(Graduate Studies in Mathematics, V. 50) Y. A. Abramovich, Charalambos D. Aliprantis-An Invitation To Operator Theory-Amer Mathematical Society (2002)

An Invitation to
Operator Theory
Y. A. Abramovich
C. D.Aliprantis
Graduate Studies
in Mathematics
Volume SO
American Mathematical Society

An to
Operator Theory
Y. A. Abramovich
Indiana University-Purdue University Indiana po/ls
C. D. Aliprantis
Purdue University
Graduate Studies
in Mathematics
Volume 50
American Mathematical Society

Providence, Rhode Island
Editorial Board
Walter Craig
Nikolai Ivanov
Steven 0. Krantz
David Saitman (Chair)
2000 Mathematics Subject Classification. Primary 46Axx, 46Bxx, 46Gxx, 47Axx, 47Bxx,
47Cxx, 47Dxx, 47Lxx, 28Axx, 28Exx, 15A48, 15A18.
Library of Congress Cataloging-in-Publication Data

Abramovich, Y. A. (Yuri A.)
An invitation to operator theory / Y. A. Abramovich, C. D. Aliprantis.
p. cm. (Graduate studies in mathematics, ISSN 1065-7339 ; v. 50)
Includes bibliographical references and index.
ISBN 0-8218-2146-6 (alk. paper)
1. Operator theory. I. Aliprantis, Charalambos D. II. Title. III. Series.
QA329.A25 2002
2002074420
Copying and reprinting. Individual readers of this publication, and nonprofit libraries
acting for them, are permitted to make fair use of the material, such as to copy a chapter for use
in teaching or research. Permission is granted to quote brief passages from this publication in
reviews, provided the customary acknowledgment of the source is given.
Republication, systematic copying, or multiple reproduction of any material in this publication
is permitted only under license from the American Mathematical Society. Requests for such
permission should be addressed to the Acquisitions Department, American Mathematical Society,
201 Charles Street, Providence, Rhode Island USA. Requests can also be made by
e-mail to reprint-permission@ams. org.
2002 by the American Mathematical Society. All rights reserved.
The American Mathematical Society retains all rights
except those granted to the United States Government.
Printed in the United States of America.
The paper used in this book is acid-free and falls within the guidelines
established to ensure permanence and durability.
Visit the AMS home page at http://www.anis.org/
10987654321 070605040302
To Alla, Julia, and Jane
YAA
To Bernadette, Claire, and Dionissi

CDA
Contents
Foreword ix
Chapter 1. Odds and Ends 1
1.1. Banach Spaces, Operators, and Linear Functionals 1.
1.2. Banach Lattices and Positive Operators 13

1.3. Bases in Banach Spaces 27
1.4. Ultrapowers of Banach Spaces 37
1.5. Vector-valued Functions 44
1.6. Fundamentals of Measure Theory 52
Chapter 2. Basic Operator Theory 69

2.1. Bounded Below Operators 69
2.2. The Ascent and Descent of an Operator 79
2.3. Banach Lattices with Order Continuous Norms 84
2.4. Compact and Weakly Compact Positive Operators 88
Chapter 3. Operators on and AM-spaces 93

3.1. AL- and AM-spaces 94
3.2. Complex B anach Lattices 103
3.3. The Center of a Banach Lattice 112
3.4. The Predual of a Principal Ideal 119
Chapter 4. Special Classes of Operators 123

4.1. Finite-rank Operators 123
4.2. Multiplication Operators 135
V
- Contents
4.3. Lattice and Algebraic Homomorphisms 142

4.4. Fredhoim Operators 155
4.5. Strictly Singular Operators 170
Chapter 5. Integral Operators 179

5.1. The Basics of Integral Operators 180
5.2. Abstract Integral Operators 193
5.3. Conditional Expectations and Positive Projections 211
Positive Projections and 228
Chapter Spectral-Properties 237

6.1. The Spectrum of an Operator 237
6.2. Special Points of the Spectrum 248
6.3. The Resolvent of a Positive Operator 253
6.4. Functional Cakulus 256
Chapter 7. Some Special Spectra 271

7.1. The Spectrum of a Compact Operator 272
7.2. Turning Approximate Eigenvalues into Eigenvalues 281
7.3. The Spectrum of a Lattice Homomorphism 287
7.4. The Order Spectrum of an Order Bounded Operator 291
7.5. The Essential Spectrum of a Bounded Operator 298
Chapter 8. Positive Matrices 315

8.1. The Banach Lattices and 316
8.2. Operators on Finite Dimensional Spaces 320
8.3. Matrices with Non-negative Entries 329
Irreducible Matrices 332
8.5. The PerronFrobenius Theorem 339
Chapter 9. Irreducible Operators

9.1. Irreducible and Expanding Operators 347
9.2. Ideal Irreducibility and the Spectral Radius 357
9.3. Band Irreducibility and the Spectral Radius 365
9.4. Krein Operators on 370
Chapter 10. Invariant Subspaces 381
10.1. A Smorgasbord of Invariant Subspaces 383
10.2. The Lomonosov Invariant Subspace Theorem 393

Contents vii
10.3. Invariant Ideals for Positive Operators 398

10.4. Invariant Subspaces of Families of Positive Operators 409
10.5. Operators 425
10.6. Positive Operators on Banach Spaces with Bases 436
10.7. A Characterization of Algebras 440
10.8. Comments on the Invariant Subspace Problem 449
Chapter 11. The Daugavet Equation 455

11.1. The Daugavet Equation and Uniform Convexity 456
11.2. The Daugavet Property in AL- and AM-spaces 467
1L3. The Daugavet Property in Banach Spaces 471
11.4. The Daugavet Property in 477
11.5. Slices and the Daugavet Property 487
11.6. Narrow Operators 493
11.7. Some Applications of the Daugavet Equation 500
Bibliography 505
Index 521
Foreword
Since the appearance of the seminal monograph of Stefan Banach [51] on

(linear) operators between normed spaces, the study of operators has been
at the heart of research in analysis. For the most part, this research was
confined to the topological structures of the spaces and to the topological
properties of the operators. At the same time there was one more aspect of
operator theory that was developed in parallel with the topological and alge-
braic study of operators. This natural aspect involved the order structures
of the spaces between which the operators acted.
Almost every classical Banach space is equipped with a natural order
that is compatible with the algebraic and topological structures of the space.
Operators that preserve the order structures are known as positive operators.
They were first introduced and studied by F. Riesz [278, 279], H.
thal [129], and L. V. Kantorovich [172, 173] in the mid 1930's. For some
peculiar reason, the investigations of the topological and order properties of
operators were not connected, and for quite a while it seemed that they had
little in common. The monographs written by L. V. Kantorovich, B. Z. Vu-
likh and A. Pinsker [1751 in the Soviet Union and by H. Nakano [251, 252]
in Japan contained most of the theoretical background on operators between
ordered spaces accumulated by 1950. A partial list of books and monographs
that deal with Banach spaces and operators between them without empha-
sizing in depth the order structures includes: [56, 85, 102, 105, 106, 113,
118, 149, 152, 199, 205, 235, 250, 266, 300, 305, 306, 317]. For
a glimpse of the modern state of the art in Banach space theory and op-
erator theory, see the special volume [164} edited by W. B. Johnson and
J. Lindenstrauss.
ix
- - -
Foreword
From the early 1960's it was realized that the topological and order
structures of operators were related and that they should be studied to-
gether. Since then much research has aimed at producing a theory that
would encompass these structures in a unified manner. The "unification"
of the order and topological properties of operators is due to the efforts of
many mathematicians in various countries. The period of this unification
can be roughly divided into two parts; the pre-1980 period and the post4 980
period. Let us describe here the basic achievements done in each period.
The pre-1980 period can be described as the period where most of the
foundations of the order structure of vectors spaces and the lattice proper-
ties of operators were laid down. In this era the theory of partially ordered
vector spaces was developed through the efforts of mathematicians from
many countries. There were several monographs written on the subject.
Those written by A. L. Peressini [264] and G. Jameson [161] and the last
part of H. H. Schaefer's book [289] were devoted entirely to ordered vec-
tor spaces. The books by W. A. J. Luxemburg and A. C. Zaanen [2253,
B. Z. Vulikh [328], and S. Kaplan [176] studied the order structure of Riesz
spaces, while D. H. Fremlin [128] and C. D. Aliprantis and 0.
shaw [29] studied topological properties of Riesz spaces. Monographs on
Banach lattices and positive operators were written by M. A. Krasnoselsky
et al. {191], H. J. Krieger [194], H. E. Lacey [197], J. Lindenstrauss and
L. Tzafriri [205, 206], H. H. Schaefer [290], and H.-U. Schwarz [298].
The post-1980 period is characterized by the study of positive operators
from all points of view. Special attention was paid to the properties of op-
erators dominated by positive operators having some compactness property.
We also find in this era a complete study of integral operators in terms of
their lattice structure. Monographs regarding positive operators were writ-
ten in this period by C. D. Aliprantis and 0. Burkinshaw [30], W. Arendt et
al. [41], A. V. Bukhvalov et al. [78], P. Meyer-Nieberg [240], and A. C. Za-
anen [343, 344]. The list of very recent monographs on the subject of
positive operators includes the books by Y. A. Abrarnovich, E. L.
son, and A. K. Kitover [21], Y. A. Abramovich and A. K. Kitover [22],
A. G. Kusraev [195], and W. Wnuk [338].
Besides its important internal role in mathematics, the theory of Riesz
spaces and positive operators has important applications to several disci-
plines. The areas where the theory of Riesz spaces and positive operators
has been found to be useful include: game theory [182]; finance [151, Chap-
ter 31]; economics [26]; nuclear reactor theory [64, 65]; statistical decision
theory [201]; and structured population dynamics [80, 248].
The objective of this book is to present at the graduate level in a self-
contained manner the theory of linear operators between Banach spaces
Foreword
and Banach lattices by exploiting their topological and order structures, as

well as the topological and order properties of the spaces of operators. The
subject matter of the book is divided into eleven chapters, which to some
extent can be read independently. Each chapter is subdivided into sections.
At the end of each section there is a list of exercises of varying degrees of
difficulty designed to help the reader comprehend the material in the section.
There are more than six hundred exercises in the book. Hints to selected
exercises are also given. Our companion book Problems in Operator Theory
contains complete solutions to all exercises.
Exercises play a considerable role in our presentation. They serve several
purposes. First of all, we often relegate some technical details of proofs
to the exercises. These details may not be critical for understanding the
proofs but they are certainly essential for their validity, and we want to
provide the student with an accurate and complete account of how such
details should be presented. Second, these exercises not only allow the
student to learn how the results presented in the book work out but they also
offer a considerable amount of additional material and further developments.
There is one more feature to these exercises. There are many useful, almost
classical, results whose proofs are not readily available in the literature (for
instance, the duality between the uniformly convex and uniformly smooth
spaces, the equivalence between the differentiability and other properties of
the norm, the explicit closed form of the resolvent of the Volterra operator,
the description of the spectrum of the averaging operator f i+ f d,u,
etc.). The reader will find detailed solutions and discussions of many of
these results among our problems.
Some of the exercises are new, and some are known. Whenever possible,
we indicate the sources of the known exercises. On the other hand, for the
standard results or for some variations of classical results, we often omit the
sources of their origins.
The prerequisites of the book are the usual graduate introductory courses
in real analysis, general topology, measure theory, and functional analysis.
to make the book as self-contained as possible, we have provided de-
tailed proofs to most theorems in the text. This, coupled with the inclusion
of the exercises, makes the book an ideal text for a graduate or advance
course in operator theory and functional analysis. Moreover, since many
chapters contain the latest results (some of which have not been published
before), we hope that the book will be useful not only to graduate students
but also to research mathematicians in operator theory and functional
ysis, as well as to scientists in other disciplines. Each of Chapters 5, 7, 9,
10, and 11 can be used for an advanced seminar and each of them will also
present the students with possibilities and directions to start on independent
Foreword
research. A brief description of the material in each chapter should give a

good idea of the topics covered in the book.
Chapter 1 presents an introduction to Banach spaces and opera-

tor theory. Also, it introduces Banach lattices, positive operators,
and functions. The chapter culminates with a very
brief but, nevertheless, rather comprehensive presentation of the
fundamentals of measure theory.
Chapter 2 reviews the basic structural properties of operators with
special emphasis on positive operators. It also studies operators
bounded from below and discusses the ascent and descent of an
operator.
o Chapter 3 deals with AM- and AL-spaces and demonstrates their
importance in operator theory. It also studies the center of a Ba-
nach lattice and presents a careful exposition of the complexifica-
tion of a Banach lattice.
o Chapter 4 presents detailed accounts regarding the following classes
of operators: finite-rank operators, multiplication operators, lattice
and algebraic Fredholm and strictly
singular operators.
o Chapter 5 dea's exclusively with integral operators. It presents
their basic properties from the measure-theoretic and abstract
points of viewS It also studies conditional expectation operators
and demonstrates their connection with positive projections.
Chapter 6 introduces the spectrum of a bounded operator and
cusses the basic geometrical properties of the special parts of the
spectrum. The study specializes in the spectral properties of
itive operators. A comprehensive study of the functional calculus
is also here.
o Chapter 7 describes the spectra of compact operators and lattice
homomorphisms. It also contains the method of turning approx-
imate eigenvalues to eigenvalues. In this chapter intro-
duces and studies the order spectrum of an order bounded operator
as well as the essential spectrum of a bounded operator.
o Chapter 8 specializes in the study of operators to finite dimensional
spaces with special emphasis on matrices with non-negative entries.
It studies irreducible matrices and presents a thorough discussion
on the classical PerronFrobenius theorem.
Foreword - xiii
0 Chapter 9 investigates irreducible operators on Banach lattices and

presents several conditions for the strict positivity of the spectral
radius. This chapter also studies Krein operators between
spaces.
0 Chapter 10 deals with the famous invariant subspace problem. It
starts with several classical invariant subspace theorems, but its
central topic is the invariant subspace problem for positive and
related operators. The results obtained for individual operators
are then generalized to collections of positive operators.
0 A bounded operator T on a Banach space satisfies the Daugavet
equation if 1 + IITM. Chapter 11 presents a comprehen.-
sive study of operators on Banach spaces that satisfy the Daugavet

equation.
We regret that the size of the book has precluded us from including a
number of relevant topics, in particular, disjointness preserving operators
and their spectral properties, interpolation of positive operators, and the
domination problem. There is one more omission that we would like to
mention. The book deals with the general theory of operators and, as a
result, Hubert space operators occur in our discussions only sporadically.
There are many excellent books devoted exclusively to Hilbert spaces and
operators on them, and so anyone interested in this particular subject will be
able to find it easily in the literature; see for instance the monographs [94,
148]. At present, the realm of operator theory is so vast that any book can
cover only a portion of it.
In writing the book we have received help and benefited from discussions
with many colleagues and friends. We are grateful and express our sincer-
est thanks to all of them: 0. Burkinshaw (for many fruitful conversations),
W. A. J. Luxemburg (for his support-and-encouragement), H. Radjavi (for
insightful comments on the invariant subspace problem), B. Randrianan-
toanina (for reading the book, correcting a number of mistakes, and pro-
viding us with many valuable suggestions), C. Sirotkin (for a number of
useful remarks and problems), T. Oikhberg (for contributing several prob-
lems), K. Podgorski (for discussions concerning conditional expectations),
S. Klimek and E. Mukhin (for helping us to incorporate graphics in LATEX),
S. I. Gelfand (for constant support at all stages during the production of this
book), the AMS Copy Editor Arlene O'Sean (for her professional reading of
the manuscript and her many recommendations for its improvement), and
two anonymous referees (for their constructive criticism and comments).
Our special thanks go to A. K. Kitover and V. C. Troitsky. They read
the entire book, corrected mistakes, modified the old and provided some new
proofs, and suggested many interesting improvements. We are also deeply
xiv Foreword
indebted to W. Geller who read most of the book and, apart from fixing
some mathematical irregularities, corrected a great number of irregularities
in our English. Without his help our writing would have been considerably
less idiomatic.
Finally, we would like to thank our wives, Alla and Bernadette, for
coping with our idiosyncrasies during the many years it took us to assemble
the material, write, and typeset this text. To accomplish this project without
their patience and understanding would not have been possible.
Y. A. Abramovich and C. D. Aliprantis

Indianapolis and West Lafayette, May 2002
Chapter 1
Odds and Ends
The material included in this chapter is reflected correctly in its title. We

introduce here some basic concepts related to Banach spaces and state a few
principal theorems of functional analysis that will be used throughout the
book. These concepts include linear operators, linear functionals, and bases
in Banach spaces.
Special attention is given to the complexification of real Banach spaces.
Section 1.5 is devoted 'to vector-valued functions and it will serve as the
background needed to study the spectral properties of operators on Banach
spaces and Banach lattices. Section 1.6 presents a brief but nevertheless
complete introduction to measure theory. Some of this material will be
necessary for a comprehensive investigation of integral operators undertaken
in Chapter 5. Measure theory is studied in conjunction with the measure-
theoretic properties of the Riesz space of all measurable functions defined
on a measure space.
We present complete proofs to many results and omit mainly the proofs
of several well-known ones. As usual, the exercises at the end of the sections
contain a wealth of supplemental material that is of independent interest in
its own right.
1.1. Banach Spaces, Operators, and Linear Functionals

Here we briefly state a few basic definitions and facts regarding normed
spaces and operators that will be used repeatedly throughout this mono-
graph. Although we shall include many proofs, for more details and exten-
sive treatments of the material in this chapter we refer the reader to the
books [30, 31, 118, 205, 317].
= 1
2 - 1. Odds and Ends
A function II X : on a (real or complex') vector space is said to

be a norm whenever it satisfies the following properties:
(1) (Positivity) lxii 0 for all x E X and = 0 if and only if
x=0.
(2) (Homogeneity) f(axlI = lxii for all x E X and all scalars
(3) (The Triangle Inequality) lix + yM lixU + y
a vector space equipped with a norm. To avoid trivi-
alities, we shall assume throughout this monograph that all normed spaces
are non-trivial, i.e., different from the trivia' vector space {0}.
Every norm on a vector space X induces a metric on X via the
formula d(x, A Banach space is a normed space which is a
y) = lix
complete metric space under the metric induced by its norm.
A mapping T: X Y between two vector spaces (over the same field)
is said to be a linear operator, or simply an operator, whenever
T(ax + /3y) = + 13T(y)
holds for all x, y E X and all scalars For simplicity, the value T(x) will
often be denoted by Tx. If X and Y are normed spaces and T: X Y is
an operator, then its (operator) norm is defined by
I!TIH= sup
T is called a bounded operator, and when IIT!1 = 00,

the operator T is called an unbounded operator. It easily follows that if
T is bounded, then
IITM = min{M 0: <MIIxJ( for each x E x}.
Also, it is easy to check that an operator T: X Y between normed spaces
is bounded if and only if T is continuous.
An operator T: X Y between two normed spaces is called an
try (or, more precisely, a linear isometry) if ijTxII = Fix!) for each x E X.
Two normed spaces are said to be isometric if there exists a surjective
linear isometry from one space onto the other.
The collection of all bounded operators from a normed space X into
another normed space Y will be denoted by (X, Y). We shall write L(X)
instead of (X,X). It turns out that (X,Y) with the usual algebraic
operations and with the operator norm is a normed space. As a matter of
fact, we have the following.
'The symbols R and C will be used exclusively to denote the set of real and complex numbers,
respectively. Also, as usual, N will denote the set of natural numbers, Q the set of rational numbers,
and Z the set of integers.
1.1. Banach Spaces, Operators, and Linear Functionals 3
Theorem 1.1. If X is a normed space and Y is a Banach space, then the

vector space (X, Y) equipped with the operator norm is a Banach space.
Recall that an algebra is a vector space A that is equipped with an
extra binary operation (x, y) xy (called multiplication) that satisfies
for all x, y, z e A and all scalars a the following properties:
(1) (xy)z = x(yz) and (ax)y = =
(2) x(y+z) = xy+xz and (x+y)z = xz +yz.
An algebra A is called unital if there exists a vector c (which is uniquely
determined and iscalled the unit .of A) such that xc cx x holds. for all
xEA.
A Banach algebra is an algebra A which is also a Banach space under
a norm that satisfies JJxyIf lxii 'ftyIf for all x,y E A.
Theorem 1.2. If X is a Banach space, then (X) with the composition

operation as multiplication is a unital Banach algebra with the identity op-
erator as its unit.
For a real complex) normed space X, the Banach space (X,R)
(resp. (X, C)) will be denoted by X* and will be referred to as the norm
dual (or simply the dual) of X. The Banach space = (X*)* is known
as the second (or double) dual of X. Considering every x e X as a linear
functional on X* via the formula x(x*) = x*(x), we get a natural isometric
embedding of X into i.e., X can be viewed as a subspace of X**.
As usual, we say that a sequence in a Banach space X converges
weakly to some vector x e X (in symbols, or if
(x) holds for each e X*. Similarly, a sequence {4} in

the dual X* of a Banach space X weak* converges to some e (in
symbols, x E X.
If X is a normed space, then the closed unit balls of X, X*, and
will be denoted by U, and U**, respectively. That is,
U= {x e X: lixil <i}, = {x* e X*: IIx*If i}, and
= {x** X**: IIx**M i}.
Ifclarity requires the normed space X to be indicated, then the above closed
unit balls will be denoted by and respectively. It follows
that for each we have
= sup x*(x)I = sup (x*,x).
xEU
Also, for each x e X we have
lxii = sup = sup (x*, x).
x*EU*
1. Odds arid Ends
The adjoint of a bounded operator T: X Y between two arbitrary

normed spaces is the bounded operator T*: + X* defined via the for-S
mula (T*y*)(x) = y*(Tx) for E and x E X. It is customary to write
the formula defining T* by means of the duality expression
(T*y*,x)
}T**
The adjoint of T*, i.e., the operator T**: defined by
(T**x**,y*) = (x**,T*y*)
for all E X** and all E is called the double adjoint of the
operator T. it is not difficult to see that [ITCI = IIT* = Indeed,
= sup = sup [sup (T*y*,x)]
= sup [
sup (y*Tx)]
IIxIl1
sup = 11TH.
IxLl1
From now on, unless otherwise stated, X and Y will denote two arbitrary
normed spaces. A subset A of (X, Y) is said to be pointwise bounded
if for each x E X there exists some > 0 (depending upon x) satisfying
for eachTEA.
Theorem 1.3 (The Uniform Boundedness Principle). Assume that X is a
Banach space and Y is a normed space. Then a subset of (X, Y) is norm
bounded if and only if it is pointviise bounded.
The Uniform Boundedness Principle is also known in the literature as

the BanachSteinhaus theorem. For bounded operators between Banach
spaces there are two more important results that will be stated next. They
too were discovered by S. Banach and are referred to as the "open mapping"
and "closed graph" theorems.
Theorem 1.4 (The Open Mapping Theorem). Any surjective operator
tween two Banach spaces is an open mapping.
Recall that the graph of a function f: A + B is the subset C1 of

the Cartesian product A x B defined by C1 = {(a,f(a)): a E A}. If
T: X + Y is an operator between vector spaces, then its graph CT is a

vector subspace of X x Y. If are normed spaces, then their
Cartesian product X1 x .. x is also a normed space under the norm
(x1,. ..,
= This normed space is also called the direct
1.1. Banach Spaces, Operators, and Linear Fun ctionals 5
sum normed space and is denoted X1 If X1, . , are all

Banach spaces, then X1 is likewise a Banach space.2
Theorem 1.5 (The Closed Graph Theorem). An operator between two Ba-
nach spaces is bounded if and only if it has a closed graph.
Corollary 1.6. (X, p) and (Y, q) be two normed spaces arid let there
Let
exist two complete norms and on X arid Y, respectively, such that
p and q. Then each (p, q)-continuous operator from X to Y is
also
Proof. Assume that T: X + Y

a (p,q)-continuous operator. To show
is
that this operator is (Ps, it suffices to show that its graph

is closed. To this end, take any sequence in X such that + 0
and qy y) 0 for some y E Y. Since dominates p, it follows that

p(Xn) and since T is (p, q)-continuous, we have
0, 0. Similarly,
y) + 0 implies y) + 0, and consequently we see that
y = 0. That is, the operator T: (Y, has a closed graph, and

hence it is continuous.
We close the section with a discussion of the complexification of a real

vector space. If X is a real vector space, then the complexification of X
is the complex vector space
x,yEX},
whose vector space operations are defined by
(x1 + zyi) + (x2 + iy2) = x1 + X2 + + Y2), and
=
The real vector space X itself is usually identified with the (real) subspace
X +z{0} of Xc. That is, by identifying every vector x E X with x + zO E
X vectors of
X a normed space with norm then we can extend the
norm to a norm on Xc via the formula
lizil = sup
OE[O,2ir]
for each z = x + zy E Clearly, lixU = lix + for each x E X. Notice

that
+ MyII) IizIl lxii +
2 other frequently used equivalent norms on the direct sum vector space X1
are 1I(xi,. = and j(xi,. = lxiii.
simplicity we write x cos 0 + y sin 0 instead of the standard expression (cos O)x + (sin O)y.
1. Odds and Ends
for all z = x + E Consequently, = x + zy in Xc if and

only if x and y in X. This implies that if X is a Banach space,
then is also a Banach space.
Every operator T: X Y between two real vector spaces gives rise
naturally to a complex linear operator Tc: Xc Yc defined via the formula
Tc(x+2y) =Tx+2Ty.
Lemma 1.7. If X and Y are norrned spaces and T: X Y is a bounded
operator, then the operator Tc: Yc is also bounded and satisfies
IITCM = ITH.
Proof. First note that the inequality ITch IT]! is obvious. On the other
hand, for each Z = x + zy and any angle 9 we have
(Tx) cos 9 + (Ty) sin = IIT(x cos 9 + y sin 0)
IITII(IlxcosO+ ysin9jl)
<
This implies ITCZI IIZII or ITch So, ITch TI.
In the sequel, quite often, we shall identify T with without any men-
tion. When we do so the identification will be clear to the reader from the
context.
It is very common to denote the vectors x + of as column vectors
[x].
With this notation in mind, it is not difficult to see that every operator
T: can be represented by a unique operator matrix of the form
T T-S where 8, T: X X are linear operators, and as usual
= T
[x]
T(x + 2y)
= =
= Tx - Sy + 2(Sx + Ty).
Notice that in this notation, if T: X X is an operator, then the operator
is represented by the matrix
TO
Any operator T: for which S = 0 is called a real operator
on Of course, an operator T: is real if and only if T = for
some operator T: X X. Also, it should be noticed that for an operator
T: to be real it is necessary and sufficient that T leaves X
invariant, i.e., T(X) c x.
Thus, every operator T E can be identified with a matrix and
also with a pair of operators T, S E (X, Y). Equivalently, every operator T
1.1. Banach Spaces, Operators, and Linear Functionals - 7
in can be identified with the vector T+ iS in (X, Y) ezL(X, Y).

Moreover, the action of T + iS on Xc can be computed formally (and
mnemonically!) by the usual multiplication procedure on complex numbers.
That is,
(T+zS)(x+zy) = (TxSy)+z(Sx+Ty).
Thus, we have established the following result.
Theorem 1.8. If X and Y are two real normed spaces, then
L(X, Y) zL(X, Y),
where every T + iS E (X, acts on via the formula

(T+zS)(x+zy) = (TxSy)+z(Sx+Ty).
If, in addition, Z is another real normed space, T1 + iS1 e
and T + iS E Ye), then the operator (Ti + zSi)(T + iS) E
is given by
(T1 + zSi)(T + iS) = (T1T S1S) +z(T1S + SiT).
There is one more important connection between complex vector spaces
and real vector spaces. To discuss this connection, let X be a complex vector
space. Then X can also be considered as a vector space over the field of real
numbers R. We shall denote this real vector space by
X p C be a linear functional. Then for each x E X we can
write
f(x) =g(x)+zh(x),
where g(x) and h(x) are real numbers. An easy argument shows that the
functions g, h: p R are in fact both linear functionals. From the ho-
mogeneity of f we zf(x) or g(zx) + zh(zx)
have f(zx) = h(x) + zg(x),
and so h(x) = g(zx). Thus, we have shown that if f: X C is a linear
functional, then there exists a unique linear functional g: + JR (denoted
Ref, i.e., g = Ref, and called the real part of f) such that
f(x) g(x) zg(zx) (*)
holds for each x E X. Conversely, if g: p R is a linear functional, then

by means of (*) the functional g gives rise to a linear functional f: X p C
such that Ref g.
Now assume that X is a complex normed space and that f: X C is
a continuous linear functional. Write f(x) = g(x) zg(zx) and note that if
x E X satisfies 1, then from the inequality
lixil
it follows that = On the other hand, if x X

satisfieslFxII 1, then write f(x) = and note that
- = =
1. Odds and Ends
Therefore, = f(x)[ and 50 11111 ugh. That is, we have

shown that each f E = liRef
satisfies MI M
The preceding discussion shows that the mapping f Ref, from

to (XR)*, is a linear isometry. In other words, we have established the
following result.
Theorem 1.9. If X is a complex normed space, then the mapping I 'f Ref
is a (real) surjective linear isometryand so, up to this linear isometry, we
can write (X*)R = (XR)*.
Using the previous result, we can also establish the following.
Theorem 1.10 (BohnenblustHahnBanach). Every continuous linear

functional defined on a subspace of a complex normed vector space X has a
continuous linear extension to all of X with preservation of its norm.
Recall that if V is then the quotient

a subspace of a vector space W,
vector space W/V is the vector space {[wJ: w e W} of all equivalence
classes, where as usual for each w E W its equivalence class is defined by
[w] = w + V = {w + v: v e V}. The addition and multiplication in W/V

are given by
[x] + [y] = [x + y} and )4xJ =
The mapping w [wJ is called the quotient map of W onto
The following result summarizes some basic properties of quotient

normed spaces. Its proof is left to the reader.
Theorem 1.11. Let Y be a closed subspace of a normed space X, and define
the function ifi. ifi. X/Y f R by
H1[xIIU = d(x,Y) inf lix

yEY
- = inf
zE[x]
IIzIL
where d(x, Y) is the distance from x to Y. Then:
(a) The function II

is a norm on the quotient vector space X/Y
(called the quotient norm).
(b) The set = {[x]: x e is the unit ball of X/Y under the
quotient norm.
(c) The quotient map is an open contraction, it is an open mapping

whose norm is at most one.
(d) If X is a Banach space, then (X/Y, I I' ifi) is also a Banach space.
If V is a vector subspace of a normed space X, then its annihilator is

the norm closed subspace of X* defined by
V' = {x* E X*: x*(v) = 0 for all v E V}.
In terms of annihilators the norm duals of a subspace and of a quotient

space are described as follows.
Theorem 1.12. For a closed subspace Y of a normed space X we have
(X/Y)*=Y and
subject to the following surjective linear isometries.

(1) The mapping Y1 (X/Y)*, defined via the formula
= x*(x), e Y', [xJ E X/Y,
is a well-defined surjective linear isometry.
(2) The mapping X*/Y defined via the formula
= (the restriction of to Y)
for each [x*J e X*/Y, is a well-defined surjective linear isometry.
Exercises
1. If two vectors x and y in a normed space satisfy Ix + = lxii +
then show that = aiIxlI+1311y11 for all scalars a,/3 0. [HINT:
If the scalars a and /3 satisfy a /3, then note that
IIax+13y11 = lIa(x+y) + a)yM .1
2. Let T: V W be a surjective one-to-one operator between two vector

spaces. Show that T' (the inverse of T) is a linear operator.
3. If T: X Y is a bounded operator between normed spaces, then show

that
= min{M 0: IITxM for all x E X}.
4. Prove Theorem 1.1.
5. Prove the following converse of Theorem 1.1. If X is a normed
space, Y is another normed space and (X, Y) is a Banach space, then
Y is likewise a Banach space.
6. Show that the complexification X under the
operations
(x1 + + (x2 + zY2) = X1 + + z(y1 + y2), and
(a + i/3)(x + zy) = ax /3y + z(/3x + ay)
is a complex vector space.
7. Let X be a real normed space with norm Show that the function
R defined via the formula
= sup lix cos 0 + y sin z = x + zy E
OE[O,21r1
1. and Ends
is a norm on that extends Also, show that
+ +
IIZlIc
for each z = x + zy E Use the preceding inequality to verify that:

(a) A sequence {Zn}, where + 2Yn, satisfies 0 if and If
only if 0 and IYnll 0.

(b) If X is a Banach space, then is also a Banach space.
8. Let X be a real vector space with complexification X zX. Estab-

lish the following properties.
(a) If V is a vector subspace of X, then = VEBzV = {x+zy: x, y E V}
is a vector subspace of xc.
(b) If X is also a normed space and V is a vector subspace of X, then
the vector V is closed in if and only if
V is a closed subspace of X.
(c) Let T: X X be an operator. A subspace V of X is said to be
invariant under T (or simply T-invariant) whenever T(V) C V.
A vector subspace V of X is T-invariant if and only if the vector
subspace V zV of xc is Ta-invariant.
9. Let X be a real vector space. It is very common to denote the vectors
x + zy of as column vectors Under this notation the vector space

operations of are given by
= [Xl+X2 X =
+ and
Y1 Y2 LY' + Y2 Y LI3X +
Now assume that X and Y are two real vector spaces. Show that
a mapping T: p is a linear operator if and only if there exist
two (uniquely determined) operators 5, T: X Y such that T has the
matrix representation T
FT-S where as usual
= [s T
T TxSy
S x
T(x + zy)
T
Sx + Ty y=
Also, show that if X and Y are normed spaces, then T is a bounded
operator if and only if both S and T are bounded operators.
10. A bounded operator T: X Y between normed spaces is said to be
invertible if there exists (a uniquely determined) bounded operator
5: Y X satisfying ST = 1x and TS = ly; the operator S is called the
inverse of T and is denoted T1.
For an operator T: X Y between real vector spaces establish the
following properties.
(a) The operator T is surjective if and only if Tc is surjective.
(b) The operator T is one-to-one if and only if
T is invertible if and only if Tc is invertible. Also,
show that if T is invertible, then (Ta)' =
11. Let X = Xi ... Xn be the direct sum Banach space. Show that
a mapping A: X p X is a bounded linear operator if and only if there
exist bounded linear operators X3 (i,j = 1, 2,. .. , n) (uniquely

determined) such that A has the matrix representation A = where
for each x x1 we have
A11 A12 ... A17., x1
A21 A22 x2
Ax= .
e e' e
=
12. Recall that a (not necessarily surjective) operator T: X f Y between
normed spaces is said to be an isometry (or more precisely a linear
isometry) if IITXII = holds for each x E X.
Show that a bounded operator T: X * Y between two normed spaces
is an isometry if and only if its double adjoint T**: y** is likewise
an isometry. [HINT: Assume that T: X Y is an isometry. Clearly,
= IIT**M = 1 and so IIT**x**M for all E X**. Now fix
X* with 1. Then the formula y*(Tx) = defines a
continuous linear functional on the range of T such that 1. Let
y* also denote a continuous linear extension of to all of Y such that
<1; clearly T*y* = Now note that for each we have
= I(T*y*,x**)I =
IT**x**II
and so Mx**II = <

13. Let T: X X be a bounded operator on a normed vector space. A glance
at the inequality T
T a unique (linear) continuous extension
T: X X, where X denotes the norm completion of X. Establish the
following properties:
(a) If T is an isometry on X, then so is T on X.
(b) If T is an invertible isometry 8, thh S is also a (linear)
isometry.
(c) If T is invertible, then so is T, and we have (11)_i =
(d) If T is invertible on X for some scalar then 1' is likewise
invertible on
(e) If is an invertible isometry on X, is T necessarily an invertible
isometry on
[HINT: For (e) let X denote the vector space of all polynomials defined
on [0, 1] equipped with the sup norm. Clearly, X = C[0, 1]. Now consider
the operator T: X X defined by Tp(t) = p(t2) for all p X and all
t E [0, 1]. Note that T is a non-invertible isometry. However, it should be
clear that C[0, 1] C[0, 11 satisfies = x(t2) for all x E C[0, 1]
and each t [0, 1], which is an invertible isometry on
41n this context, as usual, the scalar stands for the operator Al.
1. Odds and Ends
14. Show that for bounded operators X Y Z between normed spaces

(all over the same field) we have 11 1811 Use this inequality
to show that:
(a) The map (8, T) 8T, from (Y, Z) x (X, Y) to (X, Z), is jointly
norm continuous.
(b) The Banach space (U, V) of all bounded operators from a normed
space U to a Banach space V is a unital Banach algebra.
15. Let X and Y be two normed spaces over the same field. Show that the
mapping T T*, from into (Y*,X*), is a linear isometry.
17. Let A be a non-empty subset of a normed space X and let x0 E X. We
say that there is a nearest (or a closest) point to in A if there exists
some a0 E A such that aoM iso all for all a E A. Establish the
following.
(a) The norm function s 'p is weakly lower semicontinuous,
in X implies linainf
(b) If X is a reflexive Banach space and A is a nonempty weakly closed
subset of X, then for every point so E X there exists a nearest point
to x0 in A.
(c) Assume that the norm of X is uniformly convex, that is, for each
0 < 2 there exists some 0 <8 < 1 such that whenever s,y E
satisfy Is , then < 1 8. A normed space is called
II
uniformly convex if its norm is uniformly convex. (It is easy to

verify that Hilbert spaces are uniformly convex.)
If A is a non-empty, convex, and closed subset of a uniformly convex
Banach space X, then for every x0 E X there exists exactly one point
in A that is nearest to xo. In other words, there exists a mapping
7r: X A defined by
='
The mapping is called the nearest point mapping from X to A.
(d) Let C be a non-empty closed convex subset of a uniformly convex
Banach space X. Show that the nearest point mapping 7r is contin-
uous.
18. For a non-empty closed convex subset C of a Hilbert space X establish
the following properties of the nearest point mapping 7r: X + C.
(a) If s E C, then 71(5) = s.
(b) If s C, then 7r(s)E 5C.
(c) ForeachxEXandeachyECwehave
(d) For each x,y E X we have (7r(x) 'ir(y),x y) 0.
(e) For all s,y E X we have 1171(x) ir(v)1I S
(f) If C is a closed vector subspace of X, then for each s E X the vector
x 71(5) is orthogonal to C, i.e., (x 'ir(x), c) =0 for all c E C.
Remark: Recall that a non-empty subset A of a topological space is
called a retract of if there exists a continuous mapping f: A,

Banach Lattices and Positive Operators 13
called a retraction of onto A, such that f(a) = a for each a E A.

The preceding properties show that ir is a retraction of X onto C.
19. (Support Points). In this problem X will denote a real Banach space,
H is a real Hilbert space, and C is a non-empty closed convex subset of
H. The function 'ir: H p C is the nearest point mapping as defined in
Exercise 17 above. Recall that a point a0 in a subset A of X is called
a support point of A if there exists some non-zero xK E X* such that
x*(ao) <x*(a) for each a E A, i.e., if attains its minimum value on A
at Establish the following.
(a) Only boundary points of a set can be support points.
(b) If a closed convex subset of X has a non-empty interior, then every
boundary point of the set is a support point.
(c) A point c0 E c9C is a support point of C if and only if there exists
some x C such that ir(x) = c0.
(d) If H = then every boundary point of C is a support point.
JR/C,
(e) (BishopPhelps [66]) The set of support points of C is dense in the

boundary 8C of C.5
(f) Give an example of a non-empty closed convex subset of 2 for which
not every boundary point is a support point.
20. (The Finite Dimensional Separation Theorem). Recall that a non-zero
linear functional E separates two non-empty subsets A and B of
a Banach space X if x*(a) for all a E A and all b E B.
Show that every pair of disjoint non-empty convex subsets of JRk can
always be separated. [HINT: Assume that A and B are two non-empty
disjoint convex subsets of IRk. Now consider the non-empty convex set
C = A B and note that 0 C. If 0 C, then = ir(0) separates A
JRk
and B, where ir is the nearest point mapping from onto C. If 0 E C,
then verify that 0 E 8C.]
1.2. Banach Lattices and Positive Operators

In this section, we review very briefly a few basic properties of ordered
vector spaces and positive operators. For omitted proofs and details we
refer the reader to the books [29, 30, 225, 240, 290, 343]. Recall that
a reflexive, antisymmetric, and transitive relation on a set is known as an
order relation. That is, a binary relation on a set is an order relation
whenever it satisfies the properties:
(1) (Reflexivity) x x for each x.

(2) (Antisymmetry) x y and y x imply x = y.
famous result of Bishop and R. R. Phelps is true for any real Banach space; for a
complete discussion of the BishopPhelps theorem see [25, Section 8.9]. V. I. Lomonosov [213]
has proven recently that the BishopPhelps theorem fails for complex Banach spaces.
1. Odds and Ends
(3) (Transitivity) x y and y z imply x z.
As usual, the symbol y x is equivalent to x y. The notation x> y (or

y < x) means x y and x y. Any set equipped with an order relation is
called a partially ordered set.
A vector space order on a real vector space X is an order relation
on X that is compatible with the algebraic structure of X in the sense that
it satisfies the following two extra properties.
(a) Ifxy, z EX.
(b) If x y, then ax ay for each scalar a 0.
A partially ordered vector space (or simply an ordered vector space)
is a real vector space equipped with a vector space order.
If X is an ordered vector space, then the set X+ = {x E X: x 0} is
referred to as the positive cone, or simply the cone, of X. The cone X+
satisfies the following properties:
(i)
(ii) a 0.
(iii) X+ (X+) = {O}.
Any subset C a real vector space X that satisfies properties (i), (ii), and
of
(iii) above is referred to as a (convex) cone of X. The following important
relationship between cones and vector space orders is true: If C is a cone
in a real vector space X, then the binary relation defined by letting x y
whenever x y E C is a vector space order on X such that X+ =
An ordered vector space X has a generating cone if X = X+,
i.e., if every vector can be written as a difference of two positive vectors (or
equivalently, if for each x E X there exists some y E satisfying y x).
Definition 1.13. An operatorT: X Y between two ordered vector spaces

is called:
(1) positive (in symbols, T 0 or 0 T) if x 0 implies Tx 0;
(2) strictly positive if x> 0 implies Tx> 0; and
(3) regular if T can be written as a difference of two positive operators.
The collection of all regular operators between two ordered vector spaces
X and Y will be denoted Y). The reader should notice that under
the pointwise algebraic operations and under the natural order generated by
the cone of positive operators the set 4 (X, Y) is, in fact, an ordered vector
space.
A subset A of a partially ordered vector space X is said to be order
bounded if there exist x, y E X such that x a y holds for each a E A.
Definition 1.14. An operatorT: X Y between two ordered vector spaces

is called order bounded if T carries order bounded subsets of X to order
bounded subsets of Y.
Clearly, every positive operator (and hence every regular operator) is
order bounded. However, an order bounded operator need not be regular;
see [30, Example 1.11, p. 10]. The collection of all order bounded operators
from X into Y will be denoted Lb (X, Y). Under the pointwise algebraic
operations and pointwise order, Lb(X, Y) is an ordered vector space that
includes the vector space of all regular operators, i.e, Lr(X, Y) Lb(X, Y).
The. positive operators are characterized by means of their.. additivity
property on the positive cone.
Theorem 1.15 (Kantorovich). Let X and Y be two ordered vector spaces
and let T: be an additive mapping, i.e., T(x +y) = Tx +Ty holds
for all x, y in X+. If the cone of X is generating and Y is Archimedean,6
then T extends uniquely to a positive operator from X to Y. Denoting this
unique linear extension by T again, we have
Tx = Tx1 Tx2,
where x x1 x2 is any representation of x as a difference of two positive
vectors.
A lattice ordered vector space E is called a Riesz space (or a vector

lattice). That is, an ordered vector space E is a Riesz space if every pair
of vectors has a least upper bound (supremum) and a greatest lower
bound (infimum). Following the standard lattice notation, the supremum
and infimum of a pair of vectors {x, y} will be denoted by x V y and x A y,
respectively. That is,
xVy=sup{x,y} and xAy=inf{x,y}.
For an element x in a Riesz space, its positive part, its negative part,
and its absolute value are defined by
x V 0, = (x) V 0 and lxi = x V (x),
respectively. We have the following two important identities:
x x and lxi = x+ + x . (*)
The functions (x, y) 'k x V y, (x,y) x A y, x x+, x x, and x 'k lxi
are referred to collectively as the lattice operations of a Riesz space. From
(*), it has a generating cone.
follows that every Riesz space
A Riesz space is said to be Dedekind (or order) complete whenever

every non-empty subset of the space that is bounded from above has a least
6An ordered vector space Y is said to be Archimedean if nx y for all m implies x 0.

16 1. Odds and Ends
upper bound. Similarly, a Riesz space is said to be o-Dedekind (or o-

order) complete whenever every non-empty countable subset of the space
that is bounded from above has a least upper bounds
When E and F are Riesz spaces with F Dedekind complete, an operator
T: E > F is order bounded if and only if it is a regular operator. The details
of this important fact are included in the next result whose proof can be
found in [30, p.
Theorem 1.16 (F. RieszKantorovich). If E arid F are two Riesz spaces
with F Dedekind complete, then the ordered vector space r(E, F) is a
Dedekind complete Riesz space satisfying r(E, F) = 4(E, F). Moreover,
its lattice operations are given by
(1)
T(x) = sup{Ty: 0 y x}, and
(3) T1(x) = sup{Ty: x y
for all T E r(E, F) and all x E
The operator ITt = T V (.T) is known as the modulus of the opera-
tor T. Formulas (1), (2), and (3) in Theorem 1.16 are called the Riesz
Kantorovich formulas. Two operators S, T E r(E, F) are said to be
disjoint if S A T =0.
When E is also Dedekind complete, there is a useful modification of
the RieszKantorovich formulas in terms of "components." Recall that if x
is a positive vector in a Riesz space, then any positive vector y satisfying
y A (x y) = 0 is referred to as a component of x.
Theorem 1.17 (Abramovich [1}). If T: E > F is an order bounded oper-
ator between Dedekind complete Riesz spaces, then
(1)
T(x) = y E E and y A (x y) = and
and yA(xy)=0}
hold true for each x E E+.
A net {Xa}aEA in a partially ordered set is said to be decreasing (in

symbols, Xa if ai a2 implies Xa2 x means
and x = The symbols Xa I and x have analogous
meanings.
Definition 1.18. A net {Xa}aEA in a partially ordered set is said to be
order convergent to an element x, in symbols, Xa > x, whenever there
exist two nets {Yfi} flED and such that:
.L2. Banach Lattices and Positive Operators 17
(1) x and j. x.
(2) For each fi E B and 'y E r there exists some ao E A (depending on
fi and 'y) satisfying z7 for all a ao.
The element x is called the order limit of the net
The following observation regarding order limits will be useful. If in the
definition of x we consider the product index set A = B x r and
define the two nets and by
and
for each = (fi, E A, then we can replace the nets and

by and respectively. That is, we can assume in the defi-
nition of order convergence that the index sets B and r are the same.
It should be clear that if x x in a partially ordered set, then
x. We list below two basic properties of order limits.
o A net in a partially ordered set can have at most one order limit.
o If a net in a partially ordered set satisfies (resp.
and x (resp. x).
(See Exercise 4 at the end of the section.) The next result deals with order
convergence in Riesz spaces.
Lemma 1.19. For a net in a Riesz space E we have:

(a) If there exists a net (with the same index set) such that
x.
0 and xI for each a E A, then
(b) If E is Dedekind complete and is order bounded, then
p---> x in E if and only if there exists a net (with the
same index set) such that 0 and xI for each a E A.
Proof. (a) Assume that there exists a net satisfying 0 and

I
x} for each a E A. Let = x and =x+ Then it
should be clear that x, x, and holds for each a E A.
Therefore, x.
Suppose that E is Dedekind complete,

(b) is order bounded,
and that Pick two nets and such that y,9
x, and for each fi B and 'y E r there exists some a(fi, E A so that
z7 for each a a(fi,y). For each a E A, let
= inf and = sup
The above suprema and infima exist since the net is order bounded
and E is Dedekind complete. Clearly, and Moreover, since for
18 1. Odds and Ends
each E B and E F we have for all a' a(1@, '-y), it follows

that for all a a(/3, 'y). So, if v0, v, then yp v holds
forall/3EBand7Er. Similarly,
x. To finish the proof notice that I xI c Wa v0, for each a and
Since order convergent sequences in Riesz spaces are automatically order

bounded, the sequential analogue of Lemma 1.19 can be stated as follows.
Lemma 1.20. In a cr-Dedekind complete Riesz space a sequence is
order convergent to a vector x if and only if there exists a sequence
such that 0 and for each n E N.
Historically, the conclusion of Lemma 1.20 is taken as the definition of
order convergence of sequences. To be consistent with the literature, we also
follow this tradition. That is, for sequences we define order convergence as
follows.
Definition 1.21. A sequence in a partially ordered set is said to be
order convergent to some element x, in symbols, x, whenever there
exist two sequence {Yn} and {zn} such that I x, x, and

for aim.
For Riesz spaces we have the following:
A sequence in aRiesz space is order convergent to a vector
x if and only if there exists a sequence such that Un 0 and
According to Lemma 1.20, for cr-Dedekind complete Riesz spaces this defi-
nition is, of course, equivalent to the general definition of order convergence
of nets as given in Definition 1.18.
Definition subset D of a partially ordered set is said to be order
1.22. A
closed (reap. cr-order closed) if D and xE D
(reap. D and Xn x imply x E D).
We now turn our attention to order continuous mappings.

Definition 1.23. A mapping f: X Y between partially ordered sets is:
(1) cr.order continuous if Xn x in X implies f(x) in Y,

(2) order continuous if x in X implies f(x) in Y.
Simple examples of order continuous functions are provided by the lattice
operations of a Riesz space. For instance, in any Riesz space Xa x implies
and lxi. (See Exercise 4 at the end of the section.)
______________-
The proof of the following characterization of order continuity of positive

operators is left as an exercise.
Lemma 1.24. A positive operator T: X Y between two ordered vector
spaces is order continuous if and only if 0 in X implies Tx0, 0 Y.
Similarly, a positive operator T: X Y is cr-order continuous if and
only if 0 in X implies 0 in Y.
The order continuity of order bounded operators is characterized as fol-

lows.
Theorem 1.25. For an order bounded operator.T: E F between Riesz
spaces with F Dedekind complete the following statements are equivalent.
(1) T is order continuous.
(2) If mE, then in F.
(3) If Xa .J. 0 in F, then infa ITXaJ = 0 in F.
(4) T+ and T are both order continuous.
(5) Ti is order continuous.
Proof. (1) (2) If 0, then 0, and so TXa 0.
(2) Assume that

(3) 0 and let some y E F satisfy y for
each a. Since TXa 0, there exists a net {uA}AEA such that 0 and
for each A E A there exists some ao satisfying for each a a0.
This implies y and so y uA for each A E A. From uA 0,
we get y 0. This shows that infa = 0 in F.
(3) It suffices to establish that
(4) is order continuous. To this
end, assume that Xa 0 in E. Let 0 u T+Xa for each a. To show that
T+ is order continuous, we must prove that u = 0.
Fix some index ao, and let x = Clearly, for each 0 y x and all
a and 0 for each 0 y x

yA 0, it follows from our hypothesis that [T(y A = 0.
Now a glance at (*) yields 0 u Ty for each 0 y x. Taking
into consideration the RieszKantorovich formula T+x = Ty, the
latter shows that u 0, as desired.
(4) (5) Obvious.
20 1. Odds and Ends
(5) (1) Let 0 in E. Pick two nets and such

that I x, x, and for each E B and E r there
exists some
'y) e A satisfying for each It is easy to see that
holds for all a '-y).
If we put A = B x rand for each ,\ = (,B,'-y), we let UA = 0,

then I
(TI(uA) holds for all a(,i3, 'y) = The order continuity
of implies ITKuA) 0, and so 0 in F. The proof is finished.
The corresponding result for a-order continuous operators is as follows.

Theorem 1.26. For an order bounded operator T: E p F between Riesz
spaces with F Dedekind complete the following statements are equivalent.
(1) T is crorder continuous.

(2) If 0 in E, then >0 in F.
(3) If 0 in E, then in F.
(4) T+ and T are both cr-order continuous.
(5) is o-order continuous.
Let E be a Riesz space. A subset A of E is said to be solid if

and y E A imply x E A. A solid vector subspace of E is called an ideal. The
ideal generated by a non-empty subset A of E is the smallest (with
respect to inclusion) ideal containing A; it coincides with the intersection of
all ideals that contain A and is given by
= E x1,. . . E A and xi,. . , E with xji}.
IfA = {x}, then = is called the principal ideal generated by the

vector x. Note that
{y e E: there exists ,\ 0 such that }.
The principal ideals will play an important role in this book. A vector e> 0
is called an order unit (or a strong unit or simply a unit) if Ee = E, i.e.,
if for each x E X there exists some 0 such that <,\e.
An order closed ideal is called a band. Since a solid subset D in a Riesz
space E is order closed if and only if D fl E imply
x E D (see Exercise 5 at the end of the section), it follows that an ideal A
IximplyxeA.
If D is a subset of a Riesz space E, then the band generated
by D is the smallest (with respect to inclusion) band BD that contains D.
If D = {x}, then = is called the principal band generated by the
vector x. The band generated by an ideal is described as follows.
Theorem 1.27. If J is an ideal in a Riesz space E, then the band Bj

generated by J is given by
Bj={XEE:
Moreover, for each x E E the principal band generated by x is given by
{yEE: I y(}.
A vector e > 0 is said to be a weak unit jf Be = E, i.e., if x A ne x
holds for each x E E+. If E is Archimedean, then a vector e > 0 is a weak
unit if and only if x 0. Clearly, every order unit is a
weak unit.
The disjoint complement of a non-empty subset A of E is defined by
Ad={XEE: xIAJal=OforeachaEA}.
The order continuity of the lattice operations guarantees that every disjoint
complement is a band. In Archimedean Riesz spaces every band is a disjoint
complement.
Theorem 1.28. In an Archimedean Riesz space, the band generated by a
non-empty set A coincides with (Ad)d, i.e., BA = In particular,
every band B in an Archimedean Riesz space satisfies B BI.
A band B in a Riesz space E is said to be a projection band whenever

E = B Bd holds. Clearly, each projection band B gives rise to a natural
projection7 with range B which is referred to as a band projection. In a
Dedekind complete Riesz space every band is a projection band.
The collection of all order continuous regular operators from an ordered
vector space X into another ordered vector space Y will be denoted by
(X, Y). Similarly, the collection of all o-order continuous regular oper
ators will be denoted by Y). Both and are vec-
tor subspaces of the vector space of all regular operators r(X, Y), and
c
Theorem 1.29 (Ogasawara). If E and F are Riesz spaces with F Dedekind
complete, then F) and F) are both bands in 4(E, F). In par-
ticular, we have the decompositions
r(E,F) = =
If E is a Riesz space, then the Dedekind complete Riesz space 4 (E, IR)
is called the order dual of E and is denoted i.e., E" = 4(E, IR). The
band = n(E, TR) is referred to as the order continuous dual of E. If
separates the points of E, then E is called a normal Riesz space.
7RecaII that a linear operator P: V V on a vector space is called a projection if P2 = P.
22 - 1. Odds and Ends
Recall also that a Riesz space E is said to have the countable sup
property if for every subset D of E having a supremum there exists an at
most countable subset C of D satisfying sup C = sup D. The countable sup
property allows us to work with sequences rather than nets. For example,
if E satisfies the countable sup property, then F) = F).
A seminorm p on a Riesz space is said to be a lattice seminorm when-
ever implies p(x) p(y). A normed Riesz space is a Riesz space
equipped with a lattice norm. A normed Riesz space which is also norm
complete (i.e., a Banach space) is called a l3anach lattice. In a normed
Riesz space, the lattice operations are uniformly continuous functions. In
particular, this implies that every band is a closed subspace. Every band
projection on a normed Riesz space is always a contraction. We also have
the following simple but very useful result.
Lemma 1.30. In a normed Riesz space P2:
(a) The cone E+ is norm (and hence weakly) closed.
(b) The order intervals are norm (and hence weakly) closed.
(c) If a net E satisfies and 0, then x.
Our next result informs us that positive operators between Banach lat-
tices are continuous; for a proof see [30, p. 175].
Theorem 1.31. Every order bounded operator from a Banach lattice to a
normed Riesz space is continuous (and hence every positive and every regular
operator on a Banach lattice is also continuous)
Now consider two Banach lattices E and F with F Dedekind complete.
For a regular operator T: E F its r-norm is defined by ITlir = ITt 1
Note that we have
IiTIir = inf{II Sit: T S }.
This formula allows us to define the r-norm for regular operators with values
in a non-Dedekind complete Banach lattice. For the proof of the next result
see [30, Theorem 15.2, p. 248].
Theorem 1.32. If E and F are two Banach lattices with F Dedekind com-
plete, then the Dedekind complete Riesz space 4(E, F) of all regular oper-
ators from E to F equipped with the r-norm is a l3anach lattice.
Sinceevery continuous linear functional on a normed Riesz space E is

automatically order bounded, it follows that E* In fact, it follows
that the operator norm coincides with the r-norm on see Exercise 20
at the end of the section. In other words, we have the following important
result.
Theorem 1.33. E is a normed Riesz space, then, its norm dual E* is a

If
Banach lattice and its lattice operations are given by
(x*)+(x) = sup{x*(y):
(x*)_(x) = sup{_x*(y): 0 y x}, and
z*}(x) = sup{x*(y): IyIx}
for each xE E* an ideal in the order

dual E a Banach lattice, then E* =
We our discussion by introducing some additional terminology.

A positive operator T: E p F between two Riesz spaces is said to be a
lattice (or a Riesz) homomorphism if T(x V y) = Tx V Ty for all x and y
in E. A one-to-one lattice isomorphism is known as a lattice (or a Riesz)
isomorphism. A lattice isomorphism T between two Banach lattices is
said to be a lattice isometry if llTxll = for each x E E. Two Banach
lattices are called lattice isometric if there exists a lattice isometry from
one onto the other.
The most useful characterizations of lattice homomorphisms are included
in the next result.
Theorem 1.34. For a positive operator T: E F between two Riesz spaces
the following statements are equivalent.
(a) T is a lattice homomorphism.
(b) = for each x E E.
(c) T(x A y) = Tx A Ty for all x, y E E.
(d) [TzJ = for each x E E.
(e) xAy=0 mE implies TxATy=Oin F.
Proof. (a) (b) We have
T(x V 0) T(x) V T(0) = T(x) V 0 = (Tx)t

(b) (c) Using the identity x A y = y (y we get
T(xAy) = T(y) = T(y) -
= T(y) - [T(y) - T(x) A T(y).
(c) (d) The identity lx y * 2(z A y) yields:
= T(x)-f-T(y)--2T(xAy)
= T(x) + T(y) 2[T(x) A T(y)] = IT(x) T(y)I.
Letting y = 0 gives IT(x)1 = for all x E E.
24 1. Odds and Ends
(d) (e) Let x A y = 0. The identity u A v = + v ju vJ) implies
T(x) A T(y) = [T(x) + T(y) IT(x) T(y)(]

= +T(y) T(IxyI)]
= IxyD) = T(xAy) = T(0) = 0.
(e) (a) From (xxAy)A(yxAy) = 0, we get T(xAy) = T(x)AT(y).

Now use the identity x V y = x + y x A y to obtain
T(xVy) = T(x)+T(y)T(xAy)
= T(x) + T(y) [T(x) A T(y)] = T(x) V T(y).
This completes the proof.
A positive operator T: E + E is said to be interval preserving if

T[0, x] = [0, Tx] for each x E E+. There are some nice duality properties
between interval preserving operators and lattice homomorphisms. They
are included in the next result whose proof can be found in [30, p. 92]; see
also Exercise 11 in Section
Theorem 1.35 (Kim [183]; AndLotz [216]). For an arbitrary positive

operator T: E + F between two Banach lattices we have the following.
(1) If T is interval preserving, then its adjoint T*: F* E* is a lattice

homomorphism.
(2) T is a lattice homomorphism if and only if T* is interval preserving.
Building on the work and ideas of D. Maharam [228, 229, 230]

on integral representations of linear operators, W. A. J. Luxemburg and
B. de Pagter [222] have also established the following remarkable result
regarding extensions of positive operators to interval preserving operators.
Theorem 1.36 (Luxemburgde Pagter). Assume that E is an Archimedean

Riesz space, F is a Dedekind complete Riesz space, and T: E + F is a
positive operator. Then there exists a Dedekind complete Riesz space L
containingE as a Riesz subspace and an order continuous interval preserving
operator T: L + F that extends T.
Exercises
1. Let C be a cone in a real vector space X. Show that the relation on
X defined by w y whenever s y E C is a vector space order such that
x+C.
3. Show that order complete Riesz spaces and normed Riesz spaces are Ar-
chimedean.
4. Establish the following properties of order convergence.
(a) A net in a partially ordered set can have at most one order limit.
(b) If 5a s or Xa s in a partially ordered set, then Xa S.
(c) If a net {Sa}aEA in a partially ordered set satisfies Wa (resp. 5a .1.)
and 5a x, then 5a x (resp. j x).
(d) A positive operator T: X Y between two ordered vector spaces

is order continuous if Wa 0 in X implies TXa .j. 0 Y.
(e) A net {Sa}aEA in a Riesz space satisfies 5a x if and only if there
exists another net {UA}AEA such that uA 0 and for each A E A there
exists some (depending on A) such that tWa u), holds for
each a
(f) The lattice operations in any Riesz space are order continuous.
5. Show that a solid subset D of a Riesz space E is order closed if and only
if {Sa} C Dfl E+ and Is in E imply SE D.

8. Show that a vector e> 0 in an Archimedean Riesz space is a weak unit
if and only if A e = 0 implies s = 0.
9. Show that a positive operator T: X Y between two partially ordered
vector spaces is order continuous if and only if 5a 0 in X implies TSa .j. 0
mY..
10. Show the following:
(a) If is a compact topological space, then in C(ffl each component
of the constant function 1 is of the form Xv for some clopen set V.
(b) If p is a finite measure, then in L1 (p) each component of the constant
function 1 is of the form XA for some measurable set A.
11. If E and F are Riesz spaces with F Dedekind complete, then show that
the lattice operations of any two operators 5, T E b(E, F) are given by
(SVT)(s) = sup{Sy+Tz: and y+z=s}, and
(SAT)(w) = inf{Sy+Tz: and y+z=s}
for each s E E+.
12. Prove Lemma 1.30.
13. Show that if p w holds in a Banach lattice, then there exist a subse-
quence of and some u 0 satisfying for each
1. Odds and Ends
m. [HINT: Pick a subsequence {yn} of {Xn} satisfying IYn for

each m and let u =

15. Show that for a given Riesz space E there is, up to an equivalence, at
most one lattice norm which makes E a Banach lattice. [HINT: Assume
that a Riesz space E is a Banach lattice under the lattice norms and
II
Then the identity operator I: (E, II 1k) f (E, II 112) is positive,
and hence (by Theorem 1.31) it must be a homeomorphism.]
16. Show that a normed Riesz space E is a Banach lattice if and only if every
increasing norm Cauchy sequence of E+ is norm convergent. [HINT:
Assume that the condition is satisfied and let {Xn} be a norm Cauchy
sequence in E. Without loss of generality, we can suppose that lXn+1
for each n. Now consider the sequences {Yn } and {zn } of E+
defined by
= and Zn
and note that both sequences are increasing norm Cauchy sequences and
Xn+1 = Zn.]
18. Give an example of a regular operator T: E + F between two Banach
lattices with F Dedekind complete satisfying 11Th <
19. Show that a band projection P on a normed Riesz space is a contraction
IPII 1).
20. If E is a normed Riesz space, then show that II = = I II
holds for each E*. [HINT: From
fx*(x)h x*h(ixi) 1
x*t 11 lIxil
we infer that IIx*11 1
On the other hand, if Ixjh I and e > 0
then from x*i(hxh) = sup{lx*(y)h: lxi we see that there
exists some y E E with lxi such that Ix*t(hxh) <
Therefore,
II
Ix*H1 =sup{Ix*1(hxt): lixil 1} tx*1I+E,
for each e > 0. This implies II
21. A Dedekind complete Riesz space E is said to be a Dedekind

tion of a Riesz space E if there exists a lattice isomorphism ir: E + E
such that
= sup{ir(v):
= inf{ir(w): vi E and 'ir(w) }
for each E E. We can identify E with ir(E) so that E can be viewed
as a Riesz subspace of E. It should be clear that only Archimedean
Riesz spaces can have Dedekind completions. As a matter of fact, every
Bases in Banach Spaces 27
Archimedean Riesz space E has a unique (up to a lattice isomorphism)

Dedekind completion for details see [225, pp. 185196].
Now assume that E is a Banach lattice (and so E is an Archimedean
Riesz space) and let be its Dedekind completion. On E5 define the
norm
x x x E E and that (Es, II. I)is a Banach

lattice. [HINT: We shall denote . II by . again. For the norm
completeness of we use Exercise 16 of this section. Assume that a
sequence in satisfies 0 I and < for all
72; For each n pick some E E+ such that 0 and-
< < If we let then is a norm
Cauchy sequence in E and so {Xn} is norm convergent to x =
in E. Clearly, x in E. In particular, if we let z72 = E
then 0. From
= j_i)
(here we assume = 0 and x0 Ii ), we see that < x for each

n. Since is Dedekind complete, there exists some such that
I . Next, note that for each n and p we have
n+p1 n+p1
So, by letting p oo, we get 0 Zn for each n. This implies

lix IZn and from this we get 0.]
22. (Nakano [251]) Let E be a Riesz space and let f The null ideal
of the functional f is the ideal N1 = {x E E: IfKIxD = 0}. The band
C1 = N is called the carrier of f.
Assume that E is Archimedean and f, g E Show that A gI = 0
if fl
and only if C1 = {0}. [HINT: We can assume that 0 f, g E
Suppose that f A g = 0. Pick 0 x E C1 = N and fix some > 0.
Next, pick a sequence [0, xl such that + g(x Sn)
holds for each n. If = then note that Yn 0. This implies
o g(x Yn) = xi)) g(x < . Therefore,
o g(x) e holds for each e> 0 or g(x) 0. Thus, C1 c Ng and so
C1flCg{O}.J
1.3. Bases in Banach Spaces

In this section X will denote a Banach space. Our objective is to review a
few basic properties of bases in Banach spaces. For extensive treatments of
bases in Banach spaces, we refer the reader to [205], [305], and [306]. For
an excellent quick review of bases see 11601.
28 1. Odds and Ends
A sequence in X is said to be a Schauder basis (or simply a

basis) in X if for every x E X there exists a unique sequence of scalars
{ such that x has the series representation x = where the
series is assumed to converge in the norm. The uniqueness of the sequence
guarantees that x,., 0 for each n and that the vectors Xi, x2, .. are
linearly independent. Every Banach space X with a basis is automatically
separable. Indeed, if we let
D={XEX: suchthat
then D is a countable set that is dense in X. This means that only sepa-
rable Banach spaces can possibly have bases. The basis problem was the
following long standing question: Does every separable Banach space have a
basis? In 1972 P. Enflo [120] produced a separable Banach space without
a basis, and so he provided a negative answer to the basis problem.
Now let be a basis in a Banach space X, and for each vector
x E X let x = be its series representation. The
functional of the basis is the linear functional defined by
= =
It should be clear that is indeed a linear functional. Also, there exists

a sequence of natural projections on X that are associated with the
given basis The range of the projection X p X is the vector
subspace generated by the set {xi, X2,. .. , and is defined by
= =
The projections and the coordinate functionals are all continuous.

Theorem 1.37. Assume that X is a Banach space with a basis and
for each vector x E X let x = = be its series
representation with respect to the basis Then:
(a) The function X R, defined by
IUxffl = sup
nEN i=i
is a norm that is equivalent to . (the norm of X).

(b) Each projection is contin'uous and .'c = <oo.
(c) Each coordinate functional is continuous and
Proof. (a) The verification that is a norm satisfying 111 U 11

is
trivial. Verifying the completeness of is cumbersome and is left as an
exercise.
(b) Since .
Ills equivalent to II there exists a constants C > 0 such
that IIxlII for each x E X. This implies
= HIXHI
and so is a continuous projection satisfying C for each ri.
fore, ic = C < 00.
(c) Observe that 1
= Furthermore, if n> 1, then
= = =
holds for each x E X. So, each is continuous and
Definition 1.38. If is a basis in a Banach space, then the positive

number ic = < oo is called the basis constant of If
Ic = 1, then is called a monotone basis.
Corollary 1.39. If is a basis in a Banach space, then there exists an
equivalent norm in which the basis is monotone.
Proof. The norm in Theorem 1.37 is the desired norm.
A basis is said to be normalized if = I for each ri. From

part (c) of Theorem 1.37 we have the following.
Corollary 1.40. If is a normalized basis in a Banach space, then
every coordinate functional of the basis satisfies
The next result presents a characterization of the sequences in a Banach

space that are bases. Its proof is left to the reader.
Theorem 1.41. A seq'aence of vectors in a Banach space X

is a basis if and only if it satisfies the following properties:
(1) The linear span of the set X2,.. .} is dense in X.
(2) There exists some K > 0 such that if m> ri and ai,. .. ,am are
arbitrary scalars, then
n m
1. Odds and Ends
A sequence of non-zero vectors that satisfies property (2) of The-

orem 1.41 is called a basic sequence. From Theorem 1.41 it should be
immediate that a basic sequence is a basis in the closed linear span of
Next, we shall prove that every Banach space has a basic sequence. To
establish this important result, we need a lemma due to S. Mazur.
Lemma L42 (Mazur). Let X be an infinite dimensional Banach space, let
Y be a finite dimensional vector subspace of X, and let 0. Then there
is a unit vector x e X such that (1 + + for all y E Y and
all scalars
Proof. Assume 0 < 1.It suffices to show that there exists some unit
vector x E X such that I (1 + ) y E Y with = 1.
Since the closed unit ball of Y is norm compact, there exist unit vectors
yi,. .
. , in Y such that for each unit vector y e Y there exists some i so
that < For each 1 i ri. choose some e X* satisfying
= I and X is infinite dimensional we can find
a unit vector x e X such that = 0 for each I i n., that is,
xe Ker Now if a is any scalar and y e Y satisfies = 1, then
choose some I <i <ri with YiM and note that
+ - + - =1-
Therefore, I (1 + + for all unit vectors y e Y, as desired. D
Theorem 1.43 (Banach). Every infinite dimensional Banach space has a

basic sequence.
Proof1 Let X be an infinite dimensional Banach space. Fix some real num-
ber 0 and then choose a sequence of positive real numbers such
that K = + I + . Pick any unit vector E X and then
use Mazur's Lemma 1.42 to select inductively a sequence of unit vectors
X2, X3,.. such that for each ri. we have
IzII < (1 + +
for all scalars and all z in the vector subspace generated by {xi,... ,
Now note that if i-n > n., then for any choice of scalars cvi,.. . , we have
n m
This shows that is a basic sequence in X.
Let be a basis in a Banach space X. Then = 8jj holds for

all i and j. For this reason the sequence of coordinate functionals is
L3. Bases in Banach Spaces 31
also called the sequence of biorthogonal functionals associated with the

basis Notice that if m > n and A1, . Am are arbitrary scalars, then
,
for each x = E X with lixM = 1 we have
K) =
= K) = m
m
m
m
lxii <
This implies and therefore is a basic

sequence in
The basic sequence need not be a basis in For instance, if
-is not separable, then the biorthogonal sequence cannot be a basis.
However, if X is reflexive, then is a basis in X.
We shall close the section with a discussion of unconditional bases. To
introduce the notion of an unconditional basis we need the concept of an
unconditionally convergent series.
Definition 1.44. A series in a Banach space X is said to
converge unconditionally if for every permutation8 ir of N the series
is norm convergent in X.
The proof of the next result that characterizes the unconditionally con-
vergent series is left to the reader.
Lemma 1.45. For a series in a Banach space the following state-
ments are equivalent.
(1) The series converges unconditionally.
(2) For any sequence of signs (i.e., 1 for each n) the series
Snxn is norm convergent.
(3) For every strictly increasing sequence of natural numbers the
series is norm convergent.
(4) For each E> 0 there exists a natural number k such that for each
finite subset A of N with mm A 2 k we have
xn a Banach space converges uncon-
ditionally, then xn = for each permutation ir of N.
that a permutation of a set is any one-to-one and surjective function it:

32 1. Odds and Ends
It is an immediate consequence of Lemma 1.45 that if a series

converges unconditionally, then for each choice of signs the series
likewise converges unconditionally.
We are now ready to introduce the notion of an unconditional basis.
Definition 1.47 (James [159]). A basis in a Banach space is said to
be unconditional if for each vector x its series expansion x = anun
converges unconditionally.
Now assume that is an unconditional basis in a Banach space X.

Let S = { 1, i}N be the collection of all signs. We consider { 1, 1} equipped
with the discrete-topology, and so S is a compact metric space.
Given an arbitrary sign sequence s = (81, s2,...) E 5, we define a
ping M5: X X by
=
where x= It should be clear thatis a linear operator. If
denotes the sequence of biorthogonal functionals with respect to the
basis and X X is the (bounded) finite-rank operator defined by
= then we have
=
By the Uniform Boundedness Principle, the operator is also bounded.
Next, we claim that for each fixed x = anum E X the mapping

s if from S to X, is continuous. To see this, assume that a sequence
{Sn} S satisfies s = (Si, 52,...) in S and let > 0.
Since the series smanum converges unconditionally, there exists some
k such that for any non-empty finite subset A of N that satisfies mm A k
we have < Next, pick some n0 > k such that =
holds for all n and all i = 1,. k. For each p let
. . ,
Ap={mEN: and
and note that for each n n0 we have
k+p
= < 2E.
ik
Letting p cx, we get for all n n0. So, for

fl no we have
M8(x)jj
______________________
This shows that the mapping s ' M5 (x) is continuous.

Since S is a compact metric space, it follows that for each x
the set of vectors {M5 (x): s e S} is norm boundedin fact, a compact
set. That is, the family of bounded operators s S} is pointwise
bounded. Consequently, by the Uniform Boundedness Principle, the family
{M5: s e S} is norm bounded.
Definition If is an unconditional basis in a Banach space, then
the positive number
= sup 1M511
sES
is called the unconditional constant of the basis
The following result reveals a fundamental property of an unconditional
basis.
Lemma 1.49. Assume that is an 'unconditional basis in a (real) Ba-
nath space with unconditional constant Let F be a non-empty finite
subset of N, let i F} be an arbitrary collection of scalars, and let
{'yj: i F} be another collection of scalars s'uch that 1 holds for each
i e F. Then we have
iEF iEF
Proof. Assume that F and the scalars cEj and 'yj are as in the statement
of the lemma. Choose a continuous linear functional E X* of norm one
satisfying (>iEF = Next, pick any sequence of
signs s (81, 82,...) such that for each i e F we have = 1 if 0
and = 1 if <0. This implies
c
iEF iEF
as desired. D
If we replace by then the conclusion of the previous lemma

remains true for complex Banach spaces as well. Indeed, letting 'yj = +
we have V <1, and hence
+
iEF iEF iEF iEF
Corollary 1.50. If is an unconditional basis in a Banach space X,
then there exists an equivalent norm in which the unconditional constant of
the basis is one.
34 1. Odds and Ends
Proof Define X =
[O,oo) by IlxMi lIMsxIl. It is easy to
verify that this formula defines a norm on X. Moreover, note that
lxiii
for each x E X. This shows that
are equivalent norms. To
and
finish the proof, notice that the unconditional constant of with respect
to the norm II 1k is one.
The unconditional bases are characterized as follows.
Theorem 1.51. For a basis in a Banach space X the following
rnents are equivalent.
(1) The basis is unconditional.
(2) For any permutation it of N the sequence is a basis.
(3) If for some sequence of scalars the series con-
verges in norm, then for any strictly increasing sequence of
natural numbers the series is norm convergent.
(4) If for some sequence of scalars the series con-
verges in norm and another sequence of scalars satisfies
I
for each n, then is norm convergent.
Proofs (1) (2) If (1) is true, then for any permutation 'it of N the
sequence is clearly an unconditional basis.
For the converse, assume that (2) is true and fix a permutation 'it of N.
Also, let and be the sequences of biorthogonal functionals with
respect to the bases and respectively. Since agrees with
on the vector subspace V generated by {u1, u2,. .} and V is norm dense
.
in X, it follows that holds for each n.

Now fix some x E X and let x = be its (unique) series
expansion with respect to the basis Also, let x = AnUir(n) be
the series expansion of x with respect to the basis Therefore,
S= = = =
This shows that is an unconditional basis.

(1) (3) This equivalence follows immediately from the equivalence
of (1) and (3) of Lemma 1.45.
(4) (1) Assume that (4) is true and let s (s1, be a se-
quence of signs. If x = then by letting = we have
= for each n, and so = is norm conver-
gent. This shows that is an unconditional basis.
1.3.Bases in Banach Spaces 35
To see that (1) implies (4) consider a norm convergent series

and let a sequence of scalars satisfy for each n. Put =
if 0 and = 0 if = 0. An easy application of Lemma L49 shows
that the series is norm convergent. To finish the proof note
that =
If a sequence of vectors in a Banach space is an unconditional basis
in the closed linear span of then this sequence is called an
tional basic sequence. It is easy to see that an analogue of Theorem 1.51
is true for unconditional basic sequences as well. It is important to note
that the analogue of Theorem L43 for unconditional basic sequences is not
true. That is, the existence of an unconditional basic sequence in a given
infinite dimensional Banach space was a long standing open problem. Only
recently W. T. Gowers and B. Maurey [13'T] have answered this problem in
the negative by constructing an infinite dimensional Banach space without
an unconditional basic sequence. For Banach lattices the situation is more
pleasant.
Lemma 1.52. Every infinite dimensional Banach lattice has an uncondi-
tional basic sequence.
Proof. Let E be an infinite dimensional Banach lattice. Take any sequence
{ of pairwise disjoint positive unit vectors in E. If and are any
scalars such that for each n, then for each 'in we have
in m m m
=
Therefore, if the series converges, then the series
converges too. Hence, in view of Theorem 1.51, is an unconditional
basic sequence.
For a special connection between unconditional bases and Banach lat-
tices see Exercise 2 in Section
Theorem 1.51 has a useful consequence regarding projections. If is
an unconditional basis in a Banach space X, then every non-empty subset
A of N gives rise to a bounded projection PA: X X defined by
PAX =
nEA
for each x = e X. If A = {1, 2,. n}, then PA is simply the
. ,
projection that was introduced at the beginning of the section. Now if

A is a non-empty subset of N and s E is the sequence of signs
defined by = I if n E A and = 1 if n E Ac, then it is easy to
see that PA = + M8), where I is the identity operator on X. This
36 1. Odds and Ends
implies + I) On the other hand, if 5 E {1, 1}N

is any sequence of signs and we put A = {n E N: 1}, then we have
M3 = PA PAC. The above imply the following inequalities.
Lemma 1.53. If is an unconditional basis in a Banach space, then
ft sup PAll 1
2sup IIPAIL
A A
where the suprema are taken over all nom-empty subsets A of N.
Exercises
1. Prove part (a) of Theorem 1.37
3. This exercise describes a basis for the Banach spaces 1] for I p <
CX). Consider the sequence of functions defined by h1(t) = I for each
t E [0, 1] and
( -i
i
I
) , 11
Ui - 2
(0 otherwise,
for k = 0,1,2,... and I 2/c. The sequence is known as the
Haar system. Establish the following.

(a) 0 for each m.
(b) The linear span of the set = {h1, h2,. . .} is norm dense in 1}.
(c) For any collection of scalars we have . . ,
fl.
(d) The Haar system is a monotone basis in each I].

4. This exercise describes a basis in the Banach space C[0, 1]. Let be
the Haar system that was introduced in the preceding exercise. Consider
the sequence of functions in C[0, I] defined by
s1 = 1 and d'r, for ri = 2,3,4

J0
The sequence is called the Schauder system. Show that is a
monotone basis for C[0,
5. (BessagaPelczynski [61]) Let be a sequence of vectors in an infinite
dimensional Banach space satisfying 0 and urn inf > 0. Show
that has a subsequence that is a basic sequence.
91t is well known that neither C[0, 1] nor L1 [0,1] has an unconditional basis; see Corol-
lazy 11.61.
.4. Ultrapowers of Bariach Spaces 37
6. Assume that X is a reflexive Banach space with a basis Show that

the biorthogonal sequence is a basis in X'
7. Let Sn be a series in a Banach space X.
(a) If X is finite dimensional, show that Sn converges uncondi-
tionally if and only if IIXnII
(b) Give an example of an unconditionally convergent series 5n
such that >12n=i IlSnIt = 00.
9. Prove Corollary 1.46.
10. Let be an unconditionally convergent series in a Banach space
X. If E then show that the series is norm
convergent. Moreover, show that the mapping T: p X, defined by
is a bounded linear operator.
1.4. Ultrapowers of Banach Spaces

In this section we shall discuss a technique that allows one to construct a
special enlargement of a Banach space that is useful in many contexts. To
do this, we need the concept of a filter.
Definition 1.54. Let be an infinite set. A non-empty collection of
subsets of is said to be afilterif:
(a)
(b) is closed under finite intersections, that is, A, B E implies
AnB
(c) is closed under supersets, i.e.,- A E and A B imply B E
Since a filter is a non-empty collection, it follows from (c) that E
and that every filter is closed under arbitrary unions.
A filter is said to be an ultrafilter if it is maximal with respect to
inclusion. That is, a filter is an ultrafilter if it is not properly included in
any other filter, i.e., if c c and is a filter, then g = It follows from
Zorn's lemma that every filter is included in an ultrafilter.
The proof of the next result that characterizes the ultrafilters among
filters is straightforward.
Lemma 1.55. For a filter the following statements are equivalent.
(1) is an ultrafilter.
(2) If A U B E F, then either A E or B E
38 1. Odds and Ends
(3) thenAEJ.
An immediate consequence of the identity A U AC = E and the
above result is the following.
Corollary 1.56. If is an ultrafilter, then for any subset A of either
AE or (exclusively) AC E
It is not difficult to see that if U is an ultrafilter and flAEU A 0, then

there exists a unique point 6 E such that
U={AcL\: 6EA}.
Any ultrafilter of the above form is called a fixed ultrafilter. An ultrafilter
U that is not fixed, i.e., flAu A = 0, is referred to as a free ultrafilter.
There is a special filter that plays a crucial role in the study of free
ultrafilters. This is the filter of all co-finite subsets of LeO, the filter
ACisafiniteset}.
A straightforward verification shows that is a filtercalled the Frchet
filter on & It turns out that every free ultrafilter includes the Frchet filter
I". To see this, let U be a free ultrafilter and let F be a finite subset of
If F = UXEF{x} E U, then it follows from Lemma 1.55 that for some x E F
we have {x} E U. This implies U = U a contradiction.
Consequently, c U for each free ultrafilter U.
The above conclusion yields the following properties for an ultrafilter U.
No finite subset of is a member of U.
By considering the collection of all filters that include the Frchet filter
(J) and using Zorn's lemma, we can easily conclude the following.
Lemma 1.57. Every infinite set admits a free ultrafilter.
For the rest of the discussion in this section U will denote a free ultrafilter
on N. By means of the ultrafilter U we now introduce a new notion of
SC(lUential convergence in a topological space.
Definition 1.58. A sequence in a topological space converges to
some x E along the ultrafilter U, in symbols, x x = limu
x the set {n E N: E V} belongs to U.
Thc point x is called a U-limit of the sequence
Here are a few properties of the convergence along an ultrafilter.

Lemma 1.59. For a sequence in a topological space 'i-) we have:

(1) If is Hausdorff, then can have at most one U-limit. That
is, in Hausdorff spaces the U-limits are uniquely determined.
(2) then
(3) If is metrizable and x, then there exists a subsequence
of such that -'-k x.
{
(4) If lies in a compact set K, then x x E K.
Proof. (1) Assume is Hausdorif, x x

and x V W = 0.
Now consider the sets
A={nEN: and B={nEN:
and note that A, B E U. This implies that 0 = A fl
B E U, which is
impossible. Hence, a sequence in a Hausdorif space can have at most one
limit along U.
(2) If V is a neighborhood of x, then there exists some rio such that
E V for all n > n0. Therefore, {1,2,.. . ,no}c {n EN: E V}. This
implies that the set {n E N: E V} belongs to U, and so limu x.
(3) Let d be a distance generating the topology 'r and assume that
For each n the set = {k E N: d(xk,X) < belongs to
the ultrafilter U. Clearly, for each n. Next, define the sequence
{ m4 of natural numbers recursively by
ml = mm A1 and = mm \ {mi, . . . , for n 1.

Note that for each n and that the subsequence {Xmn } of
satisfies d x) < for each n. In particular-c
(4) Let be a sequence in a compact set K. If does not converge
along U to any point in K, then for each y E K there exists a neighborhood
of y such that the set = {n E N: E Vj,} does not belong to U. Since
K is compact, there exist points Yl, Y2,. . in K such that K
,
This implies U must be true for at

least one i, which is impossible. This contradiction shows that the sequence
{ converges along U to some point in K.
It should be noted that a need not be a topological limit. As

a matter of fact, according to statement (4) of the preceding result, ev-
ery bounded sequence in has a unique U-limit. For instance, consider
the sequence of JR defined by = (_i)fl. From
{ 1, 3, 5,. .} U {2, 4, 6,
.
} = N, it follows that either { 1, 3, 5,.. } E U or
. . .
.
1. Odds and Ends
{2,4,6,...} EU. If {1,3,5,...} EU, then and if {2,4,6,...} EU,

then
In topological vector spaces the U4imits also satisfy the following
erties.
Lemma 1.60. If in a topological vector space
x and ,Ax for each scalar \.
Now we turn our attention to Banach spaces. Let X be a (real or

complex) Banach space. As usual, we denote by the vector space of
all bounded sequences in X. That is, x = (xi, x2,...) belongs to if
and only if is a bounded sequence. Under the sup norm
nEN
the vector space is a Banach space. If X is a Banach lattice, then

under the pointwise lattice operations (X) is likewise a Banach lattice.
We shall denote by the vector space of all sequences in that
converge to zero along the free ultrafilter U. That is,
= {x = (xi,x2,...) E
A glance at Lemma 1.60 guarantees that is a vector subspace of oo(X).

Moreover, we claim that is also a closed To see this, assume
that some x = (xi, x2,...) belongs to the closure of and let c > 0.
Choose some y = (Yi, Y2,...) E with lix <c. Since y E JVu, the
set A = {n E N: < c} belongs to U. Now use the inequalities
IIxn - ymIj + ltYmII jx - +
YIIoo <C + IlYmII
to conclude that A {n E N: <2c}. So, {n E N: <2c} E U.

This shows that x E .iVu. Thus, is a closed subspace.
It should be clear that when X is a Banach lattice, the closed subspace
.Nu is also an ideal. In particular, if X is a Banach lattice, then the quotient
Banach space is likewise a Banach lattice.
Definition 1.61. The quotient Banach space is called the ul
trapower of X with respect to the free ultrafilterU and is denoted i.e.,
Here is an important property of the quotient norm of an ultrapower.

As usual, for eath x = (Wi, x2,...) E we shall denote its equivalence
class in by [x].
Lemma 1.62. If x = (xi, x2, ...) E 4,0(X), then = limu

Proof. Since for each u = (Ui, u2,...) the sequence of real num-
bers is bounded (and so it lies in a compact subset of R), it follows
from Lemma 1.59(4) that limu exists in JR.
Now fix x (x1,x2,...) and put a = limu Let y [x].
From the fact that preserve inequalities
(see Exercise 5 at the end of the section), and y x it follows that
= [txnii] =
This implies limu IIyn $ limu IIxri By the symmetry of the situation,
limu limu also true, and so limu
is = limu a for
each y [x]. From we also get a = limu ttYLloo. Since
y {x] is arbitrary, we conclude that a
For the reverse inequality fix f > 0. It follows that the (infinite) set
B = {n N: a < < a + f} belongs to U. Now consider the
sequence y = ,c,(X) defined by Yn = if ri B and = 0
if n B. From the inclusion
Bc{nEN:
we see that {n N: = o} U. This implies x y and so
y {x]. Now notice that
= sup = sup = sup a + .

nEN nE.B nE.B
Since f> 0 is arbitrary, we get [x] a. Thus, {x} = a, and the proof is
finished.
There is a natural embedding of X into its ultrapowers. This is the

mapping x '4 [(x, x,. .)]. From Lemma 1.62, it easily follows that the
mapping x '4 [(x, x, . .)], from X to
.
is a linear isometrywhich is also
a lattice isometry if X is a Banach lattice. This means that if X is identified
with its image in then X becomes a closed vector subspace of In
other words, we can view the ultrapowers of X as enlargements of X. We
summarize the above observations in the following result.
Theorem 1.63. Any ultrapower of a Banach space X contains X as a

closed vector subspace. If X is also a Banach lattice, then any ultrapower of
X contains X as a closed vector sublattice.
The importance of ultrapowers to operator theory stems from the fol-

lowing result.
42 1. Odds and Ends
Theorem 1.64. If T: X p Y is a bounded operator between two Banach

spaces, then for each free ultrafilter U on N the mapping Xu
defined for each (xi, x2,...) E by
X2,. {(Txi, Tx2,...)],

is a bounded operator that extends T and satisfies = II TM.
In other words, the mapping T from (X, Y) to is a
linear isometry.
Proof. Assume T: X Y is a non-zero bounded operator. Let us verify
first that is well defined. To this end, fix x = (x1,x2,...) E and let
y E [x]. If 0, then the set A = {n E N: < th } belongs
to U. From the inclusion A {m E N: <}, follows that
{n E N: <c} E U. So, [(Txi, Tx2, . .
= [(Ty1, Ty2,. .)] in .
and consequently the mapping is well defined.

It should also be obvious that the mapping Xu is a linear
operator. Moreover, if x E X, then from ([(x, x,.. .)]) = [(Tx, Tx,. .)], .
we see that the operator T: X

is bounded, notice that according to Lemma 1.62 for
each x = (xi, x2,...) E we have
= = urn lirn IlxntI
= lirri =
This shows that is a bounded operator and that ITuII Now if
x E X satisfies <1, then
IITxII = J[(Tx,Tx,. . = ITuIf f[(x,x,. . . 1[TuJI
Hence, ITxJJ c fTuIJ is also true, and so = 1TII.
Regarding bounded operators between Banach lattices we have the foh

lowing result whose straightforward proof is left for the reader.
Theorem 1.65. For a bounded operator T: X p Y between two Banach
lattices the following statements are true.
(1) T is a positive operator if and only if
T is a lattice homomorphism if and only if is a lattice homo-
morphism.
(3) If T is interval preserving, then so is
Corollary 1.66. If X is a Banach space, then the mapping T ' from
(X) to is an algebraic homomorphism and a linear isometry which
is a positive operator if X is a Banach lattice.
I..I. Ultrapowers of Banach Spaces 43
i'toof. This conclusion follows immediately from Theorems 1.64 and 1.65
the (easily checked) fact that if X Y Z is a scheme of bounded
between Banach spaces, then = Sj,,'Tu.
Notice also that ('x)u = This implies that if T (X) is an
j
operator, then E is likewise an invertible operator and
(TuY'. D
Exercises

2. Show that if an ultrafilter U on a set satisfies flAEU A 0, then there
8EA}.
3. Show that every filter coincides with the intersection of all ultrafilters
that include it.
5. Let be a space, let A be a subset of and let be a
sequence in such that for some N E U we have E A for each n E N.
Show that if x E A.
Use this conclusion to show that the U-limits preserve inequalities
in the following sense: If in a Banach lattice x, y, and for
some N E U we have for each n E N, then x y.
6. For a non-zero bounded operator T: X Y between two Banach spaces
and a sequence in X establish the following.
(a)
If 0 in Y 0 in X, then there exist some 77 > 0
and a subsequence of satisfying 77 for each n
and 0.
7. Let Z be a normed space and let U be a free ultrafilter on N. If, as usual,
we let = { (z1, Z2,...) E o}, then is a closed vector
subspace of Z, and so = is a normed space. That is, in
the definition of the ultrapowers we do not need Z to be complete.
Suppose that X is a normed space, Y is a subspace of X, and U is
an ultrafilter on N. Also, let = {(yl,y2,...) E
and = {(x1,x27...) E and define the mapping
J: by
for each y = (yr, y2,...) E Show that J is a linear isometryand

so, under this isometry, Y7.j can be considered as a vector subspace (which
is norm closed if Y is a'so norm closed) of Xu.
44 1. Odds and Ends
9. For a bounded operator T: X Y between two Banach spaces establish

the following properties.
(a) If T is surjective, then Tu is surjective.
(b) If T has a closed range and T is surjective.
(c) The operator T is invertible if and only if is invertiblein which
case we have (Ta)' =
1.5. Vector-valued Functions

In this section, X and Y will denote arbitrary complex Banach spaces. We
shall review here a few basic properties of the X-valued functions that are
defined on subsets of the complex plane C. shall pay special attention
to the results that will be employed in this book. They are the analogues
of the results for complex-valued functions and their proofs can often be
deduced from the scalar casesfor more details regarding Banach space-
valued functions see [108]. We shall refer to the basic results about complex-
valued functions of one complex variable as "standard." These results can
be found in any book on complex analysis; see for instance [671.
We start our discussion with the fundamental properties of formal series
in complex Banach spaces.
Theorem 1.67. Let {ao, a1,. .} be a sequence of vectors in X, let A0 E C,
.
and consider the formal series
Also, let R = urn sup

1
, where we adhere to the conventions = oo and
1 = Then:
(1) The series converges uniformly in norm on every compact subset
of the opem disk DR(Ao)."
(2) The series does not converge in norm whenever IA AoJ > R.
(3) When A = R, the series may or may not converge in norm.
Proof. (1) We consider only the case 0 < R < oo and leave the other two
cases for the reader. Let K be an arbitrary compact subset of DR(AO).
If 6 = IA AoJ > 0, then we have A A0f 6 < R for every
A E K, and so 6. < 1. Therefore, if we fix some such
that 6 . < < 1, then there exists some n0 such that
10As in the standard case, the number 0 <R < 00 is referred to as the radius of conver-
gence of the series.
11As usual, if r >0, we define = E C: IAAol <'r}. Occasionally, the open disk
Dr(A) will also be denoted
< a for all n n0. In particular, we have IA <a for

all n rio. So, A for all n no and all A E K, and the
desired conclusion follows.
(2) In this case, there exists some n0 such that IA 1 for all
n flo So, JA 76 0 and this shows that for IA Aol > R the
series cannot converge in norm.
(3) The usual examples can be used to establish these claims.
Definition 1.68. A function f: (9 + X, where (9 is an open subset of C,
is said to be differentiable at a point A0 E (9 provided there exists some
(uniquely determined) x0 E X such that
f(A)f(Ao)
lim -- = 0. (*)
AA0
As usual, the vector x0 is called the derivative of f at A0 and is denoted
df(Ao)
, i.e., "
j Oi 01 dA
A function f: (9 X is said to be analytic (or holomorphic) at a
point A0 E (9 whenever f is differentiable at every point of an open disk about
A0. If f is differentiable at every point of (9, them f is called an analytic
function.
IfS a subset of C, then a function g: S C is said to be analytic
is
on S if g has an analytic extension to some open set 0 containing S.
The definition of the (Riemann) integral f(A) dA of a piecewise

tinuous function f: C * X over a continuous rectifiable curve C lying in C
is defined as in the standard case. (The reader should notice that the only
property required to guarantee the existence of the integral is the complete-
ness of the range space.) We shall state below the most important results
for Banach space-valued functions.
The first theorem relates the Banach space-valued integral to continuous
operators and also gives a method of reducing Banach space-valued integrals
to standard ones.
Theorem 1.69. Let T E (X, Y), let C be a continuous rectifiable curve
lying in C, and let f: C X be a continuous function. Then
T(ff(A) dA) = fT(f(A)) dA.

In particular, for each E we have
= f x*(f(A)) dA.
46 1. Odds and Ends
Using the definition of the integral, one can prove Cauch3r's classical
theorem in a similar manner. We state it below for future reference.
Theorem 1.70 (Cauchy's Integral Theorem). Let 0 be a bounded open sub-
set of C whose boundary 30 consists of a finite number of rectifiable simple
curves, say 30 = Assume also that each curve is oriented in
the positive directionin the usual sense of the theory of complex variables.
If U is an open subset of C such that (9 C U and f: U X is an
analytic function, then
ff(A)dA= 0.
Following the standard terminology, we say that a function f: S +
where S is a subset of C, has a power series expansion at some interior
point A0 E S whenever there exist vectors ao, ai,... in X and some positive
number r> 0 satisfying Dr(A0) c S and
f(A)
=
A where (as always) the convergence of the series is
assumed to be in the norm of X.
Lemma 1.71. A function can have at most one power series expansion at
any given point.
Proof. Let A0 E 5, where S is a subset of C, and assume that
A E C S. If then
Ao)nx*(am) =
A satisfying IA < r. From the standard uniqueness theorem

)0
for complex series, we get x*(an) = for each n and each
This implies = for each n.
The following results can also be proven exactly as in the standard case.
Theorem 1.72. A function f: X, where (9 is an open subset of C, is
analytic if and only if f has a power series expansion at every point in 0
Theorem 1.73. Let f: (9 X be an analytic function on a open set 0.

Then:
(1) The function f has derivatives of all orders.

(2) If a open subset V of 0 satisfies V 0 and its
(positively oriented) boundary consists of a finite number of rectifi-
able simple curves, then the derivatives of f are given by Cauchy 's
classical formulas
1
2iri
Joy
foranyn=O,1,2,... and anyA0EV.
(3) At any A0 E V the series expansion of the function f is given by
f(A) = where
1 1 dA
2iri
The above power series converges at least for each A E D(Ao, r),
where r = d(Ao, the distance from A0 to QC = C \ (9.
As in the standard case, a function which is defined and is analytic in C

is called an entire function.
Theorem 1.74 (Liouville). A bounded entire function is constant.
Proof. Let f: C X be a bounded entire function. Then for each x* E X*

the function o f: C f C is a bounded entire function. So (by Liouville's
classical theorem), o f
is a constant for each E X*. Consequently,
x*(f(A) f(O)) = 0 for each x* E X* and each A E C. Hence, f(A) = f(O)
for each A E C, i.e., f is a constant function. D
In the same manner, we can establish the following uniqueness result for
analytic functions.
Theorem 1.75. If two analytic functions f, g: 0 X coincide on a subset
of 0 which has an accumulation point in 0, then f(A) = g(A) for each
AEO.
Let us say that a function f: 0 X is weakly analytic whenever
the (standard) function x* o f: 0 C is analytic for each E Every
analytic function is obviously weakly analytic. Remarkably, as the next
theorem shows, the converse is also true.
Theorem 1.76. A function f: 0 X defined on an open subset 0 of the
complex plane C is analytic if and only if it is weakly analytic.
48 1. Odds and Ends
Proof. Assume that f:

X is weakly analytic. Fix A0 (9 and then
(9
choose r > 0 such that C = {A C: A Aol = 2r} C (9. Since for each
X* the (standard) function a f is continuous, it follows that the
set {x*(f(A)): A c} is bounded. Hence, by the Uniform Boundedness
Principle, there exists some M> 0 satisfying IIf(A) M for each A C.
Now let A0 with A0 for each ri. By deleting a finite number
of terms if necessary, we can suppose that Aol <r for each n. Also, let
72
- ,
AU
Next, fix some with lIx*(I 1. From Cauchy's classical formula, we

have
x *( \
= -

d
Ic
where the circle C is considered positively oriented. This implies
X
*1
XTTi)
1 1 r 1 1
d
2ir Ic sEAm
i d
2ir
Ic
< . IAnAmI
2ir 2r
lvi
Therefore,

= sup 1x*(xn Am!,
IIx*II1
and so is a norm Cauchy sequence. Let x in X, and note

that (since the function f is weakly analytic) for each E we have
x*(x) = x*(xn) = (x* 0 f)'(Ao).
Now if another sequence (9 satisfies p A0, then as above
lim
7j_4QQ
=y X.
Moreover, x*(x) = x*(y) = (x* o f)'(Ao) for each
x =y and that
ff(Ao)
lim
AA0 =x
i.e., f is differentiable at A0. Since A0 is arbitrary, f is analytic.
An open annulus centered at a point A0 E C is any open set of the

form
ARI,R2(A0) = {A E C: R1 <IA Aol <R2},
where R1 = 0 and R2 = 00 are allowed. A formal Laurent series in X
about A0 E C is any series of the form
where X for each rt = 0, 1, 2,.... We say that the Laurent series

E
converges at some A E C whenever the two series (A A0) and
both converge in norm, in which case we write
= +
The proof of the following fundamental theorem concerning convergence

of Laurent series is similar to that of Theorem 1.67.
Theorem 1.77. Let be a formal Laurent series in a
Banach space X and put
R1 = lim sup and R2
=
where we adhere to the conventions = oo and = 0.
If R1 < R2, then the Laurent series converges uniforrnly in norm on
every compact subset of the open annulus ARI,R2(AO).
A function f: A X on an open annulus centered at a point A0 is said

to have a Laurent series expansion if there exists a two-sided sequence
of vectors .. . ,c_1, co, Ci, C2,... in X such that
f(A) -
A E A, where the convergence is again assumed to be in the
norm.
Theorem 1.78 (Laurent). Let f: (9 X be an analytic function defined
on an open set and let ARI,R2 (Ao) (9 be an open annulus (where
need not necessarily lie in (9). Then f has a unique Laurent series ex-
pansion on AR1 ,R2 (Ao). That is, there exist (uniquely determined) vectors
c_2, c1, cO, ci, c2,... in X such that
f(A) = - (*)
1. Odds and Ends
holds for each A AR1,R2(Ao). Moreover, if C is any positively oriented

circle with center at A0 and radius R satisfying R1 <R < R2, then
= dA
L
for each ri = 0, 1, 2,....
In addition, the Laurent series expansion given by (*) is true for any A
in the largest open annulus in (9 centered at A0 that contains AR1,R2(AO).
As in the standard case, if an analytic function f: (9 X has a Laurent

series expansion f(A) = in an annulus, then for every
simple closed rectifiable positively oriented curve C lying in the annulus we
have
dA =
Exercises
1. Establish the following "product rule" for derivatives: Consider two func-
tions g: 0 C and f: 0 X (where 0 is an open subset of C
and X is a complex Banach space). If both functions f and g are
differentiable at some A0 e 0, then show that the X-valued function
A [gf](A) = g(A)f(A) is also differentiable at A0 and
(gf)'(Ao) = g'(Ao)f(Ao) + g(Ao)f'(Ao).

2. Morera's theorem is a converse of Cauchy's theorem and it states that:
If a continuous function f: 0 C (where 0 is an open subset of C)
satisfies f(A) dA = 0 for each simple closed rectifiable curve C which
lies with its interior in 0, then f is an analytic function.
Show that Morera's theorem is valid for Banach space-valued func-
tions.
3. Let C be a rectifiable carve lying in C, and let X and Y be two complex
Banach spaces. Also, let A: C (X, Y) be a continuous
valued function. Show that for each x E X we have
= fA(A)xdA.
[HINT: Since the vectors in both sides belong to Y, it suffices to verify

that for each E we have
In view of Theorem 1.69, the right-hand side equalsy*(A(A)x) dA. Now

consider the functional x on (X, Y) and recall that for each T in
(X, Y) we have x, T) = (y*, Tx). So, on one hand
x, L A(A) dA) = (x)),

[f
and on the other hand, applying x to the operator A(A) dA and
using again Theorem 1.69 we get
x, A(A) dA) = f(y* x, dA.
Therefore,
x,f A(A)dA)
= [f A(A)dA](x) =
4. This is a generalization of Theorem 1.76. Let X be a complex Banach
space and let C be a (not necessarily closed) forming subspace of X*,
that is, there exists a constant 'y > 0 such that
E C and IX*II 1} 711X11
for each S X.
Show that a continuous function f: 0 X defined on an open subset
O of the complex plane C is analytic if and only if the scalar-valued
function o f is analytic for each C.
5. For a continuous function A: 0 * (X, Y), where 0 is an open subset
of C, show that the following statements are equivalent.
(a) The function A A(,\) is analytic on (9.
(b) For each x E X the Y-valued function q52,(,\) = A(A)x is ana-
lytic on (2
(c) For every x X and E Y* the standard complex-valued function
(A) = (A(A)x, = (A(A)x)
is analytic on (2
6. For each r > 0 let = {A C: IAI > r}. Assume that X is a complex
Banach space and for some r > 0 Ar + X is analytic
a function f:
having the Laurent series expansion f(A) = on the open
annulus Show that f(A) 0 if and only if = 0 for each
n = 0,1,2
7. Assume that two points a and b in a topological space X are joined by a
curve C. If C is covered by a collection V of open sets, then show that
there exist (not necessarily distinct) sets Vi, V2,... , V such that:
(a) a V1 and b E
(Such a collection V1, V2,. . ., of open sets is called a chain of open

sets covering the curve C.)
52 - I Odds and Ends
1.6. Fundamentals of Measure Theory

The purpose of this section is to present some measure theory that will
be needed throughout this book. This material will be employed especially
heavily in Sections 5. 1 and 5.2 for the development of a self-contained theory
of integral operators. The material selected is quite standard and can be
found in many texts on measure theory. The book [311 our primary
reference here.
We begin by establishing some notation and terminology. The generic

measure space in our discussion will be denoted by yr), where is a
set, is a or-algebra of subsets of and -ir is a or-additive measure on A
function f: is said to be (or simply measurable) if
fI (B)E for each Borel subset B of The collection of all measurable
functions is a vector space. Two functions f, g: ci * are said to be
ir-almost everywhere equal (in brief ir-a.e. or simply whenever
f(w) = g(w) holds for ir-almost all w. That is, f and g are equal almost
everywhere if there exists some A E with ir(A) = 0 such that f(w) = g(w)
for all w A. Almost everywhere equality defines an equivalence relation
on the vector space of all measurable function. It is easy to see that under
the algebraic and lattice operations defined ir-a.e., the set of all equivalence
classes is a Riesz space. This is the very important Riesz space Lo(ir).
Definition L79. The Riesz space of all equivalence classes of measurable

real functions on ci is denoted by or simply Lo(ir).
As usual, we identify the equivalence classes with functions and agree

that two arbitrary measurable functions f, 9: IR are equal if and only if
f (w) = g(w) for ir-almost all w. With this agreement in mind, a measurable
function f can be changed (or can even be left undefined) on a set of ir
measure zero and still represent the same class as f.
For a proof of the next result, see Exercise 5 at the end of the section.
Theorem 1.80. If ir is or-finite, then Lo(ir) is a Dedekind complete Riesz
space with the countable sup property.
Theorem 1.80 shows that when we deal with or-finite measure spaces, the
various order related notions in Lo(ir) can be described by their sequential
analogues. For instance, the concept of or-order continuity (for functionals
and operators) coincides with the concept of order continuity. When dealing
with Riesz spaces satisfying the countable sup property, it is a common
practice to use the general standard terms (like order continuous operators,
order continuous functionals, order closed sets, etc.) while only sequences
are needed to verify their definitions. Let us recall three basic types of
convergent sequences.
L6. Fundamentals of Measure Theory 53
Definition 1.81. A sequence in L0(7r):

(1) converges pointwise 7r-a. e. to some x E Lo(7r), in symbols,
x x 7r-a.e., if x(w) for 7r-almost all w E
(2) converges relatively uniformly to some x E LO(7r) if there ex-
ists a positive function u E LO(7r) (called the regulator of the
convergence) such that for each E> 0 there exists some k so that
xI EU holds for all n k, and
(3) converges in measure to some x E LO(7r), in symbols,

if for each E> 0 and for each A E with 7r(A) <oo we have
urn 7r({w E A: E}) = 0.12
00
Convergence in measure is characterized as follows.
Theorem 1.82. Let be a crfinite measure space. For a sequence in

L(7r) the following are equivalent.
(1)
(2) Every subsequence of has a subsequence that converges point-

wise 7r-a.e. to zero.
(3) For each A E with 7r(A) <oo we have lim

fl+00 f +Xn dir = 0.
Proof. We can assume that is a positive sequence.

(1) (2) Fix some A E with 7r(A) < oo, and let be a subse-
quence of Also, for each E > 0 let = {w E A: yn(w) E}. Since
0, it follows that for each E > 0 we have 0. In particular,
an inductive argument shows that there exists a strictly increasing sequence
{ } of integers such that 7r (k)) < holds for - all m If we let
= (i), then it is easy to see that 0 This implies
yk8(w) 0 for all w E A. That is, we have shown that for each
A E with 7r(A) <oo and for each subsequence of there exists a
subsequence of that converges 7r-a.e. to zero on A.
Next, pick a sequence with <oo for each n and I

and let be a subsequence of Then, using the preceding conclusion,
an easy inductive argument shows that for each n there exists a sequence
{ k = 1, 2, .. .} such that:
{x1,k} is a subsequence of
'2The reader should be warned that this definition of convergence in measure is less restrictive
than the standard definition, where a sequence {x8} in Lo(ir) converges to x E Lo(ir) in measure
if for each E > 0 we have ir({w E 2 }) = 0. In practice, only sets of
finite measure are important and this is why we restrict our definition to sets of finite measure.
- - ---
1. Odds and Ends
. : k= 1, 2, . . .} is a subsequence of : k 1, 2, . .
Xn,/c on
If we let Xfl,n (the diagonal sequence), then is a subsequence of
{ that converges to zero on ft
(2) Fix some A E E with 7r(A) < oo, and assume contrary to
(3)
our claim that IA d7r 74 0. Then there exist a subsequence of
and some 0 such that IA d7r 2 for each n. The validity
of (2) guarantees that there exists a subsequence of such that
0. Since 0 and 0 < < 1 on A, it follows from the
Lebesgue Dominated Convergence Theorem that IA d7r 0. However,
this contradicts c for each n and the validity of (3) has been
established.
(3) (1) Assume that some A E E <oo and let > 0.
satisfies 7r(A)
Also, let = {w E A: c}. From the fact that t c if and only
if it follows that An {w E A: }. This implies
and consequently
= f f 0.
Thus, 0, and this proves that 0.
Corollary 1.83. Pointwise convergence in Lo(ir) implies convergence in

measure, i.e., x in LQ(7r) implies x.
The next few results contain some basic properties of sequences in LQ(7r).
These results as well as some additional properties can be found in Section 71
of the book by W. A. J. Luxemburg and A. C. Zaanen [225], or in Chapter VI
of the book by B. Z. Vulikh [328].
Lemma 1.84. If cr-finite, then there exists a sequence
Lo(ir) and ir is
of positive scalars such that the sequence is order bounded in

L0(7r).
Proof. Let be a sequence in such that 1' and <oo for

each n, and fix some sequence in LQ(7r). Replacing by we
can assume that each Xn is a positive function. Since for each fixed Ti we
have A k Ik 7r-a.e., it follows from Egorov's theorem that for each n
there exists some E E with An c such that \ An) < and the
sequence A k: k 1} converges uniformly to on In particular,
there exists a strictly increasing sequence of natural numbers such that
A Xn(W)J < 1 for each w E This implies 1 + kn for
each w E By letting = get for each w E
Now, if we let A = urn A E E and for each

w E A we have (w) 0. Moreover, for each n, r E N we have
= U Ak)
m=Ik=rn k=n+r
= U
k=n+r k=n+r
\
k=n+r
As r 1 is arbitrary, \ A) = 0, and this guarantees that A) = 0.
\
Therefore, if we let x(w) = n 1}, then x E Lo(7r) and
O<Anxn<xforeachn.
We continue with one more property of pointwise convergent sequences.
Theorem 1.85. If Xn 0 holds in Lo(7r) and 7T is cx -finite, then there exists

a sequence {An} of positive real numbers such that I oo and 0.
Proof. Pick a sequence E such that I and <oo for

each n, and let a sequence in L0(7r) satisfy 0. By Egorov's the-
orem, for each k there exists some Ak E such that 7r@1k \ < and
the sequence converges uniformly to zero on Ak. An inductive argu-
ment shows that there exists a strictly increasing sequence of natural
numbers such that < for all n and all w E Ak. Let =1
if n < ni and An = k if fl < Then An oo and clearly if
w E Ak and k rn, then for all n This implies
0 for each w E In particular, we have 0
for each w E A = lim inf An = Ak. Since (as in the proof of
Lemma 1.84) we have \ A) = 0, the latter shows that AnXn 0.
In L0 (7r) the notions of pointwise convergence, order convergence, and

relative uniform convergence coincide.
Corollary 1.86. If -ir is cr-finite, then for a sequence Lo(7r) and

some s E Lo(7r) the following statements are equivalent.
(1) The sequence is pointwise convergent to x, i.e.,
(2) The sequence is relatively uniformly convergent to s.
(3) The sequence is order convergent to x,
56 1. Odds and Ends
Proof. (1) (2) Let x. By Theorem 1.85, there exists a sequence

{)'n } of positive scalars satisfying )'n oo and )'n (Xn x) 0. Now
sider the function y = sup{)tnlxn xl: ri E N}. Then yE Lo(rc) and
I for each ri. This shows that {Xn} converges relatively
formly (with regulator y) to zero.
(2) Pick a function u E Lo(7r) and a sequence {kn} of strictly
(3)
increasing natural numbers with k1 = I such that IXm xl for each
rn kn. Now for each ri let mn be the unique natural number such that
kmn fl < and put = Clearly 0 and xl EnU holds
J.
for each ri. This shows that x.

Obvious.
In particular, note that f in Lo(7r) is equivalent to saying that

fn (w) 1 f(w) for all w E IL The L0 also satisfy the
diagonal property.
Theorem 1.87 (The Diagonal Property). Suppose that 71 is u-finite
and {Xn,k} is a double sequence in Lo(7r) such that for each n we have
Xn,k Then there exists a sequence {Yn}
0. and a strictly
creasing sequence {kn} of natural numbers such that Yn 0 and iXn,kI Yn
for all k kn and each n.13
Moreover, if Xn,kl x holds for all ri and k, then we can take the
sequence {Yn} to satisfy Yn x for each ri.
Proof. By Corollary 1.86, for each k there exists some un E such that
the sequence {xn,k: k 1} is relatively uniformly convergent to zero with
respect to the regulator un. Now, by Lemma 1.84, there exists a sequence
{ )'tn } of positive scalars and some u E that )'tnun u for each ri.
such
Put Yn = and note that for each fixed natural number ri there exists an
integer kn such that lxn,kI < < for all k kn. An easy inductive
argument shows that can be taken to be strictly increasing. If x
holds for all ri and k, then let Yn = x A (ku).
The order bounded subsets of L0 are characterized as follows.

Theorem 1.88. When 71 is u-finite, then a subset D of Lo(rr) is order
bounded if and only if for each sequence {xn} D and for each sequence
{'\n} of scalars with )'tn 0 we have )'tnxn 0.
A Riesz space satisfying this property is referred to as a Riesz space with the diagonal
property.
Fundamentals of Measure Theory 57
Proof. Let D be a non-empty subset of If D is order bounded, 0

in R and D, then clearly 0. For the converse recall that (by
Corollary 1.86) order convergence and pointwise convergence of sequences
in coincide. Now assume that the set D satisfies the property:
(P) If D ,j.
0 inR, then 0 holds for 'Tr-almost
allwEft
We must show that D is order bounded in Since D is order bounded

if and only if the set iD I = {Ix!: x E D} is order bounded and the set ID I
clearly has property (P), replacing D by Di, we can assume that D c
Also, since D is order bounded if and only if the set 1+ D = {1 + x: x E D }
is order bounded and the set 1+ D has property (P), replacing D by 1+ D,
we can assume that each x E D satisfies x(w) I for it-almost all w E ft
Furthermore, the set D' = {x A k: x E D and k E N} satisfies property (P)
and has the same upper bounds as the set D. So, replacing D by D', we
can also assume that D consists of bounded functions and that each k E N
dominates some function in D.
Next, let C denote the set consisting of all finite suprema of the functions
in D. Clearly, D and C have the same upper bounds, and it can be easily
shown that C likewise satisfies property (P); see Exercise 7 at the end of the
section. Thus, replacing D by C, we can assume furthermore that D is also
closed under finite suprema.
Now for each k let Dk = {x E D: x k}. Since each Dk is
der bounded (and non-empty by the preceding discussion), it follows from
the countable sup property of (see Theorem 1.80) that there
ists a sequence n = 1, 2,.. .} Dk such that I sup Dk. Let
= E D. Clearly, y defined
by y(w) belongs to Lo(ir). If this is not the case, then there
exists a B E with ir(B) < 00 such that 00 (or, equivalently,
.j,
0) for each w E B. By Egorov's theorem, we can assume that
converges uniformly to zero on B. This implies that there exists a strictly
increasing sequence such that
of natural numbers for all
w E B and all n. That is, (w) > 1 holds for all w E B and all n.
However, the latter conclusion contradicts property (P), according to which
0. This contradiction shows that y E Lo(ir). To finish the proof
note that y is an upper bound of D; in fact, y = sup D in Lo(ir).
58 - 1. Odds arid Ends
Corollary 1.89. For an operator T: E f Lo(ir'), where ir and ir' are two
cr-finite measures and E is an ideal in Lo(ir), we have the following.
( a) If T is cr-order continuous, then T is regular.
(b) If for each order bounded sequence in E that satisfies 0
we have T is a-order continuousand hence it is
a regular order continuous operator.
Proof. (a) Fix x E E+. For any sequence [0, x} and any sequence
of scalars with j 0, we have 0 A72x j 0. This shows that
0 in E, and so by the a-order continuity of T we get 0.
From Corollary 1.86, we obtain and hence, by Theorem 1.88,
0,
the set T[0, x] is order bounded in Lo(ir'). Therefore, in view of Theo-
rem 1.16, T is a regular operator.
(b) Let 0 in E. This implies that is an order bounded se-
quence and 0. By our hypothesis, 0. Now a glance at Corol-
lary 1.86 shows that T
a is a-finite, then (as we saw in Theo-
rem 1.80) the Riesz space is Dedekind complete, and hence the ideals
in L0 (7r) are Dedekind complete Riesz spaces in their own right. In particu-
lar, the classical (0 , ir) measure space and J is an order
dense ideal14 in Lo(ir), then for each function x E Lo(ir) there exists a
sequence JflLi(ir) such that IxnI IxI for each n and x(w)
for ir-almost all w E IL
If, in addition, x then the above sequence can be chosen
in fl Li(ir) and increasing, i.e., satisfying 0 x in Lo(ir).
Recall that if x: 12 IR is a function, then the support of x is the set

Supp x = {w E x(w) 0}.
Corollary 1.91. If -ir is a a-finite measure and J is an order dense ideal

in Lo(ir), then for each function x E there exists a disjoint sequence
c > such that J, <oo for each ri and Supp x =
In particular, there is a disjoint sequence in>2 such that <00
for each m, = and each E J.
14Recall that a Riesz subspace 0 of a Riesz space H is order dense if for each 0 <x E H
there exists some 0 < y E C satisfying 0 < y <x. If H is Archimedean, C is order dense in H if
and only if for each x E H+ we have {y E 0+: y x} x in H; see Exercise 6 at the end of the
section.
Proof. Choose a sequence in > satisfying <oo for each n and

12 = Suppose first that x E J. If = {w E 12: Ix(w)I
then Suppx = From mfxl, we get E E for each m.
Now let x E L0 (ir). By Lemma 1.90 there exists a sequence E+
such that I we can assume that for each w 1' E 11. It
follows that
Supp x = Supp = = n
n=1 n=lk=1 n=lk=lrn=1
Since fl and E E, we see that E E for all n,
k, and m. Thus, we can write Supp x = U=1 with E E for each n.
Now if V1 = B1 and = \ for n 1,2,..., then E J
and <oc for each n, Vk fl = 0 fork n, and Suppx =
The validity of the last statement of the corollary follows from the fact
that 1 E L0(7r).
We continue with a description of the ideals and bands in Lo(7r).
Theorem 1.92. For a cr-finite measure space (12, >, it) and an ideal E in
Lo(7r) we have the following:
(a) There exists a smallest (with respect to rr-a.e. inclusion) set CE in
> such that every x E E vanishes rra.e. on 12 \ CE.
(b) If E is a band, then
E={XEL0(ir): x=0 rr-a.e. on
(c) Two bands B1 and B2 in Lo(71) coincide if and only if C21 = C22
Proof. (a) We consider the finite and afinite cases separately.

CASE I: 'it is a finite measure.
In this case, consider the collection of measurable sets
F={AE>: x=Om-a.e.onAforeachxEE}.
Clearly, j closed under countable unions. Put m = sup{7r(A): A E J}.
is
Since is a finite measure on >, we have 0 <m <oo. If for each n we choose
some E J such that m then the set A U=1 E and
satisfies 'ir(A) = m. We claim that the measurable set CE = 12 \ A satisfies
the desired properties.
To see this, note first that for each x E E we have x = 0 'ir-a.e. on
CE = \ A. Now assume that for some measurable set C we have x 0

7r-a.e. on CC = 12 \ C for each E. Then CC E and from the identity
CC = (CC fl A) U (CC \ A), we get CC \AE The definition of A implies
60 1. Odds and Ends
7r (CC
\ A) = 0, and from this we get CC C A Hence, CE = AC c C,
and the proof of this case is finished.
CASE II: IT is a measure.
Fix a disjoint sequence E with = and < 00
for each n. Also, let E E} and = f E}.
Then each can be considered as an ideal in it). Therefore, from
Case I, for each m there exists a unique minimal measurable set
such that each function x E vanishes outside of the set
CE = then a direct verification shows that this measurable
set CE satisfies the desired properties. -
Assume that E is a band, and let
(b)
B = {x E Lo(it): x = 0 it-a.e. on c2\CE}.
Clearly, B is a band in Lo(7r) and .E c B. Now let 0 < f B and
consider the set C = {w E f(w) > 0}, and note that -w(C) > 0 and
C CE. From the definition of the set CE, it follows that there exists
some g E such that g 0 it-a.e. on C. This means that the measurable
set C = {w E C: g(w) > 0} satisfies it(C) > 0. Moreover, the sequence
{f A ng} c E satisfies f A ng I I xc. Therefore, f xc E .E, and thus
0 < fxc < f. This shows that E is order dense in the band B. Since E is
a band, E = B. (Notice also that E can be identified with Lo(CE,IT).)
(c) This follows immediately from part (b).
Definition 1.93. If E
is an ideal in Lo(it), then the it-a.e. unique measur-
able subset CE determined in part (a) of Theorem 1.92 is called the carrier
(or the support) of the ideal E.
It should be noticed that the order dense ideals in L0(i'r) are precisely
the ideals whose carriers coincide (it-a.e.) with
Now let E be an ideal in Lo(it). Since the measurable sets of the measure
space (CE, ECE, ii), where ECE = {A n CE: A E E}, are precisely the
restrictions of the measurable subsets of to CE (see [32, Problem 15.7,
p. 125]), it follows that the band generated by E in Lo(it) can be identified
with LO(CE, it) = LO(CE). In other words, we have the following result.
Lemma 1.94. If E is an ideal in Lo(it) and is its carrier, then E is
order dense in LO(CE).
We proceed with a brief discussion of Carathodory's classical extension
of measures defined on semirings. Recall that a collection F of subsets of a
set X is said to be a semiring if:
(1)
(2) A,B E F imply AnB EF.

(3) If A, B E F, then A \ B can be written as a finite disjoint union of
sets from F.
A (set) function rn: F [0, oo] defined on a semiring F is called a measure
if rn() = 0 and rn is cr-additive. As usual, the latter property means that if
A E F is a disjoint union of a sequence in F, i.e., if A = U=1 and
flAk = 0 for n k, then rn(A) =
rn (or the outer measure generated by rn) is the outer
measure rn*: 2X [0, oo] defined by
00 00
rn*(A) = C F and
with the convention that inf 0 = oo. Recall that a subset E of X is said to
be rn-measurable if
rn*(A) = rn*(AflE)+rn*(AflEC)
for each A c X. It is well known that the collection
{A E 2X: A is rn-measurable}
is a cr-algebra. The relationships between this cr-algebra and the initial
semiring F as well as between rn* and rn are summarized in the next classical
theorem due to C. Carathodory; for details see [31, Chapter 3].
Theorem 1.95 (Carathodory). Let rn: F [0, oo] be a measure on a
semiring. Then the Carat hAodory extension m* of rn restricted to is
a measure that extends rn, i.e., rn*(A) = rn(A) for each A E F and every
member of F is measurable, i.e., F Moreover, we have the following
three important properties.
(a) The measure rn*: >m f [0,oo] is com4plete, i.e., rn*(A) = 0 implies
(b) If another semiring F1 satisfies r F1 and we consider rn*

restricted to F1, then the outer measure generated by the measure
rn*: F1 [0, oo] is precisely rn*; in particular, (rn*)* = rn*.
(c) If m is cr.jinite and F1 is another semiring satisfying F C F1 C
then the restrictiqn of rn* to F1 is the only extension of rn to a
measure on F1.
In order to avoid introducing unnecessary symbols, it is customary to

denote (with some abuse of notation) the outer measure rn* on by rn
again. Accordingly, the symbols L1 (rn) and L1 (m*) denote the same Riesz
space L1(X, Y2m, rn*). Similarly the symbols Lo(m) and Lo(m*) denote the
Riesz space Lo(X, rn*).
62 1. Odds and Ends
The next result presents a very useful characterization of non-negative

functions in terms of their integrals.
Lemma 1.96. Let rn: F [0, ooj be a measure on a semiring. Then for a
function f Li(rn) we have:
(1) f 0 rn-a.e. if and only if fA f drn 0 holds for all A E F with
rn(A) <oo.
(2) f = 0 if and only if fAfdrn = 0 holds for all A E F with
rn(A) <oo.
Inthe study of integral operators, two basic results from the theory of
iterated integrals will be needed. To state these results we need to recall the
definition of the product measure of two measures. To do so, let (8,
and (T,>22,v) be two measure spaces.
Any subset of S x T of the form A x B, where A E and B E is
called a rectangle. The collection of all rectangles is denoted by x >2.
This is a semiring and the familiar "length x width" formula
(,u x v)(A x B) =
defines a measure on X This measure is called the product measure
of and ii. The Carathodory extension procedure extends the measure
,u x ii from X to the measure x on the o-algebra of all
,u x v-measurable sets. In accordance with our agreement, we will often write
,u x ii instead of (,u x
The o--algebra generated by x >22 is denoted by and is referred
to as the product We have the following obvious inclusions:
>< >22 0 >22 For a detailed discussion of product measures
see [31, Section 26].
Consider now a function f: S x T R. We say that the iterated
integral f7 f8 f(s, t) exists if the following are true:
(1) For v-almost all t T the function f(', t) is
(2) The function g(t)= f8 f(s, t) is over T.

