Topological Vector Spaces

Graduate Texts in Mathematics 3
Editorial Board
S. Axler F.W. Gehring K.A. Ribet
Springer Science+Business Media, LLC

Graduate Texts in Mathematics
TAKEUTilZARING. Introduction to
Axiomatic Set Theory. 2nd ed. 33 HIRSCH. Differential Topology.
2 OXTOBY. Measure and Category. 2nd ed. 34 SPITZER. Principles of Random Walk.
3 SCHAEFER. Topological Vector Spaces. 2nded.
2nded. 35 ALExANoERIWERMER. Several Complex
4 HILTON/STAMMBACH. A Course in Variables and Banach Algebras. 3rd ed.
Homological Algebra. 2nd ed. 36 KELLEy/NAMIOKA et al. Linear
5 MAC LANE. Categories for the Working Topological Spaces.
Mathematician. 2nd ed. 37 MONK. Mathematical Logic.
6 HUGHESIPIPER. Projective Planes. 38 GRAUERT/FRITZSCHE. Several Complex
7 SERRE. A Course in Arithmetic. Variables.
8 TAKEUTI/ZARING. Axiomatic Set Theory. 39 ARVESON. An Invitation to C*-Algebras.
9 HUMPHREYS. Introduction to Lie Algebras 40 KEMENy/SNELL/KNAPP. Denumerable
and Representation Theory. Markov Chains. 2nd ed.
10 COHEN. A Course in Simple Homotopy 41 APoSTOL. Modular Functions and Dirichlet
Theory. Series in Number Theory.
11 CONWAY. Functions of One Complex 2nded.
Variable I. 2nd ed. 42 SERRE. Linear Representations of Finite
12 BEALS. Advanced Mathematical Analysis. Groups.
13 ANDERSON/FuLLER. Rings and Categories 43 GILLMAN/JERISON. Rings of Continuous
of Modules. 2nd ed. Functions.
14 GoLUBITSKY/GUILLEMIN. Stable Mappings 44 KENDIG. Elementary Algebraic Geometry.
and Their Singularities. 45 LoEVE. Probability Theory I. 4th ed.
15 BERBERIAN. Lectures in Functional 46 LoEVE. Probability Theory II. 4th ed.
Analysis and Operator Theory. 47 MOISE. Geometric Topology in
16 WINTER. The Structure of Fields. Dimensions 2 and 3.
17 ROSENBLATT. Random Processes. 2nd ed. 48 SACHslWu. General Relativity for
18 HALMOS. Measure Theory. Mathematicians.
19 HALMOS. A Hilbert Space Problem Book. 49 GRUENBERo/WEIR. Linear Geometry.
2nded. 2nded.
20 HUSEMOLLER. Fibre Bundles. 3rd ed. 50 EowAROS. Fermat's Last Theorem.
21 HUMPHREYS. Linear Algebraic Groups. 51 KLINGENBERG. A Course in Differential
22 BARNEs/MACK. An Algebraic Introduction Geometry.
to Mathematical Logic. 52 HARTSHORNE. Algebraic Geometry.
23 GREUB. Linear Algebra. 4th ed. 53 MANIN. A Course in Mathematical Logic.
24 HOLMES. Geometric Functional Analysis 54 GRAVERIWATKINS. Combinatorics with
and Its Applications. Emphasis on the Theory of Graphs.
25 HEwITT/STROMBERG. Real and Abstract 55 BROWN/PEARCY. Introduction to Operator
Analysis. Theory I: Elements of Functional
26 MANES. Algebraic Theories. Analysis.
27 KELLEy. General Topology. 56 MASSEY. Algebraic Topology: An
28 ZARISKI/SAMUEL. Commutative Algebra. Introduction.
Vol.I. 57 CROWELLlFox. Introduction to Knot
29 ZARISKI/SAMUEL. Commutative Algebra. Theory.
Vol.II. 58 KOBUTZ. p-adic Numbers, p-adic
30 JACOBSON. Lectures in Abstract Algebra I. Analysis, and Zeta-Functions. 2nd ed.
Basic Concepts. 59 LANG. Cyclotomic Fields.
31 JACOBSON. Lectures in Abstract Algebra II. 60 ARNOLD. Mathematical Methods in
Linear Algebra. Classical Mechanics. 2nd ed.
32 JACOBSON. Lectures in Abstract Algebra 61 WHITEHEAD. Elements of Homotopy
III. Theory of Fields and Galois Theory. Theory.
(continued after index)

H.H. Schaefer
With M.P. Wolff
Topological Vector Spaces
Second Edition
Springer
H.H. Schaefer
M.P. Wolff
Eberhard-Karls-Universitat Tiibingen
Mathematisches Institut
Auf der Morgenstelle 10
Tiibingen, 72076
Germany
Editorial Board
S. Axler F. W. Gehring K. A. Ribet
Mathematics Department Mathematics Department Mathematics
San Francisco State East Hall Department
University University of Michigan Universityof
San Francisco, CA 94132 Ann Arbor, MI 48109 California at
USA USA Berkeley
Berkeley, CA 94720-3840
USA
Mathematics Subject Classification (1991): 46-01, 46Axx, 46Lxx
Library of Congress Cataloging-in-Publication Data

Schaefer, Helmut H.
Topological vector spaces. - 2nd ed. / Helmut H. Schaefer in
assistance with M. Wolff.
p. cm. - (Graduate texts in mathematics ; 3)
Includes bibliographical references and indexes.
ISBN 978-1-4612-7155-0 ISBN 978-1-4612-1468-7 (eBook)
DOI 10.1007/978-1-4612-1468-7
1. Linear topological spaces. 1. Wolff, Manfred, 1939-
II. Title. III. Series.
QA322.S28 1999
Sl4'.32-dc21 98-53842
Printed on acid-free paper.
First edition © 1966 by H. H. Schaefer. Published by the Macmillan Company, New York.
© 1999 Springer Science+Business Media New York

Originally published by Springer-Verlag New York in 1999
Softcover reprint ofthe hardcover 2nd edition 1999
AlI rights reserved. This work may not be translated or copied in whole or in part without the
written permission of the publisher Springer Science+Business Media New York, except
for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any
form of information storage' and retrieval, electronic adaptation, computer software, or by
similar or dissimilar methodology now known or hereafter developed is forbidden.
The use of general descriptive names, trade names, trademarks, etc., in this publication, even if
the former are not especialIy identified, is not to be taken as a sign that such names, as under-
stood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely by
anyone.
Production managed by Alian Abrams; manufacturing supervised by Jeffrey Taub.
9 7 8 6 54 32 1
ISBN 978-1-4612-7155-0
PREFACE TO THE
SECOND EDITION
As the first edition of this book has been well received through five printings
over a period of more than thirty years, we have decided to leave the mate-
rial of the first edition essentially unchanged - barring a few necessary up-
dates. On the other hand, it appeared worthwhile to extend the existing text
by adding a reasonably informative introduction to C* - and W* -algebras.
The theory of these algebras seems to be of increasing importance in math-
ematics and theoretical physics, while being intimately related to topological
vector spaces and their orderings-the prime concern of this text.
The authors wish to thank J. Schweizer for a careful reading of Chapter
VI, and the publisher for their care and assistance.
Tiibingen, Germany H. H. Schaefer

Spring 1999 M. P. Wolff
v
Preface
The present book is intended to be a systematic text on topological vector

spaces and presupposes familiarity with the elements of general topology and
linear algebra. The author has found it unnecessary to rederive these results,
since they are equally basic for many other areas of mathematics, and every
beginning graduate student is likely to have made their acquaintance. Simi-
larly, the elementary facts on Hilbert and Banach spaces are widely known
and are not discussed in detail in this book, which is :plainly addressed to those
readers who have attained and wish to get beyond the introductory level.
The book has its origin in courses given by the author at Washington State
University, the University of Michigan, and the University of Ttibingen in
the years 1958-1963. At that time there existed no reasonably ccmplete text on
topological vector spaces in English, and there seemed to be a genuine need
for a book on this subject. This situation changed in 1963 with the appearance
of the book by Kelley, Namioka et al. [1] which, through its many elegant
proofs, has had some influence on the final draft of this manuscript. Yet the
two books appear to be sufficiently different in spirit and subject matter to
justify the publication of this manuscript; in particular, the present book
includes a discussion of topological tensor products, nuclear spaces, ordered
topological vector spaces, and an appendix on positive operators. The author
is also glad to acknowledge the strong influence of Bourbaki, whose mono-
graph [7], [8] was (before the publication of Kothe [5]) the only modern
treatment of topological vector spaces in printed form.
A few words should be said about the organization of the book. There is a
preliminary chapter called "Prerequisites," which is a survey aimed at
clarifying the terminology to be used and at recalling basic definitions and
facts to the reader's mind. Each of the five following chapters, as well as the
Appendix, is divided into sections. In each section, propositions are marked
u.v, where u is the section number, v the proposition number within the
vi
PREFACE vii
section. Propositions of special importance are additionally marked

"Theorem." Cross references within the chapter are (u.v), outside the chapter
(r, u.v), where r (roman numeral) is the number of the chapter referred to.
Each chapter is preceded by an introduction and followed by exercises. These
"Exercises" (a total of 142) are devoted to further results and supplements, in
particular, to examples and counter-examples. They are not meant to be
worked out one after the other, but every reader should take notice of them
because of their informative value. We have refrained from marking some of
them as difficult, because the difficulty of a given problem is a highly subjective
matter. However, hints have been given where it seemed appropriate, and
occasional references indicate literature that may be needed, or at
least helpful. The bibliography, far from being complete, contains
(with few exceptions) only those items that are referred to in the text.
I wish to thank A. Pietsch for reading the entire manuscript, and A. L.
Peressini and B. J. Walsh for reading parts of it. My special thanks are
extended to H. Lotz for a close examination of the entire manuscript, and for
many valuable discussions. Finally, I am indebted to H. Lotz and A. L.
Peressini for reading the proofs, and to the publisher for their care and
cooperation.
H.H.S.
Tiibingen, Germany
December, 1964
Table if Contents
Preface to the Second Edition v

Preface vi
Prerequisites
A. Sets and Order 1
B. General Topology 4
C. Linear Algebra 9
I. TOPOLOGICAL VECTOR SPACES

Introduction 12
Vector Space Topologies 12
2 Product Spaces. Subspaces. Direct Sums.
Quotient Spaces 19
3 Topological Vector Spaces of Finite Dimension 21
4 Linear Manifolds and Hyperplanes 24
5 Bounded Sets 25
6 Metrizability 28
7 Complexification 31
Exercises 33
II. LOCALL Y CONVEX TOPOLOGICAL

VECTOR SPACES
Introduction 36
Convex Sets and Semi-Norms 37
2 Normed and Normable Spaces 40
3 The Hahn-Banach Theorem 45
viii
TABLE OF CONTENTS ix
4 Locally Convex Spaces 47

5 Projective Topologies 51
6 Inductive Topologies 54
7 Barreled Spaces 60
8 Bornological Spaces 61
9 Separation of Convex Sets 63
10 Compact Convex Sets 66
Exercises 68
III. LINEAR MAPPINGS

Introduction 73
Continuous Linear Maps and Topological
Homomorphisms 74
2 Banach's Homomorphism Theorem 76
3 Spaces of Linear Mappings 79
4 Equicontinuity. The Principle of Uniform Boundedness
and the Banach-Steinhaus Theorem 82
5 Bilinear Mappings 87
6 Topological Tensor Products 92
7 Nuclear Mappings and Spaces 97
8 Examples of Nuclear Spaces 106
9 The Approximation Property. Compact Maps 108
Exercises 115
IV. DU ALITY
Introduction 122
Dual Systems and Weak Topologies 123
2 Elementary Properties of Adjoint Maps 128
3 Locally Convex Topologies Consistent with a
Given Duality.The Mackey-Arens Theorem 130
4 Duality of Projective and Inductive Topologies 133
5 Strong Dual of a Locally Convex Space. Bidual.
Reflexive Spaces 140
6 Dual Characterization of Completeness. Metrizable Spaces.
Theorems of Grothendieck, Banach-Dieudonne, and
Krein-Smulian 147
x TABLE OF CONTENTS
7 Adjoints of Closed Linear Mappings 155

8 The General Open Mapping and Closed Graph Theorems 161
9 Tensor Products and Nuclear Spaces 167
10 Nuclear Spaces and Absolute Summability 176
11 Weak Compactness. Theorems of Eberlein and Krein 185
~~~ m
V. ORDER STRUCTURES
Introduction 203
Ordered Vector Spaces over the Real Field 204
2 Ordered Vector Spaces over the Complex Field 214
3 Duality of Convex Cones 215
4 Ordered Topological Vector Spaces 222
5 Positive Linear Forms and Mappings 225
6 The Order Topology 230
7 Topological Vector Lattices 2:34
8 Continuous Functions on a Compact Space. Theorems
of Stone-Weierstrass and Kakutani 242
Exercises 250
VI. C*-AND W*-ALGEBRAS

Introduction 258
Preliminaries 259
2 C*-Algebras. The Gelfand Theorem 260
3 Order Structure of a C* -Algebra 267
4 Positive Linear Forms. Representations 270
5 Projections and Extreme Points 274
6 W*-Algebras 277
7 Von Neumann Algebras. Kaplansky's Density Theorem 287
8 Projections and Types of W* -Algebras 292
Exercises 299
Appendix. S PEe T R ALP R 0 PER TIE S 0 F

POSITIVE OPERATORS
Introduction 306
1 Elementary Properties of the Resolvent 307
2 Pringsheim's Theorem and Its Consequences 309
TABLE OF CONTENTS xi
3 The Peripheral Point Spectrum 316
Index of Symbols 325
Bibliography 330
Index 339
To my wife
PREREQUISITES
A formal prerequisite for an intelligent reading of this book is familiarity

with the most basic facts of set theory, general topology, and linear algebra.
The purpose of this preliminary section is not to establish these results but
to clarify terminology and notation, and to give the reader a survey of the
material that will be assumed as known in the sequel. In addition, some of
the literature is pointed out where adequate information and further refer-
ences can be found.
Throughout the book, statements intended to represent definitions are
distinguished by setting the term being defined in bold face characters.
A. SETS AND ORDER
1. Sets and Subsets. Let X, Y be sets. We use the standard notations x EX

for" x is an element of X", Xc Y (or Y:::l X) for" X is a subset of Y",
X = Y for " Xc Y and Y:::l X". If (p) is a proposition in terms of given
relations on X, the subset of all x E X for which (p) is true is denoted by
{x E X: (p)x} or, if no confusion is likely to occur, by {x: (p)x}. x ¢: X means
" x is not an element of X". The complement of X relative to Y is the set
{x E Y: x ¢: X}, and denoted by Y ~ X. The empty set is denoted by 0 and
considered to be a finite set~ the set (singleton) containing the single element
x is denoted by {x}. If (Pt), (P2) are propositions in terms of given -relations
on X, (Pt) => (P2) means" (Pt) implies (P2)", and (PI) ~ (P2) means" (Pt) is
equivalent with (P2)". The set of all subsets of X is denoted by ~(X).
2. Mappings. A mapping f of X into Y is denoted by j: X --+ Y or by
x--+f(x). Xis called the domain off, the image of Xunderf, the range off;
the graph of/is the subset GJ = {(x,f(x»: x E X} of Xx Y. The mapping of
the set ~(X) of all subsets of X into ~(Y) that is associated with f, is also
denoted by f; that is, for any A c X we write f(A) to denote the set
1
H. H. Schaefer et al., Topological Vector Spaces

© Springer Science+Business Media New York 1999
2 PREREQUISITES
{f(x) : x E A} c Y. The associated map of ~(Y) into ~(X) is denoted by

f- t ; thus for any BeY, f-t(B) = {x E X:f(x) E B}. If B = {b}, we write
f-t(b) in place of the clumsier (but more precise) notation f-t({b}). If
f: X ~ Y and g: Y ~ Z are maps, the composition map x ~ g(f(x» is
denoted by 9 0 f
A mapf: X ~ Yis biunivocal (one-to-one, injective) iff(xt ) = f(x 2 ) implies
X t = X2; it is onto Y (surjective) if f(X) = Y. A map fwhich is both injective
and surjective is called bijective (or a bijection).
Iff: X ~ Yis a map and A c X, the map g: A ~ Y defined by g(x) =f(x)
whenever x E A is called the restriction off to A and frequently denoted by fA.
Conversely, f is called an e"tension of 9 (to X with values in Y).
3. Families. If A is a non-empty set and X is a set, a mapping C/. ~ x(C/.)
of A into Xis also called a family in X; in practice, the term family is used for
mappings whose domain A enters only in terms of its set theoretic properties
(i.e., cardinality and possibly order). One writes, in this case, x" for x(C/.) and
denotes the family by {x,,: C/. E A}. Thus every non-empty set X can be viewed
as the family (identity map) x ~ x(x E X); but it is important to notice that
if {x,,: ix E A} is a family in X, then C/. #- f3 does not imply x,,#- xp. A sequence
is a family {x n : n EN}, N = {I, 2, 3, ... } denoting the set of natural numbers.
If confusion with singletons is unlikely and the domain (index set) A is clear
from the context, a family will sometimes be denoted by {x,,} (in particular, a
sequence by {xn }). .
4. Set Operations. Let {X",: C/. E A} be a family of sets. For the union of this
family, we use the notations U{X,,: C/..E A}, U X"' or briefly U"X" if the
"eA
index set A is clear from the context. If {Xn: nE N} is a sequence of sets we
'" k
also write U X n, and if {Xt' ... , X k } is a finite family of sets we write UXn or
t t
Xt u X2 U ... U X k • Similar notations are used for intersections and Car-
tesian products, with U replaced by () and TI respectively. If {X,,: C/. E A} is
a family such that X" = X for all C/. E A, the productTI"X" is also denoted by
XA.
If R is an equivalence relation (i.e., a reflexive, symmetric, transitive binary
relation) on the set X, the set of equivalence classes (the quotient set) by R is
denoted by XI R. The map x ~ x (also denoted by x ~ [x)) which orders to
each x its equivalence class x (or [x)), is called the canonical (or quotient) map
of X onto XIR.
5. Orderings. An ordering (order structure, order) on a set X is a binary
relation R, usually denoted by ~, on X which is reflexive, transitive, and anti-
symmetric (x ~ y and y ~ x imply x = y). The set X endowed with an order
~ is called an ordered set. We write y ~ x to mean x ~ y, and x < y to mean
x ~ y but x#- y (similarly for x > y). If R t and R2 are orderings of X, we say
that R t is finer than R2 (or that R2 is coarser than R t ) if x(Rt)y implies
x(R 2 )y. (Note that this defines an ordering on the set of all orderings
of X.)
§A] SETS AND ORDER 3
Let (X, ~) be an ordered set. A subset Aof X is majorized if there exists

a o E X such that a ~ a o whenever a E A; a o is a majorant (upper bound) of A.
Dually, A is minorized by ao if a o ~ a whenever a E A; then a o is a mi~orant
(lower bound) of A. A subset A which is both majorized and minoFized, is
called order bounded. If A is majorized and there exists a majorant a{) such
that ao ~ b for any majorant b of A, then ao is unique and called the supremum
(least upper bound) of A; the notation is ao= sup A. In a dual fashion, one
defines the infimum (greatest lower bound) of A, to be denoted by inf A. For
each pair (x, y) E X X X, the supremum and infimum of the set {x, y} (when-
ever they exist) are denoted by sup(x, y) and inf(x,y) respectively. (X, ~) is
called a lattice if for each pair (x, y), sup(x, y) and inf(x, y) exist, and (X, ~) is
called a complete lattice if sup A and inf A exist for every non-empty subset
A c X. (In general we avoid this latter terminology because of the possible
confusion with uniform completeness.) (X, ~) is totally ordered if for each
pair (x, y), at least one of the relations x ~ y and y ~ x is true. An element
x E X is maximal if x ~ y implies x = y.
Let (X, ~) be a non-empty ordered set. X is called directed under ~
(briefly, directed ( ~» if every subset {x, y} (hence each finite subset) possesses
an upper bound. If Xo.E X, the subset {x EX: Xo ~ x} is called a section of X
(more precisely, the section of X generated by xo). A family {y~: a E A} is
directed if A is a directed set; the sections of a directed family are the sub-
families {y~: ao ~ IX}, for any 1X0 E A.
Finally, an ordered set X is inductively ordered if each totally ordered
subset possesses an upper bound. In each inductively ordered set, there exist
maximal elements (Zorn's lemma). In most applications of Zorn's lemma,
the set in question is a family of subsets of a set S, ordered by set theoretical
inclusion c.
6. Filters. Let X be a set. A set ty of subsets of X is called a filter on X if
it satisfies the following axioms:
(1) ty # 0 and 0 rj ty.
(2) FE ty and F c G c X implies G E ty.
(3) FE ty and G E ty implies F (\ G E 3'.
A set ~ of subsets of X is a filter base if (1 ') ~ # 0 and 0 rj ~, and (2') if
B1 E ~ and B2 E ~ there exists B3 E ~ such that B3 c B1 (\ B2 • Every filter
base ~ generates a unique filter ty on X such that FE ty if and only if
Be F for at least one B E ~; ~ is called a base of the filter ty. The set of all
filters on a non-empty set X is inductively ordered by the relation tyl c ty2
(set theoretic inclusion of Ij3(X»; \j1 c \j2 is expressed by saying that ty1 is
coarser than ~2' or that ~2 is finer than ty1' Every filter on X which is maximal
with respect to this ordering, is called an ultrafilter on X; by Zorn's lemma,
for each filter ~ on X there exists an ultrafilter finer than 0:: If {x~: a E A}
is a directed family in X, the ranges of the sections of this family form a filter
base on X; the corresponding filter is called the section filter of the family.
4 PREREQUISITES
An elementary filter is the section filter of a sequence {xn : n EN} in X (N

being endowed with its usual order).
Literature. Sets: Bourbaki [1], Ha1mos [3]. Filters: Bourbaki [4], Bushaw
[1]. Order: Birkhoff [1], Bourbaki [1]. .
B. GENERAL TOPOLOGY
1. Topologies~ Let X be a set, (fj a set of subsets of X invariant under finite

intersections and arbitrary unions; it follows that X E (fj, since X is the inter-
section of the empty subset of (fj, and that 0 E (fj, since 0 is the union of
the empty subset of (fj. We say that (fj defines a topology;r on X; structurized
in this way, X is called a topological space and denoted by (X, ;r) if reference
to ;r is desirable. The sets G E (fj are called open, their complements F = X '" G
are called closed (with .respect to ;r). Given A c X, the open set A (or int A)
which is the union oLall open subsets of A, is called the interior of A; the
closed set ..4, intersection of all closed sets containing A, is called the closure
of A. An element XE A is called an interior point of A (or interior to A), an
element x E..4 is called a contact point (adherent point) of A. If A,. B are subsets
of X, B is dense relative to A if A c Ii (dense in A if B c A and A c Ii). A
topological space X is separable if X contains a countable dense subset; X is
connected if X is not the union of two disjoint non-empty open subsets
(otherwise,' X is disconnected).
Let X be a topological space. A subset UC X is a neighborhood of x if
x E 0, and a neighborhood of A if x E A implies x E 0. The set of all neigh-
borhoods of x. (respectively, of A) is a filter on X called the neighborhood
filter of x (respectively, of A); each base of this filter is a neighborhood base
of x (respectively, of A). A bijectionf of X onto another topoiogical space Y
such that f(A) is open in Y if and only if A is open in X, is called a homeo-
morphism; X and Yare homeomorphic if there exists a homeomorphism of
X onto Y. The discrete topology on X is the topology for which every subset
of X is open; the trivial topology on X is the topology whose only open sets
are 0 and X.
2. Continuity and Convergence. Let X, Y be topological spaces and let
f: X -+ Y. f is continuous at x E X if for each neighborhood V of y = f(x),
f- 1(V) is a neighborhood of x (equivalently, if the filter on Y generated by
the base feU) is finer than m, where U is the neighborhood filter of x, m the
neighborhood filter of y). fis continuous on X into Y (briefly, continuous) if
fis continuous at each x E X (equivalently, if 1- 1(G)}s open in X for each
open G c Y). If Z is also a topological space and I: X -+ Yand g: Y -+ Z are
continuous, then 9 0 f: X -+ Z is continuous.
A filter ~ on a topological space X is said to converge to x E X if ~ is finer
than the neighborhood filter of x. A sequence (more generally, a directed
family) in X converges to x E X if its section filter converges to x. If also Y
§B] GENERAL TOPOLOGY 5
is a topological space and 3' is a filter (or merely a filter base) on X, and if
/: X -+ Y is a map, then / is said to converge to y E Y along 3' if the filter
generated by 1(3') converges to y. For example, / is continuous at x E X if
and only if/converges to y = /(x) along the neighborhood filter of~. Given a
filter 3' on X and x E X, x is a cluster point (contact point, adherent point) of
3' if x E F for each FE 3'. A cluster point of a sequence (more generally, of a
directed family) is a cluster point of the section filter of this family.
3. Comparison of Topologies. If X is a set and ~l' ~2 are topologies on X,
we say that ~2 is finer than'~l (or ~l coarser than ~2) if every ~l-open set
is ~2-open (equivalently, if every ~l-closed set is ~2-closed). (if (fjl and (fj2
are the respective families of open sets in X, this amounts to the relation
(fjl c (fj2 in ~(~(X».) Let {~..: (X E A} be a family of topologies on X. There
exists a finest topology ~ on X which is coarser than each ~..(ex E A); a set G
is ~-open if and only if G is ~..-open for each ex. Dually, there exists a coarsest
topology ~o which is finer than· each ~..(ex E A). Jf we denote by <»~ the set
Of all finite intersections of sets open for some ~.. , the set (fjo of all unions of
sets in (fj~ constitutes the ~o-open sets in X. Hence with respect to the relation
"~2 is finer than ~l ", the set of all topologies on X is a complete lattice;
the coarsest topology on. X is the trivial topology, the finest topology is the
discrete topology. The topology ~ is the greatest lower bound (briefly, the
lower bound) of the family {~.. : ex E A}; similarly, ~o is the upper bound of the
family {~..: ex E A}.
One derives from this two general methods of defining a topology (Bourbaki
[4]). Let X be a set, {X.. : ex E A} a family of topological spaces. If {fa: ex E A}
is a family of mappings, respectively of X into X .. , the projective topology
(kernel topology) on X with respect to the family {( X .., fa): ex E A} is the coarsest
topology for which eachfa is continuous. Dually, if {g,,: ex E A} is a family of
mappings, respectively of X .. into X, the inductive topology (hull topology)
with respect to the family {(X"' g.. ): ex E A} is the finest topology on X for
which each g .. is continuous. (Note that eachj~ is continuous for the discrete
topology on X, and that each g .. is continuous for the trivial topology on X.)
If A = {I} and ~l is the topology of Xl' the projective topology on X with
respect to (Xl,ii) is called the inverse image of ~l under ii, and the inductive
topology with respect to (X1o gl) is called the direct image of ~l under g1'
4. Subspaces, Products, Quotients. If (X, ~) is a topological space, A a
subset of X, Jthe canonical imbedding A -+ X, then the induced topology on
A is the inverse image of ~ under f. (The open subsets of this topology are
the intersections with A of the open subsets of X.) Under the induced
topology, A is called a topological subspace of X (in -general, we shall avoid
this terminology because of possible confusion with vector. subspaces). If
(X, 1:) is a topological space, R an equivalence relation on X, 9 the canonical
map X --+ X/R, then the direct image of ~ under 9 is called the quotient
(topology) of ~; under this topology, X/R is the topological quotient of
XbyR.
6 PREREQUISITES
Let {Xa: IX E A} be a family of topological spaces, Xtheir Cartesian product,

fa the projection of X onto Xa' The projective topology on X with respect to
the family {(Xa,ja): IX E A} is called the product topology on X. Under this
topology, X is called the topological product (briefly, product) of the family
{Xa: IX E A}.
Let X, Y be topological spaces,j a mapping of X into Y. We say that f is
open (or an open map) if for each open set G c X,j(G) is open 'n the topo-
logical subspace f(X) of Y. f is called closed (a closed map) if the graph of
fis a closed subset of the topological product X x Y.
5. Separation Axioms. Let X be a topological space. X is a Hausdorff (or
separated) space if for each pair of distinct points x,y there are respective
neighborhoods Ux , Uy such that Ux n Uy = 0. If (and only if) X is separated,
each filter ~ that converges in X, converges to exactly one x EX; x is called
the limit of!y. X is called regular if it is separated and each point possesses a
base of closed neighborhoods; X is called normal if it is separated and for
each pair A, B of disjoint closed subsets of X, there exists a neighborhood U
of A and a neighborhood V of B such that Un V = 0.
A Hausdorff topological space X is normal if and only if for each pair
A, B of disjoint closed subsets of X, there exists a continuous function f on
X into the real interval [0, I] (under its usual topology) such that f(x) =
whenever x E A,f(x) = 1 whenever x E. B (Urysohn's theorem).
°
A separated space X such that for each closed subset A and each b ¢ A,
there exists a continuous functionf: X -+ [0,1] for whichf(b) = I andf(x) =
whenever x E A, is called completely regular; clearly, every normal space is
°
completely regular, and every completely regular space is regular.
6. Uniform Spaces. Let X be a set. For arbitrary subsets W, V of X x X,
we write W- 1 = {(y, x): (x, y) E W}, and VoW = {(x, z): there exists y EX
such that (x, y) E W, (y, z) E V}. The set ~ = {(x, x): x E X} is called the
diagonal of X x X. Let W be a filter on X x X satisfying these axioms:
(1) Each WE W contains the diagonal ~.

(2) WE W implies W- 1 E W.
(3) For each WE IlU, there exists V E IlU such that V 0 V c W.
We say that the filter W (or anyone of its bases) defines a uniformity (or
uniform structure) on X, each WE W being called a vicinity (entourage) of
the uniformity. Let (f) be the family of all subsets G of X such that x E G
implies the existence of WE IlTI satisfying {y: (x, y) E W} c G. Then (f) is
invariant under finite intersections and arbitrary unions, and hence defines
a topology::! on Xsuch that for each x E X, the family W(x) = {y: (x, y) E W},
where W runs through W, is a neighborhood base of x. The space (X, IlU),
endowed with the topology ::! derived from the uniformity 1113, is called a
uniform space. A topological space X is uniformisable if its topology can be
§B] GENERAL TOPOLOGY 7
derived from a uniformity on X; the reader should be cautioned that, in

general, such a uniformity is not unique.
A uniformity is separated if its vicinity filter satisfies the additional axiom
(4) n {W: WE[O} = A.
(4) is a necessary and suffiCient condition for the topology derived from the
uniformity to be a Hausdorff topology. A Hausdorff topological space is
uniformisable if and only if it is completely regular.
Let X, Y be uniform spaces. A mappingf: X -+ Y is uniformly continuous
if for each vicinity Vof Y, there exists a vicinity U of X such that (x,y) E U
implies (f(x), fey»~ E V. Each uniformly continuous map is continuous. The
uniform spaces X, Yare isomorphic if there exists a bijection f of X onto Y
such that bothfandf-1 are uniformly continuous;fitselfis called a uniform
isomorphism.
If [01 and [02 are two filters on X x X, each defining a uniformity on the
set X, and if [01 c [02' we say that the uniformity defined by [01 is coarser
than that defined by [02. If X is a set, {Xa: ex E A} a family of uniform spaces
andh(ex E A) are mappings of X into X a , then there exists a coarsest uniformity
on X for which each h(ex E A) is uniformly continuous. In this way, one
defines the product uniformity on X = n~Xa to be the coarsest uniformity for
which each of the projections X -+ X" is uniformly continuous; similarly,
if X is a uniform space and A c X, the induced uniformity is the coarsest
uniformity on A for which the canonical imbedding A -+ X is uniformly
continuous.
Let X be a uniform space. A filter 0: on X is a Cauchy filter if, for each
vicinity V, there exists FE 0: such that F x F c V. If each Cauchy filter
converges (to an element of X) then X is called complete. To each uniform
space X one can construct a complete uniform space g such that X is
(uniformly) isomorphic with a dense subspace of g, and such that g is
separated if X is. If X is separated, then g is determined by these properties
to within isomorphism, and is called the completion of X. A base of the
vicinity filter of X can be obtained by taking the closures (in the topolog-
ical product X x g) of a base of vicinities of X. A Cauchy sequence in
X is a sequence whose section filter is a Cauchy filter; if every Cauchy
sequence in X converges, then X is said to be semi-complete (sequentially
complete).
If X is a complete uniform space and A a closed subspace, then the uniform
space A is complete; if X is a separated uniform space and A a complete
subsp~ce, then A is closed in X. A product of uniform spaces is complete if
and only if each factor space is complete ..
If Xis a uniform space, Ya complete separated space, Xo c X and/: Xo -+ y
uniformly continuous; then f has a unique uniformly continuous extension
J:X o -+ Y.
7. Metric and Metrizable Spaces. If X is a set, a non-negative real function
d on X x X is called a metric if the following axioms are satisfied:
8 PREREQUISITES
(1) d(x, y) = 0 is equivalent with x = y.

(2) d(x, y) = dey, x).
(3) d(x, z) ;;;::; d(x, y) + dey, z) (triangle inequality).
Clearly, the sets Wn = {(x, y):d(x, y) < n- 1 }, where n eN, form a filter base
on X x X defining a separated uniformity on X; by the metric space (X, d) we
understand the uniform space X endowed with the metric d. Thus all uniform
concepts apply to metric spaces. (It should be understood that, historically,
uniform spaces are the upshot of metric spaces.) A topological space is
metrizable if its topology can be derived from a metric in the manner indicated;
a uniform space is metrizable (i.e., its uniformity can be generated by a
metric) if and only if it is separated and its vicinity filter has a countable base.
Clearly, a metrizable uniform space is complete if it is semi-complete.
8. Compact and Precompact Spaces. Let X be a Hausdorff topological
space. X is called compact if every open cover of X has a finite subcover.
For X to be compact, each of the following conditions is necessary and
sufficient: (a) A family of closed subsets of X has non-empty intersection
whenever each finite subfamily has non-empty intersection. (b) Each filter
on X has a cluster point. (c) Each ultrafilter on X converges.
Every closed subspace of a compact space is compact. The topological pro-
duct of any family of compact spaces is compact (Tychonov's theorem). If X
is compact, Ya Hausdorff space, andf: X ~ Y continuous, thenf(X) is a com-
pact subspace of Y. Iff is a continuous bijection of a compact space X onto a
Hausdorff space Y, thenf is a homeomorphism (equivalently: If (X, ~l) is com-
pact and ~2 is a Hausdorff topology on X coarser than ~b then ~l = ~2).
There is the following important relationship between compactness and
uniformities: On every compact space X, there exists a unique uniformity
generating the topology of X; the vicinity filter of this uniformity is the
neighborhood filter of the diagonal A in the topological product X x X. In
particular, every compact space is a complete uniform space. A separated
uniform space is called precompact if its completion is compact. (However,
note that a topological space can be precompact for several distinct uni-
formities yielding its topology.) X is precompact if and only if for each
vicinity W, there exists a finite subset Xo c X such that X c U {W(x): x e X o}.
A subspace of a precompact space is precompact, and the product of any
family of precompact spaces is precompact.
A Hausdorff topological space is called locally compact if each of its points
possesses a compact neighborhood.
9. Category and Baire Spaces. Let X be a topological space, A a subset of
X. A is called nowhere dense (rare) in X if its closure A has empty interior;
A is called meager (of first category) in X if A is the union of a countable set
of rare subsets of X. A subset A which is not meager is called non-meager (of
second category) in X; if every non-empty open subset is nonmeager in X,
then X is called a Baire space. Every locally compact space and every complete
§C] LINEAR ALGEBRA 9
metrizable space is a Baire space (Baire's theorem). Each non-meager subset

of a topological space X is non-meager in itself, but a topological subspace
of X can be a Baire space while being a rare subset of X.
Literatl,f,re: Berge [1]; Bourbaki [4], [5], [6]; Kelley [1]. A highly recom-
mendable introduction to topological and uniform spaces can be found in
Bushaw [1].
C. LINEAR ALGEBRA
1. Vector Spaces. Let L be a set, K a (not necessarily commutative) field.

Suppose there are defined a mapping (x, y) -+ x + y of L x L into L, called
addition, and a mapping (A, x) -+ Ax of K x L into L, called scalar mUltiplica-
tion, such that the following axioms are satisfied (x, y, z denoting arbitrary
elements of L, and A, Jl arbitrary elements of K):
(1) (x + y) + z = x + (y + z).
(2) x + y = y + x.
(3) There exists an element 0 E L such that x + 0 = x for all x E L.
(4) For each x E L, there exists z E L such that x + z = O.
(5) A(X + y) = AX + Ay.
(6) (A + Jl)x = AX + JlX.
(7) A(JlX) = (AJl)X.
(8) Ix = x.
Endowed with the structure so defined, L is called a left vector space over
K. The element 0 postulated by (3) is unique and called the zero element of L.
(We shall not distinguish notationally between the zero elements of Land
K.) Also, for any x E L the element z postulated by (4) is unique and denoted
by -x; moreover, one has -x = (-l)x, and it is customary to write x - y
for x + (-y).
If (1)-(4) hold as before but scalar multiplication is written (A, x) -+ XA and
(5)-(8) are changed accordingly, L is called a right vector space over K. By
a vector space over K, we shall always understand a left vector space over K.
Since there is no point in distinguishing between left and right vector spaces
over K when K is commutative, there will be no need to consider right vector
spaces except in CA below, and Chapter I, Section 4. (From Chapter II on,
K is always supposed to be the real field R or the complex field C.)
2. Linear Independence. Let L be a vector space over K. An element
AIXI + ... + AnXn," where n EN, is called a linear combination of the elements
n
Xi E L(i = 1, ... , n); as usual, this is written L AiXi or LiAiXi' If {xa: or: E H}
i= 1
is a finite family, the sum of the elements xa is denoted by L Xa; for reasons
aeH
of convenience, this is extended to the empty set by defining L x = O. (This
xe0
10 PREREQUISITES
should not be confused with the symbol A +.B for subsets A, B of L, which
by A.2 has the meaning {x + y: x E A, Y E B}; thus if A = 0, then A + B = 0
for all subsets BeL.) A subset A c L is called linearly independent if for every
non-empty finite subset {Xi: i = 1, ... , n} of A, the relation ~:>{iXi = 0 implies
Ai = 0 for i = 1, ... , n. Note that by this definition, the empty subset of L is
linearly independent. A linearly independent subset of L which is maximal
(with respect to set inclusion) is called a basis (Hamel basis) of L. The existence
of bases in L containing a given linearly independent subset is implied by
Zorn's lemma; any two bases of L have the same cardinality d, which is called
the dimension of L (over K).
3. Subspaces and Quotients. Let L be a vector space over K. A vector
subspace (briefly, subspace) of L is a non-empty subset M of L invariant under
addition and scalar multiplication, that is, such that M + M c M and
KM eM. The set of all subspaces of L is clearly invariant under arbitrary
intersections. If A is a subset of L, the linear hull of A is the intersection M of
all subspaces of L that contain A; M is also said to be the subspace of L
generated by A. M can also be characterized as the set of all linear com-
binations of elements of A (including the sum over the empty subset of A).
In particular, the linear hull of 0 is {O}.
If M is a subspace of L, the relation" x - y E M" is an equivalence
relation in L. The quotient set becomes a vector space over K by the definitions
~ + y = x + y + M, AX = AX + M where ~ = x + M, y = y + M, and is
denoted by Lj M.
4. Linear Mappings. Let L I , L2 be vector spaces over K. I: Ll --+ L2 is
called a linear map if I(AIXI + A2X2) = At!(X1 ) + Ad(X2) for all AI' A2 E K
and XI' x 2 ELI. Defining addition by (II + j~)(x) =!t(x) +/2(X) and scalar
multiplication' by (lA)(X) = I{AX)(X ELI)' the set L(Ll' L 2) of all linear maps
of Ll into L2 generates a right vector space over K. (If K is commutative, the
mapping x --+ I(AX) will be denoted by AI and L(Ll' L 2 ) considered to be a left
vector space over K.) Defining (lA)(X) = I(X)A if L2 is the one-dimensional
vector space Ko(over K) associated with K, we obtain the algebraic dual L'j'
of L 1 • The elements of L'j' are called linear forms on L 1 •
L 1 and L2 are said to be isomorphic if there exists a linear bijective map
I: Ll --+ L 2 ; such a map is called an isomorphism of Ll onto L 2. A linear
injective map I: Ll --+ L2 is called an isomorphism of L1 into L 2 •
If I: Ll --+ L2 is linear, the subspace N = 1- 1(0) of Ll is called the null
space (kernel) off I defines an isomorphism fo of Ld N onto M = I(L1 ); 10
is called the bijective map associated with f If ¢ denotes the quotient map
Ll --+ Ld Nand", denotes the canonical imbedding M --+ L 2 , thenl = '" 10 ¢
0 0
is called the canonical decomposition off

5. Vector Spaces over Valuated Fields. Let K be a field, and consider the
real field R under its usual absolute value. A function A --+ IAI of K into R+
(real numbers ~ 0) is called an absolute value on K if it satisfies the following
axioms:
§C] LINEAR ALGEBRA 11
(1)IAI = 0 is equivalent with A= O.

(2)IA + III ~ IAI + 11l1·
(3) IAIlI = IA/lIlI.
The function. (Je, Il) -+ IA - III is a metric on K; endowed with this metric and
the corresponding uniformity, K is called a valuated field. The valuated field
K is called non-discrete if its topology is not discrete (equivalently, if the
range of A-+ IAI is distinct from {O,I}). A non-discrete valuated field is neces-
.sarily infinite.
Let L be a vector space over a non-discrete valuated field K, and let A, B be
subsets of L. We say that A absorbs B if there exists Ao E K such that B c AA
whenever 1.1.1 ~ 1.1. 0 1. A subset U of L is called radial (absorbing) if U absorbs
every finite subset of L. A subset C of L is circled if AC c C whenever 1.1.1 ~ 1.
The set of radial subsets of L is invariant under finite intersections; the
set of circled subsets of L is invariant under arbitrary intersections. If A c L,
the circled hull of A is the intersection of aJI circled subsets of L containing A.
Let f: Ll -+ L2 be linear, Ll and L2 being vector spaces over a non-discrete
valuated field K. If A eLI and Be L2 are circled, thenf(A) andf-l(B) are
circled. If B is radial then f-1(B) is radial; if A is radial and f is surjective,
thenf(A) is radial.
The fields Rand C of real and complex numbers, respectively, are always
considered to be endowed with their usual absolute value, under which they
are non-discrete valuated fields. In addition, R is always considered under its
usual order.
Literature: Baer [1]; Birkhoff'-MacLane [1]; Bourbaki [2], [3], [7].

PREREQUISITES

ences can be found.
A. SETS AND ORDER

1

§1] VECTOR SPACE TOPOLOGIES 13
products. Loosely speaking, these axioms require addition and scalar multi-
plication to be (jointly) continuous. Since, in particular, this implies the
continuity of (x, y) -+ x - y, every t.V.S. is a commutative topological group.
A t.v.s. (L, Z) will occasionally be denoted by L(Z), or simply by L if the
topology of L does not require special notation. "-
Two t.V.S. Ll and L2 over the same field K are called isomorphic if there
exists a biunivocal linear mapu of Ll onto L2 which is a homeomorphism;
u is called an isomorphism of Ll onto L 2 • (Although mere algebraic isomor-
phisms will, in general, be designated as such, the terms "topological iso-
mQrphism" and "topologically isomorphic" will occasionally be used to
avoid misunderstanding.) The following assertions are more or less immediate
consequences of the definition of a t.v.S.
1.1
Let L be a t.V.S. over K.

(i) For each Xo ELand each AO E. K,AO # 0, the mapping x -+ AOX + Xo is
a homeomorphism of L onto itse/f.
(ii) For any subset Aof L and any base 11 of the neighborhood/ilter of 0 E L,
n
the closure A is given by A = {A + U: U E U}.
