertical force F (y) is appl ied at the point y causing a displacement u (x) at
each point x of the string. Let d = u (y) (Figure 1). The relative displacement of
a point on the string from a fixed end is independent of the point. Thus

F(y)
c
T

x y

Figure 1

u(x) d
- - = - ,0 S x S y
x y

u (x ) d
-- = -- y < X < i. (1.1)
i -x i-y ' - -
Let us suppose that the string is subject to small displacements and that the tension
T is constant. Since the sum of F and the vertical components of T is zero (the
tension acts in the direction of CO and C i ), we have

F = T sin O + T sin e.
220 Chapter VII Oscillations ofan Elastic String

From the assumption that the displacements are small, it follows that F is
approximately
de
Ttan8 + Ttancp = T
y(e - y )
.

Solving for d in (1.1) and substituting this value of d in the above equation yields

u(x) = k(x, y)F(y), (1.2)

where
X(f - Y )

!
-----yy- , O~x~y
k(x, y) = y (f - x ) (1.3)
Tf '
y~x ~ e.
Recall that for T = e = 1, k(x, y) is the Green's function corresponding to - ::2
(cf. (2.7) in Section VI.2.). We note that the displacement of the point x arising
from the action of a force applied at the point y is equal to the displacement of
the point y arising from the same force applied at the point x. In other words , if
F(x) = F(y), then u(x) = u(y) . This can be seen from (1.2) and (1.3) since
k(x, y) = k(y, x) and therefore

u(y) = key , x)F(x) = k(x , y)F(y) = u(x) .

Let us consider the more general situation where the string is subjected to a con-
tinuously distributed vertical force with density distribution f (y). The force acting
on the segment between y and y + ~y is approximately f(y)~y (~y is "small").
Thus if we take points 0 < Yl < Y2 < . . . < Yn = e, with ~y = Yi+l - Yi,
then the force F(Yi) at Yi is approximately f(Yi)~Y. Now each F(Yi) gives rise
to a displacement at x approximately equal to k(x, Yi)f(Yi)~y (Equation (1.2)) .
The displacement u(x) is approximately the sum L::7=1 k(x, Yi )f(Yi)~y of these
displacements. Ifwe let ~y -+ 0, we arrive at the formula

u(x) = 1 f
k(x , y)f(y) dy. (1.4)

7.2 Basic Harmonic Oscillations

An important type of oscillation is a simple harmonic oscillation. This is one in

which each point x on the string, when released from its initial position, moves
vertically so that its displacement u (x , t) is a sinusoidal function of time t with
frequency co which is independent of the point, i.e.,

u(x, t) = u(x) sinwt , w> O. (2.1)

Let us start with the assumption that there is no external force acting on the string.
If p2(x) is the density ofthe string, then by Newton's second law of motion (force
7.2 Basic Harmonic Oscillations 221

is the product of mass and acceleration), the inertial force acting on the section of
the string between y and y + /:!,. y is

(2.2)

It follows from (2.1) and (2.2) that the density distribution f ey ) is given by

(2.3)

Equations (1.4) and (1.3) yield

U(x , t ) = u (x ) sinwt

= (w21 i p2( y)k(x, y)u(y) d Y) sinwt.

If the density p2 is assumed to be constant, then

(2.4)

or
(2.5)
where K is the integral operator on L 2[0, e] with kernel function k defined in (1.3)
of Section 1.
In view of (2.5), our immediate goal is to determine the eigenvalues and eigen-
vectors of K . It follows from the discussion in VI.2 that ip is an eigenvector of K
with eigenvalues A if and only if A i= 0 and

1
ep" + -ep = 0
AT
ep (O ) = ep ee ) = o. (2.6)

The general form of a solution to (2.6) is given by

x . x
ep(x) = a cos r;p:;- + b sm r;p:;-' A> O. (2.7)
'\ITA '\ITA

ep(x) = ae Pn. + beTh, A < O. (2.8)

If (2.8) holds and ep (O ) = epee) = 0, then ep = O. Thus every eigenvalue of K is
positive and therefore K is positive , as noted in the beginn ing of Section IY.9. In
order to determine these positive eigenvalues, we note that (2.6) and (2.7) imply

A _ -I (- e )2 n = 1,2, ... (2.9)

n - T nJr '
222 Chapter VII Oscillations ofan Elastic String

are the eigenvalues of K. The eigenvectors corresponding to An are of the form

. nrrx
s, sm -i-' bn f. O. (2.10)

In particular,

(2.11 )

is an eigenvector of K corresponding to An, and lI«Pn II = 1. Moreover, «PI, «P2, . ..

is an orthonormal basis for L2 ([0 , i]) as seen in Section VI.2. It follows from
(2.5), (2.9) and (2.10) that for n = 1, 2, . .. ,

. b ' nrrx . JTnrr

un(x , t) = un(x) smwnt = n sm - - sm - - t
i pi
is the general form of the displacement of the point x.
Motions for which u (x, t) has the above form are called basic harmonic motions .
The numbers W n = V;;;trare called characteristic frequencies.

7.3 Harmonic Oscillations with an External Force

Suppose there is an external simple harmonic force which is continuously dis-

tributed and acts vertically on the string during motion. To be specific, let the
density distribution of the force be

hey) sinwt. (3.1)

Let us assume that the density p2 of the string is constant. Then by (2.3) in
Section 2 and (3.1), the density distribution fey) of the sum of the external force
and intertial force is given by

fey) = [hey) + p2w2u(y)] sinwt,

and from (1.4) in Section 1 we get

u(x , t) = u(x) sinwt = [g(X) + p2w21f k(x, y)u(y) d Y] sinwt ,

where
g(x) = i f k(x,y)h(y)dy.
Hence
(3.2)
7.3 Harmonic Oscillations with an External Force 223

If (pw) -2 is not an eigenvalue of K, i.e., w =1= .JT~~, n = 1, 2, . . . , then by

Theorem IV,8.1 , Equation (3.2) has the unique solution

An
+L
00
u(x) = g(x) (pw)-2 _ An (g, f{Jn}f{Jn(x),
n=l

where An and f{Jn are defined in (2.9) and (2.11) of Section 2. The series converges
uniformly and absolutely on [0, i). This follows from the uniform and absolute
convergence of L~! An (g , f{Jn }f{Jn (x), which is guaranteed by Theorem V,U and
the boundedness of the sequence {( pw )~2 - An ) .
Thus
00 .
'""" sm nitx
u(x) = g(x) + L...J an i ' (3.3)
n=!
where

2 An if . ntt y

r,
an = - 2 g (y) sm - - dy,
i (pw)- - An 0 i

An = ~ (n~ t» =1= v7:~, n= 1,2, . ...

We have seen that harmonic oscillations which arise from an external force,
with a density distribution of the form h (y) sin on, is decomposable into basic
harmonic oscillations.
Since the operator K with kernel function k, defined in (1.3) of Section 1 was
shown to be positive, it follows from Theorem V.3.1 that
00

k(x , y) = LAnf{Jn(X)<Pn(Y)
n=!

= --
u L -1s m
00
. nitx . nit y
--sm-- . (3.4)
T;r2
n=!
n2 e e
The series converges uniformly and absolutely on [0, i] x [0, i].
For a thorough and rigorous treatment of the equations describing large vibra-
tions of strings, as well as additional references, we refer the reader to [A].
Chapter VIII
Operational Calculus with Applications

The spectral theory which was studied in the preceding chapters provides a means
for the development of a theory of functions of a compact selfadjoint operator. We
now present this theory with applications to a variety of problems in differential
equations.

8.1 Functions of a Compact Self Adjoint Operator

Suppo se A E L(1t) is a compact self adjoint operator with a basic system of

eigenvectors {({In) and eigenvalues {A.n}. Let Po be the orthogonal projection from
'H onto Ker A. Then for each x E 'H,

x = Pox + ~)x , ({Jk }({Jk

k

Ax = L:).dx , ({Jd({Jk
k
2
A x = L A~(X , ({Jk}({Jk
k

A nx = LAk(X, ({Jk}({Jk . (1.1)

k

For any polynomial p(z) = 'LZ=o akZk, it is natural to define p (A ) = 'LZ=o akAk .
Therefore it follows from (1.1) that

p(A )x = p (O) Pox +L p(Ak)(X, ({Jk}({Jk . (1.2)

k

Let a(A ) be the subset of C consisting of zero and the eigenvalues of A . Let f be
a complex valued function which is bounded on a (A) . Guided by (1.2), we define
the operator f (A ) on'H by

f( A)x = f (O) Pox +L f (Ak)(X , ({Jk}({Jk· (1.3)

k
226 Chap ter VIII. Operational Calculu s with App lications

The operator f (A) does not depend on the choice ofthe eigenvectors {rpk }. Indeed,
let J-L I , J-L2, . .. be the distinct non-z ero eigenvalues of A and let Pn be the orthog-
onal projection onto Ker (J-LnI - A). Suppose tpj , tpj+ I , .. . , rpk is a bas is for Ker
(J-Ln I - A). Then
k
r ;» = L (x, rpi )rpi
i=j

and it follows from ( 1.3) that

f(A)x = f( O) Pox +L f( J-Ln) Pnx. (1.4)

n

Next we show that

IIf(A )1I = SUp{l f(A )1 : A an eigenvalue of A). (1.5)

Since Pox is orthogonal to each rpk, it follow s from (1.3 ) that

2 2I1 2
II f (A )xIl = If (O)1 PoxIl + L If(Ak)121(x, rpk )12
k

s IIxll 2 sup{f(Ad I2 + If (O)1 2I1PoIl 2} (1.6)

k

and

IIf(A )1I ~ If (A )rpnl = If(An)J,

II f(A)xll ~ If (O) llI x ll , x E KerA. (1.7)

It is easy to see that (1.6) and ( 1.7) impl y ( 1.5).

The operator f (A ) - f (O )Pox is compact and selfadjoint if and only if {f(Ad }
is a sequence of real numbers which is either finite or converges to zero .
At first glance the following result does not seem to be important. However, as
we shall see, it does have some interesting appl ications .

Theorem 1.1 Suppose F and g are comp lex valuedfun ction s which are bounded
on a (A). Then

(i) (a f + f3 g )(A ) = af (A ) + f3 g (A ), a , f3 E C.
(ii) (f.g)(A) = f (A )g (A ).
(iii) /f{fn } is a sequence ofbounded comp lex valuedfunctions which converges
uniformly on a (A ) to the boundedfunction f , then

IIfn(A) - f( A) II --+ O.
8.1 Functions ofa Compact SelfAdjoint Operator 227

Proof: Let {C{>n}, P.n} be a basic system of eigenvectors and eigenvalues of A . It

is clear from the definition of f(A) and g(A) that (i) holds . (ii). By definition,

f(A)g(A)x = f(O)Pog(A)x +L f('Ad(g(A)x , C{>k}C{>k . (1.8)

k

Since C{>k 1.. Ker A, 1 :::: k,

(g(A)x • • » ~ (g(O)POX + ~ gt): )(x• • j j I.j • • , )

= g('Ak)(X, C{>k} (1.9)

and

Pog(A)x = Po [g(O)POX + ~ g('Ad(x, C{>k}C{>k]

= g(O)Pox. (1.10)

Property (ii) follows immediately from (1.8), (1.9) and (1.10) . (iii). By (i) and
Equation (1.5),

IIfn(A) - f(A) II = II Un - f)(A) II

:::: sup Ifn('A) - f('A) I ~ 0 asn ~ 00.
J,.E a (A)
o
Corollary 1.2 Given a compactpositive operator A E £('H), there exists a unique
compact positive operator K such that K 2 = A. Infact,

Kx = L A (x, C{>n}C{>n, (1.11)

n

where {'An}, {C{>n} is a basic system ofeigenvalues and eigenvectors of A.

Proof: If K is given by (1.11), then K is compact positive and

Suppose B is a compact positive operator such that B 2 = A. To prove that B = K ,

it suffices to show that for all 'A > 0,

Ker(AI - B) = Ker(AI - K) . (1.12)

But (1.12) is clear since

(AI + K)(AI - K) = 'A 2I - A = (AI + B)(AI - B)

and -'A is not an eigenvalue of either B or K . o

228 Chapter VIII Operational Calculus with Applications

Theorem 1.3 Suppose A E £(1-l) is compact and self adjoint. If f(z) =

I:f:,o akz k isanalyticona closed disc containing a (A), then f(A) = I:f:,o ak A k.
The series converges in norm.

In particular, if f(z) = e Z = I:f:,o t, then

Ak
L:-.
00
A
e == f(A) =
k=O k!

The series converges in norm and

(1.13)

Moreover, we have from Equation (1.5) that

ileA II = sup{le AI : Aan eigenvalue of A}.

A simple application of Theorem 1.1 yields the following results.

Suppose f and g = t
are bounded on a(A) . Then

f(A)g(A) = g(A)f(A) = (g .f)(A) = I.

Thus f(A) is invertible and f(A)-1 = g(A) . In particular, if J1 ¢ a(A), let
f(A) = J1- A. Then f and g(A) = JL~A are bounded on a(A). Hence J1I - A is
invertible and

1
I
(J1I - A)- x = g(A)x = -Pox + L~ - -1( x , C{Jk)C{Jk.
J1 k J1 - Ak

By Equation (1.5),

1I(J1I - A)-III = IIg(A)1I = sup { _ l _ : A an eigenvalue of

1J1- AI
A} . (1.14)

Examples: 1. Let us consider the operator B = fJI + A . Now B = f(A), where

f(A) = fJ + A, and C{J is an eigenvector of A with eigenvalue A if and only if C{J is
an eigenvector of B with eigenvalue fJ + A. Therefore, if {1)I, 1)2, . .. } is the set of
eigenvalues of B which does not contain fJ , and Q) is the orthogonal projection
onto Ker (1)) I - B), then it follows from Equation (1.4) that Qi Q) = 0, i i= j
and
Bx = fJPox + L:
1)) Q)X, (1.15)
)
8.1 Functions ofa Compact SelfAdjoint Operator 229

where Po is the orthogonal projection onto Ker A = Ker (fH - B ) . Furthermore,

II B II = sup{l,8 + AI : A an eigenvalue of A}
= sup {IIi I : Ii an eigenvalue ofB}. (1.16 )

If {11k} is an infinite sequence, then it converges to ,8 since the sequence of

eigenvectors of A converges to zero.
While the series in (1.15) converges pointwise to B , it does not converge in norm
if,8 ::/= 0 (and there are infinitely many 11j). Indeed, given any integer N , choose
n so large that n > Nand l11n I > I~I . Then for cP E Ker (11n - B ) , II cpli = 1,

N
B - ,8 Po - L r/j Q j
j=l

2. Let K be an integral operator on H = Lz ([a , b D with kernel function K E

L z([a, b) x [a, bD. Suppose {An}, {CPn} is a basi c syst em of eigenvalues and eigen-
vectors of K . If f is a complex valued function defined at zero and L n I(An)IZ <
00 , then for all v in H ,

(f( K) v)(t) = f (O )(Pov)(t ) + ik b

(t , s) v (s) ds a.e. ( 1.17)

where
( 1.18)
n

convergence is with respect to the norm on L2([a, b) x [a, b)) We shall prove the
following more general result
Let {Vrn} be an orthonormal basis for H = Lz ([a, bD. If A E £ (H ) and
L i,j I(AVri , Vr j }!z < 00, then A is the integral operatoir with kernel fun ction

k(t, s) = L( AVri' Vrj}Vri(t ) Vrj (s ); (1.19)

i,j
convergence is with respect to the norm on Lz([a, b) x [a, b))
To prove this, put Cij = (AVri , Vrj) and <Pij(t, s) = Vri (t) Vrj (s ). Since {<pij}
is orthonormal in Lz([a , b) x [a, bD , Li,j II Cij<pij llz = Li,j Cb < 00 . Hence
L i,j Cij<Pij converges to some k E L z ([a, b) x [a, b)). Let K be the integral
operator with kernel function k and let K n be the integral operator with kernel
function z, = L 7, j=l Cij<Pij. Since IIKn -KII:::; IIkn -kll ~ O,astraightforward
computation verifies that
230 Chapter VIII Operational Calculus with Applications

Thus for all u, v in H = sp{1/Jj},

(Ku, v) = (Au , v) ,

which implies that A = K .

Returning to our original integral operator K, we choose A = f(K) - f(O)Po.
Our choice of an orthonormal basis for H is {CPn} together with any orthonormal
basis for Ker K = sP{CPn}.L . Since Au = 0, U E Ker K, and (Acpi , CPj) = f(A i)Oij,
Equations (1.17) and (1.18) follow from (1.19).
3. As a special case , let K be the integral operator on L2([-rr, rrD with kernel
function k(t , s) = h(t - s), where h is a continuous complex valued function of
period 2rr. It was shown in IV.1 that

An = rr
J-rr
h( s)e- ins ds n = 0, ±1 , ...

are the eigenvalues of K with corresponding eigenvectors

Since {CPn} is an orthonormal basis for L2([-rr, rrD , {CPn}, {An} is a basic system
of eigenvectors of K and Ker K = O. Hence for any positive integer p , equations
(1.17) and (1.18) applied to f(A) = )..P give

(KPv)(t) = l b
£'p(t - s)v(s) ds a.e.,

where
00
- 1 ~
kp(t) = 2rr L.... (convergence in L2 ([ -ot ; rr])).
n=-oo

8.2 Differential Equations in Hilbert Space

As we shall see in Section 3 and 4, certain integro-differential equations and

systems of differential equations give rise to a differential equation of the form

y'(t) = Ay(t), -00 < t < 00

yeO) = x E H, (2.1)

where A is in £(H) , y maps (-00 ,00) into Hand

y
'et) _ li yet
- un
+ h) - yet) '1..J
E'L.
u-;» h
Though the main results in this section are valid for an arbitrary operator A E
£(H), we shall assume that A is compact and self adjoint. This enables us to
8.2 Differential Equations in Hilbert Spa ce 231

express the solution to (2.1) as a series involving a basic system of eigenvectors

of A .
Ifwe consider the one dimensional case that A is a number, then the solution to
(2.1) is y(t) = etAx . It turns out thatthis is also the solution when A is an operator.

Theorem 2.1 Let A E L(1t) be compact and selfadjoint with a basic system of
eigencectors {qJd and eigenvalues (Ad. For each x E 'H, there exists a unique
solution to

y'(t) = A y (t ) , -00 < t < 00 (2.2)

y(O) = x.

The solution is
y(t ) = etAx = Pox + I:> Akt(X,qJk)qJk,
k

wh ere Po is the orthogonal projection from 'H onto Ker A.

Proof: From Theorem 1.1 and Equation 1.13 in Section I , if follows that if
y (t) = etAx , then

:s II xll il i AII IIAII t

k=l
Ihlk~:lIk
.

= IIxlili etAIlIIA II(e lh lli AIl - 1) ---+ 0 as h ---+ O.

Thus y' (t) exists and

y' (t) = Ay (t).
Also,
y(O) = eOx = x.
Suppose v (t ) is also a solution to (2. 1). Let

w(t) = v(t ) - y(t)

Wk(t) = (w(t) , qJkl.

Then

and
wdO) = (w(O ) , qJk) = (0, qJk) = O.
232 Chapter VIII Operational Calculus with Applications

Hence Wk(t) = O. Therefore

w'(t) = Aw(t) = I ).kWk(t)«Pk = O.

k

But this implies that yet) - vet) = wet) = O. Indeed, for any z E 'H,
d ,
dt (w(t) , z) = (w (t) , z) = O.

Hence the complex valued function (w(t), z) is a constant function. Since

(w(O), z) = (0, z) = 0, (w(t), z) = 0 for all t and all z. Thus wet) = 0 for
all t . 0

8.3 Infinite Systems of Differential Equations

Let us consider the infinite system of differential equations

Y; (r)= allYl (t) + a12Y2(t) + .
y~(t) = a2lYl (t) + a22Y2(t) + .

with boundary conditions

{Yi (O)} = {x;} E £2.

Suppose L0=1 laijl2 < 00 and aij = aji·

In order to solve this system, we define A E £(£2) by
A(al , a2, .. .) = (fh, fh, ·· .),
where f3i = L~l aijaj . Since A is compact and selfadjoint (II.16, Example 2,
11.11 .1 Example 6), it has a basic system of eigenvectors {«Pk} and eigenvalues
{Ad. By Theorem 2.1, if x = (Xl , X2, .. .) E£2, then

yet) = Pox +L e
Ak t
(x, «Pk)«Pk E £2
k

satisfies
y'(t) = Ay(t), - 00 < t < 00
yeO) = x.

Let Yj(t) be the ph component ofy(t). Then Yj(O) = Xj and

IYj(t+h~-Yj(t) _ ~ajkYk(t)1 S IIY(t+h~-Y(t) -AY(t)11 ~ 0

as h ~ O. Hence {Yj(t)}~l is the desired solution.

8.4 Integra-Differential Equations 233

8.4 Integro-Differential Equations

Let us consider the integro-differential equation

-acp (t , x) =
at
l
a
b
k (x , s) cp(t, s) ds, - 00 < t < 00, a ~ x ~ b (4.1)

with boundary cond ition

cp (O, x) = g(x ), (4.2)
where k is continuous on [a , b] x [a , b] , k (x , s) = k (s, x) and g is a given function
in L z([a, b]). Equations of this type arise, for instance , in the theory of energy
transport.
In order to use Theorem 2.1 to solve the equation, we express (4.1) in the form
y' (t ) = Ky(t) as follows :
Define y: (-00, 00) ~ L z([a , b]) by

yet) = cp(t , ).

That is to say, for each t , y (t) is the function cP(t , x) considered as a function of
x. Let K be the integral operator on L z([a, b] ) with kernel function k . Then (4.1)
can be written in the form

let) = Ky (t ).
yeO) = g. (4.3)

We must be careful here . Equation (4.3) means that

cp(t + h, ) - cp (t , ) yet + h) - y(t)
h h
converges, with respect to the norm on L z([a , b]) , to the function Kcp(t , ) as
h ~ O. Let us temporarily disregard all questions concerning convergence . We
know that the solution to (4.3) is

+L
00
tK Akt
cp(t, x) = (e g)(x) = ( POg)(x ) e (g, CPk}CPk(X)
k=!
00

= g(x) +L (e
Akt
- I)(g , CPk}CPk(X) , (4.4)
k= !

where {cpd , {Ad is a basic system of eigenvectors and eigenvalues of K.

To prove that the series converges and that tp is the desired solution, we recall
from the Hilbert-Schmidt theorem that

(K g)(x) = LAk( g , CPk}cpkCx ) (4.5)

k
234 Chapter VIII Operational Calculus with Applications

converges uniformly and absolutely on [a, b]. Thus, for any finite interval J , the
series in (4.4) converges uniformly on J x [a, b] since the mean value theorem
implies
leAkt - 11 ~ Pl.klC, 1 ~ k,
for some constant C and all t E J . Ifwe differentiate the series in (4.4) termwise
with respect to t, we get the series L~! AkeAkt (g , <Pk)<Pk(X) with converges
uniformly on J x [a , b]. Hence %'f (t , x) exists and

alp
-(t,x) " A
00

= L....Ake kt (g ,<Pk)<Pk(X),
at k=!

which is easily seen to equal the right side of (4.1). Obviously, <p(0 , x) = g(x).

Exercises VIII

1. For the following operators A find eiA , cos A and sinA.

(a) A : 12 ---?> 12 : A~ = (,.j2ii~2, ,.j2ii~! , 2Jr~3, 2Jr~4 , . ..)

I: I:
(b) A : L2[-Jr , Jr] ---?> L2[-Jr.Jr] ;

(A<p)(t) = t <p(x) cosx dx + cost <p(x)x dx

(c) A : L2[-Jr , Jr] ---?> L2[-Jr .Jr] ; (A<p)(t) = J::'n(t - s)2<p (s) ds
(d) A: L2[-Jr, Jr] ---?> L2[-Jr.Jr]; (A<p)(t) = J::'n cos t(t - s)<p(s) ds.
2. For each of the operators A in problem 1, check that e iA = cos A + i sin A
and cos 2 A + sin 2 A = I . Which is true , sin 2A = 2 cos A sin A or sin 2A =
2sinA cos A?
3. Solve the following integro-differential equations with the general boundary
condition

<p(O ,x) = f(x) E L2[-Jr,Jr] .

(a) -It<P(t, x) =cosxJ::'n<P(t,~)cos~d~

(b) -It<p(t, x) = J::'n(x - s) 2<p(t , s) ds
(c) -It<P(t, x) = J::'n <p(t, s) cos t(x - s) ds.

4. Let A be a positive compact operator on a Hilbert space. Check that the

differential equation ~ + Ay = 0 with boundary conditions yeO) = yo,
y' (0) = Xo has the solution

r: sintJA
yet) = cos tv Ayo + v'A xo,
Exercises VIII 235

where SiJt is the operator j(A) with j(A) = Sin;/! for A > °
and
j(O) = t.
5. Solve the following equations with boundary conditions

gJ(O , x) = °
gJ'(O ,x) = j(x) E L2[-JT,JT]

(a) ;t22gJ (t , X) =cosxJ~1Ccos~gJ(t.~)d~

(b) itzgJ(t, x) = J~1C cos(x - s) gJ (t , s) ds.

6. Find a formula for a solution to the nonhomogeneous equation

fr gJ (t , x) = J: k(x, s)gJ(t, s) ds + h(t , x)

{ gJ(O , x) = g(x) E L2[a, b],

where k is continuous on [a , b] x [a, b] andk(t, s) = k(s, t). Hint: Consider

a solution to y' - ay = j, where a is a number, and generalize as in VIA.
7. Solve the following nonhomogeneous equations for gJ(O , x) = g(x) E
L2[-JT,JT] .
(a) /tgJ(t, x) = cosx J~1C cos ~ gJ(t,~) d~ + cos x
(b) frgJ(t, x) = J~1C (x - s)2gJ(t, s) ds + tx
(c) frgJ(t , x) = J~1C cos! (x - s)gJ(t , s) ds + x .
Chapter IX
Solving Linear Equations by Iterative Methods

In this chapter we return to the problem of solving the equation x - Ax = y. One

of the themes of this book is to devise algorithms which generate a sequence of
vectors that converges to a solution. This chapter further develops this theme . The
main results rely on the spectral theorem.

9.1 The Main Theorem

Let H be a Hilbert space. Given A E £(H) and y E H, choose any Xo E Hand

define
Xn+l = Y + Axn, n = 0,1, . ... (1.1)
If {x n } converges, say to z, then it follows from the continuity of A that z = y + Az.
Thus z is a solution to the equation

x - Ax = y. (1.2)

The important question is, when does {x n } converge? Now

Xl Y + Axo
X2 y+AXI =y+Ay+ A2xo

Xn Y + Ay + A 2y + ...+ An-ly + Anxo.

Thus, if II A II < 1, then AnXo ~ 0 and
00

lim Xn = " Aky = (I - A)-ly.

n--->oo ~
k=O

With no restrictions on the norm of A, a sequence can be constructed which con-

verges to a solution of (1.2). We shall now prove this result under the assumption
that A is compact.

Theorem 1.1 Suppose K E £(H) is compact. Assume that for a given y E H

and a given complex number A, the equation

h+Kx=y (1.3)
238 Chapter IX Solving Linear Equations by Iterative Methods

has a solution. For any xQ E 11., define

Xn+l = (1- aB* B)xn + «tr». n = 0,1, . .. ,

where
B = AI +K and 0 < a < 2I1BII- 2 •
Then the sequence {xn } converges to a solution of(1.3).
If11. isfinite dimensional or if): 1= 0 and
0< a < min (2I1BII- 2 , 21)..1- 2 ) ,

then the convergence is uniform as XQ ranges over any bounded set.

The proof depends on the following results which are interesting in thier own
right.

9.2 Preliminaries for the Proof

Lemma 2.1 Given A E £(11.) and y E 11., suppose that the equation x - Ax = y
has a solution . Given XQ E 11., define

Xn+l = Y + Axn, n = 0, 1, ....

The sequence {x n} converges (to a solution of x - Ax = y) for every XQ E 11. if
and only if{A n x} converges for every x E 11..
The sequence {xn } converges uniformly as XQ ranges over any bounded set if
and only if{A n} converges in norm.

Proof: Let s; = Lk=Q A k . Then

(2.1)

Suppose {x n} converges for every XQ E 11.. Then for XQ = O. {Bny} converges

and (2.1) implies {AnxQ} converges. Let Lx =limn--*oo Anx, x E 11.. If {x n}
converges uniformly as XQ ranges over the I-ball S of11., then it follows from (2.1)
that Anx -+ Lx uniformly on S. Thus IIAn - LII -+ O.
Assume {An X} converges for every x. The convergence of {x n} will follow from
(2.1) once we prove that {B n y} converges. By hypothesis, there exists aWE 11.
such that Bny = Bn(w - Aw) = w - An+1w . Since {Anw} converges, so does
{Bny}. If {An} converges in norm, then it follows readily from (2.1) that {x n}
converges. uniformly as XQ ranges over any bounded set. 0

Theorem 2.2 Let A = f3I + K, where K is a compact selfadjoint operator on

11. and f3 is a real number. The sequence {An X} converges for every x E 11. ifand
only if II A II ~ 1 and Ker(1 + A) = (0) . In this case, the operator P defined on 11.
by Px = lim n--+oo AnX is the orthogonal projection onto Ker(1 - A).
If, in addition, 1f31 < 1 or 11. isfinite dimensional, then II An - P II -+ O.
9.2 Preliminaries for the Proof 239

Proof: Suppose {Anx } converges for each x E H . If).. is an eigenvalue of A with

eigenvector tp; then )..ncp = Ancp converges and x is real since A is selfadjoint.
Hence - 1 < ).. ~ 1. Thus Ker(l + A) = (0) and, by VIII. 1, Equation (1.16) ,

IIAII = sup (I)..I : ).. an eigenvalue of A} ~ 1.

Suppose IIAII ~ land Ker(l + A) = (0) . Given x E 'H, there exists a unique
U E Ker(l - A) and a unique v E Ker(l - A).L such that x = U + v. Let P be
the orthogonal projection onto Ker(l - A) . Then for n = 1,2, .. . ,

and
nx nvll nllvll nllxll
IIA - Pxll = IIA ~ IIAMll ~ IIAMll , (2.2)

where AM is the restriction of A to M = Ker(l - A).L.

The idea ofthe proof is to show that II A Mil < 1 if 1.8I < 1 and Anv --+ 0 for all
v E M . When this is accomplished, the theorem is a direct consequence of (2.2).
First we note that AM eM . Indeed, if v E M and U E Ker(l - A) , then

(Av, u) = (v, Au) = (v , u) = o.

Thus AM is in £(M) and AM = 1M + KM, where 1M and KM are the restrictions
of I and K to M, respectively. Hence it follows from VIII.1, Equation (1.16) ,
that IIAM II is 1.81 or IILI, where IL is eigenvalue of AM with largest modulus. Now
IILI ~ IIAII ~ 1 and from the definitionofMand the assumption Ker(l+A) = (0),
it follow that u :f. ±1. Thus IILI < 1 and therefore IIAMII < 1 if 1.81 < 1. Hence
we have from (2.2) that II An - P II --+ O.
If 'H is finite dimensional, then A is compact and self adjoint. Thus we may
apply the above result to 01 + A to conclude that IIAn - P II --+ o.
Suppose 1.81 > 1. Then -K = .81 - A is invertible since IIAII ~ 1. But then
'H is finite dimensional due to the compactness of K . Hence II An - P II --+ O.
Finally, assume 1.81 = 1. If M = (0), then A = I and the theorem is trivial.
Suppose M :f. (0). There exists an orthonormal set {cpkl C M and eigenvalues
{lLkl of AM such that for each v E M and n = 1,2, . . . ,

Let s > 0 be given. There exists an integer N such that

L I(v, CPk}1
2
< «n. (2.4)
k >N
240 Chapter IX Solving Linear Equations by Iterative Methods

We have shown above that the modulus of any eigenvalue of AM is less than 1.
Hence, for all n sufficiently large,
N
L IJLdnl(v, ({lk)1 2 < s/2. (2.5)
k=l

Equations (2.2)-(2.5) imply An x --+ P x . o

9.3 Proof of the Main Theorem

We are now prepared to prove Theorem 1.1. Let A = I - a B* B. The first step in
the proof is to verify that {Anx} converges for each x by showing that A satisfies
the conditions in Theorem 2.2. It is easy to see that

A = I - aB*B = I - a(iI + K*)(U + K) = (l - aIAI 2 ) I + Kv, (3.1)

where K 1 is compact and self adjoint. Given x E H, IIx II = 1,

(Ax, x) = (x - a Ii" Bx, x) = 1 - allBxll 2 .

If therefore follows from the conditions on a that

-l«Ax,x)~1.

Hence II A II .s 1 by Theorem IVA. 1, and

11(/ + A)»] 2: ((/ + A)x, x) = 1 + (Ax, x) > 0,

which implies that Ker(/ + A) = O. Moreover, the equation

x-Ax=aB*y (3.2)

has a solution. Indeed, by hypothesis, there exists aWE H such that B W = Y and

w - Aw = «s: Bw = «tr».
Therefore, we have from Theorem 2.2 and Lemma 2.1 that {x n } converges to a
solution v of Equation (3.2). The convergence is uniform as Xo ranges over any
bounded set ifH is finite dimensional or a satisfies the requirements ofthe theorem.
Now
«s: Bv = v - Av = aB*y,
hence
Bv - Y E Ker B* = 1m Bl.. .
Since Bw = y,

Bv - Y = B(v - w) Elm B n 1m Bl.. = (0)

9.3 Proofofthe Main Theorem 241

or
AV +Kv = Bv = y.
The iterative method in Theorem 1.1 does not have any specific relation to the
spectral theorem, although the proof is strongly based on the theorem.
Even if Equation (1.3) in Theorem 1.1 does not have a solution, nevertheless
{x n } converges for an appropriate ex when A =I O.

Lemma 3.1 * If B = J-LI + K , where K E £(1i) is compact and J-L =I 0, then for
any y E 'H, the equation B* Bx = B*y has a solution.

Proof: It is easy to see that A = BB* - 1J-L1 2 I is compact and self adjoint. Let
{A j }, {qJ j} be a basic system of eigenvaluoes and eigenvectors of A . For each
y E H, there exists a v E Ker A such that

y = v + L(y , qJj}qJj.
j

Nothing that B*qJj = 0 if j E N = {j : 1J-L1 2 + Aj O}, a straightforward

computation verifies that B* Bx = B*y, where

2 B *v
x = -1 + 'L.J
" 21 (y , qJj}B *qJj.
IJ-LI U N IJ-LI + Aj
o
Theorem 3.2 Let {x n } be the sequence defined in Theorem 1.1. If A =I 0, then
forO < ex < 2I1BII-2, the sequence {x n } converges. ForO < ex < min (2I1BII-2 ,
2IAI-2 ) , {x n } converges uniformly as Xo ranges over any bounded set.
Ifv = lim x. ; then (AI + K)v = Py, where P is the orthogonal projection
from 'H onto Im (AI + K) . Thus
inf IIAx +Kx - YII = IIAv
XEH
+ Kv - yll.
Proof: The proof of the main theorem relied on the fact that a B" Bx = x - Ax =
exB*y has a solution. But this we know from Lemma 3.1. An inspection of
the proof of the main Theorem 1.1 shows that {x n } converges to some v and
Bv - y E Im e-. Thus
(AI + K)v = Bv = Py .

The last statement of the theorem follows from the properties of P . o

* Lemma 3.1 still holds if, more generally, we only require that Im B be closed. Indeed,
H = Im B EB lm B.l.. = lm B EB Ker B* .
Thus y = B w + z for some w E Hand z E Ker B* . Hence B* y = B* Bw.
Theorem XV.4.2 shows that every operator of the form u.I + K, where K is compact and iL 'f 0
has a closed range.
242 Chapter IX. Solving Linear Equations by Iterative Methods

9.4 Application to Integral Equations

The iteration procedure in Theorem 1.1 can be readily applied to compact integral
operators as follows :
Let K be an integral operator with kernel function k E Lz([a, b) x [a, b]) .
Given g E Lz([a, b]), suppose that for some complex number A, the equation

Af(t) + l b
k(t,s)/(s)ds=g(t) a.e. (4.1)

has a solution in Lz([a , b]) .

Now for any IE Lz([a, b), we have from Theorem 11.11.2that

(K* Kf) = l {l b
k(s, t)
b
k(s, ~)/(~) -l ds

= ll b b
k(s, t)k(s, ~)/(~)d~ ds.
Given 10 E Lz([a , b]), define, for n = 0,1 , ... ,

In+ ! (I) = (l - a lAlz)ln (I) - a [l b

(Xk(t , s) + Ak(s, t»ln(s) ds

+ llb b
k(s, t)k(s, ~) In (~) d~ dS]
+ aXg(t) + a l b
k(s, t)g(s) ds,

where
0< a < 211A! + KII- z.
Since K is compact, it follows from Theorem 1.1 that Un} converges in Lz([a , b)
to a solution of (4.1).
Let Lf« = limn-->oo In, i.e., IILlo - In II -* O. Suppose A =I O. If a satisfies
(4.2) and a < 2IAI- z, then for each B > 0, there exists an integer N such that for
all 10 in the r-ball of Lz([a, b),

IILlo - In II < B, n > N.

The results in this chapter are based on the works of Krasnosel 'skil [Kra) and
Fel'dman [F).
Chapter X
Further Developments of the Spectral Theorem

In this chapter we continue to develop the spectral theory. Simultaneous

diagonalization, normal operators, and unitary equivalence are the main topics.

10.1 Simultaneous Diagonalization

Suppose A and B in .cO-£) can be diagonalized simultaneously, i.e., there exists

an orthonormal system of eigenvectors {'Pn} of both A and B with corresponding
eigenvalues {A.n} and {JLn} , respectively, such that for all x,

Ax = L:)"n(X, 'Pn}'Pn, Bx = LJLn(X, 'Pn}'Pn .

n n

Clearly, AB = BA.
An important result is that the converse holds if A and B are compact and self
adjoint.

Theorem 1.1 Suppose A and B are compact self adjoint operators in .c(H). If
AB = B A, then A and Bare diagonalizable simultaneously.

Proof: Let {A.n} be the basic system ofeigenvalues of A. It was shown in the proof
of the Spectral Theorem IY.5.1, that if 'PI E Ker ()"I - A), II 'PI II = 1 and

'Pn E {'PI, . .. , 'Pn_d..l n Ker (An - A), n = 2, 3, . . . , (Ll)

then {'Pn} is a basic system of eigenvectors of A. It follows from the assumption

AB = BA that Ker (An - A) is invariant under B . Thus, since B is compact and
self adjoint, we could choose 'Pn so that 'Pn satisfies (Ll) and 'Pn is an eigenvector
of B, say B'Pn = JLn'Pn . Given x E H , there exists a u E Ker A such that

x = u + L(x, 'Pk}'Pk . (1.2)

k

If B Ker A = (0) , then

(1.3)
244 Chapter X Further Developments ofthe Spectral Theorem

Suppose B Ker A :j:. {OJ. Since Ker A is invariant under B, there exists an
orthonormal system {o/j} C Ker A and a sequence of eigenvalues {1J j } of B such
that for all u E Ker A,
Bu = L 1Jj (u, o/j }o/j. (104)
j

It is clear from (1.2), (1.3) and (104) that either {qJn} or {o/I, qJI, O/Z , qJz , . . .} diag-
onalizes both A and B .
The theorem above shows that there exists an orthononnal basis {qJ j } for
M = 1m A such that the matrices corresponding to the restrictions AM, BM and
{qJj } are diagonal matrices. 0

10.2 Compact Normal Operators

Analogous to the complex numbers, every A E £('H) can be expressed in the form

where
Al * , Az =
= 2'1 (A + A) 2i1 (A - A *).

The operators A I and Az are self adjoint and are called the real and imaginary
parts of A , respectively.
It A is compact, then A I and Az are compact self adjoint. Therefore, A I and
A2 are each diagonalizable, but, in general, A is not. However, if AIA2 = A2A I ,
then by Theorem 1.1, A 1 and Az are simultaneously diagonalizable and therefore
A is diagonalizable.
A simple computation verifies that AIAz = AZAI if and only if AA* = A* A .
This leads us to the following definition.

Definition: An operator A E £('H) is normal if AA* = A* A .

Obviously, every self adjoint operator is normal ,
The following result is a generalization of the spectral theorem.

Theorem 2.1 An operator A E £('H) is compact and normal if and only if it has
a basic system ofeigenvectors and eigenvalues, where the sequence ofeigenvalues
ofA converges to zero if it is infinite.

Proof: Suppose A is compact and normal, Let A I and Az be the real and imaginary
parts of A . Then by the above discussion, there exists an orthonormal system {qJn}
and corresponding real eigenvalues {An} and {j.L n } of A I and Az, respectively such
that for each x E 'H,

Ax = Alx + iAzx = L(An + iJ.Ln)(x, qJn}qJn (2.1)

n

and An + i J.Ln -+ 0 if {qJn} is an infinite set.

10.2 Compact Normal Op erators 245

Suppose (2.1) holds and TJk = Ak + i !J.,k ~ O. Then

n
Anx = L TJd x , ifJk }ifJk
k=1

is compact and II A n - A II ~ O. Hence A is compact. Moreover,

A*x = L ijdx , ifJk }ifJk

k

smce

Therefore,
A* Ax = L ITJd(x , ifJk}ifJk = AA *x
k
for all x E 11, which means that A is normal. o
Application. Let K be an integral operator with kernel function k E L 2([a , b) x

i b fib ~)f(~) d~}

[a , b)). For f E L 2([a , b)),

( K* Kf)(t) = k (s , t) k (s, ds

= ii b b
~)f(~)
k (s , t )k(s , dsds

i i k(t , s)k(~, s)f(~)d~ds.

and
b b
(KK*f)(t) =

Hence a nece ssary and sufficient condition that K be normal is that for almost
every (t,~) E [a, b) x [a, b) ,

i k(s,t)k(s, ~)ds i k(t,s)k(~ , s)ds.

b
=
b
(2.2)

Since K is compact, we know from Theorem 2.1 that if (2.2) holds , then K has a
basic system of eigenvectors {ifJn} and eigenvalues {An} . Thus
2

ll
b b n
lim
n---> oo a
k (t , s)f(s) d s - I:a~jifJj (t ) dt = 0,
a j =1

l
where
b
a j = Aj f(s)ifJj(s) d s.

Moreo ver, Aj ~ 0 if K is of infinite rank.

246 Chapter X Further Developments ofthe Spectral Theorem

In view of Theorem 2.1, an operational calculus for compact normal operators

can be developed in the same way as for compact self adjoint operators.

10.3 Unitary Operators

It was shown in Theorem 1.18.1 that any two complex Hilbert spaces HI, H2 of
the same dimension are equivalent in the sense that there exists a U E £(H I , H2)
such that U is surjective and II U x II = IIx II for all x E HI .
An operator U which has the property that II U x II = IIx II for all x is called an
isometry.
Isometries have some very special properties as seen in the next theorem.

Theorem 3.1 Let U be in £(HI, H2). Thefollowing statements are equivalent.

(i) U is an isometry.
(ii) U*U = hil ' the identity operator on HI.
(iii) (Ux, Uy) = (x, y) for all x and y E HI .

IfU is an isometry and 1m U = H2 , then U U* = hi2' the identity operator on

H2 .

Proof: (i)::::} (ii) . For all x E HI ,

Thus U* U - hi l = 0 by Corollary IV.4.2.