The value of the iterated integral f7 f8 f(s, t) d,u(s)dv(t) is computed by
starting with the innermost integration and then continuing with the outside
integration as follows:
f f f(s, t)
= f [f f(s, t) dy(s)]
The meaning of the iterated integral f7 f(s, t) is analogous.
That is.
f f f(s, t) =
f [f f(s, t) dv(t)] dy(s).
Recall that a function f: S x T is measurable (or, more precisely,

,ax v-measurable if reference to the measure is necessary) if f is measurable
with respect to the u-algebra that is, f'(B) E for each Borel
subset B of IR. A function f: S x T is said to be jointly measurable
if f is measurable with respect to the u-algebra 0 That is, f'(B) E
>i >2 for each Borel subset B of
The iterated integrals of x vmeasurable functions are associated with
a very important classical result known as Fubini's theorem. For a proof
see [31, p. 167].
Theorem 1.97 (Fubini). If xT IR is a x v-integrable function,
then both iterated integrals exist and
fSxT
f x v)
= fTSf f(s, t) = ffST f(s, t)
A very useful converse to Fubini's theorem is known as Tonelli's theorem;

see [31, p. 168] for a proof.
Theorem 1.98 (Tonelli). Let (5, ,u) and (T, >2, v) be two u-finite mea-
sure spaces, and let f: S x T IR be a x v-measurable function. If one
of the iterated integrals
f is the other iter-

ated integral exists as well and
fSxT
The
I xv)
= ff
TS = ffST
next result presents some basic measurability properties of ji x v-
measurable functions.
Lemma 1.99, Assume that and are two u-finite mea-
sure spaces and let f: $ x T be a x v-measurable function. Then:
(1) For v-almost all t T the function f(., t) is
(2) For all s E S the function f(s,.) is v-measurable.
1ff: S x is jointly measurable, then f(.,t) is Y1-measurable for
each t c T and f(s,.) is v-measurable for each s E S.
Proof. We establishthe validity of (1) only. Since J L1(,a x v) is an

order dense ideal in Lo(,a x v), Lemma 1.90 is applicable to the function f
and so there exists a sequence J and a subset A of S x T such that
(,a x v)(A) = 0 and
f(s,t) for all (s,t) A. (*)
64 1. Odds and Ends
By F'ubini's Theorem 1.97, it follows that for each n function t) is

for a set N1 E such that
v(Ni) = 0 and t) is ,a-measurable for all t Ni.
Since (,a x v)(A) = 0, it follows (from F'ubini's theorem again) that the
sets = {s E S: (s, t) E A} are ,a-null sets for v-almost all t. That is,
there exists some N2 E such that v(N2) = 0 and = 0 for each
t N2. Then N=N1UN2 E and v(N)=0.
Fix any t N. For each s we have (s, t) A and so from (*), we
get f(s, t), i.e, the sequence t)} of ,a-measurable functions
satisfies f(s,t) for each s Thus, f(.,t) E L0(p) for each
t N, as desired. E
For a proof of the next result see [32, Problem 20.16]. See also Exer-
cise 16 at the end of this section.
Lemma 1.100. If (S, p) and (T, v) are two arbitrary rnea-
sure spaces and f: S x T R is a ,a x vmeasurable function, then there
exists a -measurable function 9: SxT R such that f = g ,axv-a.e.
A real number M is said to be an essential upper bound of a function

[oo, ooj if f(w) M holds for 7r-almost all w E
f (denoted esssupf or f(w)) is defined by
ess sup f = inf{M E R: M is an essential bound of f}.
If f does not have any essential upper bounds, thenas usualwe let
ess sup f = inf 0 = cc. The function f is essentially bounded above
(or simply bounded above) if it has an essential upper bound (or, equiv-
alently, if ess sup f < cc). The essential infimum of f (denoted ess inf f or
ess f(w)) is defined analogously. We have ess inf f = ess sup ( f).
The essential sup norm (or simply the sup norm) of a function
denoted by
[oo, 00] is and is the essential supremum of
the function If I. That is,
If = esssup (f I = inf{M 0: M is an essential bound of
A function f is said to be essentially bounded (or simply bounded) if

If ( has an essential upper bound or, equivalently, if If < 00. It is easy
to verify that the sup norm satisfies the three axioms of a norm. That is,
for arbitrary functions f, g: [oo, cc], we have:
(1) If 0 and = 0 if and only if f =0

(2) = if for II each scalar
(3) hf + gIIcx 1(fIk +

The collection of all essentially bounded functions in Lo(7r) is denoted

'it) or simply The Riesz space equipped with the
norm is a Dedekind complete Banach lattice.
Lemma 1.101. The essential sup norm satisfies the following properties.
(1) If 0 fTh in Lo(ir), then 11111100.
(2) If c is a non-empty order bounded from above subset of

then we have L5Upgg
sup [sup g] (w) = sup [ess sup g(w)].
Proof. (1) Let 11fm1100 I . Clearly, 1 f = 0 'ir-a.e., then the

conclusion is trivial. So, assume that f > 0, and let (5 < Obviously,
the measurable set D = {w E positive measure. If
f(w) > (5} has
<(5holds for each n, then it is easy to see that (5for it-almost
all w E D. This shows that f(w) for it-almost all w E D, which is a
contradiction. Hence, 6 IIfnM00 must hold for some n. Since (5 <
is arbitrary, the latter implies that f Consequently, and =f
hence I 11f1100.
(2) Let f supYEg f in 0 g f, we get

for each g E Consequently, supYEg If If 0, then the
conclusion is trivial. So, assume that If 1100 > 0.
Fix 0 < iIfU00. Then the measurable set D = {w E f(w) > 8}
must have positive measure. If 6, then let h = fXDc +
and note that g f
h < holds for each g E contradicting the fact that
f= g. Hence, supYEg > (5 for each 0 < (5 < If This shows
that supYEg and so supYEg 1g1100 = If
A normed Riesz space E is said to satisfy the Fatou property whenever

0 Xa X implies I lxii. When E has the countable sup property, the
Fatou property is equivalent to the cr-Fatou property: if 0 I x, then
IIxnII I Lemma 1.101 shows that L00(7r) has the Fatou property. In
general, every AM-space with unit (as defined after Definition 3.1) satisfies
the Fatou property.
66 1. Odds and Ends
Exercises
Let E be a Riesz space with the countable sup property, and let D be a
subset of E such that the supremum u = sup D exists. Does there exist
a countable subset D1 of D such that for each x E D there is an element
a E D1 such that a x? (Any subset D1 of D with this property is called
cofinal in D.)
2. Let E be an Archimedean Riesz space satisfying the countable sup prop-
erty and having a weak unit. Show that for each non-empty subset D of
E there exists a countable subset D0 of D such that D0 has the same set
of upper bounds as D. That is, whenever u E E satisfies u x for each
XE Do,thenuxforeachxED.
3. Let iv) be a measure space, and let of all subsets
be the collection
of that can be written in the form where A E and M c N
for some N E with iv(N) = 0. (As always, = (A\ M) U (M\ A)
is the symmetric difference of the sets A and M.) Verify that:
(a) is a cr-algebra and c
(b) The function p [0, oc], defined by = iv(A), is well
defined and extends IV from to a measure on
(c) If some A E satisfies 7r* (A) < oc, then A E In particular,
is a complete measure, that is, if some subset A of ci satisfies
= 0, then AE
(d) If IV is cr-finite, then =
4. Let J be an ideal in some Lo(iv)-space and let be an order bounded
sequence in J. Show that x in J if and only if p u(w) for
iv-almost all w E [HINT: Assume u for each n and some u E J

and x. Let = zr, = A1 and = Then
j. 0 holds in J and xl for each n
5. Prove Theorem 1.80. [HINT: Choose a sequence such that
0 <oc for each n and and define the function
p: [0,oc) by
p(x) =
Now let D be a bounded from above non-empty subset of Lt(iv). If

yE satisfies p(y) = supXED p(x), then show that y = sup D.]
6. Recall that a Riesz subspace G of a Riesz space L is said to be order
dense if for each 0 < x E L there exists some 0 < y E G satisfying
Show that a Riesz subspace G of an Archimedean Riesz space L is

order dense if and only if for each x E we have {y E y x} I x
in L. [HINT: Let C be an order dense Riesz subspace of an Archimedean
Riesz space L, and let x E L+. Now assume by way of contradiction that
there exists some z <x such that y E C and y x implies y z. Put
u = x z> 0, and then select some v E G with 0 <v u <x. This
implies v z and so 2v = v + v u + z = x. By induction we get nv x

for each n. Since L is Archimedean, the latter yields v = 0, which is a
contradiction.
7. Consider the following property (P) for a subset D of a Riesz space:
(P) If {x71} C D and A71 .J. 0 in IR, then A71x71 0.
Assume that a subset D of a Riesz space E satisfies the above property
(P). Prove that the set consisting of all finite suprema of the collection
DI = d E D} also satisfies (P).
8. For a sequence {x71} in some Lo(ir)-space establish the following.
(a) If x71 x and x71 y in Lo(ir), then x = y.
(b) If x71 in Lo(ir) and is a positive sequence, then x 0.
9. We say that a sequence {x71} in some Lu(ir) is *convergent to x (in
symbols, x) if every subsequence of {x71} has a subsequence that
converges pointwise 'ir-almost everywhere to x.
Let E, ir) be a u-finite measure space and let A be a measurable
set of finite measure. Also let A R be a ir-integrable function such
that > 0 for ir-almost all w E A. Show that for each sequence {B71}
of measurable subsets of A the following statements are equivalent:
(a)
(b)
(c) fBcbdlr_-*0.
(d) ir(B71)*0.
10. Assume that (SI, E, ir) is a u-finite measure space and that a sequence
E satisfies <oo for each n and = ft Suppose
also that b: [0, oc) p [0, oc) is a continuous strictly increasing bounded
function with = 0; for instance, let = or I
Show that a sequence {x71} L0(ir) satisfies x71 if and only if
lim k
dir = 0 for each k.
11. Let E, ir) be a u-finite measure. Fix a sequence in such that
o <oo for each k and = Also, define the function
p: Lo(ir) p [0,oc) by
p(x) =
f
Show that for a sequence {x71} Lo(ir) and a function x E Lu(ir) the
following statements are equivalent.
(a)
(b) For each k we have Em f dir = 0.
(c) lim p(x71 x) = 0.
ri*oo
12. Let (S,E1JL) and (T,E2,v) be two measure spaces. 1ff: S x T p

is jointly measurable, then f(., t) is E1-measurable for each t E T and
f(s,.) is v-measurable for each s E S.
68 1. Odds and Ends
13. Prove Lemma 1.90. [HINT: Note that L1 (iv) is an order dense ideal in
Lo(ir) and that the intersection of two order dense ideals is an order dense
ideal.
15. Let F be a semiring of subsets of set and let r p [0,oo) be two
set functions such that 0(A) holds for each A e r. If 0 is finitely
additive and e is a measure (i.e., countably additive), then show that 0 is
also a measure.
16. Establish the following property of the product measure: If the measures
,i and v are a-finite, then for each set C E and for each E> 0 there
exists a set of the form D = x where E E3 and E E2
for each n, such that x <E.
17. If / E Lo(iv), then show that
max{ess sup /, ess sup (/)} = ess sup I/I =
18. Assume that (11, E, iv) is a measure space and let Lt(iv). Also
let x = be the ir-a.e. pointwise supremum of the sequence
Show that x L1 (iv) if and only if there exists a pairwise disjoint sequence
{ATh} E such that div = 00.
Chapter 2
Basic Operator Theory
This chapter presents the general background needed to study (linear) op-
erators. It gives a concise presentation of the basic structural properties of
Banach spaces and Banach lattices and discusses the fundamental properties
of operators between these spaces. We pay special attention to the structure
of Banach lattices and operators acting on them. Banach lattices are
duced as special cases of ordered vector spaces. We take this opportunity
and review the fundamental properties of compact and weakly compact op-
erators between Banach spaces. We also state the basic compactness results
on operators dominated by compact or weakly compact operators.
We assume the reader is familiar with a basic course in functional anal-
ysis that includes an elementary introduction to operator theory.
2.1. Bounded Below Operators

Recall that a mapping f: A + B is said to be surjective (or onto) when-
ever its range R(f) coincides with B, i.e., whenever for each b E B there
exists some a E A with f(a) = b. In this section, we shall discuss some
important properties of surjective operators between normed spaces.
Definition 2.1. An operator T: X Y between normed spaces is said to

be bounded below if there exists a constant 0 such that
IITxII
for each x E X. Equivalently, T is bounded below if there exists some y> 0

satisfying tlTxM 'y for all x E X with = 1.
= 69
2. Basic Operator Theory
The negation of the preceding definition can be formulated in terms of

sequences as follows.
Lemma 2.2. An operatorT: X v Y between two normed spaces fails to be
bounded below if and only if there exists a sequence of unit vectors in
X such that 0.
The proof of the next result is left for the reader.

Lemma 2.3. An operator T: X v Y between two real norrned spaces is
bounded below if and only if Tc: v Yc is likewise bounded below.
In the next simple lemma we collect some properties of bounded below

operators which will be useful in our work.
Lemma 2.4. Let X and Y be normed spaces and let T E (X, Y) be a
bounded below operator. Then:
(1) T is one-to-one.
(2) If S E (X, Y) is close enough to T, then S is also bounded below,
i.e., the set of all bounded below operators from X to Y is an open
subset of (X,Y).
Proof. (1) Clearly Tx = 0 implies x = 0.
(2) Let for each x E X and suppose that another operator
S in (X, Y) satisfies Jf S Now if x is a unit vector in X, then
= ((S - T)x +TxJJ - fJ(S- T)xM
which shows that S is also bounded below.
The bounded below operators between Banach spaces are characterized

as follows.
Theorem 2.5. A continuous operator T: X Y between Banach spaces
is bounded below if and only if T is one-to-one and has a closed range.
Proof. Let T: X v Y be a continuous operator between Banach spaces.
Assume first that holds for all x E X and some 'y> 0. Clearly,
T is Next, if y in X, then from
- TXmIf = - Xm)jI
we see that is a norm Cauchy sequence. If v x, then p Tx,
and so y = Tx, proving that T has a closed range.
For the converse, assume that T has a closed range and is By
the Open Mapping Theorem, the operator T: X R(T) has a continuous
2. 1 . Bounded Below Operators 71
inverse. So, for some c > 0 we have for all y e R(T). Letting
y=Tx,weget forallxEX.
Corollary 2.6. If a bounded operator T: X Y between Banach spaces
has a closed range and a finite dimensional null space, then every operator
S E (X, Y) close to T has also closed range and a finite dimensional null
space such that dim N(S) dim N(T).
In particular, the collection of all bounded operators in (X, Y) having a
finite dimensional null space and a closed range is an open subset of (X, Y).
Proof. Assume that a bounded operator T: X Y between Banach spaces

has a closed range and a finite dimensional null space. Since N(T) is finite
dimensional, it is complemented in X; see Exercise 6 of Section 4.1. So, there
exists a closed subspace V of X such that N(T) V = X. In particular,
T: V R(T) is an isomorphism. Fix some c> 0 such that
for each x E V. Now let S E L(X, Y) be an operator satisfying lIT SII <.
Then for each x E V we have
JSxlI SxM =
This shows that the operator S is bounded below on V, and so its
range S(V) is a closed subspace of Y and S is olie-to-one. Now note that
S(X) = 8(N(T)) + 8(V) and that 8(N(T)) is a finite dimensional vector
subspace. Since 8(V) is closed, it follows that 8 has a closed range; see
Exercise 1 in Section 3.4. Finally, a look at Exercise 6 at the end of the
section shows that dim N(S) <dim N(T).
The notion of an isomorphism between Banach spaces is as follows.

Definition 2.7. A bounded surjective one-to-one operator between two Ba-
nach spaces is called an isornorphisrn. Two Banach spaces are said tobe
isomorphic if there is an isomorphism from one space onto the other.
It is interesting to know that two complex Banach spaces can be isomor-

phic as real Banach spaces but they may fail to be isomorphic as complex
Banach spaces; see [70, 169].
Two characterizations of isomorphisms are given in the next result.
Lemma 2.8. For a bounded operatorT: X Y between two Banach spaces
the following conditions are equivalent.
(1) T is an isomorphism, i.e., T is surjective and one-to-one.
(2) T is bounded below and R(T)=Y.
(3) T is bounded below and its range R(T) is dense in Y.
Proof. The implications (1) (2) (3) are obvious, and (3) (2)
follows from Theorem 2.5.
Now let c L(X,Y) be a sequence of surjective operators between

two Banach spaces, and assume that f
T in
L(X, Y). It should be clear that T may easily fail to be onto. For example,
if S is a surjective operator, then the sequence of surjective operators
converges to zero. However, as we shall show next, the situation is different
if we know that T is bounded below. The next result is very important
and it will be used quite often in this monographit can be found in [170]
and [19].
Theorem Let T: X Y be a bounded below operator between Banach
spaces. If a sequence C (X, Y) of surjective operators converges in
norm to T, then T is necessarily surjective (and hence an isomorphism).
Proof. Fix constant > 0 such that
some for each x E X.
Without loss of generality we can assume that [IT II 'y for each n. This
easily implies for each x E X; see the proof of Lemma 2.4.
Next, fix an arbitrary y E Y. To establish that T is surjective, we need to
show that there exists some x E X satisfying Tx = y.
Start by choosing a sequence in X satisfying = y for each
n. From we see that We claim that
{ is a Cauchy sequence. To see this, note that
IITXnTXmIJ IETXnyM+IIYTXmIJ
= + IITrnXrn TXmJI
< [lIT + IT TmM] >0.
Let x in X and so Tx. On the other hand, from
- = - TnxnlI - . 0,
it follows that Tx = lim = y, as desired.
The next result (which can also be found in [170] and [19]) asserts that
the set of all surjective operators is an open set. It should be noticed that
its proof resembles the proof of the Open Mapping Theorem.
Theorem 2.10. If X and Y are Banach spaces, then the set of all surjective
operators from X to Y is an open subset of (X, Y).
Proof. Let T E (X, Y) be a surjective operator. By the Open Mapping
Theorem we know that T is an open mapping. Therefore, there exists some
0 <c < I such that 2cUy or
cUy C (*)
2.1. Bounded Below Operators
To complete the proof, we shall show that every operator S E (X,Y)

satisfying liT < c is surjective. To this end, let S E (X, Y) satisfy
I1T-SII<c.
Fix Yo E cUy. By (*), there exists some such that Txo = Yo.
Let y' = Then 11 = SxoJ} lIT < c. So, by (*)
again, there exists some Xi E such that Tx1 = Yo Put Y2 = SXj.
and note that IlYo Yi Y211 = l1Txi SxilI
Now continuing this way, by using (*) and an inductive argument, we

can construct vectors xo, Xi,... X and Yo, E Y such that
yflSxfl_i, and
for n = 1,2,....
Since X is a Banach space, the series is norm convergent in X,
say x = Moreover, from = 0, we see that
00 00 00
Yo =>.yi = >5(xj_i) = =Sx.
This proves that S is a surjective operator.
Corollary 2.11. The collection of all isomorphisms between two Banach

spaces X and Y is an open (possibly empty) subset of (X, Y).
Proof. Let T: X Y be an isomorphism. If an operator S (X, Y) is
close enough to T, then S is surjective by Theorem 2.10 and bounded below
by Lemma 2.4(2). Consequently, S is an isomorphism by Lemma 2.8.
Corollary 2.12. If X and Y are Banach spaces, then the set of all surjective
and non-invertible operators fromX toY is an open-subset-of (X,Y-).- -
Proof. Consider the collection of bounded operators
C = {T (X, Y): T is surjective and not invertible}
and fix some T C, the operator T (X,Y) is surjective and

invertible. By Theorem 2.10, T is an interior point in the set of all surjective
operators from X to Y. Now suppose by way of contradiction that T is not
an interior point of C. This means that there exists a sequence of
invertible operators from X to Y satisfying T
T is not invertible there exists a non-zero vector x X such that
Tx 0. Let V be the (one-dimensional) vector subspace generated by x,
and let W be a closed subspace of X such that V W = X; see Exercise 6
of Section Clearly, the operator Tlw E (W, Y) is surjective and none
of the operators (W,Y) is surjective since each X Y is
injective. However, from TJw in (T4/, Y) and Theorem 2.10, it

follows that eventually the operators must be surjective, which is a
contradiction. This contradiction establishes that C is an open subset of
(X,Y).
If X is finite dimensional and dimY dimX, then C = 0. If X is finite

dimensional and dim Y < dirnX, then the collection C is and
coincides with the set of all surjective operators from X onto Y.
We shall finish the section with a few results concerning the closedness
of the range of a bounded operator.
Theorem 2.13. If T: X Y is a bounded operator between norrned
spaces, then {R(T)]-'- = N(T*). In particular, if the operator T has closed
range, then
dim [Y/R(T)}* = dim [R(T)]' = dim N(T*)
Proof. For the first part note that:

E [R(T)]' = T*y*(x) = 0 for all x E X
=0 E N(T*).
The second part follows from Theorem 1.12.
Although the range of a bounded operator between Banach spaces need

not be closed, there is a way to define a "stronger" norm on the range so
that it becomes a Banach space. The details are as follows.
Theorem 2.14. Let T: X Y be a bounded operator between Banach
spaces, and let denote the quotient norm of the quotient Banach space
X/N(T). If for each y = Tx E R(T), we let
IIyIkr' = I J
[z]
then we have the following:
(a) The function defines a norm on the range R(T) of T.

(b) (R(T), is a Banach space.
(c) The operator T: (X, 1) (R(T), IT) is continuous. 1
(d) The operator T has closed range (i.e., R(T) is Il-closed) if and
only if the norms . and are equivalent on R(T).
(e) Every operator S: R(T) R(T) which is 1-continuous is also
I
111T-contznuous. In particular, every operator S E (Y) that leaves
R(T) invariant is ItT-continuous on R(T).
21. Bounded Below Operators 75
Proof. (a) For each x E X we have

I j
= inf{ z 1: z E X and Tz = Tx }.
So, if Tu = Tv, then = 11I[vIIH. This easily implies that It IT is a
well-defined norm on R(T).
(b) Let be a sequence in R(T). Choose a sequence
in X such that = From
MYn - YrnIIT = Ilyn - YTnM + -
it follows that is a sequence in Y and is a
Cauchy sequence in X/N(T). Pick Y and [x] X/N(T) such tEat
II - 0 and - {X] 0.
From the definition of the quotient norm, it follows that there exists a
quence in X such that 0 and = Tx = Tx.
Since Ilyn 0 and + 0, we see that y = Tx. So, y E R(T) and
IIYn-YIITIIYnYII+1[Xn]-[X]11H0.
This implies that (R(T), is a Banach space.
(c) This follows immediately from Corollary L6.
(d) The identity operator I: (R(T), (R(T), 1
j) is clearly
continuous. R(T) is closed in X, then (R(T), '11) is a Banach space. But
If
then, according to the Open Mapping Theorem, I is a homeomorphism and

this guarantees that the norms and are equivalent.
Conversely, if the norms and are equivalent, then, in view of
1
(b), the range R(T) is 1-complete and, hence, closed.
(e) Again, this follows immediately from Corollary 1.6.
The next consequence of the preceding theorem includes several elemen-

tary characterizations of the closedness of the range of an operator.
Corollary 2.15. For a bounded operator T: X + Y between two Banach
spaces the following statements are equivalent.
(1) The range of T is closed.
(2) There exists a constant c > 0 such that for each y E R(T) there
exists some x E X satisfying y = Tx and Ix 1.
(3) For any sequence {yn} in R(T) satisfying lynII + 0, there exists a
sequence in X such that 0 and = for each m.
Proof. (1) (2) Since R(T) is closed, the norms and are equiv-
alent on R(T). That is, there exists some 5 > 0 such that
76 2. Basic Operator Theory
holds for all y R(T) In particular, if z X satisfies Tz = y, then we

.
have [z] JJ where [z] = z + N(T). Now notice that there exists some
x E X with y = Tx and lxii (1 +
Obvious.
(3) Assume that a sequence {YrJ R(T) satisfies
(1) f 0.
Pick a sequence in X satisfying n and * 0.
From the inequality
= JlxniI
we get IYn fIT + 0. So, the identity operator I: (R(T), II. f (R(T),
is continuous. By part (d) of Theorem 2.14, R(T) is closed.
Another simplebut very usefulcharacterization of the closedness of
the range of a bounded operator is the following.
Theorem 2.16. A bounded operator T: X * Y between two Banach spaces
has a closed range if and only if there a closed subspace Z of Y such
that R(T) n Z = {0} and the vector subspace R(T) Z of Y is closed.
Proof. If R(T) is closed, then the closed subspace Z = {0} of Y satisfies
the stated properties. For the converse, assume that there exists a closed
subspace Z of Y such that R(T) n Z = {0} and the vector subspace V =
R(T) Z of Y is closed. Clearly, the vector space VIZ equipped with its
quotient norm is a Banach space. Moreover, the mapping J: R(T) V/Z,
defined by J(y) = y + Z, is a surjective linear isomorphism.
Since fJyIJ holds for each y E R(T), it follows that
the operator J: (R(T), . * (VIZ, II) is continuous, and hence it
is an isomorphism. In particular, there exists a constant c > 0 satisfying
for all y E R(T). This implies
c(IylIT
for all y E Y. That is, the norms . I! and H lIT are equivalent on R(T).
Therefore R(T) is a closed subspace of Y.
Corollary 2.17. Let T: X f Y be a bounded operator between two Banach
spaces. If the quotient vector space Y/R(T) is finite dimensional, then the
range of T is closed.
Our final characterization of the closedness of the range of a bounded

operator is due to S. Banach [51, pp. 149150].
Theorem 2.18 (Banach). A bounded operator T: X * Y between Banach
spaces has closed range if and only if its adjoint operator T*:
likewise has closed range.

2.1. Bounded Below Operators 77
Proof. Assume first that R(T) is closed. Since the operator T: X R(T)
is a surjective operator between Banach spaces, there exists (by virtue of
the Open Mapping Theorem) some constant c > 0 such that CUR C
where UR denotes the open unit ball of R(T).
Let R(T) satisfy ft Pick a sequence satisfying
= for each n, and let be the restriction of to R(T). From
= sup = sup = sup
YEUR
and + 0, it follows that * 0. Let be a linear extension of

Y such that = Then = holds for all n
and * 0. By Corollary 2.15(3), the range of the operator T* is closed.
For the converse, assume that R(T*) is closed. Put Z = R(T) and
consider the operator S: X Z defined by Sx = Tx for each x E X. Let
y* be an arbitrary linear extension of to all of Y. Then we have
S*z*(x) = z*(Tx) = y*(Tx) = T*y*(x) for all x E X, i.e., S*z* = T*y*.
This implies that R(S*) = R(T*). In other words, replacing T by S if
necessary, we can suppose without loss of generality that the range of T is
also dense. Under this extra assumption, we must show that R(T) = Y.
Since R(T) = Y, it easily follows (see Exercise 10 at the end of the
section) that T* is one-to-one. Since R(T*) is assumed to be closed, by
Theorem 2.5, the operator T* is bounded below. That is, there exists some
constant C > 0 satisfying CJIy* tIT*y* 1 for all E Next, we claim
that we have {y E Y: < C} If this is done, then (as in the
proof of the Open Mapping Y: C T(U),
from which it follows that R(T) = Y.
To establish the inclusion {y E Y: jyfj <C} pick any E Y
satisfying < C and assume by way of contradiction that y
Then, by the Separation Theorem, there exists some E such that
y*(y) > sup{Iy*(Tx)I: x E = IIT*y*lt
Since y*(y) the preceding_inequality implies > c, a contra-

diction. So, {y E Y: <c} c is true, as desired.
Exercises
1. For a bounded below continuous operator T: X Y between normed
spaces establish the following:
(a) If X is not a Banach space, then T need not have a closed rangeS
(b) The unique continuous linear extension of T to the norm completion
of X is a'so bounded below.
[HINT: Let Y be an arbitrary Banach space, X a non-closed subspace of
Y and define the operator T: X Y by Tx = x for each x E X. Then
R(T) = X is not closed in Y.]
2. Prove Lemma 2.3.
3. Let T: X p Y be a bounded, one-to-one, and surjective operator betwee
two Banach spaces. Show that T is an isornetry if and only if T and T'
are both contractions. (Recall that a bounded operator between two
normed spaces is called a contraction if its norm is no more than one.)
4. Let T E (X) be bounded below but not invertible. Show that there
exists some > such that the operator A T
0 is not surjective whenever
AJ <. [HINT: Use Theorem
5. This is a generalization of the previous exercise. Assume that X and
Y are Banach spaces and let T E (X, Y) be below but not
invertible. Then show that there exists some such that the operator
T S is not surjective whenever S E (X, Y) satisfies <E. That is,
show that the set of all bounded below and non-surjective operators from
X to Y is an open subset of (X, Y).
6. Establish the following property that was used in the proof of Corol-
lary 2.6. Assume that X = V W, where V is finite dimensional. If Y
is an arbitrary vector space, then any operator T: X Y that is one-to-
one on W satisfies dimN(T) <dim V.
7. Show that the assumption dim N(T) <oc in the last conclusion of Cor-
ollary 2.6 cannot be replaced by the assumption that N(T) is comple-
mented.
8. Prove Corollary 2.17. [HINT: Let {[yi], .,be a basis of

Y/R(T). If V is the vector space generated in Y by {yl,y2,. .
then V is closed, R(T) fl V {O}, and R(T) V = Y. Now invoke

Theorem 2.16.]
9. Let T: X Y be a bounded operator between Banach spaces such that

its range R(T) is not closed. If X V W, where V is closed and W is
finite dimensional, then show that T: V Y is not bounded below.

10. For a bounded operator T: X Y between Banach spaces establish the
following properties regarding the ranges of T and its adjoint T*.
(a) The range R(T) of the operator T is dense if and only if the adjoint
operator T*: X* (defined as usual via the duality identity
(Tx, = (x, T*y*)) is one-to-one. [HINT: Use the Hah.nBanach
theorem.
(b) If the adjoint operator T* is one-to-one and has closed range, then
show that T is surjective. [HINT: Use the preceding part in connec-
tion with Theorem 2.18.]
(c) If T carries norm bounded closed sets onto closed sets, then T
has a closed range. Moreover, all powers of T a'so have closed
ranges. [HINT: We can assume that T is one-to-one. (Other-
wise, we can replace X by X = X/KerT and T by k * Y
defined by T[x] = Tx.) Now assume * y. If has
a norm bounded subsequence, say then y belongs to the
closed set T({zi, Z2,. .}). The case oo cannot occur.
For if 11 00, then * 0, and consequently zero belongs
to T({x E X: lxii = 1}), contrary to the fact that T is one-to-one.]

11. Let X and Y be two Banach spaces, and let I(X,Y) denote the open
subset of (X, Y) consisting of all isomorphisms from X to Y. Show that
the mapping T T',
from I(X. Y) to I(Y, X), is (norm) continuous.
[HINT: If 8, T E I(X, Y), then note that
T' = T'(T S)8' = [(T' S') + S'](T

= (T' S')(T 8)8' + 8'(T S)8' .1

12. For a bounded operator T: X Y between Banach spaces establish the

following.
(a) The T has a closed range if and only if R(T*) =
(b) If T has a closed range, then X*/R(T*) = [N(T)]*.
2.2. The Ascent and Descent of an Operator

Recall that if T: X + Y is an operator between two vector spaces, then
N(T) and R(T) denote the null space and the range of T, respectively.
That is,
Tx=O} and R(T)={Tx:
The null space of T is, of course, another name for the kernel of T.
In this section, we shall discuss the notions of ascent and descent of an
operator on a vector space. So, fix an operator T: X X on an arbitrary
vector space. The null space of the operator Tk is a T-invariant subspace
of X, that is, T(N(Tk)) C Indeed, if x E N(Tk), then Tkx 0,
and therefore = T(Tkx) = 0, Tx E N(Tk). We also have the
following subspace inclusions:
The range R(Tk) = Tk (X) of each operator Tk is clearly another T-invariant

subsp ace of X. Moreover:
R(T) R(T2) R(T3)
1
If X is a norrned space and T is continuous, then is also a T-invariant closed subspace.
Lemma 2.19. For an operator T: X X on a vector space we have the

following.
(1) If N(Tk) = holds for some k, then N(T') N(Tk) for

all n k.
(2) If R(Tk) = holds for some k, then = for
allnk.
Proof. We establish the validity of (1.) only. To this end, assume that
equality = holds for some k. The proof is by induction
on n. For the induction step, suppose that N(Tk) = for some
n1. then we have = = 0 or,
equivalently, TThx e N(Tk+l) = N(Tc). This implies = Tk(TThx) =0
or x E Consequently, =
We now introduce the ascent and descent of an operator.

Definition 2.20. Let T: X X be an operator on a vector space.
(1) The ascent a(T) of T is the smallest natural number k such that
= N with
then we let a(T) = oo.
(2) The descent 8(T) of T is the smallest natural number k such that
R(Tk) = If there is no k e N with =
then we let 8(T) oo.
turns out that if the ascent and descent of an operator are finite, then
It
they are equal. This useful result is stated next.

Lemma 2.21. If an operator T: X X on a vector space has finite ascent
and descent, then they must coincide, i.e., a(T) = 8(T) = p < 00, and
X is a Banach space and T e (X), then is a closed

subspace.
Proof. Let k = a(T) <oo and rn = 8(T) <oo. We claim that

fl R(Tk) = {0} for p = 1, 2, 3,... (*)
and
for q=1,2,3,... . (**)
To see (*), fix p e N and let y e fl Then there exists some
x E X such that y = Tkx. This implies = = 0 or, equivalently,
x e = Hence, y = Tkx = 0, and the validity of (*) has
been established.
______-
For (**) fix q E N and let x E X. Then Ttmx E R(Tm) =

and so there exists some z E X such that Tmx = This implies
Tm(x = 0 or E N(Tm). Now notice that x = (x
belongs to N(Tm) +
Letting p = rn in (*) and q = /c in (**), we get
X = N(Tm) R(Tk). (***)
Fix u E From (* * we can write U = + U2 with the vectors
U1 N(Tm) c N(Tm+l) and U2 E R(T'). But then from (*), we get
U2 = U Ui E N(Tm+l) fl = {0} or U2 = 0. Hence, U = Ui E
and so N(Tm+l) = N(Tm). Thus, /c <rn.
Next, let v E R(Tk). Using (**) we can write v = v1 + V2 with
in N(Tm) and v2 E But then it follows from (*) that
the vector = V satisfies E R(Tc) fl N(Tm) = {0} or v1 = 0.
Consequently, v = v2 E and so R(Tk) = This implies
rn < /c. Therefore, rn = k p < 00. The closedness of follows from
Theorem 2.16.
Let T: X X be a bounded operator on a Banach space. A pair (V, W)

of closed subspaces of X is said to be a reducing pair for the operator T
(or it reduces the operator T) if X = V e W and both V and W are
T-invariant. We also say that a closed subspace V of X reduces T if there
exists another closed subspace W of X such that the pair (V, W) reduces T.
Theorem 2.22. A closed subspace V of a Banach space X reduces a bounded
operator T: X X if and only if there exists a projection P E L(X) such
that its range is V and P commutes with T, i.e., TP PT.
Proof. Let T: X + X be a bounded operator on a Banach space X. As-
sume first that (V, W) is a reducing pair for T, and let P E L(X) be the
continuous projection with range V along W. If x v + w E V W X,
then we have
PTx = P(Tv + Tw) = Tv = TPx,
and so PT = TP.
For the converse, assume that V is the range of a continuous projection
P on X such that PT = TP. Let W = (I P)(X) and note that W
is a closed subspace of X satisfying V W = X. Now if v E V, then
Tv = TPv = P(Tv) E V, and so V is T-invariant. Finally, for each x E X
we have
This shows that W is T-invariant too. Therefore, V reduces T.

Recall that an operator T: U U on a vector space is said to be

nhlpotent if Tc = 0 for some k. The smallest k satisfying Tc = 0 is called
the index of nilpotence of the operator T.
Theorem 2.23. A bo'unded operator T: X X on a Banach space has

finite ascent and descent if and only if T has a reducing pair of closed sub-
spaces (V, W) such that the operator T: V ' V is nilpotent and T: W ' W
is invertible.
Moreover, if p = a(T) 8(T) < oo, then the pair (V,W), where V
N(TP) and W R(TP), is the only reducing pair for the operator T such
that T is nilpotent on V and invertible on W.
Proof. Let T: X X be a bounded operator on a Banach space. If the

ascent and descent of T are both finite, then (by Lemma 2.21) they must
coincide. Consequently, if we let a(T) = 8(T) p < oo, then we have
X = N(TP) R(TP). Clearly, (N(TP), R(TP)) is a reducing pair for T.
Moreover, T is nilpotent on N(TP) and invertible on R(TP).
For the converse, assume that (V, W) is a reducing pair for T such that
T is nilpotent on V and invertible on W. Let p be the smallest natural
number such that = 0 on V, the index of nilpotence of T. We claim that
p = a(T).
To see this, let x E Since X = V e W, there exist 'v E V and
wEWsuchthatx=v+w.
fact that the invertibility of T on W implies the invertibility of on W,
weinferthat w=0. Thus, x=vEV, and soN(TP)cV.
on V, the inclusion V is trivially true, and hence N(TP) = V.
A similar argument shows that = V for all n 0. This implies
a(T) <p < oo. Now assume that some k E N satisfies _
Then = N(TTh) for all n k, and so V. In particular,
Tk = 0 on V, and therefore k p. This establishes that p = GE(T).
To finish the proof, we must show that 8(T) = p and that W = R(TP)
Start by observing that if x = v + w E V W = X, then the vector
= + =
belongs to W, and so R(TP) W = TP(W) R(TP). That is, R(TP) = w.

Similarly, we have R(Tm) = for all rn p. In particular, we have
8(T) p < oo. Since 8(T) is finite, it follows that 8(T) = = p, and
the proof is finished. D
Exercises
1. Prove part (2) of Lemma 2.19.
2. Show that every operator on a finite dimensional vector space has finite
ascent and descent.
3. Consider the operator T: R3 R3 defined by the matrix
001
T=010.
000
Find the ascent and descent of T.
4. Give examples of bounded operators T: X X on a Banach space such
that:
(i) a(T) <co and 8(T) = co.
(ii) a(T) = co and 8(T) <Co.
(iii) = co and 8(T) = Co.
5. Let T: X X be an arbitrary bounded operator on a Banach space such

that p = = 8(T) <co. Show that the adjoint operator T*:
also has finite ascent and descent and that a(T*) = 8(T*) = p. [HINT:
Let V = N(TP) and W = and note X = V W. Then
= W*, and T* is nilpotent on V* and invertible on W*.
6. Show that a bounded operator T: X X on a Banach space is nilpotent

if and only if for each x E X there exists some n E N such that = 0.
[HINT: Note that X = U=1
7. If a nilpotent operator T: V V on a vector space has index p, then
show that a(T) = 8(T) = p.
8. For an operator T: V V on a vector space establish the following
algebraic properties.
(a) For each n, m E N we have C fl N(Tm).
(b) For each n, m E N the operator n N(Tm)
is surjectiveand consequently the vector spaces
and fl N(Tm) are linearly isomorphic.
(c) If the ascent of T is finite, then R(TP) fl = {0} for each
nENandallpa(T).
(d) If for some p, n E N we have R(TP) fl = {0}, then T has
finite ascent and a(T) p.
9. Let T: V V be an operator on a vector space. Show that for each

pair n, m E N the mapping J: V/[N(TTh) + R(Tm)]
defined by
J[v} = +
is a well-defined surjective linear isomorphism.

__
2.3. Banach Lattices with Order Continuous Norms

An important connection between the order and topological structures of a
Riesz space is provided by the notion of order continuity.
Definition 2.24. A lattice seminorm p on a Riesz space is said to be order
continuous if j 0 implies j 0. A Banach lattice has order
continuous norm if its norm is order continuous.2
Banach lattices with order continuous norms play a very important role
in analysis. Before stating their most useful characterizations, we need to
recall a few basic concepts pertaining to Riesz spaces.
First of all, in a partially ordered vector space any subset of the form
[x,y] = {z: x z y}, where x y, is referred to as an order interval.
A vector subspace A of a Riesz space E is said to be a Riesz subspace
(or a vector sublattice) if for every x, y E A the supremum x V y and the
infimum x A y (taken in E) belong to A.3 Every ideal is a Riesz subspace.
If E is a normed Riesz space, then E (embedded naturally in its double
dual E**) is always a Riesz subspace of E**. In other words, we have the
following important result.
Theorem 2.25. The natural embedding of a normed Riesz space into its
double norm dual is a lattice isometry.
We are now ready to state the most important characterizations of Ba-
nach lattices with order continuous norm; for details see [30, pp. 179183J.
For an extensive treatment of Banach lattices with order continuous norms
we refer the reader to the well-written monograph of W. Wnuk [338].
Theorem 2.26. For a Banach lattice E the following statements are
alent.
(1) E has order continuous norm.
(2) Every continuous linear functional on E is order continuous.
(3) E is an ideal in E**.
(4) The order intervals in E are weakly compact.
(5) Every order bounded disjoint sequence in E is norm convergent to
zero.
Since every ideal in a Dedekind complete Riesz space is a Dedekind com-

plete Riesz space in its own right, the following result should be immediate.
2 "order continuity" concept is very general A topology T on a partially ordered set is
said to be order continuous if Xa x implies Xa X.
31t should be noted that if the supremum of every two vectors in an ordered vector space
exists, then the infimum of every pair and y also exists and x A y = [(x) V (y)J.
2.3. Banach Lattices with Order Continuous Norms 85
Corollary 2.27. A Banach lattice with order continuous norm is Dedekind

complete.
An immediate consequence of Theorem 2.26 is that reflexive Banach

lattices have order continuous norm. The class of reflexive Banach lattices
is a proper subclass of the KB-spaces. Recall that a Banach lattice
E is said to be a KB-space (KantorovichBanach space) if every increasing
norm bounded sequence is norm convergent. For details see [30, Section 14].
It is well known that for Banach spaces X and Y the mapping T T*
from L1(X, Y) to (Y*, X*) is a linear isometry. In the case of Banach lattices
this mapping has additional properties related to the order structure. The
following result (which is a special case of a more general theorem due to
U. Krengel [193] and J. Synnatzschke [313]) is quite useful and describes
these properties; for a proof see [30, p. 651.
Theorem 228 (KrengelSynnatzschke). If E and F are Banach lattices
and F has order continuous norm, then the mapping T T*, from F)
to r(F*, E*), is a lattice isometry, that is,
T*I = ITI* and IIT*lIr
for each T F).
Proof. We shall sketch a proof when E is in addition Dedekind complete.

So, assume that E is Dedekind complete, F has order continuous norm, and
let T E
From T [TI, it follows that T* ITI*, and so
To establish the reverse inequality ITI* IT*I we must show that

(ITI*x*,x)
holds for all 0 e F* and all x E
Tothisend,flxOx*EF*andxEE+. SinceTxeFcF**andFis
a Riesz subspace of F**, it follows from Theorem 1.16 that for each z e
we have
= sup (y*,Tz) = sup (T*y*,z)

1y*Ix*
sup
Iy*
That is, for each z E E+ we have
(*)
Now let C be the collection of all vectors in F of the form I
where A = 0 for i j and = x. It is easy to see that the set

is directed upward. Moreover, from ITIxi = ITIx and
Theorem 1.17, it follows that C I lTIx. The order continuity of the norm of
F guarantees that
{x*(c): CE C} I lTix) = (**)
From (*), it follows that for each c = E C we have
x*(c) =
(IT*Ix*,x).
From this and (**), we get (jTl*x*,x) (IT*Ix*,x).
For the norm identity note that

= = =
and the proof is complete.
We continue with the definition of an atom.

Definition 2.29. A vector u > 0 in a Riesz space is said to be:
(1) an atom whenever 0 x u, 0 y u and x A y = 0 imply that
either x = 0 or y = 0 and
(2) a discrete vector whenever for eaCh 0 <x u there exists some
A 0 such that x = A'u, whenever [Q u} = {Au: 0 A I }.
For the proof of the next result see Exercise 2 at the end of this section.
Lemma 2.30. In an Archimedean Riesz space, a positive vector is an atom
if and only if it is a discrete vector. Moreover, in this case, the vector space
{ Au: A E JR} generated by an atom u is a projection band.
A Riesz space is said to be atomless (or non-atomic) if it does not have

any atoms. Our next lemma describes two duality properties of atomless
Banach lattices.
Lemma 2.31. For a Banach lattice E the following statements hold:
(1) If E has order contin'uous norm and is atomless, then its norm dual
E* is likewise an atomless Banach lattice.
(2) If E* is atomless, then E is also atomless.
2.3. Banacli Lattices with Order ContinuousNorms 87
Proof. (1) Let E be an atomless Banach lattice with order continuous norm.
Assume, by way of contradiction, that E* has an atom, say 0 < q5
So, whenever some E E* satisfies 0 q5, then there exists a constant
A 0 such that = Aq5.
The order continuity of the norm of E implies that the linear functional
q5 is order continuous, and so its null ideal = {x E E: q5(IxI) = o} is a
bandand, of course, we have E = = The assumption
q5 > 0 implies {O} (and, moreover, 0 < x E implies q5(x) > 0).
Thus, there exists some 0 <u E Cgo. We claim that u is an atom of E.
To see this, let 0 <x u and 0 y u satisfy x A y = 0. We shall
show that y = 0. Let denote the band projection of E onto the band
generated by x in E. Clearly, q5 a E E* and 0 q5 q5. So, there is
some A 0 such that = Aq5. From 0 <q5(x) = = Aq5(x), we
infer that A = 1 and so q5 = Therefore, q5(y) = = q5(O) = 0,
and by the strict positivity of q5 on its carrier we see that y = 0. Hence,
u is an atom of E which contradicts our hypothesis. Consequently, E* is an
atomless Banach lattice.
(2) Now let E* be atomless, and assume that some 0 < U E is an
atom. Then the vector subspace B= {Au: A E JR} = JRu is a projection
band of E; see Lemma 2.30. So, if A = then E = A B holds. (Both
bands A and B are, of course, closed vector subspaces.) It easily follows
that E is topologically and lattice isomorphic to the Banach lattice A JR,
and so E* is lattice isomorphic to A* JR. Since A* JR has atoms, we see
that E* likewise has atoms, which is a contradiction. Hence, E must be an
atomless Banach lattice.
Exercises
1. Let be a compact Hausdorff space. Show that the Banach lattice C(12)
has order continuous norm if and only if is a finite set. [HINT: Fix any
w E and consider the set
= {x 0 x 1, x(w) 1, and x =outside
0
of some open neighborhood of w}.

Now note that if w is an accumulation point, then 0 holds in
J.
2. Prove Lemma 2.30. [HINT; Assume that u> 0 is an atom in an Archi-

medean Riesz space E and let 0 <x U. Put a = sup {j3 0: j3x U}.
Since E is Archimedean, we see that I a < co. Moreover, ax U, and
we claim that ax = U.
To see this, assume by way of contradiction that v = ax > 0.
U
Since (v v (here we use again that E is Archimedean), it follows
that (v >0 for some k with <cr. Clearly,
On the other hand, [u + > 0. (Otherwise, [u + = 0

implies + u, contrary to the definition of of.) Next, observe that
<
So, if y = + 1)x] + and z = [u + then we have
o <y u, 0 < z u and y A z = 0, contradicting the fact that u is an
atom. Hence, = u.]
3. Show that C[0, 1]-and L1 [0, 1] are atomless Banach lattices.
4. This exercise establishes that in general the norm dual of an atomless

Banach lattice need not be atorniess; compare this with Lemma 2.31.
Show that the norm dual of the atomless Banach lattice E = C[0, 1]
has atoms. [HINT: Every Dirac measure is an atom of E*.
5. If g, h: E are two disjoint non-zero positive linear functionaks on a

Riesz space, then show that there exist disjoint vectors u,v E+ such
that g(u) > 0 and h(v) > 0.
6. Let E be a Riesz space and let 0 < f E". Show that f is an atom
in if and only if f is a lattice homomorphism, that is, if and only if
f(x Vy) = max{f(x),f(y)} for all x,y E.
7. Let S, T: E F be two regular operators between Banach lattices with
F Dedekind complete. Show that:
(a) S T in r(E, F) implies 5* in r(F*, E*).
(b) IT*I ITI*.
Also, give an example where T*f < ITJ* holds true. (Compare this with
Theorem 2.28.)
8. Show that a Banach lattice E is a KB-space if and only if E is a band in
its double dual E**.
9. (DoddsFremlin [109]) Let E and F be two Banach lattices with F having
order continuous norm. If x E+, then show that the set
B = {Tr(E, F):
E T[0, x] is norm totally bounded }
is a band in r(E,F).
2.4. Compact and Weakly Compact Positive Operators

We shall review here some basic properties of compact and weakly compact
operators on Banach spaces and especially on Banach lattices. We start
with the definition of compact and weakly compact operators.
Definition 2.32. An operator T: X Y between Banach spaces is called:
(1) compact if T carries norm bo'unded subsets of X to norm totally
bounded subsets of Y and
Compact and Weakly Compact Positive Operators 89
( 2) weakly compact if T carries norm bounded subsets of X to rela-.

tively weakly compact subsets of Y.
It should be clear that every compact operator is weakly compact and

that weakly compact operators are automatically continuous. It is well
known that a bounded operator T: X Y between Banach spaces is com-
pact (resp. weakly compact) if and only if for every norm bounded sequence
{ of X, the sequence has a norm convergent (resp. a weakly con-
vergent) subsequence in Y.
We summarize below the basic properties of compact and weakly com-
pact operators between Banach spaces. Proofs can be found in any book on
functional analysis; see, for instance [30, Section 16].
o Sums and scalar multiples of compact (resp. weakly compact) oper-
ators are compact (resp. weakly compact) operators.
o The composition of operators at least one of which is compact (resp.
weakly compact) is a compact (resp. weakly compact) operator.
o The norm limit of a sequence of compact (resp. weakly compact)
operators is a compact (resp. weakly compact) operator.
o (Gantmacher [132]) An operator is compact (resp. weakly compact)

if and only if its norm adjoint is compact (resp. weakly compact).
The collection of all compact (resp. weakly compact) operators from a
Banach space X into another Banach space Y will be denoted by 1C(X, Y)
(resp. 1'V(X, Y)). Clearly, 1((X, Y) and W(X, Y) are both closed subspaces
of (X, Y).
Now we introduce the notion of domination between operators.
Definition 2.33. Let T: E F be a positive operator between Riesz spaces.
We shall say that an operator S: E F is dominated by T (or that T
dominates S) whenever lSx! <T(lxI) holds for each x E E.
The following simple facts are immediate:
(1) A positive operator T dominates S if and only if T S T.
(2) If S is positive, then T dominates S if and only if 0 S T.
(3) If F is Dedekind complete, then S is dominated by T if and only
if ISIT.
There is an important collection of results regarding compactness
erties of operators between Banach lattices that are dominated by compact
or weakly compact operators. We list them below and for proofs and details
we refer the reader to [30]. There is no parallel for these results on general
Banach spaces.
Theorem 2.34 (AliprantisBurkinshaw [27]). If in the scheme of continu-

ous operators between Banach lattices E F C H each opera-
tor Si is dominated by a compact positive operator, then 535251 is a compact
operator.
Corollary 2.35 (AliprantisBurkinshaw [27]). If an operator 5: E + E

on a Banach lattice is dominated by a compact positive operator, then is
a compact operatoralthough need not be compact.
Theorem 2.36 (AliprantisBurkinshaw [27]). If in the scheme of contin-

uous operators between Banach lattices E F C each operator
is dominated by a compact positive operator, and either E* or C has order
continuous norm, then S2S1 is a compact operator.
Corollary 2.37 (AliprantisBurkinshaw [27]). Assume that E is a Banach
lattice such that either E or E* has order continuous norm. If an operator
5: E + E is dominated by a compact positive operator, then its square 52
is compactalthough S need not be compact.
Theorem 2.38 (DoddsFremlin [109J). If E and F are Banach lattices

such that E* and F both have order continuous norms, then every operator
S: E + F that is dominated by a compact positive operator is compact.
Theorem 2.39 (AliprantisBurkinshaw [28]). If in the scheme of contin-

uous operators between Banach lattices E F C each operator
is dominated by a weakly compact positive operator, then S2S1 is a weakly
compact operatoralthough neither nor 82 need be weakly compact.
Theorem 2.40 (Wickstead [335]). If E and F are Banach lattices such

that either E* or F has order continuous norm, then every positive operator
5: E + F that is dominated by a weakly compact positive operator is weakly
compact.
Exercises
1. Show that a positive operator T: E F between two Riesz spaces dom-
inates an operator 8: E F if and only if T S T.
2. Let T: E F be a positive operator between two Banach lattices. Show
that for an operator 8: E F the following statements are equivalent.
(a) T dominates S.
(b) T* dominates 5*
(c) T** dominates
[HINT: Use Exercise I above.
2.4. Compact and Weakly Compact Positive Operators 91
3. Prove the following basic properties of compact operators.

(a) Sums and scalar multiples of compact operators are colulnct,
ators.
(b) The composition of two (or more) bounded operators at ()[le of
which is compact is likewise a compact operator.
(c) Bounded finite-rank operators are compact. (Recall that
between two vector spaces is said to be finiterank if itH range is
finite dimensional.)
(d) A bounded operator is compact if and only if its adjoint likewise
a compact operator.
(e) The norm limit of a sequence of compact operators is
operator.
[HINT for (d): Assume that T: X Y is compact and conslitpr Y as
a closed vector subspace of C(Ufl, where is the closed ball of
equipped with its w*..topology. Now fix e and kt > 0.
Since is a norm totally bounded subset of Y, it Iroin the
classical AscoliArzel theorem that is an equicontinuouH I)Ounded
subset of C(Ufl. Consequently, there exists a of yK
satisfying [Tx(y*) Tx(y*)I = x(T*y* < for all e and
all x E Thus, T*z*11 = supXEUX Ix(T*y* T*z*)I
for all in If yr,.. .
, E are chosen so that
then T* (U*) E), proving that T* (U*) is a non ii totally
bounded subset of
4. (Gantmacher [132]) For a bounded operator T: X Y Banach

spaces show that the following statements are equivalent.
(a) T is weakly compact.
(b) T**(X**) c Y.
(c) T*: (Y*,o.(y*,Y)) 9 (X*,o.(X*,X**)) is continuous.
(d) T*: is weakly compact.
5. Prove properties similar to the ones listed in Exercise 3 above for weakly
compact operators.
6. Show that the range of an arbitrary compact operator is separable.

[HINT: Let T: X Y be a compact operator between Banacli spaces.
Then is a norm totally bounded subset of Y and hence a separa-
ble metric space. If A is a norm dense countable subset of T((Jv), then
B = nA is a norm dense countable subset in R(T).]
7. Show that a subset of a normed space is norm totally bounded if and only
if it is contained in the closed convex hull of a sequence that coxiverges to
zero.
8. (Terzioglu [318]) Show that an operator T: X Y between t,w) Banach

spaces is compact if and only if there exists a sequence }
isfying 0 and $ x E X. Use

the preceding exercise.]
9. Show that an operator T: X Y between Banach spaces is compact if

and only if T factors with compact factors through a closed subspace of
CO, i.e., if and only if there exists a closed vector subspace Z of CO and
compact operators X Z Y such that T = RS.
10. Let 'r: ci be a continuous mapping on a compact Hausdorff space,
and let T: be the composition operator defined as usual by
Tf = foT. When is T a compact operator? [HINT: Use the AscoliArzel
theorem.]
11. Let T: Y be a bounded operator, where Y is a Banach space. Show
that T is weakly compact if and only if the set n = 1,2, . . .} is
a relatively weakly compact subset of Y (where denotes the sequence
whose entry is one and every other is zero).
12. For a Banach space X prove the following.
(a) An operator T c (X, 4) is weakly compact if and only if T is
compact.
(b) An operator T c (CO, X) is weakly compact if and only if T is
compact.
13. Let {Yi, Y2,. . .} be a countable relatively weakly compact subset of a
Banach space Y; in particular, this is so if is a weakly convergent
sequence in Y. Show that the operator T: Y, defined by
is weakly compact.
Chapter 3
Operators on and
spaces
Banach lattices possess a property that has no parallel in the general theory
of Banach spaces. It can be described by saying that: Every Bariach lattice
is locally "classical." Specifically, if E is a Banach 'attice and x is any vector
in E, then the principal ideal generated by x, i.e., the vector subspace
can be identified with a This important and fundamental

resultwhich is known, as the KakutaniBohnenblustKrein representation
theoremallows one to bring into the study of operators the structure of
In particular, one can view the vectors of a Banach lattice
as functions and introduce a "local" multiplication operation between the
vectors of a Banach lattice. These multiplication operations define various
classes of "local" multiplication operators that in many instances can be
extended to operators on the whole space. This special feature of having
many "local" multiplication operators on a Banach lattice is discussed in
detail in this chapter.
Abstract Banach lattices that behave like the classical L1(p)- and
spaces are called and AM-spaces, respectively. The purpose of this
chapter is to introduce AL-spaces ("Abstract Lebesgue" spaces) and AM-
spaces ("Abstract Maximum" spaces) and to study their structures as well as
the operators acting between these spaces. During our investigation, we shall
demonstrate that many important operators between Banach lattices behave
"locally" as operators between L1Ci4- and In particular, we
shall consider abstract multiplication operators and lattice homomorphisms.
93
94 3. Operators on AL- and AM-spaces
As we shall see, these operators are in essence the "building blocks" for the
general theory of positive operators.
3.1. AL- and AM-.spaces

Two important classes of Banach lattices are provided by the AL- and
spaces. As mentioned before, these Banach lattices play a fundamental role
not only in the study of positive operators but in functional analysis in
They are defined as follows.
Definition 3.1. A Banach lattice E is said to be:
(1) an if = IIxll+liy1l for allx,y E with xAy = 0
and
(2) an if lix V = max{IJxJI, ffyjJ} for all x,y E with
xAy= 0.
Recall that in an ordered vector space X a vector e E X+ is said to
be an order unit (or a strong unit, or simply a unit) if for each x E X
there exists some A 0 stich that x Ae. Notice that if X is a Riesz space,
then a vector e E X+ is an order unit if and only if for each x E X there
exists some A 0 such that fxl Ae. (See also Exercise I at the end of this
section.)
A Banach lattice E is said to be an AM-space with unit if in addition
to being an AM-space, E also has an order unit. If E is an AM-space with
a unit e, then the formula
IxlIoo =inf{A>0: lxi Ae}=min{A 0: xJ< Ae} (*)
defines an eqriivalent lattice norm on E. Under this new norm, E remains an
AM-space and its closed unit ball coincides with the order interval [e, e];
see Exercise 2 at the end of this section. Unless otherwise stated, we will
a'ways assume that an AM-space with unit is equipped with the norm given
by (*) for some order unit e.
The norm of a positive operator acting on an with unit can
be comptited easily.
Lemma 3.2. If E is an AM-space with unit e, F is an arbitrary Banach
lattice and T: E F is a positive operator, then ITII = llTeJI.
The and the AM-spaces are dtial to each other. For the proof
of the first part of the next theorem see [30, p. 188].
Theorem 3.3. A Banach lattice E is an AL-space (resp. an AMspace) if
and only if its norm dual E* is an AM-space (resp. an AL-space). More-
over:
3.1. AL- and 95
(a) If E is an AL-space, then E* is a Dedekind complete AM-space

with unit e*, where e*(x) = 1x1 [xl for each x E E.
(b) If E is an AM-space with unit e, then E** is also an AM-space
with unit e.
P roof. We prove parts (a) and (b) only.
(a) Let E be an first part of the theorem, we know
By the
that E* is an AM-space. Now consider the function e*: E+ p JR defined by
e*(x) = Since E is an AL-space, it follows that e* is additive, and so (by
Theorem 1.15) it extends uniquely (via the formula e*(x) = lxi)
to a positive linear functional on E, i.e., e* E We leave the details to
the reader to verify that e* is an order unit of E*.
(b) Assume that E is an AM-space with unit e. We claim that e is also a
unit in the AM-space E**. To see this, note first that Mx*1I = x*(e) = e(x*)
for each E E then for each E we have
x**(x*) <
and so x** < proving that e is a'so a unit in E**.
Recall that the principal ideal generated by a vector u in a Riesz space

E is the vector subspace
= {x E E: there exists some A 0 such that xj $
It turns out that when E is a Banach lattice, then each principal ideal has
the structure of an AM-space. For a proof of the next result see [30, p. 187].
Theorem 3.4. Let E be a Banach lattice and let be the principal ideal
generated by a vector u E E. Then under the norm
= inf{A?0: I
is an AM-space with unit
The next two classical results allow us to consider the vectors of AL-
and AM-spaces as functions in some familiar function spaces.
Theorem 3.5 (KakutaniBohnenblustNakano). A Banach lattice is an

AL-space if and only if it is lattice isometric to an L1 (ii) -spaceS
Theorem 3.6 (KakutaniBohnenblustKrein). A Banach lattice E is an
AJvI-space with unit e if and only if E is lattice isometric to some
space for a unique (up to homeomorphism) compact Hausdorff space in
such a way that the unit e is identified with the constant function 1 on
Partherrnore, a Banach lattice E is an APvI -space if and only if E is
lattice isometric to a closed vector sublattice of a
96 Operators onAL- and AM-spaces
The reader should keep in mind that for a compact Hausdorif space the
Banach lattice is Dedekind complete if and only if is extreinally
disconnected or Stonean (i.e., if and only if the closure of every open
subset of is also open) . For details and more on these classical theorems
we refer the reader to [30, pp. 192195].
Corollary 3.7. Every AL-space has order continuous norm.
Lemma 3.8. If E is either a or a Dedekind complete AM-space
with unit, then there is a positive projection P: E** with range E.
Proof. If E is a KB-space, then E is a band in E** (see Exercise 8 in
Section 2.3) and the band projection from E** onto E is the desired positive
projection.
Now assume that E is a Dedekind complete AM-space with unit. By
Theorem 3.3 the unit of E is a unit of E** too, and so E majorizes1 E**.
In particular, the identity operator I: E E has a positive extension to
all of E**; see [30, Theorem 2 p. 26]. Any such extension is a positive
projection on E** with range E.
Theorem 39. Assume that either F is a Dedekind complete
with unit or E is an AL-space and F is a KB-space. Then every continuous
operator from E to F is regular, i. e., (E, F) =
(E, F) has a modulus [TI.
Proof. Assume first that F is a Dedekind complete AM-space with unit,
and take any continuous operator T: E F. Since T maps the unit ball
UE of E to a norm bounded subset of F, the set T(UE) is necessarily order
bounded, and thus T is an order bounded operator. As F is Dedekind
complete, T is a regular operator and its modulus TI exists.
Assume next that E is an a KB-space. Consider an
arbitrary T E (E, F). Then T*: F* and E* is a Dedekind complete
AM-space with unit. According to the preceding case, the operator T*
is regular. This implies that T**: E** F** is also a regular operator.
Pick two positive operators A, B: E** F** such that T** = A B. By
Lemma 3.8 there exists a positive projection P from F** onto F. It remains
to notice that PA, PB are positive operators from E** to F and the equality
T = PA PB holds on E. This shows that T is a regular operator. D
Corollary 3.10. Assume that T: E F is a continuous operator, where
either F is a Dedekind complete AM-space with unit or E is an AL-space
and F is a KB-space. Then = lTIIr, the operator norm and the
r-norm of T coincide.
Recall that a vector subspace Y of an ordered vector space Z xnajorizes Z if for every
z E Z there exists some y E Y such that z < y.
3.1. AL- and AM-spaces
In particular, in these cases, (E, F) equipped with the operator norm.

is a Dedekind complete Banach lattice.
Proof. Assume first that F is a Dedekind complete AM-space with unit

and so, by Theorem 3.6, F = for some compact Hausdorif Space
As established in Theorem 3.9, the operator TI: E f F exists and
= JTj sup ITIx
lix II <1
Fix
Since
0 and pick some unit vector x E such that JTjx L> .
=
we can find a non-empty open subset V of such that
(lTIx)(w) > E (*)
for each w E V. The RieszKantorovich formula
ITIx=sup{Tu: x<u<x}
implies the existence of some u E [x, xl and some point wO E V such that
(ITiu)(wo) >
(Otherwise the values of the function Tix on the set V would be bounded
from above by E, contradicting (*).) This implies >
and hence 11Th > As 0 is arbitrary, it follows that 11Th
The converse inequality 11Th is always true, and consequently we
have the desired equality 11Th = 1ITJr.
-Now assume that E is an AL-space and F is a KB-space. The modulus
TI: E F exists by Theorem 3.9. Furthermore, Theorem 2.28 guarantees
that ITI* = T*f. Since E* is a Dedekind complete AMspace with unit, the
first case yields 1
= JIT*II and so
= lIT) 1= JT!* = 1! = = 11Th.

The proof is finished.
The preceding result combined with Corollary 3.7 and Theorem 2.28
yields the following.
Corollary 3.11. If E is an AL-space and F is a KB-space, then the map-
ping T T*, from .C(E, F) to (F*, E*), is a lattice isometry.
We continue with some basic properties of weakly compact operators on

AL- and AA'I-spaces.
Theorem 3.12. The following two statements are true.

(1) Every weakly compact operator between AL-spaces has a weakly
compact modulus.
(2) Every weakly compact operator from an AM-space to a Dedekind
complete AM-space with unit has a weakly compact modulus.
Proof. Statement (1) is due to C. D. Aliprantis and 0. Burkinshaw [28].

For a proof see [30, Theorem 17.14, p. 292].
We prove (2) (which is due to K. D. Schmidt [296]) by using the fac-
torization technique outlined below. Let T: E F be a weakly compact
operator from an AM-space E to a Dedekind complete AM-space F with
unit. Note that IT**1: E*
F* is weakly compact, part (1)
antees that its modulus IT*t: F* E* is also weakly compact. Now by
Theorem 2.28, we have = IT*1*, and so T**f is weakly compact. To
complete the proof, we shall show that the modulus TI of T admits a fac-
torization of the form
JTI
E F
N F**
with S a positive weakly compact operator and P a positive projection.

The projection P: F** F** is the one guaranteed by Lemma 3.8. The
operator 5: E F** is the restriction of the weakly compact operator
E** + F** to E.
To see that the equality = PIT**f holds, let x E
E E satisfies lvi x, then from the inequality Ty = T**y T**Ix, we
obtain Ty = PTy PIT**1x E F, and so
TIx =sup{Ty: yE E and JyJ x}
where the supremum is taken in F. On the other hand, if E E** satisfies
x, then there exists (see Exercise 10 at the end of the section) a net
{ in the interval [x, xJ such that Xa ->
From the inequality
T**Xa = TXa Tlx, we get T**x** ITIx and so
I
T**lx = E E** and <x} ITJx,
where the supremum is now taken in Thus, PIT**lx =

and the desired equality ITIx PIT**ix follows.
3.1. AL- and AM-spaces 99
In order to obtain the corresponding results for compact operators, we

need a lemma.
Lemma 3.13 (Krengel [193J). If A is a norm totally bounded subset of an
AM-space E, then.'
(1) The subset D of E consisting of the suprema of the finite subsets
of A is likewise norm totally bounded.
(2) supA exists in E.
(3) sup A D, the norm closure of D in E.
Proof. By Theorem 3.6, we can assume that E is a closed Riesz subspace

of some with compact and Hausdorif. Thus, A can be viewed
as a norm totally bounded subset of C(Q), and hence, by the AscoliArzel
theorem, A is norm bounded and equicontinuous; see for instance [31, p. 66].
Now let g = fj E D, where E A for each i. Since A is equicon-
tinuous on given 0 and w there exists a neighborhood V of w
such that If(t) f(w)1 < for all t V and all f A. This implies
- g(w)1 max Ifj(t) - fj(w)I
for all t V and all ,q E D. But then, by the AscoliArzel. theorem again,
D is also a norm totally bounded set.
Clearly, D is an upward directed net, and hence D (being norm totally
bounded) has a norm convergent subnet to some g E E. Since D is upward
directed, it follows that D is norm convergent to g. Hence, ,q D and
= sup D sup A; see Lemma 1.30(c).
We are now ready to state and prove the analogue of Theorem 3.12 for
compact operators.
Theorem 3.14. Let T: E + F be a compact operator between two Banach

lattices. If either
(a) F is an AM-space, or else
(b) E an AL-space and F is a KB-space,
then T has a compact modulus that is given by the RieszKantorovich for-
mula
ITIx=sup{Ty: yEE and IyIx}.
Proof. (a) Let T: E F be a compact operator. Assume first that F is
an AM-space. Notice that Lemma 3.13 guarantees that in the formula
TIx=sup{Ty: yEE and
100 3. Operatorson and
the supremum exists in F for each x E . This establishes that the

operator T has a moduluseven when F is not Dedekind complete!
On the other hand, if U is the closed unit ball of E and D denotes the
subset of F consisting of the suprema of the finite subsets of the norm totally
bounded set T(U), then (by Lemma 3.13) D is a norm totally bounded
subset of F. Moreover, the preceding formula implies that for each x
we have ITIx D, i.e., TI is
that is a KB-space. By
Theorem 3.9, the modulus TI of T exists. Next, notice that T*: F*
a compact operator and E* is a Dedekind complete AM-space with unit. By
the preceding case, T* has a compact modulus IT* I. Finally, Theorem 2.28
implies ITI* IT* , and so ITI* is a compact operator. The latter guarantees
that itself is a compact operator.
We close the section by introducing two special classes of positive op-

erators acting on AM.- or AL-spaces. These are the classes of Markov and
stochastic operators.
Definition 3.15. A positive operator T: E F between two AM-spaces
with units u and v, respectively, is said to be a Markov operator if T'u = v.
A positive operator T: E +F between two AL-spaces is called a sto-
chastic operator if ITxII = lxii for each x
For some simple examples of Markov and stochastic operators see
cise 15 at the end of this section. The Markov and stochastic operators are
dual to each other in the following sense.
Theorem 3.16. A positive operator between two AM-spaces with units is
a Markov operator if and only if its norm adjoint is a stochastic operatorS
Similarly, a positive operator between two AL-spaces is a stochastic op-
erator if and only if its norm adjoint is a Markov operator.
Proof. Assume first that T: E F is a Markov operator between two

AM-spaces with units u and v, respectively. So, T is positive and Tu = v.
Now if 0 y* F*, then according to Lemma 3.2 we have
= IIT*y*(u)II = =
This shows that T* is a stochastic operator.
Next, suppose that T: E F is a stochastic operator between two
That is, T is positive and 1TxM = lxii for each x By
Theorem 3.3, the norm duals E* and F* are both AM-spaces with units
and AM-spaces 101
and given by u*(x) = for x E B and v*(y) = 11y11

for y E F. Thus, if x E then
T*v*(x) = v*(Tx) = ITxli = fJxjI = u*(x).

Therefore, T*v* = and so T* is a Markov operator.
Now let T: E F be a positive operator between two AM-spaces with

units u and v, respectively, and assume that T*: F* is stochastic.
Then T**: E** F** is a Markov operator. Since E** is an AM-space
with unit u and F** is an AM-space with unit v (see Theorem 3.3), it
follows that Tu = T**u = v. Hence, T is a Markov operator.
Finally, assume that the norm dual T* of a positive operator T: E F
between AL-spaces is a Markov operator. Then T**: E** F** is a sto-
chastic operator. Consequently, if x E then lxii = IIT**xII = IlTxlI
proving that T is a stochastic operator. The proof of the theorem is now
complete.
Exercises
1. A vector u in a subset A of a vector space X is called an internal point
of A if for each x E X there exists some > 0 such that u + Ax E A for
all Al A0.
Show that in an ordered vector space X the strong units are precisely
the internal points of the positive cone
2. If E is an AM-space with a unit e, then show that the formula
= inf{A>0: xI Ae} = min{A 0: JxJ Ae} (*)
defines an equivalent lattice norm on E. Under this norm E remains an

AiVb-space and its closed unit ball coincides with the order interval [e, e].
3. Prove Lemma 3.2.
7. Show that a Banach lattice E is:

(a) an AL-space if and only if lix + yE and
(b) an AM-space if and only if for all x, y E
8. Show that the lattice operations in any AiVI-space E are weakly se-
quentially continuous. [HINT: We can assume (by Theorem 3.6) that
E= for some compact Hausdorif space ft By the Riesz Represen-
tation Theorem, we know that a sequence in satisfies f
if and only if is norm bounded and f(w)
f in then also
3. Operators on AL- and AM-spaces
9. Show that the lattice operations in an AL-space need not be weakly se-
quentially continuous.
10. Let E be a Banach lattice. For x E
E E**: x x}.
Establish the following:
(a) [x, xl and ftx, 4 are both convex sets.
(b) [x, x] is a(E, E*)_closed and is a(E**, E*)_compact.
(c) [x,x] is a(E**,E*)_dense in
[HINT: To see (c), assume that [x, x] is not a(E**, E*)_dense in ftx, x}J.
Then, by the Separation Theorem, there exists some E E* satisfying
x*(y) = 0 for each y E [x,x] and x*(y**) 0 for some E It
follows that
Ix*I(x) = sup{x*(y): yE E and x <y x} 0.
But then 0 < < = 0, and so x*(y**) = 0, a

contradiction.]
11. Show that every regular operator from an AM-space with unit to any
p-space, where 1 p < oo, is compact. [HINT: Use the fact that the
order intervals of 4-spaces with 1 p < oo are norm compact sets.]
12. Let E be an AM-space. Extend the norm of E to its Dedekind completion
as in Exercise 21 of Section 1.2. Show that the Banach lattice E'5 is
also an AM-space. If E is an AM-space with unit e, then Ea is also an
AM-space with unit e.
13. Assume that S and Q are two compact metric spaces, is an arbitrary
(finite) Borel measure on 5, and K: S x Q + is a continuous function.
Also let T: C(S) C(Q) denote the hitegral operator with kernel K,
i.e., T is defined by
Tx(w) = fK(s, w)x(s) dy(s)

for each x E C(S) and each w E ft Establish the following.
(a) T is a compact operator.
(b) The modulus TI: C(S) C(Q) of the operator T (which exists
according to Theorem 3.14) is also an integral operator satisfying
TIx(w) = for all xE C(S) and alIw EQ.
(c) The norm of T is given by
= sup
f 5K(s,w)1 dy(s).
[HINTS: (a) Let d and p be the metrics on S and Q, respectively, and
let c > 0. Fix some S > 0 such that d(s1, S2) + p(wi, W2) < S implies
K(si,wi) < c. Now if x E C(S) satisfies 1 and
w1,w2 in Q satisfy p(w1.w2) <5, then
Tx(wi) f K(s,wi) x(s)I dy(s)

3.2. Complex Ban ach Lattices 103
This shows that the norm bounded set {Tx: 1} is an equicon-

tinuous subset of
(b) Fix any function 0 x E C(S) and let w E ft Clearly, we
have the inequality TIx(w) Denote by a the
sign function of K(.,w), i.e., if K(s,w)
a(s) = 1 0 and a(s) = 1
if K(s,w) < 0, and then choose a sequence C(S) such that
a(s) for ji-almost all s. Replacing by V (1)] A 1, we
can assume that 1 for each n. Since x(s) for
each s E S and n, it follows from the Lebesgue Dominated Convergence
Theorem that
= fK(s, w)a(s)x(s)
= lim I K(s,
Js
(c) Clearly, w)I dji(s). For the converse inequal-
ity pick some w0 E with wo)I dji(s) = w)I d,ii(s).
Denote by a the sign function of K(.,wo) and choose a sequence

in C(S) such that 1 for each n and a(s) for ji-almost
all s. Now use the Lebesgue Dominated Convergence Theorem to see that
= lim I
TL+OO
14. Show that every Markov operator carries order units to order units.
15. Every Euclidean space can be considered either as an AM-space with
unit e = (1, 1,. .. , 1) or as an AL-space under the usual norms
and
l<i<TL
i=1
respectively. An rn x n withnon-negative real entries
is called Markov (resp. stochastic) if = 1 for each 1 < i rn
(resp. = 1 for each 1 j n).
Now let A be an m x n matrix with non-negative entries and consider
A as an operator from R7L to Show the following:
(a) The positive operator A: . II II
is a 1\'Iarkov
operator if and only if the matrix A is a Markov matrix.

(b) The positive operator A: (W1, is a stochastic.
operator if and only if the matrix A is a stochastic matrix.
3.2. Complex Banach Lattices

Apart from the standard norm introduced in Section 1.1, there is another
way of defining an equivalent norm on the complexification Ec of an arbi-
trary real Banach lattice E. This new norm is very important because, as we
shall see, it is closely related to the lattice structure of E. To introduce this
104 3. Operators on and
new norm, assume first that E C(f1), the Banach lattice of all
continuous functions on a compact Haiisdorff space ft2 An easy argument
shows that for each pair f,g C(Q) and each w E Q, we have
sup {f(w) cos 9 + g(w) sin 9] sup [f(w) cos 9 + g(w) sin 9]
OE[O,21rj OER
= + [g(w)}2.
Conseqriently,
sup [f cos 9 + g sin 9] = sup [f cos 0 + g sin 9]

OE[O,2ir] OER
=
where the supremum is taken in in the usual lattice sense.
We now turn our attention to the general case of a real Banach lattice
B. Let x, y E E and consider the ideal generated in E by the element
u = lxi + By Theorems and 3.6, we know that is lattice isometric
to a This implies (in view of the preceding discussion) that the
supremum
sup [x cos 9 + y sin 9] = sup [x cos 9 + y sin 9]
OE[O,2ir] OER
exists in (and hence in E). The above sripremrim is called the modulus
of the element z = x+zy E That is, the modulus of an arbitrary element
z = x + zy E is the positive element izl of B defined via the formula
= sup [x cos 9 + y sin 9].
OE[O,21r1
Notice that
lzt= xzyj = = (=
It can be shown that for arbitrary z, z1, z2 E and A C, we have:
(1) zI 0 and IzI=0 if and only ifz=O;
(2) = Al . Izi; and
(3) < + IZ2I.
So, if we define the function 11 Ic: JR via the formula

IIzlIc = , (*)
21f we want to emphasize that we deal with real-valued functions rather than with complex-
valued functions, then we shall denote by CR(cl). The symbol Cc(cl) will denote the complex
Banach space of all complex-valued continuous functions on ci equipped with the supremum norm.
Complex Banach Lattices 105
then the above properties (1), (2), and (3) guarantee that is, in fact,
a norm on Now notice that if z = x + iy E then the inequalities
xI Izi, yI and Izl jxf + y[ imply that
+ lylf) lxii +
Therefore, not only is Ic a norm on but it is also equivalent to the
standard norm on defined by
I
z = sup lix cos 0 + y sin 0
OE[O,2ir)
The reader should also observe that = jIziI for all z E E and Izil
implies z1 Ic
From now on, we will always assume that on the complexification
E the norm is defined by (*). Specifically, we have the
following definition.
Definition 3.17. Any complex Banach space of the form

E is a real Banach lattice, will be referred to as a complex Banach
lattice.
The rest of this section will demonstrate the advantages of this norm. A
linear isometry T: between two complex Banach lattices is said to
be a lattice isometry if ITzI = for each z E Two complex Banach
lattices = B zE and F zF are said to be lattice isometric if
there exists a surjective linear isometry T: satisfying = TIzI
for all z E We have the following.
Lemma 3.18. If T: B F is a surjective lattice isometry between two

real Banach lattices, then Tc: k is also a surjective lattice isometry.
Proof. Assume that T: E + F is a surjective lattice isometry. To show

that is a lattice isometry, pick z = x + E If D denotes the
subset of all vectors of the form (x cos + Y sin where is a real
number for each i, then D I Since T is order continuous, it follows that
T(D) I in F. On the other hand, if w = + YsinOi) E D,
then T(w) = + and so T(D) I TczI in F.
Hence, Tizi for each z E Ec. Finally, from
11
1 = IIzII
we see that Tc is indeed a surjective lattice isometry.

Definition 3.19. A complex AM-space F with unit is the complexifica-

tion of a (real) AM-space E with unit, i. e., F = E iE. The unit of
E will also be referred to as the unit ofF.
Two complex AM-spaces F and C with units u and v, respectively, are
said to be lattice isometric if there exists a surjective lattice isometry
T: F C satisfying Tn = v.
From Lemma 3.18, it should be clear that the complex version of the
KakutaniBohnenblustKrein Representation Theorem 3.6 can be stated as
follows.
Theorem 3.20. If E iE is a complex AM-space with unit e, then
there exists a compact Hausdorff space (which is uniquely determined up
to homeomorphism) such that is lattice isometric to the complex AM-
space = and e corresponds to the constant function I
Corollary 3.21. If u is a positive vector in a real Banach lattice E, then

the complexification of the ideal
EuEBiEu={zEEc:
is lattice isometric to a complex Banach lattice of the form
C z the
modulus of z in the
complex Banach lattice E
E and F be two real Banach lattices. For an operator T in
(E, F) its norm ITlic is, by definition, the operator norm of the operator
Tc: where as usual + iy) = Tx + iTy.3 That is,
IlTIIc sup iITczJIc.
IIzIIc1
It should be obvious that for any bounded operator T: E E we have
the estimates <211Th. Moreover, for positive operators the two
norms and coincide. This is a consequence of the following result.
Lemma 3.22. If T: E F is a positive operator between (real) Banach
lattices, then ITczI for all z E
Proof. Let z = x + iy E Then Tcz = Tx + iTy and from
(Tx) cos 8 + (Ty) sin 8 = T(x cos 8 + y sin 8)
we get = [(Tx) cos8 + (Ty) sin 8] as claimed.
3To avoid unnecessary notation, we shall use the symbol T instead of whenever this does
not cause any ambiguity.
3.2. Complex Banach Lattices 107
Corollary 3.23. If T is a positive operator, then JiTlic = IITIJ.
In general, strict inequality < is possible even in the finite

dimensional case; see Exercise 9 at the end of this section.
The rest of the section is devoted to the concept of a regular operator
on a complex Banach lattice. So, for our discussion here, let E and F
be two real Banach lattices. Recall that (E, F) zr(E, F)
since every bounded operator from to is of the form T + iS, where
T,sEr(E,F), and
(T + zS)(x + zy) = (Tx Sy) + z(Sx + Ty)
for z = x + zy E
all We shall say that an operator T + iS Fc) is
regular if both operators T, S (E, F) are regular. The collection of all
(complex) regular operators will be denoted 4(Ec, Fc). Clearly, the vector
space 4(Ec, Fc) coincides with the complexification of 4(E, F), that is,
4(Ec, Fc) = F) F).
From now on, let us assume that F is also Dedekind complete. Then,
by the RieszKantorovich Theorem 1.16, 4(E, F) is a Dedekind complete
Riesz space which is a Banach lattice when equipped with the r-norm; see
Theorem 1.32. In particular, 4(Ec, F) F) is a
plex Banach lattice. Moreover, every regular operator T + iS has a modulus
IT + zSJ in Fe). Remarkably, the modulus JT + zSI is given (just as in
the real case) by the RieszKantorovich formula. The details are included
in the next result.
Theorem 3.24. Let E and F be two real Banach lattices with F Dedekind
complete, and let T = T + iS E 4(Ec, Fc) = F) irr(E, F). Then
the modulus ITI of T satisfies the RieszKantorovich formula, i.e., for each
u E E+ we have
= JT + iSfu = sup (T + zS)zI = sup (T + zS)(x + zy)J.

kL<u
Proof. Let T = T + iS 4(Ec, and fix u E Et It should be clear

that a = Tzl exists in F. We must show that a = JTfu.
Recall =
that ITI cos 0+8 sin 0], where the supremum is taken
in the Dedekind complete Riesz space 4(E, F). Consider the subset D of
F consisting of all vectors of the form v = cos + sin
where Ok,.. R and ul,.
. , are positive vectors in E satisfying
. . ,
+'' = u. It is well known that sup D = TIu; see [30, Theorem 1.16,
p. 15].
Now let v = cos + sin Put x =

E D. cos
and y = sin and consider the vector z = x iy E From
x cos
cos sin that

Tz (T + zS)(x zy) = (Tx + Sy) + z(Sx Ty) v+ z(Sx Ty)
implies
v < JvJ < 1v+z(SxTy)1 = ITzI a
for all v E D. Since supD = Tju, we see that
ITJua= sup TzI. (*)
For the reverse inequality, we shall need the following two (easily
checked) simple facts:
(1) If w E satisfies 1w! v for some v E then TwJ +
[SJ)v.
(2) If v E then JTvj lTlv.

Now let z E satisfy u, and fix 0. Then there exist positive
vectors u1,. ,in E and complex scalars cu,.. such that
. , 1
for each i = 1,. .. , n, = 1, and Iz see Exercise 12
at the end of the section. Letting R = JTj + 181, we have
ITzt =
T(z + +
Ru + E a ITlu
is also true. This inequality, combined with (*), establishes the validity of
the RieszKantorovich formula.
Here are three useful consequences of the preceding results.

2. Complex Banach Lattices 109
Corollary 3.25. If E and F are two Banach lattices with F Dedekirtd

complete, then every regular operator T E satisfies
Tzl < [T1([zf)

for all z E
E is a real Banach lattice, then the norm dual of the

complexification E coincides with the complexification of E*, i.e.,
= Moreover, if f E then its mothll?ls satisfies the familiar
RieszKantorovich formula, i.e.,
If 1(u) = sup
for each u E E+.

Corollary 3.27. If E is a Dedekind complete Bartach lattice, then the com-
plex Banach lattice under the r-norm is a complex unital Banach
algebra.
Proof. By Theorem 1.32, we know that = 4(E) i12#r(E) is a

complex Banach lattice. Now let 8, T E and fix some z E
satisfying Izl u for some u E E+. Then we have
= IS1 (ITzl) SI(ITlu) = (ISHTDu,
and from this and Theorem 3.24 we get 1ST) TI. Consequently,
= IIISIITIII M18)M = .
Hence, is a complex unital Banach algebra.

Finally, we can extend the notion of domination to operators between
complex Banach lattices. We shall say that a positive operator T: E F
between (real) Banach lattices dominates an operator 8: (or that
S is dominated by T) if rSzl TIzI for all z E It should be noted
that if an operator between two complex Banach lattices is dominated by a
positive operator, then it is automatically continuous.
Exercises
1. If E is a Banach lattice and z = x + zy E then show that
= sup x cos 0 + y sin 01 = sup Ix cos 0 + y sin

OE[O,27r1
2. If E is a Banach lattice, then for arbitrary z. z1, Z2 E and C show

that:
(a) 0, and zj = 0 if and only if z = 0.
(b)
3. Operators on AL- and
(c)
Use the above properties to verify that the function [1. + R, defined
by =1 is a norm.
3. Show that the complexification of coincides with Moreover,
show that for each 1,9 we have If + = +g2 and hence
If k/f2
4. Show that the complexification of R72 coincides with C72. Also, show that
all norms on (or R72) are equivalent. [HINT: Let . be a norm on
C12 and let . denote the 1-norm, i.e., Iz Iii = An easy
argument shows that the identity mapping I: _ (C72, 1) 11 1
is continuous. This implies that the set K = {z E C12: = 1} is II
a Il-compact subset of C72. The continuity of the mapping z i JIzII

guarantees that m = minZEK > 0 and M = IlzII <oo. Now
use the identity = to get mIIzIIi A'IIfzIIi for all
z E Ca,]
5. Show that the complexification of the real coincides with the com-
plex Also, show that for all real-valued functions f and g we have
If+zgl = and that
II Iii- and on C12 are defined via the formulas
and max
i=1
If A is an n x n matrix, then the and the of A

are defined by
IIAII1= sup IAzIk and sup
Show that:
IlAlk and
[HINT: We shall indicate how to establish the second formula only. Let
= If < 1 and y = Az, then
= >
for each i, so that Next, fix some m such that =

and then for each j pick some 9j such that amjezOi = amjl, where z
denotes as usual the imaginary unit. Then z = (e291,.. satisfies
IIzIIoo = 1 and if y = Az, then Ym = = =
proving that = cT.]
7. Let E be a real Banach lattice with complexification A vector sub-
space J of is said to be an ideal whenever f
1z2 and z2 E J imply
z1 E J.
Complex Banach Lattices 111
(a) Show that the ideals in are precise'y the vector subspaces of the
form J J0 'iJ0, where J0 is an ideal in E.
(b) A bounded operator T: F F on a (real or complex) Banach lattice
is said to be ideal irreducible if there is no non-trivial c'osed ideal
in F which is T-invariant. That is, T is ideal irreducible if and only
if T(J) J and J a closed ideal in F imply either J = {0} or J F.
If T: E E is an operator on a real Banach 'attice, then show that
T is ideal irreducible if and on'y if is likewise ideal
irreducible.
[HINT: Assume that J is an ideal of
E E with x + zy E J}.
E
Verify that J0 is an ideal in E satisfying J = Jo EBzJ0. For part (b) note

that an ideal J J0 iJ0 is closed in if and on'y if J0 is a c'osed ideal
in E.]
8. Show that all "domination properties" included in the results 2.342.40
are also true for comp'ex Banach lattices. [HINT: Let and be
the of two real vector spaces and let T: be
a linear operator. In view of Exercise 9 in Section 1.17 we know that
T can be represented by a matrix of the form T
T-Swhere
= T
T, S: X Y are uniquely determined linear operatorswhich are both
bounded if and only if T is a bounded operator. Now, assume that
a bounded operator S = B
A-B between two complex
A
Banach lattices is dominated by a positive operator T: E F. Show
that both A and B are dominated by T.}
9. (Wickstead) Let E = 1R2 be equipped with the sup norm 1
(i.e.,
1J(x, = max{lxI, and let F 1R2 be equipped with the 1-norm
ii (i.e., Ii(x,y)1Ii si + IyD. Now consider the operator T: E F

defined by
T(x,y) (x+y,xy).
Show that the operator T is regu'ar satisfying JJT(JC 2=
(Note that this does not contradict Corollary 3.23 since T is not positive.)
10. Let H be a real Hilbert space with inner product (.,.). Extending the
inner product to via the formula
(Zl,Z2) = +IY1,X2 +iY2) = [(51,52) + (y1'y2)] +z[(y17x2)
we tarn into a (complex) Hilbert space. Show that the standard norm
on defined by
= + = sup IS cos0 +
OER
isnot in general an inner product normand hence does not coincide

with the norm induced by the above inner product.
11. Let X be a real Banach space and define N: X x X p R by
4)
J\I(x,y) = +
3.Operators on AL- arid AM-spaces

(a) The real vector space X e X normed with N is a Banach space.
(b) The same function N now considered defined on the complexification
= XezX, that is, N(x+zy) = + is a norm on
X a real Hubert space.
12. Let be the complexification of a real Banach lattice E and assume that
z u for each E> 0 there exist
positive vectors Ui,. . . ,in E and complex scalars a1,. . ,
. such that
1 for each i = 1,.. .,m, = 1, and z
[HINT: We can assume that = and 1. Since is a com-
pact set, there exist w17... , such that E).
So, if = then each V2 is a non-empty open subset
of and = V. Next, pick a partition of unity {u1, U2,.. , .
subordinate to the open cover {V1,.. ,. see [31, Theorem 10.9, p. 85].
Put = and note that
- = -
(w) [z(w) - (w) E .
3.3. The Center of a Banach Lattice

The purpose of this section is to discuss an important class of operators that
generalizes the class of multiplication operators. This is the class of central
operators. For details and complete proofs of the results stated here, we
refer the reader to [30, Sections 8 and 15].
Definition 3.28. An operator T: E f E on a real or a complex Riesz space
is called central if it is dominated by a multiple of the identity operator.
That is, T is a central operator if and only if there exists some scalar,\ > 0
such that holds for all x E E.
The collection of all central operators is denoted by 2(E) and is referred
to as the center of the Banach lattice E.
Clearly, each central operator is regular and every positive central oper-
ator is a lattice homomorphism. Moreover, it is easy to see that the center
2(E) is an ideal in 4(E), the vector space of all regular operators on E.
The central operators are special examples of operators known as ortho-
morphisms. An order bounded operator T: E E on a Riesz space is said
to be an orthomorphism if A = 0 implies xl A ITyl 0. Recall
also that an operator T: E E on a Riesz space is band preserving if
T(B) B for every band B of E.
3.3. The Center of a Banach Lattice 113
It is worth pointing out that in general a band preserving operator does

not need to be order bounded; see [23, 342]. However, for Banach lat-
tices the situation is better: Namely, the band preserving operators and the
central operators coincideS This is included in the next result; for a proof
see [30, Theorem 15.4, p.
Theorem 3.29 (AbramovichVekslerKoldunov [23]). For a given operator
T: E p E on a Ban ach lattice the following statements are equivalent.
(1) T is a central operatorS
(2) T is an orthomorphism.
(3) T is a band preserving operator.
In spite of the fact that for an arbitrary Riesz space E, the partially
ordered vector space need not be a Riesz space, the center Z(E) is
always a Riesz space. The details are in the next result.
Theorem 3.30 (BigardKeimel [62}; ConradDiem [93]). If E is a Riesz
space, then Z(E) is a Riesz space. Its lattice operations are given by
(S V T)(x) = 5(x) V T(x) and (S A T)(x) = 5(x) A T(x)

for all 5, T E Z(E) and all x E
E Z(E), then its modulus TI satisfies
ITI(lxI) = T(IxI)I =
for each x E E.
a proof of Theorem 3.30 see [30, pp. 107109]. In fact more is true.
For
The center of a Banach lattice is always an AM-space with unit.
Theorem 3.31 (Wickstead [334]). For a Banach lattice E the following
two statements are trae.
(1) The center 2(E) equipped with the operator norm is an AM-space
with unit. The unit is the identity operator I, so that for each
operator T E 2(E) we have
11Th = II 1
=inf{A 0 TJ
Moreover, 2(E) is a unital Banach algebra.
(2) If E is Dedekind complete, then the center 2(E) coincides (as a
Riesz space) with the band generated by the identity operator I in
4(E); and so
Now let us discuss the center of an AM-space with unit. To start with,
recall that (by Theorem 3.6) an AM-space with unit e is lattice isometric to
some C(Ifl-s pace so that the unit e corresponds to the constant function 1
on ft Every function x E gives rise to a central operator (also called
a multiplication operator) defined by = xy. A
moment's thought reveals that
= sup sup
JIyIh,ol
This shows that x is a linear isometry from to It

turns out that this linear isometry in fact, a surjective lattice isometry.
Theorem 3.32 (Zaanen [342}). If is a compact Hausdor'ff space, then the
mapping x from C(11) to is a surjective lattice isometry;
and so, under this identification, we have Z(C(Ifl) =
Proof. To see that x 'p is a lattice isometry, fix y E with

y 0. Then from Theorem 3.30, we get
V = V = (uy) V (vy) = (u V v)y =
This implies V and so x is a lattice isometry.
The only thing missing to complete the proof is the fact that x 'p is
surjective. To establish this we need the following simple property of central
operators. -
. A central operator T: C(11) f C(11) is zero if and only if Ti = 0.

To see this, assume that a central operator T satisfies Ti = 0. Now if
xE then choose some A > 0 satisfying xI Ai, and then use
Theorem 3.30 to get
T(x)I = = AITiI = 0.
Hence T(x) =0 for each x E and so T = 0.
let T E Z(C(Ifl) and put u = Ti E Since the multipli-
cation operator is the operator S = T is also central and
satisfies Si = Ti = 0.
u By the 5 0 or T = This shows
that x ip is a surjective lattice isometry.
Next, consider an AM-space E with unit e. Since E is lattice isometric to
a (unique) we can view the vectors in E as continuous functions
on ft In other we can equip E with a unique multiplication having
e as a unit. In this setting, we can consider every element x E E as a
central operator on E defined via the formula = xy. Moreover,
by Theorem the mapping x 'p is a lattice isometry from E onto
3.3. TheCenterofaBanachLattice
Z(E). That is, subject to the above identification, we have the following
result. 0 (fr V i(:: , (1
Theorem 3.33. If E is an AM-space with unit, then Z(E) = E.

Now consider an arbitrary Banach lattice E and let 0 < u E. By
Theorem 3.4, we know that the principal ideal under the norm
,
= inf{A 0: fxj Au},
is an AM-space with unit u. So, every element x defines a central

operator on via the formula = xy. As before, every central
operator on is continuous with respect to the norm of E (in fact
= and hence itextends uniquely to a u-continuous operator

on, the 11-closure of in E. We shall denote this extension by
again. If happens to be dense in E, then the unique continuous extension
of is also a central operator of E. Remarkably, in this case, it turns out
that each central operator on E is of this type. This is an important result.
Theorem 334. Let E be a Banach lattice and let 0 < u E E satisfy

= E. Then the mapping x i+ from to 2(E), is a surjective
lattice isometryand so, in this case, the center Z(E) can be identified with
the AM-space
Proof. The principal ideal is lattice isometric to some By virtue

of Theorem 3.32, we know is lattice isometric to via the mapping
x As before, we again denote by the unique . 1t-.continuous
linear extension of to all of E. This extension is also a central operator
and preserves the norm of i.e., = By the continuity of the
lattice operations and the 1-denseness of in E, it follows that x
is also a lattice isometry from to Z(E). To finish the proof, we shall
show that x is also surjective.
To this end, let T E Z(E). Then T restricted to
is a central operator on By Theorem 3.32, T = for some x E By
the of both T and it follows that T = This shows
that x is also surjective, and so we can identify 2(E) with
There is one more description of the center of a Banach lattice which is

very useful in many contexts. To understand it, we need some facts from the
representation theory of Riesz spaces. Recall that if Q is a Stonean space,
i.e., a compact Hausdorif space such that the closure of every open set is
open, then the collection C(Q) of all extended continuous func-
tions f: Q [oo, Do] for which the open set {q E Q: <oo} is dense
in Q, under the pointwise algebraic and lattice operations, is a Dedekind
3. Operators on and AM-spaces
complete Riesz space and an algebra. It turns out that for every Archi-
medean Riesz space E there is a unique (up to homeornorphism) Stonean
space Q and a lattice isomorphism 'iv from E into C (Q) such that for each
f
0 < E C(Q) there exists some 0 < x e E such that 0 < 'ir(x) < f, i.e.,
'ir(E) is order dense in C(Q). If E has a weak unit e, we can assume that
iv carries e to 1, i.e., 71(e) 1, the constant function one on Q. The unique
Stonean space Q is referred to as the Stone space of E. For all
purposes, we can identify E with 71(E) and consider the elements of E as
extended real-valued functions on Q. For details about the representation
of Riesz spaces we refer the reader to [225, Chapter 7]. Summarizing the
discussion above, we have the following important result.
Theorem 3.35 (MaedaOgasawaraVulikh). If E is an Archimedean Rie$z
space, then there exists a uniqueup to homeomorphismStonean space Q
and an order dense Riesz subspace E of the Dedekind complete Riesz space
C(Q) that is lattice isomorphic to E. Moreover:
(1) If E has a weak unit e, then the lattice isornorphism can be chosen
to carry e to the constant function 1 on Q.
(2) If E is Dedekind complete, then E is an order dense ideal in
C(Q).
(3) If E is Dedekind complete with a strong unit e, then E = C(Q).
It is customary to identify E with E and we shall follow this custom. The
unique Dedekind complete Riesz space C(Q) is called the MaedaOgsa-
waraVulikh completion of E. It is also known as the universal com-
pletion of the Riesz space E. For details about the universal completion
see [29, Chapter 7].
From the general theory of orthomorphisms we also have the following
characterization of the center of a Banach lattice. For a proof see [30,
Theorem 8.28, p. 121].
Theorem 3.36. If Q is the Stone space of a Banach lattice E, then
Z(E)_{feC(Q): fgeE for all gEE}.
Moreover, if E is Dedekind complete, then Z(E) = C(Q).
We point out here an important difference between a Dedekind complete
Banach lattice and a general Banach lattice. From Theorem 3.36, it follows
that the center Z(E) of a Dedekind complete Banach lattice E is always
a non-trivial AM-space, that is, 2(E) {AI: A an arbitrary scalar}.
However, a non-Dedekind complete Banach lattice can have a trivial center.
An example of such a Banach lattice was provided first by
A. Goullet de Rugy [142]; see also [336].
3.3. The Center ofa Banach Lattice 117
By Theorem 3.31, we know that if E is a Dedekind complete Banach

lattice, then Z(E) is a band in 4(E) and
4(E)
That is, every regular operator T: E E can be decomposed uniquely in
the form T T1 T2, where T1 is a central operator and T2 is an operator
disjoint from the identity operator. This shows that the formula (T) = T1
defines a natural projection defined on 4(E) and having the center Z(E)
as its range. This projection is known as the diagonal projection of 4(E)
and is called the diagonal part of T. As a band projection, is
a contraction with respect to the r-norm. Remarkably, J. Voight has shown
that is also a contraction with respect to the operator norm.
Theorem 3.37' (Voight [327']). If E is a Dedekind complete Banach lat-

tice, then the diagonal projection from 4(E) onto the center Z(E) is a
contraction with respect to the r-norrn.
Proof. We must establish that IIPz(T)II 11Th for each T E To

this end, fix some T E 4(E). Since Z(E) is a band in 4(E), we can write
TZ+R, where EZ(E) and IRIAI=0.
We can assume Z 0. Fix 0 <'y < 1 and then select some 0 (This is possible since J = holds
for all z E E; see Theorem 3.30.) Next, we denote by the band generated
by u in E and let (Q) be the MaedaOgasawaraVulikh completion of
where u corresponds to the constant function 1 on Q. Since
(see Theorem 3.29), there exists (by Theorem 3.36) some h E satisfying
Z(y) = hy for each y E From IIZuhI > we get >
So, if V is the closure of the open set {q E Q: > 'ylIZIf}, then Visa
non-empty clopen subset of Q and xv E E.
Now fix 0 < 8 < Using Exercise 10 at the end of this section, we
know that there exists a component a xw of xv (where W is
a clopen subset of V) such that Pa( I RIa) a, where Pa denotes the
projection onto the band generated by a in E. From
(Z+R)aI =
= ha+Pa(Ra)j ('yS)IjZIJa,
we get lIZ + RIh 'Itall 1II(Z + R)aIM * 8)IIZhl . haiL Consequently, we

have Z + RI 8) h for all 8> 0 with 0 < < < 1. This implies
11Th = liz + as desired.
3. Operators on AL- and
Exercises
1. Show that a positive operator T : E E on a Riesz space is an ortho-
morphism if and only if I + T is a lattice homomorphism.
2. Assume that E is a Dedekind complete Banach lattice. Use Theorem 1.17
to present an alternate proof of Theorem 3.30.
3. Show that every central operator T: E E on a Banach lattice leaves
the principal ideals invariant. That is, show that holds for
each u E E+.
5. If is a finite measure, show (with appropriate identifications) that
= for each 1 p 00.
6. Prove the second part of the MaedaOgasawara--Vulikh Representation
Theorem 3.35 by establishing the following general result for Riesz spaces.
If F is an order dense Riesz subspace of an Archimedean Riesz space
E and F is Dedekind complete in its own right, then F is an ideal in E.
7. Show that a central operator on a Banach lattice commutes with all band
projections.
8. This exercise justifies the name "diagonal projection" for the band
tion of Lr(E) onto the center of the Banach lattice E. Let E = the
Euclidean ri-dimensional (Riesz) space. Then (E) = 4(E), the vector
space of all ri x ri matrices with real entries. If D is a diagonal matrix
with diagonal entries d1, . . . then we write D = diag (d1,.. .
Establish the following properties.
(a) If A = and .8 = are two ri x ri matrices, then A .8 holds
in 4(E) if and only if for all i and j.
(b) If A = and .8 = are two ri x ri matrices, then
= AV B = V and A A B = A
(c) The center of E consists of all diagonal matrices.

(d) If A = is an ri xri matrix, then Pz(A) = diag (a11, a22,.. . ,
9. Prove the second part of Theorem 3.36. That is, show that if E is a
Dedekind complete Banach lattice, then Z(E) = C(Q), where Q is the
Stone space of E.
10. Recall that the principal band generated by a vector x in a Riesz
space E is the smallest band that contains x. From Theorem 1.27, we
know that = {y E E: yIAriIxI I If E is Dedekind complete, then
is always a projection band, i.e., = E. The band projection
of E onto will be denoted by
Now let E be a Dedekind complete Banach lattice and let T, Z: E E
be two disjoint regular operators (i.e., ITI A ZI = 0) with 0 3/: Z E 2(E).
Show that for each 0 < x E E and each 0 < E < there exists a
non-zero component a of x (i.e., a A (x a) = 0) such that Pa(lTIa) Ea.
[HINT: We can suppose that T and Z are positive operators with Z
satisfying 0 < Z I. Fix 0 < E < Replacing E by we can
assume that x is a weak unit. Let Q be the Stone space of E. We

identify E with the ideal C(Q) of C(Q) so that x corresponds to the
constant function 1 on Q. By Theorem 3.36, there exists some 1?, e C(Q)
satisfyingZ(z) = hz for all z E E and = Now notice that the
set V = {q e Q: h(q) > E} is a clopen subset of Q and so y = Xv e E.
From Theorem 1.17, we know that
= (TA Z)y = inf{Tz + Z(y z): z e
0 and z A (y z) = 0}.
This implies that there exist a component Xw of y = Xv (where W is a
clopen subset of Q) and some e V such that
+ h(qo)xv\w(qo) <E.
In view of h(q) e for each point q e V, we see that q0 e I'V. From
the inequality T(xw)(qo) < we infer that there exists some clopen set
W1 W satisfying 0 T(xw1)(q) T(Xw)(q) for each q E W1. If
we let a = Xw1 0, then Pa(Ta) = = Ea.]
3.4. The Predual of a Principal Ideal

We shall describe in this section an interesting lattice property of the ideal
generated by a positive linear functional in the order dual of a Banach lattice.
So, for our discussion, let E be a real Banach lattice and let 0 0 such that }.
The ideal with its where 1Ix* = inf{\ 0:
1x* <
is an AM-space with unit cb. Our goal is to describe an AL-space whose
dual is E.
The linear functional b defines a lattice seminorm on E via the
formula
= cb(IxI).
This is an in the sense that (besides being a lattice seminorm)
it also satisfies Jx + = IlxiLt + for all x, y E If we consider the
null ideal of b, i.e., the closed ideal Nq5 = {x e E: cb(JxI) O}, then the
quotient seminorm induced on the quotient Riesz space E/Nq5 by 1
i.e.,
the lattice seminorm
= inf = cb(IxI)
yE [s]
is in fact an L-norm. The norm completion of the normed Riesz space

(E/Nq5, 1
an AL-space, which we shall denote by E(cb). Also, we
is
shall denote the norm dual of E(cb) by E(cb)*. Keep in mind that E(b)*
an AiVI-space with unit e* defined by the norm of the E(cb), i.e.,
e*([x]) = cb(x) for each x E E.
There is a natural mapping or i* defined by
([x]) = (x), (*)
for all x E E. To see that is well defined we argue as

follows. Fix a linear functional E and then select a scalar A > 0 such
that Ix*I <Ab. Now assume that y E [x}, [yJ = [x}, or cb(fx = 0.
From the inequality 1x*(x) x*(y)f Ix*J(Jx yI) = 0, we
x*
see that (x) = (y), and so (*) is a well-defined linear functional on the
quotient Riesz space E/NQ. From
= Ix*(x)t < Jx*f(pxI) < = = , (t)

it follows that is a continuous linear functional on Therefore,
has a unique continuous linear extension to the norm completion E(cb)*
of which we shall denote by again. From (t) it should be clear
that = IIx*lIoc holds for each E E.
We now leave the details of the proof of the following result to the reader.
Lemma 3.38. The mapping as defined by (*) is a surjective

lattice isometry.
Next, assume that T: E + E is a positive operator such that its norm

dual T*: E* leaves the principal ideal invariant, i.e., T* (E;) E.
Then there is a (unique) way to define a positive operator p
such that the diagram in Figure 1 is commutative.
T*J,
Figure 1
The operator is defined as follows. For each [xJ E we let
= [Tx]. (**)
To see that the value TQ([x}) in (**) is well defined, let y E [x], i.e., let
[y] = [x} or = 0. Since T* leaves invariant, there exists some
A0 satisfying T*b This implies
0< =0,
and hence TyI) = 0. This shows [Tx] = [Ty], and so (**) defines a
linear operator from E/Nj, to E/Ncb. From
= =
= =
we see that E/Nj, E/Nj, is also continuous. Let denote the unique
continuous extension of T1,: E/Nj, + to all of To see that the
diagram shown in Figure 1 commutes, note that for all
x E, we have:
= = =
= x*(Tx) = (Tx,x*) = (x,T*x*), and
([x], = ([x], = T*x*(x) (x,T*x*)
- Thus, we have established the first part of the following result.
Theorem 3.39. Assume that T: E E is a positive operator such that for

some 0 e E* the adjoint operatorT*: E* leaves the principal ideal
invariant. Then there exists a unique positive operator Tb:
whose adjoint makes the diagram in Figure 1 commutative. Moreover, if T

is a lattice homomorphism, then so is
Proof. We need to prove only that Tj, is a lattice homomorphism if T is

a lattice homomorphism. So, assume that T is a lattice homomorphism.
Then, using the fact that the canonical projection x h> [x] of E onto E/N1,
is a lattice homomorphism, for each x, y e E we have
V [y]) = V y]) = [Tx V Ty] = [Tx] V [Ty] = Tcb([x]) V T1,([y]).
This shows that Tb: E/Nj, + E/Nj, is a lattice homomorphism, and from
this (by continuity) it follows that is also a lattice homo-
morphism.
Exercises
1. Establish the following useful result: If Y and Z are closed subspaces of a
normed space X, one of which is finite dimensional, then Y+ Z is a closed
subspace of X. [HINT: Assume that Y is closed, Z is finite dimensional,
and let it: X X/Y denote the quotient map. Then ir(Z) is a finite
dimensional subspace of the normed space X/Y, and so 7r(Z) is a closed
subspace of X/Y. Now note that Y + Z ir1 (-ir(Z)).]
2. Let E be a Riesz space and let J be an ideal in E. Establish the

ing:
(a) The set C = {[x]: x E is a (convex) cone in E/J.
(b) The vector space E/J ordered by the cone C is a Riesz space.
(c) The quotient map x [x] is a lattice homomorphism from E onto
E/J.
3. Let E be a norined Riesz space and let J be a closed ideal in E. Establish
the following.
(a) The quotient norm is a lattice normand so E/J is a normed Riesz
space.
(b) If E is a Banach lattice, then E/J is also a Banach lattice.
4. Show that the norm completion of a normed Riesz space E is a Banach
lattice containing E as a Riesz subspace. [HINT: Let E be a normed
Riesz space with norm completion E. Show that the norm closure of E+
in E is a cone under which E is a Banach lattice.]
5. Let E be an arbitrary M-space, that is, E is a normed Riesz space
such that lix V max{tIxIf, holds for all x,y E Show the
following.
(a) The norm completion of E is an AM-space.
(b) If E has a unit e and the closed unit ball of E coincides with the
order interval [e, e], then the norm completion of E is an AM-space
with unit e.
6. Let E be an arbitrary that is, E is a normed Riesz space such
that + = lxii + I
holds for all x, y E E with x A y = 0. Show that
the norm completion of E is an AL-space.
7. Let X be a normed space and let X be its norm completion. For each
f E X*, let I denote the unique continuous (linear) extension of f to
Show the following.
(a) The mapping f '+ f is a surjective linear isometry from onto
(and so, subject to this linear isometry,
X is a normed Riesz space, then f / is a surjective lattice
isometry from onto
8. Let p be an arbitrary lattice seminorm on a Riesz space E, and define its
null ideal by = {x E E: p(x) = O}. Establish the following.
(a) is indeed an ideal in E.
(b) The function ,: IR defined by ,({x]) = p(x) is a lattice
norm on E/N, and 13([x]) = p(y).
(c) If p is an M-seminorm (resp. an L-seminorm), then is an M-noriii
(resp. an L-norm).
9. Complete the details of the proof of Lemma 3.38.
Chapter 4
Special Classes of Operators
This chapter studies the properties of several important classes of operators

that were discussed in a fragmented way in the previous chapters. These
classes are: finite-rank operators, multiplication operators, lattice and alge-
braic homomorphisms, Fredhoim operators, and strictly singular operators.
We pay special attention to the order-theoretic properties of these classes of
operators.
The class of finite-rank operators is considered in conjunction with the
approximation property of Banach spaces. The presentation of multiplica-
tion operators is in the framework of By means of the classical
KakutaniBohnenblust--Krein Representation Theorem 3.6, this framework
allows us to study multiplication operators on AM-spaces. The section on
lattice and algebraic homomorphisms deals with lattice homomorphisms be-
tween a few basic results regarding extension of
lattice homomorphisms. The section on Fredholm operators presents their
basic properties including proofs of the Index Theorem and of the continuity
of the index function.
The last section deals with strictly singular operators. An operator
between two Banach spaces is said to be strictly singular if its restriction
to any infinite dimensional closed subspace is not an isomorphism. After
establishing some basic properties of these operators, we investigate their
relationships with compact and Fredholm operators.
4.1. Finite-rank Operators

We discuss here some basic properties of finite-rank operators. An operator
T: V W between two vector spaces is said to be a rank-one operator
123
124 - 4. Special Classes of Operators
if its range is one dimensional. So, if T : V W is a rank-one operator

and w is a non-zero vector in the range of T, then there exists a (uniquely
determined) linear functional f on V satisfying T(v) = f(v)w for each v V.
Following the standard tensor-product notation, we shall denote the value
f(v)w by (f w)(v), i.e., we shall write T = f w.
Definition 4.1. Am operator T: V W between two vector spaces is said
to be a finite-rank operator if its range is finite dimensional.
Thus, if T: V is a finite-rank operator and w1,. .
TV is a basis . ,
in the range T(V), then there exist (uniquely determined) linear functionals
fi, , on V such that T(v) = i.e., T =
It is easy to see that an operator T: V > W between two vector spaces
is a finite-rank operator if and only if there exist (not necessarily linearly
independent) vectors UI.,. in W and linear functionals ga,. . , on V
. . , .
such that T = gi The dimension of the range of a finite-rank

operator is called its rank. Thus, an operator of rank k is a finite-rank
operator whose range has dimension k. The zero operator is, of course, a
operator with zero rank. Here are two immediate properties of
finite-rank operators.
o Sums and scalar products of finite-rank operators are finite-rank.
a If in a scheme of operators V W T U between vector spaces
either S or T is finite-rank, then TS is likewise finite-rank.
We now turn our attention to continuous finite-rank operators between
Banach spaces. In this case, a bounded operator T: X > Y between two
Banach spaces is a finite-rank operator if and only if there exist continuous
linear functionals on X and vectors Yi, .
. , in Y such that ,
T= yi. The latter expression is referred to as a representation

of T. From now on we shall consider only continuous finite-rank operators
between Banach spaces.
The reader should verify the following property of finite-rank operators;
for details see Exercise 3 at the end of this section.
Lemma 4.2. Assume that T: X > Y is an operator of rank k between
Banach spaces. If ui,. . .,Uk is a basis in the range of T, then T has a
unique representation T = x with the continuous linear
als . , linearly independent.
Observe that if T = Y is a rank-one operator between

y: X >
Banach spaces, then its norm adjoint is T* = y where now y is viewed
as a continuous linear functional on This easily follows from the duality
identity
((x* y)x, y*) = x*(x)y*(y) = (x, (y x*)y*)
Finite-rank Operators 125
The above identity yields the following property of finite-rank operators.

Lemma 4.3. If T: X + Y is an operator ofrank k between Banach spaces
with a representation T = then its norm adjoint T* :
is also an operator of rank k and T*
The following result also should be obvious.

Lemma 4.4. Every finite-rank operator between Banach spaces is compact.
The collection of all finite-rank operators from a Banach space X into
another Banach space Y will be denoted by J(X,Y). As usual, .2(X)
stands for .2(X, X). Clearly, .2(X, Y) is a vector subspace of (X, Y) and
moreover .2(X, Y) IC(X, Y). It was a long standing question whether or
not every compact operator between two Banach spaces was the norm limit
of a sequence of finite-rank operators. That is, it was not known whether or
not .2(X, Y) was norm dense in IC(X, Y). In 1971, P. Enflo [120] answered
problem in the negative. A. Szankowski [315] showed that even when
X and Y are Banach lattices, the vector space .2(X, Y) need not be norm
dense in IC(X,Y).
There is another commonly used notation pertaining to finite-rank op-
erators. Let X and Y be two Banach spaces and let be a subspace of
X*. Then the symbol Y denotes the vector space of all finite-rank
operators T: X Y which have a representation T 0 with
yr,. .
. , E Y and . , E In this notation .2(X, Y) =
The vector space of finite-rank operators plays an important role in
analysis and so do the spaces of operators that contain it. Next, we shall
discuss its role in connection with the approximation property in Banach
spaces.
Definition 4.5. A Banach space X is said to have the approximation
property if for every compact subset C of X and every 0 there exists a
finite-rank operator T E (X) such that Tx < holds for each x E C.
In other words, a Banach space X has the approximation property if
and only if the identity operator on X can be approximated uniformly on
every compact set by finite-rank operators. The following resu't, stated in
Exercise 7 in Section 2.4, is used frequently in the study of the approximation
property.
Lemma 4.6. A subset of a normed space is totally bounded if and only if it
is contained in the closed convex hull of a sequence that is norm convergent
to zero.
Theorem 4.7. A Banach space with a basis has the approximation property.
126 - 4. Special Glasses of Operators
Proof. Let X be a Banach space with a basis As usual, let

denote the sequence of projections on X defined by = for
each x = a number M > 0 such that M for
each m.
Assume that C is a compact subset of X, and let c > 0. According to
Lemma 4.6 there exists a sequence of X satisfying C X2,.
and + 0. Next, pick nO such that < for each n > So,
for each n> no and for all m we have
- < + = (1+ <E.
Since tIPrnX holds for each x E X, there exists
0 some k such that
m k implies 1 n for each n
and all m k we have Xn[j <E.
Now pick any y E co{x1,x2,.. .} and write y = where
Aj >Ofor each 1 <j = 1. Thenfornkwehave
m rn m
= <>
This implies y E C and all n k. Since each is a
finite-rank operator, it follows that X has the approximation property. D
The validity of the converse of Theorem 4.7, i.e., of the statement "every
separable Banach space with the approximation property has a basis," was
another long standing open problem that was solved in the negative by
P. Enflo in [120]. The situation does not improve even when X and Y are
Banach lattices; see [315].
Now let us discuss the connections between the approximation property
and the topology of uniform convergence on compact sets. Fix two Banach
spaces X and Y, and let denote the collection of all non-empty norm
compact subsets of X. For each C E define the seminorm on (X, Y)
via the formula
pc(T) = max
xEC
Itshould be clear that the family of seminorms generates a Blaus-
dorif locally convex topology Y) on (X, Y). Also, it should be obvious
that a net (X, Y) satisfies T in (X, Y) if and only
if converges uniformly to T on every compact subset of X. For this
reason, the topology Y) is called the topology of uniform conver-
gence on the compact subsets of X. We write for X).
Two easy characterizations of the approximation property in terms of
the topology of uniform convergence on compact sets are included in the
next result.
Lemma 4.8. For a Banach space X the following statements are equivalent.
(1) X has the approximation property.
(2) The identity operator I: X > X belongs to the of
the vector subspace of all finite-rank operators in (X).
(3) If an arbitrary linear functional on (X) sat-

isfies /.(x4 0 x) = 0 for all x E X and all E X*, then /.(I) = 0.
For a more useful characterization of the approximation property, we

need to describe the topological dual of (L(Y, X), X)).
Theorem 4.9. A linear functional on (Y, X), where X and Y are Ba-
nach spaces, is X)-continuous if and only if there exist two sequences
Y and c satisfying and
for eachT E (Y,X).

Proof. Let X and Y be two Banach spaces and let be a linear functional
on (Y,X). Assume first that there exist two sequences {yn} c Y and
X* satisfying <oc and =
for each T E (Y,X). Pick an arbitrary sequence of positive scalars
such that > oc and M = < Without loss of
generality we can assume that 0 for each n. Next, let = AYII and
note that > 0. Therefore, the set C = {0,ul,u2,u3,. .} is a compact .
subset of Y. Now, if T E (Y, X), then
= =
IIYnlI] = Mpc(T).
This shows that is dominated by a seminorm, and so

is itself X)-continuous.
For the converse, suppose that is X)-continuous. This means
that there exists some constant fri > 0 and some non-empty compact
subset C of Y such that Mpc(T) holds for each T E (Y, X).
In view of Lemma 4.6, we can suppose without loss of generality that
11f > 0 holds for each and <oc, then choose a strictly increasing sequence
of natural numbers such that < Letting = 1 for 1 i <Icr and A = n > 0
for i < we see that co and <co.
128 4. Special Classes of Operators
C= . . .} for some sequence {yn} of Y satisfying 0. Now

we define a mapping 8: X) (X e X e X e )o via the formula
8(T) = (Tyi, Ty2, Ty3,...).

Clearly, 8 is linear and from J8(T) = supflEN 1
11Th,
we see that 8 is a bounded operator. Next, on the range of 8 define the
linear functional on 8 X)) via the formula
= q5(T).
Note that if T, R (Y, X) satisfy (Ty1., Ty2,...) = (Ry1, Ry2,. .), then.
T R =
on C,
0 and
so 1q5(T) b(R)j = R)J Mpc(T R) = 0
implies q5(T) = q5(R). This shows that is a well-defined linear functional.
From
Ty2,. . = = MII(Tyi, Ty2,. .
it follows that is also norm continuous. By the HahnBanach theorem,

has a norm continuous linear extension (denoted by again) to all of
(XeXeXe. )o, i.e., (X e X e X e )i; see
Exercise 3 in Section 7.2. This means that there exists a sequence
satisfying and
= Ty2,...))
=
Since is a bounded sequence, we also have . <ce, and
the proof is finished.
Corollary 4.10 (Grothendieck A Banach space X has the approx-

irnation property if and oniy if for all sequences X* and x
that satisfy and = 0 in (X), we have
= 0.
Proof. Assume first that X has the approximation property, and let two
sequences X* and X satisfy IIxJI . < cc and
=
Also, let 0. denote the linear functional
on (X) defined by q5(T) Then for each rank-one operator
x we have
x) = = =
= x*(0)=O.
4. 1 . Finite-rank Operators 129
Therefore, b vanishes on the vector subspace of all finite-rank operators,

and so it vanishes on their Since X has the approximation
property, this contains the identity operator I, and hence
b(I) = 0.
For the converse, assume that the condition is satisfied, and let be
a linear functional on (X) that vanishes on the vector
subspace of all finite-rank operators. By Theorem 4.9 there exist two se-
quences X* and X with < oo such that
b(T) = holds for each T E If x is an arbitrary
rank-one operator, then we have
0= =
x E X are arbitrary, it follows that 0.

So, from our hypothesis, we have b(I) = =
This shows that
0.
identity operator I belongs to the of X* X, and so by
Lemma 4.8 the Banach space X has the approximation property.
Corollary 4.11 (Grothendieck [141]). If the norm dual X* of a Banach
space X has the approximation property, then X itself has the approximation
property.
Proof. Assume that has the approximation property, and let two se-
quences X* and X satisfy . < oo and
= 0. Then, considering each as a vector in it follows
from Lemma 4.3 that = = 0. Since the
Banach space has the approximation property, Corollary 4.10 implies
= 0. But then the same corollary guarantees
that X has the approximation property.
A. Grothendieck has also shown that a Banach space X has the ap-
proximation property if and only if for each Banach space Y the finite-rank
operators from Y to X are dense in K(Y, X), the closed vector subspace of
compact operators. This result is stated next.
Theorem 4.12 (Grothendieck [141]). A Bartach space X has the approxi-
mation property if and only if for each Bartach space Y the vector subspace
of all finite-rank operators X is norm dense in X).
Proof. Let X be a Banach space with the approximation property and

let Y be an arbitrary Banach space. Fix a_compact operator T: Y X
and some 0. Since the set C T(Uy) is a compact subset of X
and X has the approximation property, there exists a finite-rank operator
130 4. Special Glasses of Operators
S = X X such that x
for each E Uy we have
- T] - = ff S(Ty) - <. (*)

=
Now notice that the operator
R=ST= Y
isa finite-rank operator, and by (*) it satisfies hR T1I This shows .
that the vector subspace of all finite-rank operators -X is norm dense

in /C(Y,X).
For the converse, assume that for each Banach space Y the vector sub-
space of all finite-rank operators X is norm dense in K(Y, X). Fix
a compact subset C of X and let 0. By Lemma 4.6, we can
sume that there exists some sequence in X satisfying 0 and
C X2, .
. .}. Scaling appropriately, we can also suppose without loss
of generality that 0 < 1 holds for each n.
Define the sequence of X by U2n and U2n_1 =

= _____
From IIU2nhI = = we see that
and therefore the
+ 0,
set U = U2,.. .} is a compact subset of X. It is also convex and circled
(i.e., 1 and x E U imply E U). From the identity =
and Uforeachn, andsoCU. -
The preceding properties of U guarantee that Y = U=1 nU isa vector

subspace of X. Moreover, if Y is equipped with the norm defined by
yE1\U},
then Y is a Banach space. Its closed unit ball coincides with U, that is,
{y E Y: 1} = U. Moreover, if we let J: Y X denote the natural
embedding (that is, Jy = y for each y E Y), then from J(U) = U, it follows
that J is a compact operator. Therefore, according to our hypothesis, there
exists a finite-rank operator 0 E (Y, X) satisfying
y y E U.
The proof would be finished from (**) if we knew that each E belongs
to Unfortunately, this might not be the case. However, the proof can be
completed if we can verify that for each E and each S > 0 there exists
some E X* satisfying I y*(x) x*(x)f <S for each x e C or equivalently
IY*@1n) X*(Xfl)I <S for each n.
To see this, fix a non-zero E and let c5 = > 0. Let Z be the

closed vector subspace generated by m E N} in X. Jf y* = 0 on
then the zero functional 0 E X* satisfies = y*(Xn) for each n. So, we
can suppose that for some Z we have y*(zo) 1.
From U2n E U it follows that for each n. This
=
implies + 0, and so (since is HI-continuous) y*(xn) 0. Pick
some k such that <q for each n> k.
Now consider the set K = n > k}, and note that K is a
convex, circled, and subset of Z. Next, let 4 denote the (finite
dimensional) vector subspace generated by the vectors Xl,.. . , xk}, and
let F = {z E Zk: y*(z) = 1} + 4 fl Ker Clearly, F is non-empty,
and (as is J-continuous on Zk) F is also a Jj-closed subset of Zk
and hence of Z. Moreover, it is easy to see that K fl F = 0. Hence, by the
classical Separation Theorem (see, for instance [25, Theorem 5.58, p. 193]),
there exist a non-zero Il-continuous linear functional on Z (we can
assume that x* is defined on all of X) and some scalar a 0 such that
Ix*(c)I <a < x*(z)
K and all z E F. (Here we assume that X is a real Banach

holds for all c E
space; the complex case can be treated similarly.) Since 4 fl Ker is a
vector subspace of 4, it follows from (***) that Zk flKer c Zk fl Ker
This shows that there exists some scalar A such that Ay* on Zk. If
A = 0, then = 0 on Zk, and from (***) we see that = 0 for n> k,
i.e., = 0 on Z, which is a contradiction. Hence, A 0, and from (***) we
get that A > 0. Therefore, replacing by we can assume that =
on F and that 0 a 1. In particular, = y*(Xn) for all 1 <k
and for n > k we have a 1 or X*(Xn)l 7], which implies
y*(xn)i < + y*(Xfl)J <7] + = = 5. The proof of the
theorem is now complete.
We close the section with two results regarding finite-rank operators

between Banach lattices.
Lemma 4.13. If u: E p F is a rank-one operator between Banach

lattices, then 0 u has a modulus and
Proof. If x E+, then we have

sup (x* u)(y)J = sup x*(y)uI sup 1x*(y)Hul
IvIx
[sup Ix*(y)I]luI Ix*1(x)luI
fyi x
= (Ix*I fuI)(x)
This implies that u has a modulus and that Ix* = 1x*1 ui.
Theorem 4.14. Every finite-rank operator between Banach lattices has

modulus which is compact but not necessarily a finite-rank operator.
Proof. Let T = E f F be a operator between

Banach lattices. If u then it is easy to see that T(E)
where is the ideal generated by u in F and is equipped with its u-norm.
Clearly, T: E (as a operator) is compact. Now replicate
the proof of Theorem 3.14 to conclude that T: E * has a compact
modulus ITt. Since is an ideal in F, the latter guarantees that TI is also
the modulus of the operator T: E * F. Since the topology generated by
the norm of is finer than the topology generated on by the norm of
F, we infer that [TI: E * F is also a compact operator.
To see that the modulus of a finite-rank operator need not be a finite rank
operator, consider the constant function 1 and the function U: [0, 1J f R
defined by u(t) = t. We leave it to the reader to verify that the
operator T = 1 u u 1 on L1 [0,1] does not have a modulus;
see Exercise 9 at the end of the section.
Exercises
1. Let f, . , be functionals on a vector space V. Show that f
linear
lies in the linear span of fi,. , if and only if
. . Kerf.
In particular, a collection of linear functionals on a vector
. ,
space V is linearly dependent if and only if Ker gj Ker gj holds

for some j. [HINT: Assume that V is a real vector space (the same
arguments show that the result is true for vector spaces over any field)
and let Ker C Ker f. We consider the vector subspace of
W_{(fl(v),...,fk(v)) vEV}
and define the linear functional W by q5(fi (v), . . , = f(v).
Now if (fl(v),...,fk(v)) (fl(u),...,fk(u)), then vu E
and so by our condition f(v) f(u), which shows that q5 is well de-
fined. Notice that q5 extends linearly to all of So, there exist
scalars . . , such that . ,

= This implies
f=
2. Show that an operator T: V W between two vector spaces is a finite-
rank operator if and only if there exist vectors w1,... in W and linear ,
functionals . . . , on V such that T = gj

3. Let T: V + W be an operator of rank k and let wi,. be a basis . . ,
of the range of T. Then there exist (uniquely determined) linear func-

tionals fl, . , on V sucia that T =
. . 0 Show that the linear
functionals fr,.. , are linearly independent. [HINT: If =
then T = 0 + and so the rank of T is at most k 1, a
contradiction.
4. Let T: V V be a operator on a vector space and assume
that T = > 0 g3 0 uj are two representations of T, where
the vectors and uj need not belong to the range of T. Show that
i.e.,the scalar tr(T) = (vi) (called the trace of T) is independent

of the representation of T. [HINT: Fix a basis w1, . , of the range
. .
of T and write T = hr 0 Wr, where the linear fimctionals hr are

uniquely determined. Now let T = 0 be another representation
of T. Assume first that fI,.. , are linearly independent. Then for each
i there exists some E V such that = holds for each j; see
Exercise I above. Hence, E T(V) for each i, and so = for
n k n
appropriate scalars So, T 0 = Arfi) 0
and from this we get hr = for each 1 r k. Thus,
> => (Wr) =>

i=I r=I
Now assume the set . , is linearly dependent. There exist lin-

.
early independent linear functionals Oi, . . . , such that = >111=I

for appropriate scalars /4, and so = 0
By the preceding case, we have
= = = .1
r=I v=I i=1 i=I i=1
5. Show that the trace of a finite-rank operator on a Banach space is simi-

larity invariant. That is, show that if T: X X is a finite-rank operator
on a Banach space, then tr(T) = for each invertible bounded
operator S: X X.
134 -- 4. SpecialClasses of Operators
6. Let X be an arbitrary Banach space and assume that , , . . . ,
and {x1, X2, X satisfy

, = for all i and j. Show that
the finite-rank operator P = has rank n, is a continuous pro-
jection, and has range the vector subspace generated by {x1,x2,..
Use this conclusion to show that every finite-dimensional subspace
Y of X is complemented, i.e., there exists a closed subspace Z of X
satisfying Y fl Z = {O} and Y Z = X. [HINT: For the second part let
{x1, X2, -. . be a basis of Y. Pick linear functionals
, , on
Y such that = 6jj holds for all i and j. Since the norm topology
of X induces the Euclidean topology on Y, each is norm continuous.
By the HahnBanach theorem, we can assume that each is defined on
all of X, i.e., E for each i. Now consider the continuous projection
P= Xj.]
7. (Grothendieck [141]) For a Banach space X establish that the following
statements are equivalent.
(a) X has the approximation property.
(b) For every Banach space Y the vector space of finite-rank operators
X is X)-dense in (Y, X).
(c) For every Banach space Y the vector space of finite-rank operators
Y is Y)dense in Y).
8. Let X be a Banach space with the approximation property. is a If Y

Banach space and T: Y X
is a compact operator, then show that
T*: }T*
belongs to the norm closure of the finite-rank operators
from X* to }T* i.e., T* E Y*.
9. Consider the rank-two operator T = 1 n n 1 on L1 [0, 1] introduced
in the proof of Theorem 4.14, where 1(s) = I and n(s) = s for each
s E [0, 1]. Also, function v E
recall that each [0, 1] can be considered
as a continuous linear functional acting on L1 [0, 1] via the duality formula
(x, v) = j'0' v(s)x(s) ds. Prove the following.
(a) Tx(t) = j'0' (t s)x(s) ds for all x E L1 [0, 1] and eacht E [0, 1].
(b) The modulus of the operator T satisfies tTlx(t) = f0 s tlx(s) ds
for all x E L1[0, 1] and t E [0, 1].
(c) is not a finite-rank operator.
[HINT: For (b), see Exercise 13 of Section 3.1. For the validity of (c) note
that if Xk(S) = (k + 1)(k + then
Tlxk(t) (k+2)t+k+1.
Since the functions are linearly independent, we infer
that ITI cannot have a finite dimensional range.]
10. An operator T: X Y between two Banach spaces is said to be a nu-

clear operator if there exist two sequences c and } Y
satisfying <00 and
T(x) =
for each x E X, i.e., T = in (X, Y). Show that:

(a) Every nuclear operator is compact.