(iii) If A is an open subset of L, and B is any subset of L, then A + B is open.
(iv) If A, B are closed subsets q( L such that every filter on A has an adherent
point (in particular, such that A is compact), then A + B is closed.
(v) If A is a circled subset of L, then its closure A is circled, and the interior
A of A is circled when 0 E A.
Proof (i): Clearly, x -+ AOX + Xo is onto L and, by (LT)l and (LT)z, con-
tinuous with continuous inverse x -+ Ai) l(X - xo). Note that this assertion,
as well as (ii), (iii), and (v), requires only the separate continuity of addition
and scalar multiplication.
(ii): Let B = n {A + U: U E U}. By (i), {x - U: U E U} is a neighborhood
base of x for each x E L; hence x E B implies tbat each neighborhood of x
intersects A, whence B c A. Conversely, if x E A then x E A + U for each
O-neighborhood U, whence A c B.
(iii): Since A + B = U {A + b: b E B}, A + B is a union of open subsets
of L if A is open, and hence an open subset of L.
(iv): We show that for each Xo rf= A + B there exists a O-neighborhood U
such that (xo - U) (") (A + B) = 0 or, equivalently, that (B + U) (") (xo - A)
= 0. If this were .not true, then the intersections (B <I- U) (") (xo - A) would
form a filter base on Xo - A (as U runs through a O-neighborhoodbase in
L). By the assumption on A, this filter base would have an adherent point
Zo E Xo - A, also contained in the closure of B + U and hence in B + U + U,
for all U. Since by (LT)l' U + U runs through a neighborhood base of 0 as
U does, (ii) implies that Zo E B, which is contFadictory.
14 TOPOLOGICAL VECTOR SPACES [Ch. I
(v): Let A be circled and let IAI ~ 1. By (LT)2' AA c A implies AA c A;

hence A is circled. Also if A. :F 0, A.A is the interior of AA by (i) and hence
contained in .A. The assumption 0 E.A then shows that A.A c.A whenever
IAI ~ 1.
In the preceding proof we have repeatedly made use of the fact that in a
t.v.s., each translation x -+ x + Xo is a homeomorphism (which is a special
case of (i»; a topology ;r on a vector space L is called translation-invariant
if all translations are homeomorphisms. Such a topology is completely
determined by the neighborhood filter of any point x E L, in particular by the
neighborhood filter of O.
1.2
A topology ;r on a vector space Lover K satisfies -the axioms (LT)l and

(LTh if and only if;r is translation-invariant and possesses a O-neighborhood
base 58 with the following properties:
(a) For each V E 58, there exists U E 58 such that U + U c V.

(b) Every V E 58 is radial and circled.
(c) There exists A E K, 0 < IAI < I, such that V E 58 implies AV E 58.
If K is an Archimedean valuated field, condition (c) is dispensable (which is,

in particular, the case if K = R or K = C).
Proof We first prove the existence, in every Lv.s., of a O-neighborhood base

having the listed properties. Given a O-neighborhood Win L, there exists a
O-neighborhood U and a real number B > 0 such that AU c W whenever
IAI ~ B, by virtue of (LTh; hence since K is non-discrete, V = U {AU: IAI
~ B} is a O-neighborhood which is contained in W, and obviously circled.
Thus the family 58 of all circled O-neighborhoods in L is a base at O. The
continuity at A = 0 of (A,xo) -+ AXo for each Xo E L implies that every V E 58
is radial. It is obvious from (LT)l that 58 satisfies condition (a); for (c), it
suffices to observe that there exists A E K such that 0 < IAI < I, since K is
non-discrete, and that AV (V E 5B), which is a O-neighborhood by (l.l) (i), is
circled (note that if 1111 ~ I then 11 = AVr l where Ivl ~ I). Finally, the top-
ology of L is translation-invariant by (l.l) (i).
Conversely, let ;r be a translation-invariant topology on L possessing a
O-neighborhood base 58 with properties (a), (b), and (c). We have to show that
;r satisfies (LT)l and (LT)2' It is clear that {xo + V: V E 58} is a neighborhood
base of Xo E L; hence if V E 58 is given and U E 58 is selected such that
U + U c V, then x - Xo E U, Y - Yo E U imply that x + Y E Xo + Yo + V; so
(LT)l holds. To prove the continuity of the mapping (A, x) -+ AX, that is
(LTh, let Ao, Xo be any fixed elements of K, L respectively. If V E 58 is given,
by (a) there exists U E 58 such that U + U c V. Since by (b) U is radial, there
exists a real number B > 0 such that (A. - A.o)xo E U whenever IA. - Aol ~ B.
Let p. e K satisfy (c); then there exists an integer n eN such that Ip.-nl =
1p.I-n ~ IAol + e; let We m be defined by W= p."U. Now since Uis circled,
the relations x - Xo e Wand IA - Aol ~ e imply that A(X - xo) e U, and
hence the identity
Ax = AOXO + (A - AO )xo + A(X - xo)
implies that AX e AoXo + U + U C AOXO + V, which proves (LTh.
Finally, if Kis an Archimedean valuated field, then 121> 1 for 2 e K. Hence
12"1 = 121" > IAol + e (notation of the preceding paragraph) for a suitable
n eN. By repeated application of (b), we can select a WI em such that
2"WI C WI + ... + WI C U, where the sum has 2" summands (2eN).
Since WI (and hence 2"WI ) is circled, WI can be substituted for W in the
preceding proof of (LT)z, and hence (c) is dispensable in this case. This
completes the proof of (1.2).
COROLLARY. If L is a vector space over K and mis aftlter base in L having
the properties (a) through (c) of (1.2), then mis a neighborhood base 01 lor
a unique topology ~ such that (L, ~) is a t.V.S. over K.
°
Proof. We define the topology ~ by specifying a subset GeL to be open
whenever x e G implies x + V c G for some Ve m. Clearly ~ is the unique
translation-invariant topology on L for which m is a base at 0, and hence
the unique topology with this property and such that (L, ~) is a t.v.s.
Examples
In the following examples, K can be any non-discrete valuated field; for
instance, the field of p-adic numbers, or the field of quaternions with their
usual absolute values, or any subfield of these such as the rational, real, or
complex number field (with the respective induced absolute value).
1. Let A be any non-empty set, KA the set of all mappings ex --+- e" of A
into K; we write x = (e.), y = (TJJ to denote elements x,y of KA. Defin-
ing addition by x + y = (e" + TJ,,}and scalar multiplication by AX = (Ae,,),
it is immediate that KA becomes a vector space over K. For any finite
subset H c A and any real number e > 0, let VH • be the subset
{x: le,,1 ~ e if ex e H} of KA; it is clear from (1.2) that' the family of all
these sets VH • is a O-neighborhood base for a unique topology under
a
which KA is t.v.S.
2. Let X be any non-empty topological space; the set of all con-
tinuous functions I on X into K such that sup I/(t) I is finite is a subset
leX
of K X , which is a vector space CCK(X) under the operations of addition
and scalar multiplication induced by the vector space K X (Example 1);
the sets U" = {f: sup II(t) I ~ n- I } (n eN) form a neighborhood base
teX
of 0 for a unique topology under which CCK(X) is a t.v.s.
3. Let K[t] be the ring of polynomials I[t] = L"ac"t" over K in one
indeterminate t. With multiplication restricted to left multiplication by
polynomials of degree 0, K[t] becomes a vector space over K. Let r be

a fixed real number such that 0 < r ;;;i! 1 and denote by V. the set of
polynomials for which Lnlcx"I' ;;;i! 8. The family {VB: e> O} is a O-neigh-
borhood base for a unique topology under which K[t] is a t.v.S.
1.3
If L is a t.V.S. and x E L, each neighborhood of x contains a closed
neighborhood of x. In particular, the family of all closed O-neighborhood forms
a base atO.
Proof For any O-neighborhood U there exists another, V, such that
V + V c U. Since y E r only if (y - V) () V is non-empty, it follows that
reV + V c U. Hence x + U contains the closed neighborhood x + r of x.
Since by (1.2) any O-neighborhood contains a circled O-neighborhood, and
hence by (l.l) (v) and (1.3) a closed, circled O-neighborhood, we obtain the
following corollary:
COROLLARY. If L is a t.V.S. and U is any neighborhood base of 0, then the
closed, circled hulls of the sets U E U form again a base at O.
(1.3) shows that every Hausdorff t.v.s. is a regular topological space. It
will be seen from the next proposition that every t.V.S. is uniformisable, hence
every Hausdorfft.v.s. is completely regular. A uniformity on a vector space L
is called translation.invariant if it has a base 91 such that (x, y) E N is equiva-
lent with (x +z,y +z) EN for each Z EL and each N Em.
1.4
The topology of any t.v.s. can be derived from a unique translation-invariant
uniformity 91. If ~ is any neighborhood base of 0, the family Ny = {(x, y):
x - yE V}, VE ~ is a basefor 91.
Proof Let (L,:t) be a t.V.S. with O-neighborhood base ~. It is immediate
that the sets Ny, V E ~ form a filter base on L x L that is a base for a trans-
lation-invariant uniformity 91 yielding the topology :t of L. If 91 1 is another
uniformity with these properties, there exists a base IDl of 91 1 , consisting of
translation-invariant sets, and such that the sets
UM = {x - y: (x,y) EM} MEIDl
form a O-neighborhood base for :to Since UMeV implies M c Ny and
conversely, it follows that 91 1 = 91.
The fact that there is a unique translation-invariant uniformity from which
the topology of a t.v.S. can be derived is of considerable importance in the
theory of such spaces (as it is for topological groups), since uniformity
concepts can be applied unambiguously to arbitrary subsets A of a t.v.S. L.
The uniformity meant is, without exception, that induced on A c L by the
uniformity 91 of (1.4). For example, a subset A of a t.v.S. L is complete if
and only if every Cauchy filter in A converges to an element of A; A is semi-

complete (or sequentially complete) if and only if every Cauchy sequence in A
converges to an element of A. It follows from (1.4) that a filter tY iii A is a
Cauchy filter if and only if for each O-neighborhood, V in L, there exists FE tY
such that F - Fe V; accordingly, a sequence {xn: n E N} in A is a Cauchy
sequence if and only if for each O-neighborhood V in L there exists no EN
such that Xm - Xn E V whenever m ~ no and n ~ no.
A t.V.s. L is a Hausdorff (or separated) topological space if and only if L
is a separated uniform space; hence by (1.4), L is separated if and only if
n {U: U E U} = {OJ, where U is any neighborhood base of 0 in L. An equiva-
lent condition is that for each non-zero x E L, there exists a O-neighborhood
U such that x 1: U (which is also immediate from (1.3».
Recall that a subspace (vector subspace, linear subspace) of a vector space
Lover K is defined to be a subset M #- 0 of L such that M + M c M and
KM c: M. If L is a t.v.s., by a subspace of L we shall understand (unless the
contrary is expressly stated) a vector subspace M endowed with the topology
induced by L; clearly, M is a t.v.s. which is separated if Lis.
If L is a Hausdorff t.v.s., the presence of a translation-invariant separated
uniformity makes it possible to imbed L as a dense subspace of a complete
Hausdorff t.v.s. L which is essentially unique, and is called the completion
of L. (See also Exercise 2.)
1.5
Let L be a Hausdorff t.V.S. over K. There exists a complete Hausdorff t.v.S.

Lover K containing L as a dense subspace; L is unique to within isomorphism.
Moreover, for (my O-neighborhood base min L, thefamily W = {V: VE m} of
closures in l is a O-neighborhood base in L.
Proof. We assume it known (cf. Bourbaki [4], chap. II) that there exists a
separated, complete uniform space L which contains L as a dense subspace,
and which is unique up to a uniform isomorphism. By (1.4) (x, y) -+ x + y is
uniformly continuous on L x L into L, and for each fixed A E K (A, x) -+ AX
is uniformly continuous on L into L; hence these mappings have unique
continuous (in fact, uniformly continuous) extensions to L x Land L,
respectively, with values in L. It is quickly verified (continuation of identities)
that these extensions make L into a vector space over K. Before showing that
the uniform space L is a t.v.S. over K, we prove the second assertion. Since
{Ny: V Em} is a base of the uniformity m of L (notation as in (1.4», the
closures Ny of these sets in L x l form a base of the uniformity 91 of L; we
assert that Nv = Ny for all V Em. But if (x, y) E Ny, then x - y E V, since
(x, y) -+ x - Y is continuous on l x l into l. Conversely, if x - Y E V, then
we have x E y + V; hence x is in the closure (taken in L) of y + V, since
translations in L are homeomorphisms; this implies that (x, y) E Ny.
It follows that 2D is a neighborhood base of 0 in L; we use (1.2) to show that

under the topology l: defined by 91, L is a t.v.s. Clearly, i is translation-
invariant and satisfies conditions (a) and (c) of (1.2); hence it suffices to
show that each Y E 2D contains a i-neighborhood of 0 that is radial and
circled. Given V E m, there exists a circled O-neighborhood U in L such that
U + U c V. The closure (U + U) - in L is a O-neighborhood by the pre-
ceding, is circled and clearly contained in Y. Let us show that it is radial.
Given x E L, there exists a Cauchy filter iY in L convergent to x, and an FE iY
such that F - F c U. Let Xo be any element of F; since U is radial there
exists A. E K such that Xo E )'U, and since U is circled we can assume that
IA.I?; 1. Now F-xoc U; hence Fcxo+ U and xEFc).(U+ U)-,
which proves the assertion.
Finally, the uniqueness of (L, l:) (to within isomorphism) follows, by
virtue of (104), from the uniqueness of the completion L of the uniform
space L.
REMARK. The completeness of the valuated field K is not required for
the preceding construction. On the other hand, if L is a complete
Hausdorfft.v.s. over K, it is not difficult to see that scalar multiplication
has a unique continuous extension to K x L, where K is the completion
of K. Thus it follows from (1.5) that for every Hausdorff t.v.s. over K
there exists a (essentially unique) complete Hausdorff t.v.s. L1 over K
such that the topological group L is isomorphic with a dense subgroup
of the topological group L 1 •
We conclude this section with a completeness criterion for a t.v.s. (L, l:1)
in terms of a coarser topology l:2 on L.
1.6
Let L be a vectQr space over K and let l:l' l:2 be Hausdorff topologies under
each of which L is a t.v.s., and such that l:l is finer than l:2. If (L, l:l) has a
neighborhood base of 0 consisting of sets complete in (L, l:2)' then (L, l:l) is
complete. '
Proof Let ml be a l:l-neighborhood base of 0 in L consisting of sets
complete in (L, l:2). Given a Cauchy filter iY in (L, l:l) and VI E ml , there
exists a set Fo E iY such that Fo - Fo C VI. If Y is any fixed element of Fo,
the family {y - F: FE iY} is a Cauchy filter base for the uniformity associated
with l:2' for which VI is complete; since y - Fo C VI' this filter base has a
unique l:2-limit y - Xo. It is now clear that Xo E L is the l:rlimit of iY. Since
V 1 is l:2-closed, we have Fo - Xo C VI or Fo c Xo + VI; VI being arbitrary,
this shows iY to be finer than the l:1-neighborhood filter of xo and thus
proves (L, l:1) to be complete.
For the reader familiar with normed spaces, we point out this example
for (1.6): -Every reflexive normed space is complete and hence is a Banach
§2) PRODUCT SPACES. SUBSPACES. DIRECT SUMS. QUOTIENT SPACES 19
space. For in such a space the positive multiples of the closed unit ball,
which form a O-neighborhood base for the norm topology, are weakly
compact and hence weakly complete.
2. PRODUCT SPACES, SUBSPACES. DIRECT SUMS. QUOTIENT SPACES
Let {La: a E A} denote a family of vector spaces over the same scalar field
K; the Cartesian product L = TIaLa is a vector space over K if for x = (x",),
y = (y",) ELand A E K, addition and scalar multiplication are defined by
x + y = (x", + y",), AX = (Ax",). If (L"" Z",) (a E A) are t.v.S. over K, then L is a
t.v.s. under the product topology Z = TI",Z",; the simple verification of
(LT)l and (LTh is left to the reader. Moreover, it is known from general
topology that L(Z) is a Hausdorff space and a complete uniform space,
respectively, if and only if each factor is. (L, Z) will be called the product of
the family {LaCZ",): a E A}.
As has been pointed out before, by a subspace M of a vector space Lover
K we understand a subset M i= 0 invariant under addition and scalar
multiplication; we record the following simple consequence of the axioms
(LT)l and (LTh :
2.1
/f(L, Z) is a t.V.S. and M is a subspace of L, the closure M in (L, Z) is again

a subspace of L.
Proof. In fact, it follows from (LT)1 that M + M c M, and from (LTh

thatKMc M.
We recall the following facts from linear algebra. If L is a vector space,
Mi (i = 1, ... , n) subspaces of L whose linear hull is L and such that Mi ( l
(L MJ = {O} for each i, then L is called the algebraic direct sum of the
j '" i
subspaces L j (i = 1, ... , n). It follows that each x E L has a unique represen-
tation x = LiXi, where Xi ELi' and the mapping (Xl' ... , Xn) --> LiXj is an
algebraic isomorphism of TIiLj onto L. The mapping Uj: X --> Xj is called the
projection of L onto L j associated with this decomposition. If each U j is
viewed as an endomorphism of L, one has the relations UjU j = ~ijUj (i,j =
1, ... , n) and LjUj = e, e denoting the identity map.
If (L, Z) is a t. v.s. and L is algebraically decomposed as above, each of the
projections Ui is an open map of L onto the t.v.s. M j • In fact, if G is an open
subset of Land N j denotes the null space of Uj, then G + N j is open in L by
(l.l) and Uj(G) = Uj(G + N j) = (G + N;) ( l M j. From (LT) I it is also clear
that the mapping 1jJ: (Xl' ... , Xn) --> LjX j of TIjM j onto L is continuous; if IjJ
is an isomorphism, L is called the direct sum (or topological direct sum if
this distinction is desirable) of the subspaces Mj(i = 1, ... , n); we write
L = Ml E9 ... E9 Mn·
2.2
Let a t.V.S. L be the algebraic direct sum of n subspaces Mi (i = 1, ... , n).

Then L = M1 $ ... $ Mn if and only if the associated projections Ui are con-
tinuous (i = 1, ... , n).
Proof By definition of the product topology, the mapping ljJ-1: x-+

(u 1x, ... , unx) of L onto TIiMi is continuous if and only if each Ui is.
REMARK. Since the identity map e is continuous on L, the continuity

of n - 1 of these projections implies the continuity of the remaining
one.
A subspace N of a t.v.S. L such that L = M $ N is called a subspace
complementary (or supplementary) to M; such complementary subspaces
do not necessarily exist, even if M is of finite dimension (Exercise 8);
cf. also Chapter IV, Exercise 12.
Let (L, 1:) be a t.v.s. over K, let M be a subspace of L, and let 4> be the
natural (canonical, quotient) map of L onto LIM-that is, the mapping which
orders to each x E L its equivalence class ~ = x + M. The quotient topology
1: is defined to be the finest topology on LIM for which 4> is continuous.
Thus the open sets in LIM are the sets 4>(H) such that H + M is open in L;
since G + M is open in L whenever Gis, 4>(G) is open in LIM for every
open GeL; hence 4> is an open map. It follows that 4>(513) is a O-neighborhood
base in LIM for every O-neighborhood base 513 in L; since 4> is linear, i is
translation-invariant and 4>(513) satisfies conditions (a), (b), and (c) of (1.2) if
these are satisfied by 513. Hence (LIM, 1:) is a t.v.S. over K, called the quotient
space of (L, 1:) over M.
2.3
If L is a t.V.s. and if M is a subspace of L, then LIM is a Hausdorff space

if and only if M is closed in L.
Proof If LIM is Hausdorff, the set {O} c LIM is closed; by the continuity
of 4>, M = 4>-1(0) is closed. Conversely, if ~ '# 0 in LIM, then ~ = 4>(x),
where x ¢; M; if M is closed, the complement U of Min L is a neighborhood of
x; hence 4>( U) is a neighborhood of ~ not containing O. Since 4>( U) contains a
closed neighborhood of ~ by (1.3), LIM is a Hausdorff space.
By (2.3), a Hausdorff t.v.s. LIM can be associated with every t.v.s. L by
taking for M the closure in L of the subspace {O}; M is a subspace by (2.1).
This space LIM is called the Hausdorff t.v.s. associated with L.
There is the following noteworthy relation between quotients and direct
sums:
§3] TOPOLOGICAL VECTOR SPACES OF FINITE DIMENSION 21
2.4
Let L be a t.v.S. and let L be the algebraic direct sum of the subspaces M, N.
Then L is the topolog ical direct sum of M and N: L = MEt> N, if and only if the
mapping v which orders to each equivalence class mod M its unique representa-
tive in N is an isomorphism of the t.V.s. LjM onto the I.v.s. N.
Proof Denote by u the projection of L onto N vanishing on M, and by N. Since N.
3. TOPOLOGICAL VECTOR SPACES OF FINITE DIMENSION
By the dimension of a t.v.S. Lover K, we understand the algebraic dimension

of Lover K, that is, the cardinality of any maximal linearly independent sub-
set of L; such a set is called a basis (or Hamel basis) of L. Let Ko denote the
one-dimensional t.V.S. obtained by considering K as a vector space over itself.
3.1
Everyone-dimensional Hausdorff t.V.S. Lover K is isomorphic with Ko;
more precisely, ,1,--+ AXo is an isomorphism of Ko onto L for each Xo E L,
Xo i= 0, and every isomorphism of Ko onto L is of this form.
Proof It follows from (LTh that ,1,--+ AXo is continuous; moreover, this
mapping is an algebraic isomorphism of Ko onto L. To see that Axo --+ ). is
continuous, it is sufficient to show the continuity of this map at 0 E L. Let
e < I be a positive real number. Since K is non-discrete, there exists ,1,0 E K
such that 0 < 1,1,01 < e, and since L is assumed to be Hausdorff, there exists a
circled O-neighborhood V c: L such that Aoxo ¢ V. Hence AXo E V implies
1,1,1 < e; for 1,1,1 ~ e would imply Aoxo E V, since V is circled, which is contra-
dictory.
Finally, if u is an isomorphism of Ko onto L such that u(l) = xo, then u is
clearly of the form ,1,--+ AXo'
3.2
Theorem. Every Hausdorff t.V.S. L of finite dimension n over a complete
valuated field K is isomorphic with Kg. More precisely, (AI, ... , A.) --+ AIX I +
... + A.x. is an isomorphism of Kg onto Lfor each basis {Xl' ... , X.} of L, and
every isomorphism of Kg onto L is of this form.
Proof. The proof is conducted by induction. (3.1) implies the assertion to
be valid for n = 1. Assume it to be correct for k = n - 1. If {Xl' ... , X.} is
any basis of L, L is the algebraic direct sum of the subspaces M and N with
bases {Xl' ... , x n- 1 } and {xn}, respectively. By assumption, M is isomorphic

with K~-l; since Ko is complete, M is complete and since L is Hausdorff,
M is closed in L. By (2.3), LIM is Hausdorff and clearly of dimension 1;
hence the map v, ordering to each equivalence class mod M its unique
representative in N, is an isomorphism by (3.1). It follows from (2.4) that
L = M ® N, which means that (Al' ... , An) -+ AIXI + ... + AnXn is an iso-
morphism of K~-l x Ko = K8 onto L. Finally, it is obvious that every
isomorphism of K8 onto L is of this form.
It is worth remarking that while (3.1) (and a fortiori (3.2)) obviously fails
for non-Hausdorff spaces L, (3.2) may fail for n > 1 when K is not complete
(Exercise 4).
Theorem (3.2) can be restated by saying that if K is a complete valuated
field, then the product topology on Ko is the only Hausdorff topology satis-
fying (LT)l and (LT)z (Tychonoff [1]). This has a number of important
con seq uences.
3.3
Let L be a t.v.s. over K and let K be complete. If M is a closed subspace of L
and N is a finite dimensional subspace of L, then M + N is closed in L.
Proof Let ¢ denote the natural map of L onto LIM; LIM is Hausdorff by
(2.3). Since ¢(N) is a finite-dimensional subspace of LIM, it is complete by
(3.2), hence closed in LIM. This implies that M + N = ¢ -l(¢(N)) is closed,
since ¢ is continuous.
3.4
Let K be complete, let N be a finite dimensional Hausdorff t.V.S. over K, and
let L be any t.V.S. over K. Every linear map of N into L is continuous.
Proof The result is trivial if N has dimension O. If N has positive dimension
n, it is isomorphic with Ko by (3.2). But every linear map on Ko into L is
necessarily of the form (AI, ... , An) -+ AIYl + ... + AnYn, where Yi E L, and
hence continuous by (LT)l and (LT)z.
We recall that the codimension of a subspace M of a vector space L is the
dimension of LIM; N is an algebraic complementary subspace of M if
L = M + N is an algebraic direct sum.
3.5
Let L be a t.V.S. over the complete field K and let M be a closed subspace of
finite codimension. Then L = M ® N for every algebraic complementary
subspace N of M.
Proof LIM is a finite dimensional t.v.s., which is Hausdorff by (2.3);
hence by (3.4), the mapping v of LIM onto N, which orders to each element
of LIMits unique representative in N, is continuous. By (2.2), this implies
L = M ® N, since the projection u = v ¢ is continuous.
0
§3] TOPOLOGICAL VECTOR SPACES OF FINITE DIMENSION 23
REMARK. It follows from (2.4) that in the circumstances of (3.5), N

is necessarily a Hausdorff subspace of L. It is not difficult to verify this
directly.
We now turn to the second important theorem concerning t.v.S. of finite
dimension. It is clear from (3.2) that if K is locally compact (hence complete),
then every finite dimensional Hausdorff t.v.s. over K is locally compact.
Conversely, if K is complete, then every locally compact Hausdorff t.V.S.
over K is of finite dimension (cf. Exercise 3).
3.6
Theorem. Let K be complete. If L"# {O} is a locally compact Hausdorff
t.v.s. over K, then K is locally compact and Lis offinite dimension.
Proof By (3.1) everyone-dimensional subspace of L is complete, hence

closed in L and therefore locally compact; it follows that K is locally com-
pact. Now let V be a compact, circled O-neighborhood in L, and let {A'n} be a
null sequence in K consisting of non-zero terms. We show first that {An V: n E N}
is a neighborhood base of 0 in L. Given a O-neighborhood U, choose a circled
O-neighborhood W such that W + We U. Since V is compact, there exist
k
elements Xi E V (i = 1, ... , k) satisfying V c U (Xi + W), and there exists
i= 1
A E K, ,1,"# 0, such that Ax i E W for all i, and 1,1,1 ~ 1. There exists n E N for
which IAnl ~ 1,1,1, and
k
An V C AV c U (AXi + A. W) c W +W c U
i=1
shows {An V: n E N} to be a neighborhood base of O.

Let p E K satisfy 0 < Ipi ~ 1/2. Since V is compact and p V is a O-neighbor-
m
hood, there exist elements Yl (I = 1, ... , m) in V for which V c U (YI + P V).
1= 1
We denote by M the smallest subspace of L containing all YI (l = 1, ... , m)
and show that M = L, which will complete the proof. Assuming that M "# L,
there exists WE L ~ M and no EN such that (w + Ano V) n M = 0; for M,
which is finite dimensional and hence complete by (3.2), is closed in L while
{w + An V: n EN} is a neighborhood base of w. Let J1. be any number in K
such that w + J1. V intersects M (such numbers exist since V is radial) and
set b = inflJ1.l. Clearly, b ~ IAnol > O. Choose Vo E Vso thaty = w + J1.oVo EM,
where b ~ IJ1.0 I ~ 3b/2. By the definition of {YI} there exists 10, 1 ~ 10 ~ m,
such that Vo = Ylo + PVI' where VI E V, and therefore
w = Y - J1.oVo = (y - J1.0Yl o) - J1.0PVl EM + J1.opV.

This contradicts the definition of b, since V is circled and since IJ1.oP I ~ 3b/4;
hence the assumption M "# L is absurd.
4. LINEAR MANIFOLDS AND HYPERPLANES

If L is a vector space, a linear manifold (or affine subspace) in L is a subset
which is a translate of a subspace MeL, that is, a set F of the form Xo + M
for some Xo e L. F determines M uniquely while it determines Xo only
mod M: Xo + M = Xl + Nifand only if M = Nand Xl - Xo e M. Two linear
manifolds Xo + M and Xl + N are said to be parallel if either MeN or
N c M. The dimension of a linear manifold is the dimension of the subspace
of which it is a translate. A hyperplane in L is a maximal proper affine sub-
space of L; hence the corresponding subspace of a hyperplane is of codimen-
sion 1. It is further clear that two hyperplanes in L are parallel if and only if the
corresponding subspaces are identical. A hyperplane which is a subspace (i.e.,
a hyperplane containing 0) is sometimes called a. homogeneous hyperplane.
For any vector space Lover K, we denote by L* the algebraic dual of L,
that is, the (right) vector space (over K) of all linear forms on L.
4.1
A subset He L is a hyperplane if and only if H = {x:/(x) = (X} lor some
(X e K and some non-zero Ie L *. I and (X are determined by H to within a
common lactor p, 0 ¥= Pe K.
Proof. If leL* is ¥=O, then M = 1-1(0) is a maximal proper subspace of
L; if, moreover, Xo eL is such that I(xo) = (x, then H = {x:/(x) = (X} =
Xo + M, which shows H to be a hyperplane. Conversely, if H is a hyperplane,
then H = Xo + M, where M is a subspace of L such that dim LjM = 1, so
thatLjM is algebraically isomorphic with Ko. Denote by <p the natural map of
L onto LjM and by g an isomorphism of LjM onto Ko; then 1= g <p is a 0
linear form ¥=O on L such that H = {x:/(x) = (X} when (X =/(xo). If H =

{X:/l(X) = (Xl} is another representation of H, then because of /11(0) = M
we must have /1 = g 1 0 <p, where g 1 is an isomorphism of LjM onto Ko; if
e e)
is the element of Lj M for which g( = 1 and if g 1(e) = p, then/1(X) = l(x)P
for all X e L, thus completing the proof.
Since translations in a t.v.s. L are homeomorphisms, it follows from (2.1)
that the closure of an affine subspace F is an affine subspace F; but F need
not be a proper subset of L if F is.
4.2
A hyperplane H in a t.V.S. L is either closed or dense in L; H = {x:/(x) = (X}
is closed if and only ifI is continuous.
Proof. If a hyperplane H c L is not closed, it must be dense in L; otherwise,
its closure would be a proper affine subspace of L, contradicting the maxi-
mality of H. To prove the second assertion, it is sufficient to show that
/ - 1 (0) is closed if and only if I is continuous. If I is continuous, /-1(0) is
closed, since {O} is closed in K. If /- 1 (0) is closed in L, then LIf- 1 (0) is a
is] BOUNDED SETS 25
Hausdorff t.v.s. by (2.3), of dimension 1; writingf = g 0 4> as in the preceding

proof, (3.1) implies that g, hence/, is continuous.
We point out that, in general, there exist no closed hyperplanes in a t.v.s.
L, even if it is Hausdorff (Exercises 6, 7).
5. BOUNDED SETS
A subset A of a t.v.s. L is called bounded if for each O-neighborhood U
in L, there exists AE K such that A C AU. Since by (1.2) the circled O-neighbor-
hoods in L form a base at 0, A c L is bounded if and only if each 0-
neighborhood absorbs A. A fundamental system (or fundamental family) of
bounded sets of L is a family mof bounded sets such that every bounded sub-
set of L is contained in a suitable member of m.
A subset B of a t.v.s. L is called totally bounded if for each O-neighborhood
U in L there exists a finite subset Bo c B such that B c Bo + U. Recall that a
separated uniform space P is called precompact if the completion P of P is
compact; it follows readily from (1.4) and a well-known characterization of
precompact uniform spaces (see Prerequisites) that a subset B of a Hausdorff
t.v.s. is precompact if and only if it is totally bounded. (We shall use the term
precompact exclusively when dealing with Hausdorff spaces.) From the
preceding we obtain an alternative characterization of precompact sets:
A subset B of a Hausdorff t. v.s. L is precompact if and only if the closure of
B in the completion l of L is compact.
5.1
Let L be a t.v.s. over K and let A, B be bounded (respectively, totally bounded)
subsets of L. Then the following are bounded (respectively, totally bounded) sub-
sets ofL:
(i) Every subset of A.
(ii) The closure A of A.
(iii) A u B, A + B, and AA for each AE K.
Moreover, every totally bounded set is bounded. The circled hull of a bounded
set is bounded; if K is locally precompact, the circled hull of every totally
bounded set in L is totally bounded.
Proof. If A, B are bounded subsets of L, then (i) is trivial and (ii) is clear
from (1.3). To prove (iii), let At and A2 be two elements of K such that
A c Al U and B c A2 U for a given circled O-neighborhood U. Since K is non-
discrete, there exists AO E K such that IAol > SUp(iAll, IA2D. We obtain
Au B c AOU and A + B c Ao(U + U); since by (1.2) U + U runs through a
neighborhood base of 0 when U does, it follows that A u B and A + Bare
bounded; the boundedness of AA is trivial. The proof for totally bounded
sets A, B is similarly straightforward and will be omitted.
Since 0 and everyone-point set are clearly bounded, it follows from a
repeated application of (iii) that every finite set is bounded. If B is totally
bounded and U is a given circled O-neighborhood, there exists a finite set

Bo c B such that Be Bo + U. Now Bo c AoU, where we can assume that
IAol ~ 1, since U is circled; we obtain B c Ao(U + U) and conclude as before
that B is bounded. The fact that the circled hull of a bounded set is bounded
is clear from (1.3). To prove the final assertion, it is evidently sufficient to
show that the circled hull of a finite subset of L is totally bounded, provided
that K is locally precompact. In view of (iii), it is hence sufficient to observe
that each set Sa is totally bounded where a ELand S = {A: IAI ~ 1}; but
this is clear from (LTh and the assumed precompactness of S (cf. (5.4)
below). This completes the proof.
COROLLARY 1. The properties of being bounded and of being totally bounded
are preserved under the formation of finite sums and unions and under dila-
tations x -+ AoX + Xo.
COROLLARY 2. The range of every Cauchy sequence is bounded.
COROLLARY 3. The family of all closed and circled bounded subsets of a t.V.s.
Lis afundamental system of bounded sets of L.
It is clear from the definition of precompactness that a subset of a Haus-
dorff t.v.s. is compact if and only if it is precompact and complete. We record
the following simple facts on compact sets.
5.2
Let L be a Hausdorff t.v.s. over K and let A, B be compact subsets of L.
Then A u B, A + B, and AA (A E K) are compact; if K is locally compact, then
also the circled hull of A is compact.
Proof The compactness of A u B is immediate from the defining property
of compact spaces (each open cover has a finite subcover; cf. Prerequisites);
A + B is compact as the image of the compact space A x B under (x, y) -+
x + y which is continuous by (LT)l; the same argument applies to AA by
(LTh. (Another proof consists in observing that A u B, A + B, and AA are
precompact and complete.) Finally, the circled hull of A is the continuous
image of S x A (under (A, x) -+ AX), and hence compact if S is compact.
COROLLARY. Compactness of subsets of a Hausdorff t.V.S. is preserved under
the formation offinite sums and unions and under dilatations.
The following is a sequential criterion for the boundedness of a subset of a
t.v.s. (for a sequential criterion of total boundedness, see Exercise 5). By a
null sequence in a t.v.S. L, we understand a sequence converging to 0 E L.
5.3
A subset A of a t.v.s. L is bounded if and only if for every null sequence {An}
in K and every sequence {xn} in A, {A.X.} is a null sequence in L.
§5] BOUNDED SETS 27
Proof. Let A be bounded and let V be a given circled O-neighborhood in L.

There exists 11 E K, 11 -:f: 0 such that I1A c V. If {An} is any null sequence in K,
there exists no E N such that IAnl ~ 1111 whenever n ~ no; hence we obtain
Anxn E V for all n ~ no and any sequence {xn} in A. Conversely, suppose that
A is a subset of L satisfying the condition; if A were not bounded, there would
exist a O-neighborhood U such that A is not contained in PnU for any se-
quence {Pn} in K. Since K is non-discrete, we can choose Pn so that IPnl ~ n
for all n E N, and Xn E A '" PnU (n E N); it would follow that P; 1 Xn ¢ U for
all n, which is contradictory, since {p;l} is a null sequence in K.
5.4
Let L, M be t.V.S. over K and let u be a continuous linear map of L into M.
If B is a bounded (respectively, totally bounded) subset of L, u(B) is bounded
(respectively, totally bounded) in M.
Proof. If V is any O-neighborhood in M, then u-I(V) is a O-neighborhood
in L; hence if B is bounded, then Be AU-I(V) for a suitable A E K, which
implies u(B) C AV. If B is totally bounded, then Be Bo + u-I(V) for some
finite set Bo c B, whence u(B) c u(Bo) + V.
The preceding result will enable us to determine the bounded sets in a
product space f1~L~. We omit the corresponding result for totally bounded sets.
5.5
If{L~: IX E A} is afamily oft.v.s. and if L = f1~L~, a subset B of L is bounded
if and only if B c f1~B~, where each B~ (IX E A) is bounded in L~.
Proof. It is easy to verify from the definition of the product topology that
if B~ is bounded in L~ (IX E A), then f1~B~ is bounded in L; on the other
hand, if B is bounded in L, then uaCB) is bounded in L~, since the projection
map u~ of L onto L~ is continuous (IX E A), and, clearly, B c f1~u~ (B).
Thus a fundamental system of bounded sets in f1~L~ is obtained by forming
all products f1~B~, where B~ is any member of a fundamental system of
bounded sets in L~(IX E A). Further, if L is a t.v.s. and M a subspace of L, a set
is bounded in M if and only if it is bounded as a subset of L; on the other
hand, a bounded subset of L/ M is not necessarily the canonical image of a
bounded set in L (Chapter IV, Exercises 9, 20).
At. v.S. L is quasi-complete if every bounded, closed subset of L is complete;
this notion is of considerable importance for non-metrizable t.v.S. By (5.1),
Corollary 2, every quasi-complete t.v.S. is semi-complete; many results on
quasi-complete t.v.s. are valid in the presence of semi-completeness, although
there are some noteworthy exceptions (Chapter IV, Exercise 21). Note also
that in a quasi-complete Hausdorff t.v.s., every pre compact subset is rela-
tively compact.
5.6
The product of any number of quasi-complete t.V.S. is quasi-complete.
The proof is immediate from the fact that the product of any number of
complete uniform spaces is complete. and from (5.5).
6. METRIZABILITY
A t.V.S. (L, :t) is metrizable if its topology :t is metrizable, that is, if there
exists a metric on L whose open balls form a base for :t. We point out that
the uniformity generated by such a metric need not be translation-invariant
and can hence be distinct from the uniformity associated with :t by (1.4)
(Exercise 13). However, as we have agreed earlier, any uniformity notions to
be employed in connection with any t.v.S. (metrizable or not) refer to the
uniformity 9l of (1.4).
It is known from the theory of uniform spaces that a separated uniform
space is metrizable if and only if its vicinity filter has a countable base. For
topological vector spaces, the following more detailed result is available.
6.1
Theorem. A Hausdorff t.v.S. L is metrizable if and only if it possesses a
countable neighborhood base of O. In this case, there exists a function x -+ Ixl
on L into R such that:
(i) 1,1,1 ~ 1 implies IAxI ~ Ixl for all x eL.
(ii) Ix + yl ~ Ixl + Iyl for all x eL, y eL.
(iii) Ixl= 0 is equivalent with x = O.
(iv) The metric (x, y) -+ Ix - yl generates the topology of L.
We note that (i) implies Ixl = I-xl and that (i) and (iii) imply Ixl ~ 0 for
all x e L. Moreover, since the metric (x, y) -+ Ix - yl is translation-invariant,
it generates also the uniformity of the t.V.S. L.
A real function x -+ lxi, defined on a vector space Lover K and satisfying
(i) through (iii) above, is called a pseudo-norm on L. It is clear that a given
pseudo-norm on L defines, via the metric (x, y) -+ Ix - yl. a topology:t onL
satisfying (LT)l; on the other hand, (LT)2 is not necessarily satisfied (Exercise
12). However, if x -+ Ixl is a pseudo-norm on L such that A.n -+ 0 implies
IA.nxl-+ 0 for each x eLand Ixnl-+ 0 implies IAxnl-+ 0 for each A. e K, then
it follows from (i) and the identity
A.X - A.oXo = A.o(x - xo) + (A. - A.o)xo + (A. - A.o)(x ..... xo)
that the topology l: defined by x -+ Ixl satisfies (LTh, and hence that (L, l:)
is a t.v.s. over K.
Proof of (6.1). Let {Vn: n e N} be a base of circled O-neighborhoods
satisfying
(n eN). (1)
For each non-empty finite subset H of N, define the circled O-neighborhood
VH by VH = LneH Vn and the real number PH by PH = LneHT". It follows
§6] METRIZABILITY 29
from (1) by induction on the number of elements of H that these implications

hold:
(2)
where n < H means that n < k for all k E H. We define the real-valued
function x ---+ Ixl on L by Ixl = I if x is not contained in any V H , and by
Ixl = inf {PH: x E VH}
H
otherwise; the range of this function is contained in the real unit interval.
Since each VH is circled, (i) is satisfied. Let us show next that the triangle
inequality (ii) is valid. This is evident for each pair (x, y) such that Ixl + Iyl ~ J.
°
Hence suppose that Ixl + IYI < 1. Let 8> be any real number such that
Ixl + Iyl + 28 < I; there exist non-empty finite subsets H, K of N such that
x E VH, Y E VK and PH < Ixl + 8, PK < Iyl + 8. Since PH + PK < 1, there
exists a unique finite subset M of N for which PM = PH + PK; by virtue of (1),
M has the property that VH + VK C VM • It follows that x + y E V M and hence
that
Ix + yl ;::;; PM = PH + PK < Ixl + Iyl + 28,
which proves (ii).
For any 8 > 0, let S. = {x EL: Ixl;::;; 8}; we assert that
(n EN). (3)
The inclusion Vn C S2-n is obvious since x E Vn implies Ixl ;::;; 2- n. On the
other hand, if Ixl ;: ; rn-l, then there exists H such that x E VH and PH < rn;
hence (2) implies that x E Vn •
It is clear from (3) that (iii) holds, since L is a Hausdorff space and hence
family {S.: 8 > o} is a neighborhood base of °

x =0 is equivalent with x E n{Vn : n EN}. Moreover, (3) shows that the
in L; since the topology
generated by the metric (x, y) ---+ Ix - yl is translation-invariant, (iv) also
holds. This completes the proof.
REMARK. It is clear from the preceding proof that on every non-
Hausdorfft.v.s. Lover K possessing a countable neighborhood b,ase of
0, there exists a real-valued function having properties (i), (ii) and (iv) of
(6.1).
If L is a metrizable t.V.S. over K and if x ---+ Ixl is a pseudo-norm generating

the topology of L, this pseudo-norm is clearly uniformly continuous; hence
it has a unique continuous extension, x ---+ lxi, to the completion L of L. We
conclude from (1.5) that this extension, which is obviously a pseudo-norm
on L, generates the topology of L.
Example. Denote by I the real unit interval and by /1 Lebesgue

measure on I. Further let !e P (p > 0) be the vector space over R of all
real-valued, /1-measurable functions for which I/I P (where III denotes the
function t -+ I f(t) Dis J.t-integrable, and let LP be the quotient space of

,!l'P over the subspace of J.t-null functions. Ifp ~ 1,
f-+ f)fl P dJ.t

is a pseudo-norm on LP, and it is easy to verify that LP is complete under
the corresponding topology. If p < 1, £P is an example of a Hausdorff
t.v.s. on which there exists no continuous linear form other than 0
(Exercise 6).
A t.v.S. L is said to be locally bounded if L possesses a bounded neighbor-
hood of 0; clearly, such a space has a neighborhood base of 0 consisting of
bounded sets. The spaces LP of the preceding paragraph are locally bounded.