(ii) ::::} (iii).
(Ux , Uy) = (U*Ux , y) = (x, y).
(iii) ::::} (i).
IIUxl\2 = (Ux , Ux) = (x, y) = IIx1\2.
Suppose U is an isometry from HI onto H2 . Given x and yin H2, there exists a
ZE H I such that Uz = y . Hence it follows from (ii) that

(UU*x, y) = (UU*x, Uz) = ur«. U*Uz)

= ur», z) = (x, Uz) = (x, y).

Since y is arbitrary, U U* x = x. o
Definition: A linear isometry which maps H onto H is called a unitary operator.
The above theorem shows that an operator U E £(H) is unitary if and only if it
is invertible and U- I = U* .
Thus a unitary operator is normal.
10.3 Unitary Operators 247

We have seen in (20.4) of Section 11.20 that of U E £(1i) is unitary, then the
spectrum
a(A) C {AIIAI = 1} . (3.1)

Examples: 1. Let aCt) be a Lebesgue measurable function on [a, b) such that

la(t)1 = I a.e. The operator U defined on L2([a , b)) by

(Uf)(t) = a(t)f(t)

is unitary.
2. Given a real number r, let fret) = f(t + r) for all t E (-00,00) . The
operator defined on L2 [( -00 , 00)] by Uf = fr is unitary.
3. The forward shift operator on £2 is an isometry which is not unitary since it
is not surjective,

Unitary operators can be used to identify compact self adjoint operators with
each other provided they have certain properties in common. The following theorem
is an illustration of this assertion.

Theorem 3.2 Let A and B be compact self adjoint operators on a sep arable
Hilbert space 'H. There exists a unitary operator U on 'H such that U* B U = A
if and only if
dim Ker (AI - A) = dim Ker (AI - B)

for all A E C.

Proof: Suppose dim Ker(A - A) = dim Ker(A - B) for all A E C. It follows from
IY.6.I (d) that A and B have the same basic system of eigenvalues {An}· Let {qJn}
and {1fJn} be basic systems of eigenvectors of A and B, respectively, corresponding
to {An}. Since dim Ker A = dim KerB , there exists , by Theorem U8.1, a linear
isometry Uo which maps Ker A onto Ker B . Define U E £(1i) as follows. Given
x E 'H, there exists a unique u E Ker A such that

x = u + ~)x, qJk}qJk .
k

Let
Ux = Uou + L(x, qJk}1fJk .
k

It is easy to see that U is a linear isometry. Moreover, U is surjective, For given

Z E 'H, there exists a unique W E Ker B such that

z = W + L(Z, l/!k}1/!k.
k
248 Chapter X Further Developments ofthe Spectral Theorem

Clearly

Given x E H,
BUx = ~)x, ({Jk}B1/!k = :L).k(X, ({Jk}1/!k
k k

= U L)..k(X, ({Jk}({Jk = UAx.

k

Since U*U = I, we have U*BU = A .

Conversely, if U*B U = A, where U is unitary, then

U*(AI - B)U = AI - A, A E C,

which implies
dim Ker (AI - B) = dim Ker (AI - A) .
The operators A and B in the above theorem are said to be unitarily equivalent.
o
10.4 Singular Values

Given Hilbert spaces HI and Hz, let A be a compact operator in £(HI, Hz). The
operator A * A is compact and positive since

(A*Ax, x) = IIAxli z ~ o. (4.1)

Therefore it has a basic system of eigenvalues

(4.2)

and eigenvectors ({JI , ({JZ, . . .

In case the sequence in (4.1) is finite , we extend it to an infinite sequence by
adding on zeros. We define the singular values of A, written s j (A), by

sj(A) = JA j(A*A), j = 1,2, .. .

note that sj(A) ~ Sj+1 (A) and Sj (A) ~ O. The importance of singular values is
seen in the following characterization of compact operators.

Theorem 4.1 If A E £(H I , Hz) is compact, then there exist orthonormal systems
{({Jj} S; HI and {1/!j} S; Hz such thatfor all x E HI,
v(A)

Ax = L sj(A)(x, ({Jj}1/!j , (4.3)

j=1
10.4 Singular Values 249

whe re v (A ) is the number ofnon-zero singu lar valu es of A , counted according to

multiplicities. Also,
v(A)
A*y = L sj (A)(y , 1{fj) ({)j , (4.4)
j=1

Proof: If A = 0, the theorem is trivial. Suppose A :;z!: O. Let {A j (A * A )} and {({) j }

be the basic system of eigenvalues and eigenvectors of A * A . For brevity we write
Aj = Aj (A *A ) and sj = Fl,
1:5 j :5 v (A ). Take 1{fj = ~ A({)j , 1:5 j :5 v( A) .
J
The sequence 1{fj is orthonormal since

Given x E HI , there exists, by IY.6.l(a), au E Ker A * A such that

v( A)

X = U + L (x , ({) j)({)j ' (4.5)

j =1

The equality Ker A * A = Ker A follow s from (4. 1).

Hence formula (4.5) gives

v( A ) v(A )

Ax = L(x , ({)j)A({)j = LSj(x , ({)j)1{fj .

j=1 j=1

Thus for all x E HI and y E H 2,

which implies equality (4.4 ). D

Theorem 4.2 Every comp ac t op erator in £(HI, H 2) is the limit in norm of a

se quence ofoperators offin ite rank.

Proof: Given a compact operator A E £ (HI , H 2), it has a representation given

by formula (4 .3) in Theorem 4.1. For each positive integer n, Define
n
A nx = I> j (A ) (x , ({)j)1{fj.
j= 1

It is easy to see that II A n - A II -+ O. D

250 Chapter X Further Developments ofthe Spectral Theorem

Corollary 4.3 Let 1i! and 1i2 be Hilbert spaces and let A be a compact operator
in £(1il, 1i2). Then the singular values

sj(A*) = sj(A).

Proof: The operator A * is compact by Theorem 11.16.2. Hence sj (A *) =

A~/2(AA*) . Let A and A* have the representations appearing in (4.3) and (4.4).
Then for y E 1i2,
v(A)

AA*y = L sj(A)(A*y, f{Jj)1/!j

j=!

and

Hence
v(A )

AA* y = L sJ(A)(y, 1/!j)1/!j

j=l

Thus

and

o
Example: Let K : L2([0, 1]) ---* L2([0, 1]) be the integral operator

(Kf)(t) = 2i 11

f(s) ds .

By taking
k(t ,S)={I,O:::;S:::;t:::;1
0, s > t
it follows from example 3 in 11.16, that K is a compact operator. Since the kernel
function corresponding to K* is k(s, t), we have that

f
I

(K* f)(t) = -u f(s) ds .

1

Hence
10.4 Singular Values 251

To find the positive eigenvalues A of K* K, we set

f (1
1
S

4 f(17)d17) ds = V(t) , A> O.

t

a straight forward computation shows that

V"+4f=0,
f(l) = 0, f'(O) = O. (i)

The general solution of the differential equation in (i) is

I I
f(t) = ae 2it / A'1 + be- 2it / ).'!
Since

we have
, 2i
0= f (0) = A1/ 2 (a - b) .

Thus
f(t) = a(e 2it / ).1/2 +e- 2it / ).1/2 ) =2acos2t/A1 / 2

and
0= fO) = 2acos2/A1 / 2 .
Hence
2/A 1/ 2 = (2k + l)rr, k = 0, 1,2, . . .

or
Sk (K) -- , 1/2 - 2 k 0 1 2
I\. - (2k + l)rr' = ", ...

It follows from equality (4.1) and the min-max Theorem IY.9.1 applied to A*A
that
Sj(A) = min max IIAxli. (4.6)
M IIxll=1
dimM=j -l Xl.M

Another important formula for the singular values is the following:

Theorem 4.4 Let 'H be a Hilbert space and let A E £(1i) be compact. Then/or
n = 1,2, . . .

sn(A) = min{IIA - Fill FE £(1i) , rank F ::s n - I}. (4.7)

252 Chapter X Further Developments ofthe Spectral Theorem

Proof: Suppose rank F = P :5 n - 1. Then dim(Ker F).l.. = dim ImF* =

dim 1m F = p . Hence by formula (4.6), with M = (Ker rv-,
sn(A) :5 sp+l(A):5 max IIAxll = max II (A - F)xll :5 IIA - FII ·
IIxll=1 IIxll=1
XEKerF XEKerF

Since F was an arbitrary operator in £(H) of rank :5 n - 1,

sn(A) :5 inf{IIA - Fill FE £(H), rank F :5 n - I}.

It remains to prove that the infimum above is attained and equals Sn (A). By
Theorem 4.1 the operator A = ,,£j<:?
sj(A)(x, CPj)Vlj, where {cpj} and {1/Jj} are
orthonormal systems. For n < v(A) + 1, define for n > I the operator Fnx =
,,£j:1sj(A)(x, CPj) 1/Jj , and Fl = O. Since rank F n = n - 1,

2
v(A)

II (A 2
- Fn)xll = L sj(A)(x, CPj)1/Jj
n

Thus IIA - F n II :5 sn(A) . Hence formula (4.7) holds for n < v(A) + 1. Thus if
m ~ v (A) + 1, then S n = 0 and rank A :5 n - 1. Therefore the infimum is attained
at F = A. 0

The next result shows that the singular values are continuous functions of com-
pact operators in £(H).

Corollary 4.5 Let A and B be compact operators in £(H) . Then

(4.8)

Proof: For any FE £(H) of rank at most n - 1,

sn(A) :5 IIA - FII :5 IIA - BII + liB - FII ·

Hence

sn(A) -IIA - BII :5 inf{IIB - Fill FE £(H),

rank F :5 n - I} = sn(B) (4.9)

by Theorem 4.4. Interchanging the roles of A and B in (4.9) yields (4.8). 0

10.5 Trace Class and Hilbert-Schmidt Operators 253

10.5 Trace Class and Hilbert-Schmidt Operators

In this section we introduce two important classes of compact operators.

Definition: A bounded linear operator A on a Hilbert space 1{ is said to be of

trace class ifit is compact and L:J=,1 sj(A) < 00.

Theorem 5.1 Let k be continuous on [a, b] x [a, b]. Suppose that for all
f E L2([a, b] x [a, b]),

ff
b b

kit, s)f(s)f(t) dsdt ~ 0. (5.1)

a a
Then the operator K defined by

f
b

(Kf)(t) = ktt , s)f(s) ds

a
is oftrace class and

f
00 b

:~::> j (K) = kit, t) dt (5.2)

J=l a
Proof: The operator K is compact as seen in Example 3 ofSection 11.16. Moreover
K is positive by (5.1). Hence K is selfadjoint and therefore the eigenvalues of
K* K = K 2 are the squares ofthe eigenvalues of K . Thus s j (K) = Aj (K) and
by Theorem V.4.1, equality (5.2), holds. 0
The compact operator

f
t

(Kf)(t) = 2i f(t)ds (5.3)

o
defined on L2([0, 1]) was shown in the example in Section 4 to have singular
values sk(A) = (2k~I)JT' k = 0,1, . . . . Therefore K is not of trace class.

Definition: A bounded linear operator A on a Hilbert space 1{ is said to be a

Hilbert-Schmidt operator if A is compact and L:J=,1 s;(A) < 00.
Every trace class operator is clearly a Hilbert-Schmidt operator.

Theorem 5.2 Suppose k E L2([a, b] x [a, b]) and k(t , s) = kis, t) . Then the
operator K defined on L2([a, b]) by

f
b

(Kf)(t) = k(t,s)f(s)ds
a
254 Chapter X Further Developments ofthe Spectral Theorem

is Hilbert-Schmidt and

ff
00 b b

~Sj(K)2 = 2
Ik(t, s)1 dsdt. (5.4)
J=! a a

Proof: The operator K is compact and selfadjoint. Hence sJ(K) = A](K) and
formula (5.4) is a consequence of Theorem Y.1.2.
o
The compact operator K defined in (5.3) is not of trace class but is a Hilbert-
Schmidt operator since

00
~st(K) = L
00 ( 2
(2k + l)7T
)2 < 00.
J=! k=!

Thorough treatments oftrace class and Hilbert-Schmidt operators appear in [GKre],

[GGKI] Chapters VII and VIII, and [DS2] .

Exercises X

1. Which of the following operators are normal and which are not?
(a) K : L2[0, 1] ~ L2[0, 1]; (K rp)(t) = fd (t - s)2rp(s) ds.
(b) V: L2[0, 1] ~ L2[0, 1]; (Vf)(t) = f~ f(s)ds.
(c) S, : f2 ~ f2; Sr~ =
(0, ~!, b .. .).
(d) A : L2[0, 1] ~ L2[0 , 1]; (Af)(t) = a(t)f(t), where a is bounded
and Lebesgue measurable on [0, 1].

2. Which of the following pairs of operators defined in Exercise 1 commute

and which do not?
(a) Sf and S~, k > 1, where Sf is the left shift on f2.
(b) A, V.
(c) K, V.

3. Let A and B be self adjoint compact operators on an infinite dimensional

Hilbert space. Let Ker A = Ker B = {OJ and AB = BA. Prove that there
exists a bijection f on the natural numbers such that the eigenvalues

4. Let A be a selfadjoint operator on a Hilbert space 'H. Define on 'H EB 'H EB 'H
an operator

B=
0
A
0
0
A)
0
(
o A 0
Prove that B is normal.
Exercises X 255

5. Let A be a compact self adjoint operator on a Hilbert space H . Let {A j },

{cp j} be a basic system of eigenvalues and eigenvectors of A . Find the
eigenvalues and the eigenvectors of the operator B defined in Exercise 4.
6. Generalize Exercises 4 and 5 to an operator of the type

0
A
0·· ·0
0 ·· ·0
A)
0
B = O . .~ .'.'.' .~ . 0 .
(
o O· · ·A 0

7. Prove that if A E £(H) is self adjoint, then A ± if are invertible and

V = (A - i/)(A + i/)-l is unitary. Hint: IIAx ± ixll z = IIAxUZ + IIxli z.
V is called the Cayley transform of A.
8. Let K be a compact normal operator on a Hilbert space. Prove that I +K
is unitary if and only if ]l +Aj(K)1 = I forallj.
9. Let A be a compact normal operator on a Hilbert space. Suppose A =
G + i F, where G and F are self adjoint operators. Prove that

eA = eG(cos F + i sin F).

Is this also true in case A is compact but not necessarily normal?
10. Let A and B be compact self adjoint operators on a Hilbert space. Prove
that if AB = B A, then ea A +f3B = ea A . ef3B .
II. Does the formula in Exercise 10 hold if AB i= BA?
12. Let A and B be compact self adjoint operators on a Hilbert space . Prove
that if AB = BA and two functions f(A) and f(B) are defined, then

f(A)g(B) = g(B)f(A).

13. Find the real and imaginary parts of the following operators.
(a) (Kcp)(t) = J::'nk(t -s)cp(s)ds ; K : LZ[-JT,JT] --+ LZ[-JT,JT],k E
Lz[-2JT,2JT].
(b) (Af)(t) = u 1/
f(s) ds; A : Lz[O, I] --+ Lz[O, I].
(c) (Af)(t) = 2J~k(t ,s)f(s)ds; A : Lz[O,I] --+ Lz[O,I],k E
Lz([O, 1] x [0, 1]).

(
~ ~z 1~~
(d) A = . Aij E £(H).
000
000
256 Chapter X Further Developments ofthe Spectral Theorem

14. Let G be a self adjoint invertible operator on a Hilbert space 'H and let
B = G 2. Define on 1t a new scalar product (' , ')B by (h, g)B = (Bh, g) .
Prove the 'H is a Hilbert space with respect to the new scalar product and
that the new norm is equivalent to the old one.
IS. Let Band (', ' )B be defined as in exercise 14. For any A E £(1t), define
A x to be the adjoint operator of A with respect to the new scalar product
( ', .) B. Prove that A x = B- A * B and that A is self adjoint with respect to
1

the new scalar product if and only if BA = A * B .

16. Let B be the operator defined in exercise 14. Let A =1= 0 be a compact
operator such that A * B = BA. Prove that
(a) A has only real eigenvalues and at least one eigenvalue of A is not
equal to zero.
(b) If cP j and CPk are eigenvectors of A corresponding to different eigen-
values, then cP j and CPk are orthogonal with respect to the new scalar
product (-, .) B .
(c) A admits a spectral representation

Ah = L ).j(h, CPj)CPj ,
j

where p.j} are the eigenvalues and the tpj form a system for which
iBip], CPk) = Sjk·
(d) Conversely, if A has the representation given in (c), then BA = A * B .

17. Let B be defined on L2[a, b] by (Bcp)(t) = p(t)cp(t) . where p is a positive

Lebesgue measurable function such that

o< inf p(t) , sup p(t) < 00.

tE[a .b] tE[a .b]

Define A on L2[a, b] by (Acp)(t) = f:

k(t , s)p(s)cp(s)ds, where k =1= 0 is
a selfadjoint Hilbert-Schmidt kernel on [a, b] x [a, b]. Prove that
(a) BA = A*B.
(b) A has eigenvalues different from zero.
(c) The eigenvectors of A corresponding to different eigenvalues are orthog-
onal in the scalar product (-, .) B .

18. Prove that A E £(1t) is normal ifand only if II Ax II = II A *x II for all x E H,

in which case H = 1m A EB Ker A.
19. Generalize Theorem 1.1 for a finite number of operators.
20. Is it possible to generalize Theorem 1.1 for an infinite set of operators ?
Exercises X 257

21. Prove that if A is a compact normal operator and M is a closed A-invariant

subspace, then M is also A*-invariant and M.l.. is A-invariant. Hint: Let
{Pn} be a sequence of polynomials which converges uniformly on <T(A) to
fez) = Z. Consider Pn(A) and f(A) restricted to M .
22. Let A be a compact normal operator. Prove that a closed subspace is
A-invariant if and only if it is An and AI invariant, where An and AI
are the real and imaginary parts of A, respectively.
23. Generalize Theorem 1.1 for two compact normal operators.
24. Generalize Theorem 1.1 for a finite number of compact normal operators.
25. Prove that any bounded finite rank operator on a Hilbert space is a Hilbert-
Schmidt operator.
26. Prove that any bounded finite rank operator on a Hilbert space is trace class.
27. Suppose that for some orthonormal basis CPl, CP2, . • • in Hilbert space H and
for some operator A E £(H) the series L~l IIAcpj 11 2 converges. Prove
that A is Hilbert-Schmidt.
28. Suppose that for some orthonormal basis CPl , CP2, •.. in Hilbert space H and
for some operator A E £(H) the series L~II(Acpj, cpj}1 converges. Is
the operator A trace-class ?
29. If the operator A in problem 28 is known to be positive, is A trace class?
30. How many non zero singular values does a bounded operator of a given
rankk have?
31. Give an example of a self adjoint compact operator which is
(a) not trace class
(b) not Hilbert-Schmidt

32. Give an example ofa compact operator which has only zero as its eigenvalue
and is not Hilbert-Schmidt.
Chapter XI
Banach Spaces

A norm can be defined on Cn in many different ways. Some of them stem from an
inner product. An example of a norm on C n which does not arise from an inner
product is lI(al, a2, · ··, an)IJ = maxl ::;i::;m lad·
This is due, for example, to the fact that the vectors (1,0, . . . , 0) and (0,1,
0, . . . ,0) do not satisfy the parallelogram law (cf . I. 3.1) in that norm. As we
shall see, all norms on C" are equivalent. This need not be the case for infinite
dimensional spaces as the situation is more complicated.
Many problems are cast in an infinite dimensional setting with norms which do
not have some of the nice geometrical properties of a Hilbert space norm. The
choice of the norm plays a vital role in applications and in solutions of certain
problems. This leads us to study, in this chapter, Banach spaces which are more
general that Hilbert spaces . In subsequent chapters we extend the theory oflinear
operators to Banach spaces .

11.1 Definitions and Examples

Let us start with the definition of a Banach space.

Definition: Let X be a vector space over Co A norm II . II on X is a real valued

function on X which has the following properties.

(i) IIxll 2: 0, IIxll = 0 implies x = O.

(ii) lIaxll = lalllxll, a E C .
(iii) IIx + yll ~ IIxll + IIYII (triangle inequality).

The space X, together with II . II, is called a normed linear spac e.

Many of the definitions which appear in the preceding chapters carry over
verbatim to normed linear spaces. For convenience we shall not restate most
of the definitions.
For example, a sequence {x n } in a normed linear space X is called a Cauchy
sequence if IIxn - X m II -+ 0 as n , m -+ 00 . X is complete if every Cauchy
sequence in X converges.
A complete normed linear space is called a Banach space.
260 Chapter XI. Banach Spaces

Examples: 1. Every Hilbert space is a Banach space.

2. Let C ([a, b]) denote the set of continuous complex valued functions defined
on the interval a S x S b. With the usual definitions of addition and scalar
multiplication of functions, C ([a, b]) is a vector space over C and

11/11 = max
tE[a ,b]
1/(t)1

defines a norm on this space.

Since II In - 1m II -+ 0, as n, m -+ 00, if and only if Un - 1m} converges
uniformly to zero on [a, b], it follows from the uniform Cauchy criterion
that Un} converges in C([a , b]). Thus C([a , b]) is a Banach space.
3. For 1 S p < 00, let i p be the set of those sequences ~ = (~1, ~2, . . .)
of complex numbers for which L~1 I~n IP < 00. Defining addition and
scalar multiplication as for i2, it can be shown ([TL] Section II.4 and [Sc]
Section 2.4) that i p is a vector space and

00 )1/P
II~II = (
~1~nIP

is a norm on i p' The completeness of i p can be established by the argument

which is used to prove that i2 is complete (1.2.1).
4. Leti oo be the set ofall bounded sequences ofcomplex numbers. With respect
to the definitions of addition and scalar multiplication as defined on i 2 , the
set i oo is a vector space and

1I(~I , ~2, " ·)11 = sup I~nl

n

defines a norm on i oo. It is easy to show that ioo is complete.

5. Let Co be the subspace of ioo consisting of sequences of complex numbers
which converge to z. It is easy to verify that Co is a Banach space .
6. Another useful source of examples is the Banach space i p (Z), 1 S p S 00,
of all "doubly infinite" sequences

~ = (.. . , ~-2 , ~-1, ~o , ~1, ~2 , . . .)

with

1I,lI p ~ (i=oo I<,I P ) l /p , IS p < 00,

1I~1I00 = sup I~nl , p = 00.

n
11.1 Definitions and Examples 261

7. Lp([a, b]), 1 ~ p < 00, is a Banach space with

IIf11 p = ( lblf (t ) ,Pdt

)I/P
,

and Loo([a, b]) is a Banach space with

IIf1100 = ess suplj (rj] .

A treatment of these spaces may be found in [R] (see also Appendix 2) .

Theorem 1.1 If X and Yare Banach spaces, then L(X, Y) is a Banach space.

Proof: Suppose {An} is a Cauchy sequence in L(X, Y). Then for x E X,

as n, m ~ 00. Since Y is complete, {Anx} converges. Define Ax = lim n --+oo Anx.

Given 8 > 0, there exists an integer N such that IIA n - Am II ~ 8 , n, m ~ N .
Therefore, if IIxll = 1 and n ~ N, then

IIA nx - Axil = lim IIA nx - Amxll ~ lim IIA n - Amll ~ 8.

m~oo m~oo

This implies that A = A - AN + AN E LeX, Y) and

IIA - An II ~ 0 n ~ 00 .

o
Excluding £2 and L2([a , b]), none of the Banach spaces in Examples 2-8 is a
Hilbert space, i.e., there does not exist an inner product ( , ) on the space such that
Ilxll = (x,x)I /2 (Exercise 8).
There are some very basic properties which Hilbert spaces possess and arbitrary
Banach spaces lack. For example, not every closed subspace of a Banach (not
Hilbert) space is complemented. Indeed, the closed subspace Co of £00 is not
complemented in £00 (cf. [W]). These differences, which we shall point out in
subsequent sections, are due to the fact that a Banach space has less structure than
a Hilbert space.
262 Chapter XI. Banach Spaces

11.2 Finite Dimensional Normed Linear Spaces

Given a finite dimensional vector space X over C, there are infinitely many norms
which can be defined on X. For example, suppose XI , ••. , X n is a basis for X.
Then for X = L:7=1 aixi,
n
IIxlli = L lail, IIxlI(r) = rlixlli , r > 0,
i=1

(2.1)

all define norms on X. However, we shall now show that all norms on X are
equivalent in the following sense.
Definition: Two norms II . II and II . III, on a vector space X, are called equ ivalent
if there exist numbers C and m > 0 such that for all X EX,

It is clear that if 1111 and 11111 are equivalent norms on X, then Xo = (X , 1111) is
complete if and only if XI = (X , II II j) is complete. Also, a sequence converges
in X 0 if and only if it converges in XI.

Theorem 2.1 Any two norms on afinite dimensional vector sp ace are equivalent.

Proof: Let XI , • . • , X n be a basis for the vector space X. For any norm II . II on X
and any ak E C, 1 ::s k ::s n,

(2.2)

We shall show that there exists an m > 0 such that for ak E C , 1 ::s k ::s n ,

(2.3)

Define the real valued function f on S = {a E en : lIa II = I} by

11.2 Finite Dimensional Normed Linear Spaces 263

Since f is continuous and S is compact, f assumes its minimum m at some point

~ E S. The linear independence of {Xl , . . . , x n } implies that m = f (~) > O.
Hence for 0 t= a = (a I , . .. , an) E en,

Since (2.2) and (2.3) are valid for any norm on X, it follows that any two norms
on X are equivalent. 0

From the properties of en and Theorem 2.1, we obtain the following result.

Theorem 2.2 If X is a finite dimen sional normed linear space, then

(a) X is complete.
(b) Every bounded sequence in X has a convergent subsequence.

Property (a) implies that every finite dimensional subspace of a normed linear
space Y is closed in Y .
Property (b) is false if X is infinite dimensional. In order to prove this we use
the following lemma.

Lemma 2.3 If M is a finite dimen sional, proper subspace of a normed linear

space X , then there exists an X E X such that

1= IIxll = d(x , M ),
where d(x , M ) is the distance from X to M .

Proof: Let z be in X but not in M. There exists a sequence (md in M such that
li z - mkil ~ d(z , M ) > O. Since M is finite dimensional and {md is bounded,
there exists a subsequence {mk'} of {md and an m E M such that mk' ~ m .
Hence

0< liz - mil = lim li z - mk'il = d(z , M) = d(z - m, M) .

k' -4 00

Th us c: z- m
lor X = li z-mil'

d(z -m , M)
1 = [x] = = d(x , M).
liz-mil
o
Theorem 2.4 Ifevery sequence in the Y-sphere ofa norm ed linear space X has a
convergent subsequence, then X is finite dimensional.
264 Chapter XI. Banach Spaces

Proof: Assume X is infinite dimensional. Choose XI E X, IIxIIi = 1. By

Lemma 2.3, there exists an X2 E X such that

After {XI, . . . , Xk} has been obtained, choose xk+ I E X such that

where M; = Sp{XI, . . . , Xk}. Now {x n } is in the unit sphere of X but the sequence
does not have a convergent subsequence since

This contradicts the hypothesis ofthe theorem. Therefore, X is finite dimensional.

o
In general, the sum of two closed subspaces of a Banach space need not be
closed (cf., Section XII.5) unless one of the subspaces is finite dimensional.

Theorem 2.5 If M is a closed subspace and N is a finite dimensional subspace

ofa normed linear space, then M + N = {u + v : u EM, v EN} is closed.

Proof: Assume dim N = 1, say N = sp{x}. The lemma is trivial if X E M.

Suppose X 1. M and Zk -+ Y, where

The sequence {ad is bounded; otherwise there exists a subsequence {ae} such
that 0 < lak,l -+ 00 . Hence

Thus X is in M since M is closed, which is a contradiction. Consequenctly, {ak} is

bounded and therefore it has a subsequence {ak'} which converges to some a E <C.
Thus

Hence y - ax is in M or y E M + N . The proof of the lemma now follows by

induction. 0

11.3 Separable Banach Spaces and Schauder Bases

Definition: A normed linear space X is called separable if it contains a countable

set which is dense in X.
11.4 Conjugate Spaces 265

Examples: 1. C([a, b]) andLp([a , b]), 1 ~ p < 00, are separable Banach spaces
since it follows from the Weierstrass approximation theorem that the count-
able set ofpolynomials with complex rational coefficients is dense in C ([a, b])
which, in turn, is dense in Lp([a , b]).

2. lp, 1 ~ p < 00, is separable since the countable set of all sequences {ak} of
complex rationals, where ak = 0 for all k sufficiently large , is dense in lp.

3. loo is not separable. For given any countable set {x n } in loo, we can
always find an x E loo such that IIx - X n II 2: 1. Indeed, suppose X n =
I ,a2(n) ,. .. ) . Let an = 0 1'f Ian(n) l>I d
( a (n) an a n = 1 + Ian(n) I ot herwise.
.
Then x = (ai , a2, . ..) is in loo and

We have seen that every separable Hilbert space 1i has an orthonormal basis
{Cfln} . Thus every x E H can be represented by x = Lk akCflk and the
representation is unique, i.e., if x = Lk fJkCflk, then ak = fJk.

For a Banach space we have the following definition.

Definition: A sequence {x n } in a Banach space X is called a Schauder basis for X if
for every x EX, there exists a unique sequence {an} in C such that x = Ln anxn .
The standard basis {en} is a Schauder basis for lp, I ~ p < 00 . C([a , b]) has
a Schauder basis of polynomials and L p ([ -Jr, zr[), 1 ~ P < 00 has a Schauder
basis consisting of {I, cos nx, sin nx} ~= I ' These results are difficult to prove .
Many of the separable Banach spaces which one encounters have a Schauder
basis. However, there is one without a Schauder basis [E].

11.4 Conjugate Spaces

Throughout the rest ofthis chapter, X denote s a Banach space,

As we shall see in subsequent sections, it is very useful to study the space of
bounded linear functionals on X. In this section we shall describe all bounded
linear functionals on a certain X. In contrast with the Hilbert space case, the
description depends essentially on the space X. Instead of a generalized Riesz
theorem (11.5 .2) for Banach spaces, it turns out that we need many theorems, one
for each space.
By Theorem 1.1, the space of all bounded linear functionals on X with norm
1If11 = sUPllxll=1 I!(x)! is a Banach space . This space is called the conjugate
space of X and is denoted by X' .
We shall now describe the conjugate space l~, 1 ~ p < 00 .
266 Chapter XI. Banach Spaces

Theorem 4.1 Given! E i~, 1 ::; p < 00, there exists a unique fJ = (fJI, fJ2, . . .) E

iq , -t + ~ = 1, (q = 00 when p = 1), such thatfor all ~ = (~I, ~2, . . .) E ip,

00
!(~) = L ~kfJk . (4.1)
k=1
Moreover, 1If11 = IIfJlI and fJ = (f(ed, !(e2) , . . .), where {ed is the standard
basis for i p -
Conversely, given {fJd E iq , (4.1) defines an I E i~ .

Proof: We shall only prove the theorem for p = 1. The proof for 1 < p < 00
appears in [TL] p. 143.
Given lEi;, let fJk = !(ek). For s = (~I, b . ..) Ell,

I(~) = I ("E ~kek) "E ~kfJk.

= (4.2)

and fJ = (fJI, fJ2, ...) is in ioo since

IIfJlloo = sup IfJkl = sup I/(ek)1 ::; 11/11· (4.3)
k k
Also,
00 00
I!(~) I ::; L l~kfJkl s IIfJlloo L I~kl = IIfJlloolI~lI·
(4.4)
k=1 k=1
Thus we have from (4.3) and (4.4) that 1If11 = 11171100 '
If there exists (th , th , . . .) E ioo such that !(~) = L~I ~k{3k for all ~ =
(~I , b ···) E ii , then I(e j) = {3j .
Inequality (4.4) implies that the functional I which is given by (4.1) is in i;.
Theorem 4.1 shows that i~ can be identified with iq in the following sense .
There exists a linear isometry J which maps i~ onto iq . The operator J is
defined by Jf = {f(ed}. For fJ = (fJI, fJ2, .. .) E iq, J-IfJ = g E i~, where
00
g(~) = L~kfJk'
k=1
The proof of the following description of L~([a, b]), 1 ::; p < 00, appears
in [R] p. 246 . 0

Theorem 4.2 For each F E L~([a , b]), 1 ::; p < 00, there corresponds a
unique (up to sets ofmeasure zero) g E Lq([a, b]), -t + ~ = 1, such that for all
! E Lp([a, b]),

F(f) = l b
!(t)g(t) dt . (4.5)

Furthermore, IIFII = IIgll·

11.5 Hahn-Banach Theorem 267

Conversely. gi ven g E Lq([a , b]) , the functional F defined by (4.5) is in

L ~ ([a , b]).

Add itional examples of conjugate spaces may be found in [OS l] .

11.5 Hahn-Banach Theorem

One ofthe fundamental results in the theory of Banach spaces is the Hahn-Banach
theorem. This section contains the proof of the theorem together with some very
important applications.
Hahn-Banach Theorem 5.1 Iff is a bounded linear functional which is defined
on a subspace M ofa normed linear space X, then f can be extended to a bounded
linear functional F defined on X such that 1If11 = IIFII.
The theorem is clear if X is a Hilbert space (see Exercise XII-5).
In order to prove the Hahn-Banach theorem, we need the following preliminary
results.
A set E is called partially ordered if there exists a binary relation, S, defined
for certain pairs (x, y) E E x E such that

(i) x S x,
(ii) if x S y and y S Z, then x S z.

Let F be a subset of E . An element x E E is an upper bound for F (with respect

to S) if y s x for every E F.
The set F is totally ordered if for every y and Z in F , either y S z or Z S y .
An element m E E is said to be maximal if m S x implies m = x .
For example, if E is the set of all subsets of a set S, let A S B mean A C B .
Then E is partially ordered with respect to s. If F is a subset of E, then UAeFA
is an upper bound for F . If F con sists of sets A l S; A2 S; . . . , then F is totally
ordered. The set E is the only maximal element in E.
The following lemma, which is equivalent to the action of choice, has many
applications in functional analysis and other branches of mathematics. The reader
is referred to [OSl] Section 1.2 for a further discussion of Zorn's lemma.
Zorn's Lemma. Let E be a non-empty partially ordered set. If every totally
ordered subset ofE has an upper bound in E , then E contains a maximal element.
The following result is a general version of the Hahn-Banach theorem.

Theorem 5.2 Suppose X is a vector space over the real or comp lex numbers. Let
p be a real-valuedfunction defined on X such thatfor all x , y in X and numbers a ,

(a) p (x + y) S p (x) + p (y )
(b) p (ax) = la jp (x) .
268 Chapter XI. Banach Spaces

Suppose f is a linearfunctional defined on a subspace M of X and

If(z)1 ::::: p(z) for all z E M.

Then f can be extended to a linear functional F defined on all of X such that

IF(x)1 ::::: p(x) for all x EX.

Proof: Let £ be the set of all linear functionals g such that the domain D (g) of g
is contained in X, g = f on M, and Ig(x)1 ::::: p(x) x E D(g) . Obviously, f is
in E, We partially order £ by letting g ::::: h mean that h is an extension of g. The
idea ofthe proof is to use Zorn's lemma to show that £ has a maximal element and
that this element is defined on all of X.
Let .1 be a totally ordered subset of E: Define the linear functional H by

D(H) = U D(g)
gE.:J
H(x) = g(x) if x E D(g).

Since .1 is totally ordered, it follows that H is in E and that H is an upper bound

of.J. Hence £ has a maximal element F by Zorn's lemma. The theorem is proved
once we show that D(F) = X.
Suppose there exists an x E X which is not in D(F) . Let Xl = sp{x} $ D(F) .
We shall show that there exists aGE £ which is an extension of F to Xl, thereby
contradicting the maximality of F . It is clear that we should define G on X 1 by

G(ax + z) = aG(x) + F(z) , z E D(F) , (5.1)

with G(x) chosen so that

IG(ax + z)1 ::::: p(ax + z), z E D(F) . (5.2)

In order to determine G(x), let us first assume that X is a vector space over the
reals. If (5.2) is to hold, then in particular,

G(x + z) ::::: p(x + z). z E D(F) (5.3)

and
G(-x - y) ::::: p(-x - y) = p(x + y), Y E D(F) . (5.4)

Thus, from (5.1), (5.3) and (5.4) we want

G(x) ::::: p(x + z) - F(z) , z E D(F). (5.5)

and
G(x) 2: -p(x + y) - F(y), Y E D(F). (5.6)
11.5 Hahn-Banach Theorem 269

It is easy to check that the right side of (5.5) is greater than or equal to the right
side of (5.6) for all z and y in D (F ). Thus, if we define

G (x ) = inf {p(x + z) - F (z) : z E D (F )},

it follows that (5.3) and (5.4) hold . But then

G(ax + z) :s ptax + z), z E D(F). (5.7)

To see this, suppose a > 0. Then (5.3) implies

G(ax + z) = aG(x + zj a ) :s apex + zj a) = ptax + z).

If a < 0, then (5.4) implies

G(ax + z) = -aG(-x - zja) :s -ap(-x - zla) = ptax + z) .

If o = 0, then
G(z) = F(z) :s p(z).
Finally, (5.7) implies (5.2) since

- G (a x + z) = G(-ax - z) :s pC-ax - z) = ptax + z).

To summarize, we have constructed aGE {; such that F :s G but F i= G . This
contradicts the maximality of F . Hence D (F ) = X and the theorem is proved for
X a vector space over the reals .
Suppose that X is a vector space over the complex numbers. We write

f (z)= mf (z )+i ~f (z) , z E M,

where m f (z) and ~ f (z) denote the real and imaginary parts of f (z), respectively.
Since ~ fez) = -m f tiz) , we have

f ez) = mfez) - i mf tiz). (5.8)

Let X, be X considered as a vector space over the reals. Now m f is a linear

functional on M, considered as a subspace of X r . Moreover,

1m f(z)1 :s If(z)1 :s p(z), z E M.

Hence, by the result we proved for real vector spaces, there exists a linear extension
G ofm f to all of X, such that IG(z)1 :s p(x), x E X», Guided by (5.8), we
define F on X by
F (x ) = G(x) - iG(ix ).
Now G is a linear extension of f to all of X. Writing F (x ) = IF (x )le i B, we get

IF (x)1 = F(e- iB x) = mF (e- iB x ) = G(e- i Bx ) :s p(e- i Bx ) = p (x ).

This completes the proof of the theorem. o

270 Chapter XI. Banach Spaces

Proof of the Hahn-Banach Theorem. Define p on X by p(x) = II I IIl1x II, and

apply the preceding theorem to I and p . 0
Now for some important applications of the Hahn-Banach theorem.

Theorem 5.3 Let M be a subspace of a normed linear space X. Given x E X

such that
inf IIx - ull = d(x, M) > 0,
ueM

there exists an I E X' such that

Ilfll = 1, I(M) = 0 and I(x) = d(x , M).

Proof: Since x f/: M, every Z E Mo = sp{x} + M is uniquely represented in the

form
Z = ax + u, a E C, u EM.
The function
g(ax + u) = ad, d = d(x, M),

is linear on Mo and for a i= 0,

lIax + ull = lallix + ~ull ~ laid = Ig(ax + u)l ·

Thus IIgll :::: 1. On the other hand, there exists a sequence {ud in M such that
IIx - Uk II -+ d. Hence

which implies that 1 :::: IIgli . Therefore, IIgll = 1. The theorem now follows by
extending g to I E X' such that 1 = IIgll = 11111 . 0

For M = (0), the above theorem gives the following result.

Corollary 5.4 Given x E X, there exists an I E X' such that 11111 = 1 and
I(x) = [x].

Corollary 5.5 For each x EX,

[x] = max{ll(x)1 : I E X', 11111 = I}.

Indeed, there exists agE X' such that IIgll = 1 and g(x) = IIxli. Thus for all
I E X', 1If11 = 1,
II(x)l::: IIxll = g(x) .
11.5 Hahn-Banach Theorem 271

Corollary 5.6 Given a linearly independent set {XI, . . . , Xn} C X, there exist
fi , h, . .. , In in X' such that

Ij(Xk) = Ojb 1 ~ j ,k ~ n.

Ifx E Sp{XI , . . . , Xn}, then x = L:J=I h(x)Xj.

Proof: Let M] = sp[ x, : i =1= j} , 1 ~ j ~ n. Since M j is a closed subspace

of X and Xj ¢ M], Theorem 5.3 ensures the existence of a gj E X' such that
gj(Xj) = dj =1= 0 and gjMj = O. The functionals Ij = fgj, 1 ~ j ~ n, have
J
the desired properties.
lfx = L:~=I ClkXk, then h(x) = L:~=l Clkh(Xk) = Clj.
The system {XI , . . . , x n }, {fJ , . . . , In} is called a biorthogonal system. 0

Definition: A closed subspace M of X is said to be complemented in X if there

exists a closed subspace N of X such that

(i) X = M +N = {u + v: u E M , v E N} and
(ii) M n N = (0) .
X is called the direct sum of M and N and is written X = M $ N .
It is clear that X = M $ N if and only if for each X E X there exists a unique
u E M and a unique v E N such that X = u + v.
A closed subspace M of a Hilbert space is complemented by M.1. However,
as we pointed out in Section 1, a closed subspace of a Banach space need not be
complemented, unless the subspace is finite dimensional.

Theorem 5.7 Every finite dimensional subspace of X is complemented in X.

Proof: Let XI, . .. , Xn be a basis for a subspace M of X . By Corollary 5.6, there

exist fJ, .. . , In in X' such that Ii (Xj) = oij. Now N = n7=1 Ker I i is a closed
subspace of X. Furthermore, M n N = (0) . For if Z E M n N, then from
Corollary 5.6 and the definition of N, we get Z = L:7=l/i(Z)Xi = O. Given
X EX, let u = L:7= I I. (X)Xi E M and v = X - u. Then X = u + v and v is in N
since
n
IkCv) = Ik(X) - Ik(U) = Ik(X) - L Ii (X)lk(Xi)
i=1
=lk(X)-lk(X) =0, 1~k~n .

Hence X = M $ N. 0

The direct complement of a subspace is not unique. For example, el = (l, 0)

is complemented in ((;2 by sp{v}, where v ¢ sp{ed .
272 Chapter XI. Banach Spaces

Exercises XI

Throughout these exercises, X and Y denote Banach spaces.

1. Check that the following vectors are in f p , for 1 :s p :s 00.

(a) ~ = (1, ~, i,k, I~"")
(b) ~ = (1, i,~, /6' }S' ···)
2. Let 1 :s PI
following :
:s P2 :s 00 and °:s a :s b < 00 . Prove or disprove the

(a) f p l S; f p2
(b) L pi [a, b] ;2 LpJa, b]
(c) II~ II PI :s II~ II p2 for ~ E f p l
(d) 1I<pllpl ~ 1I<pllp2 for <p E LpJa , b].

3. Let 1 s PI s P2 :s 00 . Prove or disprove the following :

(a) L p I (0, 00) S; t-», (0,00)
(b) L pz (0 ,00) S; L pi (0, 00)

4. Show that the intersection of the unit ball in foo with the plane

{(~I , ~2, 0, 0, . ..) I ~I, ~2 E ci

is the square

5. Find the intersection of the unit ball in f p , 1 :s P :s 00, with the plane

6. Prove that the unit ball in a nonned linear space is

(a) convex
(b) symmetric (if it contains x, then it also contains -x)
(c) closed

7. Suppose that in an n-dimensional nonned linear space X there is given a

convex, bounded, closed symmetric set S which is not contained in a proper
subspace of X . Define a function

II ·II:X---.+R+ by Ilxll=min{IAI:iES}.