(b) If X and Y are two arbitrary Banach lattices, <oo and
<co, then the nuclear operator T = has
a compact modulus.
4.2. Multiplication Operators

In this section, we shall discuss multiplication operators on -spaces.
We start with the following result.
Lemma 4.15. For a positive element u > 0 in a Banach lattice E the
(1) The principal ideal generated by u is norm dense in E, i. e., = E.
(2) For each x E E+ we have jx A nu 0.
(3) If 0 E E*, then x*(u) > 0, i.e., the vector u acts as a strictly
positive linear functional on E*.
Proof. (1) Let x E

(2) and fix 0. Since is dense in E,
there exists some y E such that <. From the lattice inequality
Ax we Ax) that we can assume
see (by replacing y by
0 <y<x. Now.fixsome Ic suchthat 0 y kuand so Hence,
if m k, then from the inequality 0 x x A nu x x A ku x y, we
see that lix x A null Therefore, x A 0.
(2) (3) Let 0 E If x*(u) = 0, then x*(x A nu) = 0 for all

n and each x 0, and so x*(x) = 0 for all x E E+. This implies = 0, a
contradiction. Hence, x*(u) > 0.
(3) Assume by way of contradiction that
(1) E. Then, by the
Separation Theorem, there exists some non-zero E E* that vanishes on
This implies 0 < lx*Ru) = sup{jx*(y)I: u} = 0, a contradiction.
Hence, = E.
Any positive vector u in a Banach lattice E that satisfies = E is
called a quasi-interior point or a strictly positive vector. The name
"quasi-interior point" originates from the fact that an order unit of a Banach
lattice is simply an interior point of the positive cone. The name "strictly
positive vector" is justified by statement (3) of Lemma 4.15. We have the
following implications:
Order Unit Point Weak Order Unit
In general no other implication is true. However, in a Banach lattice with
order continuous norm a positive vector is a quasi-interior point if and only
if it is a weak order unit. We shall say that a complex Banach lattice
E zE has a quasi-interior point if the Banach lattice E has a

interior point.
How marty points can a Bartach latti'ce have ? It might
have none! For instance, the Banach lattice M [0, 1] of all Borel measures of
bounded variation on [0, 1] does not have quasi-interior points; see Exercise 5
at the end of the section. On the other hand, if a Banach Ilattice E has a
quasi-interior point, say u, then E has quite many more: Au for each A > 0
and any positive vector w satisfying w Au for some A > 0 are all quasi-
interior points. It follows that the set of all quasi-interior points of a Banach
lattice is either empty or else is dense in
However, in spite of the above remarks, if a Banach lattice has a quasi-
interior point, then it is essentially "unique" in the sense that if u and v
are two quasi-interior points, then the principal ideals and are lattice
isometric. This easily follows from Theorem 3.34. Let us elaborate on this
relationship a little bit more.
Let u> 0 be a positive vector in a Banach lattice E. Then (by Theo-
rem 3.6) the principal ideal generated by u is an AM-space under the
norm
= inf{A 0: IxJ Au}.
Moreover, it is lattice isometric to a concrete 0(Q)-space, where the compact
Hausdorff space Q is uniquely determined up to a homeomorphism. Now
assume that u and v are two quasi-interior points in a Banach lattice E such
that and have the representations 0(Q) and 0(e), respectively. Then,
by Theorem 3.34, 0(Q) and 0(e) are lattice isometric and so Q and e are
homeomorphic. The unique (up to a homeomorphism) compact Hausdorff
space Q is called the structure space of E and it plays a significant role
in the representation theory of Banach lattices. For details, we refer the
reader to the book of H. H. Schaefer [290, 4, pp. 165168]. It should be
noted that the structure space of a Banach lattice need not coincide with its
Stone space. For instance, if E = 0[0, 1], then the structure space of E is
[0, 1] while the Stone space Q of E is extremally disconnectedand hence
different from [0, 1].
As discussed in Section 3.3, as soon as we represent a principal ideal
as a 0(Q)-space, we get at our disposal the "operator structure" of 0(Q).
For instance, we have a plethora of multiplication and composition operators
on 0(Q) which can be viewed as operators on Enthis important feature
has no parallel in the general theory of Banach spaces. In other words, we
can transfer via the lattice isometry the (pointwise) multiplication of 0(Q)
to E gives rise to a multiplication
operator via the formula
Multiplication Operators 137
Moreover, is continuous with respect to the original norm

restricted
to (in fact = ltvIk). When u is a quasi-interior point, the multipli-
cation operators extend (by virtue of their uniform continuity) to continuous
operators on all of E.
The next two results that will be used extensively to study the invariant
subspace problem indicate the importance of quasi-interior points.
Lemma 4.16. For a (real or complex) Banach lattice E with a quasi-interior
point u > 0 we have the following.
(1) For each non-zero vector y E there exists an operator lvi domi-
nated by the identity operator (and hence M is a central operator)
satisfying M(y) > 0.
(2) For every element v satisfying 0 v u there exists an operator
T: E+ E dominated by the identity operator and such that Tu = v.
Proof. Let u > be a quasi-interior point in a (real or complex) Banach
0
lattice E. As before, we can identify with a in such a way
that the vector u corresponds to the constant function one on ft
(1) Fix y E with y 0 and view y as a continuous function on
By scaling, we can suppose that = E = 1.
Now consider the function (the complex conjugate of y) and
denote by lvi the multiplication operator defined by the function on
(and hence on En). That is, M(x) = for each x E Clearly,
JVi(y) = > 0 and lfrixl lxi holds for each x E Since = E, the
(unique) continuous extension of M to E satisfies the desired properties.
(2)Again, as above, we view v as a continuous function on and consider
the multiplication operator T defined on (and hence on by v, i.e,
T(x) = vx for each v and ITxJ lxi for each
x E The (unique) continuous extension of T to all of E also satisfies
Txi < x E E.
Lemma 4.17 (de Pagter [260]). Let E be a (real or complex) Banach lattice
with a quasi-interior point and let lxi Then there exists a sequence
of operators on E such that 0 and each is dominated
by the identity operator.
Proof. Let u > be a quasi-interior point of E. Replacing u by u + yi,
0
we can assume that u. Now let us view as a with u
corresponding to the constant function 1 on ft So, x,y E
For each n consider the multiplication operator p generated
by the function = Le., for each z E The unique
continuous linear extension of to all of E will be denoted by again.
Since 1 in C(c2), we see that z each

is dominated by the identity operator. Moreover, an easy computation
shows that xI holds in This implies
andso
The remainder of the section is devoted to some commutativity proper-
ties of multiplication operators on Accordingly, for the rest of
this section, will denote a fixed Hausdorff compact space. As usual, each
E generates the multiplication operator defined
by = We start our discussion with a useful property of weakly
compact multiplication operators.
Lemma 4.18. If a multiplication operator is weakly compact, then the
function vanishes at every accumulation point of ft
In particular, if does not have isolated points, then the only weakly
compact (and hence the only compact) multiplication operator on is
the zero operator.
Proof. Assume that is weakly compact. Suppose by way of

tion that 0 for some accumulation point wO E ft Therefore, there
is an open neighborhood V of wo and some 0 such that for
all v E V. By the UrysohnTietze extension theorem there exists a function
E such that for each w E V, the closure of V.
=
From = we see that is a weakly compact operator on
C(Ifl Since the function equals one at each point of V, it follows (again
by the Urysohn.Tietze extension theorem) that the identity operator I on
C(V) is weakly compact. This implies that C(V) is reflexive, and hence
(by Theorem 2.26) its norm must be order continuous. But then V must be
a finite set (see Exercise I in Section 2.3), and so w0 is an isolated point,
which is a contradiction.
Recall that if T: X X is an operator on a Banach space, then its

commutant {T}' consists of all bounded operators on X that commute
with T, that is,
{T}' = {s E (X): ST = TS }.
A direct verification shows that the commutant of an operator T is a unital
sub algebra of (X). In general, it is a very difficult task to describe the
commutant of a given operator. Our next goal is to describe some classes of
operators that commute with a given multiplication operator. The rest of
the material in this section is taken from [15].
For a function E we shall denote by Act, the closed unital (i.e.,
containing 1) algebra generated by in where the product of functions
Multiplication Operators 139
in C(12) is always understood as the pointwise product. This algebra is

the uniform closure of the subspace {p(q5): p is a polynomial}. That is,
= {p(q5): p is a polynomial}.
It should be noted that besides being an algebra, Ath is also a Riesz subspace
of
Lemma 4.19. For a function q5 e C(12) and an arbitrary bounded operator
T: 0(e) C(12) commuting with the following statements are true.
(1) There exists a function h E 0(Q) satisfying Tx = hx for each x in
the closed unital algebra
(2) TMf = MfT holds for each f E
(3) If is one-to-one, then T itself is a multiplication operator. In
q5
other words, if q5 is one-to-one, then the commutant of consists

of all multiplication operators, i.e, = {Mf: f
Proof. (1) The identity TMj, = means that T(bx) = bTx for each
x E C(12). By induction, we get = bmTx for each n 2 0 and all
x E C(12). Thus,
T(p(b)x) = p(b)Tx (*)
for each polynomial p(t) of one real variable t and all x E
Letting x = 1 and h = Ti, it follows from (*) that T(p(b)) = hp(b) for
each polynomial p. In other words, Tx = hx for all x in the unital algebra
generated by q5. Consequently, by continuity, Tx = hx for all x in the closed
unital algebra generated by q5.
(2) Let f Aq5. If x E C(12), then from (*) we see that

TMf(X) = T(fx) = fT(x) = MfT(X),
so that TMf = MfT.
(3) If q5 is one-to-one, then the StoneWeierstrass Approximation Theo-
rem guarantees that = C(12). So, by part (1), T = Mh. D
If q5 is not one-to-one, then the corumutant ofmight include operators

other than multiplication operators. For an example take = [1, 1] and
= An easy argument shows that consists of all even functions. If
weconsidertheoperatorT: C[1, 1] C{1, 1] =
then T is not a multiplication operator. Clearly,
The next surprising result asserts that a continuous function is constant
on an open set if and only if its multiplication operator commutes with a
non-zero finite-rank operator.
Theorem 4.20. For a multiplication operator generated by a function

E C(Q) the following statements are equivalent.
(1) The function is constant on some non-empty open subset of ft
(2) commutes with a rank-one positive operator.
(3) commutes with a rank-one operatorS
(4) commutes with a non-zero finite-rank operatorS
Proof. (1) (2) Suppose that each E V, where V is a

= c for
non-empty open subset of ft Choose a non-zero positive function E 0(Q)
such that = 0 for all V and fix some E V. Now define the
operator K: 0(Q) via the formula Kx =
We claim that and K commute. Indeed, for x E C(Q) and E Q,
we have
= = =
On the other hand, since vanishes off V and = c on V, we also have
that
= = =
Therefore, K is also a positive operator.
(2) (3) (4) Obvious.
(4) (1) Assume that the multiplication operator commutes

with a non-zero finite-rank operator T = where E 0(Q)
and E 0* (Q) for each i. We can suppose that , are linearly
independent and that 0 for each i. From Lemma 4.19(2), we know that
MfT = TMf for each function f in the closed unital algebra generated
by in C(Q). That is,
k = k
(*)
holds for all f e and for each E where as usual M: (Q) + (Q)
is the adjoint operator of Mf.
We claim that C Ker for each i and each function

f e Aj,. To see this, fix f E and let x E Ker Then (*) yields
M = 0, and from the linear independence of , it
follows that Mx7 (x) = 0 for each i. Thus, Ker Ker
Consequently, each linear functional belongs to the vector subspace
2Here we use the conclusion of Exercise 1 in Section 4.1: A linear functional f belongs to
the span of the linear fun fit.. if and only if
. Ker C Ker f.
generated by xi,. . . , In particular, for each i the linear functionals

Mx7, M2 xi,. Mk1 . . , must be linearly dependent. That is, for each
i there exist constants , not all zero such that
k+1
ai,jIVIq5iXi = =
j=1
where pj is the non-zero polynomial defined by = ajjAi.
Next, consider the non-zero polynomial p = Pk and note that
= [fl =0
for each i. Letting f = in (*), we get = 0 or
for allx and each wE ft

Now if we choose x E C(Q) with (x) then it follows from the linear
independence of .. , that Q for some wo E ft

Therefore, 0 for all w in an open neighborhood V of w0.
But then (**) yields
= =0 for all w E V. (* *
Since A0 = is a root of the polynomial p, and since p has a finite
number of roots, there exists an open neighborhood N of A0 that does not
contain any other roots of p. Finally, if we consider the open neighborhood
W = V n (p a of wO, then it should be clear from (* * *) that
= A0 for each w E W, and the proof is finished.
Exercises
1. Prove the following implications:
Order Unit Quasi-interior Point Weak Order Unit
Also show that in general no converse implication is true.
2. Show that a Banach lattice E has a weak unit if and only if E has an
at most countable maximal subset of pairwise disjoint vectors. [HINT: If
{ x1, .} is a maximal countable set of pairwise disjoint positive unit
vectors, then x = is a weak unit]
3. Show that every separable Banach 'attice has a quasi-interior pointand
hence it also has a weak unit.
4. In a Banach lattice with order continuous norm, show that a positive

vector is a quasi-interior point if and only it is a weak unit.
Prove that the Banach lattice M[O, I], of all signed Borel measures on
[0, 1] of bounded variation, does not have quasi-interior points. [HINT:
Notice that since M[0, IJ has order continuous norm, the notions of weak
unit and quasi-interior point coincide. Assume that 0 0
for each x E [0, 1]; otherwise p({x}) = 0 implies p A = 0, where
is the Dirac measure supported at x, which is impossible. However, the
latter conclusion shows that [0, 1] is countable, a contradiction.]
6. Show that if E is a Banach lattice, then its set of quasi-interior points is
either empty or else it is dense in E+.
7. If T: E F is a strictly positive operator between two Banach lattices,
then show that the adjoint operator T*: F* carries quasi-interior
points to quasi-interior points.
8. Fix a function q5 E (where is compact and Hausdorif) and let
denote the closed unital algebra generated by q5 in That is,
= {p(q5): p is a polynoinial}.
Show that is a Riesz subspace of
9. Let E, p) be a cr-finite measure space and let q5 E Show that
for the maltiplication operator (p) f where I p CX),
(a) The function q5 is constant on some measurable set of positive mea-
sure.
(b) Met, commutes with a rank-one positive operator.
(c) commutes with a rank-one operator.
(d) Met, commutes with a non-zero finite-rank operator.
10. Let -r: be a continuous mapping on a compact Hausdorif space
For the composition operator defined by the
formula (x) x o -r, establish the following.
(a) is a contractive positive operator satisfying = 1.
(b) commutes with a rank-one positive operator.
11. Show that for a continuous map -r: where is compact and
Bausdorif, the following statements are equivalent.
(a) The map 'r is surjective.
(b) The composition operator C.,- is an isometry.
(c) The composition operator C.,- is one-to-one.
4.3. Lattice and Algebraic Homomorphisms

The material presented so far has already demonstrated the usefulness of
in the study of positive operators. As explained earlier, the
importance of stems, on one hand, from the fact that every
principal ideal in a Banach lattice is an AM-space with unit and, on the
other hand, from the classical KakutaniBohnenblustKrein Theorem 3.6 on

the representation of with unit. In this section, we shall present
the basic characterizations of lattice and algebraic homomorphisms between
as well as some extension properties of lattice homomorphisms.
We start with the following result.
Theorem 4.21. Let E be a Banach lattice, let 2 be an at most countable
subset of E and let C be an at most countable collection of regular' operators
on E. Then there exists some u > 0 such that:
(1) The set J lies in i.e., J
(2) Every operator in C leaves the principal ideal invariant, that is,
T E C.
Proof. The proof will be done in three steps.
Step I: If J is at most countable, then J is included in a principal ideal.
Assume that J is a countable set, say J = {Xi, x2,.. .}. We can assume
that 0 for each i. If u = then note that J
Step II: If J is an at most countable subset of E and T E L(E) is a regular
operator, then there exists some u> 0 such that and
T a non-zero regular operator. So, we can
is
write T=T1 T2 with T1,T2 0. Put 8= +T2 >0.

By Step I, there exists some w > 0 such that J En,. Next, consider
the positive vector u = Since w u, it follows that J c
It is now a routine matter to verify that
Step III: The general case.
Assume that C is countable, say C = {T1, .}, where
.
0 for each
i. Write T2 = with 0, and let = + > 0.
Next, consider the positive operator S = By Step II, there
exists some u > 0 such that J C and Next, fix any x E
and pick some A > 0 such that lxi Au. Then for each k we have
<
= Cl
Ah
ci
So, Tkx E that is, holds for each k.
The important conclusion from Theorem 4.21 is the following. When one
represents as a then all vectors in J as well as all vectors TkX,
xE can be viewed as continuous functions on the compact Hausdorif
space We can then use the properties of the continuous functions to
144 -- 4. Special Classes of Operators
study the properties of positive operators. In addition, expressions like &,

etc. (where x and y are vectors in as well as if x 0 make
sense as functions in and hence they are well-defined vectors in E!
It should also be noticed that since is an ideal in E, its complex-
ification can be identified with the Banach space of all
complex-valued continuous functions on
We now turn our attention to the properties of lattice homomorphisms
between As we shall see, they are closely related to algebraic
homomorphisms that are defined next.
Definition 4.22. An operatorT: C(Q) is saidtobe an algebraic
homomorphism (or a multiplicative operator) if
T(fg) = T(f)T(g)
holds for all f,g e
We first characterize the lattice homomorphisms from a to
the real numbers.
Lemma 4.23. A non-zero linear functional R is a lattice ho-
momorphism if and only if there exists some c> 0 and some wo E (both
uniquely determined) such that =
c 0, thenis clearly a lattice homomorphism,
that is, A g) = for all f, g e
For the converse, assume that R is a lattice homomorphism.
Clearly, is a positive linear functional. Hence, by the Riesz Representation
Theorem, there exists a unique regular Borel measure p on satisfying
ff fe We claim that the support of ,a is a
singleton. To see this, assume that two distinct points, say s and t, belong
to the support of ji. Pick two functions f, g e satisfying f A g = 0 and
f(s) = g(t) = 1. This implies
0 Ag) = = min{ffd/2,fgd/2} > 0,
which is a contradiction. Hence, the support of consists of a single point,
say wo. This implies
f =
f e C(cl).
The algebraic homomorphisms from a to the complex num-
bers are characterized in the next result.
Lattice and Algebraic Homomorphisms 145
Lemma 4.24 (I. M. Gelfand). A non-zero linearfunctional q5: C

is multiplicative if and only if there exists a (uniquely determined) point
c,t'o E such that =
Proof. If then = = holds

for all f, gE That is, q5 is a multiplicative linear functional.
For the converse, assume that q5 is a non-zero multiplicative linear func-

tional on C(ffl. Since = = and is non-zero, it fol-
lows that 1 is the constant function one on ft Let
M = {x E = o} be the kernel of the functional We claim
that there exists a point E such that M {x E = o}.
If there is no such point, then for each we can find a function in
M such that 0. Since is continuous, there exists a neighborhood
such that
of the point ci.' 0 for each t By the compactness
of there exist neighborhoods covering the whole of ft Next,
,
consider the function x = where, as usual, denotes the

complex conjugate of a function z E By our choice of the functions,
> 0 for each ci.' E and hence y = Now the identity
xy = 1 coupled with the multiplicativity of yields = 1, and
so 0. On the other hand, as = 0 for each a.', we certainly
have q5(x) = 0, a contradiction. Thus, there exists some such that
Mc{xEC(cfl: x(L0)=0}.
Next, we shall show that {x E C(Ifl: = o} M. To this end,
let x E satisfy = 0. If u = q5(x) 1 x, then clearly u E lvi, and so
o = = Hence x E lvi, and thus lvi = {x E
x x M, E and so =0 or
(x) = as desired. The uniqueness of follows from the separation
We are now ready to characterize the lattice homomo
Theorem 4.25. A positive operator T: 0(Q) is a lattice homo-

morphism if and only if there exist a mapping 'r: Q * and some weight
function w E C(Q) such that for each f C(Ifl and each q E Q we have
Tf(q) = w(q)f('r(q)).
Moreover, in this case, w = T1c2 and the mappimg 'r is uniquely determined
and continuous on the set {q E Q: w(q) > 0}.
Proof. If T is of the above form, then clearly T is a lattice homomorphism.

For the converse, assume that T is a lattice homomorphism. Then for each
q E Q, we have
(8q 0 T) (f) = 8q(Tf) = [Tf] (q).
Since T is a lattice homomorphism, the linear functional 8 a T: C(11) R
is a lattice homomorphism. Thus, by Lemma 4.23, there exists a unique
constant w(q) 0 and some (not necessarily unique) point T(q) E such
that
{Tf](q) = (8q o T)(q) = w(q)f(r(q)).
Letting f = lci, we obtain w = E C(Q). Moreover, if w(q) >
then 0,
T(q) is uniquely determined. We leave it as an exercise to verify that 'r is
continuous on the set {q E Q: w(q) > 0}.
Here is a connection between lattice and algebraic homomorphisms.

Theorem 4.26. Every algebraic homomorphism between real
is a lattice homomorphism. However, the converse is false.
Proof. Let T: C(Q) be an algebraic homomorphism. Clearly, T is

a positive operator. Indeed, if f 0, then Tf = T((v"7)2) = [T(v"7)}2 0.
Now if f E then
T(f)f2 = [T(f)]2 = T(f2) = =
and so =
The algebraic homomorphisms are characterized as follows.
Theorem 4.27. An operator T between real or complex C(K)-spaces is an
algebraic homomorphism if and only if there exist a unique clopen subset V
of Q and a mapping 'r: Q which is continuous on V such that
or)
for each f The mapping r is also uniquely determined on V.
Proof. If T has the above form, then T is clearly an algebraic homomor-

phism. For the converse, assume that T is an algebraic homomorphism. We
shall consider the real and the complex cases separately.
CASE I: Assume that T: C(Q) C(Q) is an algebraic homomorphism
between real C(K)-spaces.
By Theorem 4.26; T is a lattice homomorphism. So, from Theorem 4.25,
it follows that there exist a mapping T: Q and some w E C(Q) such
that for each f E and each q E Q we have
Tf(q) = w(q)f(T(q)).
Lattice and Algebraic Homomorphisms 147
Moreover, we know that w = and that the mapping r is uniquely

determined and continuous on the set {q E Q: w(q) > O}. Now use the fact
that T is an algebraic homomorphism to get
= = = = w.
This implies that w = xv for a unique clopen subset V of Q.
CASE II: Assume that T: Cc(Q) is an algebraic homomorphism
between complex C(K)-spaces.
For each q E Q, we have
(1) = 8q(Tf) = [Tf] (q).
(q 0 T)
Since T is an algebraic homomorphism, it follows that the linear func-

tional 8 o T: C is likewise an algebraic homomorphism. Thus,
by Lemma 4.24, there exists some point 'r(q) E ci such that
[TfJ(q) = o T)(q) = f(T(q))
for each f Cc(ci). From = = it follows that

= Xv for a unique clopen subset V of Q. To finish the proof, notice
that Tf = = = xv(f o T).
For Markov operators between C(ci)-spaces, the notions of algebraic
homomorphism, lattice homomorphism, and composition operator coincide.
Specifically, we have the following result whose proof follows easily from the
preceding discussion and is left to the reader.
Theorem 4.28. For a Markov operator T: C(ci) C(Q) the following
(1) T is an algebraic homomorphism.
(2) T is a lattice homomorphism.
(3) T is a composition operator.
The proof of the next resultwhich characterizes in terms of complexi-
fications the lattice homomorphismsis left as an exercise. (See Exercise 5
at the end of the section.)
Theorem 4.29. A positive operator T: E F between two Banach lattices
is a lattice homomorphism if and only if the operator Tc: f satisfies
!Tczl = TIz) for all z E
Our section deals with the "automatic continuity" of

next result in this
the multiplicative functionals on Banach algebras. In Lemmas 4.23 and 4.24
we did not assume the continuity of the functional a5. The continuity of
was a consequence of the lattice or algebraic properties. In the case of a
148 4.Special Classes of Operators
real algebraic homomorphism R the continuity of is a

quence of its positivity. Moreover, as we shall see next, the continuity of
these functionals is a very special case of the continuity of multiplicative
functionals on Banach algebras. The phenomenon of the automatic con-
tinuity of multiplicative functionals and multiplicative homomorphisms on
and on many other algebras is well known and very important. We
refer to [95, 199, 304] for details.
Theorem 4.30. A multiplicative linear functional on any Banach algebra
is continuous and has norm at most one.
Proof. Let be a multiplicative functional on a (commutative or not)

nach algebra A. A thought shows that whether is discontinuous
or has norm greater than one, there exists some a e A with 11aM < I and
cb(a) = 1. If we let b = then ab = b a, and the multiplicativ-
ity of yields = That is, = 1, which is
impossible.
A few remarks regarding the existence of non-trivial multiplicative func-

tionals and their automatic continuity are in order. There are two similar
instances of automatic continuity in Banach lattices. The first one has al-
ready been discussed earlier and it asserts the continuity of any positive
operator on a Banach lattice (Theorem 1.31). This makes the automatic
continuity phenomenon of positive operators different from that of multipli-
cative homomorphisms, since, as shown by II. Dales and J. Esterle [96}, for
an arbitrary multiplicative homomorphism on a the automatic
continuity may fail. Suiprisingly, band preserving operators on Banach lat-
tices are also automatically continuous; see Theorem 3.29.
The existence of non-trivial multiplicative functionals on Banach alge-
bras is another delicate problem. As noted above each Banach algebra
carries a multitude of multiplicative functionalsfor each point w E there
corresponds a continuous multiplicative functional 8k,. For the Banach alge-
bra (X) of all bounded operators on a Banach space X this problem has
received special attention. If I < dimX < oo, then the zero functional is
the only multiplicative functional on L(X). Moreover, if Y is an arbitrary
Banach space and X = Y Y ... Y, that is, X is the direct sum of
ri> 1 copies of then again only the trivial multiplicative functional exists
on L(X); see Exercise 7 at the end of this section. For a general Banach
space X the situation is quite different. To the best of our knowledge it was
B. S. Mityagin and I. C. Edelstein [244] who were the first to exhibit two
types of Banach spaces X with non-trivial multiplicative functionals on the
Banach algebra L(X). For instance, they showed that the famous Banach
space J constructed by R. C. James [159} can be taken for such an X. This
result was generalized by A. Wilansky [337]. Nowadays, using the deep re-
suits of \'V. T. Cowers and B. Maurey from [137] and [138], it has become
possibie to produce easily a Banach space X such that (X) has non-trivial
multiplicative functionals; see Exercise 7 in Section 7.1 and [200]). All the
Banach spaces X with non-trivial multiplicative functionals on (X) men-
tioned above are non-reflexive. It was P. Mankiewicz [231] who constructed
a reflexive (even a superreflexive) Banach space X with infinitely many mul-
tiplicative functionals on (X). See also [97] for another construction of a
Banach space with a similar property.
liVe shall close the section with an interesting extension property of lat-.
tice homomorphisms. In order to describe this property, we need some pre-
liminary discussion regarding extensions of positive operators. Recall that a
mapping p: X Z, from a (real) vector space to an ordered vector space,
is said to be sublinear if:
(a) p(x + y) p(x) +p(y) for all x,y E X.
(b) p(Ax)=Ap(x)foraliAOandalixeX.
An almost verbatim repetition of the proof of the classical HahnBanach
theorem yields the foliowing result regarding extensions of linear operators.
Theorem 4.31 (HahnBanachKantorovich). Let Y be a vector subspace
of a (real) vector space X, let F be a Dedekind complete Riesz space, and
let p: X F be a sublinear mapping. If an operator T: Y F satisfies
Ty p(y) for all y e Y, then there exists an operator 8: X F that
extends T and satisfies Sx <p(x) for all x E X.
As an appiication of this theorem we present the foliowing usefui result.

Theorem 4.32 (Kantorovich). LetY be a majorizing vector subspace of an
ordered vector space X. Then every positive operator from Y to a Dedekind
complete Riesz space has a positive linear extension to all of X.
Proof. Assume that Y is a majorizing vector subspace of an ordered vector

space X, and let T: Y F be a positive operator from Y to a Dedekind
complete Riesz space F. Define the mapping p: X + F via the formula
p(x)=inf{Ty: yEYandyx}.
To see that the above infimum exists in F, we argue as foilows. Since Y is
majorizing, there exists some z E Y with z x or z x. Simiiarly, there
exists some y E Y with y x and for any such y the positivity of T implies
Tz = T(z) Ty. That is, the subset {Ty: y E Y and y x} of F is
non-empty and bounded from below in F. The Dedekind completeness of
F guarantees the existence of its infimum.
4. Special Classes of Operators
A straightforward verification shows that p is a sublinear mappilig sat-

isfying Ty = p(y) for each y Y. Theorem 4.31 guaraiitees that T has a
linear extension S to all of X satlsf3Tlng Sx p(x) for each x E X. Now if
xE then x 0 and so Sx = S(x) p(x) T(0) = 0. Thus,
Sx 0, and hence S is a positive linear extension of T.
If Y is a vector subspace of an ordered vector space X and T: Y F

is a positive operator, where F is a Riesz space, let us denote by &(T) the
(possibly empty) collection of all positive linear extensions of T to all of
X. It should be clear that &(T) is a convex subset of the vector space
Er (X, F). When F is a Dedekind complete Riesz space, the extreme points
of the convex set &(T) have been characterized by Z. Lipecki, D. Plachky,
and W. Thomsen [208] as follows.
Theorem 4.33 (LipeckiPlachkyThornsen). Let Y be a vector subspace of

a Riesz space E, let F be a Dedekind complete Riesz space, and let T: Y F
be a positive operator. Then an operator S &(T) is an extreme point if
and only if for each x E E we have
yEY} =0.
Proof. Fix S E &(T). Assume first that S is an extreme point of &(T).
Define the mapping p: X F by
yEY}.
Clearly, p is a sublinear mapping that satisfies 0 p(x) = p(x) for
eachxEEandp(z)=Oforeach zEY.
To see that p(x) = 0holds for each x E E, assume by way of contradic-
tion that p(u) > 0 for some u E E. Let Z = {Au: A E R} and define the
operator R: Z p F by R(Au) = Ap(u). From R(u) = p(u) > 0, it follows
that R is non-zero. Also, it is easy to check that R(Au) p(Au) for all A E
and so by Theorem 4.31 there exists a positive linear extension of R to all of
E (denoted by R again) such that Rx p(x) for each x E E. Since y E Y
implies R(z) = R(z) <p(z) = p(z), it follows that
Rz = 0 for each z E Y. Now notice that for each x E X+ we have
R(x) p(x) Sx and R(x) = R(x) p(x) SI = Sx.
This implies S R 0 and S + R 0. Since Ry = 0for each y E Y, we
see that S R, S + R E &(T). Now notice that S + R 5, 5 R 5, and
S= + R) + R) contradict the assumption that S is an extreme
point of &(T). So, p(x) = 0 for each x E E.
For the converse, assume that S satisfies inf {SIx yI: y E Y} = 0 for
each x E, and let two operators A, B E &(T) and some 0 < a < I satisfy
S = aA + (1 a)B. Then for each x, y e E we have
In particular, if y e Y, then Ay Ty = Sy, and so for each x e E we have
From inf {SIx yl: y e Y} = 0, we see that Ax = Sx for each x e E. This

implies A = B = 5, and thus S is an extreme point of E(T).
Corollary 4.34. Let C be a Riesz subspace of a Riesz space E, let F be
aDedekind completeRieszspace,--and letT:G 4F be-a lattice
phisrn. Then each extreme point of the convex set E(T) is also a lattice
homomorphism.
Proof. Let S be an extreme point of the convex set E (T). According to

Theorem 4.33 for each x e E we have
inf{SlxyI: yEG}=O. (*)
Since T is a lattice homomorphism, = TyJ = TIyJ = for each

y E C. Therefore for each x E E and each y e C we have
ISxI fSyJj + SIxII
This implies, using (*), that SxI = for each x E E, that is, S is a
lattice homomorphism.
The next result of Z. Lipecki [207] presents a condition that guarantees

the existence of an extreme point in E(T).
Theorem 4.35 (Lipecki). If Y is a mnajorizing vector subspace of a Riesz
space E, F is a Dedekind complete Riesz space, and T: Y E is a positive
operator, then the non-empty convex set E(T) has an extreme point.
Proof. The proof is a clever application of Zorn's lemma. Observe first

that by Theorem 4.32 the convex set E(T) is non-empty. According to
Theorem 4.33 we must establish the existence of an operator S e E(T)
satisfying
inf{Slx_yl:
for each x e X. Start by considering pairs of the form (Z, A), where Z is
a vector subspace majorizing E and A: Z F is a positive operator. For
any such pair (Z, A), we define the mapping Z
Z z ? x}.
A direct verification shows that PZ,A is a sublinear mapping such that

PzA (z) = Az for each z E Z. Moreover, if two pairs (Z1, A1) and (Z2, A2)
satisfy Z1 Z2 and A2 = A1 on Z1, then PZ2 A2 (x) PZ1,A1 for all x E E.
Now let C be the collection of all pairs (Z, A) such that:

(a) The vector subspace Z contains Y.
(b) The positive operator A: Z F coincides with T on Y.
(c) For each z E Z we have inf{pZA(Iz yE Y} = 0.
Since (Y, T) E C, the set C is non-empty. Next, define an order on C by:
(Z2, A2) >- (Z1, A1). if Z2 J Z1 and A2 = A1 on Z1.
It is easy to check that is indeed an order on C, and a routine argument
shows that every chain in C has an upper bound. So, by Zorn's lemma, there
exists a maximal element, say (V, 8). To complete the proof, it suffices to
show that V = E. If this is established, then it should be clear that S =
and Theorem 4.33 guarantees that S is an extreme point of
If V E, then there exists some x E E such that x V. Consider the
vector subspace W = {v + v E V and E R} and define the operator
B: TV F by = Clearly, V is a proper subspace of
W, B = S on V, and we claim that B is a positive operator. To see this, let
0 u+ E V. If,\ > 0, then x and this implies
from which it follows that B (v + = Sv + (x) 0. The case < 0
is similar and the case = 0 is obvious. To obtain a contradiction, we shall
establish that (W, B) satisfies property (c) above.
To this end, notice first that the sublinearity of PW,B implies that the set
is a vector subspace of E satisfying V V1. We claim that x E V1. To

see this, observe first that v E V and v x imply V x E TV and so
Pw,B (v x) = B(v x) = Sv Therefore,
o vEV}
< inf (v x): v E V and v
inf { Sv (x): v E V and v x }
inf { Sv: v E V and v (x) 0.
This shows that x E V1, and consequently W V1. Now fix w E TV and put
a = inf {PwB y yE Y each E
a Pw,B(1WY1) PWBftW vI)

PW,B(kI) vI) yl).
Since (V, 8) E C, we know that y E Y} = 0, and hence

0 a PW,B(Iw vi) for each v E V. Taking into account that w E V1,
the latter inequality implies a = inf vi): v E v}
= 0 for
each w E I'V. This establishes (W,B) E C. However, (W,B) (V,S)
and (W, B) (V, 8) contradict the maximality of (V, 8), and the proof is
finished.
The final result on extensions of lattice homomorphisms (due to

Z. Lipecki [207] and W. A. J. Luxemburg and A. R. Schep [223]) is a
simple consequence of the preceding discussion.
Corollary 4.36 If C is Riesz
subspace of a Riesz space E, then every lattice homomorphism from C to a
Dedekind complete Riesz space has a lattice homomorphic extension to all
of E.
Proof. By Theorem 4.32 the set E(T) is non-empty and by Theorem it
has an extreme point 8. Since T is a lattice homomorphism, Corollary 4.34

guarantees that 8 is a lattice homomorphic extension of T.
For more on extensions of lattice homomorphisms see [81, 82, 83].
Exercises
1. By Theorem 4.25, for any given lattice homomorphism T: C(Q) C(Q)

there exist a mapping r: Q and some w E C(Q) such that for each
f and each qEQ we have
Tf(q) w(q)f(r(q)).
Show that w = and that the mapping r is uniquely determined and
continuous on the set {q Q: w(q) > O}.
2. Give an example of a lattice homomorphism which is not an algebraic
homomorphism.
4. Let and Q be two compact Hausdorif spaces, and consider the convex
set C of all Markov operators from
C ={TE (C(12),C(Q)): T O and
T a lattice homomorphism if and only if
T is an extreme point of C.
5. Prove Theorem 4.29. [HINT: Assume that T: EF is a lattice homo-
morphism between real Banach lattices. Let z = x + zy E and put
EF. Then and
this shows that we can assume without loss of generality that E C(1l)
and F = C(Q) and that T is a Markov operator. Now use Theorem 4.28
and view T: C(Q) as a composition operator.]
6. For a multiplicative linear functional on a Banach algebra A establish

the following.
(a) If a A is an idempotent element (i.e., a2 a), then show that
q5(a) = 0 or
q be two idempotent elements in A. Then show that
the linear functional defined on A by = b(qap) is also a
multiplicative functionaleven though the mapping a 'p qap need
not be an algebraic homomorphism.
7. Fix a Banach space X and let V = X e X E13 X be the direct
sum of n > I copies of X; in particular, let V = or V = where
n> 1. Show that the Onl3T multiplicative linear functional on the Banach
algebra C(Y) is the zero functional. (For an example of a non-trivial
multiplicative function on C(X) see Exercise 7 in Section 4.5.) [HINT:
Let (/) be a multiplicative linear functional on C(Y). It suffices to show
that b(I) 0. Define the bounded operator J: V V by
(xfl,xl,x2,...
and note that Jfl = I. Therefore, in order to show that cb(I) = 0, it suffices
to establish that q5(J) = 0. Now for each 1 i < n define the bounded
operator E /2(V) by .. = (0,... ..,0), where

the vector occupies the (i + 1)-position. Also, define E /2(V) by
X2,.. . = 0, 0, . . . , 0). Clearl3r, T2 = 0 for all 1 i n
and
8. For a Banach lattice E, a majorizing vector subspace X of E, and an arbi-
trar3r normed Riesz space F establish the following automatic continuit3r
propert3r: Every positive operator T: X F is continuous. [HINT: B3T

Exercise 21 of Section 1.2 we can assume that F is Dedekind complete.
Now, by Kantorovich's Theorem 4.32, the operator T has a positive linear
extension T: E F which is continuous by Theorem 1.31.]
9. (Arendt [40]) Let F be an arbitrary Dedekind complete Riesz space and
let R: E H be a positive operator between two Riesz spaces. Then
the positive operator R gives rise to a positive operator T F* TR from
Cr(H, F) to Cr(E, F). Establish the following.
(a) If R is a lattice homomorphism, then T F* TR is interval preserving.
(b) If R is interval preserving, then T TR is a lattice homomorphism.
10. (AliprantisBurkinshaw--Kranz [33]) Let R: H F be a positive op-
erator between two Dedekind complete Riesz spaces and let E be an
arbitrary Riesz space. For the positive operator T F* TR, from /Cr(H, F)
to Cr(E, F), establish the following.
(a) If R is an order continuous lattice homomorphism, then T F* RT is
a lattice homomorphism.
(b) If the operator T F* RT is a lattice homomorphism and the order
dual is non-trivial, then R is a lattice homomorphism.
4.4. Fredholm Operators 155
11. Prove Theorem 1.35. [HINT: Use the preceding two exercises by observing
that if T: E F is a bounded operator between two Banach lattices,
then T*y* aT for each E F*.]
12. Let E = and let c denote the Riesz subspace of all convergent se-
quences. Show that the limit functional Lim: c p defined by
im(x) = x= (x1, x2,...) E c, is a lattice ho-
momorphism that has a lattice homomorphic linear extension to
4.4. Fredhoim Operators

The objective of this section is to the basic properties of the class of
Fredhoim operators. They were introduced by I. Fredhoim [127] in his study
of integral equations. These are the operators that have "small deviations"
from isomorphisms. To formalize the term "small deviation," we need some
preliminary discussion.
Let T: X
- Y be an operator between two vector spaces. As usual, we
denote by N(T) and R(T) the null space and the range of T, respectively.
That is,
Tx=O} and R(T)={Tx: xEX}.
Clearly, T is a linear isomorphism from X onto Y if and only if N(T) = {O}
and R(T) = Y or equivalently Y/R(T) = {O}. That is, T is an isomorphism
if and only if dimN(T) = 0 and dimY/R(T) = ft
Assume next that for a bounded operator T: X Y we can find a
closed vector subspace V of X and a closed vector subspace W of Y such
that X = N(T) V and Y = R(T) W. Then Y/R(T) is isomorphic to W
and T: V R(T) is an isornorphism. Moreover, it should be clear that the
"measure of deviation" of T from being an isomorphism is associated with
the subspaces N(T) and W. The dimension of the null space N(T) is called
the nullity of T and is denoted n(T), i.e., n(T) = dim N(T). The dimension
of W (or, equivalently, the dimension of the quotient space Y/R(T)) is called
the defect of T (6i' the codimension of T or even the corank3 of T) and
is denoted d(T). That is, d(T) = dimY/R(T). If n(T) or d(T) is finite, then
the extended real number
i(T) = n(T) d(T)
is called the index of the operator T.
31f W is a vector subspace of a vector space V, then the codirnension of W is the dimension
of the quotient vector space V/I'V If T: V U is a linear operator between vector spaces, then
the rank of T is the dimension of the range R(T) and the corank of T is the codimension of the
range R(T).
Definition 4.37. A bounded operator T: X Y between two Banach

spaces is said to be:
(1) serni-Fredholm if it has a closed range and either its nullity or
its defect is finite and
(2) Fredholrn if its nullity n(T) and defect d(T) are both finite.
Clearly, every linear isomorphism between two Banach spaces is a Fred-
hoim operator of index zero. Also, any operator between finite dimensional
spaces is automatically a Fredholm operator. It is customary to impose in
the definition that a Fredholm operator has a closed range. However, this
is a redundant assumption.
Lemma 4.38. Any operator T: X Y between Banach spaces with a finite
defect has a closed range and satisfies
i(T) = dimN(T) dim N(T*)
In particular, every Fredhoim operator has a closed rangeand hence it

is also a semi-Fredhoim operator.
Proof. The fact that the operator T has closed range follows from Corol-
lary 2.17. For the index formula notice that according to Theorem 2.13 we
have d(T) = dim [Y/R(T)] = dim N(T*).
In our discussion we shall use the following lemma whose proof follows
from Exercise 6 in Section 4.1. Many authors use the conclusion of this
result as the definition of a Fredhoim operator.
Lemma 4.39. If T: X Y is a Fredholm operator between Bamach spaces,
then there exist a closed vector subspace V of X and a finite dimensional
vector subspace TV of Y such that
X = N(T) V and Y = R(T) W.
In particular, the surjective operator T: V R(T) is an isomorphism.
Regarding Fredholrn operators between real Banach spaces we have the

following result whose straightforward proof is left as an exercisesee Ex-
ercise 1 at the end of the section.
Lemma 4.40. A bounded operator T: X Y between real Bamach spaces
is a Fredhoim operator if and only if the operator Tc: Yc is likewise
a Fredhoim operatorin which case we have i(Tc) i(T).
The sum of two Fredhoim operators need not be a Fredholm operator.

For instance, on an infinite dimensional Banach space the identity operator
I and I are Fredhoim operatorsboth of index zero. However, their sum

is the zero operator which is not a operator.
The classes of compact operators and Fredhoim operators are disjoint.
Lemma 4.41. If X is an infinite dimensional Banach space, then no com-
pact operator with domain X can be a Fredhoim operator.
Proof. Let T: X Y be a compact operator between two Banach spaces

with X infinite dimensional. Assume by way of contradiction that T is also
a operator. This means that dimN(T) < oo and there exist a
closed subspace V of-X and a finite dimensional subspace W of Y
X = N(T) V and Y = R(T) W. Clearly, V is infinite dimensional and
the operator S = Tlv: V R(T) is an isomorphism that is also compact.
This implies that V is finite dimensional, a contradiction.
Informally speaking, the preceding discussion shows that a compact op-

erator can be thought of as an "almost" finite-rank operator while a Fred-
hoim operator should be perceived as an "almost" isomorphism. As ex-
pected, the adjoint of a Fredholm operator is also a Fredholm operator.
Theorem 4.42. A bounded operator T: X Y between two Banach spaces
is a Fredhoim operator if and only if its adjoint operator T*:
is also a Fredhoim operator. Moreover, if this is the case, then we have
ri(T*) = d(T) and d(T*) ri(T); and so i(T*) = i(T).
Proof. Assume T is a Fredholm operator. Choose a closed subsp ace V of X

and a finite dimensional vector subspace W of Y such that X = N(T) V
and Y = W R(T). Also, let S = TIv, and note that 5: V R(T) is
an isomorphism. Consequently, T can be represented by the direct sum
-operator T = 0 5: N (T) V W R(T). Taking adjoints, it follows
that T* = 0 5*: W* [R(T)]* [N(T)}* This implies N(T*) = W*
and R(T*) = V*. Therefore, T* is a Fredhoim operator satisfying
n(T*) = dimW* = dimW = d(T)
and
d(T*) = dim [X*/R(T*)] = dim [N(T)]* = dim N(T) = n(T).
the converse, assume that T* is a Fredholm operator. Then, by
For
the previous case, T**: is also a Fredholm operator. From
N(T) N(T**), it follows that dimN(T) < oo. Now recall that R(T)
is closed if and only if R(T*) is closed; see Theorem 2.18. Therefore, by
Theorem 2.13 we have dim [Y/R(T)] = dimN(T*) <oo. This shows that T
is likewise a Fredholm operator.
A composition of Fredhoim operators is also a Fredhoim operator and

the index of the composition is the sum of the indices of the operators. This
is known as the Index Theorem.
Theorem 4.43 (The Index Theorem). For a scheme X Y Z of
bounded operators between Banach spaces we have the following.
(1) IfS and T are both Fredholm operators, then ST is also a Fredhoim
operator and i(ST) = i(S) + i(T).
(2) If ST is a Fredhoim operator, then T is a Fredhoim operator if and
only if S is a Fredhoim operator.
Proof. The proof is "combinatorial" in nature and is based upon the fol-
lowing simple algebraic fact:
o If W isa vector subspace of a vector space V, then there exists an
algebraic complement of W, i.e., there exists some vector subspace
U of V such that W e U = V.
(For a proof see Exercise 3 at the end of the section.)
The key object for our proof is the subspace N1 R(T) fl N(S) of Y.
Pick a subspace N2 of Y such that N(S) = N1 N2. So, we have
n(S) = dimN1 +dimN2. (*)
Next, choose a vector subspace Y1 of Y such that R(T) = Y1 N1. Since

N2 fl Y1 = {O} (why?), there exists another vector subspace Y2 of Y such
that Y = Y1 e N1 e N2 e Y2. In particular, we have
d(T) = dim N2 + dim Y2. (**)
Since N(T) N(ST), there exists a vector subspace X1 of X such that
N(T) e = N(ST) T'(N(S)) = n R(T)) = T'(N1).

We claim that the operator T: X1 is a linear isomorphism.
N1
Clearly, T is one-to-one. On the other hand, if y = Tx E N1 N(S), then
STx=OorxEN(ST),andsowecanwritex=y+zwithyEN(T)and
z in X1. Thus, y = Tx = Tz T(X1), and this proves that T: X1
E N1 is
surjectiveand hence it is a linear isomorphism. From N(T)eX1 = N(ST),
we see that n(T) + dimX1 = rt(ST) or
n(ST) = n(T) + dim N1. (***)
Next, observe that S is one-to-one on Y1 Y2. This implies
R(S) = S(Y1 = 5(Yi) e S(Y2) = S(Y1 N1) S(Y2)
= S(R(T)) = R(ST)
4.4. Fredholrn Operators 159
Consequently, if Z1 is a subspace of Z satisfying Z = Z1 e R(S), then

Z = e R(ST) e 5(Y2). So, d(ST) = dim Z1 + dim S(Y2) = dim Z1 + dim 1'2
d(ST) = d(S) + dimY2. (f)

With this setup in place we are now ready to prove the theorem.
(1) Assume that S and T are both Fredhoim operators. This implies that
all subspaces appearing in (*), (*-k), (* * *), and are finite dimensional.
In particular, ST is a Fredholm operator. Using (*-*) in connection with (*)
and (***), we get
dimY2 = d(T) d(T) n(S)
= d(T) - n(S) + [n(ST) - n(T)]
=
Now a glance at yields
d(ST) = d(S) + dimY2 = d(S) + rt(ST) ri(S) i(T)
or rt(ST) d(ST) = i(T)
+ ri(S) d(S). That is. i(5T) i(T) + i(S).
Assume first that ST and S are Fredholm operators. From the

(2)
inclusion N(T) c N(ST), we see that rt(T) < oo. Also, from (*), we get
dim Y2 <oo, and (*-*) yields d(T) <oc. Thus, T is a Fredhoim operator.
Finally, suppose that ST and T are Fredhoim operators. Then (*-*)
yields that dim N2 <oo and dim Y2 <oo, and (* * *) implies dim N1 <oo.
But then from (*) and we obtain ri(S) <oo and d(S) <oo. That is, S
is a Fredholm operator, and the proof is finished.
In order to obtain more properties of Fredhoim operators, we need the
following useful lemma on the existence of "almost perpendicular" vectors.
Lemma 4.44 (F. Riesz). If Y is a proper closed subspace of a Banach space
X, then for each 0 < 1 there exists some unit vector x E X such that
IIyxit1 holds for allyEY.

Moreover, if Y is finite dimensional, then there exists a unit vector x
in X such that the distance from x to Y is one, i.e., d(x,Y) = 1.
Proof. By the HahnBanach theorem there is a linear functional E
with = I satisfying x*(y) = 0 for all y E Y. Pick a unit vector x E X
such that Ix*(x)1 I . Now note that if y E Y, then
=
For the last part, assume that Y is finite dimensional. Pick another finite
dimensional vector subspace Z of X such that Y is a proper subspace of Z.
By the preceding case, for each i-i. there exists some unit vector E Z such
that I Since the unit sphere of Z is compact, by passing

to a subsequence, we can assume that x Z. Then lxii = 1 and the
continuity of the distance function z ' d(z, Y) yields
I d(x,Y) = lim lim (1
fl+OO
= 1.
This implies d(x,Y) = 1. m
Wecontinue with an important property regarding compact perturba-

tions of the identity operator.
Lemma 4.45. If K: X X is a compact operator, then the operator IK:
(1) has finite ascent and descent, and
(2) is a Fredhoim operator of index zero.
Proof. Let K: X X be a compact operator on a Banach space. We first

claim that the vector space N(I K) is finite dimensional. To see this,
notice that if x E N(I K), then Kx = x. That is, on N(I K) the
compact operator K coincides with the identity operator. This implies that
N(I K) is finite dimensional.
Next, we shall show that I K has a closed range. To see this, pick
a closed complement V of N(I K), i.e., X = N(I K) V. (Since
N(I K) is finite dimensional, closed complements of N(I K) always
exist; see Exercise 6 of Section 4J We claim that I K is bounded below
on V. If this is done, then the operator I K: V R(I K) must be an
isomorphism, and so R(I K) must be a closed subspace of X.
To establish that I K is bounded below on V, assume contrary to
our claim that there exists a sequence of unit vectors V such that
K is compact, by passing to a subsequence we can
suppose that v v
v and v
V vE the latter conclusion
contradicts the fact N(I K) fl V = {0}. Thus, I K must be bounded
below on V.
(1) We shall prove only that K) <oc. The claim 8(1 K) <oc
can be proven in a similar manner. We already know that the vector space
N(I K) is finite dimensional. Since (I can be written as I K1
for some K1 E IC(X), the above conclusion shows that N((I is also
finite dimensional for each ii.
Now assume by way of contradiction that K) = oo. This means
that N ((I K)12) is a proper subspace of N((I for each ii. Conse-
quently, by Riesz's Lemma 4.44, there exists a sequence of unit vectors
in X such that E N((I for each ii and xli I for all
x E N((I K)Th). Since each subspace N((I is invariant under

I K, it follows that Xm + (I (I E N((I for each
1 rn <n. Consequently, we have
= Ixn [Xm + (I (I K)Xm] I
for all n-i and n. This implies that the sequence does not have any
convergent subsequence, contradicting the compactness of K. This contra-
diction establishes that I K has a finite ascent.
(2) From the preceding conclusions we already know that N(I K*) is
finite dimensional: -Therefore,- by Theorem 2.13 we have-
d(I K) = dim [X/R(I K)] = dimN(I K*) <00.
K is a Fredholm operator.
This implies that I
To show that the index of I K equals zero, we proceed as follows.
By part (1) the operator I K has finite ascent and descentwhich by
Lemma 2.21 coincide. Let p = K) = 5(1 K). Then Theorem 2.23
guarantees that X = N((1 K)P) e R((1 K)P). This identity implies that
the Fredhoim operator satisfies
i((1 K)P) = dimN((1 K)P) dimN((1 K)P) = 0.
Now applying the Index Theorem, we obtain that i((1K)P) = p[i(1K)],
and so i(1 K) = 0. The proof is finished.
The next result, due to F. V. Atkinson [44], characterizes the Fredhoim

operators.
Theorem 4.46 (Atkinson). For a bo'unded operator T: X Y between
Banach spaces the following statements are eq'uivalent.
(1) T is a Fredholm operator.
(2) There exist an operator S E L(Y, X) and two finite-rank projections
P1 EL(X) and P2 EL(Y) sothatST=IxPi and
(3) There exist compact operators K1 E L(X) and K2 E L(Y) and an
operator SE L(Y,X) s'uch that ST = Ix K1 andTS = ly K2.
(4) There exist compact operators K1 E L(X) and K2 E 1C(Y) and
51, S2 E L(Y, X) s'uch that S1T = K1 and T52 = ly K2.
Proof. (1) Assume that T is a Fredholm operator. Then N(T) <00

(2)
and there exist a closed subspace V of X and a finite dimensional subspace
W of Y such that X = N(T) V and Y = W R(T). Also, let PN be the
projection of X onto N(T) along V and let be the projection of Y onto
W along R(T). Clearly, and are bounded finite-rank projections.
Now let Si = TIv: V R(T) = T(V), and note that is continuous.

Since the range R(T) is closed, it follows that the operator Sj': T(V) V,
defined by Sj' (Tv) = v, is also continuous. Next, consider the operator
S Sj4(IyPw) E (Y,X), and note that ify w+Tv E =Y
with v E V, then
Sy {Sj'(Iy Sj'(Tv) = v.
+Tv) =
This implies that for each x = m + v E N(T) e V = X we have
STx = BT(m+v) = Sj'Tv = v = (Ix PN)(Th+V) = (Ix PN)X.
Similarly, for each y = w + Tv E TV T(V) = Y with v E V we have
TSy=Tv=(IyPw)y.
So, ST = and TS = ly and the validity of (2) has been
established.
(2) (3) (4) These implications are obvious.

(4)(1) Assume that S2 E (Y, X) and the compact operators
K1 E (X) and K2 E IC(Y) satisfy T 1x K1 and TB2 = K2. ly
K1 is a compact operator, it follows from Lemma 4.45 that
Since
N(51T) = K1) is finite dimensional. From N(T) N(51T), we
see that dim N(T) <co. The same argument applied to =
shows that N (T*) is also finite dimensional. From Theorem 2.13, it fol-
lows that if R(T) is closed, then dim [X/R(T)] = dim N(T*) < oo. So, to
complete the proof, it remains to be shown that T has a closed range. -
To this end, note that Lemma 4.45 implies that ly K2 is a Fredholrn
operator. So, there exists a finite dimensional subspace Z of Y such that
R(Iy K2) Z = Y. Now from R(Iy K2) = R(T52) R(T), it follows
that R(T) + Z = Y. This implies that there exists a vector subspace Z1 of Z
(and hence Zi is finite dimensional) such = see that R(T)
Exercise 3
at the end of the section. Now a glance at Theorem 2.16 guarantees that T
has a closed range, and the proof is finished.
Corollary 4.47. If T E (X, Y) is a Fredhoirn operator, then for each
K E /C(X,Y) the operator T + K is also Fredholrn and i(T + K) = i(T).
Proof. Let T E (X, Y) be a Fredholm operator and fix some K E 1C(X, Y).
By Theorem 4.46 there exist an operator S E (Y, X) and two compact
operators K1 E (X) and K2 E (Y) such that 51T = 1x K1 and
T52 = ly K2. It follows that
51(T + K) = Ix - (K1 51K) and (T + K)52 = ly (K2 KS2).
Since the operators K1 51K and K2 KS2 are compact, another appli-
cation of Theorem 4.46 shows that T + K is a Fredholm operator.
4.4. Fredhoim Operators 163
From the Index Theorem we know that is a Fredhoirn operator, and

Lemma 4.45 guarantees that i(S1) + i(T) = i(81T) = K1) = 0. Thus,
i(T) = i(Si). Similarly, i(T + K) = i(Si), and so i(T + K) = i(T).
The perturbation of an operator T by an operator S is T S. In

this terminology, Corollary 4.47 states that: A compact perturbation of a
Fredhoirn operator remains Fredhoim and does not change its index.
If Y) denotes the collection of all Fredhoim operators from X to
Y (where, as usual, we write instead Of X)), then the index
function i: Jred(X, Y) Z is defined by i(T) for each T E Jred(X, Y).
It turns out that the index function is continuous when Z is equipped with
its discrete topology. The details are included in the next important result.
Theorem 4.48. If X and Y are Banach spaces, then Jred(X, Y) is an open
subset of (X, Y) and the index function i: Jred(X, Y) Z is continuous.
in particular, Jred(X) is an open unital subsemigroup of L(X) and the
index function 'i: Z is a continuous algebraic homomorphism.
Proof. If X is finite dimensional, then Jred(X, Y) = 0 provided Y is in-

finite dimensional and Jred(X, Y) = L(X, Y) provided Y is finite dimen-
sional. So, we can assume that T: X Y is a Fredholm operator between
two Banach spaces with X infinite dimensional. By Theorem 4.46 there
exist an operator S E (Y, X) and two compact operators K1 E IC(X) and
K2 E 1C(Y) such that
and TS=IyKi. (*)
Since X is infinite dimensional, it follows that S 0. By Lemma 4.45 the

operator 1x K1 E L(X) is a Fredholm operator of index zero. So, by the
Index Theorem 4.43 the operator S is also a Fredhoim operator, and fronr
i(S) + i(T) i(ST) = K1) = 0 we get
i(T) = -i(S). (**)
We know that if A E and B E (Y) satisfy ItIx All < 1 and

IIy <1, then 1x A and ly B are invertible; and hence
they are
both isomorphisms. Let C E (X, Y) be any operator such that
Then the operators SC E and CS E (Y) satisfy ISCII < 1 and
ICSII < 1. So, the operators 1x + SC E L(X) and ly + CS E L(Y) are
isomorphisms, and consequently they are both Fredholm operators of index
zero. From Corollary 4.47 it follows that Ix + SC K1 and ly + CS K2
are likewise Fredholm operators of index zero. Moreover, from (*) we get
S(T+C) 1x+SC-Ki and (T+C)S= Iy+CS-K2.
Now a glance at the Index Theorem 4.43(2) shows that the operator T + C
is Fredhoim and satisfies
i(S)+i(T+ C) = i(S(T+C)) = + SCK1) = 0.
Thus, i(T + C) = i(S), and from (**) we get i(T + C) = i(T).
Therefore, we have established that if C E C(X, Y) satisfies MCI! < iih'
then T + C is a Fredhoim operator and i(T + C) = i(T). This shows
that Jred(X, Y) is an open subset of L(X, Y) and that the index function
i: Jred(X, Y) Z is continuous. The proof is finished.
There is a similar result regarding semi-Fredholm operators. To establish
it, we need some preliminary discussion.
Lemma 4.49. For a Banach space X we have the following.
(1) If there exist x1, , xE

such that < 1, then X is finite dimensional.
(2) Let Z be a closed vector subspace of X. Assume that there exist

vectors Ui, . . , E such that for each x E we can find
some satisfying d(x Z) < 1. Then the quotient vector space
X/Z is finite dimensional.
Proof. (1) Let Y be the linear span of x1, x2, ... , We claim that X = Y.
If X Y, then by Riesz's Lemma 4.44 there exists some unit vector x E X
such that d(x,Y) = 1. In particular, we have Ix xjJJ d(x,Y) = 1 for
each i, a contradiction. Hence, X = Y.

(2) Let ir be the quotient mapping from X to X/Z, and consider the vectors
ir(ui), ir(u2),. . in X/Z. This is a finite collection in the closed unit
,
ball of the Banach space X/Z. From our assumption, it follows that for each
xE there exists some i such that jJ7r(x) < 1. By part (1), the
vector space X/Z is finite dimensional.
The next result deals with the codimension of an operator with a closed
range.
Lemma 4.50. For a bounded operator T: X Y between Banach spaces
with a closed range we have the following.
(1) If d(T) = 0 and N(T) is finite dimensional, then every operator
in C(X, Y) close to T also has a closed range, infinite defect, and
a finite dimensional null space.
(2) If n(T) = cc and R(T) has a finite codimension, then every oper-
ator in C(X, Y) close to T also has a closed range, infinite nullity,
and a finite codimension.
Fredhoim Operators 165
Proof. (1) The fact that for each operator S E (X, Y) close enough to T
the range R(S) is closed has been established in Corollary 2.6. The only
thing we need to verify is that d(S) = cc and dim N(S) < oo for all operators
S E (X, Y) close to T.
We assume first that T is one-to-oneS In this case, since R(T) is closed,
for some > 0 and all x E X. Pick some f E (0, 'y) such that
<1. We claim that every operator S E (X, Y) with IS <f has
a closed range, an infinite defect, and a finite dimensional null space. To
establish this, fix some S E (X, Y) satisfying HS <
From IISxII ('y that S is bounded
6)fIxfI for each x E X, we see
below on X. This implies that N(S) = {0} and that S has a closed range.
Moreover, for each y E R(S) the vector x = 5'y satisfies the estimate
(*)
To establish that R(S) has infinite codimension. assume contrary to our

claim that dimY/R(S) < 00. That is, let the Banach space Y/R(S) be
finite dimensional. Fix some S with 0 < S < 1 and then choose
vectors ui, ,. in the open unit ball of Y such that for each u E
there exists some satisfying 117r(u) < 8, where ir is the quotient
map of Y onto Y/R(S).
We claim that u2,... , satisfy the conditions of Lemma
with Z = R(T). To see this, fix any y E and consider the vector 7r(y).
Pick some i such that I < S. But then, by the definition of the
quotient norm, we can find a vector rE R(S) such that < S.
Clearly, and so IIrII
Since r E R(S), there exists some x E X with Sx = r. From (*), we get
< Now consider the element z = Tx E R(T). Clearly,
rz=Sx---Tx,andso
llr-zH = ISx-TxH =
Keeping in mind that <5, we obtain that
Therefore, the collection of vectors Ui, U2, satisfies the condition of

. . . ,
Lemma 4.49(2). Consequently, the vector space Y/R(T) must be finite di-
mensional, which is impossible.
For remains to notice that the above assumption
the general case, it
that T is one-to-one is not essential. Indeed, the null space N(T) is, by
hypothesis, finite dimensional and so it is complemented in X. Let V be a
closed vector subspace of X such that X = N(T) V. Then the restriction
of T to V is one-to-one, and consequently the previous case applies to this
Hence, if S e
restriction. Y) is close to T, then 8(V) has an infinite
codimension in Y. Since 5(X) = S(N(T)) + 5(V) and S(N(T)) is finite
dimensional, it follows that 5(X) also has an infinite dddimension in Y.
(2) Assume that a bounded operator T: X Y between Banach spaces
has a closed range and satisfies dim N(T) = oo and dim Y/R(T) <oo. Con-
sider the adjoint operator T*: X* and note that from Theorem 2.13
we have dim N(T*) = dim {Y/R(T)]* < oo. Moreover, since R(T) is closed,
the identity X*/R(T*) = [N(T)J* holds; see Exercise 12 in Section 2.1.
From this identit3r it follows that dim X*/R(T*) 00. By the preceding
case, there exists some E> 0 such that every operator A E (Y*, X*) with
All has a closed range, dim N(A) <oo and dim X*/R(A) = Do.
Now let an operator S e (X, Y) satisfy liT < Then we have
LIT* < This implies that S* has a closed range (and hence S
likewise has a closed range), dim N(S*) < Do and dim X*/R(S*) =
From dim [Y/R(S)]* = dimN(S*) < Do, we infer that dimY/R(S) < Do.
Also, from X*/R(S*) = [N(S)]*, it follows that dim N(S) = Do. The proof
is finished.
Let us denote by Fred(X, Y) the collection of all serni-Fredholm oper-

ators between the Banach spaces X and Y. That is, Fred(X, Y) consists
of all bounded operators from X to Y for which the index is a well-defined
extended integer. Clearly, we have the inclusion Fred(X, Y) Fred(X, Y).
Combining Theorem 4.48 and Lemma 4.49, we obtain the following gener-
alization of Theorem 4.48. -
Theorem 4.51. If X and Y are Banach spaces, then Fred(X, Y) is an open

subset of L(X,Y) and the index function
i: Fred(X,Y) ZU{oo,oo}
is continuous. Moreover, Fred(X,Y) is an open subset of Fred(X,Y).
We continue with a useful result regarding invertibilit3r of operators.

Lemma 4.52. For a bounded operator T: X + Y between two Banach
spaces the following statements are equivalent.
(1) There exist continuous projections P e (X) and Q E such

that R(P) = N(T) and R(Q) = R(T).
(2) There exist closed subspaces V and W of X and Y, respectively,
such thatX=N(T)eV andY=R(T)eW.
(3) There exists a bounded operator 5: Y X such that TST T.
Proof. (1) (2)In this case, the closed vector subspaces V = (Ix P)(X)
and TV = (ly Q) (Y) satisf3r the desired properties.
(2) (3) Assume that the closed subspaces V and W of X and Y,

respectively, satisfy X = N(T) V and Y = R(T) W. From Theorem 2.16,
we know that R(T) is a closed subspace of Y. Consequently, the operator
A TIv: V R(T) is an isomorphism. Next, pick a projection Q e (Y)
with range R(T) and consider the operator S = A'Q: Y + X. Clearly,
S e (Y, X). Now, if x = u + v e N(T) V = X, then
TSTx = TA1 QTx = TA' Tx = TA 1Tv = Tv = Tx,
and so TST = T holds.
(3) Assume that TST = T for some operator S e (Y, X). Then
(1)
(ST)2 = S(TST) = ST, and similarly (TS)2 = TS. That is, the operators
Pi = ST e (X) and Q = TS e (Y) are projections on X and Y,
respectively. We claim that N(P1) = N(T) and R(Q) = R(T).
For the first identity, note that N(T) c N(ST) and that x e N(ST)
implies Tx = T(STx) = 0, i.e., N(ST) N(T) is also true. For the second
identity, observe that the inclusion R(TS) R(T) is obvious and that for
each x E X we have Tx = TS(Tx) e R(TS). If we let P = P1, then P
is a bounded projection on X and satisfies P(X) = R(P) = N(T).
The operators S that satisfy statementof(3)

the preceding lemma are
referred to as pseudoinverses or as relative inverses of T.
Definition 4.53. A pseudoinverse of a bounded operator T: X Y

between Banach spaces is any operator S e (Y, X) satisfying TST = T.
Fredholm operators always have pseudoinverses that are also Fredholm

operators.
Theorem 4.54. Every Fredhoim operator T: X Y between Banach

spaces has a pseudoinverse. Moreover, every pseudoinverse 5: Y + X of T
is also a Fredholm operator and satisfies i(S) = i(T).
Proof. The first part follows immediately from Lemma 4.52. For the second
part, let S: Y + X be a pseudoinverse of T, Le., TST = T. Then, by part
(2) of Theorem 4.43, the operator ST is Fredholm. But then, by the same
argument, S is a operator. Now part (1) of Theorem 4.43 yields
i(T) = i(TST)= i(T) + i(S) + i(T). This implies i(S) = i(T).
we shall close the section with another useful property of Fred-

Finally,
hoim operators.
Lemma 4.55. For any Fredholm operator T E (X) we have the following.
(1) The vector subspace Y = R(T72) is a T-invariant closed sub.
space of X satisfying T(Y) = Y.
(2) If A 0, then N(A T) = N((A T)Jy) = N(A Tjy).
(3) There exists some > 0 such that:
(a) For each A T is Fredhoim.
(b) The nullity n(A T) and the defect d(A T) are constant on
the set {A e C: 0 < AJ <}.
Proof. (a) Let T: X p X be a Fredhoirn operator. A glance at the Index

Theorem 4.43 shows that T72 is a'so Fredhoirn for each n. In particular, by
Lemma 4.38, the range is closed for each n. This implies that Y is a
closed subspace of X. Clearly, Y is T-invariant.
To see that T(Y) = Y, observe first that there exists some k such that
R(T12) n N(T) = n N(T) holds for all n k. This should be obvious
in view of the fact that N(T) is finite dimensional and C R(T72)
for each n. Now fix some y e Y and let ri k. Pick some x e and
some z E such that y = Tx = Tz. Therefore, T(x z) = 0
and so x z e fl N(T) R(T12). This implies x e R(T71) for each
n k. Consequently, x e Y and y = Tx. This proves that T(Y) = Y.
(b) Fix a non-zero scalar A. Clearly, N(A Tly) N(A T). To see
that N(A T) N(A is also true, let x e N(A T). This implies
Tx = Ax, and so T12x = for each n. Consequently, x = E
for each n. Thus, x e Y and so x e N(A
According to Theorems 2.10 and
(c) there exists some > 0 such
that for each Al < :
o The operator A T E (X) is a Fredholrn operator and its index
satisfies i(A T) = i(T) = i(T).
The operator (A T)fy = A T(y e (Y) is surjective and its
index satisfies i(A Tjy) =
Now fi1x any A e C with 0 < <. Since the operator A e 12(Y) is
surjective, it follows that n(A = i(A Tjy). Now (b) yields
n(AT) =
=
Thus, the function A 'p n(AT) is constant on the set {A e C: 0 < Al <}.
Since the function A i(A T) = n(A T) d(A T) is also constant on the
set {A e C: 0 < <}, follows that the defect function A 'p d(A T)
is also constant on {A e C: 0 < Al < }. The proof is finished.
Exercises
L Prove Lemma 4.40. [HINT: If T: X + Y is a bounded operator between
two real Banach spaces, then note that = =
and R(T) =
2. Give a direct proof of Lemma 4.44 without using the HahnBanach
theorem. Also, present an example of an infinite dimensional proper
closed subspace Y of a Banach space X such that for each unit vector
x E X \ Y we have d(x, Y) < 1. [HINT: FLx some x0 E X \ Y and let
d = d(xo, Y) > 0. Next, select some Yo E Y with Mxo poll < and let
= Clearly, I and for any y E Y we have
> i E.
d
x
y IIxo-yoII >
IIxo-voII IIXo-yolI
3. This exercise presents some elementary algebraic properties that have

been used in the proofs of the results in this section. Establish the fol-
lowing.
(a) If W is a vector subspace of a vector space V, then W has an
algebraic complement, i.e., there exists a vector subspace U of V
satisfying WflV={0} and W+U=V.
(b) If U and T'V are subspaces of a vector space V, then there is a vector
subspace W1 of W satisfying V fl W1 = {0} and U + W1 = U + W.
(c) If T: V W is a linear operator between two vector spaces, then
the quotient vector space V/N(T) is linearly isomorphic to R(T).
[HINT: For (a) and (b) use the fact that any collection of linearly inde-
pendent vectors can be extended to a Hamel basis. For (c) consider the
linear operator T: V/N(T) R(T) defined by = Tx.]
4. be an rn-dimensional vector
Let X be an n-dimensional vector space, let Y
space, and let T: X Y be an operator. If r = dim R(T), then show
that n(T) = nr and d(T) = mr; conclude from this that i(T) = nrn.
[HINT: Use Exercise 9 in Section 8.2.]
5. For a bounded operator T: X * Y between two Banach spaces define
T: X/N(T) * Y via the formula = Tx. Let iv: X X/N(T) be
the quotient map and J: R(T) Y be the natural embedding. Now
consider the scheme of bounded operators
X X/N(T) R(T) Y,
and note that T = JT7r. If T is a Fredhoim operator, then show that all
operators in the above scheme are Fredholm operators and compute their
indices. Also, verify directly that i(T) = i(J) + i(T) + i(7r).
6. Establish the following property that was left out from the proof of
Lemma 4.45: If K: X * X is a compact operator on a Banach space,
then the operator I K has a finite descent.
7. (The Fredholm Alternative) Let X be a Banach space and K E K(X).
Show that for each A 0 exactly one of the following two exclusive
statements is true:
(a) For each y E X the equation x AKx = y has a unique solution.

(b) The equation. x AKx = 0 has a non-zero solution.
When (b) is true, show also that the equation x AKx = 0 has a finite
number of linearly independent solutions.
8. Let T E (X, Y) and let S E (Y, X) be a pseudoinverse of T. Fix some
y E Y and consider the equation
Tx=y. (*)
Show that equation (*) has a solution if and only if the vector Sy is a
solution of (*).
9. Show that a bounded operator between Hilbert spaces has a pseudoinverse
if and only if it has a closed range.
4.5. Strictly Singular Operators

In this section we shall discuss the properties of the class of strictly singular
operators that are related to compact and operators. They were
introduced by T. Kato [178].
Definition 4.56 (Kato). A bounded operator T: X * Y between Banach
spaces is said to be strictly singular if there is no infinite dimensional
closed subspace Z of X such that T: Z p T(Z), the restriction of T to Z,
is an isomorphism.
Clearly, the concept of a strictly singular operator is of interest only

when the domain space is infinite dimensional. Notice also that a bounded
operator T: X * Y between Banach spaces is strictly singular if and only
if T is not bounded below on any infinite dimensional closed subspace of X.
The next result presents some connections between compact operators and
strictly singular operators.
Lemma 4.57. Let T: X Y be an isomorphism between two Banach
spaces. If the restriction of T to a closed subspace Z of X is compact, then
Z is finite dimensional. In particular, every compact operator is strictly
singular.
Proof. Let Z be a closed subspace of X such that T: Z T(Z) is compact.

Since T: X Y is an isomorphism, it follows that T: Z T(Z) is likewise
an isomorphism. If S = (TJz)', then I = TS: Z Z is a compact
operator. This implies that Z is finite dimensional.
A strictly singular operator need not be a compact operator.

Theorem 4.58. If 1 p <r < oc, then the natural embedding J: 4
is a non-compact strictly singular operator.
Proof. To establish this result we need the following simple property whose
proof is left as an exercise.
(P) If X is an infinite dimensional vector subspace of 4, then for each
m the vector subspace
Xm={xEX:
X is also infinite dimensional.
We denote by the standard projection of onto the first n coordinates
of 4 That is, = (yi,. .. , ye We also let
If denotes the basic unit vector in then 1 for each

n and lien = 21/r for n m. This shows that J is not a compact
operator. To verify that the embedding J: 4 * r is strictly singular, it
suffices to show that for each infinite dimensional closed subspace X of
there exists a sequence X satisfying 76 0 and 0.
To this end, let X be an infinite dimensional closed subspace of 4. Fix
some 0 < E < 1 and let 8 = Start by choosing any vector u1 e X
with = 1, and then pick some m1 such that > 1 8 and
< Let Vi = Pm1U1 and w1 = Qm1U1. By Property (P), the
vector subspace Xm1 of X is infinite dimensional. Pick some U2 E Xmi
with 1u2 = 1 and then select m2 > m1 such that 1Pm2 > 1 8 and
IIQrn2U2Mr <
An inductive argument along these lines guarantees the existence of a
sequence {uk} X of unit vectors and a strictly increasing sequence of
natural numbers {mk} such that for each k we have Uk e (where
mU = 0 and Xo X) and the vectors vk PmktLk and wk QmkUk
satisfy:
(a) lIVkIIp> 18, and
(b) <
Clearly, {vk} is a disjoint sequence and Uk = vk + wk for each k.
Now let = E X. We claim that for each n
and that 0. Indeed, using the p-additivity of the norm, we get
= + -
-
>
On the other hand, the r-additivity of the norm yields
= +
nu/pTh
This is the desired conclusion.

It is interesting to note that all bounded operators acting in the opposite
direction, that is, from 4 to where 1 p < r oo, are compact and
50 L(6, p) = This is a well-known theorem due to H. R. Pitt;
for a proof see [205, Proposition 2.c.3]. For a complete characterization of
the pairs (p,), (v)) for which (p,), (v)) = "P2 (v))
see [89, Theorem 4.7]. o
It should be mentioned that the adjoint of a strictly singular operator

need not be strictly singular. For an example see Exercise 6 at the end of
the section.
Many properties of strictly singular operators depend on the following
technical result. -
Lemma 4.59. For a bounded operator T: X + Y between two Banach

spaces with X infinite dimensional we have the following.
(1) If the restriction of T to any closed subspace of finite codimension is
not an isomorphism, then for any closed subspace Xi of X of finite
codimension and each c > 0 there exists an infinite dimensional
closed subspace X2 of X1 such that T: X2 + Y is compact and
< .
(2) IfT is strictly singular, then for each infinite dimensional closed
subs pace X1 of X there exists an infinite dimensional closed sub-
space X2 of X1 such that T: X2 Y is compact and IfTIx2fJ
Proof. (1) Assume that X1 is a closed subspace of X of finite codimen-

sion and so the restriction of the operator T E (X, Y) to Xi is not an
isomorphism. Replacing X by X1, we can assume without loss of generality
that X1 = X. We claim that in this case the following version of Mazur's
Lemma 1.42 is valid:
If Z is a finite dimensional vector subspace of X, then for each

E> 0 and for each 8 > 0 there exists a unit vector x E X such that
JTxlI <8 and (1 + E)Mz + for each z E Z and each
The proof of this claim is similar to that of Lemma 1.42. Following that
proof, we notice that the closed subspace Ker has finite codimension,
and so (in view of our assumptions) the operator T cannot be bounded below
on this Therefore, we can choose a unit vector x E Ker
satisfying IITxlI
Using the above version of Mazur's lemma and arguing as in the proof
of Theorem we see that for each E > 0 there exists a normalized basic
sequence in X with basis constant less than 2 such that <*
for each n. Let X2 be the infinite dimensional closed subspace generated by
We claim that X2 satisfies the desired properties.
First, we show that < To see this, let x = E X2
satisfy 1. Then = 4 for each n; see Corollary L40.
This implies
MTxII = =
and so 11Th = IITxM <

To see that the restriction S = T[x2: X2 Y is a compact operator, let
X2 X2 be the natural projections on the first n coordinates associated
with the basis Since the operator X2 Y has finite-rank, it
is automatically compact. Now notice that if x = E X2 satisfies
lxii < 1, then
= MS(1x2 = =
i=n+1 i=n-f-1
in+1 i=n+1
Therefore, and so S
S= is a compact operator, and the proof of the first part is finished.
(2) Assume that T is a strictly singular operator. Then the restriction
of T to any infinite dimensional subspace of X is not an isomorphism, and
under this condition all arguments of the above proof remain valid for each
closed infinite dimensional subspace X1 of X.
174 - 4. Special Classes of Operators
Corollary 4.60. If a bounded operator T : X Y between two Banach

spaces does not have a closed range, then for each closed subspace X1 of X
of finite codimension and for each 0 there exists an infinite dimensional
closed subspace X2 of Xi such that T: X2 Y is compact and lITIx2Il <
Proof. It is enough to verify that T satisfies the hypotheses of Lemma 4.59,

i.e., that the restriction of T to any closed subspace of X of finite codimen-
sion is not an isomorphism. To see this, assume by way of contradiction
that V and W are closed subspaces such that X = V e W, dim W < 00,
and T: V T(V) is an isomorphism. So, T(V) is a closed subspace of Y
and = T(V) + T(W). Since T(W) is finite dimensional, it follows that
R(T) is a closed which is a contradiction; see Corollary 2.17.
We are now ready to present T. Kat&s [178] characterization of strictly

singular operators.
Theorem 4.61 (Kato). For a bounded operator T: X Y between two Ba-

nach spaces with X infinite dimensional the following statements are equiv-
alent:
(1) The operator T is strictly singular.

(2) For each infinite dimensional closed subspace X1 of X and each
0 there is an infinite dimensional closed subspace X2 of X1
such that the operator T: X2 Y is compact and JIT[x21( <c.
(3) For each infinite dimensional closed subspace X1 of X there ex-
ists an infinite dimensional closed subspace X2 of X1 such that the
operator T: X2 f Y is compact.
Proof. The implication (1) (2) follows immediately from part (2) of
Lemma 4.59. The implication (2) (3) is obvious.
(3) (1)Assume that the operator T is not strictly singular. This

means that there exists an infinite dimensional closed subspace X1 of X
such that T: X1 f T(X1) is an isomorphism. Then, by (3), there ex-
ists an infinite dimensional closed subspace X2 of X1 such that the oper-
ator T: X2 Y is compact. However, the latter conclusion contradicts
Lemma 4.57 according to which X2 must be finite dimensional. This estab-
lishes that T is strictly singular.
A consequence of the preceding theorem is the following result that de-

scribes two multiplicative properties of singular operators.
Corollary 4.62. The collection of all strictly singular operators from a

Banach space X to another Banach space Y is a closed subspace ofL(X, Y).
Moreover, if in a scheme of bounded operators X Y Z between
Banach spaces either S or T is strictly singular, then TS is strictly singular.
Proof. Let 8, T E (X, Y) be two strictly singular operators. It should be
clear that is also strictly singular for each scalar What needs a proof
is the fact that S + T is strictly singular.
To see this, let Z be an infinite dimensional closed vector subspace of
X. Since S is strictly singular, there exists (by Theorem 4.61) an infinite di-
mensional closed vector subspace Zi of Zsuch that 5: Z1 Y is
operator. Using Theorem 4.61 once more, we see that there exists an infinite
dimensional closed subspace Z2 of Z1 such that T: Z2 Y is a compact
operator. It follows that S + T: Z2 Y is a compact operator. Now a
glance at Lemma 4.57 guarantees that (S + T)Iz cannot be an isomorphism.
Therefore, the operator S + T is strictly singular.
To see that the vector subspace of all stricfly singular operators from X
to Y is closed, assume that a sequence of strictly singular operators
from X to Y satisfies T in L(X, Y). Also, let Z be an arbitrary
infinite dimensional closed subspace of X. If T is bounded below on Z, then
by Lemma 2.4 there exists some such that is bounded below on Z for
all n flu. But this contradicts the fact that no
T bounded below on T is strictly singular.
Now consider a scheme of bounded operators X Y Z between
Banach spaces. Assume that TS is an isomorphism on an infinite dimen-
sional closed vector subspace V of X. This guarantees that S is bounded
below. In particular, S cannot be strictly singular and 5: V 8(V) is an
isomorphism. Using again the fact that T: S(V) TS(V) is an isomor-
phism, we infer that T: S(V) Y must be bounded below. This shows
that T is not a strictly singular operator either. Thus, we have shown that
if ST is not strictly singular, then neither S nor T is strictly singular, and
the desired conclusion follows.
As in the case of compact perturbations of Fredholm operators, a strictly

singular perturbation of a Fredholm operator remains Fredholm and has the
same index.
Theorem 4.63. Let 5, T: X Y be two bounded operators between Ba-
nach spaces with X infinite dimensional. If T is Fredhoim and S is strictly
singular, then T + S is a Fredholm operator and i(T + 5) = i(T).
Proof. Since T is a Fredholm operator, there exist closed subspaces V of
X and W of Y such that V has a finite defect, W is finite dimensional,
4. Special Classes of Operators
X = N(T) V and Y = R(T) W. In particular, the surjective operator

T: V + R(T) is an isomorphism. Pick some c> 0 such that
IITvI[ 2cJJvIf (*)
holds for each v E V.

To see that dim N(T + 5) < 00, assume by way of contradiction that
dim N(T + 5) = 00. Consequently dim {N(T + 5) fl V] oo, since V has
a finite defect. Now, for each x in the infinite dimensional closed subspace
Xi N(T+S)nV of X we have I1Sx1l = 11TxII This implies that
S is invertible on X1, which is a contradiction. Hence, dim N(T + 5) <00.
Next, we claim that the range of T + S is closed. To establish this,
assume again by way of contradiction that R(T + 5) is not closed. Then,
by Corollary 4.60, there exists a closed infinite dimensional subspace Z of V
such that (T + S)Iz <c. Therefore, using (*), we see that for each z E Z
we have
ISzIl = * (T + S)zII 11TzI1 - 1I(T + S)zIl 2cMzII - cIizJI =
This shows that S restricted to Z is an isomorphism, which is a contradic-

tion. Hence, R(T + 8) is closed, and so T + S is a semi-Fredholm operator.
Since .AS is a strictly singular operator for each 0, it follows from
the last conclusion that T + ,AS is also a operator for each
E JR. Finally, in view of Theorem 4.51, the mapping i(T + from JR
to Z U {oo, oo} is continuous, and hence it must be constant. This implies
that i(T) for each ER. Letting = 1 yields i(T+S) = i(T).
This guarantees that T + S is a Fredholm operator.
We conclude the section by mentioning that there exists an interesting

analogue, due to F. L. Hernndez and B. Rodriguez-Salinas [150], of strictly
singular operators tailored to utilize the order structure of Banach lattices.
The corresponding operators are called disjointly strictly singular and
their definition is given in Exercise 9 at the end of the section. The reader
is referred to {123, 124] for some recent developments regarding disjointly
strictly singular operators.
Exercises
1. Show that a bounded operator T: X Y between Banach spaces is
strictly singular if and only if T is not bounded below on any infinite
dimensional closed subspace of X.
2. Let (i = 1, 2,.. , n) and (j = 1, 2, .. , m) be all real (or all complex)

. .
Banach spaces and consider a bounded operator

T: X1 X2 .. e e. Ym.
As in Exercise 11 of Section 1.1, it is easy to see that there exist bounded

operators p Y3 (i = 1, 2,. , n; j = 1, 2,.
. m) uniquely deter-
. . ,
mined such that T can be represented by the m x n matrix

T11 T12 ...
I'21 I'99 ...
T is strictly singular if and only if each operator is

strictly singular.
3. Show that a bounded operator T: X i Y between two real Banach spaces
is strictly singular if and only if * is strictly singular. [HINT:
Use Exercise 2 above.]
4. Prove the following property of Lp-spaces that was used in Theorem 4.58:
If X is an infinite dimensional vector subspace of any 4-space, then for
each m E N the vector subspace
Xm = {x = (Xi,X2,...) E X: xi = = = = O}
isalso infinite dimensional.
5. Assume that for a pair of Banach spaces (X, Y) we have the following:
(a) Each infinite dimensional closed subspace of X contains a closed
vector subspace that is isomorphic to X.
(b) No closed vector subspace of Y is isomorphic to X.
Show that each operator in (X, Y) is strictly singular. (It is known
that the pair where 1 <p < oc, satisfies properties (a) and (b);
see {205, Proposition 2.a.2, p. 53].)
6. This exercise presents an example of a strictly singular operator whose
adjoint is not strictly singular. Establish the following.
(i) If Y is a separable Banach space, then there exists a bounded sur-
jective operator T: Y whose adjoint T*: Y* is a linear
isometry.
(ii) If T: p
where 1 <p < oo, is any bounded operator satisfying
(i), then its adjoint operator T*: Cq where +=
1, is not
strictly singular.
[HINT: For (i) let Y2, . .} be a countable norm dense subset in the
.
closed unit ball of Y and then define the operator T: Y by the

formula T(A1, A2,...) For (ii) use Exercise 5 above.]
7. (T. Oikhberg) Show that there exists an infinite dimensional Banach space
X with a non-trivial multiplicative linear functional on the Banach alge-
bra (X). [HINT: An infinite dimensional Banach space X is called
hereditarily indecomposable, in brief X is an HI-space, if no infinite
dimensional subspace of X is isomorphic to a direct sum of two infinite di-

mensional Banach spaces. The existence of was established by
W. T. Gowers and B. Maurey [137]. Moreover, it was shown in [137, 138]
that: If X is an HI-space and T (X), then there exist a scalar a and
some strictly singular operator S (X) (both uniquely determined) such
that T = aI+S, where I is the identity on X. The linear functional q5 on
defined by q5(T) = q5(cJ +5) = a is non-zero and multiplicative.]
8. Show that if p and q satisfy 1 q <p <zoo, then the natural embedding
J: [0, 1] + Lq [0, 1] is not a strictly singular operator. [HINT: Let X
denote the closed vector subspace generated by the standard Rademacher
functions on [0, 1]. It is well known that the two norms and 12 are
equivalent on X; see [106, Theorem 1.10] or [205, Theorem 2.b.3]. Now
establish that J restricted to X is an isomorphism.
9. (HernndezRodriguez-Salinas [150]) A bounded operator T: E
from a Banach lattice to a Banach space, is said to be disjointly strictly
singular if for each disjoint sequence } of non-zero vectors in E the
restriction of T to the closed vector subspace generated by is not
an isomorphism. Clearly, every strictly singular operator from a Banach
lattice to a Banach space is disjointly strictly singular.
Show that if p and q satisfy 1 q < p < oc, then the natural
embedding J: 1] + Lq[O, 1] is a disjointly strictly singular operator.
Chapter 5
Integral Operators
This chapter studies the classical integral operators in the framework of

function Let E and F be two vector subspaces of L0 (5, E1, and
respectively. A linear operator T: E F is said to be an in-
tegral operator if there exists a jointly measurable function K(.,.) (called
the kernel of the operator) such that for each x E E we have
Tx(t) = I K(s, t)x(s)

is
for v-almost all t T. As we shall see, the existence of the kernel imposes on
the operator several interesting and desirable features that are not present
for an arbitrary operator.
The material in this chapter is divided in three sections. In the first

section we present the basic properties of integraloperators with emphasis
on their lattice structure. Among the established order properties, we show
that the vector space of all regular integral operators between two Banach
function spaces is a band in the Riesz space of all regular operators. Ab-
stract integral operators are introduced in the second section and their close
connection with the regular integral operators is thoroughly investigated.
It is shown here that the classes of abstract integral operators and regular
integral operators coincide. In addition, we present a proof of Bukhvalov's
theorem that allows us to recognize when a given operator is an integral
operator. As an application of this characterization, we obtain Dunford's
famous integral representation theorem for continuous operators from an
L1-space to an
o If and v are ci-finite measures, then every continuous operator

from L1(,a) to with I <p < oo is an integral operator.
179
180 5. Integral Operators
The third section of this chapter studies conditional expectation opera-

tors on ideals in L1 (,a)-spaces. The major result here is the characterization
of the positive projections that fix the constant functions as conditional
expectation operators. The last section of the chapter deals with positive
projections and lattice-subspaces.
5.1. The Basics of Integral Operators

In this section, we shall present the basic order-theoretic properties of in-
tegral operators. For extensive treatments on integral operators, see the
monographs [78, 107, 118, 149, 174, 175, 188, 191, 343]. Throughout
this section we shall assume that:
(8, ji) and (T, ii) are two cr-finite measure $paCes.
o E and F are two ideals in Lo(,a) and Lo(v), respectively.
We start our discussion with the definition of an integral operator.
Definition 5.1. Let X and Y be linear subspaces of Lo(,a) and Lo(v), re-
spectively. An operator T: X Y is said to be an integral operator if
there exists a ,a x v-measurable function T(.,.): S x T R such that for
each x E X the function s 'f T(s, t)x(s) belongs to L1 (,a) for v-almost all
t E T and the equality
Tx(t)= I (*)
Js
holds for v-almost all t E T.
Before proceeding further, a clarification is needed regarding the defini-

tion of an integral operator. The vector spaces X and Y are subspaces of
Lo(,a) and Lo(v), respectively. Hence, any operator R: X Y is in fact
an operator between equivalence classes of functions, and so for each equiv-
alence class [xJ E X the element R[x] E Y is likewise an equivalence class.
The definition of the integral operator T tacitly assumes that if x E {x],
then the function y(t) = f8T(s,t)x(s)d,a(s) is a representative from the
class T[x], i.e., T[x] = [y]. In actuality, something stronger is happening
here. Notice that we assumed that for a given function x E [x] there exists
a v-null set B such that the function s i T(s, t)x(s) is frmntegrable for all
t B. Now if x1 E [x], then it should be clear that for each t B the
function s ' T(s,t)xi(s) is also frmntegrable and
f T(s, t)xi (s) dy(s)

= f T(s, t)x(s)
That is, the function
y(t)=
JS
is the same for any x1 E [x]. This not only explains why T[x] = [y] but it
also shows that we can identify the class T[x] with the function y and write
y = Tx. In sum, an integral operator T: X Y gives rise automatically to
an operator from equivalence classes to functions, i.e., it is a "lifting" or a
"selector."
It should be emphasized that the v-null set of points outside of which the
equality (*) holds depends upon the function x; see Exercise 6 at
the section. The x v-measurable function T(.,.): S x T is referred to
as the kernel of the integral operator T.' It is a common practice to denote
the kernel of an integral operator T by the same letter T, and we will follow
this practice. This is very efficient and when the reader gets accustomed to
this convention, it will cause no problem. We shall also say that a x v-
measurable function T(.,.): S x T defines an integral operator
from X toY if for each x E X the function T(.,t)x(.) belongs to L1(1i) for
v-almost all t E T and the function Tx as given by (*) in Definition 5.1
belongs to Y.
We shall also adhere to the following agreement: An arbitrary point of
S will be denoted exclusively by the letter s and an arbitrary point of T
by the letter t. Accordingly, f(s) indicates a function f: S TR and g(t)
denotes a function g: T ilL
We assumed at the beginning of this section that E and F are ideals in

L0 (ii) and L0 (v), respectively. If we consider the supports and 71 of E
and F, respectively, then it is easy to see that an operator T: E F is an
integral operator if and only if there exists a function K E L0 T1, x v) ><
such that for each x E E the function s ' K(s, t)x(s) belongs to L1 (Si,
for v-almost all t T and the equality
Tx(t)= /
Js1
holds for v-almost all C E T1. In other words, when we deal with integral
operators, we can replace S by and T by 71 and assume without loss of
generality that:
o E and F are order dense ideals in L0(1i) and Lo(v), respectively.
Integral operators enjoy the following important continuity property.2
For this reason, the integral operators are also known as kernel operators
2As we shall see in Theorem this property characterizes the integral operators.
Lemma 5.2. Let T : E Lo(ii) be an integral operator. If an order bounded

sequence in E satisfies 0, then 0.
Proof. Assume x an order bound for

Also, let be the kernel of the operator T. In view of Definition 5.1,
we can find some B E E2 with v(B) = 0 such that for each t B and for
each n the functions and belong to L1(,u).
By Theorem 1.82 we know that 0 if and only if every sub-
sequence of has a subsequence that converges pointwise to
zero. This implies that for each t B every subsequence of the sequence
{ T(., has a subsequence that converges pointwise to zero.
Since and holds
for all t B, the Lebesgue Dominated Convergence Theorem easily implies
that 0 for v-almost all t E T.
Since any sequence that converges pointwise to zero also converges in

measure (see Corollary 1.83), the following is an immediate consequence of
the preceding result.
Corollary 5.3. Let T: E Lo(v) be an integral operator. If is
an order bounded sequence in E such that then In
particular, by Corollary 1.89, the operator T is regular and order continuous.
The next result is also a consequence of Corollary 1.89.
Corollary 5.4. Every positive integral operator T: E F is order contin-
uous.
As expected, the positive integral operators from E to F are precisely

the ones with positive kernels.
Theorem 5.5. Suppose that E and F are order dense ideals. An integral
operator T: E F is positive if and only if its kernel satisfies T(s, t) 0
for ,u x v-almost all (s,t) ES x T.
Proof. Let T: E F be an integral operator having the ,u x v-measurable
function T(.,.): S x T JR as its kernel. Assume first that T(s, t) 0 for
all (s, t) A, where A is ,u x v-null subset of S x T. Fix some x E and
then select a si-null set B E such that x(s) 0 for all s B. Next,
choose a v-null set C E E2 such that for all t C we have
Tx(t) = I T(s, t)x(s) dy(s).
is
Put D = A U (B x T) U (S x C), and note that (,u x zi)(D) = 0. Now,
notice that Fubini's Theorem 1.97 guarantees the existence of a v-null set
C1 E E2 such that the set = {s E 5: (s,t) E D} satisfies = 0
for each t C1. Now assume that t C U C1 and s Then (s, t) D,

and so s B (in which case we have x(s) 0 and T(s,t) 0), and t C
(which guarantees that Tx(t) = T(s, t)x(s) Thus, if t does not
belong to the v-null set C U C1, then T(s, t)x(s) 0 for each s and
=0 yields
Tx(t)
= f
$
T(s, t)x(s)
f
= $\Dt
T(s, t)x(s) 0.
This proves that T is a positive operator. Clearly, this proof remains valid
for arbitrary subspaces E and F of Lo(,a) and Lo(v), respectivelyand also
for arbitrary measures and ii.
For the converse, assume that T is a positive operator. We must show
that T(s,t) 0 for x v-almost all (s,t) E S x Since S and T are
both o-finite measure spaces, we can assume without loss of generality that
both measures are Moreover, since E and F are order dense ideals in
and Lo(v), respectively, we can also assume in view of Corollary 1.91
that Xs E E and XT E F. Our last simplifying assumption is that the
kernel can be assumed to be x v-integrable. To see this, note
that since the function s F-4 T(s,t) = belongs to L1(,a) for
v-almost all t E T, by deleting a v-null set from T we can suppose that
J'5 T(s, t)) d,a(s) <oo for all t T. So, we can write T = Bk, where
Bk = {t E T: JT(s, d,a(s) k}. Since the v-measurable function
g(t) T(s, t)I d,a(s) is bounded on Bk, we can conclude from Tonelli's
theorem that T(.,.) is integrable over S x Bk for each k. Thus, we can
assume that the function T(.,.) is x v-integrable over S x T.
To show that T(.,) 0, it suffices to verify (see Lemma 1.96) that
fAxB
for each A E and B E Since XA E E and T is a positive operator,
we have TXA 0 and consequently fB TXA(t) dv(t) 0. By Definition 5.1
the equality
TXA(t) =
holds for v-almost all t. Therefore, using Fubini's theorem, we obtain
f x v)
= L [f T(s, t)
dv(t) L TXA(t) dv(t) 0,
as desired.
Recall that the carrier CE of an ideal E in L0 (ii) is the "smallest"

measurable set outside of which all functions of E vanish; see Definition 1.93.
Corollary 5.6. An integral operator T: E * F is positive if and only if its

kernel satisfies T(s, t) 0 for x all (s, t) E CE x CF.
Corollary 5.7. Two integral operators 8, T: E F satisfy S T if and
only if S(s, t) T(s, t) for x v-almost all (s, t) E CE x CF.
Anotherimportant consequence of Theorem 5.5 is that the kernel of an
integral operator is uniquely determined on CE x CF.
Corollary 5.8. Two integral operators 5, T: E F coincide (i.e., S = T)
if and only if S(s, t) = T(s, t) for x v-almost all (s, t) E CE x CF.
Positive operators that are dominated by integral operators are neces-

sarily integral operators. This basic result is due to A. V. Bukhvalov [77]
and A. R. Schep [294].
Theorem 5.9 (BukhvalovSchep). Every positive operator from E to F
that is dominated by an integral operator is itself an integral operator.
Proof. The proof belowbased on the Theoremis due

to A. R. Schep [294].
Assume that two operators 5, T: E * F satisfy 0 S T and that T
is an integral operator with kernel T(., .). Let
F={AxB:xAEE,xBEF,andT(.,.)isintegrableoverAxB},
and note that F is a semiring of subsets of S x T. From Corollary it
follows that the cr-algebra generated by F is the product cr-algebra E,1 E11,
where for simplicity we write instead of (E1),1 and instead of
Using the hypothesis that T is an integral operator, we can define the
set function '7rT: F * [0, oo) via the formula
x B)
= f x v)
= L [f T(s, t) dv(t)
= f TXA(t)dv(t).
JB
A direct verification shows that 7rT is a measure on the semiring F.
Next, using the operator 8, we define a new set function 7rs: F * [0, oo)
via the formula
irs(A x B) L SXA(t) dv(t) dv(t).
= IT XB(t)SXA(t)
Clearly, 0 We claim that '/rs is a measure. To verify this, it
suffices to show (since irs is dominated by the measure lrT) that is finitely
additive; see Exercise 15 in Section 1.6.
To this end, assume that A x B = A1 x is a disjoint union, where

A x B and all the x are non-empty and lying in F. Fix r E B and let
= {i E {1,. , n}: r E B1 }. Then for each fixed T E B we have
XA(S) = XA(s)XB(r) = XA1(S)XB1 (r) = XA1(S)

i=1 iEN.I-
for all s E S. The linearity of S implies that for each r E B we have
(SXA)(t) XB(r)(SXA)(t) = = (*)
i=1
for u-almost all t T. Since the collection r E B} is finite, there

exists a u-null set C such that (r)(SXA) (t) = (r) (SxAJ (t) is
valid for each T E T and all t C. But then for all t C we have
XB(t)(SXA)(t) = This implies
B)
= f XB(t) (SXA)(t) =
IT
XB1 (t)(SXA1) (t) du(t)
= >'/rs(AixBi).
Therefore, IrS is indeed finitely additiveand hence it is a measure. A
straightforward verification shows that the outer measures and satisfy
0 and that, when restricted to 0 they are both cr-finite.
Next, observe that the measure restricted to the cr-algebra
is absolutely continuous with respect to the cr-finite measure x ii. From
it follows that is also absolutely continuous with respect to
x ii. But then, by the classical theorem, there exists
a function S(.,.): S x T > [0, oo) (and hence
is also x ui-measurable) such that = S x ii) holds for each
UE In particular, for each A x B E F we have
x B)
= f (SXA)(t) du(t)
= f S(s,t) x v)
= du(t).
Now fix some A E with (A) <oo and consider the collection
A = {B E is integrable over A x B}.
Clearly,A is a semiring such that the cr-algebra generated by this semiring
coincides with see the proof of Theorem From this and Lemma 1.96,
it follows that for each A E with /1* (A) <oc we have
SXA(t)
for v-almost all t E T.

= f S(s, t) dy(s)
= f S(s, t)XA(S) dy(s)
To finish the proof, note first that the operator S is order continuous.
(This follows from the inequality 0 S T and the fact that T
order is
continuous by Corollary 5.4.) This and (**) imply immediately that for each
x E E we have Sx(t) = f8 S(s, t)x(s) d,ti(s) for all t E T. That is,
S is an integral operator, as desired.

From the definition of an integral operator T: E F with kernel
it follows that for each x E E function
the t) Ix(.) is 1u-integrable for
all t, and therefore we can define (v-almost everywhere) the func-

tion t f5 T(s,t)Ix(s) d,a(s). It is easy to verify that this function is
see Exercise 2 at the end of the section. In other words, every
integral operator T: E F defines automatically via the formula
MTX(t)
a new positive integral operator MT: E Lo(v) with kernel It

should be emphasized that in general this new integral operator need not
map .E into F. We will specify below under what conditions MT(E) F
and show that in this case MT coincides with the familiar modulus ITI of
the operator T introduced in Section 1.2.
If we consider T as an operator acting from E to L0 (v), then it should
be clear that T MT, and therefore T: E L0(v) is a regular operator.
Since Lo(v) is Dedekind complete, T has a modulus given by the familiar
RieszKantorovich formula
zEE and zlx}
for each x E E+ (see Theorem 1.16). As expected, this formula for the
modulus of T and the one given by (***) coincide.
Theorem 5.10. If T: E F is an integral operator, then T as an operator

from E to Lo(v) is a regnlar operator and its modulus [TI: E Lo(v) for
each x E E satisfies
for
[TIx(t)
v-almost all t. That is,

=
TI = MT.
f
T(s, t) Ix(s) dy(s)
Proof. As mentioned before, MT is a positive integral operator satisfying

T MT. Therefore Ti MT. Assume that another positive operator
L0 (v) satisfies T S. Consider the operator R = MT A S. By
5. 1 . The Basics of Integral Operators 187
Theorem 5.9, the operator R is an integral operator as 0 < R Also

T R and therefore, in view of Corollary the kernel R(. .) of the ,
operator R satisfies T(s,t) R(s,t) 1T(s,t)I for x v-almost all (s,t)

in S x T. This implies that R(s, t) = T(s, t) for 1ux u-almost all (s, t) E S x T
or MT A S = MT. That is, MT 5, which establishes that JTJ = MT.
It turns out that the regularity of an integral operator T is precisely
the property needed to guarantee that MT maps E into F. This important
result is due to W. A. J. Luxemburg and A. C. Zaanen [226].
Theorem 511 (Luxemburg--Zaanen). An integral operator T: E F is
regular if and only if the operator MT defined by(***) has its range
In this case MT coincides with the usual modulus TI of T in 4 (E, F) .3
Proof. Assume first that T is a regular operator. Since F, as an ideal in

L0 (ii), is a Dedekind complete Riesz space, the modulus IT!: E + F exists
in 4(E, F) and is given by the RieszKantorovich formula. That is, the
equality
zEE and zIx}
holds in F for each x E E+. Moreover, using again that F is an ideal in
Lo(v), we see that the above supremum remains the same if it is taken in
Lo(v). That is, the modulus of T calculated in 4(E, F) or in 4(E, Lo(v))
is the same. But, by Theorem 5.10, the modulus of T in 4 (E, Lo(v))
coincides with A'IT. Consequently, MT(E) F.
the converse, suppose that the operator MT maps E to F. Since
For
T MT and MT is positive, it follows that T is a regular operator, and the
proof is finished.
From now on, regardless of whether an integral operator T: E F is

regular or not, we will always use the symbol TI rather than MT and we
will refer to TI as the modulus of the integral operator. It will always be
clear from the context between which spaces the modulus acts, and so this
will cause no ambiguity.
Corollary 5.12. Every regular integral operator T: E F is order contin-
uous. Without the assumption of regularity, the integral operator T is order
continuous as an operator from E to Lo(v).
Proof. Let T: E
F be a regular integral operator. Then, by Theo-
rem 5.11, its modulus TI: E F is also an integral operator, and so by
Corollary 5.4 the operator T itself
is also order continuous.
to this, the regular integral operators are also known as absolute integral operators
188 - 5. Integral Operators
The last statement of the corollary follows from the previous part since
each integral operator T: E F considered as an operator from E to L0 (ii)
is regular.
Let T: E F be an integral operator. As stated in Theorem 5.11,

the modulus of T maps E to F if and only if T is a regular operator.
By Corollary 5.3, we know also that T, as an operator from E to Lo(v), is
always regular. Below we present an example of an integral operator from
E to F whose modulus does not map E to F.
Example 5.13. Consider [0, 1] equipped with its Lebesgue measure and let
S = T = [0, 1]. Also, for each m let = [1 1 The length of
will be denoted by For our ideals we take E = L1(S) and
F= E Lo(T): E L1 (T) and

f IY(t)I dt 0
The kernel T(., on S x T is defined as follows. For each s e [0, 1] and for
each t e where m = 1, 2,..., we let
T(s, t) sin(2'irns).
It is obvious that T(.,.) is a bounded x v-measurable function on S x T.
First, we verify that the integral operator T generated by this kernel
maps E to F. To see this, note that for each x E E and each t e we
have Tx(t) = f5 x(s) sin(27rns) ds, and so the RiemannLebesgue lemma
yields
dt x(s) ds 0
IT =
Thus, y = Tx e F.
Next, we claim that the integral operator TI with kernel T(., .)J does not
map E to F. To this end, it suffices to show that the function v =
does not belong to F. Indeed, for each t e we have
v(t)
= f sin(27rns) ds f sin2 (2irns) ds
and hence v = JTI(xs) F.

If we equip F with the lattice norm = then both
E and F are Banach lattices with order continuous norms. That is, TI may
fail to map E to F even if both E and E are Banach lattices with order
continuous norms.
The vector space of all regular integral operators from E to F will be

denoted by F). Its order structure is described by the following result.
Theorem 5.14 (BukhvalovLuxemburgSchepZaanen). The vector space

K(E, F) of all regular integral operators from E to F is a band in 4(E, F).
Moreover, for each 5, T E K(E, F) and for each x E E we have
[(S V T)x] (t)

= f max{S(s, t), T(s, t)}x(s) dy(s), and
[(SA T)x](t) = dy(s)
for v-almost all t T.
Proof. We can assume without loss of generality that E and F are order
dense in Lo(1i) and Lo(v), respectively. Theorems 5.9 and 5.11 guarantee
that K(E,F) is an ideal in 4(E,F) and that the above formulas for cal-
culating S V T and S A T are true. The only thing left to be shown is that
K(E,F) is order closed in 4(E,F).
To see this, assume that a net {Ta} of positive integral operators sat-
isfies Ta I S in 4(E,F). By Corollary 1.91, we can assume without loss
of generality that is E E, and so Ta15 I S's in F. Since F has the
countable sup property, there is an increasing sequence of indices
such that Taml5 I Sls, that is, f5Tam(5,t)dIl(S) I Sls(t) for v-almost
all t E T. By Theorem 5.5, the kernels Tan (s, t) satisfy Tan (s, t) I. This
implies that there exists a function T(.,.) such that Tan (s, t) I T(s, t) for
x v-almost all (s, t) E S x T. Without loss of generality, we can suppose
that Tan(S,t) I T(s,t) for each (s,t) E S x T. Next, we claim that T(s,t)
is finite for x all (s, t) E S x T.
To establish this, consider the x v-measurable set

D={(s,t)ESxT: T(s,t)=oo},
and assume by way of contradiction that (/2 x v)*(D) > 0. -Since /i-x v-is-a-
a-finite measure, we can also suppose that x v)*(D) <oo. But then, from
Fubini's theorem, there exists a v-measurable subset B of T with v(B) > 0
such that for each t B the set = {s E 5: (s, t) E D} is ti-measurable
and satisfies > 0. Therefore, for each t B we have
(Sis)(t) lim f Tan(S,t)XAt(S)d/2(S) =1 T(s,t)d/2(s) =00.

s
This shows that F, which is impossible. Consequently, the function

T(.,.) belongs to x ii).
Now ifx E then Tan(S,t)X(S) I T(s,t)x(s) for all (s,t),
and so for v-almost all t E T we have
Sx(t) = lim
12-400
= fl-400
lim f Tam(S, t)x(s) d/2(s)
= f T(s, t)x(s) d/2(s).
5. Integral Operators
This shows that S is indeed an integral operator, and so F) is a band

in the Dedekind complete Riesz space 4 (E, F).
We conclude the section by providing a convenient criterion for an op-

erator to be an integral operator.
Theorem 5.15. Let T: E F be an operator s'uch that T: P2 Lo(ti) is
cr-order contin'uo'us. If the restriction of T to some order dense ideal in P2
is an integral operator, then T: P2 + F itself is an integral operator.
Proof. Start by observing that the operator T: P2 F is an integral oper-

ator if and only if T: P2 L0 (ii) is an integral operator. Therefore, without
loss of generality, we can suppose that F = Lu (ii). We can also assume that
P2 is order dense in
Suppose first that T is a positive operator and that T is an integral
operator on an order dense ideal P20 in P2. Denote by T0 (.,.) the kernel of
the operator T0 = TIEO. By Theorem 5.5, the kernel To(.,.) is a non-negative
ji x ti-measurable function. Consequently, we can assume that T0 (s, t) 0
holds for each (s, t) E S x 7. We shall finish the proof be proving that T is an
integral operator with the same kernel. That is, according to Definition 5.1,
we shall show that for each x E P2 the function s To(s, t)x(s) belongs to
Li(,u) for ti-almost all t T and that the equality
Tx(t) / To(s, t)x(s) dy(s)

Js -
holds for ti-almost all t E T.

To this end, it suffices to consider an arbitrary non-negative function
x E P2. So, fix a function x E P2+ and then pick a sequence in P20
satisfying 0 x. Hence, Tx since T is cr-order continuous. In
particular, Tx(t) for u-almost all t E T. Let D be the set of points
at which there is no convergence or where Tx(t) = 00. Thus, zi(D) = 0.
Since E P20, there exists a ti-null set such that for each t
(1) The function s ' To(s, belongs to L1(,u).
(2) = f8 To(s, dji(s).
Let B = D U U=1 and observe that v(B) = 0. For each t B we have
To(s,t)x(s) for ,a-alrnost all s E 5, and
= Tx(t) <00. (*)
So, by B. Levi's theorem, f8To(s,t)x(s)dji(s) Tx(t) < oo, that is, the
function s To(s, t)x(s) belongs to L1 (ji) for ti-almost all t E as required.
Furthermore, using once more that To(s, I To(s, t)x(s), we see that
f To(s,t)x(s)
The latter and (*) yield Tx(t) = f5 To(s, t)x(s) each t B. That for
is indeed a kernel operator with the same kernel as T0.
is, T
To show that the assumption that T is a positive operator is not essential,
we proceed as follows. Since T: E Lo(zi) is u-order continuous, T is
regular (see Corollary 1.89(a)), and hence the operator T+: E Lo(zi)
exists and is u-order continuous. By Theorem 5.10, the integral operator
T0: E0 Lo(z') is also an integral operator
with kernel .). It remains to note that the restriction of to E0
coincides with Thus, instead ofT we can consider the positive operators
T+ and T,
and the desired conclusion follows.
Exercises
1. Assume that ,a and ii are a-finite measures. Show in detail that if the
kernels of two integral operators 8, T: E F satisfy 8(s, t) =T(s, t) for
x v-almost all (s, t) c S x T, then 8x = Tx for each x C E.
2. Let (5, ,a) and (T, ii) be two a-finite measure spaces, and consider
a ,a x v-measurable function K: S x T p R. Assume that a function
f C Lo(ji) satisfies K(.,t)f(.) C L1(ji) for v-almost all t C T. Show that
the function g: T R defined v-almost everywhere by the formula
g(t)
= f K(s, t)f(s) dy(s)
is v-measurable, i.e., g C Lo(v). [HINT: By Lemma 1.90 there exists a se-

quence of ,a x v-integrable functions satisfying IK(s, t)1
for ,a x v-almost all (s, t) C S x T, and K(s, t) for ,a x v-almost
all (s, t). Since each is ,a x v-integrable, Fubini's Theorem 1.97 guaran-
tees that the formula = d,a(s) defines a v-integrable
function. Now a straightforward application of the Lebesgue Dominated
Convergence Theorem yields g(t) for v-almost all t C T.]
3. Let T(.,.) be a ,a x v-measurable function such that:
(a) T(s,t)I 1 for v-almost all t C T.
(b) jT(s, t)I dii(t) 1 for ,a-almost all s C S.
Prove that the integral operator T with kernel T(.,.) is a bounded oper-
ator from L2(ji) to L2(v).
4. ([78, Section 4.1]) Let T: L2(ji) L2(v) be an integral operator. Assume
that there exists a subset B C such that v(B) = 0 and for each function
x C L2(ji) we have f5 T(s, t)x(s)I dji(s) <co for each t B. That is, the
192 ---- 5. Integral Operators
exceptional set in the definition of the integral operator T is independent

of the "input" function x.
Prove that T(s, < cc for v-almost all t T. (Operators
that satisfy this condition are known as Carleman operators.)
5. Assume that (S, , ,u) and (T, are two o-fithte measure spaces,
,
and let 1 <p, q < satisfy +=1. Show that an integral operator
T: maps the closed unit ball of to an order bounded
subset of Lo(v) (i.e., there exists a function g E Lo(v) such that g
for each s e with 1) if and only if there exists a v-null set
BE such that
f
for each t E T\B.
In particular, an integral operator T: L2 (ii) + (ii) is a Carleman
operator if and only if T maps the closed unit ball of L2(1i) into an order
bounded subset of Lo(v).
6. Give an example of an integral operator on L2 [0, 1J which is not a Car-
lernan operatorwhere [0, 1] is equipped with its Lebesgue measure. In
view of Exercise 4, this demonstrates that in general the exceptional set in
the definition of an integral operator depends on the function the operator
is acting upon. [ifiNT: Use Exercise 3 above.]
7. Let be an increasing sequence of regular integral operators and let T
be another regular integral operator. Show that T holds in F)
if and only if T(s, t) t) x CF.
8. Let E be an ideal in Lo(,u), where (8, is a non-atomic cr-finite mea-
sure space. If T: E + E is a regular integral operator, then show that T
is disjoint from the identity operator, i.e., ITt A I = 0.
9. (AbramovichAliprantisZame [20]) Show that for a Dedekind complete
AM-space E with unit e the following two statements are equivalent.
(a) There exists a probability measure space and a surjective
lattice isometry T: E + with Te 1.
(b) E admits a strictly positive order continuous linear functional.
[HINT: Assume that E + R is a strictly positive order continuous
linear functional. We can assume that gb(e) 1. Clearly, the formula
Us HI = defines an L-norm on E. If E is the norm completion of
(E, I
then E is an AL-space (see Exercise 6 of Section 3.4) having
e as a weak unit. Now, according to the KakutaniBohnenblustNalcano
Representation Theorem 3.5, E is lattice isometric to some L1 (7r)-space,
where e corresponds to the constant function 1. Show that is a proba-
bility measure and that E is lattice isometric to I
10. Let (8, ,u) and (T, v) be two arbitrary measure spaces, and assume
that 1 <p, q < cc satisfy + = 1. If a function T(.,.) e x
then show that the formula
Ts(t) t)s(s) dy(s)
= f T(s,
defines a regular integral operator from to such that its norm

satisfies IIT(., .)llp. (When p = q = 2, such an operator T is called
a HubertSchmidt operator.)
11. Show that each HilbertSchrnidt operator is a Carleman operator.
12. Assume that (S, are (T, ii) are two u-finite measure spaces, and
assume that (S, is atomicwhich in this case means that S is a
union of at most countably many atoms.
Show that a linear operator T: E + Lo(v), where E is an ideal in
L0 (u), is an integral operator if and only if T is order continuous.
5.2. Abstract Integral Operators

This section is a continuation of the previous one. The main objective here
is to present two important characterizations of integral operators (Theo-
rems 5.28 and 5.30). Also an integral representation of order continuous
linear functionals (Theorem 5.26) will be obtained that is very useful in
applications.
Recall that if L is a Riesz space, then denotes the order continuous
dual of L, i.e., the band in L" consisting of all order continuous linear
functionals. If L is a normed Riesz space, then we shall denote by the
ideal in L* of all order continuous and norm bounded linear functionals on
L, that is, L is a Banach lattice, then of course =
and = Riesz spaces with "small" order continuous duals should be
viewed as "pathological" objects from the point of view of the theory of
Riesz spaces.
Definition 5.16. A Riesz space L is said to be a normal Riesz space if
its order continuous dual separates the points of L, that is, if for each
mon-zero x E L there exists some E satisfying q5(x) 0.
Let L and M be two Riesz spaces with lvi Dedekind complete. If is a

vector subspace of then the operators from L to A/I are the
operators in the band generated in 4(L, M) by A/i. The
operators =
(i.e., when are called abstract integral operators. They
were introduced by G. Ya. Lozanovsky in [219] under the name "almost
integral operators." Here is their exact definition.
Definition L and M be two Riesz spaces with iVi
5.17 (Lozanovsky). Let
Dedekind complete. An order bounded operator T: L IVi is said to be an
abstract integral operator (or an almost integral operator) if it is an
-integral operator, i. e., if T belongs to the band generated by
lvi in 4(L, M). The band (La' is called the band of abstract
integral operators.
194 -- 5. Integral Operators
Thus, a positive operator T : L M is an abstract integral operator if

and oniy if there exists a net {Ta} of positive operators such that T hi
4(L, M) and for each there exist 0 < E and 0 Va E M satisfying
< Non-trivial abstract integral operators exist provided that
{0}. For instance, when L is a normal Riesz space, then there are
many abstract integral operators.
From the preceding definition and Theorem 1.29, the following result
should be immediate.
Lemma 5.18. Every abstract integral operator is order continuous.
As in the previous section, E, ir) denotes our "generic" measure space,
(S, E1, and (T, E2, ii) denote o-finite measure spaces. A function space
E is simply an ideal in some Lo(ir).4 In this case we shall also say that E is
associated with Lo(ir) or with the measure it.
Definition 5.19. A function space is said to be:
(1) A normed function space if G carries a lattice norm. If G is
complete in this norm, then G is a Banach function space.
(2) A normal function space if G is normal, i.e., if separates
the points of G.
Not every function space is normal. For instance, we have the following
result that is essentially due to M. M. Day [101].
Lemma 5.20 (Day). If E, it)
is a non-atomic ci-finite measure space,
then for each 0 p < 1 we have L(ir) = {0}.
Proof. For the proof of the 0 < < I case, see [25, Theorem 12.31].
p
To see that = {0}, let 0 E We claim that is order
continuous. To see this, assume that 0 in Lo(ir). By Corollary 1.86,
the sequence is relatively uniformly convergent to zero. So, there
exists a null sequence of real numbers and some x E (it) satisfying
0 < for each n. This implies 0 for each n, and
so 0. Thus,is order continuous.
Now, by the first part, = 0 on each ideal with 0 <p < 1. Since
each of these ideals is order dense in Lo(ir) and / is order continuous, it
follows that = 0 on Lo(ir), as desired.
If an integral operator T: .E F between Banach function spaces is
positive, then it is automatically norm continuous. As mentioned earlier, if
T is not positive, then the modulus of T does not necessarily act from .E
to F; and so the norm continuity of T does not follow from the automatic
4Many authors define a function space to be simply a vector subspace of some Lo(ir)space.
continuity of positive operators. Nevertheless, integral operators are always

norm continuous.
Lemma 5.21. Let C be a Banach function space associated with some
Lo(7r). If xli 0 holds in G, then has a subsequence which
is relatively uniformly convergent to x (and hence also has a subse-
quence which is pointwise convergent to x).
In particular, if 0 for each n and the series x is norm
convergent, then x(w) for 7r-almost all w E
Proof. Assume that HXrL xif 0. Then, by a general property of Banach

, has a subsequence which is relatively uniformly
x; see Exercise 13 in Section 1.2. This subsequence automatically converges
pointwise to x.
Corollary 5.22. Every Banach function space C associated with a
measure space has a weak order unit, i.e., it contains a function
that is strictly positive on the carrier of C.
In particular, if C is order dense in Lo(7r), then there exists some u E C+
satisfying u(w) > 0 for 7r-almost all w E IZ.
Proof. Clearly, we can assume without loss of generality that the ideal G
is order dense in Lo(7r). By Corollary 1.91 there exists a disjoint sequence
C E such that = and the function belongs to C
for each n E N. Then the series u = >t1 where
denotes the norm of C, is norm convergent in C. By Lemma 5.21, we have
u(w) = (w) for 7r-almost all w E 11. This implies that
II)
u is a weak unit. D
Theorem 5.23. Every integral operator between Banach function spaces is
norm continuous.
Proof. Let T: E F be an integral operator, where E and F are Banach

function spaces. It suffices to show that T has a closed graph; see Theo-
rem 1.5. To this end, assume that xli 0 in E and 0
in F. By passing to appropriate subsequences (using Lemma 5.21), we can
assume without loss of generality that is order bounded and converges
u-almost everywhere to x and that converges everywhere
to y. Now, Corollary 5.3 implies Tx = y.
Our next goal is to establish the important fact (due to G. Ya. Lozanov-
sky) that each normed function space is necessarily normal. To prove this,
we need the following technical result due to W. A. J. Luxemburg and
A. C. Zaanen [224, Lemma 31.1].
Lemma 5.24 (LuxemburgZaanen). Let b be a positive linear functional

on a normed Riesz space L. Assume that {un} is an order bounded sequence
in satisfying < oo. Then for each 0 there exist two
sequences and {wn} in L+ such that 0, cb(wn) < for each ri,
and un=vn+wn for eachri.
Proof. By hypothesis 0 u holds for some u E L and for each ri. Fix
0, and for each n and m let
un,m =
u and the positivity of b, we see that cb(un,m) Tm b(u).
So, there is a strictly increasing sequence of natural numbers such
that
0 cb(un,m) <* (*)
for all m Next, for each ri let
Zn = ul,m1 A 'LL2,rn2 A
From 0 we get p 0, and so

I
0. Obviously Zn j, and 50 j 0. Now for each ri let
=
Yn (ui%mfl uj,mj) 0. -
Since 1 i n implies mn, it follows from (*) that
b(yn) = <.
claim that un
We + Yn for each n. To establish this inequality, it
suffices to prove that un Uj,rrij + Yn for each j = 1, . , n. For this, note
.
that ifl j then
+ Yn = + Ui,mj)
+ Uj,mj) =
Finally, using the Riesz Decomposition Property, we can write each
as un = +Wn, where 0 Vn Zn and 0 Wn Yri. Clearly, the sequences
and {Wn} satisfy the desired properties.
Theorem 525 (Lozanovsky). If C is a normed function space, then
separates the points of C. In particular, C is a normal RiesZ space.
Proof. Let H = Gfl Li('rr) and let = xd'rr for x We define

the function p: [0, cx) by the formula
yEH and Oy
where denotes the lattice norm of C.
We claim that p(x) > 0 for each 0 < x E H. To see this, assume by way
of contradiction that p(x) = 0 for some 0 < x E H. Consequently, there
exists a sequence [0, x] satisfying + b(x < for each
n. From Jun11 < and Lemma 5.24 (applied to it follows that
given any > 0 there exist two sequences and such that 0,
<e and = + for each n. Then for each n we have
=
<
Letting n we get 0 < b(x) for each 0, which is impossible.

Hence,p(x)>Oforeach 0<xEEH.
A direct verification shows that the function x ' p(Lxl) defines a lattice
norm on H. Clearly, p(lxl) llxM and p(IxI) for each x E H.
To prove that separates the points of G, it suffices to show (see
Exercise 4 at the end of the section) that for each 0 0. So, fix some u > 0. Since Li(7r) is an
order dense ideal in L0 (71) (this is the only place in the proof where we use
the assumption that 71 is o-finite), it follows that H is an order dense ideal
in G. Therefore, we can assume without loss of generality that u E H. In
particular, u also belongs to L1 (7r).
By the HahnBanach theorem, there exists a positive linear functional
on H such that = p(u) > 0 and p(IxD for each x E H.
In particular, p(x) b(x) for each x E and hence is order
continuous on H. Notice next that the formula
yEH and
defines an additive function from H+ to JR. Therefore, by Theorem 1.15,
the formula = extends to a positive linear
functional on G. Clearly, for each x E G
and so is norm continuous on G. Moreover, it is easy to see that
q' implies that is also order continuous,
i.e., E To finish the proof, note that = > 0.
The importance of order continuous linear functionals has been already

demonstrated many times in our discussions. Our next result adds a new
5. Integral Operators
feature to this story by providing an extremely useful integrad represen-

tation for these functionals.5 In the special case of spaces of bounded
functions this representation was obtained by B. Z. Vulikh and in gen-
eral by G. Ya. Lozanovsky. We refer to the paper by B. Z. Vulikh and
G. Ya. Lozanovsky [329] for additional historic details.
Theorem 5.26 (Lozanovsky). Assume that C is a function space associated
with a cr-finite measure Then for each E there exists a function
E LO(7r) such that for each x E C we have the integral representation
The function
= f
is uniquely determined on the carrier of C.
If C is an order dense normal function space in LO(7r), then the func-
tion is unique and the mapping from to Lo(7r), is a lattice
isomorphism whose image C' is the order dense normal function space in
LO(7r) described by
C' = {g E L0(7r): gx E Li(7r) for all xE C} 6
Moreover, C' can be identified with where each function g E C' acts on
C as an order continuous linear functional via the formula
(x, g) = g(x) = I g(w)x(w)

Jci
Proof. Consider the collection I' = {A E E: XA E C}. Clearly F is a

semiring of subsets of 12 and, by Corollary 1.91, the cr-algebra generated by
F coincides with E.
Fix some 0 E and then define the set function e: F [0, oo) by
= cb(XA). The order continuity of implies easily that e is a measure.
Consequently, the usual Carathodory extension of restricted to E, is a
measure on E. Next, observe that the measure is obviously afinite (since
e is cr-finite) and that is absolutely continuous with respect to the cr-finite
measure 7r. Indeed, if A E E satisfies 7r(A) = 0, then XA = 0 E C, and so
(A) = = cb(xA) 0. Therefore, by the classical
theorem (here we use again that is o-.-finite), there exists some function
0 E L('ir) such that = holds for all A E E.
Consequently, if a function f is a E-step function (that is, of the
form f = with each E E), then b(f) = d7r(w).

Now let x E C+. Then there exists a sequence of E-step functions {f,-j such
thisreason the order continuous linear functionals are also called normal integrals.
6 ideal is often referred to as the Kthe dual of C.
that 0 f71 I x holds in C. The order continuity of and Beppo Levi's

theorem imply that
This
= lim
f
= lim
establishes the representation of
= f
The verification of the uniqueness
of on the carrier of C is straightforward.
Assume now that C is normal and order dense. The uniqueness of gi,
follows easily from the fact that in this case the carrier of C is all of
Moreover, it should be clear that the mapping 'k is linear and that its
image C' is an ideal in L0(ir). To see that 'k is a lattice isomorphism,
fix E Then for each x E we have
= sup = sup
f = f
This implies that = and so 'k is a lattice isomorphism.
-For the last part, we need to verify that is order dense in Lo(ir).
To this end, take any 0 < y E L0(ir), and then choose some x E E+
satisfying 0 < x y. Since the order continuous dual separates the
points of C, there exists some 0 0. This implies
o Ax x y, and from Ax E it follows that is order dense
in L0(ir). The proof is finished.
From the preceding theorem and Corollary we get the following.