We shall encounter further examples in Chapter II, Section 2.
6.2
Every locally bounded Hausdorff t.V.S. ;s metrizable.
Proof Let V be a bounded O-neighborhood in L and let P.n} be a sequence
of non-zero elements of K such that lim An = O. If U is any circled neighbor-
hood of 0, there exists A E K such that V C AU, since V is bounded; if n is
such that lAnAI ~ 1, then AnV c U, since U is circled. It follows that {AnV:
n E N} is a O-neighborhood base, whence Lis metrizable by (6.1).
A quasi-complete, locally bounded t.V.S is complete, since it possesses a
complete neighborhood of O. We observe that the converse of (6.2) is false;
an example is furnished by the product of a countably infinite number of
one-dimensional t.V.s. which is metrizable (see below), but not locally
bounded by (5.5).
Clearly, every subspace M of a metrizable t.v.s. is metrizable; if x -+ Ixl is a
pseudo-norm on L generating its topology, the restriction of x -+ Ixl to M
generates the topology of M. Let L = TInLn be the product of countably
many metrizable t.v.s. Since the product topology is metrizable, (6.1) implies
that it can be generated by a pseudo-norm. Such a pseudo-norm can be
constructed explicitly if, on each factor Ln(n EN), a generating pseudo-norm
x -+ Ixl n is given: Writing x = (xn),
x -+ Ixl = ~..!:. Ixnl n

n=1 2n1 + IXnln
is a generating pseudo-norm on L. It is not difficult to verify conditions
(i)-(iv) of (6.1); for (i) and (ii), recall that u -+ uf(l + u) is monotone for
u ~ 0 and that
a+b a b
----<--+--
l+a+b=l+a l+b
for any two real numbers a, b ~ O. We leave it to the reader to verify that
x -+ Ixl generates the product topology on L.
§7] COMPLEXIFICATION 31
For a quotient space LIM of a metrizable t.V.S. to be metrizable, M must

necessarily be closed by (2.3); this condition is also sufficient. In terms of a
generating pseudo-norm on L, we prove the following more detailed result.
6.3
The quotient space of a metrizable t.v.s. L over a closed subspace M is
metrizable, and if L is complete then LIM is complete. If x --t Ixl is a pseudo-
norm generating the topology of L, then (with ~ = x + M)
~ --t I~I = inf{lxl: x E ~}
is a pseudo-norm generating the topology of LIM.
Proof We note first that ~ --t I~I satisfies (i)-(iii) of (6.1). Clearly, 101 = 0;
if Ixl = 0, then 0 E~, since M is closed. For (ii), let e > 0 be given; then
Ixl < I~I + e, Iyl < I.PI + e for suitable x E ~,y E P; now,
I~ + PI ~ Ix + yl ~ Ixl + Iyj
+ IPI + 2e.~ I~I
(i) follows from the corresponding property of x --t Ixl on L, since the quotient
map x --t X is linear.
Let Vn = {x: Ixl < n-1}(n EN). {Vn} is a O-neighborhood base in L; hence
{¢( Vn )} is a O-neighborhood base in LI M, since the natural map x --t X = ¢(x)
is both open and continuous. We set Vn = {x: I~I < n- 1 } and claim that
Vn = ¢(Vn) for n E N. Clearly, ¢(Vn) c: Vn. Conversely, if x E Vm there exists
x E x such that x E Vn; hence ¢ -l( Vn) c: Vn + M, which implies Vn c: ¢( Vn).
Thus X --t I~I generates the topology of LIM.
There remains to show that LIM is complete when L is complete. Given a
Cauchy sequence in LjM, there exists a subsequence {~n} such that
IXn+l - xnl < 2 -n-l(n EN). Hence there exist representativesYn+ 1 E Xn+1 - ~n
such that IYn+ll <2- n• Let X1EX 1 be arbitrarily chosen; then Xn=Xl +
n
L Yv E xn for all n ~ 2. Using condition (ii) of (6.1), it is readily verified
v=2
that {xn } is a Cauchy sequence in L, hence convergent to some x E L. Since
¢ is continuous, {xn} converges in LIM; thus the given Cauchy sequence
converges, which shows LjM to be complete.
We point out that if L is a non-metrizable, complete t.v.s. and if M is a
closed subspace of L, the quotient space LIM is, in general, not complete
(cf. Chapter IV, Exercise 11).
7. COMPLEXIFICATION
In this section we consider vector spaces over a more restricted type of
fields K than were admissible so far: We assume that either K is a subfield of
R, or else that K is a subfield of C containing the imaginary unit i and in-
variant under conjugation; in both cases, K is understood to carry the in-
duced absolute value.
Any such field can be written as K = H + iH if it contains the imaginary

unit i; H = K n R is a subfield of R. If H is, on the other hand, a subfield of
R, let us denote by H(i) = H + iH the complex extension of H. If L is a
vector space over H, can scalar multiplication in L be extended to K = H(i)?
If it can, then L possesses an automorphism u such that u 2 = -e (e the
identity mapping); namely, x ~ ix is such an automorphism. Conversely,
if u is an automorphism of L (over H) satisfying u 2 = -e, then the definition
(A, f.1 E H)
(A + if.1)x = Ax + f.1u(x) (1)
extends scalar multiplication to K = H + iH, which can be quickly verified.

Similarly, if L is a t.v.S. over H and if u is a (topological) automorphism of
L such that u2 = -e, then (1) makes L into a t.v.s.over K.
7.1
If L is a t.v.s. o.ver HeR, scalar multiplication in L has a continuous exten-

sion to H(i) x L into L if and only if L permits an automorphism u satisfying
u2 = -e.
Conversely, if L is a vector space (or t.v.s.) over a field K = H(i) containing
i, then the restriction of scalar mUltiplication to H x L turns L into a vector
space (or t.v.s.) Lo over H. Lo will be called the real underlying space of (or
associated with) L. A real linear form on L is a linear form on L o, and a real
hyperplane in L is a hyperplane in Lo. Accordingly, a real subspace (real
affine subspace) of L is a subspace (affine subspace) of L o.
Let L be a vector space over K = H + iH and let f E L * be a linear form on
L. Then f= g +ih, where g, h are uniquely determined real-valued (more
precisely, H-valued) functions on L; obviously g and hare real linear forms on
L, called the real and imaginary parts of/, respectively. Since g(ix) + ih(ix) =
f(ix) = if(x) = ig(x) - h(x) for all x E L, we have
f(x) = g(x) - ig(ix) (x E L). (2)
Conversely, if g is any real linear form onL, then/, defined by (2), is a member
of L* (verification is left to the reader), and obviously the only one with
real partg. Moreover, if L is a t.v.s. over K, then (2) shows thatfis continuous
if and only if g is continuous. We have proved:
7.2
Let L be a t.v.S. over K and let Lo be its real underlying space. The mapping
f ~ g defined by (2) is an isomorphism of(L*)0 onto (Lo)*, carrying the space of
continuous linear forms on L onto the space of continuous linear forms on Lo.
For hyperplanes in L, we have the following result:
EXERCISES 33 .
7.3
Let L be a t.V.S. over K. Every (closed) hyperplane in L is the intersection of
two uniquely determined, (closed) real hyperplanes.
Proof. By (4.1), a hyperplane G in L is of the form G = {x:f(x) = y},
where f E L * and y = ex + iP E K = H + iH. If 9 is the real part of j, then,
clearly, G = G1 n G2 , where G1 = {x: g(x) = ex}, G2 = {x: g(ix) = - Pl. Since
f is determined by G to within a non-zero factor, G1 and G2 are unique.
Moreover, by (4.2) and (7.2), G1 and G2 are closed if and only if G is closed
inL.
If L is a vector space over a field H c: R, there does not always exist an
automorphism u of L satisfying u2 = -e; examples are furnished by rear
vector spaces of finite odd dimension. It is still often desirable, especially for
the purposes of spectral theory, to imbed L isomorphically into a vector
space over K = H(i); the following procedure will provide such an imbedding.
Consider the product L x Lover H. The mapping u: (x, y) -+ ( - y, xl is an
automorphism (which is topological if L is a Lv.s. over H) of L x L satisfying
u 2 = - e; thus scalar multiplication can be extended to K x L x L into
L x L by 0). Thus iCY, 0) = (0, y), and if we agree to write (x, 0) = x for all
x E L, then each Z E L x L has a unique representation Z = x + iy with
x E L, y E L. If L is a t.v.s. over H, then L x Lover K is a t.V.S. such that
(L x L)o = L $ iL. This type of imbedding is called the complexification of a
vector space (or t.v.s.) defined over a subfield of R.
It can be shown (Exercise 16) that every vector space over a conjugation
invariant field K c: C such that K contains i, is algebraically isomorphic to the
complexification of anyone of its maximal properly real subspaces.
EXERCISES
1. Let {LIZ: ex E A} be a family of Hausdorff t.v.s. over K and denote
by 'lJ the family of subsets of the vector space L = n~1Z obtained by
forming all products V = nlZVIZ , where VIZ(ex E A) is any member of a
O-neighborhood base in LIZ' Let X denote the unique translation-
invariant topology on L for which 'lJ is a neighborhood base of O. Let M
be the subspace of L containing exactly those elements x E L which
have only a finite number of non-zero coordinates (M is denoted by
$IZLIZ and called the algebraic direct sum of the family {LIZ})'
(a) If an infinite number of the spaces LIZ are not reduced to {O},
(L, X) is not a t.v.s.
(b) (M, 'X) is a Hausdorff t.v.s. which is complete if and only if each
L" is complete.
(c) A subset of M is bounded in (M, X) if and only if it is contained
in a set of the form n"eHBIZ x {O}, where H c: A is finite and BIZ is
bounded in L" for or; E H.
2. Let L be a t.v.S. which is not a Hausdorff'space, and denote by N
the closure of {O}.
(a) The topology of the subspace N is the trivial topology whose only
members are Nand 0. If M is any algebraic complementary subspace
of N in L, then L = M Ef) Nand M is isomorphic with the Hausdorff
t.V.S. associated with L (use (2.4).)
(b) Deduce from this that every t.V.S. L is isomorphic with a dense
subspace of a complete t.v.s. over the same field.
(c) Show that a subset A of L is totally bounded if and only if the can-
onical image of A in L/N is precompact.
3. Give an example of a finite-dimensional t.v.s. L over a (non-
complete) field K such that the completion of L is infinite-dimensional
over K, and locally compact.
4. Let Q be the rational number field under its usual absolute value
and let L = Q + Q -/i. Show that L, under the topology induced by R,
is a t.v.s. over Q not isomorphic with Qo x Qo.
5. Let B be a subset ofa t.v.S. such that every sequence in Bhasa clus~
ter point; then B is totally bounded. (For a given circled O-neighborhood
V, let Bo be a subset of B such that x E B o, y E B o, and x i= y imply
x - y ¢ V, and which is maximal with respect to this property (Zorn's
lemma); then B c Bo + V. Show that the assumption" Bo is infinite"
is absurd.)
6. Let £P(O < p < I) be the vector space over R introduced in
Section 6; under the topology generated by the pseudo-norml -+ 1/11 =
JIIIP dJl. Show that £P is a complete t.V.S. on which there exist no non-
zero continuous linear forms. (If u i= 0 is a continuous linear form, then
Iu(f) I = 1 for some IE £P. Denote by z. (0 ;;; s ;;; I) the characteristic
function of [0, s] c [0, 1]; there exists t such that liz til = 1/(1 - Zt) 11
= tl/ll' For at least one of the functions IZt and 1(1 - Zt), call it til,
one has lu(t/l)1 ~ t. Moreover, 1/111 = 2p- l l/lt. By induction, define a
sequence {f,.} such that lu([")1 ~ 1 and Ifnll = 2n(p-l) 1/11') (M. Day [1],
W. Robertson [1]).
7. Let L be a t.v.S. Show these assertions to be equivalent:
(a) Every subspace of finite codimension is dense in L.
(b) There exist no closed hyperplanes in L.
(c) No finite-dimensional subspace has a complementary subspace in L.
8. Construct a decomposition L = M + N of a t.v.s. L such that
M + N is an algebraic, but not a topological, direct sum (use Exercise 4
or Exercises 6, 7).
9. The dimension of a complete metrizable t.v.s. over a complete
field K is either finite or uncountably infinite (use Baire's theorem).
10. Let {L.. : ex E A} be a family of metrizable t.v.s. The product
ILL.. is metrizable only if A is countable, and the direct sum EB ..L.. (the
space (M, l:) of Exercise l(b» is metrizable only if A is finite.
11. Deduce from Exercise 10 an example of a complete Hausdorff
t.v.s. of countable dimension which is not metrizable.
12. (a) Let L be a vector space over K and let d be a translation-
invariant metric on L such that metric space (L, d) is complete. Suppose,
in addition, that An -+ 0 implies d(Anx, 0) -+ 0 for each x ELand that
d(xll' 0) -+ 0 implies d(Axn' 0) -+ 0 for each A E K. Show that under the
EXERCISES 35
topology generated by d, L is a complete t.v.s. (Use Baire's theorem to

show that (LT)2 holds.)
(b) Let L be the vector space over R of all real-valued continuous
functions on R. Show that 1 - sup I/(t)I/O + I/(t)1) is a pseudo-norm
teR
on L, and that under the topology generated by this pseudo-norm, L is
a complete topological group with respect to addition, but not a t.v.s.
13. The metric d(x, y) = tan-1lx - yl generates the unique topology
on R under which Ro is a Hausdorfft.v.s., but the uniformity generated
by d (under which R is precompact) is distinct from the uniformity of
the t.v.s. Ro.
14. Let d be a metric on a vector space L such that under the topology
~ generated by d, L is a t.v.s., and such that the metric space (L, d) is
complete. Then the t.v.s. (L, ~) is complete. (V. L. Klee [1].)
15. Show that on the vector spaces R2n+l(n EN) there exists no auto-
morphism u satisfying u2 = -e.
16. Let L be a vector space over a subfield K = H + iH of C, where
H is a subfield of R. Calla real subspace N of L properly real if N n iN =
{O}. There exists a properly real subspace M of L such thatL = M + iM.
(Use Zorn's lemma.) Cf. Chapter IV, Exercise 3.
17. Every t.v.s. (Hausdorff or not) over R or C is connected and
locally connected .
.18. Find a formula relating the cardinality of a vector space Lover K
with its dimension. Prove that if dim L ~ card K, then dim L * =
(card K)dimL.
PREREQUISITES

ences can be found.
A. SETS AND ORDER

1

§1] CONVEX SETS AND SEMI-NORMS 37
familiar with their elementary theory, which is now part of every first course
in abstract analysis; we confine ourselves to some basic results and a review
of the most frequent examples of normed spaces, including some facts on
Hilbert space that are recorded for later use. Section 3 proves the Hahn-
Banach theorem in its two forms called by Bourbaki [7] the geometrical
and analytical forms, respectively. This is the central result of the chapter and
fundamental for most of what follows later; it lends power to the notion of
locally convex space (due to J. von Neumann [1]), defined in Section 4.
Continuous semi-norms constitute an analytical alternative for the use of
convex circled O-neighborhoods which is illustrated by the two forms of the
Hahn-Banach theorem; but while applications often suggest the use of semi-
norms, we feel that their exclusive or even preferred use does not support the
geometrical clarity of the subject.
The separation properties of convex sets, all consequences of the geometri-
cal form of the Hahn-Banach theorem, could logically follow Section 4; we
have preferred to place them at the end of the chapter so that the reader
would first have a survey of the class of spaces in which those separation
results are valid. Following a method extremely useful even in general
topology (cf. Prerequisites), we hope to give the reader an efficient way to
organize the various means of generating new locally convex spaces from
those of a given family, by simply distinguishing between projective and
inductive topologies. With the exception of spaces of linear mappings and
topological tensor products (Chapter III), Sections 5 and 6 of the present
chapter give all standard methods for constructing locally convex spaces.
It is interesting to observe «5.4), Corollary 2) that every locally convex space
can be obtained as a subspace of a suitable product of Banach spaces. Two
classes of spaces particularly frequent in applications are discussed in
Sections 7 and 8. Section 9 furnishes the standard separation theorems
which are constantly used later. The chapter closes with a rather compressed
approach, following Bourbaki [7], to the Krein-Milman theorem. This is a
beautiful and important theorem of which everyone interested in topological
vector spaces should be aware; however, it has little bearing on the theory
to be presented here, and we refer to Klee [3]-[5] for a deep analysis and the
many ramifications of this result.
1. CONVEX SETS AND SEMI-NORMS
A subset A of a vector space L is convex if x E A, YEA imply that AX +

(1 - A)Y E A for all scalars A satisfying 0 < A < 1. The sets {Ax + (1 - A)Y:
o ~ A ~ I} and {Ax + (1 - A)Y: 0 < A< I} are called the closed and open
line segments, respectively, joining x and y. It is immediate that convexity
of a subset A c: L is preserved under translations: A is convex if (and only
if) Xo + A is convex for every Xo E L.
38 LOCALLY CONVEX TOPOLOGICAL VECTOR SPACES [Ch. II
1.1
Let A be a convex subset of a t.v.s. L. If x is interior to A and y in the closure
of A, the open line segment joining x and y is interior to A.
Proof Let A, 0 < A < 1, be fixed; we have to show that AX + (1 - A)y E A.
By a translation if necessary we can arrange that AX + (1 - A)y = O. Now
y = cxx where cx < o. Since w --+ cxw is a homeomorphism of L by (I, 1.1)*
and x E A, YEA, there exists a Z E A such that cxz E A. Let fl = cx/(cx - 1);
then 0 < fl < 1 and flZ + (1 - fl)CXZ = O. Hence
U = {.uw + (1 - .u)cxz: WE A}
is a neighborhood of 0 since w --+ .uw + (1 - .u)cxz is a homeomorphism of L
mapping Z E A onto O. But WE A and cxz E A imply U c A, since A is convex;
hence 0 E A.
1.2
Let L be a t.v.s. and let A and B be convex subsets of L. Then A, A, A +B
and cxA(cx E K) are convex.
The convexity of A is immediate from (1.1); if A is fixed, 0 < A < 1, then
AA + (1 - A)A c A whence AA + (1 - A)A c A by (LT)l and (LT)z (Chapter
I, Section 1); thus A is convex. The proof that A + Band cxA are convex is
left to the reader.
1.3
If A is convex with non-empty interior, then the closure A of A equals the
closure of A, and the interior A of A equals the interior of A.
Proof Since A c A, (1) c A holds trivially. If A is convex and A non-
empty, (1.1) shows that A c (1). To prove the second assertion, it suffices to
show that 0 E (1) implies 0 E A if A is convex with non-empty interior.
There exists a circled neighborhood V of 0 such that V c A. Since A = (1),
o is in the closure of A; hence A and V intersect. Let YEA n V. Since V c A
and V is circled, we have - YEA and it follows now from (1.1) that 0 E A,
since 0 = !Y + t( -y).
A cone C of vertex 0 is a subset of a vector space L invariant under all
homothetic maps x --+ AX of strictly positive ratio A; if, in addition, C is
convex, then C is called a convex cone of vertex O. Thus a convex cone of
vertex 0 is a subset of L such that C + C c C and AC c C for all A > O. A
(convex) cone of vertex Xo is a set Xo + C, where C is a (convex) cone of
vertex o. It is a simple exercise to show that the interior and the closure of a
(convex) cone of vertex 0 in a t.v.s. L are (convex) cones of vertex 0 in L.
* Roman numerai refers to chapter number.

§1] CONVEX SETS AND SEMI-NORMS 39
For subsets of a vector space L, the properties of being circled or con-

vex are invariant under the formation of arbitrary intersections. Since L has
both properties, every subset A c: L determines a unique smallest subset
containing A and having anyone or both of these properties, respectively:
the circled hull, the convex hull, and the convex, circled hull, of A. The
circled hull of A is the set {Aa: a E A and IAI ~ I}; if A "1= 0, the convex hull
of A is the set {2)vav}, where Av > 0, 2)v = 1 and {a v } ranges over all non-
empty finite subsets of A (Exercise I). The convex, circled hull of A, denoted
r
by A, is the convex hull of the circled hull of A (Exercise 1). By rIXAIX' we
denote the convex, circled hull of the union of a family {A IX: IX E A}.
If L is a t.v.s., the properties of being circled or convex (or both) can be
combined with the property of being closed; obviously, the resulting notions
are again intersection-invariant. In particular, the closed, convex hull (some-
times referred to as the convex closure) of A c: L is the closure of the convex
hull of A, by (I .2); similarly, the closed, convex, circled hull of A is the closure
of the convex, circled hull of A (Exercise 1).
We turn to the investigation of convex, radial subsets of a vector space L;
certainly the convex hull of a radial set is of this type. If M is any radial
subset of L, the non-negative real function on L:
x --> PM(X) = inf{A > 0: x E AM},
is called the gauge, or Minkowski functional, of M. Obviously, if M is a radial
set in Land M c: N, then PN(X) ~ PM(X) for all x E L, that is PN ~ PM'
We define a semi-norm on L to be the gauge of a radial, circled and convex
subset of L; a norm is a semi-norm P such that p(x) = 0 implies x = O. The
following analytical description of semi-norms is often used as a definition.
1.4
A real-valued/unction P on a vector space L is a semi-norm if and only if
(a) p(x + y) ~ p(x) + p(y) (x, y E L)
(b) p(AX) = IAlp(x) (x E L, A E K).
Proof Let p be a semi-norm on L, that is, let p be PM' where M is radial,
circled, and convex. If x E L, y E L are given and Al > p(x), A2 > p(y), then
x + y E AIM + A2M. Since M is convex,
Al M + A2M = (AI + A2) [A

+ z
/\'1
M +A
A + /\'1 A2
M] c: (AI + A2)M ;
this implies p(x + y) ~ Al + A2 ; hence p(x + y) ~ p(x) + p(y). For (b),
observe that h E 11M is equivalent with IAlx E 11M, since M is circled; hence
if A "1= 0,
p(h) = inf{11 > 0: x E IAI-lI1M} = inf {IAII1: x E 11M} = IAlp(x);
,,>0
this proves (b), since p(O) = O.
Conversely, assume that p is a function satisfying (a) and (b), and let
M = {x: p(x) < I}. Clearly, M is radial and circled, and it follows from (a)
and (b) that M is convex. We show that p = PM. It follows from (b) that
{x: p(x) < A} = AM for every A> 0; hence if p(x) = IX, then x E AM for all
A> IX but for no A < IX, which proves thatp(x) = inf {A> 0: XE AM} = PM(X).
Simple examples show that the gauge function p of a radial set MeL
does not determine M; however, if M is convex and circled, we have the
following result, whose proof is similar to that of (1.4) and will be omitted.
1.5
Let M be a radial, convex, circled subset of L; for the semi-norm p on L to be
the gauge of M, it is necessary and sufficient that Mo c Me M 1 , where
Mo = {x:p(x) < I} and Ml = {x:p(x);;::; I}.
If L is a topological vector space, the continuity of a semi-norm p on L is
governed by the following relationship.
1.6
Let p be a semi-norm on the t.v.s. L. These properties ofp are equivalent:
(a) p is continuous at 0 E L.
(b) Mo = {x: p(x) < I} is open in L.
(c) p is uniformly continuous on L.
Proof. (a)=>(c), since by (1.4), Ip(x)-p(y)1 ;;::;p(x-y) for all X,YEL.
(c) => (b), since Mo = p-l[( - 00, 1)]. (b) => (a), since eMo = {x: p(x) < e} for
all e > O.
A subset of a t.v.s. L that is closed, convex, and has non-empty interior is
called a convex body in L. Thus if p is a continuous semi-norm on L, Ml =
{x: p(x) ;;::; I} is a convex body in L.
2. NORM ED AND NORMABLE SPACES

By the definition given in Section I, a norm p on a vector space L (over R
or C) is the gauge of a convex, circled, radial set which contains no subspace
of L other than {O}; frequently a norm is denoted by II II. We recall from
(1.4) that a norm II lion L is characterized by these analytical properties:
(i) IIAxII = IAI Ilxll for all A E K, x E L.
(ii) Ilx+yll;;::; Ilxll+ IlyllforallxEL,YEL.
(iii) Ilx II = 0 implies x = o.
We define a normed space to be a pair (L, II II), with the understanding
that L carries the topology generated by the metric (x, y) ...... Ilx - YII. Under
this topology, L is a t.v.s. (This is clear from the discussion following (1,6.1),
since a norm is also a special case of a pseudonorm (not only of a semi-
§2l NORMED AND NORMABLE SPACES 41
norm); in view of this, it is possible to define normed spaces over arbitrary,

non-discrete valuated fields (Exercise 5), but we do not follow this usage.) By
contrast, a normablespace is a t.v.s. L whose topology can be obtained from
a norm II II on L via the metric (x, y) -+llx - y II; such a norm is, of course,
not unique (Exercise 5). A complete normed space is called a Banach space,
or briefly (B)-space. A norm preserving isomorphism of one normed space
onto another is called a norm isomorphism, and two normed spaces are called
norm isomorphic if there exists a norm isomorphism between them. The set
{x: IIx II ~ I} is the (closed) unit ball of (L, II ID.
There is a simple necessary and sufficient condition for a (necessarily
Hausdorff) t.v.s. to be normable; the result is due to Kolmogoroff [1].
2.1
A Hausdorff t.v.s. L is normable if and only if L possesses a bounded, convex

neighborhood ofO.
j
Proof. The condition is necessary, for if x -+ IIx II generates the topology

of L, VI = {x: IIxll ~ I} is a convex neighborhood of 0 which is bounded,
since, by (i) above, the multiples {n- I VtJ(n E N) form a O-neighborhood base
in L. Conversely, if V is a convex, bounded O-neighborhood in L, there exists
a circled neighborhood contained in V whose convex hull U is bounded (since
it is contained in V). Clearly, the gauge p of U is a norm on L. Now the
boundedness of U implies that {n-IU}(n E N) is a O-neighborhood base,
whence it follows that p generates the topology of L.
The completion L of a normable space Lis normable, for if V is a bounded,
convex O-neighborhood in L, its closure V in L is bounded by (I, 1.5) and
convex by (1.2). If (L, p) is a normed space, then p, which is uniformly con-
tinuous on L by (1.6), has a unique continuous extension jj to L that generates
the topology of L; (L, jj) is a Banach space. It is obvious that a subspace of a
normable space is normable and that a closed subspace of a Banach space is a
Banach space. However,
2.2
The product of a family of normable spaces is normable if and only if the

number offactors :#= {OJ is finite.
Proof. This follows quickly from (2.1), since, by (I, 5.5), a O-neighborhood
in the product IT"L", can be bounded if and only if the number of factors
L,,:#= {OJ is finite.
REMARK. A norm generating the topology of the product of a finite
family of normed spaces can be constructed from the given norms in a
variety of ways (Exercise 4).
42 LOCALL Y CONVEX TOPOLOGICAL VECTOR SPACES [Ch. II
2.3
The quotient space 01 a normable (and complete) t.V.S. L over a closed
subspace M is normable (and complete). II (L, II II) is a normed space, the
norm x -> Ilx II = inf{ Ilx II: x E x} generates the topology 01 L/ M.
Proof Since M is closed, L/ M is Hausdorff by (I, 2.3); since the natural
map </> of L onto L/ M is linear, open, and continuous, </>( V) is a convex
O-neighborhood in L/ M which is bounded by (I, 5.4) if V is a bounded,
convex O-neighborhood in L; thus L/M is normable by (2.1). By (1,6.3) L/M
is complete when L is, and the pseudonorm x -> Ilx II, which is easily seen to be
a norm, generates the topology of L/ M.
It is immediate that the bounded sets in a normed space L are exactly those
subsets on which x -> Ilx II is bounded. Thus if L, N are normed spaces over
K, and u is a continuous linear map on L into N, it follows from (I, 5.4) that
the number
Ilull = sup{llu(x)ll: x EL,llxll ~ I}
is finite. It is easy to show that u -> Ilu II is a norm on the vector space !E(L, N)
over K of all continuous linear maps on L into N. !E(L, N) is a Banach space
under this norm if N is a Banach space; in particular, if N is taken to be
the one-dimensional Banach space (Ko, 1I) (cf. Chapter I, Section 3) then
L' = !E(L, K u), endowed with the above norm, is a Banach space called the
strong dual of L.
Examples
The following examples are intended to present some principal types of

Banach spaces occurring in analysis. As normed spaces in general, these
spaces have been widely covered in the literature (e.g., Day [2], Dunford-
Schwartz [I], Kothe [5]), to which we refer for details.
1. Let X be a non-empty set. Denote by R(X) the vector space over
K(K = R or C) of all K-valued bounded functions; under the norm
1-> IIIII = sup{l/(t)l: t EX}, R(X) is a Banach space. If Xu is any
subset of X, the subset of all I E R(X) vanishing on Xu is a closed sub-
space.
If:E is a a-algebra of subsets of X (cf. Halmos [1]), let M(X, :E) be the
set of all :E-measurable functions in R(X); M(X, 1:) is a closed subspace
of R(X).
If X is a topological space and Ifb'(X) the set of all continuous functions
in R(X), Ifb'(X) is a closed subspace of R(X).
An example of particular importance is the space Ifb'(X) when X is a
compact (Hausdorff) space. Using the fact that all (except the second)
of the preceding spaces are vector lattices of a particular type, it can be
proved (Chapter V, Theorem 8.5) that each of them is norm-isomorphic
to a space Ifb'(X), where X is a suitable compact space.
§2] NORMED AND NORMABLE SPACES 43
2. Let (X, I:, p.) be a measure space in the sense of Halmos [1], so
that p. is a non-negative (possibly infinity-valued), countably additive
set function defined on the CT-algebra I: of subsets of X. Denote by 2 P
the set of all I:-measurable, scalar-valued functions I for which I/IP
(1 ~ p < 00) is p.-integrable; the well-known inequalities of HOlder and
Minkowski (cf. Halmos, I.c.) show that 2 P is a vector space and that
(J
1-+ I/I Pdp.)1 /P is a semi-norm on 2 P• If ,Alp denotes the subspace of
2 P consisting of all p.-null functions and [I] the equivalence class of
I e2P mod ,Alp, then
is a norm on the quotient space 2 P/,AI P' which thus becomes a Banach
space usually denoted by LP(p.).
2 00 commonly denotes the vector space of p.-essentially bounded
I:-measurable functions on X; a I:-measurable function is p.-essentially
bounded if its equivalence class mod ,Alp contains a bounded function.
Thus 2 00 /,AI p is algebraically isomorphic with M(X, I:)/(,AI p n M(X, I:»;
the latter quotient of M(X, I:) is a Banach space usually denoted by
L 00 (p.). L oo(p.) is again norm-isomorphic to ~(X) for a suitable compact
space X.
3. Let X be a compact space. Each continuous linear form I -+ p.o(f)
on ~(X) is called a Radon measure on X (Bourbaki [9], Chapter III).
For each P.o, there exists a unique regular signed (respectively, complex)
Borel measure p. on X in the sense of Halmos [1] such that p.o(f) =
Jldp. for all I e ~(X). The correspondence flo -+ p. is a norm isomor-
phism of the strong dual Jt{X) of ~(X) onto the Banach space of all
finite signed (respectively, complex) regular Borel measures on X, the
norm lip. II being the total variation of p.. Because of this correspondence,
one frequently writes p.o(f) = Jldp.o.
Returning to the general case where (X, I:, p.) is an arbitrary measure
space (Example 2, above), let us note that for 1 < P < + 00, the strong
dual of LP(p.) can be identified with Lq(p.), where p-1 + q-1 = I, in the
sense that the correspondence [g] -+ ([I] -+ Jig dp.) is a norm iso-
morphism of U(fl) onto the strong dual of LP(p.). In the same fashion,
the strong dual of L1(fl) can be identified with L OO{p.) whenever p. is
totally CT-finite. (For a complete discussion of the duality between
LP(p.) and Lq(p.), see Kelley-Namioka [1], 14. M.)
4. Let Zo denote the open unit disk in the complex plane. Denote by
HP(I ~ P < + 00) the subspace of CZo consisting of all functions which
are holomorphic on Zo and for which
IIfllp = sup
0<r<1
(f 2" If(re
0
it ) IPdt
)1/P
is finite; 1-+ III lip is a norm on HP under which HP is a complex Banach
space. Similarly, the space of all bounded holomorphic functions on Zo
is a complex Banach space H OO under the norm I -+ 11/1100 = sup{I/(c)I:
, e Zo}; for details we refer to Hoffman [1].
44 LOCALLY CONVEX TOPOLOGICAL VECTOR SPACES [Ch."
The preceding examples can be substantially extended by considering

Banach spaces of functions taking their values in an arbitrary (B)-
space F.
5. An especially important class of Banach spaces is the class of Hil-
bert spaces. The presence of an inner product distinguishes Hilbert
spaces quite drastically from general Banach spaces; the theory of
Hilbert spaces, and in particular of their linear transformations, is
elaborate and the literature is very extensive. For later use, we record
here only the definition and the most elementary properties of Hilbert
space; see also Bourbaki ([8], chap. V), Halmos [2], and Sz.-Nagy [1].
Let H be a vector space over C and let (x, y) -+ [x, y] be a complex-
valued function on H x H such that the following conditions are satis-
fied (0(* denoting the complex conjugate of 0( E C):
(i) For each y E H, x -+ [x, y] is a linear form on H.
(ii) [x, y] = [y, x]* for all x E H, y E H.
(iii) [x, x] ~. 0 for all x E H.
(iv) [x, x] = 0 implies x = O.
The mapping (x, y) -+ [x, y] is called a positive definite Hermitian form
(or inner product); x -+ ",,[x, x] is a norm II lion H, and (H, II II) is called
a pre-Hilbert (or inner product) space. The inner product satisfies
Schwarz'inequality: ![x, y]1 ;;;; Ilxll Ilyll· If the normed space (H, II II)
is complete (hence a Banach space), it is called a Hilbert space. The cor-
responding notion of real inner product space or real Hilbert space,
respectively, is obtained by assuming (x, y) -+ [x, y] to be real valued
and Hto be a real vector space. AfunctiononH x Hsatisfying(i)through
(iii) but not necessarily (iv) is called a positive semi-definite Hermitian
form; in these circumstances, x-+ ",,[x, x] = p(x) is a semi-norm on H,
and the quotient space Hlp-l(O) is an inner product space under (~, y)
-+ [x, y], where x -+ ~ denotes the canonical map of H onto Hlp -1(0).
It is clear that the property of being an inner product (respectively,
Hilbert) space is inherited by subspaces (respectively, by closed
subspaces). More important, every closed subspace M of an inner prod-
uct space H possesses a (topologically) complementary subspace: the
subspace Ml. = {x E H: [x, y] = 0 for all y EM}, called the subspace
of H orthogonal to M, satisfies the relation H = M EB Ml.. The projec-
tion of H on to M thatvanishes on Ml. is called the orthogonal projection
of H onto M. Hence for every closed subspace M of H, the quotient
space HIM, being norm isomorphic with Ml., is an inner product space.
A subset {x,,: 0( E A} of H is orthonormal if [x"' x p] = [)"p for all 0(, PEA;
any total orthonormal subset is called an orthonormal basis of H (cf.
Chapter III, Section 9 and Exercise 23). The existence of orthonormal
bases in every complete inner product space H is implied by Zorn's lem-
ma, and it can be shown that every orthonormal basis of H has the same
cardinality d; d is called the Hilbert dimension of H.
Every Hilbert space H is self-dual in the following sense : Iff is a con-
tinuous linear form on H, there exists a unique element Z E H such
that f(x) = [x, z](x E H); f -+ Z is a norm-preserving, additive map of
§3l THE HAHN-BANACH THEOREM 45
the strong dual H' onto H under which IX! is mapped onto IX*Z(IX E C),
and which is therefore called conjugate-linear. (f .... Z is a norm iso-
morphism if H is a real Hilbert space.) This has an immediate conse-
quence: If H 1, H2 are Hilbert spaces and u is a continuous linear map
of H1 into H 2, then the identity [u(x), yh = [x, U*(Y)]1 on H1 x H2
defines a continuous linear map u* of H2 into H 1, called the conjugate
of u. It is easy to· see that Ilu II = Ilu* II and that u .... u* is a conjugate-
linear map of !l'(H1, H 2) onto !l'(H2, H 1).
The most important concrete examples of Hilbert spaces are the spaces
L 2 (/1) (Example 2 above) with inner product I!g* d/1; special instances of
the latter are the spaces [~ (or 12(A», defined to be the subspace of CA
(or of RA in the real case) of all families {ea: IX E A} for which Lleal 2
< + 00, A being a set of cardinality d. The inner product on these
spaces is defined to be [e, 11] = Lae al1!; for each pair (e,l1) the family
{eal1~: IX E A}, which has lat most countably many non-zero members,
is summ~ble by SchwarJ. inequ~lity. ~Cf. C~a~ter III, ~xer~ise 23)
Every HIlbert space of ljillbert dImenSIOn d IS IsomorphIC with I:; If
{xa: IX E A} is any orthonormal basis of H, the mapping x .... {[x, x,,]:
IX E A} is an isomorphism of H onto I~.
Finally, the method used in constructing the spaces [2(A) can be
applied to any family {Ha: IX E A} of Hilbert spaces. Consider the sub-
space H of TIaH" consisting of all elements (xa) such that the family
{llx,,11 2 : IX EA} is summable; then [x,y] = La[Xa,Ya] defines an inner
product under which H becomes a Hilbert space; H is called the Hilbert
direct sum of the family {Ha: IX E A}. In particular, if A is finite, then
H = TI"Ha and the topology of the Hilbert direct sum is the product
topology (cf. Exercise 4); hence each finite product of Hilbert spaces is a
Hilbert space in a natural way.
3. THE HAHN-BANACH THEOREM
In the preceding section we have seen that the strong dual of a Banach
space is a Banach space; however, we could not assert that this space always
contains elements other than O. We have also seen (Chapter I, Exercise 6)
that there exist metrizable t.v.s. on which 0 is the only continuous linear
form. It is the purpose of this section to establish a theorem guaranteeing
that on a large class oft.v.s. (Section 4), including the normable spaces, there
exist sufficiently many continuous linear forms to distinguish points. This
result, the theorem of Hahn-Banach, is undoubtedly one of the most im-
portant and far-reaching theorems in functional analysis.
We begin by establishing a lemma that contains the core of the Hahn-
Banach theorem.
LEMMA. Let L be a Hausdorff t.v.s. over R of dimension at least 2. If B is an

open, convex set in L not containing 0, there exists a one-dimensional subspace
of L not intersecting B.
Proof. Let E be any fixed two-dimensional subspace of L. If E ('\ B = 0,

the result is immediate; so let us assume that Bl = E ('\ B is non-empty.
BI is a convex, open subset of E not containing o. By (I, 3.2) we can identify
E with R~ (the Euclidean plane). Project BI onto a subset of the unit circle
C of E by the mapping
I: (x, y) --+ ( ; , n, r = (x 2 + y2yt.
Since B I , being convex, is connected,f(BI ) is connected, forlis continuous

on B I ; moreover,f(BI } is an open subset of C. Hence/(B t } is an open arc on
C which subtends an angle ~ 1t at 0; otherwise, there would exist points in B t
whose images under I are diametrical, contradicting the hypothesis 0 ¢ B I ,
since BI is convex. Consequently, there exists a straight line in E passing
through 0 and not intersecting B t .
The following theorem is sometimes called Mazur's theorem (cf. Day [2]),
and is virtually the geometrical form of the Hahn-Banach theorem (Bourbaki
[7]).
3.1
Theorem. Let L be a t.v.s., let M be a linear manifold in L, and let A be
a non-empty convex, open subset oiL, not intersecting M. There exists a closed
hyperplane in L, containing M and not intersecting A.
Proof. After a translation, if necessary, we can have 0 E M, so that M is a

subspace of L. Consider the set 9Jl of all closed real subspaces of L that con-
tain M and do not intersect A; 9Jl is non-empty, since M E 9Jl.
Order 9Jl by inclusion c. If {M~} is a totally ordered subset of 9Jl, the

closure of U~M~ is clearly its least upper bound; hence by Zorn's lemma there
exists a maximal element Ho of 9Jl. If Lo denotes the real underlying space of
L (Chapter I, Section 7), the quotient space Lo/Ho is Hausdorff by (I, 2.3),
for Ho is closed. Because of A"" 0, Lo/Ho has dimension ~ I; suppose that
Lo/Ho is of dimension ~2. Since the natural map cp of Lo onto Lo/Ho is
linear and open, B = cp(A) is a convex, open subset of Lo/Ho, not containing
0, since Ho does not intersect A. Hence by the preceding lemma, there exists
a one-dimensional subspace N of Lo/Ho not intersecting B; this implies
that H = cp -leN) is a closed subspace of Lo containing Ho properly and not
intersecting A. This contradicts the maximality of Ho in 9Jl; hence Lo/Ho
has dimension 1, and Ho is a closed, real hyperplane containing M and not
intersecting A. This completes the proof when L is a t.v.s. over R.
If L is a t.v.s. over C, then M = iM (assuming 0 EM), since M is a subspace
of L. Consequently HI = Ho ('\ iHo, which is a closed hyperplane in L not
intersecting A, contains M, and the proof is complete.
§4] LOCALLY CONVEX SPACES 47
COROLLARY. /f L is a t.v.s., there exists a continuous linear form f i= 0 on L

if and only if L contains a non-empty convex, open subset A i= L.
Proof. Iff i= 0 is a continuous linear form onL, the subset A = {x: If(x) I< I}
is i=L, convex, and open. Conversely, if the convex set A c L is open and
Xo ¢ A, Xo is contained in a closed hyperplane (not intersecting A) by (3.1)
which by (I, 4.2) implies the existence of a non-zero continuous linear form
onL.
We deduce now from (3.1) its analytic equivalent, the theorem of Hahn-
Banach; for a more general form valid in real vector spaces, see Exercise 6.
3.2
Theorem. Let L be a vector space, let p be a semi-norm on L, and let M
be a subspace of L. /f f is a linear form on M such that If(x)1 ~ p(x) for all
x E M, there exists a linear formfl extending f to L and such that IfI (x) I ~ p(x)
for all x EL.
Proof. Since the case f = 0 is trivial, we assume that f(x) i= 0 for some
XEM. By (1,1.2), the convex, circled sets Vn={xEL:p(x)<n- I }, nEN,
form a O-neighborhood base for a topology :.t under which L is a t.v.s.
Define H = {x E M:f(x) = lL then H is a hyperplane in M and a linear
manifold in L. Let A = VI; A is open in L(:.t) by (1.6) and An H = 0,
since p(x) ~ I for x E H. By (3.1) there exists a hyperplane HI in L, containing
H and not intersecting A. Since HI n M i= M (for 0 ¢ HI) and HI => H, it
follows that HI n M = H, since H and HI n M are both hyperplanes in M.
By (I, 4.1), we have HI = {x:fl(x) = I} for some linear formfl on L, since
o¢ HI. Now H = HI n M implies thatf(x) = j~(x) for all x E M; that is,/1
is an extension of/to L. From HI n A = 0, it follows that Ifl(x)1 ~ p(x) for
all x E L, thus completing the proof.
COROLLARY. /f(L, II If) is a normed space, M is a subspace of L, andfis a
linear form on M such that If(x) I ~ IIxil (x EM), thenf has a linear extension
fl to L satisfying IfI (x)1 ~ IIx II (x E L).
This is the classical form of the theorem for normed spaces.
4. LOCALLY CONVEX SPACES
A topological vector space E over R or C will be called locally convex if it

is a Hausdorff space such that every neighborhood of any x E E contains a
convex neighborhood of x. Equivalently, E is a locally convex t.v.s. or briefly
locally convex space (I.c.s.) if the convex neighborhoods of 0 form a base at
o with intersection {O}. A topology on a vector space over R or C, not neces-
sarily Hausdorff but satisfying (LT)I and (LT)2 (Chapter I, Section 1) and
possessing a base of convex O-neighborhoods, is called a locally convex
topology. It will be convenient to have this distinction, since the majority
48 LOCALLY CONVEX TOPOlOGICAL VECTOR SPACES [Ch. II
of the valuable results produced by convexity (such as Corollary I of (4.2),

below) holds only in Hausdorff spaces, while it is sometimes necessary (such
as in the proof of (3.2), above) to consider locally convex topologies that
are not separated.