Prove that II . II is a norm and (X, II . II) is a Banach space in which S is the
unit ball.
Exercises Xl 273

8. Prove that none of the following spaces is a Hilbert space, i.e., It IS

impossible to define an inner product on the space so that (x, x)I /2 is the
original IIx II.
(a) lp , P i= 2
(b) Lp[a, b], p i= 2
(c) qa , b].

9. Prove that a Banach space X is a Hilbert space ifand only iffor all x , Y EX,
2 2 2 2.
IIx + YII + IIx - YII = 211xll + 211Y1I

Hint: (x , y) = 1I!(x+Y)1I2-1I!(x-Y)1I2+ ill!(x+iY)1I2- ill!(x-iY)1I2 .

10. Prove that a Banach space X is a Hilbert space if and only ifthe intersection
of any plane sp{x, y}, x, Y E X with the unit sphere of X is a circle or an
ellipse.

11. Given the ellipsoid {( Xl ," " Xn ) E en : L7=II~t.s 1},introducean

inner product on en such that this ellipsoid becomes the unit ball.
12. Let F be the set of all ~ = (~l, ~2, . ..) E loo with ~l = O. Check that the
vectors yare closest in F to x.
(a) x = (2,0,0, ), Y is any element in F with IYj I .s 2, j 2: 2.
(b) x = (1, 1,0, ), Y is any element in F with IY2 -11 .s 1 and IYj I .s 1,
j = 3,4, .
(c) x = (1, !, ~, . . .), Y is any element in F with IYn - 2n~1 I .s 1, n 2: 1.

13. Given two vectors (1,1,0,0, . . .) and(~l, b ~3, ~4, 0, 0, ...) inl oo, where
o .s I~jl .s 1, j = 1,2,3,4, prove that the distance in loo between those
vectors does not exceed 2 and that there exist ~l, ~2, ~3, ~4 such that the
distance is exactly 2.
14. Show that there exist two vectors x and yin loo such that they are linearly
independent, IIxll = IIYII = 1 and IIx + YII = 2.
15. Prove that ifthe unit sphere ofa normed linear space contains a line segment,
then there exist vectors x and Y such that II x + Y II = II x II + II Y II and x, Y
are linearly independent (a line segment is a set ofthe form {au + (1 - a) v:
o .s a .s 1}.
16. Prove that if a normed linear space X contains linearly independent vectors
x and Y such that IIx + YII = IIxll + lIylI, then there is a line segment
contained in the unit sphere of X .
17. Prove that there are no line segments contained in the unit sphere of a
normed linear space if and only if the closest element in a subspace to a
given vector is unique.
274 Chapter XI. Banach Spaces

18. Prove that in £p, 1 < p < 00, there are no line segments contained in the
unit sphere.
19. Find the intersection ofthe unit ball in qo, 1] with the following subspaces:
(a) sp{t}
(b) sp{1, t}
(c) sp{l, t , t 2 }
(d) sp{l - t, t} .

20. Given two spheres IIx - Yoll = IIYoll and IIx + Yo II = IIYoll in a nonned
linear space, how many points can these spheres have in common?
21. Let ~ = (~I , ~2, . . .) be a vector in Co (the subspace of £00 consisting of
sequences which converge to zero). By renumbering the entries of ~ we
obtain ~* with I~il 2: 1~21 2: . .. . Let XI = {~ E Co : L~I tl~jl < oo}.
Define anonnon XI by IIHI = L~I tl~jl .
Prove that XI is a Banach space with respect to this norm, Let X(I) =
Ln_II;~1 (I)
{~ E col sUPn i~=, ~ < oo} and define a norm on X by

Prove that X(I) is a Banach space with respect to this norm. Prove that X I
and X(l) are not Hilbert spaces.
22 . Let M be a closed subspace of X. Define the quotient space XI M to be
the vector space consisting of the cosets [x] = {x + z : Z E M} with
vector addition and scalar multiplication defined by [x] + [y] = [x + Y],
a[x] = [ax]. Define lI[x]1I1 = d(x , M), the distance from x to M . Prove
that II . III is a norm on X 1M and that (X1M, II . III) is complete. We shall
refer to this space in XII exercise 17.
23. A Banach space is called uniformly convex if IIx n II = llYn II = 1 and
II~(Xn + Yn) II --+ 1 imply Xn - Yn --+ O.
(a) Prove that every Hilbert space is uniformly convex.
(b) It can be shown that the following inequalities hold for every ! and g
in Lp[a ,b].

II! + gliP + II! - gliP ~ 2 P- I(II!IIP

+ IIgIlP), 2 s p < 00 .

II! + gllq + II! - gllq ~ 2(lIf11 P + IIgIlP)q-l , 1 < p ~ 2,

.!.+1.=1.
P q

Prove that Lp[a, b], 1 < p < 00, is uniformly convex.

Exercises XI 275

(c) Give an example of a 2-dimensional Banach space which is not

uniformly convex.
24. Let S be a bounded closed convex set in the uniformly convex space X.
Prove that there exists a unique So E S smallest norm, i.e., lIso II < lis II for
all s E S, s i- so. Hint: Choose Sn E S so that Sn ---+ infsEs lisII and show
that {sn} is a Cauchy sequence.
°
25. Let X be uniformly convex. Show that for each I i- in X' , there exists
a unique x E X such that IIxll = 1 and 11/11 = I/(x)l .
26. Prove that the conjugate space of Co is l I.

27. Prove that a linear functional on X is bounded if and only if its kernel is
closed.
28. Prove that if X is separable, then there exists a countable set 11,12, ...
in X' with the property that if x i- 0, then there exists an Ij such that
h(x)i- 0.
29. Let rj = (a , (J, 0, . . .), Xz = (0, a , (J , 0, . . .), X3 = (0,0, a , (J , 0, .. .), . . . ,
where I~ I > 1. Show that {x j } is a Schauder basis for l z.
30. Let XI = (a, 0, . . .), Xz = (a, (J, 0, . ..), X3 = (0, a , (J , 0, . . .), . . . , where
I~I < 1. Show that {Xj} is aSchauderbasis forlz.

31. Let A E £(X) be invertible. Prove that {x j} is a Schauder basis for X if

and only if {Ax j} is a Schauder basis for X.
Chapter XII
Linear Operators on a Banach Space

Many of the definitions , theorems and proofs concerning operators in £(1iI , 1i2)
carry over verbatim to operators in £(X , Y) , where X and Yare Banach spaces.
Chapter II, Sect ions 1,3,8 and Theorems 16.1, 16.3 are illustrations of this asser-
tion . We shall refer to these results within the framework of Banach spaces even
though they were stated for Hilbert spaces .
There are also very substantial differences between operators on Banach and
Hilbert spaces. For instance, in general, there is no notion ofa selfadjoint operator
on a Banach space which is not a Hilbert space.

12.1 Description of Bounded Operators

The following result characterizes operators in £(£1) which stem from matrices.

Theorem 1.1 Let (ajkrj,k=1 be a matrix ofcomplex numbers. Define A on £1 by

A(~I , ~2 , . ..) = (lJi, 1/2, .. .) ,

where
00

TJj = L>jk~b j = 1,2, . . ..

k=1
The operator A is in t: (£ d if and only if
00

m = sup L lajkl < 00 ,

k j=1

in which cas e IIAII = m.

Proof: If A is in £(£d and {ekJ is the standard basis for £1, then
00

IIAII ::: IIAekll = l)ajkl, k = 1,2, . ...

j=1

Thus IIA II ::: m.

278 Chapter XII. Linear Operators on a Banach Space

On the other hand, if m < 00, then given; = (;\, ;2, ... ) E e\,

Hence IIAII ::: m. o

Theorem 1.2 Given a continuous complex valued/unction k defined on [a, b] x
[a, b], let K be defined on X = C([a, b]) by

(Kf)(t) = l b
k(t, s)f(s) ds.

The operator K is in L(X) and

IIKII = maxlb t a
Ik(t,s)lds.

Proof: It follows from the uniform continuity of k that K is a linear map from X

l
into X. Let
b
m = max Ik(t, s)] ds.
t a

l
Since
b
I(Kf)(t)1 ::: Ik(t, s)IIf(s)1 ds .s mllfll,
we have 11K II ::: m , Now

yet) = l b
Ik(t,s)lds

is continuous on [a, b] and therefore y attains its maximum at some point to E

[a, b]. Let
keto, s)
g(s) = ifk(to,s) =1= 0
Ik(to, s)1
and zero otherwise. Since g is bounded and Lebesgue measurable, it is in L \ ([a , b]) .
It follows from the fact that C ([a, b]) is dense in L \ ([a , b]) that there exists a
sequence {gn} in C([a, b]) such that

IIgnll = max Ign(s)l::: 1

se[a ,b]

and {gn} converges in L \ ([a, b]) to g. Hence

12.2 Closed Linear Operators 279

Therefore,
b
II K II ::: l k (to,s )g(s )ds= lb,k(to ,s),ds=m.

Let K be an integral operator with kernel function k E L 2([a , b] x [a, b]) .

Suppose that {gln}, {.A.n} is a basic system of eigenvectors and real eigenvalues of
K . The Hilbert-Schmidt Theorem V.l.l states that if

C = sup
t
I a
b
2
Ik(t , s)1 ds < 00 ,

then
(Kf)(t) = LAj(f, glj}glj(t) a.e., (1.1)
j

and the series converges uniformly and absolutely on [a, b]. We shall give the
following simple proof of this result.
Since
b )1/2
j(Kf)(t)j:::: (llk(t, s)12ds 1If112 s Cllfb
it follows that K is a bounded linear map from L 2([a , b]) into L oo([a , b]) . Now
we know that
n
sn (t) = (Pof) (t) + L ( f, gl j}glj (t)
j= l

converges in L 2([a , b]) to f . Therefore,

n
(Ksn )(t ) = LA j(f, glj}glj (t)
j =1

converges in L oo([a, b]) to Kf. The uniform convergence of the series in (1.1)
follows from the definition of the norm on L oo([a . b]). The absolute convergence
follows from the fact that the argument does not depend on the arrangement of the
terms of the series . 0

12.2 Closed Linear Operators

In Chapter VI we introduced closed linear operators and proved some results in the
setting of Hilbert spaces. These definitions and the proofs of the corresponding
results carry over verbatim to Banach spaces. For example, let X and Y be Banach
spaces. A linear operator A with domain V (A) ~ X and range 1m A ~ Y,denoted
by A (X ~ Y ), is called a closed operator ifit has the property that whenever {x n}
is a sequence in V(A ) satisfying Xn ~ x in X and AX n ~ Y E Y , then x E V(A)
and Ax = y .
280 Chapter XII. Linear Operators on a Banach Space

Examples: 1. Let X = Y = C [0, 1] and let A be defined by

D(A) = U E XII' E X}, AI = I' .

The linear operator A is unbounded. For if In (t) = t", then

II In II = max I/n(t)1 = 1 and IIAlnll = 1I/~1I = n.

tE[o.ll

To see that A is closed, suppose that In ---+ I and Aln ---+ g . Then it follows from
the definition of the norm on C ([0, 1]) that Un} and U~} converge uniformly on
[0, 1] to I and g, respectively. By considering J~ I' (s) ds, it is easy to verify that
I is differentiable on [0,1] and AI = I ' = g . Thus A is closed.
2. If A is invertible, then A is closed. The proof of this assertion is exactly the
same as the proof of Theorem VLl.l.
3. Let X = Y = Lp([O, 1]), 1 .:::: p < 00 and let A be defined by D(A) =
U E XII absolutely continuous on [0,1], I ' E X, 1(0) = o}, AI = I '.
A is obviously one-one and Im A = X since given g EX,

(A -I g)(t) = 1 t
g(s) ds.

Now by Holder's inequality,

lA-I g(t)1 .:::: ~

I Ig(s)1 ds.::::
(I)
~ Ig(sW
l ip
= IIgli p .

Hence

Thus A is closed by Example 2.

For a through treatment ofclosed differential operators in a Banach space setting,
the reader is referred to [G], [GGKl] and [DS 2].
Define the graph norm III1A on D(A) by

IIxli A = IIxll + [Ax],

If A is closed, then the argument used in Section VI.3 shows that D(A) with norm
II . II A is a Banach space.
The operator A is closed if and only if its graph G(A) = {(x, Ax)lx E D(A)}
is a closed subspace of X x Y with norm II (x , Y)II = IIxll + IIYII.
12.3 Closed Graph Theorem 281

12.3 Closed Graph Theorem

The closed graph theorem is another fundamental result in operator theory.

The theorem cannot be fully appreciated unless one sees some important applica-
tions - which we shall present in the next section.
Closed Graph Theorem 3.1 A closed linear operator which maps a Banach space
into a Banach space is continuous.
It should be stressed that the closed graph theorem requires that the domain
of the operator be complete. The differential operator given in Example 1 of the
preceding section is closed but unbounded. Its domain is a proper dense subspace
of the Banach space C([O, 1]) .
To prove the theorem we need the Baire category theorem (Theorem 3.2 below)
and a lemma. In the proof we use the following notation. Given Xo in a normed
linear space X and r > 0,

S(xo,r) {X: IIx - xoll < r}

S(xo ,r) {x: IIx -xoll.::: r}.

For Z C X and a E C
a Z = {a z : z E Z}.
A vector Xo is called an interior point of a set Z C X if S(xo , r) C Z for some
r > O.
Baire Category Theorem 3.2 If a Banach space is the union of a countable
number ofclosed sets, then at least one ofthe closed sets has an interior point.

Proof: Let the Banach space X = UnCn, where each C; is a closed set. Suppose
that none of the C, has an interior point. Choose XI E Cl. Since S'(xj , 1) ~ Cl
and Cl is closed, there exist X2 and ri , 0 < r: < ~, such that

Since S(X2, r2) rt. C2 and C2 is closed, there exist X3 and r3, 0 < rs < ~, such
that
and S(X3, r3) n C2 = 0.
*'
S(X3, r3) C S(X2, r2)
Continuing in this manner, we obtain sequences {x n } and {r n }, 0 < r« < such
that
S(Xn+l, rn+l) C S(x n, rn) and S(Xn+l, rn+l) n c, = 0. (3.1)
The sequence {x n } is a Cauchy sequence. For if n > m, then we have from (3.1)
that X n E S(xm , rm ) , i.e.,

I
II x n -xmll < Tm < -. (3.2)
m
282 Chapter XII. Linear Operators on a Banach Space

Hence {x n } converges to some x E X. Fix m and let n --+ 00 in (3.2). Then

IIx - Xm II :s rm, i.e., x E S(xm, rm) which is disjoint from Cm-I. m = 2, ... .
But this is impossible since X = Un Cn . 0

Lemma 3.3 Suppose C is a convex set in X and C = (-l)C . /fC has an interior
point, then zero is also an interior point ofC.

Proof: Suppose S(xo , r) C C. If [x] < 2r, then

x = (xo +~) - (xo -~) E S(xo, r) + (-l)S(xo , r) C C + C.

But C +C = 2C. Indeed, given u and v in C,

Since C is convex. We have shown that S(0,2r) C 2C, which implies that
S(O,r) c C. 0

Proof of the Closed Graph Theorem: Let T be a closed linear operator which
maps the Banach space X into the Banach space Y . Define Z = {x: IITxll < I}.
First we prove that the closure 2 of Z has an interior point.
Since D(T) = X and T is linear, X = U~lnZ. It follows from the Baire
Category Theorem that there exists a positive integer k such that kZ = kZ has
an interior point. Therefore, 2 has an interior point. It is easy to verify that 2
is convex and 2 = (-1)2. By Lemma 3.3 S(O, r) C Z for some r > 0, which
implies
S(O, ar) C a2 = «z , a > 0. (3.3)

°
Given < e < I and IIx II < r, we have from (3.3) that x is in 2 . Therefore , there
exists an XI C Z such that IIx - xIII < er. Since x - XI E S(O, r) C e Z, there
exists an X2 E e Z such that

Inductively, there exists a sequence {x n } such that

(3.4)

Let Sn = L:Z=I xi , From (3.4) and the definition of Z we get

(3.5)
12.4 Applications ofthe Closed Graph Theorem 283

Now {Tsn } is a Cauchy sequence since (3.5) implies that for n > m,
00 00 m
IITsn-Tsmll~ L IITXkll~Lck=lc_c--+O
k=m+l k=m

as m --+ 00. Hence, by the completeness of Y, TSn --+ Y for some y E Y . So, we
have Sn --+ x and TSn --+ y . Since T is closed, Tx = y. Thus

00 I
IITxll = IIYII = n--+oo
lim IITs nll ~"IITxkll ~
L...J
--,
I - e
k=1

whenever IIx II < r. In particular, if II v II = I, then

Thus
2

12.4 Applications of the Closed Graph Theorem

In this section we present a number of applications of the closed graph theorem.

The first concerns invertibility,
Recall that an operator A E L(X, Y) is invertible if each of the conditions (a)
Ker A = (0) (b) Im A = Y (c) A-I is bounded hold . The next theorem shows
that (a) and (b) imply (c).

Theorem 4.1 Suppose X and Yare Banach spaces. If A(X --+ Y) is closed with
properties Ker A = (0) and 1m A = Y, then A -I is bounded on Y .

Proof: SinceG(A) = {(x , Ax): x E X}isclosedinXxY,G(A- I) = {(Ax , x) :

x E X} is closed in Y xX, i.e., A -I is a closed linear operator mapping Y into X.
Therefore, A -I is bounded by the closed graph theorem. 0

Corollary 4.2 Suppose 11111 and 11112 are norms on the vector space X such that
(X, 1111 d and (X, 11112) are complete. If there exists a number C such that

IIxlll ~ ClJxll2 forall x E X,

then 11111 and 11112 are equivalent, i.e., there exists CI > 0 such that CJllx112 :s
IIxlll :s ClJx1l2.
284 Chapter XII Linear Operators on a Banach Space

Proof: Let / be the identity map on X . Now / is a bounded linear map from
(X , II liz) onto (X , II lid since
II/xIII = IIxlll ::s Cllxllz.
Hence l : I is bounded by Theorem 4.1 , and
I lllllxlli .
C"x11I s l
IIxliz = IIr xllz::s IIr
o
We now give two proofs of the following fundamental result. One proof relies
on the closed graph theorem.

Theorem 4.3 (Uniform Boundedness Principle). Given Banach spaces X

and Y, suppose F is a subset of£(X , Y) with the property thatfor each x E X ,
SUPAeF II Ax II < 00. Then sUPAeF IIAII < 00.
Proof: Suppose sUPAeF IIAII = 00. Then for each positive n there exists an
operator A E F, denoted by An, such that II An II ~ n. However, the exact proof
of Theorem II.14 .3 with the Hilbert spaces 11. I , Hz replaced by the Banach spaces
X, Y , respectively, shows that sUPn II An II < 00 . This contradicts IIA nll ~ n,
n = 1,2, . . . .
Another proof of Theorem 4.3 is the following.
Let B(F, Y) denote the vector space of functions f which map F into Y and
have the property that
1If11 = sup IIf(A)1I < 00.
AeF
Then II II is a norm on B (F , Y) and by the argument which was used to prove
that £(X, Y) is complete (Theorem XU.I), we have that B(F, Y) is complete.
Define T : X --+ B(F, Y) by
(Tx)A = Ax , x E X.
It is not difficult to verify that T is a closed linear operator on X . Hence the
boundedness of T is ensured by the closed graph theorem. Thus for all A E F
and x E X,
IIAxll = II(Tx)AII ::s IITxll ::s IITllllxll,
which shows that
sup IIAII ::s II TIl ·
AeF
o
Corollary 4.4 Let X and Y be Banach spaces. Suppose {An} C £(X, Y) is a
sequence such that {Anx} converges for each x E X. Then the operator
Ax = lim Anx , x E X
n->oo
is linear and bounded and the sequence {II An II} is bounded.
12.4 Applications ofthe Closed Graph Theorem 285

Proof: For each x E X, the sequence {Anx} is bounded since it converges . Thus
by the uniform boundedness principle, sUPn IIA n II = m < 00 and

[A»] = lim IIA nxll :::: mllxll ,

n->oo

which shows that IIAII :::: m. o

A simple application ofthe uniform boundedness principle yields the following
result.

Theorem 4.5 Suppo se that S is a subset ofa Banach space X such that for each
I E X',
sup I/(x)1 < 00 .
XES

Then S is bounded.

Proof: For each X E S , define the linear functional F, on the conjugate space X'
by F; I = I (x) . Clearly, F; is linear and by Corollary XI.5.5

IlFxll = sup IFA!) I = sup I/(x)1 = [x]: (4.1)

"1"=1 "1"=1
Thus F; is a bounded linear functional on the Banach space X' and, by hypo-
thesis, for each g E X ',

sup IFx(g)1 = sup Ig(x)1 < 00.

XES XES

Hence by (4.1) and Theorem 4.3,

sup IIxll = sup IIFxll < 00.

XE S X ES

o
Another application of the uniform boundedness principle gives the following
result. If fJ = {fJl, fJ2, .. .} is a sequence of complex numbers such that the
series L~1 fJA j converges for every {~j} E lp, 1 .s p < 00, then fJ is in
lq, t+~ = 1. To prove this, define I on lp by I({~j}) = L~1 fJj~j. Let
In({~j}) = LJ=1 fJj~j, n = 1,2, . . . . Clearly, In is in the conjugate space l~
and In(x) -+ I(x) for each x E i . Hence I is in l~ by Corrollary 4.4 and, by
Theorem XI.4 .1, fJ is in lq with IIfJll q = Ilfll.

Theorem 4.6 An operator A E £(X , Y) is one-one and has a closed range if and
only if there exists an m > 0 such that

IIAxl1 2: mllxll forall x E X.

The proof is the same as the proof of Corollary 11.14.2.

286 Chapter XII. Linear Operators on a Banach Space

12.5 Complemented Subspaces and Projections

Just as every closed subspace of a Hilbert space has a projection associated with
it, so does every closed, comp lemented subspace of a Banach space.
Definition. Let M be a subspace of X. An operator P is called a projection from
X onto M if it is a bounded linear map from X onto M and p 2 = P .
If x is in M, then P x = x . Indeed, there exists a z E X such that x = P z.
Hence Px = p 2z = P': = x .
It is easy to see that if P is a projection then Q = I - P is also a projection and
Im P = Ker Q, Ker P = Im Q. Hence Im P is closed.

Theorem 5.1 A closed subspace M ofa Banach space X is comp lemented in X

if and only if there exists a projection from X onto M .
If M is complemented by the closed subspace N , then there exists a c > 0 such
that
lIu + vII 2: cllull, for all u EM andv E N.
Proof: Suppose X = M $ N. Then given x EX , there exists a unique u E M
and a unique v E N such that x = u + v. The operator P defined by P x = U is
a linear map from X onto M and p 2 = P . We now prove that P is bounded by
showing that it is closed.
Suppose
Xk = Uk + Vk --+ x, Uk EM, Vk E N
and PXk = Uk --+ y . Now Y is in M since M is closed. Therefore y = Py and
Vk --+ x - y. Since N is also closed, x - y is in Nand

0= P(x - y) = Px - P y = Px - y .

Thus P is closed. Hence P is bounded and

IIPliliu + vII 2: IIP(u + V) II = lIulI ·

Conversely, if P is a projection from X onto M, then N = Ker P is a closed
subspaceofX. Furthermore, M n N = (0) since x E M n N implies x = Px = O.
Now x = Px + (x - Px) and x - Px is in N. Thus X = M $ N . 0
Example: Let 1£be a Hilbert space . We shall now construct two closed subspaces
L 1 and L2 with the following properties:

and
L1 $ L2 = H, L1 $ L2 i- H,
where the left side means the closure of L1 $ L2.
12.5 Complemented Subspaces and Projections 287

This construction allows one to construct a linear operator no such that no is

densely defined, n 5 = no and no is not bounded . In other words, no could be
considered as an unbounded projection.
Indeed, let us assume that the subspaces L) and L z are already constructed.
Define the operator no by the equality

no (x + y) = x, X E L) , Y E L z .

It is obvious that
n5 = TIo.

Now we have to prove that no is unbounded on L) EB Lz. We prove it by

contradiction. Assume that no is bounded . Then

IIxll = II no(x + y)1I s Il TIollllx + yll ,

also
lIyll = 11(1 - no) (x + y)1I :s III - nollllx + yll .
Let U n = X n + Y n» X n EL I , Y n E L z , n = 1, 2, . . . be a convergent sequence from
LI EB L z. Denote by U its limit (in norm). From the above inequalities it follows
that the sequences {xnlf and {Ynlf are also convergent. Denote by x E L) and
Y E L z their respective limits. It is clear that u = x + Y and hence L I EB L z is
closed and coincides with H. This contradict the properties of L I and L z . We
conclude that no is unbounded.
Now let us return to the construction of L ) and L z.
Let H be a separable Hilbert space, and let CPI , cpz , . .. be an orthonormal basis
in H . Define L I to be the closed linear span of the vectors

cp ), CP3, CPs , ··· ,

and let Lz be the closed linear span of the vectors

It is clear that the first set of vectors is an orthonormal basis for L), while the
second set is an orthogonal basis for L z .
The intersection LI n Lz consists of zero only. Indeed, let f E L) n L z . Then
f can be represented as
00
1
f = 2::>jCPZj-) , f = fJj(CPZ j-1 + 2.CPj) ,
j=) j=) ]
288 Chapter XII Linear Operators on a Banach Space

with the two series converging in the norm of H . From these equalities it follows
that
00 00 1
L(aj - Yj)({J2j-1 = LYj 2j({J2 j .
j=1 j=1

Notice that in this equality the left hand side is in L I while the right hand is
orthogonal to LI . This leads to aj = Yj and Y) = 0 for j = 1,2, . . .. Thus
f = 0, and hence LI n L2 = {OJ.
Also, LI 6' L2 = H. For suppose this is not the case . Then there exists a u =1= 0
in H such that u l.. L, and u l.. L2. But then u l.. ({JI , ({J2, . .. which implies that
u = O. Now assume L I 6' L2 = H. Take

X n = ({J2n-l , n = 1,2, . ..

and
1
Yn = ({J2n-1 + 2n ({J2n, n = 1,2, . .. .
It is obvious that Ilxn II = 1, llYn II = /1 + 2~n and

· IIX - Yn II = l'1m - 1 = O.
11m
n--+oo
n n
n--+oo 2
But this is impossible since by Theorem 5.1, there exists a c > 0 such that

IIx n - Ynll 2: CIIxn ll 2: C, n = 1,2, ...

12.6 One-Sided Invertiblity Revisited

In this section we extend the results of Section 11.15 to Banach spaces X and Y.
The definition of one-sided invertibility for an operator on a Banach space is the
same as for an operator on a Hilbert space (cf. 11.15)

Theorem 6.1 A necessary and sufficient condition for an operator A E £(X, Y)

to have a left inverse is that Ker A = {OJ and 1m A is closed and complemented in
Y . In this case, the operator AL is a projection onto 1m A and Ker AL = Ker L,
where L is a left inverse ofA.

Proof: The proof of the necessity is exactly the same as the proof of
Theorem II. 15.2. SupposeKer A = {OJ and Y = 1m A6'M, where 1m A andM are
closed subspaces of Y. Let P be the projection from Y onto 1m A with Ker P = M
and let A I be the operator A considered as a map from X onto the Banach space
1m A . Then Al is invertible by Theorem 4.1 and All P is a bounded left inverse
of A since AllPAx = All Ax = x . 0
12.7 The Projection Method Revisited 289

Theorem 6.2 A necessary and sufficient condition for an operator A E L(X, Y)

to have a right inverse is that 1m A = Y and Ker A is complemented in X. In this
case, 1 - D (- I ) D is a projection onto Ker A , where D (- 1) is a right inverse ofA.

Proof: The proof of the necessity is exactly the same as the proof of
Theorem II.I5.4. Suppose now that 1m A = Y and X = Ker A E9 N , where
N is a closed subspace of X. Let P be the projection of X onto N with Ker
P = Ker A . Let Al be the operator A restricted to N. Then Al is injective and
AIN = AX = Y . Hence Al is invertible and AAily = y.
The proofs of the following theorems are exactly the same as the proofs of the
corresponding Theorems II.I5.3 and II.15.5. 0

Theorem 6.3 Let A E L(X, Y) have a left inverse L E L(Y, X) . If B is an

operator in L(X, Y) and II A - B II < II L -III-I, then B has a left inverse L , given
by

LI ~ L(l- (A - B)L)-I ~ L (t,[(A - B)Ll) k (6.1)

Moreover,
codim 1m B = codim 1m A . (6.2)

Theorem 6.4 Let A E L(X, Y) have a right inverse R. If B is an operator in

L(X, Y) and II A - BII < IIRII - I, then B has a right inverse RI E L(Y, X) given
by
RI = (1- R(A - B»-I R. (6.3)
Moreover,
dim Ker A = dim Ker B. (6.4)

12.7 The Projection Method Revisited

Definition: Let X and Y be Banach spaces and let {P n ), {Qn} be sequences of

projections in L(X) and L(Y), respectively, with the properties that Pi,» -+ x for
all x E X and QnY -+ y for all y E Y. Given an invertible operator A E L(X, Y),
the projection method for A seeks to approximate a solution to the equation Ax = y
by a sequence {x n } of solutions to the equations

QnAPnx = QnY , n = 1,2, . . .. (7.1)

The projection method for A is said to converge if there an integer N such that for
each y E Y and n ::: N , there exists a unique solution X n to equation (7.1) and, in
addition, the sequence {x n} converges to A -I y . We denote by IT(P n, Qn) the set
of invertible operators for which the convergent method converges.
The projection method for A was presented in Section II.I7 for X = Y a Hilbert
space and Qn = Pn.
290 Chapter XII. Linear Operators on a Banach Spa ce

Theorem 7.1 Let A E L(X, Y) be invertible. Then A E IT(Pn , Qn) ifand only if
there exists an integer N such that for n 2: N,

(i) the restriction of the operator QnAPn to 1m Pn has a bounded inverse on

1m Qn, denoted by (QnAPn)-I , and
(ii) sUPn :::N II (QnAPn)-1 II < 00.

Proof: By arguing as in the proof of Theorem 11.17.1 with PnAPn replaced by

QnAPn, we obtain Theorem 7.1. 0

Corollary 7.2 Let {P n} be a sequence ofprojections on a Banach space X with

II Pn II = 1, n = 1,2, .. .. Assume Pnx -+ x for all x E X. Suppose A E L(X)
with III - All < 1. Then A E ITn(Pn , Pn) .

The proof of the corollary is exactly the same as the proof of Corollary 11.17.2.
The proof of the next theorem is analogous to the proof of the corresponding
Theorem 11.17.6 with PnAPn replaced by QnAPn.

Theorem 7.3 Suppose A E IT(Pn, Qn)' There exists a y > 0 such that if B E
L(X, Y) and liB II < y , then A + BE IT(Pn , Qn) .

12.8 The Spectrum of an Operator

In this section we extend the notion of the spectrum of an operator (which was
introduced in Section 11.20 for Hilbert space operators) to a Banach space setting.
Throughout this section X is a Banach space.
Definition: Given A E L(X), a point x E C is called a regular point of A if
AI - A is invertible. The set peA) of regular points is called the resolvent set of
A . The spectrum a(A) of A is the complement of peA) .

Theorem 8.1 The resolvent set ofany A E L(X) is an open set in C containing
().. 11)..1
> IIA II}· Hence a(A) isa closed bounded set contained in {).. 11)..1 ~ IIA II}.
Furthermore.for): E oa(A), the boundary ofo (A) , the operator AI- A is neither
left nor right invertible.

Proof: The proof is exactly the same as the proof of Theorem 11.20.1 0

Examples: 1. If A is a compact self adjoint operator on an infinite dimensional

Hilbert space, then a(A) consists of zero and the eigenvalues of A by Theorem
IY.S.1. The formula for (AI - A)-l is given in the same theorem.
2. Let X = C([c, d]) . Define A E L(X) by

(Af)(t) = aCt) f (r) ,

12.8 The Spectrum ofan Operator 291

where a(t) E X . For A =1= a(t), t E [c, d], it is clear that X E p(A) and

1
«U - A)-lg)(t) = g(t).
A - a(t)

If A = a(to) for some to E [c, d] , then A E a(A) . Indeed, suppose A E p(A).

Then for some f E X and all t E [c, d],

1 = «U - A)f)(t) = (a(to) - a(t»f(t) .

But this is impossible since 1imH'o(A - a(t»f(t) = O. Hence

a(A) = {a(t) , c :s t :s d} .
3. For X = lp, 1 :s p < 00, let A E £(X) be the backward shift operator,

A(al, a2, . . .) = (a2, a3,· · .).

Clearly, IIA II = 1. Thus if IAI > 1, then A E p(A) by Theorem 6.1. If AI < 1,
then x = (l, A, A2 , . . .) is in Ker (U - A) . Therefore, A E a (A) . Since a (A) is
a closed set, it follows that a(A) = {A : IAI :s 1} .
If IAI > 1, then

k
1 ( A)-l 00 A
(U - A)-l = - 1 - - = " - .
A A LJ Ak+l
k=O

Thus given y = ({31, {32, .. .) EX, (U - A)-l Y = (ai , a2, .. .), where

The spectrum of the backward shift operator on lp, 1 :s p < 00, was shown
above to be independent of p . The following example shows that the spectrum
can vary as the space varies.
4. Given a number q > 0, let ll(q) be the set of those sequences x =
(ai, a2, . ..) of complex numbers such that

00

II xll q = L lajl q-j < 00.

j=l

with the usual definitions of additions and scalar multiplication, II (q), together
with the norm IIl1q , is a Banach space.
Let A be the backward shift operator on II (q). We shall show that a(A) =
{A : IAI :s q}. If IAI < q, then x = (l, A, A2 , ... ) is in Ker(U - A) . Thus
292 Chapter XII. Linear Operators on a Banach Space

A E a(A). Suppose IAI > q. Given y = <fh,lh, .. .) E .fl(q), we wish to

determine x = (ai, az, ...) such that (AI - A)x = y . If this x were to exist, then

In particular,

az = Aal -Ih, a3 = Aaz -Ih = AZal - Af31 - f3z·

By induction,

n-l
an+l = Anal - LAkf3n-k (8.1)
k=O
and
n-l
al = A., -n an+l + ""'
L..... A., k-n pn-k·
R (8.2)
k=O
Since x is to be in.fl (q) and IAI > q,

n-l n-l 00

L IA k-nf3n_kl ::s L lf3n-kll- n ::s L If3Jlq -J < 00. (8.4)

k=O k=O j=O
It follows from (8.2)-(8.4) that we should take
00

al=LA- Jf3J (8.5)

j=l

and an , n > 1, as defined in (8.1). This argument also shows that for each
y E .fl(q) there exists a unique x E .fl(q) such that (AI - A)x = y, provided
IAI > q. Formulas (8.1) and (8.5) give (AI - A)-l y. Thus, since a(A) is closed
and contains {A : IAI < q},

a(A) = {A : IAI ::s q}.

We know that ifjA II < 1, then I -A is invertible, i.e., 1 E peA) and (l-A)-l =
L~o A k . It is possible for the series to converge even though II A II ~ 1, in which
case I - A is invertible. This may be seen in the next section. Let us start with
the following theorem.

Theorem 8.2 Suppose A is in LeX) and L~o A k converges in LeX). Then I - A

is invertible and (l - A)-l = L~o A k .
12.9 Volterra Integral Operator 293

Proof: Let Sn = L~=o A k and S = lim n ---+ oo Sn. Then (l - A )S = lim n ---+ oo
(l - A )Sn = lim n ---+ oo Sn(l - A) = S (l - A ).
Since (l- A )Sn = I - A n+1 --+ I ,

(l - A )S = S(l - A) = I.
Thu s
00

L A k = S = (l- A )-I.
k=O

Corollary 8.3 Suppose A E £ (X ) and IIAPII < 1 for some p ositive integer p .
Then L~o A k converges, I - A is invertible and (l - A )-I = L~o A k.

Proof: Let
00

Sk = LII Anp+k ll , k= l , 2, ... , p.

n=O
Each Sk is finite since
00 00

L II Anp+k II ::: IIA k ll L II A P li n < 00 .

n=O n=O

Hence
00

L IIAj II = SI + S2 + ... + Sp < 00 .

j= 1

Thus L 1==OA j converges since £ (X ) is complete (Theorem XU .1) and

n n
L Aj ::: L IIA jll --+ 0 as m ,n --+ 00 .
j=m j=m

The corollary now follows from Theorem 8.2. D

12.9 Volterra Integral Operator

Suppose k (t , s) is continuous on [0, 1] x [0, 1]. We shall show that the equation

Af(t) -1 1
k (t , s) j(s) d s = g(t)
294 Chapter XII Linear Operators on a Banach Space

has a unique solution in C ([0, 1]) for every g E C ([0, 1]) if).. i= O. In other words,
if Y is the Volterra integral operator on C ([0 , 1]) defined by

(Yf)(t) 1 1

k(t, s)f(s) ds,

then a(Y) = {O}.

First we note that Yf is in C([O, 1]) for each f E C([O, 1]). This follows from
the uniform continuity of k(t , s) and the equation

(Y f)(td - (Y f)(t) = Jot

l
[k(tl, s) - k(t , s)]f(s) ds + III
I k(t , s)f(s) ds .

For /-L = max0:9.s:::;1 Ik(t, s)l,

I(Yf)(t)1 S 1 1

Ik(t, s)llf(s)1 ds S /-Ltllfll

By induction,

Hence
k
lIy
kll
S ~! ~ O.

Thus it follows from Corollary 8.3 that I - Y is invertible and (I - y)-I

L~o v- .
Ifwe replace k(t, s) by tk(t, s),).. i= 0, we get the invertibility of 1- t . Thus
H - Y is invertible and

1( y)-I yk
vr:'
00
(H - = -).. 1 - )..
- = " '-.
~ )..k+1
k=O

To find a formula for v', let kl (t, s) = k(t, s), if a ss Stand zero otherwise.

1
Then
1
(Yf)(t) = kl(t,s)f(s)ds

and y k can be obtained from n .lO, formula (10.3).

t
Note that II II can be made as large as we please, yet L~o( t)k converges.
12.10 Analytic Operator Valued Functions 295

12.10 Analytic Operator Valued Functions

Definition : An operator valued function A(A) which maps a subset of'C into £(X)
is analytic at AO if

A(A) = Ao + (A - Ao)AI + (A - AO)2 A2 + ...,

where each Ak is in £(X) and the series converges for each A in some neighborhood
ofAo.

Theorem 10.1 Thefunction A(A) = (AI - A)-I is analytic at each point in the
open set p(A).

Proof: Suppose AO E p(A) . Now

AI - A = (AoI - A) - (AO - A)I = (AoI - A)[I - (AO - A)A(AO)] · (10.1)

Since p(A) is open, we may choose e > 0 so that IA - Aol < e implies A E p(A)
and II(A - Ao)A(Ao)1I < 1. In this case, it follows from (10.1) that
00

A(A) = [I - (AO - A)A(AO)rl(AOI - A)-I = L (AO - A)kA(AO)k+I. (10.2)

k=O

From the series representation of A(A), it is not difficult to show that

d
-(AI - A)-I = -(AI - A)-2.
dA

The function A(A) = (AI - A)-I is called the resolventfunction of A , or simply,

the resolvent of A . 0

Theorem 10.2 The spectrum ofany A E £(X) is non-empty.

Proof: Suppose a(A) = 0. Then by Theorem 10.1 formula (10.2), with AO = 0,

00

(AI - A)-I = L(-l)k Ak(A- I l +I for all A E Co (10.3)

k=O

By Corollary XI.5.4, there eixst x' E X' and x E X such that x'x f= O. From
(10.3) it follows that
00

f(A) =x'(AI - A)-Ix = L(-l)k Akx'(A- I)k+l x, A E Co

k=O
296 Chapter XII. Linear Operators on a Banach Space

Hence I (A) is a complex valued function which is analytic on C, i.e., I is an entire

function. Moreover I is bounded. Indeed , since

II(Al - A)»] 2: (IAI- II AII) IIxll,

it follows that for IAI > IIAII,

II(Al - A)-Ixll < IIxll ~ 0 as IAI ~ 00 . (1004)

- IAI-IIAII
Since I(A) is continuous on {A I IAI :::: IIAII}, it is bounded on this set. Thus
I is entire and bounded on Co By Liouville 's theorem, I is a constant function.
Therefore x'x = 1(0) = I(A) ~ 0 by (10.3) . Hence x 'x = 0 which contradicts
our choice of x and x' . We have shown that a(A) =1= 0. 0

Exercises XII

Throughout these exercises, X and Y denote Banach spaces.

I . Let B be a Banach space with a given norm II . II. Check that the following
norms II . III are equivalent to II . II ·
(a) B = l2 ; 1I~lIi = L~II(1 + J)~jI2.
(b) B = L2[O, I] ; III/II = 11(2/ + V) !II , where VI(t)
= f~ I(s) ds .
(c) B = lp(Z), I :::: p :::: 00 ; U : lp(Z) ~ lp(Z) is given by

U(... ~- I,~O,~I, ... ) = (. .. ~-2 ,~-I,~O,~I, . ..)

t t
O-place O-place
1I~1I1 = IIU~II ·

(d) Let S be a bounded invertible operator on B and let IIx III = II Sx II .

2. Let {x n } be a Schauder basis for X. For k = I , 2, . . . , define the coor-
dinate functional Ik on X by Ik(X) = CXk, where x = Ln CXnXn . Thus
x = Ln In (x)x n, x E X. Prove that each /k is bounded by establishing
the following :
(a) Given x = Lk CXkXk, define IIx 1100 = sup, I LZ=I CXkXk I . Then II . 1100
is a norm and (X, II . 1100) is complete.
(b) The norms II . II and II . II 00 are equivalent.
(c) Each Ik is linear and bounded on (X, II . II).

3. Let A be a linear operator which maps a Hilbert space H into H . Prove

that if (Au , v) = (u , Av) for all u, v in H , then A is continuous.
Exercises XII 297

4. Let A be a linear operator which maps X into f. Prove that A is continuous

if and only if y' 0 A is continuous for every y' E f' .
5. Assuming X is a Hilbert space, prove that the Hahn-Banach Theorem can
be generalized as follows : If M is a subspace of X and A is in L(M, f),
then A can be extended to an operator A E ax, Y) such that II AII = II A I .
6. Let M and N be closed subspaces of X with M n N = (0) . Prove that
M EB N is closed if and only if there exists a number c > 0 such that
II u + v II ~ ell u II and II u + v II ~ cII v II for all u EM, v EN.
7. Let Z be a Banach space and let B E 'c(f, Z) be injective. Suppose A
is a linear operator which maps X into f and has the property that BA is
bounded. Prove that A is bounded.
8. Let T be in 'c(X , Z) and let A be in 'c(f, Z), where Z is a Banach space.
Suppose that for each x EX, there exists a unique y E f such that T x =
Ay. Define Bx = y . Prove that B is in 'c(X , f).