Corollary 5.27. If C is a Banach function space associated with a a-finite
measure ir, then the order continuous dual has a weak unitand so it is
a principal band in
We are now ready to present a characterization of integral operators
as abstract integral operators. This basic result is due to G. Ya. Lo-
zanovsky [219].
Theorem 5.28 (Lozanovsky). If E and F are function spaces and E is
normal, then the bands of abstract integral operators and regular integral
operators from E to F coincide, i.e., ,(E, F) = 0
Proof. By Lemma 1.94, we can assume without loss of generality that E
and F are order dense ideals in L0 and L0 (ii), respectively. For simplicity
we denote the band by B. Take an arbitrary 0 e E 0 F. By
Theorem 5.26 there is some g E such that b(x) = f5g(s)x(s)
holds for all x E E. Therefore, for each x E E we have
e)x](t) =
[f g(s)x(s) =
200 Integral Operators
for v-almost all t e T. Consequently, is a regular integral operator with

kernel g(s)e(t), i.e., e e k(E, F). Since, by Theorem 5.14, the regular
integral operators form a band in 4 (E, F) , it follows that 13 (E, F).
For the converse we must establish that F) Take a x ii-
measurable subset U of S x T with x v)*(U) < oo and assume that the
/2 X v-measurable function Xu defines an integral operator T acting from E
to F; so, in particular, Tx(t) = e F for each x e E.
Since x (U) <oo, there exists a sequence of u-measurable subsets
of S and a sequence of v-measurable subsets of T such that
UCUAnXBn. (t)
Now recall that the order denseness of E implies (according to Theorem 5.26)
that is also order dense. In particular, it follows from Corollary 1.91 that
for each m there exist sets As',... with E and
v-measurable sets Bk,... with e F such that =
and = This shows that in the union in (f), we can suppose that
E and e F for each n. Next, put = Ufl Ak x Bk) and
note that the integral operator with kernel satisfies the inequalities
o xBk Since XAk E 13 for each k, it follows that
e 13. From I Xu we get T T E 13; see Exercise 7 in
Section 5.1.
Finally, let T: E F be a positive integral operator with kernel T(.,.).
In view of Theorem 5.5, we can assume without loss of generality that
T(s, t) > 0 for all (s, t) E S x T. Since the measure x ii is o-finite, there
ists a sequence of x v-step functions such that 0 t) I T(s, t)
for each (s, t) E S x T. Now notice that if = is a x
function satisfying 0 T, then each x v-measurable set satis-
fies the conditions of the previous part. Consequently, the integral operator
generated by belongs to 13, and so the integral operator with kernel
also belongs to 13. Thus, the integral operator with kernel 'J!Th belongs
to 13. Since 13 is a band, it follows (again by Exercise 7 in Section 5.1) that
T e 13. Therefore, F) 13 is also true, and the proof is finished.
Recall that if T: L + M is an order continuous regular operator be-
tween two Riesz spaces with M Dedekind complete, then its order adjoint
operator M" 17, defined for each x E L and each y e via the
duality identity
(Tx, y) = (x, =
carries order continuous linear functionals to order continuous linear func-
tionals, i.e., holds. To see this, let y E and assume
Xcy 0 holds in L. Then the order continuity of T implies TXa 0 in M,
________-
and so = y(TXa) 0 in JR. This shows that is order con-
tinuous. (If L and M are Banach function spaces, then of course =
Now let T: E p F be a regular integral operator between two normal

function spaces. Then, by Corollary 5.12, T is order continuous, and so
C That is, if we consider the restriction of to
then we have another order continuous regular operator
between normal function spaces. It turns out that this operator is also an
integral operator.
Theorem 5.29.. If T: E F is a regular integral operator between two
normal function spaces, then its adjoint operator is also a
regular integral operator, and its kernel satisfies = T(s,t)
fort' x ,u-almost all (t, s) E Tx S. That is, if y E then
= I T(s, t)y(t) dv(t)

Jr
holds for p,-almost all s E S.
Proof. Let T: E F be a regular integral operator between two normal

function spaces. Without loss of generality, we can assume that E and F are
order dense ideals in Lo(,u) and L0(v), respectively, and that T is a positive
operator. Fix some 0 y E Then for each x E F+ we have
(Tx, y)
= L [f T(s, t)x(s) y(t) dv(t)
dv(t).
So, by Tonelli's Theorem 1.98, the function (s, t) T(s, t)x(s)y(t) is x ii-
integrable. In particular, for each fixed x E the function T(s, .)x(s)y(.)
is integrable over T for 'u-almost all s E 5, and the function
g(s) = LT(5,t)Y(t)x(5)dt'(t) =
is Since E is order dense in L0 (n), it easily follows that the

function h(s) = is 'u-measurable. Moreover, for each
x E E and each y E it follows from Fubini's Theorem 1.97 that
(Tx,y) =
f t)y(t) dv(t)]x(s) dy(s)

= f h(s)x(s)
It now readily follows from Theorem 5.26 that h E and that

= h(s) = T(s, t)y(t) dv(t) holds for n-almost all s E S. That
is, is an integral operator whose kernel satisfies s) = T(s, t) for

ii x all (t,s) e Tx S.
We are now ready to state and prove another important characterization

of integral operators that is due to A. Bukhvalov [77].
Theorem 5.30 (Bukhvalov). For an operator T: E F between two func-

tion spaces the following four statements are equivalent.
(1) T is an integral operator.

(2) If E is order bounded and satisfies then
T satisfies the two conditions:

(a) If C E is order bounded and 0, then 0.
(b) If xA E E for some measurable set A of finite measure, then for
every sequence of measurable subsets of A with 0
we have
(4) The operator T: E Lo(v) is u-order continuous and satisfies
condition 3(b) above.
Proof. (1) (2) This follows from Lemma 5.2.

(2) (3) Condition 3(b) is a special case of (2) and condition 3(a)
follows immediately from Corollary 1.83.
(3) (4) This follows from Corollary 1.89.
(4) This proof is quite involved. Before proceeding with it, we
(1)
need to make a few simplifications. First of all, note that we can assume
without loss of generality that F = Lo(v) and (in view of Lemma 1.94) that
E is order dense in Hence, by Corollary 1.89, the u-order continuous
operator T: E Lo(v) is regular and order continuous.
Let us show that we can also assume that the operator T is positive.
Indeed, observe that since T is order continuous, T+ and T are both order
continuous. So, our claim will be justified if we show that satisfies 3(b)
(in which case T = T satisfies 3(b) automatically as a difference of
two operators satisfying this condition).
To see this, assume that is a sequence of ia-measurable sets such
that 0 and A for some measurable subset A C S such that
pJ(A) <oc and XA E E. Now fix a ia-measurable subset C of A and note that
the obvious identity C\Cn(A\BTh) = implies f 0.
Abstract Integral Operators 203
Therefore, using condition 3(b) , we get 0. Consequently,
= liminf
<
lim sup
ntoo
Since, by Theorem 1.17, = sup{Txc: C E E1 and C }, the

previous inequality yields
urn sup
fl4 00
whence lim sup 0. Since the inequality 0 lim inf is

trivially true, it follows that urn sup = lim inf = that is,
0,
0 indeed holds, and so T+ satisfies 3(b). Therefore, we can as-
sume without loss of generality that T is also a positive operator.
Next, we claim that E can be assumed to be a normal function space.
To see this, consider the order dense ideal E0 = E fl (u). Clearly, E0
is a normal function space since the order continuous linear functionals on
are order continuous on E0 as well. Now the operator T0 = TIE0,
the restriction of the operator T to E0, obviously is order continuous and
satisfies 3(b) because the operator T satisfies it. If we can show that T0 is
an integral operator, then Theorem 5.15 will guarantee that T is also an
integral operator. Thus, replacing E with E0 and T with T0, we can also
assume without loss of generality that E is a normal function space.
We now proceed with the rest of the proof. From Theorem 5.28 we
know that k(E,F) = 0 F)dd. Consequently our operator T can be
written as T = R + S with R E F) and S I 0 F). This shows
that T is an integral operator if and only if S = 0. From 0 S T and
the order continuity of T, it easily follows that S is also order continuous
and satisfies 3(b). Therefore, to finish the proof, we need to establish that
the zero operator is the only positive operator S that satisfies condition
3(b), is order continuous and is disjoint from 0 F. Assume by way of
contradiction that S > 0 is such an operator.
Since (in view of Theorem 5.26) is order dense in ('u), it follows
that EflE" is an order dense ideal. In particular, is order dense in E.
Therefore, the order continuity of S implies that there exists a measurable
set A of finite 1u-measure such that XA E E fl and e SXA > ft So,
there is some > 0 and some subset D of Y with v(D) > 0
such that e(t) > for each t D, i.e., e = SXA
Consider b e E 0 F, where we let b = XA By our assumption

S A (b 0 e) = 0, and so (by Theorem 1.17) we obtain
0 = [SA(cboe)](xA)
= + (fA\C dii) e: C E and C C A}
= and CCA}.
Notice that for each measurable subset C of A we have
SXC+ii(A\C)eSXA+ii(A)e [1+ii(A)]e. (t)
F has the countable sup property, there is a sequence
Since
such that C A for each n and
inf + \ E and C A} =0. (*)
We claim now that there exists a subsequence } of such that

0 and
inf
mEN
+ ii(A \ )e} =0.
If this claim is established, then we get a contradiction as follows. Since
\ Cam) 0,it follows from 3(b) that SXA\Cnm 0 or, equivalently,
that 8XCnm SXA e. Consequently, by Egorov's theorem, there exists

a v-measurable subset D1 of D with v(Di) > 0 and some m1 such that
e(t) (t) <e holds for all m and all t E D1. Therefore (since
e(t) 2 for each t E D), we can conclude that
5XCTLm (**)
for all m Notice now that no matter how large the integer might
be, there exists some El > 0 such that
inf + \ Cnm)e} E1XD. (* * *)
mfl <mi
To see this, fix 1 <m < If ii(A \ Cam) = 0, then = A ii-a.e., and
so
+ \ Cnm)e = SXCThm = 8XA = e 2EXD.
Otherwise, if ii(A \ Cam) > 0, then
5XCnm + ii(A \ Cnm)e \ Cnm)e.

Since we have only a finite number of terms, the existence of a desired > 0
is established. Prom (**) and (***), it follows that
inf + ii(A \ Cnm)e} A )xD1 > 0,

mEN
which is a contradiction.
It remains to prove the existence of a subsequence of with

the required properties. For each k E N let Nk = {n> k: p1(A \ <
and note that for each n Nk we have
In view of (f) and (*), this implies
inf
flENk
and cA} =0.
In particular, we have that infflENk = 0, infflENk \ = 0, and
that each set Nki5 infinite. Let Nk = {mk,1, mk12,. .}, where mk,n
.
for each n.
Now for each k and r in N put
Yk,r
=
From = it follows that for each k we have Yk,r
0, 0. Conse-
quently, by the diagonal property (see Theorem 1.87) there exists a sequence
{Yk} in Lo(v) with Yk t 0 and a strictly increasing sequence {rk} of natural
numbers such that Yk,rk Yk for each k. Let It'I = {mk,1,. .. ,
and note that M is a countable subset on N. Therefore, M = {ni, n2,. .}, .
where n1 <fl2 and clearly it(A \ Cam) = 0.

We claim that is the desired subsequence, i.e, that
To see this, fix 6> 0, and then pick some mU such that p..(A \ Cam) <6 for
each m mU. Then for each k we have:
g < inf
mmo
+ \ Cnm)e} inf + 6e Yk + 6e,
where the last inequality follows from the definition of functions Yk,r and the
fact that Yk,rk But then Yk 0 yields g <6e. Since 6 > 0 is arbitrary,
we get g = 0. This completes the proof of the theorem.
Some historical comments regarding Bukhvalov's Theorem 5.30 are in

order. It was proven by A. V. Bukhvalov in [77] under the assumption
that the domain space is normal. The basic ideas of the proof are due to
H. Nakano [253J, where he established that a positive bilinear form lies in
the band generated by the finite-rank bilinear forms if and only if it satisfies
a continuity property similar to the one of statement (2). Bukhvalov's the-
orem was also reproved independently by A. R. Schep [294]. Removing the
assumption that E is not particularly important since the
function spaces with trivial order dual may carry only trivial integral opera-
tors. For instance, it follows from a well known theorem by S. Kwapien [1963
(see also [170, Theorem 8.4]) that the function space Lo[0, 1] admits only
the trivial integral operator.
The first simple application of Theorem 5.30 describes an important
class of Banach function spaces on which every continuous operator is an
integral operator.
Corollary E be a Banach f'anction space with order continuous

5.31. Let
norm. Then every continuous operator T: E is a regular integral
operator. That is, =
Proof. Since is a Dedekind complete with unit, we know

from Theorem 3.9 that (E, = 4 (E, So, it remains to
show that every positive operator T: E is an integral operator.
Assume that E is associated with Lo(j4, and let an order bounded se-
quence E satisfy 0. The order continuity of the norm in E
easily implies (see Exercise 8 at the end of the section) that 0,
and so 0. In particular, we have T satisfies
condition (2) of Theorem 5.30, and so T is an integral operator.
Corollary F is a reflexive Banach function space associated with

5.32. If
Lo(zi), then every continuous operator T: L1(,a) F is a regular
operator. That is, (L1(14, F) = k (Li(14, F).
Proof. Let T: L1(ji) F be a continuous operator. By Theorem 3.9 we

know that the operator T is regular. Also, T is order continuous since
L1(,a) has order continuous norm. Taking the adjoint operator, we see that
T*: F* Li(14* = Since the Banach function space F is reflexive,
both F and F* have order continuous norms, and hence F* So, by
Corollary 5.31, the operator T*: F* Li(,a)* = is a regular integral
operator.
Since = L1(1i) (see Exercise 6 at the end of the section), The-
orem 5.29 implies that T = T**: = L1(,a) (F3fl* = F must be a
regular integral operator.
The two classical representation theorems of N. Dun-

ford [117] are special cases of the preceding results.
Corollary 5.33 (Dunford's Theorem). We have the following.

(1) If 1 p < 00, then every continuous operatorT: f
is a regular integral operator.
(2) If 1 <p <oo, then every continuous operator T:

is a regular integral operator.
We shall close the section with one more characterization of integral

operators between Banach function spaces when the measures ,u and ii are
both finite. To state the result, we need some preliminary discussion.
If D is any subset of Lo(ir) and u e Lo(ir), then let uD = {ux: x e D}.
Any subset of the form uD with u satisfying u(w) > 0 for it-almost all w e
is called a twist of D. It is not difficult to see that a twist of an ideal is an
ideal and a twist of an order dense ideal is also order dense. Moreover, if G
is a Banach function space, then every twist ziG of G is a Banach function
space. (To see this, define the lattice norm on ziG by IlluxUl = lIxM for each
xEG.)
Regarding twists, we have the following remarkable result of G. Ya. Lo-
zanovsky [220, Theorem whose lengthy and technical proof is omitted.
Theorem 5.34 (Lozanovsky's Twist Theorem). Assume that G is an order
dense I3anach function space associated with a finite measure it. Then there
exists a twist ziG of G such that C uG C L1(ir).
By means of Lozanovsky's Twist Theorem we can now establish the

following representation theorem for integral operators. This representation
theorem is also due to G. Ya. Lozano.vsky and it can be found in [329].
Theorem 5.35 (Lozanovsky). Assume that (8, E1, ,u) and (T, >2, ii) are
two finite measure spaces, and let E and F be two order dense I3anach
function spaces associated with L0 (,u) and L0 (ii), respectively. If T: E F
is a regular integral operator, then there exist a strictly positive function
zi e Lo(14, a strictly positive function v e Lo(v), and a ,u x v-integrable
function K such that for each x E E we have
Tx(t) = v(t)
f K(s, t)u(s)x(s)
for v-almost all t e T.
Proof. Without loss of generality, we can assume that T is a positive

gral operator. Suppose first that E and F L1(v). Then e
and so T1s(.) = f5 T(s,.) d,u(s) E Li(v). But then Tonelli's theorem im-
plies that T(.,.) e L1 (,u x ii), and our conclusion holds true with u =
v = 1T, and K = T(., .).
208 - - 5. Integral Operators
Now consider the general case. By Lozanovsky's Twist Theorem 5.34,

two strictly positive measurable functions u E Lo(,u) and w in
Lo(v) such that uE and wF Li(v). Next,
define the positive operator K: uE wF via the formula K(ux) = wTx.
From Theorem 5.30 it follows immediately that K is an integrable op-
erator and, by the preceding case, its kernel K(.,.) E L1 (,u x 71). Now if we
let v = then v is a strictly positive function in L0 (ii), and for each x E E
we have
isI
Tx(t) = v(t) K(s,t)u(s)x(s) dy(s)
for v-almost all t T.
We conclude the section with one comment regarding the cr-finiteness

assumption that has been used throughout this and the previous sections.
There exists a considerably weaker condition (the so-called direct sum
property) that is usually sufficient instead of the a-finiteness. Here is its
definition: A measure space satisfies the direct sum property
if there exists a family of pairwise disjoint sets in such that
0 <00 for all A E A, and for each A E > with finite measure we
can find an at most countable family A0 A and a ,u-null set N E such
that A = UAEAO(AflIIA)UN.
Exercises
1. Let E and C be two Banach function spaces associated with Lo(7r). Show
that E fl C is also a Banach function spacewhich is order dense if both
E and F are order dense ideals. [HINT: If IE and denote the
1
norms on E and C, respectively, then consider the lattice norm Ii on

En C defined by lxii = lix lIE + .1
2. Let C be an ideal in a Riesz space E, let F be a Dedekind complete Riesz

space, and let T: C F be a positive operator. Assuming that T has a
regular extension to all of E, establish the following.
(a) The operator T has a minimal positive extension to all of E. That
is, there exists a positive extension T such that if
8: E F is any positive extension of T, then S.
(b) The minimal extension T is given by the formula
Tc(x)=.sup{Ty:
each x E E+. In particular, Tc(x) = 0 for each x E
for
(c) If T is order continuous, then is also order continuous.
3. By means of Nakano's theorem (see Exercise 22 of Section 1.2), show
directly that if E is a Banach function space associated with a cr-finite
measure ii, then its order continuous dual has weak units.
Abstract Integral Operators 209
4. Let E be a normed Riesz space. Show that separates the points of E

if and only if for each 0 < x E E there exists some 0 0.
5. Verify the uniqueness of the function g in the Representation Theo-
rem
6. Let 71) be a u-finite measure space. Show that = L1(7r).
[HINT: Let E = By Theorem 5.26, E is a Banach function space
associated with Lo(7r). Clearly, Ll(7r) E. Now assume that f E. This
means that E R for each x in In particu'ar,
if x Signf, then x E and If Jd7r = f(w)x(w)d71(w) E R.
This shows that f Ll(-,'r) and = Li(7r).]
7. Let a Banach function space
be a u-finite measure space. If E is
in Lo(7r), then show that the center of E coincides with that is,
2(E) =
8. For an order bounded sequence in a Banach function space with
order continuous norm consider the following convergence properties:
(a) (b) (c) (d)
Show that (a) (b), (c) (d), and (b) (c). Does (d) (b)?
[HINT: Use Exercise 4 in Section 5.1 and Lemma 1.30(c).]
9. If is an order bounded positive sequence in some Banach function
space E in L0(71), then prove that 0 is equivalent to 0,
i.e., to the property: Xn(W)y(W) d7r(w) 0 for each y E
Is this equivalence true if is not assumed to be a positive se-
quence?
10. If E and F are Banach function spaces, then show that is a
principal band in Lr(E, F). [HINT: We can assume that E and F are
order dense ideals. If e and cb are weak units in F and respectively,
then show that =
Prove Corollary 5.31 using Theorem 5.28 instead of Theorem 5.30.
Show that Dunford's Theorem (see Corollary 5.33) is not valid for continu-
ous operators between arbitrary Li-spaces. [HINT: The identity operator
I: L1 [0, 1] L1 [0, 1] is disjoint from every regular integral operator; see
Exercise 8 in Section 5.1.]
13. Assume that (8, ,u), (T, ii), and (Q, 0) are three u-finite mea-
sure spaces and that E, F, and G are ideals in L0(1i), Lo(v), and L0(71),
respective'y. Assume also that in the scheme E F C the operators
S and T are both regular and order continuous.
If either S or T is an integral operator, then show that TS is likewise
an integral operator.
14. Assume that (8, ,u) and (T, ii) are two u-finite measure spaces
with (8, ,ii) non-atomic. Also, suppose that E and F are ideals in
L0(8, and L0(T, ii), respectively, and letS: E F be a regular
integral operator.
Let F0 denote the order ideal in F generated by the range space of

the operator S and assume that there is an operator T: F0 E such
that TS = 'E,the identity operator (in a sense
on E T is a right inverse
of the operator 5). Show that T cannot be an integral operator.
15. Assume that (8, (T, ii), and (Q, 8) are three a-finite mea
sure spaces and that E, F, and G are ideals in L0 (/2), Lo (ii), and L0 (8),
respectively. Suppose also that in the scheme E F G both S and T
are regular integral operators with kernels and T(., '), respectively.
Show that the composition operator R = TS is also an integral
ator whose kernel is given by the convolution of the two kernels S(.,.)
and T(., .), that is,
R(s, q) = / S(s, t)T(t, q) dv(t).

JT
16. Let (8, and (T, ii) be two a-finite measure spaces with (8,
and let 0 < E Lo(,a x 71). Do there exist a function
o < f E Lo(/2) and a subset B E such that:
(a) v(B)>0.
(b) For each t B we have K(s, t) f(s) for /2-almost all s E 8?
17. (BenedekPanzone [54]) Let (8, and (T, ii) be two a-finite
sure spaces. For each p E [1,oo) we denote by x ii) (or simply by
the collection of all functions f Lo x z,) such that
11! Ilp,oo = ess sup If(s, dv(t)] <00.

As usual, the equivalent functions are identified. Similarly, we denote by
x ii) (or simply by the collection of all functions f E Lo(/2 x ii)
such that
dv(t)] dy(s) <00.

= f [f If(s,
The non-negative real numbers and If are called the mixed
norms of f. Show that:
(a) The function is well defined, if f, g E satisfy f = g
/2 x v-a.e., then =
(b) The collections LP700 and (with their mixed norms) are both
Banach function spaces associated with Lo(/2 X v).
(c) If 1 <p, q <oo satisfy + 1, then =
18. The result of this exercise can be found in many monographs; see for
instance [107, 118, 174, 175]. For the notation used here see Exercise 17
above.
Assume that (8, and (T, are two a-finite measure spaces.
Show that a /2 x v-measurable function T(.,.): 8 x T R defines, via
the usual formula
Tx(t) = [T(s, t)x(s) d/2(s),

Js
5. 3. Conditional Expectations and Positive Projections 211
an integral operator from to for some 1 < p < oo if and only

if .) belongs to Moreover, prove that in this case the integra'
operator T determined by the kernel T(. . .) is a regular operator and its
norm is given by IT(.,
19. (Lozanovsky [221]) Dunford's theorem stated in part (2) of Corollary 5.33
asserts that: If (S, ,a) and (T, ii) are cr-finite measure spaces and
1 < p < 00, then every bounded operator T : L1(,a) is a regular
integral operator. Present an alternate proof of Dunford's theorem using
Exercise 18 above.
20. Let E and F be two Banach function spaces with F having order con-
tinuous norm. Show that every regular integral operator from E to F
carries order bounded subsets of E to norm totally bounded subsets of
F. That is, show that if x E E+, then for each regular integral operator
T: E F the set T[O, x] is norm totally bounded. [HINT: Use Exercise 9
of Section 2.3.]
5.3. Conditional Expectations and Positive Projections

In this section >, 1a) will denote a fixed probability measure space. The
constant function one on will be denoted by 1. For each 1 p oo the
familiar Dedekind complete (real) Banach lattice >, 1a) will also be
denoted by As usual, we view the equivalence classes as functions.
If A is a cr-subalgebra of > and 1 p oo, then by we shall
denote the Dedekind complete Banach lattice A, This Banach lat-
tice consists of the equivalences classes of all A-measurable functions whose
absolute powers are Also, (A) consists of all essen-
tially bounded A-measurable functions. That is, an function
f: R belongs to if and only if <oo, where
If = inf {M >0: bL({w E ft 11(w)! > M}) = o}.
Whenever the cr-subalgebra under consideration is clear from the con-
text, we will simply write instead of if or 111 Clearly, each
is a vector sublattice of If 1 p <oo and f then
it should be obvious that the In-norm of f in
.
f
coincides with the
I
lIp-norm of f in The same conclusion is true for p cc. That is,
Ill = for each function I Thus, for each 1 ) containing
the constant functions. We shall see later in this section that every closed
vector sublattice of an that includes the function 1 is of this
type.
Before proceeding further, let us make one more cautionary comment
about the Banach lattice where A is a cr-subalgebra of >1 If we
consider the probability measure space A, p) in its own right, then we
2 12 5. Integral Operators
can think of (A) as the usual associated with the measure space
( A, ,u). On the other hand, if we consider as a vector sublattice
of then we view the classes of as classes in That is,
we tacitly identify (A) with its image in under the lattice isometry
[hA The following example clarifies the situation. Let
. = [0, 1],
let be the of all Lebesgue measurable subsets of [0, 1], and let
,u be the Lebesgue measure. Consider also the a-subalgebra of given by
A = {, [0, 1], [0, 1] }. Then it is easy to see that
= {clX[OI} + C2 E
and so each class of (A) consists of one function. However, if we consider

(A) as a vector sublattice of then we can view the classes in (A)
as classes in To clarify it better, consider the function I =
If we consider as an in its own right, then the function f
is identified with the class [hA = {f} in (A). On the other hand, if
we consider as a vector sublattice of then the function f is
identified with the class [f}E = {g E g = ,u-a.e.} in f
Now fix a o--subalgebra A of By the classical RadonNikodym theo-
rem for each f there exists a unique function h E Li(A) such that
IA h d1ii = fA f for each A E A. The function h is called the
tional expectation of f with respect to A and is denoted 5(1 IA). That
is, 5(1 IA) is a ,u-a.e. uniquely determined A-measurable and ,u-integrable
function satisfying the identity
for each A E A. Clearly, if f, g E then for all 18 E JR we have

E(a'f + = +
That is, E(IA) defines an operator on whose range is Li(A).
Definition 5.36 (Kolrnogorov [187]). The operator E('IA):

is called the conditional expectation operator induced by A.
The proof of the next result is straightforward and is omitted.

Lemma 5.37. For every cr-subalgebra A of the conditional expectation
operator E(. JA): L1 L1 is a strictly positive order continuous
traCtive projection with range L1(A) that leaves invariant. That is..
(1) 1ff E and f> 0, then E(f A) > Li(A).
0 in
(2) If 0 in then 0 in Li(A).7
7RecalI (see Theorem 1 80 and the paragraph thereafter) that since we deal here with a finite
measure space, each ideal E in Lo(E) satisfies the countable sup Accordingly, a positive
operator T: E + E is order continuous if and only if 0 implies 0.
(3) The operator has norm one, i.e., 1IE('IA)IIi = 1.

(4) If g E Li(A), then = g, and so
E(1A)2=E('JA) and
(5) The operator E('IA) leaves invariant and = 1.
Not only do the conditional expectation operators leave invariant,

but they also leave invariant all the
Lemma 5.38. For each u-subalgebra A of and for each I < p < oo
the conditional expectation operator leaves invariant and is
contractive on this space.
Proof. From the definition of the conditional expectation operator and from
Lemma 5.37 it follows that we need to consider only I 1 Take any non-negative function f and apply to
it Jensen's inequality for conditional expectation operators with the convex
function = see Exercise 5(b) at the end of the section. This gives
o (t)
Since fP E L1(>J), the function belongs to Li(A) There-
fore, from (t) we get E L1(>J), that is, E(f[A) E This
proves that E(.[A) leaves invariant.
Now, according to the definition of the conditional expectation we have

d,a
= fP dp, and so
=
=
Therefore, E(. A): is a contraction, as claimed.
Following the work of G. Birkhoff [63], we shall say that a bounded
operator T: L1(>J) Li(>J) satisfies the averaging property (or that
the operator T is an averaging operator) if for each pair h, f
such that hTf E L1 [in particular, this is so whenever h E (>J) and
f we have
T(hTf) =ThTf.
Conditional expectations enjoy the averaging property.
8 Positive projections leaving the function 1 fixed are sometimes called Markov projections.
This is in agreement with our Definition 3.15 of a Markov operator
Theorem 5.39. Every conditional expectation operator is an averaging op-

erator. That is, if A is any ci-subalgebra of and fE satisfy
hE(f A) E L1(E), then
E(hE(fIA)IA) = E(hIA)E(fIA).
Proof. In view of the order (or norm) continuity of E(.IA) it is enough to
verify that
= (*)
for each B, C Since IA) is an essentially bounded non-negative
A-measurable function, there exists a sequence of non-negative A-step
functions satisfying IA). This and the fact that A) = for
each n imply at once that in order to prove (*), it suffices to establish that
E(XBxDIA) = &(xBIA)xD
for each B E and each D E A. To do so, we just have to prove (by the
uniqueness of the derivative) that
I
JA JA
for each A E A or, equivalently,
/
JAnD
XB = f
JAnD
E(xB IA)
Since A fl D E A, the last equality holds true automatically by virtue of the

definition of E(jA), and the proof is finished.
Our next objective is to establish that the properties of the conditional

expectation operators in Lemma 5.37 basically characterize the Markov pro-
jections not only on L1 but also on a much broader class of sub-
spaces of
We begin with a characterization of the order closed vector sublattices
of Lo(1a) containing the constant function 1.
Lemma 5.40. If E is an ideal in containing the constants, then for a
vector sublattice Y of E with 1 E Y the following statements are equivalent.
(1) Y = E fl Lo(A) for some ci-subalgebra A of E.
(2) Y is order closed in P2.
(3) Y is closed under ,a-a. e. pointwise monotone sequential limits in P2.
Proof. The implications (1) (2) (3) are trivial. Hence, only the
implication (3) (1) needs a proof. So, assume that (3) is true. Consider
the collection of sets
A={AETh XA"}
We claim that A is a a-subalgebra of Using the identities
XACIXA,
we conclude at once that A is a subalgebra of E. To see that A is a a-algebra,
assume that some sequence A satisfies I A. This certainly
implies that A E and XA E E, our condition (3)
guarantees that XA E Y. Thus, A E A, and so A is a a-subalgebra of
Since the vector sublattice of all A-step functions is included in Y and
is order dense in E fl L0 (A), it follows from (3) that E fl L0(A) C Y. Now
let f Y and consider the set A {w E f(w) > Since Y is a
vector sublattice and 1 E Y, it follows that A 1} Y and moreover
mf+ A I I Using once more (3), we see that Y and so A E A.
Finally, note that since I ci E Y for each scalar c, the preceding case
implies
{w E 11: f(w) > c} = {w E ft [1 cl](w) > 0} E A.

So, f is A-measurable and hence f Lo(A). This implies Y c E n Lo(A).
Consequently, Y = E fl L0(A), as desired.
In a Banach lattice with order continuous norm, a vector sublattice is

order closed if and only if it is norm closed. In the absence of the order conti-
nuity of the norm, the assumption of the order closedness of Y in Lemma 5.40
cannot be replaced with the norm closedness of Y; see Exercise 6 at the end
of the section.
An algebraic analogue of Lemma 5.40 for is as follows.
Corollary 5.41. For a subalgebra Y of the Banach algebra the fol-

lowing statements are equivalent.
(1) Y = for some a-subalgebra A of
(2) Y contains the constants and is order closed in
Proof. It should be clear that (1) implies (2). Now assume that (2) is
true. Since norm convergence in implies order convergence, it easily
follows that Y is a normed closed subalgebra of But then it is well
known from the proof of the StoneWeierstrass theorem (see for instance [31,
Theorem 11.5, p. 89]) that Y is a vector sublattice. The validity of (1) now
follows from Lemma 5.40. 0
The next result presents two basic order properties of contractive oper-
ators on
Lemma 5.42. For a contractive operator T: k 1 p 00,

satisfying Ti = 1 we have the following:
(a) 1f1_<p<oo, then T*1=1.
(b) If p = 1 or p = 00, then T is a positive operator.
Proof. (a) Let q be the conjugate index of p, i.e., + 1. Since T
is contractive, T* is also contractive, and consequently 1. From
Ti = 1 and Holder's inequality, we obtain
1 (1, i) = (Ti, i) = (1,T*1) IliIfp. 1.
So, in this case, Holder's inequality is the equality
= = = 1.
This guarantees the existence of some constant c> 0 such that = ci;
see for instance [32, Problem 31.4]. Clearly, c = 1, and from this and (**)
we get T*1 = 1.
(b)Consider first the case p = oo. Pick any 0 x e Dividing
by if necessary, we can assume that 0 x $ 1, and 50 0 1 x 1.
Taking into account that = 1 and that the closed unit ball of
is the order interval [1, 1], it follows that 1 Tx = T(1 x) e [1, 1].
Therefore, 1 Tx 1 and so Tx 0. This shows that T is positive.

Let now p = 1. It follows from (a) that T* 1 = 1. Therefore, by the
previous case the adjoint operator T*: is positive. This
implies that the second adjoint T** of T is also a positive operator. Since
T** agrees with T on L1 (E), it follows that T itself is positive too.
Exercise 7 at the end of the section shows that for other values of p
the conclusions of the preceding lemma do not hold in general. Recall that
a Banach lattice E is said to have strictly monotone norm if xl <
implies lix 11 < IIylI. The Lu-spaces with 1 p < 00 have strictly monotone
norms.
Lemma 5.43. Let T: E * E be an operator on a vector lattice and assume
that, either
(a) T is a strictly positive projection, or
(b) E is a Banach lattice with strictly monotone norm and T is a con-
tractive positive operator.
Then the fixed space 2(T) = {x E: Tx = x} of the operator T is a
e
vector sublattice of E (which, in case (b), is also norm closed).
Conditional Expectations' and Positive Projections 217
Proof. (a) To begin with, notice that since T is a projection, the fixed space
of T simply coincides with the range T(E) of T.
Now take any x E T(E). Then lxi TIxj, and so TIxI lxi 0.
Since T2 = T, it follows that jxj) = 0, and so the strict positivity
of T implies xl, that is, lxi E Therefore, = T(E) is a
vector sublattice of E.
(b) Clearly, 2(T) is a closed vector subspace of E. To see that

is a vector sublattice, pick any x E J(T). Then lxi = Tx Ti xl. If
lxi <Tixi were true, then the strict monotonicity of the norm would imply
< and so 11Th > 1, a contradiction. Hence lxi = i.e.,
lxi E J(T). This shows that is a vector sublattice of E.
Our next result is essentially due to J. L. Kelley [180], G. L. Seever [299],

and D. E. Wulbert [341]. It asserts that any contractive projection on
with range LOC(A) has the averaging property.
Lemma 5.44 (KelleySeever--Wulbert). If A is a ci-subalgebra of then

any contractive projection P: with range is an
averaging operator, i.e., it satisfies
P(fP(g)) = P(f)P(g)
for each f,g E
Proof. Since I E we have P1 = I and hence 11PM = 1. By

Lemma 5.42(b) the operator P is positive.
Since Px = x for each x E and since the step

functions are norm dense in (A), to prove the desired averaging property
it suffices to show that P(fx3 = (Pf)XA for all f E and A E A.
So, let f E and AE A. Without loss of generality we can
assume that the function f satisfies 0 f 1. Then 0 fxA and
hence 0 P(fXA) = Similarly, 0 P(fXAC) Therefore,
Supp P(fXA) C A and Supp P(fxAC) AC.
These inclusions should, of course, be understood to be valid ,a-a.e. with

respect to A. For instance, the inclusion Supp P(fXA) A means that the
set D = {w E w $ A} belongs to A and ,a(A) = 0. Since
+ fxAC f, it follows that
P(fXA) + = Pf = (Pf)XA + (Pf)XAC,
and from this and (t) we get P(fXA) = (Pf)XA, as desired.

It is worthwhile pointing out that Lemma combined with the or-

der continuity of the conditional expectation operator provides an alternate
proof of Theorem 5.39.
We are now ready to present a characterization due to A. A. Mekler [236]
of the order continuous positive projections that preserve 1. The charac-
terization was also proven independently by P. G. Dodds, C. B. Huijsmans,
and B. de Pagter in [110]. Recall that (by Theorem 5.26) for any ideal E in
(E) its order continuous dual can be identified with an ideal in Lo
as follows:
In particular, notice that:

o If 1 e E, i.e., if E, then
Ec then c c
In the sequel, keep in mind that an "ideal E in (E) containing the
constants" is simply any ideal E in Lu satisfying E c L1 (E).
Theorem 5.45 (Mekier). If an ideal E in L1(E) contains the constants,
then for an operator F: E E the following statements are equivalent.
(1) P is an order continuous positive projection, its range is a vector
sublattice of E, and P1 = 1.
(2) There exist a A of E and a non-negative function
in such that = 1 and Px = E E.
Proof. (1) The order continuity of the projection P implies imm&
(2)
diately that the vector sublattice Y = F(E) of E satisfies statement (3) of
Lemma 5.40. So, according to this lemma, there exists a A of
such that Y = Efl Lo(A). Since E L1(E), we have Y = En L1(A).
Next, we shall consider two cases.
CASE I: E =
In this case we have Y = fl L1(A) = (Why?) By
Lemma 5.44 the operator P has the averaging property on i.e.,
P(xiP(x2)) P(xi)P(x2) for each x1,x2 e
Now consider the positive linear functional R defined by
f
= P(x) dii. The order continuity of the operator P implies that is
order continuous, that is, e = L1 (E). Therefore, there exists a
non-negative function L1(E) such that =
In particular, for each A e A and each x e E we have
fd1i for each x e E.
IA = f = =
f P(xxj (*)
Conditional Expectations and Positive Projections 219
Using that P satisfies the averaging property and that = XA7 we get
P(xxA) = P(XPXA) = P(x)P(xA) = P(x)XA
for each x E E. From this and (*), we obtain
f
JA J
[ JA
Since Px is and for each x E E, it follows from
the definition of the conditional expectation that
E E. In particular, E(q5[A) = = P1 = 1. This establishes the
desired representation of P in the special case when E
CASE II: E is an ideal such that (E) E L1 (E).
The hypotheses that P is positive and P1 = 1 guarantee that P leaves
invariant and = 1. So, if denotes the restriction of P to
then is a contractive projection on satisfying = 1.
We claim that the range of is a vector sublattice of (s). To see
this, let x, y and put z PxVPy E Since the range of the
operator P: E E is a vector sublattice of E, there exists some u E E such
that Pu = z. But then the function z E satisfies z Pz
This shows that the range of is a vector sublattice of
According to Case I, there exist a a-subalgebra A of and a non-
negative function q5 E such that = E
This means that the desired representation
Px E(bxjA)
is valid for each x E Since is an order dense ideal in L1(>), it
follows that is also an order dense ideal in E. Now use the fact that
P and are both order continuous on E to infer that Px =
E E.
The obtained representation implies, in particular, that for each x E E
the function q5x is integrable. This means that q5 E Eq'.
(2) Assume that there exist a a-subalgebra A of and a non-
(1)
negative function q5 such that = 1 and Px
E E. In particular, notice that = q51 E Lemma 5.37
guarantees that P is a positive order continuous operator. We also have
P1 = = = 1 by the condition on
Let us verify that P is a projectiom Keeping in mind that E(' IA) satisfies
the averaging property, for each x E E we obtain
P2x =
= = = Px.
It remains to verify that Y = P(E), the range of P, is a vector sublattice

of E. To this end, let y E Y and then choose a vector x E E such that
y = Px = Let A = {w E y(w) O}. Then A E A because y is
A-measurable, and so S(XA A) = Notice also that = Using this
and the averaging property of the operator e(.IA) once more, we obtain
YXA = JA) = S(q5xS(XAIA)IA)
= S(q5XxAIA) = P(XXA).
Since the function belongs to E, it follows that E This shows
that 1' is a vector sublattice of E, and the proof is finished.
Theorem 5A5 has a number of important consequences.
Corollary 5.46. Let E be an ideal in L1(>) containing the constants. For
an operator P: E E the following conditions are equivalent.
(1) P is an order continuous strictly positive projection and P1 = 1.
(2) There exist a o--subalgebra A of > and a strictly positive function
q5 E such that S(q5(A) = 1 and Px = S(q5xlA) for each x E E.
Proof. Only the implication (1) (2) needs a proof. Since P is strictly
positive, it follows from Lemma 5.43(a) that the range Y = P(E) is a
vector sublattice of E. Therefore, Theorem 5.45 applies and guarantees the
existence of a a non-negative function q5 E such
that S(q5JA) = 1 and Px = S(q5xIA) for each x E E. To see that b is strictly
positive, let A e E with > 0. Then 0 < XA E E, and by the strict
positivity of P we get > 0. Therefore,
IA =
holds for all A E E with
f = f [A)
= f PXA >0
> 0. This guarantees that q5 is a strictly

positive function.
Corollary 5.47. For an operator P: where 1 p < 00,
(1) P is a strictly positive projection such that P1 = 1.
(2) There exist a A of > and a strictly positive function q5
in Lq(>), where = 1, such that = 1 andPx =
for each x E
Proof. The equivalence of these two statements follows immediately from
Corollary 5.46 by observing that the order continuity of the norm implies
that every positive operator on any with 1 p < oo is automatically
order continuous. D
Let E be an ideal in L1(E) containing the constants. Then we know

that L1 (E). This means that we can consider (E) as an
ideal in the order dual With this observation in mind, we can state the
following result.
Lemma 5.48. For an ideal E in L1(E) containing the constants and a
positive operator T: E E we have the following.
(1) The order adjoint of T fixes 1, i.e., = 1, if and only if for

each x E E we have fTxd1i = fxd,u.
(2) If T is strictly positive and order continuous.
Proof. (1) Assume first that I = I. Then for each x E E we have
= (x,I) = = (Tx,I)
f
For the converse, assume that f x d1i = Tx d1i holds for each x E E.
Rewriting this identity in duality form and taking into account that I e
we see that
(x,1) = (Tx,1) =
(x, I I) = 0 for each x E E. This implies I = I.
(2) We show first that T is strictly positive. To this end, let 0 < x E E.
From part (1), we have f Tx = f x dpi> 0. This implies Tx > 0, and so
T is strictly positive.
For the order continuity of T, assume 0 in E. This implies 0
in L1(E), and so from part (1) we get
/
Jci
t,. 0 ,u-a.e. Consequently,
T is order continuous.
We now present the definition of a bistochastic projection.
Definition 5.49. Let E be an ideal in L1 (E) containing the constants. A
positive projection P: E B is said to be bistochastic if both P and its
order adjoint operator are Markov, that is, P1 = I and
From Lemma 5.48 we have the following.

Corollary 5.50. Every bistochastic projection is strictly positive and order
continuous.
We continue with a few more representations of projections my means
of conditional expectation operators.
Corollary 5.51. Let E be an ideal in Li(E) containing the constants. For

an operator F: E + E the following statements are equivalent.
(1) P is a bistochastic projection.
(2) There exists a unique cr-subalgebra A of such that Px = E(xJA)
for each x E E.
Proof. Again only the implication (1) (2) is non-trivial. Assume that
P is a bistochastic projection. By Corollary 5.50, we know that P is strictly
positive and order continuous. Therefore, in view of Theorem 5.45, there
exist a u-subalgebra A of E and a function q5 E c L1 (E)
such that E(q5JA) = 1 and Px = for each x E E. The only thing
left to verify is that q5 = 1.
Assume to the contrary that q5 1. Since = 1, it follows that
there exist a constant > I and a set B E E with ,u(B) > 0 such that
c/5XB > 'YXB Using that P a bistochastic operator and the definition of
is
the conditional expectation operator e(.IA), we obtain that
=
which is a contradiction. Therefore, q5 1 and so P = e(.IA).
If P is a contractive projection on L1(E) satisfying P1 = then by 1,
Lemma 5.42(a) the projection P is bistochastic, and so the preceding corol-
lary immediately yields the classical result of R. G. Douglas [112].
Corollary 5.52 (Douglas). For an operator P: Li(E) L1(E) the follow.-
ing statements are equivalent.
(1) P is a contractive projection satisfying P1 = 1.
(2) P is a conditional expectation operator. That is, there exists a
unique A of E such that P = e(.jA).
Corollary 5.52 was generalized to for I <p < oo with p 2
by T. And Under the assumption that the contractive projection is
positive the restriction p 2 is redundant. We present below a very simple
proof of this result.
Corollary 5.53 (And). For a positive operator P: where
I <p < 00, the following statements are equivalent.
(1) P is a contractive projection satisfying P1 = 1.
(2) P is a conditional expectation operator. That is, there exists a

cr-subalgebra A of E satisfying Px = E(xfA) for each x E
Conditional Expectations and Positive Projections 223
Proof. Notice that for p = 1 the result is simply Douglas' result (Corol-
lary 5.52). So, we can assume that 1 < p < oo.
(
1) The projection P is certainly order continuous since the
(2)
norm on is order continuous. In view of Lemma 5.43(b), the range
of P is a vector sublattice of since the operator P is contractive
and the norm on is strictly monotone. Hence, by Theorem 545, the
representation Px = holds. We want to show that = 1. For
this to be true it suffices to know that P*1 = 1. But this is guaranteed by
Lemma
(2) (1) The implication was established in Lemma 5.38.
Regarding the contractive projections on some historical com-

ments are in order. A version of Douglas' theorem (Corollary for
positive operators appeared in the book by J. Neveu [256, Exercise IV.3.1,
p. 123] that was published in 1963. Several other versions of Corollary 5.52
(with different assumptions about the operator) go back even earlier, namely,
to the papers by S-T. C. Moy [246] and G. C. Rota [285]. As we have already
mentioned, Douglas' theorem was generalized to Lu-spaces by T. And [34].
Subsequently, the description of positive projections on Banach function
spaces via conditional expectations was studied by many authors; for details
see [59, 110, 236, 237] and the references therein. The proofs presented
here are adaptations and refinements of those in [9, 110, 236, 237]. It
seems that C. Moy [2461 was the first who used the conditional expec-
tations for representations of some operators that were not assumed to be
projections. Some considerable generalizations of his results were obtained
by R G. Dodds, C. B. Ruijsmans and B. de Pagter [110]; see for instance
Exercise 10 at the end of the section. For a current survey on contractive
projections-on Banach spaces see the article by B. Randrianantoanina [274].
Our next comment relates some of the results obtained in this sections
with the interpolation of operators on Banach function spaces. For this
comment we need some terminology from the theory of Banach function
spaces.
o Two functions x, y E Lo(,u) are said to be equirneasurable if they
generate the same distribution function, if the equality
ft x(w) <t}) = E ft y(w) t})
for each t R.
holds
o A Banach function space E associated with some is
called rearrangement invariant if every y E L0 that is
to some x E E belongs to E and = lxii.
0 A rearrangement invariant Banach function space E is called an

interpolation space if each bounded operator on L1 that leaves
invariant and is bounded on also leaves E invariant
and is bounded on E.
Regarding, interpolation spaces we have the following classical result
known as the RieszThorin theorem [56, Theorem 1.1.1].
Theorem 554 (RieszThorin). Each is an interpolation space.
The RieszThorin theorem provides, of course, another proof of Lem-

ma 5.38, according to which the conditional expectation operator
leaves each invariant, that is, IA) for each
A of it is important to point out that this is not
However,
true for an arbitrary ideal E in For some algebras A the inclusion
E may hold and for some it may not. Remarkably, as was proved
C
by A. A. Mekler [238}, if E
E is an interpolation space.
We continue with two examples showing that all conditions in the pre-
ceding corollaries are essential. Our first example shows that without the
assumption P1 = 1 the conclusion of Theorem 5.45 is no longer true.
Example 555. Let E = 1], where [0, 1] is equipped with the Lebesgue
L1[O,
measure. We define an operator P: L1 [0, 1] L1 [0, 1] by letting
Px =
A direct verification shows thatpositive

a strictly P is projection on
L1[0, 1] with P1 = and range 1]. We claim that this projection
P does not have a representation of the form Px = for some in

= 1] and some A
To see this, assume contrary to our claim that Px = for some
e [0, 1] and some A This would imply (see Exer-
cise 9 at the end of the section) that were a characteristic function.
However, = P1 = a contradiction. Therefore, P cannot
have a representation of the form Px =
On the other hand, if we define Q: L1[0, 1] L1[O, 1] by Qx =
then Q is also a positive projection with range L1 1] satisfying Q1 =
Moreover, it is easy to see that if
A= { [0, U A: A is a Lebesgue measurable subset of 1] }
and = then Qx = for each x E E.

Our next example shows that in general a 1\'Iarkov projection can easily
fail to be contractive and its range can fail to be a vector sublattice.
Example 5.56. We shall establish the existence of a positive projection
P: L1[O, 1] L1[O, 1] such that:
(1) P1=1.
(2) >1.
(3) The range of P is not a vector sub1attice of L1 [0, 1].
Consider the sets A B= and C = and let A be

the cr-subalgebra of the Lebesgue a-algebra >J of [0, 1] generated by these
three sets. It follows that
= {c1XA +
Li(A) C2XB + c3 E R },
and the positive operator Li(>2) is given by
= 3[f x(s) ds]XA + 3[f x(s) ds]XB + 3[fx(s)
and has L1 (A) as its range. (See Exercise 4 at the end of the section.)
Next, define Q: Li(A) Li(A) by
Q(c1XA + c2XB + = c1XA + c2XB +

and note that Q is a positive projection satisfying Q1 = 1.
If we define P: L1[0, 1] L1[0, 1] by P = P is a positive

then
projection satisfying P1 = 1. In particular, P is a Markov projection. Now
notice that if x = then IfxfIi = 1, = 3XA and Px = 3XA +
This implies tIPxIIi = > 1, and so > 1 holds.
To see that the range of P is not a vector sublattice, notice that the two
functionsy1 3XA andy2 = XB satisfy Py1 = and Py2
This implies Py1 V Py2 = 3XA + XB + R(P), and so the range of P is
not a vector sublattice of L1[O,
It is worth noting that for a Markov projection on L1 [0, 1] condition

(3) in the preceding example is in general stronger than condition (2); see
Exercise 13 at the end of the section.
We conclude the section with one more important property of the con-
ditional expectation operator asserting that the operator S('IA) is basically
seif-adjoint.
Theorem 5.57. For each a-subalgebra A of the cr-algebra the adjoint
operator S(.IA)*: satisfies
E That is, coincides with the restriction of the conditional
ezpectation operator S('IA) to the space
Proof. In this proofthe projection : will be denoted

for brevity by P. Recall that P is bistochastic, that is, = 1.
The operator P*: > is also a bistochastic projection on
and therefore by Corollary 5.51 there exists a u-subalgebra 13 of
such that for each y e we have = That is, on the space
we have the two (positive order continuous) bistochastic projections
Obviously, the projections E(.IA) and E('IB) can be assumed acting on

L2 To complete the proof, it suffices to show that =
for each y e Since E(JA) and E('IB) are projections on it
is enough to prove that and have the same range and the
same kernel. Since (by Lemma 5.38) both projections and
are contractions on it follows that E(.jA) and are orthogonal
projections; see Exercise 14 at the end of the section. In particular, if
E('IA) and have the same range, then they automatically have the
same kernelwhich is the orthogonal complement of the common range.
Consequently, E(' A) = E(. if and only if E (. IA) and E have the same
range as projections on To verify the latter, it suffices to establish
that XA E L2(B) for each A e A and XB E L2(A) for each B E 13.
Let A e A. Using the definition of the adjoint operator and the fact
that PXA XA, we have:
(E (XA 13), XA) = XA, XA) = (XA, PXA) = (XA, XA) = (A).
At the same time, using Holder's inequality, we obtain
(E(xA113),xA) = (P*XX) IIXA III

IIXAMI = 1u(A).
That is, we have obtained the equality in the Holder inequality, and therefore
the function E(xA 13) must be equal to 1 on the set A. Similarly, we can show
that E(XAC 13) must be equal to 1 on the set Since both functions E(XA jB)
and E(XAC 13) are non-negative and E(XA [B) + E(XAc = E(1113) = 1, we
can conclude that XA = e(XA 13) e L2(13).
Next, take any B e 13. We have
(PXB,XB) = (xB,P*xB) = (XB,E(xB113)) =

Using Holder's inequality once again, we obtain
(PXB,XB) IX = IIPXB III MXBIII = 1u(B).
As above, this implies XB = e L2(A). The proof is finished.

p.3. Conditional Expectations and PositiveProjections 227
Exercises
1. Let E, p) be a probability space, and assume that A is a usubalgebra
of E. Show that for each function f we have = 1ff
2. This exercise shows that every Banach lattice with order continuous norm
and a weak unit can be represented by a Banach function space that
"lives" between and L1.
Let E be a Banach lattice with order continuous norm and a weak
unit e. Show that there exists a probability measure space E, p) and
an ideal F in L1(E) containing the constants, i.e., F L1(p),
and a surjective lattice isomorphism T: E + F such that T(e) = 1.
[HINT: Let E JR be a strictly positive 'inear functional satisfying
cb(e) =1. (Since E has order continuous norm, such a functional always
exists;see [30, Theorem 12.14, p. 183].) Now consider the L-norm . J
on E defined by Then E is an ideal in the norm completion

E of (E, and e is a weak unit in E. Finally, use the Kakutani
BohnenblustNakano Representation Theorem 3.5 to identify E with an
L1 (p)-space for some probability measure space E, p) with e corre-
sponding to the function 1. This implies that (p) E L1 (p).]
4. Assume that {IIi}iEI is an at most countable partition of a set That
is, I is at most countable, 0 for each i E I, n = for i
and = Establish the following.
(a) The u-algebra generated by the family is given by
Jci}.
(b) A function f: JR is A-measurable if and only if it is of the form
I =
(c) If is at most countable, then every u-algebra of subsets of is
generated by a unique at most countable partition of ft
(d) Now suppose that E, p) is a probability measure space and that
{ is an at most countable measurable partition of i.e, be-
sides being a partition of we also have E for each
iEI.
have
E(f IA) =
iEI
5. For an arbitrary convex function JR > JR, any function f
E E establish the following
inequalitiesknown as Jensen's
(a) fcbofdp.
(b) A)) E('iL' o in L1(A).
6. Give an example of an ideal E in L1(E) such that E is a Banach lattice
and contains a non-trivial norm closed order dense ideal in E.
7. Show that the conclusions of Lemma 5.42 do not hold in general for the
remaining values of the indices.
8. Show that the conditional expectation operator satisfies the follow-

ing property: If a function f E L'(E) has its support in some set A A,
them the function also has its support in A.
9. (DoddsHuijsmans--de Pagter [110]) Suppose that E is an ideal in L1 (E)

containing the constants, and let 0 b E Assume also that for
some a-subalgebra A of E the operator P: L1(E) L1(E), defined by
Px E(cbxlA), leaves E invariant.
Show that P is a projection if and only if there exists a subset A E A
such that Supp b A and E(cbJA) in L1(A).
10. (Moy [246]; DoddsHuijsmansde Pagter [110]) This exercise is a

ification of Theorem 5.45 for linear operators that are not necessarily
projections. Let E be an ideal in L1(E) containing the constants, and let
T: E E be a linear operator such that:
(a) If in E, then
(b) T leaves invariant.
(c) For each h E (E) and each g E E we have T(hTg) = ThTg.
Then there exists a A of E and a function b E so that
Tx = E E.
11. Give an example of a Markov projection on L1[0, 1] that is not bistochas-
tic. Show that such a projection cannot be contractive.
12. For a vector subspace Y of some with 1 p < oo establish the
following.
(a) If Y is a closed vector sublattice, then Y has a weak unit.
(b) Y is a closed vector sublattice if and only if it is the range of a
contractive positive projection on
13. According to the previous exercise, if a Markov projection P on iome
(E) with 1 p <oo is contractive, then its range is a vector sublattice
of Show that the converse is not true. That is, give an example of
a Markov projection P on such that its range is a vector sublattice
of but 1IPII > 1.
14. Show that a projection on a Hilbert space is an orthogonal projection if
and only if it is a contraction.
15. Use Theorem 5.57 to obtain an alternative proof of Lemma 5.38. That
is, show that for each o--subalgebra A of E and for each 1 p oo the
operator leaves the space invariant and is contractive on
this space.
5.4. Positive Projections and Lattice-subspaces

In spite of the somewhat "negative" nature of Examples 5.55 and 5.56, there
is a considerable amount of interesting information that can be obtained
about general positive projections and their ranges. This will be done in
this section. As our starting point, we shall use Example 5.56. Though the
range Y of the projection Q is not a vector sublattice of X, nevertheless, Y
5.4. Positive Projections and Lattice-subspaces 229
considered in its own right is a vector lattice! This special feature possessed
by certain subspaces is formalized in the next definition.
Definition 5.58. A vector subspace Y of an ordered vector space X is said
to be a laUice-subspace if Y equipped with the ordering induced by X is a
vector lattice. Equivalently, Y is a lattice-subspace of X if Y equipped with
the cone Y+ = Y fl is a vector lattice.
If Y is a lattice-subspace of an ordered vector space X, then the lattice
operations in the vector lattice (Y, will be denoted by V and A (or by
Vy and Ay, respectively). That is, if yi, E Y, then the supremum of the
set {y1,y2} in the vector lattice (Y,Y) is denoted by ylVy2 or Yl
Similarly, the infimum of the set {yi, Y2} in the vector lattice (Y, 4) is
denoted by ylAy2 or y Y will be
denoted by
It should be obvious that a vector sublattice of a vector lattice is auto-
matically a lattice-subspace. The converse is not true in general; namely,
the class of lattice-subspaces is considerably larger than the class of vector
sublattices, and so, an arbitrary lattice-subspace need not be a vector sub-
lattice. For instance, if we consider the Riesz space E = C[O, 1] and let Y
be the vector subspace of E consisting of all linear functions, i.e., Y consists
of all functions of the form y(t) = mt + b, then Y is a lattice-subspace of E
that fails to be a vector sublattice.
The next result is essentially due to H. H. Schaefer [290, p. 214] and it

reveals several order properties of the range of a positive projection.
Theorem 5.59 (Schaefer). The range F of a positive projection F: E E
on a vector lattice satisfies the following properties.
(1) F is a lattice-subspace of E and its lattice operations are given by
xVy = P(xVy), xAy = P(xAy), and XIF =
(2) If e is a strong unit in E, then Fe is a strong unit in F.
(3) If E is Dedekind complete, then F is Dedekind complete.
(4) If P is strictly positive, then F is a vector sublattice of E.
(5) If E is a normed lattice, then the norm
HxIEl = = XE F,
is a lattice norm on F and lxii Ilixill for each x E F.
(6) If E is a Banach lattice, then the restriction of Ito F is equivalent
to .
and consequently (F, t II I 1)
is a Banach lattice.
ProofS (1) Clearly, F is a cone. We must show that this cone
induces a lattice ordering on F. We shall prove only the supremum formula.
230 - 5. Integral Operators
Tothisend, pickanyx,yEF. ThenfromxxVyandy xVywe

see that x = Fx P(x V y) and y F(x V y). That is, F(x V y) is an upper
bound in F for the set {x, y}. To see that this is the least upper bound,
assume x z and y z for some z E F. Then x V y z, and the positivity
of F implies F(x V y) Pz = z. Hence, xVy = P(x V y) in F.
(2) Assume that some element e E E is a strong unit, and take any
x E F. There exists A > 0 such that x Ae. The positivity of F implies
that x = Px APe. This shows that Fe is a strong unit in F.
(3) Assume that B is Dedekind complete and let 0
x in B. Clearly, Px PXa =
for each a, that is, Px in a lower bound of in F. On the other hand,
if y holds in F for some y E F and each a, then x y in B, and so
Px Py = y. Therefore, Wa j Px in F and this shows that F is Dedekind
complete.
(4) This follows from Lemma 5.43(b).
(5) Clearly, IXIF in F implies = IUXIFII =
Consequently, isf
a lattice norm on F. Now if x e F, then we have
IxI = PIxI = IXIF, and so IIIXIFII = HxW.

(6) If B is a Banach lattice, then the positivity of P guarantees that P
is norm continuous, and hence the projection P has closed range. From (5),
for each x E F we have
= .
which means that the two norms IH and are equivalent on F. In

particular, (F, II is a Banach lattice.
A vector subspace Y of an ordered vector space E is said to be posi-

tively complemented if there exists a positive projection on B with range
Y. In this terminology, the main conclusion of Theorem 5.59 states that:
Every positively complemented subspace of a vector lattice is a

lattice-subspace.
The next result, due to T. And (personal communication) and inde-
pendently to N. Ghoussoub [135, Lemma 111.3], shows that the range of
a positive projection under some mild conditions is lattice isomorphic to a
vector sublattice.
Theorem 5.60. 1fF: B B is an order continuous positive projection
on a Dedekind complete vector lattice, then theTe exists a positively com-
plemented vector sublattice F of B that is order isomorphic to the lattice-
subspace P(E).
5.4.Positive Projections and Lattice-subspaces 231
If E is also a Banach lattice, then the vector sublattice F is closed.

Proof. We know that the range Y = P(E) is a latticesubspace of E and
we claim that Y is order isomorphic to a positively complemented vector
sublattice of E. Consider first N = {x E E: Pixi = o}, the null ideal
of P. Since P is order continuous, N is a band in E. Hence, E = N
Nd. We denote by Q the band projection from E onto Nd. Notice that if
Now let P1 = QP. Clearly, P1 0 and

P12 = (QP)(QP) = Q(PQ)P = QPP = QP = P1.
That is, Pi is a positive projection on E. Obviously, P1 leaves Nd invariant.

We claim that P1 is strictly positive on the disjoint complement Nd. Indeed,
ifO<x ENd and P1x=QPx=0, thenPx EN, and so Px=P(Px)=O,
contrary to the assumption that x E Nd. That is, 0 < x E Nd implies
P1 x> 0. Clearly, P1 (E) = Pi (Nd). Therefore, letting F = P1 (E) and
using Theorem 5.59(4), we see that F is a positively complemented vector
sublattice of Nd and hence of E. In particular, notice that F is a closed
vector sublattice if E is a Banach lattice.
Finally, we claim that P1: Y > F is an onto order isomorphism. To see
this, notefirstthatx EYandP1x=OimplyQx=QPX=Pix=Oor
x E N, from which it follows that x = Px = 0. That is, P1 is one-to-one.
Also, it should be obvious that Pi is positive and onto. To finish the proof,
it remains to be shown that F Y is also positive. To this end, fix
ThisimpliesoP1x=QPx=Qx. Now
write x = xi + x2 E N Nd and note that X2 Qx 0. Consequently,
= x = Px = Px1 + Px2 = Px2 0, and the proof is finished.
Theorem 5.59 does not characterize lattice-subspaces. That is, not every
lattice-subspace of a vector lattice is the range of a positive projection. This
is demonstrated by the next example taken from [18].
Example 5.61. Consider the positive functions xl, x2 E C[0, 1] defined by
xi(t) =
1(t_1)2
and x2(t) =
1!_t
2
ifO<t<1
2
1 .
Let F be the vector space generated in C[0, 1] by the linearly independent

functions and X2. We claim that F is a lattice-subspace of C[0, 1].
To prove this, it clearly suffices to show that if + 0, then
0 and '\2 0. So, assume that + ,\2x2 0. This implies that for
0< t < we have \1(t 0, or t) +\2 0. Letting
232 -- 5._ Integral Operators
t we obtain A2 0. Similarly, from A1(t + A2(t 0 for

each < t 1, we infer that Ai 0.
Next,we claim that there is no positive projection on C[0, 1] whose

range is F. To see this, assume by way of contradiction that there exists a
positive projection F: C[O, 1] C[0, 1] such that P (CEO, 1]) = F. Consider
the positive linear functional f: F defined by
f(A1x1 + A2x2) = A1.
Then foP is a positive linear functional on C[O, iJ. So, by the Riesz Repre-
sentation Theorem, there exists a (unique regular Borel measure
v on [0,1] such that [f o P}(x) f[01] x dv for each x E C[O, 1]. Letting
x = X2, we get 0 = [f o P](x2) = f[01] x2 dv, from which it follows that
Supp v = { }. Finally, letting x = Wi, we obtain
1=[foP](xi)=f =0,
[071]
a contradiction. So, F cannot be the range of a positive projection.
In spite of the of positive projections from a vector

tice E onto its Iattice-subspaces, there is an important projection associated
with any lattice a posi-
tive projection defined on the vector sublattice 'R.(Y) of E generated by Y
and having Y as its range. This result is not trivial since even for a two
dimensional Y the vector sublattice can be infinite dimensional; see
Exercise 6 at the end of the section. To introduce these special projections,
we need to describe the vector sublattice generated by a vector subspace.
For every non-empty subset A of a vector lattice, the symbol AA will
denote the collection of all vectors that can be written as infima of finite
subsets of A. That is, a vector a belongs to A" if there exist vectors
ai, a2,. . ,
. a= Similarly, A" is the set consisting
of all suprerna of the finite subsets of A. We write AVA for (AV)A and AAV
for (A")". It always turns out that AVA = AAV. This is a consequence of
the following simple result whose proof is based upon the distributive law
of the lattice operations in vector lattices and is left as an exercise.
Lemma 5.62. Let i = {1,2,...,'n}; j J = {1,2,... ,'m}} be a
finite subset of vectors in a vector lattice. Then
72 n m m mm
A V Xi,r(i) VA and V A Xr(j),j = A V
rEFil ilj1
where J1 demotes the (fimite) collectiom of all mappings from I to J.
Positive Projections andLattice-subspaces 233
If A is a non-empty subset of a vector sublattice E, then the collection

of all vector sublattices of B that include A is non-empty. The intersection
of this collection of vector sublattices is the smallest vector sublattice that
includes A. This vector sublattice is called the vector sublattice (or the
Riesz subspace) generated by A and will be denoted The next
result presents a description of
Lemma 5.63. The vector sublattice generated by a vector subspace
Y of a vector lattice coincides with and also with yVA That is,
yAV = yVA
Consequently, the vector sublattice generated by a non-empty subset A

of a vector lattice is 7Z.(A) = where [A] denotes the linear span of A.
Proof. The verification of this identity follows easily from Lemma 5.62. We
leave the details as an exercise.
The relationship between lattice-subspaces and positive projections is

completely revealed by the following result of S. Miyajima [245].
Theorem 5.64 (Miyajima). A vector subspace Y of a vector lattice B is a
lattice-subspace of E if and only if there exists a (unique) positive projection
F: 7?,(Y) 7Z.(Y) whose range is Y. This projection P is defined by
=
Moreover, the operator F: p Y is a surjective interval preserving
Proof. Let Y be a vector subspace of a vector lattice B such that there

exists a positive projection F: 'R.(Y) 7Z.(Y) whose range is Y. By The-
orem 5.5 9(1) the vector space Y is a of 7?.(Y), and conse-
quently it is also a lattice-subspace of B.
For the converse, assume that Y is a lattice-subspace of E. We start by
establishing that the mapping as described in the theorem is well defined.
That is, we shall prove that
(*)
where the vectors and ypq belong to Y. To establish this, we need to
show first that if 0, then AT=ixij 0.
So, assume that 0. From Lemma 5.62 it follows that
0 for each 'r E j1. Since V 0,
applying Lemma 5.62 again (but this time to the vector lattice (F, Ft)), we
obtain that 0.
Now assume that = Ypq From the inequalities

Ypq Ypq for each p, it follows that
= Ypq 0,
andhence
V A A for each p, and thus
\'j/n
W i=1 //\\ 'W p_I /I\\ q_iYpq'
By symmetry, the reverse inequality is also true, and so the validity of (*)
has been established. In other words, the mapping P is well defined.
Next, observe that if x = = Ypq, then
ID
j //\\p1 W q=iYpq'
This follows easily from Lemma 5.62. We are now ready to establish the
basic properties of the mapping P.
o The mapping P is additive.
To see this, let x = and y Ypq, and note that
P(x + y) = + VL Ypq)
= + Ypq))
/ i
'Wi=1 -

Wj=l//\\j=lXijl Wp=1//\\q=1YPq
= P(x)+P(y).
o The mapping P is homogeneous.
Let x = and fix some scalar If 0, then it should be
clear that P(cEx) = aP(x). If < 0, then note that
= = VT=ia'xij
1 DI

a surjective interval preseruing

If x,y then it should be clear that P(x V y) = P(x)VP(y), and
therefore P is also a lattice homomorphism. To see that P is an interval
preserving operator, assume that two vectors 0 x E and y E Y
satisfy 0 y Px. If we consider the vector z = x A y then
0 z x holds and P(z) = P(x A y) = P(x) A P(y) = P(x) A y = y. Thus,
P[0,x] = [0,Px].
54. Positive Projections and Lattice-subspaces 235
The positive projection F: is the only positive pro-

jection on having Y as its range.
To establish this claim, let Q: be another positive projection
with range Y. In view of Theorem 5.59 we know that for each finite collection
of vectors xl, X2,.. , .in Y we have
= and =
Now fix some vector x = in Then for each i we have
x Xij, and the positivity of Q yields Q(x) =
for each i. This implies Q(x) V A X&j F(x). Since the vector
x can also be written in the form x = a similar argument
shows that Q(x) F(x). Therefore, Q(x) = P(x) for each x E and
the proof is finished.
If Y is a lattice-subspace of a vector lattice, then the unique positive

projection F: , R,(Y) with range Y determined in Theorem 5.64
will be referred to as the Miyajima projection. The reader can find more
about lattice-subspaces in [18, 268, 269, 270].
Exercises
1. If P is a positive projection on an for some 1 p <oo, then
show that its range is again order isomorphic (up to an equivalent norm)
to an for some measure space (Il1, E1,v).

4. Show that a vector subspace of a finite dimensional vector lattice is a
lattice-subspace if and only if it is the range of a positive projection.
5. For a vector subspace Y of some C(K)-space establish the following.
(a) Y separates the points of K if and only if separates the points
of K.
(b) Assume that the induced cone Y = Y fl is generating in
Y, i.e., = Then Y majorizes C(K) if and only if

majorizes C(K).
(c) Assume that:
(i) Y a Dedekind complete Iattice-subspace. (In particular,
is
this is so if Y is a finite dimensional lattice-subspace.)

(ii) Y separates the points of K and 1 Y.
Then the Miyajima projection P: f extends uniquely
to a positive projection on C(K) with range Y.
6. Let Y bethe vector subspace of C[O, 1} generated by the functions 1 and

s, where s(t) t for each t E [0, 1]. That is, Y consists of all linear
functions. Establish the following.
(a) Y is a Iattice-subspace of E.
(b) consists of all piecewise linear continuous functions.
(c) The Miyajirna projection F: is given by
Fx(t) = tx(1) + (1 t)x(0)
for each x E and each t [0, 1].
(d) The operator Q: C[0, 1] C[0, 1] defined by
Qx(t) = tx(1) + (1 t)x(0)
for each x E 1] and each t E [0, 1} is the only extension of the
Miyajima projection to a positive projection on all of C[0, 1].
(e) There are infinitely many bounded projections on C[0, 1] with
range Y.
7. If Y is a majorizing lattice-subspace of a Banach lattice, then show that
the Miyajima projection F: is continuous.
8. This exercise presents an application of lattice-subspaces to harmonic
analysis. Recall that a C2-function n: 0 where 0 is an open
R,
subset of R2, is said to be harmonic if it satisfies Laplace's equation
V2n + = 0 at every point of 0.
Now let D = {(x, y): x2 + y2 < i} be the open unit disk. We shall
denote by fl the vector space of all continuous functions defined on the
closed unit disk 1) = {(x, y): x2 + y2 < i} and harmonic on D. The
classical Dirichlet theorem states that if f is a continuous function on
the unit circle 3D = {(x, y): x2 + y2 = 1 }, then there exists a unique
continuous extensionf of f to all of b which is harmonic on D. The
harmonic extension f in polar coordinates is given by Poisson's classical
formula
J(r, 0) = (r, 0) E [0,1) x [0,
Establish the following.
(a) A function f satisfies 0 if and only if fOD 0, where
faD denotes the restriction of the function f to 3D.
(b) The mapping Q: defined by Qf = is a positive
projection whose range is 7-(.
(c) 7-( is a lattice-subspace of C(15) and Q is the unique extension of
the Miyajima projection to
(d) 7-( is an with unit 1, the constant function one, and the
mapping f ii f, from C(3D) to 7-(, is a surjective lattice isometry.
Chapter 6
Spectral Properties
This chapter introduces the spectrum of a bounded operator and studies its
basic properties. We start with a general discussion of the spectrum, of the
resolvent set, and of the spectral radius. Subsequently, we introduce and
study the resolvent function. In particular, we pay special attention to the
resolvent function of a positive operatorS
The various important parts of the spectrum (the point spectrum, the
residual spectrum, the continuous spectrum and the approximate point spec-
trum) are all introduced and some of their interrelationships are obtained.
The chapter culminates with an extensive treatment of the "Functional Cal-
culus" with a special emphasis on the Spectral Mapping Theorem and some
of its applications.
The Spectral Mapping Theorem is used in conjunction with the notion of
a spectral set to construct reducing pairs of subspaces for a given operator.
We also study the isolated points in the spectrum that are poles of the
resolvent function.
6.1. The Spectrum of an Operator

In this section, unless otherwise stated, X will denote a non-trivial (i.e.,
X {O}) complex Banach space. If a real Banach space X is given, then
X should be replaced by its complexification Xc. As mentioned earlier,
the symbol (X, Y) will denote the Banach space of all bounded (linear)
operators from a Banach space X into another Banach space Y and (X)
stands for (X, X). Unless otherwise stated, the letter I will denote the
identity operator on X. Identifying each complex number with we can
consider A as an operator on X.
237
238 6. Spectral Properties
Definition 6.1. The specLrum cT(T) of a bounded operator T: X X is

the set of all complex numbers such that the operator T is not invertible
on X, that is,
cT(T) = e C: T)' does not exist}.
The complement of the spectrum is called the resolvenL seL of T and is
denoted by p(T), that is,
p(T) = C\a(T) = e C: is an invertible operator from X onto X}.
It should be noticed immediately from the above definition that both

the spectrum cT(T) and the resolvent set p(T) depend only on the topology
of X and not on any particular norm generating the topology.
Let us make a comment regarding the resolvent and the spectrum of an
operator defined on a real Banach space. So, let T: X f X be a bounded
operator on a real Banach space X. Then, identifying T with T + 10, we
can consider T as a bounded operator on the complexification Xc of X, i.e.,
T + i0 E Now if,\ = a + is a complex number,
then the operator T can be identified with the operator (a T) + zf3I
on where I is the identity operator on X. This means that T
is an invertible operator (on Xc) if and only if there exist two operators
A, B e satisfying
[(aT)+ifiI](A+iB) = (A+iB)[(aT)+i/31] = I+z0
or [(aT)A--I3B]+i[(aT)B+I3A] = [A(aT)I3B]+i[B(a--T)+iiAf=
I + i0. Thus, we have the following result regarding the spectrum of a
bounded operator on a real Banach space.
Lemma 6.2. Let X be a real Banach space and let T e A complex
number = a + belongs to the resolvent ofT, i.e., e p(T), if and only
if there exist two operators A, B e satisfying
A(a-T) = (a-T)A =I+,1313 and = B(a-T) =
When ,\ = a + i,@ is a real number (i.e., when = 0),the preceding
lemma yields A(a T) = (a T)A = I. In view of Definition 6.1, there
exists a function from p(T) into This function is called
the resolvent function (or simply the resolvent) of T and is denoted by
R(.,T), i.e.,
T) = T)', e p(T).
If there is no need to indicate the operator T, then we shall simply write
instead of T). Notice that the assumption X {0} guarantees
that 0 for each e p(T).
6J. The Spectrum of an Operator 239
Theorem 6.3. If T E (X) and > 11Th, then,\ E p(T) and the resolvent
T) is given by the Laurent series
wherethe series converges in the operator norm in the Banach space (X).
Moreover, for each ,\1> 11Th we have
and so, in particular, T) = 0.

Proof. The norm convergence of the series follows from < 1. We
need to verify that the sum of the series, i.e., the bounded operator
S= coincides with T). To see this, note that
= lim
in-*oo
lim
m=0
in in
= lim I
L
lim I
Similarly, S(,\ T) = I, and so S = T). For the norm inequality
note that
00 00
= = IAI-UTII'
I I(
and the proof is finished.
The series expansion R(,\, T) = ,\_(m+1)Tm is usually referred to

as the Neumann series. It appeared for the first time in the work of
C. Neumann [254] devoted to potential theory.
Lemma 6.4. For an operator T E (X) and any p e p(T) we have:
(1) R(14R(\).
(2) IfS e (X) satisfies ST = TS, then SR(.A,T) for
each E In other words, if T belongs to the commutant of
an operator S, then T) belongs to it too. In particular, T and
R(,\,T) commuteS
(3) R(.A) R(p) =
Proof. (1) Notice that T) = (,\ T). So, by taking

inverses on both sides, we get =
(2) Fix E p(T). Then we have
T) = T)S = S =
and so by T, we get R(,\, T)S = SR(,\, T).

(3) Using part (2) we get
=
= =
and the desired identity follows.
Definition 6.5. For a given operator T E (X), the identity
R(,u)
is called the resolvent identity (of T).
A glance at Corollary 2.11 guarantees that the resolvent set of an opr-

ator is always open. Our next theorem shows that in actuality much more
is true.
Theorem 6.6. If T E (X) and )'o E p(T), then the open disk
D (,\o, is included in p(T) and for each E we have
00 00
= =
Moreover, letting 'y 1)' '\ol fR(,\o)11 <1, we have
(*)
and
S . (**)
Proof. Let A Aol Then 'y = A Aol' IR(Ao)i1 <1 and so the
series is norm convergent. Now note that
(A T) A)Th(A0
= [(Ao A) + T)]
(A0
[(Ao (A0
urn (A0 A)m+1R(A0)m+I] = I.

Tfl-'OQ
Sim:ilarly, A)Th(A0 T) = I.
establishes
This
the representation for R( A). In view of Theorem 1.72, the function R( is
analytic in p(T).
To establish the inequality (*), observe that
IIR(1\)11 = -
IA0 -
=
= IIR(Ao)1!.
Finally, to establish (**), use the identity
R(A) R(A0) =
in connection with the triangle inequality.

Corollary 6.7. For each operator T E iC(X) the set p(T) is an
open subset of the complex plane and the operator-valued function
A R(A,T) =
from p(T) to iC(X), is
In particular, the resolvent function A R(A) is continuous.
Corollary 6.8. For every operator T E (X) the derivative of its resolvent
function is given by

dA
Proof. By Corollary 6.7 the resolvent function is continuous. So, using the
resolvent identity, we see that
dR(A,T)
= urn = urn =
as claimed.
Here is another important consequence of Theorem 6.6.

Corollary 6.9. For an operator T E (X) we have:
(1) If E p(T), then where d(\,a(T)) is the
distance from to the spectrum cr(T).
(2) Assume that a sequence p(T) satisfies E C. Then

)'o E ci(T) if and only if 00.
Proof. (1) Fix,\ E p(T). If E cr(T), then according to Theorem we

see that (Otherwise, < 1IR(A)II
implies E p(T), a
contradiction.) So,
d(\, = inf{1\ E a(T)}
(2) If E cr(T), then the conclusion follows from the relation
Conversely, if T) 00, then be true.

E cr(T) must
erwise, if E p(T), then by the continuity of the resolvent function
R(,T): p(T) X we would have ER.
It turns out that the spectrum of a bounded operator is always nonempty

and compact. This is a basic result.
Theorem 6.10. The spectrum of a bounded operator on a (non-trivial)
complex Banach space is a non-empty compact subset of the complex plane.
Proof. Let X be a complex Banach space and let T E (X). Assume by

way of contradiction that cr(T) = 0 or equivalently that p(T) = C. This
means that the resolvent function R(.) is an entire function. The estimate
(for > established in Theorem 6.3, implies
= 0. The latter shows that is bounded at infinity. Since
R(.) is analytic, it follows that R(.) is a bounded entire function.
Consequently, by Liouville's Theorem 1.74, R(.) is a constant function.

In particular, notice that
(T)'=R(O)= Em
A'oo
which implies that X = {O}, a contradiction. Hence, 0.

From = C \ p(T), we see that is a closed set. On the other
hand, since Al > 11Th implies A E p(T), we also see that is a bounded
subset of C. So, is a non-empty compact subset of C.
It is interesting to observe that every non-empty compact subset K of

the complex plane C is the spectrum of a bounded operator. To see this,
choose a countable dense subset {A17A2,. .} of K and then consider the
.
linear operator T: 2 2 defined by
T(zi, z2,...) = (Aizi, A2z2,...).

If is the sequence whose nth component is one and every other is zero,
then we have T) (en) = 0 and so E o(T) for each ri. Now an easy
verification shows that if A K, then (A T)'
exists, and from this it
follows that = K. For another example see Exercise 15 at the end of
the section.
Definition 6.11. The spectral radius r(T) of an arbitrary operator T in

(X) is the smallest non-negative real number r for which the closed disk
{A E C: Al r} contains the spectrum That is,
r(T) A A E c7(T)}.
Clearly,r(T) and strict inequality is possible. As noticed before,

the spectrum of an operator T e (X) does not depend on any particular
equivalent norm on X. Therefore, the spectral radius r (T) is also indepen-
dent of any particular equivalent norm on X. Nevertheless, there is a very
important formula due to I. M. Gelfand that allows one to calculate the
spectral radius of an operator in terms of the norm.
Theorem 6.12 (Gelfand [133]). If T E (X), then
r(T)
In addition, for each A EC satisfying > r(T) the resolvent R(A,T) is

given by the Neumann series
R(A, T)
=
244 6. Spectral Prop erties
Proof. Let T e (X). Ife o-(T), then we claim that e cr(TT1). Indeed,
if then there exists some bounded operator 8: X X such
that (,V' TTh)8 = 8(,V' TTh) = I. Write
= +
and note that the operators A + + + + Ta') 8 and
B= satisfy T)A = B(.AT) = I.
This implies )'.. e p(T), which is a contradiction.1
Now choose some E cr(T) with = r(T). From e we see
that and so = r(T) holds for each m. It follows
that
r(T) inf <liminf (*)
Note that the resolvent function R(,\, T) is analytic on the open annulus
A= e C: > r(T)} and that A is the largest open annulus centered
at zero that is included in p(T). By Laurent's Theorem 1.78 we must have
= for each > r(T), and from Theorem 1.77
we see that r(T) = lim A glance at (*) guarantees that
r(T) = =
Werefer to Exercise 6 at the end of this section elementary proof
of the existence of the limit *.
Definition 6.13. An operator T E (X) is said to be:
(1) nilpotent if = 0 for some positive integer k and
(2) quasinilpotent if r(T) = = 0.
Clearly, every nilpotent operator is quasinilpotent. Also notice that an
operator T E (X) is quasinilpotent if and only if a(T) = {0}.
Recall that if T: X Y is a bounded operator between Banach spaces,
then its adjoint operator T*: is defined via the duality formula
(x,T*y*) = (Tx,y*)
for all x E X and all E It is well known that the mapping T
from (X, Y) to (Y*, X*), is a linear isometry.
Theorem 6.14. The spectrum of an operator T E (X) coincides with the
spectrum of its adjoint, i.e.,
cr(T) = cr(T*).
In particular, we have r(T) = r(T*).
1
we use the fact that if A, B, C E (X) satisfy CA = BC = I, then A = B = C'.
To see this, note that A = IA = (BC)A = B(CA) = BI = B.
Proof. The proof is a consequence of the following two facts:

(1) for all eC and
(2) An operator S e (X) is invertible if and only if its adjoint 8*
invertible.
To verify (2), notice first that SA = AS = I implies = A*8* = 1*
Conversely, if 5*: is invertible, then by the previous remark the
operator S**: is invertible and so 8 (which is the restriction of
to X) is one-to-one and carries X onto a closed subspace 8(X) of X.
If 8(X) X, then by the Separation Theorem there exists some non-zero
e X* which vanishes on 8(X). That is,
(Sx, = (x, 8*x*) =0
holds for all x E X, and so 8*x* = 0. Since is invertible, we infer that
= 0, which is a contradiction. Hence, 8(X) = X and so S is invertible if
and only if 5* is invertible.
We close the section with the definition of the peripheral spectrum of an

operator.
Definition 6.15. For an operator T E (X), the non-empty compact set

)\1=r(T)}=cJ(T)fl{)\EC: 1\l=r(T)}
is known as the peripheral spectrum of the operatorS That is, the periph-
eral spectrum of an operator is the non-empty part of the spectrum that lies
on the circle centered at zero with radius r(T).
Finally, let us indicate that the preceding definition and Tresults

operators can be extended verbatim to complex unital Banach algebras. To
see this, let A be a complex Banach algebra with unit e. If x is a vector in
A, then the spectrum of x is defined by
cr(x) = E C: x is not invertible in A}
The resolvent p(x) of x is the complement of Le., p(x) = C \

The spectral radius of x is defined as usual by r(x) =
Similarly, the resolvent function R(., x): p(x) A of x is defined by
R(\,x) = (\ex)1.
It is easy to see that all results concerning operators that were obtained
in this section are also true for vectors in complex unital Banach algebras.
For instance, we have the following theoremS
Theorem 6.16. If x is a vector in a complex unital Banach algebra, then:

(a) Its spectrum cr(x) is a non-empty compact subset of C.
(b) Its spectral radius is given by r(x) = urn txtm II = inf 11*.
fl400 fl
Exercises
1. Establish the following identityknown as the second resolvent iden-

tity. If S,T E (X) and A E p(T) np(S), then
R(A,S) R(A,T) R(A,S)(ST)R(A,T).
2. Generalize Corollary 6.8 by proving that the derivatives of the resolvent
function of an operator T E (X) are given by the formulas
=
dATh
for n = 1,2,.... [HINT: Use Theorems 1.73(3) and 6.6.1

3. If X is a Banach space and T E (X), then show that lim AR(A, T) I.
[HINT: Use either Exercise 11 of Section 2.1 or Exercise 6 of Section 1.5.1

4. For any operator T E (X) show that [R(A,T)]* = R(A,T*) holds for
each A E p(T).
5. Let T E (X). If A0 E p(T), then show that R(A0, T) cannot be quasi-
nilpotent. [HINT: If S E (X) is invertible, then 0 o(T).]
6. Show directly that if T: X f X is a bounded operator on a (real or
complex) Banach space, then 1
always exists in R and
= inf
[HINT: Let a II 0. From a

1
a. Next, let E> and then fix some k such that
0
<a + . Now for any positive integer n write n = kin + r, where
Notethat
IITThII = IITMT IITkMm = 11ThT'
= M(a+)Th,
where = Thus, from which we
get for all 0, or a.]
7. For any operator T E (X) show that
r(T)=inf{a>0:
8. If 5, T: X X are two commuting bounded operators on an arbitrary
Banach space, then show that r(ST) <r(S)r(T). [HINT: If ST = TS,
then (ST)Th = for each n.]
The Spectrum of an Operator 247
9. Let T : X p be a bounded operator and assume that for some non-zero

a E C the series 00 'T nx converges in norm for each x X.
.
'
Show that a E p(T). [HINT: Use the Uniform Boundedness Principle.]
10. For an arbitrary operator T E (X) show that r(T) =0
if and only if * = 0 holds for each x E X. [HINT: Use the
previous exercise.]
11. If X is a Banach space and T E (X), then show that
a(Tk)= {Ak: A E
for each k (and conclude from this that r(Tk) = holds).
Use the above conclusion to establish that an operator T E (X)
satisfies 0 if and only if r(T) < 1. [HINT: For the second part,
note that if 1T9 0, then 1 for some k and therefore we have
r(T) = < 1. On the other hand, if r(T) = tJTfhI* <1,
then fix r(T) <6 < 1 and note that iT7t II for all m
12. If T: X X is an invertible bounded operator on a Banach space, then
show that a(T') = {A': A E cr(T)}.
13. If T E (X) and IITkII < 1 for some k, then show that T7t = 0.
[HINT: Note that r(T) = inf7t < I and use Exercisell above.]
14. Let be a compact Hausdorif space and let p be a Borel measure on ft
Assume that either E = or E is a Banach function space that is
order dense in L0(p). If p C is a continuous function, then show
that the spectrum of the multiplication operator E p E coincides
with the range of i.e., = = {q5(w): w E
15. For a non-empty compact subset K of C establish the following.
(a) There exists a bounded operator T: 2 p such that a(T) = K.
(b) If S: p Cc(K) is the multiplication operator defined by
(Sx)(z) = zx(z), then o(S) = K.
(See also Exercise 11 in Section 7.1.)
16. If an invertible operator T E (X) satisfies liT7t M < oc for all
1
integers n (such an operator is called a Cesaro operator or a doubly

power bounded operator), then show that a(T) { E C: (Al = 1}.
[HINT: If Al <1, consider the operator S and note
that
17. Let I,... ,ri) be bounded operators on Banach spaces.
(i
If X = X1 . . is the direct sum Banach space, then show that:
.
(a) The direct sum operator T = T1 . .. X X, defined by
is invertible if and only if each is invertiblein which case the

inverse operator satisfies T' = ... T;'.
(b) The spectrum of the direct sum operator T is the union of the spectra
of the operators That is,
18. Show that an operator T: X + X on a complex Banach space is nilpotent

if and only if it is quasinilpotent and has finite descent.
19. If 8, T E (X), then show that cr(ST) U {O} = cr(T8) U {O}. Conclude
from this that:
(a) cr(ST) = {O} if and only if cr(TS) = {O}.
(b) r(ST) = r(T8).
Give an example of two operators S, T (X) such that cr(ST) cr(T8).
[HINT: If A, B, C e (X) satisfy the identity (I AB)C = C(I AB) =
I, then notice that (I + BCA)(I BA) = (I BA)(I + BCA) = I. For a
counterexample, notice that if 8 and T denote the backward and forward
shift operators on 4, then = {1} while = {O, 1}.]
20. Let T e (X) be an operator and Ilet z0 e C be a point such that every
Ae \ {z0} is an isolated point of cr(T). Show that is at most
countable and that if is countable, say ci(T) = {A1,A2,...}, then
z0. [HINT: For each c > 0 let = {A e JA c}.
Establish that each is a finite set.]
6.2. Special Points of the Spectrum

As before, X will denote a non-trivial complex Banach space and T e (X)
a fixed operator.
A complex number A is an eigenvalue of T whenever there exists a
nonzero vector x (called an eigenvector corresponding to A) such that
Tx = Ax. Notice that the eigenvalues of T are exactly the complex numhers
A for which A T is not one-to-one, and so they belong to the spectrum of
T. The set of all eigenvalues of T is known as the point spectrum of T
and is denoted by That is,
= {A e A T is not one-to-one }
= {AEu(T):
Recall that if an operator is one-to-one, then (in view of the Open Map-
ping Theorem) it fails to be invertible if and only if it is not surjective.
Accordingly, when a noninvertible operator A T is one-to-one, we distin-
guish two cases:
(1) The range of the operator AT is not dense, i.e., (A T)(X)
(2) The operator A T has dense range, i.e., (A T)(X) = X.
In case (1) we say that A belongs to the residual spectrum of T.
That is, the residual spectrum is defined by
= {A e \ (A - T) (X) x}.
In case (2) we say that A belongs to the continuous spectrum of

T. That is, the continuous spectrum is defined by
{A E a(T) \ (A T)(X) = X
Clearly, (T), and
o"(T) = U U
Another important part of the spectrum is the approximate point

spectrum o"a(T) that defined as follows.
O"a(T) = {A E C: V > 0 x with lxii = I and TxII <c}.
The reader should be able to verify easily the following equivalent ways of
defining the approximate point spectrum.
O"a(T) = {A E C: A T is not bounded below }
= {A E C: c X with I V n and -4 0
}
= {A E C: X satisfying 74 0 and 0 }
{A E C: CC with A and C X with

I for all n such that 0
}.
The points of the approximate point spectrum are also known as

proximate eigenvalues. Clearly, c
Theorem 6.17. The continuous spectrum of an operator T E (X) is in-
cluded in the approximate point spectrum of T, i.e., C cra(T). In
parlicular, we have o"(T) = o"a(T) U crr(T).
Proof. Take an arbitrary A E By the definition of the continuous
spectrum, the range of A T is dense and consequently (by Lemma 2.8) the
operator A T is not bounded below. Therefore, A E cra(T).
The important topological properties of the approximate spectrum are

included in the next theorem.
Theorem 6.18. For an operator T E L(X) we have the following:
(I) The approximate point spectrum O"a(T) is a closed set.

(2) Every boundary point of the spectrum belongs to the approximate
point spectrum. That is, &r(T) o"a(T) (and so o"a(T) 0).
(3) The complement of O"a(T) in o"(T), i.e., the set
o"(T) \ o"a(T) = {A E C: A T is bounded below

and
is an open subset of the complex plane.
Proof. (1) Take any a 0 Then the operator a T is bounded below.

For each A e C we have I(A T) (a T)IJ = JA aJ. So, if A e C is
close enough to a, then by Lemma 2.4(2) the operator A T is also bounded
below. That is, for all A sufficiently close to a we have A 0 cTa(T) proving
that cra(T) is closed.
(2) Let A e cr(T). Then there exists a sequence p(T) satisfying
A A T were bounded below, then
Theorem 2.9 would imply that A T is invertible, contrary to A e cr(T).
Hence A e cTa(T), and so Oo"(T) O"a(T).
(3) Let a cra(T) and cra(T) is closed, there
exists-some >--0-such fl-cia(T) = 0. Also, from Ocr(T) cra(T),
we know that a Ocr(T), and hence a' is an interior point of cr(T). It follows
that there exists some 0 < r < such that c \ cra(T). That is,
is an interior point of cr(T) \ cra(T) and so cr(T) \ cra(T) is an open subset
of the complex plane C.
It is important to know that each point of the residual spectrum of an
operator is an eigenvalue of its adjoint.
Theorem 6.19. If T L(X), then
e
crp(T*)={AEC:
In particular, we have (T) (T*).
Proof. The proof will be based on the duality identity

=(x,(A_T*)x*). (*)
If AT does not have dense range, then by the Separation Theorem there
exists some e satisfying ((A T)x,x*) = 0 for all x e X.
But then (*) implies (A T*)x* = 0 or T*x* = Ax*, i.e., A is an eigenvalue
of T*. Conversely, if T*x* = Ax* holds for some e X*, then (*)
implies ((A T)x, x*) = 0 for all x e X, and so the range of A T cannot
be dense in X. D
Regarding eigenvalues we have the following interesting result.
Theorem 6.20. If an operator T: X X on a complex Banach space is
surjective but not invertible, then there exists some > 0 such that every A
in the open disk D(0, E) is an eigenvalue ofT.
Proof. According to Corollary 2.12 there exists some E > 0 such that every
operator S e (X) that satisfies < E is surjective and non-invertible.
In particular, for each A e D(0, E) the operator T A is surjective but not
invertible. This implies that each A e D(0, ) is an eigenvalue of T.
Let us use the shift operator as an example to illustrate the various parts
of the spectrum. The (forward) shift operator T: 4 (1 <p < oc)
is defined by
T(zi, Z2,...) = (0, zi, Z2,...).
For 1 p <oc we have = eq, where + = 1, and the duality eq)
is defined by
(z,e) =
for all z = (zl,z2,...) E 4 and all e = (el,e2,. .) E It is easy to verify

that the adjoint .T*: q _> q of the shift operator is given by
T*(zi,z2,...)=(z2,z3,...),
which is also known as the backward shift operator.

Example 6.21. Let 1 4 be the shift operator.
Notice that lITThzIIp = holds for all z E and so ITThII = 1 for each n.
Therefore,
1
r(T) = lim
Th-+oo
=1=
We shall establish that a(T) = o.(T*) = {A E C: <i}.
1. The eigenvalues of the forward and backward shift operators.
Obviously, 0. We claim that ap(T*) = E C: < 1}.
Indeed, if < 1, then the vector ZA = (1, A, A2,...) E eq satisfies
(Compare this conclusion with Theorem 6.2ft) An easy argument shows

that no A E C with A! = 1 can be an eigenvalue of T*. Taking into account
that 11Th = 1, we conclude that -
o.(T)=o.(T*)={AEC:
2. The residual spectra of the forward and backward shift operators.

Since = 0, it follows that = ap(T*) = E C:< 1};
see Theorem 6.19. Also, we have ar(T*) = 0. Indeed, Theorem 6.19 applied
to the adjoint operator T*: eq _> implies
cTr(T*) C ap(T**) = =0.
3. The continuous spectra of the forward and backward shift operators.
We have
= a(T)\ = {A E C: = i}
and
crc(T*) = a(T*) \ [crp(T*) U Ur(T*)] E C: = i}.
=
4. The approximate point spectra of the forward and backward shift
tors.
Here we claim
AI=1} and aa(T*)=a(T*)={,AEC: JAI<i}.

To see that ={#\ E C: = i}, note first that
Da(T) = i} c cra(T).
E C:
To establish that equality holds, now observe that for a! < 1 and x E 4
with lix lip = 1, we have
TxIIp = I
I
that an operator T (X) is not one-to-one if and only if 0 is an

eigenvalue of T.
2. Let T: V V be a linear operator on a vector space. Show that a
collection of eigenvectors of T any two of which correspond to distinct
eigenvalues is a linearly independent set.
3. Let T: X + X be a bounded operator. Show that if for some C
there exists a sequence of unit vectors such that 0,
then cr(T), i.e., show that oa(T) C cT(T).
4. Show that a complex number is an approximate eigenvalue for an operator
T (X) if and only if there exists a sequence in X satisfying
74 0 and 0.
5. Let T: X X be a bounded be'ow operator on a Banach space. If 'y> 0
satisfies fTxfj 'yIIxM for each x X, then show that
c7a(T)fl{.AEC:
In particular, if the operator T: X + X is an isometry on a Banach space,
then show that (T) fl {.A C: < 1 } = 0. [HINT: If J.Af <'y and
xE Xsatisfies = 1, then lTx)txjI >
6. Let T: X + X be a non-invertible bounded operator on a Banach space.
If some'y > 0 satisfies for each x X, then show that
D(0, c Clearly, D(0,

We must show that 'y. To establish this claim, assume by way of
6.3. The Resol vent of a Positive Operator 23
contradiction that /3 < 'y. This implies the existence of a sequence of

complex numbers satisfying /3 < </3+ and for
each n and A. From IA! = /3, it follows that A a(T). Now notice
that
- lITxIl - - lIxlJ
holds for each x X. Since each operator T has dense range, it foL
lows from Lemma 2.8 that each T is an invertible operator. But then,
by Theorem 2.5, the operator A T is itself invertible, a contradiction.]
7. If P e (X) is a non-trivial projection (i.e., if P2 = P and P 0, 1),
then show that a(P) = = {0, 1}.
8. Let S e (X) and T E (Y) be two bounded operators, where X and
Y are two complex Banach spaces. The operators S and T are said to
be similar whenever there exists an invertible operator 1? (X, Y)
satisfying T = RSR1. In other words, S and T are similar if and only if
there exists an isomorphism (i.e., a surjective, one-to-one, and continuous
operator) 1?: X Y such the diagram shown in Figure 1 is commutative.
If S and T are similar operators, then show that
a(S) = o-(T), ar(S) = ca(S) = and ar(S) =
x
sj jT
x
Figure 1
9. For two operators 5, T (X) establish the following.

(a) A non-zero complex number is an eigenvalue of ST if and only if it
is an eigenvalue of TS.
(b) If X is finite dimensional, then show that cr(ST) = a(TS); compare
thiswith Exercise 19 of Section 6.1 and Exercise 7 of Section
[HINT: (ST)x = Ax implies (TS)Tx = ATx.1
6.3. The Resolvent of a Positive Operator

In this section we shall exhibit a few additional properties of the resolvent
function when our Banach space X is the complexification of a Banach
lattice. So, let E be a (real) Banach lattice. As before, if T E (E), then by
the spectrum, the resolvent, the resolvent function, the spectral radius, etc.
of T, we shall always understand the corresponding notions for the extension
T of T E (E) is the
set
= {A E C: A is not invertible on
Since the two norms and

on defined by
IJzII= sup and
OE[O,2ir]
are equivalent, the following result should be immediate from Lemma 1.7.
Lemma 6.22. If T: E p E is a bounded operator on a real Banach lattice,
then
1 1
r(T) max Al = lim ]n lim
C
Concerning the resolvent of a positive operator, we have the following.
Lemma 6.23. For a positive operator T: E E on a real Banach lattice

the following statements hold.
(1) If A E C satisfies IAI > r(T), then
IR(A,T)zl R(IAJ,T)lzl for all z E
(2) If A1 and A2 are real numbers and A1 > A2 > r(T), then
o R(A2,T).
R(A1,T)
(3) If A is a real number satisfying A > r(T), then

o TR(A,T) AR(A,T).
Proof. (1) Using Lemma and Exercise 1 at the end of the section,
see that
00 00
=
Assume that A is real and satisfies A > r(T). Then each operator
(2)
A_(k+1)Tk is positive, and so the operator is positive for

each n. Therefore, if x E E+, it follows from the closedness of the positive
cone E+ that
R(A, T)x A +1)Tk] x

= k=O k=O
So, R(A, T) a positive operator. The other inequality now follows from
is
the resolvent identity by observing that
R(A2,T) R(A1,T) = (Ai A2)R(A1,T)R(A2,T) 0.
6.3. The Resolvent of a Positive Operator - 255
(3) Here we have

00 00
0 TR(A, T) >A(n+l)Tn+l =
=A < AR(A, T),
We close the section with an important property of certain positive op-

erators.
Lemma 6.24. Let T: E E be a positive operator on a Banach lattice.
Assume av <Tv for some 0 <v E E and some scalar > 0. Then for
each real number A > r(T) there exists some 0 r(T), we know
that R(A, T) = We claim that the vector u = R(A, T)v has the
desired properties. Indeed,
00 0000
Applying Lemma 6.23(3), we get
Tu = TR(A,T)v AR(A,T)v = Au.
Finally, from R(A,T) we obtain u = R(A,T)v This implies
Exercises
1. Assume that exists in E, where = + e for each n.
Show that exists in and 1 where the
series are assumed to converge in norm. [HINT: If k and 9 are fixed, then
xn) cos 9+ cos 9 = cos 9+ sin 0)
Hence, cos 9 + sin 9 holds for all 0, and

from this the desired inequality follows.
- ----- - Properties
2. If 8, T : E E are positive operators, then show that

r(S+T) max{r(S),r(T)}.
3. Let T : E p E be a positive operator on a Banach lattice. If T is a-
order continuous (resp. order continuous), then show that R(A, T) is also
a-order continuous (resp. order continuous) for each A > r(T).
4. Let T: E E be a positive operator on a Banach lattice. If a scalar a > 0
satisfies ax Tx for some x > 0 or ax* < T*x* for some 0 < E E*,
then show that <r(T).
5. Let T: X X be a bounded operator on a complex Banach space arid
let (T) denote the unbounded (open) component in C of the resolvent
p(T). Establish the following.
(a) If a(T) is at most countable (in particular, if every non-zero point
in a(T) is an isolated point), then = p(T).
(b) A closed subspace Y of the Banach space X is T-invariant if and
only if Y is R(A, T)-invariant for each A E
(c) If a closed subspace Y of X is Tinvariant and
T to Y, then c and = R(A,T)Jy for
each A E
(T) = p(T), then for every T-invariant closed subspace Y of X
we have a(T[y) c a(T).
(e) Assume that an open annulus ARL,R2 (A0) lies entirely in (T) and
that R(A,T) = is the Laurent series expansion
of the resolvent in AR1 ,R2 (A0). If Y is a T-invariant closed subspace

X, then Y is also An-invariant for each n E Z.
of
6. Give an example of a positive operator T: E p E on a Banach lattice

and some T-invariant closed ideal J in E such that:
(a) There is some A E p(T) for which J is not R(A, T)-invariant.
(b) The spectrum a(TIj) is not included in a(T), where Tlj denotes
the restriction of T to J.
(Compare these conclusions with Exercise 5 above.)
6.4. Functional Calculus

In the course of the proof of Theorem 6.12 we have established that if T is
an operator on a complex Banach space and A E cr(T), then E for
each n. More generally, if p(z) = a polynomial and
p(T) denotes the operator obtained by the formal substitution of an operator
T E (X) for z in p(z), then one can show easily (see Exercise 1 at the end
of the section) that cr(p(T)) = p(cr(T)). This result is a very special case of
a powerful theory that allows one to associate with each operator T E
a broad class J(T) of analytic functions (that includes all polynomials) such
that for each f J(T) an operator f(T) can be defined whose spectrum
cr(f(T)) coincides with f(cr(T)).
6.4.Functional Calculus 257
The present section is an introduction to this "functional calculus."

A rigorous treatment of this theory depends heavily on the theory of
Banach-valued functions of a complex variable, some elementary proper-
ties of which were presented in Section 1.5. Detailed presentations of func-
tional calculus can be found in the classic monographs by N. Dunford and
J. T. Schwartz [118] and E. hue and R. S. Phillips [152]. See also [1131
and [317].
As before, X will denote a complex Banach space and
T: X * X a bounded operator. Recall that a neighborhood of a sub-
set A of C is any open set C such that A c C. We shall denote by
functions that-are analytic on-neighbor--
hoods of the spectrum cr(T) of T. That is, f E F(T) if and only if there
exists a neighborhood C (depending on f) of o(T) such that f: C f C is
analyticthe neighborhood C can always be assumed to be bounded.
To be more precise, consists of pairs (f, V), where V is an open
bounded set including u(T) and f: V * C is an analytic function. If we
define the relation on by letting (f, V) (g, W) whenever f = g
on some neighborhood U of u(T) satisfying U c V fl W, then is an
equivalence relation on F(T). With this equivalence relation in mind, we
shall interpret the elements in as equivalence classes rather than as
individual functions. The proof of the next result is straightforward and is
omitted.
Lemma 6.25. For every operator T E (T) the of all
collection
eqtLivalence classes of f'anctions eqtLipped with the pointwise operations of
addition and m'altiplication is a 'anital algebra.
It is a very important fact that to each function f E we can

associate an operator f(T) E (X). The definition of the operator f(T)
will be described next. But first, let us say that a subset C of C is a Jordan
contour if there exists a finite number of pairwise disjoint (closed simple)
rectifiable Jordan curves Ci, C2, . ,Cm in C such that C =
. . Cm.
Definition 6.26. LetV be a neighborhood of a s'abset A of C. A Jordan

conto'ar C in V surrounds A if there exists an open s'abset W of V s'ach
thatACWCWCV andC=OW.
The Jordan conto'arC, viewed as the bo'andary of W, is always ass'amed
to be positively oriented.
The following is a very useful result. Although it is intuitively clear,

its proof is cumbersome and requires some lengthy technical arguments; for
details see [317, p. 310].
Lemma 6.27. Every neighborhood V of a compact set A admits a Jordan

contour in V surrounding A.
Now let T E (X) and f Pick any neighborhood V of cr(T)

on which f is analytic, and let C be a Jordan contour in V surrounding
o-(T). Then the operator-valued function A 'p T) is continuous on
C, and hence it is Riemann integrable over C. Moreover, Cauchy's Integral
Theorem 1.70 shows that if C1 is another Jordan contour surrounding cr(T)
in the domain of analyticity2 of f, then
f f(A)R(A, T) dA
= f T)
Thus, the integral f(A)R(A, T) dA is well defined, i.e., it is independent of

the Jordan contour C surrounding o-(T) in the domain of analyticity of f.
This observation gives rise to the operator f(T) as follows.
Definition 6.28. Let T E (X) and fix f E J(T). If C is any Jordan con-
tour surrounding cr(T) in the domain of analyticity of f, then the operator
f(T) in (X) is defined by
f(T)=
Thealgebraic properties of the operators f(T) are included in the next

result. These properties are referred to collectively as "functional calculus."
Theorem 6.29 (The Functional Calculus). The mapping f -+ f(T), from
F(T) to (X), is an algebraic homomorphism. That is, for each pair f, g
in F(T) and all scalars a and in C we have
(af + ,8g) (T) af(T) + ,8g(T) and (fg)(T) f(T)g(T).
Moreover:
(1) If S E (X) commutes with T, then S commutes with f(T). In
particular, Tf(T) = f(T)T for each f E J(T).
(2) 1ff E J(T), then f and f(T*) = f(T)*.
(3) If a function f satisfies f(A) = a neighbor-
hood of cr(T), then f E F(T) and f(T) =
E a neighborhood V of a(T)
such that both f and g are analytic on V, and let C1 be a Jordan contour
2 means that there exists some neighborhood U of cr(T) such that C1 U and f
analytic on U.
6.4. Functional Calculus 259
in V surrounding cr(T). Then for each fi E C we have
+fig(T)
= f
= + T) = + fig)(T).
Let W be the neighborhood of cr(T) such that cr(T) W C W C V and

C1 = EThV. Now pick another Jordan contour C2 in W surrounding o-(T).
By Cauchy's Integral Theorem 1.70 we have = for each
and
E C2 dA = 0 for each E C1. Therefore, using Fubini's
theorem and the Resolvent Identity (see Definition 6.5), we see that
f(T)g(T)
=
f
C2 f
C1
: EL2
T)
T) = (fg)(T).
1C2
The above establish that the mapping f 'k 1(T), from 1'(T) to (X), is
algebraic homomorphism. We shall finish the proof by verifying the
remaining properties.
(1) From ST = TS and Lemma 6.4 we see that SR(1\, T) = R(1\, T)S for
each e p(T). Now the commutativity of S and f(T) follows immediately
from the definition of the Riemann integral of a Banach-valued function.
(2) Theorem 6.14 asserts that o-(T) = o(T*), and so =
Also, we know that T*) = T)]* for all A E p(T); see Exercise 4 of
Section 6.1. Now the desired conclusion follows from the definition of the
Riemann integral by observing that:
f(T*)
=
=
(3) Consider first an entire function of the form h(A) = where k E N,

and fix any positively oriented circle C centered at zero with radius greater
than Using that the Neumann series (see Theorem 6.3) converges
uniformly on any compact subset outside of the closed disk centered at zero
with radius we see that
h(T) =
f AkR(A T) dA =
f dA
This conclusion combined with the first part shows that for every polynomial
p(A) = we have p(T) =
We now consider the general case. Since the series f(A) = E=O
a neighborhood of o-(T), there exist 0 < ri < such that
cr(T) B(O, ri) and the series converges for each A E r2).
In particular, the series f(A) = converges uniformly on B(0, r1).
That is, if pk(A) = then the sequence of polynomials {Pk}
verges uniformly to f on B (0, ri). This implies (see Exercise 5 at the end
of the section) that
f(T) = lim Pk (T) = lim

koo k'oo
n=O n=O
holds in (X), and the proof is finished. D
From part (3) of the preceding theorem we get the following.

Corollary 6.30. If T E C(X) arid V is a meighborhood of a(T), then for
any Jordam comtour C in V surroumdimg cr(T) we have
I.
We are now ready to state and prove the Spectral Mapping Theorem.
Theorem 6.31 (The Spectral Mapping Theorem). If T E (X), them for
every fumction f E we have cr(f(T)) = f(a(T)). That is,
a(f(T)) = {f(A): AE
Proof. Let f E J(T) and choose a neighborhood V of cr(T) on which f is

analytic. Fix E cr(T) and define the function g: V C by
f
Functional Calculus 261
Clearly, the function g is analytic on V and satisfies f(A) = A)g(A)

f(T) =
for all A E V. From Theorem 6.29, we get T)g(T). Since
E u(T) and commutes with g(T), it follows from the previous identity
that the operator f(T) cannot be invertible, i.e., e o-(f(T)).
Hence, f(cr(T)) C cr(f(T)).
For the reverse inclusion, let E a(f(T)) and assume by way of con-
tradiction that f (u(T)). Consider the function h(A) and note
=
that h E J'(T). From f(A)) = 1 for each A E V and Theorem 6.29,
it follows that h (T) 1(T)] = 1(T)] h (T) = I, contrary to the non-

invertibility of Therefore, E I (a(T)), and so a (1(T)) I (u(T)).
Consequently, f(a(T)) = u(f(T)).
The rest of the section deals with properties of spectral sets. To illustrate
their role, we begin with a simple example. Let (i = 1,2) be two
bounded operators defined on complex Banach spaces. Also let =
be the spectrum of the operator Consider now the direct sum Banach
space X = Xi X2 and the direct sum (continuous) operator T: X f X
defined by T(xi, x2) = T1x1 T2x2. It is easy to check that u(T) = al U a2
(see also Exercise 17 of Section 6.1) and that X1 and X2 are complemented
T-invariant closed subspaces of X. These decomposition properties are the
basic features behind the concept of a spectral set of an operator. As we
shall see, whenever the spectrum of an operator T can be represented in the
form al U = u(T), where al and a2 are closed and disjoint, then there
exists a unique pair (Y, Z) of T-reducing subspaces such that u(Tfy) = al
and a(Tlz) = Therefore, in such a case, the operator T: X f X can be
reduced to two "simpler" parts.
The above discussion can be formalized by introducing the concept of a
spectral set of an operator.
Definition 6.32. Let T: X X be a continuous operator on a Banach
space. A subset a of cr(T) is called a spectral set of T if u is a closed and
open subset of a(T) in the relative topology of cr(T).
Equivalently, a closed subset a of a(T) is a spectral set whenever its
relative complement a' = u(T) \ cr is also a closed set.
Now let a be a non-empty spectral set of an operator T e such

that a' = a(T) \ a is also non-empty. If we let 26 = d(a, u') > 0, then the
two non-empty open sets V0. = UAE0. B(A, 6) and = UAE0.' B(A, 6) satisfy
acV0., a'V0.', and
V0. a
V0. a of u(T). Next, consider the
2(2 6. Spectral Properties
i'iiictjons fe,., V0. U Vt,.' + C defined by

Ii if
and
10 if
= if E = E
Clearly, f0., fe,.' E and

)2
(f)2 = ia, (icy' = icy', Icr icr' = 0, and icr + icr' = 1.
Consequently, according to Theorem 6.29 we have
[f0.(T)]2 = f0.(T), = =0, and

fcr(T)+fcr'(T) = I.
It follows that f0.(T) and (T) are both bounded projections on X satisfy-
ing X R(f0.(T)) R(fcr'(T)). Moreover, the Spectral Mapping Theorem
yields
ci(f0.(T)) =f0.(cT(T)) ={0,1} and {O,1}.
This shows that f0.(T) and both non-trivial bounded projec-

are
tions. Since f0.(T) and i0.' (T) commute with T, we see that the pair
(R(fcr (T)), R(fcr' (T))) of complemented closed subspaces reduces T.
We emphasize that the projection f0.(T) is independent of the open set
Vcr. This projection plays an important role in operator theory and has a
special name.
Definition 6.33. The projection f0.(T) is called the spectral projection
associated with the spectral set ci and is denoted P0.(T), i.e., Pcr(T) = icr(T).
We can also consider the extreme cases ci = o(T) and ci = 0. In this
cases, we have = I and = 0.
A fundamental property of the spectral projection P0.(T) is included in
the next result.
Theorem 6.34. Assume that a bounded operator T: X + X on a Banach
space has a non-trivial spectral set ci. Then T admits a unique (non-trivial)
reducing pair (Y0., Zcr) such that
and cr(TIza)=cr(T)\cr,
where and denote the restrictions oi T to Y0. and Z0., respectively.
Moreover, the spectral projection P0.(T) is the projection of X onto Y0.

along Z0..
Proof. As above, let V0. and be two disjoint non-empty neighborhoods

of o- and o-' = o-(T) \ o-, respectively, and let i0. and i0.' also be defined as
above. If Y0. and Z0. are the ranges of the non-trivial projections i0.(T) and
i0.' (T), then the pair (Y0.: Z0.) of non-trivial closed subspaces reduces T; see
Theorem 2.22. First, we shall verify that = o and = cr. By

symmetry, it suffices to establish that = a. For simplicity, for each
R E L(X) we shall denote the restriction of R to Yr7 by Rr7, i.e., Rn =
Let and assume by way of contradiction that a. We can
suppose that V0.; otherwise we replace V0. by Vn, fl (C \ Define the
function g: V0. U Vc,.' k C by g(1Lt) = if E Vc, and g(ji) = 0 p, E
Clearly, g E = f0.(1Lt) and g(,u)f0.(,u) = g(,u) for each
E Vc, U Consequently, from Theorem 6.29 we get
T)f0.(T) = (.\ = and (*)
g(T)f0.(T) == g(T). (**)

Now (**) shows that g(T) leaves Yr7 invariant and a glance at (*) guarantees
that T0.) = (..\ T0.)g(T)0. = Therefore, is invertible on
Yny, and hence a contradiction. This contradiction implies E

and so a(T0.) C a.
-For the reverse inclusion, let E a and assume that a(T0.). So,
there exists some S E L(Y0.) such that S(..\ = (..\ = 10.. Next,
consider the analytic function h: U k C defined by h(,u) = 0 if
E and = if E V0.'. Clearly, = for each

E So, Theorem 6.29 implies h(T)(.\ T) = (.\ T)h(T) =
Consequently,
[sf0.(T) + h(T)] T) = T) [Sf0.(T) + h(T)]
= (T) + (T) = I.
This shows that the operator T is invertible on X or
a(T), contrary
to the assumption that E a a(T). Thus, ..\ a(T0.), and so a(T0.) = a.
Finally, we shall prove the uniqueness of the reducing pair. To this end,
assume that (Y Z) is another reducing pair for T satisfying
a and a Jordan
contour in surrounding and C2 a Jordan contour
in V0.' surrounding a(T). Now note that if y E Y, then we have
f =
This implies y E Y0., and so Y Y0.. Similarly, Z c Z0.. If v Y0., then
writev=y+zEYEDZ=X, andnotethatz=vy EY0.nZ0.={O}.
That is, z = 0 or v y E Y. Thus, Y0. Y, and hence Y = Similarly,
Z= and the uniqueness of the reducing pair has been established.
264 -- - - -- 6. Spectral Properties
It should be noticed that in the extreme cases a = cr(T) and a = 0 we

have = (X,{o}) and (Y0,z0) = ({O},X).
A special class of spectral sets is provided by the isolated points of
the spectrum of an operator T L(X). If is an isolated point in the
spectrum u(T) of an operator T L(X), then there exists some r > 0 such
that r) fl cr(T) = In this case, we know that R(. , T) has a unique
Laurent series expansion on the open annulus {..\ E C: 0 < <r}.
The unique series expansion of R(., T) in such an annulus will be referred
to as the Laurent series expansion of the resolvent around
Adopting the standard complex-variable terminology, we shall say that
an isolated point in the spectrum cr(T) of a bounded operator T E L(X)
is a pole of order k 1 of the resolvent R(., T) if in Laurent expansion
R(.\,T) = around we have A_k 0 and = 0 for
all n> k. If k = 1, then is called a simple pole. If 0 for infinitely
many n E N, then we shall say that the resolvent R(., T) has an essential
singularity at Regarding isolated points, we have the following results.
Lemma 6.35. If T E L(X) and is an isolated point in cr(T), then is
either a pole or an essential singularity of the resolvent T).
Proof. Consider the Laurent series R(.\, T) = (.\ around
the point If were not a singularity, then = 0 for each n 1 and
so T) = A0 E L(X). But (.\T)R(.\, T) = R(.\, T)(.\T) = I,
and so letting we get T)A0 = T) = I, contrary to
the hypothesis that cr(T).
Lemma 6.36. Let T E L(X) and let If the operator T has

E cr(T).
finite ascent and descent, then is an isolated point of the spectrum of T.
Proof. Assume that cr(T) is such that T) = T) = p < oo.
Let Y = T)P) and Z = T)P). By Theorem 2.23 the T-
invariant closed subspaces Y and Z satisfy X = Y Z and the operator
T is quasinilpotent on Y and invertible on Z.
The fact that T is quasinilpotent on Y implies that the spectrum
of the operator TIy in L(Y) is {O}. Therefore, the spectrum of T

restricted to Y is That is, =
Since the operator T is invertible on Z, it follows that there exists
a neighborhood V of such that the operator Tjz is invertible on Z

for each E V; see Corollary 2.11. This implies that for each E V \
the operators and are invertible on Y and Z, respectively.
But then for each V \ {.\o} the operator ,\ T TIy)
(..\ (..\ TIz)
E13
is also invertible on X; see Exercise 17 in Section 6.1. Therefore, is an

isolated point in cr(T).
The coefficients of the Laurent series expansion of the resolvent around

an isolated point of the spectrum of an operator satisfy some useful identities
that are included in the next lemma.
Lemma 6.37. Let T: X + X be a bo'unded operator on a Banach space
and let A0 be an isolated point of the spectrum of ci(T). If
R(A, T) =
is the La'urent series expansion of the resolvent function R(., T) aro'und A0,
then the coefficients of the expansion satisfy the identities:
A_1 = I+(TAo)Ao,
= (T for each n 0, and
= for each n.
In partic'ular, for each n 1 we have
(T = A0 and = (T
Proof. Multiplying R(A, T) = A T
and using that (A T)R(A, T) = I, we obtain
I= (AT)CAAo)mAn
fl 00
= A0) + (A0 T)J(A
00 00
= + T)(A AO)71A7.,
= + (Ao
The uniqueness of the Laurent series expansion yields I = A_1 + (A0 T)Ao
and + (A0 T)A72 = 0 for each n 0. These are the first two desired
identities.
If we multiply R(A, T) = A
then as above we can obtain that = for each n. The last two
identities are straightforward.
If A0 is an isolated point of the spectrum, then the set a = {A0} is clearly
a spectral set. For simplicity we shall write P\0 (T) rather than (T).
In this case, we shall also write and instead of the more laborious
symbols and
Following the standard terminology from the theory of complex vari-

ables, we shall call the operator A_1 in the Laurent series expansion
R(A,T) the residue operator at A0. Remark-
ably, when A0 is an isolated point, the residue operator coincides with the
spectral projection (T). The details are included in the next result.
Lemma 6.38. If A0 is an isolated point in the spectrum of an operatorT in
L1(X), then in the Laurent series expansion R(A, T) =
around A0 the residue operator A_1 is a projection that coincides with the
spectral projection associated with {Ao}. That is, A_1 = PA0(T).
Proof. Since A0 is an isolated point of the spectrum u(T), -there exists

somer>Osuchthat{AEC: O<IAAoI <3r}flcr(T)=. LetCbe
any positively oriented circle with radius less than 3r centered at A0. By
Theorem 1.78 the coefficients of the Laurent series expansion of the resolvent
function R(A, T) around A0 are given by
=
f dA for n =0, 1, 2
In particular, we have
A1 =
For the rest of the proof the symbol C5 will denote the positively oriented
circle centered at A0 and having radius 8 > 0. Now consider the open set
V = D(Ao, r) U {z E C: 2r < 1z Aol <p}, where p > 0 is large enough so
that u(T) V. Clearly, 3V = Cr U Cp U (C2r), where denotes the
circle oriented negatively.
Let f be the function defined by
f(A)cf0 if
if
Clearly, f belongs to 2(T) and is identically 1 on the open disk D(A0, r).
So, by the definition of the spectral projection P,\0 (T) associated with the
spectral set {Ao}, we see that f(T) = PA0(T). It remains to notice that
A_1 =
f R(A, T) dA = f(A)R(A, T) dA = f(T) = PA0(T).
The proof is finished.
We are now ready to characterize the isolated points of the spectrum

that are poles of the
Functional Calculus 267
Theorem 6.39. Let A0 be an isolated point in the spectrum of an operator

T E and let RCA, T) = be the Laurent series
expansion of the resolverit around A0. Then the following statements are
equivalent.
(1) The point A0 is a pole of the resolvent of order k.
(2) We have (T 0 and (T = 0.
(3) The ascent and descent of the operator A0T are both finiteand
in this case they are equal to the order of the pole A0.
Proof. (1) (2) We have A_k 0 and = 0 for all n > k. But
then it follows from Lemma 6.37 that (T = 0 and
(T A0)kA_1 = A_(k+1) = 0.
(2) (3) From Lemma 6.38, we know that A_1 = PA0(T). So, we
have (T 0 and (T = 0.
We shall verify first that N((Ao = N((A0 We already
know that N ((Ao C N ((Ao To prove the reverse inclusion,
let x E N((Ao and assume by way of contradiction that x does
not belong to that is, the vector y = (Ao 0. It
follows that (Ao T)y = (Ao = 0 or Ty = Aoy. This implies
(see Exercise 3 at the end of this section) that f(T)y = f(,\o)y for each
f In particular, letting f = f0., where f0. is the function defined
before Theorem 6.34, we obtain that = y. Consequently,
0 = (T Ao)kPAO(T)x = PA0(T)(T Ao)kx = = y 0,
which is a contradiction. Hence, N((Ao T)k) = N((Ao This
shows that the operator A0 T has finite ascent and that p = a(Ao T) <k.
In particular, we have N((A0 = N((Ao for all n p.
Next, notice that (T 0 guarantees the existence of
some vector x in the range of the projection PA0(T) such that
(T = (T 0.
From (T Ao)kx = (T = 0, it follows that
- N((Ao -
This shows that p k is also true. Thus, p= k.
Next, we consider the unique decomposition X =YA0 ZA0 described
in Theorem 6.34, where ZA0) is a reducing pair for T. ) = {A0}
and cr(TIZA ) = cr(T) \ {Ao}. In particular, the operator A0 T is invertible
on ZA0. This implies that for each n E N the operator (A0 T)Th is also
invertible on ZA0.
Since PA0(T)(X) = YA3, the identity (T (T) = 0, implies that

= 0 on YA3. Consequently,
= T)k) =
Thus, T has finite descent, and Lemma guarantees that the descent
T) = k.
(3) (1) Assume thatT) = T) = p < oc. If we let

Y= T)P) and Z = T)P), then by Theorem both

Y and Z are T-invariant closed subspaces satisfying X = Y Z, and the
operator T is quasinilpotent on Y and invertible on Z.

From the identity T = we get a(T) = a(TIy) U a(TIz).;

see Exercise 17 of Section Since a(Tlz) and = we
see that = a(T) \ This conclusion implies that Y = YA3 and
Z= see Theorem 6.34. In particular, PA3(T) is the projection of X
on Y along Z. So, if we take an arbitrary n > p and x E X, then from
PA3(T)X E Y = T)P) = it follows that
= (T =0.
That is, = 0 holds for all n > p, and so is a pole of the resol-
vent R(., T) whose order k is less than or equal to p. But then from the
implication (2) (3), we see that necessarily k = p.
Corollary 6.40. If T E (X) and E a(T) is a pole of order k of the
resolvent function R(., T) around then:
(1) is an eigenvalue of T.
(2) YA0 _T)k) and Z.A3 = _T)k).
Proof. Let T E (X) and assume that E a(T) is a pole of order k of
the resolvent function T) around According to Theorem 6.39 we
have (T 0 and (T =
If we choose any vector
0.
x E X such that y = 0, then clearly the vector y satisfies
Ty = Therefore, is an eigenvalue.
The identities YA3 = T)k) and ZA3 = follow
immediately from Theorems 6.34 and 2.23.
Corollary 6.41. If X is a finite dimensional complex Banach space and
T e 12(X), then every point in the spectrum of T is a pole of the resolvent
function
Proof. Fix T E 12(X). From dim X < Do it follows that the spectrum of T
is finite, it consists of eigenvalues, and obviously for each E C the operator
T has finite ascent and descent. The conclusion now follows immediately
from Theorem
Functional Galcul us 269
Exercises
1. Show directly (LeO, by avoiding the Spectral Mapping Theorem) that
if p(z) = a polynomial, then for every bounded operator
T: X X on a Banach space we have
a(p(T)) = p(a(T)) = {p(A): A e a(T) }.

[HINT: Fix ji C and write ji p(z) = an(z zi)(z Z2)' (z zn),
where z17. , are the n solutions of the equation p(z) = Then
=
Now notice that ji e a(p(T)) if and only if E (7(T) for some
2. Assume that f is an analytic function defined on a neighborhood V of CX),
that is, the function f is defined and is analytic on an open set of the form
V = {A e C: > R} for some R> Suppose also that f vanishes
at in V surrounding the spectrum of an
operator T E (X), then show that Jc f(A)R(A, T) dA = 0.
3. Let T: X X be a bounded operator on a Banach space. Fix any
function f e and let C be a Jordan contour surrounding cr(T) in
the domain of analyticity of f. Show that
f(T)x =
f f(A)R(A, T)x dA
holds for each x E In particular, if Tx = ax for some x e X and a

scalar a E C, then show that f(T)x = f(a)x.
4. Let T e (X), let f e and let g e J(f(T)). Establish that the

composition go f belongs to and that (go f)(T) = g(f(T)).
5. Let T: X X be a bounded operator on a Banach space, and let V be
an open neighborhood of a(T). Assume that a sequence of analytic
functions defined on V converges uniformly to some function f: V C.
Show that f e J(T) and that
6. Let T: X X be a bounded operator on a Banach space, and let a be
a spectral set of Also, let (Y, Z) be the pair of reducing subspaces as
determined in Theorem 6.34. If h E J(T), then show that
=
7. Let X be a Banach spaceS Show that there do not exist operators 5, T in
(X) such that ST TS = I, the identity operator on L [HINT: Use
Exercise 19 in Section
8. Let T: X X be a bounded operator and let a be a spectral set of T.

Also, let (Ya, Za) be the unique T-reducing pair of subspaces such that
X= Zr,, a(TIyj = a, and a(TIza) = a(T) \ a; see Theorem
If we let Ta = then show that
an ar(Tc7) = afl ar(T), and = afl

9. Let be an isolated point in the spectrum of an operator T E L(X).

Show that the resolvent of T has an essential singularity at Acj if and only
if either the ascent or the descent of T is infinite.
11. Let T e L(X) and let be an isolated point in the spectrum a(T) such
that every neighborhood of contains points that are not eigenvalues of
T. Show that is a pole of the resolvent R(., T) if and only if A0 T
has finite descent.
12. If a quasinilpotent operator T e L(X) satisfies p = <00, then show
that = 0. [HINT: Use the precediiig exercise.]
13. For a spectral set ci of an operator T e L(X) establish the following.
(a) If 0 a, then there exists an operator S E L(X) such that
= TS = ST.
(b) If 0 a, then there exists an operator S E L(X) such that
=I+TS=I+ST.
[HINT: Let be the unique T-reducing pair such that = X,
and let Pa(T) be the projection of X onto along Assume 0 o.
This means that the operator T: f is invertible. Let R E
be the inverse of and put S = (T).]
14. Let a be a spectral set of an operator T e C(X). If there exists some
r >0 such that a {A eC: Al <r} and a(T)\a {A E C: Al >r},
then show that the range of the spectral projection is given by
15. (R. Melton) For an entire function f: C f C having the series expansion
f (A) the following statements are equivalent.
(a) Each coefficient of the expansion is non-negative.
(b) For each positive operator T on any Banach lattice E the operator
f(T) = is positive.
(c) There exists an infinite dimensional Banach lattice E such that f(K)
is positive for each positive compact operator K on E.
Chapter 7
Some Special Spectra
This chapter is a continuation of the previous one. Our objective here is to

investigate the structure of the spectra of several classes of operators. The
first- section provides a complete description of the spectrum of compact
and strictly singular operators. We establish here that each strictly singular
(in particular, each compact) operator has an at most countable spectrum
whose non-zero points are isolated eigenvalues. If the spectrum of such an
operator is countable, then it forms a sequence that converges to zero. These
results are illustrated with several examples of compact integral operators.
The second section of this chapter presents a useful method of converting
approximate eigenvalues of an operator into eigenvalues by enlarging the
Banach space upon which the operator is acting. The third section studies
the structure of the spectrum of lattice homomorphisms. The spectrum
of these operators exhibits a remarkable degree of symmetry, and the main
results of this section show that the spectrum of each lattice homomorphism
and of each interval preserving positive operator is cyclic. The fourth section
introduces the order spectrum of an order bounded operator and studies
some basic properties the order spectrum.
The fifth and final section of this chapter is devoted to the essential
spectrum of an operator. The essential spectrum of an operator T E (X) is
the non-empty compact subset of the spectrum of T consisting of all complex
numbers for which the operator T fails to be a Fredhoim operator. This
is the same as the spectrum of T in the Calkin algebra (X)/IC(X)
and, therefore, there exists a possibility of studying the essential spectrum
of T in the context of Banach algebras. A discussion regarding connections
between the essential spectrum and measures of non-compactness closes the
section.
271
272 7. Some Special Spectra
7.1. The Spectrum of a Compact Operator

Recall that an operator T: X Y between two Banach spaces is said to
be compact whenever the image under T of the unit ball of X is a norm
totally bounded subset of Y (or, equivalently, whenever the image under
T of each bounded sequence in X has a convergent subsequence in Y). A
compact operator is, of course, automatically continuous.
The objective of this section is to establish that the spectrum of a com-
pact operator is an at most countable set. As before, X denotes a complex
Banach space.
An operator T E (X) is said to be power compact if there exists
some k E N such that is a compact operator.
Theorem 7.1. If T E (X) is a power compact operator (in particular, if
T is compact) and X is infinite dimensional, then 0 cr(T).
Proof. Assume that some power of T, say S = is a compact operator.
If 0 cr(T), then T is invertible and so S = Tk is also invertible, i.e.,
51 exists. But then the identity operator I = 88_i: X X is compact.
Hence, X must be finite dimensional, a contradiction.
The next result is crucial for the description of the spectrum of a compact
operator.
Lemma 7.2. If T E (X) is a compact operator, then for each > 0 the
set of all eigenvalues of T with modulus greater than , i.e., the set
{AEap(T): AI>},
is finite (and possibly empty).
Proof. Assume by way of contradiction that there exists a sequence of
distinct eigenvalues satisfying > > 0 for all n. For each n pick a unit
vector such that = Clearly, the set of eigenvectors {xi, X2,. .} .
is a linearly independent set; see Exercise 2 at the end of Section 6.2.