By the topological dual (or briefly dual) of a t.v.s. L, we understand the
vector space L' of continuous linear forms on L; L' is a subspace of the
algebraic dual L* of, L. If E is a I.c.s., its dual E' separates points in E; that
is, for any two elements x, y E E, x f= y, there exists an fE E' such that
f(x) =1= f(Y). (Equivalently, for every non-zero x E E there exists fEE' with
f(x) f= 0.) This important result is an immediate consequence of (3.1), and
formally contained in (4.2), Corollary 1.
If E is a vector space, a locally convex topology on E can geometrically be
defined by selecting a filter base min E, consisting of radial, convex, ~ircled
sets and such that V E m implies -t V Em; since -t V + ! V = V by the
convexity of each V, the corollary of (I, 1.2) implies that mis a O-neighbor-
hood base for a unique locally convex topology. Conversely, every I.c.
topology on E can be so defined; for example, the family of all closed,
convex, circled O-neighborhoods is a base at O.
Analytically a locally convex topology on E is determined by an arbitrary
family {Pa: r:J. E A} of semi-norms as follows: For each r:J. E A, let Ua =
{xEE:paCx);;:; I} and consider the family {n-1U}, where n runs through all
positive integers and U ranges over all finite intersections of sets U a (a E A).
This family msatisfies the conditions indicated above and hence is a base at
o for a locally convex topology ::t on E, called the topology generated by the
family {Pa}; equivalently, {Pal is said to be a generating family of semi-norms
for ::to Conversely, every locally convex topology on E is generated by a
suitable family of semi-norms; it suffices to take the gauge functions of a
family of convex, circled O-neighborhoods whose positive multiples form a
subbase at O. It is clear from (1.6) that every member of a generating family
of semi-norms is continuous for ::t, and it i~ easy to see that ::t is Hausdorff
if and only if for each non-zero x E E and each family &> of semi-norms
generating ::t there exists P E &> such that p(x) > O. We can leave it to the
reader to prove that, for a given I.e. topology ::t, the smallest cardinality of a
base at 0 is identical with the smallest cardinality of a generating family of
semi-norms, except when the latter is 1.
It is a direct consequence of the definitions (Chapter I, Section 2) that
induced, quotient, and product topologies of locally convex topologies are
locally convex; accordingly, subspaces, separated quotients, and products of
I.c.s. are again I.c.s. These will be discussed in the subsequent sections. Here
we confine ourselves to a few simple facts concerning metrizable I.c.s.
In view of (I, 6.1) a I.c.s. is metrizable if and only if it possesses a countable
base at 0 consisting of convex, circled sets, and hence a base which consists of
the members of a decreasing sequence {Un} of convex, circled sets. Equiva-
lently, a I.c.s. is metrizable if and only if its topology is generated by a
countable family of semi-norms, and hence by a sequence of semi-norms
§4] LOCALLY CONVEX SPACES 49
{Pn} which is increasing. A complete metrizable l.c.s. is called a Frechet-space,

or briefly (F)-space. Clearly, every complete normable space (and hence
every Banach space) is an (F)-space; the simplest example of an (F)-space
which is not normable is furnished by the space K~ of all numerical se-
quences under the product topology (K~ is not normable by (I, 5.5) and (2.2».
It follows from the results of Chapter I, Section 6, that every closed subspace
and every separated quotient of an (F)-space is an (F)-space, and so is every
countable product of (F)-spaces.
As a simple example for the definition of a locally convex topology by
families of semi-norms, let E be any vector space with algebraic dual E*,
and suppose that M is a subset of E* (M #- 0). The semi-norms x ---+ If(x) I
(f E M) generate a locally convex topology on E under which E is a l.c.s.
if and only if M separates points in E.
4.1
The completion E of a l.c.s. E is a l.c.s., whose topology is generated by the
continuous extensions to E of the members of any generating family of semi-
norms on E.
Proof If p is any member of a family [!Il of generating semi-norms on E,
p has a unique continuous extension,v to E by (1.6). If U = {x E E: p(x) ~ I},
then D = {x E E: ,vex) ~ I} is the closure of U in it It follows from (I, 1.5)
that E is a l.c.s. (since D is convex) and that {p: p E [!Il} is a generating family
of semi-norms on E.
The following consequence of the Hahn-Banach theorem reflects a basic
property of locally convex topologies:
4.2
Theorem. Let E be a t.v.s. whose topology is locally convex. If f is a
linear form, defined and continuous on a subspace M of E, then f has a con-
tinuous linear e]Ctension to E.
Proof Since fis continuous on M, V = {x: If(x) I ~ I} is a O-neighborhood
in M. There exists a convex, circled O-neighborhood U in E such that
U 11 M c: V; the gauge p of U is a continuous semi-norm on E such that
If(x)1 ~ p(x) on M, since U 11 M c: V. By (3.2) there exists an extension f1
off to E such that If1 (x)1 ~ p(x) on E;f1 is continuous, since 1ft (x) - f1 (y)1 ~ 8
whenever x - y E 8U (8) 0).
COROLLARY 1. GivelJ n (n E N) linearly independent elements Xv ofa I.c.s. E,
there exist n continuous linear forms fp. on E such that fp.(x v ) = Op.v (fl, v =
1, ... , n).
Proof Denote by M the n-dimensional subspace with basis {x.} in E.
By (J, 3.4), the linear forms gp. (fl = 1, ... , n) on M which are determined
by gix.) = 0p.., where op.p. = 1 and Op.v = 0 for fl #- v (fl, v = 1, ... , n), are
continuous. Any set {fl'} of continuous extensions of the respective forms gI'
to E has the required properties.
COROLLARY 2. Any finite-dimensional subspace M of a l.c.s. E has a comple-
mentary subspace.
Proof Let M have dimension n and let {fl'} be n continuous linear forms on
E whose restrictions to M are linearly independent (cf. Corollary 1). Then
nf;I(O) is a closed subspace of E and an algebraic complementary
n
N =
1'=1
subspace of M, and the assertion follows from (I, 3.5).
If E is a t.v.s. whose topology is locally convex, the definition of bounded-
ness (Chapter I, Section 5) implies that the convex hull of a bounded subset
of E is bounded. In particular, the family of all closed, convex, and circled
bounded subsets is a fundamental system of bounded sets in E.
4.3
In every locally convex space, the convex hull and the convex, circled hull of
a precompact subset is precompact.
Proof Since the circled hull of a precompact set is clearly precompact
(cf. Chapter I, 5.1), it is enough to prove the assertion for convex hulls; the
reader will notice, however, that (with the obvious modifications) the follow-
ing proof also applies to convex, circled hulls. Observe first that the convex
hull P of a finite set {a;: i = 1, ... , n} is compact, for P is the image of the
n
compact simplex {(AI' ... , An): A;;:;; 0, LA; = I} c R n under the continuous
1
map (Al, ... ,An)~ LA;a;. (This is a special case of (10.2) below.) Now let
BeE be precompact, C the convex hull of B, Van arbitrary convex neighbor-
hood of 0 in E. Supposing B to be non-empty, there exist elements a; E B
n
(i = 1, ... , n) such that Be U(a; + V). The convex hull P of {aJ is compact,
1
and C c P + V, since P + V is convex and contains B; hence we have
m
Pc U(b j + V) for a suitable finite subset {b/i = 1, ... , m) of Pee, and it
1 m
follows that C c U(b j + 2 V), which shows C to be pre compact.
I
COROLLARY. If E is a quasi-complete l.c.s., then the closed, convex hull and

the closed, convex, circled hull of every precompact subset of E are compact.
However, if E is not quasi-complete, then the closed, convex hull of a
compact subset of E can fail to be compact (Exercise 27).
For the construction of locally convex spaces from those of a given class a
substantially more general approach proves fruitful than we have discussed
in Chapter I, Section 2, for arbitrary topological vector spaces; the two
following sections are concerned with two basic methods of generating
locally convex spaces.
§5] PROJECTIVE TOPOLOGIES 51
5. PROJECTIVE TOPOLOGIES
Let E and EIZ (IX E A) be vector spaces over K, letfoe be a linear map on E
into E", and let Z" be a locally convex topology on E" (IX E A). The projective
topology Z on E with respect to the family {(E", ZIZ,foe): IX E A} is the coarsest
topology on E for which each of the mappings foe (IX E A) on E into (E", Z,,) is
continuous.
Clearly, Z is the upper bound (in the lattice of topologies on E) of the
top.ologies f;l(ZIZ) (IX E A); if x E E and XfZ = foe(x) E E IZ , a Z-neighb.orh.o.od
base of x is given by all intersections n
f;l(UIZ ), where U" is any neighbor-
lZeH
ho.od of X IZ with respect to ZIZ' and H is any finite subset of A. Since the foe
are linear maps and the Z" are l.ocally convex t.opol.ogies .on the respective
spaces, EIZ , Z is a translati.on-invariant top.ology .on E with a base of convex
O-neighborhoods satisfying conditions (a) and (b) of (1, 1.2); hence Z is a
locally convex top.ology on E.
5.1
The projective topology on E with respect to the family {(EIZ' Z",foe): IX E A}
is a Hausdorff topology if and only if for each non-zero x E E, there exists an
IX E A and a O-neighborhood U" c EIZ such that fix) t/: UIZ .
Proof If Z is Hausd.orff and 0 ¥= x E E, there exists a O-neighborhood U in

(E, Z) not c.ontaining x; since there exist O-neighb.orhh.ods U" c EIZ and a
finite subset H c A with n
f;l(UIZ ) c U, we must havefoe(x) t/: UIZ for some
"eH
H. C.onversely, foe(x) t/: UIZ implies x t/:f;l(UIZ ) which sh.ows Z to be a
IX E
Hausdorff t.op.ol.ogy.
5.2
A mapping u of a topological space F into E, where E is endowed with the

projective topology defined by the family {(E", ZIZ,foe): IX E A}, is continuous if
and only iffor each IX E A, foe 0 u is continuous on F into (EIZ , ZIZ).
Proof If u is continu.ous, then, clearly, so is eachfoe 0 u (IX E A). Conversely,

let GIZ be any open subset of E IZ ; then u- 1[f;1(GIZ ) ] is an .open subset of F.
Now each .open subset G .of E is the union of a suitable family of finite
intersecti.ons .of sets f; 1 (GIZ ), whence it f.oll.ows that u- 1(G) is open in F, and
hence u is c.ontinu.ous.
The reader will have' n.oticed that n.o vector space concepts are needed in
the preceding result; in fact, it reflects a general property .of pr.ojective
t.op.ol.ogies (cf. Prerequisites).
We pr.oceed t.o enumerate the most imp.ortant examples .of pr.ojective
locally c.onvex topol.ogies.
Subspaces. Let M be a subspace of the I.c.s. (E, Z); the topology of M

(i.e., the topology induced on M by :t) is the projective topology on M with
respect to the canonical imbedding M --t E.
Products. Let {(E", Z,,): oc e A} be a family of t.v.s., each Z" being a locally
convex topology. The product topology Z on E = fl"E" (Chapter I, Section 2)
is evidently locally convex; :t is the projective topology on E with respect to
the projections E --t E" (oc e A). In particular, the product of any family of
l.c.s. is a I.c.s.
Upper Bounds. Let {Z,,: oc e A} be a family oflocally convex topologies on a
vector space E; their least upper bound Z (in the lattice of topologies on E)
is a locally convex topology which is a projective topology; in fact, Z is the
projective topology with respect to the family {(E, Z", e): oc e A}, where e is
the identity map of E.
Weak Topologies. Let E be a vector space over K and let F be a subset of
E* that is non-empty. Set Ef = Ko for every fe F; the projective topology
on E with respect to the family {(Ef,J):fe F} is called the weak topology
generated by F, and is denoted by (1(E, F). Since F can be replaced by its
linear hull in E* without changing the corresponding projective topology,
F can be assumed to be a subspace of E*. By (5.1), E is a I.c.s. under (1(E, F)
if and only if F separates points in E.
In particular, when (E, Z) is a locally convex space, then its dual E'
separates points in E by (4.2), Corollary 1; (1(E, E') is called the weak topology
of E (associated with :t if this distinction is necessary). On the other hand, E'
is a I.c.s. under aCE', E), called the weak dual of (E, Z); here E is to be
viewed as a subspace of (E')*.
Projective Limits. Let A be an index set directed under a (reflexive, transi-
tive, anti-symmetric) relation " ~ ", let {E,,: oc E A} be a family of I.c.s. over
K, and denote, for oc ~ 13, by g/%p a continuous linear map of Ep into E".
Let E be the subspace of fl"E" whose elements x = (x/%) satisfy the, relation
x" = g/%p(xfJ) whenever oc ~ 13; E is called the projective limit of the family
{E,,: oc E A} with respect to the mappings g"p(a, 13 E A; a ~ 13), and denoted by
lim g"pEp' It is evident that the topology of E is the projective topology on E
+--
with respect to the family {(E", Z/%,J,,): a E A}, where Z" denotes the topology
of E", and f.. denotes the restriction to E of the projection map PIX of flpEp
onto E".
5.3
The projective limit of a family of quasi-complete (respectively, complete)
locally convex spaces is quasi-complete (respectively, complete).
Proof. Let E = lim g"pEp, F = fl"E". If every E" is complete, then F is
+-
complete; if every E" is quasi-complete, so is F by (I, 5.6). Hence the propo-
sition will be proved when we show that E is a closed subspace of F. Denote,
§5] PROJECTIVE TOPOLOGIES 53
for each pair (0:, f3) E A x A such that 0: ~ f3, by V"p the subspace {x: Xa -
gap(xp) = O} of F. Since Ea is Hausdorff and V",p is the null space of the
continuous linear map Pa - gap 0 Pp of F into Ea, V",p is closed; thus
E= n
a~p
VaP is closed in F.
The product fl",E", of a family {E",: 0: E A} of I.c.s. is itself an example of a
projective limit. If {H} denotes the family of all non-empty finite subsets of
A, ordered by inclusion, EH = fl Ea and gH,A denotes the projection of
aeH
EA onto EH when H c: A, then fl",Ea = ~ gH,AEA' Other examples of
projective limits are provided by the duals of inductive limits (Chapter IV,
Section 4), for which concrete examples will be given in Section 6. In the
proof of (5.4) below we construct, for every complete locally convex space E,
a projective limit of Banach spaces to which E is isomorphic. Finally, we
point out that, in general, there is nothing to prevent a projective limit of
I.c.s. from being {O}; but it can be shown (Exercise 10) that if A is countable
and certain additional conditions are satisfied, then p",(E) = E", (IX E A).
SA
Every complete l.c.s. E is isomorphic to a projective limit ofafamily of Banach
spaces; this family can be so chosen that its cardinality equals the cardinality
of a given O-neighborhood base in E.
Proof. Let {U"': 0: E A} denote a given base of convex, circled neighbor-
hoods of 0 in E. A is directed under the relation IX ~ p, defined by "IX ~ f3 if
Up c: U"'''' Denote by P'" the gauge of U'" and set Fa = Ej V"" where V", = p; 1(0)
(0: E A). If.x", is the equivalence class of x E E mod Va, then x", -+ IIx", II = pix)
is a norm on Fa generating a topology Z'" which is coarser than the topology
of the quotient space Ej V",. If 0: ~ p, every equivalence class mod Vp, xp,
is contained in a unique equivalence class mod V"" x"" since Vp c: V"'; the
mapping gap: Xp -+ x", is linear, and continuous from (Fp, Zp) onto (Fa, Z"')'
since IIx",11 ~ Ilxpli.
Let us form -the projective limit F = lim gafJFp(Zp). The mapping x -+ (xa)
+--
of E into F is clearly linear, and one-to-one, since E is Hausdorff. We show
that this mapping is onto F. Let H be any non-empty finite subset of A and let
Z = (z"') be any fixed element of F. There exists XH E E such that the equiva-
lence class of XHmod Va is Za for every 0: E H (if f3 E A is such that f3 ~ 0:
for all 0: E H, then any x E E with xp = zp will do). Now, if 0: E HI and IX E H 2 ,
then XH, - XH2 E U'" which shows the filter of sections of {XH}, where for each
finite H "# 0 an X H has been selected as above, to be a Cauchy filter in E.
Since E is complete, {XH} has a limit yin E for which Ya = z",(ex E A), because
x -+ (x",) is clearly continuous. Moreover, x -+ (xa) is readily seen to be a
homeomorphism, and hence an isomorphism of E onto F. Now F is a dense
subspace (proof!) of the projective limit lim gapFp, where F", (IX E A) is the
completion of (F"" Za) and g",P the continuous extension of gap (ex ~ f3) to Fp
54 LOCALL Y CONVEX TOPOLOGICAL VECTOR SPACES [Ch."
with values in P". But F, being isomorphic with E, is complete, and hence
F = lim B"pFp, which completes the proof of the theorem.
+--
1. Every Frechet space is isomorphic with a projective limit of a
COROLLARY
sequence of Banach spaces.
COROLLARY 2. Every locally convex space is isomorphic with a subspace of

a product of Banach spaces.
6. INDUCTIVE TOPOLOGIES
Let E and E" (0( E A) be vector spaces over K, let g" be a linear mapping of
E" into E, and let ~" be a locally convex topology on E" (0( E A). The in-
ductive topology on E with respect to the family {(E", ~'" g,,): 0( E A} is the
finest locally convex topology for which each of the mappings g,. (0( E A) is
continuous on (E", ~,,) into E.
To see that this topology is well defined, we note that the class fT of I.c.
topologies on E for which all g,. are continuous is not empty; the trivial
topology (whose only open sets are 0 and E) is a member of fT. Now the
upper bound ~ of .0/" (in the lattice of topologies on E) is clearly a locally
convex topology for which all g" are continuous, and hence ~ is the topology
whose existence was to be verified. (As an upper bound, ~ is also a projective
topology, namely the projective topology with respect to the family fT and
the identity map on E.) ~ need not be separated, even if all ~,. are. (We leave
it to the reader to construct an example.) A O-neighborhood base for ~ is
given by the family {U} of all radial, convex, circled subsets of E such that for
each 0( E A, g;;1(U) is a O-neighborhood in (E", ~,,). If E is the linear hull of
U g,.(E,,), such a base can be obtained by forming all sets of the form r"giUrz),
where U,. is any member of a O-neighborhood base in (E", ~,,).
6.1
A linear map v on a vector space E into a l.c.s. F is continuous for an in-
ductive topology on E if and only if each map v 0 g" (0( E A) is continuous on
(E,., ~,,) into F.
Proo}: The condition is clearly necessary. Conversely, let each v g" be0
continuous (0( E A) and let Wbe a convex, circled O-neighborhood in F. Then

(v gJ-1(W) = g;;1[V- 1(W)] is a neighborhood of 0 in E" (0( E A), which
0
implies that v- 1 ( W) is a neighborhood of 0 for the inductive topology on E.

We consider now the most important instances of inductive topologies:
Quotient Spaces. If E is a I.c.s. and M is a subspace of E, it is immediate
from the discussion in Chapter I, Section 2, that the topology of ElM is
locally convex. If ~ denotes the topology of E, the quotient topology is the
inductive topology with respect to the family {(E,~, rfJ)}, where rfJ is the
§6] INDUCTIVE TOPOLOGIES 55
natural (or canonical) map of E onto ElM. By (1,2.3) this topology is Haus-
dorff if and only if M is closed in E. The quotient topology can be generated
by a family of semi-norms derived from certain generating families of semi-
norms on E through a process analogous to that used in (2.3) (Exercise 8).
Locally Convex Direct Sums. If {Ea: ex E A} is a family of vector spaces over
K, the algebraic direct sum EEl",Ea (Chapter I, Exercise 1) is defined to be the
subspace of TI..E", for whose elements x all but a finite number of the pro-
jections x", = p..{x) are O. Denote by g", (ex E A) the injection map (or canonical
imbedding) E,. --+ EEl pEp. The locally convex direct sum of the family {E,.(Z,.):
ex E A} of I.c.s. is defined to be EEl",E,. under the inductive topology with
respect to the family {(E"" Z"" g,.): ex E A} and, when reference to the topologies
is desired, denoted by E(Z) = EEl..EaCZ",). Since Z is finer than the topology
induced on E by n"E",(Z",), Z is a Hausdorff topology and hence E(Z) is a
I.c.s. From the remarks made above it follows that a O-neighborhood base of
Ee..E"{Z",) is provided by all sets of the form U = r",giU",); that is
U = {L A",g,.(X",): L Il",1 ~ 1, x" E U a }, (*)
where {U",: ex E A} is any family of respective O-neighborhoods in the spaces
E",. For simplicity of notation, we shall often write x,. in place of g"(x,,,), thus
identifying E", with its canonical image g,,(E,,) in EElpEp. (Note that each got is
an isomorphism of EiZ,,) into EElpEp(Zp).)
6.2
The locally convex direct sum EEl,.E" of a family of l.c.s. is complete if and only
if each summand E" is complete.
Proof. Let E = Ee"E", be complete. Since each of the projections Pot of E
onto E" is continuous, every summand E" is closed in E and hence is com-
plete.
Conversely, suppose that each E" is complete (ex E A). Denote by ~l the
unique translation-invariant topology on E for which a O-neighborhood base
is given by the sets En V, where V = TI", V" and V" is any O-neighborhood in
E". Then Zl is evidently coarser than the locally convex direct sum topology
Z on E, and (E, Zl) is a t.v.s. (Chapter I, Exercise 1) which is complete. (In
fact, the sets V form a O-neighborhood base in F = TI"E", for a unique trans-
lation-invariant topology, under which F is easily seen to be a complete
topological group with respect to addition. If Z E F is in the closure of E,
then for each V there exists x E E satisfying x - Z E V; this implies Z E E,
and hence E is a closed subgroup of F.) To prove that (E, Z) is complete
it suffices, in view of (I, 1.6), to show that the Z-closures U are ~l-closed,
where U = r"U", and U'" is any convex, circled O-neighborhood in E,.. Let U
be given, and let (fj be a ~rCauchy filter on U with ~l-limit x = (x",). Denote
by H the finite set of indices {ex: x,. =1= O}, and write EH = EEl{E",: ex E H}
(setting EH = {O} if H = 0), and ED = €e{Ep: PEA", H}; finally, let p be the
projection E -+ EH that vanishes on ED. Clearly, p is Xrcontinuous and

X-continuous, and satisfies p(U) c:: U. Thus p(6)) is a filter base on D n EH
that Xl-converges to x = p(x). Since, H being finite, the topologies X and Xl
agree on EH , p(Q» also X-converges to x; it follows that xED, completing the
proof.
We remark that the locally convex sum of a finite family ofl.c.s. is identical
with their product.
Example. Let E be a vector space over K and let {x..: ex E A} be a basis
of E. Obviously E is isomorphic with the algebraic direct sum Ee ..K .. ,
where K.. = Ko(ex E A) and Ko is> the one-dimensional vector space
associated with K. By (I, 3.4), the imbedding map on K.. into E (K..
being endowed with the topology of K) is continuous for any topology
on E under which E is a t.v.s.; hence the locally convex direct sum
topology X on E = E9 ..K.. is the finest locally convex topology on E. X is
consequently a Hausdorff topology under which E is complete; it is
generated by the family of all semi-norms on E. Equivalently, a O-neigh-
borhood base in E(X) is formed by the family of all radial, convex.
circled subsets of E (hence X is sometimes called the convex core top-
ology). For further properties of this topology see Exercise 7.
6.3
A subset B of a locally convex direct sum Ee{E..: ex E A} is bounded if and
only if there exists a finite subset H c:: A such that p ..(B) = {O} for ex ¢ H, and
p ..(B) is bounded in E.. if ex E H.
Proof. Let B be bounded in $ ..E... Since, as we have noted above. the
projection P.. of the direct sum onto the subspace E .. is continuous, p..{B) is
bounded for all ex E A by (I, 5.4). Suppose there is an infinite subset B c:: A
such that p..{B) :j: {O} whenever ex E B; then B contains a sequence {exll} of
distinct indices. There exists a sequence {y(II)} c:: B such that y~:) :j: 0 for all
n E N,and hence, since E .." is Hausdorff, such that yt) ¢ nVII , where VII is a
suitable circled O-neighborhood in E..". Now if U is a O-neighborhood of type
(*) in Ee ..E.. such that U.." = VII for all n, then n- l y<") ¢ U for any n E N,
which contradicts the boundedness of B by (I, 5.3). Conversely, it is clear
that the condition is sufficient for a set B to be bounded.
COROLLARY. The I.c. direct sum of a family of quasi-complete l.c.s. is quasi-
complete.
Inductive Limits. Let {E..: ex E A} be a family of l.c.s. over K, where A is an
index set directed under a (reflexive, transitive, anti-symmetric) relation" ;a; "
and denote, whenever ex ~ p, by hp.. a continuous linear map of E.. into Ep.
Set F = eaE.. and denote (9.. being the canonical imbedding of E .. in F) by
H the subspace of F generated by the ranges of the linear maps g .. - Up 0 hp..
of E.. iilto F, where (ex, P) runs through all pairs such that ex ~ p. If the quotient
space F/H is Hausdorff (equivalently, if H is closed in F), th~ l.c.s. F/H is

called the inductive limit of the family {EO(: 0( E A} with respect to the map- .
pings h/l,., and is denoted by lim hp,.E,.. It appears to be unknown whether
H is necessarily closed in F. -+
The requirements for the construction of an inductive limit are often
realized in the following special form: {E,.: 0( E A} is a family of subspaces ofa
vector space E such that E,. "I: E/I for 0( =F p, directed under inclusion and
satisfying E = U,.E,.; then A is directed under" 0( ~ p if E,. c Ep ". Moreover,
on each E,. (0( E A) a Hausdorff locally convex topology l:,. is given such that,
whenever 0( ~ p, the topology induced by l:/I on E,. is coarser than l:,.. De-
noting by g,.(O( E A) the canonical imbedding of E,. into E and by hp,. the
canonical imbedding of E,. into E/I (0( ~ P), and supposing that the inductive
-
topology l: on E with respect to the family {(E,., l:,., g,.): 0( E A} is Hausdorff,
it is easy to see that the inductive limit lim hp,.E,. exists and is isomorphic with
E(l:). In these circumstances, E(l:) is called the inductive limit of the family
{E,.(l:,.): 0( E A} of subspaces. An inductive limit of a family of subspaces is
strict if l:p induces l:,. on E,.whenever 0( ~ p.
Examples
1. The locally convex direct sum of a family {E,.: 0( E A} of l.c.s. is
itself an example of an inductive limit. If {H} denotes the family of all
non-empty finite subsets of A, ordered by inclusion, EH = <:9 E,. and
"eH
hA,H the canonical imbedding of EH into EA when H c A, then E9,.E,. =
~hA'HEH.
2. Let R!' (n E N fixed) be represented as the union of countably many
compact subsets Gm{m EN) such that Gm is contained in the interior of
Gm + 1 for all m. The vector space !7}{Gm ) over C of all complex-valued
functions. infinitely differentiable on R" and supported by Gm , is a
Frechet-space under the topology of uniform convergence in all deriva-
tives; a generating family of semi-norms Pk{k = 0, 1,2, ... ) is given by
f -+ Pk(f) ;:::; 8upldfl,
the sup being taken over all t E R" and all derivatives of order k. If !7}
denotes the vector space of all complex-valued infinitely differentiable
functions on R" whose support is compact (but arbitrarily variable with
f), then the inductive topology l: on !7} with respect to the sequence
n
{!7}(Gm of subspaces is separated, for l: is finer than the topology of
uniform convergence on compact subsets of R", which is a Hausdorff
topology. Thus !7}(l:) is the strict inductive limit of a sequence of sub-
spaces; its dual !'J' is the space of complex distributions on R" (L.
Schwartz [I]).
3.. Let X be a locally compact space and let E be the vector space of
all continuous, complex-valued functions on X with compact support.
For any fixed compact subset C c X, denote by (Ee. l:c) the Banach
space of functions in E that are supported by C, with the uniform norm

generating ~c. Ordering the family of compact subsets of X by inclusion
and denoting by gc the canonical imbedding of Ec into E, the inductive
topology ~ on E is readily seen to be Hausdorff: ~ is finer than the
topology of compact convergence. Hence E(~) is the inductive limit of
the subspaces (Ec, ~d; the dual E(~)' is the space of complex Radon
measures on X (Bourbaki [9], chap. III).
The preceding definitions of inductive limit, and even of strict inductive
limit of a family of subspaces, are too general to ensure a great number of
interesting results (cf. Komura [I]). The situation is different for the strict
inductive limit of a sequence {En: n E N} of subspaces of E, with N under its
natural order. We prove some of the most important results which are due to
Dieudonne and Schwartz [I], and Kothe [2].
The strict inductive limit of an increasing sequence of (B)-spaces will be
called an (LB)-space, and that of (F)-spaces an (LF)-space. Example 2 above is
an (LF)-space. Example 3 is an (LB)-space, provided the locally compact
space X is countable at infinity (i.e., a countable union of compact sub-
spaces).
6.4
If {En(~n): n E N} is an increasing sequence of l.c.s. such that the topology
~n+l induces ~nfor all nand if the vector space E is the union of the subspaces
En (n EN), then the inductive topology on E with respect to the canonical im-
beddings En ~ E is separated and induces ~n on En (n EN).
The proof is based on this lemma:
LEMMA. If E is a l.c.s., M a subspace of E, and U a convex, circledO-neighbor-
hood in M, there exists a convex, circledO-neighborhood V in E with U = V n M.
If Xo E E is not in M, then V can be chosen so that, in addition, Xo if: V.
Proof. Let Wbe a convex, circled O-neighborhood in E such that W n M c: U.
Then V = r(W u U) satisfies V n M = U. Obviously U c: V n M; if
Z E V n M, then Z = AW + f.1U, where WE W, U E U, IAI + 1f.11 ~ 1, and AW =
Z - f.1U E M implies either A = 0 or WE M; in both cases we have Z E U,
whence V n M c: U. If Xo is not in the closure M of M, then in the preceding
construction W can be so chosen that (xo + W) n M is empty, whence it
follows that Xo if: V, for Xo = AW + f.1U E V would imply Xo - AW E M and
Xo - AW E Xo + W, which contradicts the choice of w.
Proof of (6.4). Let n be fixed and let Vn be a convex, circled neighborhood
of 0 in En(~n). Using the lemma above, we can by induction construct a
sequence {Vn+k}(k = 1,2, ... ) of subsets of E such that Vn+k is a convex,
circled O-neighborhood in En+k and Vn+k+l n E n +k = Vn+k for all k;;:; o.
Clearly, V = U Vn+k is a O-neighborhood for the inductive topology ~ on
k:2:0
E such that V -n En = Vn• It follows that the topology on En induced by ~ is
both finer and coarser than X", and hence is identical with Xn• Since E = UnEn,
it is also clear that X is separated.
-
A similar construction enables us to determine all bounded subsets in
E(X) when E(X) = lim EiXn ) is the strict inductive limit of a sequence of
subspaces.
6.5
Let E(X) = lim EiXn) and let En be closed in E n + 1(Xn + 1) (n EN). A subset
--+
BeE is bounded in E(X) if and only if for some n E N, B is a bounded subset
of En(Xn)·
Proof By (6.4) the condition is clearly sufficient for B to be bounded.
Conversely, assume that B is bounded but not contained in En for any n E N.
There exists an increasing sequence {k 1 , k 2 , ... } eN and a sequence {xn} c B
such that Xn E E kn + 1 , but Xn if. Ek n (n EN). Using the lemma in (6.4), we
construct inductively a sequence {Vd of convex, circled O-neighborhoods in
Ekn' respectively, such that n-1xn if. Vkn + 1 and Vkn + 1 fI Ek n = Vkn for all n E N.
00
Again V = U Vkn is a O-neighborhood in E(X), but n-1xn if. V (n EN), which

n=1
is impossible by (I, 5.3); hence the assumption that B be not contained in any
En is absurd.
6.6
The strict inductive limit of a sequence of complete locally convex spaces is
complete.
-
Proof Let E(X) = lim E.(Xn) be a strict inductive limit of complete I.c.s.
If ~ is a Cauchy filter in E(X) and U is the neighborhood filter of 0, then
~ + U = {F + U: F E ~, U E U} is a Cauchy filter base in E(X) which con-
verges if and only if ~ converges. We show that there must exist an no E N for
which the trace of ~ + U on Eno is a filter base; if so, this trace converges in
Eno(Xno ), since Eno is complete, and hence ~ converges in E. Otherwise, there
exists a sequence Fn E ~ (n E N) and a decreasing sequence of convex, circled
O-neighborhoods Wn in E(X) such that (Fn + Wn) fI En = 0 for every n.
=r
00
Now U (Wn fI En) is aO-neighborhood inE(X); we show that (Fn + U) fI

n= 1 p
En =0 for all n. Let y E En fI (Fn + U); then y = Zn + L: A;X;,
i= 1
where
L:!A'il ;;;;; 1, Xi E Wi fI E; (i = 1, ... , p) and Zn E Fn, hence
Since Wi C Wn for i > nand Wn is circled and convex, the right-hand member
of the last equality is in Fn + Wn, while the left-hand member is in En, which
is impossible; thus (Fn + U) n En = 0 for all n. Since iY is a Cauchy filter,

there exists FE iY such that F - FeU. Let w E F, then wEEk for some
kEN. Let v E Fk n F; then W = v + (w - v) E V + (F - F) c Fk + U, which is
contradictory.
COROLLARY. Every space of type (LB) or (LF) is complete.
7. BARRELED SPACES
In this and the following section, we discuss the elementary properties of
two types of locally convex spaces that occur frequently in applications, and
whose importance is largely due to the fact that they include all Frechet (and
hence Banach) spaces and that their defining properties are invariant under the
formation of inductive topologies.
A barrel (tonneau) in a t.v.s. E is a subset which is radial, convex, circled,
and closed. A l.c.s. E is barreled (tonne16) if each barrel in E is a neighborhood
of O. Equivalently, a barreled space is a l.c.s. in which the family of all
barrels forms a neighborhood base at 0 (or on which each semi-norm that is
semi-continuous from below, is continuous).
7.1
Every locally convex space which is a Baire space is barreled.
Proof. Let D be a barrel in E; since D is radial and circled, E = U nD.
neN
Since E is a Baire space, there exists no E N such that noD (which is closed)
has an interior point; hence D has an interior point y. Since D is circled,
- y ED; hence 0 = ty + t( - y) is interior to D by (1.1) because D is convex.
COROLLARY. Every Banach space and every Frechet space is barreled.
The property of being barreled is, in general, not inherited by projective
topologies; for instance, there exist (non-complete) normed spaces which
are not barreled (Exercise 14), and even a closed subspace of a barreled space
is, in general, not barreled (Chapter IV, Exercise 10). The same is true for
projective limits (cf. (5.4». However, it can be shown that the completion of
a barreled space is barreled (Exercise 15), and that the product of any family
of barreled spaces is barreled (Chapter IV, Section 4). Moreover, any induc-
tive topology inherits this property.
7.2
If'X is the inductive topology on E with respect to a family of barreled spaces
(and corresponding linear maps), then each barrel in E is a O-neighborhoodfor
'X.
Proof. Let 'X be the inductive topology on E with respect to the family
{(Eel' 'X~, gel): IX E A}, where 'Xa (IX E A) is a barreled l.c. topology on Ea. If D
§8] BORNOLOGICAL SPACES 61
is a barrel in E(Z), then g;;l(D) is a barrel D~ in Ea(Za) for each a E A, and

hence a neighborhood of 0; it follows that D is a O-neighborhood in E(Z).
COROLLARY 1. Every separated quotient of a barreled space is barreled; the
locally convex direct sum and the inductive limit of a family of barreled spaces
are barreled.
COROLLARY 2. Every space of type (LB) or (LF) is barreled.
Since a space of type (LF) (a strict inductive limit of a sequence of Frechet
spaces) is not a Baire space, there exist barreled spaces which are not Baire
spaces. Examples of such spaces are given in Section 6.
8. BORNOLOGICAL SPACES
A locally convex space E is bornological if every circled, convex subset
AcE that absorbs every bounded set in E is a neighborhood of O. Equiva-
lently, a bornological space is a l.c~s. on which each semi-norm that is
bounded on bounded sets, is continuous.
8.1
Every metrizable l.c.s. is bomological.
Proof If E is metrizable, there exists a countable O-neighborhood base
{Vn: n E N} by (I, 6.1), which can be chosen to be decreasing. Let A be a
convex, circled subset of E that absorbs every bounded set; we must have
Vn c nA for some n E N. For if this were false, there would exist elements
Xn E Vn such that Xn ¢ nA (n EN); since {xn} is a null sequence, it is bounded
by (I, 5.1), Corollary 2, and hence absorbed by A, which is contradictory.
It can be shown (Chapter IV, Exercise 20, and Kothe [5], §28.4) that a
closed subspace of a bornological space is not necessarily bornological. It is
not known whether every product of bornological spaces is bornological, but
the answer to this question depends only on the cardinality of the set of factor
spaces (Exercise 19). Thus since K~o is bornological, every countable product
of bomological spaces is bornological; more generally, the theorem of
Mackey-Ulam (Kothe [5], §28.8) asserts that every product of d bornological
spaces is bornological if d is smaller than the smallest strongly inaccessible
cardinal. (It is not known if strongly inaccessible cardinals exist; a cardinal do
is called strongly inaccessible if (a) do > ~o (b) ~)d~: a E A} < do whenever
card A < do and da < do for alia E A (c) d < do implies 2d < do. For details
on strongly inaccessible cardinals see, e.g., Gillman-lerison [1].) In particular,
it follows from the Mackey-Vlam theorem that Kg is bornological when
d =~ or d = 2K , where ~ is the cardinality of the continuum.
We note from (8.1) that every Frechet space (and hence every Banach
space) is bornological. Moreover, the property of being bornological is
preserved under the formation of inductive topologies.
8.2
Let ~ be the inductive topology on E with respect to a family of bornological
spaces (and corresponding linear maps); each convex, circled subset of E ab-
sorbing all bounded sets in E(~) is a O-neighborhood for ~.
Proof. Let A be such a subset of E and let ~ be the inductive topology with
respect to the family {(Ea, ~a, ga): c( E A}. If Ba is bounded in (E., ~a), then
gaCBa) is bounded in E(~) by (I, 5.4); hence A absorbs ga(Ba ), whence g; leA)
absorbs Ba, and so g; leA) is a O-neighborhood in (Ea, ~a)' Since this holds for
all c( E A, A is a O-neighborhood in E(~).
COROLLARY 1. Every separated quotient of a bornological space is borno-
logical; the locally convex direct sum and the inductive limit of any family of
bornological spaces is bornological.
In conjunction with (8.1) we obtain:
COROLLARY 2. Every space of type (LB) or (LF) is bornological.
Essentially by virtue of (I, 5.3), bomological spaces E have the interesting
property that continuity of a linear map u into a l.c.s. F is equivalent to the
sequential continuity of u, which is in tum equivalent to u being bounded on
bounded sets. This latter property, stating that the continuous linear maps of
E into any l.c.s. are exactly those linear maps that preserve bounded ness,
actually characterizes bomological spaces (see Exercise 18).
8.3
Let E(~) be bornological, let F be any l.c.s., and let u be a linear map on E
into F. These assertions are equivalent:
(a) u is continuous.
(b) {u(xn)} is a null sequence for every nuN sequence {xn} c: E.
(c) u(B) is bounded for every bounded subset B c: E.
Proof. (a)=>(b) is obvious. (b) => (c): If {u(x.}} (XnEB, nEN) is any se-
quence of elements of u(B), then {A.x n } is a null sequence in E for every null
sequence of scalars An E K, by (I, 5.3); hence {Anu(Xn)} is a null sequence in
F by (b), and repeated application of (I, 5.3) shows u(B) to be bounded in F.
(c) => (a): If B is any bounded set in E, and V is a given convex, circled 0-
neighborhood in F, then V absorbs u(B); hence u- 1 (V) absorbs B. Since B
was arbitrary and E is bomological, u- 1 (V) is a O-neighborhood for ~.
This holds for any given V which implies (since the topologies of E and Fare
translation-invariant) the continuity of u.
Let E(~) be any I.c.s., m the family of all bounded subsets of E(~). The
class fI of separated I.c. topologies on E for each of which every B E ~ is
bounded is non-empty, since ~ E fl. The upper bound ~o of fI Un the lattice
of topologies on E) is a projective topology (Section 5), and the finest I.c.
§9] SEPARATION OF CONVEX SETS 63
topology on E whose family of bounded sets is identical with lB. Clearly,

E('1: o) is bomological; E('1: o) is called the bornological space associated with
E('1:). Thus E('1:) is bomological if and only if '1: = '1:0.
The bomological space E(1: o) associated with E('1:) can also be defined as
follows: Let lB denote the family of all closed, convex, and circled bounded
subsets of E('1:), ordered by inclusion. For a fixed BE lB, consider the sub-
space EB = U nB of E; if PB denotes the gauge function* of Bin E B, PB is a
neN
norm (since B is bounded and E('1:) is Hausdorff) and (EB, PB) is a normed
space whose topology is finer than the topology induced by '1:. (If B is com-
plete, it follows from (I, 1.6) that (EB , PB) is a Banach space.) Let g B denote the
canonical imbedding of EB into E; it is evident that the inductive topology
on E with respect to the family {(EB, PB' g B): BE lB} is the bomological
topology '1:0 associated with '1:. Hence:
8.4
Every bomological space E is the inductive limit of a family of normed
spaces (and of Banach spaces if E is quasi-complete); the cardinality of this
family can be chosen as the cardinality of any fundamental system of bounded
sets in E.
COROLLARY. Every quasi-complete bomological space is barreled.
This is immediate from (7.2), Corollary 1. Since there exist normed spaces
which are not barreled (Exercise 14), a bomological space is not necessarily
barreled; conversely, Nachbin [1] and Shirota [1] have given examples of
barreled spaces that are not bomological.
We conclude this section with a remark that refines the last corollary and
follows easily from the preceding discussion, and which will be needed later
on (Chapter III, Section 3).
8.5
Let E be any l.c.s. and let D be a barrel in E. Then D absorbs each bounded
subset BeE that is convex, eire/ed, and complete.
Proof (EB, PB) is a Banach (hence barreled) space whose topology is finer
than the topology induced on EB by E. Thus D n EB is a barrel in (EB, PB),
which implies that D absorbs B.
9. SEPARATION OF CONVEX SETS

Let E be a vector space over K and let H = {x:f(x) = IX} be a real hyper-
plane in E; the four convex sets, Fa = {x:f(x) ;£ IX}, pa = {x:f(x) ~ IX},
Ga = {x:f(x) < IX}, Ga = {x:f(x) > IX}, are called the semi-spaces determined
* For B = 0, set EB = {O} and PB = O.

by H. (Note that rand F", are closed, G'" and G,. are open for the finest
locally convex topology on E.) If E is a t.v.s. and H is a closed real hyperplane
in E(equivalently,fis a continuous real linear form #0), then rand F", are
called closed semi-spaces, and G"', G", are called open semi-spaces. Two non-
empty subsets A, B of E, are said to be separated (respectively, strictly
separated) by the real hyperplane H if either A c F,. and B c F"', or B c F", and
A c r (respectively, if either A c G", and BeG"', or BeG" and A c G"').
If A is a subset ofthe t.v.s. E, a closed real hyperplane H is called a supporting
hyperplane of A if A n H # 0 and if A is contained in one of the closed semi-
spaces determined by H.
Theorem (3.1) is a separation theorem; it asserts that every convex open
set A # 0 and affine subspace M, not intersecting A, in a t.v.s. can be
separated by a closed real hyperplane. We derive from (3.1) two more
separation theorems (for the second of which it will be important that E is a
I.c.s.) that have become standard tools of the theory.