9. Let {{3k} be a bounded sequence ofcomplex numbers. Define T : l p --+ l p,

1 S p S 00, by T (ai, az , . . .) = ({31 ai , {3zaz, ...). Determine aCT) ·
10. Given A E 'c(X), prove that
(a) limp.l---+oo(U - A)-I = 0 and
(b) limlAI---+oo )..(U - A)-I = I .
11. Suppose {An} is a sequence ofinvertible operators in 'c(X) which converges
to A. Prove that A is invertible if and only if {II A; III} is bounded, in which
case A;I --+ A -I.
12. Given A E 'c(X) and given a polynomial p, prove that a(p(A)) = {p()..) :
).. E a (A)} == pa(A) . Hint: Writep()..)-J-L = cIl7=1()..-)..i) and consider
Ili=1 (A - )..i l).

13. Let T : II --+ II be given by T(al, az , . . .) = (0, at, ~2 , i ,...).

(a) Find IIT nll, n = 1,2, . . .
(b) Determine aCT) .

14. Given A E 'c(X) , let A()") = ().. - A)-I, ).. E peA) .

(a) Prove that A()") - A(J-L) = (J-L - )..)A()..)A(J-L), ).. , J-L E peA) .
(b) Prove that fAA()") = -A()..)z .

15. Let A()") and B()") be analytic operator valued functions defined on a set
S C Co Find the derivatives of A()..)-I, A()")B()") and A ()..)-I B()") at the
points where the derivatives exist.
16. Suppose A E 'c(X) is injective . Prove that the range of A is closed if and
only if there exists a y > 0 such that II Ax II ~ y IIx II for all x EX.
298 Chapter XII. Linear Operators on a Banach Space

17. For A E L(X, Y) define A : X/Ker A --+ Y by A[x] = Ax (cf. IX,

exercise 22, for the definition of X/ M).
(a) Prove that A is linear, injective and IIAII = IIAII.
(b) Prove that the range of A is closed if and only if there exists a y > 0
such that [Ax] ::: yd(x, Ker A) for all x E X.
18. (a) Let T be in L(X, Y). Suppose there exists a closed subspace N of
Y such that the direct sum Im(T) $ N is closed. Prove that 1m T is
closed. Hint: Define To on X x N by To(x, z) = Tx + z; then use
exercise 17(b).
(b) Prove that if A E L(X, Y) and codim 1m A < 00, then 1m A is closed.
Chapter XIII
Compact Operators on Banach Spaces

In this chapter, we present some basic properties ofcompact operators in L(X, Y) ,

where X and Y are Banach spaces. The results enable us to determine conditions
under which certain integral equations have solutions in Lp([a, b]), 1 < p < 00
or C([a , b]) .
An operator K E L(X, Y) is compact iffor every sequence {x n } in X, IIxn II = 1,
the sequence {Kx n } has a convergent subsequence.
Theorems 11.16.1, 11.16.3 and the corresponding proofs remain the same when
the Hilbert spaces are replaced by Banach spaces.

13.1 Examples of Compact Operators

1. Every K E L(X, Y) which is offinite rank is compact. For if {x n } is a sequence

in the l-sphere of X, then {Kx n } is a bounded sequence in the finite dimensional
space. 1m K . Hence {Kx n } has a convergent subsequence by Theorem XI.2.2.
2. Let {13k} be a sequence of complex numbers which converges to zero .
Define K E L(f p ) , 1 S P < 00 by K({ad) = {f3kad . Let Kn(aj, a2, . ..) =
(131 , 132, ... , f3n, 0, 0, ...). Since K n is a bounded linear operator on f p of finite
rank, it is compact. Moreover,

IIK-Knllssuplf3kl~O asn~oo.
k>n

Hence K is compact.
3. Suppose k is continuous on [a, b] x [a, b] . Let K : C([a, b]) ~ C([a, b])
be the linear integral operator defined by

(Kf)(t) = l b
k(t , s)/(s) ds.

We shall now prove that K is compact.

It follows from the uniform continuity of K that 1m K C C([a , b]) . K is
bounded since

11K/II = max I(Kf)(t) I ~ IIf11(b - a) max Ik(t, s)l . (1.1)

a=st::s.b a::s.s,t::s.b
300 Chapter XIIl Compact Operators on Banach Spaces

Since k is continuous, there exists a sequence {Pn } of polynomials in t and s

which converges uniformly to k on [a, b] x [a, b]([R], p. 174). Define K; on

l
C([a, b]) by
b
Kn/ = e.«. s)/(s) ds.
It is easy to see that K n is a bounded linear operator of finite rank on C([a , b]) .
Hence K n is compact. Replacing k by k - Pn in (1.1) gives

11K - Knll S (b - a) max Ik(t, s) - Pn(t, s)1 --+ 0

a ~s,t~b

as n --+ 00. Therefore, K is compact.

t
4. For 1 < p ; q < 00, letk(t, s) be in Lr([O, 1] x [0, 1]), wherer = max(p' , q),
+ ~ = 1. We shall now show that the integral operator K defined by
(Kf)(t) = 1 1
k(t ,s)/(s) ds

is a compact linear map from L p ([0 , 1]) into L q ([0, 1]).

Let ~ = 1 -~ . Then r' S p and for / E Lp([O, 1]), Holder's inequality
applied to I/Ir! E Lp/r'([O , 1]) and the constant function 1 implies 1I/lIr' S II/lip
and

IIKfll~ = III
11 k(t , s)/(s) ds I
q

dt

1
S IIf11;,
I ( I
1lk(t , sW ds
)q/r
dt. (1.2)

Let
g(t) = 11Ik(t,sWds. (1.3)

Since k is in Lr([O, 1] x [0, 1]), g is in LI ([0 , 1]). Therefore, it follows from

Holder's inequality applied to Iglq /r E L r/ q and the constant function 1 that

I
1Ig(t),q/r dt S
(I1 Ig(t),
)q/r
(1.4)

From (1.2), (1.3), (1.4), and the observation 1I/llr' S Ilfll p , we obtain
I I ) I/r
II Kfll q S II/lip ( 11Ik(t , s)]" dsdt (1.5)

Thus K is a bounded linear map from Lp([O, 1]) into Lq([O, 1]).
J3. J Examples ofCompact Operators 301

The set of polynomials in s and t is dense in L r([O , 1] x [0, 1]). Thus there
exists a sequence {Pn (t, s) } ofpolynomials which converges in L; ([0, 1] x [0, 1])
to k (t , s) . Define kn : L p([O , 1]) ~ Lq([O , 1]) by

(Knf)(t) = 1 1
Pn(t , s ) f (s ) ds.

Ifwe replace k in (1.5) , first by Pn and then by k - P«, we see that K; is a bounded
linear operator of finite rank and

Hence K is compact.
It follows from the example that the Volterra operator V defined in XIL9 is
compact linear map from L p([O , 1]) into Lq([O , 1]), 1 < p , q < 00 .
Ifthe restriction 1 < p , q < 00 is removed, then the result in Example 4 is false.
In [G], p. 90, a bounded Lebesgue measurable function defined on [0, I] x [0, 1]
is constructed such that the corresponding integral operator is not compact on
L1 ([0, 1]).
5. The Lebesgue integral operator

I
b B (t , s )
(Kf)(t) = f(s) ds
a It - sla
is compact on C ([a, b]), where B(t , s) is continuous on [a, b] x [a, b] and a < 1.
Also ,
IIKII = max [IB (t , s) j ds. (1.6)
IE [a ,b] Ja It - s la

First we show thatKf is in C ([a, b]) for each f E q a , b]. Let M = maXS,IE[a,b]
IB(t , s)] and II f11 00 = maXIE[a,b] lf (t )l·
Then

[b 2M
I(Kf) (t)1 ~ Mllflloo J It - s l-a ds ~ 1_ a (b - a)l-a IIf11oo.
a

Next we show that K is com pact on C([a, b]). Letk (t , s ) = ,~~;,2 and define for
n = 1, 2, . ..

l
o, It -s l -< ...!..
2n
kn(t , s) = (2n lt - s l - l )k (t , s), 2~ ~ It - s] ~ ~ . (1.7)
k (t , s ), It - sl 2: ~
302 Chapter XlII Compact Operators on Banach Spaces

The function kn is continuous on [a, b) x [a, b) and therefore the operator

(Knf)(t) = l b
kn(t, s)f(s)ds

is compact on C([a, b)) by Example 3. Let Sn = {s 1 It - s I :s ~}.

Since
o, s ¢ s,
Ik (t , s) - k n(t, s)1 = { :s Ik()1
t,s ,SE Sn »
we have

I«K - Kn)f)(t)1 :s l b
Ik(t, s) - kn(t, s)llf(s)lds

s IIflloo Ln Ik(t, s)lds :s Mllflloo 1_1nIt -

t+1

n
sl-ads

2 (2)1-a
s Mllflloo 1 _ ex -;; (1.8)

It follows that Kf E C([a, b)) and 11K - s,

II :s M l~a (~)a -+ 0.

J: 1~~;12Ids
Thus K is compact on C([a, b)). Finally, we prove equality (b). The function
F(t) = is in C([a, b)). This follows from the above result with
IB(t , s)1 in place of B(t , s) and F(t) =
1 on [a, b) . Thus
J: 1~~;,21 f(s)ds, where f is identically

max
t JaI" Ik(t, s)lds = Jra Ik(to, s)lds (1.9)

for some to E [a, b]. Since Ikn(t, s)] :s Ik(t, s)] for each nand s -:F t, we have
from (1.9) and Theorem XII .1.2 that

Jr Jar Ik(t,s)lds
b b
IIKII = lim IIKn ll = lim max Ikn(t,s)lds:s max
n--->oo n ---> oo t a t

l;r Ik(to, s)lds = n--'>oo l;r Ikn(to, s)lds:s

b b
= lim lim IIKn II
n--'>oo
= IIKII·
Thus (1.6) is established. 0

13.2 Decomposition of Operators of Finite Rank

The following theorem is very useful when dealing with operators of finite rank.

Theorem 2.1 IfK is an operator in £(X) offinite rank, then there exist subspaces
Nand Z ofX such that N is finite dimensional, Z is closed,
X=NEElZ, KNCNandKZ=(O).
13.3 Approximation by Operators ofFinite Rank 303

Proof: Let {KVl, ... , Kv n} be a basis for 1m K . By Corollary XI.5 .6, there exist
!J, ... ,In
in X' such that l iKvj = oij, 1 ::::: i, j ::::: n, and for each x E X,
n
Kx = L Ii (Kx)Kvi. (2.1)
i= \

Let N = sp{V\ , .. . , Vn , Kvi ; ... , Kv n}. Clearly, KN C 1m KeN and from

(2.1),
n
X- L Ii (Kx)Vi E Ker K.
i=\
Hence,
X = N +Ker K. (2.2)
Since N is finite dimensional, it follows from Theorem XI.5 .7, that there exists a
closed subspace Z such that
KerK = (N n KerK) $ Z. (2.3)
Equations (2.2) and (2.3) imply X =N $ Z and KZ = (0). o
Every K E £.(X) of finite rank can be written in the form
n
Kx = Lgi(X)Wi, gi E X ', Wi Elm K.
i= \

Indeed, from (2.1) in the proof ofTheorem 2.1, we see that we can choose gi (x) =
li(Kx) and ui ; = KVi .

13.3 Approximation by Operators of Finite Rank

It was shown in Theorem X.4 .2 that every compact linear operator which maps a
Hilbert space into a Hilbert space is the limit, in norm, of a sequence of operators
of finite rank. While this result need not hold for Banach spaces, we do have the
following results.

Theorem 3.1 Suppose {Tn} is a sequence ofoperators in £.(Y) such that Tn y --+
Tyfor all Y E Y. If K is a compact operator in £.(X, Y) , then
IITnK - TKII --+ O.
The proof is exactly the same as the proof of Lemma 11.17.8

Corollary 3.2 Suppose there exists a sequence {Pn} ofoperators offinite rank in
£.(Y) with the property Pny --+ Y for every Y E Y . Then every compact operator
K E £.(X , Y) is the limit, in norm , ofoperators offinite rank. In fact,
IIPnK - KII --+ o.
304 Chapter XIII. Compact Operators on Banach Spa ces

The operators Pn definedon zj , 1 S p < 00, by Pn(al, az , . . .) = (al, . . . , an,

0,0, . . .), n = 1,2, . . . satisfy the hypotheses of the theorem. Hence if K E
LeX , l p) is compact, then IIPn K - KII ---* 0.
Looking at this another way, let {ej} be the standard basis for l p' Given
x = (ai, az, .. .) E lp, x = L:~I ajej . Define !J(x) = aj and Pn(x) =
L:J=I !J(x)ej . Since {Pn} satisfies the hypotheses of Corollary 3.2, IIPn K -
KII ---* 0.
More generally, if {Yi} is a Schauder basis for Y, it can be shown (Exercise
XII-2) that the linear functional Ij defined on Y by

Ij(Y) = «i- where Y = I>jYj

j

is in Y' . Hence, defining Pny = L:J=I !J (y) Yj, it is clear that {Pn} satisfies the
hypotheses of Corollary 3.2. Thus we have following result.

Corollary 3.3 If the Banach space Y has a Schauder basis , then every compact
operator in LeX, Y) is the limit, in norm, ofa sequence ofoperators offinite rank.

Corollary 3.4 Let (ajk)hk=1 be a matrix of complex numbers. The operator A

defined on l I by
00

A(al , az, ...) = (f3I, fJz, ... ), f3j = L ajkak,

k=1

is a compact operator in £(l I ) ifand only iffor each 8 > 0, there exists an integer
N such that
sup L lajkl < 8. (3.1)
k j >N

Indeed, suppose (3.1) holds. For each positive integer n define Pn E £(ld by
Pn (ai , az , . . .) = (al , . . . , an , 0, 0, . . .). It follows from Theorem XILl.l and

L lajkl ---* °as n ---* 00 .

(3.1) that
IIA - PnAIl = sup
k j> n

Conversely, assume A is compact. Then by Corollary 3.2, there exists, for each
8 > 0, an integer N such that

sup L lajkl = IIPNA - All < 8. (3.2)

k j>N

In Section XIL7 the convergence of the projection method for an invertible

operator in LeX , Y) was discussed. Recall that A E I1(Pn, Qn) if A is invertible
and the projection method for A converges . Using an argument analogous to the
proof of Theorem 11.17.7, where PnAPn is replaced by QnAPn, we obtain the
following stability theorem.
13.4 First Results in Fredholm Theory 305

Theorem 3.5 Suppose A E IT(Pn , Qn) and suppose K is a compact operator in

L(X, Y) with the property that A + K is invertible . Then A + K E IT(Pn , Qn) .

13.4 First Results in Fredholm Theory

We recall from linear algebra that if A is a linear operator defined on a finite

dimensional vector space X, then

dim Ker A = codim ImA (4.1)

This useful result is proved as follows.

Let {Xl, . . . ,XK} be a basis for Ker A. Extend it to a basis {Xl , .. . , x n } for
X. Then {AXK+l , ... ,Ax} is a basis for 1m A . Now X = M E9 ImA for some
k- dimensional subspace M of X.
As a generalization we have the following theroem.

Theorem 4.1 Given K E L(X), suppose there exists afinite rank operator Ko E
L(X), such that 11K - Koll < 1. Then I - K has a closed range and

dim Ker(I - K) = codim Im(I - K) < 00.

In particular, the equation

(I - K)x = Y

has a unique solution for every y E X if and only if the equation

(I - K)x = 0

has only the trivial solution x = o.

Proof: First, let us assume that K is of finite rank . By Theorem 2.1, there exist
closed subspaces Nand Z of X such that N is finite dimensional, X = N E9 Z,
KN C Nand KZ = {O}. Let (I - K)N be the restriction of 1- K to N. Then

Ker(I - K) = Ker(I - K) N (4.2)

Im(I - K) = Im(I - K)N E9 Z . (4.3)

To see this, let x E X be given. There exist u E Nand Z E Z such that X = u + z.

Hence (I - K)x = (I - K)u + z and (I - K)u is in N. Thus (4.3) follows.
If (I - K) x = 0, then z = u - ku E N n Z = (0). Thys X = u E Nand
(I - K) x = (I - K)NX .
306 Chapter XIII Compact Operators on Banach Spaces

Since (/ - K)N is in £(N) and N is finite dimensional, equations (4.1), (4.2)

and (4.3) imply

00 > dim Ker(/ - K) = dim Ker(/ - K)N = codim Im(/ - K)N

= codim Im(/ - K) .

Also, it follows from Theorem XI.2.5 and (4.3) that Im(/ - K) is closed.
More generally, assume 11K - Koll < 1, where Ko is of finite rank. Now
B = I - (K - Ko) is invertible and

1- K =B- Ko = (/ - KoB-I)B .

Hence

Im(/ - K) = Im(/ - KoB- I), (4.4)

B Ker(/ - K) = Ker(/ - KoB- I). (4.5)

Since KoB- 1 is of finite rank , it follows from what we have shown, together with
(4.4) and (4.5), that Im(/ - K) is closed and

00 > dim Ker(/ - K) = dim Ker(/ - KoB- I)

= codim Im(/ - KoB- I) = codim Im(/ - K) .

o
Corollary 4.2 If K E LeX) is the limit, in norm, of a sequence ofoperators of
finite rank, then I - K has a closed range and

dim Ker(/ - K) = codim Im(/ - K) < 00 .

Corollary 4.3 If X is a Banach space with a Schauder basis and K E LeX) is

compact, then I - K has a closed range and

dim Ker(/ - K) = codim Im(/ - K) < 00.

Proof: Corollaries 3.3 and 4.2.

The last statement of Theorem 4.1 is referred to as the Fredholm alternative.
In Chapter XV Theorem 4.2 we shall show that Corollary 4.2 is valid when it is
only assumed that K is compact 0

13.5 Conjugate Operators on a Banach Space

Corresponding to the adjoint ofan operator defined on a Hilbert space, we introduce

the concept of the conjugate of an operator defined on a Banach space.
13.5 Conjugate Op erators on a Banach Space 307

Definition: Given A E £ (X , Y ), the conjugate A' : Y' ---+ X' of A is the operator
defined by A' I = f o A, lEY' .
lt is clear that A' is linear. Furthermore, II A' II = II A II . The proofofthis assertion
is as follows.
IIA'/II = II/ o Ail s IIflill AII·
Hence II A' II ~ II A II. On the other hand , by Corollary XI.5.5

IIAx il = max I/ (Ax)1 = max I(A ' f)xl .s II A'II IIx ll·
"1"=1 "1"=1
Thus II AII ~ IIA'II·
lt is easy to verify that if A and B are in £(X , Y ), then

(A + B )' = A' + B' and (aA)' = aA'.

IfC is in £(Y, Z) , then (CA )' = A'C' E £(Z', X') .

Examples: 1. Suppose K E £ (X , Y) is of finite rank. Let Yl, .. . , Yn be a basis

for ImK. For each i . define the linear functional I j on X by I j (x) = a j , where
Kx = L:J=1
a j Yj . Then for g E Y ' ,

( K'g)x ~ gKx ~ ~ Ij (x)g(Yj) ~ (~g(Yj)/j ) x.

Thu s
n
K 'g =L g (Yj )/j.
j =l

Choo sing gk E Y ' such that gkYj = 0kj, 1 ~ i . k ~ n , we obtain

Ik = K ' gk EX' .

t
2. Let K be the integral operator in Sect ion 1, Example 4. If we identify
L ~ ([a , b]) with L pl([a , b]) , + ~ = 1, as in Theorem XI.4.2 , then for F E L~
and g E L p([a, b]) ,

i {i
(K' F)g = F (Kg) =
b
F(t )
b
k(t , s )g (S)dS} dt

=i {i b b
g(s) k (t ' S)F(t)dt} ds .

i
Hence
b
( K' F )(s ) = k (t , s )F(t )dt. (5.1)

To be more prec ise, we have shown that if if in L~ ([a , b] ) corresponds to

F E L ql([a , b]) , Then K 'if E L~ ([a , b] ) corresponds to the function defined
by (5.1).
308 Chapter XIII. Compact Operators on Banach Spaces

Analogous to the relationships between the ranges and kernels of an operator

and its adjoint, we have the following results for the conjugate operator.

Definition: For M a subset of X and N a subset of X ', define

M1.. = {f E X' : I(M) = O}

and
1.. N = {x EX : g(x) = 0 forallg EN}.

Theorem 5.1 For A E L(X, Y) ,

(i) (Im.A):': = Ker A'

(ii) IrnA =1.. (Ker A')
(iii) Ker A =1.. (Im A')
(iv) (Ker A)1.. :J 1m N.

Proof: We shall only prove (ii). The proofs of the remaining relationships are
similar. These we leave to the reader.
Suppose IE Ker A'. For any x EX,

I(Ax) = (A' f)x = O.

It therefore follows from the conitnuity of I that I(lm A) = O. Thus
1m A C 1..(Ker A'). Assumex E 1..(Ker A') but x ¢ 1m A . Then by Theorem XI.5 .3
there exists agE X' such that

g(x) -=F 0 and g(lm A) = O. (5.2)

Hence for all z EX,

(A 'g)z = gA z = O.
Thus g is in Ker A' and since x is in 1..(Ker A') , g(x) = O. But this contradicts
(5.2) . Hence 1..(Ker A'), C 1m A. 0

Theorem 5.2 Given A E L(X, Y), suppose 1m A is closed and codim 1m A < 00.
Then
dim Ker A' = codim 1m A.

Proof: By hypothesis, there exists a finite dimensional subspace N such that

Y = 1m A EEl N. Let YI, . . . , Yn be a basis for N. Since M] = Sp{ydi;6J EEl 1m A
is closed and Yj ¢ M], Theorem XI.5.3 ensures the existence of an !J E y' such
that
13.5 Conjugate Operators on a Banach Space 309

Hence
f iYj = oij and fi E (ImA).L = KerA', 1 ~ i ~ n. (5.3)
We need only show that {II, .. . , fn} is a basis for Ker A' . It follows readily from
(5.3) that the /; are linearly independent. Given f E Ker A' and Y E Y, there
exists a u E 1m A and a v E N such that Y = u + v. Since f is in (1m A).L and
{YI, ... , Yn} is a basis for N, (5.3) implies that

fey) = f(v) = f (~fi(V)Yi) = ~/;(Y)f(Yi) = (~f(Yi)/;) (y).

Thus f = :L7=1 f(Yi)/;, which concludes the proof that fJ, .. . , fn is a basis for
KerA' . 0
Exercise XII-I 8 shows that codim 1m A < 00 implies 1m A is closed.
Since (l - K)' = 1 - K', the next result is an immediate consequence of
Theorems 4.1 and 5.2 .

Theorem 5.3 Given K E L(X), suppose there exsist afinite rank operator Ko E
L(X) such that 11K - Koll < 1. Then
00 > dim Ker (l - K) = codim Im(l - K) = dim Ker(l - K').
We shall now prove the main result in this section, namely, that the conjugate
ofa compact operator is compact. The proofrelies on the following theorem.

Theorem 5.4 Let W be a subset ofa Banach space X. Thefollowing statements

are equivalent.
(a) Every sequence in W has a convergent subsequence.
(b) For each e > Othereexistsafinite set{wI , W2, . . . , wn}ofvectorsin Wwith
the property that given W E W, there exists a W k such that II W - W k II < e.
Proof: Defining S( z, r) = {x E Xlllx - z] < r}, statement (b) means that
We Uk=1 S(Wk, e) . We call S'(z, r) a ball of radius r .
Assume (a) . If (b) is not valid, then for some e > 0, W S; S(u , e) where UI
is any vector in W . Thus there is a U2 E W , U2 1= S(UI, e), i.e., lIu2 - u IIi ~ e.
Since W S; S(UI, e) U S(U2, e), there is a U3 E W , U3 1= S(UI , e) U S(U2, e), i.e.,
lIu3 - uill ~ e, i = 1,2. Continuing, we obtain a sequence {ukl ~ W such that
lIum - Un II ~ e, m i= n . Thus {Un} does not have a convergent subsequence. This
contradicts (a).
Assume (b) . Let {Vk} be a sequence in W . To prove that {Vk} has a convergent
subsequence, it suffices to assume that Vm i= Vn, m i= n. Since VI = {VI, V2, . . .}
is an infinite subset of Wand W is contained in the union of a finite number of
balls of radius i,
at least one of these balls contains an infinite subset V2 ~ VI.
Since V2 is an infin ite subset of VI ~ W, it is contained in the union of a finite
310 Chapter XIII Compact Operators on Banach Spaces

number of balls of radius ~ . Hence there is an infinite subset V3 S; V2 which

is contained in a ball of radius ~ . Continuing in this manner; we obtain subsets
VI ;2 V2 ;2 V3 ;2 . . . , and each Vk is an infinite subset of a ball of radius Thus t.
there exists vn k E Vb nk > nk-I.
Since n j > nk implies V n j and vn k are in Vk and Vk is contained in a ball of
radius t, it follows that {v n k } is a Cauchy sequence which therefore converges
inX. 0

Theorem 5.5 If K E LeX, Y) is compact, then K ' E Ley', X') is compact.

Proof: Define S = {x E Xlllxll :s 1}. Since K is compact, every sequence inKS
has a convergent subsequence.
Thus, by Theorem 5.4, for each 8 > 0, there exists a finite set {KXI, KX2, . . . ,
K x n } ;2 KS with the property that given XES, there exists a Kx, such that
(5.4)
Let A be the map from Y' to en given by
Ay' = (y' KXI , . . . , y' Kx n ) .
Since Y' is complete and A is compact, it follows from Theorem 5.4 that there
exist y;, y~ , y:n in Sy ' = {y ' E Y'IIIy'1i :s 1} such that for each y' E Srr, there is
a yj for which
IIA'y' - A'yjll < 8/3 .
Thus,
Iy' KXi - yjKx;l .:'S IIA'y - A'yjll < 8/3, l:s i :s n . (5.5)
For x EX, IIxli = 1, we have from (5.4) and (5.5),
IK'y'x - K'yjxl :s Iy'Kx - y'Kx;I + ly'Kxi - yjKx;I + IYjKxi - yjKxl
:s IIKx - Kxill + 8/3 + IIKxi - Kxll < 8.
Hence IIK'y' - K'ylll < 8. The compactness of K' is now a consequence of
J
Theorem 5.4.

13.6 Spectrum of a Compact Operator

As we pointed out in Section XII.8 the spectrwn of an operator A E LeX), where

X is finite dimensional, consists of a finite number of points, each of which is
an eigenvalue of A. Before proving the following extension of this result, we
note that eigenvectors VI , •.. , V n corresponding to distinct eigenvalues AI, ••. , An
of a linear operator T are linearly independent. For, if not, there exists a Vk
13.6 Spectrum ofa Compact Operator 311

such thatVI, . .. ,Vk-I are linearly independent and Vk E sp{VI , •.• , Vk-I}, say
"k-I S·
Vk = L"j=1 CXjVj . mce

k-l
0= AkVk - TVk = LCXj(Ak - Aj)Vj,
j=1

We have a j = 0, 1 ~ j ~ k - 1. But then Vk = 0 which is impossible. 0

Theorem 6.1 Let X be infinite dimensional, and let T E LeX) be the limit in norm
ofa sequence ofoperators offinite rank. Then the spectrum ofT is a countable
set {AI, A2, . . .} which includes A = O. IfAj i= 0, then it is an eigenvalue ofT. If
{Aj} is an infinite set, then Aj -+ O.
Proof: Since X is infinite dimensional, 0 E aCT), otherwise I = TT- I is com-
pact. Suppose 0 i= A E aCT) . Then by Corollary 4.2,

codim(U - T) = dim Ker(U - T) = dimKer (I - f) i= (0),

i.e., A is an eigenvalue of T . To prove the rest of the theorem, it suffices to show

that for each B > 0, the set S = {A E aCT) : IAI ~ B} is finite. Suppose that
for some B > 0, S contains an infinite set AI, A2, . . .. Let Vj be an eigenvector
of T with eigenvalue Aj and let M n = sp{VI, ... , vn }. Since the Vj'S are linearly
independent, Mn-I is a proper subspace of M n. Hence by Lemma XI.2.3, there
exists a W n E M n such that

(6.1)
Now

It is easy to see that AnWn - TW n and TW m are in Mn-I, provided n > m . Hence
it follows from (6.1) and (6.2) that for n > m,
IITwn - TW m II ~ IAnld(w n, Mn-J> = IAnl ~ B.

But this is impossible since {Tw n } has a convergent subsequence. 0

In Chapter XV, Theorem 4.4, the theorem above will be extended to any compact
operator T E LeX) without the assumption that T is the limit in norm ofoperators
of finite rank.
The above proof of Theorem 6.1 is short, but it does not explain the analytical
motivation which is behind the theorem. It turns out that the proof of this theorem
can be derived from the well known fact from complex analysis that if a non-zero
analytic function vanishes on a compact set Z, then Z is finite .
We shall prove Theorem 6.1 under the assumption that for each B > 0 there
exists a K E LeX) of finite rank such that II T - K II < B. As before, it suffices
312 Chapter XIII Compact Operators on Banach Spaces

to show that S = {}.. E aCT) : I}..I ~ s] is finite . This is done by constructing a

non-zero analytic function which vanishes on S. For A = T - K,
(AI - T) = AI - A - K (6.3)
and AI - A is invertible whenever X E S since II ~ II < 1. Thus S C peA) .
Let e be an eigenvector ofT corresponding to z, E S and let A(}") = (AI -A)- I .
From (6.1) we get
O=}..cp - Tcp = A(}")[(AI - T)cp] = cp - A(}")Kcp. (6.4)
Since 1m K is finite dimensional, we know from the remark following Theorem 2.1
that there exist gl, . . . , Sn E X' and a basis YI, .. . , Yn for 1m K such that for all
x EX, Kx = L:J=I gj(x)Yj. Thus it follows from (6.4) that
n
cp = Lgj(cp)A(}..)Yj . (6.5)
j= 1
Hence
n
gkCcp) = Lgj(CP)gk(A(}..)Yj) , 1::: k::: n .
j=1
Writing
(6.6)

we get the system of equations

n
~k - I ) jaAj(}..) = 0, 1::: k ::: n. (6.7)
j=1
Not all ~j are zero; otherwise cp = 0 by (6.5), which is impossible since cp is
an eigenvector. Hence, the system of Equations (6.7) has a non trivial solution,
namely ~j = gj (e) , Therefore,
h(}") = det(o/g' - akj(}..» = 0, }.. E S.
Conversely, if h(}") = 0, there exists ~I, ... .u,
not all zero, such that (6.7) holds .
Guided by (6.5) and (6.6), we define cp = L:J=I ~jA(}..)Yj . Since YI, . .. , Yn are
linearly independent and A(}") is invertible, it follows that cp i= O. With the use of
(6.7), a straightforward computation verifies that (AI - T)cp = O.
Thus, h(S) = 0 and h is not identically zero since I~I > IIAII + 1IT11 implies
~ E peA) n peT), whence h(~) i= O. Moreover, h is analytic on peA). To see
this, we recall from Theorem XILlO.l that A(}") is an analytic operator valued
function on peA) . It follows readily from the linearity and continuity of gk that
a/g'(}..) = gk(A(}..)Yj) is a complex valued function which is analytic on peA) .
Hence h is analytic on peA). A basic result in complex analysis states that if a
non-zero analytic function vanishes on a compact set Z , then Z is finite. Since S
is compact and h(S) = 0, S is finite. This completes the proof of the theorem.
13.7 Applications 313

13.7 Applications

In Section 1 it was shown that the operators appearing below are limits in norm of
a sequence of operators of finite rank. Hence the conclusions below follow from
Corollary 4.2 and Theorem 6.1.
1. Suppose k is continuous on [a, b] x [a, b]. The equation

Af(t) -l b
k(t ,s)j(s)ds = g(t), ).. #: 0 (7.1)

has a unique solution in C ([a, b]) for each g E C ([a , b]) if and only if the homo-
geneous equation

Af(t) = l b
k(t, s)j(s) ds = 0, ).. #: 0, (7.2)

has only the trivial solution in C([a, b]).

Except for a countable set of ).., which has zero as the only possible limit point,
Equation (7.1) has a unique solution for every g E C([a, b]). For ), #: 0, the
equation (7.2) has at most a finite number oflinear independent solutions.
2. For 1 < p < 00, letk(t, s) be in L,([O, 1] x [0, 1]), where r = max(p, p i),
i ?
+ = 1. The conclusions in the example above remain valid when we replace
C([a, b]) by Lp([O, 1]).
3. Let (ajk)},'k=1 be a matrix of complex numbers such that
00

lim sup L lajkl = O.

n--+oo k .
J=n

The infinite system of equations

00

)..Xj-Lajkxk=Yj, j=I,2, ... , ),,#:0 (7.3)

k=1
has a unique solution (XI, xz, . . .) E il for every (YI, YZ, . . .) E il if and only if
the homogeneous system of equations
00

Axj-Lajkxk=O, j=I ,2, . . . , ),,#:0 (7.4)

k=1
has only the trivial solution in i I .
Except for a countable set of ).., which has zero as the only possible limit point,
Equation (7.3) has a unique solution (XI, xz, . .. ) E il for every (YI, YZ, .. .) E il .
For X #: 0, Equation (7.4) has at most a finite number of linear independent
solutions.
4. If L},k=1 lajd < 00, the conclusions in the above example remain valid if
we replace il by iz . Here we also need Example 2 following Theorem 11.16.3.
314 Chapter XIII. Compact Operators on Banach Spaces

Exercises XIII

Throughout these exercises, X and Y denote Banach spaces.

1. Let{,Bn} be a bounded sequence of complex numbers. Define T : p --+ e ep»

1 ~ p < 00, by T(al , az, . . .) = (,Blal, ,Bzaz, . . .). Find T' .
2. Leth be a bounded complex valued Lebsegue measurable function on [a, b).
Define A : Lp[a, b) --+ Lp[a , b), 1 .s p < 00, by (Af)(t) = h(t)f(t) .
Find A'.
3. Let T : Lp[O, oo) --+ Lp[O,oo), 1 ~ p < 00, be given by (Tf)(x) =
f(x + 1). Find T' .
4. Suppose P E .c(X) is a projection from X onto M. Prove that P' is a
projection with kernel M 1-. What is the range of P'?
5. Prove that ifT E .c(X, Y) is invertible , then T' is also invertible. What is
the relationship between (T') -I and (T- I )'?
6. Suppose A E .c(X, Y) and 1m A = Y . Prove that A is an isometry if and
only if A' is an isometry and 1m A' = X'.
7. Prove that a projection is compact if and only if it is of finite rank.
8. Let K E L(X) be compact. Suppose that I - K is either left or right
invertible. Prove that I - K is invertible and I - (I - K) -I is compact.
9. Suppose Un} c x', {Yn} C Y and L~ I IIfn II llYn II < 00. Prove that the
operator K : X --+ Y given by Kx = L~I fn(x)Yn is compact. In this
case K is called a nuclear operator.
10. Let K E L(X, Y) be compact. Suppose {x n } is a sequence in X with the
property that there exists an x E X such that fnx --+ f(x) for all f EX'.
(a) Prove that KX n --+ Kx .
(b) Give an example which shows that (a) need not hold if K is not compact.

11. Suppose K E .c(X) is compact and I - K is injective. Prove that if SeX

and (I - K)S is a bounded set, then S is bounded.
12. Let K E .c(X) be compact. Given A =1= 0, prove that K}.. = AI - K has a
closed range by establishing the following:
(a) Ker K}.. is finite dimensional (Hint: K = AI on Ker K}..). Therefore
Ker K}.. is complemented by a closed subspace M .
(b) Given Y E 1m K}.. there exists a sequence {vn} in M such that
K}..vn --+ y .
(c) The sequence {vn} is bounded.
(d) The sequence {v n } has a subsequence which converges to some x E X
and Kvx = y .
Exercises XIII 315

13. Let CPI, CP2, .. . be an orthonormal basis for a Hilbert space 1-£. Define T E
£(1-£) by T x = (x , CP3}CP2 + (x , CPS}CP4 + (x, CP7}CP6 + .... Show that T is
not compact but T 2 = O.
14. Suppose A E £(X) and AP is compact for some positive integer p .
Prove that the Fredholm Theorems Xl-4.1, 5.3 and 6.1 hold for A . Hint:
Choose n roots of unity 1, ~I, ... , ~n-I for some n ~ p so that ~i E p(A),
1 ~ i ~ n - 1 (why do such ~i exist?) Then I - An = (I - A)
(~I - A) . .. (~n-I - A) .
Chapter XIV
Poincare Operators: Determinant and trace

A linear operator P defined in the standard basis of £p (l :s P < 00) by a matrix

of the form

with Pjk E C and

00

L IPjkl < 00
j, k= l

we call a Poincare operator, H. Poincare [P] introduced and studied this class in
connection with problems in celestial mechanics . One of the important advantages
of this class is that for these operators P it is possible to introduce a trace and for
operators I - P a determinant that are natural generalizations of the trace and
determinant of finite matrices. To keep the presentation of the material in this
chapter simpler we restrict ourselves to the Hilbert space £2 . However it can be
generalized for many Banach spaces, in particular, for £p(l :s P < 00).

14.1 Determinant and Trace

Denote by P the set of all Poincare operators on £2. Recall that these operators P
are defined in the standard basis of £2 by the matrix

with the property

00

L Ip jkl < 00.

j .k= l

It is easy to see that those operators can be represented in the form

00

P = L P jkUjk.
j ,k= l

where
318 Chapter XN. Poincare Operators

and el, ez . . .. is the standard basis in £2. Taking into account that
00

IIPII :5 L IpjklllUjkll
j,k=1

and
IIUjkll = 1,
we obtain that P is a bounded operator and

where
00

IIPlip = L Ipjkl·
j ,k= 1

For any operator P = (p jkr::k=1 E P we denote by n

P an operator defined on en
by the equality:
Pn = (Pjk)'j ,k=I'

We call this operator the n-th section of P. Denote by p(n) E P the operator

p(n) = (Pjt»)hk=I '

where pjt) = Pjk for j, k = 1,2, . . . , n and pjt) = 0 for the rest of the indices j
and k. Note that the following relations hold

liP - p(n ) II :5 liP - p(n) lip = liPlip -IIP(n )lIp.

The last difference converges to zero when n --+ 00 . From here it follows that P
is a compact operator on £2.
For any operator PEP the following inequality holds:

where

TI(Pn ) = .~
J-I
(1 + i: k=1
IPjkl)

Indeed, let
v = (Vjk)j ,k=1
be a matrix with entries Vjk E C. According to the definition of the determinant

where the coefficient Yi, .h. ....i « is equal to one of three values +1, -1 and O.
14.1 Determinant and Trace 319

Let the entries Vjk have the form

It is clear that
[det VI s nrv),
where
n
n(V) = n (1 + IUjl1 + IUj21 + . . . + IUjnl) .
j=1

Let some set L of entries Ujk in the matrix V be replaced by zeros and the rest is
not changed. Denote the new matrix by W . It is easy to see that

[det V - det WI :s n(V) - neW) . (1.2)

From this inequality follows inequality (Ll). Indeed, introduce Ln as the set
of Ujk for which at least one of the indices j or k is > n . Replace Ujk by 0, for
i . k E Ln' then the inequality (Ll) follows from (1.2).
We are now in a position to define the determinant detp(/ - P) for any PEP.
Let Pn(n = 1,2, . . .) be the n-th finite section of P:

From the relation (1.1) follows that the sequence of complex numbers

det(/n - Pn) (n = 1,2, . ..)

is convergent. The limit we denote by detp(/ - P) and call it the Poincare

determinant of the operator I - P :

detp(/ - P) = lim det(/n - Pn) .

n--->oo

Also note that

Idetp(/ - P)I :s n
00

j=1
(l + Iplj 1+ IP2j 1+ .. .).

Using the inequality 1 + x :s e' we obtain also

320 Chapter XIV. Poincare Operators

The trace is introduced for operators PEP by the formula

00

ts-p P = LPjj,
j=l

where P = (Pjk)hk=l' The trace is a linear functional on P and

ts-p P = lim trPn .

n---+oo

The trace is continuous in P in the sense

and hence it is a Lipschitz function on all of P:

ItrpP' - trpP"1 ::: liP' - P"lIp .

Example: 1. Let the operator V E P be given by the infinite matrix

-~10 "')
UI U2 Un-l
0 0 0 O· . .
V =
(
-V2 0 0 o·..
0 0 -V3 0 0 ··.

We shall assume that the series

00 00

Llujl and Llvj!

j=l j=l

are convergent. It is obvious that V E P. Now we shall calculate detp (/ - V) .

The starting point is the calculation of the determinant of the n-th section

If we expand this determinant along the last column we obtain

det (/ - V n) = (_l)n+I Un VI V2, ... , Vn + det (/ - Vn-t}.

Hence

det (/ - V n) = (_l)n+IUnVIV2 Vn +
(-ltU n-IVIV2 Vn-l + ... + UIVI + 1.
From here it follows that
00

detp (/ - V) = L(_l)n+IUnVIV2 ... Vn +1

n=l
14.2 Finite Rank Operators, Determinants and Traces 321

where the series converges absolutely. Note that

00

detp (I - JLV) = LJLn+I(_1)n+IUnVI , V2 ··· Vn + 1

n= 1

and
trp V = O.
Note also that the operator V is not of finite rank, generally speaking .

14.2 Finite Rank Operators, Determinants and Traces

Let F(f2) denote the space of bounded operators of finite rank on f2 . Due to
Theorem XIII.2.l, for any operator F E F(f2) there exist closed subspaces MF
and N F of f2 with the following properties:

a) MF(JJNF=f2
b) F M F ~ M F, N F ~ Ker F
c) dim M F < 00 .

With respect to the decomposition a), the operator F can be represented by

where FI = FIMF' Hence

I _ F _ (!J - FlO)
- 0 h

where I , !J, h stand for the identity in f2, M F, N F respectively. Since M F is a

finite dimensional space, the functionals tr FI and det (!J - FJ) are well defined.
We use this to define the tr F and det (I - F) by the equalities

tr F = tr FI and det(I - F) = det(!J - FJ).

Note that these definitions do not depend on the particular choice of the subspace
M F. To see this we recall from linear algebra that

tr FI =
m
~::).j(FI)
j=1
and det(!J - FJ) = n
m

j=1
(l - Aj(FJ))

where m = dim MF and Al (FI), A2(FJ) , ... , Am(FI) are the eigenvalues of FI
counted according to their algebraic multiplicities. In the two identities the zero
322 Chapter XN. Poincare Operators

eigenvalues do not give any contribution. It is also easy to see that the operators
F and Fl have the same nonzero eigenvalues counting multiplicities. Thus

tr F = I>'j(f)
j
and det(I - F) = n(l-
j
Aj(F)). (2.1)

Now we shall prove the following four properties of the trace and determinant.
For any two operators F and K E F(l2) the following equalities hold

tr(exF + 13K) = ex tr F + 13 tr K, ex, 13 E <C

trFK = trKF
det«(I - F)(I - K)) = det(I - F) det(I - K)
det(I - FK) = det(I - KF)
We shall show that all these properties follow from the corresponding properties
in linear algebra . Suppose
p q

Fx = L(x, CPj)1/!j and Kx = L(x, fj)gj,x E l2 .

j=l j=l

We set

M = span{1/!j, CPj, i = 1,2, . . . , p; ik, gb k = 1,2, .. . , q}

and
N=Ml.. .

Then
F(M) C M, K(M) eM

NcKerF , NcKerK.
Now let

The subspace M has a finite dimension and hence all four properties for F M and
KM follow from linear algebra. This statement is proved in the same way as the
first statement in this section. The next theorem gives a useful representation of
the trace and determinant.