Let be the linear span of {x1,... , Then:
o Each is invariant under T, i.e., c
o is a proper closed subspace of for each n 2.
By Lemma 4.44, for each n there exists a unit vector E such that
forall
We claim that E Indeed, if = ajxj, then
= E Va_i.
=
Therefore, for n > m we have Tym + )'tnyn E and hence

Tym = + Tym)
= I Yn (Tym + Tyn)
This shows that {Tyn} cannot have any convergent subsequence, contrary
to the compactness of T. This completes the proof.
And now we come to a complete description of the spectrum of a compact

operator.
Theorem 7.3. For any compact operator T e we have:
(1) The spectrum of T is at most countable.
(2) Every non-zero point of the spectrum is an eigenvalue.
(3) If the spectrum of T is countable and )'2,. .} is any enumer-
.
ation of cr(T), then 0. In particular, every non-zero point in
the spectrum of T is an isolated point.
Proof. Let T e be a compact operator. From Lemma 7.2 it follows

that T has at most countably many eigenvalues and that if {'\17 )'2,. .} is
.
any sequence of distinct eigenvalues, then 0.
Our next goal is to show that each non-zero boundary point of the
spectrum is an eigenvalue, i.e., each e 9a(T) \ {0} belongs to
So, let,\ e 9a(T) \ {0}. By Theorem 6.18(2), we know that belongs to
the approximate point spectrum of T. So, there exists a sequence of
unit vectors such that T guarantees
that some subsequence of {TXn} (which we shall denote by {TXn} again)
converges. From = we see that the sequence {Xn}
is also norm convergent, say x. Since Ifr'nM = 1 for each n, we infer
that lxii = 1. Using once again that TXn 0, we obtain Tx = 0,

or Tx = ,\x. Thus, every non-zero boundary point of 9cx(T) is an eigenvalue
of T, and hence 9a(T) is at most countable.
To finish the proof, we shall show that a(T) has no interior points.
Indeed, if o-(T) had an interior point then it is easy to see that a(T)
would contain a boundary point in every direction from i.e., 9cr(T) would
be uncountable. So, o-(T) = 9a(T), and the proof is finished.
If T: X X is a compact operator and X is infinite dimensional, then

zero is always a boundary point of the spectrum of T. This implies that when
X is infinite dimensional, zero belongs to the approximate point spectrum
of T.
Corollary 7.4. For a power compact operator T E L(X) we have:
(1) Its spectrum is at most countable.

(2) If the spectrum of T is a countable set and {A1, A2,. .} is an arbi-
.
trary enumeration of a(T), then p 0.
Proof. The conclusions follow immediately from Theorem 7.3 and the fact
that = A E cr(T)}. (Use here either the Spectral Mapping
Theorem 6.31 or Exercise 11 in Section 6.1.)
Corollary 7.5. If T E L(X) is a compact operator, then each non-zero
A0 E cr(T) is a pole of the resolvent of T.
Proof. Assume that T: X p X is a compact operator on a Banach space

and let A0 be a non-zero point in the spectrum of T. By Theorem 7.3 we
know that A0 is an isolated point of cr(T). Also, by Lemma 4.45, the operator
A0 T has finite ascent and descent. Now a glance at Theorem 6.39 sho
that A0 is a pole of the resolvent of T.
As expected every finite-rank operator has a finite spectrum.
Theorem 7.6. The spectrum a finite-rank operator T: X p X on a
of
Banach space is a finite set and consists of eigenvalues.
If X is infinite dimensional and T = 0 is the unique rep-
resentation of T with respect to a basis ui, .. , . of the range of T, then
the spectrum of T consists of zero together with the eigenvalues of the k x k
square matrix A (ui)].
Proof. If X is finite dimensional, then the result is obvious. So, assume
that X is infinite dimensional. Clearly, T is not one-to-one and so Tx = 0
for some x 0, which shows that 0 is an eigenvalue of T.
Assume that Tx = Ax for some A 0 and x 0 and let T =
be the representation of T with respect to a basis {ui,... uk}
, of T(X). It
follows that
j=1,...,k.
From x = we get x E T(X). Thus, x = appropriate

for
(uniquely determined) scalars Now if denotes the column vector with

coordinates cu,. . and A = then it is easy to see that the
equality Tx = Ax is equivalent to = Aa. This shows that the non-zero
eigenvalues of T are precisely the non-zero eigenvalues of the k x k square
matrix A.
The next result describes a large class of compact integral operators.
Theorem Assume that (5, d) is a compact metric space and bt is a

Borel measure on S. If K: S x S p R is a continuous function, then the
integral operatorT: defined by
Tx(t) = [K(s, t)x(s)

is
is a compact operator for each 1 p oc.
Proof. Assume 1 p oc. Fix some to E and let ft By the

S uniform
continuity of K on S x S there exists some 8 > 0 such that d(to, t) <8 implies
K(s,t) K(s,to)f < for each s E S. So, if d(to,t) < 8 and f
satisfies If lip 1, then Holder's inequality implies
Tf(t) -Tf(to)1 = f[K(s,t)
where += 1. Hence, {Tf: If 1} is a and equicon-

tinuous subset of C(S). From the classical AscoliArzel Theorem (see [31,
Theorem 7.9, p. 60]), we obtain that {Tf: 1} is a II
bounded subset of C(S), and consequently a Ii lip-totally
bounded subset
of In other words, the integral operator T is compact.
Here is another important example of a compact integral operator.

Example 7.8 (The Volterra Operator). Consider the jointly measurable
function K: [0, 1] x [0,1] R defined by
lo if 0<t<s<1
K(s,t)=<
if
and let V: 1] 1] (1 p oc) be the integral operator with
kernel K. That is, for each x E 1] we have
r1 rt
Vx(t)= / K(s,t)x(s)ds= / x(s)ds.
Jo Jo
This integral operator is known as the Volterra operator. According to
Theorem 5.29 its adjoint operator V*: Lq{O, 1] Lq[O, 1] is given by
fl I'1
V*y(s)=J y(t)dt,
0 S
where 1 q oc satisfies + = 1.
The Volterra operator is compact. Indeed, for each x in 1] it follows

from Holder's inequality that
Vx(t1) - Vx(t2)[ = f
Jt1
It1
1
for all t1, t2 E [0, 1]. This implies that { Vx: x I } is a bounded and
equicontinuous subset of C[0, 1], and from this we infer that V is a compact
operator.
The spectrum of V consists of the zero element alone, i.e., a(V) = {0}.
The fact that 0 E a(V) follows immediately from Theorem 7.1. On the
other hand, if 0, then for every y E 1], it is easy to check that the
equation
(\V)x=y
has a unique solution x (see Exercise 3 at the end of the section) given by
x(s)
This shows that V)' exists for each 0, and therefore a(V) = {O},
V is a quasinilpotent operator. Finally, it should be noticed that 0 is
not an eigenvalue of V.
We continue with an important result (known as the KreinRutman

theorem) regarding the spectral radius of a compact positive operator. Since
the peripheral spectrum of an operator is a non-empty compact set, there
always exists a point in the spectrum that lies on the circle centered at zero
with radius r(T). In the case of a positive operator, we get some extra
information about the peripheral spectrum. Namely, as we shall see next,
the spectral radius of a positive operator always belongs to the spectrum of
the operator.
Theorem 7.9. The spectral radius of a positive operator on a Banach lattice
belongs to the spectrum of the operator.
Proof. Let T: .E + .E be a positive operator on a Banach lattice. For

each n put = r(T)+ Clearly, p(T) and
r(T) E a(T), it suffices to show (in view of Corollary 6.9) that
= 00.
To see this, start by fixing some E a(T) with = r(T) and let
= for each n. Then E p(T), = for each n, and
By Corollary 6.9, we have = oc. Next, for each n pick
a unit vector E such that
131 '7'\ 1 ni rr'\

The Spectrum of a Compact Operator 277
From Lemma 6.23, we have T) Hence
and so T)tIc = 00, as desired. D
And now we come to a classical result due to M. C. Krein and M. A. Rut-

man [192].
Theorem 7.10 (KreinRutman). If T: E E is a compact positive oper-

ator on a Banach lattice with r(T) > 0, then the spectral radius r(T) is an
eigenvalue having a positive eigenvector. That is, there exists some x > 0
such that Tx = r(T)x.
Proof. Let T: E E be a compact positive operator on a Banach lattice

with r(T) > 0. From Theorems 7.9 and 7.3 it follows that r(T) is an
eigenvalue of T. What we do not know yet is that this eigenvalue has a
positive eigenvector. We shall accomplish this next. The elegant proof
below is due to F. F. Bonsall [68].
Let = r(T) + and so .j. r(T). Now Theorem and Corol-
lary 6.9 guarantee that T)tIc = oc. Since T) 0 holds
for each n, it follows from Corollary 3.23 that T) = T)
Consequently, for each n there exists a positive unit vector E E such that
= > 0.
Put = and note that is a positive unit vector. From
1 ( n' jYn
= r(T)]xn =
we obtain that
1 2 /
+
By the compactness of T, the sequence has a convergent subsequence,
which we shall denote by again. From r(T) > 0 and (*), it follows
that x for some positive unit vector x. Using (*) once again, we obtain
Tx = r(T)x, as desired.
We close the section by proving that the structure of the spectrum of a

strictly singular operator is identical to that of a compact operator. How-
ever, the proof for strictly singular operators is much more involved and for
this reason we have chosen to keep the much simpler proof of Theorem 7.3
for compact operators as well.
278 -- - Special Spectra
Theorem 7.11. If T (X) is a strictly singular operator on an infinite

dimensional Banach space, then:
(1) Its spectrum cr(T) is at most countable,
(2) The point zero belongs to the spectrum and is the only possible ac-
cumulation point of ci(T).
( 3) Each point of the spectrum is an eigenvalue.
Proof. We begin by observing that =

belongs to cr(T). Indeed, if
0
o a(T), then the operator T would be invertible, which is impossible since
T is strictly singular.
From Theorem 4.63 we know that for each non-zero E C the operator
T is Fredholrn and that T) = il7J) = 0. By virtue of part (c) of
Lemma 4.55, there exists some open disk centered at and such that
for each ,a E the operator ,a T is Fredholm and its nullity n(,a T)
and defect d(,a T) are constant on the set \ Moreover, we claim
that T) = d(,a T) = 0 on with the possible exception of the
point
To establish this claim, fix some 0 and then choose any point
with > and such that the closed line segment L = joining
and does not include zero. Since the operator ,a T is invertible whenever
11TH, it follows that n(,a T) = d(,a T) = 0 for all ,a close to and
consequently for all ,a E \ by our choice of the disk
Thecompactness and connectedness of the segment L guarantee the

existence of a finite number of disks . , such that they cover
.
L and fl 0 for 1 j k 1. (See also Exercise 7 in

Section 1.5.) Since the functions n(,a T) and d(,a T) constant on are
each set \ it follows that they are constant on the union of the
disks D(e3) except possibly for the points e',. Moreover, since on
\ } both functions n(,a T) and d(,a T) are equal to zero, it
follows that T) = d(,a T) = 0 for all ,a E \ In
particular, this implies that n(,a T) = d(,a T) = 0 for all ,a E \
In other words, the operator ,a T is invertible for all ,a in with a
possiMe exception of the center Since 0 is arbitrary, this proves that
each non-zero point in the spectrum a(T) is an isolated point.
Next, we shall show that the set cr(T) is at most countable. For this, it
suffices to verify that for every E > 0 the set = E cr(T): PtI E} is
finite (and possibly empty). Since is compact and does not contain zero,
there exists a finite collection of disks D(ei),. . . , as described earlier
that cover As shown above, only the centers ti,. . of these disks
can belong to the spectrum, and so c {ei,. . ,
Finally, let us prove that each non-zero E cr(T) is an eigenvalue. Recall

that is an isolated point in the spectrum, and consequently we can consider
the spectral subspace XA corresponding to ). By Theorem 6.34 the closed
subspace XA is T-invariant and cT(TA) = where TA denotes as usual the
restriction of T to XA.
We claim that dim XA <oc. To see this, assume by way of contradiction
that dim XA = oo. Since T is strictly singular on X, it follows that the
operator TA is strictly singular on XA. But then the conclusion at the very
beginning of the proof implies 0 E cT(TA) = contrary to 0. Hence,
the Banach space XA is finite dimensional.
Now notice that the operator TA: XA -4 necessarily has an eigen-
value. As is the only point in the spectrum cr(TA), we conclude that is
an eigenvalue of TA, and therefore of T as well.
Exercises
1. Show that the only compact operators with closed range are the finite-
rank operators.
2. Consider the interval [0, 1] equipped with its Lebesgue measure and for
each 1 p oo define the positive integral operator T: 1] + 1]
by
Tf(t) (s2 + t2)f(s) ds.

= j
0
Show that T is a finite-rank integral operator and find its spectrum.

3. Consider the Volterra operator V: 1] 11(1 p < oo) defined
by
Pt
Vf(t)= / f(s)ds.
JO
Show that V is quasinilpotent, i.e., o(V) = {0}, by proving that for each
0 and each g [0, 1] the equation V) f g has a unique
solution given by
f(s) =
f + kg(s).
[HINT: Consider the operator 8: 11 + 1] defined by
Sf(t) =
f +
Prove that S is a bounded operator satisfying = =f

for each continuous function Use the fact that C[0, 1] is norm dense in
1] to conclude that S =
4. Show directly (i.e., without using the previous exercise) that the Volterra
operator has no eigenvaluesand conclude from this that the Volterra
operator is quasinilpotent.
5. Let T: E + E be a positive contraction on a Banach 'attice. Show that
1 belongs to the approximate point spectrum of T if and only if r(T) = 1.
6. Show that the hypothesis r(T) > 0 cannot be dropped from the Krein
Rutman Theorem 7.10.
7. Show that the compactness hypothesis of the operator in the Krein
Rutman Theorem 7.10 cannot be dropped. [HINT: Use the shift operator
in Example 6.21.]
8. Let T: E * E be a positive operator on a Banach lattice and let p be a
polynomial with non-negative coefficients. Show that r(p(T)) = p(r(T)).
[HINT: Use Exercise 1 of Section
9. Prove the following generalization of the KreinRutman Theorem 7.10.
If a power compact positive operator T: E * E satisfies r(T) > 0, then
there exists some vector x > 0 such that Tx = r(T)x. [HINT: Assume
that r = r(T) > 0 and that Tk is compact for some k> 1. According to
Theorem 7.10 there exists some x > .0 such that Tkx = rCx. Now notice
that if S = then y = Sx> 0 and
Tyry=TSx--rSx=(Tr)Sx= (Tic _rc)x=O,
proving that Ty = ry.]
10. Prove Lemma 4.18 using the KreinRutman theorem. That is, show by
using the KreinRutman theorem that if ci is a compact Hausdorif space
without isolated points, then the only weakly compact multiplication op-
erator on C(ci) is the zero operator. [HINT: Assume that there exists
a non-zero weakly compact multiplication operator on C(ci). Then
= is a non-zero compact positive multiplication operator;
see [30, p. 337]. So, replacing q5 by q52, we can suppose that q5> 0 and
that is compact. Clearly,
1 1
r = = lim
1
= = >
and so by the KreinRutman Theorem 7.10 there exists some 0 < x E

C(ci) such that Mq,x = thx = rx, or (q5 r)x 0. The latter guarantees
that q5(w) = r for all in a non-empty open subset V of It follows that
the closed unit ball of C(V) is compact. This implies that V (and hence
V) must be a finite set. But then this guarantees that ci has isolated
points.]
11. Show that there exist a Banach space X and a non-empty compact subset
K of C such that u(T) K for each T E L(X). (See also Exercise 15 in
Section 6.1.)
7.2. Turning Approximate Eigenvalues into Eigenvalues 281
12. Establish the following generalization of Theorem 7.11: If S (X) is

strictly singular and an operator T E (X) has an at most countable
spectrum, then the spectrum of S + T is at most countable and the points
of o(T) and zero are the only possible accumulation points of o-(S + T).
13. (T. Oikhberg and V. Tftoitsky) Let q5 be a non-zero multiplicative func-
tional on the Banach algebra (X), where X is an arbitrary Banach space
with dimX > 1. Show that:
(a) For each T (X) we have q5(T) cr(T).
(b) If for T (X) there is a reducing pair of closed subspaces (V, W)
such that dim V <oo and T vanishes on W, then q5(T) = 0.
(c) If T (X) is a strictly singular operator (in particular, if T is
compact), then q5(T) = 0.
7.2. Turning Approximate Eigenvalues into Eigenvalues

The purpose of this section is to present a method that allows us to treat
the-approximate eigenvalues of an operator as eigenvalues. The method is
based on the following result that will be proven in this section: If X is a
(real or complex) Banach space, then X can be enlarged to a new Banach
space X such that every bounded operator T: X >X can be extended to a
bounded operatorT: X k X satisfying cr(T) = cr(T) and cra(T) =
We start with some preliminary discussion regarding sums of sequences
of Banach spaces. Let Xi, X2,... be a sequence of Banach spaces all over
the same field or C). For each I p < oc we define the of the
sequence by
(Xi,x2,...) EHXn:
We also define
(Xi = (xl,x2,...) E flXn: <0},

n=1
and
limIlxnII=O}.
fl*oO
n=1
It is also customary to denote these Banach spaces by 4(Xi
and co(Xi X2 E13 ) In the case X = = we shall write
and
It is a routine matter to verify that under the coordinatewise operations

of addition and scalar multiplication, each 4-sum is a Banach space. If
-- 7. SorneSpecial Spectra
each is also a Banach lattice, then under the coordinatewise ordering

and lattice operations the are likewise Banach lattices.
Clearly, the c0-surn of the sequence } is a closed subspace of its
sum. If, in addition, each is a Banach lattice, then the of the
sequence is a closed ideal in the of the sequence. Moreover,
if each is an AM-space with unit then the of the sequence
is also an AM-space with unit e = (ei,e2, .).
Definition 7.12. If X is a Banach space, then we let
=
the quotient Banach space of over the closed subspace co(X).
Clearly, X is a Banach lattice if X is a Banach lattice. The vectors of

X are equivalence classes and, as usual, for each x = (xi,x2,...) E
its equivalence class will be denoted by {x]. It should be obvious that two
vectors x = (xi, x2,...) and y = (Yi, in are equivalent (i.e.,
[x] = {y]) if and only if = 0. This implies that for each
x = (xi, x2,...) E the norm of the vector [xJ E X is given by
jj[x]jj = inf = lim sup
yE[xJ
Our next objective is to show that the Banach space X has the properties
stated at the beginning of the section. -
First of all, notice that there is a natural mapping x from X into

X, defined by = (x, x, x,. .). That is, is the equivalence class of the
constant sequence (x, x, x,. .). It is a routine matter to verify that the
mapping x is a linear isornetrywhich is also a lattice isornetry if X
is a Banach lattice. Thus, identifying X with its image X, we can assume
without loss of generality that X is a closed subspace of the Banach space
X. In case X is a Banach lattice, X is also a Banach lattice and X is a
closed Riesz subspace of X.
Next, assume that T: X X is a bounded operator. Let [xJ E X,
and pick an arbitrary sequence (xi, x2,...) E [x]. Since (Xi, x2,...) is a
bounded sequence in X and T is continuous, the sequence
is also bounded. Moreover, if y = (yi, y2,..) E [x], then x7,, > 0,
and so by the continuity of T we get = + 0. This
implies that the two sequences (Txj, Tx2,...) and (Tyi, Ty2,...) belong to
the same equivalence class, which we shall denote by T([x]). That is, the
bounded operator T gives rise to a mapping T: X X via the formula
.)]) = [(Tx1,Tx2,. . .)}.
7.2. Thrning Approximate Eigenvalues into Eigenvalues 283
A direct verification shows that T: X X is a linear operator, which is

also bounded. As a matter of fact, ITIL = IITI.
To see this, start by observing that for each x = (x1, x2,...) we have
= tI[(Txi,Tx2, . = limsup
< limsup
1 1
T is that Now if
x E X satisfies < 1, then = lxii 1, and so 11Th = ITxM.
This implies
ITII sup 1Txl = tTII
lix
and hence = 11Th.
Next, notice that if x E X and, as before, denotes the equivalence class

with representative (x, x, .), then the constant sequence (Tx, Tx, Tx,...)
.
is a representative of the cilass T. This shows that for each x E X we have

Tth = [(Tx, Tx, Tx,. .)]. That is, T is a bounded linear extension of T from
.
X to all of In other words, we have established the following theorem.

Theorem 7.13. If X is a Banach space and T: X X is a bounded
operator, then T: X X is a linear extension of T to all of X satisfying
ITI1 = IT1. Moreover, when X is a Banach lattice, T is positive if and only
if T is positive.
More properties of the mapping T T are included in the next results.
Theorem 7.14. If X is a Banach space, then the mapping T F-4 T, from
L1(X)to L1(X), is a linear isometry that is also an algebraic homomorphism,
i. e., ST = ST for all 5, T E (X).
Proof. It should be clear that the mapping T T is linear and (in view of
Theorem 7.13) it is also an isometry. To see that this mapping is an algebraic
homomorphism, note that if 5, T E (X) and x = (x1, x2,...) E
then
= [(STxi, STx2,. .)] = .
Tx2,..
=
Therefore, ST = ST, and so T T is an a'gebraic homomorphism.
Theorem 7.15. A bounded operator T: X X on a Banach space is
invertible if arid only if T: X X is invertible. Moreover, if T is invertible,
then (T)' (T').
Proof. Assume first that T is invertible, and so TT' = T1T = I. Since

S is an algebraic homomorphism, we have TT-' = T-'T I = I.
This shows that T is invertible and that (T)' = (T-').
For the converse, suppose that is invertible. Since T is an extension of
T, it follows that T is one-to-one. To see that T is surjective, fix some y E X.
Pick some x = (xl,x2,...) E such that T[x] = = [(y,y,.
This means that y, and consequently T is
invertible, we see that is a Cauchy sequence in X, and so is a
Cauchy sequence in X. Since X is a Banach space, we have x in X,
and consequently T in X. This implies y,i = T, or y = Tx. So, T
is also surjective, and hence it is an invertible operator.
We are now ready to establish the desired the relationships between the
spectra of the operators T and T. We start with the complex version of this
result.
Theorem 7.16. If T: X X is a bounded operator on a complex Banach
space, then:
(1) p(T) and =
(2) = = cTp(
(3) R(A,T) = R(A,T) holds for each A E p(T).
Proof. (1) If A E C, then it follows from Theorems 7.14 and 7.15 that the
operator A T is invertible in L1(X) if and only if A T is invertible in (X).
This shows that p(T) = p(T), and so o(T) = C \ p(T) = C \ p(T) = o(T).
(2) To prove that 0a(T) = notice that:
A E 0a(T) c X such that 74 0 and I1AZTh 0
[z] = [(zi, Z2, . .)] 0 such that T[z] = A[z]
AE 0a(T), and let > 0. This implies that there exists some
vector [z] = [(zi, z2, .)] E X satisfying
1
[z] = lirn sup =I and I
= lim sup
Th-400
Then there exists some k such that < holds for all n k. Since
= 1, there are infinitely many m k such that >
This shows that A E 0a(T), and so 0a(T) Since 0a(T) 0a(T)
obviously true, we see that 0a(T) = c7a(T).
7.2. Turning EigenvaJues into EigenvaJues 285
(3) If A E p(T), then a glance at Theorem 7.15 shows that
- =
The real version of Theorem 7.16 can be formulated as follows.
Corollary 7.17. Let T: X X be a bounded operator on a real Banach

space and let T the complexifi cation X given
by the formula + zy) = Tx + iTy. Then we have:
(1) and a(Tc)=a(Tc).
(2) cra(Tc) = =
(3) holds for each e p(Tc).
We close the section by mentioning that the Banach space X satisfying

the properties of Theorem 7.16 is not unique. Instead of X we can use any
ultrapower of X. In this case, the corresponding analogue of Theorem 7.16
takes the following form (whose proof is left for the reader).
Theorem 7.18. If T: X p X is a bounded operator on a Banach space,

then for every ultrafilter U on N we have:
(1) a(T)=cr(Tu).
(2) cra(T) = =
Exercises -
1. Show that if X is a Banach space, then for each x = (x1, x2,...)

the norm of the vector [x] X is given by
= inf =
yE[xJ
2. Assume that is a sequence of real Banach spaces. Show that (with

the obvious identifications) for each I <p < oo we have
Also show that
a...) = co(X1 X2 ico(X1 X2 a...).

3. Let be a sequence of real (or complex) Banach spaces, and assume

that 1 <p, q <oc satisfy +=1. Show that
where the equality is understood subject to the natural duality
(xx*)
for all x = (Xl,X2,...) in (X1 EDX2 and all = in

)q. Likewise show that
= and
=
4. Let be a sequence of all real (or complex) Banach spaces. Fix any
ForeachT=(T1,T2,...)in
(L(X1) define the mapping X * X by
= (T1x1,T2x2,...).
(a) is a bounded operator on X, i.e., E (X).
(b) The mapping T from (L(X1) .. to (X), is
a linear isometryand so for each 1 p < oc the Banach space
(L(X1) (X2) is a closed subspace of the Banach space
5. Verify that the mapping x from X to defined by = [(x, x, x,..

,
is a linear isometrywhich is also a lattice isometry if X is in addition a
Banach lattice.
6. If X is a Banach lattice, then show that abounded operator T: X * X
is positive if and only if the operator T: X X is positive.
7. If T: X +X is a surjective operator on a Banach space, then show that
T: X + X is also surjective. Conversely, if T has a closed range and T
is surjective, then show that T is also surjective.
8. Show that a positive operator T: E * E on a real Banach lattice is
a lattice homomorphism if and only if T: E * E is likewise a lattice
homomorphism.
9. Let X be a real Banach space with complexification X Es-
tablish the following:
(a) A sequence + is bounded, that is,
z= (x1 + + 7'y2,...) E
if and only if the sequences x = (xi, x2,...) and y = (Yi, Y2,...) in X

are both bounded, i.e., x, y In this case, write z = x +
(b) Show that the mapping T: X defined by
T([x + [x] +
isasurjective
(topological) isomorphismwhich is a lattice isomorphism if X is
also a Banach lattice.
[HINT: For the continuity of T note that
fT({x + = 11[xI + = sup cos9 + [y] sin 911
9ER
= sup 1 [(xi cos 9 + sin 9, cos 9 + Y2 sin 9,. .
= sup [urn sup cos9 + sin
= IL[x+zy]IJ.]
00
IT I
Figure 1
10. Let T: X* X be a bounded operator on a real Banach space, and let

T: X p X be its extension to the Banach space X. If X iX is
the complexificationof X, then in the preceding exercise we defined the
mapping T: X iX by
T([x + zy]) = [x] + z[y]
for all x (xi, X2,...) andy (yl,y2,...) in 00(X) and claimed that it
was a surjective (topological) isomorphism. Now establish the following
properties.
(a) The diagram shown in Figure 1 is commutative.
(b) The spectra of the operators T, T + iT coincide. That is,
a(T) = = a(T+zT), = = and
cia(T) = =
(c) If X is a Banach lattice and T is a lattice homomorphism, then
=
holds for all z E
7.3. The Spectrum of a Lattice Homomorphism

The objective of this section is to establish that the spectrum of a lattice
homomorphism has a nice symmetric property known as cyclicity. In our
discussion here we shall use the letter r to denote the unit circle, i.e.,
r={AEC: A(=1}.
Keep in mind that the unit circle is a multiplicative group.
Definition 7.19. Let A be a subset of complex numbers. We shall say that

a complex number is a cyclic point of A if E A for each
integer k E Z. The set A is called cyclic if every point of A is cyclic.
All open and closed discs centered at zero are cyclic sets and likewise
so is any set consisting of roots of unity for some n, that is, any set of the
form {z E C: = 1} = {e p = 0, 1,. , n I }. More generally, any
. .
subgroup of I' is a cyclic set. Also, according to this definition, every subset
of real numbers considered as a set of complex numbers is a
cyclic set.
H. B. Schaefer [290, Proposition 4.2, p. 324] has shown that the periph-
eral spectrum of a Markov operator enjoys the following "cyclicity" property.
Theorem 7.20 (Schaefer). Let E be an AM-space with unit e, and let
T: E E be a Markov operator. Then the set of eigenvalues
E={AET: TcZ=AZforsomeZEEc with IZI=e}
is a subgroup of the unit circle C (and hence E is a cyclic set).
Proof. Since Te = e, we have 1 E E. By Theorem 3.20, we can assume that
E = C(11) and = for some compact Bausdorif space where
the unit e corresponds to the constant function 1 on ft Also, we know
that E* = ca and = ca zca where ca is the
of all regular (finite) Borel measures on Since T is a Markov operator,
T* carries regular Borel probability measures to regular Borel probability
measures. For simplicity, we shall denote by T.
Now ffx w E and let = T*&,. Assume that there exist Z E
andAETsuchthat ZJ=landTz=Az. Thenwehave
and so f
AZ(w) = (TZ)(w) = (TZ, = (Z, = (Z, /2w)
=
= 1. Since P\'[Z(w)]'Z(t)I =
f Z(t)
1 for each
t E it follows (how?) that ?C'Z(t) [Z(w)j' = 1 for each t lying in the
support of This implies Z(t) = AZ(w) for all t in the support of
Similarly, if another E I' satisfies TZ1 = A1Z1 for some E
with JZi 1 1, then Zi(t) = A1Z1(w) for all t in the support of and so
= for all tin the support of Hence,
AA1Z(W)Zi(W)
= (TZZ1,
= f Z(t)Zi(t)
= (TZZ1)(w).
= = T*8W)
Thus, T(ZZ1) = AAizzi, so that A similar argument shows that if

A E E, then A E E, where A is the complex conjugate of A. Hence, E is a
subgroup of T.
Corollary 7.21. Let T: E + E be a positive operator on a Banach lattice

having a positive eigenvector u corresponding to an eigenval'ue r > 0, i.e.,
T'u = ru. If T has another eigenval'ue of the form A = corresponding
to an eigenvector z E with Izi u, then A is a cyclic point of the point
spectrum of T.
Proof. Let S = and note that the positive operator 8: E + E satisfies

S'u 'u. This implies that 8: * E
and = u, it follows from Theorem that eikO is an
eigenvalue of S for all k E Z. Therefore, the complex number A = is a
cyclic point in the point spectrum of the operator T.
Corollary 7.22. The point spectrum of a lattice homomorphism is cyclic.
Proof. Let T: E + E be a lattice homomorphism, and let Tz = Az for

some non-zero z E E a glance at
Theorem 4.29 shows that
Now use Corollary 7.21 to conclude that E for each k E Z.
The next remarkable result is due to E. Scheffold [293].
Theorem 7.23 (Scheffold). The spectrum of a lattice homomorphism is

cyclic.
Proof. Let T: E + E be a lattice homomorphism. Combining Exercise 8

of Section 7.2 with Corollary 7.17, we can assume without loss of generality
that = Also, from Corollary 7.22, we know that is cyclic.
So, what needs to be shown is that if A = Ale' E or(T), then AlethO E a(T)
for each k E Z.
To this end, let A = AlezO E crr(T). Then A E crp(T*); see Theorem 6.19.
Therefore, there exists some non-zero vector E = satisfying
T* Az*. This implies
IAI . z*1 = T*Iz*I.
Now if we fix some real number A0 > r(T), then according to Lemma 6.24
there exists some 0 < q5 E E* such that the principal ideal generated by q5
in E* is E and A0q5. By Theorem 3.39,
there exists a unique lattice homomorphism Tgt,: + whose adjoint
makes the diagram in Figure 2 commutative. (The construction of the
space is described in Section 3.4.)
Figure 2
From the construction of the space and the fact that T is a lattice
homomorphism, for each x E E we have
AJ =
=ci(TIxI)
= =
=
This implies for each z E i.e., is a bounded
below operator.
Now observe that if 0 then T is a surjective lattice isomorphism,
and since is a lattice isometry, T*: is also a lattice isomorphism.
Since A E it follows from Corollary 7.22 that
E o(T)
for each k E Z.So, in this case, A is a cyclic point of the spectrum cr(T).
Next, assume that 0 E Since T17!, is bounded below, it follows (see
Exercises 5, 6, and 8 of Section 6.2) that
D(O, Aj) = cYp(T*JE;) =
That is, the open disk centered at zero with radius Al lies in o(T), and so
A is automatically a cyclic point of o(T), and the proof is finished.
Corollary 7.24. The spectrum of an interval preserving positive operator
is cyclic.
Proof. If T: E E is an interval preserving operator, then T*: E* _4

is a lattice homomorphism; see Theorem 1.35. Therefore, by Theorem 7.23,
the set o(T) = is cyclic.
We conclude the section with a brief comment on some recent devel-

opments regarding the symmetry of the spectrum of an operator. Recall
that the spectrum of an operator is said to be rotation invariant if for
each point A in the spectrum the circle centered at zero with radius tAt lies
entirely in the spectrum.
If T: E E is a disjointness preserving operator on a Banach 'attice,

then there are several general results asserting rotational invariance of cr(T).
For example, this is so if E is a Banach function space and Titm A Titm = 0
whenever ri Tn; for details see Chapters 12 and 13 in [21].
Exercises
1. Let G = {z1,.. . , be a subset of the unit circle F consisting of p distinct

numbers. If G is closed under multiplication, then show that
G= {z C: zP = 1} = k
[HINT: Fix some z E G. Since each of the p+l elements z, z2, .. . , is

in G, necessarily zm = for some 1 m <e p+l. This implies =1
for some 1 <ri <p. If we let ii = min{ri N: 1 <m < k and = 1},
then H = {1, z,. . . , z"'} is a subgroup of G and so p is divisible by ii.
This guarantees that = 1.1
2. Fix an angle 0 E R and let G = n E N }. Establish the following.
(a) If 0 is a rational multiple of 2ir, then G is a finite subgroup of the
unit circle F (and so G consists of roots of unity).

(b) If 0 is not a rational multiple of 2ir, then G is dense in F.
3. (Kronecker's theorem) Let G be an arbitrary subgroup of the unit circle
F. Show that either G is dense in F or else it consists of roots of unity,
i.e., G = k = 0,1,.. .,pl} for some p N.
4. Establish the following result that was used in the proofof Theorem 7.20.
is a regular Borel probability measure on a compact Hausdorff space
and some Z satisfies = 1 and Z(t) d,u(t) = 1, then Z(t) = 1
for each t in the snpport of a.
5. Let T: E p E be an interval preserving positive operator. If A = A1e26

belongs to the residualspectrum of T, then show that for each integer m
the complex number is an eigenvalue of the adjoint operator T*.
6. Assume that for some continuous function p on a compact Hau-
sdorff space the multiplication operator Ivia: p is a lattice
homomorphism.
Show directly (i.e., without using Theorem 7.23) that the spectrum
of the multiplication operator is cyclic.
7. Let T: X X be a bounded operator on a Banach space. If the ap-
proximate point spectrum of T is rotation invariant, then show that the
spectrum of T is also rotation invariant.
7.4. The Order Spectrum of an Order Bounded Operator

The lattices can be used to define a new concept
lattice structure of Banach
of a spectrum for a bounded operator that allows a deeper look at the order
292 7. Some Sped a] Spectra
properties of the inverse operator. This is the order spectrum which is

defined as follows.
Definition 725. Let T : E E be a bonnded operator on a (real) Banach
lattice. The order spectrum o0(T) of T consists of all complex nnmbers A
for which the operator A T does not have an order bonnded inverse on the
complexification of E. That is,
ci0(T) = {A E C: A T does not have an order bonnded inverse on Ec}.
Notice that if E is a Dedekind complete Banach lattice and T: E E

is an order bounded operator, then T belongs to the complex unital Banach
algebra and its spectrum in 4(Ec) obviously coincides with the
order spectrum of T. So, when E is Dedekind complete and T: E E is
order bounded, the order spectrum o0(T) is a non-empty compact subset of
C. Letting r0(T) = Al, we have
r0(T) lim
(See Theorem 6.16 and the discussion preceding it.) It should be clear that
for any bounded operator T we have the inclusion o(T) a0(T).
In what follows, we will restrict our attention to order bounded oper-
ators. Let us mention at once that in general the order spectrum a0(T)
can properly include the usual spectrum a(T) (see Example 7.36 below).
Nevertheless, there exists a method that allows us to reduce the notion of
the order spectrum to the usual spectrum. More precisely, each Dedekind
complete Banach lattice E can be embedded into a larger Dedekind com-
plete Banach lattice E in such a way that each order bounded operator T
on E will have an order bounded extension D to this enlargement satisfying
= a(T). The purpose of this section is to describe this method.
We begin by assuming that E is a Riesz space and that A is an arbitrary
fixed index set. Only this is necessary to define the extension E.
It is well known, and easy to check, that the Cartesian product EA under
the pointwise operations is also a Riesz space. That is, EA is a Riesz space
under the operations defined by
[f + g}(a') + g(a), = and [f V
g all E E A. The positive cone of EA consists
of all positive functions, that is, functions f E EA with E for each
There is a natuxal embedding of E into This embedding simply

assigns to each x E E the constant function A E with value x, i.e,
= x for each E A. It should be clear that x is indeed a lattice
isomorphism and so E (identified with its image in EA) can be viewed as
a Riesz subspace of Ek avoid introducing unnecessary notation, from

To
now on, every vector x E E will play a double role: On the one hand, it can
be viewed as an element in E and, on the other hand, it can be considered
as the constant function x on A; that is, if x is considered as a function on
A, then x(a) = x for each a E A. If clarity requires to differentiate between
the vector x E E and the constant function x on A, then we shall use the
symbol The next result is left as an exercise.
Lemma 7.26. A Riesz space E is Dedekind complete if and only if the Riesz
space EA is Dedekind complete.
The ideal generated by E in will be our "enlargement" space.

Definition 7.27. The ideal generated by E in E4 will be denoted E, i.e.,
E= {f E EA: E snch that f(a)I <x for each a E A}.
Assume now that is a lattice norm on the Riesz space E. There is a
natural way to extend . a lattice norm on E. Namely, for each f E E
we let
XEE and (*)
It is a routine matter to verify that (*) indeed defines a lattice norm on E
that extends the original norm of E. In actuality, as the next result shows,
more is true.
Lemma 7.28. If E is a Banach lattice, then E eqnipped with the lattice
norm given by (*) is a Banach lattice containing E as a closed vector snb-
lattice.
For any mapping T: E f E there is a natural extension D: EA EA

given by the formula
= T(f(a)).
where f Notice that T is linear if and only if T is linear.
E A.
The reader should verify also that if S, T: E E are two operators and A
is an arbitrary scalar, then
(S + = + t, = At, and =
Regarding invertibility of operators we have the following.
Lemma 7.29. Let T: E p E be an operator on a Riesz space. Then:
(1) T is one-to-one if and only if T: E E is one-to-one.
(2) T is snrjective if and only if 1': E'1 EA is snrjective.
Proof. (1) Assume that the operator T is one-to-one and let Tf = 0. This
implies = T(f(a)) = 0 for each E A. Since T is one-to-one, it
follows that = 0 for each a E A, f = 0. Rence, is one-to--one.
Now suppose that T is one-to--one on E, and let Tx 0. So, = 0,
and thus = 0. This implies x = 0, and hence T is one-to-one.
(2) Assume that T: EA pjA
is surjective, and let y E E. Pick some
f E EA such that Tf = and then note that for each index E A we have
= T is surjective.
For the converse, assume that T is surjective, and let f E EA. For each
a E A pick some E E such that = Then the function
gE EA satisfies Tg = f, so that is surjective.
Several important order properties of operators transfer to their exten-

sions.
Lemma 7.30. An operator T: E E on a Riesz space is:
(1) positive if and only if T is positive,
(2) a lattice homomorphism if and only if T is a lattice homomorphism,
(3) interval preserving if and only if T is interval preserving.
Proof. (1) Assume that T is positive, and let 0 f E P2k Then for each
E A we have Tf T
a is positive and x E P2+, then for each
index we have 0 = Tx, proving that T is a positive operator.
(2) Assume first that T is a lattice homomorphism, and let f, g E P2k
Then we have
= T([f V =
= V = V
for each A. Thus, i'(f V g) = Tf V and so is a lattice

phism.
For the converse, assume that is a lattice homomorphism, and let
x,y E P2. If is any index, then
Tx V Ty = V = V = V
= V = V T(x V y).
This shows that T is a lattice homomorphism.
(3) Assume that T is interval preserving, and let f 0 and g E
satisfy 0 g Tf. That is, 0 = holds for each
a' E A. ,Since T is interval preserving, for each E A there exists some
7A. The Order Spectrum of an Order Bounded Operator 295
o h(a) f(a) such that T(h(a)) = g(a). Then the function h E EA

satisfies 0 h f and Th = g. This shows that T is interval preserving.
Finally, assume that T is interval preserving, and let x, y E satisfy
o x Ty. This implies 0 Tv', and so there exists some f EA
satisfying 0 I and Tf = . Now take any index a and note that
o f(a) y and T(f(a)) = Tf(a) = = x. This shows that T is
interval preserving.
For the remaining results in this section, we need to add a technical

condition ensuring that the index set A is not too small. Namely, we shall
assume that:
The cardinality of the index set A is at least as large as the cardi-
nality of the Riesz space E.
The next result characterizes the order bounded operators on E as those
whOse extensions to EA leave E invariant.
Theorem 7.31. An operator T: E * E on a Riesz space is order bounded
if and only if the operator T: EA EA leaves E invariant. In this case T
is also order bounded.
Proof. Assume first that T is order bounded, and let f E E. Pick some
xE satisfying x for each a E A. Since T is order bounded,
there exists some y E such that T[x,x] [y,y]. Now note that
Tf(a)1 = y for each a E A. Therefore, Tf E E, and so T
leaves E invariant. Moreover, notice that for each g E E satisfying f
we have Tg This shows that T: E E is indeed an order
bounded operator.
For the converse, assume that T leaves E invariant. Fix some x E E+.
Since the cardinality of the index set A is at least as large as the cardinality
of E, there exists a function f: A E whose range is the entire order
interval [x, x]. In view of our hypothesis, Tf E E and so there exists an
element y E such that Tf I Therefore T[x,x] [y,y}. This
proves that T is order bounded, and the proof is finished.
We are now ready for the basic results of this section.

Theorem 7.32. Let T: E E be an order bounded operator on a Riesz
space. If T: E E also invertible and its inverse
is invertible, then T is
satisfies (11)_i = and is order bounded.
Proof. Let S: E E be the inverse of E E, i.e., ST = =

Part (1) of Lemma 7.29 guarantees that T is one-to-one and the proof of
296 - 7. Some Special Spectra
part (2) of the same lemma shows that T is also surjective. Consequently,
T' existsand, of course, is a linear operator.
From TT1 = T'T = I. it follows that = = I on
Ek Since 1' leaves E invariant, it follows that S = (T)' = (T1)"; see
Exercise 1 at the end of this section. Finally, observing that (T)' leaves
E invariant, we from Theorem 7.31 that T' is also an order bounded
operator.
Recall that every linear operator on E zE is of the form S + iT,
where S and T are linear operators on E. Let us say that an operator S +zT
is order bounded if both S and T are order bounded. Also, if S + iT is an
operator on let us write (S + + zt; clearly, this operator acts
on (EA)c. Notice that if S and T are order bounded, then S + iT leaves the
vector subspace = E zE of (EA)c invariant. The complex version of
Theorem 7.32 is now immediate.
Theorem 7.33. Let 5, T: E E be two order bounded operators on a
Riesz space. If (S + zT)": i is invertible, then S + iT: i is
also invertible, and its inverse is order bounded and satisfies
[(S + [(S + = +
Theorem 7.33 implies at once the result announced at the beginning of
the section about reducing the order spectrum to the standard spectrum.
Theorem 734. If T is an order bounded operator on a Dedekind complete
Banach lattice, then cr0(T) = = cr(T).
Corollary 7.35. Lattice homomorphisms and interval preserving positive
operators have cyclic order spectra.
Proof. Let T: E i E be a positive operator which is either interval pre
serving or a lattice homomorphism. According to Theorem 7.34 we have
cr0(T) = cr0(T) = cr(t). Also, from Lemma 7.30, we know that T: E E is
either interval preserving or a lattice homomorphism, respectively. To com-
plete the proof, it remains to notice that by Theorem 7.23 and Corollary 7.24
these types of positive operators have cyclic spectra.
We conclude the section with the promised example, due to T. And
(see [291]), of an order bounded operator for which the order spectrum is
larger than the usual spectrum.
Example 7.36 (And). For each k consider the finite dimensional real
Hilbert space Ek = of Ic factors of with the stan-
dard coordinatewise inner product. With the usual coordinatewise order
becomes a Dedekind complete Banach lattice.
Consider now the 2-sum E = (E1 the sequence {Ek}. It )2 of

is a Dedekind complete Banach lattice which is a (real) Hubert space under
the inner product defined by
(x,y) =
for all x = (Xi, X2,...) = (yi, Y2,...) in E.

and y
Next, for each k E N we consider the 2 x 2 orthogonal matrix

cost sink
Mk=
for each m E N.
I
cost
Therefore, the modulus of (considered as an operator on R2) is given
by the 2 x 2 symmetric matrix
cos f
sin
k

It is easy to see that the eigenvalues of are f
cos + sin So, the
Eucidean norm of is = cos + sin j.
f
Next, for each k E N we define the operator Tk: Ek Ek via the formula
Tk = Mk . Clearly, Tk is a unitary operator on Ek which is
also order bounded. This implies fITkII = 1. Moreover, the modulus of
is given by ED jA'Ij E
Now consider the operator T = (T1,T2,...) E 4(E); see Exercise 4

in Section 7.2. It should be noticed that T is indeed an order bounded
operator which is also unitary. This implies that T as an operator on
(the complexification of E) is also unitary, and so r(T) = 1.
Fix any 0 < and
sin > for each n no. This and Exercise 4 in Section 7.2 imply that
for alln no we have
= sup
kN
=
>
=
Consequently, we have
r0(T) = n*oc
lim T)
lim [(i + = e> 1= r(T).
So, the spectrum of T is a proper subset of the order spectrum of T.
Exercises
1. Let T: V V be an invertible operator on a vector space, and let
V V be its inverse. Assume that W is a T-invariant subspace of
V such that T: W TV is also invertible. Then show that T' leaves TV
invariant and that the inverse of Tjw (the restriction of T to TV) is the
restriction of T' to W, i.e., (Tlw)1 T11w.
2. Describe the Banach lattice B when B = R.
7. Establish the following elementary properties that were used in Exam-
ple 7.36. Consider the matrix
acting on R2, where is an arbitrary angle. Then:

(a) For each ii E N we have
sin
(b) The modulus of is the matrix
cos nq5j sin
I
sin J
cos nq5j
(c) The eigenvalues of the symmetric matrix are

I cosn15j I sin
(d) The Euclidean norm of the matrix is
= cos nq5J + sin

J
7.5. The Essential Spectrum of a Bounded Operator

In this section, unless otherwise stated, X will denote an infinite dimensional
complex Banach space. As we already know, the vector space K(X) of all
compact operators on X is a closed algebraic ideal in L(X). It is easy to see
that the quotient vector space (X)/K(X) under the algebraic operations
o [S]+[T]=[S+T],
o A{T] = [AT], and
o [S][T] = [STJ
is a unital algebra with the element [I] as its unit. Moreover, the quotient
vector space L(X)/IC(X) under the quotient norm
= S -T e
is a Banach space. Since the quotient norm also satisfies the properties
and 1['IM 1,
L(X)/IC(X) is a unital Banach algebra. This Banach algebra is called the
Calkin algebra' of X and is denoted i.e., = L(X)/IC(X).
The quotient map (also called the canonical projection) of (X) onto
will be denoted by ir. That is, it: (X) is defined by
From the inequality it (T) fj

lIT it follows that it is a contraction. The
above properties are summarized in the following result.
Theorem 7.37. If X is an infinite dimensional Banach space, then the
Calkin algebra t(X) is a unital Banach algebra. The canonical projection
it: L(X) is a contraction and an algebraic homomorphism.
There is standard terminology regarding properties of operators as el-

ementsin the Calkin algebra. It is customary to say that an operator
T e L(X) possess a property (P) essentially if its canonical projectio
it(T) satisfies property (P) in the Calkin algebra
For instance, we say that an operator T E L(X) is essentially in-
vertible if it(T) is invertible in the Calkin algebra Incidentally, the
essentially invertible operators are precisely the Fredholm operators. This
is simply a restatement of Theorem &46.
Theorem 7.38. An operator T e L(X) is essentially invertible if and only

if it is a Fredholm operator.
We now introduce the essential spectrum of an operator.

Definition 7.39. The essential (or the Fredholm) spectrum of an oper-
ator T e L(X) is the spectrum of it(T) in the algebra That is,
oess(T) = {A e C: A it(T) is not invertible in
= {A E C: A T is not a Fredholm operator}.
was introduced by J. W. Calkin in [84].
7. Some Special Spectra
An immediate consequence of Theorem 6. 16 is the following.

Lemma 7.40. The essential spectrum of an operator T (X) is a non-
empty compact subset of o(T).
Since for each T E (X) and each E C we have T* = T)*,

the following result is an immediate consequence of Theorem 4.42.
Lemma 7.41. The essential spectrum of any operator T E (X) coincides
with the essential spectrum of its adjoint, i.e., Oess(T) = Oess(T*).
The next result presents some operators whose essential spectrum

cides with their spectrum.
Theorem 7.42. If an operator T E (X) does not have any non-trivial
closed invariant subspace, then the essential spectrum and the spectrum of
T coincide, i.e., Oess(T) = o(T).
Proof. Assume that T has no non-trivial closed invariant subspaces and let
o'(T). Clearly, the closed vector subspace T) is T-invariant and
hence either T) = X or T) = {O}. If T) = X, then
T = \I, which contradicts our hypothesis. Therefore, T) = {O}.
Now the assumption E guarantees that R(..\T) X, and clearly

T) {O}. Since T) is T-invariant, it follows that T)
is dense in X. Therefore, T does not have a closed range, and hence

it fails to be a Fredholm operator. This implies that the operator T
is not essentially invertible. That E ess(T). Consequently, we have
o(T) c 0ess(T) C o(T) or Oess(T) = cr(T). D
The next result identifies a special class of isolated points in the spectrum
of a bounded operator.
Lemma 743. Let T E (X) and )'o E \ If there is a path
lying outside of ess (T) and joining with a point in the resolvent set p (T),
then is an isolated point of the spectrum o(T).
Proof. Fix E \ and let C be a path joining with a point
E p(T) such that CflO'ess(T) = 0. So, for each z E C the operator zT is
a Fredholrn operator. Now, according to Lemma 4.55, for each z E C there
exists an open disk D(z) centered at z such that the nullity n(( T) and
the defect T) are constant on D(z) \ {z}.
Since the curve C is a compact set, there exists a finite subcollection

{D(zi), D(z2), . .. , from these disks that covers C. By Exercise 7 in
Section 1.5, we can assume without loss of generality that this collection is
also a chain such that = and = 1u. Since for each ( D(zn)\{zn} we
have n(( T) = d(( T) = 0,

easily follows that n(( T) = d(( T) = 0
it
in the open disk D(zi) with the exception of ( = This implies that T
is invertible for all ( \ So, is an isolated point of o(T).
Theorem 7.44. Let be an isolated point in the spectrum of an operator
TEL(X). If then:
(1) The spectral projection P,\0(T) is of finite-rank.
(2) The operator )'o T has finite ascent and descent.
(3) The point is an eigenvalue of T and is a pole of the resolvent of
T whose order equals the ascent (or the descent)
Proof. Let be an isolated point in the spectrum of an operator T E (X)

such that (T). So, T is a Fredholm operator.
(1) Let (T) be the spectral projection corresponding to the spectral

set and let = R(P,\0(T)) and X2 = N(P,\0(T)). Then X =
and (in view of Theorem both X1 and X2 are invariant under T.
Moreover, letting T1 = T1x1 and T2 = we know (by Theorem 6.34
again) that cr(T1) = and cr(T2) = Since \oIx2 T2 E (X2)
is one-to-one and onto, for each n 1 we have
N((\o = Ti)tm) e {0}, and (*)
T)Th) e x2. (**)

Now notice that there exists a (not necessarily closed) vector subspace Y of
X1 such that R(\oIx1 Ti) e Y = Xi. Clearly,
X=X1EBX2 -T1)eYeX2
Since T is Fredholm, Y is finite dimensional, and so from
dim Ti) dim T) <Do,

we see that the operator )'olx, T1 E (X1) is Fredholm, LeO, )'o
But then from cress(Ti) c cr(Ti) = we obtain c7ess(Ti) = 0. This implies
that X1 is finite dimensional. Thus, PA0 (T) is of finite-rank.
(2) and (3) Since X1 is finite dimensional, it follows that \oix1 T1 has
finite ascent and descent. This, in conjunction with (*) and (**), guaran-
tees that T likewise has finite ascent and descent. Now a glance at
Theorem shows that is a pole of the resolvent of T.
Since cr(T1) = {\o} and dim X1 <oo, there exists some non-zero vector
x E Xi such that Tx = T1x = Thus, is an eigenvalue of T. The last
conclusion follows immediately from Theorem o
Corollary 7.45. For an isolated point in the spectrum of an operator

T E (X) the following statements are equivalent.
(a) The spectral projection is of finite-rank.
(b) The operator T is Fredholm, i.e., )'o aess(T).
Proof. The implication (b) (a) has been established in Theorem 7.44.
Now assume that (a) is true.
Let Xi = (T)) and X2 = N(PA0 (T)). to our hypothesis
According
X1 is finite dimensional, X = X1 e X2, and the closed subspaces X1 and
X2 of X are both invariant under T. If T to then
cr(T1) = and cr(T2) = cr(T) \ Since cr(T2), it follows that
T2 E (X2) is an isomorphism.
This implies, on the one hand, T) = N(\o T1) X1, and so

n(T) = dim T) <oo. On the other hand, from X2 c R(\o 1'), it
easily follows that d(T) = dim X/R(\o T) < 00. Therefore, T is a
Fredholm operator.
We now introduce the essential spectral radius of a bounded operator.
Definition 7.46. The essential spectral radius of an operator T E (X)
is the spectral radius of the element ir(T) in the Calkin algebra i.e,
ress(T) r(ir(T))
= E cress(T)}
= lim
flP 00
=
inf jJ TTh j*.
The essentially quasinilpotent operators are defined as follows.
Definition 7.47. An operator T E (X) is said to be essentially
nilpotent (or a Riesz operator) if aess(T) = {0}, or equivalently, if its
essential spectral radius is zero.
The next result presents a useful characterization of the essentially quasi-
nilpotent operators.
Theorem 7.48. An operator T E (X) is essentially quasinilpotent if and
only if every E a(T) \ {0} is:
(1) an isolated point of a(T), and
(2) its spectral projection has finite-rank.

Proof. Assume that T is essentially quasinilpotent, i.e., = {O}, and

let E u(T) \ {O}. Clearly, can be joined with any point of p(T) by a
path lying outside = {o}. So, according to Lemma 7.43, is an
isolated point of u(T). By Theorem 7.44 the spectral projection has
finite-rank.
For the converse, assume that every point E cr(T) \ {O} satisfies (1)
and (2). Now fix E u(T) \ {0}. By (1), we know that is an isolated
point of u(T). Therefore, Corollary 7.45 applies and shows that

Since cress(T) c cr(T) and cress(T) it follows that cress(T) = {0}, i.e., T
is essentially quasinilpotent.
There are several interesting consequences of the preceding result. We

list some of them next.
Corollary 7.49. If an operator T e (X) is essentially quasinilpotent, then
for each\ E u(T) \ {0}:
(a) is a pole of the resolvent of T,
(b) \ is an eigenvalue of T,
(c) the operator T has finite ascent and descent, and
(d) the order of the pole equals the ascent (or the descent) of the
operator T.
Proof. Assume that an operator T E satisfies Uess(T) = {0}, and let

E u(T) \ {0}. By Theorem 7.48, we know that is an isolated point
of u(T). Theorem 7.44 implies that is an eigenvalue of T and that the
operator T has finite ascent and descent and, consequently, from part
(3) of the same theorem is a pole of the resolvent of T whose order equals
the ascent (or the descent) of the operator )'o T.
Corollary 7.50. The spectrum of an essentially quasinilpotent operator is
at most countable and if it is countable, say {)'i, .}, then 0.
Proof. By Theorem 7.48, we know that every non-zero point in the

trum of T is an isolated point of u(T). This implies that u(T) is at most
countable. The rest follows immediately from Exercise 20 of Section 6.1.
Corollary (X) is essentially quasinilpotent, then its resol-
7.51. If T E
vent set p(T) is always a connected set.
Proof. This follows from Theorem 7.48 in conjunction with Exercise 5(a)
of Section 6.3.
Corollary 7.52. The restriction of an essentially quasinilpotent operator

to any invariant closed subspace is again essentially quasinilpotent.
304 - 7. Some Special Spectra
Proof. Suppose that T (X) is essentially quasinilpotent and that Y is

a closed T-invariant subspace of X. According to part (d) of Exercise 5 in
Section 6.3, we have C cr(T).
Now let E \ {O}. By Theorem 7.48, is an isolated point
of cr(T). Let = be the Laurent series ex-
pansion of T) around By Exercise 5(e) in Section 6.3, we know
that the subspace Y is An-invariant for each n. In particular, we have
= and so = =
Since (T) has finite-rank, it follows that (TJy) also has finite-rank
By Theorem 7.44, TJy is essentially quasinilpotent.
The next result characterizes the essentially quasinilpotent operators in

terms of a decomposition property of their resolvent functions.
Theorem 7.53. An operator T E (X) is essentially quasinilpotent if and

only if its resolvent can be written in the form
=
where the functions B: C \ {O} (X) and K: p(T) IC(X) are both
analytic and = 0.
Proof. Assume first that R(\, T) has the stated decomposition. Then the
analytic function B has the Laurent series expansion B(\) =
(see Exercise 6 of Section 1.5) and = 0 (see Theorem 1.77).
From the Neumann series expansion of the resolvent, it follows that
each > r(T) we have = = where we let
for eachri0. Nowifwetakeanyr>r(T) and
consider the positively oriented circle with center at zero and radius r,
then from the Cauchy formula
and the definition of the Riemann integral, we infer that each is a com-
pact operator. In particular, we have = for
each n 0. From = 0, we get
ress (T) lim [Ta]

J
=0.
T is essentially quasinilpotent.
For the converse, assume that the operator T is essentially quasinilpo-

tent. This guarantees the existence of compact operators K0, K1, K2,
such that KThI1 0. Put = TTh for each n 0 and note
that the series = converges for each E C \ {0}. So, B
defines an analytic function on C \ {0}. Since R(\, T) = holds
for each P'I > r(T), it follows that the series = converges
for each .AJ > r(T) to a (necessarily) compact operator. In particular, we
have R(.A, T) B(\) + K(,\) for each > r(T).
The function K: p(T) > (X) defined by =
is clearly analytic. Since = for each > r(T) and p(T) is
connected, it follows that K is the unique analytic extension of K to p(T).
To finish the proof, it remains to be shown that is a compact operator
for each E p(T).

To see this, notice that if (X) > is the canonical projection
of (X) function.f: p(T.).
defined by = 'it T)) K(I)\)), is clearly analytic. iVioreover, the
function g: p(T) > defined by = is also analytic and
holds for each > r(T). Since p(T) is connected, it follows
that = for each e p(T). This implies K(,\)) 0 for each
e p(T). That is, is a compact operator for each e p(T), and the
proof is finished.
Our next goal is to establish that the Calldn algebra can be represented
as a subalgebra of bounded operators on an appropriate Banach space. To
do this, we need some preliminary discussion.
Let U be an ultrafilter on N and let X11, be the ultrapower of X deter-
mined by the ultrafilter U; see Section 1.4. If T E L(X), then there exists a
natural extension of T to an operator T11, on X11, defined by
TU[(Xi,X2,. [(Tx1, Tx2,.. = (Tx1, Tx2,...) +
foreach (xl, x2,. .)e . Because X is a closed vector subspace of Xe,,,

the quotient space X = is a Banach space.
Next, we shall indicate how each operator T E (X) gives rise to an
operator T e L(X). To do this, we need to introduce some notation. First
of all, if x = (xi, x2,...) e (X), let us write
x denote the equivalence class of in X
in place, a sequence x = (xi, x2,. . in
satisfies
0 exists in X.
limu
In particular, notice that for a sequence y = (yl,y2, ..) e we have
e [xv] if and only if there exists some v E X such that v
Each operator T E (X) gives rise to a bounded operator T: X > X

defined via the formula
for each (xi, x2,...) E (X).

Let us verify that the above formula is well defined and that it defines a
bounded operator on X. Fix E [xv]. This means that there exists some
u E X such that u 0. Consequently, Tn 0, and
so (Ty1, Ty2, E [(Ty1, Ty2, . .)u]. This shows that T is well defined,
.
and clearly it is linear. Moreover, for each E [xv] we have
=
= .)u]M
S ITuII
and from this we easily get
(
That is, T: X X is indeed a bounded operator.
The basic algebraic properties of the operators T are included in the
next result. These properties (as well as many others) have been obtained
by J. J. Buoni, R. Harte, and A. W. Wickstead [79], J. J. Chadwick and
A. W. Wickstead [87], and M. P. W. Wolff [340].
Lemma 7.54. For an operator T E (X) we have the following.
(1) T = 0 if and only if T is a compact operator.
(2) T is one-to-one if and only if T is a semi-Fredhoim operator with
finite nullity, i.e., n(T) < 00. -
(3) T is surjective and T has closed range if and only if T is a semi-

Fredholm operator with finite defect, i.e., d(T) <00.
(4) T is invertible if and only if T is a Fredholm operator.
Proof. (1) Fix an operator T E C(X). Assume first that T = 0 and let
{ be a sequence in If we let x = (xi, x2,...) E then from
= [(Tx1, Tx2,. . = 0, it follows that there exists some z E X such
that z. In particular, it follows (see part (3) of Lemma 1.59) that
there exists a subsequence of such that p z in X. This
implies that T is a compact operator.

For the converse, suppose that T is a compact operator and pick any
vector x = (x1, x2,...) E Since the sequence lies in a compact
set, it follows from part (4) of Lemma 1.59 that there exists some z E X such
that z or z 0. This implies that 0 E [(Tx1, Tx2,.
and so T=
T is a semi-Fredholm operator with dim N(T) <00.
By the definition of we know that the range R(T) of
7.5. The Essential Spectrum ofa BoundedOperator 307
T is closed. Since N(T) is finite dimensional, there exists a closed vector

subspace V of X such that X = N(T) V. In particular, the operator
S= : V p R(T) is an isomorphism.
Now assume that a sequence x = (x1 , X2, . .) E
. satisfies = 0.
Clearly, this is equivalent to saying that there exists some z E X such that
Now for each n write = + with E N(T) and E V,
and note that both sequences and are bounded. Since N(T) is
finite dimensional, the sequence lies in a compact set, and so there
exists some y E N(T) such that y. From = z in X, it
follows that z in R(T), and hence S'(z) in X.
Consequently, = y + 8'(z) holds in X. This shows that
0, and so T is one-to-one.
For the converse, assume that the operator T is one-to-one. Let a se-
quence in N(T) satisfy 1 for each n. Now consider the sequence
x = (x1, x2,. E and note that TFfi = [(Tx1, Tx2,. .)u] = 0. Since
the operator T is one-to-one, we get = 0. This means that there exists
some y E X such that y. By Lemma 1.59(3), there exists a subse-
quence of that is norm convergent to y. This shows that the closed
unit ball of N(T) is a compact set, and hence the vector space N(T) is finite
dimensional.
Next, pick a closed vector subspace W of X such that X = N(T) W.
To establish that T has closed range, it suffices to show that T is bounded
below on W. If this is not the case, there exists a sequence of unit
vectors in W such that U
p 0. In particular, we have 0,
and from this we infer that the vector w = (wl,w2,...) E satisfies
TW = 0. Using again that T is one-to-one, we obtain ii = 0. This means
that there exists some u E X such that u. By Lemma 1.59(3), we can
assume (by passing to a subsequence if necessary) that liwn uM + 0. Thus,
U E W and = 1. Now the continuity of T implies IITwm
Tu T u = 0, contrary
to = 1. This contradiction establishes that T is bounded below on W,
and so R(T) is closed.
(3) Assume first that T is surjective and that T has closed range. Also,
suppose by way of contradiction that X/R(T) is infinite dimensional. Then,
using Riesz's Lemma 4.44 in conjunction with Exercise 1 of Section 3.4, we
see that there exists a sequence {yn} of unit vectors such that for each Ti
and ally E R(T) Span{yi, ,
we have
(*)
Put y = (yi, Y2,...) E Since T is surjective, there exists some

x = (x1,x2,...) E such that = This means that for some
vector a E X we have a 0. Consequently, by Lemma 1.59(3),
there exists a strictly increasing sequence of natural numbers such that
I aJJ 0. By passing to another subsequence if necessary we
can also assume that
aM
<1 (**)
for all m. Now using (*) and (**), we see that
< +
= II
< aM +
which is impossible. Therefore, X/R(T) is finite dimensional.

For the converse, assume that T is a serni-Fredhoim operator such that
X/R(T) is finite dimensional. By the definition of semi-Fredholrnness the
range R(T) is closed and so there exists a finite dimensional vector subspace
Z of X such that X = R(T) Z. Let y = (yl,y2,. To finish
the proof, we must show that there exists some x = (x1, X2,...) E
such that = -
To this end, for each m let = + where E R(T) and E Z.

Since R(T) is closed, it easily follows that both sequences r = (r1, r2,...)
and z = (z1, z2,...) belong to (X). Since lies in Z (which is finite
dimensional) and is bounded, it also lies in a compact set, and so according
to Lemma 1.59(4) there exists some b E X such that b or, equivalently,
= 0. Next, notice that (in view of Corollary 2.15) there exists a sequence
x = (xi, x2,...) E (X) such that = for each m. Finally, notice
that the identities = + are equivalent to = + This
implies
= = + = Tx + = Tx,
and tlie proof of this part is finished.
(4) This follows immediately from properties (2) and (3)
To continue our discussion, we now need to introduce the notion of the

measure of non-compactness for sets and operators.
Definition 7.5 5. If A is a bounded subset of a metric space, then the mea-

sure of of A is the imfimum of allr > 0 such that
A can be covered by a finite number of open balls of radius r. That is,
X(A)=inf{r>0: such that
A a metric space we let = 00.
If X is a Banach space and T E (X), then the measure of non-
compactness of T is defined by
=
The next result contains several elementary properties of the measure of
non-compactness that foflow easily from the preceding definition.
Lemma 7.50. For the measure of non-compactness we have:
(1) For each subset A of a metric space we have = x(A).
(2) A subset A of a metric space is relatively compact if and only if
=0.
(3) A bounded operator T E (X) is compact if and only if = 0.
(4) If X, Y and Z are Banach spaces, T E (X,Y) and S E (Y,Z),
then
the following formula, due to V. G. Troitsky [322], re-

We also have
garding the measure of non-compactness for bounded subsets of a separable
Banach space.
Lemma 7.57 (Troitsky). If A is a bounded subset of a separable Banach
space, then
x=(xl,x2,...)EAN}.
Proof. Let A be a bounded subset of a Banach space X. Fix p>
and then pick a1, a2,. . . E A such that A B(ak, p). Now let x
(xi, x2,...) E AN and for each 1 <k m put Nk = {n E N: E B(ak, p)}.
Clearly, N = Nk. So, according to Lemma 1.55, for some 1 <k <m
we have Nk E U. If we let y = (Xi ak, x2 ak,...) E e(x), then E and
so (by Exercise 5 of Section 1.4) we have IIiiL(II = limu <p.
This implies
If = i.e., if A is a norm totally bounded set, then

0, 0 for all
X = (X1,X2,...) E So, we can assume that > 0, i.e., that A is not
norm totally bounded.

Since X is a separable Banach spaceS there exists a countable dense
subset Z = z2,. . .} of X. Fix r such that 0 < r < This implies
that for each n the set A is not a subset of So, for each n
there exists some x E We
shall finish the proof by showing that =
To see this, start by observing that from Jjxn zkII r for each n> k
it follows that zkII = lirnu r, and so zkjl for
each k. Since Z is dense in X, we have 4 for each z E X.
But then we have
d(x,X)
That is, = as
The importance of the measure of non-compactness of an operator is

highlighted in the next result.
Corollary 7.58 (Troitsky [322]). If X is a separable Bariach space, then

for each operator T E (X) we have =
Proof. Assume that X is a separable Banach space and let T E (X).

Using Lemma 7.57, we see that
=
=
= }
=
=
The preceding corollary is also valid for non-separable Banach spaces,

provided that the chosen ultrafilter U satisfies a certain "saturation"
tion. For details see [322}.
The proof of the next lemma is a direct consequence of the preceding
results and is omitted.
Lemma 7.59. If X of non-
is a separable Bamach space, them the
compactness x: (X) * R is a serninorrn. This seminorm induces a norm
on the Calkin algebra
And now we come to the identification of the Banach algebra as

an algebra of bounded operators on a Banach space.
Theorem 7.60 (Troitsky [322]). If X is a separable Banach space, then

the mapping [TJ T, from to (X), is a well-defined algebraic
isomorphismandso can be considered as an algebra of operators on
the Banach space X.
In addition, if is equipped with the norm then the mapping
[T] 'p T is a linear isometry from into (X).
Proof. From part (1) of Lemma it follows immediately that the map-
ping [Tj 'p T, from to (X), is well defined. Clearly, this mapping
is also linear.
To see that it is an algebraic homomorphism, let 8, T E (X) and fix
From
[((ST)xi,(ST)x2,...)u}
= [(S(Txi),S(Tx2),. .
= Tx2,. .)u] .
=
we obtain that ST = ST so that the mapping [T] 'p T is also an algebraic
homomorphism.
The fact that [TJ T is a linear isometry when is equipped with
the norm follows from Corollary 7.58. This guarantees that the operator
{T] T is also an algebraic isomorphism.
Next we will establish an interesting formula, due to R. D. Nuss-
baum [258], for calculating the essential spectral radius of a bounded oper-
ator.
Theorem--7.61--(Nussbaum). -For a bounded operator T: X X on a
Banach space we have the following.
(1) The spectrum of T coincides with the essential spectrum of T, i.e.,
a(T) aess(T).
(2) If X is separable, then the essential spectral radius of T satisfies
Tess(T) = r(T) = lim = Em
Proof. (1) Using part (4) of Lemma 7.54 and the definition of the essential
spectrum, we see that:
AE A T is not invertible
A T is not Fredholm
AEc.Tess(T).
This implies o(T) cress(T).
(2) This follows immediately from part (1) and Corollary 7.58.
We close the chapter with a discussion regarding a decomposition prop-

erty of the essentially quasinilpotent operators. If T = Q + K for some
K E K(X) and some quasinilpotent operator Q E (X), then we have
= 0.
This implies that 7r(T) is quasinilpotent in the Calkin algebra i.e., T
is essentially quasinilpotent.
T. T. West [333] has proven that the converse of the preceding conclusion
is true for Hubert spaces. We state this result below. Its proof is beyond
the scope of this monograph and is omitted.
Theorem 7.62 (West). A bounded operator T: H H on a Hubert space is
essentially quasinilpo tent if and only if there exist a quasinilpotent operator
Q E (H) and a compact operator K E i'C(H) such that T = Q + K.
We shall say that an operator T E (X) has a West Decomposition

if T can be written in the form T = Q + K, where Q is quasinilpotent and
K is compact. As noticed before, if an operator has a West decomposition,
then it is essentially quasinilpotent.
We shall also say that a Banach space X has the West Decomposition
Property if every essentially quasinilpotent operator on X has a West
decomposition. In this terminology, Theorem 7.62 states that:
Every Hilbert space has the West Decomposition Property.
The following question arises here:
o Does every Banach space have the West Decomposition Property?
The answer to this question for general Banach spaces is unknown.
K. R. Davidson and D. A. Herrero [99] have shown that the Banach spaces
CO and 4 for 1 <p < 00 have the West decomposition property.
There is another special part of the spectrum of an operatorknown
as the Weyl spectrumthat is related to the essential spectrum and to the
West decomposition property. Its definition is as follows.
Definition 7.63. The Weyl spectrum of an operator T E (X) is
the intersection of the spectra of all compact perturbations of T, i. e.,
fl u(TK).
KEIC(X)
7.5. The Essential Spectrum ofa Bounded Operator 313
Clearly, c o-(T).
it is not known if there is a particular
compact operator K such that cr(T K) = That is:
Is the Weyl spectrum of a bounded operator equal to the spectrum
of a compact perturbation of the operator?
A positive answer was obtained by J. G. Stampfli [311] for Hubert spaces

and by K. R. Davidson and D. A. Herrero [99] for co and 4-spaces. For other
results regarding the West Decomposition Property and the Weyl spectrum
in general Banach spaces see [99, 163, 346, 347, 348]. For more about the
Calkin algebra see the monograph by S. R. Caradus, W. E. Pfaffenberger,
and B. Yood [85].
Exercises
1. For the identity operator I on a Banach space X establish the following:
(a) Xis infinite dimensional if and only if = 1.
(b) X is finite dimensional if and only if [I] = 0.
2. Let A be a unital Banach algebra, and let G(A) denote the non-empty
collection of all invertible elements of A. Show that G(A) is an open
subset of A and that the mapping x 'p x1, from G(A) to G(A), is
continuous. In particular, G(A) is a topological group. [HINT: Note that
if e is the unit element of the algebra A and if < 1, then the element
e a is invertible and (e a)' =
a that some x E A satisfies tx < 1a1IV Then
from Ie xa'tj IRa x)a'IL <1, it follows that
xa1 = e (e xa') is an invertible element. This implies that x is
invertible. For the continuity of x x1 see Exercise 11 of Section 2.1.]
3. Let 5: 4 where 1 p < Do, be the forward shift defined by
S(xi,x27x3,...) = (0,xI,X2,x3,...).
Show that the essential spectrum of S coincides with the unit circle, that
is,
4. Show that if an operator T: X X on an infinite dimensional Banach

space satisfies cr(T) = {A0}, then = {A0}.
5. Show that the power compact operators on a Banach space are precisely
the essentially nilpotent operators.
Deduce from this that every bounded operator on a Banach lattice
that is dominated by a compact positive operator is essentially nilpotent.
6. Let A0 be an isolated point in the spectrum of an operator T E
Show without using Theorem 7.44 that if the spectral projection (T)
is of finite-rank, then A0 is a pole of the rcsolvciit of T.
7. Some Special Spectra
7. Give an example of an operator T e L(X) and some A e a(T) \

that is not an eigenvalue of T. [HINT: Consider the forward shift operator
8: 2 2, Le, 8(xi,x2,..) = (O,Xl,X2,.. .), and note that 0 is not an
eigenvalue of S while 0 e cr(S) \ aess(S).
8. Show that every strictly singular operator on an infinite dimensional Ba-
nach space is essentially quasinilpotent. [HINT: Assume that T E (X)
is strictly singular and that A E a(T) \ {0}. From Theorem 7.11 we know
that A is an isolated point of a(T). Now according to Exercise 13 of Sec-
tion 6.4 there exists an operator S e (X) such that (T) = TS = ST.
This implies that the operator (T) E L(X) is strictly singular. Since
(T) is the identity operator on its range, it follows that (T) has
finite-rank. Now use Theorem 7.48.]
Chapter 8
Positive Matrices
It is well known that operators on or can be represented by square

matrices with real or complex entries. The spectrum of an operator on
or coincides, of course, with the spectrum of the matrix representing the
operator and therefore it consists of the eigenvalues of the matrix. iViatri-
ces with non-negative entries enjoy many special spectral properties. For
instance, the spectral radius of a matrix with non-negative entries is always
an eigenvalue corresponding to a positive eigenvectoralthough, of course,
the spectral radius of an arbitrary matrix need not be an eigenvalue.
The purpose of this chapter is to present a general discussion of the
spectral properties possessed by square matrices with non-negative entries.
Matrices with non-negative entries are also known as positive (or
negative) matrices. These are precisely the matrices corresponding to
positive operators on TRY1 or Because of their importance and wide
range of applications, positive matrices have been studied extensively by
many people in various disciplines. For instance, in engineering positive
matrices appear quite often in control systems of the forrri
(t) = Ax(t) + Bu(t),
where A and B are appropriate matrices, x(.) is the function and

u(.) is the control function. In economics, positive arc associated
with production. If 0 is the amount of the I to produce
one unit of the good (where i, j 1,... , ri), then th' matrix
A= is called the production matrix or the inpiikfflitPut matrix of
the economy. Now if x 0 is the production vector the component
denotes the amount of the good produced) and i' 0 deSignates the
consumption vector (where the component is the ait ,iiiit of the good
315
316 8. Positive Matrices
used for consumption) , then the "supply = demand" equilibrium condition

yields
x=Ax+c or (IA)x=c.
Predicting the consumption vector c (input), the above equation allows us
to find the needed production vector x (output). For extensive treatments
on positive matrices and their applications, we refer the reader to the mono-
graphs [57, 58, 157, 198, 243]. For a rigorous study of finite dimensional
vector spaces the reader can consult [145].
When positivity is combined with the notion of irreducibility, it results
in several remarkable spectral properties the most important of which are
described by the celebrated PerronFrobenius theorem. As usual, a square
matrix is called irreducible if (considered as an operator) it does not leave
invariant any non-trivial ideal. For positive matrices this is equivalent to
saying that for each i and j there exists some natural number k such that the
(j,j)th entry of the matrix Ak is positive. The PerronFrobenius theorem
now can be summarized as follows.
0 Every positive irreducible matrix has a positive spectral radius and
a peripheral spectrum that consists of a finite number of complex
numbers that are equally spaced along the spectral circle. Moreover,
the spectral radius is an eigenvalue with algebraic and geometric
multiplicity one.
8.1. The Banach Lattices and

The inner product of two vectors x, y E will be denoted, as usual, by
(x,y), i.e.,
(x,y) =
The unit sphere of C7' under the Euclidean norm will be denoted by S,
i.e., S = { x E = = 1 }. The positive part of S will be
denoted by i.e. {x E 5: x O}. Both S and are compact
subsets of Whenever necessary, 5+ will be considered as a compact
subset of RTh. Recall that a positive vector x > 0 in is said to be strictly
positive, in symbols x >> 0, whenever each component of x is positive.
We shall consider any given n x n matrix (with complex or real entries)
as an operator acting on via the usua' matrix multiplication formula.
Recall that throughout this book the vectors in C7' are usually denoted as
row vectors; for instance, we write x = (Xi, X2,.. . , However, whenever
we encounter the expression Ax, where A = is an n x n matrix, then
8J. The B4anach Lattices and 317
it will be understood that the vector x is flow considered as a column vec-

xl
tor, i.e., x = and the vector Ax is computed via stafidard matrix
,
multiplication. That is,
Ax= =
arnjxj
If n matrix
[ajj] is au withcomplexeuitries, A = [iT]
conjugate of A and At where = is the transpose of A.
Clearly, At = At. We have the fundamental duality ideuitities:
and
for-all x,y e
Au n x n matrix A = is said to be Hermitian if At = A. A Her-

rnitian matrix with real efitries is known as a symmetric matrix. Notice
that for every n x n matrix A the matrix AtA is always a Hermitian matrix.
The proofs of the next few results are left as exercises.
Lemma 8.1. An n x n matrix A is Hermitian if and only if (Ax, y) =

(x,Ay) for all
x n matrix Uis called unitary if U' = _Ut. A unitary matrix with
real entries is referred to as an orthogonal matrix. The unitary matrices
are characterized as follows.
Lemma 8.2. For an n x n matrix A the following are equivalent.
(a) A is unitary.
(b) The rows of A form an orthonormal basis in
(c) The columns of A form an orthonormal basis in

(d) A is inner product preserving, (Ax, Ay) = (x, y) for x, y e
(e) A is Euclidean norm preserving, for x, y e
lAxit = lixil Ctm.
(f) A carries an orthonormal basis to an orthonormal basis.
Recall that the characteristic polynomial of an n x n matrix A is the

polynomial of degree n defluied by
= A).
The spectrum of A couisists precise'y of the roots of the po'ynomial PA, Le.,
PA(\)0}.
It is well known that each A E o(A) is an eigenvalue of the matrix A, that is,
cr(A) = If A1, A2,.. , are the roots of the characteristic polynomial
.
PA (counted with their multiplicities), then it is easy to see that
= tr(A) and ft det(A).
Recall that the sum of the diagonal entries of A is known as the

trace of A.
Each matrix "satisfies" its characteristic polynomial. This important
result is known as the CayleyHamilton theorem.1
Theorem 8.3 (CayleyHamilton). If an n x n matrix A has characteristic
polynomial PA(A) = Atm + + . + aiA + ao, then
+ + + + aiA + aol = 0.
The operator (or the Eudlidean) norm of an n x n matrix A will be

denoted by IA JJ, i.e.,
x ECTh and JJxH i}.
If A1 is the largest eigenvalue of the Hermitian matrix AtA, then JJA(l =

see Exercise 9 at the end of this section.
The vector space of all n x n matrices with real (resp. complex) entries
will be denoted by (resp. Clearly, is a Banach lattice
and is the complexification of (R), i.e., =
The Banach lattice is lattice isomorphic to and the complex
Banach lattice to
Exercises
1. If A is an n x n matrix with real or complex entries, then show that
t
(Ax,y) (x,A y)

for all x, y E In particular, show that A is Hermitian if and only if

(Ax, y) = (x, Ay) for all x, y E
2. Prove Lemma 8.2.
3. As in the case of operators, two m x n matrices A and B are said to be
similar if there exists an invertible matrix C such that A = C'BC.
Recall also that the trace tr(A) of an n x n matrix A = is the sum
of its diagonal entries, i.e., tr(A) = Establish the following.
1
This famous theorem was proved in 1853 for some special matrices by William Rowan Hamil-
ton (18051865). A few years later, Arthur Cayley (18211895) announced (without providing a
proof) the validity of the theorem for all matrices.
8.1. The Banach Lattices and 319
(i) Two similar matrices have the same characteristic polynomialand

hence they have the same eigenvalues counted with their multiplici-
ties.
(ii) This part deals with the trace of matrices.
(a) Show that if A and B are two arbitrary ri x ri matrices, then
tr(AB) = tr(BA).
(b) Show that the trace of an ri x ri matrix A coincides with
the sum of its eigenvalues. That is, if are the
eigenvalues of A (counted with their multiplicities), then
tr(A) =
(c) Show that two similar Ti x ri matrices have the same trace,
i.e., the trace of an ri x ri matrLx is similarity invariant.
(d) Consider an ri x ri matrix A with column vectors c1, .. , .
Also, let denote the unit vector and consider as a

linear functional acting on (or on via the formula
= = Show that A, as an operator on
(or on can be identified with the finite-rank operator
T=
(e) Using the notation of part (b), show that tr(A)
(Compare this with Exercise 4 of Section 4d.)
4. Show that each eigenvalue of a unitary matrix has modulus one. [HINT:
If U is a unitary matrix and a unit vector z E satisfies Uz = Az, then
= z) = (Az, Az) = (Uz, Uz) (UtUz z) = (z, z) = 1.
See also Exercise 5 of Section 6.2.]

5. Show that every Hermitian matrix has real eigenvalues and that eigen-
vectors corresponding to distinct eigenvalues are orthogonal. [HINT: If A
is a Hermitian matrix and a unit vector z E satisfies Az = Az, then
3; )(z, z) = (z, Az) = (z, Az) = (TAtz, z)
= (Az,z) = (Az,z) =
is called diagonalizable if there exists an
A
invertible Ti X Ti matrix R (called a diagonalizing matrix for A) such
that the matrix R'AR is diagonal. Establish the following.
(a) A matrix A is diagonalizable if and only if A has Ti linearly indepen-
dent eigenvectors.
(b) Every Hermitian matrix is diagonalizable by a unitary matrix.
7. An Ti X Ti matrix A is said to be positive semidefinite if (Az, z) 0 for
all z E Ctm. Establish the following.
(a) If A is a positive semidefinite matrix, then A has non-negative eigen-
values.
(b) A Hermitian matrix A is positive semidefinite if and only if it has
non-negative eigenvalues.
[HINT: For (b) use the fact that a Hermitian matrix is diagonalizable by
a unitary matrix.]
8. For each k let Ak = be an n x n matrix. It is easy to see that the

sequence of matrices {Ak} converges to a matrix A = [aij] in the Banach
lattice if and only if for each i and j.
Show that an n x n matrix A satisfies Ac 0 in if and only if
every eigenvalue of A has modu'us less than one. [HINT: See Exercise 11
of Section 6.1.]
9. Consider a matrix A If is the (non-negative) largest eigen-
value of the Hermitian matrix AtA, then show that hAil = In
particular, if A is Hermitian, then show that its Euclidean norm coin
cides with its spectral radius. [HINT: Let x 6 satisfy = 1. From
the identity (x,AtAX) = (Ax,Ax) ILAxII2 0, we infer that the Her-
mitian matrix AtA is also so-it has non-negative
elgenvalues, say 0. If U is a unitary matrix that
diagonalizes 4tA and x = Uy, then 1 and
IhAxll2
= (x,AtAX)
=
This implies hAll =
10. Show that a square matrix is quasinilpotent if and only if it is nilpotent.
(Compare this with Exercise 18 of Section 6.1.)
11. Let A = 6 If we let = + and consider the real
matrices B = and C = then we can write A B + iC and view
A as a vector in the complexification = of the
real Banach lattice
Show that the modulus of A = B + iC in is the
positive matrix Al = [IajjhI.
12. Prove the Cayley_Hamilton Theorem 8.3.
8.2. Operators on Finite Dimensional Spaces

In this section, we shall discuss the structure of operators on finite dimen-
sional spaces and their connections with square matrices. Unless otherwise
stated, X will denote a finite dimensional complex vector space of dimension
n. We start our discussion by mentioning (without proofs) a few standard
facts that will be used throughout this book regarding finite dimensional
vector spaces and operators acting on them.
For an operator T: X X on a finite dimensional vector space the
(a) T is invertible.
(b) T is one-to-one.
(c) T is surjective.
o If {ei,. . ,
. is a basis of X, then (A1, A2, . . . , is a
surjective linear isomorphism between and X. Under this mapping,
2. Operators on Finite Dimensional Spaces 321
we can identify X with CT1. Moreover, we can define the Euclidean

norm on X relative to the basis {ei,.. . , by
= =
There exists a unique Hausdorff linear topology on X, called the Eu-

clidean topology on X. In particular, all norms defined on X are equiv-
alent.
Every linear operator T: X X is continuous.
Let T: X X be an operator, and fix a basis {ei,. in X. If for .
each j we write = and consider the n x n matrix A=
i.e., the column of A is the column vector

:
then for each vector
A1
x= = = A, we have
Tx =
:1=1 j=1 i=1
ei) = E (E
i=1 j=1
where AA is the usual matrix multiplication of the matrix A with the

column vector A. The matrix A is called the representing matrix of T
with respect to the basis {ei,.. . ,
In other words, for each fixed basis, the operator T can be identified
with an n x n matrix. In particular, we can identify (X) with (C).
This means that once a basis of X has been chosen, then every n x n
matrix definesas abovea unique operator on X.
If we consider another basis {fi, f2,.. fn} in X and let B be the repre-
. ,
senting matrix of the operator T: X X relative to this new basis, then

B is similar to A. That is, there exists an invertible matrix C such that
B = C'AC. This implies that A and B have the same characteristic
polynomial. Therefore, all matrices representing T have the same charac-
teristic polynomial. This common characteristic polynomial is called the
characteristic polynomial of the operator T and is denoted by pT(A).
It should be clear that the characteristic polynomial of T has degree
n, the dimension of the vector space X.
The spectrum of an operator T: X X consists of eigenvalues. These

eigenvalues are the roots of the characteristic polynomial of T. That is.
cr(T) = {A E C: PT(A) = O}.
In particular, every operator on a finite dimensional complex vector
space has an eigenvalueand hence an eigenvector.
Recall that a pair (Y, Z) of (closed) subspaces of X is reducing for an

operator T: X X, or that T is reduced by the pair (Y, Z), if Y and
Z are both T-invariant and X = Y Z.
If (Y,Z) is a reducing pair of subspaces for an operator T: X X,
then
PTP")
If T: X X is an operator on a finite dimensional vector space, then
dim(Ker(T)) + dim(R(T)) = dim(X).

Having listed all these basic properties above, we now proceed to describe
the structure of linear operators on finite dimensiona' vector spaces. We
start with the following result.
Theorem If T: X X is an operator on an n-dimensional complex
vector space, then there exist n subspaces Vi, V2,.. . , of X such that:
(a)
(b) Each 14 has dimension i.
(c) Each 14 is T-invariant, c .
Proof. The proof is by induction on the dimension n of the complex vector

space X. For n = 1 the conclusion is obviously true. So, for the induction
step, assume that our claim is true for some n and let T: X X be an
operator on an (n + 1)-dimensional complex vector space. Then the topo-
logical dual X* of X is also an (n + 1)-dimensional complex vector space,
and so its adjoint operator T*: has an eigenvector. That is,
there exists some non-zero vector E X* and some complex number )\
such that T*x* = Ax*. Fix some vector e E X with x*(e) = 1 and put
V = Kerx* = {x e X: x*(x) = O}. We claim that VEBCe = X.
To see this, note first that if x e X, then the vector y = x x*(x)e
belongs to V and so x = y + x*(x)e e V + Ce. So, V + Ce = X. On the
other hand, if x e V fl Ce, then x = ,ae and so = (iie) = 0. This implies
x = 0, and therefore V e Ce X. Clearly, V is an n-dimensional complex
vector space.
Next, we claim that V is T-invariant. To see this, let v E V and note
that from
(Tv, x*) = (v, T*x*) = (v, = x*) = Ax*(v) = 0,
it follows that Tv e V. So, by our induction hypothesis, there exist sub-

spaces V1, 1/2,.. . , V that satisfy properties (a), (b), and (c) for the
operator T: V V. It follows that the subspaces I/i, Va,. =X . . ,
satisfy properties (a), (b), and (c) for the operator T: X X, and the
proof is finished.
The structure of operators on finite dimensional vector spaces is closely

related to the structure of nilpotent operators. Recall that an operator
T: X X on a vector space is said to be nilpotent if = 0 for some
Ic. The natural number q = min{k E N: = 0} is called the index of
nilpotence of T.
The concepts of nilpotence and invertibility of operators on finite dimen-
sional spaces are in a way "opposite" notions. This is immediate from our
next result that is the finite dimensional version of Theorem 2.23. (Keep in
mind that every operator on a finite dimensional vector space automatically
has finite ascent and descent.)
Theorem 8.5. If T: X X is an operator on a finite dimensional vector
space, then there exists a unique T-reducing pair (Y, Z) of vector sub spaces
such that T: Y Y is nilpo tent and T: Z Z is invertible.2
Therefore, in order to describe the structure of operators on finite di-

mensional vector spaces, we begin by studying the structure of nilpotent
operators. We start with the following result.
Lemma 8.6. Let T: X X be a linear operator. If for some q 1 a
vector u E X satisfies 0 and = 0, then the q vectors
u, Tu, T2u, .. .
,
are linearly independent. In particular, if X is finite dimensional and T is

nilpotent with index q, then q dim(X).
Proof. To see that the q vectors u,Tu,T2u,.. . are linearly inde-

pendent, let
a1u + a2Tu + a3T2u + . . + =0. (1)
Now suppose by way of contradiction that aj 0 for some i, and consider
the natural number j = min{1 i q: 0}. Then, applying the
operator to (1), we get a contradiction. Hence,
= 0 for each i, and so the vectors u, Tu, T2u, are linearly
. . . ,
independent.
Now assume that X is finite dimensional and that T is nilpotent with
index of nilpotence q. Hence, there exists at least one vector u E X such
that 0. This implies that q dim(X). D
The next result describes a key "reducibility" property of nilpotent op-

erators.
2This is also expressed by saying that: Every operator on a finite dimensional vector space
is the direct sum of a nilpotent operator and an invertible operator.
Theorem 8.7. Let T: X X be a riilpoterit operator on a finite dimeri-

sional vector space with index of nilpotence q, and let a vector u E X satisfy
0.If U is the T-invariant subspace generated by the linearly
independent vectors u, Tu,.. , then there exists a T-invariant com-
plement W to U, i.e., U 14' = X and T(W) W.
Proof. The proof is by induction on the index of nilpotence q. If T is
nilpotent with index q = 1 (i.e., T = 0), then the conclusion should be obvi-
ous. So, for the induction step, assume that the claim is true for nilpotent
operators of index q 1.
Now let X be a finite dimensional vector space and let T: X X
be a nilpotent operator with index q + 1. Also, let u be a vector of
X satisfying 0, and let U be the (q + 1)-dimensional
subspace generated by the vectors 'u, Tu,.. . We denote by V the
,
T-invariant subspace generated by the linearly independent

vectors Tu, T2u,. ..,
Clearly,
T(U) = V.
The range R(T) of T is a T-invariant subspace and the operator
T: R(T) R(T) is nilpotent with index of nilpotence q. So, app'ying our
induction hypothesis to the subspace V of R(T), we see that there exists a
T-invariant vector subspace Y of R(T) such that
VEBY=R(T). (*)
Now consider the vector subspace Z T' (Y) of X. Since T(Y) Y, it
follows that Y C Z and that Z is T-invariant. We claim that
U+Z=X. (**)
To see this, let x E X. Then, by (*), there exist some u' E U and some
y E Y such that Tx = Tu' + y or T(x u') = y E Y. Therefore, the vector
z = x u' belongs to Z. This implies x = u' + z E U + Z, and the validity
of (**) has been established.
Next, we claim that Y fl (U fl Z) = Y fl U = {0}. To see this, let
x E Y fl U and choose scalars such that x = Then Tx E Y
and Tx E T(U) = V, and so from (*) we infer that Tx E V fl Y = {0}, i.e.,
Tx = = 0. This implies = 0 for all 0 i q 1, and so
Finally, consider the vector subspace Y (U fl Z) of Z and let Y1 be an

arbitrary complement of Y (U fl Z) in Z, that is,
=Z.
To finish the proof, we shall show that if W = Y Y1, then the pair of
subspaces (U, W) reduces the operator T. To establish this, note first that
x e Un (Ye Y1) implies x e (Un Z) fl (Ye Yi) {O}, x= Next, from

0.
UeYeY1 Ye(UnZ)eY1 = Z, UeYeY1 U, and (**), it follows that
U e Y e Y1 = X. Moreover, T(Z) Y implies T(Y e Y1) Y Y e Y1.
This shows that the pair (U, W) reduces the operator T, and the proof is
finished. D
An easy consequence of the preceding result is the following.

Corollary 8.8. Let T: X X be a nilpotent operator on a finite dimen-
sional vector space. Then there exist natural numbers
and vectors uI,u2 , uj- satisfying:-
(1) and
0 0 for each i.
(2) The collection of vectors
,. . TU2, u2,
.. Tub, uk
,
is a basis of X.
In particular, the matrix A representing T with respect to this basis has
the following form.' Every entry of A outside the diagonal above the main
diagonal of A is zero. On the diagonal above the main diagonal, we have
(starting from the top) a string of qi I ones followed by a zero if k 1.
Then there is a string of 1 ones followed by a zero if k 2. This pattern
follows until we exhaust all the qj.
And now we come to the basic structuralproperty-of-operators-on-finite

dimensional complex vector spaces.
Theorem 8.9. Let T: X X an operator on a finite dimensional com-
be
plex vector space. Assume that )'i, .. are the distinct roots of the charac-
,
teristic polynomial of T with multiplicities mi,... , rnj, respectively. Then

there exist T-invariant subspaces X1, ,Xk of X such that:
.
(1) for each i=1,2,...,k.

(2)X=XleX2e...eXk.
(3) The operator T: is nilpo tent for each i = 1, 2, .. . , k.
Proof. According to Theorem 8.5, for each i = 1. ... k there exists a pair of
,
subspaces (Xi, that reduces the operator such that

T is invertible. Clearly, each pair (Xi, Yj) also
- 8. Positive Matrices
reduces T. Indeed, if x E then Tx = (T + E and similarly

E Yj yields Ty E This implies that pT(A) = PTIX. ()t).
We claim that PTIy 0. Otherwise, there would exist some non-
zero vector y E such that (T = 0, contradicting the
ity of T Aj: Now this conclusion, coupled with the identity
= + = implies that (A) = (A
This guarantees that has

Next, we claim that if i j, then n X3 = {O}. To see this, assume
that for some i j we have n Xj {O}. Since fl X3 is T4nvariant,
it follows that there exists some A E C and some non-zero z E fl Xj
such that Tz = Az. This implies A = = contrary to Hence,
= {O} holds true for i j. Finally, a simple dimensionality argument
shows that X1 e X2 e = X, and the proof is finished.
Now let A be an n x matrix with distinct eigenvaJues A1,. .. , counted
n
with their multiplicities m1,. .. , respectively. Then Theorem 8.9 cou-
pled with Corollary 8.8 informs us that A is similar to a matrix B with the
following properties:
(a) The entries on the main diagonal of the matrix B are the eigenval-
ues of A repeated according to their multiplicities. Starting from
the position (1, 1), we have the eigenvalue A1 repeated m14imes
followed by the eigenvalue A2 repeated m2-times, and so on.
(b) All entries of B which are not on the main diagonal or on 'the
diagonal above the main one are zero.
(c) On the diagonal above the main diagonal the entries are only zeros
and ones and have the following pattern: On the part above the
string of the Aj, the entries start with a zero followed
by a string of 1)-ones, and sO on.

The matrix B is known as the Jordan form matrix A.
Exercises
1. Show that for an operator T: X X on a finite dimensional vector space
(a) T is invertible.
(b) T is one-to-one.
(c) T is surjective.
2. A square matrix A is said to be upper triangiilarizable if there exists

an invertible matrix C (called an upper triangularizing matrix for A)
such that C'AC is an upper triangular matrix.
Operators on Finite Dimensional Spaces 327
Show that every matrix A is upper triangularizable and that if C is an

upper triangularizing matrix for A, then the diagonal entries of C'AC
consist of the eigenvalues of A (counted with their multiplicities) . [HINT:
Use Theorem 8.4.]
3. Let A be an n x n matrix having distinct eigenvalues A1, . , Ak with
. .
multiplicities m1 , , mk , respective'y. If q(') is a polynomial, then show

. . .
that the characteristic polynomial of the n x n matrix q(A) is

Pq(A)(') = [A q(A1)]m'[A q(A2)}m2 [A
[
HINT: By Exercise 2 above, we know that every matrix is similar to an
upper triangular matrix, and so we can suppose that A is itself an upper
triangular-matrix. In particular, the diagonal entries of A are the eigen-
values Aj; each appearing timesthe multiplicity of the eigenvalue
It follows that q(A) is also an upper triangular matrix having diagonal
entries q(A1),.. . , with each appearing rnj times.]
4. This exercise describes the minimal polynomial of an operator T: X X
on an n-dimensional complex Banach space.
(a) Show that the spectrum of T is finite and consists of
Moreover, if we let o(T) = {A1,... then show that k <
, n.
(b) Show that there exists a non-zero p(A) such that p(T) =0.
(c) Prove that some root of the polynomial p(A) in (b) must belong to
the spectrum of T; and so the spectrum of T is non-empty.
(d) Assume that 7r(A) is a polynomial of minimal degree and leading
coefficient one satisfying 7r(T) = 0. Show that
7r(A) =
where, as in (a), o-(T) = . . , and each is the ascent of
the operator T. (This establishes that the polynomial 7r(A) is
uniquely determined; it is referred to as the minimal polynomial
of the operator T.)
(e) Show that a q(A) satisfies q(T) = 0 if and on'y if q(A) is
divisible by the minimal po'ynomial 7r(A) of T. In particu'ar, prove
that two polynomials and satisfy ''(T) = if and only
if there exists a polynomial 9(A) such that = + 9(A)7r(A).
[HINT: Let {x1,.. , . be a basis of X. Since for each i the n + I vec-
tors .
, must be linearly dependent, there exists a non-zero
polynomia' such that = 0. If p(A) = p1(A)p2(A) . .
then p(A) is a non-zero polynomial satisfying p(T) 0.1

5. Let T: X f X be an operator on a finite dimensional vector space. If a
pair of subspaces (Y, Z) reduces T, then show that PT(A) =pTIy(A)pTIZ(A).
6. Let T: X f X be an operator on an n-dimensional vector space. If two
n x n matrices A and B represent T with respect to two different bases
of X, then show that A and B are similar.
7. If 5, T: X X are two operators on a finite dimensiona' vector space,
then show that ST and TS have the same characteristic polynomialand
conclude from this that if A and B are two n x n square matrices, then
AB and BA have the same characteristic polynomial. [HINT: Prove first
the result for the special case when S is a projection, Le, 82 = S. For
the general case show that there exists an invertible operator R: X X
such that RS is a projection, and then apply the preceding conclusion to
the pair of operators RS and TR'.
8. Let V be a finite dimensional vector space. If A is a non-empty subset of
V, then its annihilator A-'- is the vector subspace of V* defined by
v*(a)=O forall aEA}.
Similarly, the annihilator of any non-empty subset B of
V defined by B-'- = {v E IT: b*(v) = 0 for all b* E B}.
Assume that T: X Y is an operator between two finite dimensional
vector spaces. As usual, we define its adjoint operator T*: via
the duality identity
(Tx,y*) = (x,T*y*) = y*(Tx)
for all x E X and all E Establish the following:
(a) = Ker
(b) [Ker
(c) dim R(T*) = dim R(T).
9. If T: X Y is an operator between finite dimensional vector spaces,
then show that
dim Ker(T) + dim R(T) = dim X.
10. Use Theorem to present an alternate proof of the CayleyHamilton
Theorem
11. Show that every square matrix is similar to its transpose.
12. Let T: Y Y be a bounded operator on. a Banach space. A vector y E Y
is said to be cyclic (or more precisely T-cyclic) if the vector subspace
generated by {y, Ty, T2y, T3y, .} is dense in the Banach space Y.
.
For an operator T: X X on an n-dimensional vector space estab-

lish the following.
(a) A vector x E X is T-cyclic if and only if {x, Tx,. , 'x} is a
basis of X.
(b) The operator T has a cyclic vector if and only if T has a matrix
representation of the form:
o o 0 0 0
1 0 0 0 0 at
o 1 0 0 0 a2
: (*)
0 0 0 0 0
0 0 0 . 1 0
0 0 0 0 1
(c) If x is a T-cyclic vector and = a3x a1Tx .

then the characteristic polynomial of T is given by
PT(A) + + +. + a1A +
Matrices with Non-negative Entries -
8.3. Matrices with Non-negative Entries

Here shall take a look at square matrices with non-negative entries. We
start by classifying various types of such matrices.
Definition 8.10. A square m x m matrix A = is said to be:
(1) positive (non-negative) i andj (equivalently, if
AxOforeachx>O),
(2) strictly positive if 0 for all i and j and every column has a
positive entry (equivalently, if Ax> 0 for each x > 0), and
(3) strongly positive if > 0 for all i andj (equivalently, if Ax>> 0
for each x> 0).
It should be clear that a strongly positive matrix is strictly positive
and that a strictly positive matrix is positiveS It should also be noticed
that our terminology regarding matrices is in complete agreement with our
terminology regarding operators introduced in Definition 1.13. That is, a
matrix A is positive if and only if A, as an operator on (or Ca), is a
positive operatorin which case we write A 0 or 0 A. The operator
notation B A (or A B) for matrices with real entries is equivalent, in
this case, to < for all i and j. Similarly, a matrix A is a strictly
positive matrix if and only if A, as an operator on (or Ca), is a strictly
positive operator. It is easy to see that every n x n matrix A = has a
modulus which is given by the matrix = [lajjlJ.
Theorem 8.11. The spectral radius of a positive matrix is an eigenvalue
having a positive eigenvector.
Proof. Let A be a positive matrix. If r(A) > 0, then the conclusion follows
immediately from the KreinRutman Theorem 7J0. However, in the finite
dimensional case, the proof of the KreinRutman theorem can be used to
cover the case r(A) = 0 as well. Indeed, in the finite dimensional case one
has to notice that the sequence of positive unit vectors always has a
convergent subsequence to some positive vector. This positive vector is an
eigenvector corresponding to the spectral radius of the matrix.
In spite of the fact that the proof of Theorem 8.11 mimics that of the
KreinRutman Theorem 7.10, there is a difference between these two results
that is worth pointing out. In the KreinRutman theorem, we assume that
the spectral radius of the compact operator is strictly positive, while in
Theorem 8.11 we do not exclude that the spectral radius can be zero. In
Theorem 8.11 the conclusion about the existence of a positive eigenvector
remains true even if the spectral radius is zeroS We refer to Exercise 10 at
the end of this section for an independent proof of this fact.
Definition 8.12. With each non-negative ri x ri matrix A, we associate the

following two sets
LA = {A 0: x> 0 with Ax Ax}

= {A0:
with )tx>>Ax}
= {\0: with Ax>>Ax}.
Notice that if 0 A1. A and E LA, then A1 E LA and, similarly,
A1 and imply E UA. Notice also that if x>-O---
satisfy Ax Ax, then
= IIAxM hAil Mxli
and so )t hAil. That is, LA c [0, IAII]. Since 0 E LA, we see that LA is
a non-empty interval in R that is bounded from above. The set LA is also
closed. To see this, assume that a sequence c LA satisfies A. For
each ri pick some E with Since is a compact set, by
passing to a subsequence (if necessary), we can assume that x E S+.
Hence, Ax = A E LA and
so LA is closed. Consequently, LA = [0, AA], where
= sup LA = maXLA.
The set UA will be described in Corollary 8.23. The properties of
non-negative number AA will be discussed in detail below. The first property
follows immediately from the definition of AA.
Lemma 8.13. If two ri x ri positive matrices A and B satisfy A B, then
AAAB.
More properties of AA are included in the next result.
Lemma 8.14. For an n x ri positive matrix A we have the following:
(1) AA = r(A), the spectral radius of A.
(2)
A is strictly positive, then AA > 0.
(4) AA is the only possible eigenvalne of A having a strictly positive
eigenvector.
Proof. (1) By Theorem 8.11 there exists some x> 0 with Ax = r(A)x and
so r(A) <AA. On the other hand, if x E and A 0 satisfy Ax Ax,
then = llAkxll for each k. Consequently,

8.3. Matrices with_Non-negative_Entries 331
for each k and so = r(A). This implies r(A), and

therefore )'A = r(A).
(2) This follows from (1) and the fact that A and At have the
spectrum.
(3) Choose some x > 0 with Ax Since A is strictly positive, we
have Ax > 0 and this implies AA> 0.
(4) Assume that Ax = holds for some E C and some x 0. From
(1) and (2) we know that there is some y> 0 such that Aty r(A)y. Since
x 0, we see that (x, y) = xiyi > 0. Now note that
[a r(A)] (x, y) = y) r(A)(x, y) (ccx, y) * (x, r(A)y)
(Ax,y) (x,Aty) = (Ax,y) (Ax,y) = 0,
from which it follows that = r(A).

The following result (whose proof is left as an exercise) characterizes the
matrices in (Ia) that are lattice homomorphisms or interval preserving
operators on
Theorem 8.15. Let A be an n x n non-negative matrix. Then:
(1) A defines a lattice homomorphism on if and only if each row
has at most one non-zero entry.
(2) A is interval preserving if and only if every column of A has at

most one non-zero entry.
Exercises
1. Let A = an n x ri matrix with real entries. Show that A (as a

be
regular operator on has a modulus and that Al = [laiji].
2. Show (by means of an example) that although the spectral radius of a
(non-zero) non-negative matrix is an eigenvalue, it need not have a strictly
positive eigenvector.
3. Assume that a non-zero positive matrix A commutes with a strongly
positive matrix. Show that if AA > 0 and Ax for some x> 0, then
Ax = AAx. [HINT: Argue as in the proof of Theorem 8.22.]
4. (Frobenius [130]) Let A = be an n x n positive matrix. For each i
let = (the sum of all elements in row i) and put p = mini
and R = Show that p T(A) <R. [HINT: Pick some x> 0
such that Ax = T(A)x. From T(A)x = Ax Rx, we get T(A) R. On
the other hand, if e = (1, 1,.. ., 1), then pe Ac.]
5. (Gershgorin [134]) This exercise presents a simple but very useful "ap-
proximation" of the spectrum of a square matrix. Let A = E (C)
and for each i let = and = {z E C: z <
Show that u(A) C2. [HINT: Let A be an eigenvalue of the ma-
trix A and pick x = 0 satisfying Ax = Ax. Choose
some k such that jxk = > 0. Now use the equation
(A akk)xk = to get IA
a E >
satisfies for each i, then
show that A is invertible. [HINT: Use the preceding exercise.
7. A non-negative n x n matrix A = is said to be a Markov matrix
whenever the sum of the elements in each row equals one, whenever
= 1 for each i. (This is equivalent, of course, to saying that A
defines a Markov operator on Wi.) Show in four different ways that every
Markov matrix A satisfies r(A) = 1 by using:
(a) Exercise 4 above,
(b) Lemma 8.14(4),
(c) Gershgorin's theorem (see Exercise 5 above), and
(d) the spectral radius formulaS
[HINT: If x = (1, 1,.. . , 1), then Ax = x, so that us an eigenvalue of
A. By Lemma 8.14(4), we see that r(A) = 1. Gershgorin's theorem also
informs us that if A is an eigenvalue of A, then for some i we have
Al IA = 1 a22.
Consequently, At (1 + = 1, which shows that the spectrum of

A lies entirely inside the closed unit disk.] -

9. If an n x n positive matrix A = satisfies u(A) = {1}, then show that
1 for all i. Use this and Theorem 8.15 to show that if a positive
matrix A with u(A) = {1} defines either a lattice homomorphism or an
interval preserving operator, then A = I.
10. Let A be an n x n positive matrix. If A is nilpotent (i.e., r(A) = 0), then
show that there exists some vector x> 0 such that Ax = 0.
8.4. Irreducible Matrices

It is easy to see that a vector subspace J of (or W') is an ideal if and
only if it is a bandand also if and only if it is of the form
for some subset I of {1,2,. . . ,n}. We are now ready to define the notion
of ideal irreducibility for matrices. In Chapter 9, this definition will be
extended to arbitrary operators; see Definition 9.2.
8.4. Irreducible Matrices 333
Definition 8.16. An n x ri matrix A is said to be reducible whenever A

leaves invariant a non-trivial ideal in That is, A is reducible if theTe
exists a vector subspace J of of the form
J = {(zi,.. . = 0 for all i E
wheTe I is a non-empty proper subset of {1,2, ,m}, such that A(J) C J.
. . .
A square matrix that is not reducible is called an irreducible matrix.3
Here is a simple but very useful property of irreducible matrices.

Lemma 8.17. Irreducible matrices have no zero rows or columns. In par-
ticular, every irreducible matrix carries strictly positive vectors to strictly
positive vectors.
Proof. Let A be an irreducible matrix. If A has a zero row, say the then
for each x E we have = 0, i.e., the component of Ax is zero.
This implies that the proper ideal J = {x E = 0} is A-invariant, a
contradictionS
On the other hand, if A has a zero column, say the and we consider
the proper ideal J {x = 0 for all i j}, then A(J) = {0} C J,
which is also impossible.
If A = [aij] is an mx n matrix, then for each k we shall write Ak

i.e., denotes the (i, j)th element of the matrix Ak. Also, let us recall that
an m x n matrLx is said to be a permutation matrix if its rows consist of
a permutation of the basic unit row vectors. That is, a permutation matrix
is a matrix which can be obtained by permuting the rows (or, equivalently,
the columns) of the identity matrix. Three important properties that are
satisfiecLhy_any m x m permutation matrix P are the following.
(1) ptp = ppt = I, that is, = pt
(2) If A is an n x rn matrix, then PA is obtained from A by permuting
the rows of A in the same manner we permuted the rows of the
identity matrix to get P. For instance, if P is the 3 x 3 permutation
matrix with rows e2, e1 and A has rows r1, r2, r3, then PA is the
matrix with rows r2, r3, ri.
(3) If A is an m x n matrix, then AP is obtained from A by permuting
the columns of A in the same manner we permuted the columns of
the identity matrix to get P.
We are now ready to characterize the irreducible positive matrices.
3Many authors refer to an irreducible matrix as an indecomposable matrix.

Theorem 8.18. For an n x n positive matrix A the following statements

are equivalent.
(1) The matrix A is irreducible.
(2) The matrix I + A is irreducible.
(3) The matrix (I + is strongly positive.
(4) For each position (i,j) there exists some integer k such that the
(k) (k)
(z, j) entry of A is positive, z. e., > 0.
(5) There is no permutation matrix P such that
C]
where B and D are square matrices.

Proof. (1) This follows from the fact that for a vector subspace
(2)
V of we have A(V) V if and only if (I + A)(V) V.
(1) (3) We claim that:
a If A is irreducible and a positive vector x has I k n I
positive components, then the positive vector (I + A)x has at least
k + 1 positive components.
If the claim is true, then from the identity
(I + A) (I + A) x
(ni) factors
and an easy inductive argument, we easily see that (I + 0 for

each x> 0, and this guarantees that (I + A)n_l is strongly positive.
To establish the above claim, fix an arbitrary x > 0 and assume that
the set I = {t E {1,. n}: xt > 0} has k elements, where I k n I.
. . ,
From (I + A)x = x + Ax, we see that [(I + A)x]t xt > 0 for each t E I.
To see that [(I + A)x]t = xt + (Ax)t > 0 for some t I, we must show that
(Ax)t > 0 for some t I. To this end, assume by way of contradiction that
(Ax)t = 0 for all t I and consider the non-trivial ideal
J={zECn: zj=0 for all
If z J, clearly Izi < Ax holds for some A > 0, and therefore from
then
lAzi AIzI AAx it follows that = 0 for all t I, i.e., Az E J. In
other words, A(J) J, contrary to the irreducibility of A, and the validity
of our claim follows.
(3) (4) Assume (I + is strongly positive. First, we observe
that A cannot have a zero row. For if a row of A is zero, say the ith row,
then the row of I + A is equal to the basic unit vector But then I + A
8.4.Irreducible Matrices 335
carries each vector whose component equals zero to a vector whose

component is also zero. Consequently, (I + carries vectors with
component equal to zero to vectors with component also zero. However,
the latter conclusion contradicts the strong positivity of (I +
A is zero, it follows that
A(I + = (n1)Ak]
=
we thatfor each i) there some integer k (in 1 k n).
such that > 0.
(4) (5) Assume by way of contradiction that there exists some
permutation matrix P such that PtAP where B and D are
square matrices. Then for each k we have
where Bk and Dk are also square matrices having the same sizes as B and
D, respectively. From the above representation of we infer that there is
at least one position (i, j) such that the (i, entry of Ak is zero for each
k. However, the latter contradicts our hypothesis and the validity of (5) is
established.
(5) Assume by way of contradiction that A leaves invariant a
(1)
non-trivial ideal J = {z E = 0 for all i I}. Let I {i1,. . ,
1 <ii <n and 1 <k n 1. The inclusion A(J) C J

means that if some z E satisfies = 0 for each j = 1,. k, then ,
= 0 also holds for each j = 1, .. , k.

Let Q be a permutation matrix whose last k rows are the (row) unit
vectors e11, . . . , Then for each z E J the has its last
column vector Qz
k components all equal to zero, i.e., 0 for each i = m k + 1, .. , n. .
In particular, for each z E J, the column vector QAz also has its last k
components all equal to zero.
Now let us look at the permutation matrix P = the transpose
of Clearly, the last k co'umns of are the column unit vectors
. , In particular, the components of the first m k column vectors
of are all zero at the coordinates So, if the columns of AQt
,
are then the column vectors ci,.. .,Ck all have zero
components at the coordinates ii, . . . , 1, .. , m k i .
the column vector must have (by the choice of Q) its last k components
equal to zero. In other words, the matrix ptAp = QAQt is of the form
where 0 is the k x (n k)-zero matrix, B is an (n k) x (n Ic) square

matrix and D is a k x Ic square matrix. To complete the proof, notice that
the last conclusion contradicts our assumption.
Corollary 8.19. Every strongly positive matrix is irreducible.
An irreducible positive matrix need not be strongly positive. For in-

stance, it is easy to verify that the non-negative matrix A is irre
=
ducible but fails to be strongly positive. The same matrix also demonstrates
two other properties:
e A power of an irreducible matrix need not be an irreducible matrix.
o For a given irreducible positive matrix A there may be no power

Atm which is strongly positive.
Corollary 8.20. Every irreducible positive matrix commutes with some

strongly positive matrix.
Proof. If A is a irreducible positive matrix, then A commutes with the

strongly positive matrix (I +
A positive;
is positive, i.e., r(A) = )'A > 0.
Proof. Let A be a positive irreducible matrix and let x > 0. If Ax = 0,

then = 0 for each k contradicting statement (4) of Theorem 8.18.
More interesting spectral properties of positive irreducible matrices are

included in the next result.
8.4. Irreducible Matrices 337
Theorem 8.22. Assume that A is a positive irreducible matrix. If some

vector x > 0 satisfies Ax then:
(1) AxAAx.
(2) The positive vector x is strictly positive, i.e., x>> 0.
(3) The eigenspace NAA = {z E Az = AAZ} is one dimensional
and is generated by the strictly positive vector x.
Proof. (1) and (2) By Corollary 8.20, there exists a strongly positive matrix
B such that AB = BA. Clearly, this matrix AB is also strongly positive.
Indeed, if z > 0, then Bz >> 0, and (since A has no zero rows) ABz>> 0.
Now let Ax AAx for some x > 0 and assume by way of contradiction
that Ax > )t44x. By Corollary 8.21 we have AA > 0. Now consider the
vector y = AAx) and note that y> 0 and Ax = AA(x + y). Hence,
ABy>> 0 and B(x + y) >> 0. It follows that

AB(x + y) = ABx + ABy >> ABx = BAx AAB(X + y).
This implies AB(x + y) (AA + E)B(x + y) for some > 0, proving that
AA + E JCA contrary to the definition of )tA. Thus, Ax = )t44x.
To see that x = (xi,.. , 0, notice that Akx = holds for each
Ic. Next, fix some j with x3 > 0. Then, if i is given, pick some k with
> 0. It follows that = > 0, and so > 0 for each

i, i.e., x>> 0.
(3) Assume first that some non-zero vector z e satisfies Az AAZ.
Then Iz[ > 0 and = fAz( <AfzI. By (1) and (2), we have
>> 0, and so each component of z is non-zero. Now consider the vector
y = xlz zix. Clearly, Ay = AAY and since the first component of y is zero,
it follows from the above discussion that y = 0. This implies z =
Corollary 8.23. Every positive irreducible matrix A satisfies the following
properties.
(1) The set UA 0: x>> 0 with >> Ax} coincides with
the open interval (AA,00), i.e., UA =
(2) If an eigenvalue A0 of A has an eigenvector y> 0, then A0 =

and y >> 0.
(3) For a matrix B satisfying 0 B A we have AB = AA if and only
if B = A (and so if B is a non-negative matrix satisfying B < A,
then AB <AA).
Proof'. (1) By Theorem 8.22, there exists x 0 such that Ax = AAX. This
implies >> Ax for each and so (AA, oc) c UA.
Conversely, assume by way of contradiction that some E UA satisfies

Pick some y >> 0 with >> Ay. Also, by Theorem 8.11 and
Theorem 8.22, there exists some z 0 such that Atz = Now note
that
(Ay, Atz) (Ay, = \A(Ay, z) = \A(y, Atz)
=
> (Ay,Atz),
which is impossible. Hence, and so UA = ()'A, oc).
(2) Assume that Ay = for some y > 0. By Corollary 8.20 there
exists a strongly positive matrix B such that AB = BA. Consequently, we
have ABy = BAy = and the vector z = By is strictly positive. This,
coupled with the equality Az = ,\oz, implies that > 0. On the other hand,
it follows that Az << +c)z for each c > 0. Hence, + e UA and so (by
the previous part) + c for all c > 0. This implies .AA Since
the inequality r(A) = is always true, we conclude that =
Now invoke Theorem 8.22(2) to get y>> 0.
(3) Assume 0 0 such that = )\BX = Bx Ax. But then Theorem 8.22(2)
implies x>> 0 and = Ax. Hence (AB)x = 0, and using that AB 0
and x>> 0, we easily get A = B.
Exercises
1. For a vector subspace J of show that the following statements are
equivalent.
(a) J is an ideal, i.e., J = Jo zJo, where Jo is an ideal in
(b) J is a band, i.e., J = B iB, where B is a band in RTh.
(c) There exists some non-empty subset I of {1, 2, . . . , n} such that
2. Show that an n x n positive matrix A is irreducible if and only if At

is irreducible. [HINT: Use the identity (I + = [(I + A)Th_h]t in
conjunction with Theorem 8.18.1
3. For positive m x n matrices A and B establish the following.
(a) If Ac is irreducible for some k 1, then A is irreducible.
(b) If A is irreducible, then A + B is irreducible.
(c) If A and B are irreducible, then AB and BA need not be irreducible.
&5. The PerronFrobenius Theorem 339
4. Show that a non-negative matrix A is irreducible if and only if the matrix

B= I + A defines a Krein operator, that is, for each x > 0 we have
Bkx>> 0 for some k.
5. Use the preceding exercise and Theorem 9.42 to present an alternate proof
of the fact that the spectral radius of a positive irreducible matrix has a
unique (up to scalar multiples) strictly positive eigenvector.
6. Let A = a positive irreducible matrix and let
be (t)] be
a positive matrix-valued function defined for each t in some subinterval
(a, b) of R such that (t) = a23 for each (i, j) and some to (a, b).
Show that limt,t0r(At) = r(A). [HINT: Fix a sequence (a,b)
satisfying + to. By Theorem 8.11 for each n there exists a positive
unit vector with By passing to a subsequence, we

can assume that x > 0. It follows that
lim
nboo
= fl+OO
lim = Ax > 0,
where the last conclusion holds true by virtue of the strict positivity of A
(Corollary 8.21). This yields = and so Ax ,\0x. By
Corollary 8.23, = = r(A) and this implies limt,t0 = r(A).
8.5. The PerronFrobenius Theorem

We start by characterizing the eigenvalues in the peripheral spectrum of a
positive irreducible matrix.
Lemma 8.24. Let A be a positive irreducible m x ri matrix with spectral
radius r = r(A). For an angle the following statements are equivalent.
(1) The complex number is an eigenvalue of A.
(2) There exists a diagonal matrix D such that ID = I and
A=
Proof. (1) Assume that

(2) is an eigenvalue of A. Therefore, there
exists some non-zero vector x = (xi,... E such that Ax =
This implies rIxI = = fAxj and Theorem 8.22 guarantees that
IzI >> 0 and
AlxI=rIxI.
In particular, 0 for each j and so the diagonal matrix
is well defined and satisfies ID! = I. Obviously, DIxI x. Next, consider

the matrix
B= [bk3] =
and notice that

BlxI = =
= = rD'x = rixi
=
At the same time, from
BIIxf = <
=
we see that Bixi = IBIIxI, that is
(*)
for each k = 1,. . , n. This implies that for each k there exists some angle
0k such that = for j = 1,.. . , m; see Exercise 1 at the end of
this section. But then (*) yields (EL1 Ixj I) = I
(xj and I
so for each fixed k either = 0 for each j or else etk = 1. In either case,
bkj Ibkjf Oforallkandj.
From the definition of B it follows that
B = IBI = A D1 = A,
and therefore akj 0 for all Ic and j. Now taking into account that
AlxI = BJxJ, we see that bkj)fxjI = 0 for all k. Since 0
for each j, the latter implies ak3 = for all k and j, i.e., A = B, and
our conclusion follows. (Another way of proving that A = B is by using
Corollary & 23(3) in conjunction with the inequality 0 <B A.)
(2) Assume that A =

(1) = This means
that A is similar to and so the matrices and A have the same
spectrum. This implies e cr(A), we
conclude that E cr(A).
Let us say that a subset of complex numbers is rotation invariant

under an angle 0 if =
The spectrum of an irreducible positive matrix exhibits the following
rotational invariance.
Corollary 8.25. Let A be a positive irreducible matrix with spectral radius

r = r(A). The spectrum of A is rotation invariant under am angle 0 if and
only if is an eigenvalue of A.
Proof. If reZ0 is an eigenvalue of A, then, by Lemma 8.24, A is similar to

This implies = = cr(A).
8.5. The PerronFrobenius Theorem 341
Conversely, if cr(A) is rotation invariant under an angle 0, then (given

that r is an eigenvalue of A) it follows that E
We are now ready to state the classical PerronFrobenius theorem re-

garding the spectral properties of an irreducible positive matrix. The theo-
rem was proven first for strongly positive matrices by 0. Perron [265] and
was generalized later to positive irreducible matrices by G. Frobenius [131].
Theorem 8.26 (PerronFrobenius). If A is a positive irreducible matrix,
then:
(1) Its spectral radius r(A) is an eigenvalue and satisfies
r(A) =
x 0 with \x>> Ax} > 0.
(2) The eigenspace of r(A) is one dimensional and is generated by a
strictly positive eigenvector.
(3) The eigenvalue r(A) is a simple root of the characteristic polyno-
mial of A.
(4) The peripheral spectrum of A is of the form
2ith
k=0,1,...,p1}
and consists of simple roots of the characteristic polynomial. More-

over, if the characteristic polynomial of A is
PA(\) = + +
where n> n1 > >0 and 0 for each j, then
p= gcd{nnl,n--n2,...,nnk}
= gcd{n ni, ni
. nk},
.
where gcd is an abbreviation for the "greatest common divisor."