9.1
(FIRST SEPARATION THEOREM). Let A be a convex subset of a t.v.s. E, such
that A # 0 and let B be a non-empty convex subset of E not intersecting the
interior A of A. There exists a closed real hyperplane H separating A and B; if
A and B are both open, H separates A and B strictly.
Proof. A is convex by (1.2) and so is A - B, which is open and does not
contain 0, since A n B = 0. Hence by (3.1), there exists a closed real hyper-
plane Ho containing the subspace {O}, Ho = {x:f(x) = O}, and disjoint from
A-B. Now f(A - B) is convex and hence is an interval in R and does not
contain 0; we have, after a change of sign in f if necessary, f(A - B) > o.
Thus if O! = inff(A), H = {x:f(x) = IX} separates A and B: A c F"', Be F",.
Since A c A and A = (.4) by (1.3), we have A c F"', since F'" is closed in E;
thus H separates A and B. If A and B are open sets,f(A) andf(B) are open
intervals in R. For,f = g c/J where c/J is the natural map of Eo onto EoIHo,
0
which is open (Eo denoting the underlying real space of E), and g is an
isomorphism of EolHo onto Ro (Chapter I, Section 4). Hence A c G", BeG",
in this case, so that H separates A and B strictly.
COROLLARY. Let C be a convex body in E. Every boundary point of C is
contained in at least one supporting hyperplane of C, and C is the intersection
of the closed semi-spaces which contain C and are determined by the supporting
hyperplanes of c.
Proof. To see that each boundary point Xo of C is contained in at least one
supporting hyperplane, it will do to apply (9.1) with A = C, B = {x o}. To
prove the second assertion, we need the lemma that no supporting hyperplane
of C contains an interior point of C. Assuming this to be true, suppose that
y ~ C; we have to show that there exists a closed semi-space containing C,
§9] SEPARATION OF CONVEX SETS 65
but not containing y. Let x E C; the open segment joining x and y contains
exactly one boundary point Xo' There exists a supporting hyperplane H
passing through xo; H does not contain y, or else H would contain x. It is
clear that one of the closed semi-spaces determined by H contains C but not y.
We prove the lemma:
LEMMA. If C is a convex body in a t.V.S. E, no supporting hyperplane of C
contains an interior point of C.
Assume that x E H () C, where H = {x:f(x) = oc} is a supporting hyper-
plane of C such that C c: Fa. There exists y E C withf(y) < oc, since H cannot
contain C. Now f[x + e(x - y)] > f(x) = oc for every e > 0; since x E C,
X + e(x - y) E C for some e > O. This contradicts C c: Fa; hence the assump-
tion H () C :F 0 is absurd.
9.2
(SECOND SEPARATION THEOREM). Let A, B be non-empty, disjoint convex
subsets of a l.c.s. such that A is closed and B is compact. There exists a closed
real hyperplane in E strictly separating A and B.
Proof. We shall show that there exists a convex, open a-neighborhood V
in E such that the sets A + V and B + V are disjoint; since these are open,
convex subsets of E, the assertion will follow at once from (9.1).
It suffices to prove the existence of a convex, circled, open a-neighborhood
W for which (A + W) () B = 0; then V = ! W will satisfy the requirement
above. Denote by U the filter base of all open, convex, circled neighborhoods
of a in E and assume that A + U intersects B for each U E U; then {(A + U) ()
B: U E U} is a filter base in B which has a contact point Xo E B, since B is
compact. Hence Xo E A + U c: A + 2U for each U E U, whence Xo E n{A + U:
U E U} = A = A, since A is closed; this is contradictory.
Since sets containing exactly one point are compact, we obtain
COROLLARY I. Every non-empty closed, convex subset of a locally convex
space is the intersection of all closed semi-spaces containing it.
This implies a very important property of convex sets in locally convex
spaces:
COROLLARY 2. In every l.c.s. E(~), the ~-closure and the weak (that is,
u(E, E')-) closure of any convex set are identical.
Proof. Since
~ is finer than u(E, E'), every u(E, E')-closed subset of E is
~-closed. Conversely, every convex, ~-closed subset of E is weakly closed,
since it is the intersection of a family of semi-spaces Fa = {x:f(x) ;i oc}, and
each Fa is weakly closed.
For non-convex sets in an infinite-dimensionall.c.s., the weak closure is,
in general, larger than the ~-closure (Exercise 22).
66 LOCALL Y CONVEX TOPOLOGICAL VECTOR SPACES [Ch. II
10. COMPACT CONVEX SETS

For compact, convex subsets of a locally convex space, a number of strong
separation results can be established that will lead to the theorem of Krein-
Milman. The theorem asserts that each compact, non-empty convex set
contains extreme points (for the definition, see below) and is, in fact, the
convex closure of its set of extreme points. We begin with a sharpening of
Corollary 1 of (9.2); our proofs follow Bourbaki [7].
10.1
If C is a non-empty, compact, convex subset of a l.c.s. E, then for each closed
real hyperplane H in E there exist at least one and, at most, two supporting
hyperplanes of C parallel to H. Moreover, C is the intersection of the closed
semi-spaces that contain C and are determined by its hyperplanes of support.
Proof Let H = {x:f(x) = y} be any closed real hyperplane in E. Since the
restriction of f to C is a continuous real-valued function, there exist points
Xo E C and Xl E C such that f(x o) = IX = inff(C) and f(x l ) = f3 = supf(C).
It is clear that Ho = {x:f(x) = IX} and HI = {x:f(x) = f3} are (not necessarily
distinct) supporting hyperplanes of C, and that there exist no further support-
ing hyperplanes parallel to H. To prove the second assertion, let y rt= C; by
(9.2), there exists a closed real hyperplane H strictly separating {y} and C.
Evidently there exists a hyperplane HI parallel to H and supporting C, and
such that y is contained in that open semi-space determined by HI which
does not intersect C.
COROLLARY. If E is a l.c.s., E~ is the space of all continuous real linear forms
on E, and C is a compact, convex subset of E, then
C = () U- 1 [f(C)]:1 E E~}.
10.2
The convex hull of a finite family of compact, convex subsets of a Hausdorff
t.V.s. is compact.
Proof In fact, if Ai (i = 1, ... , n) are non-empty convex subsets of E, it is
n n
readily verified that their convex hull is A = {L A;ai: ai E Ai, Ai ;?: 0, L Ai =
i::::: 1 j;;;: 1
1
(i = 1, ... , n)}. Thus if Pc Rn is the compact set {(AI' ... , An): A; ~ 0, LA; = I},
A is the continuous image of the set P x [1iAi C R n x En; hence A is compact
if each Ai is (i = 1, ... , n).
We need now a generalization of the concept of a supporting hyperplane
for convex sets. If A is a convex subset of a I.c.s. E, a closed, real linear
manifold is said to support A if M n A # 0, and if every closed segment
SeA belongs to M whenever the corresponding open segment So intersects
M. In other words, M supports A if SeA and So n Mol 0 together imply
§10] COMPACT CONVEX SETS 67
S c: M, supposing in addition that M n A =1= 0. An extreme point of A isa

point Xo E A such that {x o} supports A.
Examples. Every vertex of a convex polyhedron A in R3 is an extreme
point of A; every straight line containing an edge of A is a supporting
manifold; every plane containing a face of A is a supporting hyperplane.
For an infinite-dimensional example of extreme points, see Exercise 29.
The following theorem, asserting the existence of an extreme point in every
hyperplane supporting a compact, convex set, is the final step toward the
Krein-Milman theorem:
10.3
If C is a compact, convex subset of a locally convex space, every closed real
hyperplane supporting C contains at least one extreme point of C.
Proof. Let H be a closed real hyperplane supporting C and denote by 9)1
the family of all closed real linear manifolds contained in H and supporting
C. 9)1 is inductively ordered under downward inclusion :::J ; for, if {Ma: a E A}
is a totally ordered subfamily, then M = naMa will be its lower bound in WI,
provided that M n C =1= 0. But the family {Ma n c: a E A}, again totally
ordered under ::::>, is a filter base consisting of closed subsets of C; since C
is compact, it follows that M n C = <naMa) n C = niMa n C) is not
empty. Hence by Zorn's lemma, there exists a minimal element Mo E 9)1.
If Mo = {x o }, then Xo is an extreme point of C contained in H; we shall
show that the assumption dim Mo ~ 1 contradicts the minimality of Mo.
Now Co = C n Mo is a compact, convex subset of the affine subspace M o,
and if the dimension of Mo is ~ 1, (10.1) implies that there exists a closed
hyperplane Ml in Mo such that Ml supports Co. We claim that Ml E 9)1, for
if S is a closed segment, S c: C, and the corresponding open segment So
intersects M I , then So intersects Mo and hence we have S c: Mo which, in
turn, implies S c: Co and therefore S c: MI' Thus MI E 9)1, which contradicts
the minimality of Mo in W1.
10.4
Theorem. (Krein-Milman). Every compact, convex subset of a locally
convex space is the closed, convex hull of its set of extreme points.
Proof. If C =1= 0 is convex and compact and B is the closed, convex hull
of the set of extreme points of C, then clearly B c: C. On the other hand,
if f =1= 0 is a continuous real linear form on E and f( C) = [a, P] there exist,
by (10.3), extreme points of Cin the supporting hyperplanesf-1(a) andf- 1 (p),
whence it follows that f(C) c:f(B). Thus f- 1 [f(C)] c:f-l[f(B)] for each
f E Eb, which implies C c: B by the corollary of (10.1).
The following supplement of (lOA) is due to Milman [1].
10.5
If A is a compact subset of a locally convex space such that the closed, convex
hull C of A is compact, then each extreme point of C is an element of A.
Proof. Let Xo be an extreme point of C, and V any closed convex O-neighbor-
hood. There exist pointsYI E A (i = 1, ... , n), with A c UI(Yi + V). Denote by
Wi the closed, convex hull of An (Yi + V). Since Wi (which are subsets of C)
are compact, by (10.2) so is the convex hull of their union which is, therefore,
n
identical with C. Hence Xo
n
= L AiWi,
i=l
where Wi E Wi' Ai ~ 0 (i = 1, ... , n)
and L Ai = 1. Xo being an extreme point of C, it follows that Xo = WI for
i=l
some i; hence Xo E WI C Yi + V, and since Yi E A, it follows that Xo E A + V.
Since V is an arbitrary member of a O-neighborhood base and A is closed,
it follows that Xo E A.
COROLLARY. If C is a compact, convex subset of a l.c.s. and cff is the set of it;
extreme points, then J is the minimal closed subset of C whose convex closure
equals C.
However, in general, cff is dense in C (Exercise 29).
EXERCISES
1. Let E be a vector space and let A #= 0 be a subset of E. The convex
n
hull (the convex, circled hull) of A consists of all finite sums L: AIXI such
n n 1
that Xi E A, Ai ~ 0 and L: Ai = 1 (such that L:IAil ~ 1); the convex,
r 1 1
circled hull A is the convex hull of the circled hull of A. If E is a
t.v.s., the convex hull of an open subset is open, and the closed, convex,
circled hull of A is the closure of A.r
2. A real-valued function ¢ on a convex subset of a vector space E
is convex if A, Il > 0 and A + Il = 1 imply that ¢(h + IlY) ~ A¢(x) +
1l¢(Y). If E is a t.v.s. and ¢ is a convex function on E, show that these
assertions are equival~nt:
(a) ¢ is continuous on E.
(b) ¢ is upper semi-continuous on E.
(c) There exists a non-empty, convex, open subset of E on which
¢ is bounded above.
Deduce from this that there exists a continuous· linear form f #- 0
on E if and only if there exists a non-constant convex function on E
which is upper semi-continuous. (Use the corollary of (3.1 ).)
3. Show that each convex, radial subset of a finite-dimensional vector
space E is a O-neighborhood for the unique separated topology on E
under which E is a t.v.s. (Chapter I, Section 3).
4. A real-valued function t/I on a vector space E is sublinear if it is
EXERCISES 69
convex and positive homogeneous (Le., if t/I(;,x) = ;,t/I(x) for all x E E

and all ;, ~ 0).
(a) Every sublinear function on Rn is continuous; examples are
x --. sup Xi> X --. £Llx1Iq]l/q (q ~ l)(i = 1, ... , n). (Use Exercises 2, 3.)
(b) If t/I is a sublinear function on R n such that t/I(x l , ... , x n) ~ 0
whenever Xi ~ 0 (i = 1, ... , n) and t/I(Xl> ... , x n) ;;;; 0 if all Xi;;;; 0, and if
Pi are continuous semi-norms on a t.V.S. E, then t/I(PI' ... ,Pn) is a con-
tinuous semi-norm on E.
(c) Assume that t/I is a sublinear function on Rn, as in (b), having the
additional property that XI ~ 0 (i = 1, ... , n) and t/I(x l , ... , xn) = 0
imply X; = 0 (i = 1, ... , n). Show that if (EI' PI) are n normed spaces,
then (Xl' ... , Xn) --. t/I[Pl(XI)' ... , Pn(Xn)] is a norm on TIiE; generating
the product topology.
5. Let L be a vector space over a complete, non-discrete valuated
field K (not necessarily R or C), and call (L, p) a normed space over K
if P satisfies conditions (i) through (iii) of Section 2. Find out to what
extent the results of Section 2 can be carried over to this more general
case. Show that the topologies generated on L by two such norms PI' P2
are identical if and only if there exist constants c, C > 0 such that
CPI (x) ;;;; P2(X) ;;;; CPI (x) for all X E L.
6. Let E be a vector space over R, let M be a subspace of E, and let
9 be a linear form on M such that g(x) ;;;; p(x) (x EM), where P is a
sublinear function on E. There exists a linear form f on E extending 9
and such that I(x) ;;;; p(x) for all x E E. (Observe that the linear forms
on E x R are the maps (x, t) --. hex) + IXI with h E E* and IX E R. Consider
the linear manifold Ho = {(x, I): g(x) - I = I} and the convex cone
C = {(x, I): p(x) ;;;; I} in E x R, and prove the existence of a hyperplane
He Ex R such that H::l Ho and H n C = 0: Cf. proof of (3.2).)
Show that this form ofthe Hahn-Banach theorem, (3.1), and (3.2) imply
each other.
7. Denote by E an infinite-dimensional vector space and by l: the
finest locally convex topology on E (Example following (6.2», Show
that E(l:) is a I.c.s. having these properties:
(a) Every linear map u on E(l:) into any I.c.s. F is continuous; hence
E(l:)' = E*, every subspace is closed, and every algebraic direct sum
decomposition of E is topological.
(b) A subset BeE is bounded if and only if it is contained in a
finite-dimensional subspace and bounded there; a subset of E is sequen-
tially closed if and only if its intersection with each finite-dimensional
subspace is closed.
(c) If E has a countable basis, a convex subset of E is closed if (and
only if) its intersection with each finite-dimensional subspace is closed.
(d) E(l:) is complete and not metrizable.
Show al~o that the property of carrying the finest locally convex top-
ology is inherited by quotients, by inductive limits and by subspaces, but
not by infinite products.
8. A family P of semi-norms on a vector space E is directed if it is
directed forthe usual order ;;;;, defined by "p(x);;;; q(x) for all x E E".
70 LOCALLY 'CONVEX TOPOLOGICAL VECTOR SPACES [Ch. II
(a) Let P be a family of semi-norms on E, and let Up,n = {x: p(x)

< n- 1 } for PEP, n EN. For {Up,n: PEP, n EN} to be a O-neighborhood
base of a locally convex topology on E, it is necessary and sufficient that
the family {cp: c > 0, pEP} of semi-norms be directed.
(b) If Po is a family of semi-norms generating the topology of the
I.c.s. E, the family P of the suprema of non-empty finite subsets of Po
is directed and generates the topology of E.
(c) Let M be a subspace of E. For a given semi-norm P on E, define
p(;£) = inf{p(x): x E x}(x E ElM); p is a semi-norm on ElM. If P is a
directed family of semi-norms on E, the family {p: pEP} generates on
--
ElM the quotient of the topology generated by the family P.
9. Let E(:!) = lim gapEp(:!p) be the projective limit of a family of I.c.s.
and suppose that gay = gap gpy whenever IX ~ P ~ y. Prove that E(:!)
0
is isomorphic with lim g{) .•E.(c:r.) (0, e E B) if B is a cofinal subset of A.

~-
A corresponding result holds for inductive limits if hya = hyp 0 hPa when-
ever IX ~ P~ y.
10. If E(:!) = lim gmnEn(:!n) is the projective limit of a sequence of
~-
I.c.s. such that gmp = gmn gnp and Em = gmiEn) whenever m ~ n ~ p,
0
then/,,(E) = En Un denoting the projection ofTIkEk onto En). The result

carries over to projective limits of countable families (use Exercise 9).
11. The I.c. direct sum of an infinite family of locally convex spaces,
each not reduced to {O}, is not metrizable. (Consider the completion of
the I.c. direct sum of a countable subfamily and use Baire's theorem.)
12. Show that if E(c:r) is the locally convex direct sum of a denum-
erable family of I.c.s., :! is identical with the topology defined in Chap-
ter I, Exercise 1.
13. If E is a metrizable I.c.s. which possesses a countable fundamental
system of bounded sets, then E is normable. (Observe that the com-
pletion of E is the union of countably many bounded subsets.) Give an
example of a non-metrizable I.c.s. that possesses a countable fundamen-
tal system of bounded sets. (Use (6.3).)
14. Let E be the vector space over R of all continuous real-valued
functions f on [0, 1] that vanish in a neighborhood (depending on f)
of t = 0, under the uniform topology. Show that D = {f nlf(n-1)1 ~ I,
n EN} is a barrel in E but not a neighborhood of 0, thus exhibiting a
normable (hence bornological) space which is not barreled.
15. Let E be a I.c.s., E its completion. (a) If E is barreled, E is bar-
reled; (b) if E isbornological, E is barreled. (For (b), use (8.5).)
16. Prove the following generalization of (8.1): Let L be a met-
rizable t. v.s. over a non-discrete, valuated field K (not necessarily R or
C); each circled subset of L that absorbs every null sequence in L is a
neighborhood of O.
17. Let E, F be I.c.s., where E is bornological, and let u be a linear
map of E into F. If for each null sequence {xn} c E, the sequence {u(xn)}
is bounded in F, then lim Xn = 0 implies lim u(xn) = O. Use this result to
derive a more general form of (8.3).
18. A I.c.s. E is bornological if and only if every linear map U on E
into any Banach space F such that u carries bounded sets onto bounded
EXERCISES 71
sets is continuous. (To establish the sufficiency of the condition, consider

the bomological space Eo associated with E, and show that the identity
map of E onto Eo is continuous by using (5.2) and (5.4), Corollary 2.)
19. Let {E«: 0( E A} be a family of l.c.s. over K.
(a) Assume that K~ is bomological. If u is a linear map on E = TI~«
into a l.c.s. F such that u carries bounded sets onto bounded sets and the
restriction of u to each of the subspaces E« = {x E E: xp = 0 for p =1= O(}
of E(O( E A) vanishes, then u = o. (Consider the restriction of u to the
bomological space TI«Kx« for each x = (x«) E E, and use (8.3).)
(b) Let F be a Banach space and let u be a linear map of E = TI«E«
into F transforming bounded sets into bounded sets. Show that u must
vanish on all but a finite number of the subspaces E«.
(c) Using (a) and (b), show that if K~ is bomological and u is a linear
map on TI~« into a Banach space F such that u maps bounded sets onto
bounded sets, there exists a finite subset H c: A such that TI~« = TI Ep
ED G and u(G) = {O}. peH
(d) Using (c) and Exercise 18, show that TI«E« is bomological if E«
(0( E A) and K~ are bomological. Deduce from this that the product of
a countable number of bomological spaces is bomological.
(e) If TI~", is bomological, then TI Ep = G is bomological for any
peB
subset B c: A. (Observe that G is isomorphic with a quotient space of
TI~,,·)
20. Let E be a vector space and let A and B be non-empty convex
subsets of E such that A n B = 0 and A has a core point Xo (i.e., a
point Xo such that A - Xo is radial). There exists a real hyperplane in E
separating A and B. (Note that a core point of A is an interior point of A
for the finest locally convex topology on E.)
21. Let A, B be non-empty, non-intersecting, convex subsets of a
vector space E. Show that there exist convex subsets C, D of E such that
A c: C, B c: D, C n D = 0 and CuD = E. (Use Zom's lemma.)
22. Let E = [2, the Hilbert space of square summable sequences
x = (Xl' X2' ... ) with IIxll = <Llxnll)t. Show that the weak closure ofthe
n
sphereS = {x: Ilxll = 1},whichisclosedinE,istheballB= {x: Ilxll ~ I}.
23. Show that in a finite-dimensional l.c.s., each pair of non-empty,
non-intersecting, closed, convex subsets is separated by a real hyper-
plane. (Represent one of the sets as the union of an increasing sequence
of compact convex subsets.) Show by an example in R~ that in this re-
sult, "separated" cannot be replaced by "strictly separated".
24. Let E be a l.c.s and let C be a closed, convex cone of vertex 0 in
E such that C =1= E.. Show that C is the intersection of the closed semi-
spaces containing it and determined by the supporting hyperplanes of C.
25. Let E be a l.c.s. over R, let C be a convex cone of vertex 0 in E,
and let C' be the subset of the dual E' whose elements are non-negative
on C. C' is a convex cone (of vertex 0) in E' which separates points in E
if and only if C n (- C) = {O}.
26. Show that in a finite-dimensionall.c.s. the convex hull of a com-
pact set is compact.
72 LOCALLY CONVEX TOPOLOGICAL VECTOR SPACES [Ch. 1/
'}.7. By establishing the following propositions show that in an infinite-

dimensional l.c.s., the convex hull of a compact subset is not neces-
sarily closed, and the closed, convex hull of a compact subset is not
necessarily compact.
(a) Denote by X the family of all real-valued continuous functions
on the unit interval [0, 1] and by E the product space R~; Eisa l.c.s.
For each fixed t E [0, 1], let CPt E E be the" evaluation map" f -+ f(t).
The set K1 = {cpt: t E [0, I]} is compact in E.
(b) The element cP E E given by the Riemann integral f -+ f(t) dt f:
is in the closure of the convex hull C of K 1, but cP ~ C; hence C is not
closed in E.
(c) Denote by F the smallest subspace of E that contains K 1 ; C is
closed in F and hence is the closed, convex hull of K1 in F, but is
not compact.
28. Let E be the Banach space of real null sequences x = (Xl> X2' ••• )
with Ilxll = sUPnlxnl. The unit ball B = {x: IIxll;;;:; I} in E is closed and
convex, but has no extreme points.
29. Let E' be the Hilbert space [2 over R. Denote by C the subset of
00
E determined by L (2nxn)2 ;;;:; 1. Then C is a convex and compact set,
n=1
and the closure of its subset tff of extreme points. (Let En be the sub-
space of E determined by Xk = 0 for k > n; the boundary points of the
ellipsoid C n En form a set tCn consisting entirely of extreme points
00
of C. Show that U tCn is dense in C. The example is due to Poulsen [1].)
n=1
30. A convex cone C of vertex 0 in a l.c.s. E has compact base if there
exists a real affine subspace N of E, 0 ~ N, such that N n C is compact,
non-empty, and C is the smallest cone of vertex 0 containing N n C.
A ray R = {A.Xo: A. ~ O}, 0 ¥= Xo E C, is extreme if x E R, Y E C, and
x - Y E C imply Y E R. Show that a convex cone with compact base is
closed, satisfies C n - C = {O}, and is the closed, convex hull of the set of
its extreme rays. (For more general results,in this direction, see Klee [5].)
PREREQUISITES

ences can be found.
A. SETS AND ORDER

1

74 LINEAR MAPPINGS [Ch. III
background of the theory of topological tensor products of which the ele-

ments are presented in Section 6 (barring the use of duality). This, in turn,
leads naturally to nuclear mappings and spaces, an important class of locally
convex spaces that are beyond the reach of Banach space theory. The com-
paratively recent results in this area are practically all due to Grothendieck
[13]; it is perhaps of interest to the expert how many of the basic results on
these spaces can be obtained without the use of duality or abstract measure
theory. The final section discusses the approximation problem, with some
emphasis on Banach spaces, and briefly the basis problem. It becomes ap-
parent here that duality is hardly dispensable, but the results on strong
duals and adjoints in (B)-spaces used here are elementary, so we have decided
to place this discussion before Chapter IV despite some technical inconve-
niences.
1. CONTINUOUS LINEAR MAPS AND TOPOLOGICAL HOMOMORPHISMS
If Land Mare t.v.s. over K and u is a linear map (an algebraic homo-
morphism) of L into M, then u is continuous if and only if u is continuous at
o E L; for if V is a given O-neighborhood in M, and U is a O-neighborhood in
L such that u(U) c V, then x - y E U implies u(x - y) = u(x) - u(y) E V.
Hence if u is continuous at 0, it is even uniformly continuous on L into M for
the respective uniformities (Chapter I, Section I). Thus if u is continuous on L
into M with M separated and complete, then u has a unique continuous
extension ii, with values in M, to any t.v.S. L of which L is a dense subspace
(in particular, to the completion L of L if L is separated); it is easy to see that
ii is linear. We supplement these simple facts by a statement in terms of semi-
norms.
1.1
Let the topologies of Land M be locally convex and let fY> be a family of
semi-norms generating the topology of L. A linear map u of L into M is con-
tinuous if and only if for each continuous semi-norm q on M, there exists a
finite subset {Pi: i = 1, ... , n} of fY> and a number c> 0 such that q[u(x)] ~
c sup; p;(x) for all x E L.
Proof The condition is necessary. Let V be the O-neighborhood {y: q(y)

~ I}, where q is a given continuous semi-norm on M. Since u is continuous
and [JJ generates the topology of L (Chapter II, Section 4), there exist 0-
neighborhoods U; = {x: p;(x) ~ 8;} (8; > 0, p; E fY>; i = 1, ... , n) such that
um;U;) c V. Hence, letting 8 = min; 8;, the relation sup;p;(x) ~ 8 implies
u(x) E V; thus q[u(x)] ~ l. Clearly, then, q[u(x)] ~ 8- 1 sup;p;(x) for all
x EL.
The condition is sufficient. If V is a given convex circled O-neighborhood
in M, its gauge function q is a continuous semi-norm on M. Thus if q[u(x)] ~
§1] CONTINUOUS LINEAR MAPS AND TOPOLOGICAL HOMOMORPHISMS 75
c SUPi p/(x), where c > o and Pi e {ljJ (i = 1, ... , n), it follows that u(U) c: Vfor
the O-neighborhood U = {x: cpJx) ~ 1, i = 1, ... , n} in L.
COROLLARY. If u is a linear map on a normed space (L, II 10 into a normed
space (M, II II), u is continuous if and only if II u(x) II ~ c IIx II for some c > 0 and
all x eL.
A continuous linear map on L into M, where Land Mare t.V.S. over K, is
called a topological homomorphism (or, briefly, homomorphism when no
confusion is likely to occur) if for each open subset G c: L, the image u(G) is
an open subset of u(L) (for the topology induced by M). Examples of topo-
logical homomorphisms are, for any subspaces Hand N of L, the canonical
(quotient) map c/>: L-+LIN and the canonical imbedding l/!: H -+L. With the
aid of these two mappings, every linear map u of L into M can be "canon-
ically" decomposed:
L -+ LIN -+ u(L) -+ M.
'" Uo '"
Here N = u-1(O) is the null space of u, and Uo is the algebraic isomorphism
which maps each equivalence class ~ of L mod N onto the common image
u(x) (x e x) of this class under u. Hence u = l/! 0 Uo 0 C/>, and we call the bi-
jective map Uo associated with u. We leave it to the reader to verify that u is
an open map if and only if Uo is open and that u is a continuous map if and
only if U o is continuous.
1.2
Let Land M be t.V.s. and let u be a linear map of L into M. These assertions
are eqUivalent:
(a) u is a topological homomorphism.
(b) For every neighborhood base U of 0 in L, u(U) is a neighborhood base
of 0 in u(L).
(c) The map Uo associated with u is an isomorphism.
Prool (a) => (b): Since u is open, every element of u(U) is a O-neighborhood,
and u(U) is a base at 0 e u(L), since u is continuous. (b) => (c): Since c/>(U) is a
O-neighborhood base in LIN, N = u-1(O), for any O-neighborhood base in L,
Uo has property (b) and is consequently an isomorphism. (c) => (a): Since C/>, uo, l/!
are all continuous and open, so is u = l/! 0 Uo 0 C/>, and hence is a topological
homomorphism.
1.3
Let u be a linear map on L whose range is a finite-dimensional Hausdorff t.v.S.
These assertions are equivalent:
(b) u- 1(O) is closed in L.
(c) u is a topological homomorphism.
Proof (a)=:> (b): Since u(L) is Hausdorff, {O} is closed and thus u-I(O)
is closed if u is continuous. (b) =:> (c): If u-I(O) is closed, Lju-I(O) is a Haus-
dorff t.v.s. of finite dimension, whence by (I, 3.4), U o is an isomorphism;
it follows from (1.2) that u is a topological homomorphism. (c) =:> (a) is clear.
COROLLARY. Every continuous linear form on a t.V.S. L is a topological
homomorphism.
This fact has been used implicitly in the proof of(H, 9.1).
2. BANACH'S HOMOMORPHISM THEOREM
It follows from (1.3) that every continuous linear map with finite-dimen-
sional separated range is a topological homomorphism; the question arises
for what, if any, larger class oft.v.s. it is true that a continuous linear mapping
is automatically open (hence a homomorphism). We shall see that this holds
for all mappings of one Frechet space onto another, and in certain more
general cases. For a deeper study of this subject, the reader is referred to
Chapter IV, Section 8. We first prove a classical result of Banach ([1], chap.
III, theor. 3) for which we need the following lemma:
Let L, M be metric t.v.s. whose respective metrics d, b are given by pseudo-
norms (Chapter I, Section 6): d(xI' x 2 ) = IXI - x 2 and b(Y1, h) = IY1 - hi·
1
We denote by Sr = {x EL: Ixi ~ r} and Sp = {y E M: IYI ~ p}, respectively,

the closed balls of center 0 and radius r, p.
LEMMA. Let L be complete and let u be a continuous linear map of L into
M satisfying
(P): For every r > 0, there exists p = per) > 0 such that u(Sr) :::::> Sp.
Then u(St):::::> Spfor each t > r.
Proof Let rand t, t > r > 0 be fixed and denote by {rn} a sequence of
positive real numbers such that r 1 = rand L rn
00
= t. Let {P.} be a null sequence

I
of numbers > 0 such that P1 = P and for each n E N, Pn satisfies u(SrJ:::::> Spn'
For each Y ESp, we must establish the existence of ZESt with u(z) = y.
We define inductively a sequence {x.: n = 0, 1, ... } such that for all n $;;: I:
(i) IX n - xn-11 ~ rn'

(ii) lu(xn) - yl ~ Pn+l'
°
Set Xo = and assume that Xl' X2' ... , Xk-l have been selected to satisfy
(i) and (ii) (k $;;: 1). By property (P), the set U(Xk-1 + SrJ is dense with respect
to U(Xk-l) + SPk' From (ii) we conclude that Y E U(X k- l ) + SPk; thus there
exists Xk satisfying IXk - Xk-ll ~ rk and IU(Xk) - yl ~ PHl'
00
Since ~>n converges, {xn} is a Cauchy sequence in the complete space Land
I
§2) BANACH'S HOMOMORPHISM THEOREM 77
thus converges to some z E L. Clearly, Izl ;;;:; t, and u(z) = y follows from the
continuity of u and (ii), since {Pn} was chosen to be a null sequence.
Let us point out that the results and, up to minor modifications, the proofs
of this section, with the exception of (2.2), are valid for topological vector
spaces L over an arbitrary, non-discrete valuated field K (Chapter I). Also the
following remark may not be amiss. A Baire space is, by definition, a topo-
logical space in which every non-empty open subset is not meager. This
implies that every t.v.s. Lover K which is non-meager (of second category)
in itself, is a Baire space. Otherwise, there would exist a meager, non-empty,
open subset of L, and hence a meager O-neighborhood U. Since L is a count-
able union of homothetic images of U (hence of meager subsets), we arrive at
a contradiction.
2.1
Theorem. Let L, M be complete, metfizable t.V.S. and let u be a con-
tinuous linear map of L with range dense in M. Then either u(L) is meager (of
first category) in M, or else u(L) = M and u is a topological homomorphism.
Proof. Suppose that u(L) is not meager in M. As in the preceding lemma, we
can assume the topologies of Land M to be generated by pseudonorms by
(1,6.1), and we continue to use the notation of the lemma. The family
{Sr: r > O} is a O-neighborhood base in L. For fixed r, let U = Sr' V = Sr/2;
00
then V + V c U and u(L) = U nu(V), since V is radial. Let us denote the

1
closure of a set A in u(L) by [A]-. Since, by assumption, u(L) is a Baire space,
there exists n EN such that [nu(V)r has an interior point; hence [u(V)r has
an interior point by (I, 1.1). Now
[u(V)r + [u(V)r c [u(V) + u(V)r = [u(V + V)r c [u(U)r;
thus [u(J:T)r is a O-neighborhood in u(L), since 0 is interior to

[u(V)r + [u00r. Hence there exists P > 0 such that u(L) n Sp c [u(U)r,
and the lemma above implies that u(L) n Sp c u(Sr+e) for every 8 > O. Thus.
{u(St): t> O} is a neighborhood base of 0 in u(L), whence by (1.2), u is a
topological homomorphism. Therefore Uo is an isomorphism of the space
Lju-1(O) which is complete by (I, 6.3), onto u(L), whence it follows that
u(L)=M.
COROLLARY 1. A continuous linear map u of a complete, metrizable I.v.s. L
into another such space, M, is a topological homomorphism if and only if u(L)
is closed in M.
Proof. The necessity of the condition is immediate, since u(L), being iso-
morphic with Lju-1(O), is complete and hence closed in M. Conversely, if
u(L) is closed in M, then it is complete and metrizable and hence can replace
Min (2.1).
COROLLARY 2. Let L be a complete, metrizable t.V.S. for both of the top-

ologies ~l and ~2' and suppose that ~l is finer than ~2' Then ~l and ~2 are
identical.
This is immediate from Corollary 1, since the identity map is continuous
from (L, ~l) onto (L, ~2)'
COROLLARY 3. If a complete, metrizable t.v.S. L is the direct algebraic sum
of two closed subspaces M and N, the sum is topological: L = M EB N.
Proof Since M and N are complete and metrizable, so is M x N; whence

it follows that the continuous mapping (Xl' X2) --+ Xl + X 2 of M x N onto L
is an isomorphism (Chapter I, Section 2).
With our present tools, we can extend Corollary 1 somewhat beyond the
metrizable case. The following extension is due to Dieudonne-Schwartz [1].
2.2
Let E be a locally convex space of type (LF) and let F be a locally convex
space of type (F) or (LF). Every continuous linear map u of E onto F is a topo-
logical homomorphism.
- -
Proof Let E = lim En be an (LF)-space and let F = lim Fn be an (LF)-
space; the case where F is a Frechet space can be formally subsumed under
the following proof by letting Fn = F(n EN). For all m, n EN, set Gm.n =
Em (') u-I(Fn); as a closed subspace of Em, Gm.n is complete and metrizable.
00
Since u(E) = F and u(Gm.n) = u(Em) (') Fm it follows that U u(Gm •n) = Fn
m=l
for each fixed n. Since Fn is a Baire space, there exists mo (depending on n)
such that u(Gmo,n) is non-meager in Fn; it follows from (2.1) that u(G mo .n) = Fn.
If U is any O-neighborhood in E, U (') Gmo .n is a O-neighborhood in Gmo .n by
(II, 6.4); hence u( U (') Gmo •n) is a O-neighborhood in Fn and a fortiori u( U) (') Fn
is a neighborhood of 0 in Fn. Since this holds for all n EN, it follows that u(U)
is a O-neighborhood in F, and hence u is a homomorphism.
Another direct consequence of Banach's theorem (2.1) is the following
frequently used result, called the closed graph theorem.
2.3
Theorem. If Land M are complete, metrizable t.v.s., a linear mapping
of L into M is continuous if and only if its graph is closed in L x M.
Proof Recall that the graph of u is the subset G = {(x, u(x)): X E L} of

L x M. Clearly, if u is continuous, then G is closed in the productspaceL x M.
Conversely, if G is closed, it is (since u is linear) a complete, metrizable
subspace of L x M. The mapping (x, u(x)) --+ x of G onto L is biunivocal,
linear, and continuous, and hence an isomorphism by (2.1). It follows that
§3] SPACES OF LINEAR MAPPINGS 79
x --+ (x, u(x»is continuous, whence u is continuous by definition of the

product topology on L x M.
3. SPACES OF LINEAR MAPPINGS

Let P be a vector space over K, let T be a set, and let 6 be a family of subsets
of T directed under set-theoretic inclusion c. (Whenever the letter 6 is used
in the following, it will denote a family of sets with this property.) A subfamily
6 1 of 6 is fundamental (with respect to 6) if it is cofinal with 6 under in-
clusion (that is, if each member of 6 is contained in some member of 6 1).
Consider the vector space pT, product of T (more precisely, of card T) copies
of P; as a set, pT is the collection of all mappings of T into P. Suppose in
addition that Pis a t.v.s., and let )D be a neighborhood base of 0 in P. When S
runs through 6, V through )D, the family
M(S, V) = {I:/(S) c V} (*)
is a O-neighborhood base in pT for a unique translation-invariant topology,
called the topology of uniform convergence on the sets S E 6, or, briefly, the
6-topology. For if V3 c V1 n V z and Sl v Sz C S3' then M(S3, V 3) is
contained in M(Sl' V 1 ) n M(Sz, V z ); hence the sets (*) form a filter base in
pT which has the additional property that M(S, V) + M(S, V) c M(S, U)
whenever V + V c U. In (*), the family 6 can evidently be replaced by any
fundamental subfamily, and likewise we note that the 6-topology does not
depend on the particular choice of the neighborhood base )D of 0 in P. Now
we have to settle the question under what conditions pT, or a subspace G of
pT, is a t.v.s. for a given 6-topology.
3.1
A (vector) subspace G c pT is a t.v.s. under an 6-topology if and only if
lor each lEG and S E 6,/(S) is bounded in P.
Proof The sets M(S, V) n G (S E 6, V E )D) form a base of the O-neighbor-
hood filter of the topology ~ induced on G by the 6-topology; in what
follows we shall denote these sets again by M(S, V) with the understanding
that now M(S, V) = {IE G:/(S) c V}. For (G,~) to be a t.v.s., by (I, 1.2) it
is necessary and sufficient that the O-neighborhood filter have a base consist-
ing of radial and circled sets, since from the remark made above, it follows
that condition (a) of (I, 1.2) is satisfied. Since M(S, AV) = AM(S, V) for
each A "# 0, M(S, V) is circled if V is circled; thus let )D consist of circled sets.
Now suppose that for each S E 6,/ E G, the set/(S) is bounded in P. Then
for given f, S and V, there exists A > 0 such that I(S) c AV and hence I E
M(S, AV) = AM(S, V); it follows that M(S, V) is radial. Conversely, if~ is a
vector space topology on G, each M(S, V) is necessarily radial; thus for given
f, S, and V, there exists A > 0 with I E AM(S, V) = M(S, AV). Hence I(S) c
AV, which shows I(S) to be bounded in P.
3.2
Let F be a locally convex space, let T be a topological space, and let 6 be a
family of subsets of T whose union is dense. If G is a subspace of FT whose
elements are continuous on T and bounded on each S E 6, then G is a locally
convex space under the 6-topology ..
Proof. If ID is a O-neighborhood ,base in F consisting of convex sets, then

each M(S, V) is convex; hence by (3.1) the 6-topology is locally convex.
There remains to show that the 6-topology is Hausdorff. Letf E G andf:F 0;
sincefis continuous and U{S: S E 6} is dense, there exists to E So E 6 such
thatf(to} :F O. Since F is a Hausdorff space (Chapter II, Section 4), we have
f(to} ¢ Vo for a suitable Vo E ID. It follows thatf ¢ M(So, Vo), and hence the
6-topology is a Hausdorff topology on G.
If T is itself a t.v.s., and each S E 6 is bounded, and G is a vector space of
continuous linear maps into F, then the assumption that eachf(S) be bounded
is automatically satisfied by (I, 5.4), and for the conclusion of (3.2) to hold,
it suffices that the linear hull of U{S: S E 6} be dense. It is convenient to
have a term for this: A subset of a t.v.s. L is total in L if its linear hull is dense
in L. With this notation, we obtain the following corollary of (3.2).
COROLLARY. Let E be a t.v.s., let F be a l.c.s., and let 6 be afamity of bounded
subsets of E whose union is total in E. Then the vector space 2(E; F) of all
continuous linear mappings of E into F is a locally convex space under the
6-topology.
Endowed with an 6-topology, the space 2(E, F) is sometimes denoted by
2 riE, F). It is no restriction of generality to suppose E separated, for if Eo is
the Hausdorfft.v.s. associated with E (Chapter I, Section 2), then 2(E, F) is
algebraically isomorphic with 2(Eo, F) (Exercise 5). Moreover, if E is Haus-
dorff and F is complete, 2(E, F) is algebraically isomorphic with 2(£, F),
where £ denotes the completion of E (Exercise 5). We shall see later that every
locally convex topology on a vector space E is an 6-topology, where 6 is
a suitable family of subsets of the algebraic dual E* (Chapter IV, Section 1).
Examples
1. Let T be a given set, let F be any t. v.s., and let 6 be the family
of all finite subsets of T. Under the 6-topology, FT is isomorphic with
the topological product of T copies of F.
2. Let T be a Hausdorff topological space, let F be a I.c.s., and let
6 be the family of all compact subsets of T. Under the 6-topology
(called the topology of compact convergence), the space of all continuous
functions on T into F is a I.c.s.
3. Let E be a I.c.s. over K with dual E'; the weak dual (E', u(E', E)
is the space 2(E, Ko) under the 6-topology, with 6 the family of all
finite subsets of E.
4. Let E, F be I.c.s. The following 6-topologies are of special import-
ance on !l'(E, F):
§3] SPACES OF LINEAR MAPPINGS 81
a. The topology of simple (or pointwise) convergence: 6 the family

of all finite subsets of E.
b. The topology of convex, circled, compact convergence: 6 the
family of all convex, circled, compact subsets of E.
c. The topology of precompact convergence: 6 the family of all
precompact subsets of E.
d.. The topology of bounded convergence: 6 the family of all bounded
subsets of E.
The families 6 in the preceding examples have the property that the union
of their members is E; such a family 6 is said to cover E.
A family 6 =F { 0} of bounded subsets of ~ I.c.s. E is called saturated
if (1) it contains arbitrary subsets of each of its members, (2) it contains
all scalar multiples of each of its members, and (3) it contains the closed,
convex, circled hull.of the union of each finite subfamily. Thus the families
6 of Ex,ample 4c and 4d are saturated, and 4a and 4b are not saturated
unless E = {OJ. Since the family of all bounded subsets of E is saturated
and since the intersection of any non-empty collection of saturated families is
saturated, a given family 6 of bounded sets in E determines a smallest
saturated family ~ containing it; ~ is called the saturated hull of 6. E and
F being locally convex, it is clear that for each family 6 of bounded subsets
of E, the 6-topology and the ~-topology are identical on !l'(E, F). (Cf.
Exercise 7.)
To supplement the corollary of (3.2), we note that if {p..: ex E A} is a family
of semi-norms generating the topology of F, the family of semi-norms
u -+ Ps,..(u) = sup P.. [u(x)]
. xeS
(S E 6, ex
A) generates the 6-topology on !l'(E, F). In particular, if E and
E
Fare normed spaces, the norm
u -+ Ilull = sup{lIu(x)II: IIxll ~ 1}
generates the topology of bounded convergence on 2(E, F) (cf. Chapter II,
Section 2).