Theorem 2.1 Let FE Fel2) be given in the form

m

Fx = L(x, Cfik)fb x E H
k=l
14.2 Finite Rank Operators, Determinants and Traces 323

Then
m
tr F = L(!k, rpk} (2.2)
k=l
and
(2.3)

Proof: Since tr F is a linear functional, it is enough to prove the first equality for
the case m = 1. In this case the operator is

Fx = {x, rpl} ft .
Let

Then
y
x = ;: ft, where y = (x, rpt)
and
y
y = ;:(fl, rpl)·

Taking y f= 0 we obtain that the only non-zero eigenvalue of F is

Now let us prove the second equality. Consider two operators X E £(£2 , C'") and
Y E £(c m , (2) defined by the equalities:

Xrp = «rp, rpl), (rp , f/!2}, . .. , (rp , rpm}),

m
Y(ah a 2 , .. . ,am) = LaJij.
j=l

It is clear that F = YX. Hence

det(I - F) = det(I - YX) .

Note that

HenceXY = «(fk,rpj})j,k=l'
324 Chapter XN. Poincare Operators

Now it remains to prove that

det(Im - XY) = det(I - YX) .

Introduce the following finite rank operator ( ~ ~) and take

which acts on em EB l2 . Using the following two Schur decompositions of A we

obtain

and

A=(~ ~)=(5 ~)(h~XY ~)(~ ~).

Computing the determinants from both sides of these equalities we get

det(Im - XY) = det(I - YX) .

This ends the proof. o

The next theorem gives a representation of the trace and the determinant for the
case when the finite rank operator is given by an infinite matrix .

Theorem 2.2 Let the operator F E F(l2) be defined in the standard basis
e 1, ei , ... of l2 by the following matrix

Then
00

tr F = Lakk. (2.4)
k=l
where the series converges absolutely and

det(I - F) = lim det(Ojk - aJ'k)J~ k=l (2.5)

n-+oo '

Proof: The operator F can be represented in the form

m

Fx = L(x, CPr)fr.
r=l
14.2 Finite Rank Operators, Determinants and Traces 325

Hence
m
ajk = (Feb ej) = L(eb rpr)(fr, ej), k , j = 1,2,. ... (2.6)
r=l
Taking into consideration that
00 00

fr = L(Jr , ej)ej, rps = L(rps, ek)ek

j=l k=l

we obtain
00

(Jr, rps) = L(Jr, ej)(({Js, ej) (2.7)

j=o
This series converges absolutely since
00

L l(Jr, ej)ll(rps, ej )1 ~ IIfr II lI<PslI·

j=l
From formulas (2.6) and (2.7) follows that
00 00 m
Lakk= LL(Jr,ek)(<Pr ,ek)
k=l k=l r=l
m 00 m
= L L(Jr , ek)(<Pr, ek) = L(Jr, <Pr) .
r=lk=l r=l
In Theorem 2.1 we proved that
m
tr F = L(Jr , <Pr)
r=l
and hence
00

tr F = Lakk.
k=l
The first part of the theorem is proved. Now we proceed with the second part .
Denote
v
(v) ~ F-
brs = ~J rk<Psk
k=l
where
frk = (Jr, ek) and <Psk = (<Ps, ek)'
When v tends to infinity b~~) tends to (Jr , <Ps):

lim b~~) = (Jr ,<ps) ,r,s = 1, . . . .m,

v--+oo
326 Chapter XN. Poincare Operators

The determinant of finite matrices of fixed size is a continuous function of its

entries . From this follows that

lim det(ors - b~s)~s=1 = det(ors - (fr, %»~s=1

V~OO ' ,
(2.8)

In Theorem 2.1 it was proved that

det(l- F) = det(ors - (fr, rps»~s=1

and this means
det(l - F) = lim det(l - R(v».
v-'>oo
where
R( V) = (b(v»m
rs r,s=I '
It is easy to see that the matrix R(v) is a product of two matrices.
R (v) -- X f ,v y*
cp ,V'
where

and
ycp,v = (rpsj)~I' }=1 .
In the proof of Theorem 2.1 it was proved that for two finite matrices X f, v
and Ycp,v
det(l- Xf,vY;,v) = det(l- Y; ,vXf,v).
The matrix Y;, X
v f ,v has the size v x v. Denote

Y;,v X [, v = (Vjk)} ,k= I'

where
m m
Vjk = LfPjrfrk = L(ej, rpr}(fr, ek), j, k = 1,2. .. . , v.
r=1 r=1

It is clear from (2.6) that

Vjk = akj j , k = 1,2, .. . , v.

It is also now clear that

det(l - R(v» = det(Ojk - ajk) },k=O'
Now we go to the limit from both sides when v -+ 00 and taking into consideration
the earlier results, we obtain
det(l - F) = lim det(l - R( v»
V-'>OO
which finishes the proof. o
14.3 Theorems about the Poincare Determinant 327

Note that so far we have introduced for Poincare operators F E F(l2) two
definitions of determinants for I - F and two definitions of trace for F.
Namely, in the first section we defined

detp(l - F) = lim det(l - Fn )

n--+oo

and
ts-p F = lim tr Fn ,
n--+oo

where Fn is the n-th section of F . In this section the determinant and the trace
were defined via the formulas

det(l- F) = nm

r=!
(l - Aj(F))

and
m
trF= LAj(F).
r=!
As a corollary of Theorem 2.2 we obtain that these two definitions coincide:

detp(l - F) = det(l - F), trp F = tr F.

In particular

detp(l- F) = n
r=!
m
(l- Aj(F)),

m
trp(l- F) = LAj(F) .
r=!
In this connection we will drop, in subsequent sections, the subscript P in the
Poincare determinant.
In conclusion we mention that it can happen that an operator has finite rank and
is not a Poincare operator; see exercise 7.

14.3 Theorems about the Poincare Determinant

In this section we shall discuss a number of important properties of the determinant.

We start with the analysis of continuity of the function det (l - X), X E P .
Note that P is a linear space. It is not difficult to see that it is a Banach space in
the P-norm II . lip·

Theorem 3.1 The function det (l - X), X E P is Lipschitz continuous. Namely,

for any operators X, YEP the following inequality holds:

Idet(l- X) - det(l- Y)I ::s [exp(IIXllp + IIY11p + l)]IIX - YIIp . (3.1)

328 Chapter XN. Poincare Operators

Proof: Let us first prove the theorem for operators X, Y of finite rank . Introduce
the function
g(A) = det (I - ~(X Y)+ +A(X - Y)).
It is obvious that g(A) is an entire function and

g(~) = det(l- Y) , g(-~) = det(l- X).

Therefore

Idet(l- X) - det(l- Y)I = Ig(~) - g(-~) I::::; sup Ig'(t)l .

2 2 _1 <1 <1
2- -2

From complex analysis we use the following Cauchy formula

g'(t) = _1 [ g({ ~ t) .u,

2JT J1{I=p {

where p is a positive number that will be chosen later and - i < t < i.
Hence
1
Ig'(t)1 ::::; - sup Ig(t + nl
p 1{I=p
and
1
Idet(l- X) - det(l- Y)I ::::; - sup Ig(A)I ·
p [).I::;p+1

Put p = IIX - YII:pl and IAI < p + i. Taking into account that
H Y)
(X + + A(X - Y)llp ::::; ~ (II(X + Y)lIp + II (X - Y)lIp) + 1,

we obtain

B (X + Y) + A(X - Y)llp ::::; IIXlip + IIY11p + 1.

Using now the formula (1.3) is Section 1, we have

Ig(A)1 = Idet (I - ~ (X + Y) + A(X - Y)) I ::::; exp(IIXlip + IIY11p + 1).

Now for any finite sections X n and Yn of X and Y, respectively. we have

Passing to the limit we obtain (3.1). o

14.3 Theorems about the Poincare Determinant 329

Corollary 3.2 The function det(I - X) is a continuous function in P in the fol-

lowing sense: For any X E P and 8 > 0, there exists a 8 > Osuch that for any
YEP which satisfies the inequality

IIX - YIIP < 8,

the following inequality holds

Idet(I - X) - det(I - Y)I < 8.

Let us remark that the functions trX and det(I - X) are not continuous on P in
the operator norm. Indeed, define operators

where
(n) -1 .
ex j =.jii ' ] = 1, 2, . . . , n

and
exjn) = 0, j = n + 1, n + 2, . ..
It is obvious that Fn E P and

lim
n-+oo
IIFn ll = 0.

Consider det(I - Fn ) and tr Fn . It is clear that

det(I - F n) (1 + In)n ~
= 00

and
tr Fn = -.jii ~ -00

forn ~ 00 .
In subsequent sections we will need the following properties of the determinant.

Theorem 3.3 The set P has the following properties: If G(l) and G(2) E P then
G = G(l)G(2) E P and

Also the following equality holds:

330 Chapter XN. Poincare Operators

Proof: The first property follows from the relations

00
'" (I) (2)
gjk = LJ gjr g rk
r= 1

and
00 00
" Ig (1) 1 'LJ
. < 'LJ
IgJkl -
" Ig (2) I,
jr rk
r=1 r=1

Hence

L
00

IIGlip = Igjkl ::: IIG(1)lIplld

2)lIp.
j ,k= 1

To deduce the second relation we first remark that

Consider the equality

det(I - G(l)
n
det(I - d n2» = det(I - G(1)
n
- d n2 ) + d 1l ) G(2»
n'

Here G~) is the n-th section of G (i) .t = I, 2. Now let us take the limit when
n --+ 00 from both sides of this equality. The limit of the left hand side is equal to

and the limit of the operator G(3)(n) = G~l) + G~2) - G~I) G~2) converges to G(O)
in the following sense

From the continuity of the function detp(I - X) it follows that

lim det(I - G(3)(n» = det(I - G(O»

n--+oo

This ends the proof. o

14.4 Determinants and Inversion of Operators

In this section we discuss the role of the determinant in the inversion of operators
of the form 1- G, G E P.
14.4 Determinants and Inversion of Operators 331

Theorem 4.1 Let A E P. Then the operator I - A is invertible if and only if

det(l - A) ~ O. If the last condition holds, then (I - A)-I - I E P.

Proof: Due to Theorem 3.1 the function det(l- X) , X E P is uniformly contin-

uous in some neighborhood of X = O. This implies that there exists a 8 > 0 such
that for all X E P with IIXlip < 8, the inequality det(l - X) ~ 0 holds . The
operator A can be represented in the form A = A (n) + R n, where II R; lip < 8 and
A (n) is the n-th section of A which has finite rank. We shall assume that 8 < 1.
Then II s, II ::: II Rn lip < 1 and the operator I - Rn is invertible . Taking into
consideration that
+L
00

(I - R n ) -I = I R~
j=1

The series

is convergent in the sense that

m
R- L R~ ~ 0, m~ 00.
j=1 p

This follows from the fact that

k+m
lim
k.m-e-co
L IIRnll~ = O.
j=k+1

Since P is complete, the operator REP. Represent I - A in the form

I - A = I - A(n) - R n = (1- Rn)(I - (1- Rn)-IA(n») .

The operator

and from Theorem 3.3 follows

det(l- A) = det(l- R n) det(l- (I + R)A(n») . (4.1)

Since A (n) is of finite rank, the operator (I + R)A(n) is of finite rank and the
operator I - (I + R)A(n) is invertible if and only if det(l - (I + R)A(n») ~ O.
Now it is easy to conclude that the operator I - A is invertible if and only if
332 Chapter XIV. Poincare Operators

det(l- A) i= O. Note also that if the last condition, holds then in the decomposition
(.2= 1m Pn EBlm(l- Pn), where Pn is the orthogonal projection onto the subspace
spanned by the first basis vectors el , ... , en, the following equality holds

Hence

I _ (I + R)A (n) = (Zl1

-Z21
0)
I '

where

From here it follows that

0)
I .

This implies

Z- 1 I
(I - A)-I - I =
((
11 ~I
Z21 ZI1
~) + I) (I - Rn ) - I
Z- I - I
--R
- n +( II
Z Z-I
21 11

Taking into consideration that

Z- I -
11 I
I 0)
0 E P and R; E P
(
Z21 Z li
we obtain (l- A)-I - IE P. The theorem is proved. o

Corollary 4.2 Let A E P and D,(v) = det(l - vA). The function D,(v) is entire
in C and D,(vo) = 0 if and only ifvo i= 0 and AO = ljvo is an eigenvalue of A.
Moreover if « is the multiplicity of the zero vo of the analytic function D,(v), then
K is the algebraic multiplicity ofthe eigenvalue AO = ljvo ofthe operator A.
14.4 Determinants and Inversion of Operators 333

Proof: All claims except the last follow from the previous theorem. For the
last statement we need knowledge about the multiplicity of an eigenvalue . This
material is outside the scope of this book. The proof of this statement can be found
in [GGKr] , Theorem 11.6.2. 0

Now we shall use the determinant det(l - G), G E P to obtain formulas for
the inverse operator (I - G) -I. Let us assume that G E P and det(l - G) I- 0.
Then by Theorem 4.1 the operator 1- G is invertible and (I - G) -I - I E P.
Represent o» on the entire space £2 = Im Pn EB Im (I - Pn ) by the matrix

(G 0)
n

° °'
where Pnx = {XI, X2, . .. , Xn , 0, 0, . . .} and G n is the n-th section of G. It is clear
that I - G(n) is invertible for n sufficiently large. We shall now prove that

lim 11(1 - G) -I - (I - G(n»)-Ilip = 0.

n---+oo
We start with the equalities

I - C<n ) = (I - G )(I - (I - G)-I (c<n ) - G))

= (I - G )(I - [(I - G)-I - I] + I) (c<n) - G)

Introduce the operator

X; = (1- G )-I (G (n) - G) = ([ (1 - G)- I - I] + I) (G (n) - G)

which is in P and

IIX nllp :s (11(1- G)-I - Ilip + l)IIG(n) - G llp.

Hence for n large enough , IIX n lip < 1 and

=L
00

(I - c<n»)-I = (I - Xn) -I (I - G)-I X~ (I - G) -I.

j =o

From here it follows that

As it is well known that for An = I - G (n),

-I 1
An = det An R n ,

where
R n = (r j kn»)oo
j ,k= 1
334 Chapter XN. Poincare Operators

is defined as follows . The matrix (rj~»'j,k=I is the transposed matrix of algebraic

complements of the entries of In - G n;
(n)
rkk = 1 for k= n + 1, ...
and rj~) = 0 for the remaining indices. It is obvious that

(detAn)A;;-I = R n.

Taking into consideration that

lim det A, = detA and lim IIA- I - A-III = 0,
n-+ oo n-+oo n

where A = I - G, we obtain that R n converges to an operator R E £(£2) which

we shall call the operator of algebraic complements of A :

lim IIR - R n II = O.
n-+ oo

In particular
. (n)
I1m. rjk = Fjk »
n-e-co jk

where R = (rjdfk=I'
It is obvious that
1
A= - - R .
detA
This equality generalizes, for operators of the form A = I - G, G E P, Cramer's
rule of inversion of finite matrices. Now we shall give another representation of
the inverse (I - J.tG)-I in terms of a series .

Theorem 4.3 Let G E P . Then the following equality holds:

(I - J.tG)-1 = I + det(I ~ J.tG) D(G, J.t),

where D(G; J.t) is an entire operator valued fun ction

= L a; (G)J.tm,
00

D(G; J.t)
m=O
of which the coefficients d m = dm(G) are operators from P which are uniquely
determined by the recurence relations

do = G, dm = G(dm-I + ~m) m = 1,2, . ...

with
L ~ J.tj .
00

det(I - J.tG) = j
j =O
14.4 Determinants and Inversion of Operators 335

Proof: There exists a 8 > 0 such that for IJLI < 8 the operator I - JLC is invertible:
00

(/- JLC)-I = 1+ L JLJC j ,

j=1

where the series converges in the nonn of P . Consider the operator I - JLc (n),
where the sequence {c (n)} converges to C :

lim IIC(n) - Clip = O.

n -->oo

There exists a number 81 > 0 (81 < 8) such that for IJL I :::: 81 ,
00

(/- JLC6n»-1 = 1+ LJL j(C 6n» j.

j=1

Using the computations preceding this theorem we come to the conclusion that
lim 11(/ - JLC(n» -1 - (l - JLC)-Ilip = 0,
n--> OO

uniformly in the disc IJLI :::: 81. We start now with the formula

(I _ C (n» - 1 _ I R ( )
JL - det(l _ JLc (n» n JL

where Rn(JL ) is the matrix of algebraic complements for I - JLC6n ) (see the
definition preceding Theorem 4.3). It is clear that Rn(JL ) is a polynomial in JL
with operator valued coefficients. After using the formula
(/- JLC(n» -1 = 1+ JLC(n)(l- JLC(n» - 1
we have
(l - JLC(n» -1 = 1+ I JL c<n)Rn (JL )
det(/ - JLc (n»
Denote c (n)Rn(JL) by D (n )(JL ). Hence
D (n)(JL ) = (l - JLC(n» -IC(n)det(l - JLC(n».

Therefore

uniformly on the disk IJLI :::: 81 .

Since D (n)(JL ) is a polynom ial in JL we may conclude that D (n)(JL ) converges
unifonnl yon IJLI :::: r , where r can be taken arbitrarily large. Thus on C there
exists an operator valued entire function D (JL ) such that

lim II D(JL) - D (n)(JL ) lip

n-->oo
= 0,
336 Chapter XN. Poincare Operators

uniformly in any disk I/LI < r . It is obvious that

Since

and
00 00

D(/L) =L /Lj d]; det(l -/LG) =L /Ln !'1 n,

j=O n=O

we obtain
00 00 00

L/Lndn - L/Ln+IGdn = GL/Ln!'1n .

n~ n~ n~

Thus, comparing coefficients, we get

do = G; and d n = Gdn-l + G!'1n

or
do = G and dn = G(!'1n +dn-lI).
Now we come to the conclusion that D(G; /L) = D(/L). This ends the proof. 0

We note that the last recurrence relation can be further simplified. More than
that, it can be solved in closed form as a difference equation. For these results we
refer to [GGKl], Ch VII Section 7 and [GGKr] ChVIII Section 1.

14.5 Trace and Determinant Formulas for Poincare Operators

The main theorem in this section gives formulas for the trace and determinant
of a Poincare operator in terms of its eigenvalues. Here we rely, in part, on
complex function theory. In particular, we shall use the following theorem, of
which the proof can be found in [Ah] Chapter 5 sec. 2.3 and [GGKl]
Theorem VII.5.l . It is a special case of Hadamard's factorization theorem.

Theorem 5.1 Let a i , a2, . . . be the zeros ofan entire function f ordered according
to increasing absloute value s with multiplicities taken into account. Assume that

1
~ lajl < 00 .
J
14.5 Trace and Determ inant Formulasfor Poincare Operators 337

Supp ose f (O ) = 1 and f or each e > 0 there exists a constant C (€) such that
IAje
If (A) 1 S C (€)e , AE C
Then f admits the representa tion

f(A) = n(1 - ~)
J

which converges uniformly on compact subsets of C.

We know from Corollary 4.2 that for A E P, the function det(l - AA ) is entire
and det(l - AoA ) = 0 if and only if AO i= 0 and l/Ao is an eigenvalue of A . The
order of the zero AO of det(l - AA) is called the algebraic multiplicity of l/Ao as
an eigenvalue of A .
The following result, wh ich we do not prove, appears in [GKre],Theorem 1.4.2.

Theorem 5.2 Suppose that {A n} is a sequence of compact ope rators in .c(1-0

(H a Hilb ert space) which converges in the uniform norm to an ope rator A E .c(H).
Then f or an app ropriate enume ration of the eigenvalues Aj (An) of the ope rators
An, we have

For the proof of the main theorem about trace and determinant we will need the
following properties of the eigenvalues of Poincare operators.

Lemma 5.3 * Let A = (ajk)'J,k=1 be an n x n matrix and let AI, A2, . . . , An be the
eigenvalues of A, counting multipli cities. Then
n n
L IAjl S L lajkl· (5.1)
j= 1 j.k=1
Proof: Let U = (UjdJ,k=1 be a unitary matrix such that UAU * is upper
triangular with AI, A2, . . . , An on the diagonal. Take B to be the diagonal matrix
diag tr}, tz . .. . , tn ) where Aj tj = IAj l, Itjl = 1, 1 S j S n . Then for S =
BUAU* = (Sjk)j,k=1 it is clear that
n n
L IAjl = L:>jj. (5.2)
j=1 j= 1

Now for U * = (Vjk)j,k=I'

n
I>jj = LL L
n n (n )
tj ujpapq Vqj.
j=1 j =l q=1 p=1

*) The authors are indebted to N. Krupnik for this lemma.

338 Chapter XN. Poincare Operators

Summing first with respect to j we get

< t (E laNI lu/p I2) 1/2 (E IV~12) 1/2 (5.3)

Since U is unitary, L:J=I IUjplZ = L:J=I IVqjlZ = 1.

Hence inequality (5.1) follows from (5.2) and (5.3). D

Theorem 5.4 Let A E P and let Al (A) , Az(A) , ... be the non-zero eigenvalues
of A counted according to their algebraic multiplicities. Then

L IAj(A)1 s IIAllp ·
j

Proof: Let An be the nth-finite section of A = (ajdfk=1 and let Al (An),

Az(A n), ... , An (An) be the non-zero eigenvalues of An counted according to their
multiplicities. Then

by Theorem 5.2. From Lemma 5.3 we have for n > k,

k n
L IAj(An)l:s L lajkl :s IIAllp·
j =1 j ,k=1

Hence
k

L IAj(A) 1:s IIAllp ·

j=1
Since k was arbitrary,
L IAj(A) 1:s IIAllp ·
j

For the application of Theorem 5.1 to the function det(I - AA) we need the
following theorem .

Theorem 5.5 Let A E P. Given e > 0, there exists a constant C(s) such that

Idet(I - AA)I :s C(s)eCIA1, A E C,

14.5 Trace and Determinant Formulas for Poincare Operators 339

D( E
Proof: Since A E P , we have from Section 1 that

Id"U - "A) I " 1 + 1"1 lai'l) , (5.4)

where L~l L~l lajkl < 00 . Given e > 0, there exists a positive integer N
such that L~N+l L~l lajkl < s/2. Let c, = L~l IUjkl, j = 1,2, . . .. Since
1 + x S e", x 2: 0, it follows from (5.4) that

Idet(l - AA)I s n
N
(l + IAICj) n
00

j=N+I
(l + IAICj) (5.5)

"D D
j=1

(I i
+ I"IC )exp (1"1 i i i Ci) " (I + l"ICi)' !"'. (5.6)

Now the function (l + Cjx)e-e/2x is bounded on [0,00). Hence it follows that

there exists a constant C(s) such that nJ=1 (l + IAICj) e-IAl e/2 S C(s) . Thus
from (5.5) we get
Idet(l- AA)I S C(s)e IA\e/2eIAl e/2 = C(s)eeIAI .
o
Now we are ready to prove the main theorem of this section.

Theorem 5.6 Let A = (ajk)'rk=l E P and let Al (A), A2(A) , . .. be the sequence
ofthe non-zero eigenvalues ofA counted according to their algebraic multiplicities.
Then
det(l - AA) = n
(l - H j (A)) .
j
(i)

The convergence ofthe product is uniform on compact subset of c. Furthermore,

00

tr A := '~::~>kk = I>j(A) . (ii)

k=1 j

Proof: (i) From the remark preceding Theorem 5.2, we know that the zeros of
the entire function det(l - AA) are AA A) of order, by definition, the algebraic
multiplicity of Aj (A) for A . Statement (i) now follows from Theorem 5.1, applied
to f(A) = det(l - AA) . The conditions of this theorem are met by Theorems 5.4
and 5.5.
(ii) Since the function det(l - AA) is entire, det(l - AA) = L~o CnAn, A E C-
Let An be the nth-finite section of A . Now
n n
det(l-AA n) = n(l-Hj(A n)) = LCkn Ak ,
j=1 k=O
340 Chapter XN. Poincare Operators

where
n n
Cln =- L:>l.k(A n ) =- Lakk.
k=1 k=1
Since det(/ - AA n ) converges uniformly on compact subsets of C to det(/ - AA),
it follows that the sequence {cln} converges to CI, i.e.,
00

CI = - Lakk . (5.7)
k=1
On the other hand, we have from (i) that the sequence of polynomials
nJ=1 (l - U) (A)) denoted by "LJ=o dinA} converges uniformly on compact sets
to det(/ - AA). Hence the sequence {din} converges to CI. Since
n
din = - LA}(A),
}=I

we obtain

By (5.6)
00

CI = - Lakk .
k=1
o
Much of the material in this chapter has its ongm in the book [GGKr],
Chapters I-III, as well as in the books [GGKl] and [GKre]. Further develop-
ments of the theory of traces, determinants and inversion can be found in these
books also along with a list of other sources on this subject.

Exercises XIV

1. Let X and Y be any operators in L:(cn) , and suppose CI (n) and C2(n) are
positive constants such that
(a) ItrX-trYI::sCI(n)IIX-YJI
(b) Idet(/ - X) - det(/ - Y)I ::s C2(n)IIX - YJI
Prove that
lim CI (n) = 00 and lim C2(n) = 00,
n--+oo n--+oo

for any operator norm.

Exercises XN 341

2. Prove that there exists a number r > 0 such that for F, G E F(f.2) with

IlFllp < r, IIGlip < r.

The following inequality holds :

3. Prove that in a small enough neighborhood IAI < 8 the equality

1
L -n tr (Fn)A n.
00
det(/ - AF) = exp
n=!

holds, where F E P.
4. Prove that for any F E F(f.2) and A E C,

C (F)
det(l- AF) = 1 + L
rank F
(_l)n_n-,_An ,
n.
n=!

where
tr F n-1 o
tr F 2 tr F n-2
Cn(F) = det .
(
tr F n

5. Prove that tr(X), X E P, is the only linear functional, continuous in II . lip,

with the properties

tr(XY) = tr(XY) X, YEP

and

1 0 . . .)
tr ( ~ ~ : : : =1.

6. Prove that det(l - X) is the only multiplicative functional, continuous in

II . lip, with the property that
det[(l - X)(l - Y)] = det(l - X) det(l - Y)

for X, YEP.
7. Construct a finite rank operator F E F(H) such that F fj. P .
342 Chapter XN. Poincare Operators

8. Let {cp j } f be an orthonormal basis in Hilbert space H . Denote by P rp the set

of operators in £(H) such that their matrices in the basis {cp j}f belong to
P . Let {1fr j } f be another orthonormal basis in H. Find an operator X E P rp
such that X rt. P1{f.
9. Let K E £(Lz(a, b» be defined by the equality

(K cp)(t) = l b
k(t , s)cp(s) ds,

with the kernel

n
k(t , s) = L !j(t)gk(S) ,
j=l
where I ). g j E Lz(a, b), j = 1,2, ... , n.

l
Prove that
b
~>.j(K) = k(s, s) ds
J

10. Denote by HSrp the set of all operators A in £(H) such that their matrices
(ajk)),'k=1 in the orthonormal basis {cpj}f satisfy the condition
00

L lajkl <
Z
00.
j,k=1
Let {1fr j } f be another orthonormal basis in H. Find an operator X E H Srp
such that X rt. H S1{f.

11. (a) Compute the determinants of the following matrix

.. .
...
)
under the conditions that
00

LlajlZ < 00,

j =l

and

L IXj -
00

11 < 00.
j=O
(b) The same problem under the same conditions for the special case when

Xj = l-a jbj.

In this case the last condition follows from the previous ones .
Exercises XN 343

12. Let {ej }~-oo be the standard orthonormal basis of lz(Z) and denote
by P (Z ) the set of all operators A E £ (l z(Z )) such that their matrice s
(ajdfk=- oo in the basis (e j } ~- oo satisfy the condition

00

II AII 'P(z) = L laj kl < 00.

j .k=- oo

Denote by A (-n ,m) the finite section

(a) Let A E P (Z), then

lim det(I - A(- n,m»)

n , m ~oo

exists. Denote this limit by

det(I - A ).

(b) Prove that det(I - X ) , X E P (Z ) is a Lipschitz function in the nonn

1I11'P(z),
(c) Prove that for A E P (Z ) the operator I - A is invertible if and only if
det(I - A ) i= O. If this condition holds, then (I - A)- I - IE P (Z ).

13. Let F E F (l z(Z )). We assume that (fjdfk =-OO is the matrix of F in the
standard basis {ej } ~ -oo of lz(Z) . Denote by F (- n,m) the operator defined
by the section
(fjdj,k=- n'
Define tr F and det(I - F ) in the same way as these quantities are defined
for F E F(l z); see Section 2.
(a) Prove that
00

tr F = lim tr F (-n ,m) = " ikk

n,m->oo L..
k=-oo

and the series converges absolutely.

(b) Prove that

det(I - F ) = lim det(I - F (-n,m»).

n,m->oo
344 Chapter XN. Poincare Operators

14. Compute the determinant det(/ - V), where the operator V is defined in the
standard basis of £z (Z) by one of the following matrices:

(a)

b-4 0 0 0 0 0
0 b-3 0 0 0 0
0 0 b-z 0 0 0
V= a-3 a-Z a-I b-I 0 0
0 0 0 ao al az
0 0 0 bi 0 0
0 0 0 0 bz 0

00

L lajl < 00;

j=-oo

ajj = Xj j = 0, ±1 , . ..

under the condition

L
00 00

L lajl2 < 00 IbjlZ < 00,

j=-oo j=-oo

and
00

L IXjl < 00 .
j=-oo

(c) V = (bjkrhk=-oo' where

bjk=ajbk j,k=O,±I, ...

under the same conditions as in b)

15. Construct a compact operator A E £(£z) defined in the standard basis of £2

by a matrix (ajdhk=1 such that neither of the following limits
n
lim "akk. lim det(8jk - ajk)J~ k-I
n-+oo L..... n-+oo '
k=1

exists .
Exercises XN 345

16. Let P(7L.) be the set of operators defined in problem 12. Suppose A =
(ajk)'fk=-oo E P(7L.) with its sequence of eigenvalues AI (A), A2(A), ...
counted according to their algebraic multiplicities. Prove that

det(l- AA) = n(1- j

AAj(A)).

17. Prove that in problem 16,

tr A = I:> j(A) .
j
Chapter XV
Fredholm Operators

Fredholm operators are operators that have a finite dimensional kernel and an
image offinite codimension. This class includes all operators acting between finite
dimensional spaces and operators of the form two-sided invertible plus compact.
Fredholm operators appear in a natural way in the theory of Toeplitz operators.
The main properties of Fredholm operators, the perturbation theorems and the
stability of the index, are presented in this chapter. The proofs are based on the
fact that these operators can be represented as finite rank perturbations ofone-sided
invertible operators.

15.1 Definition and Examples

Let X and Y be Banach spaces. An operator A E L(X, Y) is called a Fredholm

operator if the numbers n(A) = dimKer A and d (A ) = codim 1m A are finite.
In this case the number
ind A = n (A ) - d (A )

is called the index of A .

Examples: I. As a special case of Theorem XIIIA.I, we have that an operator

I - F with F E L (X) of finite rank is Fredholm with index zero.
2. Given any positive integers rand k , there exists an operator A on lz with
n (A ) = k and d (A ) = r. Indeed, define A on lz by

A (al , a z , . . .) = (0, 0, ... , 0, a k+l , a k+ Z, . . .) .

'-..-'
r

Then Ker A = sp [ej , . .. , ekJ and l z = 1m A $ sp {el ' . . . , e. }, where el, ez, .. .
is the standard orthonormal basis in l z. Thus

n (A ) = k, d (A ) = r, ind A =k- r.

15.2 First Properties

Theorem 2.1 The range ofa Fredholm operator A E L (X , Y) is closed.

348 Chapter xv. Fredholm Operators

Proof: Since n(A) < 00 and d(A) < 00, there exist, by Theorem XI.5 .7, closed
subspaces M and N such that

x =M EB Ker A, Y = 1m A EB N .
Let M x N be the Banach space with norm lI(u, v) II = lIuli + IIvll . Define the
operator A on M x N by A(u , v) = Au + v. Since A is an injective bounded
linear map from M x N onto Y , we have that A is invertible by Theorem XII.4.1.
Thus for any u E M,

(2.1)

Hence 1m A = 1m AIM is closed by Theorem XII.4.6. o

Theorem 2.2 Suppose A E £(X , Y) and B E £(Y, Z) are Fredholm operators.
Then BA is a Fredholm operator with

ind (BA) = ind B + ind A .

Proof: Since Ker A is a finite dimensional subspace of Ker BA , there exists a
closed subspace M of Ker BA such that

Ker BA = Ker A EB M . (2.2)

It is easy to verify that the restriction of A to M is a one to one map from M onto
1m AnKer B . Hence

dim M = dim (1m AnKer B) :s n(B) < 00.

Thus by (2.2),

n(BA) = n(A) + nt, nt = dim Ns , N; = 1m AnKer B . (2.3)

Since N, is a subspace of the finite dimensional space Ker B , there exists a sub-
space N: of Ker B such that

Ker B = N, EB Nz . (2.4)

Then
n(B) = nt + nz, nz = dim Nz. (2.5)
Since d(A) and dim N: are finite and 1m A n N: = (0), there exists a finite
dimensional subspace N3 such that

Y = 1m A EB Nz EB N3 (2.6)

and
d(A) = nz + n3 < 00, n3 = dim N3. (2.7)
15.2 First Properties 349

From
Ker B = Nl EB N2 C 1m A EB N2
and equalities (2.4) and (2.6), it follows that B is one to one on N3 and

1m B = 1m BA EBBN3. (2.8)

Since Z = 1m B EB W for some finite dimensional subspace W, we have from

(2.8) that
Z = 1m BA EBBN3 EB W ,
which implies that

d(BA) = dimBN3 + dim W = dim N3 + d(B) = n3 + d(B) < a. (2.9)

From (2.3), (2.5), (2.7) and (2.9), we have that the operator BA is a Fredholm
operator with

ind (BA) = n(A) + nl - d(B) - n3 = n(A) + n(B) - n2 - d(B) - n3

= n(A) + ind (B) - d(A) = ind (A) + ind B .
D

Theorem 2.3 An operator A E .c(X, Y) is a Fredholm operator if and only if A

has the representation
A = D+F, (2.10)
where F has finite rank and D is a one-sided invertible Fredholm operator.

Proof: Let A have the representation (2.10) . Assume that D(-l) is a right inverse
of D. Then
A = D(1 + D(-l) F) .
Notice that by Theorem XIII.4.1 the operator I + D(-l) F is Fredholm and has
index zero. Hence A is Fredholm and ind A = ind D by Theorem 2.2. A similar
argument shows that A is Fredholm if D has a left inverse.
Suppose that A is Fredholm. There exists a linear independent set {Yl , Y2, . . . , Yn}
and a closed subspace M such that

Y = 1m A EB SP{Yl , Y2, . . . , Yn}, (2.11 )

and
x = MEBKer A .
Let {Xl, X2, . . . x m } be a basis for Ker A . By Corollary XI.5 .6, linear functionals
Ii, 12,···, t- , which are bounded on X, may be chosen so that !j(Xk) = Ojk,
j , k = 1,2, ... , m, where 0j k is the Kronecker delta and for every X E Ker A,
m

X= L h(x)Xj .
j=l
350 Chapter xv. Fredholm Operators

Define a finite rank operator F E LeX, Y) by

min(m ,n)
F(x) = L !j(x)Yj' (2.12)
j=l

Setting D = A - F, we have

1m D = 1m A EB 1m F. (2.13)

Indeed, it is obvious from (2.11) and (2.12) that 1m A n 1m F = (0) . Clearly

1m D C 1m A EB 1m F . Given Ax + L:f=l <XjYj E 1m A EB 1m F, where p =
min(m.n), it follows from F(Xk) = Yk, k = 1,2, . . . , p and Xk E Ker A that
p

Du=Ax+ L<XjYj ElmAEBImF,

j=l

where u = x - L:f=l h(x)Xj - L:f=l <XjXj . Thus equality (2.13) holds. If

n :s m, then from (2.13), (2.11) and F(Xk) = Yk, 1 :s k :s n , we have

Y = 1m D , n z: m.
It is easy to see that Ker D = Sp{Xn+l, ... ,xm } for n < m and

Ker D = {OJ, n >: m . (2.14)

Thus Ker D is complemented in X. These results combined with Theorem XII .6.2
show that D is right invertible when n :s m , i.e., d(A) :s n(A) . For n ::: m,
i.e.,d(A)::: n(A),wehavefrom(2.13),(2.14) and TheoremXII.6.1 thatn(D) = 0
and d(D) :s d (A) < 00, i.e., D is a left invertible Fredholm operator. Thus (2.10)
follows. 0

Summarizing the above results, we have the following corollary.

Corollary 2.4 Suppose A E LeX, Y) is a Fredholm operator. Then A has the

representation
A=D+F
where F has finite rank and D is a one-sided invertible Fredholm operator.
Moreover,

(i) if ind A ::: 0, then D is right invertible;

(ii) if ind A :s 0, then D is left invertible;
(iii) if ind A = 0, then D is invertible.
J5.2 First Properties 351

Lemma 2.5 Given A E LeX, Y) and B E Ley, Z), where X, Y and Z are Banach
spaces, suppose that AB = F and BA = G are Fredholm operators. Then A and
B are Fredholm operators and

ind (AB) = ind A + ind B.

Proof: Since 1m F C 1m A and Ker AcKer G, it follows that d(A) :::: d(F) <
00 and n(A) :::: neG) < 00. Thus A is a Fredholm operator. An analogous
argument shows that B is a Fredholm operator. Now apply Theorem 2.2. 0

Theorem 2.6 Suppose A E LeX, Y) is a Fredholm operator. Then

(a) the conjugate operator A' is a Fredholm operator,
(b) 1m A' = (Ker A)..L, 1m A = ..L(Ker A '),
(c) n(A') = d(A), d(A') = n(A),
(d) ind A' = -ind A.

Proof: (a) Since A is a Fredholm operator, A = D + F, where D is a one-

sided invertible Fredholm operator and F has finite rank. We shall show that
the conjugate operator D' is a Fredholm operator. Let D(-I) denote a one-sided
inverse of D . If D(-I) is a left inverse of D, then D(-I) D = I and I - DD(-I)
is a projection onto a subspace complementing 1m D by Theorem XII.6.1 . Since
d(D) < 00, it follows that I - DD(-I) = FI is a finite rank operator. Taking
conjugates we see that D'(D(-I)), = I and (D(-I)), D' = 1- F{ . Since FI has
finite rank, so does F{ by Example 1 in Section XIII.5 . Hence D '(D(-I)), and
(D(-I)), D' are Fredholm operators by Theorem XII1.4.1. Lemma 2.5 shows that
D' is a Fredholm operator. If D(-I) is a right inverse of D, then DD(-I) = I and
by Theorem XII.6.2, I - D (-I) D = F2 is a finite rank operator with range the
finite dimens ional space Ker D . Arguing as above, we see that D' is a Fredholm
operator. Since (D(-I)), is a one-sided inverse of D', it follows from Theorem 2.3
that A' = D ' + F' is a Fredholm operator.
(b) Since A and A' are Fredholm operators, there exist finite dimensional
subspaces Yo, W and a closed subspace Xo such that
X = Xo EEl Ker A, Y = 1m A EEl Yo, X' = 1m A ' EEl W.
Theorem XIII.5 .1 shows that 1m A = ..L(Ker A') and 1m A ' ~ (Ker A)..L. To
prove (b) it remains to show that (Ker A)..L C Im A ' . Suppose x ' E (Im A)..L.
Define the functional f on 1m A = AXo by

f(Axo) = x '(xo), Xo E Xo.

Since A is one to one on Xo and AXo is closed, it follows from Theorem XII.4.6
that there exists m > 0 such that IIAxoll ~ mllxoll for Xo E Xo. Thus
, 1 ,
If(Axo)l:::: IIx IIl1 xoll :::: -lix III1 Ax oll·
m
352 Chapter xv. Fredholm Operators

Hence f is a bounded linear functional on 1m A . By the Hahn-Banach theorem,

f has a bounded linear extension g to all of Y. Now for Xo E Xo and z E Ker A,
A'g(xo + z) = g(Axo ) = f(Axo) = x' (xo) = x' (x + z) .
Therefore x ' = A' g E 1m A' and (b) is proved.
(c) Equality n(A') = d(A) appears in Theorem XIII.5 .2. We have

X' = 1m A' E9 W = (Ker A).l E9 W. (2.15)

Define qJ : W -+ (Ker A)' by

where W R is the restriction ofthe functional w to Ker A . The functional qJ is linear

and one to one, for if qJ( w) = W R = 0, then W E Ker A.l n W = (0). Moreover, qJ
maps W onto (Ker A) ' . Indeed, given h E (Ker A) ' , there exists a bounded linear
extension g' of h to X , by the Hahn-Banach theorem. From equality (2.15), the
functional g E X' has the representation

g = u + v, u E (Ker A).l , v E W.

Hence g - u = V E Wand for x E Ker A ,

qJ(V) = VR, VR(X) = g( x) - u(x) = g(x) = hex).

Thus qJ(v) = h. Hence, since Ker A is finite dimensional,

d(A') = dim W = dim qJ(W) = dim (Ker A)' = dim Ker A = n(A).

Thus (c) is proved. The index formula (d) is immediate . o

The following corollary follows directly from Theorem 2.6

Corollary 2.7 Let A E LeX, Y) be a Fredholm operator. Then Ax = y has a

solution if and only if A' qJ = 0 implies qJ(y) = O. Also, the equation A'1{r = f
has a solution if and only if Au = 0 implies f (u) = O.

15.3 Perturbations Small in Norm

Theorem 3.1 Given a Fredholm operator A E £(X , Y), there exists a y > 0 such
that if B E £( X, Y) and II B II < y, then A + B is a Fredholm operator with
ind(A + B) = indA .

Furthermore,
n(A + B) :::: n(A), d(A + B) :::: d(A) .
15.3 Perturbations Small in Norm 353

Proof: Since A is Fredholm , we have from Theorem 2.3 that A has the
representation A = D + F , where D is a one-sided invertible Fredholm oper-
ator and F has finite rank. Let D (-I ) denote a one-sided inverse of D. We know
from TheoremsXII.6.3 andXII.6.4, that for IIB II < IID(- l) II-I, D+B is invertible
from the same side as D and

n(D + B) = n(D) < 0, d(D + B) = d(D ) < 00. (3.1)

Suppose D (-I ) is a right inverse of D . Then

A +B = D +B +F = (D + B)(J + (D + B)-I F). (3.2)

Hence by Theorem 2.2 and equalities (3.1) and (3.2),

ind (A + B) = ind (D + B) = ind D = ind (D(J + D - I F» = ind (A). (3.3)

A similar argument shows that (3.3) holds if D(-I) is a left inverse of D.

Next we show that n(A + B) ~ n(A) . Suppose ind (A) ~ O. We see from
(2.12) and (2.13) in Section 2 that the operator F may be chosen so that

dim Im F = n(A) . (3.4)

We have from Corollary 2.4 that the operator D is left invertible and, as noticed
above, D + B is also left invertible. Thus Ker (D + B) = {OJ and

A + B = D + B + K = (J + F(D + B )(-I)(D + B ). (3.5)

Since Ker (J + F(D + B )(-l » c 1m (F(D + B) (-I », it follows from (3.4) and
(3.5) that

n(A + B) = dimKer (A + B)
~ dim Ker (J + F(D + B)(-I » + dim Ker (D + B)
= dimKer (J + F(D + B)(-I» ~ dim 1m (F(D + B (-I »)
~ dim 1m F = n(A) . (3.6)

Since ind (A + B) = ind A , we have that d(A + B) ~ d(A). Now suppose

ind A 2: O. Then ind A' = -ind A ~ 0 by Theorem 2.6. Hence the above result
applied to A', together with Theorem 2.6, yield

d(A + B) = n(A' + B') ~ n(A') = d(A) . (3.7)

Since ind (A + B) = ind A , inequality (3.7) implies n(A + B) ~ n(A) . 0

The above theorem yields the following corollary.