(5) The spectrum of A is rotation invariant under the angle
(6) If,a e is non-zero, then the p distinct complex numbers
4-r1 2(pl)ir
/i,,LteP z
in are roots of the polynomial of A having the

same multiplicity.
Proof. (1) This is part (1) of Lemma 8.14 and part (1) of Corollary 8.23.
(2) This follows immediately from Theorem 8.22(3).
(3) The polynomial PA = det(AI A) is the characteristic polynomial
of A. Let us denote by Ak the (ii 1) x (ii 1) matrix which is obtained
from A by deleting its Icth row and kth column. Also, let Pk ( ) denote the
characteristic polynomial of Ak, i.e., det(AI Ak). An easy corn-
putation (using the familiar formula for differentiating a determinant whose
entries are functions) shows that
Nowlet Bk denote the n x n matrix obtained from A by making all

entries in the /cth row and kth column of A zero. Then 0 Bk < A and
Bk A, and from Corollary 8.23(3), we know that )'Bk <AA = r(A). Since
the characteristic polynomial PBk of Bk satisfies PBk = it easily
follows that )"A is greater than any real root of any one of the polynomials
Since each Pk has leading term it follows that > 0 for each
k. A glance at (**) shows that > 0 and this implies that )"A is a
simple root of the characteristic polynomial PA of A.
and (5) Let c7per(A) =
(4) k = 0,... ,p1} be the periph-
eral spectrum of A. By Corollary 8.25, the set oper(A) is rotation invariant
under This guarantees that the subset of complex numbers
k=0,...,p1}
is a group, and consequently G = {z E C: = 1}, i.e., G has the desired
form; see Exercise I in Section 7.3. Now invoke Corollary 8.25 once again to
obtain that the whole spectrum o-(A) is rotation invariant under the angle
The simplicity of the roots can be established either directly or by
employing the conclusion of part (6) below.
To see that p can be computed by the described formula, note first
27r
that it follows from Lemma 8.24 that the two matrices A and e are
similar, and so they have the same characteristic polynomial. Computing
the characteristic polynomials yields
PAP') = + + ... + aflk.A
27r(nn1) 21r(nnk)
= + p
+ aflke
2ir(n_n,)
This implies = or e =I for each j. Therefore, p
divides n nj for each j.
To establish that p is the gcd of the numbers n we distinguish two
cases: p = I and p> I. Assume first that p = I, i.e., cTper(A) = {r(A)}. In
this case, if m > I is a divisor of each n then (as above) the matrix
has the same characteristic polynomial as A. This guarantees that
is a point of the peripheral spectrum of A which is different from
a contradiction. Now consider p > I. If m > I is a divisor of each
8. 5. The PerronFrobenius Theorem 343
Th then (as in the previous case) the spectrum of A is rotation invariant

,
under the angle By Corollary 8.25, is an eigenvalue lying in

the peripheral spectrum of A. In particular, or p m, and the
formula for p has been established.
(6) Assume that are the distinct eigenvalues of A with multi-
. ,
plicities m1, . respectively. It follows from Exercise 3 of Section 8.2

,
2ir 2ir . . .
that the matrix has the distinct eigenvalues with ,
multiplicities respectively. But since the matrices A and e

. . ,
are similar, are also the distinct eigenvalues of A with

. . . ,
multiplicities m1,... respectively. In other words, a multiplication of

,
an eigenvalue by does not change its multiplicity, and our final conclu-
sion follows.
Corollary 8.27. Let A = [aj.:j] be an n x n irreducible positive matrix. If
0 for some i, then crper(A) = {r(A)}.
Proof. In this case tr(A) = > 0 and the characteristic polynomial

of A is given by pA(\) + + det(A) Therefore, by part
(4) of Theorem 8.26, the peripheral spectrum is the singleton {rA}.
A special class of irreducible positive matrices is provided by the class
of primitive matrices.
Definition 8.28. An irreducible n x n positive matrix A is said to be prim-
itive if its peripheral spectram is the singleton {r(A)}.
The primitive matrices are precisely the irreducible positive matrices
which have a strongly positive power. The details are included in the next
result.
Theorem 8.29. For an n x m irreducible positive matrix A with spectral
radius r = r(A) the following statements are equivalent.
(1) The matrix A is primitive, i.e., crper(A) = {r}.
(2) For some m the matrix Atm is strongly positive.
(3) There is a strongly positive matrix S such that S.
Proof. (3) Since S is strongly positive and

(2) S, it follows
that there exists some m such that each entry of is greater than zero
for all m in. Hence, Atm is strongly positive.
(2) Assume that Atm is strongly positive for some m. By Cor-
(1)
ollary 8.19, Atm is irreducible and since it is also strongly positive, it follows
from Corollary 8.27 that rtm is the only point in the peripheral spectrum
of Atm and, by part (3) of Theorem 8.26, rm is a simple root of the

acteristic polynomial of Atm. This implies that 5per(A) = {r}. Indeed, if
the peripheral spectrum of A were not a singleton, then (by Exercise 3 of
Section 8.2) either the peripheral spectrum of Atm would fail to be a single-
ton or else rm would be an eigenvalue of Atm of multiplicity greater than
one, which contradicts in both cases our hypothesis. Hence, A is a primitive
matrix.
(1) (3) Replacing A by we can assume that r = 1. The Jordan
form of A guarantees the existence of an invertible matrix B such that
(*)
g],
where C is an (m x (m 1) upper triangular matrix having along its main
1)
diagonal only eigenvalues of A with moduli strictly less than one. The latter
implies 0 (see Exercise 8 of Section 8.1). Since =B B'
for each m, we see that
see that S is strictly positive, let b = (b,, Li2,. ,

To denote the first
. .
row of B' and = (ci, c2,. .

ct be the transpose of the first column c
. ,
of B. Clearly, both Li and c are non-zero vectors. From (*), we have
At(B_i)t = (B_i)t [1 U]
AB= and
and from this it easily follows that Ac = c and Atbt = Lit. Since A and At are
irreducible and both have spectral radius 1, we infer that all components of
the vectors Li and c are non-zero; see Theorem 8.22(2). Now notice that the
(i, j)th entry of S satisfies = 0 and this guarantees that indeed
> 0 for all i and j.
Corollary 8.30. If A is a primitive matrix with r = r(A), then for each
x > 0 the limit y = exists, is strictly positive, and satisfies
Ay=ry.
Proof. Fix x > 0. By Theorem 8.29, we have Sx = y>> 0.
Moreover, note that Ay = lim = r lirn = ry.
In view of the preceding results one might think that it would be easy
to figure out which finite subsets of complex numbers are the spectra of
irreducible positive matrices. Surprisingly, this is a very difficult problem
and a complete answer is not yet known. Even when we place a bound on
&5. The PerronFrobenius Theorem 345
the number of points in a subset o of C, the irreducible positive matrices

that have the set as their spectrum can be of very large sizes. For more
detailed information see [72] and [184].
We close the section by mentioning that there is extensive literature
regarding extensions of the PerronFrobenius theory to matrices that leave
a cone invariant. For details we refer the reader to the survey article by
Tam [316] and the references therein.
Exercises
1. Show that m complex numbers Zi,... satisfy zj
=
I
if and only if there exists some angle 9 such that z,, = for each
i=1,...,m.
2. Consider the polynomials:
(i) 1, (ii) ,\3 and (iii) X' + 1.
Which of the above polynomials are the characteristic polynomials of

positive irreducible matrices? [HINT: Consider the matrices
010 0 0
0 0 1 and 0
100 0 2
0
3. Consider the following Markov matrix

nn
""2 1
2 1
A
0100.
1
0010
0 0 0
Show that A is an irreducible positive matrix satisfying {1}.

[HINT: Note that PA = and that A' is strongly positive.]
4. This exercise illustrates the PerronFrobenius theorem. Consider the pos-
itive matrix
0 0 0 0 1 0 0 0
00000010
0 0
00000001
0 0 0 1 0 0
11100000
11010000
10110000
01110000
(a) Show that A is irreducible.

(b) Show that the characteristic polynomial of A is
PA(A) = (A2 1)2(A + 1)2(A2 1).
Conclude from this that on the unit circle there are six eigenvalues:
1 with multiplicity two, 1 with multiplicity two, and z.
(c) Show that the spectrum of A is rotation invariant under 180.
5 Let A be a primitive matrix and let 8, where r = r(A). Show
that each column of S is an eigenvector corresponding to the eigenvalue r.

6. Show that an n x n positive matrix is primitive if and only if it defines a
Krein operator.
7. We have two glasses; call them 1 and 2. Glass 1 contains 100 grams
of water and glass 2 contains 100 grams of wine. We transfer 10 grams
from glass 1 into glass 2 and stir the mixture very well. Subsequently, we
transfer 10 grams from glass 2 to glass 1 and again stir the mixtures in
both glasses very well. We repeat this process ad infinitum.
Assume that and denote the amounts (in grams) of wine in
glasses 1 and 2, respectively, after the stage of this process. Show
that ntoo
urn lim
n,00
= 50.
IQ
[HINT: Note that = . This, in conjunction with
[w;] [1].
Corollary 8.30, imPlies Now use that
100.]
Chapter 9
Irreducible Operators
This chapter extends the discussion about irreducible matrices to infinite

dimensional spaces. Among other things, we shall present conditions which
guarantee that the spectral radius of a positive operator is non-zero. The
conditions are known as ideal and band irreducibility. An operator T on a
Banach lattice is said to be ideal band) irreducible if there is no
non-trivial closed ideal (resp. band) which is T-invariant. The following list
contains samples of several results that will be obtained in this chapter.
o Every ideal irreducible compact positive operator has a positive
spectral radius (B. de Pagter).
o If two positive operators commute with each other and one is ideal
irreducible and the other is compact, then both have positive spec-
tral radii (AbramovichAliprantisBurkinshaw).
o A band irreducible positive kernel operator has a positive spectral
radius (AndKrieger).
o If a positive operator on a (where is a compact
Hausdorif space) has a positive eigenvalue A having an order unit
as an eigenvector, then its adjoint also has A as an eigenvalue
(M. C, Krein).
In the process of establishing these results, we will also introduce several
new classes of positive operators and study their spectral properties.
9.1. Irreducible and Expanding Operators

Before presenting some general results regarding the spectral radius of posi-
tive operators, we shall pause for a while to discuss various classes of positive
347
9. Irreducible Operators
operators and their relationships. We shall group these operators into three
categories according to their special properties. The first two categories of
operators will be defined in terms of certain invariant subspacesthe ideals
and bands.
Let T: X X be a bounded operator on a Banach space. Recall that
a vector subspace V of X is invariant under T (or simply T-invariant)
if T(V) V. A vector subspace V of X is said to be hyperinvariant
for T (or simply T-hyperinvariant) whenever V is invariant under every
bounded operator on X that commutes with T, i.e., whenever S E (X)
and ST = TS imply 5(V) V. If X is a Banach lattice and T is a positive
operator, then a vector subspace V of X is called lattice hyperinvariant
for T (or simply whenever V is invariant under every
positive operator on X that commutes with T. As usual, a vector subspace
V of a vector space X is non-trivial if V {O} and V X.
The null ideal NT of a positive operator T: B > F between two Riesz
spaces is the ideal in B defined by
NT={XEB:
T= 0 if and only if NT = E. It should also be obvious that
if B and F are Banach lattices, then the null ideal NT is closed in B.
Moreover, observe that a positive operator T: B F is strictly positive (see
Definition 1.13) if and only if NT = {0}. If T: B F is a positive operator
between two Riesz spaces, then the positive operator T: B/NT F, defined
by T[xJ = Tx, is strictly positive. -
Lemma 9.1. If T: B B is an arbitrary positive operator on a Banach

lattice, then its null ideal and the closed ideal generated by its range are both
thyperinvariant.
Proof. Let T: E B be a positive operator on a Banach lattice and let

5: E B be a positive operator commuting with T. If x E NT, then
from 0 T(Sfxl) = S(TIxI) = 0, it follows that = 0, and
so Sx E NT. That is, S(NT) NT and this proves that NT is a closed
-hyperinvariant ideal for T.
Next, notice that the closed ideal JT generated by the range of T is
described by
JT={XEB:
Again, assume that S E (B) is a positive operator that commutes with T.
If Ty for some y 0, then ISxf S(Ty) = T(Sy), which shows
that Sx E JT. Thus, JT is S-invariant, and so JT is a closed -hyperinvariant
ideal for T.
We are now ready to define the classes of ideal and band irreducible
operators. They extend the corresponding notions for matrices; see Defini-
tion 8.16.
Definition 9.2. A bounded operator T: E E on a Banach lattice is:
(1) ideal irreducible if T has no invariant non-trivial closed ideals,
that is, whenever a closed ideal J satisfies T(J) J, then either
J_{O} orJ=E, and
(2) band irreducible if T has no invariant non-trivial bands, i.e.,
whenever a band B satisfies T(B) B, then either B = {O} or
B=E.
The next result indicates that "ideal irreducibility" is a relatively strong
property. In particular, as we shall see, only Banach lattices with quasi-
interior points admit ideal irreducible positive operators.
Theorem 9.3. If T: E E is an ideal irreducible positive operator on a
Banach lattice, then:
(1) T is band irreducible.
(2) T is strictly positive, i.e., x > 0 implies Tx > 0.
(3) P2 has quasi-interior points.
(4) T carries points to quasi-interior points.
Proof. (1) The conclusion follows from the fact that every band is a closed
ideal.
(2) Let T: P2 E be an ideal irreducible positive operator on a Banach
lattice. The null ideal NT of T is closed and T-invariant. Since T is ideal
irreducible, it follows that NT = {0}, which means that T is strictly positive.
(3) By Theorem 4.21 there exists some u > 0 such that
This implies T is ideal irreducible
and {0}, we see that = E. So, u is a quasi-interior point.
(4) Let u> 0 be a quasi-interior point of P2 and let JT denote the closed
ideal generated by the range of T. That is,
JT={xEE: ay>Osuchthatlxl<Ty}.
By Lemma 9.1, JT is a closed -hyperinvariant ideal for T, and so = P2.
Since u> 0 is a quasi-interior point, the principal ideal generated by u
is norm dense in P2. Clearly,
T(E) c JT.
This implies = JT = E, i.e., Tu is also a quasi-interior point of P2.
350 9. Irreducible Operators
A similar result holds true for band irreducible positive operators; see
Exercise 10 at the end of the section. We now introduce the classes of
"expanding" operators.
Definition 9.4. A positive operator T: B + B on a Banach lattice is said

to be:
(1) strongly expanding if for each x > 0 the vector Ti is a quasi-

interior point,
(2) expanding if for each x > 0 the vector Tx is a weak unit, and
(3) weakly if for each x > 0 there exists some n (depend-
ing upon x) such that the vector is a weak unit.
There are some other classes of operators that are closely related to the
classes of operators introduced above. These are the classes of Krein and
strong Krein operators that were introduced in [8]. Their precise definition
is as follows.
Definition 9.5. A positive operator T: B + B on an arbitrary

with unit is said to be:
(1) a strong Krein operator whenever T maps each non-zero positive

vector to a (strong) unit, and
(2) a Krein operator whenever for each x > 0 there exists some n
(depending upon x) such that the vector TThx is a (strong) unit.
The properties introduced in Definitions 9.4, and 9.5 are quite

strong, and positive operators with such properties need not exist on an
arbitrary Banach lattice. For instance, Krein and strongly expanding op-
erators require the presence of strong units and of quasi-interior points,
respectively, while expanding operators can possibly exist only on spaces
with weak units, andas we shall seeideal irreducible positive operators
can exist only on Banach lattices with the countable sup property. Some
interesting examples of compact integral Krein operators on
will be presented in Example 9.39.
The next result shows schematically the relationships between the vari-
ous classes of positive operators introduced above.
Theorem 9.6. We have the following implication scheme for positive op-
erators on a Banach lattice:
Strongly Krein
V
Krein Strongly Expanding
Expanding
Ideal Irreducible Weakly Expanding
Band Irreducible
V
In general, no other implication is true.
Proof. We leave the details of proving the indicated implications as an

exercise. Several examples below demonstrate that in general no other
plication holds.
Example 9.7. An order continuous band irreducible positive operator that
is neither ideal irreducible nor weakly expanding.
Consider the positive operator T: defined by
T(xi,x2,...) =
Clearly, T(co) CO so T is not ideal irreducible. Also, it is easy to

and
see that T is order continuous. Now, if x = (1, 0, 0,. > 0, then it easily
follows that is not a weak unit for each k, and hence T is not weakly
expanding.
To see that T is band irreducible, assume that T(B) B for some band
B {0}. If 0 <x E B, then the element Tkx E B has its first k components
non-zero. This means that for each k the band B contains a positive vector
whose kth component is non-zero. So, if some v E satisfies v 1. B, then
v = 0, which shows that B = i.e., T is a band irreducible operator.
Example 9.8. A Krein operator which is neither a strong Krein nor an
operator.
The operator T: defined by
T(xi, X2, x3, S..) = (xi + X2, X1 + a;3, + X4,
is a Krein operator which is neither strong Krein nor expanding. El
Example 99. An expanding operator which is not strongly expanding.

Consider the operator T : C[O, 1] + C{O, 1] defined by the formula
Tx(t) = / sin(st)x(s)ds.
JO
It is easy to verify that T is an expanding compact operator. Since Tx(O) = 0
for each x E C[0, 1], this operator fails to be strongly expanding. Also,
notice that if J = {x e C[0, 1]: x(0) = 0}, then J is a closed
ideal and T(J) J holds. This shows that T is not ideal irreducible; for
more properties see Exercise 8 of Section 9.2.
Example 9.10. An ideal irreducible operator which is not weakly expanding.
Consider the operator T in Example 9.7 restricted to co. As noticed
there, T(co) co. Since CO has order continuous norm, the notions of ideal
irreducibility and band irreducibility coincide, and so T: co * co is ideal
irreducible. The final argument in Example 9.7 shows that T: co + co is
not weakly expanding.
Example 9.11. A weakly expanding operator which is neither expanding
nor ideal irreducible.
Consider first the positive operator 8: co p CO defined by
8(xi, X2,...) = (x2 + X1, + +
It should be immediate that S is weakly expanding but fails to be expanding.
Arguing as in the last part of Example 9.7, we see that S is ideal irreducible.
Moreover, since co has order continuous norm, it follows that S is also order
continuous.
Next, extend S to the Banach lattice c of all convergent sequences by
letting
T(xi, x2,...) = S(xi x2 +
where = and e = (1, 1, 1,. .). The reader can verify that
.
T: c + C 15 a strictly positive operator, it is weakly expanding, and (by

virtue of T(co) C CU) it fails to be ideal irreducible. In addition, it should
be noted that T is not u-order continuous.
Example 9.12. A rank-one expanding order continuous positive operator
which fails to be ideal irreducible.
Consider the sequence u = (1, i,...) and note that u is a weak unit
of Also, let R be a strictly positive order continuous linear
functional; for instance we can define by = Now consider
the rank-one order continuous positive operator T = 0 u: *
Clearly, T(co) co, and so T is not ideal irreducible. However, if 0 e

then T(x) = is a weak unit, and so the compact positive operator T is
9. 1 . Irreducible and Expanding Operators 353
expanding. Finally, it is interesting to note that if we replace u by a strong

unit (for instance, let u = (1, 1, 1,. .)), then u becomes a strong Krein
.
operator on which is not ideal irreducible.
It is clear that each expanding operator is also strictly positive. The

next lemma presents two useful sufficient conditions for the converse to be
true.
Lemma 9.13. Let 8: E * E be a strictly positive operator. If an ideal
irreducible (resp. a u-order continuous band irreducible) positive operator
T: E * E satisfies TS < AS for some > 0, then TkS and STk are
strongly expanding (resp. expanding) for each k = 0, 1, 2
Proof. We shall consider only the case when T is positive and ideal irre-
ducible. The other case is similar.
The proof is by induction on k. Assume first k = 0. Let x > 0 and put
u = Sx > 0. From T(Sx) ASx (i.e., Tu we see that c
By the ideal irreducibility of T, we get = E. That is, u = Sx is a quasi-
interior point, and so S is strongly expanding. Now assume that is
strongly expanding. Since = T(Tc_lS) = and
(by the induction hypothesis) is strictly positive, the above argument
shows that is strongly expanding.
Finally, since S is strongly expanding and Tk is strictly positive, it should
be immediate that is also strongly expanding.
It is natural to ask here if the order of operators in Lemma 9.13 is

essential, i.e., if the assumption TS AS can be replaced by ST AS. We
refer to Exercise 17 at the end of this section for examples demonstrating
--that the order is essential.
Corollary 9.14. If T: E E is an ideal irreducible positive operator,
then for each A > r(T) and each k =0, 1,2,... the operator R(A, T)Tk is
strongly expanding (and hence ideal irreducible).
Similarly, if T: E E is a ci-order continuous band irreducible positive
operator, then for each A > r(T) and each k = 0,1,2,... the operator
R(A, is expanding.
Proof. Note that for each A > r(T) the operator R(A, T) is strictly positive
and satisfies T) AR(A, T); see Lemma 6.23(3).
The existence of an ideal irreducible positive operator on a Banach lattice

E imposes a severe restriction on E.
Lemma 9.15. If a Bariach lattice E admits a compact strictly positive op-

erator, then E also admits a strictly positive linear functional. In particular,
in this case, E has the countable sup property.1
Proof. Let T: E E be a compact strictly positive operator. Fix a count-
able set {cbi, cb2,. .} of continuous non-zero linear functionals in the range
.
R(T*) of T* that is norm dense in R(T*) and put = 0.

(This is possible since T*: E* E* is a compact operator.) We claim that
is strictly positive on E.
To this end, let x> 0. If cb(x) = 0, then = 0 for each n and so
= = 0 for each Hence, Tx = 0, contrary to the strict
positivity of T. Thus, cb(x) > 0 and so is strictly positive. This implies
that E has the countable sup property.
Corollary 9.16. If a Banach lattice E admits a compact-strictly positive
operator, then for positive operators on E the notions of u-order continuity
and order continuity coincide.
Corollary 9.17. If a Banach lattice E admits an ideal irreducible positive
operator, then E has the countable sup property.
Proof. Let T: E E be an ideal irreducible positive operator. Fix some

scalar A > r(T) and note that (by Corollary 9.14) the positive operator
Q = R(A, T) is strongly expanding. Now fix 0 < E* and 0 0, it follows
that Kx = b(x)u > 0. Hence, KQ: E E is a compact strictly positive
operator and so, by Lemma 9.15, E has the countable sup property.
Exercises
1. Verify the Scheme" of Theorem 9.6.

2. Show that every weakly expanding operator is strictly positive.
3. Let 8, T: E E be two arbitrary positive operators on a Banach lattice.
If 0 S T and S is band (resp. ideal) irreducible, then show that T is
likewise band (resp. ideal) irreducib'e.
4. If S and T are two positive operators on a Banach lattice and one of them
is band (resp. ideal) irreducible, then show that S+T is band (resp. ideal)
irreducible.
1
Recall that a Riesz space E is said to have the countable sup property whenever for
every subset D of E having a supremurn there exists an at most countable subset C of D
sup C = sup D. Any Riesz space that admits a strictly positive linear functional has the countable
sup property; see [29, Theorem 2.6, p. 12].
9.1. Irreducibleand Expanding Operators 355
5. Give an example of a band irreducible positive operator T that is not a

supremum of a family of ideal irreducible positive operators dominated
byT.
6. We know that each central operator leaves every ideal invariant while an
ideal irreducible operator does not leave invariant any non-trivial closed
ideal. So, the following conjecture seems plausible: If S is a central
positive operator and T is an ideal irreducible positive operator, then S A
T = 0. Disprove this conjecture.
7. If a positive operator T on a Banach lattice is either strongly expanding,
or expanding, or strongly Krein and S is a strictly positive operator, then
show that TS also possesses the corresponding property of T.
8. If a Banach lattice E has order continuous norm, then show that a positive
operator on E is expanding if and only if it is strongly expanding.
9. Show that a positive operator on a Banach lattice with order continuous
norm is band irreducib'e if and only if it is ideal irreducible.
10. If T: E * E is a a-order continuous band irreducible positive operator,
then show that:
(a) T is strictly positive.
(b) T carries weak units to weak units.
(c) E has weak units.
[HINT: For (a) if for some x> Owe have Tx = 0, then = {O}
(where denotes the band generated by x), and so = E which
implies T= 0. For (b) fix a weak unit u > 0, and note that Tu > 0
implies {O}. Now if x> 0, then x Amu x implies Tx A mTu Tx,
and so Tx E So, is T-invariant, and this implies = E. For
(c) mimic the proof of Theorem 9.3(4).]
11. For a positive operator T: E E on a Banach lattice establish the
following properties.
(a) If Tc is ideal (resp. band) irreducible for some k > 1, then T is also
-- ideal (resp. band) irreducible.
(b) If T is ideal irreducible, then T2 need not be ideal irreducible.
[HINT: For (b) consider T: JR2 R2 defined by T(x,y) = (y,x).]
12. For a Markov operator T: E E on an AM-space with unit e establish

the following.
(a) There exists some 0 < E E* satisfying = cb. (Note that if
E= and e = 1, then is a regular Borel measure on known
as an invariant measure for T.)
(b) If 0 < E E* satisfies = then the rank-one positive operator
S=be satisfies STTS=S.
[HINT: For (a) consider the convex subset
of Note that the adjoint operator T*: E* is

and satisfies T*(C) C. Thus, by the classical Tychonoff Fixed Point
Theorem, there exists some E C such that T* =
13. Consider the positive operator T: C[O, 1] C[O, 1] defined by
Tf(t) = (*)
forallO<t<landTf(O)=f(O). Showthat:
(a) T defines indeed a positive operator on C[O, 1].
(b) T is one-to-oneand hence strictly positive.
(c) T is a Markov operator.
(d) T is not band irreducibleand hence not ideal irreducible either.
(e) The point spectrum of T is the set
[HINT: Notice that if IA < then the continuous complex

I
valued function fECc[O, 1], given by satisfies TfAf.]

(f) T is not a compact operator.
14. Let 1 0 and Tf(0)=f(0). Show that (**) defines a one-to-one strictly
s
positive operator from oc) to oo) such that =
15. Prove the converse of Corollary 9.14. That is, show that if T: E E
is a positive operator on a Banach lattice such that T) is strongly
expanding for some A > r(T), then T is ideal irreducible. [HINT: If a -
closed ideal J is T-invariant, then J is also R(A, T)-invariant.]

16. For a positive operator T: E E on a Banach lattice show the follow-
ing.
(a) The operator T is ideal irreducible if and only if {0} and E* are
the only bands in E* that are invariant under the adjoint
operator T*: E* E*.
(b) If the adjoint operator T* is band irreducible, then T is ideal ir-
reducible. Give an example of an ideal irreducible operator whose
adjoint is not band irreducible.
(c) If E has order continuous then T is ideal irreducible if and
only if T* is band irreducible.
[HINT: For each ideal J of E let
= E E*: 0 for all XE J}
and for each band B of E* let B = {x E E: (x, x*) = 0 for all E B}.
For (b) notice that the adjoint of any positive operator on C[0, 1] cannot
be ideal irreducible since the norm dual of C[0, 1] does not have any quasi-
interior points. To establish (c), use (a) in conjunction with the fact that
a Banach lattice has order continuous norm if and only if every band in
E* is see [29, Theorem 9.1, p. 60].]
17. Give examples of two positive operators S and T on atomic and non-
atomic Banach lattices such that T is ideal irreducible, ST 5, and
S is a non-expanding strictly positive operator. (This shows that the
inequality TS AS in Lemma 9.13 cannot be replaced by ST AS.)
9.2. Ideal Irreducibility and the Spectral Radius

The KreinRutman Theorem 7.10 asserts that when the spectral radius of a
compact positive operator is positive, then it is an eigenvalue corresponding
to a positive eigenvector. However, the KreinRutman theorem does not
tell us when the spectral radius of a compact positive operator is positive!
A glance at the Volterra operator shows that the spectral radius can be
zero. The purpose of this section is to present conditions that guarantee the
positivity of the spectral radius for compact positive operators. To avoid
trivialities, we shall assume that all operators appearing in the statements of
the results below are non-zero and that the dimensions of the Banach lattices
are greater than one.
We start our discussion with the notion of local quasinilpotence which
was introduced first in [10].
Definition 9.18. A bounded operator T: X X on a Banach space is said
1
to be (locally) quasinilpotent at a vector xO if lim n = 0.
fl+OO
It is well known that an operator T: X X on a Banach space is

quasinilpotent if and only if it is locally quasinilpotent at each point of X;
see Exercise 10 in Section 6.1. On the other hand, a one-to-one operator can
easily be locally quasinilpotent at some points without being quasinilpotent;
see Exercise 4 at the end of this section.
As we shall see later, non-zero compact operators on infinite dimensional-
Banach spaces always have nontrivial closed hyperinvariant subspaces; see
Theorem 10.15. The next basic result, due to B. de Pagter, guarantees the
existence of non-trivial closed -hyperinvariant ideals for locally quasinilpo-
tent compact positive operators. B. de Pagter's proof is a skillful adapta-
tion to the Banach lattice setting of H. M. Hilden's proof of Lomonosov's
Invariant Subspace Theorem 10.19 as presented by A. J. Michaels in [241];
see Theorem 10.15. The fact that quasinilpotence can be replaced by local
quasinilpotence has been observed in [7].
Recall that if T: X X is a bounded operator on a Banach space, then
its commutant {T}' is the subalgebra of L(X) consisting of all bounded
operators on .K that commute with T, i.e.,
{T}'={Ser(X): ST=TS}.
Theorem 9.19 (de Pagter [260]). Every mom-zero compact positive operator
which is locally quasimilpotent at a mom-zero positive vector has a mom-trivial
closed -hyperimvariamt ideal.
Proof. Let T: E E be a (non-zero) compact positive operator on a

nach lattice. We can assume that T is strictly positive; otherwise (according
to Lemma 9.1) the null ideal NT of T is a non-trivial closed hyperinvariant
ideal. By Lemma 9.1 the non-zero closed ideal JT generated by the range
of T,
JT={XEE:
is lattice hyperinvariant for T. If E, then JT is the desired closed
-hyperinvariant ideal for T. So, assume that JT = E.
The compactness of T implies that the range of T is separable; see
Exercise 6 in Section 2.4. So, there exists a countable subset {wi, w2, . . of
the range T(E) consisting of vectors that is norm dense in T(E).
Let w= > 0. The obvious inclusion {wi, w2, .} c
. implies
T(E) c Hence, E = JT and so = E. In other words, in this
case, we can also suppose that E has quasi-interior points. Next, consider
the collection of operators
A={AEL(E): IAXIBlxl for allxEE}.
Clearly, A is a subalgebra of L(E) containing the center of E, i.e., 2(E) A.
Now for each x e E, let
Ax={Ax: AeA}.
We claim that Ax is an -hyperinvariant vector subspace of E. To see
this, let 0 S e L(E) satisfy ST = TS. Choose any A e A and consider
the vector y = Ax e Ax. Pick 0 B E {T}' satisfying Az! for all
z E E, and note that SAz < SIAzJ (SB)Jz for each z e E and that
SB e {T}'. This implies SA e A and so Sy = SAx e Ax. In other words,
Ax is -hyperinvariant for each x e E. Consequently, the norm closure Ax
is likewise an hyperinvariant subspace of E for each x.
Since I e A, we have x = Ix e Ax, and so Ax is non-zero for each
zero vector x E E. Next, we claim that is, in fact, a closed ideal in E for
each x. To see this, let y e E satisfy JAjxl, where each E A.
Since (in view of the Riesz Decomposition Property [29, Theorem 1.2, p. 3])
we can write Y = yj with for each i, we can assume that
A e A. Now a glance at Lemma 4.17 guarantees the
existence of a sequence c 2(E) c A such that y. Since
e A for each m, we immediately get that y E Ax, and so Ax is an ideal.
To finish the proof, it suffices to show that Ax E for some x 0. To

this end, assume by way of contradiction that Ax = E for each x 0. Let
1
u> 0 be a vector of local quasinilpotence of T, i.e., )lT72ulIn = 0.
The strict positivity of T implies Tu > 0. So, by the continuity of T,
there exists an open ball U centered at u such that
0 U and 0 0 T(U).____ (*)
From our assumption that Ax = E for each x E T(U), we see that u e Ax
for each x E T(U).
Now for each e T(U) pick some operator A such that e U,
and then choose some open ball centered at x such that U.
The compactness of T_implies that_T(U) is a norm compact set. Since the
open balls x E T(U)} cover T(U), there exist Xi, E T(U) such
that T(U) For simplicity let us write instead of and
instead of Next, notice that there exists some operator 0 R E {T}'
satisfying IAjzt for all i = 1,. k and all z e E.
. ,
Since Tu E T(U), there exists some index ji E {1,. , k} such that

Tu E B31. This implies e U and so e T(U). Repeating this
argument inductively, we obtain a sequence of indices from {1,. k} . . ,
such that
E U
for each ri. From IAjzI and RT = TR we get for each
ri. The latter implies 1IRI1Th Hence,
1 1
0,
and so > 0. E U for each ri, it follows that 0 E U, contrary to
Since
(*). This contradiction completes the proof of the theorem.
Corollary 9.20 (de [260]). An ideal irreducible compact positive
Pagter
operator on a Banach lattice has a positive spectral radius.
It should be pointed out that the compactness assumption in the above
corollary is essential; see the comments at the end of Section 9.3. We are
now in a position to sharpen the previous result by proving the positivity of
the spectral radii of two commuting positive operators.
Corollary 9.21 (AbramovichAliprantisBurkinshaw [7]).If an ideal ir-
reducible positive operator T: E * E commutes with a compact positive
operatorS: E E, then r(T) > 0 and r(S) > 0.
Proof. Let 5, T: E E be two positive operators on a Banach
lattice. Assume that S is compact, T is ideal irreducible, and ST = TS.
From Theorem 9.3, it follows that ST = TS > 0.
Assume that r(ST) = 0. Then, by Theorem 9.19, the compact positive

operator ST has a non-trivial closed -hyperinvariant ideal J. Since the
operator T commutes with ST, we see that T(J) J, which contradicts
the ideal irreducibility of T. Hence, r(ST) > 0. It remains to observe that
(in view of the commutativity of S and T) we have 0 <r(ST) r(S)r(T);
see Exercise 8 of Section 6.1. This implies r(T) > 0 and r(S) > 0.
The next results show that the compactness assumption in Corollary 9.20
can be considerably relaxed.
Corollary 9.22 (AbramovichAliprantis--Burkinshaw [7]). If T is a com-

pactly dominated ideal irreducible positive operator on a Banach lattice, then
every positive operator S in the commutant of T has a positive spectral ra-
dius, i.e., r(S) > 0.
Proof. Let 0 T K for some compact operator K and let S be a positive

operator commuting with the ideal irreducible operator T. By Theorem 9.3
the operator T is strictly positive. Hence, by Corollary 2.35, the strictly pos-
itive operator T3 is compact. Theorem 9.3 shows that E has quasi-interior
points and that T carries quasi-interior points to quasi-interior points. This
implies that the positive operator ST3 is non-zero. Furthermore, ST3 is
compact and commutes with T. Finally, Corollary 9.21 implies r(5T3) > 0,
and so from r(5T3) r(S)r(T3) we obtain r(S) > 0.
Recall that a Banach space X is said to have the DunfordPettis

Property if 0 in X and 0 in imply 0. It is well
known that AL- and AM-spaces have the DunfordPettis Property; see [30,
Section 19].
Lemma 9.23. If X 8 y T Z are weakly compact operators between

Banach spaces and Y has the DunfordPettis property, then TS is a
pact operator.
Proof. See [30, Theorem 19.8, p. 337].

Corollary 9.24. Let T: E E be an ideal irreducible weakly compact
positive operator on a Banach lattice. If E has the DunfordPettis property
(in particular, if E is an AL- or an AM-space), then T has a positive spectral
radius.
Proof. In this case T2 is a strictly positive compact operator which com-

mutes with the ideal irreducible operator T.
B. de Pagter's Theorem 9.19 guarantees that a compact quasinilpotent

positive operator T: E p E on a Banach lattice has a non-trivial -hyper-
invariant ideal J. What this result does not guarantee is that T(J) {O}.
Here is a simple example.
Example 9.25. Consider the positive operator T: defined by the

formula T(x, y) = (y, 0). Clearly, T2 = It is easy to verify that the
ideal (in fact, the band) J = {(x, 0): x E is the only non-trivial closed
hyperinvariant ideal for T. However, T(J) = {0}.
Surprisingly, if we exclude the case T2 = 0, then the lattice hyper-

invariant ideal J guaranteed by Theorem 9.19 can be chosen to satisfy
T(J) {0}.
Theorem 9.26. If a compact quasinilpotent positive operator T: E + E on

a Banach lattice satisfies T2 0, then there exists a non-trivial closed -
hyperinvariant ideal J for T such that T(J) {0}.
Proof. Let T: E + E be a compact quasinilpotent positive operator on a

Banach lattice such that T2 0. Let N = {x E E: = 0}, the null
ideal ofT, and define the operator T: E/N E/N by T[x] = [Tx]. Clearly,
T is a quasinilpotent compact positive operator. We also claim that T is
non-zero. Indeed, if we pick some x> 0 satisfying T2x> 0, then Tx N,
and hence T[x] = [Tx] > 0 in E/N, which shows that T> 0.
By Theorem 9.19 there exists a non-trivial closed -hyperinvariant ideal
J for T. Let J = {x E E: [x] E J}. Since the quotient map x [x] is
a lattice homomorphism, it follows that J is a non-trivial closed ideal in E
that obviously is invariant under T. We claim that T(J) {0}. To see
this, pick some 0 < x E J satisfying 0 < {x] E J and notice that Tx > 0.
Otherwise, Tx = 0 implies x E N, that is, [x] = 0, which is a contradiction.
Finally, notice that J is an -hyperinvariant ideal for T. To verify this,
take a positive operator 8: E E that commutes with T. If [x] = [y], i.e.,
ifxyEN,thenfromOT18x8y1 =0, we
see that 8x Sy E N, that is, [Sx] = [Syl. Therefore, S defines an operator
S 0 on E/N via the formula S[x] = [Sx]. As S commutes with T, the
inclusion 8(J) C J holds and from this we infer that 8(J) C J.
'\'Ve conclude the section with a general result concerning strongly ex-
panding operators.
Theorem 9.27. Let Q, 5: E E be two positive operators on a Banach

lattice such that:
(1) Q is a strongly expanding operator.
(2) S dominates a compactly dominated positive opeTatorin particu-
lai', this is so if S itself is dominated by a compact operator.
Then r(QS) > 0, i.e., the spectral radius of the operatoT QS is positive.
Proof. Let Q S satisfy the properties stated in the theorem. Pick a

and
positive operator L and a compact operator K satisfying 0 < L S and
LK.
Consider the positive operator C = QL and note that C is non-zero
since Q is strongly expanding. Note also that if Ckx > 0 for some x > 0 and
some k, then Ckx is a quasi-interior point. (This follows from the fact that
Ck can be written in the form Ck = QR for some R > 0.) This guarantees
that Ck > 0 for each k. Since 0 < C = QL QK, Corollary 2.35 implies
that C3 is a compact operator.
Assume that r(C) = 0. Then r(C3) = [r(C)]3 = and taking into
0
account that (C3)2 = C6 > 0, Theorem 9.26 guarantees the existence of a
non-trivial closed -hyperinvariant ideal J for C3 with C3(J) {0}. Hence,
for some x > 0 the quasi-interior point C3x belongs to the closed ideal J,
which implies that J = E, contradicting our assumption that r(C) = 0. So,
r(QS) r(C) > 0.
Corollary 9.28. If the commutant of a positive operator T contains an

ideal iTreducible positive operator S and a compact positive operator K, then
r(T) > 0.
Proof. Let TS = ST and TK = KT, where S > 0 is ideal irreduc-

ible and K > 0 is compact. Fix A > T(S) and note that the op-
erator R(A, 5) is strongly expanding by Corollary 9.14. This implies
TR(A,S)K > 0. Since both R(A,S) and K commute with T, we get
T[R(A,S)K] = [R(A,S)K]T> 0. Consequently, TK = KT> 0. Next,
notice that the operators R(A, 5) and TK satisfy the hypotheses of Theo-
rem 9.27, and so
0 <T(R(A,S)TK) = r(T[R(A,S)K]) T(T)r(R(A,S)K).
Hence, r(T) > 0.
Corollary 9.29. Let 5, T: E i E be two commuting positive operators on

a Banach lattice. If T is ideal irTeducible and some power of S dominates
a compactly dominated positive operator, then r(S) > 0 and T(T) > 0.
Proof. Assume sm dominates some compactly dominated positive

tor. FLx some A > r(T) and let Q = R(A,T)T. By Corollary 9.14, Q is
strongly expanding. Now from Theorem 9.27, it follows that
o < r(QSm) = r(R(A, T)TSm) r(R(A, T))r(T) [r(S)]m,
and so r(S) > 0 and r(T) > 0 both hold.
The last four results are due to the authors and 0. Burkinshaw [7].
Exercises
1. Let E be a Banach lattice and let T: E * E be a positive operator

having a non-trivial closed invariant ideal. Show that there exists some
o < u E E satisfying E and [HINT: Let J be a non-
trivial closed T-invariant ideal and fix some 0 <x E J. Next consider the
0O
element u 1
2. Let T: X p X be a bounded operator on a Banach space. If A is an eigen-

value of T, then show that its eigenspace NT(A)={x E X: Tx = Ax} is
a closed T-hyperinvariant subspace.
3. (ErcanOnal [122]) Let T: X p X be a bounded operator on a Banach
space and let x be a vector in X. Show that T is locally quasinilpotent
at x if and on'y if (Tax) I 0 for each E
X is a Banach lattice, then show that T is locally quasinilpotent
at x if and only if Ix* (Tax) 0 holds for each 0 E
4. This exercise presents a one-to-one positive operator which is locally
quasinilpotent at some positive vector but fails to be quasinilpotent. Con-
sider the positive operator T: 2 2 defined by
T(xi,x2,x3,...) =
Show that T is one-to-one, is locally quasinilpotent at e2, and fails to be
quasinilpotent. [HINT: From = for all n 2, we see that
= (n44)' Hence = (n-i)!' from which it follows that
limn_,00 = 0. On the other hand, 1 implies
1
1 for each n.j
5. Let T: X p X be a continuous operator on a Banach space and let QT

denote the set of all points at which T is locally quasinilpotent, i.e.,
QT={xEX:
Show that QT is a T-hyperinvariant vector subspace. 2 [HINT: Let x, y
in QT and fix 0. Pick n0 such that and for
2
some applications and historical comments regarding QT see [233, 234].
1 1
all rt no. So, ITT1(x + + n <2E for all n n0.

Now let x3 e QT, let S e (X) satisf3T ST = TS, and note that
1 1 1 1
=
6. For the notation used here refer to Exercise 5 above. For a bounded op-
erator T: X X on a Banach space establish the following properties.
(a) QT = {O} is possible.
(b) QT can be dense without being equal to X.
[HINT: For (a) notice that every isometry T satisfies QT = {O}. For (b)
look at the backward shift.]
7. This exercise presents a pair of positive operators that satisfy the hy-
potheses of Corollary 9.21. Consider the positive operator T:
defined by
Tt'x 1, x2,x (X1+X2
2 ' 2 ' 2

(a) The operator T is a Krein (and hence an ideal irreducible) non-
compact positive operator.
(b) The spectral radius of T equals one, i.e., r(T) = 1.
(c) There exists some 0 e satisfying =q5(e) = 1 and T*q5 =
where e denotes the constant one sequence, i.e., e = (1, 1, 1,. .
(d) The positive operator S = on satisfies TS = ST = S.

[HINT: Note that T is a Markov operator and then use part (b) of Exer-
cise 12 of Section 9.1.]
8. Consider the operator T: C[0, 1] C[0, 1] defined by the formula
p1
Tx(t) I sin(st)x(s) ds.
Jo
Show the following:
(a) T is one-to-one and strictly positive.
(b) T is a compact operator.
(c) If J = {x e C[0, 1]: x(0) = 0}, then J is a non-trivial closed ideal
and T(J) C J holdsand so T is not ideal irreducible.
(d) For each x > 0 the function Tx is a quasi-interior point of J. In
particular, the operator T: J J (the restriction of T to J) is
strongly expandingand hence ideal irreducible.
(e) The spectral radius of the operator T: C[0, 1] C[0, 1] is positive
(i.e., r(T) > 0) and coincides with the spectral radius of the operator
T: J * J (the restriction of T to J).
[HINT: For (a) notice that if Tx = 0, then sin(st)x(s) ds = 0 for
each n = 0, 1 Now observe that the algebra generated by {1, s2} is
(by the StoneWeierstrass theorem) norm dense in C[0, 1].]
9. Generalize Corollary 9.21 as follows: If an ideal irreducible positive oper-
ator T: E E commutes with a compact positive operator 5: E f
then neither T nor S is locally quiasinilpotent at any positive vector.
9.3. Bandlrreducibility and the Spectral Radius 365
10. Generalize Corollary 9.22 as follows: If T: E f E is a compactly domi-

nated ideal irreducible positive operator on a Banach lattice, then every
positive operator on E that commutes with T (in particular, T itself)
cannot be locally qiiasinilpotent at any positive vector of E.
11. Show that in Corollary 9.22 the hypothesis "S commutes with T" cannot
be replaced by "S commutes with some power of T" even when T is
compact. [HINT: Consider the pair of positive operators S, T: }R2
defined by S(x,y) = (y,O) and T(x,y) = (y,x). Then T2 = I and so
ST2 T28. However, r(S) = 0.]
12. Show that if an ideal irreducible positive operator T: E f E on a Banach
lattice has a compact power (in particular, if T is dominated by a compact
operator), then r(T) > 0. [HINT: Use Corollary 9.21.1
13. (de Pagter [260]) Show that if a power of an ideal irreducible posi-
tive operator T dominates a compactly dominated positive operator,
then r(T) > 0. [HINT: Use Corollary 9.29.]
9.3. Band Irreducibility and the Spectral Radius

In this section, we shall discuss the spectral radius of band irreducible pos-
itive operators. As we shall see, the results are analogous to the ones of
the previous section but not identical. When we deal with band irreducible
positive operators, quite often we must assume that they are also a-order
continuous. As in the previous section, we shall suppose that all operators
appearing in the statements of the results are non-zero and they act on
Banach lattices of dimension greater than one.
Lemma 9.30. Let T: E + E be a a-order continuous, compact, and strictly
positive operator on a Banach lattice. Then every positive operator on E that
commutes with T is order continuous.
Proof. By Corollary 9.16, we know that every a-order continuous positive
operator on E is automatically order continuous. Now assume that a positive
operatorS: E E satisfies ST = TS. Let 0. Then 0 and by the
compactness of T, we infer3 that II 0. So, 0. On the other
hand, if x 0 holds for each n, then from = Tx 0
and 0, we get Tx = 0. Since T is strictly positive, the latter shows
that x = 0, and so S
S is order continuous.
Let us mention that without the compactness of the operator T
Lemma 9.30 is not true even if T is in addition ideal irreducible; see Exer-
cise I at the end of the section.
Here is the argument. The compactness of T implies that the sequence has a norm
convergent subsequence, say Since 1. 0, Lemma 1.30(c) guarantees IJ
0,
and from this we obtain 1 0.
- ----
The analogue of Theorems 9.26 and 9.19 for lattice hyperinvariant bands
is as follows.
Theorem 9.31. If T : E E is an order continuous compact quasi-
riilpoterit positive operator on a Bariach lattice, then. there exists a non-trivial
-hyperinvariant band B for T. Moreover, if T2 0, then the band B can
be chosen to satisfy T(B) {O}.
Proof. Ifthe null ideal N = {x E: TIrcI = o} is non-zero, then N is (by
Lemma 9.1) a non4rivial T-hyperinvariant band. So, we can assume that
NT = {O} , and hence T is strictly positive. Let JT denote the closed ideal
generated by the range of T in E, i.e.,
JT={XEE:
By Lemma 9.1, JT is a closed lattice hyperinvariant ideal for T. Since
the operator T: JT JT is strictly positive and quasinilpotent, it follows
from Theorem 9.26 that T, as an operator on JT, has a non-trivial closed
-hyperinvariant ideal J JT. Since JT itself is -hyperinvariant, it fol-
lows that J, considered as an ideal of E, is also -hyperinvariant. Now let
B be the band generated by J in E. By Lemma 9.30, every positive op-
erator commuting with T is order continuous. This implies that B is an
-hyperinvariant band for T. To finish the proof, it suffices to show that
B E. To see this, assume by way of contradiction that B = E. Hence,
for each x > 0 there exists a net in J such that 0 x. By the
order continuity of T, we get I Tx, and by the compactness of T we
see that the increasing net is norm convergent in E. It follows that
0. Since J is J, and thus Tx E J for
all x E E. This implies J = contradicting our assumption J
To establish the second part, assume now that T2 0. In this case, we
consider the operator I': E/N E/N defined by i'[x] = [Tx], where N is
again the null ideal of T. Since T is order continuous, N is a band. We
claim that T> 0. Indeed, if we pick x > 0 such that T2x > 0, then Tx N
and so T[x] > 0. Clearly, T is quasinilpotent and compact. Furthermore,
we claim that T is order continuous.
To see this, let J. 0 in E/N. Replacing } by the net of all finite
infima of the net we can assume that 0 j holds in E. Since E is
Archimedean, there exists a net {y,), } of E satisfying 0 y%, for each
a and each A and 0; see [225, Theorem 22.5, p. 115]. From
J. 0, we see that [y,)j = 0, i.e., = 0 for each A. Therefore, replacing
{ by we can assume without loss of generality that J. 0 in
E. By the order continuity of T, it follows that Tx0, J. 0. So, by the order
continuity of the quotient map x 'p [x] (see Exercise 2 at the end of this
section), we obtain T is order continuous.
Band Irreducibility and the Spectral Radius 367
By the first part of the proof, there exists a non-trivial -hyperinvariant

band in E/N. Let D = {x E : [x} i3} . The order continuity of the
quotient map x [x] implies that D is a non-trivial -hyperinvariant band.
To complete the proof, it suffices to show that T(D) {o}. To see this, pick
any 0 <x E D satisfying 0 < [x] E i3. We claim that Tx> 0. Otherwise,
Tx 0 would imply that x E N, i.e., [x] = 0, which is a contradiction.
Using the preceding theorem we will show now that every order
ous compact positive operator lying in the commutant of a band irreducible
positive operator has a spectral radius.
Corollary 9.32. Every order continuous compact positive operator T that
commutes with a band irreducible positive operator has a positive spectral
radius, i.e., r(T) > 0.
Proof. Assume that T commutes with a band irreducible positive operator
S. If r(T) = 0, then by Theorem 9.31 the operator T has a non-trivial hyper-
invariant band B. In particular, 8(B) c B must hold, which is impossible.
So, r(T) > 0.
Corollary 9.33 (SchaeferGrobler). Every ci-order continuous band irre-
ducible compact positive operator on a Banach lattice has a positive spectral
radius.
Proof. Let T: E E be a ci-order continuous band irreducible compact

positive operator. Then T is strictly positive (see Exercise 10 in Section 9.1),
and so (by Lemma 9.30) T is also order continuous. By Corollary 9.32,
r(T)>0.
Under the assumptions of Corollary 9.33 not only T but also every pos-
itive operator in the commutant of T has positive spectral radius.
Corollary 9.34. Let T be a ci-order continuous band irreducible compact
positive operator. Then every positive operator S in the commutant of T has
positive spectral radius, i.e., r(S) > 0.
Proof. Let 0 <5 E {T}'. Observe first that ST is a ci-order continuous
positive operator. To see this, note that since T is strictly positive, we have
ST = TS > 0. Next, let j 0. Then j 0 and by the compactness
of T, we get IITxnII j 0. So, from 0 $ j and 0, we obtain
ST ST commutes
with T, Lemma 9.30 implies that ST is also order continuous.
Now notice that since the compact positive operator ST commutes with
the band irreducible positive operator T, Corollary 9.32 guarantees that
0 <r(ST) r(S)r(T). Consequently, r(S) > 0. M
The analogue of Theorem 9.27 for bands is as follows.

Theorem 9.35. Let Q, 8: E p E be two positive operators such that:
( 1) Q is u-order continuous and expanding.
( 2) S dominates a compactly dominated u-order continuous positive
operatorin particular, this is the case if S itself is u-order con-
tinuous and dominated by a compact operator.
Then r(QS) > 0, the spectral radius of the operator QS is positiveS
Proof. The proof is a modification of the proof of Theorem 9.27. Pick a

u-order continuous operator L > 0 and a compact operator K> 0 such that
L <S and L K.
Define the operator C QL and note that if for some x> 0 we have
C3x > 0, then C3x is a weak unit. Now notice that (LQ)3 = L(QL)(QL)Q
is a compact strictly positive operator. By Lemma 9d5, the Banach lattice
E has the countable sup property, and so the operators Q and L are order
continuous. On the other hand, from C3 = Q(LQ)(LQ)L and the order
tinuity of L, we see that C3 > 0 and C6 > 0. To complete the proof, invoke
the arguments employed in the proof of Theorem and use Theorem 9.31
instead of Theorem
Now let E be a Dedekind complete Banach lattice. Recall that the

Dedekind complete Riesz space of all regular operators on E is denoted by
4(E) and that the Riesz space of all abstract integral operators is the
band generated by the finite-rank operators E in 4(E).
From this definition it follows that for each positive abstract integral oper-
ator S there exist a finite-rank operator K E (E) and a positive operator
L E (E) satisfying 0 < L S and L K. In particular, K is a com-
pact operator and so S dominates a compactly dominated order continuous
operator. In other words, each positive operator S E)dd satisfies
condition (2) of Theorem We are ready now to formulate and prove
the next result.
Theorem 9.36. Let E be a Dedekind complete Banach lattice. If a u-order
continuous band irreducible positive operator T: E p E commutes with a
positive abstract integral operator 5: E E, then r(T) > 0 and r(S) > 0.
Proof. Fix some A > r(T) and consider the operator Q = R(A, T)T.
Clearly, Q commutes with 5, is u-order continuous, and, by Corollary
is expanding. Therefore, the operators Q and S satisfy the hypotheses
of Theorem 9.35, and thus r(QS) > 0. Finally, using the inequality
r(QS) r(R(A,T))r(T)r(5), we obtain r(T) > 0 and r(S) >
9.3. Band Irreducibility and the Spectral Radius 369
The abstract formulation of the celebrated AndKrieger theorem [34,

194] is a special case of the preceding theorem.
Corollary 9.37 (AndKrieger) . If

T is a band irreducible abstract integral
positive operator on a Dedekind complete Banach lattice, then r(T) > 0.
We close the section with a brief historical comment. The AndKrieger

theorem above preceded most results in Sections 9.2 and 9.3. It initiated the
study of ideal and band irreducible operators on Banach lattices. The first
challenging task was to establish an analogue of Corollary 9.37 for cT-order
continuous band irreducible positive operatorsrecall that integral opera-
tors are automatically order continuous; see Corollary 5.12. H. H. Schae-
fer [292} was the first to do so assuming that the operator was compact
and that the cT-order continuous dual of the space was separating. Later
J. J. Grobler [139] removed the last assumption, and their final result is
Corollary 9.33.
The second challenging problem was whether or not each compact ideal
irreducible positive operator has a strictly positive spectral radius. The
affirmative answer was provided by the remarkable result of B. de Pagter
(Corollary 9.20).
We refer to [288] for an example showing that the compactness assump-
tion in the above results is essential and to [7] for an example of a nilpotent
compact (in fact, band irreducible positive operator.
Exercises
1. Show that the conclusion of Lemma 9.30 is false if the "compactness" of
T is replaced by "ideal irreducibility" of T. That is, give an example of
two commuting positive operators T, 8: E E on a Banach lattice such
that:
(a) T is strictly positive, order continuous, and ideal irreducible.
(b) S is not order continuous.
2. Let J be an ideal in a Riesz space E. Show that J is a band if and only if
the quotient map x [x], from E onto E/J, is order continuous. [HINT:
Assume that J is a band, that .1. 0 in E, and that [x] in E/J
for all a and some x E E+. Now use the fact that the quotient map
is a lattice homomorphism (see Exercise 2 of Section 3.4) to obtain that
0
3. Prove the following analogue of Corollary 9.28: Let T: E E be a a-
order continuous positive operator on a Banach lattice. If the commutant
of T contains a cr-order continuous band irreducible positive operator and
also a cr-order continuous compact positive operator, then T(T) > 0.
370 - 9. Irreducible Operators
4. Prove the following analogue of Corollary 9.29: Let S, T : E E be

two commuting positive operators on a Banach lattice. If T is a crorder
continuous band irreducible operator and some power of S dominates
a continuous compactly dominated positive operator, then both
operators have positive spectral radii, i.e., r(S) > 0 and r(T) > 0.
5. Complete the details in the proof of Theorem 9.35.
6. (Schaefer [292]; Grobler [139]) If T: E + E is a cr-order continuous band
irreducible quasinilpotent positive operator and 5: E E is a compact
positive operator, then show that [0, 5] fl [0, = {0} for each k 1.
[HINT: Let and
then from Theorem 9.35 it foHows that 0 <r(QTC) and so
r(T) > 0, a contradictionS]
7. (Caselles [86]) A Harris operator is a positive operator T: E + E on
a Dedekind complete Banach lattice such that some power of T is not
disjoint from the band of abstract integral operators.
Prove that every cr-order continuous band irreducible Harris operator
has a positive spectral radius. [HINT: Use Theorem 9.35.]
9.4. Krein Operators on C(12)-spaces

We shall discuss here several remarkable properties of Krein operators on
(real) C(11)-spaces, where denotes a compact Hausdorff space. As usual,
the constant function one on will be denoted by 1. The vector space
C(12) is a Banach lattice under the sup norm. Its complexification is the
Banach lattice Cc(12) of all continuous complex-valued functions. Most of
the results in this section are due to M. G. Krein. For a general discussion
of Krein operators general partially ordered vector spaces we refer the
reader to [8].
Recall that a function u E is called a strong unit, or simply a
unit, if for each x E C(12) there exists some A> 0 such that x Au. Clearly,
a function u e C(12) is a unit if and only if u(w) > 0 for each w E ft The
units of C(11) are precisely the interior points of the positive cone C(12)+.
Notice that if u is a unit, then so are: for each > 0, u + x for each
x E C(1l)+, and each v E C(12) with v u. It should also be clear that a
non-negative linear functional on is the zero functional if and only if
it vanishes at some unit. The boundary of C(12)+ with respect to the sup
norm is denoted
Lemma 9.38. If 0 < x E C(1l) amd y 0, them there exists some scalar
such that x + E
Proof. If x E 3 [C(12)+], then let = 0. Therefore, we can assume that

xE i.e., that x is a unit. Let A = E R: x + E
9.4. Krein Operators on -spaces 371
and put
,@=infA and 'y=supA.

Since x E we see that,@ < 0 <'v. We claim that either ,@ or -y is
finite. Indeed, if = oo and -y = oo, then (by the convexity of it
follows that x + ay E for all a E 1R. In particular, we have +ny x
for each n. So y = 0, which is a contradiction.
Now note that if,@ E then E and similarly if E 1R,
then X+7y E a[c(c1)+].
Recall that a positive operator T: is said to be:

o a strong Krein operator whenever T carries non-zero positive
elements to units, i.e., if 0 <x E C(cl), then Tx is a unit, and
a Krein operator whenever for each x > 0 there exists some

positive integer n (depending on x) such that is a unit.
Clearly, every strong Krein operator is a Krein operator. It should also be

noticed that every Krein operator is strictly positive.
Example 9.39. Here are some examples of Krein operators.
(1) Let A = be a real n x n matrix such that > 0 for all (i,j).
Then A is a strong Krein operator from to
(2) Assume that C(11) admits a strictly positive linear functional f.
(This is equivalent to assuming that there exists a regular Borel
measure whose support is If u is a unit, then the rank-one
operator T = f u is a strong Krein operator.
More generally, any finite-rank operator T =
where each is a strictly positive linear functional on C(11), uj > 0
for each i, and is a unit, is a strong Krein operator.
(3) Let be a non-zero regular Borel measure on a compact Hausdorif

space 11, and let T: C(11) C(11) be a positive integral operator
with a continuous strictly positive kernel. That is, there exists a
continuous real-valued function K on x 11 satisfying K(s, t) > 0
for all (s,t) Eli x 11, and
Tf(t) =
Jci
for each f E and all t E ft Then T is a strong Krein compact

operator.
(4) Let K: [a, b] x [a, b] [0, oo) be a continuous function such that
K(s,t) > 0 for all s < , where 0 is fixed. Then the
operator generated by the function K(s,t), i.e., the opera-

integral
tor T: C[a,b] p C[a,b}, defined by Tf(t) = ds, is a

compact Krein operator.
(5) This example is an important special case of the preceding example.
Consider a continuous function K: [a, b] x [a, b} [0, oo) satisfying
K(t, t) > 0 for all a t b. Then the integral operator generated
by K on C[a, b] is a compact Krein operator. (See Exercise 4 at
the end of the section.)
(6) If T: is a Krein operator and 8: is
an arbitrary positive operator, then S + T is a Krein operator.
Two useful properties of Krein operators are included in the next lemma.
Lemma 9.40. If T: p C(11) is a Krein operator, then:
(1) T carries ttnits to units.
(2) TTh is a Krein operator for each n 1 (and so TTh Al for each
n 1 and all scalars A 0).
Proof (1) Let a unit. Since T is a Krein operator, there exists some
be
positive integer k 1 such that v = is also a unit. Next, choose some
> 0 such that $ 'yu and note that v = 7Th holds. This
implies that T'u is a unit.
(2) Using (1) and an inductive argument, we see that carries units
to units for each k. Now fix n 1 and let x> 0. Choose some k such that
is a unit. But then (TTh)kx = is a unit, which proves that
is a Krein operator. El
Notice that from (1) above, it follows that if T is a Krein operator and we
let v = TI, then the operator x is a Markov operator. Regarding
the spectral radius of a Krein operator, we have the following.
Theorem 9.41. If T: p is a Krein operator, then:
(1) T has a positive spectral radius, i.e., r(T) > 0.
1
(2) r(T) = lim for each x> 0.
Proof. (1) By Lemma 9.40, Ti is a unit. This implies r(T) > 0; see
Exercise 2 at the end of this section.
(2) Let x > 0. Fix some k such that 'ii = TJcx is a unit. This implies
that the formula = inf{A 0: AtL} defines an equivalent norm
Krein Operators on -spaces 373
on C(f1). In particular, we have =

1 1 _Lnk
= = r(T).
1
Since and are equivalent norms, we get n = r(T)
We now turn our attention to eigenvalues of Krein operators. In light of

the fact that Krein operators are ideal irreducible, the next result parallels
Theorem 8.22.
Theorem 9.42 (M. G. Krein). If a Krein operator T: C(11) has a
positive eigenvector, say Tx0 = A0x0 for some xO > 0, then:
(1) The eigenvalue A0 coincides with the spectral radius of the operator
T, that is, A0 = r(T) > 0.
(2) The eigenvector xO is a unit and (up to a scalar multiple) is the

only positive eigenvector of T.
Proof. (1) Assume that Tx0 = for some xO > 0. Since T is a Krein
A0x0
operator, there exists some positive integer Ic such that is a unit. From
Tkxo = (Ao)kxo, it follows that A0 > 0 and that xO is a unit.
Now from Theorem 9.41, we obtain
1 1 1
r(T) = lim
fl-400
= lim = A0 lim
TL400
= A0.
Assume Tx = Ax for some x > 0 and some scalar A. By part (1),

(2)
we get that A = A0. Furthermore, by Lemma there exists some E IR
such that xO + E Clearly, 0. If x0 + > 0, then from
T(xo + = Ao(xo + and part (1), we infer that xO + is a unit, that
is, + E Int(C(Ifl+), which is impossible. Thus, xO + = 0 and hence
x=
Corollary 9.43. If T: C(Ifl is a compact Krein operator, then the
spectral radius r(T) is the only eigenvalue of T having a positive eigenvector
(which is necessarily a unit).
Proof. By Theorem 9.41, we have r(T) > 0. So, by the KreinRutman

Theorem 7.10, r(T) is an eigenvalue of T having a positive eigenvector. Our
conclusion now follows from Theorem 9.42. m
We now state a classical result, due to P. Jentzsch [162], regarding

integral operators.
Theorem 9.44 (Jentzsch). Let K: L2{0, 1] L2[0, 1] be an integral opera-

tor defined by
ci
Kx(t) = I K(s,t)x(s)ds,
Jo
where K: [0, 1] x [0, 1] R is a strictly positive continuous function. Then:
(1) The positive operator K is compact and has a positive spectral ra-
dius.
(2) The spectral radius r(K) is an eigenvalue of K and has a unique
(up to a scalar multiple) strictly positive eigenfunction.
(3) The peripheral spectrum of K consists only of the point r(K), that
is, if o(K) and r(K), then <r(K).
Notice that the compactness of K is a consequence of Theorem 7.7.

Also, according to Example 9.39(3), the operator K: C[O, 1] C[O, 1] is a
strong Krein operator. Since the range of the operator K: L2 [0, 1] L2[O, 1]
lies in C[0, 1], properties (1) and (2) in Theorem 9.44 follow immediately
from Theorem 9.42. However, property (3) is not that easy to establish
and requires some extra machinery. For a proof of property (3) we refer
the reader to [343, pp. 630633], where the reader will also find several
generalizations of Jentzsch's theorem to Banach lattices and Banach function
spaces.
Our next objective is to discuss interrelationships between fixed points
and eigenvectors of families of commuting positive operators on
We start by clarifying a crucial connection between fixed points and eigen-
vectors.
Lemma 9.45. Let T: be a positive operator, let u be a unit
in and consider the non-empty convex and w*..compact set
Also, consider the continuous mapping F: (Ca, w*) (Ca, w*) defined by
Then:
(1) Up to a scalar multiple, the positive eigenvectors of the adjoint

operator T* are precisely the fixed points of F.
(2) The set of fixed points of F is non-empty and
Proof. (1) Let E If = then =

if = then 0 and
= = =
_______
9.4. Kreiri Operators on C -spaces 375
(2) Clearly, is a non-empty, convex, and subset of

A direct verification shows that F indeed maps into and that
F: (Ct, w*) (Ca, w*) is a continuous function. Therefore, by Tychonoff's
Fixed Point Theorem (see, for instance [25, Corollary 14.51, p. 483]) there
exists some C such that =
Corollary 9.46 (M. C. Krein [192]). The adjoint of an arbitrary positive
operator on a -space has a positive eigenvector corresponding to a no ri-
negative eigenvalue.
A proof of the above result that does not use a fixed point argument can
be found in [185]. The next result generalizes the previous corollary to a
family of commuting operators.
Theorem 9.47 (M. G. Krein). Let be a family of pairwise com-
muting positive operators on some C(11)-space. Then the family of
a common positive eigenvector. That is, there exist
o < and a family of non-negative scalars such that
Proof. Consider again the non-empty, convex and set

b(1)=1}.
For each index let Fa: (C,w*) denote the continuous mapping
defined by
F *

By Lemma 9.45(1), every fixed point of Fa is an eigenvector of Tc corre-
sponding to a non-negative eigenvalue. Let Da denote the set of all fixed
points of Fa, i.e., let
By Leimna 9.45(2) each is a non-empty and subset of C

To complete the proof, we must show that 0, or equivalently
that the family of sets {Da: A} has the finite intersection
property.
The proof goes by induction. As mentioned above, each is non-
empty. So, for the induction step, assume that the intersection of any n
members of the family E A} has a non-empty intersection and let
be n + I indices. We must show that
To this end, start by choosing some 0 < E fl . fl Conse-
quently, for = 1) 0 it follows that = for i = 1, . , n. .
Now for each I <i <n define the set

Clearly, each is convex, and non-empty (since E Ci). Next,

we claim that (Ci) for each 1 i n. Indeed, if 'i/i E then
/ I ' T*
/ T* /
T* "F T* ( crrL+191 \

A1________

which shows that E

C C is a convex, and (in view of
E C) non-empty subset of C such that (C) C. By Tychonoff's
Fixed Point Theorem there exists some 0 < q5o C with = q5o.
Since C2, we have = and hence = for each
I i $ n + 1. Therefore, E Da2. Thus, Daj 0, and the
proof is finished.
As mentioned before, for every unit u E C(11) the formula

= inf{A > 0: Jxf Au}
defines a lattice norm on C(12) which equivalent to the sup norm.
is
We are now ready to establish that if a family of pairwise commuting

positive operators on C(12) fixes a unit, then the family of their adjoints also
fixes a positive linear functional.
Theorem 9.48 (M. G. Krein). Let {Ta}aEA be a family of pairwise com-
muting positive operators on a C(12)-space. If for some unit u E C(12) we
have TaU = u for all a, then there exists some 0 < E C(12)* such that
= for all a.
Proof. Assume that for some unit u we have TaU = u for each a E A. Since
tu x< tu implies tu = = tu, it follows
that IITaXIIu for each x E C(12) and each E A. This implies that
if .. . , are natural numbers and a1, ... , are arbitrary indices, then
'ar'a2
.
*
for each x E C(12).
Next, we consider the vector subspace Y of C(12) defined by
= E E A, xj with x=
We claim that Y flU = 0, where U = To see this, assume by

way of contradiction that Y flU 74 0. So, there exist indices ... , and
9.4.Krein Operators on -spaces -- - 377
functions xl,. , x7,, E

. such that the vector e (Ta. xj is a
unit. Fix these indices al,.. and for each m define the operator
. ,
m m m
Tm = ... .. .
k1=dk2=l
where we write instead of Clearly, each Tm is a positive operator

and TmU = 'a. Since the family consists of pairwise commuting
operators, we see that
it m m m
Tme = .. - xi)]
i=l k1=l k2=1
it m m m m rn
i=l k1=l
it m m m m
= ... fJ -
i=1 k1=l
If we let c = max{ i= 1, . . . , n}, then from (*) it follows that
flrpkj(rpm+1 m
U
=
Consequently, from the triangle inequality, we infer that
um rn
Since e is assumed to be a unit, there exists some 0 such that u
This implies
1= = IITm(U)tIu <T/IITrn(e)IIu
for each rn, which is impossible. Therefore, Y fl U = 0.
Since U is open, the Separation Theorem guarantees the existence of
some non-zero b E satisfying = 0 for all y E Y and 0 for
all x e U. Taking into account that U = we conclude that q5> 0.
From x E Y, it follows that
(x, = (Tax x, 0,
for each x e Thus, = b for each a E A, and this shows that
0 is a fixed point for the family of operators
Corollary 9.49. Let T: be a positive operator. If some A > 0
is an eigenval'ae of T having an eigenvector which is a 'anit, then the adjoint
operator T* also has A as an eigenval'ae with a positive eigenvector.
If all operators in the previous result are contractions, then we have the
following companion to Theorem 9.48.
Theorem 9.50. Let {Ta} be a family of pairwise commuting positive con-
tractions on C(cfl. If there exists a non-zero vector v e such that
v for each a, then there exists a non-zero positive linear functional
e satisfying = for each a.
Proof. Start by fixing some non-zero vector v E C(12) such that =v
for each a. Next note that if k1, k2,.. , are arbitrary natural numbers
and al, a2,. . . , are arbitrary indices, then
. lxii,
where denotes the sup norm.
Now we consider the same vector subspace Y of as in the proof of
Theorem 9.48. Assume that e E Y fl U, and then for each m construct the
operator Tm. If we let c = rnax{ Jjxj!i: i = 1,.. , n}, then it follows that
Now if > 0 satisfies IvI then
for each m, which is impossible. The rest of the proof is the same as the
proof of Theorem 9.48.
Corollary 9.51. If a positive contraction T: has 1 as an
eigenvalue, then there exists some 0 <4 e such that T*q5 =
Our next result is a famous theorem of A. A. Markov [232] and it is a

surprisingly simple consequence of Theorem 9.48. In order to formulate it,
we need some preliminary discussion.
Recall that if is an arbitrary non-empty set, then denotes the
Banach lattice of all bounded functions on With the sup
norm is an AM-space with unit 1 (the constant function one) and,
in particular, it is a C(cfl-space. Also, recall that the dual can be
identified with the space of all finitely additive measures on see [25,
11.40, p. 394]. Now if 'i-: is a mapping, then 'r defines a
positive operator CT: via the formula
= (x or)(8) = x(r(8)), x 8 E
The operator is, of course, the familiar composition operator determined

byr.
Theorem 952 (Markov). Let a family of pairwise commuting
be
mappings on an arbitrary non-empty set into itself Then there exists
some 0 < E such that = 4(x) for all a and all x E
9.4. Krein Operators on -spaces 379
Proof. As mentioned above = C(Q) for some compact Hausdorif

space Also, the family }aEA consists of pairwise commuting positive
operators on and = 1 for each The conclusion now is a
direct consequence of Theorem 9.48.
Exercises
1. Let T: be a positive operator. Show that for each unit
uE we have
1
= and r(T) =
2. Consider a positive operator T: p and put

m = minTl(w) and M = maxTi(w).
Show that m r(T) M and conclude from this that if Ti >> 0, then
the operator T has a positive spectral radius.
3. Show that the operator defined in Example 9.39(3) is indeed a compact
Krein operator.
4. Show that the operator defined in Example 9.39(4) is indeed a compact
Krein operator.
[HINT: To see this, note first that T2 is also an integral operator with
kernel given by the formula K2(t, s) K(t, u)K(u, s) du. Now observe
that K2 (t, s) > 0 for all < 2. To this end, let s, t E [a, b] satisfy
s < t and t s Is <2. the midpoint of the closed
interval [s, t], and put 6 Clearly, 0 <6 < , and if fUQ uI <6,
then ul Is uoI + Iu uol < +6 and similarly It uI <.
So, K(t, u) > 0 and K(u, s) > 0 for all u E (u0 6, u0 + 6). This implies
p5 /'UD+c5
K2(t, s) K(t, u)K(u, s)du K(t, u)K(u, s) du> 0.
= Ja J
Repeating the above process, we see that there exists some positive integer
n such that s) > 0 holds for all (t, s) E [a, b] x [a, b}. It follows that
T is a compact Krein operator.]
5. By Theorem we know that a Krein operator can have only one (up to
a scalar multiple) positive eigenvector corresponding to a positive eigen-
value. Give an example of a Krein operator whose eigenvalues are all
positive.
6. Show that the product of two commuting Krein operators is also a Krein
operator. Give an example of two non-commuting Krein operators whose
product is not a Krein operator. [HINT: Notice that the positive matrices
and B both define Krein operators on R whale

1 1 = 1 0
their product AB is not a Krein operator.

7. Theorem 9.47 was originally proved by M. G. Krein under the extra hy-
pothesis that C U for each a, where U = Theo-
rem 9.47 implies, of course, that version. Show that in turn the original
Krein theorem also implies Theorem 9.47. [HINT: For each 0 < 1
let = + 1, where I is the identity operator on C(cI). Clearly,
U holds, and so for each 0 there exists some 0 < E C
(where C {'eb E = 1}) and a family of non-negative scalars
such that
+ = (*)
The net (where 62 means 61 of C has aw*_convergent
subnet. We can assume that holds in C. Now fix a E A and
then pick some t> 0 with tl. From (*) it follows that
0 = = + t + 1.
This implies that the net has an accumulation point 0.
Now one more glance at (*) guarantees that =
8. Let be a family of pairwise commuting positive operators on
a If there exist some A > 0 and some unit it such that
Tofu Au for each a, then show that there exists some 0 <
such that = Ad, for each a. [HINT: Consider the family of positive
operators and use Theorem 9.48.]
9. Show that Theorem 9.50 does not follow from Theorem 9.48. [HINT: If
R2 and T: is the positive projection represented
by the matrix then T has 1 as an eigenvalue with eigenvector
However, T does not have any unit an eigenvector.]
10. Show that the commutativity assumption regarding the family of

erators in Theorems 9.48 and 9.50 cannot be dropped. [HINT: Let P
denote the family of all (Borel) probability measures on [0, 1]. For each
measure /i E 7' consider the positive operator C[0, 1] C[0, 1]
fined by = i.e., = ,u 1. Clearly, =
holds for all ii E P. In particular, it follows that {T,1: P} is
a family of non-commuting positive projections on C[0, 1J. In addition,
notice that = 1 for each E 7'. Assume now by way of
dictjoii that there exists some 0 < E M{0,1] = such that
= holds for each we can suppose that 1]) = L Hence,
= Tb = (1 = (1, = for each ,u E 2, which is impossible.]
Chapter 10
Invariant Subspaces
One of the most famous and still unsettled problems in functional analysis
and modern mathematics is the so-called invariant subspace problem
that is described by the following simple question:
o When does a bounded operator on a Banach space have a non-trivial
closed invariant subspace?
To recall some basic definitions related to the above question, let us
suppose that T: X X is a bounded operator on a Banach space and V is a
subspace of X.' Then V is non-trivial if V {O} and V X. We say that
V is invariant under T (or simply T-invariant) if T(V) V. Also, V is
said to be hyperinvariant for T (or simply T-hyperinvariant) whenever
V is invariant under every bounded operator on X that commutes with T,
i.e., if S e (X) and ST = TS imply that 5(V) V. If X is a Banach
lattice and T is a positive operator, then V is called lattice hyperinvariant
for T (or simply -hyperinvariant) if V is invariant under every positive
operator on X that commutes with T.
It is very easy to show (see the beginning of the next section) that for
non-separable Banach spaces the invariant subspace problem has an imme-
diate affirmative answer. For finite dimensional complex Banach spaces the
invariant subspace problem has an affirmative answer too. To see this, let
T: X X be a bounded operator on a finite dimensional complex Banach
space (of dimension greater than one). There is nothing to prove if T is a
multiple of the identity operator, since then each subspace is invariant. So,
1
Unless otherwise stated, throughout this chapter X will denote an arbitrary real or complex
Banach space of dimension greater than one. Similarly, E will denote a real Banach lattice of
dimension greater than one.
381
382 - I Invariant Subspaces
assume that T is not a multiple of the identity. Now if A is an eigenvalue

of T, then its eigenspace NA = {x X : Tx = Ax} is a non-trivial closed
hyperinvariant subspace (see Lemma 10.5). In view of the above comments
the invariant subspace problem should be formulated as follows.
C does a bounded operator on an infinite dimensional separable
ViThen
Banach space have a non-trivial closed invariant subspace?
For algebras and semigroups of operators the existence of a common non-
trivial closed invariant subspace is of interest even in the case.
P. Enflo [121] was the first to present an example of a bounded op-
erator on a real or complex separable non-reflexive Banach space without
closed invariant subspaces. Subsequently, C. J. Read [275] (see
also [276] and [277]) presented an example of a bounded operator on the
real Li without non-trivial closed invariant subspaces. Thus, these examples
have established that the invariant subspace problem in its general form has
a negative answer. However, for various important classes of Banach spaces
and operators the invariant subspace problem remains open. For instance,
it is not known if every bounded operator on a separable Hilbert space (or
on a separable reflexive Banach space) has a closed invariant
subspace.
There exists a vast literature dedicated to this problem but it is im-
possible to mention here even the most significant publications. From the
multitude of results on the existence of invariant subspaces we have chosen
only two topics to discuss in detail in this chapter. Our choice does not
necessarily depend on the importance of these topics but rather on their
proximity to the main theme of this bookpositivity, and to some extent
of course on the personal tastes of the authors. The topics chosen are:
(1) Lornonosov's theorem on the existence of hyperinvariant subspaces
for compact operators and some of its applications.
(2) The existence of invariant subspaces for positive operators, for op-
erators related to them, and for collections of positive operators.
As we will see, the invariant subspaces for such operators very often
have a special geometrical structurethey are ideals.
For the invariant subspace problem in Hilbert spaces and its history, see
the works of H. Bercovici [55], 5. Brown [76], P. R. Halmos [148, pp. 100
101], N. K. Nikolskii [257], C. Pearcy and A. L. Shields [262], H. Radjavi and
P. Rosenthal [272], and P. Rosenthal [284] and the references therein. For a
general treatment of the invariant subspace problem see the monograph by
B. Beauzarny [52]. The invariant subspace problem for positive operators
has been studied extensively by Y. A. Abrarnovich, C. D. Aliprantis, and
0. Burkinshaw [7, 8, 10, 11, 12, 14].
I 0. 1 A Smorgasbord of Invariant Su bspaces 383
10.1. A Smorgasbord of Invariant Subspaces

The purpose of this section is to present several basic theorems about
variant subspaces of operators on general Banach spaces. It should be clear
that if a vector subspace is invariant or hyperinvariant for a bounded oper-
ator, then so is its norm closure. We start with some useful terminology.
If T: X X is an operator on a Banach space and x E X, then
the orbit (more precisely the T of vectors
{ x, Tx, T2x,.. .}. The T-orbit space of x is the Vector subspace generated
by the orbit of x and it will be denoted by OT(x). Clearly, OT(x) {O}
for each x 0. A vector x is said to be cyclic (more precisely T-cyclic)
whenever its orbit space is dense. That is, a vector x is cyclic if OT(x) = X.
It should be obvious that each orbit of T is invariant under T. So, if
OT(x) X for some x 0, then OT(x) is a non-trivial closed T-invariant
subspace. Therefore, if V is a non-trivial closed T-invariant subspace of
then for each non-zero x E V the closure of the orbit space, OT(x), is
included in V and thus it is a non-trivial closed T-invariant subspace. In
other words, we have the following simple result.
o A bounded operator on a Banach space has a non-trivial closed
invariant subspace if and only if it has a non-zero vector that is not
cyclic.
Since each space OT(x) is clearly separable, it follows that every bounded
operator on a non-separable Banach space has a non-trivial closed invariant
subspace.
Definition 10.1. A bounded operator on a Banach space is said to be a
scalar operaLor if it is a multiple of the identity operator. A bounded
operator which is not a scalar operator is called a non-s calar operator.
Every non-scalar operator is automatically and different from
the identity operator. Also, notice that only Banach spaces of dimension
greater than one can admit non-scalar operators.
We now continue with an elementary result describing some special in-
variant subspaces.
Lemma 10.2. If T: X X is a bounded operator on a Banach space,
then its kernel and its range are subspaces.
Proof. Assume that an operator S E JC(X) commutes with the operator T,

i.e., ST = TS. If x e KerT, then from T(Sx) = S(Tx) = 0, we see that
S(Ker T) C Ker T, i.e., Ker T is a closed T-hyperinvariant subspace. Next,
notice that the equation S(Tx) = T(Sx) valid for each x e X shows that
S(T(X)) c T(X), which means that the range of T is T-hyperinvariant.
384 Invariant Subspaces
Corollary 10.3. If a bounded operatorT commutes with a non-zero bounded

operator which fails either to be one-to-one or to have a dense range, then
T has a non-trivial closed invariant subspace.
In particular, each bounded operator which is not one-to-one or does not
have a dense range has a non-trivial closed invariant subspace.
The preceding conclusion shows that as far as the invariant subspace

problem is concerned, not only can we suppose that a given operator T
is and has dense range but also that every non-zero operator
commuting with T is one-to-one and has dense range. That is, Corollary 10.3
can be restated as follows.
Corollary 10.4. If an operator T E (X) not have a non-trivial closed
does
invariant subspace, then every operator that commutes with T (including T
itself) is one-to-one and has a dense range.
Recall that an eigenspace of a bounded operator T: X X is any
subspace of the form NA = {x E X: Tx = Ax}, where A is an eigenvalue
of T. If zero is an eigenvalue of T, then N0 is precisely the kernel of T,
i.e., Ker T = N0. Clearly, the eigenspaces are non-zero closed subspaces and
when T is a non-scalar operator, they are different from X. The eigenspaces
are also hyperinvariant.
Lemma Every eigenspace of any operator T E (X) is T-hyper-
invariant. Moreover, if T is non-scalar, then every eigemspace of T is
Proof. Let A be an eigenvalue of a bounded operator T: X > X, and let

NA = {x E X: Tx = Ax} = Ker (T A).
By our hypothesis Ker (T A) 5L X. Since every operator which commutes
with T also commutes with T A, it follows from Lemma 10.2 that the
closed subspace NA is T-hyperinvariant.
The next two results deal with the existence of invariant subspaces for
operators on finite dimensional vector spaces. As we shall see, the situation
is different between the real and complex cases.
Corollary 10.6. Every non-scalar operator on a finite dimensional complex
Banach space has a non-trivial hyperinvariant subspace.
Proof. Let T: X X be a non-scalar operator on a finite dimensional
complex Banach space, and consequently T has an eigenvalue A e C. Using
that T is a non-scalar operator, we infer that the eigenspace NA is non-trivial.
By Lemma 10.5, this eigenspace NA is T-hyperinvariant.
Corollary 10.7. For a non-scalar operator T: X + X on a finite dimen-

sional real Banach space we have the following.
(1) If dim X is an odd number, then T has a non-trivial hyperinvari ant
subspace.
(2) If dim X is an even number greater than two, then T has non-trivial
invariant subs paces that need not be hyperinvariant.
(3) IfdimX= 2, then T need not have a non-trivial invariant subspace.
Proof. Let T: X X be a non-scalar operator on a finite dimensional real
Banach space.
(1) Since X is a finite dimensional real Banach space of odd dimension,
it follows that T has a real eigenvalue A E JR. But then, as in Corollary 10.6,
the eigenspace NA is a non-trivial T-hyperinvariant subspace. In particular,
dimX> 1.
(2) According to Corollary 10.6 the operator Tc: has an eigen-
value a E C. Pick a non-zero vector z = x + iy such that Tcz = az. If we
consider the vector subspace V of X generated by x and y, then the
sumption dim X> 2 guarantees that V is a non-trivial vector subspace of
X. It can be easily checked that V is T-invariant.
For an example of an operator on JR4 without non-trivial hyperinvariant
subspaces see Exercise 2 at the end of the section.
(3) Let us consider on JR2 a rotation operator through an angle satis-
fying 0 < <-ir. That is, is the operator represented by the orthogonal
.
matrix . . It is easy to check (since the non-trivial subspaces
of JR2 are the straight lines passing through the origin) that Rj, does leave
any non-trivial subspace invariant.
If the adjoint of an operator has an eigenvalue, then the operator itself
has a non-trivial closed invariant subspace.
Theorem 10.8. Let T: X X be a non-scalar operator. If T or its
adjoint T* has an eigenvalue, then both T and T* have non-trivial closed
hyperinvari ant subspaces.
Proof. Notice that T is a non-scalar operator if and only if its adjoint

T* is likewise non-scalar. If T has an eigenvalue A, then by Lemma 10.5 its
eigenspace NA = {x E X: Tx Ax} is a non-trivial closed T-hyperinvariant
subspace.
Now assume that T*x* = Ax* for some scalar A and some non-zero
in Let V = (T A)(X), the closure of the range of the operator T
I 0. Invariant Subspaces
Since T Al, we see that V {O}. iVioreover, since Q and since
((T A)xx*) (x, (T* _ A)x*) (x,O) = 0
for each x X, it follows that V X. That is, V is a non-trivial closed

subspace of X. We claim that the subspace V is T-hyperinvariant. Indeed,
if an operator S (X) satisfies ST = TS, then for each x E X we have
S[(T A)x] = (T A)(Sx) E V, which implies that S(V) V.
To complete the proof, notice that if A is an eigenvalue of T, then A is
also an eigenvalue of T**, and therefore the preceding case implies that T*
has a non-trivial closed hyperinvariant subspace.
Recall that a closed subspace V of a Banach space X is said to be

complemented if there exists another closed subspace TV of X such that
V TV = X. The subspace W is called a complement of V. As usual, an
operator P: X X is said to be a projection if P2 = P. A projection P
is proper if P 0 and P I. The Closed Graph Theorem easily implies
that a closed subspace V of a Banach space X is complemented if and only
if V is the range of a continuous projection on X.
The following simple lemma describes a necessary and sufficient condi-
tion for a complemented subspace to be invariant under a bounded operator.
Lemma 109. The range of a projection P E (X) is invariant under an
operator T E (X) if and only if TP = PTP.
Proof. (X) and Y = P(X), where P E (X) is a projection.
Let T E
Assume first that Y is T-invariant. If x E X, then TPx E Y, and so
TPx = P(TPx). That is, TP = PTP.
For the converse, suppose that TP = PTP and let y E Y. It follows
that Ty = TPy = P(TPy) E Y. This shows that Y is T-invariant.
LetT: X X be a continuous operator on a Banach space. Recall that
a non-trivial closed T-invariant subspace V of X is said to be reducing the
operator T if V has a closed complement TV which is also T-invariant. In
this case the pair (V, TV) is called a pair of reducing subspaces for T. The
existence of reducing pairs was characterized in Theorem 2.22. We restate
this theorem below for convenience.
Theorem 10.10. A operator T: X X on a Banach space has
continuous
a reducing (closed) subspace if and only if there exists a proper continuous
projection P: X X that commutes with T, i.e., TP = PT.
A vector subspace A of (X), where X is a Banach space, is called
an algebra of operators (or a subalgebra of (X)) whenever S, T E A
imply ST E A. An algebra of operators A is said to be unital if the identity
191L A Smorgasbord oflnvariant Subspaces
operator I belongs to A. The dual algebra of an algebra of operators A

in (X) is the subalgebra A4 of (X*) defined by A* {T* : T e A}. An
algebra of operators A is said to be closed if A is closed in (X).
Every subset of operators S is contained in a smallest (with respect to

inclusion) subalgebra, called the algebra generated by S. This algebra
coincides with the intersection of all subalgebras containing S (note that
(X) is one of them). The algebra generated by a single operator T e (X)
will be denoted by AT.A moment's thought reveals that
AT = {p(T): p is a polynomial without constant term }.
The unital algebra generated by a set S is the algebra generated by the
set S U {I}. The unital algebra generated by a single operator T e (X) is
precisely the collection of operators
{p(T): p is an arbitrary polynomial }.
If A a subalgebra of (X), then a subspace V of X is said to be

is
A-invariant whenever V is invariant under every operator in A, i.e., if

A(V) C V for each A A.
Recall that a collection of functions from a topological space to
itself is called transitive if for each pair Wi, W2 e and any neighborhood
V of w2 there exists some f e such that f(wi) E V. Moreover, is said
to be strictly transitive if for any two points w1, w2 E there exists some
f E such that f(wi) = w2. It is easy to see that an algebra of operators
A (X) does not have non-trivial closed invariant subspaces if for any
pair x, y e X with x 0 and every 0 there exists some operator A e A
such that IIAx <. This discussion motivates the following definition.
Definition 10.11. An algebra A of operators is said to be:
(1) transitive if it has no non-trivial closed A-invariant subspaces and
(2) non-transitive if it has a non-trivial closed A-invariant subspace.
It should be clear that a subspace V of X is invariant under an operator

T e (X) if and only if V is invariant under the algebra (unital or not)
generated by T in (X). As usual, for an algebra A and any x e X we let
Ax={Ax: AeA}.
Obviously, Ax is a subspace of X which is also A-invariant. Moreover, it
is easy to see that a unital a'gebra A is non-transitive if and only if there
exists some x e X such that the subspace Ax is non-trivial.
The next result is a folklore characterization of the non-transitive alge-
bras.
-- - - 10. Invariant Subspaces
Theorem 10.12. For an algebra A of bounded operators on a Banach space

X the following statements are equivalent.
(1) The algebra A is non-transitive.
(2) There exist a non-zero vector x e X and a non-zero linear func-
tional E X* such that for each A e A we have
(Ax, x*) = (x, A*x*) = 0
Proof. (1) (2) Assume that W is a non-trivial closed A-invariant sub-

space. Fix a non-zero vector x E W and consider the closed vector subspace
V = Ax. Clear'y, V c W, and so V X. Therefore, there exists some
non-zero x* e that annihilates Le., (Ax, x*) = 0 for all A E A
(2) Suppose that (Ax, x*) = 0 for all A e A and some non-zero
(1)
x e X and e It follows that the closed subspace Ax is not
equal to X. Now, if Ax {0}, then Ax is a non-trivial closed A-invariant
subspace. On the other hand, if Ax = {0} holds, then note that the one-
dimensional vector subspace V = {Ax: A scalar} is non-trivial, closed, and
A-invariant.
Corollary If an algebra of operators A on a Banach space X is
non-transitive, then its dual algebra A* is likewise non-transitive.
Proof. There exist non-zero x E X and x* e satisfying (Ax, x*) = 0 for

all A e A. Viewing x as a vector in we get (x*, (A*)*x) = 0 for each
A E A Now Theorem 10.12 guarantees that the subalgebra A* of is
non-transitive.
The next result describes the hyperinvariant subspaces of an operator in

terms of an algebra of operators. Its straightforward proof is omitted.
Lemma 10.14. If T e (X), then:
(1) Its commutant {T}' {S e L(X): ST = TS} is a closed unital
subalgebra of (X).
(2) T has a non-trivial closed hyperinvariant subspace if and only if
{ T}' is a non-transitive algebra.
The existence of a non-trivial closed invariant subspace for each coin-

pact operator on an (infinite dimensional) Hilbert space was established
by J. von Neumann in 1935. N. Aronszajn and K. T. Smith [42] gener-
alized this to Banach spaces in 1954. It took another nineteen years until
I. Lomonosov [211J improved this by replacing an invariant subspace
with a hyperinvariant subspace. In fact, as we will see in the next section,
V. I. Lomonosov [211] proved more, but at this point we want to present a
10.1. A Smorgasbord of Invariant Su bspaces 389
rather elementary proof of this result. The proof below is a simplification

of Lomonosov's original proof [211] and is due to M. Rilden [241].
Theorem 10.15 (AronszajnSmithLomonosov). Each non-zero compact
operator on an infinite dimensional real or complex Banach space has a
non-trivial closed hyperinvariant subspace.
Proof. Let K: X p X be a non-zero compact operator on an infinite

dimensional real or complex Banach space. We can assume without loss
of generality that 1
1. First, we shall establish the result for the
special case when the operator K is quasinilpotent, i.e., we shall assume
that r(K) = lirn
Put A = {K}'. We must show that the algebra A is non-transitive. For
this it suffices to show that there exists some x 0 such that the subspace
Ax = {Ax: A E A} is not dense in X. So, assume by way of contradiction
that
Ax = X for each x 0. (*)
Since K is non-zero, it easy to see that there exists some xo E X such

that > 1 and > 1. Let U0 denote the closed unit ball centered
at xo, i.e., Uo = {x E X: xoII i}. Since 1, llxoII > I and
> 1, we see that
0 U0 and 0 K(Uo). (**)
In view of (*) we know that for each x 0 there exists an operator A E A

such that Ax E U0, i.e., lAx xojI < I (and clearly A 0). Consequently,
when A runs over A, the open sets OA = {y E X: xoIl < i} cover
X\{0} and, in particular, they cover the compact set K(U0)._So, we can find
a finite number of operators A1,... ,Ak E A such that K(U0) c
Let c = and note that c> 0.
Since Kx0 E K(Uo), there exists some 1 ji k such that Kx0 E
and thus x1 = Kx0 E U0. Then Kx1 K(Uo) and so there exists some
I k such that x2 = E (J0. Proceeding inductively, we can
select a sequence of indices in {I, 2,.. k} such that the sequence
. ,
{ of vectors defined via the recursive formula

n=0,1,2,...
lies in U0. Recalling now that each operator commutes with K, we can
rewrite the previous expression for Xn+1 as follows:
= ..
=
390 --- --- I 0. Invariant Subspaces
The quasinilpotence of K implies clearly the quasinilpotence of cK, and so

I
0. Since 11C'Ajm 1 for each rn = 1, . . , n , we infer from
M
.
the above identity that the sequence Uo satisfies 0, contrary

to the first condition in (**). This contradiction proves that Ax cannot be
dense for all x 0 and so, in this case, K has a non-trivial hyperinvariant
subspace.
It remains to consider the case where the compact operator K is not
quasinilpotent. We consider the complex case first. In this case, we know
that the spectrum of K has some non-zero point A E C. By Theorem 7.3,
we also know that A is an eigenvalue of K and since K equals Al on the
eigensp ace it follows that N,\ is necessarily finite dimensional. Now
Lemma 10.5 guarantees that is a non-trivial closed K-hyperinvariant
subspace of X.
Finally, assume X to be real and consider the operator Kc: Xc Xc.
Clearly, is a non-quasinilpotent compact operator. By the preceding
case, there is an elgenspace N,\ in that is Since
is finite dimensional, there exists a basis {zi, z2,.. . , for Write
zk = xk + for each k and let V be the finite dimensional vector subspace
generated in X by the vectors {x1,. . . ,
, Since X is infinite
dimensional, V is a non-trivial closed vector subspace of X. To complete
the proof, it remains to check that V is also K-hyperinvariant.
Corollary 10.16 (Lomonosov). If a bounded operator T on an infinite di-
mensional real or complex I3anach space commutes with a non-zero compact
operator, then T has a non-trivial closed invariant subspace.
Recall that a bounded operator T: X X is said to be polynomially
compact whenever there exists a non-zero polynomial p such that p(T)
is a compact operator. A. R. Bernstein and A. Robinson [60] proved using
non-standard analysis that polynomially compact operators have non-trivial
closed invariant subspaces; a "standard" proof of this result was provided
later by P. R. ilalmos [146].
Corollary 1017 (BernsteinRobinson [60]). Each polynomially compact
operator on an infinite dimensional real or complex Banach space has a
non-trivial closed invariant subspace.
Proof. Let T: X X be a bounded operator on an infinite dimensional
Banach space. If T is a scalar operator, then T has non-trivial closed in-
variant So, suppose that T is a non-scalar polynomially compact
operator. Pick a non-zero polynomial p(t) = + +... + a1t + ao
such that p(T) is a compact operator.
If p(T) 0, then (according to the AronszajnSmith--Lomonosov The-
orem 10.15) the non-zero compact operator p(T) has a non-trivial closed
p(T)-hyperinvariant subspace, say V. Since T commutes with p(T), it fol-

lows that V is also T-invariant.
Now assume that p(T) = 0. In this case, we have T72 =
Fix any non-zero vector x e X and let V denote the non-zero closed subspace
generated by the set {x, Tx,. . , Since X is infinite dimensional,
we have V X, and a simple verification shows that V is T-invariant.
Exercises
1. (Schaefer [287]) Show that every linear operator on an infinite dimen-
sional vector space has a non-trivial invariant subspace. [HINT: Let
T: X X be a linear operator on an infinite dimensional vector space.
Fix some x 0 and consider the T-orbit space (9T of x. If (9T
then (9T is a non-trivial T-invariant subspace. So, assume that = X.
In this case, it follows that the set of vectors {x, Tx, T2x,. . is linearly
independent. This means that every vector y E X has a unique repre-
sentation of the form y = witl1 the )'tj equa' to zero for all
but finitely many i. Now define the non-zero linear functional on X by
q5
It should be easy to verify that Ker q5 is a non-trivial

T-invariant subspace.]
2. Give an example of a non-zero operator on 1R4 without non-trivial hyper-
invariant subspaces. [HINT: Let 0 be a non-zero angle that is an irra-
tional multiple of 2ir and consider the operator T: 1R4 defined by
T(x17x2,x37x4) = (R0(x1,x2),R0(x3.x4)), where R0: 1R2 denotes

the rotation operator through the angle 0. Show that the invariant sub-
spaces of the operator T are
Ve={(xl,x2,0,0): xi,x2ER} and Vr={(0,0,xl,x2): xi,x2ER}
and the ones given by
= x17x2 E
where E [0, 2ir) and > 0.
If we consider the operator S: 1R4 p 1R4 defined by
S(x1,x27x3,x4) =
then S commutes with T and leaves neither nor Vr invariant. On
the other hand, if a is a non-zero angle such that + is an irrational
multiple of 2ir, then the operator R: 1R4 defined by
R(xi,x2,x37x4) =
commutes with T and is not R-invariant for all 'y > 0.1
3. Let T: X X be a bounded operator on a Banach space. Denote
by Lat(T) the set of all closed T-invariant subspaces of X. Show that,
under the standard set inclusion. Lat(T) is an order complete lattice with
smallest and largest elements.
392 10. Invariant Subspaces
4. For a bounded operator T: X X on an infinite dimensional real Banach

space establish the following.
(a) The operator T has a non-zero finite dimensional invariant subspace
if and only if the operator likewise has a non-zero
finite dimensional invariant subspace.
(b) The operator T has a non-zero finite dimensional hyperinvariant
subspace if and only if the operator likewise has a
non-zero finite dimensional hyperinvariant subspace.
5. Let T: X X be a bounded operator on a Banach space. If X is not
reflexive, then show that the double adjoint T**: has a non-
trivial norm closed subspace.
6. Let be a compact Hausdorif space. Fix some q5 E and consider

the multiplication operator on If is a subset of
and V ={f f(w) =0 for aliw E then show that Visa
closed Mcp-invariant subspace of
7. Let be a compact Hausdorif space. Fix a continuous mapping -r:

and consider the composition operator on defined, as usual, by
C-rf = for. If is a nonempty r-invariant subset of (i.e.,
then show that the vector subspace
V={fE C(cl): f(w)=0 for all wE
is a closed C7-invariant subspace of

9. If X is a Banach space, then show that the only I-hyperinvariant sub-
spaces are {0} and X.
10. Let A be an arbitrary algebra of operators on a vector space. Show that

A1 = {cJ + A: is a and A E A} is the unital algebra generated
by A. (This implies that any algebra of operators A and the unital algebra
generated by A have the same invariant subspaces.)
11. Show that an algebra of operators A (X) has no non-trivial closed
invariant subspaces if and only if for any pair x, y E X with x 0 and
every E> 0 there exists some operator A E A such that
A be a unital algebra of continuous operators on a Banach space X
and, as usual, let Ax = {Ax: A E A}. Show that A is non-transitive if
and only if A has a non-trivial closed vector subspace of the form Ax.
13. Let T: X X be a bounded operator on a Banach space and consider
the set
V= {xEX:
(a) V is a T-hyperinvariant subspace.
(b) If T is power bounded, then V is also closed.
Lomonosov Invariant Subspace Theorem 393
14. Let 8, T : X X be a pair of bounded operators on a Banach space. As

usual, the direct sum of S and T is the operator SeT: XeX XEfX
defined by (S EE T)(x y) Sx eTy. If a bounded operator R: X X
satisfies RS = TR, then show that its graph GR = {x Rx: x E X} is
a closed S e T-invariant subspace of X e X.
15. Show that the shift operator on any er-space has a non-trivial closed
hyperinvariant
16. Show that the converse of Corollary 10J3 is not true. That is, show
that the dual algebra A can be
non-transitive.
17.. Let q5: 1Rbe a continuous function separating the points of a compact
Hausdorff space and let denote the usual multiplication operator
on defined by M6(f) = bf. Show that {M6}' = {M9: 9 E
[HINT: Let T E and put h = Since TMj, = implies
= fE it follows that = and by
induction we see that for each m 0. The latter implies
= for each polynomial p. By the StoneWeierstrass Ap-
proximation Theorem the algebra p a polynomial} is norm dense
in and from this we conclude that Tf = hf for each f E
10.2. The Lomonosov Invariant Subspace Theorem

Many invariant subspace theorems are for operators that are
related in one way or another with the compact operators. It is generally
acclaimed that the most famous of these results is Lomonosov's Invariant
Subspace Theorem [211j. It asserts that:
o If an operator T: X p X on a complex Banach space commutes
with a non-scalar operator S e (X) which in turn commutes with
a compact operator, then T has a non-trivial closed in-
variant subspace.
The objective of this section is to prove this theorem. To do so, we need the
following result which is of importance in its own right.
Theorem 10.18 (Lomonosov [211]). If A is a transitive algebra of oper-
ators on a real or complex Banach space, then for each non-zero compact
operator K there exists some A e A such that the compact operator AK has
a non-zero fixed point, ie., AK'u = u for some u 0.
Proof. Let A (X) be a transitive algebra and let K be a non-zero

compact operator. We can assume = 1.
Pick some xo e X such that lIKxojl > 1 and let U0 denote the closed
unit ball centered at x0. That is, Uo = {x e X: Ilx xojI 1}. An easy
argument shows that

and (*)
Since A is an algebra, (the norm closure of Ax) is A-invariant for each

x E X. We claim that {O} for each x 0. Indeed, if = {0} for
some x 0, then the iion-trivial closed subspace a' scalar} is clearly
A-invariant, contrary to the hypothesis. Moreover, the absence of non-trivial
closed A-invariant subspaces guarantees that Ax = X for each x 0.
In particular, for each x E K(Uo) there exists some A E A such that
lAx xoII < 1. So, the family of open sets {y E X: < 1}AEA is
an open cover of K(Uo), that is,
K(Uo) U {y E X: lAy - xoII <1}.
AEA
Since (in view of the compactness of K) the set K(Uo) is compact, there
exist operators A1,... , A such that
IIAiy-xo(1<1}.
Next, for each i define the continuous function X [0, oo) by

= max{0, 1 xoIl}.
Clearly, > 0 if and only if xolI < 1. From (**), we see that

f(z) = > 0 for each z E and so the formulas gj(z)

=
define non-negative continuous functions on K(Uo) such that gj(z) = 1
for each z E K(Uo). In particular, = 1 for each x E Uo.
Now consider the function Uo X defined by
Since x E Uo and (1(x) > 0 imply E Uo, it follows from the convexity
of U0 that Uo. In addition, because is a convex combination
of the vectors A1Kx,.. and E we see that
belongs to C = But is a norm totally
bounded set (since each is a compact operator), and so by Mazur's

theorem its closed convex hull C is likewise a compact set; see, for exam-
ple [30, Theorem 9.4, p. 1311.2 Hence, C fl U0 and C fl U0 is a
non-empty convex compact subset of X.
2jf X is a complex Banach space then when taking the closed convex hull of C, we consider
X as a real vector space.
By Tychonoff's Fixed Point Theorem, the map CnU0 CnU0 has a

fixed point, say u. Since u E Uo, it follows from (*) that u 0. To complete
the proof, consider the operator A = E A, and note that
AKu = = = u.
And now we are ready to state and prove the two famous Lomonosov
invariant subspace theorems.
Theorem 10.19 (Lomonosov [211]). If a non-scalar operator T on a com-
plex Banach space commutes with a non-zero compact operator, then T has
a non-trivial closed hyperinvariant subspace.
Proof. Assume that T E (X) a non-scalar operator and that for some
is
non-zero compact operator K E (X) we have TK = KT. To establish the
existence of a non-trivial closed T-hyperinvariant subspace, according to
Lemma 10.14, it suffices to show that the corninutant of T is non4ransitive.
To verify this, assume by way of contradiction that the commutant {T}"
is a transitive algebra. So, by Theorem 10.18, there exists some A E {T}"
such that the operator AK has a non-zero fixed point. Let
F={xEX: AKx=x}.
Clearly, F is a non-zero closed AK-hyperinvariant subspace. Since AK is
a compact operator and it is the identity on F, it follows that F is finite
dimensional. Since T commutes with AK and F is AK-hyperinvariant, we
see that F is So, from T(F) F and the fact that F is a finite
dimensional complex Banach space, it follows that the operator T: F F
has an eigenvalue C.3 Now let
NA={XEX:
and note that NA is a non-zero closed {T}"-invariant subspace which is dif-
ferent from X since T is a non-scalar operator. However, this contradicts
our assumption and the proof of the theorem is complete.
The assumption that the Banach space X is complex is essential. As

we know from Corollary 10.7, on a finite dimensional real Banach space of
even dimension there exist operators with non-trivial invariant subspaces but
without non-trivial hyperinvariant subspaces. It should be noted that for
complex Banach spaces Theorem 10.19 implies immediately Theorem 10.15
and Corollary 10.16.
This is the place where we use that X is a complex Banach space. When X is a real
Banach space, there is no guarantee that the operator T: F + F has a real eigenvalue.
396 -- 10. Invariant Subspaces
Theorem 10.20 (The Lomonosov Invariant Subspace Theorem). Assume

that for a continuous operator T: X + X on a complex Banach space there
exists an operator S E L(X) satisfying the following three properties:
(1) S is a non-scalar operator.
(2) S commutes with T.

(3) S commutes with a non-zero compact ojerator.
Then T has a non-trivial closed invariant subspace.

Proof. Assume that the operators T, S E satisfy the conditions of
the theorem. Then S satisfies the hypotheses of Theorem 10.19. Therefore,
S has a non-trivial closed hyperiuvariant subspace. Since T commutes with
5, this subspace is automatically invariant under T.
Surprisingly, Theorem 10.20 fails to be true if the "chain" connecting
the given operator with a compact operator contains more than one inter-
mediate link. V. G. Troitsky [321] exhibited an example of three operators
T, 51, 52 E such that:
(a) T does not have a non-trivial closed invariant subspace.
(b) Neither 52 nor S3 is a scalar operator.
(c) T commutes with 51 and Si commutes with 52, i.e., 51T = T51
and 5152 = 5251.
(d) S2 commutes with a non-zero compact operator.
We shall say that an operator T E is a Lomonosov operator if
it satisfies the assumptions of Theorem 10.20. That is, T is a Lomonosov
operator if there exists an operator S E L(X) such that:
(1) S is a non-scalar operator.
(2) 5 commutes with T.
(3) 5 commutes with a compact operator.
In this terminology Theorem 10.20 can be stated as follows.
Every Lomonosov operator on a complex Banach space has a

trivial closed invariant subspace.
It should be immediate from Theorem 10.20 that operators without non-

trivialclosed invariant subspaces fail to be Lomonosov operators. After
the appearance of the Lomonosov Invariant Subspace Theorem it was not
clear whether the Lomonosov operators were the only operators on Banach
spaces with non-trivial closed invariant subspaces. It turns out that not
every operator with a closed invariant subspace is a Lombnosov
operator. Subtle examples of non-Lornonosov operators on complex Bariach
spaces with non-trivial closed invariant subspaces can be found in [143]

and {144].
Exercises
1. Give an example of a polynomially compact operator which is not com-
pact.
2. If is a compact Hausdorif space without isolated points and a function
E C(Q) is one-to-one, then show that the multiplication operator
does not commute with any non-zero weakly compact operator.
3. Show that every multiplication operator on a is a Lomonosov
operator.
4. Assume that (cl, >, is a cr-finite measure space such that there exist
two disjoint sets A, B E with 0 <cc and 0 <cc. Show
that every multiplication operator on an with 1 p oo is
a Lomonosov operator.
5. Show that every Markov operator T on an AM-space with unit commutes
with a rank-one positive operatorand hence T is a Lomonosov operator.
6. Consider the operator T: C[1, 1] + C[1, 1] defined by Tx(t) = x(t)
for each x E C[1, 1] and all t E [1,1].
(a) Show that T is a non-scalar surjective lattice isometry which is also
a Markov operatorand so, by Exercise 5 above, T commutes with
a rank-one positive operator and is a Lomonosov operator.
(b) Exhibit a rank-one operator on C[1, 1] that commutes with T.
(c) Show that the compact operator 8: C[ 1, 1] + C[ 1, 1], defined by
Sx(t) = I x(s) sin(st) ds,

J1
commutes with the operator T.

(d) Show that 1 and 1 are the only eigenvalues of T.
(e) Exhibit non-trivial closed invariant subspaces different from the
eigenspaces N1 and N_1.
7. This exercise is a combination of results obtained by W. B. Arveson
and J. Feldman [43], C. Pearcy and N. Salinas [261], and P. Meyer-
Nieberg [239]. Let X be a complex Banach space and for an operator
T E (X) let denote the norm closure in (X) of the collection of
all operators q(T), where is a rational function (i.e., a quotient of two
polynomials with complex coefficients) that is analytic in a neighborhood
of the spectrum cr(T). Clearly, is a commutative unital algebra of
operators.
Show that if an operator T E (X) is non-scalar and contains
a non-zero compact operator, then is non-transitive.
10.3. Invariant Ideals for Positive Operators

The invariant subspace problem when specialized to positive operators on
Banach lattices acquires some special and desirable features. These special
features will be the subject of our discussion in this section. As we shall
see, not only do some broad classes of positive operators have non-trivial
closed invariant subspaces but these invariant subspaces also have a very
nIce geometrical structurethey are ideals. As a matter of fact, a first
result of this type has already been proved in Theorem 9.19 asserting the
existence of a closed ideal for every compact
positive operator which is quasinilpotent at a non-zero positive vector.
We begin by formulating from the outset the main (and still open) con-
jecture regarding the invariant subspace problem for positive operators.
Conjecture: Every positive operator on a Banach lattice of dimen-

sion greater than two has a non-trivial closed invariant subspace.
The results in this section (as well as the results in the subsequent sections)
will provide a considerable amount of evidence to expect the affirmative an-
swer to this conjecture. For a while it was thought that the modulus of some
of the quasinilpotent operators on (the real) 4 without non-trivial closed
invariant subspaces constructed by Read [275, 276, 277] might provide
a counterexample to this conjecture. However, V. G. Tftoitsky [319] has
shown (see Corollary below) that the modulus of any quasinilpotent
Read operator on has an eigenvalue, and so its modulus necessarily has
a non-trivial closed invariant subspace.
Definition 10.21. If A, B: X X are two bounded operators on a Banach

space, then their commutator is the operator [A, B] defined by
[A,B] =AB-BA.
Clearly, two operators A, B E (X) commute if and only if [A, B] = 0.

The commutant of an operator T E (X) can be described in terms of
commutators as follows: {T}' = {s E (X): [T,S] = o}.
If the operators act on a Banach lattice E, then the existence of the
order structure allows us to introduce some new and important collections
of operators related to the commutator. Namely, we can relax the condition
[A, B] = 0 by considering operators whose commutators are positive or neg-
ative. To do this. we will define the super right- and super left-commutants
of a positive operator.
Definition 10.22. Let B: E E be a positive operator on a Banach lattice.

The super right-commutant [B) is the collection of all positive operators
A: E + E such that [A, B] 0. That is,
[B)={AEL(E)+: AB-BA0}.
Similarly, the super left-commutant (B] is the collection of all positive
operators A: E p E such that [A,B] 0. That is,
(B]={AEL(E)+: AB-BA<0}.
Notice that A E [B) if and only if B E (A], and also (B] fl [B) =
the positive part of the commutant of B.
Consider a positive operator B: E p E on a Banach lattice. As usual,
the null ideal of B is defined by NB = {x E E: = 0 }. Clearly, NB
is a closed ideal in E. The range ideal of B is the ideal generated by the
range of B and it will be denoted by RB. A moment's thought reveals that
RB={yEE: suchthat yt<Bx}.

We are now ready to state an analogue of Lemma 10.2 for positive op-
erators. Its proof is straightforward and is omitted. (See also Exercise 4 at
the end of the section.)
Lemma 10.23. If B: E p E is a positive operator on a Banach lattice and
A E [B), then.'
(1) The null ideal NB is
(2) The range ideal RA is B-invariant.
In particular, the ideals NB and RB are both -hyperinvariant.
The next basic result describes a new large class of operators having
non-trivial closed ideals.
Theorem 10.24 (AbramovichAliprantisBurkinshaw [11]). Let B: E p B
be a positive operator on a Banach lattice and assume that there exists a
positive opera,tor 5: E p E such that:
(1) S E (B], that is, SB BS.
(2) 5 is quasinilpotent at some xO > 0, i.e., fl-400
Em = 0.
(3) 5 dominates a non-zero compact operator.
Then the operator B has a non-trivial closed invariant ideal.
Proof. The proof is similar to that of Theorem 10.15. Let B, 5, and x0
the properties stated in the theorem, and let K be a non-zero compact
satisfy
operator dominated by 5, i.e., Kxl for each x E E. Obviously, we
ilL Invariant Subspaces
can assume that IIBJI < 1. Therefore, the series A = converges

and defines a positive operator on E. Let us show next that the hypothesis
SB BS implies that the inequalities SBk BkS and SAk AkS hold
for each k. The first inequality follows by induction. Indeed, if SBk BkS
is true for some k, then = SBkB BkSB BkBS = Next
note that
00 00 00 00
This implies at once that SAk AkS for all k 0.

For each x > 0 we shall denote by the principal ideal generated by
Ax. That is,
Since x the ideal non-zero. We claim that

is is B-invariant.
Indeed, if y then we have yJ for some > 0, and so
jByJ BJyJ
This implies By So, is a closed B-invariant ideal.

The proof will be finished if we can show that E for some z> 0.
To establish this, assume by of contradiction that
= E for each x> 0.
Next, we claim that without loss of generality, we can suppose Kx0 0.
To see this, consider the ideal If Ky = 0 for each 0 < y
then we have K = 0 on and consequently by (*), we get K = 0, a
contradiction. Therefore, Kyo 0 for some 0 < yo Jxo. Since the
operator S is quasinilpotent at xo and yo for some)\ > 0, it follows
that S is also quasinilpotent at yo. Indeed, 572yo
implies ) 0. Hence, replacing xO by yo if
necessary, we can assume that Kx0 0.
By scaling, we can suppose that = 1. Also we can suppose (replac-
ing xo by axo for an appropriate scalar a> 1) that IIxoM > 1 and iIKxoJI > 1.
Now let Uo = {z E: Ixo zil 1}, the closed unit ball centered at xo > 0.
By our choice of xo, we have
0 Uo and 0 K(Uo). (**)
From (*), we know that = E for each x 0. In other words,
is a quasi-interior in E and consequently for each y 0 the
sequence {y A mA(JxI)} is norm convergent and y A riA(Ix[) =
see Lemma 4.15. In particular, for each x 0 there is some m such that
lixo - xO A <1.
Since the function z Fk xO A riA(lzI) is continuous, we see that the set
{z E E: <i} is open for each ri. In view of 0
the above arguments guarantee that
k(Uo)cU{zEE: <1}.
Since the sets {z E E: IxoxoAriA(lzt)1l < i} are increasing as ri increases,
the compactness implies that
K(Uo) {z E E: lIxo x0 A rnA(lzl)11 <i}
for some m. In other words, there exists some fixed rn such that x E K(Uo)
implies that xO A rnA(lxI) E Uo.
In particular, we have x1 = E Uo. So, Kx1 E K(Uo) and
this implies x2 = xO A E Proceeding inductively, we obtain a
sequence of positive vectors in U0 defined by xlHi = xO A
Now we claim that the inequality
0
holds for each ri. The proof is by induction. For ri = the desired inequality
1
= xo A mA mnA(Sxo) is trivially true. For the induction stej

assume that 0 < < is true for some ri. From this and the
inequality we obtain:
0< Xrj+1 = xO A
<
<
as desired.
Therefore, and so . for
each m. Since lim * = 0, it follows that lim = 0, and hence
fl-900
lim = 0. However, Uo implies that 0 E U0 = Uo, contrary to
riboo
(**), and the proof of the theorem is finished.
Recall that an operator T: E F between two Banach lattices is said to

be AM-compact, provided that T maps order bounded sets to norm totally
bounded sets, i.e., if T[a, b} is a norm totally bounded set for each order
interval [a, b}. It should be clear that each compact operator is necessarily
but the converse is not true; see Exercise 7 at the end of the
section. Theorem 10.24 can be easily generalized to AJ1i1-compact operators.
Moreover, condition (2) can also be relaxed. The proof of the next result is
very similar to that of Theorem 10.24 and is omitted.
Theorem 10.25. Let B: E E be a positive operator om a Bamach lattice
amd assume that there exists a positive operator 5: E + E such that:
(1)
1
(2) There exists some xO > 0 satisfyimg liminf = 0.
11-4(X)
(3) S domimates a mom-zero AM-compact operator.
Them the operator B has a mom-trivial closed imvariamt ideal.
The assumption in Theorems 10.24 and 10.25 that S E (B] is consider-

ably weaker than the assumption of commutativity of operators B and 5;
see Exercise 3 at the end of the section. It is still an open question if an
analogue of either of these two theorems is true for S e [B), that is, if we
replace the inequality SB BS by the reverse inequality SB BS. If
we assume that S is quasinilpotent (rather than just being locally quasinil-
potent) then, as our next theorem demonstrates, such a version is possible.
The proof is quite similar to that of Theorem 10.24 and is left as an exercise.
Theorem 10.26. Let B: E E be a positive operator om a Bamach lattice
amd assume that there exists a positive operator 5: E + P2 such that:
(1) 5 e [B), that is, SB BS.
(2) 5 is quasimilpotemt.
(3) 5 domimates a mom-zero AM-compact operator.
Them the operator B has a mom-trivial closed imvariamt ideal.
The next result is a strong companion of Theorems 10.2410.26. It
shows that we can distribute properties (2) and (3) between the commuting
operators.
Theorem 10.27. If B, 5: E + E are two commutimg mom-zero positive
operators om a Bamach lattice such that:
(a) ome of them is quasimilpotemt at a mom-zero positive vector, amd
(b) the other domimates a mom-zero AM-compact operator,
them B amd S have a commom mom-trivial closed imvariamt ideal.
Proof. Assume that S is locally quasinilpotent at some vector xO > 0 and

that B dominates a non-zero AM-compact operator K. If S is not strictly
positive, then the null ideal is a desired common non-trivial closed
variant ideal. So, suppose that S is strictly positive. Without loss of gener-
ality, we can also assume that JB + SJj <1. Put A = + 5)Th and
Invariant Ideals for Positive Operators 403
let J denote the ideal generated by Ax0, i.e.,
J={yEE: A>Osuchthatlyl<AAx0}.
Clearly, J is a (B + 8)-invariant ideal. Since 0 B, S S + B, it
follows that this B and S.
ideal is invariant under both
If J E, then J is a non-trivial closed ideal invariant under both B

and S. So, it remains to consider the case when J = P2. In this case, since
K 0, there exists some 0 < yo E J such that Kyo 0. A straightforward
verification shows that S remains quasinilpotent at
Since Kyol Byo E J, it follows from Lemma 4.16 that there exists an
operator M: E f E such that
MKy0 > 0 and MxJ lxi for each x E P2.
Now consider the operator SiVIK. Clearly, this operator is AM-compact

because K is AM-compact. Since B is strictly positive and IVIKyo > 0, it
follows that SMKy0 > 0 and so SMK is non-zero. Also, we have
ISMKxI S(IKxl)
SB dominates the non-zero AIt'I-compact opera-
tor SMK, is quasinilpotent at yo, and commutes with the positive operator
S+ B. By Theorem 10.25, there exists a non-trivial closed (S + B)-invariant
ideaL Clearly, this ideal is invariant under both B and S.
Corollary 10.28. Assume that a positive operator B: E P2 on a Banach
lattice satisfies the following two conditions:
(1) B is locally quasinilpotent at a non-zero positive vector.
(2) Borne power of B dominates a non-zero AIVI-compact operator.
Then B has a non-trivial closed invariant ideal.
Proof. It suffices to apply Theorem 10.27 to the operators B and S = Btm,

where dominates a non-zero AM-compact operator.
For operators on discrete spaces Theorem 10.24 can be improved.

Theorem 10.29. Let B: P2 f E operator on a discrete Banach
be a positive
lattice and assume that there exists a non-zero operator S E (B] such that
S is locally quasinilpo tent at a non-zero positive vector. Then B has a nom-
trivial closed invariant ideal.
Proof. As E is discrete, there exists a set F such that E is an order dense

Riesz subspace of IRt' (Ct' in the complex case) and the atoms =
belong to E for each 'y E F.
10. Invariant Subspaces
If S is not strictly positive. then consider the null ideal N8. It is a non-
trivial closed S-invariant ideal in E. Since S e (B], we know that B E [5)
and so Ng is also B-invariant: see Lemma 10.23.
Consequently, we can assume that S is strictly positive. Therefore, for
any -y e F the operator K is non-zero, where R,. denotes the stan-
dard rank-one projection onto the one-dimensional subspace spanned by
Clearly, K is compact and 0 <K < 5. Now apply Theorem 10.24.
Theorem 10.29 will be improved even further in Corollary 10.48.

Corollary 10.30. Let B: 4 (1 <p < be a positive operator. If
oo)
there exists a non-zero positive operator 5: 4 * 4, such that:
(1) 5 commutes with B, and
(2) 5 is locally quasinilpo tent at a non-zero positive vector,
then B has a non-trivial closed invariant ideal.
Corollary 10.31. If a positive operator B: 4, 4, (1 p < oc) is quasi-
nilpo tent at some positive vector, then B has a non-trivial closed invariant
ideal.
We are now ready to present an elegant result of V. C. Troitsky [319]

regarding the modulus of regular operators on 4,-spaces.
Theorem 10.32 (Troitsky). If B: 4 (1 p<
is an ideal irre-
oc)
ducible regular quasinilpotent operator such that B is compact, then: -
(1) The modulus IBI of B is not quasinilpotent.
(2) The spectral radius of is an eigenvalueand so IBI has a non-
(3) The positive part B is not quasinilpotent and the spectral
radius of B+ is an eigenvalueconsequently B+ has a non-trivial
closed invariant subspace.
Proof. We begin by observing that the operator IBI cannot be quasinilpo-

tent. Indeed, if it were, then by Corollary 10.31 the operator IBI would have
a non-trivial closed invariant ideal. This ideal would also be B-invariant con-
trary to our assumptions. Thus, r(IBI) > 0. In view of Theorem 7.9 this
guarantees that r(IBI) belongs to the spectrum of the operator IBI. It is
also clear that IBI cannot be a scalar operator, since otherwise the operator
B would again have a non-trivial closed invariant ideal.
'We will show next that ress( (B) = 0, that is, lB is an essentially quasi-
nilpotent operator. To prove this, notice that IBI = B + 2BTh
and so for each n we have
B is compact, whence
HBIThM ess = lBtm + <
IIBThII.
Consequently,
ress(JBI) = urn ( lim = r(B) =0.

This implies that IB! is an essentially quasinilpotent operator and, as shown
above, 0 <r(JBI) E
Now a glance at Corollary 7.49 guarantees that r(IBI) is an eigenvalue
of the operator IBIc in This implies that r((Bf) is also an eigenvalue of
I
in E, and so the corresponding eigenspace in E is a desired non-trivial
closed invariant subspace of B1.
Finally, consider the operator If it is quasinilpotent, then
ollary 10.31 guarantees that has a non-trivial closed invariant ideal.
So, it remains to consider the case when B+ is not quasinilpotent. Since
= B + B, the same argument as in the first part of the proof shows
that ress(B+) = 0. So we have 0 = ress(B+) < and this again implies
that B+ has a non-trivial closed invariant subspace. D
It is worth mentioning that, in fact, there exists a positive eigenvector

corresponding to the eigenvalue r(IBI). This follows from the following useful
version of the KreinRutman Theorem 7.10 established by P. P. Zabreiko
and S. V. Smitskikh [345] and independently by R. Nussbaum [259]:
o IfS is a positive operator on a Banach lattice and ress(S) <r(S),
then r(S) is an eigenvalue of S having a positive eigenvector.
immediately that the modulus of each quasinil-

Theorem 10.32 implies
potent Read operator on Li has a non-trivial closed invariant subspace.
Corollary 10.33 (Troitsky [319]). If T: is a quasinilpotent opera-
tor without non-trivial closed invariant subspaces, then:
(1) Neither ITt nor has a non-trivial closed invariant ideal.
(2) Each of the operators and T+ has an eigenvectorand hence
each of these operators has a non-trivial closed invariant subspace.
Proof. According to Theorem 3.9 the operator T is regular. The validity

of the first statement is now immediate from the hypothesis that T does
not have any non-trivial closed invariant subspace. The second statement
_____________________
406 10. Invariant Subspace,s
We shall close the section with two interesting cases where the existence
of an invariant subspace of a positive operator can be obtained without any
compactness assumptions.
Theorem 10.34. If is a compact Hausdorff space, then every non-s calar
positive operator on has a non-trivial closed hyperinvariant subspace.
In particular, each Krein operator has a non-trivial closed hyperinvari ant
subspace.
Proof. Let T: be a non-scalar positive operator. By Corol-

lary 9.46 the adjoint operator T* has an eigenvalue, and the first conclusion
The second conclusion follows from the fact that a Krein operator on a
cannot be a scalar operator. w
Recall that if fl is a topological space, then C0(fl) denotes the Banach
lattice of all continuous functions f: fl JR such that for each 0 there
exists a compact subset K of Q (depending on with < for all
w E fl \ K. Clearly, Co(Q) is a closed ideal in Cb(fl), the Banach lattice
of all bounded continuous real-valued functions on (as usual and
are equipped with the sup norm). If is a compact space, then
= If is a locally compact Hausdorif space and is the
one-point compactification of 11, then = {x E x(oo) = 0 }.
When is a locally compact Hausdorif space, an extension of the pre-
vious theorem to positive quasinilpotent operators on was obtained
by H. Schaefer [288]. His original proof was rather involved and used the
resolvent operator and the Spectral Mapping Theorem. The proof below is
more direct and requires only local quasinilpotence.
Theorem 10.35 (Schaefer). Let B: be a positive operator,
where is a locally compact Hausdorff space. If B is locally quasinilpo tent
at some non-zero positive vector, then B has a non-trivial closed invariant
ideal.
Proof. Let B be locally quasinilpotent at some xO > 0. If Bx0 0, then

there is nothing to prove, since the null ideal NB = {x BIxI = 0}
is a non-trivial closed invariant subspace.
So, assume that Bx0 > 0. Without loss of generality we can assume
that the support set F = {t e fl: xo(t) > 0} of the function xO is compact.
To see this, fix a compact subset K of on which xO does not vanish. Since
is a locally compact Hausdorif space, there exists a continuous function
y: [0, 1] with compact support such that y(w) = 1 for each w E K. If
we let yo = yxo, then 0 < yo XO, the function has compact support,
and the operator B is quasinilpotent at
We claim next that there exists a point to E such that

(t0) = 0 (*)
for all n E N. To see this, assume to the contrary that (*) does not hold.
Then for each t E there exists some E N such that (t) 0. In
particular, there exist some Ct > 0 and an open neighborhood Gt of t such
that
[Bfltxo] (s) ct > 0 for each s E
Pick a finite cover . ,G}
of the compact set F, and let .
be the corresponding powers of B for these points, i.e., flj = Now con-
sider the operator
S= + + Bn'k, (**)
and note that for each t E F we have
= k
[Sxo] (t) (t) min{ct17. . .
> 0.
Therefore, the function Sx0 is strictly positive on the compact set F. In

particular, there exists some > 0 such that 7Xo. The latter implies
easily that S is not quasinilpotent at see Exercise 1 at the end of the
section. On the other hand, (**) guarantees that S is quasinilpotent at
Indeed, from (**) we see that
=BR==RB,
where R = B and R commute, we have
Snixo = = and so
.
for eacir n E This- 0, which is

impossible. Consequently, there exists a point to E that satisfies (*).
Finally, consider the following ideal in
J = {x E A E R+ and m2, . . . , EN
such that lxi }.
Obviously, J {0} as 0 Bs0 E J. It is also obvious that J is B-invariant.
It remains to notice that J cannot be dense in since x(to) = 0 for each
x E J. Therefore, the norm closure J of J is a non-trivial closed
ideal, and the proof is finished.
Most of the results in this section were obtained by Y. A. Abramovich,

C. D. Aliprantis and 0. Burkinshaw [10, 11, 16]. For quasinilpotent positive
operators, Corollary 10.31 was also proven in [91].
Exercises
1. Let S: E E be a positive operator on an arbitrary Banacli lattice such
that Sxo holds for some vector x0 > 0 and some scalar 'y > 0.
Show that S is not locally quasinilpotent at
2. Let J be an ideal in a Banacli lattice E. If J is invariant under some
positive operator B: E + E, then J is also invariant under every operator
dominated by B.
3. Give an example of a pair of non-commuting positive operators A and B
such that A E [B), that is, AB BA> 0.
4. For a positive operator B: E + E on a Banach lattice establish the
following.
(a) The range ideal RB of B (i.e., tile ideal generated by the range of
the operator B) is given by
RB = {y E E: 2 x E such that Bx }.
(b) The range ideal RB is -hyperinvariant.
(c) For each A E [B) the range ideal RA is B-invariant.
(d) The null ideal NB is A-invariant for each A E [B). In particular,
NB is L?hyperinvariant.
(e) Present an exaniple of an operator A E (B] for whiicli tue ideal NB
fails to be A-invariant.
(f) If B is and [B) contains an ideal irreducible operator, then
B is strictly positive.
5. Assume that K: E + E is a compact operator on a Banach lattice suchi
that Kx0 0 for some x0 E E. Show that there exists a compact operator
C such that Cx0 > 0 and for each x E E. [HINT: Consider
the range ideal RK of K. Using the compactness of K, establish that the
norm ciosure of RK iias a quasi-interior point. Tiien apply Lemma 4.16.]
6. For a positive operator T: E + E on a Banacii lattice let
QtT = {A E (E): B such that AxJ BtxI for all x E
denotes as usual tue set of ail positive operators in the com-

mutant of T. Show that:
(a) The center of E is included in QLT, i.e., 2(E) c QtT.
(b) QtT is a unital subalgebra of (E).
(c) An ideal J is L?-hyperinvariant for T if and only if J is QtT-invariaut.
7. Give an example of an AM-compact operator which is not a compact
operator.
8. Prove Theorem 10.25. [HINT: Tue proof is similar to that of Thieo-
reni 10.24. In this case we cannot ciaim tiiat the set K(Uo) is compact.
However, it follows that the set K(Uo fl [0, Xe]) is compact. The inclusion
K(U0 fl [0,xo]) E E: lixo <1}

1 0. Invariant Subspaces of Families of Positive Operators 409
remains true, and so, in view of the compactness of K(U0 fl [0, xo]), there
exists some m such that
K(Uon {z E: X0AmA(IzI)II < 1}.
Now the sequence defined recursively by = and
xn+1 = A n 1, 2, . . .
satisfies E (Jon [0, x0] for each n. The rest of the proof remains basically
the same.]
Prove Theorem 10.26.
10. Recall that a bounded operator T: X Y between Banach spaces is a
DunfordPettis operator if 0 in X implies 0 in Y.
Show that every positive DunfordPettis operator on a Banach lattice
which is quasinilpotent at a non-zero positive vector has a non-trivial
closed invariant ideal. [HINT: Let B: E E be a positive Dujiford
Pettis operator on a Banach lattice which is quasinilpotent at a non-zero
positive vector. We can assume that B is strictly positiveotherwise
NB is a non-trivial closed invariant ideal. Now note that B carries order
intervals onto relatively weakly compact sets [30, Theorem 19.12, p. 339],
and so B2 is a non-zero AM-compact operator.
10.4. Invariant Subspaces of Families of Positive Operators

The objective of this section is to present several invariant subspace results
for collections of positive operators on Banach lattices. In particular, several
theorems obtained in Section 10.3 will be extended to semigroups of oper-
ators on Banach lattices. Among other things, it will be shown that each
multiplicative semigroup of quasinilpotent compact positive operators has a
common non-trivial closed invariant subspace. Though many ideas of the
proofs remain the same as in Section 10.3, to apply them to semigroups of
operators, some additional techniques must be developed. This has been ac-
complished by R. [114] whose interesting work we adopt here. As
noted in [114], some of these methods are based on the works by V. S. Shul-
man [301] and Y. V. Throvskii [324]. As before, throughout the section
X will denote a real or a complex Banach space and E will denote a real
Banach lattice.
We begin our discussion with the definition of a semigroup of operators.
Definition 10.36. A S of bounded operators on a Banach space
is said to be a multiplicative (resp. an additive) semigroup if for each
pair 8, T E S the operator ST (resp. S + T) also belongs to S.
To avoid trivialities, we tacitly assume in this section that all collections

or families of operators under consideration in this section are non-empty.
We continue by introducing some notation and terminology regarding a

collection C of operators in L(X). For a subset D of a Banach space, we
let lID 11 = supXED lxii. Accordingly, ItCh = ItCh. As usual, for each
x E X we let Cx = {Cx: C E C}, and therefore IICx!I C E C}.
For any A E (X) we define
AC={AC: CEC} and CA={CA: CEC}.
For each n E N, we shall also use the notation
and hope that whenever this notation is used it will not cause any confusion
with the standard Cartesian product notation. The commutant of C is the
unital algebra of operators defined by
C'={Aer(X): AC=CA for all CEC}.
A subspace V of X is said to be C-invariant if V is C-invariant for each
C E C. A collection C of operators is said to be non-transitive if there
exists a non-trivial closed C-invariant subspace. Otherwise, the family C is
called transitive.
We now extend the concept of local quasinilpotence to collections of
operators.
Definition 10.37. A family C of operators in (X) is said to be:
(1) (locally) quasinilpotent at a point x E X if lim 0
and
(2) finitely quasinilpotent at a point x E X if every finite subcol-
lection of C is (locally) quasinilpotent at x.
Finitely quasinilpotent algebras of operators were considered by V. S. Shul-

man {301J. Recall that if T: X X is a bounded operator on a Banach
space, then Qr denotes the subset of X consisting of all vectors at which T
is locally quasinilpotent, i.e.,
Qr={xEX:
It was shown in Exercise 5 of Section 9.2 that QT is a T-hyperinvariant
subspace of X. To generalize this to a collection of operators C, we let
= {x E X: C is finitely quasinilpotent at x}.
As in the aforementioned exercise, we can prove the following result.
Lemma 10.38. The set is a vector snbspace of X and this subspace is
both C-invariant and C'-invariant.
UI 4. Invariant Subspaces of Families of Positive Operators -- 411
For the rest of the discussion in this section we shall assume that:
C denotes a non-empty collection of positive operators on a Banach
lattice E.
The presence of the order structure on E leads naturally to a modification
of the set Namely, we let
It follows immediately from Lemma 10.38 that is an ideal in E.