Returning to a more general setting, let E, F be Hausdorff t.v.s. over K,
let 6 be a (directed) family of bounded subsets of E, and let f£(E, F) be the
vector space over K of all continuous linear maps on E into F. We turn our
attention to the subsets of f£(E, F) that are bounded for the'6-topology.
3.3
Let H be a subset 0/ f£(E, F). The/ollowing assertions are equivalent:
(a) H is bounded/or the 6-topology.
(b) For each O-neighborhood V in F, n u- (V) absorbs every S
ueH
1 E 6.
(c) For each S E 6, U u(S) is bounded in F.
ueH
Proof. (a) ~ (b): We can assume V to be circled. If H is bounded; it is

absorbed by each M(S, V); hence u(S) c: A.V for all u E H and some A. > 0,
which implies SeA n u -l( V). (b) => (c): If S E 6 and a circled O-neighbor-
given, then SeA n u-I(V) implies u(S) c AV for all
NeH
hood V in F are
NeH
U E H; hence U u(S) is bounded in F. (c) => (a): For given S and V, the
NeH
existence of A such that u(S) c AV for all u E H implies He AM(S, V);
hence H is bounded for the 6-topology.
A subset of !l'(E, F) is simply bounded if it is bounded for the topology of
simple convergence (Examples I and 4a above). It is important to know
conditions under which simply bounded subsets are bounded for finer 6-
topologies on !l'(E, F).
3.4
Let E, F be I.c.s. and let 6 be the family of all convex, circled subsets of E
that are bounded and complete. Each simply bounded subset of !l'(E, F) is
bounded for the 6-topology.
Proof If H is simply bounded in !l'(E, F) and V is a closed, convex, circled
O-neighborhood in F, then D = n
u-I(V) is a closed, convex, circled subset
NeH
of E which is radial by (3.3)(b), and hence a barrel; thus by (II, 8.5) D absorbs
every S E 6, which implies, again by (3.3), that H is bounded for the 6-
topology.
COROLLARY. If E, Fare l.c.s. and E is quasi-complete, then the respective
families of bounded subsets of !l'(E, F) are identical for all 6-topologies such
that 6 is a family of bounded sets covering E.
Proof When E is quasi-complete, the family 6 of (3.4) is a fundamental
system of bounded sets in E; in other words, the 6-topology of (3.4) is the
topology of bounded convergence. The assertion is now immediate.
4. EQUICONTINUITY. THE PRINCIPLE OF UNIFORM BOUNDEDNESS

AND THE BANACH-STEINHAUS THEOREM
If T is a topological space and F is a uniform space, a set HeFT is equi-

continuous at to E T if for each vicinity (entourage) N c F x F, there exists
a neighborhood U(to) of to such that (f(t),!(to)] EN whenever t E U(t o) and
f E H; His equicontinuous if it is equicontinuous at each t E T. If Tis a uniform
space as well and if for each vicinity N in F there exists a vicinity Min T such
that (tl' t 2 ) EM implies [f(tI),!(t 2 )] EN for all f E H, then H is called
uniformlyequicontinuous. It is at once clear that if T = E is a t.v.s., and if F
is a t.v.s., a set H of linear mappings of E into F is uniformly equicontinuous
(for the unique translation-invariant uniformities associated with the topol-
ogies of E and F, respectively (Chapter J, Section 1» if and only if H is
equicontinuous at 0 E E; that is, if and only if for each O-neighborhood Yin F,
there exists a O-neighborhood U in E such that u(U) c V whenever u E H.
§4] EQUICONTINUITY 83
Of course an equicontinuous set of linear mappings of E into F is a subset of

2(E, F).
As before, we shall denote by 2(E, F) the space of all continuous linear
maps of E into F, E and Fbeing Hausdorfft.v.s. over the same field K, and by
21$(E, F) the same space under an 6-topology with 6 a,(directed) family of
bounded subsets of E whose union is total in E. Finally, L(E, F) will denote
the vector space of all linear maps (continuous or not) of E into F.
The proof of the following statement is quite similar to. the proof of (3.3)
and will be omitted.
4.1
Let H be a subset of 2(E, F). Thefollowing assertions are equivalent:
(a) His equicontinuous.
(b) For each O-neighborhood V in F, n u- (V) is a O-neighborhood in E.
ueH
1
(c) For each O-neighborhood V in F, there exists a O-neighborhood U in E

such that U u(U) c V.
ueH
(4.1)(b) implies (3.3)(b), hence:

COROLLARY. Each equicontinuous subset of 2(E, F) is bounded for every
6-topology.
The converse of this corollary is not valid (Exercise 10), but there are im-
portant instances in which even a simply bounded subset of 2(E, F) is neces-
sarily equicontinuous.
4.2
Theorem. Let E, F be l.c.s. such that E is barreled, or let E, F be t.v.s.
such that E is a Baire space. Every simply bounded subset H of 2(E, F) is
equicontinuous.
Proof We give the proof first for the case where E is barreled and F is any
l.c.s. If V is any closed, convex, circled O-neighborhood in F, W = n
ueH
u-I(V)
is a closed, convex, circled subset of E which, by condition (b) of (3.3),
absorbs finite subsets in E; thus W is a barrel and hence a O-neighborhood in
E, whence His equicontinuous by (4.1)(b).
If E is a Baire space, F is any t.v.s., and V is a given O-neighborhood in F,
select a closed, circled O-neighborhood VI such that VI + VI C V. By (3.3)(b),
W=
ueH
nU-1(V1 ) is a closed, circled subset of E which is radial, whence
00
E = U n W. Since E is a Baire space, n W must have an interior point for at

I
least one n; hence W must have an interior point, whence U = W + W is a
neighborhood of 0 in E. Now u(W) C VI and hence u(U) c V for all U E H,
which proves H to be equicontinuous.
An immediate consequence is the following classical result due to Banach-

Steinhaus [1], and known as the principle of uniform boundedness:
COROLLARY. Let E be a normed space, let F be a normed space, and let
H be a subset of !l'(E, F) such that sup{ Ilu(x) II: u E H} is finite for every
x E M, where M is not meager in E. Then sup{lIull: u E H} isfinite.
Proof. The linear hull EM of M, which is clearly dense in E and a Baire
space, since M is not meager in E, has the property that Ho is simply bounded
in !l'(EM' F), where Ho is the set obtained by restricting all u E H to EM'
Hence by (4.2), Ho is equicontinuous and thus norm bounded in !l'(EM' F).
Now since the unit ball of EM is dense in the unit ball of E, the mapping
u -+ Uo (uo the restriction of u E !l'(E, F) to EM) is a norm isomorphism of
!l'(E, F) into !l'(EM, F); hence H is norm bounded as asserted.
Before we can prove the Banach-Steinhaus theorem (see (4.6) ;below) in
appropriate generality, we have to gather further information on equicon-
tinuous sets which will also be needed in Chapter IV. We note first that the
subspace L(E, F) of FE is closed in FE for the topology of simple convergence
(which is the topology of the product of E copies of F(Section 3, Example 1»:
Since Fis assumed to be Hausdorft" and sincef-+f(x) is continuous on FE into
F for each x E E, it follows that the set
M(x, y, )., J.l) = {f E FE:f().x + J.ly) - ).f(x) - J.lf(y) = O}
is closed for each fixed quadruple (x, y, )., J.l), and L(E, F) = n M(x, y, )., J.l)
where (x, y, )., J.l) ranges over E x E x K x K.
4.3
If H c: !l'(E, F) is equicontinuous and H1 is the closure of H in FE for the
topology of Simple convergence, then H1 c: !l'(E, F) and H1 is equicontinuous.
Proof. If U1 E Hlo then U1 E L(E, F) by the preceding remark. Since H is
equicontinuous, there exists aO-neighborhood U in E such that for all u E H,
u(U) c: V, where V is a given O-neighborhood in F which can, without re-
striction of generality, be assumed closed. From the continuity of f -+ f(x)
on FE into F, we conclude that U1(X) E V for all U1 E H1 and x E U. Thus H1
is equicontinuou!) in !l'(E, F).
Combining this result with Tychonov's theorem on products of compact
spaces, we obtain the following well-known result, known as the theorem of
Alaoglu-Bourbaki.
COROLLARY. Let E be a t.v.s. with dual E'; every equicontinuous subset of E' is
relatively compact for (1(E', E).
Proof. The weak topology (1(E', E) is the topology of simple convergence
on E' = !l'(E, Ko) and hence induced by the product topology of Kg. By the
Tychonov theorem, a subset Hc: Kg is relatively compact if (and only if) for
each x E E, {f(X):fE H} is relatively compact in Ko. Now if H c: E' is
§4] EQUICONTINUITY 8S
equicontinuous, there exists a O-neighborhood U in £ such that Iu(x) I ~ 1 for

all u E H and x E U; thus if Xo E £ is given, there exists A. > 0 such that
Axo E U, whence Iu(xo) I ~ A. -1 for all u E H. Thus the closure H1 of H in
Kg is compact; but since H1 c E' by (4.3), H1 agrees with the closure H of H
in (E', q(E', E)), so H is weakly compact, which proves the assertion.
4..4
If F is quasi-complete and 6 covers E, every closed, equicontinuous set is
complete in ~ 1$(£' F).
Proof. Let H c It'1$(E, F) be closed and equicontinuous. If ij is a Cauchy
filter on H, it is a fortiori a Cauchy filter on H for the uniformity associated
with the topology of simple convergence; hence for each x E £ the sets
{<D(x): <D E ij} are bounded and a base of a Cauchy filter in F (for H is bounded
and u --+ u(x) is linear and continuous). Since F is quasi-complete, this filter
base converges to an element U1(X) E F and by (4.3), x --+ u1(x) is in It'(E, F).
Moreover, ij being a Cauchy filter for the 6-topology, there exists <D E ij such
that u(x) - vex) E V for all u E <D, v E <D and XES, where S E 6 and the
O-neighborhood V in F can be preassigned. Hence if V is chosen to be closed,
it follows that u(x) - U1(X) E Vfor all u E <D and all XES, implying that U1 =
lim ij for the 6-topology.
COROLLARY. If E, F satisfy the assumptions of(4.2) and F is quasi-complete,
then It' aCE, F) is quasi-complete for every 6-topology such that 6 covers E.
For another condition guaranteeing quasi-completeness or completeness
of It'1$(E, F) for certain 6-topologies, see Exercise 8.
4.5
Let H be an equicontinuous subset of 2(E, F). The restrictions to H of the
following topologies are identical:
1. The topology of simple convergence on a total subset of E.
2. The topology of simple convergence (on E).
3. The topology of precompact convergence.
Proof. Each of the three topologies is finer than the preceding one. The
result will be established if we can show that when restricted to H, topology 1
is finer than topology 3. Let A be a total subset of E. We have to show that
for each Uo E H, O-neighborhood V in F, and precompact set SeE, there
exist a finite subset So c A and a O-neighborhood Vo in F such that
[uo + M(So, Yo)] n H c Uo + M(S, V),
where the notation is that employed in Section 3. Select a O-neighborhood
W in F such that W + W + W + W + We V, and a circled O-neighborhood
UinEsuch that w(U) c Wwhenever w E H. S (#= 0) being precompact, there
exist elements Yi E S (i = 1, ... , m)for which S c Ui (Yi + U). Since the linear
hull of A is dense in E, there exist (supposing that E f= {O}) elements xij E A
m
and scalars Aij (i = 1, ... , m; j = 1, ... , n) such that Yi E L Aijxij + U. It
j=1
follows that
Choose a circled O-neighborhood Vo in F with L (Aij Yo) c W, and denote

i, j
by So the finite set {xii i = 1, ... , m;j = 1, ... , n}. If v E M(So, Yo), then
v(xij) E Vo for all i, j and
m n
v(S) C U L (AijVo) + v(U)+ v(U)c W +v(U)+v(U).
i=1j=1
Now let Uo E Hand w E H () [u o + M(So, Vo)]; then w = Uo + v, where
v E M(So, Yo). Since v = w - Uo, v(U) c w(U) + uo(U) c W + W, since U is
circled. Thus v(S) c V, w = Uo + V E Uo + M(S, V), and the proof is complete.
The preceding results make it possible to prove the following theorem,
called the theorem of Banach-Steinhaus, in substantial generality (cf. Bour-
baki [8], chap. III). For briefness we call a filter ~ on a t.v.s. E bounded if
~ contains a bounded subset of E.
4.6
Theorem. Let E, F be l.c.s. such that E is barreled; or let E, F be t.v.s.

such that E is a Baire space. If~ is afilter in 2'(E, F), boundedfor the topology
of simple convergence and which converges pointwise to a mapping u1 E FE, then
U1 E 2'(E, F) and ~ converges uniformly to U 1 on every precompact subset of E.
Proof. Let <1> be an element of ~ bounded for the topology of simple con-
vergence; by (4.2), <1> is equicontinuous. If <1>1 denotes the closure of <1> in FE,
then U1 E <1>1 by hypothesis and by (4.3), <1>1 is contained in 2'(E, F) and equi-
continuous. Since by (4.5) the topologies of SImple and precompact conver-
gence agree on <1>1' the theorem is proved.
The theorem applies, in particular, to a sequence {un} such that for each
x E E, {unCx)} is a Cauchy sequence in F, provided that F is quasi-complete.
More generally, it applies when ~ is a filter (not necessarily bounded) with
countable base (Exercise 11). The following corollary is an extended version
of the classical form of the theorem (cf. Banach [1], chap. V, theor. 3-5.)
COROLLARY. Let E, F be Banach spaces, and let Me E be a subset not
meager in E. If {un} c 2'(E, F) is a sequence such that {un(x)} is a Cauchy
sequence in F for every x E M, then {un} converges to an element u E 2'(E, F)
uniformly on each compact subset of E.
§5] BILINEAR MAPPINGS 87
Proof Let EM be the linear hull of M ; EM is a non-meager subspace of E,

and hence a Baire space. Denoting by Un the restriction of Un to EM (n EN),
it follows from (4.6) that lim un(x) = u(x) for all x E EM' where U E !l'(EM, F).
By the corollary of (4.2), the sequence {un} is norm bounded in !l'(E, F) and
hence equicontinuous; thus if we denote by U the unique continuous extension
of U to E, the set H = {un: n E N} u {u} is still equicontinuous. It follows now
from (4.5) that lim Un = U uniformly on every precompact (or equivalently,
since E is complete, on every compact) subset of E.
We conclude this section by giving conditions under which an equicontin-
uous set He !l'(E, F) is metrizable and separable. Recall that a metric space
possesses a countable base of open sets if and only if it is separable.
4.7
If H c !l'(E, F) is equicontinuous, if E is separable and ifF is metrizable, then
the restriction to H of the topology of simple convergence is metrizable. If, in
addition, F is separable, then H is separable for this topology.
Proof In view of (4.5) it is sufficient to prove the theorem for the topology,
restricted to H, of simple convergence on a total subset of E. Since E is sepa-
rable, there exists an at most countable subset A = {xn} of E which is linearly
independent and total in E. Take {Vm } to be a countable O-neighborhood base
in F, and let Sn = {xk:k ~ n}. Clearly, the sets M(Sn> V m ), (n, m) E N x N,
form a neighborhood base of 0 in !l'(E, F) (notation as in Section 3) for the
topology of simple convergence on A; hence this topology is metrizable by
(I, 6.1) and so is its restriction to H.
In view of the remark preceding (4.7), the second assertion will be proved
when we show that the 6~topology, 6 = {Sn: n EN}, on L(E, F) (which is, in
general, not a Hausdorff topology) possesses a countable base of open sets.
To this end, extend A to a vector space basis B of E, let Y = {Yn: n E N} be a
dense subset of F, and define Q to be the set of elements u E L(E, F) such that
u(z) = 0 for all z E B except for finitely many Xv E A (v = 1, ... , n) for which
u(xv) = Yn., where {YnJ is any non-empty, finite subset of Y. Q is clearly
countable, and dense in L(E, F) for the 6-topology: If u E L(E, F) and u +
M(Sn' V m) is a given neighborhood of u, then we choose U o E Q such that uolX)
E u(x) + Vm for each x E Sn (which is possible, since Y is dense in F), whence
it follows that U o E u + M(S", V m). Thus if we denote by M(S", Vm)O the inter-
ior of M(S", Vm ) in L(E, F), it is immediate that the countable family {u +
M(Sn' Vm)o: U E Q, (n, m) EN x N} is a base of open sets for the 6-topology.
5. BILINEAR MAPPINGS
Let E, F, G be vector spaces over K; a mappingf of Ex F into G is called
bilinear if for each x E E and each Y E F, the partial mappings fx: Y --+ f(x, y)
and fy: x --+ f(x, y) are linear. If E, F, G are t.v.s., it is not difficult to prove
that a bilinear map f is continuous if and only if f is continuous at (0, 0)
(Exercise 16); accordingly, a family B of bilinear maps is equicontinuous if and

only if B is equicontinuous at (0, 0). A bilinear mapf is said to be separately con-
tinuous if all partial mapsfx andJ;, are continuous; that is, iffx E 2(F, G) for
all x E E andfy E 2(E, G) for all y E F. Accordingly, a family B of bilinear
maps of E x F into G is separately equicontinuous iffor each x E E and each y E F
the families {fx:f E B} and {j~: fEB} are equicontinuous. Finally, if G = Ko,
then a bilinear map of E x F into G is called a bilinear form on E x F.
The following important result is a special case of a theorem due to
Bourbaki [8] (chap. III, §3, theor. 3):
5.1
Theorem. Let E, F be metrizable and let G be any t.V.S. If E is a Baire
space or if E is barreled and G is locally convex, then every separately equi"
continuous family B of bilinear mappings of E x F into Gis equicontinuous.
Proof. In view of the identity (f E B)
f(x, y) - f(x o, Yo) = f(x - x o, y - Yo) + f(x - x o, Yo) + f(xo, y - Yo)
and the separate equicontinuity of B, it is sufficient to prove the equicontin-
uity of Bat (0,0). Denote by {Un}, {Vn} decreasing sequences that constitute
a O-neighborhood base in E, F respectively; {Un X Vn } is a O-neighborhood
base in E x F. Now if B were not equicontinuous at (0, 0), there would
exist a O-neighborhood Wo in G and sequences {x n }, {Yn}, with Xn E Un>
Yn E Vn (n EN) such that for all n, /,,(xn> Yn) r/= W o, where {j~} is a sequence
suitably chosen from B. We shall show that this is impossible. Since for each
fixed x E E, the family {Ix: fEB} is equicontinuous, by the corollary of (4.1)
it is bounded for the topology of compact convergence on 2(F, G); thus
{fx({Yn}):f E B} is bounded in G by (3.3), since {Yn}, being a null sequence in
F, is relatively compact. Therefore by (3.3)(c) the family {x --+ fix, Yn): n EN}
of linear maps is simply bounded in 2(E, G) and hence is equicontinuous
by (4.2); it follows that /,,(U, Yn) c: Wo (n EN) for a suitable O-neighbor-
hood U in E, which conflicts with the assumption that /,,(xn' Yn) r/= W o (n EN),
since {xn } is a null sequence in E.
COROLLARY 1. Under the assumptions made on E, F, Gin (5.1), every sep-
arately continuous bilinear mapping on Ex F into G is continuous.
COROLLARY 2. In addition to the assumptions made on E, F, G in (5.1),
suppose that F is a Baire space or (if G is locally convex) that F is barreled.
If B is afamity of separately continuous bilinear maps of E x F into G such that
{f(x, y):f E B} is bounded in G for each (x, y) E E x F, then B is equicontinuous.
The proof of Corollary I is obtained by applying (5.1) to the family B
consisting of a single elementf; the proof of Corollary 2 is also easy, since
by (4.2) the assumptions imply that B is separately equicontinuous.
§S] BILINEAR MAPPINGS 89
As simple examples show (Exercise 17), in general, a separately continuous

bilinear map is not continuous; it has thus proved fruitful to introduce an
intermediate concept that is closely related to the notion of an 6-topology.
Let E, F, G be t.v.s., let 6 be a family of bounded subsets of E, and letfbe a
bilinear map on E x F into G. f is called 6-hypocontinuous iff is separately
continuous and if, for each S E 6 and each O-neighborhood W in G, there
exists a O-neighborhood Yin F such thatf(S x V) c W. By (4.1) it amounts
to the same to require that for each S E 6 the family {fx: XES} be equicon-
tinuous. The 1:-hypocontinuity of f is analogously defined if 1: is a family
of bounded subsets of F: f is 1:-hypocontinuous if, for each TE1:, {fy: YET}
is equicontinuous, and iffis separately continuous. Finally, a bilinear map is
(6, Z)-hypocontinuous if it is both 6-hypocontinuous and Z-hypocontinuous.
Note that separate continuity emerges as a particular case when 6 and Z are
the families of all finite subsets of E and F, respectively.
5.2
If F is barreled and G is locally convex (or if F is a Baire space), every
separately continuous bilinear map f of E x F into G is 93-hypocontinuous,
where 93 is the family of all bounded subsets of E.
Proof The separate continuity off is obviously equivalent to the assertion
that the linear map x -. fx of E into L(F, G) maps E into f.e(F, G) and is con-
tinuous for the topology of simple convergence on f.e(F, G). Thus if BeE
is bounded, {fx: x E B} is simply bounded in f.e(F, G) and hence is equicon-
tinuous by (4.2); this establishes the proposition.
5.3
Let 6, Z be families of bounded subsets of E, F, respectively, and let f be a
bilinear map of Ex F into G, where E, F, G are t.V.S. Iffis 6-hypocontinuous,
thenfis continuous on S x F for each S E 6; iffis (6, Z)-hypocontinuous, then
f is uniformly continuous on S x T for each S E 6 and T E Z.
Proof The first assertion is an immediate consequence of the 6-hypo-
continuity off and the identity
f(x, y) - f(xo, Yo) = f(x, y - Yo) + f(x - xo, Yo)
to be applied for x, Xo E Sand y, Yo E F. To prove the second assertion,
allow x, x to be variable in Sand y, y to be variable in T. Sincefis (6, Z)-
hypocontinuous, for a given O-neighborhood Win G, there exist O-neighbor-
hoods U, Yin E, F, respectively, such thatf(S x V) c Wandf(U x T) c W.
If x - x E U, Y - Y E V, it follows that
f(x,y) - f(x, y) =f(x,y -y) + f(x - x, y) E W + W;
hencefis uniformly continuous on S x T.
The preceding result is useful for the extension of ($, Z)-hypocontinuous

bilinear maps.
Let E, E 1 , F, F1 be t.v.s. such that E is a dense subspace of E1 and F is a
dense subspace of Fl' Suppose that $ is a family of bounded subsets of E with
the property that $1 covers E 1 , where $1 denotes the family of the closures,
taken in E 1 , of all S E $; suppose further that Z is a family of bounded sub-
sets of F, such that the corresponding family Zl of closures covers F l ;
finally, let G be a quasi-complete Hausdorff t.V.s. Under these assumptions,
the following extension theorem holds:
5,4
Every ($, Z)-hypocontinuous bilinear mapping of Ex F into G has a unique

extension to El x Fl (and into G) which is bilinear and($l' Zl)-hypocontinuous.
Proof. As before (Section 3), we suppose $, Z to be directed under" c "
(which is, incidentally, no restriction of generality); then so are the families
{S x T} and {Sl x T 1 }. Since S x T is dense in the uniform space Sl x T l ,
the restriction fS.T of the bilinear map f to S x T has by (5.3) a unique
(uniformly) continuous extension lSI.TI to Sl X Tl with values in G (since
}(S x 1) is bounded and G quasi-complete). Since the family {Sl x T l :
S E $, T E Z} is directed and covers E1 x F l , it follows that in their totality,
the extensions lSI.TI define an extension 1 off to E1 x Fl' This argument also
shows that a possible extension off with the desired properties is necessarily
unique; it remains to show that 1 is bilinear and ($1' Zl)-hypocontinuous.
Let x EEl be given; there exists S1 with x E S1' The map rPx: y -+ lex, y)
(y E F) is an element of .P(F, G) by (4.3) since, f being $-hypocontinuous,
{Ix: XES} is equicontinuous in !l'(F, G). Now rPx has a unique continuous
extension to F1 with values in G, which must necessarily agree with lx:
ji -+ lex, ji), since G is separated (uniqueness of limits). Henceeachlx (and by
symmetry, each ly) is linear and continuous, which showsfto be bilinear and
separately continuous.
Sincefis $-hypocontinuous, for each S E $ there exists a O-neighborhood
Vin F such thatf(S x V) c W, W being a given O-neighborhood in G which
can be assumed closed. Denoting by V 1 the closure of V in F1 (V1 is a 0-
neighborhood in F l , cf. (I, 1.5», it follows from the separate continuity of 1
that l(S x V1) c Wand, repeating the argument, that j(S1 x VI) C W. Thus
lis $l-hypocontinuous and (by symmetry) Zrhypocontinuous, which com-
pletes the proof.
We remark that if E and G are locally convex, an $-hypocontinuous bi-
linear map of E x F into G is also ~-hypocontinuous, where ~ denotes the
saturated hull of $ (Section 3), with a corresponding statement holding under
($, Z)-hypocontinuity.
The set of all bilinear mappings of E x F into G is a vector space B(E, F; G)
which is a subspace of GEXF ; the subspaces of B(E, F; G) (supposing E, F, G
§5] BILINEAR MAPPINGS 91
to be t.v.s.) consisting of all separately continuous and all continuous bilinear

maps, respectively, will be denoted by !B(E, F; G) and by f!J(E, F; G). The
corresponding spaces of bilinear forms will be denoted by B(E, F), !B(E, F),
and f!J(E, F).
If 6 and ~ are families of bounded subsets of E and F, respectively, and
if D is a subspace of B(E, F; G), we consider the topology of 6 x ~-conver
gence on D (Section 3), that is, the topology of uniform convergence on the
sets S x T, where S E 6 and T E~. We recall that D is a t.V.s. under this
topology if (and only if) for all S E 6, T E ~ and fED, f(S x T) is bounded
in G; this is, in particular, always the case when D c f!J(E, F; G) (cf. Exercise
16). If the preceding condition is satisfied and G is locally convex, then the
6 x ~-topology is locally convex. We leave it to the reader to verify that if G
is separated and D c !B(E, F; G), the 6 x ~-topology is a Hausdorff top-
ology whenever 6 and ~ are total families (that is, families whose union is a
total subset of E or F, respectively).
The following is a general condition under which !B(E, F; G) is a I.c.s. for
an 6 x ~-topology.
5.5
Let E, F, G be locally convex spaces; denote by 6 a total, saturatedfamily of

bounded subsets of E such that the closure of each S E 6 is complete, and denote
by ~ a total family of bounded subsets of F. Then !B(E, F; G) is a locally convex
space under the 6 x ~-topology.
Proof. We have to show that for each f E !B(E, F; G) and all sets S E 6,
T E ~,f( Sx T) is bounded in G; since 6 is saturated, we can suppose S to be
closed, convex, and circled. Now since T is bounded in F and since (by the
separate continuity off) the linear map y --+ fy is continuous on Finto !feE, G)
when 2(E, G) carries the topology of simple convergence, the set {fy: YET}
is simply bounded in !feE, G). Thus if W is a closed, convex, circled O-neigh-
n
borhood in G, the set U = {f; 1( W): YET} is closed, convex, circled, and by
(3.3) radial; hence U is. a barrel in E. It follows from (II, 8.5) that U absorbs
S, whence we have f(S x T) c .:tW for a suitable scalar .:t. Since W was an
arbitrary element of a O-neighborhoood base in G,/(S x T) is bounded.
The conditions of the preceding proposition are, in particular, satisfied if
E and F are replaced by the weak duals (Chapter II, Section 5) E; =
(E', (f(E', E» and F; of two arbitrary I.c.s. E and F, and if 6 and ~aretaken to
be the families of all equicontinuous subsets of E' and F', respectively; for 6
and ~ are saturated families of bounded sets whose closed members are
compact (hence complete) in E; and F;, respectively, by the corollary of
(4.3). This 6 x ~-topology is called the topology of bi-equicontinuous con-
vergence (Grothendieck [13]), and under this topology, !B(E;, F;; G) is a
locally convex space which will be denoted by !BlE;, F;; G).
6. TOPOLOGICAL TENSOR PRODUCTS

Let E, F be vector spaces over K and let B(E, F) be the vector space of all
bilinear forms on E x F. For each pair (x, y) E E x F, the mappingf -+ f(x, y)
is a linear form on B(E, F), and hence an element UX " of the algebraic dual
B(E, F)*. It. is easily seen that the mapping x: (x, y) -+ ux ,,, of E x F into
B(E, F)* is bilinear. Thelinearhullofx(E x F) inB(E,F)*is denoted by E® F
and is called the tensor product of E and F; X is called the canonical bilinear
map of E x F into E ® F. The element UX ,1 of E ® F will be denoted by
x ® Y so that each element of E ® F is a finite sum LA;(x; ® y;) (the sum over
the empty set being 0). We shall also find it convenient to write A ® B =
X(A xB) for arbitrary subsets A c: E, B c: F, although this usage is inconsist-
ent with the notation E ® F. Ambiguity can be avoided if, only for subspaces
M c: E, N c: F, the symbol M ® N denotes the linear hull of X(M x N)
rather than the set X(M x N) itself.
OneverifieswithoutdifficultytherulesA(x ® y) = (Ax) ® y = x ® (AY)(A EK),
(Xl +x2)®y=xl®y+x2®y,andx®(Yl +Y2)=X®Yl +x®Y2· Hence
each element U E E ® F is of the form U = LX; ® Yl' Obviously, the
representation of u is not unique, but it can be assumed that both" sets
{x;} and {Yi} are linearly independent sets of r(~ 0) elements. The number r is
uniquely determined by u and called the rank of u; it is the minimal number of
summands by means of which u can be represented (Exercise 18).
One of the principal advantages of tensor products lies in the fact that they
permit us to consider vector spaces of bilinear (more generally, of multilinear)
maps as vector spaces of linear mappings. We recall this more precisely:
6.1
Let E, F be vector spaces over K and let X be the canonical bilinear map of
E x F into E ® F. For any vector space Gover K, the mapping u -+ u X is an
0
isomorphism of L(E ® F, G) onto B(E, F; G).
Proof It is clear that u -+ u 0 X = f is a linear map of L(E ® F, G) into

B(E, F; G), which is one-to-one, since f = 0 implies u(x ® y) = f(x, y) = 0
for all x E E and Y E F, hence u = O. It remains to show that the map is onto
B(E, F; G). For any f E B(E, F; G) define u(Lx; ® Yi) = ~f(x;, Yi); it is clear
that the definition is consistent, that u is linear on E ® F into G, and that
f= uO X.
COROLLARY. The algebraic dual of E ® F can be identified with B(E, F);

under this identification, each vector space of linear forms on E ® F is a vector
space of bilinear forms on E x F, and conversely.
In particular, E* ® F* can be identified with a space of bilinear forms on
Ex Fby means of(x* ® y*)(x, y) = x*(x)y*(y), and hence with a subspace of
(E ® F)*; it is readily seen that E* ® F* separates points in E ® F.
16] TOPOLOGICAL TENSOR PRODUCTS 93
In order to define useful topologies on E ® F when E, Fare t.v.s., we

restrict our attention to locally convex spaces E, F and locally convex top-
ologies on E ® F. Consider the family !T of all locally convex topologies on
E ® F for which the canonical bilinear map (E, F being I.c.s.) on E x F into
E ® F is continuous: The upper bound Zp of!T (Chapter II, Section 5) is a
locally convex topology, called the projective (tensor product) topology on
E ® F. It is immediate that when U, ID are O-neighborhood, bases in E, F,
respectively, the family of convex, circled hulls {r(U ® V): U e U, Ve ID} is a
neighborhood base of 0 for Zp; thus the projective topology is the finest
locally convex topology on E ® F for which the canonical bilinear map is
continuous. We shall see at once that Zp is always a Hausdorff topology.
6.2
Let E, F, G be locally convex spaces and provide E ® F with the projective
topology. Then the isomorphism u ~ u 0 X of(6.1) maps the space of continuous
linear mappings !l'(E ® F, G) onto the space of continuous bilinear mappings
!fI(E, F; G).
Proof It is clear that the continuity of u implies that of u X, since X is
0
continuous. Conversely, if W is a convex, circled O-neighborhood in G and

f= u X is continuous, then f-l(W) contains a O-neighborhood U x V in
0
Ex F. It follows that u- 1(W) contains U® V. Since u- 1 (W)is convex and

r(
circled, it contains U ® V), which proves the continuity of u.
COROLLARY. The dual of E ® F for the projective topology can be identified
with the space PlJ(E, F) of all continuous bilinear forms on E x F. Under this
identification, the equicontinuous subsets of (E ® F)' are the equicontinuous sets
of bilinear forms on E x F.
This corollary implies that Zp is necessarily Hausdorff, for evidently
E' ® F' c f4(E, F); thus if we show that E' ® F' separates points in E ® F,
it follows that (l(E ® F, PlJ(E, F» and a fortiori Zp is a Hausdorff topology.
r
Now if u e E ® F is of rank r ~ 1, say, u = LXi ® Yi> then {xil and {y;} are
i=1
linearly independent, whence by (II, 4.2), Corollary 1, there exist linear forms
f1 e E' and g1 e F' such thatf1(x i) = (ju and g1(Yi) = (ju (i = 1, .'" r), and it
follows that!t ®gl(U) = Lj~(xi)g1(Yi) = 1. For a description ofZp by semi-
norms, we need the following result:
6.3
Let p, q be semi-norms on E, F respectively, such that p is the gauge of
U c E, and qis the gauge of V c F. Then the semi-norm on E ® F,
u ~ r(u) = inf {L;;P(X;)q(Yi): u = LXI ® Yi},
is the gauge ofr(U ® V) and has the property that rex ® y) = p(x)q(y) jar all
xeE,yeF.
Proof. It is immediate that r is a semi-norm on E ® F; let M 0 =

{u: r(u) < I}, Ml = {u: r(u) ~ I}. To prove that r is the gauge off(U® n,
it suffices to show that Mo c f(U ® V) c MI' If u E f(U ® V), then u =
~)i(Xi ® Yi), where Xi E U, Yi E V for all i and LIAil ~ 1. Now u = LXi ® Yb
where Xi = AiXb whence r(u) ~ LP(Xi)q(Yi) = LIAilp(Xi)q(Yi) ~ 1. On the other
hand, ifu E M o, then u = LXi ® Yi' where LP(Xi)q(y;) < 1. Thus there exist real
numbers 8i > 0 such that LJ.Li < 1, where J.Li = [P(Xi) + 8 iHq(Yi) + 8;] for all i.
Set Xi = X;/[P(Xi) + 8;] and Yi = y;/[q(Yi) + 8i]; then Xi E U, Yi E V, and hence
u = LJ.L;(x i ® Yi) E f(U ® V).
To prove the second assertion, let Xo E E, Yo E F be given; we conclude from
(II, 3.2) that there exist linear forms f E E*, g E F* such that f(xo) = p(x o),
g(yo) = q(yo), and If(x) I ~ p(x), Ig(Y)1 ~ q(y) for all X E E, Y E F. [Define f on
the subspace generated by Xo by f(Ax o) = Ap(XO) and extend to E.] It is
immediate that for the linear form f® g on E ® F and u = LXi ® Yb we
have If®g(u)1 ~LP(Xi)q(Yi)' whence If®g(u)1 ~ r(u); hence p(xo)q(yo) ~
r(xo ® Yo). Since clearly r(xo ® Yo) ~ p(xo)q(yo), the proof is complete.
The semi-norm r is called the tensor product of the semi-norms p and q, and
is denoted by p ® q. It is not difficult to prove that p ® q is a norm on E ® F
if and only if p and q are norms on E and F, respectively (Exercise 20). A
family P of semi-norms on E is directed if, for each pair PI' P2 E P, there exists
P3 E P such that sup (Pl' P2) ~ P3; if P is directed, the sets Up,n = {x E E:
p(x) ~ n -l} (p E P, n EN) form a neighborhood base of 0 for a locally convex
topology on E (Chapter II, Exercise 8). Thus we obtain this corollary of (6.3):
COROLLARY. Let E, F be locally convex spaces and let P and Q be directed

families of semi-norms generating the topologies of E and F, respectively.
The projective topology on E ® F is generated by the directed family
{p ®q: (p, q) EP X Q}.
In particular, if E and Fare normed spaces, then the tensor product of the
respective norms generates the projective topology on E ® F (Exercise 21).
If E, F are any l.c.s., then (E ® F, :l:p) is a l.c.s. as we have seen above;
hence by (I, 1.5) it can be imbedded in a complete l.c.s. which is unique (to
within isomorphism) and will be denoted by E ~ F. It results from the corol-
lary of (6.3) that if E, Fare metrizable, then E ® F is an (F)-space. It is one of
the fundamental results (also due to Grothendieck [13]) of the theory to have
an explicit representation of E ® F, when E, Fare metrizable l.c.s. (For the
definition of an absolutely convergent series in a t.v.S. E, see Exercise 23.)
6.4
Theorem. Let E, F be metrizable l.c.s.; each element u E E ® F is the
sum of an absolutely convergent series,
00
u = LAixi@Yi,
i=l
where LilA;! < + 00, and {x;}, {y;} are null sequences in E, F, respectively.
§6] TOPOLOGICAL TENSOR PRODUCTS 95
PTOOf. (The following simple proof is due to A. Pietsch.) Let {Pn}, {qn} be
increasing sequences of semi-norms generating the topologies of E, F re-
spectively, and denote by Tn the semi-norm Pn ® qn (n EN); the sequence {Tn}
generates the projective topology of E ® F and is increasing. Denote by
rn(n EN) the continuous extension of Tn to E ®F. .
Given U E E 0 F, there exists a sequence {un} in E ® F such that
it
i'n(U - Un) <n- 2 2-(n+I). Let L AiX, ® Yi be any representation of U lO and set
i= 1
Vn = Un + 1 - Un for all n EN.
We have
We conclude from (6.3) that there exists a representation

in + I
vn = L
i~i"+l
AiXi®Yi
such that Pn(Xi) ~ n-l, qn(x;) ~ n- 1 whenever in < i ~ in+l' and such that
in+1
I IAil ~rn.
i=i l1 +1
+ L vn = L AiXi ® Yi'
00 00
Therefore, we have U = Ul where the sequences

1 1
{xJ, {Y;}, and {A;} have the desired properties, and the proof is complete.
We shall now consider a general example of a projective tensor product.
Let (X, ~, fl) be a measure space (Chapter II, Section 2, Example 2' so that fl
is a positive measure on X, and Ll(fl) the Banach space of (equivalence classes
modulo fl-null functions of) real-valued fl-summable functions on X, with
11/11 = Sill dfl. Let E be any Banach space over R and let SE be the vector
space over R of all E-valued simple functions; that is, functions ¢ of the
n
form t -+ L t/I i(t)X;, where t/I i are the characteristic functions of n sets Si E :r
i= 1
such that fl(Sj) < + 00, and Xi are arbitrary elements of E. It is clear that
¢ -+ SII¢ I dfl is a semi-norm P on SE; now L1(fl) is defined to be the com-
pletion of the Hausdorff (hence normed) space (SE, p)/p- 1 (O). The space
L1(fl) is called the space of (classes of) E-valued fl-summable functions.
We show that the Banach space L1(fl) is norm-isomorphic with Ll(fl) ® E.
There exists a natural imbedding u -> ii of Ll(fl) ® E into L1(fl) such that for
U = Lli ® Xi' ii is the class containing the function t -+ Lh(t)x j ; evidently
u -+ ii is linear and maps S ® E onto SE' where S denotes the subspace of
L1(fl) whose elements contain a simple function, and where SE = SElp- 1 (O).
Denote by T the tensor product of the norms of Ll(fl) and E, respectively;
since
holds for all representations Igj ® Yj of a fixed element u, it follows that

p(U) ;;;; r(u). On the other hand, if u E S ® E, we can choose a representation
u = I'" i ® Xi such that the characteristic functions", i have disjoint carriers
S" which implies r(u) ;;;; I II'" i 1IIIxi II = p(u). Thus u ~ u is a norm isomorphism
of S ® E onto SE and the assertion follows from the fact that S ® E is dense in
L 1 (Jl) ®E, since S is dense in L 1(Jl).
In the preceding considerations, it is not essential that E be a Banach (or
even normable) space. If E is any I.c.s. with P a family of semi-norms genera-
ting the topology of E, we define SE as before and a locally convex topology
on SE by means of the semi-norms <p ~ Jp[<P(t)] dll(t)(P E P); the completion
of the associated Hausdorff t.v.s. then serves to define L1(Jl), and we prove,
as before, that u ~ uis an ismorphism of S ® E (under the projective topology)
onto SE' Hence:
6.5
The natural imbedding of L 1(1l) ® E into Ll(ll) induces an isomorphism of
L (1l)® E onto L1(1l) which is norm-preserving if E is a Banach space.
1
We have seen above that the projective topology l:p on E ® F is the finest
I.c. topology for which the canonical bilinear map is continuous. Another
topology of importance on E ® F is the inductive (tensor product) topology
l:;, defined to be the finest I.c. topology on E ® F for which the canonical
bilinear map is separately continuous. l:. is an inductive topology in the sense
of Chapter II, Section 6; in analogy to (6.2) above, we show that for every
I.c.s. G, the isomorphism of (6.1) carries the space of l:;-continuous linear
maps into G to the space of all separately continuous bilinear maps on E x F
into G. In particular, the dual of (E ® F, l:;) is the space !B(E, F) (Exercise
22). We shall not be further concerned with l:;, for which we refer the reader
to Grothendieck [13] as well as for other topologies on E ® F whose defini-
tion is based on the (6, l:)-hypocontinuity (Section 5) of the canonical bi-
linear map X. Let us point out that under the assumptions of (5.1), l:p and l:i
agree on E ® F.
A topology on E ® F of considerably greater importance than l:i is the top-
ology l:e of bi-equicontinuous convergence; viewing E ® F as a space of linear
maps on E' ® F' by virtue of X ® y(x' ® Y') = x'(x)y'(Y), l:e is the topology
of uniform convergence on the sets S ® T, where S, T are arbitrary equi-
continuous subsets of E', F', respectively. l:e can be equally characterized
as the topology induced on (the subspace) E ® F by !Be(E;, F;) which is a
I.c.s. (see end of Section 5). The completion of (E ® F, l:e) will be denoted by
E ®F. It is not difficult to see that l:e is coarser than l:p on E ® F; for a
successful study of this topology, we need a number of results on duality
(Chapter IV). For the moment, we mention only that if E, F are complete
I.e.s., then !Be(E;, F;) is complete, whence in this case, E® F can be identified
with the closure of E ® Fin !Be(E;, F;).
§7] NUCLEAR MAPPINGS AND SPACES 97
7. NUCLEAR MAPPINGS AND SPACES

If E is a vecto:J:' space over K and V is a convex, circled, and radial subset of
E, then {n -1 V: n E N} is a O-neighborhood base for a locally convex topology
Z" on E. The Hausdorff t.V.S. associated with (E, Zv) is the quotient space
(E, Zv)/p-l(O), where p is the gauge of V; this quotient space is normable
by the norm ~ -+ II~II = p(x), where x E~. We shall denote by E" the normed
space (E/p-1(0), II II) just introduced, and by Ev its completion, which is a
Banach space. If E is a I.c.s. and V is a convex, circled neighborhood of 0, the
topology of the quotient space E/p -1(0) is (in general, strictly) finer than the
topology of Ev. Thus the quotient map (called the canonical map) is continu-
ous on E into E,,; this map will be denoted by cPv.
Dually, if E is a I.c.s. and B =F 0 a convex, circled, and bounded subset of E,
00
then E1 = U nB is a (not necessarily closed) subspace of E. The gauge func-
n=1
tion p B of B in E1 is quickly seen to be a norm on E1 ; the normed space
(E1 , p B) will henceforth be denoted by E B' It is immediate that the imbedding
map I/IB: EB -+ E (~gain called canonical) is continuous. Moreover, if B is
complete in E, then EB is a Banach space by (I, 1.6). We finally note that no
confusion can arise if V = B is a convex, circled subset of E which is radial
and bounded, for in this case the spaces Ev and E B are identical.