354 Chapter XV Fredholm Operators

Corollary 3.2 The spectrum ofthe forward shift S on l p, 1 ::s p ::s 00, is equal
to the closed unit disc. Furthermore,for IAI < 1 the operator AI - S is Fredholm
with
n(AI - S) = 0, d(AI - S) = 1, ind (AI - S) = -1 , (3.8)
and AI - S is not Fredholmfor IAI = 1.

Proof: Since IISII = 1, we have a(S) C {A I IAI s I}. To prove equality,

it suffices to show that AI - S is not two-sided invertible for IAI < 1. To do
this, let A be the backward shift on lp. Then AS = I and IIAII = 1. Now take
IAI < 1 = II A II-I . Thus, by Theorem XII. 6.3, the operator AI - S is left invertible
and
codim 1m (AI - S) = codim 1m S = 1.
It follows that AI - S is not two-sided invertible. Also we have proved (3.8).
Fix IAI = 1. Suppose AI - S is Fredholm. Then, by the previous theorem, for
IA - ,81 sufficiently small the operator ,81 - S is Fredholm and

ind (AI - S) = ind (,81 - S).

But this is impossible because ind (,81- S) = -1 for 1,8 I < 1 and ind (,81- S) =0
for 1,81 > 1. Thus AI - S is not Fredholm for IAI = 1. 0

Examples: 1. As a further illustration ofTheorem 3.1 consider the backward shift

A on l p, 1 ::s p ::s 00, i.e., A(al, a2, . . .) = (a2, a3, . ..). Then, as was shown in
Example 3 ofSectionXII.8, the spectruma(A) = {A IIAI ::s I}. IflAI = 1, then
A - AI is not a Fredholm operator. To see this, suppose A - AI were Fredholm.
Then if 1,81 > 1 and,8 "close" to Awe have ,8 E peA) and

0= ind (A - ,8I) = ind (A - AI) . (3.9)

If 1111 < 1 and 11 " close" to A, then it is easy to see that n(A - 11I) 1 and
,8(A - 111) = O. Hence

1 = ind (A - 11I) = ind (A - AI) .

which contradicts equality (3.9).

2. Let X = C([O, 1]). Define A E L(X) by (Af)(t) = tf(t) . It was shown in
XII.8, Example 2 that the spectrum of A is given by

a(A) = [0, 1].

If A E [0, 1], then A - AI is not a Fredholm operator. Indeed, suppose A - AI

is a Fredholm operator. Then for,8 rt [0, 1] and,8 "close" to A, ,8 is in peA)
and 0 = ind (A - ,8I) = ind (A - AI). It is clear that A - AI is one to one.
Hence d(A - AI) = ind (A - AI) = O. But this shows that A E peA), which is a
contradiction.
15.4 Compact Perturbations 355

15.4 Compact Perturbations

The main result in this section is the following theorem.

Theorem 4.1 Let A E £ (X , Y) be a Fredholm operator and let K E £ (X , Y) be

compact. Then A + K is a Fredholm ope rator with
ind (A +K) = ind A
The basic step in the proof of the theorem relies on the following result.

Theorem 4.2 If K E £ (X , Y) is compact, then I - K is Fredholm with index

zero.
Proof: The subspace Ker (/ - K ) is finite dimens ional. Indeed, if {Xk} is a
sequence in the l-ball of Ker (/ - K ), then {xd = {K xd has a convergent
subsequence and therefore Ker (/ - K ) is finite dimens ional by Theorem XI.2.4 .
Thus X = Ker (/ - K ) EB M , where M is a closed subspace of X. We now prove
that Im (/ - K ) = (/ - K )M is closed. It suffices, by Theorem XII.4.6, to show
the existence of m > 0, such that
11(/ - K )ulI ~ m llull for all U EM. (4.1)
If(4. 1) fails for every m > 0, there exists a sequence {un} C M such that II Un II = 1
and (/ - K )u n ~ O. Since K is compact, there exists a subsequence {un'} of {U n}
such that {K un'} converges. Hence lim n--+ o Un' = lim n--+ oo K Un' = U for some
U E M . Thus lI uli = I and

(/ - K )u = n--+
lim (/ -
oo
K )u n, = O.
But this is impossible since I - K is one to one on M and U =1= o.
Next we show that d(/ - K ) < 00. Assume this is not the case. Then for any
positive integer n , ther e exist linearl y independent vectors YI , Y2, .. . , Yn such that
1m (/ - K) n Sp{YI , Y2, ... , Yn} = {OJ. As we saw in the proofofTheoremXI.5.3,
there exist functionals /J, h, . . . , In in 1m (/ - K ).l.. such that Ii (y j) = oij ' Now
the set {II, h , . . . , In } is linearl y independent since 0 = L J= I a jI j implies
0= (LJ=I a j/j) (Yd = as, k = 1,2 , .. . , n . Hence n ~ dim Im (/ - K ).l.. =
dim Ker (/ - K ') by Theorem XIII .5.1. Since n was arbitrary, dim Ker (/ -
K ' ) = 00 . But this is impossible, for K ' is compact and, as we observed above,
Ker (/ - K ' ) is finite dimensional. Hence d(/ - K ) < 00 .
We have shown that I - K is Fredholm. Finally we prove that ind (/ - K ) = O.
For each J... E C, we have seen that I - J.. K is a Fredholm operator. Define
the mapp ing tp on the interval [0, I] by cp(J...) = ind (/ - J...K ). It follows from
Theorem 3.1 that cp is continuou s. Since cp is integer valued, ip must be a constant
function. Hence
0= cp(O) = cp (l) = ind (/ - K ).
This completes the proof. o
356 Chapter xv. Fredholm Operators

Theorem 4.2 is the full generalization of Corollary XIV.4.2.

Proof of Theorem 4.1: Since A is Fredholm, A = D + F, where D is a one-sided
invertible Fredholm operator and F has finite rank. Assume that D has a right
inverse D (-I) . Then
A +K = D +F +K = D(/ + D(-I)(F + K». (4.2)
Hence A + K is Fredholm and
indA + K = indD = ind (D(/ + D(-I) F» = ind (D + F) = indA
by Theorems 2.2 and 4.2 and equality (4.2). A similar argument shows that ind
(A + K) = ind A if D has a left inverse.
As a consequence of Theorems 2.3 and 4.1 we have the following result.

Corollary 4.3 Let A be in £(X, Y) . Thefollowing statements are equivalent :

(a) A is a Fredholm operator
(b) A has the representation A = D + F, where F has finite rank and D is a
one-sided invertible Fredholm operator
(c) A has the representation A = D + K , where K is compact and D is a
one-sided invertible Fredholm operator.

Using Theorem 4.2, the proof of the following theorem is the same as the proof
of Theorem XIII. 6.1.

Theorem 4.4 Let T be a compact operator in £(X), where X is an infinite dimen-

sional Banach space. Thenfor the spectrum a (T), one ofthefollowing equalities
hold:
a) a(T) = {OJ
b) a(T) = {A.I, A2 ,· ··, AN; O}
c) a(T) = {AI , A2, . . . ,; OJ.
All Aj are different from zero and are eigenvalues of T . In case c) we have
Iim Aj = O.
J-+OO

15.5 Unbounded Fredholm Operators

Theorems 3.1 and 4.1 can be extended to closed operators in the following manner.
First we extend the definition of a Fredholm operator.
A linear operator T (X ---+ Y) is called a Fredholm operator if the numbers
n(T) = dim Ker T and d(T) = codim Im T are both finite. In this case the
number ind T = n(T) - d(T) is called the index of T . The Sturm-Liouville
operator discussed in Section VI.5 is a closed Fredholm operator.
15.5 Unbounded Fredholm Operators 357

If T is a closed operator, we introduce the graph norm II . II T on the domain

V (T) ofT by
IIxliT = II xll + II Txll·
By the argumen t given in Section VI.3 we see that V (T ) with norm II . II T is a
Banach space which we denot e by V T. Obviously T is bounded on V T.

Theorem 5.1 Let T (X ---+ Y ) be a closed Fredh olm operator. There exis ts a
y > 0 such that for any operator B (X ---+ Y ) with V (B ) :J V (T ) and

IIBxll s y ( lI x ll + IITxll ) , x E V (T) , (5.1)

the operator T +B is closed and Fredholm with

(a) ind (T + B) = ind (T)

(b) n (T + B) S n (T )
(c) d (T + B ) S d (T)

Proof: Notice that V (T + B ) = V (T) . Since T is a bounded Fredholm operator

on V T , Theorem 3.1 shows that there exists y > 0 such that (5.1) implies that
statements (a), (b) and (c) hold. Without loss of generality we may assume that
y<l.
It remains to show that T + B is closed . Given x E V (T) = V (T + B ), we
have
II( T + B )x ll s yllx ll + (1 + y ) IITx ll (5.2)

and
II( T + B )xll 2: IITx ll- IIBx ll 2: (1 - y ) II T xll- Y llx ll ·
Hence
1
IITx ll s - - (yllxll
l- y
+ II (T + B )xll )· (5.3)

Suppose Xn ---+ x , Xn E V (T + B ), and (T + B)xn ---+ y. Then from (5.3) we see

that {T xnl is a Cauchy sequence which therefore converges to some v E Y . Since
T is closed, x E V(T ) and T x = lim n - H JO TX n = v . Hence by (5.2),

II (T + B) (x n - x )1I S Yllx n - x II + (1 + y ) IIT(x n - x) II ---+ o.

Therefore, (T + B )x = y which proves that T + B is a closed operator. 0

Theorem 5.2 Let T (X ---+ Y) be a closed Fredholm operator, and let K (X ---+ Y)
be a comp act operator on V T . Then T + K is a close d Fredh olm operator with

ind (T +K) = ind T.

358 Chapter xv. Fredholm Operators
Proof: Since T is a bounded Fredholm operator on VT and K is compact on V T,
the operator T + K is Fredholm with ind (T + K) = ind T by Theorem 4.2 . It
remains to prove that T + K is closed. Since n(T + K) < 00, there exists a closed
subspace M of VT such that VT = M ffi Ker (T + K). Now the restriction of
T + K to M is bounded on VT , is one-one and has range 1m (T + K) which is
closed because T + K is a Fredholm operator. Thus, by Theorem XII.4.6, there
exists an m > 0 such that
II(T + K)vlI ~ mllvllT, v E M. (5.4)
To see that T + K is closed, suppose X n ~ x and (T + K)x n ~ y . Now
Xn = V n + Zn , V n EM, Zn E Ker (T + K) and (T + K)v n = (T + K)x n ~ y .
It follows from (5.4) that the sequence {II V n II r l is bounded. Since K is compact
on VT , there is a subsequence {v n' } of {v n} such that K Vn' ~ w for some w E Y.
Hence Tx n, ~ Y - w. Since T is closed, x E V(T) and Tx = y - w. Moreover,
w = Kx . Indeed,
IIxn, - xllT = IIxn' - xII + II Tx n, - Txll ~ O.
Thus Kx n, ~ Kx = wand Tx = y - Kx or (T + K)x = y. Hence T + K is a
closed operator. 0

Example: Let H = L2([0, 1]). Define A(1t ~ 1t) by

V(A) = {f E 1t I I is absolutely continuous on[O, 1],1(0) = 0, I' E 1t} ,

AI=I'
Let b be a bounded Lebesgue measurable function on [0, 1]. Define B E £(1-{) by
(Bg)(t) = b(t)g(t) .

Then B is A-compact and A + B is a Fredholm operator with ind (A + B)

ind A = O. Indeed, we have from Section VII. 1 that A is an invertible operator

I
with inverse
t
(A-Ig)(t) = g(t)ds .

Thus A-I is a compact operator on 'H, To see that B is A-compact, suppose {gn}
is an A-bounded sequence, i.e., IIgn II + IIAgn II :s M, n = 1,2, ... . Then {Ag n}
is a bounded sequence in H, and since A -I is compact, {gn} = {A -I Agn} has
a convergent subsequence. Hence {Bg n} has a convergent subsequence which
shows that B is A-compact. From Theorem 5.2 we get that A + B is a closed
Fredholm operator with
ind(A + B) = indA = O.

For a thorough treatment of perturbation theory, the reader is referred to the paper
[GKrel], and the books [G], [GGKl] and [K]. Applications of Theorems 5.1 and
5.2 to differential operators appear in [G].
Exercises XV 359

Exercises XV

I. Let U be the operator on ep (I ::: p < 00) defined by

U(xo , xl , X2,· . .) = (0,0, . . . , 0, xk , Xk+l, . ..) .
'-..-'
r
As we have seen in Section I , the operator U is Fredholm and

n(U) = k, d(U) = r, indU = k - r.

(a) Prove that for the operator
U+(Xo, Xl, X2 , ... ) = (0,0, . . . ,0, x - , Xr+l, ... )
'-..-'
k

the operators U+ U - I and UU+ - I define projections on p and e

compute their action.
(b) Prove that for IAI > I the operator AI - U is invertible and that it is a
Fredholm operator with ind (AI - U) = k - r for IAI < 1.
(c) Let k i- r. Show that for IAI = I the operator AI - U is not Fredholm.
What happens in this case when k = r?
(d) Compute n(AI - U) and d(AI - U) for lA/ < 1.
(e) Determine the spectrum of AI - U for each A.
(f) Represent U as the sum of a one-sided invertible operator and a finite
rank one.
2. Let U and U+ be as in the previous exercise with k i- r, and consider the
operator
A = str: +bI + c U .
(a) Show that A is Fredholm if and only if
a(A) =aA- 1 +b+d i- 0 (IAI = I ) .

(b) If the latter condition on a holds, show that ind A = (k - r)m, where m
is the winding number relative to zero of the oriented curve t ~ a(e it ) ,
with t running from -71: to 71:.
Determine the Fredholm properties of the operator A when k = r .
3. Find the spectrum of the operator A defined in the previous exercise (with
k i- r) .
4. Let U and U+ be as in the first exercise with k i- r , and consider the operator
N M
B = I>- j (U +) j +aoI + 'La ju j ,
j= l j=l

where a-N , . . . , aM are complex numbers.

360 Chapter XV Fredholm Operators

(a) Show that B is Fredholm if and only if

M
f3(A) = L ajA
j
i= 0 (IAI = 1).
j=- N

(b) If the latter condition on f3 holds, show that ind B = (k - r)m, where m
is the winding number relative to zero of the oriented curve t --+ f3(e i ! ),
with t running from -77: to 77:.

5. Find the spectrum of the operator B defined in the previous exercise.

6. If I - An is a Fredholm operator, prove that I - A is Fredholm.
7. Let the Banach space X be a direct sum of the closed subspaces X I and X2,
i.e.,
X = XI E9X2.
In that case a bounded linear operator A on X is represented by a 2 x 2
operator matrix

where Aij is a bounded linear operator from X j into Xi (i, j = 1,2). In fact,
(*) means that for XI E XI and X2 E X2 we have A(xi + X2) = YI + Y2,
where
Yi = Ail XI + A i 2X2 E Xi (i = 1,2).

(a) If A is invertible and the operators A12 and A21 are compact, prove that
A II and A22 are Fredholm and

ind Aj , = -ind A22 .

(b) Is the previous statement true if only one of the operators A 12 and A21
is compact?
(c) If the answer to the previous question is negative, what conclusions can
nevertheless be made?
(d) If the operator A lion X I is Fredholm, show that there exist operators
A12, A21 and A22 such that A12 and A21 have finite rank and the operator
A given by (*) is invertible.
(e) Let All on XI and A22 on X2 be Fredholm operators, and assume that
ind All = -ind A22 . Showthatthere exist finite rank operators A12 and
A21 such that the operator A given by (*) is invertible.
Chapter XVI
Toeplitz and Singular Integral Operators

In this chapter we develop the theory ofLaurent and Toeplitz operators for the case
when the underlying space is lp with I :s p < 00. To keep the presentation as
simple as possible we have chosen a special class ofsymbols, namely those that are
analytic in an annulus around the unit circle . For this class of symbols the results
are independent of p . We prove the theorems about left and right invertibility and
derive the Fredholm properties. Also, the convergence ofthe finite section method
is analyzed. In the proofs factorization is used systematically. The chapter also
contains extensions ofthe theory to pair operators and to a simple class of singular
integral operators on the unit circle .

16.1 Laurent Operators on lp(2)

In this section we study the invertibility of Laurent operators on the Banach space
f p (2), where 1 :s p < 00. Such an operator L assigns to an element x = (x j) j EZ
in l p (2) an element y = (y j ) j e Z via the rule

L
00

Lx = y, Yn = an-kXk (n E Z) . (1.1)
k=-oo

In what follows we assume that there exist constants c :::: 0 and 0 :s p < 1 such
that
(1.2)

From (1.2) it follows that the operator L in (1.1) is well-defined and bounded. To
see this, let Y be the bilateral (forward) shift on f p (2), that is, Y is the operator
given by
(1.3)

Notice that Y is invertible and II y n II = 1 for each n E 2 . We claim that the action
of the operator L in (1.1) is also given by

L
00

Lx = a v yvx , x E lp(2). (1.4)

v=-oo
362 Chapter XVI. Toeplitz and Singular Integral Operators

Indeed, the series in the right hand side of (1.4) converges in fp(Z) to a vector
which we shall denote by Lx. Since the map (x j) j ell: --+ Xn is a continuous linear
functional on fp(Z), we have

00 00 00

(Lx)n = L av(VVx)n = L avxn-v = L an-kXk = (Lx)n.

v=-oo v=-oo k=-oo

This holds for each n, and thus Lx = Lx, and (1.4) is proved. Now,

(1.5)

Thus L is bounded and II L II :s L~-oo la v I.

With the operator L in (1.1) we associate the complex function
00

a(A) = L n
anA , A E 'Jr. (1.6)
n=-oo

Notice that condition (1.2) is equivalent to the requirement that a is analytic in an

annulus containing the unit circle . Furthermore, we can recover the coefficients
an from (1.6) via the formula

an = -1 j1f a(e't)e-
. mt. dt, n E Z. (1.7)
2n -1f

We shall call a the symbol of the Laurent operator L.

In the sequel we let A denote the set of all complex-valued functions a on the
unit circle that are analytic on some (depending on a) annulus containing the unit
circle. Notice that the set A is closed under the usual addition and multiplication
of functions. Also, if a E A and a does not vanish on the unit circle, then a-I,
where a-I (A) = Ija(A) for each A E T, belongs to A.

Theorem 1.1 Let La be the Laurent operator on f p(Z) with symbol a E A Then
La is invertible if and only if a does not vanish on the unit circle, and in that case

(1.8)

Proof: We split the proof into two parts. In the first part we assume that a does
not vanish on T and we prove (1.9) . The second part concerns the necessity ofthe
condition on a .
Part 1. It will be convenient to show first that for each a and f3 in A have

(1.9)
16.1 Laurent Operators on l p (Z) 363

To prove (1.9) it suffices to show that LaLfJej = LafJej for each vector ej =
(Ojn)nE'Z., where 0jn is the Kronecker delta. Notice that LfJej = (b n- j )nE'Z. , where

bn = _I_i
2rr
Jr

-Jr
{3(eit)e-int dt (n E Z).

Let an be given by (1.7) . Then according to (1.1) we have

00

(LaLfJe j)n = L an-kbk-j ' (1.10)

k=-oo

The right hand side of (1.10) is equal to the coefficient Yn- j ofAn- j in the Laurent
series expansion ofY(A) = a(A){3(A) . Thus the right hand side of (1.10) is also
equal to (LafJej)n. This holds for each n , and thus (1.9) is proved.
Now assume that a does not vanish on 1I', and put {3 = a-I. Then (1.9) yields

where I is the identity operator on lp(Z) . Thus (1.8) holds .

Part 2. Assume La is invertible . We have to show that a(A) =1= 0 for each A E 1I'.
This will be done by contradiction using arguments similar to those used in the
second part of the proof of Theorem 4.1 in Chapter III. So assume a(Ao) = 0 for
some AO E 1I', and put 8 = IIL;;III-I. By Chapter XII, Theorem 6.3 an operator I
on l p (Z) will be invertible whenever

III- La II < 8. (1.11)

Condition (2) implies that L~-oo Ian I is convergent. We choose a positive integer
N so that Lln! >N lanl < 8/2. Put

N
aN(A) = L n
anA ,
n=-N

and let LaN be the Laurent operator with symbol aN. Then

liLa - L aN II = L lanl < 8/2, (1.12)

Inl>N

la(A) - aN(A)1 S L lanl < 8/2 (A E 1I'). (1.13)

Inl>N

Put I = L aN -an (Ao) I. From (1.13) and a(Ao) = 0 we see that /a N(ao) I < 8/2.
By combining this with (1.12) we get that II La -III < 8, and hence I is invertible.
Now, put
364 Chapter XVI. Toeplitz and Singular Integral Operators

Notice that to is a trigonometric polynomial, and hence w is analytic in an annulus

containing T. Let L w be the corresponding Laurent operator. Since i. is the
Laurent operator with symbol aN(A) - aN ()"o), formula (1.9) and the fact that
(A - AO)W(A) = aN(A) - aN(Ao) imply that
i. = (V - AoI)L w = Lw(V - AOI). (1.l4)
Since L is invertible, we conclude that V - AoI is both left and right invertible,
and thus AO is not in the spectrum of V . But this contradicts the fact that lAo I = I
because a(V) = T. Thus a(A) =1= 0 for all A E T . D
The arguments used in the second part of the proof ofthe previous theorem also
yield the following result.

Theorem 1.2 Let L be the Laurent operator on l p (Z) with symbol a E A Then
L is invertible if and only if L is Fredholm.
Proof: We only have to establish the "if part" of the theorem . So assume L is
Fredholm . In view of Theorem l.l it suffices to show that a(A) =1= 0 for each
A E T. Assume not. So a(Ao) = 0 for some AO E T . By the stability theorems
for Fredholm operators there exists 8 > 0 such that an operator i. on l p (Z) will
be Fredholm whenever
ilL - ill < 8.
Now, using this 8, construct i. and W as in the second part of the proof of
Theorem 1.1. Then i. is Fredholm and (1.14) holds. This implies that the operator
V - AoI is Fredholm, which is impossible because IAol = 1. Thus a(A) =1= 0 for
A E T, and hence L is invertible. D

16.2 Toeplitz Operators on f p

In this section we study the invertibility of Toeplitz operators on the Banach space
l p» where 1 ~ P < 00 . Such an operator T assigns to an element x = (x j )1=0 in
lp an element Y = (Yj)1=O via the rule
00

Tx = Y, Yn = Lan-kXk, n = 0,1,2, . .. . (2.1)

k=O
Throughout we assume that the coefficients an are given by

an = -
1 . m. dt,
jJr a(e1t)e- t
nEZ, (2.2)
2rr -Jr

where a E A, that is, a is analytic in an annulus containing the unit circle T, or,
equivalently, for some c ~ 0 and 0 ~ p < 1 we have
Ian I ~ cp1n l (n E Z) . (2.3)
The function a is called the symbol of T.
16.2 Toeplitz Operators on l p 365

From (2.3) it follows that T is a bounded linear operator on lp . To prove

this, notice that l p may be identified with the subspace of l p (Z) consisting of all
(Xj)jEZ such that Xj = 0 for j < O. Let P be the projection of lp(Z) onto lp
defined by
P(... , X-I , XO, XI , .. .) = (.. . , 0, 0, XO , XI, .. .). (2.4)
Thus II P II = 1, and T = PL 11m P, where L is the Laurent operator on l p (Z) with
symbol a . Thus
00

(2.5)
n=-oo
Let eo, el , ei, . . . be the standard basis of lp. Then (2.1) means that

:~ a~1 :=~ :::] [~~] [~~ ]

[ ai
·
··
aI ao
.' .
. '.
X2 = Y2
..
.
'

with the usual matrix column multiplication. Therefore, instead of (2.1) we some-
times simply write

ao a_I a-2 ]
al ao a_I .
T = a2 al ao .
[
··· ..
.
To study the invertibility of Toeplitz operators we shall use the notion of a
Wiener-Hopf factorization for a function from A. So let a E A. A factorization
(2.6)
is called a Wiener-Hopffactorization of a if K is an integer, the functions a : and
a+ both belong to A and do not vanish on 11', and
(i) a+O±1 extends to a function which is analytic on the unit disk,
(ii) a.: O±I extends to a function which is analytic outside the unit disk
including infinity.
Since a E A, the conditions (i) and (ii) imply that the functions a+O±1 and
a_O±1 are also in A. Furthermore, from (i) and (ii) we see that a+O±1 and
a_O±1 admit series expansions of the following form :
00 00

a+(A) = L.a)A j, a+(A)-1 = L. yfAj (IAI s 1),

j=O j =O
00 00

a_(A) = L.a=jA-
j, a_(A)-1 = L. Y'=-jA- j (IAI :::: 1).
j =o j=o
366 Chapter XVI. Toeplitz and Singular Integral Operators

Thus the Toeplitz operators with symbols a+O and a+O- 1 are represented by
lower triangular matrices, and those with symbols a-O and a-O- I by upper
triangular matrices.
For a Wiener-Hopf factorization to exist it is necessary that a does not vanish
on T . The next theorem shows that this condition is also sufficient.

Theorem 2.1 A function a in A admits a Wiener-Hopffactorization if and only

if a does not vanish on T, and in that case the integer K in (2.6) is equal to the
winding number ofa relative to zero.

Proof: We have already seen that a(A.) =1= 0 for x E T is necessary. To prove
that this condition is also sufficient, assume that it is satisfied, and let K be the
winding number of a relative to zero . Put w ()..) = ).. -K a ()..). Then w is analytic
in an annulus containing the unit circle, co does not vanish on T, and the winding
number of w relative to zero is equal to zero. It follows from complex function
theory that w(A.) = exp f()..) for some function f that is analytic on an annulus
containing T . In fact, we can take f = log co for a suitable branch ofthe logarithm.
Thus for some 8 > 0 the function f admits the following Laurent series expansion:
00

f()..) = L fn)..n , 1-8 < 1)..1 < 1 +8.

n=-oo

Now put
00

f+()..) = Lfn)..n, 1)..1<1+8,

n=O
-I
f-(A.) = L fn)..n, 1)..1> 1 - 8,
n= - oo
and set
a+()..) = exp f+ ()..), a_()..) = exp f-()..).

Then w()..) = a_()..)a+()..), and hence

(2.7)

From the analytic properties of f+ and f- it immediately follows that (2.7) is a

Wiener-Hopf factorization of a.
We conclude by showing that in a Wiener-Hopf factorization of a the integer
K is uniquely determined by a, and thus by the result of the previous paragraph
K will always be equal to the winding number of a relative to zero . So consider
the Wiener-Hopf factorization of a in (2.6), and let a()..) = a_ ()")A.JLa+(A.) be a
second Wiener-Hopffactorization. Assume u. > K . Then

(2.8)
16.2 Toeplitz Operators on lp 367

The right hand side of(2.8) extends to a function which is analytic on 1.1..1 < 1 +0+
for some 0+ > 0, and the left hand side of (2.8) extends to a function which
is analytic on 1.1..1 > 1 - L for some L > O. Moreover, the left hand side of
(2 .8) is analytic at infinity and its value at infinity is zero. By Liouville's theorem
(from complex function theory), both sides of (2.8) are identically zero, which is
impossible. Thus JL cannot be strictly larger than K. In a similar way one shows
that K cannot be strictly larger than JL. Thus K = JL, and K is uniquely determined
by « . 0

We are now ready to state the main result of this section.

Theorem 2.2 Let T be the Toeplitz operator on lp, 1 ~ P < 00, with symbol
a E A Then T is left or right invertible if and only ifate'"; =1= Ofor-n ~ t ~ tt .
Assume the latter condition is satisfied and let

(2 .9)

be a Wiener-Hopffactorization ofa. Then

(i) T is left invertible if and only if « ::: 0, and in that case codim Im T = K,

(ii) T is right invertible if and only if« ~ 0, and in that case dim Ker T = -K ,

and in both cases a left or right inverse is given by

(2.10)

where Ta +- I andTa _- I are the Toeplitzoperatorswithsymbols 1ja+(A) and 1ja_(A),

respectively, and

s(n)=!sn , n=0,1,2 , ,
(2.11)
(s( -I)) -n, n = -1 , -2, ,

with S the forward shift and S(-I) the backward shift on lp.

Formula (1.9) in the previous section for the product of two Laurent operators
does not hold for Toeplitz operators. This fact complicates the study of Toeplitz
operators. However, we do have the following intermediate result.

Theorem 2.3 Let Ta and Tf3 be the Toeplitz operators on lp, 1 ~ P < 00, with
symbols a and f3 from A If a extends to a function which is analytic outside the
unit disk including infinity or f3 extends to afunction which is analytic on the unit
disk, then
(2.12)
368 Chapter XVI. Toeplitz and Singular Integral Operators

Proof: Let La and L/3 be the Laurent operators on lp(Z) with symbols ex and f3,
and let P be the projection of lp(Z) defined by (2.4) . By identifying 1m P with
l p we see that
t; = PLallm P, T/3 = PL/3llm P.
Put Q = I - P . Since L a/3 = L aL/3' we have

PL a/3P = PL aL/3P = PLa(P + Q)L/3P

= (PL aP)(PL/3P) + (PLaQ)(QL/3P) ,
Now observe that a extends to a function which is analytic outside the unit disk
including infinity if and only if

n=I,2 , ....

It follows that a has this analytic property if and only if PL a Q = O. Similarly, f3

extends to a function which is analytic on the unit disk if and only if QL /3 P = O.
Thus our hypotheses imply that the operator (PL a Q) (QL /3 P) is equal to zero.
Thus PL a/3P = (PL aP)(PL/3P), which yields (2.12). 0

The order ofthe factors in (2.12) is important, because in general Ta T/3 i= T/3 Tex.
However, as we shall see later (Theorem 2.5 below), for a and f3 in A the difference
Ta T/3 = T/3 Ta is always compact. We are now ready to prove Theorem 2.2.

Proof of Theorem 2.2: We split the proof into two parts. In the first part we prove
the necessity of the condition ex (A) i= 0 for A E T . The second part concerns the
reverse implication and the proof of (2.1 0).
Part 1. Assume T is left or right invertible, and let ex (.1..0) = 0 for some .1..0 E T .
We want to show that these assumptions are contradictory. To do this we use the
same line of reasoning as in the second part of the proof of Theorem 1.1. Since T
is left or right invertible, we know from Section XII .6 that there exists e > 0 such
that f on l p is left or right invertible whenever liT - f II < £ . Choose a positive
integer N such that
L lanl < £/2 , (2.13)
Inl >N
and set
N N
TaN = L anS(n), exN(A) = L anAn . (2.14)
n=- N n=-N

Here s(n) is defined as in (2.11). Then

liT - TaNII = L anS(n) < £/2,

Inl >N
16.2 Toeplitz Operators on f p 369

and

IaN ()"O)! = la(Ao) - aN(Ao)1 = L anAo < e/2.

Inl >N

PutT = Ta N - a N (Ao)1. Then liT-Til < s.und hence T is left or right invertible,
depending on T being left or right invertible. Put

Then

(2.15)

Both WI and £02 are trigonometric polynomials, and hence WI and £02 belong
to A. Let Tw ) and TW2 be the corresponding Toeplitz operators. Notice that T is
the Toeplitz operator with symbol a NO - aN(Ao). Thus (2.15) and Theorem 2.3
yield
- - (-I)
T = Tw ) (S - Ao1), T = (1 - AOS )TW2 ' (2.16)
We have already proved that T is left or right invertible . Thus (2.16) shows that
S - A01 is left invertible or 1 - AOS(-I) is right invertible. Both are impossible
because AO E T. Indeed, the spectra of Sand S(-I) are equal to the closed unit
disc, and according to the stability results of Section XII.6 the operators Ai - S
and Ai - S(-I) are not one-sided invertible for any A E aa(T) = T. We conclude
that for T to be left or right invertible it is necessary that a(A) i= 0 for each A E T.
Part 2. We assume that a does not vanish on T. So, by Theorem 2.1, the
function a admits a Wiener-Hopf factorizat ion. Let this factorization be given by
(2.9). From the properties of the factors a.: and a+ in the right hand side of(2.9)
we may conclude (using Theorem 2.3) that

(2.17)

From Theorem 2.3 and the fact that 1

a::
extends to a function which is analytic
outside the unit disk including infinity it also follows that

Thus T«. is invertible and (Ta_)-I = Ta-::l. Similarly, Ta+ is invertible and
(Ta+)-I = Ta-, . The fact that in the right hand side of(2.17) the factors T«: and
+
Ta + are invertible, allows us to obtain the invertibility properties of T from those
of S(K) .
From (2.11) we see that

S(-K) S(K) =I (K ::: 0), S(K) S(-K ) = 1 (K :'S 0) . (2.18)

370 Chapter XVI. Toeplitz and Singular Integral Operators

Therefore T is left invertible if and only if K ~ 0, and in that case

dimKer T = dimKer SK = K.

This proves item (i) of the theorem; item (ii) is proved in a similar way. Finally,
from (2.17) and (2.18) it follows that the operator T(-l) defined by (2.10) is a left
or right inverse of T.
The following theorem is the second main result of this section.

Theorem 2.4 Let T be the Toeplitz operator on l p, 1 :::: p < 00, with symbol
a E A Then T is Fredholm on l p if and only if its symbol a does not vanish
on the unit circle, and in that case ind T is equal to the negative of the winding
number ofa relative to zero.

For the proof of Theorem 2.4 we need the following addition to Theorem 2.3.

Theorem 2.5 Let Ta and Tf3 be the Toeplitz operators on l p with symbols a and
{3 from A Then the commutator Ta Tf3 - Tf3 Ta is compact.

Proof: Let La and Lf3 be the Laurent operators on lp(Z) with symbols a and {3,
respectively, and let P be the projection of l p(Z) defined by (2.4). Put Q = 1- P.
We have already seen in the proof of Theorem 2.3 that

PL aLf3P = (PL aP)(PLf3P) + (PL aQ)(QLf3P).

By interchanging the order of a and {3 we get

PLf3L aP = (PLf3P)(PL aP) + (PLf3Q)(QL aP).

According to formula (1.9) in the previous section, L aLf3 = Lf3L a, and hence

(PLaP)(PLflP) - (PLflP)(PLaP)
= (PLf3Q)(QL aP) - (PL aQ)(QLf3P). (2.19)

Recall that 1m P may be identified with l p» and then

t; = PLallm P , Tf3 = PLf3llm P .

Thus in order to prove that Ta Tf3 - Tf3 Ta is compact it suffices to show that the
right hand side of (2.19) is compact. In fact, to complete the proof it suffices to
show that for a E A the operators PLa Q and QL a P are compact.
Put aN 0.. ) = L~=-N anA.n, where an is given by (2.2), and let LaN be the
Laurent operator on l p (Z) with symbol aN . Then

liLa - La NII:::: L lanl -+ 0 (N -+ 00) .

jnl >N
16.2 Toeplitz Operators on l p 371

It follows that for N ~ 00 we have

Thus, in order to prove that PL a Q and QLa P are compact, it is sufficient to show
that PLaNQ and QLaNP are finite rank operators. But the latter is a consequence
of the fact that aN is a trigonometric polynomial. Indeed, for each x = (x j ) j e Z
in l p (Z) we have
N- n
(PLaN QX)n = L an+kX-k. n = 0, . .. , N - 1
k=1

and (PLaNQX)n = ootherwise. Hence the rank of PLaN Q is at most N . Similarly,

rank QLaNP S N. 0

Proof of Theorem 2.4: Assume a does not vanish on the unit circle . Then we
can use Theorem 2.2 (i) and (ii) to show that T is Fredholm, and that

ind T = dim Ker T - codim T = -K ,

where K is the integer appearing in (2.9). According to the second part of
Theorem 2.1 this integer K is equal to the winding number of a with respect
to zero. So it remains to show that T is Fredholm implies that a does not vanish
on 11'. This will be proved by contradiction using the same line of arguments as in
the first part of the proof of Theorem 2.2
Assume T is Fredholm and a(Ao) = 0 for some AO E 11'. According to
Theorem 3.6 there exists e > 0 such that T on lp is Fredholm whenever
liT - Til < £. Choose a positive integer N such that (2.13) holds, and define
TaNand a N as in (2.14). Let T be the Toeplitz operator with symbol aN -aN(Ao) .
Then liT - I'll < S, and hence l' is Fredholm. Put

(2.20)

and let Tw be the Toeplitz operator with symbol w. From Theorem 2.3 and (2.20)
we see that l' = Tw(S-A01) , where Sisthe forward shifton z • . Thus Tw(S-Ao1)
is Fredholm. Theorem 2.5 tells us that (S - Ao1)Tw is a compact perturbation of
Tw(S - A01) . According to Theorem 4.1 in the previous chapter this yields that
(S - Ao1)Tw is also Fredholm. Thus both Tw(S - Ao1) and (S - Lo1)Tw are
Fredholm. But then we can use Lemma XV.2.5 to show that S - A01 is Fredholm.
However, by Corollary XY.3.2, the latter is impossible because AO E 11'. So we
reached a contradiction, and hence a(A) =1= 0 for each A E 11' whenever T is
Fredholm. 0
The following theorems show that Theorem 2.4 remains true for Toeplitz oper-
ators with continuous symbols provided we take p = 2. The result is a further
addition to Theorem III.4.1.
372 Chapter XVI. Toeplitz and Singular Integral Operators

Theorem 2.6 Let T be the Toeplitz operator on £2 with continuous symbol a .

Then T is Fredholm on £2 if and only if its symbol a does not vanish on the unit
circle, and in that case ind T is equal to the negative ofthe winding number ofa
relative to zero.
Proof: In view ofTheorem IlIA. 1we only have to show that T is Fredholm implies
that its symbol a does not vanish on T . We shall prove this by following the same
line of arguments as in Part 2 of the proof of Theorem IlIA. 1.
Let T be Fredholm, and assume a(Ao) = 0 for some Ao E T . By Theorem 3.1
from the previous chapter there exists e > 0 such that the operator T on £2 is
Fredholm whenever liT - T II < s. Now use the second Weierstrass approximation
theorem (see Section 1.13) to pick a trigonometric polynomial such that a
. . 1
la(e lf ) - a(elf)1 < '2s , -Jr ~ t ~ n, (2.21 )

and let T be the Toeplitz operator with symbol = a a-

a(Ao) . From (2.21) we
see that la(Ao)1 < ts,
and hence using Theorem 111.2.2 we see that

II T - TIl = max la(e it ) - a(Ao) - a(e it ) I < s.

- rr g :::;:rr

Thus T is Fredholm. Notice that a E A. Hence, Theorem 204 applied to T

and a, shows that a does not vanish on T, which contradicts the fact a(Ao) =
a(Ao) - a(Ao) = O. Therefore T is Fredholm implies a(A) =1= 0 for all A E T . 0

16.3 An Illustrative Example

In this section we use the theory developed in the previous section to solve in £p
the following infinite system of equations:

L
00

eip ij - k lxk - Axk = qj (j = 0,1,2, . ..). (3.1)

k=O
Here A is a spectral parameter, q and p are complex numbers with Iq I < 1 and
:sp > O.
The above equations appear in a natural way in the study of propagation of
electromagnetic waves in a medium of periodic structure; see [Kre], § 13 and the
references given there . In this particular case 1- A is purely imaginary and q = eiP .
In the sequel we set c = e iP . We have c =1= 0 and [c] < 1 because :sp > O. Let
T be the Toeplitz operator on £p given by

1 C C2 • •• ]
c c ...
T= c2cI . (3.2)
[ · ..
· · .
16.3 An Illustrative Examp le 373

Since lei < 1, the symbol a of T is given by

{ E T.

Notice that a is analytic in an annulus around the unit circle T, and hence the
operator T is well-defined and bounded on i p (l ::::; p < 00) .
Using the operator T the equation (3.1) can be rewritten as

T x - AX = y , (3.3)

where y = (l ,q ,q2, . . .) and x = (XO,XI,X2, . . .) . To solve (3. 1) or (3.3) we

distinguish between A = 0 and A =1= o.

Theorem 3.1 Let A = O. Then equation (3.1) has a unique solution in i p (l ::::;
p < 00), namely,
e- i p _ q . . 1
Xo = e - ip
Xj = xo(q - e'P)q J-, j = 1,2, .... (3.4)
- eip '
Proof: First let us prove that the operator T is invertible . To do this, notice that
its symbol a can be rewritten as
e e- 1
a(n = -{ ---e - -{ ---e--""""I { E T.

It follows that a admits the factorizat ion

{ E T, (3.5)

with the factors being given by

a-en = (e - e- 1) _{_ = (e - e -1)(l - eC 1) - I, (3.6)

{ -e
1 - I
a+(n = 1 = - e(l - en . (3.7)
{ - e-
Recall that e is inside the unit disc. Thus a+O±1 extends to a function which
is analytic on the unit disc, and a_ (.)±1 extends to a function which is analytic
outside the unit disc including infinity. Thus (3.5) is a Wiener-Hopffactorization
with K = O. But then we can apply Theorem 2.2 to show that T is invertible .
Moreover,
T - 1 = T - I T _ I. (3.8)
Q'+ Q'-

Now, let us use (3.8) to comput e the solution of (3.1). From (3.6) and (3.7) we
see that
1 I 1
TQ'-I = - I (I - eS), T - I = - - (1- eS) ,
- e-e Q'+ e
374 Chapter XVI. Toeplitz and Singular Integral Operators

where S' and S are the backward shift and the forward shift on i p» respectively. It
follows that
T- I = _1_ (I - cS -
l-c 2
es
+ c2SS'),
and hence the matrix ofT- 1 with respect to the standard basis of i p is tridiagonal:

-c 0 0
-c 1 +c2 -c 0
T- I = _1_ 0 -c 1 + c2 -c
1 - c2
0 0 -c 1 + c2

Put Y = (l , q, q2, . ..), and let x = (xo, XI, X2, ...) be T- I y . Then

1 c- I - q
Xo = --2(l-cq) = -I '
l-c c -c
and for k ~ 1 we have

_ (q -c)(l-cq) j_1 _ c- I -q( _ ) j _ 1

- 1 2 q - -I q C q .
-c c -c

Since c = e'", this yields (3.4). o

Now fix A i= O. Then the symbol co of the Toeplitz operator T - AI is given by

(c-c-I)s
w(S) = a(S) - A = (s _ c)(s _ c- I ) - A

AS 2 - A(C + c-I)s - (c - c- I)s +1

(s - c)(s - c- I )

AS 2 - 2(ACOSp + i sinp)S +A
sET.
(s - c)(s - C l)

Here we used the convention that for c = e ip we have

c + C -I = 2 cos p, c- C -I = 2i sin p . (3.9)

16.3 An Illustrative Example 375

In the above expression for w({) the numerator is a quadratic polynomial and
the product of its two roots ZI and Z2 is equal to one. SO Z2 = Z- I, and we can
assume that IZII s I. Thus for ), .:p 0 the symbol co of T - ).../ is of the form

w({) = _ )... (~ - zJ>(~ - ZJI),

(3.10)
~ E T,
(~ - c)(~ - c- I )

where 0 < Izii s 1.