Recall that for a positive operator C: E E on a Banach lattice, we
denote by [C) the collection of all positive operators A: E E such that
[A,C} 0. That is,
[C) = {A e AC - CA> 0 }.
In accordance with this notation we also let
[C)={AEL(E)+:ACCAOforeachCEC}.
That is, [C) = flcc[C). Similarly, we let
of positive operators in (E) the set [C)

C
is a norm closed additive and multiplicative sernigroup in and contains
the zero and the identity operators.
Proof. The fact that [C) is norm closed is obvious. Also, it should be
clear that the operators 0 and I belong to [C). Now take two arbitrary
operators 5, T in [C). Then for each operator C E C we have SC CS and
TC CT. Adding up the two inequalities, we get (S + T)C C(S + T),
that is, S + T E [C). Finally, note that
STC=S(TC) SCT=(SC)T CST.
Therefore, ST E [C) and the proof is finished.
Lemma 10.40. The ideal is both C-invariant and [C)-invariant.
Proof. Fix x E This means that 0 for each finite subset
of C. We must prove that Cx and Tx belong to for each C E C and

each T E [C). To this end, fix C E C, T E [C) and let = {C1,. be
. . ,
a finite subset of C.
If we let = {C1, , C}, then
. . . = {C1, . . , C
n E N. Therefore, from
1
n+1
< =
we get Cx f
0. This shows that Cx E and hence the ideal
is C-invariant.
Now note that for each 1 k. This implies at once that
i
for each operator A E we have AT TA, and therefore
<
Consequently, hF Tx E This proves that the ideal

is also [C)-invariant.
We need to introduce two additional collections associated with an arbi-

trary collection C of positive operators on K The first of these collections is
the multiplicative semigroup generated by C in It is the small-
est (with respect to inclusion) semigroup of operators that contains C and
itwill be denoted by It should be clear that consists of all finite
products of operators in C. In other words,
The second collection, denoted by is also a large collection of positive

operators that is defined as follows:
and are presented in the next lemmas.
Lemma 10.41. If family of positive operators, then the collection

C is a
is an additive and multiplicative semigroup in (E).
Proof. Pick any two operators D1 and D2 in Hence,

for some E [C) and e where j = 1, 2. The fact that D1 + D2
belongs to follows immediately from the definition of Let us verify
that D1D2 E DC. Indeed,
flj 722 flj 722
D1D2 T1,kSl,k] =
k=1 i=1 k=i i=1
Since [C), it follows that (see Exercise 5 at the end of the
section) and hence Therefore,
D1D2
k=i
10.4. Invariant Subspaces of Families of Positive Operators 413
Since [C) and are semigroups, we have that E [C), S1,k52,i E Sc

and this proves that the operator D1D2 belongs to Dc.
Recall that for x E E the symbol DcX denotes the orbit of x under the
action of that is, D E VC}. To continue our discussion, we
shall introduce one more notation.
. For each x E E we will denote by [DC x] the ideal generated by DCX.
As we shall see, the ideals [Dcx] will provide a source of invariant subspaces.
Lemma 10.42. Each ideal [DCX] is both C-invariant and [C)-invariant.
Proof. Take any y E [DCX]. Since is an additive semigroup, it follows

that ADx for some scalar A and some D E DC. By the definition of
DC there exist operators E [C) and Si E 5C (i = 1,2,... ,n) such that
D and so <A
Next, we shall show that [Dc x] is C-invariant. To this end, fix C E C and
consider the vector Cy. From for each i, we see that
ICyL
Since E 5C for each i, we see that A = E DC. Therefore,

Cy E [DCXJ.
To get that [DCX] is [C) -invariant, let T E [C). Since [C) is a multiplicative
semigroup, E [C) for each i, and hence the operator B =
belongs to Now notice that
Ty E [DCX] and the proof is finished.
Lemma 10.43. If C is finitely quasinilpotent at some XO E then both
and Dc are likewise finitely quasinilpotent at XO.
Proof. We begin by showing that the semigroup 5C is finitely quasinilpotent

at x0. To this end, fix 0 < 1 and a finite subset of By the
definition of 5C' there exists a finite subset g c C and an m E N such that
9c Since C is finitely quasinilpotent at XO and the set is finite,
we have for each n Now for n > no we have
rn mn
<ffUgk]flxfl =
= max < max =
c n n0. This shows that 5C is finitely

quasinilpotent at xO.
Next, we shall verify that each operator D E DC is quasinilpotent at

XO. By the definition of D where
E [C) and E for each i = 1,. ..,m. Put c = maxl<j<m IITdI and
= . , Srn}, and then a direct verification shows that
If
<
This implies cm 0, since is finitely quasinilpo-

tent at xO. Hence, D is quasinilpotent at xO.
The fact that is finitely quasinilpotent at xO can be demonstrated in
a similar manner.
Theorem 10044 If a collection C of positive operators
(1) is finitely quasinilpbtent at some positive non-zero vector, and
(2) some operator in C dominates a non-zero operator,
then C and [C) have a common non-trivial closed invariant ideal.
Proof. Let C0 E be an operator that dominates a non-zero AM-compact

C
operator K. The basic steps of the proof follow those of Theorem 10.24. By
the hypotheses the ideal is non-zero, and by Lemma 10.40 we know that
is invariant under C and [C). If is not dense, then the norm closure of
is a common non-trivial closed invariant ideal. So, we can assume that
is a norm dense ideal.
From K 0, it follows that there exists some 0 < xO E such that
Kx0 0. Without loss of generality we can assume that
IIKII = 1, IXOII >1, and IIKxoII >1.

Let Uo be the closed unit ball centered_at xO and let F = U0 fl [0, XO}. The
set F is order bounded and hence K(F) is_compact since the operator K is
AM-compact. It should be clear that 0 K(F).
Recall that for each x 0 we denote by [Dcx] the ideal generated by the
orbit x = {Ax: A E } of the semigroup Since the identity operator
belongs to it follows that x E x] and so x} is non-zero provided
x> 0. As shown in Lemma 10.42, this ideal [Dcx] is invariant under both
C and [C), and so we are done if there exists some x} which is not norm
dense in E.
Therefore, we can assume without loss of generality that xJ is dense
in E for each x > 0. This implies that for each y 0 there exists an operator
I 0. 4. Invariant Subspaces of Families of Positive Operators 415
A such that lixo XO A < 1. Because 0 K(F), it follows that
K(F)c U {yEE:
AEVc
All sets in this union are obviously open, and so we can find a finite subcover.
That is, there exist operators A1,.. , E such that
k(F) c {y E E: xO AAj(lyI)JI <1 }.
Let A = A1 + ' + Am + C0. Since I E [C), E C, and = IC0, we

see that A E and thus, in fact, K(F) is covered by a single set. That is,
K(F) C {yEE: <i}.

Now consider y = Kx0. Then y E K(F) so that lixo A A(IKxoI)I1 < 1,
x0
and also 0 < x0 A < Hence the vector x1 = x0 A A(IKxoI)
belongs to F, and therefore Kx1 E K(F). Repeating this argument with
y = Kx1, we will get x2 = x0 A A(IKxiI) E F. Proceeding by induction, at
the ri + 1 step we define = X0 A and obtain E F fl Uo.
Moreover, we have
<
Since x0 E Lemma 10.43 guarantees that the operator A E is
quasinilpotent at x0, and so 0. In particu-
lar, 0. However, the latter conclusion contradicts that C U0
and 0 U0 = U0, and the proof of the theorem is complete.
Let B: E E a positive operator satisfying the conditions of The-

be
orem 10.24. That is, assume that some positive operator S E (B] is quasi-
nilpotent at a positive non-zero vector and dominates a non-zero compact
operator. Consider the collection C = {S} consisting of a single operator S.
Then clearly C satisfies the hypotheses of Theorem 10.44 and hence there
exists a non-trivial closed ideal invariant under C and [C). The assumption
S E (B] implies that B E [C), and so this ideal is certainly B-invariant.
In other words, Theorem 10.44 implies Theorem 10.24. However, we have
decided to keep both proofs in order to separate the details essential for an
individual operator from the details essential for collections of operators.
The semigroups of positive operators having a common non-trivial closed
invariant ideal are often referred to as ideal reducible semigroups.
cordingly, the semigroups without a common non-trivial closed invariant
ideal are referred to as ideal irreducible semigroups. This is in agree-
ment with the terminology of Chapter 9. Using Theorem 10.44 we can easily
prove that various sernigroups of positive operators are ideal reducible. We

begin with additive semigroups.
Corollary 10.45. If for an additive semigroup S of positive operators
(1) each operator in S is quasinilpotent at some vector > 0 and
(2) some operator in S dominates a non-zero AM-compact operator,
then S and [5) have a common non-trivial closed invariant ideal.
Proof. We will apply Theorem 10.44 to the collection S of positive opera-

tors. The only condition we need to verify is that S is finitely quasinilpotent
at > 0. To this end, let F = Sk} be a finite subset of S and
. .
,
consider the operator S = S

S S S is quasinilpotent at xO. That is, IISnxo p 0.
Since each operator A E ..P' is the product of n operators from F, it

follows that A < and hence lAx0 II IlSnxolI. Therefore, we have
1 1 1
= sup
and so 0.This implies that the additive semigroup S is finitely

quasinilpotent at By Theorem 10.44 the semigroups S and [5) have a
common non-trivial closed invariant ideal.
To get an analogue of Corollary 10.45 for multiplicative semigroups, we

need the following deep result due to Y. V. Turovskii in Corollary 7]
Theorem 10.46 (Tarovskii). Each multiplicative semigroup of compact
quasimilpo tent operators on a Banach space is finitely quasinilpotent.
Corollary 10.47. If S is a non-zero multiplicative semigroup of
potent compact positive operators on a Banach lattice, then S and [5) have
a common non-trivial closed invariant ideal.
Proof. According to Theorem 10.46 the multiplicative semigroup S is

finitely quasinilpotent. Now apply Theorem 10.44.
Corollary 10.47 can be viewed as an extension of Theorem 10.27 to semi-

groups of compact positive operators. As we shall see later (Example 10.53),
the compactness assumption in Corollary 10.47 is essential.
Corollary 10.48. Let B: E i E be a non-zero positive operator on a dis-
crete Banach lattice. If B is locally quasinilpotent at some non-zero positive
vector, then B has a non-trivial closed -hyperinvariant ideal.
1 0. 4. Invariant Su bspaces of Families of Positive Operators 417
Proof. In the proof of Theorem 10.29 we indicated that the operator B

dominates a non-zero compact operator. Therefore, the singleton collection
C = {B} satisfies the hypotheses of Theorem 10.44. Consequently, there
exists a non-trivial closed ideal J that is invariant under C and [C) But.
each positive operator that commutes with B belongs to [C) Thus, J is

.
-hyperinvariant for B.
Some special cases of Corolaries 10.45, 10.47, and 10.48 have been ob-
tamed by M. Jahandideh [158]. Our next goal is to extend Theorem 10.27
to collections of positive operators. This requires some preliminary work.
Recall that a subset 5o of a multiplicative semigroup S is said to be a
(two-sided algebraic) semigroup ideal in S if for each T E So and for
each S E S the operators TS and ST belong to Clearly, any algebraic
semigroup ideal is a multiplicative semigroup in its own right. Since the
intersection of any family of algebraic semigroup ideals is again a semigroup
ideal, it follows that for any collection of operators C there exists a smallest
semigroup ideal containing this subsetthis semigroup is called
the semigroup ideal generated by C.
The following simple result provides a well-known method of proving
the existence of invariant subspaces for semigroups of operators; see [271,
Lemma 1] and [114, Lemma 4.6].
Lemma 10.49. Let S be a multiplicative group of continuous operators on
a Banach space and let 5o be a non-zero semigroup ideal in S. If So has
a (common) non-trivial closed invariant subspace, then S also has a non-
Moreover, ifS consists of positive operators on a Banach lattice and the
semigroup ideal S
has a closed invariant ideal.
Proof. Let V be a non-trivial closed subspace invariant under a non-zero

semigroup ideal 5o of the semigroup S. Consider the closed vector subspace
W of V spanned by the set G = {Tx: x E V and T E So}.
First, we claim that W is invariant under S. To establish this, choose
S E 5, T E 5o, and x E V. We must show that the vector S(Tx) remains
in W. But the operator ST belongs to since So is a semigroup ideal in
5, and hence S(Tx) = STx E G W. This shows that if W {0}, then
W is a closed S-invariant subspace.
So, let us assume that W = {0}. This means that each operator in
is zero on V. Consider N = flTESOKer T. Clearly, N is a closed subspace,
V N and also since So is non-zero, N is not equal to the whole space.
That is, N is a non-trivial closed subspace. It remains to notice that this
418 - 10. Invariant Subspaces
subspace is invariant under S. Indeed, take any x E N and any S E S. We

want to show that Sx E N, that is, TSx = 0 for each T E S. But this is
obvious, since the operator TS belongs to the semigroup ideal So and thus
indeed TSx = 0.
The second statement can be proven in a similar manner.
Theorem 10.50 [114]). Assume that a non-zero collection C of
positive operators on a Banach lattice:
(1) finitely quasinilpotent at a non-zero positive vector and
is
(2) its corrtmutant C' contains an operator that dominates a non-zero

compact operator.
Then C arid [C) have a common non-trivial closed invariant ideal.
Proof. It follows that the ideal is and, by Lemma 10.40, we
know that is invariant under C and [C). Therefore, we can assume that
is norm dense in E (otherwise its closure is a non-trivial closed ideal
invariant under C and [C)).

Now consider the intersection of all null ideals generated by the
operators in C, that is, = flcEc Clearly, this is a closed ideal and
a straightforward verification shows that it is invariant under both C and
[C); see Exercise 4 at the end of the section. Therefore, if is
then it is a common non-trivial closed ideal invariant under C and [C).
Consequently, we can also assume that = {O}.
Let K be a non-zero compact operator that is dominated by some pos-
itive operator T0 E C'. Since is norm dense in E, we can find a vector
0 < xO E such that Kx0 0.
we consider the semigroups Sc and
Next, associated with our collec-
tion C. Recall that is the multiplicative semigroup generated by C and
is the multiplicative and additive semigroup defined as follows:
D }.
We will also need the multiplicative semigroup SCU[c) generated by the dollec-
tionC U [C). We claim that Indeed, take an arbitrary operator
A E 8cu[c) and recall that the elements of the semigroup SCU[C) are just the
finite products of the operators in C U [C). From the definition of [C) we
also know that for any operators C E C and T E [C) we have CT TC.
This immediately implies that the operator A is dominated by an operator
of the form T1 .. TmCi . Ck for some T1,... ,Tm E [C) and C1,.
. Ck E C.
. ,
But the operator T = T1 .. belongs to [C) since [C) is a semigroup, and

Invariant Subspaces of Families of PositiveOperators 419
similarly the operator S = Ci belongs to the semigroup

' Ck In other .
words, we have shown that each operator A in 8CU[C) is dominated by an

operator of the form TS with T E [C) and S and consequently A
belongs to
Notice next that for each C E C the operator T0C belongs to 5CU[C)'
and therefore we can consider the (two-sided) semigroup ideal in 8CU[C)
generated by T0C. That is, is the semigroup ideal generated by the set
{T0C: CEC}.
From 8Cu[C) c we obtain So c Since C is finitely quasinilpotent
at know from Lemma 10.43 that is also finitelyquasinilpotent at
xO, and therefore is likewise finitely quasinilpotent at x0.
us verify that
Let contains some operator that dominates a non-zero
compact operator. As we know (see Exercise 5 in Section 10.3), there exists a
compact operator K1 such that IKixl for each x E and K1x0 > 0.
Since = {0}, we can find some C E C satisfying CK1x0 > 0. Hence,
CK1 is a non-zero compact operator on E. Moreover, for each x E E we
have < CIKxI CT0 IxI, establishing our claim since
CT0 E
We have shown by now that the collection of positive operators is
finitely quasinilpotent at x0 > 0 and contains some operator dominating
a non-zero compact operator. That is, satisfies all the hypotheses of
Theorem 10.44. Therefore, according to this theorem, has a non-trivial
closed invariant ideal. But 5o is a semigroup ideal in 5Cu[C)' and so from
Lemma 10.49 we see that 5cu[C) also has a non-trivial closed invariant ideal.
Recalling that C U [C) c 8Cu[C)' we infer that C and [C) have a common
closed invariant ideal. This completes the proof.
Recall that in Theorem 10.35 we have proved that if a positive operator

B: C0 C0 is locally quasinilpotent at some vector x0 > 0, where
is a locally compact Hausdorif space, then B has a non-trivial closed
invariant ideal. The next theorem is a semigroup version of this result. The
proof is very similar to that of Theorem 10.35 and we will just outline it.
Theorem 10.51. Let be a locally compact Hausdorff space and let S be a

non-zero multiplicative semigroup of positive operators on which are
quasinilpotent at some xO > 0. If S is also an additive semigroiip, them S
has a common closed invariant ideal.
Proof. If Sxo = 0 for each S E 5, then there is nothing to prove, since

the intersection of the null ideals J = Sjxj = o} is a
non-trivial closed ideal.
So, we assume that > 0 for some.80 E S. Without loss of generality

we may assume that the support set F = {t xO (t) > 0} of the function
xO is compact. Next, as in the proof of Theorem 10.35, it is easy to show
that there exists a point t0 such that
(to) = 0
for each S S. Finally, consider the ideal
j = {x xJ where S and
}
Clearly, J {O} as 0 S0x0 E J. It should also be obvious that the ideal J

is S-invariant. Furthermore, J cannot be dense in since (to) = 0
for each operator S E S. Therefore, the norm closure J of J is a
closed &invariant ideaL
It comes as a surprise that the assumption of additivity in Theorem 10.51

is essential. Moreover, without this assumption even the quasinilpotence of
each operator in the semigroup is not enough to guarantee the existence of a
common closed invariant ideal. To present such an example, we
need some preliminary discussion. Below, I' will denote the unit circle, and
z= I' will be a fixed complex number whose argument is an irrational
multiple of 2ir. It is well known that for each t e I' the orbit {tzm: m N }
isdense in I'; see Exercise 2 in Section 7.3. With each function w E C(T)
we associate the weighted composition operator C(T) C(I') defined
for each x C(T) by
(*)
We will make use of the following remarkable formula, due to A. K. Ki-

tover [186] and independently A. B. Antonevich and A. V. Lebedev [38],
for calculating the spectral radius of the operator
Theorem 10.52 (AntonevichLebedevKitover). If is the operator de-
fined by (*) above, then its spectral radius is given by
in dO
=
Proof. For simplicity let T = Assume first that there exists some
> 0 such that w(t)J for each t E I'. So, the function t ' is
continuous. Clearly, for n E N, x E C(T), and t I' we have

= w(t)w(tz) . . .
10.4. Subspaces of Families of Positive 421
This implies that

sup =
<1
for some E IT. If we consider the probability measure on I' defined by

= then it is easy to see that
= (*)
Now notice that the sequence lies in the closed unit ball of C(F)*,
which is w*_compact. So, there exists a subnet } of such that
for some measure This and (*) imply at once that
r(T) = n-+oo
lim lim
a
We will establish next that the measure ji is the normalized Lebesgue
measure on I'. To this end consider the operator S: C(F) * C(T) defined
by Sx(t) = x(tz). From (x, = =
it follows that IIS*/#Ln This implies IS*,an in
particular we get 0. From ji, we get
and so = The latter is equivalent to
Jr/
Sxd/i= /xd/i
Jr
for each x E C(F).
Using (**) we can easily show that the measure is rotation invariant.
To this end fix an arbitrary s E I'. Since the argument of z = is an
irrational multiple of 27r, there exists an increasing sequence of natural
numbers such that s. It follows that x(st),aud-
so from (**) and from the Lebesgue Dominated Convergence Theorem we
get
f rx(st) f x(t) d/i(t)

= r
for all x E C(r). Clearly (t) is also true for each x E L1(F, In particular,
it follows that
=
for each Borel subset A of I'. But then, by the uniqueness of the Haar
measure, we conclude that must be the normalized Lebesgue measure on
I'. Thus, in the special case when wi is separated away from zero, we have
established that
= efr In Iw(t)J d1i(t) = In lw(&)I dO
For the general case, consider the sequence of continuous functions {wk}
given by wk = wi + and let Tk = Twk. From for all n we
get r(T) r(Tk) = efr IWkkiIi
for each k. Letting k oo yields
r'(T) (tt)
In particular, if the measurable function t ln Iw(t) is not Lebesgue

h1 Iwl d,i
grable, i.e., if f1, ln wi d,u = oo, then r(T) = efr = 0.
remains to consider the case when the function t
It In Jw(t)J is
grable. In this case, using (t) we get
fin [Iw(t)w(tz) . . .
(]
=
ln I
'Jr
Consequently, lITTh II = efr In Iw(t)I d[L(t) for each n and hence
r(T)
Combining this with (if), we see that the formula r(T) = efr In wi d,i is true
in this case too, and the proof is finished.
It is worth pointing out that Theorem 10.52 is a special case of the

following result obtained by the same authors.
o Let b: p horn eornorphisrn of a compact Hausdorff space,
be a
let w E C(IZ), and consider the operator C(IZ) * C(IZ) defined
by = w(w)x(cb(w)). Then the spectral radius of the operator
Tw is given by
r(Tw) = sup
where the supremurn is taken over all b-invariant Borel probability

rneasures on ft
From Theorem follows that Tw is quasinilpotent if and only if
10.52 it
the Lebesgue measurable function t ln w(t)I is not integrable over I'.
For our final result let us fix some function wo E with the following
properties:
10.4. Invariant Subspaces of' Families of Positive Operators 423
(a) O<wo(t)<lforeachtEF.
(b) t 1 is the oniy point at which wO vanishes.
(c) wo (t) = 1for all t outside of a small neighborhood of t = 1.

(d) The function wO drops rapidly enough at t = 1, so that the function
t lnwo(t) is not integrable.4
These conditions imply that the operator is positive and quasinilpotent.

Example 10.53. There exists an ideal irreducible multiplicative semigroup
S of quasinilpotent positive operators on C(F).
Using the function wO satisfying the properties (a)(d) preceding the

statement of the theorem, we introduce the following collection of weights:
W {w E C(F): 0 w(t) wo(ts) for each t E F and some s E F}.
Next, for each w E W and n E N define the operator C(F) C(F) by
(t) = w(t)f(tz72), f E C(F), t E F,
where z = E F is a fixed complex number whose argument b is an
irrational multiple of 27r. Let S denote the collection of all these operators.
That is,
S={Tw,n: wEWaridnEN}.
A glance at (*) shows that each operator E S is quasinilpotent. To
verify that S is a multiplicative semigroup, let E 5, and so for
each f E C(F) and each t E F we have
(t) =
where i = 1,2. A direct computation shows that

(t) = w2(t)wi(tzTh2)f(tzflh+fl2) = w(t)f(tzfh+Th2)
where w(t) w2(t)wi(tzTh2). Clearly w E W since 0 w w2, and
therefore E S.
We claim that this semigroup S is ideal irreducible. If not, then there
exists a non-trivial closed ideal J in C(F) which is T-invariant for each
T E S. Consequently, there exists a closed set F c F such that J consists
of all functions that vanish on this set F. Clearly, F cannot be dense in I',
otherwise J would equal {O}. Also, F is non-empty, otherwise J = E. So,
there exists a point to E F such that some neighborhood of t0 does not meet
F. Pick a non-negative function f E C(F) such that f vanishes on F (and
therefore f belongs to J) and f is greater than zero on some open set
containing to.
instance, the function wo(t) = about t= 1 (or 0 = 0) satisfies this

non-integrability condition.
424 - 10. Invariant Subspaces
Now fix some t1 E F. Since the set m E N} is dense in F, there

:
exists some k E N such that the number t2 = tizk belongs to Next

choose s E F such that wo(t1s) = 1, and let w(t) = wo(ts) for t E F. Then
w E C(F), 0 w(t) wo(ts), and w(ti) = 1. Finally, consider the operator
ES. Since J if T-invariant, E J, and so = 0. On the
other hand
(t1) = w(ti)f(tizk) = f(t2) > 0,
a contradiction. This establishes that the multiplicative semigroup S is ideal
irreducible.
An example of an irreducible multiplicative semigroup of nilpotent pos-

itive operators on L2 [0, 1] was constructed by Y. Zhong [349]. Moreover, as
shown in [115], it is possible to obtain in addition that the square of each
operator in the semigroup is zero and to construct such a semigroup on any
1] with 1 p < 00. Both Example 10.53 and the above mentioned
examples show also that the assumption of compactness in Corollary 10.47
is essential.
Exercises
1. Let S be a multiplicative semigroup of operators on a Banach space and
let S U {I}, where I is the identity operator. Show that
S and have the same in-
variant subspaces and therefore, when dealing with invariant subspaces
of a multiplicative semigroup, we can always assume that the semigroup
is unital.)
2. Let S be an additive sernigroup of operators on a Banach space and let
= S U{0}, where 0 is the zero operator. Show that is again an addi-
tive semigroup. (Clearly, S and have the same invariant subspaces and
therefore, when dealing with invariant subspaces of an additive semigroup,
we can always assume that the seniigroup contains the zero operator.)
4. Let C be a collection of positive operators on a Banach lattice E and let
= flCECNC. Show that is invariant under C and [C).
5. Let C be a collection of positive operators on a Banach lattice and let
be the multiplicative semigroup generated by C. Establish the following
identities: [C) = [Se), (C] = (Se], and C' =
6. Let C be a commutative collection of positive operators on a Banach
lattice. If C is quasinilpotent at some vector > 0, then show that C is
finitely quasinilpotent at
7. Finish the details of the proof of Theorem 10.51.
10.5. Compact-friendly Operators 425
8. Let C be a norm bounded collection of operators in (X). The joint (or

the RotaStrang) spectral radius of C is defined by
r(C) = inf
nEN
where we let = {C1 C2 E C} and 11= S E Cn}.

Show that r(C) lim
fl4 00
9. Assume that a positive matrix defines a continuous operator on some

4, where 1 p <oc, and this operator is quasinilpotent at some non-zero
positive vector. Consider also the collection W of all double sequences
w = i, j = 1, 2, . .} of scalars such that the operator
. defined
by the weighted matrix is a regular operator on 4. (Clearly, W
contains all bounded double sequences.)
Show that the family of operators w E W} has a common
non-trivial closed invariant ideal.
10.5. Compact-friendly Operators

As we have already seen, the compact positive operators enjoy many special
properties concerning the invariant subspace problem. In this section we
will show that the compactness assumption can be relaxed considerably
while still preserving these properties. The corresponding class of operators
was introduced in [11] under the name of compact-friendly operators. The
connections between compact-friendly operators and compact operators are
brought about via the order structure.
Definition 10.54. A positive operator B: B > E on a Banach lattice
is said to be compact-friendly if there exists a positive operator in the
commutant of B that dominates a non-zero operator that is dominated by a
compact positive operator.
That is, a positive operator B is compact-friendly if and only if there
exist three non-zero operators R, K, C: E E such that R, K are positive,
K is compact, RB = BR, and for each x B we have
and
On a finite dimensional Banach lattice every operator is compact and

so the notion of a compact-friendly operator is of substance only on infi-
nite dimensional Banach lattices. Clearly, every power (even every polyno-
mial with non-negative coefficients) of a compact-friendly operator is also
compact-friendly. Here are some other examples of opera-
tors.
o Compact positive operators.

Positive operators in the commutant of a non-zero compact positive
operatOr.
e Positive operators that dominate non-zero compact positive
o Positive integral operators (see Lemma 10.58 below).
The above examples show that a compact-friendly operator need not

be compact. The identity operator (which clearly is is
perhaps the best example to illustrate this. The relevance of the compact-
friendly operators to the invariant subspace problem is demonstrated by the
following basic theorem, which is an improvement of a result from [11J.
Theorem 10.55 (AbramovichAliprantisBurkinshaw). If a momzero com-
pact-friendly operator B: E f E on a Banach lattice is quasinilpotent at
some xO > 0, then B has a non-trivial closed invariant ideal.
Moreover, for each sequence in [B) there exists a mom-trivial closed
ideal that is invariant under B and under each
Proof. Without loss of generality we can assume that B < 1. Pick ar-
bitrary scalars > 0 that are small enough so that the positive operator
T= exists and lB + TIl <1. Since [B) is a norm closed additive
semigroup, it follows that T E [B). The same arguments show that for each
n E N the operator (B + T)Th also belongs to [B) and, consequently, -the
positive operator A = + T)Th belongs to [B) too.
For each x > 0 we denote by the principal ideal generated by Ax,
that is,
Jx{YEE: ly!AAxforsomeA>0}.
The obvious inequality x Ax implies that x E and so this ideal is
non-zero.
Observe next that the ideal is (B + T)-invariant. Indeed, if y E
then AAx for some 0 and so we have
00 00
J(B+T)yI < A(B+T) = ,\Ax,
as claimed. Because 0 B, T B + T, it follows that is invariant under

B and T. Clearly this ideal is also Ta-invariant for each n. The preceding
arguments imply that if for some vector x > 0 the norm closure
then is a desired nontrivial closed ideal which is invariant under B and
under each
Operators 427
Consequently, to proceed with the proof of our theorem, we can assume

from now on that
(*)
for each x > 0, that is, Ax is a quasi-interior point in E for each x > 0. (In
the terminology of Chapter 9, A is a strongly expanding operator.)
Fix three non-zero operators R, K, C: E E with K, R positive, K
compact, and satisfying
RBBR, ICxIR(lxI) and foreach
Since C 0, there exists some x1 > 0 such that Cx1 0. Then

Al Cx1 a point satisfying Al Cx1 Cxj. I' and it follows
from Lemma 4.16(1), that there exists an operator M1: E E dominated
by the identity operator such that x2 M1Cx1 > 0. Put = M1C, and
note that the operator is dominated by the compact positive operator K
and by the operator R.
By (*) we have = E and consequently, since C 0, there exists
o < y Ax2 such that Cy 0. Hence (again by (*)) the element AICyI is
a quasi-interior point. Since CyI there exists (by Lemma 4.16(1)
again) an operator M2: E E dominated by the identity operator such that
X3 = M2Cy > 0. Since Ax2 and Ax2 is a quasi-interior point, it follows
from Lemma 4.16(2) that there exists an operator lvi: E E dominated by
the identity operator such that MAx2 y. So, x3 = M2Cy JvI2CMAx2.
Let 112 = M2CMA and note that 112 is dominated by the compact positive
operator KA and by the operator RA.
If we repeat the preceding arguments with the vector x2 replaced by x3,
then we obtain one more operator 113: E E that satisfies 113x3 > 0 and
that is dominated by the compact positive operator KA and by RA.
Consider the operator 113112111. From 113112111x1 113x3 > 0, we see
that 113112111 is a non-zero operator. Moreover, as shown above, each
erator (i = 1,2,3) is dominated by a compact operator and, therefore,
Theorem 2.34 guarantees that the operator 113112111 is necessarily compact.
A straightforward verification also shows that
RARARIxI
for each x E E.
Let S RARAR. We will verify that the operator S e [B). To see this
recall that A [B) and that the operator R also belongs to [B) because
it commutes with B. Since [B) is a multiplicative semigroup, the operator
S = RARAR necessarily belongs to [B).
428 " 10. Invariant Subspaces
It remains to notice that the collection C = {B} consisting of the single

operator B satisfies the conditions of Theorem 10.50. Indeed, C is quasi-
nilpotent at xO > 0 and [C) = [B) contains the operator S that dominates
a non-zero compact operator. By this theorem, there exists a non-trivial
closed ideal that is invariant under [B). This completes the proof.
It might be worth mentioning that some additional assumption on a
compact-friendly operator is necessary if one wants to establish the exis-
tence of an -hyperinvariant ideal. Indeed, the identity operator is compact-
friendly but does not have any non-trivial closed -hyperinvariant ideal.
Corollary 10.56. Let 5: E E be a positive operator on a Banach lat-
tice and suppose that (5] contains a non-zero compact-friendly operator that
is quasinilpotent at some positive vector. Then S has a mon-trivial closed
invariant ideal.
Proof. Let .8 E (5] be a non-zero compact-friendly operator that is quasi-
nilpotent at some positive vector. As we know, the condition B E (5] is
equivalent to the condition S E [B) and consequently the desired conclusion
For Dedekind complete Banach lattices we can improve Theorem 10.55
by proving that there always exists a non-trivial closed [.B)-invariant ideal.
The proof is a modification of that of Theorem 10.55.
Theorem 10.57. If a non-zero compact-friendly operator B: E f E on
a Dedekind complete l3anach lattice is quasinilpotent at some xO > 0, then
there exists a non-trivial closed ideal that is invariant under [B).
Proof. For each x > 0 we denote by the ideal generated by the orbit
{B)x, that is,
Jx={yEE Iy1AxforsomeAE[B)}.
It is obvious that is a non-zero ideal. Observe next that the ideal is
[.B)-invariant. Indeed, if y E then Ax for some A E [B) and so for
any A1 E [B) we have
IAiyI AiIyI A1Ax.
Since [B) is a multiplicative semigroup, the operator A1A belongs to [B)
and we see that A1y E as claimed. Therefore, if for some vector x > 0
the ideal is not norm dense in E, then its norm closure is a desired
non-trivial closed [B)-invariant ideal.
Consequently, to proceed with the proof of our theorem, we can assume
from now on that
(*)
for each x > 0. Our next step is to show that the assumption (*) implies
that [B) contains an operator that dominates a non-zero compact operator.
Fix three non-zero operators R,K,C: E * E with K,R positive, K
compact, and satisfying
RB=BR, ICx!R(IxI) and foreach XEE.
Since C 0, there exists some > 0 such that Cx1 0. This means
that at least one of the vectors (Cxi)+ or (Cxi)_ is non-zero, and therefore,
since E is Dedekind complete, there exists an operator M1 dominated by
the identity operator such that x2 = M1Cx1 > 0. Put = M1C and note
that the operator ii is dominated by the compact positive operator K and
by the operator R.
Since is norm dense in E and C is non-zero, there exists some y e
and an operator A1 E [B) such that 0 < y A1x2 and Cy 5L 0. Using again
the hypothesis that E is Dedekind complete, we can find operators M and
M2 dominated by the identity operator and such that y = MA1x2 and
X3 = M2Cy = IvI2CMA1x2 > 0. Let 112 = and note that the
operator 112 is dominated by the compact operator KA1 and by the operator
RA1.
If we repeat the preceding arguments with the vector x2 replaced by X3,
then we obtain an operator A2 E [B) and an operator 113: E f E which
satisfies 113X3 > 0 and which is dominated by the compact operator KA2
and by the operator RA2.
Consider the operator 113112111. It is non-zero as 113112 = 113X3 > 0.
Moreover, as shown above, each operator (i = 1,2,3) is dominated by a
compact operator and therefore Theorem 2.34 guarantees that the operator
113112111 is necessarily compact. A straightforward verification also shows
that
113112111x1 RA2RA1R(IxI)
for each x E E.
Finally, consider the operator S = RA2RA1R. We claim that S e [B).
To see this, recall that A1, A2 e [B) and also R E [B) because R commutes
with B. Since [B) is a multiplicative semigroup, the operator S = RA2RA1R
necessarily belongs to [B).
It remains to notice that the collection C = {B} consisting of a single
operator B satisfies the conditions of Theorem 10.50. Indeed, C is quasinil-
potent at > 0 and [C) = [B) contains the operator S that dominates a
non-zero compact operator. By this theorem there exists a non-trivial closed
ideal that is invariant under [B). This completes the proof.
Inorder to proceed further, let us recall briefly some terminology and

notation introduced in Sections 4.1 and 5.2. A (continuous)
on a Banach space X is any operator of the form
where E X (and n(x) = x*(x)n for x E X). Any operator of
the form is called a finite-rank operator. As usual, X* X
denotes the vector space of all continuous finite-rank operators on X.
Suppose now that E is a Dedekind complete Banach lattice. Then,
by the RieszKantorovich Theorem 1.16, the space of all regular operators
T (E) is a Dedekind complete vector lattice. For an arbitrary vector sub-
space of E* the operators in the band (4 0 generated by E in
4(E) are called operators. When = the
operators are referred to as abstract (or almost) integral operators.
The terms are justified by Lozanovsky's Theorem 5.28 asserting that on a
Banach function space an operator is almost integral if and only if it is a
regular integral operator.
The next characterization of E*_integral operators follows immediately

from their definition:
A positive operator T: E + E on a Dedeki rid complete Banach

lattice is an E* operator if arid only if there exists a net
{Ta} of positive operators on E such that each Ta is dominated by
a finite-rank operator and 0 TaX I Tx for each x 0.
Letting R = T and C = Ta in the above characterization and taking for
K any finite-rank operator dominating Ta, we obtain the following result.
Lemma 10.58. Every positive operator is compact-friendly.
It is useful to know that the third power of a strictly positive
operator dominates a compact positive operator.
Lemma 10.59. If E is a strictly positive
8: E + operator on
a Dedekind complete Bariach lattice, then for each x0 > 0 there exists a
compact positive operator K: E + E such that 0 K and Kx0 > 0.
Proof. Fix any > 0. Since 8: E E is a positive operator,

there exists a net {Sa} of positive operators such that 0 Sa I S and each
Sa is dominated by a positive finite-rank operator. By Corollary 2.35, each
S is strictly positive, we have > 0. From thisand from

SaXO I Sx0, we see that there is some index so that SaXQ > 0 for all
a > Likewise from S(SaiX) > 0 and Ia S(Sa1XO) we get
> 0 for all a a2 Finally, from > 0 and
Sa(Sa2SaiX) Ia S(Sa2SaiX), it follows that Sa(Sa2SaiX) > 0 for all
a a3 ai.
10.5. Operators 431
Now let K = and note that K : E E is a compact positive

operator satisfying 0 K and Kx0 > 0.
We are now ready to present an invariant subspace theorem for

integral operators.
Theorem 10.60. Let S : E + E be a non-zero positive operator
on a Dedekind complete Banach lattice and let B : E E be any non-zero
positive operator commuting with S. If either S or B is quasinilpotent at
some non-zero positive vector, then the operators S and B have a common
non-trivial closed invariant ideal. Moreover, this ideal can be chosen to be
invariant either under [8) or under [B).
Proof. If S is not strictly positive, then we know that the non-trivial null
ideal N3 is invariant under [5); see Exercise 4(d) in Section 10.3. So, we can
suppose that S is strictly positive. In this case, it follows from Lemma 10.59
that 53 dominates a non-zero compact positive operator.
Assume first that S is quasinilpotent at some xU > 0. Consider the
collection C = {S} consisting of the single operator S. Then C satisfies th
conditions of Theorem 10.50. Indeed, C is quasinilpotent at xU > 0 and [5)
contains the operator that dominates a non-zero compact operator. By
this theorem there exists a closed ideal that is invariant under
[5). As S and B commute, the operator B belongs to [5).
Assume now that B is quasinilpotent at xU. Then consider the collection
C = {B} consisting of the single operator B. Again C satisfies the condi-
tions of Theorem 10.50 since C is quasinilpotent at xU and [B) also contains
the operator A second application of Theorem 10.50 guarantees the
existence of a non-trivial closed ideal that is invariant under [B).
Corollary 10.61. Every positive E* -integral operator that is quasinilpotent
at a non-zero positive vector has a non-trivial closed -hyperinvariant ideal.
T. And [34] and H. J. Krieger [194] proved that each positive irreduc-
ible integral operator on an has a positive spectral radius. In the
"invariant subspace" terminology this means that every positive quasinil-
potent integral operator on an Lu-space has a non-trivial closed invariant
subspace; see Corollary 9.37. This was the first result on the invariant
space problem in the framework of Banach lattices. It should be clear that
Corollary 10.61 is a substantial improvement of the AndKrieger theorem.
For Banach function spaces CorollarylO.61 specializes to the following result.
Corollary 10.62. Let E be a Banach function space associated with a
finite measure space ,i) and let 5: E E be a positive integral oper-
ator with kernel S(.,.). If S is quasinilpotent at some positive vector, then
432 1 0. Invariant Subspaces
the family of integral operators T: E E, defined by
Tx(t) = xE E,
where w(.,.) is an arbitrary bounded x ,a-measurable function, has a com-
mon non-trivial closed invariant ideal.
Proof. Observe that each integral operator T defined by a kernel is

dominated by a multiple of the operator S, and the conclusion follows from
Corollary 10.61 and Exercise 2 in Section 10.3.
We shall conclude the section with a characterization obtained in [17] of

compact-friendly multiplication operators on spaces of continuous functions.
To simplify the exposition, we need to introduce the following terminology
taken from [15].
Definition 10.63. A continuous function b: where is a topolog-
ical space, has a flat if there exists a non-empty open set G such that is
constant on G.
Similarly, a measurable function b: R, where E, is now a
measure space, has a flaL if cb is constant on some A E with > 0.
We continue with some useful properties of operators dominated by com-

pact positive operators.
Lemma 10.64. For a norm bounded sequence in a Banach lattice E
and a bounded operator A: E E that is dominated by a compact operator
we have following.
(1) The sequence has an order bounded subsequence.
(2) If E has order continuous norm and the sequence is disjoint,
then lIAenII 0.
Proof. (1) Let K: E E be a compact positive operator dominating A,

i.e., (AxJ Kixi for each x e E. Since K is compact and is norm
bounded, we can extract from {K(lenJ)} a convergent subsequence. Without
loss of generality we can assume that the sequence [)} itself converges,
that is, there exists some y e E such that y. By passing to another
subsequence if necessary, we can also assume that < for
each n. Therefore, the series e = I
y is norm convergent and
defines a positive vector. Clearly, e, whence e+
(
J
e + (yj.
(2) Assume that is a disjoint sequence and let {fTh} be a subse-
quence of By part (1) there exists a subsequence {gm} of such
that the disjoint sequence is order bounded. Since E has order

tinuous norm, Theorem 2 .26 guarantees that 0. Thus, we have
shown that every subsequence of has a subsequence that is norm
convergent to zero. This implies that
positive multiplication operator
A on a
where is a compact Hausdorff space, is compact-friendly if and only if the
multiplier has a flat.
Proof. Assume first that a multiplier q5 E is constant

on a non-empty open subset of ft Then, as shown in Theorem 4.20, the
operator commutes with a non-zero positive rank-one operator. This
implies immediately that the operator is compact-friendly.
For the converse, assume that is compact-friendly. Therefore, there
exist three non-zero operators R, K, A: with R, K positive,
K compact, = and such that for each x E we have
jAxI<RlxI and AxI<KIxI.

Taking adjoints, we obtain that = and that
A*x*1 and I
K*1x*1
for each E Exercise 2 in Section 2.4. Now the following three
see
properties can be verified in a rather straightforward way.
(1) For each the support of the measure is contained in the
set = where &i denotes the s-measure concentrated

at the point This follows immediately from the identity
= =
(2) Since A* is dominated by R*, it followsiminediately from (1) that
for each E the signed measure A*&1 is also supported by
(3) Fix any function h E such that IIhII = 1 and Ah 5L 0. Next,
choose a non-empty open set U on which for some
> 0. Then for each E U we have . To see this, note
that
I(A*&j,h)1 = (*)
From the definition of the sets we see that = ft To
complete the proof, assume by way of contradiction that for each E
the set has an empty interior. Then the non-empty open set U, chosen
in (3), must meet infinitely many sets Pick a sequence in U
satisfying for rn n, and let for each n. Then
. Furthermore, since each is supported by the set and
the sequence is pairwise disjoint, the sequence in is also
disjoint. Therefore, since the norm in is order continuous, it follows

from part (2) of Lemma 1ft64 that 0. However, the last conclusion
contradicts (*), and the proof of the theorem is complete.
A similar characterization of compact-friendly multiplication operators

on was obtained in [17]. G. Sirotkin [307] has managed to extend
Theorem 10.65 to arbitrary Banach function spaces.
Exercises
1. Prove that the following operators are all compact-friendly:
(a) compact positive operators,
(b) positive operators commuting with non-zero positive compact
ators,
(c) positive operators that dominate non-zero compact positive opera-
tors,
(d) positive E*_integral operators (in particular, abstract integral opera-
tors), where E is a Dedekind complete Banach lattice,
(e) positive integral operators.
2. Let B: E * E be a positive operator on a Banach lattice. If there exists
a non-trivial Binvariant projection band L\ in E such that Bx = 0 for
each x E L\d, then show that B is compact-friendly.
3. (Sirotkin) For a non-zero positive operator T: E E on a Banach lattice
establish the following.
(a) If K E (E) is a non-zero compact positive operator, then T + K is
compact-friendly.
(b) There exists a operator B E (E) such that T B.
(c) If E is Dedekind complete, then there exists a compact-friendly op-
erator S such that 0 < S < T.
4. Let be a positive one-to-one function in C(c1), where is a compact
Hausdorif space without isolated points. Show directly, i.e., without us-
ing Theorem 10.65, that the multiplication operator Mc1, is not compact-
friendly.
5. A positive operator B: E E on a Dedekind complete Banacli lattice
is called a generalized Harris operator if some power of B is not
disjoint from the band (E* of operators on E, i.e.,
if some power of B dominates a non-zero positive operator.
Recall that if some power of B is not disjoint from the band of
abstract integral operators, then B is called a Harris operator. Of course,
each Harris operator is a generalized Harris operator.
For a generalized Harris operator B: E E establish the following.
(a) B is compact-friendly.
(b) If B is quasinilpotent at a non-zero positive vector, then B has a
non-trivial closed ideal.
6. Show that Theorems 10.55 and 10.60 remain true if we replace the as-
1 1
sumption lim = 0 with limiaf =
7. Let 4 p < co) be a continuous operator with modulus.
(1
Assume that there exists a non-zero positive operator 5: 4 4 such
that:
(a) S commutes with the modulus of T.
(b) S is quasinilpotent at a non-zero positive vector.
Show directly, i.e., without using Theorem 10.24, that T has a non-trivial
closed invariant ideal.
8. (Sirotkin) Let B: E + E be a compact-friendly operator on a Banach
lattice such that no positive operator commuting with B dominates a
non-zero compact operator. Show that for each at most countable family
of positive operators commuting with B there exists a non-trivial closed
ideal that is B-invariant and ,F-invariant.
9. [319]) Show that every essentially quasinilpotent compact-
friendly operator on a Banach lattice has a non-trivial closed invariant
subspace.
10. (Troitsky [319}) Let 5: E E be a quasinilpotent integral operator on
a Banach function space such that S is compact. Then each of the
operators Sj and 5+ has a non-trivial closed invariant subspace.
11. Show that the set of all compact-friendly operators on a Banach lattice
E need not be closed or open in (E).
12. This exercise presents an example of a HubertSchmidt operator on
L2 [0, 1] which is one-to-one, positive, locally quasinilpotent at a posi-
tive vector, but fails to be quasinilpotent. The construction of such an
operator will be done in three steps as follows.
(a) Let 0 a <b < 1 and 0 c 1 [see Figure 1(a)] and consider
the integral operator T: L2 [a, bJ r L2 [c, d] defined by
b
Tx(t)=f sin(st)x(s)ds,
Then the operator T is one-to-one and compact.

(b) Assume that 0 a < b 1. Then the positive integral operator
T: L2[a, b] + L2 [a, b], defined by
b
Tx(t)=f sin(st)x(s)ds, a<t<b,
has a positive spectral radiusand so T is a one-to-one compact

positiveintegral operator which is not quasinilpotent.
(c) Consider the function K: [0, 1] x [0, 1] JR defined as follows: the
value K(s, t) is equal to sin(st) if (s, t) lies in each one of the shaded
rectangles shown in Figure 1(b) and is zero everywhere else. The
points a0, ai, a2,... shown in Figure 1(b) are defined by
a0=1 and for n1.

k=0
The length of the interval is equal to = n!2e

As usual, we shall denote by K the operator on 112 [0, 1] defined be
the kernel K(s, t). That is, the operator K: L2[0, 1] 112 [0, 1] is
defined by
,nl
Kx(t) K(s,t)x(s)ds, 0 t 1.
=j 0
Then the integral operator K: L2[0, 1] 112 [0, ljis a

compact positive operator which is quasinilpotent at some positive
vectors but fails to be quasinilpotent.
t
1
t a1,
a2
a3
1
a b
IT '11'
a3 a2 al
(a) (b)
Figure 1. The kernel of the operator K
10.6. Positive Operators on Banach Spaces with Bases

Some of the results on the existence of invariant subspaces of positive op-
erators on Banach lattices can be extended to Banach spaces ordered by
the cone generated by a basis. This study was initiated in [12] and in this
section we present the main results from that paper.
Recall that a sequence } in a Banach space X is called a Schauder
basis (or simply a basis) of X if for every x E X there exists a unique
sequence of scalars such that x = where the series is
assumed to converge in the norm. Every basis gives rise to a natural
10.6. Positive Operatorson Banach Spaces with Bases 437
closed cone C defined by
C = = : o for each n = 1, 2, .
. .}.
The cone C will be referred to as the (closed) cone generated by the

basis For an extensive discussion concerning the cone generated by a
basis see [305].
As discussed in Section 1.3, with every basis we associate the standard
sequence of coefficient functionals. Let be a basis of a Banach space
X. Then the linear functional defined for each x = by
= aTL,
is a continuous linear functional on X. Observe that each is automatically

positive with respect to the cone generated by the basis The sequence
of continuous linear functionals satisfies
-An operator T: X X on a Banach space with a basis is said to

be positive (with respect to this basis) if T(C) C, where C is the closed
cone generated by
If we fix a basis for a Banach space X, then every continuous
operator T: X X can be identified in the usual manner with an infinite
matrix [tjj}. In this context, we can also say that an infinite matrix
defines an operator on X. Note that an operator T: X X with matrix
is a positive operator if and only if 0 for each pair (i,j).
We are ready to present a version of Theorem 10.24 that holds true for
positive operators on a Banach space with a basis.
Theorem 10.66. Let X be a Banach space with a basis and let T: X X
be a continuous positive operator. If T commutes with a non-zero positive
operator that is quasinilpotent at a non-zero positive vector, then T has a
non-trivial closed invariant subs pace.
Proof. Let be a basis of the Banach space X and let be the

sequence of coefficient functionals associated with the basis
Assume that a non-zero positive operator A: X p X satisfies TA = AT
and is quasinilpotent at a non-zero positive vector Yo, i.e., lim
n+oo
= 0.
If Ayo = 0, then the kernel of A is a non-trivial closed vector subspace that
is invariant under T. Thus, we can suppose that Ay0 is non-zero. Hence,
0 for some k. By an appropriate scaling of Yo, we can assume that
o Xk Yo'
Now let P: X p X denote the continuous projection onto the vector

subspace generated by xk defined by Px = Clearly, 0 Px x
for each 0 x e X. We claim that
PTmAXk = 0 (*)
for each m 0. To see this, fix m 0 and let PTmAXk = axk for some non-
negative scalar a 0. Since P is a positive operator and the composition
of positive operators is a positive operator, it follows that
0 aThxk = (PTmA)nXk = TtmThAThxk TtmThAThy0.
Using the positivity of the linear functional we obtain

0< =
mn
Consequently, 0 *
JJCklf hA you, and so
1 1
0 a
1
From lim = 0, we conclude that a = 0, and thus condition (*)
must hold.
Finally, consider the subspace Y generated by {TmAxk: 7n = 0, 1,..
Clearly, Y is invariant under T and since 0 Axk e Y, we see that Y {0}.
Also, it follows from (*) that for each y E Y we have
= = 0,
and consequently (y) = 0 for all y e Y. The latter shows that Y is a
closed subspace of X invariant under the operator T.
Corollary 10.67r. If T: X k X is a quasinilpotent positive operator on
a Banach space with a basis, then T has a non-trivial closed invariant sub-
space.
For the next result we need to recall the notions of the strong operator
topology on (X, Y). The strong operator topology on (X, Y) is the
Hausdorff locally convex topology 'r8 generated by the family of seminorms
{Px}XEX defined by = JJTxJJ for each T e (X, Y). Notice that a net
{TA} (X, Y) satisfies TA T in (X, Y) if and only if JJTAx k 0
for each x E X.
One can add arbitrary weights to the matrix representing a quasinil-
potent positive operator and still be guaranteed that a non-trivial closed
invariant subspace exists.
Theorem 10.68. Let X be a Banach space with a basis. Assume that a
positive matrix A = defines a continuous operator on X that is quasi-
nilpo tent at a non-zero positive vector. If for a double sequence } of
scalars, the weighted matrix defines a continuous operator B on X,
then the operator B has a non-trivial closed invariant subs pace.
10. 6. Positive Operators on Banach Spaces withBases 439
Proof. Let be a basis of the Banach space X and let be the

sequence of coefficient functionals associated with the basis Assume
that the positive operator A = satisfies urn 0 for some
Ti+oo
non-zero positive vector Yo An appropriate scaling of Yo shows that there
exists some k satisfying 0 < xk If Axk = 0, then an easy argument
shows that BXk = 0, and thus the kernel of B is a closed invariant
subspace (here we assume, of course, that B 0). Thus, we can suppose
that Axk is non-zero.
Now let P be the positive projection on X defined by Px =
Arguing as in the proof of Theorem we can establish that PAmxk 0
for each m 1. In particular, we have = 0. So, for m 1 and
each positive operator S: X p X satisfying 0 S Atm we have
=0. (**)
Next, consider the vector subspace Y of X generated by the set
{Sxk: SEL(X)and O<S<Atm forsome m1}.

Clearly, Y is invariant under each operator R that satisfies 0 < R < A.
Also, it follows from (**) that
for all y E Y. The latter shows that Y is a non-trivial closed vector subspace
of X that is invariant under each operator R: X X satisfying 0 R A.
For each pair (i,j) consider the operator defined by =
and Ajj(Xm) 0 for m j. Since this operator satisfies 0 $ A,
it follows that Y is invariant under each operator Therefore, for each
n E N the vector subspace Y is invariant under the operator
=
i=1 j=1
It remains to notice that the sequence of operators converges in the
strong operator topology to B. Therefore, 13(Y) Y, and thus the operator
B has a non-trivial closed invariant subspace.5
Corollary 10.69. Let X be a Banach space with a basis. that a
positive matrix A defines a operator on X which is quasi-
nilpotent at a non-zero positive vector. If a operator T: X + X
is defined by a matrix T = satisfying 0 whenever = 0, then
the operator T has a non-trivial closed invariant subspace.
5As a matter of fact, Y is a common invariant subspace for all operators generated by the
weighted matrices satisfying the condition of the theorem. Compare this with Exercise 9
in Section 10.4
A further generalization to Banach spaces with a Markushevich basis

can be found in [122]. The paper [49] contains some results on invariant
subspaces of positive operators on general ordered Banach spaces.
Exercises
1. Let be a basis in a Banach space X and let
C= 0 for each n = 1,2,...

}
=
Show that C is a closed cone in X.
2. If a Banach space X has an unconditional basis then show that:
(a) The cone C generated by the basis } is a lattice cone.
(b) The Riesz space (X, C) has weak units.
(c) After an equivalent renorming the ordered Banach space (X, C) be-
conies a Banach lattice with order continuous norm.
3. Let X be a Banach space with an unconditional basis If a linear
operator T: X + X is positive with respect to the cone generated by the
basis then show that T is a bounded operator.
4. (Spalsbury {310J) Consider the Volterra operator defined in Example
Show that there is no basis in L2[0, 1] with respect to which the Volterra
operator is positive. [HINT: According to [168] every non-trivial closed
subspace of L2 [0, 1] which is invariant under the Volterra operator is of
the form = {f L2[0, 1]: f = 0 on [0, a}} for some 0 < a < 1.
Use this result in conjunction with Theorem 10.66.]
10.7. A Characterization of Non-transitive Algebras

Recall that an algebra A of operators on a Banach space is said to be non-
transitive if it has a non-trivial closed A-invariant subspace.
The objective of this section is to present some characterizations
of non-transitive algebras of operators obtained by Y. A. Abramovich,
C. D. Aliprantis, and 0. Burkinshaw [14]. These characterizations are based
on the works of V. I. Lomonosov [212] and L. de Branges [75].
Throughout this section X will denote an arbitrary real or complex
Banach space. As usual, the closed unit ball of X* will be denoted by U*.
That is,
= {x* EX*: IIx*M i}.
We assume that is equipped with its weak* topology. Recall that
a weak* compact subset of X*.
10.7. A Characterization of Non-transitive Algebras 441
Definition 10.70. The vector space of all continuous functions from

into X*, when both U* and are equipped with the weak* topology, will be
denoted by C(U*, X*). Occasionally, C(U*, X*) will also be denoted by Y,
i.e., Y=C(U*,X*).
Clearly, the vector space C(U*, X*) equipped with the norm
= sup lf(x*)I1, f E C(U*7X*),
is a Banach space. The Banach space C(U*, X*) was studied first in [212]
and Observe that for each operator T E L(X) the restriction of the
adjoint operator T* to is an element of C(U*, X*).
V. I. Lomonosov [212], inspired by L. de Branges' proof of the
sical StoneWeierstrass theorem [7'3], characterized the extreme points
of the closed unit ball of the norm dual of C(U*, X*). Subsequently,
L. de Branges [75] presented an analysis of these extreme points and
obtained an abstract version of the StoneWeierstrass Theorem. The
Lomonosovde Branges analysis will be employed later on in our charac-
terization of the invariant subspace problem.
As usual, C(U*) denotes the Banach space of all continuous real-valued
(or complex-valued if X is a complex Banach space) functions defined on
Each function a E C(U*) defines a multiplication operator on C(U*, X*) via
the formula
(af)(x*) = a(x*)f(x*), f E C(U*,X*), E
Clearly,
IIaf II
Every pair (x**, x*) E x gives rise to a linear functional
on C(U*,X*) defined for each f C(U*,X*) by
(f,x** x*) = (x** x*)(f) = (f(x*),x**) = x**(f(x*))
Clearly, is a norm continuous linear functional on C(U*, X*). The

vector space generated by the family {x** E and E U*}
in the norm dual of Y = C(U*, X*) will be denoted by Y', i.e.,
xr E X** and E U* for each i = 1,. . , ri}.

=
Clearly7 Y' separates the points of Y, and so (Y Y") is a dual system.
Besides the norm topology on Y, we shall also consider two other locally
convex topologies. They are the topologies and that are defined as
follows.
(1) The topology on Y is the locally convex topology generated by

the family of seminorms {px**,x*: E X** and E }, where
=
f Y = C(U*, X*), i.e., is simply the weak topology

cr(Y, Y').
(2) The topology r8 on Y is the locally convex topology generated by
the family of seminorms {px*: }, where
(f) =
for fEY=C(U*,X*).
The alert reader will recognize immediately that these topologies are similar
to the usual weak and strong operator topologies on (X) and this justifies
our choice of the subscripts w and s. Moreover, in analogy with the classical
weak and strong operator topologies [118, Theorem 4, p. 477], the topologies
and m have the same continuous linear functionals. That is, they are both
consistent with the dual system (Y Y').
Theorem 10.71. The locally convex topology r8 is consistent with the dual
system (Y,Y'). That is, we have the inclusions
i-8 'r(Y,Y'),
where, as usual, r(Y, Y') denotes the Mackey topology on Y.
Proof. Clearly, a(Y, Y') r8. So, it suffices to establish r8 'r(Y, Y'). To
this end, fix E We must show that the set
{f E Y: = lIf(x*)II 1}
is a 'r(Y, Y')-neighborhood of zero. Consider the operator
R: (X**,cr(X**,X*)) (Y',cr(Y',Y)),
defined by R
0 in X**, then (f, = (f, = (f(x*), 0 for
each f E Y, and so 0. Since is a set,
it follows that the convex circled set D = R(U**) = {x** 0 E }
is a(Y', Y)-compact. So, its polar
D = {f E Y: 0 f)I I for all E U**}
= {f E Y: x**(f(x*))I < 1 for all E U**}

= {fEY: lIf(x*)MI}
is a r(Y, Y')-neighborhood of zero, and the proof is finished.
The standard duality theory result yields the following.

Corollary 10.72. The topologies and on C(U*, X*) have the same
closed convex sets.
We are now ready to establish that if a subspace M of C(U*, X*)

then the elements outside of the closure of M can be sep-
arated by linear functionals of the form
Lemma 10.73. Let M be a vector subspace of C(U*, X*) which is invariant

under multiplication by elements of C(U*). Then a function fo E C(U*, X*)
does not belong to the r3-closure of M if and only if there exist x** E
and x* E U* such that
(x** fo) = I and (x** 0 f) =0
for all f E M.
Proof. The "only if" part needs verification. Therefore, suppose that fo
does not belong to the 'r3-closure of M. Then there exists some 'r3continuous
linear functional on C(U*,X*) such that 0 and = 0 for
each f E M. By Theorem 10.71 we know that the linear functional
is of the form = 0 where for i j. From
= 0, it follows that there exists some k
(4*, fo) =
fying 0. We can assume fo(xfl) = 1.
Now, by Urysohn's lemma, there is some a E C(U*) such that =I
Since afEMforeachf EM, wehave
1)
= f(xfl) = = =0
for all f E M, and the proof is finished. U
Now let A be a The next result gives a necessary and

subalgebra of (X).
sufficient condition for the existence of a common non-trivial
closed subspace.
Theorem 10.74. For an arbitrary subalgebra A of (X), the following two
(1) There exists a non-trivial closed subspace.
(2) There exists an operator B E (X) such that its adjoint operator
B* does not belong to the r3-closure in C(U*,X*) of the vector
space generated by the set {aT*: a E C(U*) and T E A).
Proof. Let M denote the vector subspace of C(U*, X*) generated by the
collection of functions {aT*: a E C(U*) and T E A), and let M denote
the closure of M in the topology 'r3.
(1) By Theorem 10.12 there exist non-zero

(2) and
satisfying (T*x*, x**) = 0 for each T E A Without loss of generality we can
suppose that E U*.
Pick b* E X such that (b*,x**) = 1 and (b,x*) = 1. We
cdaim that the operator B = b* b E satisfies B* To
this, note that
(x** aT*) = a(x*)(T*x* x**) 0
for each T E A and each a E C(U*). That is, the 'r5-continuous linear
functional vanishes on M. On the other hand, we have
(x** x* = (x** b b*) = (b*, x**)(b x*) = 1,
bind so Lemma 10.73 implies that B*
(2) Pick some B E
(1) such that B* M. Since M is invariant
under multiplication by elements of C(U*), it follows from Lemma 10.73
that there exist E X** and E U* such that
(x** B*) = (B*x*, x**) = 1 and (T*x*, x**) 0
f'or all T E A. Since (B*x*,x**) = 1, it follows that 0 and 0,
and by Theorem 10.12 the algebra A* is non-transitive.
Our next goal is to obtain a similar characterization for the invariant
subspace problem in terms of the norm topology. To accomplish this, we
need to introduce a new class of functions.
Definition 10.75. A function f Y = C(U*, X*) is to be completely
continuous if it for the weak* topology on U* and the norm
topology on
The vector of all completely functions of the Banach
C(U*, X*) will be denoted by 1C(U*, X*).
In other words, a function f X*) is completely continuous if

and only if in U* implies f(x*) p 0. It is easy to see
that if T: X Xis a compact operator, then T*: (the restriction

of T* to U*) is completely continuous, i.e., T* E 1C(U*, X*).
Clearly, 1C(U*, X*) is a norm closed subspace of C(U*, X*). As a vector
subspace of C(U*, X*), the space 1C(U*, X*) inherits the three topologies
considered on C(U*, X*); the norm topology, the topology and the to-
'r5. There is no reason to believe that 'rw or r8 is consistent with the
norm topology on 1C(U*, Nevertheless, for subspaces
of 1C(U*, X*) the situation is different. Using the Lomonosovde Branges
technique, we can show that for any subspace of 1C(U*, X*)
its closures in these topologies coincide.
Theorem 1O.r6. For each ant subspace M of 1C(U*,X*) the

(1) M is in /C(U*,X*).
(2) M
is norm closed.
Proof. Clearly, (1) (2) (3). It remains to establish (3) (1).

To this end, let fo E /C(U*, X*) belong to theof M. To show
that fo E M, it suffices to prove that fo belongs to the norm closure of
M. By the HahnB anach theorem, it suffices to verify that for any norm
continuous linear functional b on C(U*, X*) the equality (b, f) = 0 for all
f M implies (b, fo) = 0. So, let (b, f) = 0 for all f E M. We can assume
that / has norm one.
For each a E C(U*), we define on Y = C(U*, X*) the continuous linear
functional by = b(af), where f E Y. Since M is invariant
under multiplication by elements of C(U*), it follows that (ba, f) = 0 for all
f E M. This certainly guarantees ('l', f) 0 for all E V and f E M, where
V is the weak* closed subspace of Y* generated by the functionals We
denote by U the intersection of V with the closed unit ball of Obviously
b E U. Now let be any extreme point of U. By the characterization of
extreme points obtained in [75, Theorem 1], it follows that there exist an
element E and an element E X** such that
= (x**x*,f) = (f(x*),x**)
for every f E /C(U*,X*). In particular, = (f(x*),x**) = 0 for all
f E M and = (fo(x*), x**). Since fo is in the ru-closure of M, it
follows that 'b(fo) (fo(x*),x**) = 0.
That is, we have proved that 0 for each extreme point 7/i' of U.
Since / E U and since, by the KreinMilman theorem, U is the weak* closed
convex hull of its extreme points, we can conclude that b(fo) = 0. This
completes the proof.
We are now ready to improve Theorem 10.74.
Theorem 10.7'?'. For a subalgebra A of (X) the following statements are

equivalent.
(1) dual algebra A* is

The
(2) There exist operators B,K E .C(X) with K compact such that the
operator B*K* does not belong to the norm closure in C(U*, X*) of
the vector subspace generated by {crT*K*: a E C(U*) and TEA}.
446 10. Invariant Su bspaces
Proof. (1) By Theorem 10.12, there exist non-zero

(2) E X** and
E satisfying (T*x*,x**) = 0 for each T E A. We can suppose that
E U*. Now, pick elements b E X and b* E X* such that (x*, b) = 1 and
(b*,x**) = 1.
Next, consider the rank-one operators K = b and B b* b, and
note that K*x* = (x*, b)x* = and B*x* = (x*, b)b* = b*. Therefore,
B*K*x* = b*. We claim that B*K* does not belong to the norm closure
of the space M generated by the set {aT*K*: a E C(U*) and T E A} in
the space Y.
To see this, observe first that the norm continuous linear functional
b = on C(U*, X*) satisfies
(b, aT*K*) = a(x*)(x**,T*K*x*) = a(x*)(T*x*,x**) =0
foreach T E A and each a E C(U*). That is, b vanishes on M. On the
other hand, cb(B*K*) = (B*K*x*,x**) = (b*,x**) = 1 and this guarantees
that B*K* M.
(2) (1) Assume that B, K E (X) satisfy the stated properties. Let
denote the norm closure in C(U*, X*) of the vector space M generated by
the set {aT*K*: C(U*) and T E A}. Clearly, M is
a E
B*K* belongs to 1C(U*,X*) and M 1C(U*,X*).
Since B*K* does not belong to the norm closure of M, it follows from
Theorem 10.76 that B*K* is not in the of M in 1C(U*, X*).
Therefore, by Lemma 10.73, there exist E X** and E satisfying
and
for all T E A. The former condition implies that K*x* 0, and hence the
latter condition shows that Theorem 10.12 is applicable to A*. Thus, the
algebra A* is non-transitive.
The preceding arguments also allow us to obtain a necessary and suffi-

cient condition for the existence of a non-trivial closed A-invariant subspace.
The only difference is that we must apply the compact operator on the left,
as opposed to the multiplication on the right in the preceding theorem.
Theorem 10.78. For a subalgebra A of (X) the following statements are
equivalent.
(I) A is non-transitive.
(2) There exist operators B, K E with K compact such that the
operator K*B* does not belong to the norm closure in C(U*, X*)
of the subspace generated by {aK*T*: a E C(U*) and T E A}.
Proof. (1) (2) By Theorem 10.12 there exist non-zero x E X and

satisfying (x*,Tx) = 0 for each T E A. We can suppose that E
Take any b* E X* such that (x, b*) = 1, and then consider the
one operator K = b* 0 x E (X). Kx = (x, b*)x = x.
Observe that
Next choose any element b in X such that (x*, b) 1 and then define the
operator B = 0 b E (X). Clearly, the adjoint operator B* satisfies
b*
B*x* = (b,x*)b* = b*.
The vector subspace generated by {aK*T*: a E C(U*) and T E A}
will be denoted by Al. We claim that the operator K*B* is not in the norm
closure of Al. To see this, note that the linear functional b = x is norm
continuous on C(U*, X*) and satisfies
a(x*)(x, K*T*x*) = x*)
a(x*)(Tx,x*)=O
for each T E A and each a E C(U*). That is, = x0 vanishes on Al.
On the other hand, the equality B*x* = b* yields
= (x, K*B*x*) = (x, K*b*) (Kx, b*) = (x, b*) = 1,
and this shows that K* B* does not belong to the norm closure of Al.
(2) Assume that the operators B and K satisfy the stated proper-
(1)
ties. Again, let Al denote the vector space generated in C(U*, X*) by the
collection of functions a E C(U*) and T E A}. Clearly, Al is
K*B* belongs to 1C(U*, X*) and Al IC(U*, X*).
Since the compact operator K*B* does not belong to the norm closure
of Al, it follows from Theorem 10.76 that K*B* is not in the of
Al in 1C(U*, X*). Consequently, by Lemma 10.73, there exist E X** and
E U* such that
(K*B*, 0 = (K*B*x*, = 1 and (K*T*x*, = 0
for all T E A. Since (K*B*x*, x**) = 1, we see that xO = K**x** 0.
Moreover, the compactness of K implies that xO = K**x** E X. So, for

each T E A, we have
(x*,Txo) = (x*,TK**x**) = (T*x*,K**x**) = (K*T*x*,x**) = 0,
and consequently Theorem 10.12 guarantees that the algebra A is non-
transitive. The proof is complete.
We close the section by emphasizing the following remarkaNe fact that

has been present throughout our discussion on the invariant subspace prob-
lem. As soon as we ask about the existence of a A-invariant
closed subspace of X (or about the existence of a non-trivial
closed subspace of X*), a compact operator seems to emerge from nowhere
though the given algebra A were not connected with any compactness
erties in any way from the outset!
Exercises
1. If 11 is a compact topological space and X is a Banach space, then
show that the space C(11,X) of all X-valued continuous functions on
equipped with the norm
If 1 = max If (w) II
is a Banach spacewhich is a Banach lattice if X is a Banach lattice.
2. Let X be a Banach space, and let the closed unit ball U* of X* be
equipped with the Establish the following.
(a) If T: X X is a compact operator, then T*: X* (the re-
striction of T* to U*) is a completely continuous function.
(b) The collection of all completely continuous functions X*) is a
norm closed subspace of C(U*, X*), and so it is a Banach space in
its own right.
3. Assume that for a subalgebra A of (X) there exists a non-zero operator
T0 E (X) such that:
(a) T0 commutes with each operator in A.
(b) For some B E (X) its adjoint operator B* is not in the re-closure of
the vector subspace spanned by a E C(U*) and T E A}
in C(U*,X*).
Show that A* is non-transitive. [HINT: By Lemma 10.73, there exist two
vectors E and E U* satisfying
(x**x*, B*) = x**(B*x*) 1 and

T c A. If = 0, then the relation (T*y*, = 0 for each
T c A coupled with Theorem 10.12 shows that A* admits a
closed subspace. Now assume that = 0, and let N
denote the null space of Since T0 commutes with every member of A,
it follows that N is an closed subspace. From (B*x*) = 1,
we infer that 0. Since E N, we see that N is non-zero. On the
other hand, if N = X*, then = 0, and so T0 = 0, a contradiction.
Therefore, N is a non-trivial closed subspace.
4. Recall that a Banach space X is said to have the Schur property if
implies 0.
If X is a Banach space with the Schur property and A is a commuta-
tive subalgebra of (X), then show that the dual algebra A* of (X*) is
non-transitive. [HINT: If each element in A is a scalar operator, then the
conclusion is trivial. So, we can assume that there exists some non-scalar
operator T E A. Fix some A0 in the approximate point spectrum of T
and consider the operator T0 = T Also, select a sequence of
10.8. Comments onthe Invariant Subspace Problem 449
unit vectors in X satisfying

= 0. (*)
Clearly, T0 commutes with every member of A. Next, consider the

space M of C(U* , X*) generated by : a C(U*) and T A}.
We claim that the identity operator : (restricted to U*) does
not belong to the r8-closure of M . To see this, assume by way of contra-
diction that 1* belongs to the of M. Therefore, if 0 and
E are fixed, then there exist E C(U*) and E A (i = 1, .. m),
satisfying 11[I* <e. This implies
(x,x*) - <6
for all x E X with tIxll 1. Taking into account that A is a commutative

algebra, it follows from (**) that (x, x*) aj(x*) (T0x, <E.
In particular, for each ri we have x x )1 <E,
which, in view of (*), yields x*) < for all sufficiently 'arge n. In
other words, the sequence converges wealdy to zero in X. Since X
has the Schur property, it follows that 0, contrary to =1
1
for each n. Therefore, the identity operator 1* does not belong to the
r8-closure of M and the conclusion follows from the previous exercise.]
5. (Honor [156]) If A is a commutative algebra of bounded operators on
then show that the dual algebra A* of operators on is non-transitive.
10.8. Comments on the Invariant Subspace Problem

The existence of non-trivial closed invariant subspaces for a given operator
presents the opportunity to analyze the structure of the operator by studying
its restrictions to the invariant subspaces. In particular, if an operator on
a Banach space has a reducing pair, then the operator can be written as
a direct sum of two operators and, as expected, the two summands might
provide some crucial information about the action of the operator. This
rationale is behind the importance of the invariant subspace problem that
has been studied so extensively over the last eighty years.
In this chapter, we have presented a number of invariant subspace re-
sults which clearly reflect the authors' interest in the invariant subspace
problem. There are, however, many other directions regarding the invariant
subspace problem that lie outside the scope of this monograph. In this sec-
tion, we shall highlight several remarkable invariant subspace results that
were considered as major breakthroughs at the time of their discovery.
We start by mentioning again that the first example of a bounded oper-
ator on a separable non-reflexive complex Banach space without non-trivial
closed invariant subspaces was constructed by P. Enflo [121] Enfl&s con-

.
struction was simplified considerably by B. Beauzarny [53]. C. J. Read [275,

276, 277] presented examples of bounded operators (some of which are
quasinilpotent) on the real without non-trivial closed invariant subspaces.
An example of a continuous operator on a nuclear Frchet space without
non-trivial closed invariant subsp aces was constructed by A. Atzmon [47].
It is still an open question whether a bounded operator on a separable re-
flexive Banach space (in particular, on a separable Hilbert space) admits a
non-trivial closed invariant subspace.
The first result that we would like to mention here deals with linear
isometries. Let T: X + X be a linear isometry on a Banach space. The
range R(T) of T is obviously closed and, therefore, if T is not surjective,
then R(T) is a closed T-hyperinvariant subspace. For surjective
isometries the conclusion remains true according to the following theorem,
that is basically due to E. R. Lorch [214]. Recall that an invertible operator
T: X X is said to be doubly power bounded (or a Cesaro operator)
if supflEZ < oo. It is obvious that each surjective isometry is doubly
power bounded.
Theorem 10.79 (Lorch). Every non-scalar doubly power bounded operator
on a complex Banach space has a closed hyperinvari ant subspace.
J. Wermer [331] produced an exceptionally strong generalization of

Lorch's theorem for the class of non-quasianallytic operators. An invert-
ible operator T: X X on a Banach space is called quasianalytie if
= oo. Accordingly, the invertible operator T is non-
quasianalytic if
nih
(*)
fl 00
Clearly, each doubly power bounded operator is non-quasianalytic.
Theorem 10.80 (Wermer). If T is non-quasianalytic and its spectrum

is not a singleton, then T has a non-trivial closed hyperinvariant sub-
space.
This theorem generated an avalanche of subsequent extensions and

provements. In this directions we mention the papers by Ju. I. Ljubi and
V. I. Macaev [210], E. M. Dynkin [119] and A. Atzmon [48]. It is still
unknown if Theorem 10.80 is true for non-quasianalytic operators with sin-
gleton spectra. We have some reasons to believe that Wermer's theorem is
true for non-quasianalybic positive operators.
For operators with a norm estimate on the norms of their powers (in
particular, for doubly power bounded operators) Wermer's theorem is true.
Theorem 10.81. If the spectrum of an invertible operator T E (X) is a

singleton and there is a k E N such that ItTtm ii & for each m E Z, then T
has eigenvalues.
A much sharper result than Theorem 10.81 was obtained by A. Atz-

mon [46]. The second spectacular result that we shall discuss next is also
related to Wermer's theorem. Let I' be the unit circle in the complex plane
and let be the normalized Lebesgue measure on F. It was shown by
A. M. Davie [100] that for each 1 p CX) the weighted composition oper-
ator (called the Bishop operator) T: defined by the
formula
= + a), 0 27r,
is non-quasianalytic provided that the weight w E is a Lipschitz func-
tion on I' and is an irrational number which cannot be approximated "too
fast" by rational numbersin the sense that is a non-Liouville number.6
Using some additional ingenious arguments, A. M. Davie [100] also proved
the following.
Theorem 10.82 (Davie). If is not a Liouville number, then every Bishop
operator T: defined by
= + a), 0
has a non-trivial closed hyperinvariant subspace.
This result of Davie was refined further by A. L. Gromov [140] and by

G. W. MacDonald [227].
There are a few major recent generalizations of Lomonosov's Invariant
Subspace Theorem to algebras and semigroups of compact operators on
Banach spaces. A very important breakthrough was obtained in 1984 by
V. S. Shulman [301] who proved the following result for Volterra algebras.7
Theorem 10.83 (Shulman). Each non-zero Volterra algebra has a non-
trivial closed hyperinvariant subspace.
Ittook more than a decade to generalize this result to Volterra semi-

groups of operators. This significant contribution has been accomplished by
Y. V. Turovskii [324] who proved the following.
6 Recall that a real number u is said to be a Liouville number if there exists a sequence
of pairwise distinct rational numbers such that (mn, = 1 and [u I< holds
for each ri. For instance, the real number u = 0.110001. is a Liouville number.
Every Liouville number is transcendental.
7A collection of operators on a Banach space is said to be a Volterra collection if each
operator in the collection is compact and quasinilpotent.
Theorem 10.84 (Throvskii). Each non-zero multiplicative Volterra semi-

group has a non-trivial closed hyperinvariant subspace.
The major step needed to prove Theorem 10.84 was the verification that
the algebra generated by a multiplicative Volterra semigroup remained like-
wise a multiplicative Volterra semigroup. This verification is very technical
and uses several subtle ideas. After that, the existence of the
variant subspace is a simple consequence of Shulman's Theorem 10.83. For
a complete exposition of Theorem 10.84 we refer the reader to the excellent
monograph by H. Radjavi and P. Rosenthal [273J.
Recall that if X is a finite dimensional Banach space, then the classical
Burnside theorem states the following.
Theorem 10.85 (Burnside). If X is a finite dimensional complex Banach
space, then each proper subalgebra A of (X) is non-transitive.
For a proof of this theorem see [273, Section 1. 2}. It is worth pointing
out that Burnside's theorem is not true for finite dimensional real Banach
spaces. To see this, let X = 2, and take any non-scalar operator T
on X without a non-trivial hyperinvariant subspacewe refer to Exercise 2
in Section 10.1 for the construction of such an operator. Consider the algebra
A = {T}'. It is transitive, since T does not have a non-trivial hyperinvariant
subspace. At the same time a straightforward verification shows that this
algebra A is necessarily proper, since T is a non-scalar operator.
We will mention two elegant ways of extending the Burnside theorem to
infinite dimensional Banach spaces. The first result is due to G. L. Litvinov
and V. I. Lomonosov [209] and it is true for locally convex spaces as well.
Notice that while discussing the transitivity of a subalgebra A of (X) we
can always assume that A is weakly closed, that is, closed in the weak
operator topology (otherwise we can replace A by its weak closure which
remains non-transitive if A were non-transitive).
Theorem 10.86 (LitvinovLomonosov). Let X be a complex Banach space
and let A be a weakly closed subalgebra of (X). If A is transitive and
contains a non-zero compact operator, then A =
If X is a finite dimensional Banach space and A is a proper subalgebra

of (X), then A is certainly closed and each operator in A is compact.
Therefore, A cannot be transitive, that is, Theorem 10.86 does imply the
Burnside theorem.
According to Theorem 10.12, Theorem 10.85 is equivalent to the exis-
tence of a non-zero x E X and a non-zero e such that (Ax,x*) = 0
for each A E A. It was this form of the Burnside theorem that was used by
V. I. Lomonosov [212] to produce another generalization to infinite dimen-

sional Banach spaces.
Theorem 10.87 (Lomonosov). Let X be a complex Banach space and let
A be a proper weakly closed subalgebra of (X). Then there exist a non-zero
e X** and a non-zero E X* such that
< IIAiIess
for each A e A, where denotes the essential norm of the operator A.
If X is finite dimensional, then lIAlless = 0 and so Theorem 10.87 does
imply Burnside's theorem. The reader is referred to [212] and to [90] for
proofs and some interesting additional developments.
V. I. Lomonosov also proposed in [212] the following conjecture regard-
ing the invariant subspace problem.
o The adjoint of a bounded operator on a Banach space has a non-
A Read operator T e without non-trivial closed invariant subspaces
seemed again to be a candidate for disproving this conjecture. However,
V. G. Troitsky [320] has shown that this operator is not co)-continuous,
and therefore T is not the adjoint to any bounded operator on c0. Moreover,
T. Schlumprecht and V. G. Troitsky [295] have recently shown that T is
not the adjoint of any bounded operator on any predual of 4.
We close the section with a few results dealing with operators on Hilbert
spaces. An excellent introductory book on Hubert space operators is [94].
Let H be a complex Hubert space. Recall that the Hubert space adjoint
of an operator T e (H) is the unique operator T* e (H) that satisfies
the identity
=
for all x, y E H.8 An operator T E (H) is said to be normal if it
commutes with T*, i.e., TT* = T*T. From the well-known classical theorem
of J. von Neumann [255] on the spectral decomposition of normal operators,
we have the following invariant subspace result. (For details see [280, p. 286]
and [317, p. 430].)
Theorem 10.88. Every non-scalar normal operator has a reducing hyper-
invariant subspace.
An operator A on a Hilbert space H is called subnormal if there exists
a Hilbert space 7-1 such that H can be identified with a closed subspace of
8The notation T* for the Hubert space adjoint of the operator T is standard in Hubert space
theory and although there is a conflict with the notation of the Banach space adjoint of T, we
hope that it will cause no problem
454 -- - 10. Invariant Sn bspaces
7-1 and there exists a normal operator B on N that leaves the subspace H
invariant and A = BIH, that is, the restriction of B to H conicides with the
initial operator A.
Regarding subnormal operators, S. Brown [76] proved the following re-
markable resultthat was viewed as a major achievement following a long
period of partial successes by many mathematicians working on the invariant
subspace problem for Hilbert space operators.
Theorem 10.89 (S. Brown). Every subnormal operator has a non-trivial
closed invariant subspace.
The following interesting invariant subspace result was proved by B. Sz.-

Nagy and C. Foias [249] and independently by L. de Branges [74].
Theorem 10.90 (Sz.-NagyFoiasde Branges). If a non-scalar operator T
in L(H) is power bounded such that neither nor (T*)fl converges strongly
to zero, then T has a non-trivial closed hyperinvari ant subspace.
We conclude the section by mentioning the important class of quasitri-
angular operators. An operator T E (H) is said to be quasitriangular if
there exists a sequence of orthogonal projections on H that converges
strongly to the identity operator and
lim
n+cxD
- 11=0.
These operators were introduced by P. R. Halmos [147], and it seemed for
a while that for quasitriangular operators the invariant subspace problem
might be more manageable. However, a remarkable result by C. Apostol,
C. Foias, and D. Vcdculescu [39] has established just the opposite.
Theorem 10.91 (ApostolFoiasVoiculescu). Every non- quasitriangular
operator T E (H) has a non-trivial closed invariant subspace.
From Theorem 10.91 it follows that if every quasitriangular operator on
a Hilbert space H has a non-trivial closed invariant subspace, then the in-
variant subspace problem has an affirmative solution for Hilbert spaces, that
is, each bounded operator on H has a non-trivial closed invariant subspace.
Chapter 11
The Daugavet Equation
In 1963, L K. Daugavet [98] proved that each compact operator T on the

Banach space C[O, 1] satisfies the equation
It turns out that various classes of operators on many other Banach spaces
satisfy this equation, which is known today as the Daugavet equation.
In 1965, C. Foias and I. Singer [126] extended Daugavet's result to arbi-
trary atomless and to several classes of operators (including
the weakly compact ones), and in 1966 G. Ya. Lozanovsky [219] established
that compact operators on L1 [0, 1] satisfy the Daugavet equation.
During the next fifteen years dr so these results were left without much
attention approximately until the beginning of the 1980's. At that time a
new wave of interest in this topic surfaced and the Daugavet equation has
been studied by many authors in various contexts. The major results of
these studies will be reported in this chapter. The introduced techniques
vary from purely order-theoretical to analytical and we intend to familiarize
the reader with the majority of these methods. However, this has its price.
Our presentation is not as efficient as it could be, since we prove several
basic results repeatedly to demonstrate different methods. For the most
part these results deal with various classes of operators on and L1 (ii;).-
spaces. These spaces are Banach lattices and we have an option to utilize
their order structures.
We hope that the resulting comprehensive treatment of the currently
available methods on the subject is an acceptable trade-off for the occasional
redundancy of a few proofs. The last section of the chapter demonstrates
how the Daugavet equation, a purely isometric property, can be used to
455
456 -- 11. The Daugavet Equation
obtain some topological conclusions regarding Banach spaces. For instance,

a Banach space cannot have an unconditional basis provided each finite-rank
operator on this space satisfies the Daugavet equation.
11.1. The Daugavet Equation and Uniform Convexity

The purpose of this section is to study the basic properties of operators that
have the Daugavet property. The material in this section is taken from [6].
Definition 11.1. A continuous operatorT: X + X on a J3anach space is
said to satisfy the Daugavet equation provided
It should be clear that a bounded operator T E (X) satisfies the Dau-

gavet equation if and only if JjI + TjJ 1+ The following result should
also be obvious.
Lemma 11.2. A bounded operator satisfies the Daugavet equation if and

only if its norm adjoint likewise satisfies the Daugavet equation.
Lemma 11.3. If the norm of a bounded operator T: X + X on a J3anach
space belongs to the spectrum of T, i.e., 11Th E cr(T), then T satisfies the
Daugavet equation.
Proof. Assume fT II E Since the spectrum of T lies inside the disk

with center at zero and radius 11Th, it follows that is a boundary point of
the spectrum. In particular, E the approximate point spectrum
of T. Pick a sequence of vectors in X such that I for each n
and limn 0. Now note that
fI+TjI
+ -fJ -
=
Taking the limit, we see that 11 + TJJ I + 11Th, and so T satisfies the
Daugavet equation.
We continue with a simple geometric property of normed spaces which

guarantees that if an operator satisfies the Daugavet equation, then so does
each positive multiple of the operator.
Lemma 11.4. If two vectors u and v in a normed space satisfy the equality
Jfu f ziJJ = huh j
then + /3v11 = + ,1311v11 for all ,6 0.

Proof. Assume that the vectors u and v satisfy lull + and let
a, ,8 0. By the symmetry of the situation, we can assume that ,8 0.
Then we have
= jcE(u+v)
-
= + Ilvil) - -
=
and the desired equality follows.
Corollary 11.5. If a bounded operator T on a Banach space satisfies the

Daugavet equation, then for each 0 the operator cET likewise satisfies
the Daugavet equation.
The next lemma describes two equivalent norm conditions the first of
which is usually taken for the definition of uniform convexity. As usual, we
denote by (Ix the closed unit ball of a Banach space X.
Lemma 11.6. For a Banach space X the following are equivalent.
(1) For each 0 < 2 there exists some 0 < (5 < 1 such that if
x,yE satisfy IIxyIl , then <1(5.
(2) If two sequences and in satisfy = 2,
then IIxn 0.
Proof. (1) (2) Let two sequences and belong to (Ix and
satisfy IIxn + 2. Now fix 0 < 2 and then pick some
0 <(5<1 that satisfies (1). Next, choosesomen0 suchthat > i(5
for all n ? n0. But then it follows from (1) that < for each
n i.e., = 0.
(2) (1) If (1) were not true, then there would exist some 0
and (Ix with > and > 1 This implies
lIxn + = 2 and hence, by (2), we have = 0,
contrary to > for each n.
We are now ready to introduce the notions of uniformly convex and

uniformly smooth Banach spaces.
Definition 11.7. A Banach space that satisfies either one of the equivalent
statements of Lemma 11.6 is said to be uniformly convex (or uniformly
rotund).
Definition 11.8. A Banach space is called uniformly smooth if for each
0 there is some (5 > 0 such that fJxII 1, 1, and lix < 6
imply tx + llxIt + Mx yll.
458 11. TheDaugavet Equation
Uniform convexity and uniform smoothness are dual notions.

Lemma 11.9 (Smulian [308]). A Banach space is uniformly convex if and
only if its norm dual is uniformly smooth.
Similarly, a Banach space is uniformly smooth if and only if its norm
dual is uniformly convex.
Fora proof of the preceding result see Exercises 13 and 14 at the end
of the section. For more information about uniformly convex and uniformly
smooth Banach spaces see [102, 105, 118] and {189, pp. 353369].
For uniformly convex Banach spaces the converse of Lemma 11.3 is true.
Theorem 11.10. If X is uniformly convex and T E (X) satisfies the
Daugavet equation, then 11Th E cT(T).
Proof. Assume that X is uniformly convex and that T E (X) is

zero and satisfies the Daugavet equation. By Corollary 11.5, the operator
S= also satisfies the Daugavet equation, i.e.,
111+811 = sup = =2.

IlxII=1
Thus, there exists a sequence of unit vectors satisfying

lim
ntoo
+ 2.
Since X is uniformly convex, we infer that limn_+oo = 0.

sequently, = 0 or IITIIxnII = 0. This
shows that 11Th is in the approximate point spectrum of T.

Since an operator satisfies the Daugavet equation if and only if its ad-
joint satisfies the Daugavet equation and since the spectrum of an operator
coincides with that of its adjoint (Theorem 6.14), we have the following
consequence of Theorem 11.10.
Corollary 11.11. A on a uniformly convex or on a
continuous operator T
uniformly smooth Banach space satisfies the Daugavet equation if and only
if its norm 11Th lies in the spectrum o'(T).
From Theorem 7.11 we know that the non-zero points of the spectrum
of a strictly singular operator on a Banach space are eigenvalues. Therefore,
the condition that the norm of an operator is an eigenvalue characterizes
the strictly singular operators that satisfy the Daugavet equation.
Corollary 11.12. A strictly singular (in particular, compact) operator T
on a uniformly convex or a uniformly smooth Banach space satisfies the
Daugavet equation if and only if its norm lIT J is an eigenvalue of T.
It is well known that every with 1 <p < oc is uniformly convex;

see Exercise 16 at the end of the section. Therefore, Corollary 11.12 applies
and it characterizes the strictly singular operators that satisfy the Daugavet
equation on for 1 <p <00.
Corollary 11.13. A strictly singular operator T: where
1 <p < oc, satisfies the Daugavet equation if and only if its norm is
an eigenvalue of T.
For the boundary values p = 1 and p = oc the situation is drastically

different and it will be discussed in Section 11.2. As we shall see, on atomless
L1- and each compact operator satisfies the Daugavet equation.
Theorem 11.14. Assume that a continuous operator T: X X on a
uniformly convex or a uniformly smooth Banach space satisfies the Daugavet
equation and let f(A) = a power series with 0 for each
n.Iff(1ITII) <oc, then the continuous operatorf(T) satisfies the Daugavet
equation and Jf(T)1I = f(IITII), i.e., we have
Ill + =1+ =1+
In particular, we have the following consequences:
(1) For each m = 0, 1,2,... the operator satisfies the Daugavet equa-
tiori and IITThII = i.e.,
1I+TThI1 = 1+ ITThU 1 + ITIITh.
(2) For any polynomial p(A) = ao + a1A . . + with non-negative

coefficients, the operator p(T) satisfies the Daugavet equation and
IIp(T)1I pGITII).
Proof. The conclusions of the theorem follow easily from the Spectral
Mapping Theorem. However, we prefer to present an independent and
direct proof. Assume that the Banach space X, the continuous operator
T: X X, and the function f(A) satisfy the stated properties. By
ollary 11.11, we know that there exists a sequence } of unit vectors
satisfying MTxn 0. Using the identity
+
and an easy inductive argument, we see that
lim 0
flp 00
for each k = 0, 1,2 Next, we claim that
lim =0.
460 11. The Daugavet Equation
To see this, let 0. Fix some m such that 11Tht < and then
pick some such that II
< holds for all m no.
Therefore, for n no we have
- f(IITI1)x4= -
m 00
i=O i=m+1
<+2
i=m+1
which shows that = 0. Consequently, the
real number 1(11Th) lies in the approximate point spectrum of f(T). Since
111(T) 1 f( 11Th) and the spectrum of f(T) lies inside the closed disk with
center at zero and radius 111(T) we infer that the equality 111(T) II = f( ITH)
must hold. Now Corollary 11.11 guarantees that the operator f(T) satisfies
We now turn our attention to operators acting on locally uniformly con-

vex Banach spaces. To discuss this notion, we need the next result whose
proof is similar to that of Lemma 11.6 and is left as an exercise.
Lemma 11.15. For a unit vector x in a Banach space X the following
(1) For each 0 < 2 there exists some 0 < < 1 such that lvll 1
imply <1s.
(2) If C Ux and lim = 1, then lim = 0.
We are ready to define the locally uniformly convex Banach spaces.

Definition 11.16. A Banach space is said to be locally uniformly convex
at a unit vector x 'if it satisfies either one of the equivalent statements of
Lemma 11.15.
A locally uniformly convex Banach space is a Banach space that is
locally uniformly convex at every unit vector.
It is well known that there exist locally uniformly convex Banach spaces
that are not uniformly convex. Therefore, it is natural to ask whether the
characterization of operators satisfying the Daugavet equation on uniformly
convex Banach spaces given in Corollary 11.11 remains valid for locally uni-
formly convex spaces. Surprisingly, as the next example shows, the answer
is negative.
Example 11.17. Fix a sequence R such that 1 and

> I for each m; for instance let = I+ For each m let =
That is, is R2 equipped with the = II + ylPm)
Clearly, is a reflexive uniformly convex Banach lattice. If = (1, 0)
and = (0, I), then
I and + =
For each m define the positive operator Tn: + by
Tn(X,y) = (y,O).
Observe that Tn(Un) = 0, Tn (va) = and IITn IL = 1 for each ri,.
Next, consider the Banach lattice X = (X1 that is, X
is the 2-sum of the Banach lattices and define the positive operator
T(xi,x2,...) = (Tixi,T2x2,...).
The conclusion that the operator T and the space X have the desired
properties will be obtained from the following three statements.
(1) The Banach lattice X is reflexive, fails to be uniformly convex but
is locally uniformly convex.
(2) The operator T satisfies the Daugavet equation.
(3) The norm = 1 does not belong to the spectrum of T.
The reflexivity of X follows from the identity
established in Exercise 3 of Section 7.2. To see that the Banach space X is

not uniformly convex, let
xn=(0,...,0,un,0,...) and (*)
where the vectors and occupy the coordinates. Then we have

IlXn 112 = = I. Moreover,
12112 = = 2' p and
ilxnynlI2
Therefore, X is not uniformly convex.
The local uniform convexity of X follows from the well-known fact that
the 2-sum of a sequence of locally uniformly convex Banach spaces is locally
uniformly convex; see, for instance [218]. Now we shall verify statement (2).
If x = (x1, x2,...) E X satisfies 11x112 1, then note that
00 00 00
=> = <1,
and so 1. On the other hand, if is the unit vector defined in (*),

then
11Th = = IIUrLIIpn = 1,
and so 1. Hence, = 1.
Similarly, if + denotes the identity operator on then
lII+TII =
= + = + =
From = 2, it follows that jjl + TJj = 2, and so T satisfies the

Daugavet equation.
Finally, note that T2 = 0, i.e., the operator T is nilpotent, and so
cr(T) = {0}. Hence, = I does not belong to the approximate point
spectrum of T.
The situation for locally uniformly convex Banach spaces is not as bad
as one might expect from the previous example. For compact operators on
locally uniformly spaces the conclusion of Corollary 11.12 remains true:
Theorem 11.18. A compact operator T: X + X on a locally uniformly
convex Banach space satisfies the Dan gavet equation if and only if its norm
is an eigenvalue of T.
Proof. The "if" part is true for each Banach space and is independent
the compactness of the operator. Indeed, if JjTjj is an eigenvalue of T, then
by Theorem 11.10 we get + TII = 1 + 11Th.
For the converse, assume that T is non-zero and satisfies the Daugavet
equation. By Corollary 11.5, we know that the compact operator S
also satisfies the Daugavet equation, and so sup11x11=1 jx+Sxlj = 1+11811 = 2.
Pick a sequence of unit vectors such that
lim + = 2. (**)
Using the compactness of S (and passing to a subsequence if necessary), we

can assume that xIJ 0 holds for some x e X. From
+ + I + 2
and (**), we infer that lix 11 = 1

= 1. Furthermore, from the
inequality
<2,
we see that + = 2, or = 1. By the
cal uniform convexity of X, we conclude that = 0. So,
liSx Sx = x, or equivalently = x.
Thus, Tx = which shows that ITIL is an eigenvalue of T.
Exercises
1. Show that if the vectors . . , in a normed space satisfy
=
then = holds for each choice of non-negative
scalars a1, . , [Hint: The proof is by induction. For rt = 1, the
claim is trivially true. So, assume the claim true for some n and let
the vectors , . satisfy it Ski = 15k Also, let ii.
a1,. . ,. be non-negative real numbers; we can suppose a1 ak

for each k. From
= IIxiII +
it follows that Ski = Now notice that the identity
n+1
= 5k] >(ai ak)xk,
in conjunction with the induction hypothesis, yields
n+1
fl>(ai ak)xkfl
= >(al-ak)IIxkII
2. This exercise presents an isomorphic version of Lemma 11.4. Let X be a

normed space and assume that a vector u E X and some unit vector v
satisfy + vu 2 E for some E > 0. Show that for all a, fi E IR+ we
have
iIau+f3vII (1E)(a+fl).
3. Let V be a vector subspace of a normed space X, and assume that there
exist a vector u E X and some 0 such that lu + 2 E holds for
each v E V with lvii = 1. Show that for each v E V and all scalars we
also have + vii (1 E) + lvii].
4. (0.-S. Lin [202]) Assume that two sequences and in a normed

space satisfy
lim + + 0
0 a
0
+/3JJvThM)
+ +
= aIIUTh + 0.]
5. (C.-S. Lin [202]) Generalize the preceding exercise as follows. Assume

that rn sequences } (k = 1,2,.. , rn) in a normed space satisfy
.
m m
lim
fl+ 00
=0.
k=1 k=1
Then show that ak = 0 for any

choice of non-negative scalars a1,.. . ,
6. (C.-S. Lin [202]) Generalize Theorem 11.10 by showing that for a uni-
formly convex Banach space X and two operators S, T E (X) the fol-
lowing statements are equivalent.
(a)
(b) There exists a sequence } of unit vectors such that
lim
fl
=1+ + 11Th.
7. (C.-S. Lin [202]) If T: X > X is an invertible operator on a uniformly

convex or uniformly smooth Banach space, then show that I
implies 11 <1. [HINT: If 11 E a(I 2T), then the operator
I (I 2T) = 2T is not invertible, contradicting our assumption that
T is invertible. So, I 2T does not satisfy the Daugavet equation, and
hence
= = II+(I2T)H =2.
This implies TIf < 1.1
8. Let T: X > X be a non-zero bounded operator on a uniformly convex or

uniformly smooth Banach space. Show that T is non-invertible if
and only if T satisfies the Daugavet equation. [HINT: Note that T
is non-invertible if and only if E
9. If a continuous operator T: X > X on a uniformly convex or uniformly
smooth Banach space satisfies the Daugavet equation, then show that for
each a 0 the continuous operator eaT satisfies the Daugavet equation
and that JCaTM = eaIlTiI holds.
10. Let E, be an arbitrary cr-finite measure space. A bounded operator

T: is called symmetric if
(Tf,g) = (f,Tg)
for all f,g E L2(14, where (.,.) denotes the usual inner product in
defined by (f, g) = Establish the following.
(a) Symmetric operators are bounded.
(b) An integral operator T on with a kernel K E x is
symmetric if and only if K(s, t) = K(t, s) holds for x all
(s,t) E X
(c) The integral operator T: L2 [0, 11 L2 [0, 1], defined by
p1
Tx(t) = / sin(st)x(s) ds,
Jo
is symmetric.
(d) Every symmetric operator T: satisfies r(T)
(e) Every symmetric positive operator T: L2 (ii) satisfies the
Daugavet equation, i.e., 1I+T11 1 + IITLI.
[HINT: For (c) note that 112 = (Tf,Tf) = (T2f,f) 112
implies or 1T112. Tk is also a

Since
symmetric operator for each k, an inductive argument guarantees that
= for each ri. Hence r(T) = = For
(d) invoke Theorems 7.9 and 11.10.1
11. Let } and be two sequences of unit vectors in a uniformly convex
Banach space X, and let be a sequence of unit vectors in X*. Jf
1 and 1, then show that 0.
12. (Milman [242]) Show that uniformly convex Banach spaces are reflexive.
13. Show that a Banach space X is uniformly smooth if and only if for each
E > 0 there exists some 'y > 0 such that for each unit vector u and for
each v E X with lvii <'y we have
+1(uvfj <2+EIivII.
14. [308]) For a Banach space X establish the following.

(a) X is uniformly convex if and only if its dual
X is uniformly smooth if if its dual X* is uniformly convex.
15. Show that a locally uniformly convex Banach space X always satisfies
the KadetsKIee property. That is, show that in a locally uniformly
convex Banach space a sequence } is norm convergent to some vector
x if and only if x and tixti.
16. Show that each with 1 <p < co is uniformly convex and uni-
formly smooth. Also show that the infinite dimensional L1- and
spaces are neither uniformly convex nor uniformly smooth.
17. This problem discusses various differentiability properties of the norm.

Let x be an arbitrary non-zero vector in a Banach space X. Establish the
following.
(a) IfO<t1 <t27thenforeachvEXwehave
< lxii <
tl t2
<lvii.

This shows that the limit

IIx+tviI lixil
t
exists and Also, = lxii and = jjxII.
(b) The function v from X to is sublinear.
(c) A linear functional E is said to be a supporting functional
at x if 1x*1I = 1 and x*(x) = lxii. We denote by the non-empty
collection of all supporting functionals at x. Show that consists
of all linear functionals on X that are dominated by i.e.,
= E X*: x*(v) for all v E x}. (*)
(d) For the function v (v) the following statements are equiva-
lent.
(i) is linear.
(ii) = for each v E X.
(iii) The norm of the Banach space X is Gateaux differentiable
at x, i.e., exists in for each v E X.
(iv) There exists a unique supporting functional at x, that is, the
set is a singleton.
(v) (.) is linear and is the only supporting functional at x.
(e) The norm of a Banach space X is said to be Frchet differentiable
at x if the limit Ix+t4- exists uniformly in v in the unit
sphere Sx = {v E X: vii = 1 }. Establish the following.
(cr) If the norm is Frchet differentiable at x, then it is also
Gateaux differentiable at x.
The norm is Frchet differentiable at x if and only if there
exists a (necessarily unique) linear functional E X* such
that: for every > 0 there is some 8 > 0 so that for each
u E X with <8 we have
I
- Mxli _x*(u)1 (**)
18. Show that a Banach space X is uniformly smooth if and only if its norm
is uniformly Frchet differentiable over the unit sphere of X, that is, the
limit
= lim
exists in JR uniformly for all unit vectors x, y E X.

I . 2. The Daugavet Property in AL- and AIVI-spaces 467
11.2. The Daugavet Property in AL- and AM-spaces

We begin with a definition that generalizes the Daugavet property for fam-
ilies of operators. It was introduced first in [166] and [167].
Definition 11.19. Let X be a (not necessarily closed) vector subspace of a
Banach space Y and let J: X Y denote the natural inclusion operator.
We say that the pair (X, Y) has ( or satisfies) the Daugavet property for
a collection S of bounded operators from X to Y provided
= 1+11Th (*)
for each T E S.
If X = Y, then J coincides with the identity operator I and (*) simply
means that each T in S satisfies the usual Daugavet equation. The most
important families of operators that we will consider are the vector space
K(X) of all compact operators on X, the vector space W(X) of all weakly
compact operators on X, the vector space of all finite-rank operators
on X, and (when X is a Banach lattice) the vector space 1d of all regular
operators disjoint from the identity operator or, equivalently, disjoint from
all central operators.
In this section we shall study the Daugavet property when X is either
an AL- or an AA'I-space and Y = X. Our approach will be based on the
lattice structure of the spaces.
It follows from Corollary 11.13 that with 1 <p < 00 do not
satisfy the Daugavet property for compact operatorsand hence do not for
weakly compact operators either. However, as we shall see next, for p = 1
and p = 00 the situation is quite different.
Theorem 11.20. The atorniess Dedekind complete AM-spaces with units
and the atorniess AL-spaces satisfy the Daugavet property for weakly compact
operators.'
The proof of this theorem will be a consequence of the next two lemmas.
First of all, we wish to point out that neither AM- nor AL-spaces with
atoms can satisfy the Daugavet property; see Exercise 2 at the end of this
1
A few remarks are in order regarding the long history of this theorem. As mentioned in
the introduction to the chapter, for compact operators on C[O, 1] the theorem was discovered by
I. K. Daugavet [98]. It was extended to weakly compact operators on where is an
arbitrary compact topological space without isolated points, by C. Foias and I. Singer [126). For
compact operators on L1 [0, 1] the theorem was established by G. Ya. Lozanovsky in 1966 and
reproved in 1981 by V. F. Babenko and A. Pichugov [50]. Further generalizations and different
proofs were obtained by Y. A. Abramovich, C. fl Aliprantis, and 0. Burkinshaw [6], P. Chau-
veheid [88], U. U. Diallo and P. P. Zabreilco [104], J. R. Holub [153, 1541, H. Kamowitz [1T1J,
M. A. Krasnoselsky [190]? K. D. Schmidt [297], and J. Synnatzschke [312, 313, 3141. The proof
presented here is taken from [6].
section. It is also important to keep in mind that if a Banach lattice E is

either an AL-space or a Dedekind complete AM-space with unit, then every
continuous operator T: E E is order boundedand hence, its modulus
exists; see Theorem 3.9.
Two classes of operators that satisfy the Daugavet equation are
sented in the next lemma.
Lemma 11.21. Let E be either an AL-space or a Dedekind complete
space with unit and let T: E E be a continuous operator. If T is either
(1) positive or
(2) disjoint from the identity operator,
then T satisfies the Daugavet equation.
Proof. Assume first that E = is a Dedekind complete AM-space with
unit. If T is positive, then from Lemma 3.2 it follows that
= 1 + supTl(w)
wEfl wEfl
=
If T is disjoint from the identity operator, then using Theorem 3.9,
Corollary 3.10 and the preceding conclusion, we see that
= = =
If E is an AL-space, then E* is a Dedekind complete AM-space with
unit. In addition, the adjoint operator T*: E* is positive if and only
if T is positive. Furthermore, if T is disjoint from the identity operator
on E, then Theorem 2.28 implies that T* is also disjoint from the identity
operator on E*. So, the preceding case yields
= = 1+ = 1+
Lemma 11.21 is, of course, false for general non-atomic see
Exercise 6 at the end of the section. The next lemma is what is missing to
complete the proof of Theorem 11.20.
Lemma 11.22. If E is either an atomless AL-space or an atomless Dede-
kind complete AM-space with unit and T: E E is a weakly compact op-
erator, then T is disjoint from the identity operatorand hence it satisfies
Proof. Let T: E E be a weakly compact operator. Assume first that
E is an atomless Dedekind complete AM-space with unit. Then, by Theo-
rem 3.12, we know that the modulus TI of T is a weakly compact operator.
11.2. The Daugavet Property in AL- and AM-spaces 469
Now from the inequalities 0 I A TI I, 0 and Theo-

IA
rem 2.40, we see that IA is a weakly compact central operator. But then
Lemma 4J8 guarantees that I A = 0.
Next, assume that E is an atomless AL-space. Since E has order contin-
uous norm, it follows from Lemma 2.31 that the norm dual E* is an atom-
less Dedekind complete AM-space with unit. Moreover, T*: E* E* is a
weakly compact operator and so, by the preceding case, we have 1* A T* J = 0.
Finally, using that the mapping S is a lattice isometry (Theorem 2.28),
we see that
(IAITI)* ==I*AITI* =1* AIT*I =0,
and so I A TI 0 holds in this case too.
We conclude this section with one more interesting property related to

the Daugavet equation for operators on AL- and AM-spaces.
Theorem 11.23. If T: E E is a continuous operator on an AL- or an
AM-space, then either T or T satisfies the Daugavet
Proof. Let T: E E be a bounded operator on an AL or an AA'l-space.

Since T satisfies the Daugavet equation if and only if its adjoint operator
T*: E* E* satisfies the Daugavet equation and since the norm dual of
an AL-space and the second norm dual of an AM-space are both Dedekind
complete AM-spaces with units, we can assume without loss of generality
that E is a Dedekind complete AM-space with unit. That is, we can assume
that E = where is an extremally disconnected compact Hausdorff
space.
From Theorem 3.9, we know that
= HTI = 1 ITI(1) Ii =sup{(JTI1)(w):
This implies that for any given > 0 there exists a non-empty clopen subset
V of such that
(ITI1)(w) > (*)
for all w e V. Furthermore, by Theorem 1.17, we have
JTJ1=sup{T(2e1):eEEandeA(1--e)=0}.
So, there exist e = Xw, where W is a clopen set, and some wO e V such
that
T(2e-1)(wo)> IITM (*k)
2 theorem has a long history. For E = C[O, it was proven first by J. Duncan, C. WI. Mc-
Gregor, J. D. Price, and A. J. White [116] and was rediscovered later by J. R. Holub [153]. It
was generalized and proven in its present form by Y. A. Abramovich [4] and K. D. Schmidt [297').
The proof presented here is taken from [41
______________
470 - - 11. The Daugavet Equation
If we put x = 2e I = xw Xwc, then x E C(f1) and (**) can be rewritten
(Tx)(wo)> ITM -
Clearly, = I and thus x(w0) = 1 or x(w0) = 1. If x(wO) = 1, then
JI+T1I
x(wO) + (Tx)(wo) = 1 + (Tx)(wo)
1+IITJI,
and if x(wO) = 1, then
lI-TIE = -x(wo) + (Tx)(wo) 1 + 11TH -

Thus, we have shown that for each E> 0 we have either
or III-TIl
and our conclusion follows.
Exercises
1. Show that a compact Hausdorif space has an isolated point if and only
if has an atom.
2. This exercise shows that a compact operator on an AM- or an AL-space

with atoms need not satisfy the Daugavet equation. Let E = where
is a compact Hausdorif space having an isolated point, say w0. Con-
sider the multiplication operator T: E r E, defined by Tf = X{w0} f
Establish that:
(a) T is compact.
(b)
(c) 1II+TIt 1<1+JITII=2.
3. If J: E E is an arbitrary isometry on an AL-space, then show that
IJ + 1 + for each operator T E (E) which is disjoint from J.
[HINT: If x E E satisfies = 1, then note that
MJ+T11 = IIIJ+T1II IlJIIxl+ITUxI1I
= II 17iIxI II;
11
1 1 1171x11 = 1 1 1171xM 1
4. Take two arbitrary non-zero bounded real measurable functions f and g

on [0, 1] and consider the rank-one operator T = f g acting on each
1]-space. That is, the operator T: 1] 1] is defined via
the formula
Th = (f g)h = (h, f)g =
for each Ii E 1].

I I . 3. The Daugavet Property in Banach Spaces 471
(a) Verify directly (i.e., without using Theorem 11.20) that the operator
T: Lc,40, 1] LOC{O, 1] satisfies the Daugavet equation.
(b) Show that T : L1 [0 , 1] + L1 [0 , 1] satisfies the D augavet equation.
(c) If (g, 1) = 0, then show that T: 1] + 1] does not satisfy
the Daugavet equation for any 1 < p < oo.
5. If ,u is a or-finite non-atomic measure, then show that every finite-rank
operatofbn where 1 p < oc, is disjoint from the identity oper-
ator. (And so, in this case, by Lemma 11.21 every finite-rank operator
on satisfies the Daugavet equation.) [HINT: Assume that
is a non-atomic finite measure space, f E and u e (a), where
1 oc and 1 <q oo satisfy = 1. Toshow IA(fu) =0, it
suffices to establish that IA (fu)(1)
To verify this, assume that some g E satisfies g IA(f'u)(l).
Fix 0. The absolute continuity of the integral guarantees the existence
of some 8> 0 such that IA f(t) d,a(t) holds for each measurable set
A with < Next, choose some n with and then select
pairwise disjoint measurable sets A1,. . , with = for each
i. Now note that g <XAC + [J'A. f(t) + ai for each i. This
implies g Since 0 is arbitrary. the latter shows that g = 0 and
from this it follows that I A (f u) =
6. Show that Lemma 11.21 is false for with 1 <p < 00. [HINT:
Consider, for instance, the rank-one operator T: L2[0,1J + L2[0,1] de-
fined by Tf = where 1 is the constant function one.
Then T is and (since
positive, compact, disjoint from the identity =1
is not an eigenvalue) it fails to satisfyby Corollary 11.13the Daugavet
equation.]
11.3. The Daugavet Property in Banach Spaces

It is well known that the classical spaces L1 (ii) and (ii) often play an
exceptional role in many problems of functional analysis. In particular, as
we have seen in the previous section, they play a special role regarding the
Daugavet property. It even seemed quite natural to expect that the Dau-
gavet property for compact or weakly compact operators would characterize
the atomless AM- and However, as it turns out, this is not the
case. The purpose of this section is to present some new classes of Banach
spaces different from the atomless AL- and AM-spaces that also satisfy the
Daugavet property for large classes of operators.
If {XA}AEA is a family of Banach spaces, then as usual we denote by
(resp. the Banach space consisting of all vectors
x = with xA E XA for each A and lxiii = IxAll < oo (resp.
= IIXA1I < oo). In the case of a finite set A = {1, 2,.. , n}, we
shall also employ the notation

and
The Banach spaces and are usually referred to
as the Li-sum and respectively, of the family of Banach spaces
{ The direct sums of sequences of Banach spaces were already
introduced and discussed in Section 7.2.
For each index A e A we denote by PA: X + XA the natural projection
x = XA from either one of the Banach spaces X =
X onto XA.
As we shall see later, the introduced Banach spaces satisfy the Daugavet
property for a given family of operators if and only if each summand XA
satisfies the Daugavet property for this family. We start our discussion with
several auxiliary lemmas.
Lemma 11.24. If u and v are arbitrary vectors in a normed space and
a,fi E [0, 1], then lIa'u +fivII max{IJu +
Proof. The desired inequality follows immediately from the identity
au+/3v=
and the triangle inequality. Notice also that if 0 a 1, then the
above identity implies + a max{Mu + vM}.
Corollary 11.25. If T: X Z and 8: Y f Z are bounded operators
between norrned spaces, then
sup{IITx+SylI: lxii 1, 1}=sup{jJTx+SyfI:
We continue by introducing the concept of an operator ideal.

Definition 11.26. Let B denote the class of all Banach spaces and let 0 de-
note the class of all bounded operators between Banach spaces. An operator
ideal I is a "mapping" I: B x B such that:
(1) For each pair of Banach spaces X and Y the collection of operators
I(X, Y) is a (not necessarily closed) linear subspace of (X, Y)
containing all finite-rank operators from X into Y.
(2) If in a scheme of bounded operators between Banach spaces
we have T e I(X, Y), then 82T81 e I(V, W).

As usual, we write 1(X) instead of 1(X, X). In the operator ideal termi-
nology we can talk about the operator ideals of compact operators, weakly
compact operators, DunfordPettis operators, and strictly singular opera-
tors. For details regarding operator ideals we refer the reader to [266].
11.3. The Daugavet Property in Banach Spaces 473
Definition 11.27. A Banach space X is said to satisfy the Daugavet

property with respect to an operator ideal I if each operator in 1(X)
satisfies the Daugavet equation.
In the sequel we shall deal with operators acting on either (X or

(X Y)1. Recall that the natural projections P1 and P2 on either one of
these two spaces are defined by Pi(x, y) = x and P2(x, y) = y. The purpose
of the next two lemmas is to establish some inequalities that will be used to
approximate + T11 via the norm 1LTx + 811 for an appropriate operator
5: X-4X.
Lemma. 11.28. Let T: (X (X be a continuous operator,
f
where X and Y are Banach spaces. Consider the operator
A=P1T:
and assume that hAil 11Th c for some c > Then there exists an
operator 5: X X such that:
(1) 11511
I j'x +811- +TII.

(3) If for some operator ideall the operatorT belongs to
then S belongs to 1(X).
Proof. If we consider the operators Si: X f X and 82: Y f X defined
by 51x = A(x, 0) and S2(y) = A(0, y), then A(x, y) = Six + S2Y. But then
from
= sup{IjSix + S2y11:
1
1 and 1}
and Corollary 11.25, it follows that there exist two unit vectors x0 e X and
Yo e Y such that ItSixo + S2y011 = IA(xo,yo)11 hAil
Next, choose a linear functional e X* satisfying = = 1, and
then define the bounded operator 5: X X by Sx = This
implies = MAll On the other hand, we have
1 and '}
= hAil,
and the validity of (1) has been established.
To prove (2), we begin by choosing a unit vector u E X such that
llu + SuM > IIIx + 811 c. Then we have
f +T1I
= Ilu+ SuM
>
474 - 11. The Daugavet Equation
For (3) consider the bounded operator R: X (X defined via

the formula Rx = and note that in the scheme of bounded
operators
we have 8 = if T E I((X
P1TR. So, for some operator ideal I,
then it follows from Definition 11.26 that 8 E 1(X). D
An Li-version of Lemma 11.28 is as follows.

Lemma 11.29. Let T: (X Y)1 be an operator, where X and
(X Y)1
Y are Banach spaces, and consider the operator A: X X Y defined
by hAil 11Th for some > 0. Then there
exists an operator 8: X * X such that:
(1) - <11811
(2) lIx +811 +TIL
(3) If T belongs to I((X for some operator ideal I. then 8
belongs to 1(X).
Proof. Define the operators A1 = P1A: X X and A2 = P2A: X Y.

Clearly, IAxJt IIAixIJ + 11A2xlI for each x E X. Now fix SO E X with
11xo11 = 1 and
lAxoM = IIAixoI + JJA2xo1f 11AM
and then choose some with =1 E and = IIA2xoII. -
Assume first that A1x0 = 0. In this case, we claim that the rank-
one operator 8: X X, defined by Sx = satisfies the desired
properties. To see this, note that
hAil < !IAxoII = IA2xoIl = = ilSxoIl 11811 trAil.
For (2), we have
llxey + 1l(xo,0) +T(xo, 0) II = 1I(xo + I(xo,A2xo)II
= lixolt + IIA2xoII = 1 + tIA2xo!l = I + IIAxoIl
1+ MAIl 1+
IIIx+SIl.
For (3), consider the scheme of bounded operators
R2>X,
where R1x = (x,0), R2(x,y)
= y))xo, and note that 8 = R2TR1.
Now assume A1x0 0. In this case, we define the bounded operator
8: X X by 8x = A1x + where VO
= We claim that
11.3. The Daugavet Property in Banach Spaces 475
S satisfies the desired properties. Let us verify first (1). For an arbitrary
x E X, we have
ISilt + + = lAxII <

whence hAl. For the other inequality note that
151 lISxoIf = = + IIA2XOMvO 11
= lIAixohI + = IIAxohI
For (2) consider the operator 1x + 8: X X and choose some u e X

with = 1 and lu + Suit > IIx + Sit Now from the definition of 5,
we see that u + Su = u + A1u + and so
+T1I +T)(u,O)I1 = [u+ AiuII + IIA2uM

lu +
+ kb(A2u)1 = +
Iu+Aiu+i,b(A2u)voII = IIu+SulI
>
Finally, to see (3) in this case, consider the scheme of bounded operators
R2>X,
where R1x = (x,O) and R2(x,y) = P1(x,y) and note that
S=R2TR1.
We are now ready to establish that 1-sums and of Banach

spaces that satisfy the Daugavet property also satisfy the Daugavet property.
Theorem 11.30 (Wojtaszczyk [339]). Let be a family of Banach
spares such that each satisfies the Daugavet property with respect to some
operator ideal I. Then the Banach spaces and also
satisfy the Daugavet property with respect to this operator ideal I.
Proof. Assume that is a family of Banach spaces such that each

satisfies the Daugavet property with respect to a given operator ideal I.
Put X = and let T belong to Consider the
operators A,\ = P,\T: and note that IITI =
Fix > 0 and an index ,a e A for which > IIT1I and let
= Then T: V V and, moreover,
the operator = V satisfies the hypotheses of
Lemma 11.28. Therefore, there exists an operator S E such that
11811 and lIIx +T11 Since satisfies the Dau-
gavet property with respect to the ideal I, it follows that =
Consequently,
IlIx + TIl 1 + 11811 E 1 + 2e 1+ 3c.
As c > 0 is arbitrary, + T satisfies the

Daugavet property.
For the 4-sum consider an operator T E and define the
operators A), = T o IrA: X), p where IrA is the natural em-
bedding of X), into Clearly, 11Th = supA 11A 11. The remaining
arguments for are similar to those for with the
exception that-instead of Lemma 11.28 we must invoke Lemma 11.29.
Consider now the space X = where = L1[0, 1] for each

n E N. By Theorem 11.20 each satisfies the Daugavet property for
weakly compact operators and so, by Theorem 11.30, X also satisfies the
Daugavet property for weakly compact operators. At the same time X is not
lattice isomorphic to an AL- or to an AM-space; see Exercise 3 at the end
of the section. This demonstrates that, indeed, there exist "non-classical"
Banach lattices satisfying the Daugavet property. The first examples of
such Banach lattices were obtained in [5], where finite sums of AL- and
AM-spaces were considered.
Exercises
2. (Abramovich [2]) Let E be a cr-Dedekind complete Banach lattice and

UE be its closed unit ball. Consider the set D of all vectors x for which
there exists some collection .. ,
} of pairwise disjoint positive vec-
tors in such that = V V ... V and let
s ED}.
Show that Eis lattice isomorphic to an AM-space if and only ifM(E) <oo.
3. Let = L1[O, 1] and Fn = 1] for each n. For the Banach lattices
E= and F establish the following.
(a) Neither E nor F is lattice isomorphic to an AL- or an AM-space.
(b) Neither (E F)1 nor (E is lattice isomorphic to an AL- or
an AM-space.
(c) The Banach spaces (E F)1 and (E satisfy the Daugavet
property for weakly compact operators.
The Daugavet Property in 477
11.4. The Daugavet Property in

In this section will denote a compact Hausdorif space. We will prove
here several results on the Daugavet equation in These results
either provide new classes of operators satisfying the Daugavet equation or
they illustrate various techniques of establishing the validity of the Daugavet
equation.
We start by describing a new class of operators satisfying the Daugavet
equation that was found by J. R. Holub [155]. In this works he proved that
every continuous operator on C[O, 1] that factors3 through CO satisfies the
Daugavet equation. Later S. I. Ansari [37] e:xended this result to arbitrary
non-atomic
One of the major order related problems in dealing with operators on
a for an arbitrary compact space (as opposed to a Stonean
space is caused by the fact that in general the ordered Banach space
is not a Dedekind complete Riesz space. Even worse, it fails to
be a Riesz space, and so the modulus of an operator does not necessarily
exist. one cannot expect to obtain the validity of the Daugavet
equation for a given class of operators on a by simply proving
that any operator from this class is disjoint from the identity operator. A
way to bypass this obstacle is of interest in itself and is described next.
Definition Let T: be two continuous operators.
We shall say that 8 and T are essentially disjoint if there exists a Stonean
space Q and a lattice isomorphism J: C(Ifl C(Q) such that:
(1) = where and lQ denote the constant one functions

in and C(Q), respectively.
(2) The operators JT and JS are disjoint in 4 C(Q)).
(3) IIJTM 11Th and = 11811.
If the mapping J above is merely a lattice homomorphism (rather than

a lattice isomorphism), then we shall call T and S almost essentially
disjoint operators.
Some remarks are in order. First, notice thatsince C(Q) is a Dedekind
complete AM-space with uniteach bounded operator R: C(Q)
is regu'ar and that I1RII = Furthermore, note that J is a lVlarkov
operator, that condition (1) implies IJII = and that in view of (2) we
have IJT + = IJT1 + IJ8L
3Recall that a continuous operator T. X Y betweenBanach spaces factors through

a Banach space Z if there exist two continuous operators R X Z and S: Z Y such that
T=SR.
___________
I I The Daugavet Equation

.
Lemma 11.32. If is Stonean, then for bounded operators on C(IZ):

( a) The notions of essential disjointness and disjointness coincide.
(b) Almost essential disjointness does not imply essential disjointness.
Proof. (a) Assume that two operators 8, T e satisfy S A T = 0.