If U, V are convex, circled, and radial subsets of E with respective gauge
functions p, q and such that U c: V, then p-1(0) c: q-1(O) and each equiva-
lence class ~ mod p -1(0) is contained in a unique equivalence class p mod
q-1(O); ~ -+ p is a linear map cPv,u, which is called the canonical map of Eu
onto Ev. Since cPv,u is clearly continuous (in fact, of norm ~ 1), it has a
unique continuous extension on Eu into 2 v , which is again called canonical,
and also denoted by cPv,u.
Likewise, if Band C are convex, circled, and bounded sets of a I.c.s. E
such that 0 =F B c: C, then EB c: Ee and the canonical imbedding I/Ie,B: EB -+ Ee
is continuous. Finally, if U, V, B, C are as before and cPu, cP", I/IB' I/Ie are the
canonical maps E -+ E u, E -+ E", E B-+ E and Ee -+ E, we have the relations
cPv = cP",u 0 cPu and I/Ie = I/IC,B 0 I/IB'
The two methods of constructing auxiliary normed spaces were syste-
matically employed by Grothendieck [13] and will be extremely useful in
what follows. We have used these methods before in Chapter II (proof of
(II, 5.4) and the discussion preceding (II, 8.4».
Let E, F be I.c.s. and let E' be the dual of E. Each element vEE' ® F
defines a linear map U E !l'(E, F) by virtue of
r
X -+ u(x) = LJi(X)YI
i=1
r
if v = Lfi ® Yi' and v -+ U is even an (algebraic) isomorphism of E' ® F into
i=1
98 LINEAR MAPPINGS [Ch. /II
2(E, F) (Exercise 18). The mappings u E 2(E, F), which originate in this
fashion from an element vEE' ® F, are called continuous maps of finite
rank; the rank r of u is defined to be the rank of v (Section 6). The mappings of
finite rank are very special cases of compact linear maps on E into F: A linear
map u on E into F is compact if, for a suitable O-neighborhood U in E, u( U) is
a relatively compact subset of F.
Suppose now that E, F are Banach spaces, and let E' be the Banach space
which is the strong dual of E (Chapter II, Section 2). Then the imbedding
v -+ u is continuous for the projective topology on E' ® F and the norm
topology (the topology of bounded convergence) on 2(E, F) : If vEE' ® F
then
r r
lIull = sup Ilu(x)1I
Ilxll~1
~ sup
IIxll.~1
Li=1Ifi(x)IIIYili ~ i=1
L 11J:lllly;l1
r
for all representations v = Ifi ® Yi; hence Ilull ~ rev), where the norm r is
i= 1
the tensor product ofthe respective norms of E and F (cf. (6.3) and its corol-
lary). Since 2(E, F) is complete under the norm topology by the corollary
of (4.4), the imbedding v -+ u has a continuous extension 't' to E' ~ F, with
values in 2(E, F). The linear maps contained in the range of 't' are called
nuclear; that is, u E 2(E, F) is nuclear if u = 't'(v) for some v E E'® F. (It is
known that 't' is not necessarily one-to-one, cf. Chap. IV, Exerc. 30.)
The definition of a nuclear map generalizes to arbitrary l.c.s. E, F as fol-
lows. A linear map u on E into F is bounded if for a suitable O-neighborhood U
in E, u(U) is a bounded subset of F(for example, every compact map is boun-
ded); every bounded map is continuous. A bounded map can be decomposed
as follows: Let U be a convex, circled, O-neighborhood in E such that u( U) c
B, where B is convex, circled, and bounded in F; then u = '" B 0 U o 0 ¢u,
where Uo is the map in 2(Eu, F B) induced by' u. If, in addition, FB is complete,
then Uo has a continuous extension Uo E 2(tu, F B ) for which u = '" B 0 iio 0 ¢u'
The definition is now this:
A linear map u of a I.c.s. E into another I.c.s. F is nuclear if there exists a
convex, circled O-neighborhood U in E such that u( U) c B, where B is bounded
with FB complete, and such that the induced mapping iio is nuclear on tu
into F B •
It follows at once that every continuous linear map of finite rank is nuclear;
plOreover, if u is nuclear in 2(E, F), there exists a O-neighborhood U in E
and a bounded, convex, circled subset of F for which F B is complete, such that
u is the uniform limit on U of a sequence of maps of finite rank in 2(E, F B)'
Hence for every 6-topology on 2(E, F), the nuclear maps are contained in
the closure of E' ® F (the latter being viewed as a subspace of 2(E, F».
With the aid of Theorem (6.4), we obtain the following explicit characteriza-
tion of nuclear maps.
7.1
A linear map u E !l'(E, F) is nuclear if and only if it is of the form
00
x --+ u(x) = L Anln(X)Yn,

n= 1
00
where L IAnl < + 00, Un} is an equicontinuous sequence in E', and {Yn} is a
n= 1
sequence contained in a convex, circled, and bounded subset B of F for which
FB is complete.
Proof. The condition is necessary. For, if u is nuclear, then u = 1/1 B 0 ito 0 <Pu,
where ito is nuclear in !{(£u, F B ), U being a suitable O-neighborhood in E,
and B being a suitable bounded subset of F for which F B is complete. Hence
ito originates from an element v of [Eu]' ~ FB , which is, by (6.4), of the
00 00
form v = L Anhn ® Yn with LIAnl < + 00 and where {h n} and {Yn} are null
n= 1 1
sequences in [Eu]' and FB , respectively. Define a sequence Un} of linear forms
on E by setting!.. = hn <Pu. Since {h n} is a bounded sequence in [Eu]', the
0
sequence {In} is uniformly bounded on U and hence is equicontinuous. It is

clear now that the mapping u = I/IB 0 r(v) 0 <Pu is of the form indicated above.
The condition is sufficient. For if u is as indicated in the proposition, let
U = {x E E: If..(x) I ~ I, n EN}; U is convex and circled and is a O-neighbor-
hood in E by the equicontinuity of {f..}. Defining hn (n E N) by!.. = hn 0 <Pu
on Eu and subsequent extension to £u, we obtain IlhnII ~ I for all n; evidentlY,
00 00
ito is the map ~ -+ L Anhn<~)Yn' Since ~]Anl IIhnII llYn II converges, the series
00 n=1 1
LAnhn ® Yn is absolutely convergent in [Eu]' ~ FB by (6.3) and its corollary,
1
and hence defines an element v E [Eu)' ~ F B ; clearly, ito = r(v), whence u is
nuclear.
REMARK. If u is of the form indicated in (7.1), we shall find it conven-
00
ient to write u = L Anln ® Yn, keeping in mind that u is not, properly

1
speaking, an element of a topological tensor product. It follows then
from the first part of the proof that for nuclear u, there exists a rep-
00
resentation u = LAn f" ® Yn such that {!..} is a sequence converging to
1
o uniformly on a suitable O-neighborhood U of E, and {Yn} converges
to 0 in a suitable Banach space F B; finally, (An) E [1.
COROLLARY I. Every nuclear map is compact.
00
Proof Let u = :LAn!n ® Yn and let U = {x E E: I!..(x) I ~ I, n EN}. In view

1
of the preceding remark, it can be assumed that {Yn} is a null sequence in a
00
suitable space FB and, in addition, that LIAnl ~ 1. It follows that the image
1
u( U) of the O-neighborhood U is contained in the closed, convex, circled hull

C of the null sequence {Yn} in F B; since {Yn} is relatively compact in FB and
FB is complete, C is compact in FB (cf. (I, 5.2) and (II, 4.3», and hence a for-
tiori compact in F by the continuity of F B -+ F.
COROLLARY 2. Let E, F, G, H be l.c.s., let u E .P(E, F), let WE 'p(G, H),
and let v be a nuclear map on F into G. Then v 0 u and w 0 v (and hence w 0 v 0 u)
are nuclear maps.
Proof It is evident from (7.1) that v u is nuclear. By Corollary 1, there
0
exists a convex, circled O-neighborhood V in F such that v( V) = B is 'compact

in G. Thus Bl = w(B) is compact in H, hence HBI is complete. It is now clear
that w 0 v is nuclear in .P(F, H).
COROLLARY 3. If U E .P(E, F) is nuclear, then u has a unique extension
ii E .p(E, F), where E is the completion of E, and ii is nuclear.
Proof The first of the stated properties is shared by u with all compact
maps on E into F. In fact, if U is a O-neighborhood in E such that u( U) c C,
where C is compact, then since u is uniformly continuous, its restriction to U
has a unique continuous extension to U (the closure of U in E) with values in
C, since C is complete. It is immediately clear that this extension is the re-
striction to U of a linear map ii of E into F which is compact, hence continuous;
that ii is nuclear is a direct consequence of the definition of a nuclear map,
or of (7.1).
We are now ready to define a nuclear space. A locally convex space E is
nuclear if there exists a base mof convex, circled O-neighborhoods in E such
that for each V E m, the canonical mapping E -+ Ev is nuclear.
It is at once clear from this definition and (7.1) that a l.c.s. E is nuclear if
and only if its completion E is nuclear. The space Kg (d any cardinal) is a
first example of a nuclear space; in fact, for any convex, circled O-neighbor-
hood V, the space Ev = Ev is of finite dimension; thus E -+ Ev is of finite rank
and hence nuclear. Further and more interesting examples will be given
below and in Section 9, Chapter IV. Let us note, however, that a normed
space E cannot be nuclear unless it is of finite dimension; for if V is a convex,
circled O-neighborhood which is bounded, then E -+ Ev is a topological
automorphism; hence if E -+ Ev is a nuclear map, it is compact by Corollary 1
above. Thus (I, 3.6) implies that E is finite-dimensional. We shall have use for
the following alternative characterizations of nuclear spaces.
7.2
Let E be a I.c.s. The following assertions are equivalent:
(a) E is nuclear.
(b) Every continuous linear map of E into any Banach space is nuclear.
(c) Each convex, circled O-neighborhood U inE contains another, V, such

that the canonical map Ev -+ Eu is nuclear.
Proof. (a) ~ (b): Let F be any Banach space and u E .P(E, F). There exists
a convex, circled O-neighborhood V in E such that ¢v: E -+ Ev is nuclear, and
such that u(V) is bounded in F. Since ¢v(E) = E v , u determines a unique
v E ft'(E v , F) such that u = v 0 ¢v, and it follows from Corollary 2 of (7.1)
that u is nuclear. (b) ~ (c) : Let U be any convex, circled O-neighborhood in E.
By assumption, the canonical map E -+ Eu is nuclear, and hence of the form
¢u = ~)'nf" ® Yn as described in (7.1), Set V = Un {x: Ifn(x) I ~ 1, n EN},
then V c: U is convex, circled, and a O-neighborhood by the equicontinuity
of the sequence Un}. Now eachfn induces a continuous linear form (of norm
~ 1) on Ev. Denote by hn its continuous extension to Ev: It is now trivial that
the canonical map ¢u,v: Ev -+ Eu is given by LAnhn ® Yn' and hence nuclear
by (7.1).
(c) ~ (a): If U is a given convex, circled O-neighborhood in E, there exists
another, V, such that ¢u,v is nuclear. Since ¢u = ¢u,v 0 ¢v, it follows from
Corollary 2 of (7.1) that E -+ Eu is nuclear, whence E is a nuclear space by
definition.
COROLLARY 1. If E is a nuclear space, then E -+ Ev is a nuclear map for every
convex, circled neighborhood V of 0 in E.
For E -+ Ev is continuous and Ev is a Banach space.
COROLLARY 2. Every bounded subset of a nuclear space is precompact.
Proof. If ~ is a neighborhood base of 0 in E consisting of convex, circled
sets, then by Corollary 2 of (II, 5.4) E is isomorphic with a subspace of
TI
Ve!8
Ev by virtue of the mapping x -+ {¢v(x): V E ~}, This isomorphism car-
ries a bounded set B c: E into the set TI¢v(B). Now if E is nuclear, each
¢v(B) is precompact in Ev by Corollary 1 of (7.1). Thus the product TI¢v(B)
is precompact, which proves the assertion.
We recall the common usage to understand by lP (1 ~ P < + (0) the Banach
space of all (real or complex) sequences x = (Xl' x 2 , ... ) whose pth powers are
(absolutely) summable, under the norm IIxlip = (~]xnlP)IIP; 100 is the Banach
space of bounded sequences with IIxJJoo = sUPnlxnl.
The following result reveals the special structure of nuclear spaces.
7.3
Let E be a nuclear space, let U be a given O-neighborhood in E, and let p be
a number such that 1 ~ p ~ 00. There exists a convex, circled O-neighborhood
V c: U for which Ev is (norm) isomorphic with a subspace of lP.
Proof. We show that there exists a continuous linear map v E (ft'E, lP) such
that v-I(B) c: U, where B is the open unit ball of [P; V = v-I(B) will be the
neighborhood in question. Assume without loss of generality that U is convex
and circled. The canonical map <Pu is nuclear by (7.2), Corollary 1. hence of
the form <Pu = IAn{" ® Yn, where we can assumethatAn > 0 (n EN), I An = 1,
00
n~ 1
IIYn.l1 = 1 in Eu (n EN) and that the sequence {In} is equicontinuous. Define
v by
vex) = U:/Xdl(X), ~/X2fZ(x), ... )
for all x E E (set .ejIn = I for all n if p = (0). By the equicontinuity of the
sequence {In}, we have vex) E IP and evidently v E 2(E, IP). Nowletp-l + q-l
= I (q = I if p = 00 and q = c:tJ if P = 1) and apply Holder's inequality to
f exnPn with exn
n~l
= .ejInf,,(x), Pn = ~In" Denoting by I II the norm in Eu, we
obtain
00 00
I <Pu(x) I = I I
n~l
Anfn{x)Ynll ~ I Anlfn(x)1 ~
1
Ilv(x)ll p ,
whence v-l(B) c U. Letting V = v-l(B), the definition of v implies that Ey
is norm isomorphic with v(E); hence Ey is norm isomorphic with the closed
subspace vee) of IP.
In the three corollaries that follow, denote by A a set whose cardinality
is the minimal cardinality of a neighborhood base of 0 in E.
COROLLARY 1. Let E be nuclear, and let {Ea: ex E A} be a family of Banach
spaces, each of which is isomorphic with a space IP(1 ~ p ~ 00). There exist
linear maps /". of E into Ea (ex E A) such that the topology of E is the coarsest
topology for which all mappings /". are continuous.
In other words, the topology of E is the projective topology with respect
to the family {(Ea'/"'): ex E A} (Chapter II, Section 5). If we apply (7.3), with
p = 2, to each element Ua(ex E A) of a O-neighborhood base in E, we obtain a
base {Va: ex E A} of O-neighborhoods such that for each ex E A, Ea =Ey is
a Hilbert space (not necessarily of infinite dimension (cf. Chapter II, Sectio~ 2,
Example 5». Now if Ea is a Hilbert space, the norm of Ea originates from a
positive definite Hermitian form (x, P) -+ [x, P]a on Ea x Ea; hence if <Pa
denotes the canonical map E -+ Ea, then (x, y) -+ [<Pix), <PiY)]a is a positive
semi-definite Hermitian form on Ex E such that x -+ [<Pix), <Pa(x)]t is the
gauge function Pa. of Va.
COROLLARY 2. In every nuclear space E there exists a O-neighborhood base
{Va: ex E A} such that for each ex E A, Ey• is a Hilbert space; hence the topology
of E can be generated by afamily of semi-norms, each of which originates from
a positive semi-definite Hermitian form on E x E.
Combining this result with the construction used in the proof of (II, 5.4),
we obtain a representation of nuclear spaces as dense subspaces of projective
limits of Hilbert spaces. Thus the completion of a nuclear space E is isomor-
phic with a projective limit of Hilbert spaces, and obviously nuclear by
Corollary 3 of (7.1).
COROLLARY 3. Every complete nuclear space is isomorphic with the projective

limit of a suitable family (of cardinality card A) of Hilbert spaces. A Frechet
space E is nuclear if and only if it is the projective limit of a sequence of Hilbert
spaces, E = lim gmnHn such that gmn is a nuclear map whenever m < n.
~
Proof. We have only to prove the second assertion. If E is a nuclear (F)-

space, by Corollary 2 there exists a base {Vn: n E N} at 0 which can be sup-
posed decreasing, and such that each En is a Hilbert space. By (7.2)(c) we can
even suppose that each of the canonical maps </>V n , V n +l: En+l ~ En is nuc-
lear. The desired representation is then obtained with Hn = En and gmn =
</>V m , Vn (m ~ n). Conversely, if E is of the form indicated and Vis a convex,
circled O-neighborhood chosen from a suitable base in E, then E ~ Ev can be
identified with the projection p of E into a finite product of spaces Hm say
m
TI H k • Denoting by Pn the projection of E into Hn (n E N) we have p =
k=l
(Plo ... , Pm); hence p = (gin Pm ... ,ginn Pn) for any n > m, which implies that
0 0
P is nuclear.
The following important theorem of permanence is also due to Grothen-
dieck (cf. [13], chap. II, theor. 9).
7.4
Theorem. Every subspace and every separated quotient space of a nuclear
space is nuclear. The product of an arbitrary family of nuclear spaces is nuclear,
and the locally convex direct sum of a countable family of nuclear spaces is a
nuclear space.
Before proving the theorem, we note the following immediate consequence:
COROLLARY. The projective limit of any family of nuclear spaces, and the
inductive limit of a countablefamily o.fnuclear spaces, are nuclear.
Proof of (7.4)
1. The proof for countable direct sums and arbitrary products will be based
00
on property (b) of(7.2). Let E = EB E;, Ei (i EN) be nuclear spaces, and let u
i= 1
be a continuous linear map of E into a given Banach space F. If Ui is the
restriction of u to the subspace Ei of E, U i is continuous and hence nuclear,
and thus of the form
00
u·I = ".i..J /l(i)h(i)

1""'11
'X' Y 11,'.
n \CI (i EN).
n=l
L 1f.l~i)1
00
Here we can assume that !!Yn,i!! ~ I in F for all (n, i) EN x N, that

n=l
~ i- 2 (i EN), and that each of the sequences {h~i): n EN} is equicontinuous
on E i. Let Vi be a O-neighborhood in Ei such thatlh~i)(xi)1 ~ 1 for all Xi E Vi
and all n EN, and define In i to be the continuous linear form on E which is
the extension of h~i) to E' that vanishes on the complementary subspace
EB Ej of E I • The family {In,i: (n, i) EN x N} is equicontinuous, for if U is the

j*i
O-neighborhood riUi in E, then x E U implies !fnix)! ~ 1 for all nand i.
Since u can be written as
00
u= L J.l~i)fn,i ® Yn,i>
n,i= 1
it follows from (7.1) that u is nuclear.

Let {E,,: IX E A} be any family of nuclear spaces, E = TI"E" and let u be a
continuous linear map of E into a given Banach space F. There exists a 0-
neighborhood V in E such that u( V) is bounded in F, and by definition of the
product topology, V contains a O-neighborhood of the form V"' x ... x V,,"
x TI Ep. It follows that u vanishes on the subspace G = TI Ep of E.
p*_ n p*_
Since E = TI E". EB G, it
1= 1
remains to show that the restriction of u to TIiE".
n
is nuclear. But this is clear from the preceding proof, since EB E", is
. n 1=1
identical with TI E", (Chapter II, Section 6).
1= 1
2. The proof of nuclearity for subspaces and quotient spaces will be based
on property (c) of (7.2) and Corollary 1 of (7.3). Let E be a nuclear space and
let M be a subspace orE. For each convex, circled O-neighborhood U in E, set
V = M n U. We show that for each V, there exists another such neighbor-
hood, VI c: V, such that the canonical map &Iv, -+&Iv is nuclear. We can
assume without loss of generality that V = M n U, where U is such that £u
is a Hilbert space. There exists a O-neighborhood U1 c: U such that the
canonical map cPu,u,: Eu, -+ Bu is nuclear; let VI = M n U1 • Now it is not
difficult to see that Mv, and Mv can be identified with closed subspaces of
£u, and £u, respectively, so that the canonical map cPv,v, is the restriction of
L At/; ® YI with {AI} sum-
00
cPu,u, to &Iv, (Exercise 3). But cPu,u, is of the form

i=1
mable, {fJ equicontinuous in [£u,]', and {yJ bounded in £u. Denote by p the
orthogonal projection of £u onto &Iv, let Wi = PYi, and denote bygitherestric-
tion of II to &Iv, (i EN). Then {g;} is equicontinuous, {w;} is bounded in &Iv,
and cPv,v, necessarily of the form L Ajg i ® Wi' and hence nuclear by (7.1).
00
i=1
We employ the same pattern of proof for quotient spaces: Let Ebe nuclear,
let M be a closed subspace of E, F = ElM (topological), and let cP be the
canonical map E -+ F. For a given convex, circled O-neighborhood V in F, we
show the existence of another, VI c: V, such that cPv,v,: Fv , -+Fy is nuclear.
For this we can suppose that V = cP(U), Bu is a Hilbert space, and U1 c: U is
such that £u, is a Hilbert space and cPu,u,: £u, -+ £u is nuclear. The point of
the proof consists now in recognizing that Py can be identified with a quotient
space of £u. In fact, Fy is isomorphic with the space BulL, where L is the clo-
sure of cPu(M) in Bu. Similarly, letting VI = cP(U1 ), Fv , can be identified with
£U,/Ll' where L1 is the closure of cPu,(M) in B u, (Exercise 3).
We note further that lPu.u, maps L1 into L, and lPv. v, is nothing else but the
map of EuJL1 into EulL induced by lPu.u, under the identification just made.
<Xl
Since lPu.u, is nuclear, it is of the form L Adi ® Yi as described in (7.1).

i= 1
We decompose E u, = L1 Ef) Lt, Eu = L Ef) L1. (orthogonal complements).
Letf/ = f; + It and Yi = yi + y7 (i E N) be the corresponding decompositions
(so that, for /;, we have li(Lt) == 17(L1 ) = {O}). Since lPu.u, maps Ll into L,
it follows that LjAif; ® Y;' vanishes on Eu , whence,
<Xl <Xl
lPU.Ur = L Adi ® Y; + L Adt ® y7·

i=l i=l
If now gi denotes the linear form on Eu,/Ll determined by /;", and Wi denotes
the equivalence class of y7 mod L (i EN), then lPv.v, (being the map induced
by lPu.u,) is of the form LA/gi ® Wi> and hence is nuclear by (7.1).
The proof of the theorem is complete.
We supplement theorem (7.4) by showing that the projective tensor product
of two nuclear spaces is nuclear. To this end, we need the concept of the
tensor product of two linear mappings: Let E, F, G, H be vector spaces over
K and u E L(E, G), v E L(F, H). The mapping (x, y) -+ u(x) ® v(y) is bilinear
on E x F into G ® H; the linear mapping of E ® F into G ® H, which corres-
ponds to the former, is denoted by u ® v, and is called the tensor product of u
and v. It is obvious that (u, v) -+ u ® v is bilinear on L(E, G) x L(F, H) into
L(E ® F, G ® H). Thus again by (6.1), to this map there corresponds a
linear map of L(E, G) ® L(F, H) into L(E ® F, G ® H) (called the canonical
imbedding), which is an isomorphism. If G = H = K o, that is, if u = J, v = g
are linear forms, then tensor multiplication in Ko ® Ko can be identified with
ordinary multiplication in K (proof!) and we have I® g(x ® y) = I(x)g(y)
so that the tensor products I ® g and E* ® F* considered earlier are special
cases of the present definition.
7.5
If E and F are nuclear spaces, the projective tensor product of E and F, as well
as its completion E ® F, are nuclear.
Prooj. Let U, V be convex, circled O-neighborhoods in E, F respectively;
set G = E® Fand W = rU® Vin G.ltis clear from (6.3) that Gwis identical
with the normed space (Eu ® Fv , r), where r is the tensor product of the re-
spective norms of Eu and Fv. Hence if lPu, lPv, lPw denote the respective can-
onical maps E -+ E u, F -+ Fv, G -+ Gw, we have 4>w = 4>u ® 4>v. Since E, F
are nuclear, (7.1) implies that 4>u = ~)i/; ® ~i' 4>v = "5).tjgj ® Pj' where {Ail,
{Jlj}, etc., have the properties enumerated in (7.1). For x E E, Y E Fwe have by
definition
106 LINEAR MAPPINGS [Ch. /II
which, as an element of Gw = Eu ®Fv , can be written

L AiIlJi(x)giY)(~i ® P),
00
cPw(x ® Y) =
i,i=l
so that cPw is represented by L Ailli(fi ® g) ® (~i ® Pj)'

i,j
Now {Aillj: (i,j)
eN x N} is a summable family, {[; ® gi} is an equicontinuous family (namely,
uniformly bounded on rU1 ® V1 for suitable O-neighborhoods U1 , V1 in E, F,
respectively), and, clearly, the family {~i ® PJ is bounded in Gw because of
II~i ® Pi II = II ~III liP] II. Hence cPw is nuclear for each element W of a O-neigh-
borhood base of the projective topology on E ® F. The nuclearity of E ®F is
immediate from (7.1), Corollary 3.
8. EXAMPLES OF NUCLEAR SPACES

1. Let T be the k-dimensional torus. The space ~T of (real- or
complex-valued) infinitely differentiable functions on T, endowed with
the topology of uniform convergence in all derivatives, is a nuclear
(F)-space. By (7.4) this implies that the space q)I of infinitely differen-
tiable functions on Rk, whose support is contained in the k-dimensional
interval I, is nuclear (notation as in Chapter II, Section 6, Example 2,
except that the domain is sometimes written in parentheses). For ~l
can be identified with a subspace of ~T by considering each/e ~l as
a k-fold periodic function on I. The method of proof will be sufficiently
exhibited by considering the case k = 1.
Denote by gkCk = 0, ± 1, ±2, ... ) the normalized trigonometric func-
tions gk(t) = (27t)-t eikt, and set [J, gk] = f:,,!(t)g:(t) dt for / e ~p It is
well known from elementary analysis that / has a Fourier expan-
sion/ = L.kQkgk that converges to/uniformly in all derivatives; that is,
/ = L.kQkgk is an expression valid in q)p The coefficients are given by
Q k = [J, gk], and by repeated partial integration it follows that
Q k = [J, gk] = (ik)-m[f(m), gk]

for all k = ± 1, ± 2, ... and all integers m ~ 0, pm} denoting the mth
derivative of f The family of norms / --+ Pn(f) = sup{lpm)(t)l: t e T,
m ~ n} generates the topology of ~T' and the O-neighborhoods
Vn = {j:Pn(f) ~ n- 1 } where n eN form a base at O. Note also that
each of the spaces E v" is algebraically isomorphic with E = ~T>
since the Pn are norms. For fixed n, say n,= m, the expansion off can
be written
Now the linear forms / --+ hk(f) = [pm+2}, gk] (k = ± 1, ±2, ... ) are
uniformly bounded on Vm + 2 and hence are equicontinuous, and the
functions Yk = (ik)-mgk (k = ±l, ±2, ... ) can be arranged to form a
bounded sequence in £v"" Hence the canonical map E --+ tv"" being
§8] EXAMPLES OF NUCLEAR SPACES 107
of the form
is nuclear. Since meN was arbitrary, !lJJ T is nuclear as asserted.

The same conclusion holds if C is any compact subset of R (and more
generally, of Rk), for then C is contained in an interval I as considered
above, and f!JJ c is a closed subspace of !lJJ1 , hence nuclear by (7.4).
2. The space f!JJ of L. Schwartz (Chapter II, Section 6, Example 2),
being the inductive limit of a sequence of spaces !lJJco is nuclear by the
corollary of (7.4).
3. Let tS be the space of infinitely differentiable complex functions
on Rk (with no restrictions on their supports), under the topology of
compact convergence in all derivatives. Let {Cn} be an increasing
sequence of convex compact sets with non-empty interior in Rk such
that UnCn = Rk. Then each compact set C c Rk is contained in some
Cn' and the topology of Iff is generated by the semi-norms
(n eN)
where IDmj(t) I stands for the sum of the absolute values (at t) of all
derivatives ofjthat have order m (~O). As f!JJ co Iff is an (F)-space which
is nuclear. We shall not verify the nuclearity of Iff directly, since it
will be a consequence of (IV, 9.7).
4. Let .1t'(C) be the space of all entire functions of one complex
variable under the topology of compact convergence. It is clear from the
elements of complex function theory that .1t'(C) is an (F)-space; more-
over, by a classical theorem of Weierstrass,' .1t'(C) is not only alge-
braically, but also topologically, a closed subspace of Iff(R 2 ). Hence by
Example 3 and (7.4), .1t'(C) is nuclear.
5. The space!/ (cf. L. Schwartz [2], chap. VII, §3), or !/(Rk),
is (algebraically) defined to be the subspace of tS(Rk) such that lim
It I ... co
Itlm D"f(t) = 0 for any derivative of j of any order n and any integer
k
meN, with It I = [L q]t denoting the Euclidean norm of t = (tl' ... , t k)
i= 1 '
e Rk. The topology of!/ is defined by the sequence of semi-norms
j -+ Pn(f) = sup{(l + W)ID mj(t)l: It I ~ n, m ~ n}

(n eN). !/ is an (F)-space, called the space of rapidly decreasing,
infinitely differentiable functions on Rk. The space !/ is nuclear. This
can be proved directly by applying the method used in Example 1 to
the expansion ofJe!/ by the functions of Hermite (cf. L. Schwartz [2],
vol. II, p. 117). Another proof uses the fact that !/(Rk) is. isomorphic
with a closed subspace of f!JJ(Sk)(I.C. p. 91), where,@(Sk) is;the space of
infinitely differentiable functions on thek.:sphere Sk. For k = 1, the
nuclearity of !lJJ(Sl) was shown in Example 1; for k > I, one can use
expansions by spherical functions.
108 LINEAR MAPPINGS [Ch. '"
In Examples 1,2, and 3 Rk can be replaced by an open subset of Rk or,

more generally, by an infinitely differentiable manifold, and C by an open
subset in Example 4. Further examples will be obtained in Chapter IV, Sec-
tion 9. For examples of nuclear sequence spaces, see Exercise 25.
9. THE APPROXIMATION PROPERTY. COMPACT MAPS

Let H be a Hilbert space. If {x/Z: ex E A} is an orthonormal basis of H
(Chapter II, Section 2, Example 5), if [ , ] is the inner product of H, and if
!..(ex E A) is the continuous linear form on H for which!..(x) = [x, x/Z], then it
is known and easy to prove that for each x E H the family {Ia(x)x/Z: ex E A} is
summable to x (for the definition of summability in H, see Exercise 23) :
x = L!..{x)x/Z.
/ZeA
The convergence of this sum can be interpreted in the following way: If for
each finite subset CI> c A, we denote by u~ the linear map x ~ ~/Ze~!..(x)x/Z
(that is, if u~ = L/Ze~!.. ® x/Z)' then {u~} converges pointwise to the identity
map e of H; the convergence being along the family of finite subsets of A
directed by inclusion c. Now since each u~, being an orthogonal projection,
is (if CI> =1= 0) of norm 1 in !l'(H), it follows from (4.6) that the convergence of
{u~} is uniform on every compact subset of H. [We write !l'(E) = !l'(E, E) and
endow, for any pair of normed spaces E, F, the space !l'(E, F) with the stan-
dard norm u~ lIull = sup{lIu(x)ll: IIxll ~ I}.]
This implies that for every Hilbert space H, the identity map e is in the clo-
sure of H' ® H c !l'(H) for the topology of precompact convergence. It can
be shown (Karlin [2]) that even a separable Banach space does not, in general,
contain an unconditional basis (see below), i.e., a sequence {x,,} for which
there exists a sequence {j,,} c E' satisfying fm(x,,) = bmn (m, n EN), and such
that {j"(x)x,, : n E N} is summable to x for all x E E. A I.c.s. E is said to
have the approximation property (a.p.) if its identity map e can be approxi-
mated, uniformly on precompact sets, by continuous linear maps of finite
rank. (This property is characterized in 9.1 below.) It had long been an open
question (approximation problem) if every I.c.s. has the a.p.; Enflo [1] gave
a negative answer by constructing a sophisticated example of a separable
(B)-space not having the a.p. (For detailed information we refer the reader
to Lindenstrauss-Tzafriri [1], [2].) In what follows we denote the topology
of precompact convergence by a subscript "c".
9.!
Let E be any locally convex space with dual E'. The following properties of E
are equivalent:
(a) The closure of E' ® E in !l' c(E) contains the identity map e.
§9) THE APPROXIMATION PROPERTY. COMPACT MAPS 109
(b) E' ® E is dense in 2c(E).

(c) For every I.c.s. F, E' ® F is dense in 2c(E, F).
(d) For every l.c.s. F, F' ® E is dense in 2c(F, E).
Proof (a) = (b): Given a u E 2(E), a pre compact set A, and a O-neighbor-
hood V in E, we have to show the existence of U o E E' ® E such that u(x) -
uo(x) E V for all x E A. Let U be a O-neighborhood in E for which u(U) c V.
By (a) there exists an eo E E' ® E such that x - eoCx) E U for all x EA.
n n
Clearly, if eo = LX; ® Xi' then the map Uo =u 0 eo =L x; ® U(Xi) satisfies
1 1
the requirement.
(b) = (c): For each fixed v E 2(E, F) the mapping U ~ v u is continuous
0
on 2c(E) into 2c(E, F). Thus since E' ® E is dense in 2c(E), E' ® vee) is
dense in a subspace of 2 c(E, F) containing v 0 e = v, which establishes the
assertion.
(b) = (d): Likewise, for each fixed w E 2(F, E), u ~ u w is continuous on
0
2 c(E) into 2 cCF, E), since for each precompact set B c F, weB) is precompact
in E (for w is uniformly continuous). It follows that w is in the closure .of the
subspace (E' ® E) 0 w of 2 c(F, E), and it is quickly seen that (E' ® E) 0 w c
F'®E.
Finally, the implications (d) = (b), (c) = (b), and (b)=(a) are trivial.
The following result reduces the approximation problem entirely to Banach
spaces. Besides, it gives a positive answer for a large class of locally convex
spaces (including all nuclear spaces).
9.2
Let E be a l.c.s. with a O-neighborhood base mof convex, circled sets such that
for each V E m, Ev has the approximation property. Then E possesses the
approximation property.
Proof Let V E mbe given, and denote by ¢ the canonical map E ~ Ev = F.
Let W = t¢( V), where the closure is taken in F. It follows that ¢ - \ W) c
n
! V + Vo c V if Vo = II. V denotes the null space of V. Note further that
A>O
Ev = ¢(E) is dense in F. Since F has the a.p. by assumption, (9.1)(d) implies
that E' ® F is dense in 2 c(E, F), whence E' ® ¢(E) is also dense in 2 c(E, F).
Hence, ¢ being in 2(E, F), for a given precompact set AcE there exists
n
wEE' ® ¢(E) such that w(x) - ¢(x) E Wfor all x EA. Let w = Lxi ® ¢(x;).
t
1
It follows that ¢[Lx;(x)x i - x] E Whence (since ¢-l(W) c V) X;(X)Xi - X

i 1
E V for all x E A. This proves the assertion, since V was any member of a
neighborhood base of 0 in E.
COROLLARY 1. Every subspace of an arbitrary product of Hilbert spaces
possesses the approximation property.
Proof. Let E = TI Ha be a product of Hilbert spaces. There exists a 0-

aeA
neighborhood base min E such that for each V E m, Ey is isomorphic with
the Hilbert space TI H a , where (J> is a suitable finite subset of A; thus Ey has
a eo!)
the a.p. If M is a subspace of E and W = M n V (V Em), then Mw can be
identified with the closure of cPy(M) in Ey (Exercise 3); hence Mw is a Hilbert
space if Ey is. .
COROLLARY 2. Every projective limit of Hilbert spaces and every subspace of
such a projective limit (in particular, every nuclear space) has the approximation
property.
This is an immediate consequence of Corollary 1 and (7.3), Corollary 2.
COROLLARY 3. If there exists a locally convex space not having the a.p., then
there exists a Banach space not having the a.p.
In view of the last corollary, we shall analyze the approximation problem
for Banach spaces somewhat further. For this we need several results on
compact maps and sets. Since these results are also of independent interest,
they will be proved in detail. We denote by a subscript" b" the topology of
bounded convergence on Ii'(E, F); recall (Section 3) that when E, Fare
normed spaces, this is the to'pology of the normed space Ii'(E, F).
9.3
Let E be normed and let F be a quasi-complete l.c.s.; the set of all compact
linear maps of E into F is a closed subspace of Ii'b(E, F).
Proof. The subset of Ii'(E, F) consisting of all compact maps is evidently
a subspace M. Let us show that M is closed. Let v E M c Ii' bee, F) and a 0-
neighborhood V in F be given, let W be a circled O-neighborhood in F such
that W + W + W c V, and denote by U the unit ball of E. There exists u E M
such that vex) - u(x) E W for all x E U, and since u is compact, u(U) is rela-
tively compact, so that u(U) c U;(b j + W) for a suitable finite subset {b i } of
u(U). Since W is circled, it follows that u(U) c v(U) + W, whence hi E aj + W
(i = I, ... , n) for a suitable subset {aJ, of v(U). Now
n n
v(U) C u(U) + We U (hi + W + W) c U (ai + V),
1 1
which shows v(U) to be precompact, and hence relatively compact, since F is

quasi-complete.
REMARK. The preceding proof shows that if E, Fare t.v.s., F is
separated and quasi-complete, and v is the limit of compact maps, uni-
formly on some non-empty open subset of E (or even on a non-meager
subset of E if E is a Baire space), then v is compact.
§9] THE APPROXIMATION PROBLEM. COMPACT MAPS 111
Although the study of adjoint maps is deferred to Chapter IV (Sections 2

and 7), since it can be handled successfully only with the aid of duality, we
shall make use in what follows of a few elementary facts with which the
reader is likely to be familiar. Recall that if E, Fare normed spaces and if
£', F' are the Banach spaces that are their respective strong duals, then every
u E !feE, F) induces a v E !f(F', E') by means of y' ~ v(y') = y' 0 u. It
is immediate that for all x EE, y' EF' Iv(y')[xll ~ Ily'llllullllxll, whence
Ilv II ~ Iluli. An application of the Hahn-Banach theorem (in its analytical form
(II, 3.2» then shows that Ilv II = Ilu II. v is called the adjoint of u, and denoted
by u'. The following result is due to Schauder.
9.4
Let E, F be normed spaces and let F be complete. A linear map u E !feE, F) is

compact if and only if its adjoint u' E !f(F', E') is compact.
Proof Let U, V be the respective unit balls in E, F and let UO, VO be the cor-
responding dual unit balls in E', F'. We show that u'(VO) = B is relatively
compact in E'. For this it suffices (since E' is metric) to show that each
sequence{x~} c Bhasaclusterpoint. Let x~ = u'(y~)(n E N), where{y~} c VO.
Since VO is equicontinuous and closed for (J(F', F), it is (J(F', F)-compact
by (4.3), Corollary. Hence {y~} has a weak cluster point y' which, by (4.5),
is also a cluster point for the topology of compact convergence. Thus if A =
u(U) is relatively compact (i.e., if u is compact), there exist infinitely many
kEN such that Iy~(ux) - y'(ux)1 ~ B for all x E U and a given B > O. Let
z' = u'(y'). It follows that Ilu'(y~) - u'(y') II = Ilx" - z II ~ B for the same k,
which shows z' to be a cluster point of the sequence {x~}. The converse is clear,
since if u' is compact, then u" is compact in !f(E", F") by the preceding, and u
is the restriction of u" to the subspace E of E".
We further need the following lemma which, as an inspection of the proof
shows, is actually valid in (F)-spaces. We shall confine ourselves to (B)-spaces
for convenience, in particular since the first assertion will be obtained for
(F)-spaces in a different context (IV, 6.3, Corollary 1).
LEMMA I. Let A be a compact subset of the Banach space E. There exists a
null sequence {x n } in E whose closed, convex, circled hull contains A; and there
exists a compact, convex, circled subset B of E such that A is compact as a
subset of E B'
L An =
00
Proof Let {An} be a sequence of positive numbers such that 1, and put
1
Bn = A; + l' Denote by P j (i E N) a finite subset of A( #- 0) such that for x E A,
00
there exists YEP j satisfying Ilx - y II < Bj; U P is dense in A. We define

clearly, j
1
a sequence {Q;} of finite subsets of E, as follows: Let Q1 be the set A1 1P 1 •
If i > I, select for each y EP j an element z E Pj - 1 satisfying Ily - zll < 8i-1'
and form x = (y - Z)/Ai; the resulting set (which has the same number of
elements as Pi) is Qi' Since each element of Qi is of norm < Ai when i > I, the
sequence {Qt, Q2' ... } defines a null sequence {xn: n EN} such that, say,
xn E QI exactly when ni ~ n < ni+1' It is now readily seen that each y E P k is
00
of the form y = AtXil + ... + AkXik and hence in the convex hull of U Q;,
which proves the first assertion. t
To prove the second assertion, it is certainly legitimate to assume A is
convex (since E is complete), and even that A is the closed, convex hull of the
range Q of a suitable null sequence {xn: n EN} in E. Notice that there exists
a sequence {ocn : n E N} of positive numbers such that OCn --+ + 00 and
{ocnxn: n E N} is still a null sequence in E; it suffices, for example, to take
OCn = I/.J Ilxn II if Xn # 0, OCn = n if Xn = 0 (n EN). Let B be the closed, convex,
circled hull of the range of {ocnxn: n E N}; B is compact. It is clear that {xn}
is a null sequence in E B; for if PB is the gauge of B, then piocnxn) ~ I, whence
PB(Xn) ~ oc,;-t for all n EN. To show that A is compact in EB denote by At
the closed, convex hull of Q in E B; At is compact in EB and, as a subset of E,
dense in A. But since EB --+ E is continuous, At is a fortiori compact in E,
and hence identical with A. The lemma is proved.
The results established so far are not quite sufficient to prove (9.5) below;
we shall have to use Proposition (IV, 1.2) in two places, and Lemma 2, below.
To be sure, Lemma 2 is an easy consequence of (IV, 2.3), Corollary 1, and
(IV, 3.3), but we shall give a direct proof which involves only the geometrical
form (II, 3.1) of the Hahn-Banach theorem. For a full understanding of (9.5)
the reader is advised to defer reading its proof until he is familiar with the
material contained in the first three sections of Chapter IV.
LEMMA 2. Let (E, l:) be a I.c.s., B # 0 a compact, convex, circled subset of E,

let Q be the subspace of [EBl' whose elements are continuous for the topology
induced by l:, and let BO be the unit ball of [EB]" Then Q (') BO is dense in BOfor
the topology of uniform convergence on all compact subsets of E B'
Proof. By (4.5) it suffices to show that Q (') BO is dense in BO for the topology
of simple convergence. Let y' E BO, e > 0 and Yi E B (i = 1, ... , n) be given.
We can assume that y'(y;) # 0 for some i. Denote by M the finite dimensional
subspace ofEB(and ofE) generated bytheYi (i = 1, ... ,n); HI = {xEM:y'(x) =
1 + e} is a hyperplane in M and a linear manifold in E, not intersecting B.
Since B is compact, there exists a convex, open set V in E containing B and not
intersecting H t . By (II, 3.1) there exists a closed hyperplane Hin E containing
H t and not intersecting V, say, H = {x E E: x'(x) = I}. Since H (') B = 0,
the element of Q which x' defines is in BO. Moreover, H (') M = HI' whence
for x E M, y'(x) = (1 +e)x'(x), which implies !y'(Yi) - x'(y;)! ~ e for all i,
since Yi E B.
We can now establish the following theorem (Grothendieck [13]) on the
a.p. of Banach spaces and their strong duals.
§9] THE APPROXIMATION PROBLEM. COMPACT MAPS 113
9.5
Theorem. Let E be a Banach space and let E' be its strong dual. Consider
the following assertions:
(a) E has the approximation property.