We claim that for x .:p 0 the operator T - ).../ is invertible if and only if IZ II < 1.
Indeed, ifT -).../ is invertible, then ca (e i I) .:p 0 for each t by Theorem 2.2, and hence
we see from (3.10) that IZII = 1 is impossible. To prove the converse implication,
assume that IZII < 1. The function w in (3.10) admits the factorization

~ E T, (3.11)

with the factors being given by

r 1 r - I
w_({) = -)...~ = -)... - Zh , (3.12)
~ - c 1 - c~-I

~ - Z-I -I 1 - ZI~
w+({) = ~ - c ~I = ZI c 1
-c
~ (3.13)

Recall that Z I and c are in the open unit disc. Thus w+O±1 extends to function
which is analytic on the unit disc, and co.: O±I extends to a function which is
analytic outside the unit disc including infinity. Thus (3.11) is a Wiener-Hopf
factorization with K = 0, and we can apply Theorem 2.2 to show that T - ).../ is
invertible . Moreover,
(T - ).../)-1 = T - I T - I , (3.14)
w+ w_

where Tw+ -I and T - I are the Toeplitzoperators with symbols Ijw+()...) and Ijw_()...),
w_
respectively.

Theorem 3.2 Let ): .:p 0, and assume IZ II < 1. Then the equation (3.1) has a
unique solution in l p (1 S p < (0), and for q = e i p this solution is given by
2i sin P j+1
x·- Z (j=0,1,2, . .. ). (3.15)
] - )...(1 - eipzJ> I

Proof: It remains to prove (3.15). Writey = (1, q, q2, ), whereq = eiP. Then,
according to (3.14), the unique solution x = (xo, XI, X2 , ) in lp (1 S p < (0)
of(3.1) is given by
X=Tw+ -IT-lY·
w_ (3.16)
Recall that Tw=1 is an upper triangular Toeplitz operator. Hence , using (3.12), we

(1)-1 1-
have
-I cq
Tw- IY = w_ - Y = -)... y. (3.17)
- q l-zlQ
376 Chapter XVI Toeplitz and Singular Integral Operators

Next, since T - I is lower triangular,

w+

(3.18)

where ao, aI, a2, .. . are the Taylor coefficients at zero of the function w+({)-I x
(1- q{)-I .
Now, take q = eip (= c) . Then we see from (3.13) that

w+({)(1 - q{)-I = zle- l (1 - ZI{)-I,

and hence aj = c- I z{+I. By combining this with (3.16H3.18) we obtain that

-I 1 - c2 -I j+1
1 -1
C- c )+1
Xj = -A c zl = - ZI' j = 0,1,2, . . . .
I-zlc AI-ZIC

Together with the second identity in (3.9) this yields (3.15) . o

Theorem 3.3 Let c = eip with ';Sp > O. Thenfor each 1 S p < 00, the spectrum
ofthe operator Tin (3.2) is given by

a(T) = { -isinp 1-1 S T S I}. (3.19)

T - cosp

More precisely, a (T) is an arc on a circle, the endpoints ofthe arc are
i sinp -i sinp
A-I = , A+I = ,
1 + cosp 1 - cosp

and it passes through the point

i sinp
AO= - - .
cosp

Proof: From the proof of Theorem 3.1 we know that T is invertible . Thus A E
a(T) if and only if A =1= 0 and the roots ZI = ZI(A) and Z2 = ZI (A)-I of the
quadratic polynomial

are both on the unit circle . For A =1= 0 put

I
_ ZI(A)+ZI(A)-I _ 1 (
T(A) - - - c+c
2
-I
2
+ c-c-
A
)
.
16.4 Applications to Pair Operators 377

Notice that a nonzero complex number z lies on the unit circle if and only if
(z + z- I )/ 2 E [-1 , 1]. We conclude that A E a (T ) if and only if A i= 0 and
r (A) E [-1 , 1], and in that case
C _c- 1
A - - - - - - --:-
- 2r (A) - c - c- I .

Here we used that [c] < 1, and hence (c + c- I )/ 2 ¢ [-1 , 1].

Next, notice that r(A) runs through [-1, 1] when A runs through a (T). Indeed,
if r E [-1 , 1], then Zr i= c + c- I , and
c -c- 1
A :=
Zr - c - c-
1 i= O.
For this A we have r (A) = r E [-1 , 1], and hence A E a (T). Finally, using (3.9)
we see that (3.19) is proved.
From what we have proved so far we see that a (T) is the image of the interval
[- 1, 1] under the map
c- c- I
r f-+ AT = 1
.
2r-c- c-
Such a linear fractional map transforms an interval into an arc of a circle (or an
interval). By taking r = - 1 and r = 1 we see that [- 1, 1] is transformed into the
arc which begins at the point A-I and ends at AI , and which passes through AO.
These three points uniquely determine the circle on which this arc lies. 0

16.4 Applications to Pair Operators

Throughoutthis section (except in some ofthe remarks) we assume that a and,8 are
functions from the class A , that is, a and,8 are analyti c in an annulus containing the
unit circle T. With these two functions we associate an operator M a.fJ on £p(71),
where 1 ::'S p ::'S 00 is fixed. The definition is as follows. Given x = (X j)j EZ the
element y = M a.fJx is the vector y = ( y j) j EZ with
00 - I

Yn = L a n- kXk + L bn -kXk, (4.1)

k=O k=-oo

i
where for each n
1C

an = - 1 a(e't)
•
e- znt
.
dt, (4.2)
2rr - 1C

We shall refer to Ma . fJ as the pair operator defined by a and ,8.

Notice that for a = ,8 the operator Ma •fJ is just the Laurent operator with
symbol a . More generally, if L a and L fJ denote the Laurent operators on f p(Z)
with symbols a and,8 , respectively, then

M a.fJ = L aP + L fJQ, (4.3)

378 Chapter XVI. Toeplitz and Singular Integral Operators

where P is the orthogonal projection of lp(Z) onto the subspace consisting of

all vectors (Xj)jez such that Xj = 0 for j < 0, and Q = 1- P . From (4.3) it
immediately follows that M ex ,f3 is a bounded operator on lp(Z). 0

Theorem 4.1 Let ex, f3 be in A . The operator M ex ,f3 on lp(Z) is Fredholm if and
only if the functions ex and f3 do not vanish on T, and in that case ind M ex ,f3 is the
negative ofthe winding number of y = ex / f3 relative to zero .

Before we prove the theorem it will be convenient first to analyze the operator
M ex ,f3 a bit better. Let P and Q be the complementary projections appearing in
(4.3). Then lp(Z) = 1m Q 61 1m P, and relative to this decomposition we can
write M ex ,f3 as a 2 x 2 operator matrix,

QLf3llm Q QLexllm P ]
Mex ,f3 = : 1m Q 61 1m P ~ 1m Q 611m P. (4.4)
[ PLf3llm Q PLexllm P

Here each element of the 2 x 2 matrix is viewed as an operator between the

corresponding spaces. For instance, QLex 11m P acts from 1m P into Im Q. Recall
that 1m P consists of all vector (x j) j e Z in l p (Z) such that x j = 0 for j < 0, and
hence Im P can be identified with lp . Using this identification we have

(4.5)

where Tex is the Toeplitz operator on lp with symbol ex. In Section 2 (see
the proof of Theorem 2.5) we have shown that the operator QLex P is compact,
and hence Q Lex 11m P is a compact operator from Im P into 1m Q. Similarly,
the operator PLf3llm Q from Im Q into Im P is compact. It remains to analyze
QLf3llm Q. Let J be the operator from c, into Im Q that transforms the vector
(Xj )}':o in l p into the vector (Yj) jeZ, where

. _ {X_j-I, j = -1, -2, -3 , . . . ,

(4,6)
Y] - 0, otherwise,

Then J : l p ~ Im Q is one-one and onto, and J is norm preserving, A straight-

forward calculation shows that

(4.7)

where Tf3# is the Toeplitz operator on l p with symbol f3# given by

A E T. (4.8)
16.4 Applications to Pair Operators 379

Indeed, let b., be the coefficient of Av in the Laurent series expansion of f3 . Then
for x = (Xj)i=o inl p and Ix = Y = (Yj)jE7l., where Yj is given by (4.6), we have
for n :::: 0

(J-I(QLpIIm Q)Jx)n = (J-IQLpY)n = (QLpY)-n-1

00 -I

= L b-n-I-kYk = L b-n-I-kX-k-1
k=-oo k=-oo
00 00

= L b-n-£x£ = L b~_£x£ ,
£=0 £=0

where be is the coefficient of AV in the Laurent series expansion of f3# . Notice that
f3# is also analytic in an annulus containing the unit circle, that is, f3# E A. We
are now ready to prove Theorem 4.1
Proof of Theorem 4.1: Consider the representation (4.4) . Since the operators
QLalIm P andPLplIm Q are compact, Ma,p and

QL pIIm Q 0 ] (4.9)
[ o PLalIm P

differ by a compact operator. Thus, by Theorem 3.1 in the previous chapter, the
operator Ma,p is Fredholm if and only if the operator (4.9) is Fredholm. But the
latter happens if and only if PLalIm P and QLpIm Q are Fredholm. Now use
(4.5) and (4.7). It follows that Ma.p is Fredholm if and only if Ta and Tp# are
Fredholm. Now apply Theorem 2.4 to Ta and Tp# . Notice that f3# does not vanish
on T if and only if the same holds true for f3 . We conclude that Ma,p is Fredholm
if and only if ex and f3 do not vanish on T.
Now, assume ex and f3 do not vanish on T, and put y = ex / f3 . By applying
Theorem 1.1 to L p we see that L p is invertible, and thus, using formula (9) in
Section 1, we have

Ma,p = Lp(LyP + Q) = Lp [ lIm0 Q QL y 11m

PLy 11m P .
P]
Again the operator QL y 11m P is compact, and, by Theorem 2.4, the operator
PLy 11m P = Ty is Fredholm and its index is equal to -K , where K is the winding
number of y relative to zero . By Theorem 3.1 in the previous chapter,

ind Ma,p = ind lIm Q 0 ]-'d _

[ 0 PLy 11m P - ill Ty - -K ,

which completes the proof. o

For the case when p = 2 Theorem 4.1 above is also true if ex and f3 are just
continuous on the unit circle . To see this recall (see Section IlL 1) that for ex and f3
380 Chapter XVI. Toeplitz and Singular Integral Operators

continuous on T the Laurent operators La and L{3 are well-defined on i2Gl;) . In

fact, in the terminology of Section III.l the operators La and L {3 are the Laurent
operators defined by the functions a(t) = a(e it) and b(t) = f3(eit), respectively.
Also, the operators QL a P and PL {3 Q are compact. For instance, to prove that
QL a P is compact, we choose a sequence aI, a2, . . . of trigonometric polynomials
which converges uniformly to a on BT. Such a sequence exists because of the
second Weierstrass approximation theorem (see Section 1.13). Now let Lan be the
Laurent operator defined by an(t) = an(e it ) . Then QL an P is an operator of finite
rank (see the last paragraph of the proof of Theorem 2.5), and

IIQL aP - QLanPIl S liLa - Lan II = liLa-an II

it
max la(e ) - an(eit)1 --+ 0 (n --+ (0).
-7r~t ~7r

Here we used formula (1.3) of Section III.1. Thus QL a P is the limit in the operator
norm of a sequence of finite rank operators, and therefore QL a P is compact. In
a similar way one shows that PL{3 Q is compact. We can now repeat in precisely
the same way all the arguments in the proof of Theorem 4.1 with the exception
of three. The first two concern the two references to Theorem 2.4 which have to
be replaced by references to Theorem 2.6 . The third concerns the reference to
Theorem 1.1 which has to be replaced by a reference to Theorem III .l .2. With
these changes the proof of Theorem 4.1 also works for the case when a and f3 are
continuous on T provided p = 2.

Theorem 4.2 Let a, f3 be in A. The operator M a,{3 is left or right invertible if

and only if a and f3 do not vanish on the unit circle. Assume the latter condition
holds, and let K be the winding number of y = a / f3 relative to zero. Then

(i) M a,{3 is left invertible if and only if K > 0, and in that case
codim 1m Ma ,{3 = K ,
(ii) M a,{3 is right invertible if and only if K < 0 and in that case
dim Ker M a ,{3 =-K .

Furthermore, ify admits a Wiener-Hopffactorization

A E T, (4.10)

then a left or right inverse ofM a,{3 is given by

M(-{3I) = (L
~ ~
- 1 P + L y_ O)(V- K
P + Q)L ~- I L{3-I. (4.11)

Here L w denotes the Laurent operator on i p (Z) with symbol co, and V is the
bilateralforward shift on i p (Z).

Proof: We split the proof in three parts. The first part contains some general facts
on the function space A.
16.4 Applications to Pair Operators 381

Part 1. Consider a E A. Let an be the coefficient of )... n in the Laurent

series expansion of a . We know that there exist constants c ~ 0 and 0 S p < 1
such that
Ian I S cp 1nl (n E Z).
Take p < r < 1, and let a r()...) = L~_ooanrn)...n . Then a; also belongs to A,
because
lanrnl S cplnlrn S cp1n l (n E Z),
where p = p/ r < 1. Next, let La and L ar be the Laurent operators on ep (Z) with
symbols a and a-, respectively. Then

liLa - L ar II ---+ 0 (p < r t 1) (4.12)

Indeed,
00 00

liLa - Larll < L

n=-oo
Ian - anrnl = c
n=-oo
L p 1n1ll _r nl

s c f
n=O
(p" - pnr n) + c f ((~)
n=l
n _ pn)

= c(_I 1_) +c(_p- - _p-) ---+ 0 (p < r t 1).

1- p 1 - pr r- p 1- P
Part 2. In this part we assume that Ma,fJ is left or right invertible. We want to
show that a and fJ do not vanish on 11'. Since Ma,fJ is left or right invertible, there
exists e > 0 such that 11M - Ma,fJ II < e implies that M is left or right invertible
and

dimKer M = dimKer Ma,fJ' codim 1m M = codim 1m Ma,fJ . (4.13)

The left or right invertibility of Ma,fJ also implies that the functions a and fJ are
not identically equal to zero. This together with the fact that a and fJ are analytic
in an annulus containing 11' shows that the zeros of a and fJ are isolated. So we
can find 0 < r' < 1 such that for 0 < r' < 1 the functions a; and fJr are in A and
do not vanish on 11'. By definition,

Ma,fJ =LaP+LfJQ, Mar ,fJr =LarP+LfJrQ ,

and hence (4.12) yields

(r' < r t 1).

But then we can use (4.13) to show that for r' < r < 1 and 1 - r sufficiently small
we have

dim Ker Mar,fJr = dim Ker Ma,fJ , codim 1m Mar,fJr = codim 1m Ma,fJ '
(4.14)
382 Chapter XVI. Toeplitz and Singular Integral Operators

However, since a; and f3r do not vanish on T for r' < r < 1, the operator Mar,fJr
is Fredholm (by Theorem 4.1), and hence the identities in (4.14) show that Ma ,fJ
is Fredholm too. But then, again using Theorem 4.1, we can conclude that a and
f3 do not vanish on T .
Part 3. In this part we assume that a and f3 do not vanish on T, and we prove
the statements (i) and (ii) and the inversion formula (4.11). Our assumption on a
and f3 imply that
(4.15)
where y = a / f3. The function y does not vanish on T, and hence it admits
a Wiener-Hopf factorization as in (4.10), where K is the winding number of y
relative to zero. Let L y _ and L y + be the Laurent operators with symbols y_ and
y+. The fact that y_ and y+ do not vanish on T yields

L y- 1 = L - I, (4.16)
- y-

and from the analyticity properties of y_ and y+ we see that

L±l P = PL±1 P. (4.17)

y+ y+

Using the factorization (4.10) we see that (4.15) can be rewritten as

where Y is the bilateral forward shift on lp(Z) . Hence, by (4.16) and (4.17),

(4.18)

Moreover, all factors in the right hand side of(4.18) are invertible with the possible
exception of yK P + Q. Now, notice that yk P + Q is left invertible if K ~ 0 and
right invertible if K ~ 0, and in both cases a left or right inverse of yK P + Q is
given by
(yK P + Q)(-l) = y-K P + Q. (4.19)
also
codim 1m (yK P+ Q) = codim 1m T )..K = K (k ~ 0), (4.20)
dimKer (yK P + Q) = dimKer T ).-K = -K (k ~ 0). (4.21)
From (4.18)-(4.21) and the invertibility of all factors in the right hand side of
(4.18) different from yK P + Q we see that statements (i) and (ii) hold true, and a
left or right inverse of Ma,fJ is given by