If we take Q = and let J: C(Q) be the identity operator, then
clearly J is a lattice isomorphism such that = lQ, =
IIJSM = [1811, and JT is disjoint from J8 in 4 C(Q)).
For the converse, assume that 8 and T are essentially disjoint. That is,
there exist a Stonean space Q and a lattice isomorphism J: C(12) p C(Q)
satisfying the properties of the definition of essential disjointness. Since
is Stonean, the operator 8 A T exists in L(C(IZ)). And now from the
inequalities 0 J(8 A T) (J8) A (JT) = 0, it follows that J(8 A T) = 0.
This implies 8 A T = 0 in
(b) To show that the essential disjointness may be stronger than the
almost essential disjointness let us fix three arbitrary disjoint non-empty
clopen subsets A, B1, B2 of such that = A U B1 U B2. (We assume here
that contains more than two points and so such A, B1, B2 do exist.) Let
E = C(B1 U B2) and let J: p E be the natural restriction operator,
i.e., for f E we let Jf = Finally, define 8, T:

by Tf = fXAUB1 and Sf = fXAUB2. It is easy to see that T and S are not
disjoint in Indeed, using Theorem 1.17, we see that
(TA = + V is a clopen subset of
= inf {X1'fl(AUB1) + XVCfl(AUB2): V clopen }
Thus, we have established that T A S > 0. Finally, a direct verification

shows that the operators JT and JS are disjoint in Lr(C(IZ), E) and also
that = = = 11811 1.
As shown above, the almost essential disjointness is a rather weak prop-

erty, but nevertheless, it is sufficient to imply an analogue of Lemma 11.21.
Lemma 11.33. If a bounded operator T on is almost essentially dis-
joint from the identity operator, then T satisfies the Daugavet equation.
Proof. Let Q be a Stonean space and J: C(IZ) C(Q) be a lattice isomor-

phism satisfying the properties of Definition 11.31 for the operators I and
T. Clearly, IV+TM as 1. On the other hand, we have
IIJI+ JTII =
I+ = 1+ = 1+ 11Th
11.4. The Daugavet Property in C(IZ)-spaces 479
Hence, 111 + Tlt 1 + 11Th, and we are done.
If1i, is a regular Borel measure on then by we will denote the natural

embedding of into That is, for each x its image in
is denoted by Obviously, is a lattice homomorphism and
1 (provided 0).
We will need below the well-known fact (due to W. Rudin [286] and
A. Pelczynski and Z. Semadeni [263]) that a perfect compact topological
space admits a non-atomic measure. For other proofs see [300, Theo-
rem 19.7.6], [24, Theorem 5.3] or [25, Theorem 10.20]. As usual, a measure
is called non-atomic if = 0 for each w E ft
Theorem 11.34. If has no isolated points, then it admits a non-atomic
regular Borel probability measure.
From now on, until the end of this section, if is a measure on then we
shall denote the u-essential supremum of a function f by if
Lemma 11.35. If has no isolated points and T: is a
bounded operator, then there exists a non-atomic regular Borel probability
measure on such that =
Proof. For each k choose a function xk E such that iIXk 1 and
IlTxklIoo> 11Th Put Uk = {w E Txk(w)l> llTxk H00 and

note that each Uk is a non-empty open set.
Since does not have isolated points, it follows that each Uk does not
have isolated points either; see Exercise 4 at the end of this section.
fore, by Theorem 11.34 there exists a non-atomic regular Borel probability
measure on Uk. We extend to a non-atomic measure on all of by
letting \ = 0. Clearly, > IT1I
Now let = Then is a non-atomic regular Borel proba-
bility measure on and 11Th holds for each
k. This implies 11Th, and consequently 11Th.
If is a metrizable compact space without isolated points, then there

exists a regular Borel probability measure on such that the
embedding L00 is a lattice isometry; see Exercise 5 at the
end of this section.
Corollary 11.36. If has no isolated points, then for every seq'aence
of bounded operators on C(IZ) there exists a non-atomic regular Borel pro b-
ability measure on such that = for each m.
Proof. By Lemma 11.35, for each m there exists a non-atomic regular Borel
probability measure on such that = Then the measure
= is a regular Borel probability measure on

with the desired property.
Lemma 11.37. If is a regular Borel measure on then for each linear

operator 8: satisfying 0 S there exists a non-
negative function h E such that Sx = hx for each x E
Proof. The image G = (C(11)) is a Riesz subspace of the Dedekind comrn

plete Banach lattice We define an operator R: G by
letting R(g) = Sx, where x E and J,2x = p. It is easy to see that R is
well defined and satisfies 0 Rg p for each p E Therefore, by [30,
Theorem 8.15] the operator R can be extended to a positive central operator
on which we shall denote by R again. However, as we already know
from Theorem each central operator on is a multiplication op-
erator. That is, there exists a function h E such that Rg hg for
each p E (u). Since R is positive, the function h is non-negative. Finally,
for each x E we have that Sx = RJ12x = Rx = hx.
Recall that the series of elements of a (real or complex)

nach space X is said to be weakly unconditionally Cauchy if for each
E X* the series of real numbers )f converges in IR. Clearly,
every subseries of a weakly unconditionally Cauchy series is also weakly
unconditionally Cauchy.
If is a weakly unconditionally Cauchy series in then by
using the functional we obtain that < oo for each
w E ft The next result informs us that not only each weakly unconditionally
Cauchy series in converges absolutely at each point of but also that
this convergence can be made uniform on subsets of positive measure.
Lemma 11.38. Let > be a weakly unconditionally Cauchy series in

If is a regular Borel measure on then for each > 0 there
exists a closed subset B of ci such that \ B) and the series
converges uniformly on B.
Proof. Since Ihn(w)I <oo holds for each w the conclusion fol-
lows from Egorov's theorem and the regularity of the measure
We continue with a factorization property of operators on

11.4. The Daugavet Property in -spaces 481
Lemma 11.39 (Holub [155]). An operatorT E (C(IZ)) factors through CO

if and only if T admits a representation of the form
(*)
where:
(a) The sequence c C(cz)* is weak* convergent to zero.
(b) The series is weakly unconditionally Cauchy in C(ffl.
(c) The series in (*) converges in the strong operator topology, that is,
is norm convergent in C(IZ) for each x E
Proof. Let denote the standard basis in co and denote its

biorthogonal sequence. We begin by establishing that each operator T in
L(C(ffl) factoring through CO admits a (*)-representation. Therefore, as-
sume that there exist two bounded operators C(IZ) C(IZ) such
T = SR. Let = and = As 0 in the topology
co), it follows that 0 in the topology
To show that >j is a weakly unconditionally Cauchy series, pick
E and consider the vector S*x* = = Then

00 00 CO 00
lx*(Sen)1 I(S*x*)(en)1 P'ml <o.
= = =
To show that the series in (*) converges in the strong operator topology, fix
any x E C(IZ), and let Rx = (ai, a2,...) = ==
It follows that
Tx S(Rx) = S(>(Rx, =
For the converse, assume that an operator T can be represented in

the form (*). Let Rx = Since * 0 in the topology
the mapping R is a continuous operator from C(fZ) to co.
Next define 5: c0 * C(IZ) by 8\ = >j This series converges in
since is weakly unconditionally Cauchy; see Exercise 3 at
the end of the section. It remains to note that the convergence condition
(c) implies that T = SR.
Theorem 11.40 (Ansari [37]). If has no isolated points, then every con-
tinuous operator on C(fZ) factoring through c0 is almost essentially disjoint
from the identity operator.
Proof. Let T: C(IZ) * C(IZ) be a bounded operator that factors through

Co. By Lemma 11.35 there exists a non-atomic Borel measure ,u on such
that = and = 1, where is the natural

embedding. Lemma 11.39 guarantees that
for some weak* null sequence in C(11)* and a weakly unconditionally

Cauchy series in C(Q). Let 'y = supflEN The
validity of our theorem will be established if we show that the operator
S = JJ/.LTJ A J/.L E is trivial, that is, S = 0.
To this end, assume contrary to our claim that S 0. Since 0 S
by Lemma 11.37 there exists a non-negative function h E such that
Sx = hx for each x E C(11). As S 0, the multiplier h is non-zero, and
hence for some > 0 we have 1i(A) > 0, where A = {w E h(w) > }.
By Lemma 11.38, we can find an integer k and a closed subset B of A with
> 0 such that
sup
wEB n=/c+1
As isnon-atomic and each /in is regular, we can find an open sef U

such that ,ii(B fl U) > 0 and < for each 1 n $ k. Using
the regularity of we can assume without loss of generality that B U
(otherwise, we can replace B by a compact subset of B fl U of positive
measure). Also pick a function x0 E C(11) such that xoJB = 1, xoJcl\u = 0,
and
Notice that Sx = hx0, and so > for each w E B. On the other
hand, we have
5x0 = J,UTI A
= xEC(11) and xjxO},
where the supremum is taken in Let us denote this supremum by

Yo We claim that Yo = G, where the function C: [0,oo] is defined by
G(w) = sup {Tx(w): x and Jxf x0}. Indeed, each Tx above is
a continuous function and JJTxIJ 11Th. Therefore, C is a bounded lower
semicontinuous function, and hence C E This implies C = yo, and
11.4. The Daugavet Property in 483
so C. Now using the representation for T, for each w E B we get
c < Sxo(w) G(w) = sup

ri=1
= sup +
IxIxo ri=1 m=k+1
which is a contradiction. This completes the proof of the theorem.

Corollary 11.41 (AnsariHolub). If has rio isolated points, them every
operator on that factors through co satisfies the Daugavet equation.
Proof1 If an operator T E factors through cO, then T is almost
essentially disjoint from the identity operator by Theorem 11.40. So, by
Lemma 11.33, the operator T satisfies the Daugavet equation.
Note that there exist non-weakly compact operators on
that factor through co; see Exercise I at the end of the section. Thus,
Corollary 11.41 describes a new class of operators satisfying the Daugavet
equation.
We proceed with a few comments regarding the role of the embedding
> for the method discussed above. Recall that
is a Dedekind complete Banach lattice and the operator of the canoni-
cal embedding j: is a lattice isometry. Consequently,
for each bounded operator T on the operator jT has a modulus in
L(C(12), C(12)**). Moreover, repeating verbatim the proof of Lemma 11.33,
one can see that if jT is disjoint from ji, where I is the identity operator
on then T satisfies the Daugavet equation. Therefore, it is natural to
ask if it is possible to use the operator j instead of an appropriate isomor-
phism Exercise 9 at the end of this section shows that the answer to
this question is negative.
The rest of this section will be devoted to a description of one more
approach to the Daugavet equation for operators on This will
allow us to describe two more classes of operators on that satisfy
the Daugavet equation. These classes of operators have been introduced by
L. Weis and D. Werner in [330] who used them to generalize Corollary 11.41
by showing that c0 can be replaced by any Banach space with a separable
dual. As we shall see, the proof is surprisingly simple and independent of
that of Corollary 11.41.
Recallthat each bounded operator T: generates the so-

called representing kernel, i.e., a family of signed regular Borel measures
{/-tw}WEII on defined by = for each w E ft From the estimate
= IIT*&1I = it follows that the family {/2w}wEll is
a norm bounded subset of the norm dual C(cI)*.
The following characterization of the operators satisfying the Daugavet
equation via the representing kernel is due to D. Werner [332}. The proof
below is similar in spirit to the original proof of I. K. Daugavet in [98]. A
more formal algebraic proof is given in [332].
Lemma 11.42 (Werner). If the representing kernel of a continu-
ous operator T: C(fZ) C(fZ) satisfies
inf = 0 (t)
wEG
for each non-empty open subset G of then T satisfies the Daugavet equa-
tion.
Proof. Let T: p be a bounded operator. Fix E> 0 and choose

some x E such that I and > E. The set
ft
I
G = {w E ITx(w)I > ITII E} is non-empty and open and hence, by

(f), we can find a point to G such that = < E. By
the regularity of the measuie I there exists an open neighborhood V of
the point t0 such that 1(V) < E. We can assume that V c G. Now for
any function y E whose support lies in V we have
= =
=
< <
y yc V we have
+ y) + T(x (x+ y)(to) + Tx(to) +Ty(to)
(x + y)(to) + 11Th
Next pick a function z E such that 0 z 1, z(to) = I and z = 0

on Vc. the function y = (1 x)z
Consider E and note that y
has support in V and that 2. From = I, it follows that
1 x 1, and so I. Hence, from
I? x(to) +y(to) =x(to)+Ix(to) = I,
it follows that x(to) + y(to) = I. Finally, note that
fl +TM 11(1 + T)(x + (x + y)(to) + 11Th E
Since > 0 is arbitrary, this implies I + T I + I That is, T satisfies

1 1.4. The Daugavet Property in C(IZ)-spaces 485
Theorem 11.43 (WeisWerner {330]). Assume that is a compact Hau-

sdorff space without isolated points and let T 1(C(IZ)) satisfy one of the
following two conditions:
(1) The adjoint of T has a separable range.
(2) T factors through a Banach space with a separable dual.
Then T satisfies the Daugavet property.
Proof. (1) Let }wEcl be the representing kernel of T. Since the range of
T* is separable, there exists an at most countable subset of such that
the collection wE is norm dense in w -.E Therefore, it
is easy to see that the set A = {w E O} coincides with the
set {w E o}. Consequently, A is at most countable.
isolated points, it follows that each is an open dense subset of From

the Baire Category Theorem (see, for instance [247, p. 294]), we conclude
that the set
fl
aEA
is also dense in This implies that condition (t) of Lemma 11.42 is satisfied,
and so T satisfies the Daugavet equation.
(2) Let T E L(C(IZ)) and assume that there exist a Banach space Z with
a separable dual and bounded operators Z satisfying
T SR. From T* = and the fact that is separable, it follows that
the range of T* is necessarily separable. Therefore, by the previous case, T
satisfies the Daugavet equation.
Exercises 7 and 8 at the end of this section indicate how one can use
Theorem 11.43 to obtain an alternative proof of the fact that each (weakly)
compact operator on an atomless satisfies the Daugavet equation
(Theorem 11.20).
We conclude this section by one more elegant application of Theo-
rem 11.43. Recall that a bounded operator T: X > Y between two Banach
spaces fixes a copy of a Banach space Z if there exists a subspace X0 of
X such that Xo is isomorphic to Z arid T: X0 T(Xo) is an isomorphism.
It should be clear that if is uncountable and a bounded operator
T: * fixes a copy of then the range of T* cannot be
separable.4 Moreover, according to a deep theorem of H. P. Rosenthal [281],
for an uncountable metrizable the converse statement is also true, i.e., if
4Notice that if w2, then 1. Therefore, if is uncountable, then the

norm dual of is not separable.
for some bounded operator T: the adjoint operator T* has a

non-separable range, then T fixes a copy of This and Theorem 11.43
yield the following nice result which is also contained in [330].
Theorem 11.44 (WeisWerner). Assume that is a metrizable compact

space without isolated If an operator T E does not fix a
copy of then T satisfies the Daugavet equation.
The natural generalizations of the are the uniform algebras,

and it does not come as a surprise that some of them satisfy the Daugavet
property for the same collections of operators as This was
established by Wojtaszczyk in [339]. Since the theory of Banach algebras is
beyond the scope of this book, we refer the reader to this paper for details.
Exercises
1. Give an example of a continuous operator on C[O, 1] that factors through
CO but that is not weakly compact. [HINT: By a well-known theorem of
A. Sobczyk [309], we know that every closed subspace X of C[O, 1] that

is isomorphic to CO is complemented in C[O, 1]. If P: C[O, 1] + C[O, 1] is
a continuous projection with range X, then P is a non-weakly compact
operator that factors through Co.]
2. For a series in a (real or complex) Banach space the
(a) The series is weakly unconditionally Cauchy.
(b) The supremum supmEN is finite.
(c) The supremum supmEN is finite.
3. Prove that a series in a Banach space X is weakly uncondi-
tionally Cauchy if and only if for each A = (A1, A2,...) E c0 the series
is norm unconditionally convergent. Furthermore, show that
the mapping T: + X, defined by
is a bounded operator that need not be weakly compact.

4. Let S be a topological space without isolated points and let C be an
arbitrary open subset of S. Show that neither C nor its closure C has
isolated pointswhere, of coarse, C and G are equipped with the induced
topology.
1 1 . 5. Slices and the Daugavet Property 487
5. For a compact Hausdorfi' space establish the following.

(a) If a regular Borel measure on
ii, is with full support (that is,
Supp ii, := then the natural embedding : 4 L00 (iij) is a
lattice isometry.
(b) If also metrizable and has no isolated points, then there exists a
non-atomic regular Borel probability measure on with full support.
[HINT: For (b) mimic the proofs of Lemma 11.35 and Corollary 11.36.1
6. Establish the following separability properties.
(a) If the adjoint of a bounded operator T: X + Y between Banach
spaces has a separable range, then T likewise has a separable range.
(b) If a Banach space X has a separable dual, then X itself is separable.
Also, give an example of a separable Banach space whose norm dual
is not separable.
7. (WeisWerner [330]) If is a metrizable compact space without isolated
points, then show that every weakly compact operator on C(11) satisfies
the Daugavet equation. [HINT: Recall that an operator T E is
weakly compact if and only if
8. (WeisWerner [330]) The BartleDunfordSchwartz theorem [118, p. 306]
states that: A non-empty subset A of = is weakly compact
if and only if it is norm bounded and there exists a measure ,u E
such that v(E) 0 uniformly in v E A when + 0.

Modify, using the BartleDunfordSchwartz theorem, the arguments
of the proof of Theorem 11.43 to show that the assumption of metriz-
ability made in Exercise 7 is not necessary. That is, show that if is
a compact Hausdorff space without isolated points, then every weakly
compact operator on satisfies the Daugavet equation.
9. (Ansari [37]) Fix q E [0, 1] and consider the rank-one positive operator
T = Sq' on C[O, 1]. That is, Tx = x(q)1 for each x E C[O, 1]. Establish
the following properties of T.
(a) T factors through CO (and hence T satisfies the Daugavet equation).
(b) The operator jT: C[0, 1] k C[0, is not disjoint from ji, where
the operator j: C[0, 1] C[O, is the canonical embedding and
I is the identity operator on C[O, 1].
(c) The adjoint operator T*: C[O, 11* C[0, 1]* is not disjoint from
the identity operator on C[0,
11.5. Slices and the Daugavet Property

The purpose of this section is to establish the following interesting result due
to V. M. Kadets, R. V. Shvidkoy, C. C. Sirotkin, and D. Werner [166, 167]:
If the operators on a Banach space satisfy the Daugavet equation,
then every weakly compact operator also satisfies the Daugavet equation. To
prove this, we shall follow the approach of the above authors which is based
on the notion of a slice of a set.
Definition 11.45. Let (X, X') be a dual pair of real vector spaces, and let
C be a weakly bo'anded s'abset of X. Also, let x' E X' and e> 0
be given, and p'at s = (c, x'). Then the slice of C determined by x'
and is the set
= {x E C: (x,x') s
The geometric illustration of a slice is shown in Figure 1. Clearly, each

slice is and the smaller the the smaller the slice.
Figure 1
The next lemma presents a convenient geometric characterization in

terms of slices of Banach spaces satisfying the Daugavet property with re-
spect to rank-one operators.
In our discussion, the dual pair X') will be, as a rule, lArhere
X is a Banach space and X* is its norm dual. The unit ball of X or of X*
will serve as the set C. Under these conditions we will also assume that the
vector x E X) is of norm one. In this case, we have
xE = Jx*1I=1 and sup{(x,x*): E = ltxlI=1.
Lemma 1146. Let X be a linear s'abspace of a Banach space Y and let
J: X Y denote the inclusion operator. Then the following statements are
equivalent.
(1) Every rank-one operator T: X Y satisfies the Da'agavet equa-
tion, JIJ+Tjl = 1 + ITIL
(2) If E Y and are unit vectors, then for each 0 < 1
there exist some linear f'anctional E of norm one and some
0 < S < 1 such that:
(a) Slice(UX,X*,8)
(b) For each x E 6) we have + 2
(3) If vo E Y and are unit vectors, then for each 0 <E <
E 1
there exist a unit vector y E Y and some 0 <6 < 1 such that:
(a) Slice(Uy*,y,6) C Slice(Uy*,yo,).
(b) For each ESlice(Uy*,y,6) we have 2.
Proof. We establish only the equivalence (1) (2). The proof of the
equivalence (1) (3) is similar and is left to the reader.
(1) (2) Assume that Yo E Y and E are unit vectors, and let
0<- <-1. Start -by- observing that the rank-one operator- T = -yo in
L(X, Y) satisfies 11Th 1. So, by our hypothesis, + = 1 + 1TII = 2,
and consequently 1IJ* + T*L1 = 2. Fix some linear functional E of
y* y*
norm one satisfying 1
+ > 2 > 0 and (Yo) 0, and let
- x *_
an d
I and 0 <6 < 1.
Fix any x E 6) = {x E x*(x) 1 6}. We claim that
y*(x) + y*(y )x*(x)
2 E. (*)
To see this, note that
y*(x)+ = (x, (J* +T*)y*) = (x, IIJ*y* +T*y*11x*)
Since y*(x) 1 and 0 <y*(yo) <1, it follows from (*) that
2__y*(x) Ie>0.
But then y*(yo) 0 yields y*(yo) > 0 and > 0. Therefore,
1 c> 0.
That is, x E ), and so 6) c ).
Now taking into account 0 < 1 and using (*) once more, we get
2 y*(x) + )x*(x) <y*(x) + = (x +yo,y*) + mit.
(2) (1) Let T = 0 Yo E L(X, Y) be a rank-one operator. We

must verify that + I'll 1+ Without loss of generality we can
assume that 1 (and so 11Th = 1).
Fix 0 < < 1. By our assumption there exist a linear functional E
of norm one and some 0 < 6 < 1 such that if x E 6), then
x e and + poll 2 E. In particular, if we take any

x E Slice(UX,x*,6), then
+ + [1
(1)(2)
for each 0 < < 1. Letting e J, 0 yields JIJ + TJI 2, as desired. D
Corollary 11.47. Let X be a linear subspace of a Banach space

If Y.
every rank-one operator from X to Y satisfies the Daugavet equation, then
any slice of the unit balls and Uy* has diameter equal to 2.
Proof. We shall show the result only for the slices of So, let E
be a unit vector, let 0 < 1, and consider the slice Also,
fix some 0 < and note that
Pick an arbitrary unit vector u E Slice(Ux, ) and let yo = u. Now
apply the second statement of Lemma 11.46 to yo, and e to get a unit
vector X* and some 6> 0 such that Slice(Ux,x*,6)
and jx + yo JJ 2 for each x e Slice(Ux, x
x u belong to it
follows that the diameter of is greater than or equal to 2
for all 0 < Hence, the diameter of ) is 2, as claimed. D
In order to proceed further, we need to recall a few facts regarding

extreme points of convex sets in Banach spaces. First, let us recall their
definition.
Definition 11.48. Let C be a non-empty convex subset of a real Banach
space Z. A point cC e is said to be:
(1) an point if c = au + (1 a)v for some u,v e
extreme C and
somea'E(O,I) implythatu=v=c,
(2) an exposed point if there exists some E Z* (called an exposing
linear functional) satisfying z*(c) > z*(z) for all z E C\{c}, and
(3) a strongly exposed point if it is an exposed point and there exists
an exposing linear functional 2* e Z* such that a sequence in
C satisfies lIzn cli 0 if and only if z*(zn) z*(c).5
It should be clear that an exposed point is an extreme point. However,

an extreme point need not be an exposed point; see Figure 2. Also, an
51n this case, we also say that the linear functional E Z* strongly exposes the point c.
exposed point need not be strongly exposed. For instance, if Z = 2 and

C= then 0 is an exposed point of C but it fails to be strictly exposed.
A non-exposed extreme point
A strongly exposed
point A strongly exposed
point
Figure 2
Regarding strongly exposed points we have the following result of J. Lin-

denstrauss [203] and S. L. Troyanski [323]. (For a geometric proof see
J. Bourgain
Theorem 11.49 (LindenstraussTroyanski). Each convex weakly compact
non-empty subset of a Banach space is the closed convex hull of its strongly
exposed points.
Now we are ready to prove the promised characterization of the Daugavet

property for weakly compact operators by means of rank-one operators.
Theorem 11.50 (KadetsShvidkoySirotkin--Werner). Let X be a linear
sub space of a Banach space Y and let J: X Y denote the inclusion opera-
tor. If every rank-one operator from X to Y satisfies the Daugavet equation,
then every weakly compact operator in L(X,Y) also satisfies the Daugavet
equation.
Proof. Assume that every operator from X to Y satisfies the

Daugavet equation, and let T: X + Y be a weakly compact operator with
= 1. Fix 0< c <1. We shall show that 2 3.
Start by observing that the set K = T(Ux), the norm closure of

is weakly compact. Therefore, according to Theorem 11.49, K coincides
with the closed convex hull of its strongly exposed points. This implies that
there exists a strongly exposed point yo in K such that
> sup 1-c.
yEK
From =
it follows that
1, Ii 1. Next pick a linear functional
y*
E that strongly exposes Yo and satisfies y*(yo) = maxYEK y*(y) = 1.
We claim that there exists some 0 < 8 < such that the slice of K
B = Slice(K,y*,8) = {y E K: y*(y) 1 6}
is included in the open ball centered at Yo with radius , i.e., B c B(yo, ). To

establish the existence of such a 8, assume by way of contradiction that this is
not the case. This means that there exist a null sequence and a sequence
,
c K satisfying 0 < < 1 and Yo IF for
each n. This implies 1
= (Yo), and so (since strongly exposes
Yo) we have IIyn YoM 0, which is a contradiction. This contradiction
establishes the existence of a desired 6.
Next, let = T*y* and note that
=sup{y*(Tx): xE = max{y*(y): yE K} = 1.
Moreover, the slice A Slice(Ux, 6) of satisfies T(A) c B. Indeed,
T(A) = {Tx: x E and x*(x) 1 8}

= {Tx:xEUxandy*(Tx)1_6}cB.
In particular, for each x E A we have I Jyo TxJJ <.
Now statement (2) of Lemma 11.46, applied to the unit vector
and the slice A, guarantees the existence of some vector XO E A satisfying
xo + 2 6 > 2 . This implies
lixo +yoH = f(xo + + (1

21-f-1>2--2.
Finally, note that
(I J + T1L + Txo 1= + Yo) - (Yo - Txo)
and the proof is finished. D

______
I I . 6. Narrow Operators 493
Exercises
1. Prove the equivalence of statements (1) and (3) in Lemma 11.46.

2. ([1671) Let X be a subspace of a Banach space Y and let J : X Y
denote the inclusion operator. Show that the following statements are
equivalent
(a) Every rank-one operator T: X Y satisfies the Daugavet equation,
i.e., + 1+
(b) For any pair of unit vectors y E Y and E and each 0 < < 1
there is a unit vector x X satisfying if(x) and + 2 E.
(c) For every pair of unit vectors y E Y and E and each 0 < 1
there is some unit linear functional E 1/* satisfying (y) 1
and I1x* + J*y*1t 2 .
3. If X is a reflexive Banach space, then X does not satisfy the Daugavet
property for rank-one operators.
4. Let C = the positive cone of 2. Show that 0 is an exposed point of
C that fails to be strongly exposed.
5. Use Theorem 11.50 to present an alternative proof of the second part
of Theorem 11.20: If E is an atomless AL-space, then E satisfies the
Daugavet property for weakly compact operators. [HINT: According to
Theorem 11.50 we must show that each rank-one operator satisfies the
Daugavet equation. Without loss of generality, we can assume that there
is an atomless measure ji such that E = L1(ji). By Lemma 11.21 it suffices
to show that if 0 E and 0 u E then IA (q5 u) = 0.
Since the functions in L1 (ji) have a-finite support, we can assume that p
is also a-finite.]
6. ([167, 181}) Let be a non-atomic measure space and let Z be
a Banach space. Show that the Banach space X L1 (,u, Z) of Z-valued
Bochner integrable functions satisfies the Daugavet property for weakly
compact operators.
7. (Kadets [165]; Khalil-[181])-Let K be-a.compacLHausdorff spacewithout
isolated points and let Z be an arbitrary Banach space. Show that the
Banach space X = C(K, Z) of Z-valued continuous functions satisfies the
Daugavet property for weakly compact operators.
11.6. Narrow Operators

In this section we shall discuss the relatively new class of narrow operators
and we shall study its relationship with the Daugavet property. Though it
is not clear to us who introduced the class of narrow operators, it should be
mentioned that the works of J. Bourgain, N. Choussoub, and H. P. Rosen-
thal [71, 136, 282, 283] on operators are closely related
to narrow operators. In our presentation we follow [267], where the reader
can also find many other interesting results.
I 1. The Daugavet Equation
Throughout this section, we shall assume that (Il, is an atomless

probability measure space and that E is a real Banach function space on
with order continuous norm and that Y is an arbitrary Banach
space. We will also assume that E L1 (u). The last assumption
is quite mild since by the Twist Theorem of C. Lozanovsky for any
Banach function space E there exists a measurable function h such that
hE = {hx: x E E} Moreover, if E is a rearrangement
invariant Banach function space,6 then the inclusions c E L1
hold automatically.
We start with a simple result that will be needed later.
Lemma 11.51. If E has order continuous- norm,-then-for--each x E E and
for each > 0 there exists some 6> 0 such that for each measurable set B
with 1u(B) <6 we have JJXXBJI <.
Proof. If this is not the case, then there exist some 0and a sequence
satisfying < and for each n. Letting
= Bk, we see that 1 1
for each
n, which is impossible.
We are now ready to introduce the notion of a narrow operator.

Definition 11.52. A bounded operator T: E Y is said to be narrow if
for each A E and each > 0 there exist some B E with B A and
p,(B) and a function x E E satisfying IxI = XB and MTxII <.
Note that since we deal with real Banach spaces, any measarable func-
tion x that satisfies xI = is necessarily of the form x = XB\c for
some measurable subset C of B. Clearly, scalar multiples of narrow opera-
tors are again narrow operators. However, the sum of two narrow operators
need not be a narrow operator; see Exercise 7 at the end of this section.
A useful characterization of narrow operators is as follows.
Theorem 11.53. A bounded operator T: E Y is narrow if and only if
for each A E and each > 0 there exists a function x E E such that
and
Jci
Proof. The "only if" part is obvious. We shall prove the "if" part. So,
assume that T is a non-zero narrow operator. We shall show first that for
Banach function space E is called rearrangement invariant if for each y E

that is equimeasurable to some x E E we have y E E and ItxfI. As usual, two functions
x, y E are said to be equimeasurable if they generate the same distribution function, i.e.,
any D E and any > 0 there exists some y E E satisfying = XD and
To see this, fix D e and 77 > 0. We claim that there exist a se-
quence .} of pairwise disjoint measurable functions and a sequence
{ Co, Ci, C2,. .} of pairwise disjoint measurable subsets of D such that for
each n 1 we have:
(a)
(b)
(c)
(d)
(e) IITynII
If this is done, then for each n let = and note that E

j
and Zn 0. From Zn and the order continuity of
the norm, it follows that the series y = is norm convergent in E.
Clearly, IITYnI[ = 77. Moreover, from

0< =
=
it follows that = that is, D = Cn. Consequently,
[YI = IYmI = XCn =
The existence of {yi, } and {C0, Ci,. .} will be established by
. .
induction. Start by letting D0

- D and Co = 0. Since T is a narrow
operator, there exist a measurable subset C1 of D with and
a measurable function Yi E E such that = xc1 and < Clearly,
= \ = = =
Now, for the induction step, suppose that the measurable functions
Yi, Yk and the measurable subsets Ci,.. of D have been chosen
. ,
to satisfy (a), (b), (c), (d), and (e) for each n = 0,1,. ..,k. Since T is a
narrow operator, there exists a measurable subset Ck+1 of Dk = D \
with and some measurable function Yk+i E E satisfying
I = XCk+1 and IITyk+1M < Moreover, we have
=
1 \ 1
This completes the induction.

496 - II . The Daugavet Equation
We are now ready to finish the proof of the theor

A e and > 0. Since 1 e E, there exists (by Lemma 11.51) some S > 0
such that IXBH = 111XB11 < whenever B e satisfies ,a(B) < S. Fix
some n with < 8, and then partition A into n subsets B1 , each . . . ,
of measure this is possible since is non-atomic. By the first part, for

each k = 1,... ,n we can find some function Uk E E such that IukI = XBk
and < Let ak = d,u. Clearly and therefore we
can find signs k = 1 such that
Let u = Clearly,
= IITuM
ui and
u(w)=1}andcl2={weA: u(w)=1},then
the last inequality implies /2(d1) We can assume that
/2(d1) /2(d12). Now select any measurable subset d13 of such that
d11
= and consider the function x E E defined by x(w) = 1
if d13 and x(w) = u(w) if w E dl \ d13. It follows that
x d,u = 0, and
<
IITxIl + 11Th' +<
Thus, x is a function with the desired properties.
Corollary 11.54. If T E L1(E, Y) is a narrow operator, then for each A in
and each n there is a measurable subset B of A such that /2(B) = 2Th/2(A)
and the function h = l)xB XA\B e E satisfies
Proof. The proof is by induction. For n = 1 the conclusion follows un-

mediately from Theorem 11.53. For the induction step, suppose that the
statement is true for some n. We must prove that it is also true for n + 1.
By the induction hypothesis, there exists a measurable subset of A
such that = and the function = E E
satisfies < By Theorem 11.53 there exists some x E E such that
= = 0 and ITxIJ
Now consider the measurable set = {w e x(w) = 1 }. Clearly,
= and the function =
in E satisfies = 2Thx. This implies
+ II2ThTx(E + + =
and the induction is complete. D
To continue our discussion, we need the concept of a generalized sequence

of Rademacher functions supported by an arbitrary measurable subset of
dl. To define such a sequence, let A be a measurable subset of dl. We
let ri = r1 (A) = Since ,u is assumed to be non-atomic, we can split
A into two disjoint measurable subsets of equal measure. That is, there
are disjoint measurable subsets A1, A2 E E such that A A1 U A2 and

= We let r2 = r2(A) XA1
The sequence will be constructed by induction following the pre-
ceding process as follows. Assume that the nth Rademacher function has
been constructed so that
= =
where fl 0 if i j, A = Bk, and = each for

1 k For each 1 k n choose two disjoint measurable subsets Bk,1
and Bk,2 of Bk = Bk and
i = 1,2. Now the (n + i)th Rademacher function is defined by
= = XBk,2)'
Clearly, each function satisfies lTn(A)t = XA and so, since 1 e E, each

necessarily belongs to E.
Definition 11.55. Any sequence of functions constructed as above
is called a (generalized) sequence of Rademacher functions sup-
ported by A.
Generalized sequences of Rademacher functions converge weakly to zero.

Lemma 11.56. Every generalized sequence ofRademacherfunctions
is weakly convergent to zero in E, i.e., -
Proof. This proof is due to Y. A. Abramovich [3, Theorem 3.1]. Fix some
measurable subset A of with > 0, and let be a generalized
sequence of Rademacher functions supported by A. We must show that
fe E*. Since is an ideal in E and E has order con-

(,u)
tinuous norm, it follows that f is an order continuous linear functional on

Taking into account that the order continuous dual of coin-
cides with Li(,u), we see that f restricted to (,u) belongs to L1 (,u). Now
observe that C (,u). So, in order to establish 0, it suffices to
prove that f E L1 (,a). In view of the 1-denseness
of in Li(,u), to establish the latter we must show that
f f
a sequence in the Hubert space
L2(,u). The construction of guarantees that
(rn, rm)
= f rnrm 0
holds for all n m and 112 = for each n. Therefore, the sequence
is an orthonormal sequence in L2(,u). Since f E L2(,a), Parseval's
inequality yields <oc. This implies
f = U, > 0,
and the proof is finished. (For a generalization of this result see Exercise 10
at the end of the section.)
Recall that a bounded operator T: X > Y between two Banach spaces

is a DunfordPettis operator if 0 in X implies > 0. We
are now ready to establish that DunfordPettis and compact operators are
narrow operators.
Theorem 11.57. Each DunfordPettis operatorand hence each compact
operatorfrom E to Y is narrow.
Proof. Assume that T E (E, Y) is a DunfordPettis operator. Fix any

A E E and let E> 0. Consider a Rademacher sequence supported by
A. By Lemma 11.56, we know that 0. Since T is DunfordPettis,
it follows that > 0. In particular, there exists some k such that
ITrk <E. But then the function x = rk satisfies the required condition of
Definition 11.52. This shows that T is a narrow operator.
It should be noted that a narrow operator need not be compact. For an

example see Exercise 7 at the end of this section.
And now we can state and prove the main result of this section.
Theorem 11.58. If E, ,a) is a non-atomic probability measure space,
then every narrow operator on L1 (,a) satisfies the Daugavet equation.
Proof. Let T: L1 E, ,a) > E, ,a) be a non-zero narrow operator,

and let E> 0. Pick a step function x = akXAk such that = 1 and
ITx Iii > 11Th E. We can assume that fl = 0 for i j. Since Tx is
absolutely continuous, there exists some S > 0 such that
1IXBTXIII <E (*)
for each B E E with p,(B) <8. Fix n E N such that <8.

Since T is narrow, for each 1 k m there exists (by Corollary 11.54)
some measurable subset Bk of Ak such that /i,(Bk) = and the
function hk = l)XBk XAk\Bk satisfies IThk Iii < (jakIi)m Now
consider the function h = akhk. Then 11Th [1' < E and it is easy to
check that x + h = 2ThakXBk. This implies

rn rn
lLx + hili >1 lakl2mIi(Bk) = >1 lxiii = 1.
k=i k=i
Next, let B = \ Bk and y XBTX. Then Tx y=

Furthermore, since
rn rn m
= /2(UBk) = >/i(Bk) = <S,
it follows from (*) that IITx Consequently,
Observe now that the functions x + h and y are disjoint, and hence
ljx+h+yjIi = llx+hlk+ lly[li = 1+ lITxlfi E.
This implies lix + h + Tx[Ii 1 + IITxIli 2. Finally,
II+T1I T)(x +h)jj1 lix + h +TxIli - jIThili
Ix+h+Txlk E 1 +jfTxIIi
Since E> 0 is arbitrary, it follows that 11+ T1I 1 + lIT and so T satisfies
the Daugavet equation. m
We conclude the section by stating without a proof an analogue of The-

orem 11.44 which is due to J. Bourgain and H. P. Rosenthal [71]. Its proof
is quite involved and can be found in [71].
Theorem 11.59 (BourgainRosenthal). If E, ,a) is a non-atomic prob-
ability measure space, then every operator on L1 (,tt) that does not fix a copy
of Li(,a) is narrowand hence it satisfies the Daugavet equation.
Exercises
1. Give an example of a weakly compact operator that is not narrow.
2. Show that each weakly compact operator on L1 (,a) is a narrow operator.
3. If in a scheme of bounded operators E Y _L Z the operator S is
narrow, then show that the operator TS: E + Z is narrow.
4. Give an example of a narrow operator that is not regular.
5. Let T: E Y be a narrow operator. Show that for each A e there
exists a sequence of Rademacher functions } supported by A such that
0.
500 I I The Daugavet Equation
.
6. (PlichkoPopov [267]) Show that the sum of a narrow and a compact

operator is narrow.
7. (PlichkoPopov [267]) Let E = 1}, where 1 < p < oo. Show that the
identity operator I is the sum of two narrow projectionsand conclude
from this that the sum of two narrow operators need not be a narrow
operator. [HINT: Use the Haar system (see Exercise 3 in Section 10.6 for
a precise definition of the Haar functions) to split E into a direct sum of
two complemented subspaces E1 and E2 such that each of these subspaces
has infinite many "packs" of the consecutive Haar functions with pairwise
disjoint supports covering [0, 1].]
8. (PlichkoPopov [267]) Show that each bounded operator on 1],
where I < p < 00, is the sum of two narrow operators. [HINT: Use
Exercises 3 and 7 above.
9. Let E, ,u) be a non-atomic probability space. Show that every gener-
alized sequence of Rademacher functions (A)} satisfies the conclusion
of the RiemannLebesgue lemma, i.e., xE
10. Let E, ,u) be a non-atomic probability space, and let E be a Banach
function space with order continuous norm such that (,u) c E L1 (/2).
Also, let be a sequence of Rademacher functions supported by a

measurable subset A of ft Establish the following.
(a) For each x E E we have
x E L1(1u) we have IIX+XTrdIi llxIk.
11.7. Some Applications of the Daugavet Equation

Recall that by its very definition the Daugavet property is isometric in na-
ture. Therefore, it comes as a surprise that this isometric property has
purely isomorphic consequences. Some of them will be discussed in this
section.
The first result of this type is due to V. Kadets [165] and it informs us
that a Banach space with the Daugavet property cannot have an uncondi-
tional basis.
Theorem 11.60 (Kadets). If a Bariach space satisfies the Daugavet prop-
erty with respect to operators, then it does not have an uncondi-
tional basis.
Proof. Let X be a Banach space satisfying the Daugavet property with

respect to rank-one operators. Assume by way of contradiction that X has
an unconditional basis
For each finite subset A of natural numbers let PA and QA denote the
natural projections on the closed subspaces generated by {ek: k E A} and
{ek: k A}, respectively. Clearly, PA + QA = I and is a finite-
rank operator. By Theorem 11.50, the operator PA satisfies the Daugavet
equation and therefore

IIQAII lIPAIl (*)
Next, consider the subsets of real numbers

V= {11PAII: A is a finite subset of N} and
W= {IIQAt1: A is a finite subset of N}.
If A is an arbitrary finite subset of N and = {1,. . n} n AC, then . ,
clearly IPAThX QAx1I 0 for each x E X. This implies that llQAIl sup V.
Therefore, sup W sup V. On the hand, it follows from (*) that
sup W 1+sup V. Consequently, sup V = sup W = 00. However, according
to Lemma 1.53, supV must be finite, a contradiction. This contradiction
establishes that X cannot have an unconditional basis.
Since we know that the Banach lattices C[0, 1] and L1 [0, 1] satisfy the
Daugavet property for rank-one operators, Theorem 11.60 provides imme-
diately a new and simple proof of the following well-known fact.
Corollary 11.61. Neither C[0, ii nor L1{0, 1] has an unconditional basis.
Another surprising isomorphic consequence forced by the Daugavet

property on a Banach space X is due to V. Kadets, R. Shvidkoy, G. Sirotkin
and D. Werner [167] and is presented next.
Theorem 11.62 (KadetsShvidkoySirotkinWerner). If a Banach space
X satisfies the Daugavet property for rank-one operators, then X contains
a copy of e1 (and hence X cannot be reflexive).
Proof. Let V be a finite dimensional subspace of X and let U be the closed

unit ball of X. If 0 < < 1 is given, then we claim that there exists a
Slice(U,x*,8) such that
lix + vii (1 c)(llx[i + jjvlI)
for each x E Slice(U, 8) and for each v E V.
To establish this claim, choose a finite set {vi,.. . , of unit vectors
that are i-dense in the unit sphere of V. An inductive argument in con-
junction with Lemma 11.46 guarantees the existence of n slices satisfying
Slice( U, Slice( U, 82) J ... Slice( U,
such that
(**)
for each k and for all x E Slice(U, 8k).
In particular, for the last (smallest) slice Slice(U, the inequal-

ity (**) holds for each Ic = 1,... , n. Moreover, if we recall that the set
is i-dense in the unit sphere of V, then (**) shows that
{ 'vl, . . ,
2E
for each x E Slice(U, and for each V E V with = 1. Therefore, by II
Exercise 3 in Section 11.1 for any x E Slice(U, we have

lAx + VII (1
E V and all scalars
Now it is easy to construct a sequence in X whose closed linear span is

isomorphic to Li. Indeed, using inductively the previous part of the proof, we
can construct a sequence of unit vectors ei, e2,... and a decreasing sequence
of slices such that E and
+ (1- En+i)(1A1 +
for each A and for each v in the linear span of {ei,... Without loss
,
of generality, we can assume that 0 for each ri. (This condition

guarantees that ey = Ek) e (0, 1).)
Next, we claim that the inequality
(t)
holds and any choice of scalars

for all ri. The proof is by induction.
,
For ri = 1 notice that II)tieill = (1 Ei)IA1I. Now assume that (t) is

true for some n. Then for any choice of scalars '\i, . )'m+i we have ,
(1- +
(1 + [11(1- Ek)]
n+i n+i
This completes the induction and establishes the validity of (t)

Since (1 Ek) ey for each n, it follows from (t) that for all ri. and
any choice of scalars Ar,. ,
we have
This shows that the sequence is equivalent to the standard unit basis
in and so X contains a copy of 4.
The relationships between the Daugavet property and the geometric

properties of the unit ball of a Banach space are still far from being under-
stood completely. Only a few results in this direction are available. The
next theorem and its consequences are due to P. Wojtaszczyk [339].
Theorem 11.63 (Wojtaszczyk). If a Banach space X satisfies the Daugavet
property with respect to the rank-one operators, then the closed unit ball of
X does not have strongly points.
Proof. Assume by way of contradiction that the closed unit ball of a Banach
space X has a strongly exposed point uO. Pick a linear functional f E
that strongly exposes uosuch that f(uo) = hf hi = 1. Also, fix any non-zero
vector v E X with f(v) = 0.
By our assumption, the rank-one operator T = f v satisfies the Dau-
gavet property, i.e., III+T11 =1 + 1 + lvii. Now take
a sequence of unit vectors such that
IIxn + = Iixn + f(xn)vIl v 1+ IIvII.
From Iixn 1 + . lvii 1 + lvii, it follows that 1.

Replacing by and passing to a subsequence if necessary, we can
assume that f strongly exposes uO, we see that uo.
Therefore,
= lim Ixn = lim = 1+ Ilvil,
flWQ
and this is true for all v E X with f(v) = 0.
show that X is isometric to the 4-sum (Ker f Ruo)i, where

The above
denotes the one-dimensional vector subspace spanned by UO. Finally,
consider the rank-one operator S = f uo and note that for each vector
x = y + Au0 E (Ker f Ruo)i, we have
1(1 - = + Au0 - f(y)uo -

= = lly+Auoll = Mxii.
This shows that II 1 < 1 + IISM, and consequently the rank-one

operator S does not satisfy the Daugavet equation, which is impossible.
This contradiction establishes that the closed unit ball of X has no strongly
exposed points.
Recall that a Banach space Z has the RadonNikodym property if

every continuous operator T: L1 [0, 1] Z can be represented as a Bochner
integral, i.e., there exists an essentially bounded strongly measurable func-

tion [0, 1] + Z such that
Tx =
for all x E Li[0, 1]. We refer to [107] for a comprehensive treatment of the
RadonNikodyrn property that plays an important role in many problems
in analysis.
Corollary 11.64. If a Banach space X has the RadonNikodym property,
then X does not satisfy the Daugavet property for rank-one operators.
Proof. The conclusion follows from Theorem 11.63 by invoking the follow--
ing well-known theorem of R. Phelps: The closed unit ball of a Banach space
with the RadonNikodym property has strongly exposed points. (See
Chapter VII, Theorem 3].)
Corollary 11.65. Every Banach space can be renorrned so that in the new
norm it fails the Daugavet property for rank-one operators.
Proof. It suffices to establish that a given Banach space X can be renormed
in such a way that its closed unit ball has strongly exposed points. To see
that this is possible, take any x0 that does not belong to the original closed
unit ball U and then renorm X by taking for the new closed unit ball the
closure of the circled convex hull of U U {XO}. It remains to be noticed that
is a strongly exposed point of the new closed unit ball.
Exercises
1. Let X be a Banach space and assume that there exists an increasing
continuous function R+ + R and a scalar y> U such that for every
finite-rank projection P we have
Show that X does not have an unconditional basis.

2. Complete the details of the proof of Corollary 11.65.
Bibliography
1. Y. A. Abramovich, Injective envelopes of normed lattices, Soviet Iviath. DokL

12 (1971), 511514.
2. Y. A. Abramovich, Some theorems on normed lattices, Vestnik Leningr. Univ. Mat.
Meh. Astronoin., no. 13 (1971), 511. (English translation: Vestnik Leningr. Univ.
Math. 4(1977), 153159.)
3. Y. A. Abramovich, Weakly compact sets in Dedekind complete topological vector

lattices, Tear. Funkcii Funkcional. Anal. i Prilozen. (Kharkov), no. 15 (1972), 2735.
4. Y. A. Abramovich, A generalization of a theorem of J. Holub, Proc. Amer. Math. Soc.
108 (1990), 937939.
5. Y. A. Abramovich, New classes of Banach spaces on which compact operators satisfy
the Daugavet equation, J. Operator Theory 25 (1991), 331345.
6. Y. A. Abramovich, C. D. Aliprantis, and 0. Burkinshaw, The Daugavet equation in
uniformly convex Banach spaces, J. Funct. Anal. 97 (1991), 215230.
7. Y. A. Abramovich, C. D. Aliprantis, and 0. Burkinshaw, On the spectral radius of
positive operators, Math. Z. 21]. (1992), 593607. Corrigendum: Math. Z. 215 (1994),
167168.
8. Y. A. Abramovich, C. D. Aliprantis, and 0. Burkinshaw, Positive operators on Krein
spaces, Acta Appi. Math. 27 (1992), 122.
9. Y. A. Abramovich, C. D. Aliprantis, and 0. Burkinshaw, An elementary proof of
Douglas' theorem on contractive projections on L1-spaces, J. lvi at h. Anal. Appl.
177 (1993), 641644.
10. Y. A. Abramovich, C. D. Aliprantis, and 0. Burkinshaw, Invariant subspaces of op-
erators on 4-spaces, J. Funct. Anal. 115 (1993), 418424.
11. Y. A. Abramovich, C. D. Aliprantis, and 0. Burkinshaw, Invariant subspaces for
positive operators, J. Funct. Anal. 124 (1994), 95ill.
12. Y. A. Abramovich, C. D. Aliprantis, and 0. Burkinshaw, Invariant subspaces for
positive operators acting on a Banach space with basis, Proc. Amer. Math. Soc.
123 (1995), 17731777.
13. Y. A. Abramovich, C. D. Aliprantis, and 0. Burkinshaw, Local quasinilpotence, cy-
cles and invariant subspaces, in: N Kalton, E. Saab, and S. Montgomery-Smith, eds.,
505
506 Bibliography
Interaction Between Functional Analysis, Harmonic Analysis, and Probability

ture Notes in Pure and Applied Mathematics, 175, Marcel Dekker, New York, 1996),
112.
14. Y. A. Abramovich, C. D. Aliprantis, and 0. Burkinshaw, Another characterization

of the invariant subspace problem, in: C. B. Huijsmans, M. A. Kaashoek, W. A. J.
Laxemburg, and B. de Pagter, eds., Operator Theory in Function Spaces and Banach
Lattices (Operator Theory Advances and Applications, 75, Birkhuser Verlag, Berlin,
1995), 1531.
15. Y. A. Abramovich, C. D. Aliprantis, and 0. Burkinshaw, Multiplication and
friendly operators, Positivity 1 (1997), 171180.
16. Y. A. Abramovich, C. D. Aliprantis, and 0. Burkinshaw, The invariant subspace
problem: Some recent advances, Rend. Istit. Mat. Univ. Trieste 29 (1998), 176.
17. Y. A. Abramovich, C. D. Aliprantis, 0. Burkinshaw, and A. W. Wickstead, A charac-
terization of multiplication operators, Indag. Math. N. S. 10 (1999),
161171.
18. Y. A. Abramovich, C. D. Aliprantis, and I. A. Polyrakis, Positive projections and
lattice-subspaces, Proc. Royal Irish Acad. Sect. A 94(1994), 237253.
19. Y. A. Abramovich, C. D. Aliprantis, and I. A. Polyrakis, Some remarks on surjective
and bounded below operators, Atti Sem. Mat. Fis. Univ. Modena 44(1996), 455464.
20. Y. A. Abramovich, C. D. Aliprantis, and W. R. Zame, A representation theorem for
Riesz spaces and its applications to economics, Econom. Theory 5 (1995), 527535.
21. Y. A. Abramovich, E. L. Arenson, and A. K. Kitover, Banach C(K)-modules and
Operators Preserving Disjointness, Pitman Research Notes in Mathematical Series,
# 277, Longman Scientific & Technical, New York, 1992.
22. Y. A. Abramovich and A. K. Kitover, Inverses of disjointness preserving operators,
Mem. Amer. Math. Soc., 679, Providence, RI, 2000.
23. Y. A. Abramovich, A. I. Veksler, and A. V. Koldunov, On operators preserving dis-
jointness, Soviet Math. Dokl. 20 (1979), 10891093.
24. Y. A. Abramovich and A. W. Wickstead, Remarkable classes of unital AM-spaces, J.
Math. Anal. Appl. 180 (1993), 398411.
25. C. D. Aliprantis and K. C. Border, Infinite Dimensional Analysis: A Hitchhikers

Guide, Edition, SpringerVerlag, Heidelberg and New York, 1999.
26. C. D. Aliprantis, D. J. Brown, and 0. Burkinshaw, Existence and Optimality of

Competitive Equilibria, SpringerVerlag, Heidelberg and New York, 1990.
27. C. D. Aliprantis and 0. Burkinshaw, Positive compact operators on Banach
Math. Z. 174 (1980), 289298.
28. C. D. Aliprantis and 0. Burkinshaw, On weakly compact operators on Banach

Proc. Amer. Math. Soc. 183 (1981), 573578.
29. C. D. Aliprantis and 0. Burkinshaw, Locally Solid Riesz Spaces, Academic Press, New
York and London, 1978.
30. C. D. Aliprantis and 0. Burkinshaw, Positive Operators, Academic Press, New York
and London, 1985.
31. C. D. Aliprantis and 0. Burkinshaw, Principles of Real Analysis, Edition, Aca-
demic Press, New York and London, 1998.
32. C. D. Aliprantis and 0. Burkinshaw, Problems in Real Analysis, Edition, Aca-
demic Press, New York and London, 1998.
Bibliography 507
33. C. D. Aliprantis, 0. Burkinshaw, and P. Kranz, On lattice properties of the composi-

tion operator, Manuscripta Math. 36 (1981), 1931.
34. T. And, Positive operators in semi-ordered linear spaces, J. Fac. Sci. Hokkaido Univ.,
Ser. 1, 13 (1957), 214228.
35, T. And, Contractive projections in Pacific J. Math. 17 (1966), 391405.

36. T. And, Banachverbnde und positiver Projektionen, Math. Z. 109 (1969), 121130.
37. S. I. Ansari, Essential disjointness and the Daugavet equation, Houston J. Math.
19 (1993), 587601.
38. A. B. Antonevich and A. V. Lebedev, Spectral properties of operators with shift, Izv.
Akad. Nauk USSR 5cr. Mat. 47 (1983), 915941.
39. C. Apostol, C. Foias, and D. Voiculescu, Some results on non-quasitriangular operators
W, Math. Pures Appi: 18 (1973), 487514.
40. W. Arendt, tlber das spektrum regularer operatoren, Ph.D. Dissertation, University
of Tbingen, 1979.
41. W. Arendt, A. Grabosch, G. Greiner, U. Groh, H. P. Lotz, U. Moustakas, R. Nagel,
F. Neubrander, and U. Schiotterbeck, One-parameter Semigroups of Positive Oper-
ators, Lecture Notes in Mathematics, 1184, SpringerVerlag, Berlin and New York,
1986.
42: N. Aronszajn and K. T. Smith, Invariant subspaces of completely continuous opera-
Ann. of Math. 60 (1954), 345350.
43. W. B. Arveson and J. Feldman, A note on invariant subspaces, Michigan Math. J.
15 (1968), 6164.
44. F. V. Atkinson, The normal solubility of linear equations in normed spaces, Mat. Sb.
(N.S.) 28(1951), 314.
45. A. Atzmon, Power regular operators, Trans. Amer. Math. Soc. 347 (1995), 3 1013109.
46. A. Atzmon, On the existence of hyperinvariant subspaces, J. Operator Theory 11
(1984), 340.
47. A. Atzmon, An operator without invariant subspaces on a nuclear space, Ann.

of Math. (2) 117 (1983), 669694.
48. A. Atzmon, Operators which are annihilated by analytic functions and invariant sub-
spaces, Acta Math. 144 (1980), 2763.
49. A. Atzmon and G. Godefroy, An applicatioirof variationalprinciple-to
the existence of nontrivial subspaces, C. R. Acad. Sci. Paris Sr. I Math. 332 (2001),
151156.
50. V. F. Babenko and S. A. Pichugov, A property of compact operators in Banach spaces,
Ukrainian Math. J. 33 (1981), 374376.
51. 5. Banach, Thorie des Operations Linaires, Monografje Matematyczne, Warsaw,
1932. (Reprinted by Chelsea Publishing Co., New York, 1955.)
B. Beauzamy, Introduction to Operator Theory and Invariant Subspaces, North-
Holland Mathematical Library, 42, North-Holland, Amsterdam, 1988.
53. B. Beauzamy, Un oprateur sans sous-espace invariant: simplification de I'exemple
de P. Euflo (French) [An operator with no invariant subspace: simplification of the
example of P. Enflo], Integral Equations Operator Theory 8 (1985), 314384.
54. A. Benedek and R. Panzone, The spaces with mixed norm, Duke Math. J. 28(1961),
301324.
55. H. Bercovici, Notes on invariant subspaces, Bull. Amer. Math. Soc. (N.S.) 23 (1990),
136.
508 Bibliography
56. J. Bergh and J. Interpolation Spaces. An Iniroduction, SpringerVerlag,

Heidelberg and New York, 1976.
57, A. Berman, M. Neumann, and R. J. Stern, Nonnegative Matrices in Dynamical Sys-
tems, John Wiley and Sons, New York, 1989.
58. A. Berman and R. J. Plemmons, Nonnegative Matrices in the Mathematical Sciences,
Academic Press, New York, 1979.
59. 5. J. Bernau and H. E. Lacey, The range of a contractive projection on an Lu-space,
Pacific J. Math. 53 (1974), 2141.
60. A. R. Bernstein and A. Robinson, Solution of an invariant subspace problem of K. T.
Smith and P. R. Halmos, Pacific J. Math. 16 (1966), 421431.
61. C. Bessaga and A. Pelczynski, On bases and unconditional convergence of series in
Banach spaces, Studia Math. 17(1958), 151164.
62. A. Bigard and K. Keimel, Sur les endomorphisnies conversant les polaires d'un group
Archimdien, Bull. Soc. Math. France 97 (1969), 381398.
63. 0. Birkhoff, Moyennes de fonctions, algebre et theorie des nombrees, Coloques Inter-
nationaux du Centre National de Ia Recherche Scientifique, no. 24 (1950), 143153.
64. 0. Birkhoff, Reactor criticality in transport theory, Proc. Nat. Acad. Sci. U.S.A.
45 (1959), 567569.
65. 0. Birkhoff, Positivity and criticality, Proc. Sympos. Appi. Math. 10, 116126. Amer.
Math. Soc., Providence, RI, 1961.
66. E. Bishop and R. R. Phelps, The support functionals of a convex set, in: V. KIee
ed., Convexity (Proceedings of Symposia in Pure Mathematics, 7, Amer. Math. Soc.,
Providence, RI, 1963), 2735.
67. R. P. Boas, Invitation to Complex Analysis, The Random HouseBirkhaser Math.
Series, New York, 1987.
68. F. F. Bonsall, Linear operators in complete positive cones, Proc. London. Math. Soc.
8 (1958), 5375.
69. 1 Bourgain, Real isomorphic complex Banach spaces need not be complex isomorphic,
70. J. Bourgain, Strongly exposed points in weakly compact sets in Banach spaces, Proc.
Amer. Math. Soc. 58 197200.
71. J. Bourgain and H. P. Rosenthal, Applications of the theory of semi-embeddings to
Banach space theory, J. Funct. Anal. 52 (1983), 149188.
72. M. Boyle and D. Handelman, The spectra of nonnegative matrices via symbolic dy-
namics, Ann. of Math. 133 (1991), 249316.
73. L. de Branges, The StoneWeierstrass theorem, Proc. Amer. Math. Soc. 10 (1959),
822824.
74. L. de Branges, Factorization and invariant subspaces, J. Math. Anal. App!. 129 (1970),
163200.
75. L. de Branges, A construction of invariant subspaces, Math. Nachr. 163 (1993), 163
175.
76. 5. Brown, Invariant subspaces for subnormal operators, Integral Equations Operator
Theory 1 (1978),
77. A. V. Bukhvalov, Integral representation of linear operators. Zap. Nauchn. Sem.
Leningrad Otdel. Mat. Inst. Steklov. (LOMI) 47(1974), 514. Also, in: J. Soviet Math-
ematics 8(1978), 129137.
Bibliography 509
78. A. V. Bulchvalov, V. B. Korotlcov, A. G. Kusraev, S. S. Kutateladze, and B. M.

Makarov, Vector Lattices and Integral Operators, Nauka, Novosibirsk, 1992. (English
translation edited by S. S. Kutateladze was published by Kiuwer Academic Publishers,
Dordrecht and Boston, 1996.)
79. J. J. Buoni, R. Harte, and A. W. Wiclcstead, Upper and lower Fredhoim spectra, I,
80. R. Burger, Mutation-selection and continuum-of-Alleles models, Iviath. Biosci.
91 (1988), 6783.
81. G. Buskes, Extension of Riesz homomorphisms. I, J. Austral. )vlath. Soc. Ser. A 39
(1985), 107120.
82. G. Buskes, On extension of Riesz homomorphisms. II, J. Miss. Acad. Sci. 33 (1988),
2947.
83. G. Buskes, Extension of Riesz homomorphisms. III, J. Austral. Math. Soc. Ser. A
43 (1987), 3546.
J. W. Calkin, Two-sided ideals and congruences in the ring of bounded operators in
Hubert spaces, Ann. of Math. 42 (1941),
85. S. R. Caradus, W. E. Pfaffenberger, and B. Yood, Calkin Algebras and Algebras of
Operators on Banach Spaces, Lecture Notes in Pure and Applied Mathematics, 9,
Marcel Dekker, New York, 1974.
86. V. Caselles, On irreducible operators on Banach lattices, Indag. Math. 48 (1986), 11
16.
87. J. J. M. Chadwick and A. W. Wickstead, A quotient of ultrapowers of Banach spaces
and semi-Fredhoim operators, Bull. London Math. Soc. 9 (1977), 321325.
88. F. Chauveheid, On a property of compact operators in Banach spaces, Bull. Soc. Roy.
Sci. Liege 51 (1982), 371378.
89. Z. L. Chen and A. W. Wickstead, Some applications of Rademacher sequences in
Banach lattices, Positivity 2 (1998), 171191.
90. B. Chevreau, W. S. Li, and C. Fearcy, A new Lomonosov lemma, Integral Equations
Operator Theory 40 (1998), 409417.
91. M.-D. Choi, E. A. Nordgren, H. Radjavi,F. Rosenthal, and Y. Zhong, Triangularizing
semigroups of quasinilpotent operators with non-negative entries, Indiana Univ. Math.
J. 42 (1993), 1525.
92. J. A. Clarkson, Uniformly convex spaces, Trans. Amer. Math. Soc. 40 (1936), 396414.
93. P. F. Conrad and J. E. Diem, The ring of polar preserving endomorphisms of an

Abelian lattice-ordered group, Illinois J. Math. 15 (1971), 222240.
94. J. Conway, A Course in Operator Graduate Studies in Mathematics, 21,
Amer. Math. Providence, RI, 2000.
95. H. G. Dales, Automatic continuity: a survey, Bull. London Math. Soc. 10 (1978),
129183.
96. H. G. Dales and J. Esterle, Discontinuous homomorphisms from C(X), Bull. Amer.
Math. Soc. 83 (1977), 257259.
97. H. G. Dales, R. J. Loy, and G. A. Willis, Homomorphisms and derivatives from 13(E),
J. Funct. Anal. 120 (1994), 201219.
98. I. K. Daugavet, On a property of compact operators in the space C, Uspekhi Mat.
Nazth 18 (1963), 157-158. (Russian)
99. K. R. Davidson and H. A. Domingo, Decomposition of Banach space operators, Indi-
ana Univ. Math. J. 35 (1986), 333343.
100. A. M. Davie, Invariant subspaces for Bishop's operators, Bull. London Math. Soc.
6(1974), 343348.
101. M. M. Day, The spaces with 0 < p < 1, Bull. Amer. Math. Soc. 46(1940),
816823.
102. M. M. Day, Normed Linear Spaces, edition, SpringerVerlag, Heidelberg and
New York, 1973.
103. D. Deckard, EL G. Douglas, and C. Pearcy, On invariant subspaces of quasitriangular
operators, Amer. Math. J. 91 (1969),
104. U. Diallo and P. P. On the DaugavetKrasnoselsky theorem, Vestnik
Acad. Nauk BSSR Ser. Fiz. Mat. Nauk, no. 3 (1987), 2631. (Russian)
105. J. Diestel, Sequences and Series in Banach Spaces, Graduate Texts in Mathematics,
92, SpringerVerlag, Heidelberg and New York, 1984.
106. J. Diestel, H. Jarchow, and A. Tong, Absolutely Summing Operators, Cambridge
Studies in Advanced Mathematics, 43, Cambridge University Press, 1995.
107. J. Diestel and J. J. Uhi, Jr., Vector Measures, Math. Surveys, 15, Amer. Math. Soc.1
Providence, Rhode Island, 1977.
108. J. Dieudonn, Foundations of Modern Analysis, Academic Press, Pure and Applied
Math Series, X, New York and London, 1960.
109. P. 0. Dodds and D. H. Fremlin, Compact operators in Banach lattices, Israel J.
Math. 34 (1979), 287320.
110. P. G. Dodds, C. B. Characterizations of conditional

and B. de Pagter,
expectation-type operators, Pacific J. Math. 141 (1990), 5577.
111. J. L. Doob, Measure Theory, Graduate Texts in Mathematics, 143, Springer-Verlag,
Heidelberg and New York, 1994.
112. Ft. G. Douglas, Contractive projections on an Li-space, Pacific J. Math. 15 (1965),
443462.
113. H. FL Dowson, Spectral Theory of Linear Operators, Academic Press, New York and
London, 1978.
114. Ft. Common invariant subspaces for collections of operators, Integral Equa-
tions Operator Theory 39 (2001), 253266.
115. Ft. D. L. Livshits, G. MacDonald, M. Omladii, and
H. Ftadjavi, An irreducible semigroup of non-negative squarezero operators, Integral
Equations Operator Theory, forthcoming.
116. J. Duncan, C. M. McGregor, J. D. Price, and A. J. White, The numerical index o(
a normed space, J. London Math. Soc. (2) 2 (1970), 481488.
117. N. Dunford, Integration and linear operations, Trans. Amer. Math. Soc. 40(1936),
474494.
118. N. Dunford and J. T. Schwartz, Linear Operators I, Wiley (Interscience), New York,
1958.
119. E. M. Dyrikin, An operator calculus based on the CauchyGreen formula, Zap.
Nauchn. Sem. Leningrad Otdel. Mat. Inst. Steklov. (LOM1) 30 (1972), 3339.
120. P. Enflo, A counterexample to the approximation property in Banach spaces, Acta
Math. 130 (1973), 309317.
121. P. Enflo, On the invariant subspace problem for Banach spaces, Seminaire Maurey
Schwarz (19751976); Acta Math. 158 (1987), 213313.
122. Z. Ercan and S. Onal, Invariant subspaces for positive operators acting on a Banach
space with Markushevich Positivity, forthcoming.
Bibliography 511
123. 1 Flores and F. L. Hernndez, Domination by positive strictly singular operators,

(1. London Math. Soc., forthcoming.
124. 1 Flores and F. L. Hernndez, Domination by positive disjointly strictly singular
operators, Proc. Amer. Math. Soc. 129 (2001), 19791986.
125. J. Flores and F. L. Hernndez, On positive strictly singular operators and domina-
tion, preprint.
126. C. Foias and I. Singer, Points of diffusion of linear operators and almost diffuse
operators in spaces of continuous functions, Math. Z. 87(1965), 434450.
127. I. Fredholm, Sur une classe d'quations fonctionnelles, Acta Math. 27 (1903), 365
390.
128. D. H. Fremlin, Topological Riesz Spaces and Measure Theory, Cambridge University
Press, London and New York, 1974.
129. H. Freudenthal, Teilweise geordnete moduin, Nederl. Akad. Wetensch. Proc. Ser. A
39 (1936), 641651.
130. G. Frobenius, Uber die Darstellung der endlichen Gruppen durch lineare Substitu-
tionen, K. Preuss. Akad. Wiss. Berlin (1897), 9941015.
13L G. Frobenius, Uber Matrizen aus nicht-negativen Elementen, Sitz. Berichte Kgl.
Akad. Wiss. Berlin, 456477, 1912.
132. V. Gantmacher, Uber schwache totaistetige operatoren, Mat. Sb. (N.S.) 7(49) (1940),
301308.
133. I. M. Gelfand, Normierte Ringe, Mat. Sb. (N.S.), 9(51) (1941), 324.
134. 5. A. Gershgorin, Uber die Albrenzung Eigenwerte einer Matrix, Izv. Akad. Nauk
SSSR Ser. Fiz.-Mat. 6 (1931), 749754.
135. N. Ghoussoub, Positive embeddings of L1, (F), and Math. Ann.
262 (1983), 461472.
136. N. Ghoussoub and H. P. Rosenthal, G5-embeddings and quotients of L1, Math. Ann.
264(1983), 321332.
137. W. T. Cowers and B. Maurey, The anconditional sequence problem, J. Amer. Math.
Soc. 6 (1993), 851874.
138. W. T. Gowers and B. Maurey, Banach spaces with small spaces of operators, Math.
Ann. 307 (1997), 541568.
139. J. J. Grobler, Band irreducible operators, Indag. Math. 48 (1986), 405409.
140. A. L. Grornov, Invariant subspaces of weighted permutation operators, Funktsional.
Anal. i Prilozhen 22 (1988), 7576. (Russian)
141. A. Grothendieck, Produits tensoriels toplogiques et espaces nuclaires, iviemn. Amer.
Math. Soc., 16, Providence, RI, 1965.
142. A. Goullet de Rugy, La structure idale des J. Math. Pures et AppI.
51 (1972), 331373.
143. D. W. Hadwin, An operator still not satisfying Lomonosov's hypothesis, Proc. Amer.
Math. Soc. 123(1995), 30393041.
144. W. Hadwin, E. A. Nordgren, H. Radjavi, and P. Rosenthal, An operator not

satisfying Lomonosov's hypothesis, J. Funct. AnaL 38 (1980), 410415.
145. P. R. Halmos, Finite-Dimensional Vector Spaces? 2nd Edition, Van Nostrand Co.,
New York and London, 1958. Also, reprinted in the series: Undergraduate Texts in
Mathematics, Springer-Verlag, New York and Heidelberg, 1974.
512 Bibliography
146. P. R. Halmos, Invariant subspaces of polynomially compact operators, Pacific J.

Math. 16 (1966), 433437.
147. P. R. Halmos, Quasitriangular operators, Acta Sci. Math. (Szeged) 29 (1968), 283
293.
148. P. R. Halinos, A Hubert Space Problem Book, 2nd Edition, Graduate Texts in Math-
ematics, 19, SpringerVerlag, New York and Heidelberg, 1982.
149. P. R. Halmos and V. S. Sunder, Bounded Integral Operators on L2 Spaces, Springer
Verlag, Berlin and New York, 1978.
150. F. L. Hernndez and 13. Rodriguez-Salinas, On "-complirnented copies in Orlicz
spaces II, Israel J. Math. 68 (1989), 2755.
151. W. Hildenbrand and H. Sonnenschein, Eds., Handbook of Mathematical Economics,
Vol. IV, North-Holland, Amsterdam and New York, 1991.
152. E. Hille and R. S. Phillips, Functional Analysis and Semigroups, Amer. Math. Soc.
Colloq. Pubi., 36, Providence, RI, 1974.
153. J. Ft. Holub, A property of weakly compact operators on C[0,1}, Proc. Amer. Math.
Soc. 97 (1986), 396398.
154. J. Ft. Holub, Daugavet's equation and operators on L' Proc. Amer. Math. Soc.
100 (1987), 295300.
155. J. Ft. Holub, Daugavet's equation and ideals of operators on C[0, 1], Math. Nachr.
141 (1989), 177181.
156. Ft. B. Honor, Density and transitivity results on and 1, J. London Math. Soc.
32 (1985), 52 1527.
157. Ft. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University Press, London
and New York, 1985.
158. M. T. Jahandideh, Decomposability and Triangularizability of Positive Operators on
Banach Lattices, Ph.D. Dissertation, Dalahousie University, Halifax, Canada,1997.
159. Ft. C. James, Bases and reflexivity in Banach spaces, Ann. of Math. 52 (1950), 518
527.
160. R. C. James, Bases in Banach spaces, Amer. Math. Monthly 89 (1982), 625640.
161. G. Jameson, Ordered Linear Spaces, Lecture Notes in Mathematics, 141, Springer
Verlag, Berlin and New York, 1970.
162. P. Jentzsch, Uber Integraigleichungen mit positivem Kern, J. Reine Angew. Math.
141 (1912), 235244.
163. C. L. Jiang, Stampfli's theorem for operators on some Banach spaces, Northeast.
Math. J. 11 (1995), 5660.
164. W. B. Johnson and J. Lindenstrauss, Eds., Handbook of the Geometry of Banach

Spaces, Vol. 1, North-Holland, Amsterdam and New York, 2001.
165. V. M. Kadets, Some remarks concerning the Daugavet equation, Quaestiones Math.
19 (1996), 225235.
166. V. M. Kadets, Ft. V. Shvidkoy, G. G. Siroticin, and D. Werner, Espaces de Banach
ayant la proprit de Daugavet, C. R. Acad. Sci. Paris Sr. I Math. 325 (1997), 1291
1294.
167. V. M. Kadets, Ft. V. Shvidkoy, G. G. Sirotkin, and D. Werner, Banach spaces with
the Daugavet property, Trans. Amer. Math. Soc. 352 (2000), 855873.
168. G. K. Kalisch, On similarity, reducing manifolds, and unitary equivalences of certain
Volterra operators, Ann. of Math. 66 (1957), 481494.
________
Bibliography 513
169. N. I Kalton, An elementary example of a Banach space not isomorphic to its complex
conjugate, Canadian Math. Bull. 38 (1995), 218222.
170. N. 3. Kalton, N. T. Peck, and W. Roberts, Art F-Sampler, London Math. Soc.
Lecture Notes Series, 89, Cambridge Univ. Press, London and New York, 1984.
171. H. Kamowitz, A property of compact operators, Proc. Amer. Math. Soc. 91 (1984),
23 1236.
172. L. V. Kantorovich, On partially ordered linear spaces and their applications in the
theory oflinear operators, Doki. Akad. Nauk SSSR 4(1935), 1316. (Russian)
173. L. V. Kantorovich, Sur les proprits des espaces semi-ordonns Iinaires, C. 1?. A cad.
Sci. Paris Ser. A-B 202 (1936), 813816.
174. L. V. Kantorovich and G. P. Akilov, Functional Analysis, Pergamon Press, Oxford
and New York, 1982.
175. L. V. Kantorovich, B. Z. Vulikh, and A. G. Pinsker, Functional Analysis in Partially
Ordered Spaces, Gostekhizdat, Moscow and Leningrad, 1950.
176. 5. Kaplan, The Bidual of C(X), I, North-Holland, Amsterdam and New York, 1985.
177. 5. Karlin, Positive operators, J. Math. Mech. 8 (1959), 907937.
178. T. Kato, Perturbation theory for nullity deficiency and other quantities of linear
operators, J. Analyse Math. 6 (1958), 273322.
179. Y. Katznelson and L. Tzafriri, On power-bounded operators, J. Funct. Anal.
68 (1986), 313328.
180. J. L. Kelley, Averaging operators on Illinois J. Math. 2 (1958), 214223.

181. R. KhaIil, The Daugavet equation in vector-valued function spaces, Panamer. Math.
J. 6(1996), 5153.
182. M. A. Khan and N. C. Yannelis (Editors), Theory in Infinite Dimensional
Spaces, Studies in Economic Theory, 1, SpringerVerlag, Heidelberg and New York,
1991.
183. J. Kim, The characterization of a lattice homomorphism, Canadian J. Math.
27 (1975), 172175.
184. K. H. Kim, N. S. Ormes, and F. W. Roush, The spectra of nonnegative integer
matrices via formal power series, J. Amer. Math. Soc. 13 (2000), 773806.
185. A. K. Kitover, The spectral properties of weighted homomorphisms in algebras of
continuous functions and Otdel. Mat.
Inst. Steklov. (LOMI) 107 (1982), 89103.
186. A. K. Kitover, On the spectrum of weighted automorphisrns and on a Karnowittz
Shineberg theorem, Funct. Anal. Appl. 13(1979), 7071.
187. A. N. Kolmogorov, Cnrndbegriffe der Wahrscheinlichkeitrechnung, Ergebnisse der
Mathematik, Vol. 2, no. 3, 1933. [English Translation: Foundations of the Theory of
Probability, Chelsea Pubi. Co., New York, 1950.]
188. V. B. Korotkov, Integral Operators, Nauka, Novosibirsk, 1983.
189. G. Kthe, Topological Vector Spaces I, SpringerVerlag, New York and Heidelberg,
1969.
190. M. A. Krasnoselsky, A class of linear operators in the space of abstract continuous
functions, Math. Notes 2 (1967), 856858.
191. M. A. Krasnoselsky, P. P Zabreilco, E. I. Pustilnik. and P. E. Sobolevsky, Integral
Operators in Spaces of Sumrnable Functions, Nauka. Moscow, 1966. (Translated from
the Russian by T. And& Monographs and Textbooks on Mechanics of Solids and
Fluids, Mechanics: Analysis, Noordhoff International Publishing, Leiden, 1976.)
514 --- .---- Bibliography
192. M. G. Krein and M. A. Rutman, Linear operators leaving invariant a cone in a

Banach space, Uspekhi Mat. Nauk 3 (1948), 395 (Russian). Also, Amer. Math. Soc.
Transl. Ser., 2, 26, Providence, RI, 1950.
193. U. Krengel, Uber den Absolutbertrag stetiger linearer Operatoren und seine Anwen-
dung auf ergodische Zerlegungen, Math. Scand. 13 (1963), 151187.
194. H. J. Krieger, Beitrge zur Theorie positiver Operatoren, Schriftenreihe der Institute
fr Math. Reihe A, Heft 6; Akad.-Verlag, Berlin, 1969.
195. A. G. Kusraev, Dominated Operators, Mathematics and its Applications, 519,
Kluwer Academic Publishers, Dordrecht & London, 2000.
196. S. Kwapien, On the form of a linear operator in the space of all measurable functions,
Btill. Acad. Polon. Sci. 21 (1973), 951954.
197. H. E. Lacey, The Isometric Theory of Classical Banach Spaces, SpringerVerlag,
Berlin and New York, 1974.
198. P. Lancaster and M. Tismenetsky, The Theory of Matrices with Applications, 2nd
Edition, Academic Press, Orlando and New York, 1985.
199. K. B. Laursen and M. M. Neumann, An Introduction to Local Spectral Theory,
Clarendon Press, Oxford, 2000.
200. N. J. Laustsen, Maximal ideals in the algebra of certain Banach spaces, University
of Leeds Preprint Series, no. 1, 2000.
201. L. Le Cam, Asymptotic Methods in Statistical Decision Theory, SpringerVerlag,
New York and Berlin, 1986.
202. C.-S. Lin, Generalized Daugavet equations and invertible operators on uniformly
convex Banach spaces, J. Math. Anal. AppI. 197 (1996), 518528.
203. J. Lindenstrauss, On operators which attain their norm, Israel J. Math. 5(1963),
139148.
204. J. Lindenstrauss and L. Tzafriri, On the complemented subspaces problem, Israel J.
Math. 9 (1971), 263269.
205. J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces I, SpringerVerlag, Berlin
and New York, 1977.
206. J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces II, SpringerVerlag, Berlin
and New York, 1979.
207. Z. Lipecki, Extension of vector-lattice homomorphisms, Proc. Amer. Math. Soc.
79 (1980), 247248.
208. Z. Lipecki, D. Plachky, and V'J. Thomsen, Extension of positive operators and ex-
treme points I, Colloq. Math. 42 (1979), 279284.
209. G. L. Litvinov and V. I. Lomonosov, Density theorems in locally convex spaces and
applications, Selecta Math. Soviet 8 (1989), 323339.
210. Ju. I. Ljubi and V. I. Macaev, Operators with separable spectrum, Mat. Sb. (N.S.)
56 (1962), 433468.
211. V. I. Lomonosov, Invariant subspaces of the family of operators that commute with a
completely continuous operator, Funktsional. Anal. i Prilozhen 7(1973), No. 3, 5556.
(Russian)
212. V. I. Lomonosov, An extension of Burnside's theorem to infinite-dimensional spaces,

Israel J. Math. 75 (1991), 329339.
213. V. I. Lomonosov, A counterexample to the BishopPhelps theorem in complex spaces,
Israel J. Math. 115 (2000), 2528.
Bibliography 515
214. E. R. Lorch, The integral representation of weakly almost-periodic transformations

in reflexive vector spaces, Trans. Amer. Math. Soc. 49 (1941), 1840.
215. H. P. Lotz, Uber das Spektrum positiver Operatoren, Math. Z. 108 (1968), 1532.
216. H. P. Lotz, Minimal and reflexive Banach lattices, Trans. Amer. Math. Soc.
211 (1975), 85100.
217. H. p. Lotz and H. H. Schaefer, Uber einen Satz von F. Niiro und I. Sawashima, Math.
Z. 108 (1968), 3336.
218. A. R. Lovaglia, Locally uniformly convex Banach spaces, Trans. Amer. Math. Soc.
78 (1955), 225238.
219. G. Ya. Lozanovsky, On almost integral operators in KB-spaces, Vestmik Leningrad
Univ. Mat. Mekh. Astronom., no. 7(1966), 3544. (Russian)
220. G. Ya. Lozanovsky, On some Banach lattices, Siberian Math. J. 10 (1969), 584599.
221. G. Ya. Lozanovsky, On a theorem of N. Dunford, Izv. Vyssh. Uchebn. Zaved. Mat.
8(1974), 5859. (Russian)
222. W. A. J. Luxemburg and B. de Pagter, Maharam extensions of positive operators
and f-modules, Positivity 6 (2002), 147190.
223. W. A. J. Luxemburg and A. R. Schep, An extension theorem for Riesz homomor-
phisms, Indag. Math. 41 (1979), 145154.
224. W. A. J. Luxemburg and A. C. Zaanen, Notes on Banach function spaces, IX, Indag.
Math. 26 (1964), 360376.
225. W. A. J. Luxemburg and A. C. Zaanen, Riesz Spaces I, North-Holland, Amsterdam,
1971.
226. W. A. J. Luxemburg and A. C. Zaanen, The linear modulus of an order bounded
linear transformation, Indag. Math. 33 (1973), 422447.
227. G. W. MacDonald, Invariant subspaces for Bishop-type operators, J. Funct. Anal.
91 (1990), 287311.
228. D. Maharam, The representation of abstract integrals, Trans. Amer. Math. Soc.
75 (1953), 154184.
229. D. Maharam, On kernel representation of linear operators, Trans. Amer. Math. Soc.
79 (1955), 229255.
230. D. Maharam, On positive operators. Conference in modern analysis and probability
(New Haven, Conn., 1982), 263277, Conternp. Math., 26, Amer. Math. Soc., Provi-
dence, RI, 1984.
231. P. Mankiewicz, A superreflexive Banach space X with L(X) admitting a homomor-
phism onto the Banach algebra C(pN), Israel J. Math. 65 (1989), 116.
232. A. A. Markov, Some theorems on Abelian collections, Doki. Akad. Naiik USSR
10 (1936), 311313.
233. M. Mbekhta, de la decomposition de Kato aux oprateurs para-
normeaux et spectraux, Glasgow Math. J. 29 (1987), 159175.
234. M. Mbekhta, Sur la thorie spectrale locale et limite des nilpotents, Proc. Amer.
Math. Soc. 110 (1990), 621631.
235. R. E. Megginson, An Introduction to Banach Space Theory, Graduate Texts in Math-
ematics, 183, SpringerVerlag, New York and Heidelberg, 1998.
236. A. A. Mekier, On a representation of stochastic projections, Systems of Processing
and Transferring of Information, Trudy (LIAP), no. 87 (1974), 115118.
Bibliography
237. A. A. Mekier, Operators Averaging Over cr-Subalgebras on Ideals in L1 (hz), Disserta-

tion, Leningrad Institute of Airapparatus Construction (LIAP) , Leningrad, 1997.
238. A. A. Mekier, Intermediate spaces and bistochastic projections, Math. Issledovaniya
10 (1975), 270275. (Kishinev).
239. P. Meyer-Nieberg, Quasitriangulierbare Operatoren und invariante Tinter-

vekrorrume stetiger linearer Operatoren, Arch. Math. (Basel) 22 (1971), 186199.
240. P. Meyer-Nieberg, Banach Lattices, SpringerVerlag, Berlin and New York, 1991.
241. A. J. Michaels, Hilden's simple proof of Lomonosov's invariant subspace theorem,
Adv. in Math. 25 (1977), 5658.
242. D. P. Milman, On some criteria for the regularity of spaces of the type (B), Doki.
Akad. Nauk SSSR, N. 5., 20 (1938), 243246. (Russian)
243. H. Minc, Nonnegative Matrices, John Wiley and Sons, Berlin and New York, 1988.
244. B. S. Mityagin and I. C. Edeistein, Homotopy of linear groups for two classes of
Banach spaces, Funkt. Analiz Priloz., (3), 4(1970), 6172. (Russian)
245. 5. Miyajima, Structure of Banach quasi-sublattices, Hokkaido Math. J. 12 (1983),
8391.
246. ST. C. Moy, Characterization of conditional expectation as a transformation on

function spaces, Pacific J. Math. 4(1954), 4763.
247. J. R. Munkres, Topology, Prentice-Hall, Englewood Cliffs, NJ, 1975.
248. R. J. Nagel, Order in pure and applied functional analysis, rrends in Functional
Analysis and Approximation Theory (Acquafredda di Maratea, 1989), Univ.
Modena, Modena, 1991.
249. B. Sz.Nagy and C. Foias, Quasi-similitude des oprateurs et sous-espaces invariants
C. R. Acad. Sci. Paris Sr. I Math. 261 (1965), 39383940.
250. B. Sz.-Nagy and C. Foias, Harmonic Analysis of Operators on Hubert Space, North-
Holland, Amsterdam and New York, 1970.
251. H. Nakano, Modern Spectral Theory, Maruzen Co., Tokyo, 1950.
252. H. Nakano, Modulared Semi-Ordered Linear Spaces, Maruzen Co., Tokyo, 1950.
253. H. Nalcano, Product spaces of semi-ordered linear spaces, J. Math. Fac. Sci. Hokkaido
12 (1953), 163210.
254. C. Neumann, Untersuchungen ber das logarithmische und Newtonsche Potential,
Teubner, Leipzig, 1877.
255. J. von Neumann, Zur Algebra der Funktionaloperationen und Theorie der normalen
Operatoren, Math. Ann. 102 (1929), 370427.
256. J. Neveu, Mathematical Foundations of the Calculus of Probability, Holden & Day,
San Fransisco London, 1965. (English translation from the 1963 French edition.)
257. K. Nikolskii, Invariant subspaces in operator theory and function theory, Mathe-
matical Analysis 12 (1974), 199412, Akad. Nauk SSSR Vsesojuz. Inst. Naun. i Tehn.
Informacii, Moscow, 1974. (Russian)
258. R. D. Nussbaum, The radius of the essential spectrum, Duke Math. J. 37(1970),
473478.
259. R. D. Nussbaum, Eigenvectors of nonlinear positive operators and the linear Krein
Rutman theorem, in: E. Fadell and G. Fourier eds., Fixed Point Theory (Sherbrooke,
Que., 1980), 886, SpringerVerlag, Berlin and New York, 1981, 309330.
260. B. de Pagter, Irreducible compact operators, Math. Z. 192 (1986), 149153.
Bibliography 517
261. C. Pearcy and N. Salinas, An invariant-subspace theorem, Michigan iViath. I

20(1973), 2131.
262. C. Pearcy and A. L. Shields, A survey of the Lomonosov technique in the theory
of invariant subspaces, in Topics in Operator Theory, 219229, i'vlath. Surveys, 13,
Amer. Math. Soc., Providence, RI, 1974.
263. A. Pelczynski and Z. Semadeni, Spaces of continuous functions. IlL Spaces C(Q) for
Q without perfect subsets, Studia A'Iath. 18 (1959), 211222.
264. A. L. Peressini, Ordered Topological Vector Spaces, Harper and Row, New York and
London, 1967.
265. 0. Perron, Zur Theorie der Matrizen, Math. Ann. 64(1907), 248263.
266. A. Pietsch, Operator Ideals, VEB, Deutscher Verlag der Wissenschaften, Berlin, 1978.
267. A. IVI. Plichko and M. M. Popov, Symmetric function spaces on atomless probability
spaces, Dissertationes Mathematicae, 306, 1990.
268. I. A. Polyrakis, Schauder bases in locally solid lattice Banach spaces, Math. Proc.
Cambridge Philos. Soc. 101 (1987), 91105.
269. I. A. Polyrakis, Lattice subspaces of C[0, 1] and positive bases, J. Math. Anal. Appi.
184 (1994), 118.
270. I. A. Polyralcis, Finite-dimensional latticesubspaces of C(Q) and curves of IWL, Trans
Amer. Math. Soc. 348 (1996), 27932810.
271. H. Radjavi, On reducibility of semigroups of compact operators, Indiana Univ. Math.
J. 39 (1990), 499515.
272. H. Radjavi and P. Rosenthal, Invariant Sttbspaces, SpringerVerlag, Berlin and New
York, 1973.
273. H. Radjavi and P. Rosenthal, Simultaneous Triangularization, SpringerVerlag,
Berlin and New York, 2000.
274. B. Randrianantoanina, Norm-one projections in Banach spaces, Taiwanese J. Math.
5(2001), 3595.
275. C. J. Read, A solution to the invariant subspace problem on the space Bull.
London iVIath. Soc. 17 (1985), 305317.
276. C. J. Read, A short proof concerning the invariant subspace problem, J. London
Math. Soc. (2) 34 (1986)7 335348.
277. C. J. Read, Quasinilpotent operators and the invariant subspace problem, J. London
Math. Soc. (2) 56 (1997), 595606.
278. F. Riesz, Sur la decomposition des operations linaires, Atti. Congr. Internaz. Mat.
Bologna 3 (1930), 143148.
279. F. Riesz, Sur quelques notions fondamentals dans Ia theorie gnrale des operations
linaires, Ann. of Math. 41 (1940), 174206.
280. F. Riesz and B. Sz.-Nagy, Fttnctional Analysis, Frederick IJngar Publishing Co., New
York, 1965.
281. H. P. Rosenthal, On factors of C([0, 1]) with non-separable dual, Israel J. Math.
13 (1972), 36 1378; Correction. Ibid. 21 (1975), 9394.
282. H. P. Rosenthal, Sign-embeddings of L', Banach spaces, Harmonic Analysis, and
Probability Theory (Storrs, Conn., 1980/1981), 155165, Lecture Notes in Mathemat-
ics, 995, SpringerVerlag, Berlin and New York, 1983.
283. H. P. Rosenthal, Some remarks concerning sign-embedding, Seminar on the geometry
of Banach spaces (Paris, 1982), 129132; PubI Math. Univ. Paris VII 16, Univ. Paris
VII, Paris, 1983.
518 Bibliography
284. P. Rosenthal, Equivalents of the invariant subspace problem, in Paul Halmos: Cele-
brating 50 Years of Mathematics, SpringerVerlag, Berlin and New York, 1991, 179
188.
285. 0. C. Rota, On the representing of averaging operators, Rend. Padova 30 (1960)7
5264.
286. W. Rudin, Continuous functions on compact spaces without perfect subsets, Proc.
Amer. Math. Soc. 8 (1957), 3942.
287. H. H. Schaefer, Spekraleigenschaften positwer linearer Operatoren, Math. Z.
82 (1963), 303313.
288. H. H. Schaefer, Topologische Nilpotenz irreduzibler Operatoren, Math. Z. 117 (1970),
135140.
289. H. H. Schaefer, Topological Vector Spaces, SpringerVerlag, Berlin and New York,
1974.
290. H. H. Schaefer, Banach Lattices and Positive Operators, SpringerVerlag, Berlin arid
New York, 1974.
291. H. H. Schaefer, On the o-spectrum of order bounded operators, Math. Z. 154 (1977),
7984.
292. H. H. Schaefer, On theorems of de Pagter and AndOKrieger, Math. Z. 192 (1986),
155157.
293. E. Scheffold, Das Spektrum von Verbandsoperatoren in Banachverbnden, Math. Z.
123 (1971), 177190.
294. A. R. Schep, Kernel operators, Indag. Math. 41 (1979), 3953.
295. T. Schlumprecht and V. 0. Troitsky. On quasi-affine transforms of Read's operator,
Proc. Amer. Math. Soc., forthcoming.
296. K. D. Schmidt, On the modulus of weakly compact operators and strongly additive
vector measures, Proc. Amer. Math. Soc. 102 (1988), 862866. -
297. K. D. Schmidt, Daugavet's equation and orthomorphisms, Proc. Amer. Math. Soc.
108 (1990), 905911.
298. H.-U. Schwarz, Banach Lattices and Operators, Tebner Texte, 71, Leipzig, 1984.
299. 0. L. Seever, Nonnegative projections on Co(X), Pacific J. Math. 17(1966), 159166.
300. Z. Semadeni, Banach Spaces of Continuous Functions, Polish Scientific Publishers,
Warsaw, 1971.
301. V. S. Shulmari, On invariant subspaces of Volterra operators, Funktsional Anal. i
Prilozhen 18(1984), 84-85.
302. A. An extension of Lomonosov's techniques to non-compact operators,
Trans. Amer. Math. Soc. 348 (1995), 975995.
303. A. Simoni, A construction of Lomonosov functions and applications to the invariant
subspace problem, Pacific J. Math. 175 (1996), 257270.
304. A. M. Sinclair, Automatic Continuity of Linear Operators, Cambridge University
Press, Cambridge, 1976.
305. I. Singer, Bases in Banach Spaces, Vol. 1, SpringerVerlag, Berlin and New York,
1970.
306. 1. Singer, Bases in Banach Spaces, Vol. 2, SpringerVerlag, Berlin and New York,
1981.
307. 0. 0. Sirotkin, Compact-friendly multiplication operators on Banach function spaces,
J. Funct. Analysis, forthcoming.
Bibliography 519
308. V. L. Smulian, Sur la structure de la sphere unitaire dans l'espace de Banach, Mat.
Sb. (N.S.) 9(1941), 545-561.
309. A. Sobczyk, Projection of the space m on its subspace CO Bull Amer. Math. Soc.
,
47 (1941), 938947.
310. A. Spaisbury, Operators not positive with respect to any basis, Quaestiones Math.
23 (2000), 489494.
311. J. ci Stampfii, Compact perturbations, normal eigenvalues and a problem of Salinas,
J. London Math. Soc. (2) 9 (1974/75), 165475.
312. J. Synnatzschke, The almost integral operators in K-spaces, Vesinik Leningrad Univ.
iVlath. Mekh. Astronom. 13 (1971), 8189.
313. J. Synnatzschke, On the adjoint of a regular operator and some of its applications to
the question of complete continuity and weak continuity of regular operators, Vestnik
Leningrad Univ. Math. Mekh. Astronom. 5(1978), 7181.
314. J. Synnatzschke, Uber eine additive Normgleichheit fur Operatoren in Ba-
nachverbnden, Math. Nachr. 117(1984), 175180.
315. A. Szankowski, A Banach lattice without the approximation property, Israel J. Math.
24(1976), 329337.
316. B.-S. Tam, A cone-theoretic approach to the spectral theory of positive linear oper-
ators: the finite dimensional case, Taiwanese J. Math. 5 (2001), 207277.
317. A. Taylor and D. C. Lay, Introduction to Functional Analysis, R. E. Krieger, Malabar,
Florida, 1986.
318. T. Terzioglu, A characterization of compact linear mappings, Arch. Math. (Basel)
22 (1971), 7678.
319. V. 0. Troitsky, On the modulus of C. J. Read's operator, Positivity 3 (1998), 257
264.
320. V. G. Troitsky, Invariant Subspace Problem and Speciral Properties of Bounded Lin-
ear Operators on Banach Spaces, Banach Lattices, and Topological Vector Spaces,
Ph.D. Dissertation, University of Illinois, Urbana-Champaign, Illinois, 1999.
321. V. G. Troitsky, Lomonosov's theorem cannot be extended to chains of four operators,
322. V. G. Troitsky, Measures of non-compactness of operators on Banach lattices,
preprint.
323. S. L. Troyanski, On locally uniformly convex and differentiable norms in certain
non-separable Banach spaces, Studia Math. 37 (1970/71), 173180.
324. Y. V. Turovskii, Volterra semigroups have invarinat subspaces, J. Funct. Anal.
162 (1999), 313322.
325. L. Tzafriri, Remarks on contractive projections in Israel J. Math. 7(1969),
915.
326. L. Tzafriri, An isomorphic characterization of and co-spaces, II, Michigan Math.
J. 18 (1971), 2131.
327. J. Voight, The projection onto the center of operators in a Banach lattice, Math. Z.
199 (1988), 115117.
328. B. Z. Vulikh, Introduction to the Theory of Partially Ordered Spaces, Wolter
Hoordhoff1 Groningen, Netherlands, 1967. (English translation from the Russian.)
329. B. Z. Vulikh and C. Ya. Lozanovsky, Representation of order continuous and regular
functionals iii partially ordered spaces, Math. USSR-Sb. 13 (1971), 323343.
520 -- Bibliography
330. L. Weis and D. Werner, The Daugavet equation for operators not fixing a copy of
C(S), J. Operator Theory 39 (1998), 8998.
331. J. Wermer, The existence of invariant subspaces, Duke Math. J. 19 (1952), 615622.
332. D. Werner, An elementary approach to the Daugavet equation, in: N. Kalton,

E. Saab, and S. Montgomery-Smith, eds., Interaction Between Functional Analysis,
Harmonic Analysis, and Probability (Lecture Notes in Pure and Applied Mathematics,
175, Marcel Dekker, New York, 1996), 449454.
333. T. T. West, The decomposition of Riesz operators, Proc. London Math. Soc. (3)
16 (1966), 737752.
334. A. W. Wickstead, Representation and duality of multiplication operators on Archi-
medean Riesz spaces, Compositio Math 35 (1977), 225238.
335. A. W. Wickstead, Extremal structure of cones of operators, Quart. J. Math. Oxford
Ser. (2) 32 (1981), 239253.
336. A. W. Wiclcstead, Banach lattices with trivial center, Proc. Roy. Irish Acad. Sect. A
88 (1988), 7183.
337. A. Wilansky, Subalgebras of B(X), Proc. Amer. Math. Soc. 29 (1971), 355360.
338. W. Wnuk, Banach Lattices with Order Continuous Norms, Polish Scientific Publish-
ers, Warsaw, 1999.
339. P. Wojtaszczyk, Some remarks on the Daugavet equation, Proc. Amer. Math. Soc.
115 (1992), 10471052.
340. P. H. Wolff, An introduction to nonstandard functional analysis. In Nonstandard
Analysis (Edinburgh, 1966), pp. 121151. Kluwer Acad. Publ., Dordrecht, 1977.
341. D. E. Wulbert, Averaging projections, Illinois J. Math. 13 (1969), 689693.
342. A. C. Zaanen, Examples of orthomorphisms, J. Approx. Theory 13 (1975), 192204.
343. A. C. Zaanen, Riesz Spaces II, North-Holland, Amsterdam, 1983.

344. A. C. Zaanen, Introduction to Operator Theory in Riesz Spaces, SpringerVerlag,
New York and Heidelberg, 1997.
345. P. P. Zabreiko and S. V. Smitskikh, A theorem of M. C. and M. A. Rutman,
Funktsional. Anal. i Prilozhen. 13 (1979), 8182.
346. H. J. Zhong, Riesz operators on the spaces have West decompositions, Northeast.
Math. J. 4 (1988), 282288. (In Chinese with an English summary.) MR 90c:47031
347. H. J. Zhong, spaces and West decomposition of Riesz operators on them,

Acta Math. Sinica 37 (1994), 563569.
348. H. J. Zhong, Tsirelson's space and West decomposition of Riesz operators on it,
Sci. China Ser. A 39 (1996), 491500. (In Chinese with an English summary.) MR
97j:46003
349. Y. Zhong, Irreducible semigroups of functionally positive nilpotent operators, Trans.
Amer. Math. Soc. 347 (1995), 30933100.
Index
AL-space, 94, 101 of a set, 328

AM-compact operator, 401 of a vector subspace, 8, 328
AM-space, 94, 101 annulus, 49
complex, 106 antisymmetry, 13
with unit, 94 approximate eigenvalue, 249
a(T), ascent of operator T, 80 approximate point spectrum, 249
A-invariant subspace, 387 approximation property, 125
AT, algebra generated by T, 387 Archimedean ordered vector space, 15
Ax, the orbit of x under A, 387 ascent of operator, 80, 270
a.e. equal functions, 52 atom, 86
absolute integral operator, 187 atomic measure space, 193
absolute value of vector, 15 atomless Riesz space, 86
abstract integral operator, 193, 368, 430 automatic continuity, 147, 154
additive semigroup, 409 averaging operator, 213
adjoint operator, 4, 244 averaging property, 213
Hilbert space, 453
algebra, 3, 386 band generated by x, 20
Ax, 387 [B), super right-commutant of B, 399
Calkin, 299 (B), super left-commutant of B, 399
closed, 387 backward shift operator, 251
dual, 387 Banach function space, 194
generatcd by set of operators, 387 associated with Lo(n) or 194

non-transitive, 387, 410 rearrangement invariant, 223, 494
Banach lattice, 22
transitive, 387
AL, 94
3, 386, 387
94
algebra of operators, 386
complex, 105
generated by a set, 387
with order continuous norm, 84
non-transitive, 387, 440
Banach space, 2
transitive, 387
direct sum, 5, 10
algebraic complement of subspace, 169 function, 194
algebraic homomorphism, 1.14
locally uniformly convex, 460
almost essentially Opcrators, 477 quotient, 8
almost integral operator, 193, reflexive, 465
analytic function, 45 uniformly convex, 12, 457, 465
annihilator, 8 uniformly rotund7 457
522 Index
uniformly smooth, 457 Closed Graph Theorem, 5

BanachSteinhaus theorem, 4 closest point to a set, 12
band, 20 codimension of vector subspace, 155
generated by a set, 20 cofinal set, 66
of abstract integral operators, 193, 430 collection
of integral operators, 188 strictly transitive, 387
principal, 20, 118 transitive, 387
projection, 21 commutant, 138, 357, 388, 398, 410
band irreducible operator, 349 of family of operators, 410
band preserving operator, 112 super left, (B], 399
band projection, 21 super rights [B), 399
basic sequence, 30 commutator of two operators, 398
unconditional, 35 compact operator, 88, 272
basis, 28, 436 compact-friendly operator, 425
in 1J, 36 complement
monotone, 29 a'gebraic of subspace, 169
normalized, 29 of a closed subspace, 386
unconditional, 32, 500 complemented subspace, 134, 386, 486
basis constant, 29 complete measure, 66
basis problem, 28 completely continuous function, 444
biorthogonal functional, 31 completion
Bishop operator, 451 Dedekind, 26
bistochastic projection, 221 MaedaOgasawara--Vulikh, 116
bound, 15 universal, 116
greatest lower, 15 complex AM-space, 106
least upper, 15 complex Banach lattice, 105
lower, 15
complex lattice isometry, 105
upper, 15
complexification, 5
bounded function, 64
of Banach lattice, 105, 110
from above, 64
component of a vector, 16
bounded operator, 2
from
composition operator, 142, 378
below, 69
conditional expectation of function, 212
C (Q),
conditional expectation operator, 212
extended continuous functions, 115
C0(cl), functions vanishing at 00, 406
cone, 14, 440
C, complex numbers, 2 generated by a basis, 437, 440
the Calkin algebra, 299 generating, 14
C-invariant collection, 410 conjugate of matrix, 317
x, convergence along an ultrafilter, 38 continuous spectrum, 249
Calkin algebra, 299 contraction, 78
canonical projection, 299 control function, 315
Carat hodory extension of measure, 61 convergence
Carathodory's theorem, 61 *, 67
Carleman operator, 192 weak* 3
carrier of ideal, 27 in measure, 53
in L0, 60 notions in Lu, 209
CayleyHamilton theorem, 318 of Laurent series, 49
center of Banach function space, 209 order, 16
center of Banach lattice, 112 pointwise, 53
trivial, 116 uniform, 53
central operator, 112 weak, 3
Cesaro operator, 247, 450 convolution of kernels, 210
chain of open sets, 51 corank of operator, 155
characteristic polynomial, 317, 327 countable sup property, 22, 354
of operator7 321 cyclic point of a set, 288
characterization of narrow operator, 494 cyclic set, 288
closed algebra, 387 cyclic vector, 328, 383
Index 523
8(T), descent of operator T, 80 equimeasurable functions, 223, 494

IID1I, norm of set D, 410 essential property, 299
d(T), defect of operator T, 155 essential singularity, 264, 270
Daugavet equation, 456, 470 essential spectral radius, 302
Daugavet property, 467, 473 essential spectrum of an operator, 299
decreasing net, 16 essential sup, 64
Dedekind complete Riesz space, 15 essential sup norm, 64
Dedekind completion, 26 essential upper bound, 64
defect of operator, 155 essentially bounded function, 64
derivative of a function, 45 from above, 64
descent of operator, 80, 270 essentially disjoint operators, 477
diagonal matrix, 118 essentially invertible operator, 299
diagonal part of operator, 117 essentially quasinilpotent operator, 302
diagonal projection, 117 Euclidean norm, 316, 321
diagonal property, 56 of matrix, 318
diagonalizable matrix, 319 Euclidean topology, 321
diagonalizing matrix, 319 expanding operator, 350
differentiable function, 45 exposed point, 490
direct sum Banach space, 10 exposing linear functional, 490
direct sum normed space, 5 extremally disconnected space, 96
direct sum of Banach spaces, 247 extreme point, 490
direct sum of operators, 393
direct sum operator, 247 .Fred(X, Y), Fredholm operators, 163
direct sum property, 208 operators, 193, 430
Dirichlet's theorem, 236 Y, operators, 125
discrete vector, 86 .Fred(X, Y), semi-Fredholm operators, 166
disjoint complement, 21 F(X, Y), finite-rank operators from X to Y,
disjoint operators, 16 125
almost essentially, 477 f(T), 258
essentially, 477 factorization of an operator, 477
disjointly strictly singular opcrator, 178 family of operators
disk, 44 finitely quasinilpotent at a point, 410
domain of analyticity of function, 258 non-transitive, 410
domination of operators, 89, 109 quasinilpotent at a point, 410
double adjoint, 4 transitive, 410
double dual, 3 Fatou property, 65
doubly power bounded operator, 247, 450 oW-, 65
dual, 3 filter, 37
double, 3 Frchet, 38
Kthe, 198 Finite Dimensional Separation Theorem, 13
norm, 3 finite quasinilpotence, 410
order, 21 finite-rank operator1 91, 124, 430, 471
order continuous, 21, 193 rank, 124
second1 3 representation, 124
dual algebra, 387 fixed ulfrafllter, 38
DunfordPettis operator, 409 198 flat of a function, 432
DunfordPettis property, ;wo Frchet differentiability of the norm, 466
filter, 38
idealgerierrttcd by x, 20 Fredhoim alternative, 169
order dual of E1 21 Fredholm operator, 156
eigenspace, 384 Fredholm spectrum, 299
eigenvalue, 248 free ultrafilter, 38
approximate, 2 19 Fubini's theorem, 63
eigenvector, 248 fun ct ion
entire function, 47 a-order continuous, 18
equation analytic, 45
bounded, 64
524 Inde
completely continuous, 444 ideal irreducible semigroup, 415

control, 315 ideal reducible semigroup, 415
defining an integral operator, 181 implication scheme, 351
differentiable at a point, 45 indecomposable matrix, 333
entire, 47 index function, 163
essentially bounded above, 64 index of nilpotence, 82, 323
harmonic, 236 index of operator, 155
holomorphic, 45 Index Theorem, 158
index, 163 inequality
jointly measurable, 63 Jensen, 227
measurable, 52 triangle, 2
onto, 69 infimum, 15
order continuous, 18 inner product preserving matrix, 317
resolvent, 238, 245 input-output matrix, 315
retraction, 13 integral
state, 315 iterated, 62
surjective, 69 integral operator, 102, 180
weakly analytic, 47 absolute, 187
function space, 194 internal point, 101
normal, 194 interpolation space, 224
functional interval preserving operator, 24, 154
biorthogonal, 31 invariant measure, 355
multiplicative, 148 invariant subspace, 10, 348, 381
supporting, 466 invariant subspace problem, 381
Functional Calculus, 258 invertible operator, 10, 298
irreducible matrix, 333
Gantmacher's theorem, 89 isolated point of the spectrum, 265
Gateaux differentiability of the norm, 466 isometric normed spaces, 2
generalized Harris operator, 434 isometry, 2, 11
generating cone, 14 lattice, 23, 105
graph of function, 4 linear, 2, 11
isomorphic Banach spaces, 71
Haar system, 36 isomorphism between Banach spaces, 71 -
harmonic function, 236 iterated integral, 62

Harris operator, 370
hereditarily indecomposable B-space, 177 Jensen's inequality, 227
Hermitian matrix, 317 joint spectral radius, 425
Hubert space adjoint, 453 jointly measurable function, 63
HubertSchmidt operator, 193 Jordan contour, 257
holomorphic function, 45 surrounding a set, 257
homogeneity of the norm, 2 Jordan form of a matrix, 326
homomorphism
algebraic, 144 KB-space, 85
lattice, 88 IC(X,Y), compact operators from X to Y
hyperinvariant subspace, 348, 381 89
Kthe dual, 198
I(X,Y), 472 KadetsKlee property, 465
ideal, 20, 110 KantorovichBanach space, 85
generated by a set, 20 kernel of integral operator, 181
null, 119, 348, 399, 408 kernel operator, 181
null of an -seminorm, 122 Krein operator, 350, 371
null of functional, 27 KreinRutman theorem, 277
operator, 472 Kronecker's theorem, 291
principal, 20, 95, 119
range of operator, 399, 408 Lspace, 122
semigroup, 417 Lo(7r), 52
ideal irreducible operator, 111, 349 210
Index 525
210 indecomposable, 333

si-sum of Banach spaces, 472 inner product preserving, 317
of Banach spaces, 472 input-output, 315
of Banach spaces, 281 irreducible, 333
(X), bounded operators on X, 2 Markov, 103, 332
(X, Y), bounded operators from X to Y, 2 non-negative, 329
K(E, F), regular integral operators from E norm preserving, 317
to F, 188 orthogonal, 317
b(X, Y), order bounded operators from X permutation, 333
to Y, 15 positive, 329
r(X,Y), regular operators from X to Y, positive semidefinite, 319
14 primitive, 343
Laplace equation, 236 production, 315
lattice homomorphism, 23, 88, 154 reducible, 333
lattice hyperinvariant subspace, 348, 381 representing, 321
lattice isometric Banach lattices, 23 stochastic, 103
lattice isometric complex B-lattices, 105 strictly positive, 329
lattice isometry, 23, 105 strongly positive, 329
complex, 105 symmetric, 317
lattice isomorphism, 23 unitary, 317
lattice operations, 15 upper triangularizable, 326
lattice seminorm, 22 upper triangularizing, 326
229 Mazur's lemma, 30
Laurent series, 49 measurable function, 52
expansion of function, 49 measure, 61
Laurent series expansion, 266 atomic, 193
limit complete, 66
U, 38 invariant, 355
functional, 155 non-atomic, 479
order, 17 on a semiring, 61
linear functional product, 62
exposing a point, 490 measure convergence, 53
strongly exposing a point, 490 measure of non-compactness
linear isometry, 2, 11 of operator, 309
linear operator, 2 of set, 309
Liouville number, 451 minimal polynomial, 327
Liouville's theorem, 47 mixed norm, 210
local quasinilpotence, 357, 410 Miyajima projection, 235
locally uniformly convex B-space, 460 modulus, 16, 104, 107
Lomonosov operator, 396 of integral operator, 102, 186, 187
of operator, 16, 96, 107, 113
M-space, 122 of vector, 104
space of rt x n square matrices, 318 monotone basis, 29
MaedaOgasawara-Vulikh completion, 116 Morera's theorem, 50
majorizing vector space, 96 multiplication operation, 3
mapping multiplication operator, 136
nearest point, 12 multiplicative functional, 148
quotient, 8 multiplicative operator, 144
sublinear, 149
Markov matrix, 103, 332 NA, eigenspace with eigenvalue 384
Markov operator, 100 19, the natural numbers, {1,2, .}, 2
Markov projection, 213 n(T), nullity of operator T, 155
matrix narrow operator, 494
diagonal, 118 nearest point mapping, 12
diagonalizable, 319 nearest point to a set, 12
diagonalizing, 319 negative part of vector, 15
Hermitian, 317 neighborhood of set, 257
526 Index
net, 16 backward shift, 251

decreasing, 16 band irreducible, 349
order convergent, 16 band preserving, 112
Neumann series, 239 Bishop, 451
nilpotent operator, 82, 244, 323 bounded, 2
non-atomic measure, 479 bounded below, 69
non-atomic Riesz space, 86 Carleman, 192
nonnegative matrix, 329 central, 112
non-quasianalytic, 450 Cesaro, 247, 450
non-scalar operator, 383 compact, 88, 272
non-transitive algebra, 387, 440 compact-friendly, 425
non-transitive family of operators, 410 composition, 142, 378
non-trivial vector subspace, 348, 381 conditional expectation, 212
norm, 2 disjointly strictly singular, 178
essential sup, 64 double adjoint, 4
Euclidean, 316, 321 doubly power bounded, 247, 450
Frchet differentiable, 466 DunfordPettis, 409, 498
Gateaux differentiable, 466 essentially invertible, 299
mixed, 210 essentially quasinilpotent, 302
of a set, 410 expanding, 350
of integral operator, 102 finite-rank, 91, 124, 430, 471
of matrix, 318 fixing a copy of a B-space, 485
order continuous, 84 Fredholm, 156
quotient, 8 generalized Harris, 434
strictly monotone, 216 Harris, 370
sup, 64 HilbertSchmidt, 193
uniformly convex, 12 ideal irreducible, 111, 349
norm dual, 3 integral, 102, 180
norm preserving matrix, 317 interval preserving, 24, 154
normal function space, 194 invertible, 10, 298
normal integral, 198 isometry, 11
normal operator, 453 kernel, 181
normal Riesz space, 21, 193 Krein, 350, 371
normalized basis, 29 lattice homomorphism, 23
normed function space, 194 lattice isometry, 23
normed Riesz space, 22 lattice isomorphism, 23
normed space, 2 linear, 2
function, 194 Lomonosov, 396
norming subspace, 51 Markov, 100
null ideal, 27, 348, 399, 408 multiplication, 136
of seminorm, 122 multiplicative, 144
of functional, 119 narrow, 494
null space of operator, 79, 155 nilpotent, 82, 244, 323
nullity of operator, 155 non-quasianalytic, 450
nonscalar, 383
onto function, 69 normal, 453
open disk, 44 nuclear, 134
Open Mapping Theorem, 4 of rank k, 124
operation order adjoint, 200
multiplication, 3 order bounded, 15, 296
operator, 2 polynomially compact, 390
AM-compact, 401 positive, 14, 437
f(T), 258 power compact, 272
abstract integral, 193, 368, 430 projection, 21, 386
adjoint, 4, 244, 453 quasianalytic, 450
a'most integral, 193, 430 quasinilpotent, 244
averaging, 213 quasitriangular, 454
Index 527
rank-one, 123, 430 Archimedean, 15

real on a complex B-space, 6 partition of set, 227
regular, 14, 107 peripheralspectrum, 245
residue, 266 permutation matrix, 333
Riesz, 302 permutation of a set, 31
Riesz homomorphism, 23 PerronFrobenius theorem, 341
Riesz isomorphism, 23 perturbation, 163
rotation, 385 of operator, 163
scalar, 383 point
semi-Fredholm, 156 closest to a set, 12
shift, 251 cyclic, 288
stochastic, 100 exposed, 490
strictly positive, 14, 408 extreme, 490
strictly singular, 170 internal, 101
strong Krein, 350, 371 nearest to a set, 12
strongly expanding, 350 quasi-interior, 135, 349
subnormal, 453 strongly exposed, 490, 503
symmetric, 465 support. 13
unbounded, 2 point spectrum, 248
Volterra, 275 pointwise bounded set, 4
weakly compact, 89 pointwise convergence, 53
weakly expanding, 350 pole, 264
operator domination, 89, 109, 425 of the resolvent, 264, 268
operator ideal, 472 simple. 264
operator norm, 2 polynomial
operator norm of matrix, 318 characteristic, 317
operator of rank k, 124 minimal, 327
orbit, 383 polynomially compact operator, 390
orbit space, 383 positive matrix, 329
order adjoint operator, 200 positive operator, 14, 437
order bounded operator, 15, 296 on a B-space, 436
order bounded set, 14 positive part of vector, 15
order closed set, 18 positive semidefinite matrix, 319
order complete Riesz space, 15 positively complemented subspace, 230
order continuous dual, 21, 193 positivityof the norm, 2
order continuous function, 18 power compact operator, 272
order continuous norm, 84 power series expansion of function, 46
order continuous topology, 84 predual of ideal generated by a functional,
order convergence, 16 119
order convergence of sequences, 18 primitive matrix, 343
order convergent net, 16 principal band, 20, 118
order dense Riesz subspace, 58, 66 principal ideal, 20, 95, 119
order dual, 21 product cr-algebra, 62
order interval, 84 product measure, 62
order limit, 17 product rule for derivatives, 50
order relation, 13 production matrix, 315
order spectrum, 292 projection, 21, 386
order unit, 20, 94 band. 21
ordered vector space, 14 bistochastic, 221
Archimedean, 15 canonical, 299
orthogonal matrix, 317 diagonal, 117
orthomorphism, 112 Markov, 213
outer measure generated by a measure, 61 Miyajima, 235
proper. 386
pair of reducing subspaces, 386 spectral, 262
partially ordered set, 14 projection band, 21
partially ordered vector space, 14 proper projection, 386
528 Inde
property residual spectrum, 248

antisymmetry, 13 residue operator, 266
approximation, 125 resolvent, 238, 245
countable sup, 22, 354 resolvent function, 238, 245, 265
Daugavet, 467, 473 resolvent identity, 240
direct sum, 208 second, 246
DunfordPettis, 360 resolvent set, 238
essential, 299 retract of a space, 12
KadetsKlee, 465 retraction, 13
RadonNikodym, 503 Riesz homomorphism, 23
reflexivity, 13 Riesz isomorphism, 23
Schur, 448 Riesz operator, 302
transitivity, 14 Riesz space, 15
pseudoinverse of operator, 167 cr-Dedekind complete, 16
cr-order complete, 16
Q, the rational numbers, 2 atomless, 86
quasi-interior point, 135, 349 Dedekind complete, 15
quasianalytic operator, 450 non-atomic, 86
quasinilpotent operator, 244 normal, 21, 193
quasitriangular operator, 454 normed, 22
quotient Banach space, 8 order complete, 15
quotient map, 8 Riesz subspace, 84
quotient norm, 8 order dense, 58, 66
quotient vector space, 8 RieszKantorovich formula, 16, 107
RieszThorin theorem, 224
R, the real numbers, 2 rotation invariant set, 340
vector sublattice generated by A, 233 rotation invariant spectrum, 290
r-norm of operator, 22 rotation operator, 385
ress(T), essential spectral radius of 302
Radernacher functions, 497 cr-Dedekind complete Riesz space, 16
generalized, 497 cr-order closed set, 18
radius of convergence, 44 cr-order complete Riesz space, 16
RadonNikodym property, 503 cr-order continuous function, 18
range, 69 o-(T), spectrum of operator T, 238
of function, 69 cra(T), approximate spectrum of T, 249
of operator, 79, 155 continuous spectrum of T, 249
range ideal of operator, 399, 408 point spectrum of T, 248
rank cir(T), residual spectrum of T, 248
of finite-rank operator, 124 Weyl spectrum of T, 312
of matrix, 328 cress(T), essential spectrum of T, 299
of operator, 155 *-convergence, 67
rank-one operator, 123, 430 scalar operator, 383
real operator on a complex B-space, 6 Schauder basis, 28, 436
real part of complex functional, 7 Schauder system, 36
rearrangement invariant B-function space, Schur property, 448
223, 494 second dual, 3
rectangle, 62 second resolvent identity, 246
reducible matrix, 333 semiFredhoim operator, 156
reducing pair of subspaces, 81, 262, 322 semigroup
reducing subspace, 81, 386 ideal irreducible, 415
reflexive Banach space, 465 ideal reducible, 415
reflexivity property, 13 semigroup ideal, 417
regular operator, 14, 107 generated by a set, 417
regulator of convergence, 53 semigroup of operators, 409
representation of finite-rank operator, 124 additive, 409
representing kernel, 484 generated by a family, 412
representing matrix, 321 multiplicative, 409
Index 529
Volterra, 451 order, 292

seminorm 22, 126 peripheral, 245
lattice, 22 point, 248
semiring of sets, 60 residual, 248
separation of points, 209 rotation invariant, 290
separation of sets, 13 Weyl, 312
Separation Theorem spectrum of vector in a B-algebra, 245
Finite Dimensional, 13 state function, 315
series stochastic matrix, 103
convergent unconditionally, 31 stochastic operator, 100
Laurent, 49 Stone space of a Riesz space, 116
Neumann, 239 Stonean topological space, 96, 115
weakly unconditionally Cauchy, 480, 486 strictly monotone norm, 216
set strictly positive matrix, 329
o-order closed, 18 strictly positive operator, 14, 408
coflnal, 66 strictly positive vector, 135, 316
cyclic, 288 strictly singular operator, 170
order bounded, 14 strictly transitive collection, 387
order closed, 18 strong Krein operator, 350, 371
partially ordered 14 strong operator topology, 438
pointwise bounded, 4 strong unit, 20, 94, 370
resolvent, 238 strongly expanding operator, 350
rotation invariant, 340 strongly exposed point, 490, 503
solid, 20 strongly exposing linear functional, 490
spectral, 261 strongly positive 329
shift operator, 251 structure space, 136
similar matrices, 318 subalgebra of operators, 386
similar operators, 253 subgroup of the unit circle, 291
simple pole, 264 dense, 291
slice of aset, 488 sublinear mapping, 149
solid set, 20 subnormal operator, 453
space subspace
AL, 94 E-hyperinvariant, 348, 381
AM, 94 complemented, 134, 386
KB, 85 hyperinvariant, 381
L, 122 invariant, 10, 348, 381
M, 122 lattice, 229
Banach, 2 non-trivial, 348, 381
KantorovichBanach, 85 norming, 51
normed, 2 positively complemented, 230
orbit, 383 reducing an operator, 81, 386
Riesz, 15 subspace invariant under an operator, 10
Stonean, 115 sum of Banach spaces, 247, 281, 472
structure of B-lattice, 136 e1, 472
Spectral Mapping Theorem, 260 472
spectral projections 262 sum of vector spaces closed, 121
spectral radius, 2 13, 2 15 sup norm, 64
essential, 302 super left-commutant, 399
joint. 425 super right-commutant, 399
125 support
spectral set, 261 of function, 58
spectrum 245 of ideal in Lu, 60
spectrum of operator, 238 point of a set, 13
approximate point. 2 P) supporting functional, 466
contimlous. 2 19 supremurn, 15
essentiti, surjective function, 69
Fredhoim, symmetric difference of two sets, 66
530 Index
symmetric matrix, 317 upper triangularizing, 326

symmetric operator, 465
system vector
Haar, 36 cyclic, 328, 383
Schauder, 36 discrete, 86
strictly positive, 135, 316
T*, the adjoint of operator T, 4 vector lattice, 15
T**, the double adjoint of operator T, 4 vector space
tensor-product notation, 124 majorizing, 96
Tonelli's theorem, 63 ordered, 14
topological space partially ordered, 14
extremally disconnected, 96 quotient, 8
Stonean, 96, 115 vector space order, 14
topology vector sublattice, 84
Euclidean, 321 vector sublattice generated by a set, 233
of uniform convergence on compact sets, Volterra collection, 451
12G Volterra operator, 275
order continuous, 84
strong operator, 438 VV(X, Y), weakly compact operators from X
trace, 133, 318 toY, 89
of finite-rank operator, 133 weak* convergence, 3
of matrix, 318 weak convergence, 3
transitive algebra, 387 weak unit, 21, 135, 350
transitive collection, 387 weakly analytic function, 47
transitive family of operators, 410 weakly compact operator, 89
transitivity property, 14 weakly expanding operator, 350
transpose of matrix, 317 weakly unconditionally Cauchy series, 480,
triangle inequality, 2 486
trivial centers 116 West decomposition, 312
twist of a set, 207 Weyl spectrum of operator, 312
U, U*, closed unit balls, 3 Xu, ultrapower of B-space X, 40

U-limit of sequence, 38
X*, norm dual of X, 3
ultrafilter, 37 X**, double norm dual of X, 3
fixed, 38 24 x, convergence along an ultrafilter, 38
free, 38
ultrapower of B-space, 40 Z, the integers, 2
unbounded component of p(T), 256

unbounded operator, 2
unconditional basic sequence, 35
unconditional basis, 32, 500
unconditional constant, 33
unconditionally convergent series, 31
uniform boundedness principle, 4
uniform convergence, 53
uniformly convex B-space, 465
uniformly convex Banach space, 12, 457
uniformly convex norm, 12
uniformly rotund Banach space, 457
uniformly smooth Banach space, 457
unit, 3, 20, 94, 370
order, 20, 94
weak, 21, 135, 350
unital algebra, 3, 386, 387
unitary matrix, 317
universal completion, 116
upper triangularizable matrix, 326
This book offers a comprehensive and reader-friendly exposition of the theory of
linear operators on Banach spaces and Banach lattices using their topological and
order structures and properties. Abramovich and Aliprantis give unique presen-
tation that includes many new and recent advances in operator theory and brings
together results that are spread over the vast literature. For instance, invariant
subspaces of positive operators and the Daugavet equation are presented in
monograph form for the first time.
The authors keep the discussion self-contained and use exercises to achieve this
goal The book contains over 600 exercises to help students master the mate-
rial developed in the texiThe exercises are of varying degrees of difficulty and
play an important and useful role in the presentation. They help to free the
proofs of the main results of technical details, which are secondary to the prin-
cipal ideas, but provide students with accurate and complete accounts of how
such details ought to be worked out. The exercises also contain a considerable
amount of additional material, and among them there are many well-known
results whose proofs are not readily available elsewhere.
Prerequisites are the standard introductory graduate courses in real analysis.
general topology, measure theory and functional analysis. The volume is suitable
for graduate or advanced courses in operator theory. real analysis, integration
theory. measure theory. function theory, and functional analysis. It will also be of
great interest to researchers in mathematics, as well as in physics. economics,
finance, engineering, and other related areas.
The companion volume, Problems in Operator Theory. containing complete solu-
tions to all exercises in An Invitation to Operator Theory. is available from the AMS
as Volume SI in the Graduate Studies in Mathematics series.
ISBN
on rue Web

(Graduate Studies in Mathematics, V. 50) Y. A. Abramovich, Charalambos D. Aliprantis-An Invitation To Operator Theory-Amer Mathematical Society (2002)

Uploaded by

Copyright:

Available Formats

(Graduate Studies in Mathematics, V. 50) Y. A. Abramovich, Charalambos D. Aliprantis-An Invitation To Operator Theory-Amer Mathematical Society (2002)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Graduate Studies in Mathematics, V. 50) Y. A. Abramovich, Charalambos D. Aliprantis-An Invitation To Operator Theory-Amer Mathematical Society (2002)

Uploaded by

Copyright:

Available Formats

An Invitation to

American Mathematical Society

American Mathematical Society

Library of Congress Cataloging-in-Publication Data

To Bernadette, Claire, and Dionissi

1.1. Banach Spaces, Operators, and Linear Functionals 1.

1.2. Banach Lattices and Positive Operators 13

Chapter 2. Basic Operator Theory 69

Chapter 3. Operators on and AM-spaces 93

Chapter 4. Special Classes of Operators 123

4.3. Lattice and Algebraic Homomorphisms 142

Chapter 5. Integral Operators 179

Chapter Spectral-Properties 237

Chapter 7. Some Special Spectra 271

Chapter 8. Positive Matrices 315

Chapter 9. Irreducible Operators

Chapter 10. Invariant Subspaces 381

10.1. A Smorgasbord of Invariant Subspaces 383

10.2. The Lomonosov Invariant Subspace Theorem 393

10.3. Invariant Ideals for Positive Operators 398

Chapter 11. The Daugavet Equation 455

Since the appearance of the seminal monograph of Stefan Banach [51] on

and Banach lattices by exploiting their topological and order structures, as

research. A brief description of the material in each chapter should give a

Chapter 1 presents an introduction to Banach spaces and opera-

0 Chapter 9 investigates irreducible operators on Banach lattices and

sive study of operators on Banach spaces that satisfy the Daugavet

Y. A. Abramovich and C. D. Aliprantis

Odds and Ends

The material included in this chapter is reflected correctly in its title. We

1.1. Banach Spaces, Operators, and Linear Functionals

A function II X : on a (real or complex') vector space is said to

T is called a bounded operator, and when IIT!1 = 00,

Theorem 1.1. If X is a normed space and Y is a Banach space, then the

Theorem 1.2. If X is a Banach space, then (X) with the composition

(x) holds for each e X*. Similarly, a sequence {4} in

The adjoint of a bounded operator T: X Y between two arbitrary

= sup = sup [sup (T*y*,x)]

The Uniform Boundedness Principle is also known in the literature as

Recall that the graph of a function f: A + B is the subset C1 of

T: X + Y is an operator between vector spaces, then its graph CT is a

sum normed space and is denoted X1 If X1, . , are all

Proof. Assume that T: X + Y

that this operator is (Ps, it suffices to show that its graph

and qy y) 0 for some y E Y. Since dominates p, it follows that

y = 0. That is, the operator T: (Y, has a closed graph, and

We close the section with a discussion of the complexification of a real

for each z = x + zy E Clearly, lixU = lix + for each x E X. Notice

for all z = x + E Consequently, = x + zy in Xc if and

in can be identified with the vector T+ iS in (X, Y) ezL(X, Y).

where every T + iS E (X, acts on via the formula

holds for each x E X. Conversely, if g: p R is a linear functional, then

it follows that = On the other hand, if x X

Therefore, = f(x)[ and 50 11111 ugh. That is, we have

The preceding discussion shows that the mapping f Ref, from

Using the previous result, we can also establish the following.

Theorem 1.10 (BohnenblustHahnBanach). Every continuous linear

Recall that if V is then the quotient

[w] = w + V = {w + v: v e V}. The addition and multiplication in W/V

= sup = sup [sup (Ty,x)]

for each [xJ e X/Y, is a well-defined surjective linear isometry.