The latter formula together with (4.16) and (4.17) yield (4.11), which completes
~~~ 0
I6.4 Applications to Pair Operators 383

We conclude this section with some further information about the pair operator
M a ,f3 for the case when this operator acts on e2(Z) and the functions a and f3 are
not required to belong to the class A but are merely continuous. In that case the
first part of Theorem 4.2 stilI holds true, that is, the following theorem holds .

Theorem 4.3 Let M a ,f3 be the pair operator on e2(Z), with a and f3 being con-
tinuous on the unit circle. Then M a ,f3 is left or right invertible if and only if a and
f3 do not vanish on T. Assume the latt er condition holds, and let K be the winding
number of y = a / f3 relative to zero . Then

(i) M a ,f3 is left invertible if and only if K > 0, and in that case
codim 1m M a ,f3 = K,
(ii) M a ,f3 is right invertible if and only if K < 0, and in that case
dim Ker M a ,f3 = -K .
In particular, M a ,f3 is two-sided invertible if and only if « = O.

Proof: We split the proof into two parts . In the first part we prove the necessity
of the condition on a and f3. The secon d part concerns the reverse implication and
statements (i) and (ii).
Part 1. Let M a ,f3 have a left or right inverse, and assume a (>"0) = 0 for some
>"0 E T . By the results of Section XII .6 there exists e > 0 such that the operator
M on e2(Z) is left or right invertible whenever 1 M- Mil < e. Now, pick a
a
trigonometric polynomial such that la (e") -a (e") I < £/ 4 for each - Jr .::: t .::: Jr .
a a- a
Put = a
(>"0). Then (>"0) = 0 and
. . I
la(e't) - a (e/t)1 < 2£' -Jr'::: t .::: tt . (4.22)

Next , choose a trigonometric polynomial ~ such that

I
1f3 (e/t ) f3 (e/t ) I < 2£'
A' •

- - Jr '::: t .::: Jr. (4.23)

Then, by formu la (1.3) in Section III.l , we have

II M a ,p- M a ,f3 11 = II (L a - La) P + (L p - L f3)QII

.::: li L a - La II + II L p - Lf311 < e.

Thus M a. , f3' is left or right invertible. Both a and ~ belong to class A . Thus
Theorem 4.2 implies that a does not vanish on T which contradicts the fact that
a (>..o) = O. Therefore, M a ,f3 is left or right invertible implies that a does not
vanish on T. The analogous result for f3 is proved in the same way.
Part 2. In this part we assume that a and f3 do not vanish on T. Let K be the
winding number of y = a / f3 relat ive to zero. Notice that

M a ,f3 = L f3 (PL y P + Q)(l + QL y P ). (4.24)

384 Chapter XVI. Toeplitz and Singular Integral Operators

The operator L f3 is invertible by Theorem III.1.2. Since (QL y p)2 is the zero
operator, we have that I + QL y P is invertible too. Next, recall (see formula
(5» that we can identify the operator PLy 11m P with the Toeplitz operator Ty.
But then we can use Theorem I1IA.1 to show that PLy P + Q is left or right
invertible. The invertibility of the factors L f3 and I + QL y P in (4.24) now yields
that Ma ,f3 is also left or right invertible. Furthermore, by applying (i) and (ii) in
Theorem IlIA. 1 to T = Ty , formula (4.24) also proves the statements (i) and (ii)
of the present theorem. 0

We conclude this section with a remark about operators that can be considered
as conjugates of pair operators . Fix 1 ~ P ~ 00, and let a and fJ belong to the
class A. Given x = (XdkEZ in lp(Z) we define

"L r: _ ooa j - kXk. j = 0,1,2, ... ,

(Ka f3 x )j =
, ! "Lr:_oobj-kXk. j = -1, -2, ....

Here an and bn are given by (4.2). It is straightforward to check that Ka ,f3x E

lp(Z), and that K a ,f3 is a bounded linear operator on lp(Z). In fact,

(4.25)

where La and Lf3 are the Laurent operators with symbols a and fJ, the operator P

(Xj)jEZ with Xj = °
is the orthogonal projection of l p (Z) onto the subspace consisting of all vectors
for j < 0, and Q = 1- P. We shall refer to K a ,f3 as the
associate pair operator defined by a and fJ.
From the representations (4.25) and (4.3) it follows that the conjugate of Ka ,f3
is the operator

where a#(A.) = a (A. -I) and fJ#(A.) = fJ(A. -I). This connection allows us to
conclude that with appropriate modifications, Theorems 4.1~.3 remain true ifthe
pair operator Ma,f3 is replaced by the associate pair operator K a,f3 ' For instance,
K a ,f3 is Fredholm if and only if the functions a and fJ do not vanish . Moreover,
in that case ind K a,f3 = -ind M a#,f3#' and hence ind K a,f3 is again equal to the
negative of the winding number of a] fJ relative to zero. In a similar way, one
derives the invertibility properties (one- or two-sided) of Ka ,f3 '

16.5 The Finite Section Method Revisited

In this section we study the convergence ofthe finite section method ofthe Laurent
operators, Toeplitz operators and pair operators considered in the previous sections .

°
We begin with the Toeplitz operators on lp. Throughout, 1 ~ P < 00.
For each n 2: let Pn be the projection of lp defined by

Pn (xo, XI , X2 , • .•) = (xo, ... ,Xn , 0, 0, ...). (5.1)

16.5 The Finite Section Method Revisited 385

Notice that II Pn II = 1, and P;» ---+ x (n ---+ 00) for each x E f.p • Here we use that
p < 00. Indeed, for x = (xo, XI, X2, . . .) we have

(n ---+ 0).

Now let T be an invertible operator on f.p . We say that the finite section method
converges for T if for n sufficiently large, n ~ no say, the operator Pn T Pn on
1m Pn is invertible and for each y E f.p the vector

x(n) = (xo(n), . . . , Xn (n) , 0, 0, . . .) = (P n T Pn)-I PnY (n ~ no)

converges to a solution x of T x = y . In other words the finite section method

converges for T if and only if T E TI {P n } .
Recall (see Section 1) that A stands for the class ofall complex-valued functions
that are analytic on some annulus containing the unit circle .

Theorem 5.1 For an invertible Toeplitz operator on f.p, 1 ~ p < 00, with symbol
from the class A, the finite section method converges.

Proof: Let a E A be the symbol of T. Since T is invertible, Theorem 2.4 shows

that a does not vanish on the unit circle, and the winding number K of a relative to
zero is equal to zero. Thus, by Theorem 2.1, the function a admits a factorization
a = a_a+, where a : and a+ belong to A and do not vanish on 11', and conditions
(i) and (ii) in the definition of a Wiener-Hopf factorization (see the paragraph after
Theorem 2.4) are satisfied. From Theorem 2.2 it then also follows that

with
(Ta_)-I = Ta=l, (Ta+)-I = Ta:;: 1.

Put F = Ta+ Ta_. Then F is invertible, and by Theorem 2.5 the operator T and F
differ by a compact operator. Therefore, by Theorem XIII.3.5, it suffices to prove

a=
that the finite section method converges for F .
From the analyticity properties of the functions a~ I and I it follows that for
each n E Z we have

PnTa+ = PnTa+Pn , Ta_Pn = PnTa_Pn , (5.2)

PnTa~1 = PnTa~1 r; t;__1 r, = r.r;__1 r; (5.3)

It is now straightforward to check that P;Ta + P; and Pn T«. P; are invertible as

operators on Im Pn and

(PnTa+Pn)-lpn = PnTa~l, (PnTa _Pn)-lpn = t;__IPn (5.4)

386 Chapter XVI. Toeplitz and Singular Int egral Operators

for each n E Z. Indeed , since P; = Pn , the first identities in (5.2) and (5.3) yield

ir, Ta+Pn )(PnTa~1 Pn ) = PnTa+PnTa~1 r; = PnTa+Ta~1 r; = r.,

and

ir; Ta~l Pn )( r; Ta+Pn ) = r; Ta~l r, Ta+r: = r; Ta~l Ta+r, = r;

Therefore Pn Ta+Pn is invertible on Im Pn and

( PnTa+pn) -1 r, = PnTa~1 PnPn = PnTa~l ,

which proves the first identity in (5.4). The invertibility of Pn Ta_ Pn on Im Pn and
the second identity in (5.4) are proved in the same way.
Notice that

PnFP n = PnTa+Ta_Pn = (P"Ta +Pn)(PnTa_P,,) , (5.5)

and hence P"FPn is invertible on Im Pn for each n . Now take y E f p . Then , by

(5.5) and (5.4),

( PnFPn)-1 Pny = ( PnTa_ Pn)- I( PnTa+Pn)-1 Pny

= ( PnTa_ Pn)- l PnTa~1 Y = t;__1Pn Ta~1 y.

But r, Ta~l y ---+ Ta~l Y if n ---+ 00. Thu s

lim ( PnFPn)-1 P" y lim Ta- 1PnTa-1y = T; ITa- 1y =

= n->oo F - 1y ,
n->oo - + - +

and hence the finite section method for F converges . o

We proceed with an analysis of the finite section method for Laurent operators
and pair operators on f p (Z), where 1 ~ p < 00 is fixed. For each n let Qn be the
projection of f p(Z) defined by

Qn« Xj )j EZ) = (. . . , 0, 0, X- n , . .. , Xn , 0, 0, . . .). (5.6)

The projections Qn have all norm one, and Qnx ---+ x (n ---+ (0) for each x E f p
(because 1 ~ p < (0). We say that the finite section method converges for
an invertible operator L on fp(Z) if for n sufficiently large, n ~ no say, the
operator QnLQn on Im Q" is invertible and for each y E fp(Z) the vector x(n) =
( QnLQn)- l QnY , n ~ no, converge s to a solution x of L x = y. In other word s,
the finite section method converges for L if and only if LEn {Q n}. We shall
prove the following theorem.

Theorem 5.2 Let L a be an invertible Laurent operator on fp(Z) , 1 ~ p < 00,

with symbol a E A (i.e., a does not vanish on 'lI'). In order that the fi nite section
method converge f or L a it is necessa ry and suffic ient that the winding number of
a with resp ect to zero is equal to zero.
16.5 The Finite Section Method Revisited 387

Proof: The fact that a does not vanish on the circle follows from the invertibility
of La (use Theorem 1.1). Let P; be the projection of i p defined by (5.1), and
define In to be the map from Im Qn to Im P2n+l given by

(5 .7)

From the definitions of Qn and P; it follows that I is a linear operator which maps
Im Qn in a one to one way onto Im P2n+l. Moreover, IIInxll = IIxll for each
x E Im Qn. A straightforward calculation shows that

where Ta is the Toeplitz operator with symbol a . Since I is one to one and onto , the
operator QnLa Qn is invertible on Im Qn ifand only if P2n+l Ta P2n+l is invertible
on Im P2n+l, and in that case

(5.9)

because In is norm preserving.

We shall also need the map I: : Im Qn --+ Irn P2n+l defined by

I:«X j)jEZ) = (x n , Xn-l . . . x.:« , 0, 0, . . .). (5.10)

Notice that the right hand side of(5.10) may be obtained from the right hand side
of (5.7) by reversing the order of the first 2n + 1 entries. Again I: is a linear
operator which maps Im Qn in a one to one way onto Im P2n+ 1, and I: is norm
preserving. Moreover,

I:(QnLaQn) = (P2n+lTa#P2n+l)I: , n - 0,1 ,2, ... , (5.11)

where Tcx# is the Toeplitz operator on i p with symbol a # given by

(5 .12)

As mentioned in the previous section (see the paragraph before the proof of
Theorem 4.1), the function a# E A. From (5.11) it follows that QnLaQn is
invertible on Im Qn if and only if the same holds true for P2n+l Ta#P2n+l on
Im P2n+l, and in that case

(5.13)

Now, assume that the finite section method converges for La . Then, by Theorem
XI!. 7.1, the operator QnLa Qn is invertible on Im Q for n sufficiently large, n 2: N
say, and sUPn ::':N II(QnLaQn)-11l < 00 . Using (5.9) and (5.13) it follows that

sup II (P2n+lTa P2n+l)-1 II < 00, sup II (P2n+l Ta#P2n+j)- 11l < 00. (5.14)
n::':N n::,:N
388 Chapter XVI Toeplitz and Singular Integral Operators

By the remark made at the end of Section 11.17 (which carries over to a Banach
space setting) it follows that both Tcx and Ta# are one to one. Let K be the winding
number of a with respect to zero. Since Ta is one to one, Theorem 2.2 implies
that K 2: O. Notice that a# does not vanish on T and its winding number relative
to zero equals to -K. Then Theorem 2.2 applied to Ta # yields -K 2: O. Hence
K = 0 as desired.
Finally, let us assume that the winding number of a relative to zero is equal
to zero. Then Ta is invertible (by Theorem 2.2), and according to Theorem 5.1
we have Ta E TI{Pn }. But then we can apply Theorem XII.7.1 to show that the
first inequality in (5.14) holds for some positive integer N , and hence (5.9) yields
sUPn>N II QnLa Qn)-lll < 00. So, using Theorern Xll.Z.l again , we can conclude
that La E TI{Qn}. 0

We conclude this section with an analysis ofthe convergence ofthe finite section
method for the pair operator M a ,/3 introduced in the previous section . Here a and
f3 belong to A. Recall that for an invertible operator L on l p (Z) the convergence
of the finite section method is defined in the paragraph preceding Theorem 5.2.

Theorem 5.3 Assume the operator Ma ,/3 on l p(Z), I ::s p < 00, is invertible
(i.e., a and f3 do not vanish on the unit circle). In order that the finite section
method convergefor Ma ,/3 it is necessary and sufficient that the winding numbers
ofa and f3 relative to zero are equal to zero.

Proof: Since M a ,/3 is invertible, we know from Theorem 4.2 that a and f3 do not
vanish on T, and the winding number of a relative to zero is equal to the winding
number of f3 relative to zero. We denote this winding number by K . It remains to
prove the statement about the convergence of the finite section method. We split
the proof in two parts.
Part 1. In this part we assume that the finite section method converges for M a ,/3 .
We have to show that the number K introduced in the previous paragraph is equal
to zero. Assume K > O. Let P be the projection of lp(Z) defined by

x } , j 2: K,
y} = [ 0, otherwise.

Put Q = 1- P . We claim that the operator HaJJ = QLa P + PLfJ Q is compact.

Here P and Q are the complementary projections appearing in (3) of the previous
section. To see this, write

Ha,fJ = QLaP + (P - P)LaP +PLfJQ + (P - P)LfJQ.

The fact that a and f3 belong to A implies that QL a P and PLfJ Q are compact (as
we have seen in the proof of Theorem 2.5) . Since P - P is an operator of finite
rank, the operators (P - p)L a P and (P - P)LfJ Q are also compact. Hence Ha,fJ
is compact.
16.5 The Finite Section Method Revisited 389

Next, we show that the operator Ma,fJ = Ma,fJ - Ha,fJ is invertible . Notice that
= PLaP + QLfJQ. Thus it suffices to show thatPLaP from Im P to Im P
Ma ,fJ
and QL fJ Q from Im Q to 1m Q are invertible . To see that PLa P is invertible from
Im P to Im P, let J : Im P -+ lp and j : Im P -+ lp be defined by

Jx = (xo, XI, X2, . . .),

for each X = (x j) j e ll in Im P and lm P, respectively. Then

(5.15)

Notice that the function <li()') = ). -K ex does not vanish on 11' and has winding
number zero relative to zero. By Theorem 2.2 this implies that the Toeplitz operator
T).,-Ka is invertible . Next, observe that both J and j are invertible . Hence (5.15)
shows that PLa P is an invertible operator from Irn P onto Im P. In a similar way
one shows that up to invertible factors the operator QL fJ Q from Im Q to Im Q is
equal to the Toeplitz operator T).,KfJ#' where f3#().) = 13(1/).). We know that x" 13#
does not vanish on 11' and its winding number relative to the origin is equal to zero.
So T).,K fJ# is invertible, and hence QL fJ Q : Im Q -+ Im Q is invertible.
Recall that the finite section converges for Ma,fJ ' By Theorem Xm.3.5, the
compactness of Ha,fJ and the invertibility of Ma,fJ = Ma,fJ - Ha,fJ imply that
the finite section method converges for Ma,fJ ' Thus for n sufficiently large, the
operator QnMa ,fJ Qn on Im Qn is invertible . We shall show that this is impossible.
Indeed, let e, = (Oij)jEZ, where oij is the Kronecker delta, and take n > K. Notice
that the vectors e_ n, e- n+ I, ... , en-I , en form a basis ofIm Q, and the matrix of
the operator QnMa,fJQn on Im Qn relative to this basis has the form:

bo b-n+1 0 0

bn-I bo 0 0
bn bl 0 0
(5.16)
bn+K-I bK 0 0
0 0 aK aK-n

0 0 an ao
Here a v and bv are the coefficients of ).v in the Laurent series expansions of ex
and 13, respectively. The matrix (5.16) is a square matrix of order 2n + 1, and it
partitions as a 2 x 2 block matrix

[~ ~],
390 Chapter XVI Toeplitz and Singular Integral Operators

where A has size (n - K + 1) x (n + 1) and B has size (n + K) X n , Thus

rank A + rank B ~ n - K + 1 + n < 2n + 1.

Therefore the matrix (5.16) is not invertible , and hence QnMa./3 Qn on 1m Qn is
not invertible. We reached a contradiction. Hence K > 0 is impossible. In a
similar way one shows that K < 0 is impossible. Thus K = 0, as desired.
Part 2. In this part we assume that the winding number of a and f3 relative to
zero is equal to zero . We want to show that the first section method converges for
M a ./3 . Let P be the projection of lp(Z) defined by

I
Xj , j=0 ,1 ,2, .. . ,
P«Xj) jEZ) = (Yj)j EZ, Yj = 0, otherwise.

Put Q = I - P . Then
M a./3 = PLaP + QL/3Q + QLaP +PL/3Q.
As we have seen in the previous section, the operator QLa P + PL/3 Q is compact
and up to invertible factors the operators PLa P on 1m P and QL /3 Q on 1m Q are
equal to the Toep1itz operators Ta and T/3#' respectively. By our assumptions on
a and f3 the Toeplitz operators Ta and T/3# are invertible, and hence the operator
PLa P + QL /3 Q is invertible too. By Theorem 5.1 the invertibility of Ta and T/3#
also implies that these operators belong to IT{Pn }, where Pn is the projection of l p
defined by (5.1). Thus for n sufficiently large, n 2: N say, the operators PnTaPn
r,
and r, T/3# on 1m r, are invertible and
sup II(PnTaPn) - 11l < 00, sup II (PnT/3 #Pn)- l II < 00 .
n?N n? N
But this implies (use formulas (5.5) and (5.7) in the previous section) that for
n 2: N + I the operator Qn(PLaP + QL/3 Q)Qn is invertible and
sup IIWn(PL aP + QL/3 Q)Qn}-11l < O.
n? N
Since PLa P + QL /3 Q is invertible , we can use Theorem XII. 7.1 to show that the
finite section method converges for this operator. Our assumptions on a and f3
also imply that the operator M a,/3 is invertible (by Theorem 4.2). Thus M a./3 is an
invertible operator which is a compact perturbation of an invertible operator for
which the finite section method converges. But then Theorem XIV.3.5 implies the
the finite section method converges for Ma ./3 .

16.6 Singular Integral Operators on the Unit Circle

In this section we study the inversion of the following singular integral equation

1
on the unit circle 1[':
1 k(A, z)
w(A)f(A) +~ - - f ( z ) dz = g(A), A E 1['. (6.1)
7T:1 1f z- A
J6.6 Singular Integral Operators on the Unit Circle 391

Here w is a continuous function on T, the kernel function k is continuous on the

torus T x T, and k is assumed to satisfy the following Lipschitz type continuity
condition:
Ik(A, JL) - k(A, A) I .:s: clJL - AI. A,JLE'][', (6.2)
for some constant c. The right hand side g of (6.1) belongs to Lz('][') which
consists of all functions f such that t H f(e i l ) is Lebesgue measurable and
square integrable on -n .:s: t .:s: n , endowed with the inner product

The integral in (6.1) has to be understood as a principal value integral. The precise
meaning of the left hand side of (6.1) will be explained by the next lemma and its
proof. 0

Lemma 6.1 Condition (6.2) implies that the operator B is given by

(Bf)(A) = (k(A , z ) fez) d z, A E 1', (6.3)

J"T
Z- A

defines a bounded linear operator on Lz(1').

Proof: First let us remark that

i -1- dz = lim
1I' z - 1 e.j.O
(1--IT 0 it
- .ie- - dt
ell - 1
+ lIT -.--
ie
+e ell -
it
dt )
1
= tt i. (6.4)

This follows directly from the identity

. .
i ell (ell - 1)-
1
= '21 i + '12 sin t (1 - eos t) - ,
1

and the fact that the second term in the right hand side of this identity is an even
function. The integral in the left hand side of (6.4) does not change if we replace
z by A-I z where A E T is fixed. Thus

{_I_ dz = it i (A E 1') . (6.5)

Jr z -
A

1
Next, notice that

-I -zm- dz= lAm, m > 0,

(6.6)
it i 1I'Z-A _Am , m < O.

To see this we first take m > 0 and write

392 Chapter XVI. Toeplitz and Singular Integral Operators

Since

i 'f
zkd z = jlC te«Lkt eitdt = {271:i ,
-lC 0,
k=-I, .
otherwise,
(6.8)

we see from (6.7) and (6.5) that (6.6) holds for m ::: O. For m < 0 we get the
desired equality by replacing (6.7) by

z" = Am + (zm A-I + zm+I A-2 + ... + z-IAm)(A - z). (6.9)

The functions cn(e it) = eint, nEZ, form an orthonormal basis of £2(1['). We let
lP' be the orthogonal projection of £2('][') defined by

l
cn, n = 0, 1,2, ... ,
lP'cn = (6.1 0)
0, otherwise.

Put Q = I - P. Formulas (6.5) and (6.6) show that for each trigonometric
polynomial <p we have

1
----;-
7I:l
i -<p(z) dz = «lP' - Q)<p)(A),
'fZ-A
A E 1'. (6.11)

In the sequel for any <p in £2 (1') the function given by the left hand side of (6.11)
will by definition be equal to (P - Q)<p. Thus we define for each tp E £2 (1')

1
----;-
7I:l
i 'fZ-A
-
<p(z)
dz := «lP' - Q)<p) (A), A E l' a.e.. (6.12)

Now, let us consider the integral in (6.3). Put

h(A , z) = k(A, z) - k(A, A) .

Z-A
The function h is continuous on {(A, z) E l' x l' I A =F z} and, according to
condition (6.2), the function h is bounded on l' x 1'. Hence the function ii(t, s) =
h (e it , eis ) is square integrable on [-71:, 71:] X [-71: , 71:] . This shows that the operator
C given by

(Cf)(A) = ~ [ k(A , z) - k(A , A) fez) dz, A E l' a.e., (6.13)

71:1 i'f z- A

is a Hilbert-Schmidt integral operator on £2(1'). In what follows we understand

the function Bf in (6.3) to be defined by

(Bf)(A) = 7I:i (Cf) (A) + k(A, A) [ _1_ f(z) d z; A E l' a.e.. (6.14)
i'f z - A
In this way B is a well-defined bounded linear operator on £2 (1') . D
16.6 Singular Integral Operators on the Unit Circle 393

The operator in the left hand side of (6.12) is usually referred to as the operator
of singular integration, and it is denoted by S']f. Notice that equation (6.1) can
now be rewritten in the following equivalent form

w(A)f(A) + k(A, A)(S']ff) (A) + (Cf)(A) = g(A), A E T a.e., (6.15)

where C is the Hilbert-Schmidt integral operator given by (6.12). Next, put

a(A) = W(A) + k(A , A), f3(A) = W(A) - k(A , A). (6.16)

Then both a and f3 are continuous functions on T, and using ST = IP' - Ql, we can
rewrite (6.14) as
a (.)IP' f + f30Ql + Cf = g.
Now put
Af = a( ·)1P' f + f30Qlf + Cf. (6.17)
In what follows we refer to A as the operator on L2 (T) associated with (6.1) .
The next theorem is the first main result of this section.

Theorem 6.2 The operator A associated with the singular integral equation (6 .1)
is Fredholm ifand only if the functions a and f3 defined by (6.16) do not vanish
on the unit circle, and in that case ind A is the negative ofthe winding number of
y = a]f3 relative to zero.

Proof: Since the operator C defined by (6.13) is compact, formula (6.17) shows
that A is a compact perturbation of the operator Ao = a (.)IP' + f3(.)Ql. But then,
by Theorem 4.1 in the previous chapter, it suffices to prove the theorem for Ao in
place of A .
Next, let U be the operator on L2 (T) that assigns to each f E L2 (T) its
sequence of Fourier coefficients (c n (f» nE'Z with respect to the orthonormal basis
int
. .. , £ - 1, £0 , £1 , . . ., where £ n (e") = e for n E 2 . Thus

It follows that Uf E £2(2), and U is a unitary operator from L2(1l') onto £2(2) .
From the definition oflP' in (6.10) and Ql = I -IP' we see that UIP'U- 1 = P and
UQlU- 1 = Q, where P is the orthogonal projection of £2(2) defined by

P«X j)jE7l.) = (. . . , 0, 0, Xo, XI , X2, .. .), (6.18)

and Q = I - P . Furthermore, for each f E L2 (T) we have

Ua(·)f = L aUf, (6.19)

where L a and LfJ are the Laurent operators on £2(2) with symbols a and f3 (or
in the terminology of Section IILl the Laurent operators on £2 (2) defined by the
394 Chapter XVI Toeplitz and Singular Integral Operators

functions a(t) = ex(eit ) and b(t) = f3(eit ) on -]7: ::: t ::: ]7:). We conclude that the
operator Ao = ex (.)IP' + f3(.)Q is unitarily equivalent to the pair operator Ma ,{J on
£2(Z).
From the remark made after the proof of Theorem 3.1 we know that for ex and
f3 continuous on T the pair operator Ma ,{J is Fredholm on £2(Z) if and only if ex
and f3 do not vanish on T, and in that case ind Ma,{J = -K, where K is the wind ing
number of a] f3 relative to zero. Since Ao and Ma ,{J are unitarily equivalent, the
same holds true for Ao, which completes the proof. 0

The operator Ao = ex(·)1P' + f3(.)Q appearing in the above proof is usually

referred to as the main part of the operator associated with (6.1). Notice that the
action of Ao is also given by

Aof = w(·)f + v(·)§·rrJ, (6.20)

where V(A) = k(A , A), the function w is as in (6.1) , and § 1!' is the operator of
singular integration. As we have seen , Theorem 6.2 remains true if A is replaced
by Ao. The next two theorems concern invertibility of the operator Ao .

Theorem 6.3 Let Ao be the main part ofthe operator associated with the singular
integral equation (6.1), and let ex and f3 be given by (6.16) . Then Ao is invertible
ifand only ifex and f3 do not vanish on the unit circle T and the winding number
of y = ex I f3 relative to zero is equal to zero. Let these conditions be satisfied, and
assume in addition that ex and f3 are analytic in an annulus containing T. Then y
admits a Wiener-Hopffactorization y = y_ y+ and the inverse of A is given by

where IP' is the projection defined by (6.10) and Q = I - IP'.

From (6.16) we see that the additional assumption requiring ex and f3 to be

analytic in an annulus containing T (which appears in the above and in the next
theorem) is satisfied if and only if the functions W(A) and k(A, A) are analytic at
each point of T. The main reason for putting this additional analyticity require-
ment is the use of the Wiener-Hopf factorization. If the functions are merely
continuous, the Wiener-Hopf factorization may not exist.
Theorem 6.3 appears as a special case of the next theorem.

Theorem 6.4 Let Ao be the main part ofthe operator associated with the singu lar
integral equation (6.1), and let ex and f3 be the functions defined by (6.16). Then
Ao is left or right invertible ifand only ifex and f3 do not vanish on the unit circle
T. Assume the latter condition holds, and let K be the winding number of y = ex I f3
relative to zero. Then
Exercises XVI 395

(i) Ao is left invertible if and only if K :::: 0, and in that case codim 1m Ao = K,

(ii) Ao is right invertible if and only if K :::: 0, and in that case dim Ker Ao =
-K.

If, in addition, a and {3 are analytic in an annulus containing the unit circle, then
y admits a Wiener-Hopffactorization

A E 11',

and a left or right inverse ofAo is given by

Here IP' is the orthogonal projection defined by (6.10) , the operator <Q = I - IP',
and SeA) = A.

Proof: Let U be the unitary operator from L2 (11') onto £2(Z) introduced in the
second paragraph of the proof of Theorem 6.2. Then U AoU- 1 = M a ,f3 , where
M a,f3 = La P + L f3 Q. Here P is the projection defined by (6.18), the operator
Q = I - P, and La and L f3 are the Laurent operators on £2 (Z) with symbols a and
{3,respectively. SinceUIP'U- 1 = P,U<QU- 1 = Q,andUw(.)U- 1 = Lwforany
continuous function on co, Theorem 6.4 immediately follows from Theorem 4.2
when the functions a and {3 belong to the class A (that is, are analytic in an annulus
containing the unit circle). To complete the proof for the general case, we use the
same unitary equivalence and apply Theorem 2.2, 0
As was noted earlier, this chapter is a continuation of the material in Chapter III.
Further developments can be found in the paper [Kre], and in the monographs
[GF], [BS], [GGK2]. The requirement that p is finite is not essential in the first
four sections. All results in these sections remain true for p = 00 with some minor
modifications in the proofs.

Exercises XVI

1. Let T be the Toeplitz operator on £p (l :::: p < 00) with symbol

Sa
a(A) = 2"A- 1
- (a 5)
+ 2 + A.

Here a is a complex parameter.

(a) Determine the values of the parameter a with lal i= 1 for which the
equation
Tx = 0, x E £p,
has a non-trivial solution. Compute all solutions of this homogeneous
equation,
396 Chapter XVI. Toeplitz and Singular Integral Operators

(b) For which values ofthe parameter a is the operator T invertible? Find
the inverse of T if it exists .
(c) What is the spectrum of T ? Does it depend on p?
2. Let T be the Toeplitz operator on l p (1 s p < (0) with symbol

Again a is a complex parameter.

(a) For which values ofa is the operator T invertible? When is Tone-sided
invertible?
(b) Solve the nonhomogeneous equation

Tx = eo,
where eo = (1,0,0 , . . .).
(c) Solve the same problem as in (b) with y = (1, q , q2, .. .), Iql < 1, in
place of eo.
3. Let T be the Toeplitz operator on lp (1 S p < (0) with symbol ex E A.
Assume ex admits a Wiener-Hopf factorization

(a) When has the homogeneous equation

Tx = 0,

a non-trivial solution? Find all solutions of this equations. Hint : use

the Taylor expansion of ex+O- 1 at zero .
(b) Prove that Ker T has a basis of the form

(Yo, Yl ,· · · , Yr-l , Yr, Yr+l, ·· .),

(0 , Yo , . .. , Yr-2 , Yr-l, r-, ...),

(0,0, . . . , 0, yo, Yl , · · .) .
'-.-'
r

Determine the number r.

(c) Solve the nonhomogeneous equation

Tx = eo, x E lp,
where eo = (1, 0, 0, . . .). Express the solution(s) in terms of the
functions ex±O-I .
Exercises XVI 397

(d) Solve the same problem as in (c) with y = (1, q, q2, . . .), Iql < 1, in
place of eo.
4. Let T be the Toeplitz operator on ip (1 ::s p < 00) with symbol a E A,
and let T' be the Toeplitz operator with symbol f3(A) = a(A -I).
(a) What is the relation between the matrices of T and T' relative to the
standard basis of i p ?
(b) Assume a admits a Wiener-Hopffactorization, and let eo = (1,0,0, .. .).
Show that T is invertible if and only if the equations

Tx = eo, T'x' = eo

have solutions in ip .
(c) Let a be as in (b), and assume the two equations in (b) are solvable in
i p ' Express the inverse of T in terms of the solutions x and x',
5. Let T be a Toeplitz operator with symbol a E A. Assume that the two
equations in (b) of the previous exercise have solutions in il. Show that
a(A) i- 0 for each A E 11' and that relative to zero the winding number of
the oriented curve t -+ a(e it ) with t running from - ] f to rr is equal to zero.
6. Let U and U+ be the operators on i p (1 ::s p < 00) defined in the first
exercise to Chapter XV, i.e.,

U(xo, XI, X2 . . .) = (0,0, . .. , 0, Xk, Xk+I, ...),

~
r
U+(Xo, XI, X2"') = (0,0, . .. ,0, Xr, Xr+I, .. .),
~
k

Consider the operator

00 00

A = I>-j(u+)j +aol + Lajuj,

j=1 j=1

where a j (j E Z) are complex numbers such that Ian I ::s cp In I, nEZ, for
some c ~ 0 and 0 ::s p < 1. Assume k i- r.
(a) Show that A is Fredholm if and only if
00

a(A)= L ajAji- O (IAI = 1).

j=-oo

(b) Ifthe latter condition on o holds, show that ind A = (k-r)m, wherem
is the winding number relative to zero ofthe oriented curve t t-+ a(e it )
with t running from - ] f to tt .
398 Chapter XVI Toeplitz and Singular Integral Operators

7. Fix 0 :s t < 00 . Let Y be the operator on Lp[O, 00) ,1 :s p < 00, defined
by
(Vj)(x) = I(x + t), x ~ O.
Also consider the operator Y + on L p (0, 00) given by

(V+ f)(x) = (/(X - t) for x ~ t,

o for O:s x < t.

(a) Compute the operators v-v and YY+ .

(b) Let a be a complex parameter, lal i= 1. Solve in L p[O, 00) the homo-
geneous equation

T
5a y+ I - (5)
+ 2: I + Vi
a = o.

8. Let Y and y+ be as in the previous exercise, and consider the operator

00 00

A = :L>-j(y+)j +aoI + I>jyj,

j=1 j=1

where a j (j E Z) are as in Exercise 6. Solve in Lp[O, 00) , I :s p < 00,

the equation AI = 0 under the assumption that L~-oo a j ).) i= 0 for each
A. E 1['.
9. Solve the same problem as in the previous exercise with Y and y+ being
replaced by the following operators:

Wf)(x) f tax) , x ~ 0,

(y+ f)(x) = (/(x/a) , x > 1,

0, O:sx:s1.

Here a is a fixed positive integer.

10. Let M = LaP + Lf3Q be the pair operator on lp(Z), I :s p < 00 , with

5
a(A.) = A.-I - 1], {3(A.) = A. - -.
2
Here 1] is a complex parameter.
(a) For which values of 1] is the operator M invertible? Find the inverse if
it exists.
(b) Determine the spectrum of M.
(c) When is M left or right invertible?
Exercises XVI 399

(d) Let 11]1 =1= 1. Solve in ep(Z) the equation Mx = 0. Also, solve
Mx = eo, with eo = (1,0,0, . . .).

11. Do the previous exercise with

a(A) = A-
I 2
-- , f3(A) = A - 1].
5

12. Let K = PLa + QLf3 be the associate pair operator on ep(Z), 1 S P < 00 ,
with a and f3 as in Exercise 10.
(a) For which values of 1] is the operator K invertible? Find the inverse if
it exists .
(b) Determine the spectrum of K.
(c) When is K left or right invertible?
e
(d) Let 11] 1 =1= 1. Solve in p (Z) the equation Kx 0. Also, solve
Kx = eo, with eo = (1,0,0, . ..).

13. Do the previous exercise with a and f3 as in Exercise 11.

14. Let A be the operator on L2 (T) associated with the singular integral equation

1]f(A) + ---:
1 1AZ
--fez)
n t or z - A
d: = g(A), A E T.

Here 1] is a complex number. For which values of 1] is the operator A

Fredholm? For these values of 1] compute the index of A.
Chapter XVII
Non Linear Operators

Linear operators are the simplest operators. In many problems on e has to consider
more complicated nonlinear operators. As in the case of linear operators, aga in
the main problem is to sol ve equations Ax = y for a nonlinear A in a Hilbert or
Banach space. Geometrically, this problem means that a certain map or operator
B leaves fixed at least one vector x, i.e. ,

x=Bx , where Bx=x+Ax- y ,

and we have to find this vector. Theorems which establish the existence of such
fixed vectors are calledfixed point theorems. There are a number ofvery important
fixed point theorems. In this chapter we present one ofthe simplest; the Contraction
Mapping Theorem. This theorem is very powerful in that it allows one to pro ve the
existence of solutions to nonlinear integral, differential and functional equations,
and it gives a procedure for numerical approximations to the solution. Some of
the applications are also included in this chapter.

17.1 Fixed Point Theorems

A function j which maps a set S into S is said to have afixed p oint if there exists
an s E S such that j (s ) = s.
Contraction Mapping Theorem 1.1 Let S be a closed subset of a Banach space
and let T map S in to S. Suppo se there exists a number a < 1 such that for all
x, y in S,
IJT x - T ylJ :5 alJx - ylJ · (Ll)
Then T has a unique fix ed point in S.

Proof: Given an arbitrary vector Xo E S , let

xn=TnxO , n=I,2, ... .

We shall show that {x n } converges to a fixed point of T. For convenience, we write

d (x , y) instead of IJx - y lJ .
By hypothesis and the definition of x n ,

d (xn, xn- d = d (T xn- l , T x n- z) :5 ad (xn- l , Xn-Z )

:5 a Zd (x n- Z, Xn- 3) :5 " ' :5 a n- 1d(xl , XO )'
402 Chapter XVII. Non Linear Operators

Hence for n > m,

d(x n, x m) ::: d(x m, xm+I) + d(xm+I, Xm+2) + ...+ d(Xn-l , x n)
00 m
s d(Xl,XO) "
LJ
a k = d(Xl,XO)_a- -+ 0
I-a
k=m
asn , m -+ 00 . Since S is a closed subset ofa Banach space, {x n } convergesto some
XES. Clearly, T is continuous. Hence we have Xn -+ x and x n+1 = T x n -+ T x,
which implies that T x = x .
If z is also a fixed point of T, then

d(z , x) = d(Tz, Tx) ::: adtr; x)

which can only be if d(z , x) = 0 or, equivalently, z = x. o

The operator T in Theorem 1.1 is called a contraction.

17.2 Applications of the Contraction Mapping Theorem

In this section the contraction mapping theorem is used to prove the existence and
uniqueness ofsolutions to certain non linear integral and differential equations. In
addition, we give a proof of the implicit function theorem.

Theorem 2.1 Let k be continuous on [a, b] x [a, b] x Co Suppose there exists a

number m such that

Ik(t, s ,~) - ktt , s , ~')I ::: ml~ - ~'I

for all~,~' E Co Thenfor IAI < m(La) ' the equation

I(t) - A l b
k(t , s, I(s)) ds = get) (2.1)

has a unique solution in C([a, b])for every g E C([a , b]).

Proof: Define T on C([a, b]) by

(Tf)(t) = A l b
k(t , s , I(s)) ds + g(t) .
It is easy to see that 1m T C C([a, b]) . T is a contraction. Indeed, for all I,
hE C([a, b]),

I(Tf)(t) - (Th)(t)1 ::: IAllb Ik(t , s, I(s)) - k(t , s, h(s))1 ds

::: IAllb ml/(s) - h(s)1 ds ::: IAlm(b - a)1I1 - hll ·

17.2 Applications ofthe Contraction Mapping Theorem 403

Thus
IITf - Thll :::: IAlm(b - a)lIf - hi!'
and IAlm(b - a) < 1. Hence T has a unique fixed point fo E C([a, b]) by
Theorem 1.1. Obviously, fo is the solution to (2.1).
Equation (1.1) includes linear integral equations of the second kind. For given
ko(t, s), let k(t, s,~) = ko(t, s)~.
Then

l b
k(t , s, f(s)) ds = l b
ko(t , s)f(s) ds.

Theorem 2.2 Let f be a complex-valuedfunction which is continuous and bounded

on some open subset 0 ofthe plane. Suppose there exists a number C such that

If(x, y) - f(x , z)1 :::: C!y - z] (2.2)

for all (x, y), (x, z) in O. Thenfor any (xo, YO) EO, the differential equation
dy
dx = f(x, y(x)) (2.3)

with intitial condition

y(xo) = YO (2.4)

has a unique solution y = y(x) on some imterval containing xo.

Proof: Equation (2.3) together with (2.4) is equivalent to the equation

y(x) = YO + r
lXQ
f(t, yet»~ dt. (2.5)

We now show that (2.5) has a unique solution on some interval containing xn, To do
this, we define a contraction map as follows. Let M = sUP(x.y)eO If(x, y)1 < 00.
Choose p > 0 such that

Cp < 1, [x - xol :::: p and Iy - rcl :::: Mp imply (x, y) EO.

Let S be the set ofcomplex valued functions y which are continuous on the interval
J = {x : Ix - xol :::: p} and have the property that

Iy(x) - Yol ::: Mp.

It is clear that S is a closed subset of the Banach space C(J). Define Ton S by

(Tg)(x) = YO + lXQ
x
f(s, g(s)) ds.
404 Chapter XVII. Non Linear Operators

Now TS C S since TS C C(J) and for g E S,

I(Tg)(x) - Yol = 11: j(S,g(S))dSI s Mp , x E J.

Furthermore, T is a contraction. Indeed, given g, h E S, (2.2) implies

I(Tg)(x) - (Th)(x)1 = 11: j(s, g(s)) - j(s, h(s)) I s Cpllg - fII,

whence

IITg - Thll s Cpllg - hll , Cp < 1.

Therefore, there exists a unique yES such that Ty = Y, i.e., (2.5) holds.
If YI is also a complex valued function on J which is a solution to (2.3) with initial
condition (2.4), then YI is in C(J) and is a solution to (2.5). Hence IYI (x) - Yol s
M p for all x E J . Thus YI is in Sand YI is a fixed point of T . Therefore, Y = YI.
D

continuous on some set Q in the plane. Suppose

and is continuous at some point (xo, YO) E Q . If
*
Implicit Function Theorem 2.3 Let j be a real valued function which is
exists at each point in Q

and
j(xo, YO) = 0,
then there exists a rectangle R : [xo - 8, Xo + 8] x [Yo - e, YO + £] contained in Q
andauniquey(x) E C([xo-8,xo +8]) such that (x , y(x)) E R, j(x ,y(x)) =0
for all x E [xo - 8, Xo + 8] and y(xo) = YO .
Proof: We shall show that for a suitable set of functions, the operator
1 aj
(TgJ)(x) = gJ(x) - - j(x, gJ(x)), m = -(XO,Yo) (2.6)
m ay
has a fixed point.
It follows from the conditions on j that there exist e > 0 and 8 > 0 such that
the rectangle R = {(x, y) : Ix - xol s 8, IY - Yol s s] is contained in Q,

1 aj 1
<2'
I
ma/ x,y)-1 (x , y) E R , (i)
1

and
(ii)
J 7.3 Generalizations 405

Let J = {x : Ix - xol ~ 8} and define

S = {g E C(J) : g(xo) = YO , Ig(x) - Yol ~ 8, x E f}.

Clearly, S is a closed subset of the Banach space C(f) . Define T on S by (2.6).

We shall show that TS c S and that T is a contraction. Now T S C C(f) and for
(x , y) E R, we have from (i) that

Iaf (I f
-
ay
y - -
m
(x, y) )I I 1af
= 1- - -
may
I 21
(x, y) < - .

Hence it follows from the Mean Value Theorem applied to the function
y - ~ f(x, y) that for g and h in S,

1
I(Tg) (x) - (Th)(x) I ~ 2 Ig(x) - h(x)l · (iii)

Thus
1
IITg-Thll ~ 2I1g-hll .

To show that T has a fixed point in S, it remains to prove that T S c S. Given

g E S,
1 1
(Tg)(xo) = g(xo) - - f(xo, g(xo)) = Yo - - f tx«, YO) = Yo·
m m
Define yo(x ) = Yo, x E f . From (iii), (ii) and the definition of S, we get

I(Tg)(x) - Yol ~ I(Tg)(x) - (Tyo)(x) I + I(Tyo)(x) - Yol

~ ~lg(X)-Yol+l~f(x ,yo)1 <8.

Thus Tg E S, which proves that TS c S. Hence T has a unique fixed point

yES. Clearly, Y satisfies the conclusions of the theorem. If Yt also satisfies the
conclusions, then Yt is in Sand TYI = Yl. Therefore, Yl = y.
For any cp E S, the sequence {Tncp} converges uniformly on J to y(x) . 0

17.3 Generalizations

The function d defined on S x S by d (x , y) = IIx - Y II has the following properties.

(i) d(x , y) > 0 if x i= Y; d(x, x) = O.

(ii) d(x, y) = dey, x) .
(iii) d(x, y) + dey, z) ~ d(x , z) .
406 Chapter XVII. Non Linear Operators

An arbitrary set S, together with a function d defined on S x S with properties

(i), (ii) and (iii), is called a metric space. The metric space (S, d) is called com-
plete if limm,n--+oo d (xn, x m) = 0 implies the existence of an XES such that
d(x n, x) ---* O.
The proof of the contraction mapping theorem shows that this theorem holds
for complete metric spaces.

Theorem 3.1 Suppose S is a complete metric space and T : S ---* S has the
property that T" is a contractionfor somepositive integern. Then T has a unique
fixed point.
Proof: Since T" is a contraction, it has a unique fixed point XES. Now

T''T» = TT?» = Tx,

i.e., Tx is also a fixed point of T" , But T" has only one fixed point. Hence
T x = x. Any other fixed point of T is also a fixed point of T" , Thus the fixed
point of T is unique. 0
Ifwe remove the requirement in Theorem 1.1 that T is a contraction and replace
it by the weaker condition

IITx - TYII < IIx - YII, for all x, yES, x =1= y,

then T need not have a fixed point. However, if, in addition, S is compact, then T
has a unique fixed point.

Definition: A subset S of a normed linear space is compact if every sequence in

S has a subsequence which converges to a vector in S.

Theorem 3.2 Let S be a compact subset ofa normed linear space and let T map
S into S. If
IITx - TYII < IIx - YII for all x , Y in S, x =1= y , (3.1)
then T has a uniquefixed point in S.
Proof: Let m = infxEsIITx - x]: There exists a sequence {xn} in S such that
IITxn - xnll ---* m , Since S is compact, {xn} has a subsequence {xn'} which
converges to some XES. Hence

IIx - Txll = lim IIxn, - Txn,II = m.

n'--+oo

Therefore T x = x, otherwise

m S IIT 2x - Txll < IITx - xII = m.

It is obvious from (3.1) that T cannot have more than one fixed point. 0
17.3 Generalizations 407

The proof shows that the theorem can be extended to compact metric spaces .
Ifthe set S in the Theorem 3.2 is also convex, then we can weaken the inequality
(3.1) and obtain the following result.

Theorem 3.3 Let S be a compact convex subset ofa normed linear space and let
T map S into S. If

IITx - TYII:s IIx - YII forall X,Y in S,

then T has afixed point.

Proof: First we assume that 0 E S. For n = 1,2, . . . , define Tn on S by E;» =

(l - ~)x . Since S is convex and 0 E S, Ti,» = (l - ~)x + ~.O E S. Clearly, Tn
is contraction. Hence by Theorem 1.1, there exists an X n E S such that

(1-~)Txn=Tnxn=xn, n=I,2, . .. . (3.2)

Since S in compact, {x n } has a subsequence {x n' } which converges to some XES.

Thus Tx n, -+ Tx and (3.2) implies

Tx = lim Tx n, = lim Xn' = X.

n'--+oo n--+oo

If 0 ¢ S, choose any Xo E S and let So = -xo + S. The theroem follows from

the result we just proved applied to So and the operator To : So -+ So defined by

To(-xo +x) = -xo + Tx, XES . o

Appendix 1
Countable Sets and Separable Hilbert Spaces

This appendix presents an alternative definition of separability of a Hilbert space.

We start with some preliminary results about countable sets .
Definition: A set S is countable if it is either finite or can be put into one to one
correspondence with the positive integers. Thus an infinite set is countable ifthere
exists a sequence whose terms consist of all the members of S.

(i) If S is a countable set, then the set consisting of all finite sequences of
members of S is also countable.
Indeed, since S can be put into one to one correspondence with a subset
of positive integers it suffices to prove that the set Z which consists of all
finite sequences of positive integers is countable.
For each ~ositive integer n, let Z; be the set of those (m I, . . . ,mk) E Z for
which Li=1 m; = n. Since Zn is a finite set, the members of Zn may be
listed as we please. Thus we can list all the members of Z = U~I Z; by
listing the members of ZI, then Z2, etc .
(ii) The set of all finite sequences of rational complex numbers (the real and
imaginary parts of the number are rational) is countable. Indeed, by iden-
tifying the complex number ~: + i ~ with (PI, ql, P2, q2), where Pj and
qj are integers, it follows from (i) th~t the set of finite sequences of rational
complex numbers is a subset of a countable set and therefore is countable.
(iii) The set of all polynomials with rational complex coefficients is count-
able since we can identify the polynomial anx n + ... + alx + ao with
(an, . '" ai , ao)·
(iv) The set of real numbers is uncountable. Indeed, it suffices to show that
any sequence of real numbers in [0, I] does not contain some real number
in [0, I].
Suppose Sl, S2 , . . . is a sequence of real numbers. Each Si has the
decimal expansion s; = oii) oii), .. . , where infinitely many of the inte-
gers 01(i) ,° (i) , . .. are not 9 . L et s = . °1°2 .. . , were
2
h Ok = O I'f a (k) -I-
k I 0
and I if ak
k
) = O. Then s is in [0, I] but s i= si for every i .
410 Appendix 1. Separable Hilbert Spaces

Theorem: A Hilbert space 'H is separable if and only if'H contains a countable
set which is dense in H,

Proof: If {WI , W 2 , .. . } is dense in 'H, then obviously Sp{WI, W 2 , . .. } is dense in

Ji. Therefore, by definition, 'H is separable. On the other hand, ifJi is separable,
then there exist vectors VI , V2 , . .. such that 'H = sp{VI, V2 , .. . }. Hence the set
S of all linear combinations of the form "L7= I (Xi Vi , where each (Xi is a complex
rational number, is dense in 'H. The countability of S now follows from (ii), 0

Corollary: Every orthonormal set in a separable inner product space is

countable.

Proof: Let n be an orthonormal set in an inner product space E and let S

{Xl , x2 , . . .} be dense in E. We define an integer valued function J on n as
follows:
For each ep E n, choose an integer j such that lIep - Xj I < 1/2 and define
J(ep) = j . Now if 1/1 is in nand 1/1 :f=. ep, then lIep -1/111 = -J2, which implies
that J(ep) :f=. J(1/I). Hence n is in one to one correspondence with a subset of the
positive integers and therefore is countable. 0
Appendix 2
The Lebesgue Integral and L p Spaces

In this section we give a very brief introduction to the Lebesgue integral which is
necessary for the description and understanding of the spaces L p .

A.2.t Lebesgue Integral

A subset Z of the real line is said to have Lebesgue measure zero if for each
8 > 0, there exists a countable set of intervals h , h . . . such that Z C U] I j and
L j JL(I j ) < 8 , where JL(Ij ) is the length of I j .
Every countable subset {Xl , X2, . •. } of the line has Lebesgue measure zero.
Indeed, given 8> 0, take I j = [Xj , Xj + 8/2 H 1).
A real valued function I which is defined on an interval J is called a step
f unction if I(x) = L~=l CXkC h (x ), where CXk is a real number, h , . . . , In are
mutually disjoint subintervals, and C h (x) = I if X E h and zero otherwise. The
step function I is Lebesgue integrable if L~= I cxkJL(h ) < 00, (0· 00 = 0). In
this case, we define the integral

[ I(x) d x =
JJ
t
k= l
cxkJ1(h )·

A non negative real valued function I defined on the interval J is Lebesgue mea-
surable if there exists a nondecreasing sequence {In} of step functions defined on
J such that In (x) ~ I (x ) for all x E J which lie outside some set of Lebesgue
measure zero (possibly the empty set). We then say that Un} converges to I
almost everywhere (a.e.). If each step function In is Lebesgue integrable, then
fJ II (x ) dx ~ f J fz(x ) d x s .... The function I is Lebesgue intergrable if
lim J. I n (x) d x < 00 . The Lebesgue integral of I is defined by
n-+oo J

[ I(x) d x =
JJ
lim
n-+oo
r I n (x) d x .
JJ
It turns out that the integral is independent of the choice of the increasing
sequence Un}.
412 Appendix 2. The Lebesgue Integral and L p Spaces

Given a real valued function f defined on J, define

f+(x) = max(f(x), 0), f-(x) = max(- f(x), 0).

Then f = r - r. f+ ::: 0, i ::: O. The function f is Lebesgue integrable

ifboth r
and f- are Lebesgue integrable. In this case we define
i f(x) dx =i f+(x) dx - i f-(x) dx.

If f is a complex valued function on J, say f = f1 + ih, where f1 and I:

are real valued functions, then f is Lebesgue integrable if !J and h are Lebesgue
integrable, in which case we define

i f( x) dx =i f1 (x) dx +i i hex) dx.

A bounded Lebesgue measurable function is Lebesgue integrable. A real

valued function which is Riemann integrable is Lebesgue integrable and the two
integrals coincide. However, a Lebesgue integrable function need not be Riemann
integrable. For example, if f is zero on the rationals and 1 on the irrationals
in [0, 1], then f is not Riemann integrable, but it is Lebesgue integrable with
fd f(x) dx = 1.
The following theorem is fundamental to the theory of Lebesgue integration.
Lebesgue Dominated Convergence Theorem. Suppose Un} is a sequence of
Lebesgue integrable functions which converges a.e. to f on the interval J . If
there exists a Lebesgue integrable function g such that

Ifn(x)1 .:s g(x) a.e. on J, n = 1,2, . . . ,

then f is Lebesgue integrable and

ifn(X)dXt--+ if(X)dX .

The theory of Lebesgue integration of complex valued or real valued functions

defined on a rectangle is very similar to the theory of Lebesgue integration of
functions defined on an interval. One need only replace, in the discussion above,
intervals by rectangles and lengths by areas . The integral of f over a rectangle R
is written fR f(x, y) dxdy.
Fubini's theorem enables one to express the integral as an iterated integral.
Fubini's Theorem. Let f be Lebesgue integrable on the rectangle R : [a, b] x
[c, d]. Thenfor almost every x E [a, b], thefunction f (x , ) is Lebesgue integrable
on [c, d] and fed f(x, y) dy is integrable on [a, b]. Moreover,

i {l b d
f(x, y) d Y} dx = 1 f(x , y) dxdy.
A.2.2 t. , Spaces 413

f:
Similarly, for almost every y E [c, d], the function f ( , y) is Lebesgue integrable
on [a, b] and f (x , y) dx is Lebesgue integrable on [c, d], Moreover,

A.2.2 L p Spaces

A complex valued Lebesgue measurable function defined on an interval J is said to

be in L p ( J) , 1 :s p < 00, if If Ip is Lebesgue integrable . With the usual definition
of addition and scalar multiplication of functions , L p (1 ) is a vector space. If we
define

IIf11 p = (l'f(X WdX)':

the II . II p has all the properties of a norm except that II f IIp = 0 only implies
f = 0 a.e. To remedy this situation, functions are identified which are equal
almost everywhere. To be specific, we decompose L p (1 ) into disjoint sets, called
equivalence classes, as follows. For each f E L p (1 ) ,

[f] = {g E L p (1 ) : f = g a.e.].
Then, either [f] = [h] or [f] n [h] = 0 . The set of equivalence classes becomes
a vector space under the operations

[f] + [h] = [f + h], a[f] = [afl

Moreover,
1I[f]lI p = (l,g(XWdX) li P ,

where g is arbitrary in [f] , defines a norm on this vector space. For the sake
of simplicity, we usually do not distinguish between functions and equivalence
classes.
It is a very important fact that L p(1) is complete . For a detailed proof we refer
the reader to [R].
A complex valued Lebesgue measurable f defined on J is called essentially
bounded if there exists a number M such that If (x )1 :s M a.e. The greatest lower
bound of all such M is denoted by II f II 00 ' If functions which are equal almost
everywhere are identified as above, then II . 11 00 is a norm on the vector space
L oo (1 ) of essentially bounded function s and L oo (1 ) is a Banach space.
The following inequal ity is essential for our treatment of integral operators on
t., spaces.
414 Appendix 2. The Lebesgue Integral and L p Spaces

HOlder's Inequality. If f E L p(J) and g E Lq(J) , where 1 S p S 00 and

i+ ~ = 1(~ = 0), then f . g is in L1 (1) and

Equality holds ifand only ifthere exists some non-zero a , f3 in <C such thatalfl P =
f3lgl q a.e.
Suggested Reading

1. Akhiezer, NJ., and Glazman, l.M., Theory ofLinear Operators in Hilbert

Space, vol. I (1961) and vol. II (1963) Ungar, New York.
2. Bottcher, A., and Gudsky, S.M., Toeplitz matrices, asymptotic linear
algebra, andfunctional analysis, Birkhauser Verlag, Basel, 2000 .
3. Bottcher, A., and Silbennann, 8., Analysis ofToeplitz operators, Springer-
Verlag, Berlin , 1990.
4. Douglas, R.G., Banach Algebra Techniques in Operator Theory, Academic
Press, New York, 1972.
5. Dunford, N., and Schwartz, J.T., Linear Operators, Part I: General Theory
(1958) Part II: Spectral Theory (1963), Interscience, New York.
6. Gelfand, LM., Functional Analysis, in: Mathematics - Its Content. Methods
and Meaning, 2nd ed., vol. 3, M.LT. Press , Cambridge, 1969,227-261.
7. Gohberg, LG., and Krein, M.G., Introduction to the Theory ofLinear Non-
SelfAdjoint Operators in Hilbert Space, Translations, Math . Monographs,
vol. 18, Amer. Math. Soc., Providence, 1969.
8. Goldberg, S., Unbounded Linear Operators, McGraw-Hill, New York,
1966.
9. Halmos, P.R., A Hilbert Space Problem Book, Van Nostrand, Princeton,
1967.
10. Kato, T., Perturbation Theory for Linear Operators, 2nd ed., Springer
Verlag, New York, 1976.
11. Lax, P.D., Functional Analysis, Wiley-Interscience, New York, 2002.
12. Riesz, E, and Sz.-Nagy, 8., Lecons d'Analyse Fonctionnelle, Academie des
Sciences de Hongue, 1955.
13. Rudin, W., Functional Analysis, McGraw-Hill, New York, 1973.
14. Schechter, M. Principles ofFunctional Analysis, Amer. Math. Soc., Pro-
vidence, R.I., 2002.
15. Shilov, G.E., An Introduction to the Theory ofLinear Spaces, Dover Pub.,
New York, 1974.
416 Suggested Reading

16. Sz.-Nagy, 8., and Foias, C., Harmonic Analysis of Operators on Hilbert
Space, North-Holland Publ. Co., Amsterdam-Budapest, 1970.
17. Taylor, A.E., and Lay, D.C., Introduction to Functional Analysis. 2nd ed.,
Wiley, New York, 1980.
18. Weidmann, J., Linear Operators in Hilbert Spaces, Springer Verlag,
New York, 1980.
19. Zaanew, AC., Linear analysis, North-Holland Publ. Co., Amsterdam, 1956.
References

[A] S.S. Antman, The equations for large vibrations ofstrings , Amer. Math.
Monthly 87 (1980), 359-370.
[Ah] t .v Ahlfors, Complex Analysis, McGraw-Hill, 1966.
[BS] A. Bottcher and B. Silberrnann, Analysis of Toeplitz Operators,
Springer-Verlag, Berlin, 1990.
[CH] R. Courant and D. Hilbert , Methods ofMathemati cal Physics, vol. I,
Interscience, New York, 1953.
[DS1] N. Dunford and J.T. Schwartz, Linear Operators, Part I: General
Theory, Interscience, New York, 1958.
[DS2] N. Dunford and J.T. Schwartz, Linear Operators, Part II: Spectral
Theory, Intersc ience , New York, 1963.
[E] P. Enflo, A counterexample to the approximation problem in Banach
spaces, Acta Math. 130 (1973), 309-317.
[F] LA. Fel'dman, Some remarks on convergence of the iterative method,
Izvestia Akad. Nauk Mold. SSR 4 (1966), 94-96.
[G] S. Goldberg, Unbounded Linear Operators, McGraw-Hill,
New York, 1966.
[GF] I. Gohberg and I. Feldman, Convolution Equations and Projection
Methods for their Solution , Transl. Math . Monograph, vol. 41, Amer.
Math . Soc., Providence, R.I., 1974.
[GG] I. Gohberg and S. Goldberg, Basic Operator Theory, Birkhauser,
Basel, 1981.
[GGKa1] I. Gohberg, S. Goldberg and M.A. Kaashoek, Classes of Linear
Operators , vol. I, Birkhauser, Basel, 1990.
[GGKa2] I. Gohberg, S. Goldberg and M.A. Kaashoek, Classes of Linear
Operators, vol. II, Birkhauser, Basel, 1993.
[GGKr] I. Gohberg, S. Goldberg and N. Krupnik, Traces and determinants of
linear operators, Birkhauser, Basel, 2000.
[GKre] I. Gohberg and M.G. Krein, Introduction to the theory oflinear non-
selfadjoint operators, Transl. Math. Monographs, vol. 18, Amer. Math.
Soc., Providence, R.I., 1969.
[GKre1] I. Gohberg and M.G. Krein, The basic propositions on defect numbers,
root numbers and indices oflinear operators, Uspekhi Math. Nauk 12,
2(74) (1957) , 43-118 (Russian); English Transl., Amer. Math. Soc.
Transl. (Series 2) 13 (1960) ,185-265.
[H] P.R. Halmos, Finite-dimensional Vector Spaces, 2nd ed., Van
Nostrand, Princeton, 1958.
418 References

[K] T. Kato, Perturbation Theory for Linear Operators, 2nd ed., Springer
Verlag, New York, 1976.
[Kra] M.A Krasnosel'skii, Solving linear equations with selfadjoint
operators by iterative method, Uspehi - Matern. Nauk, 15,3 (1960),
161-165 .
[Kre] M.G. Krein, Integral equations on a half-line with kernel depending
upon the difference of the arguments, Uspekhi Math. Nauk 13 (5)
(1958), 3-120 (Russian); English Transl., Amer. Math. Soc. Trans/.
(Series 2) 22 (1962), 163-288.
[P] H. Poincare, Sur les determinants d'ordre infini, Bull. Soc. Math.
France 14 (1886), 77-90.
[PS] C. Pearcy and AL. Shields, A survey of the Lomonosov technique
in the theory of invariant subspaces, Topics in Operator Theory,
Mathematical Surveys, No. 13, American Mathematical Society, Prov-
idence, 1974.
[R] H.L. Royden, Real Analysis, 2nd ed., Macmillan, New York, 1968.
[S] G. Strang, Linear Algebra and Its Applications, 2nd ed., Academic
Press, New York, 1980.
[Sc] M. Schechter, Principles offunctional analysis, Amer. Math. Soc.,
Providence, R.I., 2002.
[TL] AE. Taylor and D.C. Lay, Introduction to Functional Analysis, 2nd
ed., Wiley, New York, 1980.
[W] R. Whitley, Projecting m onto co, Amer. Math. Monthly, 73 (1966),
285-286.
List of Symbols

A' adjoint of the operator A, 77

A' conjugate of the operator A , 307
A IM restriction of the operator A to M
A (HI -+ Hz) operator A with domain a subspac e of HI and
range a subspace of Hz , 204, 279
C field of complex numbers
en complex Euclidean n-space, 1
C([a , bj) the space of complex valued continuous functions on [a , bj , 260
COOUa, b)) the space of infinitely differentiable complex
valued functions which vanish outside an open subinterval of [a , bj , 208
codim M codimension of a subspace M, 88
d(A) codimension of the range of A , 347
as boundary of a set S
dim M dimension of a subspace M
d (v, s ) distance from a vector v to a set S, 8
0jk Kronecker delta, 33
V(A) domain of an operator A, 203 , 279
det p (l - P) determinant of I - P , with P a Poincare operator, 319
:F(lz) space of bounded linear operators of finite rank , 321
g(YI , Yz , · · · , Yn) Gram determinant, II
G(A) graph of the operator A, 207
1m A range of an operator A, 63, 279
;) Cl imaginary part of a complex number Cl
(', -)A inner product on the graph of A, 207
ind A index of the operator A , 347
K winding number, 143

Ker A kernel of A, 65
lz , lz (2) the Hilbert spaces of square summable sequences
with entries in C, 3, 5
lz(w) weighted lz space, 40
lp ,1 =:: p s 00,260
Lp([a , bj) , 1 =:: p =:: 00,261 ,413
£(H) space of bounded linear operators on H , 52
£(HI , Hz) space of bounded linear operators from HI into Hz, 52
Ma ,{3 pair operator, 377
n(A) dimension of the kernel of A, 347
n(Pn ) , n(Pn , Qn) , 97, 289
420 List ofSymbols

P space of Poincare operators, 317

p(n ),318
p eA) resolvent set of the operator A, 110,290
!Ra real part of a complex number a
r eT) spectral radiu s of the operator T , 147
S.l orthogonal complement of S
and also the annihilator of S, 17, 308
.IN, 308
sp S span of the set S, 6
S closure of the set S, 16
a(A ) spectrum of the operator A , 110,290
S j ( A) j -th singular value of the operator A , 248
'][' unit circle in the complex plane
tr A trace of the operator A , 320
X' conjugate space of the space X, 265
Z set of integers
Index

adjoint operator 77, 208 conjugate of an operator 307

algebraic multiplicity 337 conjugate space 265
almost periodic 38 continuity of an operator 57
analytic operator valued contraction 402
function 295 contraction mapp ing theorem 40 I
associate pair operator 384 convex set 16
countable set 409
backward shift 56
Baire category theorem 281 defining function 136
ba116,309 diagonalization 178
Banach space 259 Dini 's theorem 197
band matrix 94 direct sum 85, 271
band operator 94 distance from a point to a set 8
basic harmonic motions 222
basic system of eigenvectors eigenvalue 172
eigenvector 172
and eigenvalues 180
equivalent norms 262, 283
basis , orthonormal 28
Schander 265
finite section method 99, 159, 163,385
standard orthonormal 28
fixed point 40 I
bilateral shift 56
forward shift 56
biorthogonal system 48, 271
Fourier coefficient 28, 31
bounded linear functionals,
Fourier series 31
representation of 265-267
Fourier transform 37
bounded linear operator 52, 277
Fredholm alternative 306
Fredholm operator 347 ,356
Cauchy-Schwarz inequality 7
functional 60
Cauchy sequence 7, 259
Cayley transform 255 Gram determinant II
characteristic frequencies 222 graph of an operator 207
closed graph theorem 281 graph norm 207, 280, 357
closed operator 203 , 279 Green's function 205
closed set 16
closure of a set 16 Hadamard's inequality 13
codimension 88 Hahn-Banach theorem 267
compact operator 91, 299 Hilbert space 7
compact set 406 Hilbert -Schmidt operator 253
complement of a subspace 85, 271 Hilbert-Schmidt theorem 193, 194
complete 406
complete inner product space 7 image of an operator 63
complete normed linear space 259 imaginary part of an operator 244
complete orthonormal system 29 implicit function theorem 404
complex n-space I index of an operator 347,356
422 Index

injective map 65 Poincare operator 317

inner product 6 determinant 319
inner product space 6 positive operator 185
integral equation of the second kind 73 product space 207
integral operator 56 projection 84, 286
interior point 281 projection method 97, 289
invariant subspace 108 modified 106
inverse operator 64, 204 Pythagorean theorem 9
invertible operator 64, 204
isometry 114,246 quotient space 274

kernel function 56 range of an operator 63

kernel of an operator 65 rank of an operator 63
Kronecker delta 33 real part of an operator 244
regular point 110, 290
Laurent operator 135, 36 1 resolvent 295
Lebesgue integral 411 resolvent set 110, 290
left inverse 85 Riesz representation
left invertible 85 theorem 61
Legendre polynomial 26 right inverse 89
linear isometry 36 right invertible 89
L p space 261, 413
Schauder basis 265
maximal element 267 Schwarz inequality 7
Mercer's theorem 197 self adjoint operator 80, 210
metric space 406 separable Hilbert space 35
modified projection method 106 normed linear space 264
multiplicity of an eigenvalue 181 simple eigenvalue 213
simple harmonic oscillation 220
non negat ive operator 185 singular values 248
norm 6, 259 span of a set 6
of an operator 52 spectral radius 147
normal operator 244 spectral theorem 178
normed linear space 259 spectrum of an operator 110, 290
nuclear operator 314 standard basis 19
nth section 159, 318 orthonormal basis 28
strictly positive operator 166
one-one operator 65 Sturm-Liouville system 211
operator of algebraic complements 334 operator 21 I
operator of singular integration 393 subspace 6
orthogonal complement 17 invariant 108
orthogonal projection 81 symbol 136, 142,362,364
orthogonal vectors 8 symmetric operator 215
orthonormal basis 28 Szego polynomials 19
stability of34
orthonormal system 9 Toeplitz matrix 20, 59
Toeplitz operator 141, 364
pair operator 377 totally ordered set 267
parallelogram law 7 trace class 253
Parseval's equality 29 trace of finite rank operator 321
partially ordered set 267 Poincare operator 320
Index 423

triangle inequality 2, 7 Weierstrass approximation theorem 29

tridiagonal matrix 94 second approximation theorem 30
Wiener-Hopf factorization 365
uniform boundedness princ iple 284 winding number 143, 152, 153
uniformly convex 274 Wronskian 211
unitarily equivalent 248
unitary operator 114,246 Zorn's lemma 267

vector space 5
Volterra integral operator 293, 294
or series
Edited by
Israel Gohberg, School of Mathematical
Sciences, Tel Aviv University, Ramat Aviv, Israel
This series is devoted to the publication of
current research in operator theory, with particular emphasis
on applications to classical analysis and the theory of integral
equations, as well as to numerical analysis, mathematical phy-
sics and mathematical methods in electrical engineering.
OT series
Edited by
Israel Gohberg, School of Mathematical
Sciences, Tel Aviv University, Ramat Aviv, Israel
This series is devoted to the publication of
current research in operator theory, with particular emphasis
on applications to classical analysis and the theory of integral
equations, as well as to numerical analysis, mathematical phy-
sics and mathematical methods in electrical engineering.