(b) For every Banach space F, the closure of F' ® E in !Fb(F, E) is identical
with the space of compact maps in !F(F, E).
(c) E' has the approximation property.
(d) For every Banach space F, the closure of E' ® F in !Fb(E, F) is identical
with the space of compact maps in !F(E, F).
Then (a) ¢> (b) and (c) ¢> (d).
REMARK. Assertion (a) is also equivalent to the following: (a') The
canonical map of E' ®E into !F(E) is one-to-one. With the aid of the
equivalence (a)¢>(a'), it can also be shown that (c)=>(a); but the
proofs require further results on duality. The interested reader is referred
to Chapter IV, Exercise 30.
Proof of (9.5). (a) =>. (b): If w E !F(F, E) is compact and V is the unit ball
of F, then A = w( V) is precompact in E. Thus there exists eo E E' ® E such
that, given e > 0, Ileo(x) - xii < e for all x EA. Thus Ilwo - wll ~ e, where
Wo = eo 0 w. Since Wo is an element of F' ® E (namely, Lw'(xD ® Xi if eo =
Ix; ® x;), the implication is proved.
(b) => (a): Given e > 0 and a compact subset AcE, we show the existence
of eo E E' ® E satisfying Ileo(x) - xii < e for x E A. By Lemma I above, there
exists a compact, convex, circled subset BeE such that A c B and A is
compact in E 8 • Letting F =E8 , (b) implies that there exists Wo E [EB ], ® E
such that Ilwo -0/811 <e/2, for the canonical map o/B is a compact map of EB
into E. Since by Lemma 2 each y' E [E8]' can be approximated, uniformly on
A, by elements x' E E', it follows that Ileo(x) - wo(x) II <e/2 for all x E A and
a suitable eo E E' ®E, which implies, by (9.1)(a), that E has the a.p.
(c) => (d) : Let u E !F(E, F) be a compact map. If we imbed, as usual, E and
F as subspaces of their strong biduals E" and F" respectively, then the second
adjoint u" E !F(E", F") is an extension of u. Now the unit ball U of E is (I(E",
E')-dense in the unit ball D of E". We obtain this result by applying Lemma
2 to the weak dual E~ = (E', (I(E', E» of E, substituting for B the unit ball
U O of the strong dual E' and using the fact that, by virtue of (IV, 1.2), E is to
be substituted for Q. It is an easy matter to verify that u" is continuous for
(I(E", E') and (f(r, F'). On the other hand, since u is compact, u(U) is con-
tained in a compact (hence a fortiori (I(F, F')-compact) subset C of F; hence
it follows that u"(U) c C, which implies that u"(E") c F. (This is also a special
case of Chapter IV, Section 9, Lemma 1.)
Now by (9.4), u' E !F(r, E') is compact; that is, the image under u' of the
unit ball of F' is contained in a compact subset A of E'. Since E' has the a.p.
by hypothesis, there exists a mapping vo = Ifi ® x; E E" ®E' such that, e > 0
being preassigned, we have Ilx' - Lfi(x')x; II < e for all x' EA. Now vo 0 u' =
L(fi 0 u') ® x; and we have fi 0 u' = U"(!i) = Yi E E by the preceding. This

implies Ilu' - LYi ® x; II < 8 and, therefore, lIu - LX; ® Yi II < 8. In view of
(9.3) the implication is proved.
(d) => (c): Let A be any compact, convex, circled subset of E'; A is norm
bounded, hence equicontinuous, and it follows that U = {x E E: /x'(x)/ ~ I
for all x' E A} is a convex, circled O-neighborhood in E. Using the fact that
the dual of(E', (J(E', E» can be identified with E, (IV, 1.2), it is quickly verified
by an application of the Hahn-Banach theorem that [E']A can be identified
with the strong dual of Eu and that under this identification l/I A is the adjoint
map of ¢u (Exercise 3). Since l/I A is compact, so is ¢u by (9.4). Hence (d)
implies that, given 8 > 0, there exists an element wEE' ® 2u satisfying
Ilw - ¢u II < 8/2. Since ¢u(E) is dense in Eu, there exists Wo E E' ® ¢u(E)
such that Ilwo - ¢u II < 8. Let Wo = LX; ® ¢u(xJ It follows that Ilw~ - l/I A II
< 8 or, equivalently, that
Ilx' - L x'(x/)x:1I < 8
whenever x' E A; this shows that E' has the a.p.

This completes the proof of (9.5).
When E, Fare normed spaces and W, V O are the respective unit balls of
the Banach spaces E", F', then the topology ofbi-equicontinuous convergence
on E' ® F is the topology of uniform convergence on W ® V O, and hence
normable. It is not hard to see (Exercise 24) that the natural norm for this
topology is identical with the norm induced by 2(E, F). Hence the following
corollary:
COROLLARY. Let E be a Banach space whose strong dual E' has the approxI-
mation property. Thenfor every Banach space F, the canonical imbedding of the
Banach space E' ~ F into 2(E, F) is a norm isomorphism onto the subspace of
compact linear maps of E into F.
For separable (B)-spaces, a stronger form of the approximation property
is obtained as follows: A sequence {x"} c E is called a Schauder basis if each
x E E has a unique representation
where the series converges in E (in the ordinary sense that its partial sums
L" IXkXk converge to x as n -+ 00; cf. Exercise 23). For example, the sequence
1
{e,,} (e" being the vector x = (Xl' x 2 , ... ) for which x" = 1, xm = 0 when
m #- n) constitutes a Schauder basis in each of the spaces IP(1 ~ p < (0) and
Co (the space of null sequences under the sup-norm); bases for most standard
(B)-spaces were constructed by Schauder [1]. We call a Schauder basis
normalized if each of its members has norm 1. Clearly, every Schauder basis
can be normalized. It is immediate from the postulated unicity of the rep-
resentation of x by a Schauder basis that the maps x -+ IX" (n E N) are linear
EXERCISES 115
forms. The following stronger result, a beautiful application of the Banach-

Steinhaus and the homomorphism theorem, was essentially known to Banach
[1].
9.6
If {xn} is a normalized Schauder basis of the Banach space E, then the co-
efficient forms x ~ OCn (n E N) are equicontinuous linear forms J" on E, and the
expansion x = Lnfn(x)xn converges uniformly on every compact subset of E.
n
Proof. Letting Ilxlll = sUPn11 L ockxkll, x ~ Ilxlll is a new norm on E under
k=1
which E is complete. Since Ilxll ~ IIxlll> the new norm also generates the
topology of E by Corollary 2 of (2.1). It follows that there exists a number
C ~ 1 such that Ilxlll :;;;; CIIxll for all x EE (cf. Chapter II, Exercise 5). Now
for each nE N,
n+ I n
locnl = lIocnxnll = II L OCkXk - k=L1 OCkXkll
k= 1
~ 211xlll ~ 2C1\xlI,
which implies that x ~ OCn = J,,(x) are equicontinuous linear forms. The
remainder is immediate from (4.6).
COROLLARY. Every (separable) (B)-space that contains a Schauder basis
possesses the approximation property.
LJ" ® x n, where the
00
This is immediate since (9.6) implies that e = series

1
converges in !l' c(E). The preceding considerations can be carried over to
separable (F)-spaces without difficulty. For an enlightening discussion of bases
in the framework of separable barreled spaces, see Dieudonne [8]. The
question if every separable (B)-space has a Schauder basis used to be called
the basis problem; like the approximation problem, it was answered nega-
tively by Enflo [1]. Note that a Schauder basis need not be unconditional,
i.e., such that for each x, {fn(x)xn : n E N} is summable to x (cf. Exercise 23
for notation); the result is due to Karlin [2]. For the many ramifications of
the basis problem in (B)-spaces, we refer to Day [2].
EXERCISES
1. Let L, M be Hausdorff t.v.s., let Lo be a dense subspace of L, and
let u be a continuous linear map of L into M whose restriction U o to Lo
is a topological homomorphism; then u is a topological homomorphism.
In particular, if u is a topological homomorphism of L into M, and it
is its continuous extension to the completion L with values in Nt, then
u is a topological homomorphism. [Note that, in general, u(L) =F Nt
even if u(L) = M'; cf. Exercise 2, below, and Chapter IV, Exercise 11].
2. Let L be a metrizable t.v.s. and let N be a closed subspace of L;
Show that the completion (L/N)- is isomorphic (norm isomorphic
if L is normed) with L/N, where N is the closure of N in L. (By the
method used in the proof of (1,6.3), show that ii(L) = (L/N)- and
that N = ii-leO), where u denotes the quotient map L-+L/N.)
3. Let E be a vector space, let M be a subspace of E, and let U be a
radial, convex, circled subset of E. Denote by ¢u the canonical map of
E into Pu. (For notation see the beginning of Section 7.)
(a) If V = Un M, there exists a natural norm isomorphism of Mv
onto the subspace ¢u(M) of Eu.
(b) Let ¢ be the quotient map E -+ E/ M and set W = ¢( U). There
exists a natural norm isomorphism of (E/M)w onto Eu/N, where N is
the closure of ¢u(M) in the normed space Eu.
(c) In addition to the hypotheses above, suppose that E is a I.c.s.
and that U is a O-neighborhood in E. Let B be the set of all x' E E' such
that Ix'(x)1 ~ I whenever x E U. There exists a natural norm isomor-
phism of the strong dual of Eu (or of the strong dual of p u) onto [E'la.
4. Let E(:!) be a l.c.s. such that :!) is the inductive topology with
respect to a family {(Ea' ga): e< E A}, where all Ea are Banach spaces and
such that E = UagiEa) (e.g., let E(:!l) be a quasi-complete bomological
space, (II, 8.4». Let F(:!2) be a l.c.s. such that :!2 is the inductive topol-
ogy with respect to a sequence {(Fn' hn): n EN}, where all Fn are Frechet
spaces and such that F = Unhn(Fn). Generalize (2.2) as follows:
(a) If v is a linear map of F onto E which is continuous, then v is a
topological homomorphism.
(b) If u is a linear map of E into F with closed graph, then u is
continuous.
(For a proof, see Grothendieck [13], Intro., theor. B.)
5. Let E be a t.v.s. over K and let F be a Hausdorff t.v.s. over K.
(a) If Eo denotes the Hausdorff t.v.s. associated with E, and ¢ the
canonical map E -+ Eo, show that x: u -+ u ¢ is an (algebraic) isomor-
0
phism of 2(Eo, F) onto 2(E, F); if6 is a family of bounded subsets

of E and6 0 = ¢(6), then Xis a (topological) isomorphism of 2 60(Eo, F)
onto 2 6 (E, F).
(b) Suppose that E is Hausdorff, F is complete, and P is the comple-
tion of E. For each U E 2(E, F) denote 19y ii the continuous extension
of u to P. u -+ ii is an isomorphism of 2(E, F) onto 2(P, F), which is
topological for the 6-topology and the @-topology, respectively, if
for a family 6 of bounded subsets of E, @ denotes the family of their
closures in E:.
(c) The isomorphism u -+ ii maps the respective families of equicon-
tinuous subsets onto each other. If 6 is a total family of precompact
subsets of E, this correspondence H -+ HI induces a set of homeo-
morphisms (and even uniform isomorphisms). (Use (4.5).)
6. Let E be a vector space over K, let {Ea: e< E A} be a family of
l.c.s. over K, and let {ga: e< E A} be a family of linear maps of Ea into
E, respectively, such that E = UagaCEa). Denote by 6 a (e< E A) a total
family of bounded subsets of Ea. Let 6 = Uaga(6 a), provide E with the
inductive topology with respect to the class {(Ea, ga): e< E A}, and let F
be any l.c.s. The 6-topology on 2(E, F) is the projective topology with
respect to the mappings u -+ u 0 ga of 2(E, F) into 2 6~(Ea, F) (e< E A).
EXERCISES 117
Deduce from this that if E = EB"E", then 2 e(E, F) is isomorphic with

TI..2 e.(E", F).
7. Let E, F be I.c.s. such that F;o!: {O}, and let 6 1 and 6 2 be saturated
families of bounded subsets of E with 6 1 c: 6 2 , If 6 1 oF 6 2 , the 6 2 -
topology is strictly finer than the 6 1 -topology on 2(E, F). Deduce from
this that, under the conditions stated, the family of 6-topologies on
2(E, F) is in biunivocal correspondence with the saturated families of
bounded subsets of E.
8. Let E be a bomological space, and let 6 be a family of bounded
subsets of E such that the range of each null sequence in E is con-
tained in some S E 6. Show that if F is a quasi-complete (respectively,
complete) I.c.s., then 2 6 (E, F) is quasi-complete (respectively, com-
plete). (Use (8.3) and Exercise 17, Chapter II.)
9. (Theorem of Osgood). Let X be a non-empty topological space
which is a Baire space.
(a) Let {[,,} be a pointwise convergent sequence of continuous func-
tions on X with values in a metric space Y; the set of points where the
sequence is equicontinuous is not meager in X.
(b) Let {f,,} be a simply bounded sequence of continuous functions
with values in F, where F is a t.v.s. possessing a fundamental sequence
of bounded sets; there exists a subset X 0 of X with non-empty interior
such that the sequence is uniformly bounded on Xo.
Deduce from this the classical versions of the principle of uniform
boundedness and the Banach-Steinhaus theorem.
10. Show that under the conditions of Chapter II, Exercise 14, the
sequence of linear forms f -+ nf(n- l )(n EN) on E is simply bounded
(and, in fact, uniformly bounded on every convex, circled subset of E
which is complete) but not equicontinuous.
11. Let T be a set. A filter tY on T is elementary if it is the section
filter of a sequence {tn} in T, that is, if the sets Fn = {t k : k ~ n}(n EN)
form a base of tY. Show that every filter (fj on T which possesses a
countable base is the intersection of all elementary filters tY::> (fj. Use
this to extend (4.6) to filters with countable base.
12. (Principle of the Condensation of Singularities. Banach-Steinhaus
[1 D. Let E, F be t. V.s. such that E is a Baire space. If H c: 2(E, F) is not
equicontinuous, show that M = {x: H(x) is bounded in F} is a meager
subspace of E. Thus if {Hn} is a sequence of subsets of 2(E, F) each of
which is not simply bounded, there exists Xo E E such that Hn(x o) is
unbounded in F for all n EN, and the set of these Xo is not meager
inE.
13. Let E = ~(l) be the Banach space of continuous complex-
valued functions on the real unit interval I = [0, 1] and let H be the
Hilbert space L 2 (/-l), where /-l denotes Lebesgue measure on I. Identify
E algebraically with a subspace of H, and denote by {Uk} any ortho-
normal basis of H which is also a total family in E; for example, the
normalized trigonometric functions. The formal series Lk[J, Uk]Uk,
where [,] denotes the inner product in H, is called the Fourier ex-
pansion of f (with respect to {Uk})' Show that for each given to E I
n
such that lim sup Ln(to) = 00, where Ln(s) = S! L9k(S) gt(t)! dJ1(t), there
1
exists a continuous function lEE whose Fourier expansion is un-
bounded (hence not convergent) at t = to. (Establish the result by apply-
ing (4.2).)
14. With the notation of Exercise 13, let P be a countable subset of I
and let 6 be a countable family of orthonormal bases in L 2 (J1), each
of which satisfies lim sup Ln(t) = 00 for all t E P.
(a) Show that there exists an lEE whose Fourier expansion with
respect to any member of6 is unbounded at each t E P. (Use Exercises
12, 13.)
(b) Generalize the foregoing result to the case where X is a metriz-
able, locally compact space, J1 a bounded positive measure on X with
p,(G) > 0 for each open G:f. 0, with E = fC(X), H = L 2 (J1) (Chapter II,
Section 2, Examples) and P a countable subset of X such that J1({t}) = 0
for every t E P.
15. Let {a mn } be a numerical double sequence such that for each
mEN there exists a sum mabie sequence {x~m): n EN} E [I for which
00
L amnx~m) is not convergent. There exists a sequence {xn } E [I such

n=1
00
that L amnXn is divergent for all mEN. (Use Exercise 12.)

n=l
16. Let E, F, G be t.v.s. over K.
(a) Show that a bilinear map on E x F into G which is continuous
at (0, 0), is continuous (everywhere).
(b) Let H be a vector space of(6, l:)-hypocontinuous bilinear maps
of Ex F into G. Show that H is a t.v.s. under the topology of 6 x l:-
convergence (it suffices that each IE H be either 6- or l:-hypocon-
tinuous).
(c) A family B of bilinear maps of E x F into G is 6-equihypo-
continuous if, for each S E 6, the family {Ix: x E S,f E B} is equi-
continuous in !l'(F, G). Define the corresponding notions of l:- and
(6, l:)-equihypocontinuity, and prove two propositions analogous
to (5.2) and (5.3).
17. Let E be an infinite-dimensional normable space and let F be its
weak dual (E', (1(E', E». The bilinear form (x,f) -+ I(x) is l:-hypo-
continuous where l: is the family of all equicontinuous subsets of
F; but it is not 6-hypocontinuous for any saturated family 6 of
bounded subsets of E other than the one generated by the finite
subsets of E, and a fortiori not continuous on E x F. (Let S be a
bounded subset of E not contained in a finite-dimensional subspace;
whatever the O-neighborhood U = {f !/(x;)! ~ I}, there exists YES
and, by the Hahn-Banach theorem, lEU such that I(y) is a given
number.)
18. Let E, F be vector spaces of respective dimensions d 1 , d2 over K.
(a) The map x ® y -+ [x* -+ x*(x)y] induces an isomorphism of
E ® F into L(E*, F); similarly, E ® F is isomorphic with a subspace of
L(F*, E).
EXERCISES 119
(b) For each U E E ® F, the minimal number k(~O) of summands

n
Xi ® Yi such that = LXi ® Yi' is identical with the dimension r of
U
1 n
the range of v: x* ~ v(x*) = L x*(Xi)Yi' In each representation of u
1
by r summands, both sets {x;} and {y;} are linearly independent.
(c) The isomorphism of (a) is onto L(E*, F) if and only if at least one
of the cardinals d l , d2 is finite.
(d) The dimension of E®Fis d 1 d 2 •
19. Let Ei(i = 1, ... n), G be vector spaces over K. A mapping
(XI' ... , xn) ~ f(XI' ... , xn) of niEi into G is multilinear (n-linear) if
each of the partial maps Ei ~ G, obtained by fixing the coor-
dinates x/j"# i) of X; is linear. Define the tensor product F = ®;Ei'
and if Ei are I.c.s. define the projective tensor product topology
on F. Formulate and prove for this case results analogous to those in
Section 6.
20. If E, F are vector spaces and p, q are semi-norms on E, F, res-
pectively, the tensor product p ® q is a norm on E ® F if and only if both
p and q are norms.
21. Let E, F, G be normed spaces with respective norms p, q, r.
A norm s on E ® F is called a cross-norm of p and q if for all
(x, y) E E x F, sex ® y) = p(x)q(y).
(a) The tensor product p ® q can be characterized as the unique
norm on E ® F such that for each normed space (G, r), the canonical
isomorphism of f£?(E ® F, G) onto fJ4(E, F; G) (cf. (6.2» is a norm
isomorphism, fJ4(E, F; G) being provided with the norm f ~ IIf II =
sup{r[f(x,y)]:p(x) ~ 1, q(y) ~ I}.
(b) If s is any cross-norm of p and q, then s ~ p ® q.
(c) Denote by E' (respectively, F') the strong dual of E (respectively,
of F) with their standard norms p', q'. Show that
u ~ s(u) = sUP{L x'(xi)y'(Yi): U = L Xi ® Yi' p'(x') ~ 1, q'(y') ~ I}
is a cross-norm on E ® Fthat generates the topology ofbi-equicontinuous
convergence (Section 6).
(d) If t is a cross-norm of p and q, the following assertions are
equivalent:
(1) s ~ t ~ p ® q, where s is the norm introduced in (c).
(2) The functional z ~ t'(z) = sup{lz(w)l: I(W) ~ I}, where w E E® F
and z E E' ® F' is a cross-norm of p' and q'. (Cf. Schatten [1].)
22. Let E, F be I.c.s., X the canonical bilinear map of E x F into
E ® F, l:i the inductive tensor product topology, and E ® F the com-
pletion of (E ® F, l:i)'
(a) For every l.c.s. G, the isomorphism of (6.1) maps 2«E ® F,
l:i)' G) onto ~(E, F; G).
(b) l:i is the inductive topology (Chapter iI, Section 6) with respect
to the family of mappings {Xx, Xy: X E E, Y E F} of F (respectively, E)
into E ® F. Deduce from this that if E, F are both bomological or both
barreled, E@ F is barreled. (Use Exercise 15, Chapter II).
23. Let E be a Hausdorff Lv.s.

00
(i) A formal series L X n, where Xn E E (n EN) is convergent to

1 n
X E E if the sequence of partial sums = L Xk
Sn (n EN) converges to x
k;l 00
in E. This is expressed by writing x = L X n •

1
(ii) A family {x~: IX E A} c E is summable to x E E if for each 0-
neighborhood U in E, there exists a finite subset <Du c A such that for
each finite set <D satisfying <Du c <D c A, it is true that L x~ EX + U.
aE<l>
This is expressed by writing x = L~X~ .If A = Nand {xn} is summable
00
(a summable sequence), the series L Xn is called unconditionally

convergent. 1
(iii) Suppose E to be locally convex. A family {xa: IX E A} E is ab-

solutely summable if it is summable in E and if for each continuous
semi-norm p on E, the family {p(x~): IX E A} is summable (in R). If
00
A = Nand {xn} is absolutely summable, the series L Xn is called

absolutely convergent. 1
(a) If E is complete, {x~: IX E A} is summable if and only if for each
O-neighborhood U in E, there exists a finite subset <Du c A such that
L~E<l>Xa E Uwhenever <D c A is finite and <D II <Du = 0. IfEis a complete
l.c.s., {x~: IX E A} is absolutely summable if for each member p ofa gener-
ating family of semi-norms the family {p(x.): IX E A} is summable.
(b) (ii) and (iii) are equivalent if E is finite dimensional (hence locally
convex).
(c) If E is a l.c.s. on which there exists a continuous norm, an ab-
solutely summable family in E cannot contain more than countably
many non-zero members.
(d) Let E be a complete l.c.s., A an index set of cardinality d > 0, and
denote by Sa the subspace of EA whose elements constitute absolutely
sum mabie families in E. Let r!J> be a generating family of semi-norms
on E; under the topology generated by the semi-norms x --+ p(x) =
L~P(X~) (p E r!J», Sa is a l.c.s. isomorphic with lJ ® E. (Use (6.5).)
(e) Let {x~: IX E A} be a sum mabie family in the l.c.s. E. Show that
for each equicontinuous set BeE', L~lx'(x~)1 converges uniformly with
respect to x' E B.
24. Let E, F be normed spaces and identify E' ® F with a subspace
of !l'(E, F) (Section 7). The topology of bounded convergence on
!l'(E, F) induces the topology of bi-equicontinuous convergence on
E' ® F (E' being the strong dual of E). (By an application of the Hahn-
Banach theorem (II, 3.2), Corollary, show that for any normed space G
and z E G, liz II = sup{lz'(z)l: z' E G', liz' II ~ I}, and use Exercise 2I(c).)
25. Let E be a gestufter Raum (Kothe [1]). Algebraically, E is defined
as follows: Let {un: n EN} be a family of sequences Un = (sCr>. S~), ... )
of real numbers, such that 0 ~ S~) ~ S~J+ 1) for all m, n EN and such
that for each m, there exists n satisfying s~) > o. Consider the subspace
of K~ for whose elements x = (Xl' X 2 , ... ) each of the sequences
EXERCISES 121
{Xms~): m E N}(n EN) is summable (Exercise 23). Provided with the

topology generated by the semi-norms
L IXms~)1
00
X -4 Pn(x) = (n EN)
m=l
E is an (F)-space. This space is nuclear if and only if for each n, there

00
exists pEN such that L sc,::)/sc,::+ p) < + 00. (Replace any quotients 0/0
m=l
by 0.) (For the proof, use (7.2)(c). Concerning the necessity of the
condition, note that the dual E' of E can be identified with the space of
sequences, each of which is absolutely majorized by some nUn. If B.
denotes the set of sequences {Ym} such that IYml ~ s~) for all m, the family
{nBn} is a fundamental family of equicontinuous subsets of E'.)
26. A family {xa: ex E A} in a t.v.s. E is topologically free iffor each
ex E A, Xa is not contained in the smallest closed subspace of E containing
the subfamily {xp: fJ =f. ex}. Let E be a separable (B)-space (more gener-
ally, a separable barreled space), and let {xn} be a maximal, topo-
logically free sequence in E. Show that there exists a unique sequence
U.} c: E' biorthogonal to {x.}, and show that {x.} is a Schauder basis
of E if for each x E E and each gEE', the numerical sequence
{± Ik(X)g(X k): n EN} is bounded.
k= 1
27. Let S be a set, Fa I.c.s. and let G be a vector space of functions on
S into F that are bounded on S. Let V be a fixed O-neighborhood in F
and let Z be a subset of G such that for each finite subset H c: Z, there
exists XES satifying I(x) rt V whenever I E H. The complements of all
these sets Z (as V runs through a base of O-neighborhoods in F) form
a base at 0 for a locally convex topology on G, called the topology of
almost uniform convergence on S (Brace [1]). If E, F are Banach spaces,
a map U E 2(E, F) is compact if and only if it is a cluster point, for the
topology of almost uniform convergence on the unit ball of E, of a se-
quence in E' ® F. (Brace [2].)
PREREQUISITES

ences can be found.
A. SETS AND ORDER

1

§1] DUAL SYSTEMS AND WEAK TOPOLOGIES 123
one l.c.s. into another. There are dual characterizations of continuous and
open linear maps, followed by several results relating to Frechet and Banach
spaces. The section also paves the way for the general open mapping and
closed graph theorems derived in Section 8. These theorems, essentially due
to Ptak, show a rather unexpected relationship between Banach's homo-
morphism theorem and the theorem of Krein-Smulian; they provide an
excellent example of the power of duality theory. Section 9 continues the
study of topological tensor products and nuclear spaces from Chapter III,
presenting several fundamental results on nuclear spaces. Section 10 is
devoted to a study of the relationship between the concepts of absolute
summability and nuclear space, and opens an approach (due to Pietsch)
to nuclear spaces independently of the theory of topological tensor products.
As a by-result one obtains the theorem of Dvoretzky-Rogers (for a sharpened
form of the theorem see Exercise 36). The chapter concludes with a section on
weak compactness, a subject that has received a great deal of attention in the
literature; included are the theorems of Eberlein and Krein in their general
versions due to Dieudonne and Grothendieck.
1. DUAL SYSTEMS AND WEAK TOPOLOGIES
Let F, G be a pair of vector spaces over K, and letJbe a bilinear form on

F x G satisfying the separation axioms:
(Sl) J(xo, y) = 0 Jor all y E G implies Xo = O.

(Sz) J(x, Yo) = 0 Jor all x E F implies Yo = O.
The triple (F, G,f) is called a dual system or duality (over K). It is also
customary to say that J places F and G in duality, or separated duality if the
validity of (Sl) and (Sz) is to be stressed. To distinguishJfrom other bilinear
forms on F x G, J is called the canonical bilinear form of the duality, and is
usually denoted by (x, y) -+ <x, y). The triple (F, G, < , »is more con-
veniently denoted by <F, G).
Examples
1. Let E be a vector space and let E* be its algebraic dual; the bilinear
form (x, x*) -+ x*(x) = <x, x*) defines the duality <E, E*).
2. If E is a locally convex space with (topological) dual E', then E'
is a subspace of E* separating points in E by Corollary 1 of (II, 4.2);
hence the duality <E, E*) of Example 1 induces a duality <E, E') on
the subspace E x E' of E x E*.
3. Let E, Fbe I.c.s. with respective duals E', F'. The algebraic tensor
products E ® F and E* ® F* are in duality with respect to the bilinear
form determined by <x ® y, x* ® y*) = <x, x*)<y, y*). It has been
shown in Chapter III (Section 6) that this duality induces a duality
between E ® F and E' ® F' (cf. Exercise 2).
124 DUALITY [Ch. IV
4. Denote by 4>d and rod' respectively, the direct sum and product of d
copies of the scalar field K. For any vector space A. over K such that
4>d c: A. c: rod' let A. x be the subspace of rod such that y = (YIX) E A. x
whenever the family {XIXYIX} is summable for every x = (XIX) E A.. The
bilinear form (x, y) -+ L.,XIXYIX places A. and A. x in duality (Exercise 5).
If <F, G) is a duality, the mapping x -+ <x, y) is, for each y E G, a linear
form!, on F. Since Y -+ /y is linear and, by virtue of (S2)' biunivocal, it is an
isomorphism of G into the algebraic dual F* of F; thus G can be identified
with a subspace of F*. This identification will be made in the following unless
the contrary is explicitly stated. Note that under this identification, the
canonical bilinear form of <F, G> is induced by the canonical bilinear form
of <F, F*) (Example 1, above).
In this and the following sections, any proposition on F can also be made
on G by simply interchanging the roles of F and G; this is immediate from
the symmetry of <F, G) with respect to F and G, and will not be repeated.
We begin our investigation with a simple algebraic observation; Dij is, as
usual, the Kronecker symbol.
1.1
Let <F, G) be a duality and let {Yi: i = 1, ... , n} (n EN) be a linearly indepen-
dent subset of G. There exist n (necessarily linearly independent) elements
Xi E F such that <Xi> Yj) = Dij (i,j = 1, ... , n).
Proof The proof is by induction with respect to n. By (S2) the assertion
holds for n = 1; if n > 1 there exists by assumption a set {Xi: i = 1, ... , n - I}
for which <Xi' Yj) = Dij (i,j = 1, ... , n - 1). Let Mn be the subspace generated
by the elements Xi (1 ~ i ~ n - 1) and let Fn = {x E F: <x, Yj) = O,j = 1, ... ,
n - I}. Clearly, F = Fn + Mn is an algebraic direct sum. Now Yn cannot
vanish on Fn , or else it would be a linear combination of {Yi: i = 1, ... , n - I}.
Hence there exists Xn E Fn such that <xn, Yn) = 1 and, defining Xi (i = 1, ... ,
n - 1) by Xi = Xi - <Xi> Yn)xn, we obtain the desired set {Xi: i = 1, ... , n}.
COROLLARY. Let F be a vector space, and letfi (i = 1, ... , n) and 9 be linear
forms on F such that the relations fi(x) = 0, i = 1, ... , n imply g(x) = (equiva- °
n fi-
n
lently, such that 1(0) c: 9 -1(0». Then 9 is a linear combination of the forms
i= 1
/; (i = 1, ... , n).
We recall (Chapter II, Section 5) that the weak topology u(F, G) is the
coarsest topology on F for which the linear forms x -+ <x, Y), Y E G are
continuous; by (S1) F is a I.c.s. under u(F, G). If B is any Hamel basis of G,
the topology u(F, G) is ~enerated by the semi-norms x -+ I<x, y)l, Y E B.
1.2
The dual of (F, u(F, G» is G; that is, a linear form f on F is u(F, G)-con-
tinuous ifand only ifit is oftheformf(x) = <x, y)for a (unique) y E G.
Proof In view of the definition of a(F, G), we have to show only that a
given continuous f can be written as indicated. By (III, 1.1) there exist ele-
ments Yi E G (i = 1, ... , n) such that If(x) I ;;;; c SUPi I<x, y)1 for all x E F and
a suitable constant c. Viewing the Yi as linear forms on F, the corollary of (1.1)
showsfto be a linear combination of the Yi' whence the proposition follows.
COROLLARY. Let <F, G) and <F, G1 ) be dual systems such that G1 c: G.
Unless G1 = G, a(F, G1 ) is strictly coarser than a(F, G).
1.3
Let <F, G) be a duality and let G1 be a subspace ofG; the canonical bilinear
form of <F, G) places F and G1 in duality if and only ifG1 is a(G, F)-dense in G.
Proof To prove the sufficiency of the condition, we have to show that the
canonical bilinear form satisfies (S1) on F x G1. If G1 is weakly dense in G
and <xo, y) = 0 for all y E G1 , then the a(G, F)~continuity of y -+ <xo, y)
implies that <xo, y) = 0 for all y E G, whence Xo = O.
For the necessity of the condition, suppose that <F, G) induces a duality
between F and G 1 • If G1 were not dense in (G, a(G, F)), there would exist a
Yo E G not contained in the closure (;1 of G1 • Define a linear formf on (;1 +
[yo] (where [yo] is the one-dimensional subspace of G generated by Yo) by
f(y) = 0 when y E (;1 andf(yo) = 1 ;fis a(G, F)-continuous on its domain by
(I, 4.2), hence by (II, 4.2) has a continuous extension i to G. By (1.2) i(Y) =
<xo, y) for ally E G and an Xo E F. Since (S1) holds on F x G1 by assumption,
it follows that Xo = 0, which conflicts with f(yo) = 1.
COROLLARY. Let F be a vector space and let G be a subspace of F*. <F, F*)
induces a duality between F and G if and only if G is a(F*, F)-dense in F*.
Let <F, G) be a duality. For any subset M of F,
MO = {y E G: Re<x, y) ;;;; 1 if x EM},
where Re<x, y) denotes the real part of <x, y), is a subset of G, called the
polar set (or polar) of M. The absolute polar of M is the polar of the circled
hull of M; it is the subset {y: I<x, y)1 ;;;; 1 if x EM} of G. The following facts
are immediate consequences of this definition:
1. 0° = G and FO ={O}.
2. If A =1= 0 and AM c: N then N° c: A-1 MO.
3. For any family {Ma} of subsets of F, [UMar = nM~.
4. If6 is any saturated family of a(F, G)-bounded subsets of F, the family
of polars {So: S E6} is a O-neighborhood base for the 6-topology on G.
(G = .!l'«F, a(F, G)), Ko).)
5. If L is a t.v.s., a subset M of its dual L' is equicontinuous if and only
if the polar MO (with respect to the duality <L, L *») is a O-neighborhood in L.
The proof of these statements as well as of the following result is left to the
reader.
126 DUALITY [Ch. IV
1.4
For any subset M c F, M O is a a( G, F)-closed, convex subset of G containing
O. If M is circled, then so is M O ; if M is a subspace of F, M O is a subspace ofG.
If M c F, the polar of M O is a subset of F, called the bipolar of M, and is
denoted by MOo; accordingly, the polar of MOO is denoted by M Ooo . The
following result, called the bipolar theorem, is a consequence of the Hahn-
Banach theorem and is an indispensable tool in working with dualities.
1.5
Theorem. Let <F, G) be a duality. For any subset Me F, the bipolar MOO
is the a(F, G)-closed, convex hull of M u {O}.
Proof It follows from (1.4) that MOO is a(F, G)-closed, convex, and con-
tains 0; obviously it also contains M. Thus MI c MOO if MI is the closed,
convex hull of M u {O}; the assertion will be proved when we show that
x ¢ MI implies x ¢ MOo. Suppose that Xo ¢ MI' By the second separation
theorem (II, 9.2), there exists a closed real hyperplane separating MI and
{xo} strictly. Since 0 E Ml> H is of the form H = {x E F:f(x) = I} for a
suitable a(F, G)-continuous real linear form f on F. It follows from (1.2)
and (I, 7.2) thatf(x) = Re<x, Yo) for all x E F and some Yo E G. Now since
o E M I , we have Re<x, Yo) < 1 if x E M I , hence Re<xo, Yo) > 1; it follows
that Yo E M~ cM o, whence Xo ¢ MOo.
COROLLARY 1. For any Me F, M OOO = MO.
COROI.,LARY 2. Let {M.: rJ. E A} be afamity ofa(F, G)-closed, convex subsets

of F, each containing 0, and let M = n.M.; then the polar of M is the a(G, F)-
closed, convex hull of U.M~.
Proof Let N be the closed, convex hull of UM~. Since M;o = M. (rJ. E A)
it follows that W = [UM~r = nM~o = nM. = M (the first of these equalities
holding by (1.5) and Corollary 1, the second by Remark 3 preceding (1.4»,
hence M O = N°O = N as was to be shown.
It is clear that for any Lv.s. L, the polars (taken in L *) of a O-neighborhood
base form a fundamental family of equicontinuous sets (cf. Remark 5 above).
For locally convex spaces, the converse is also true:
COROLLARY 3. If E is a l.c.s., then the polars (taken with respect to <E, E'»
of any fundamental family of equicontinuous sets in E' form a neighborhood
base of 0 in E.
Proof Let 6 be a fundamental family of equicontinuous subsets of E' and
let U be a given O-neighborhood in E; since E is locally convex, U can be
assumed closed and convex, hence a(E, E')-closed by (II, 9.2), Corollary 1.
Since VO is equicontinuous, there exists SEe with vo c S; it follows that

So c voo = V, which proves the assertion.
In slightly greater generality, the last corollary can be restated as follows:
COROLLARY 4. If (and only if) E is a t.v.s. whose topology l: is locally convex,
l: is the topology of uniform convergence on the equicontinuous subsets of E*.
It follows from the bipolar theorem that for subspaces M c F, M = MOO
if and only if M is closed for O'(F, G). Hence the mapping M --+ MO is one-to-
one from the family of all O'(F, G)-closed subspaces of F to the family of all
O'(G, F)-closed subspaces of G. More precisely, M --+ MO is an anti-isomor-
phism of the lattice of closed subspaces of F onto the lattice of closed sub-
spaces of G, the lattice operations being defined by inf(M1 , M 2) = Ml (1 M2
and SUp(Ml' M 2) = (Ml + M 2 )-. Foritis immediate from (1.5) and its corol-
laries that (inf(Ml' M 2)r = sup(M~, M~) and (SUp(Ml' M 2)t = inf(M~, Mn.
It is customary to call the polar MO of a subspace M c F the subspace
of G orthogonal to M (with respect to the duality (F, G». If F =M1 + M2
is the algebraic direct sum of the closed subspaces Ml and M 2, then
G = (M'l + Mn- by the preceding; it will be seen below (Section 2) that
the sum is O'(F, G)-topological, F = Ml Ee M 2 , if and only if G = M~ ® M~
for O'(G, F).
The most important and most frequent dualities are the systems (E, E,),
where E is a given l.c.s. (Example 2 above). Note that every dual system
(F, G) can be interpreted in this way; by (1.2) it suffices to endow F with
O'(F, G) and consider G as the dual F' of F. Recall that the weak dual of a l.c.s.
(more generally, of a t.v.s.) E is the l.c.s. (E', O'(E', E»; we shall find it con-
venient to denote this space by E;. The section is concluded with two useful
results on E;.
1.6
If E is any l.c.s., the family of all barrels in E and the family of all bounded
subsets of E: that are closed, convex, and circled, correspond to each other by
polarity (with respect to (E, E'».
Proof. Let D be a barrel in E, then DO is bounded in E:. Since by (II, 9.2),
Corollary 2, D is closed for O'(E, E'), it follows from (1.5) that D = DOO
with respect to (E, E'); it is hence sufficient to show that for each bounded,
closed, convex, circled subset B of E;, B O is a barrel in E. In view of(1.4) there
remains to prove only that B Ois radial. Let x E E; then {x}O is a O-neighbor-
hood in E:, hence there exists A > 0 such that Be A-l{xr = {hr. This
implies AX E BO, completing the proof.
In view of Corollary 3 of (1.5) there results the following dual charac-
terization of barreled spaces (for a strengthened form, cf. Section 5).
COROLLARY. A I.c.s. E is barreled if and only if each bounded subset of E; is
equicontinuous.
128 DUALITY [Ch. IV
1.7
Let E be a separable t.V.S. Every closed equicontinuous subset of E; is a
compact metrizable space (under the induced topology); if, in addition, E is
metrizable then E; is separable.
Proof. The first assertion is a special case of (III, 4.7) for F = Ko. If E is
metrizable, denote by {Un: n E N} a neighborhood base of O. Since each
polar U~ is equicontinuous and closed in E;. U~ is a compact metrizable,
hence separable space for the topology induced by a(E', E). Let An be a
countable, dense subset of U~ (n EN). Since, clearly, each x' E E' is contained
in some U~ (i.e., since {U~}, being fundamental, covers E'), it follows that the
co
countable set A = UAn is dense in E;.
1
2. ELEMENTARY PROPERTIES OF ADJOINT MAPS
Let F and F1 be vector spaces over K and let u be a linear map of F into Fl'
For each element y* of Ft, the mapping y* u: x -+<UX, y*) is a linear form
0
x* E F* (we write ux in place of u(x» commonly denoted by u*y*. Thus we

have the identity
<ux, y*) = <x, u*y*)
on F x F;; the map y* -+ u*y*, evidently linear on Ft into F*, is called the
algebraic adjoint u* of u.
2.1
Let <F, G) and <F1, G1 ) be dualities over K. A linear map u of F into Fl is
continuous for a(F, G) and a(Fl' G1) if and only ifu*(G1 ) c G. In this case the
restriction u' of u* to G1 is continuous for a(G1 , F1 ) and a(G, F), and u" =
(u')' = u.
Proof. If u*(G1 ) c G, then x -+ <ux, y') =<x, u'y')(x E F, y' E G1 ) is
continuous for a(F, G), which implies, by definition of a(F1> G1 ), the con-
tinuity of u for these two topologies. Conversely, if u is continuous for
a(F, G) and a(F1> G1 ), then x -+ <ux, y') = <x, u*y') is continuous for
a(F, G), whence u*y' E G by (1.2). By virtue of the identity
<ux, y') = <x, u'y') (x E F, y' E G1 )
it is now obviously true that u' is continuous for a(G1> F1 ) and a(G, F); the
final assertion follows by symmetry.
If u*(G1 ) c G,the mapping u' is called the adjoint (transpose, dual map) of
u with respect to the dualities <F, G) and <F1 , G1 ). Sometimes the relation
u*(G1 ) c G, equivalent with the weak continuity of u, is expressed by saying
that the adjoint of u "exists". Notice also that every linear map u of F into
Fl is continuous for a(F, F*) and a(F1> Ft), these topologies being the weak
§2] ELEMENTARY PROPERTIES OF ADJOINT MAPS 129
topologies associated with the finest locally convex topology on F and Fl.
respectively, (Chapter II, Section 6 and Exercise 7).
We consider an example. Suppose that <F, G) is a duality and
F = MI + M2 an algebraic direct sum; denote po lars with respect to
<F, G) by 0, and polars with respect to <F, F*) by". Let p be the
projection of F onto MI that vanishes on M 2. It follows readily from the
identity <px, x*) = <x, p*x*) on F x F* that p* is a projection with
range M; and null space M;; hence F* = M~ + M; is an algebraic
direct sum; if M I , M2 are (J(F, G)-closed, Mf + M~ is (J(G, F)-dense in
G by (1.5), Corollary 2. In order that G = Mf + M~ it is necessary and
sufficient thatp*(G) = M~ which, by virtue of M~ = M; 11 G, is equiva-
lent with p* (G) c G; this, in turn, is equivalent with the (J(F, G)-continu-
ity of p. Hence an algebraic direct sum F = MI + M2 of two closed
subspaces is (J(F, G)-topological if and only if G = Mf + M~; in that
case, this decomposition of G is (J(G, F)-topological, and the projec-
tions p: F ~ MI and p': G ~ M~ are mutually adjoint.
The proof of the following simple result, recorded for reference, is omitted.
2.2
Let <Fi , G i ) (i = 1,2,3) be dualities over K, and denote by Ui a weakly
continuous linear map of Fi into F i + l (i = 1,2). The adjoint of w = U2 ° Ul
is W' = u~ °uz.
2.3
Let <F, G) and <FI , GI ) be dualities, let u be a weakly continuous linear map
of F into FI with adjoint u', and let A, B be subsets of F, Fl respectively. Then
the following relations hold:
(a) [u(AW = (u')-l(AO).
(b) u(A) c B implies u'(BO) c AO.
(c) If A and B are weakly closed, convex sets containing 0, then u'(BO) c AO
implies u(A) c B.
Proof (a) [u(A)]O = {y' E G I : Re<ux, y') ~ 1 if x E A} = {y' E G1 :
Re<x, u'y') ~ 1 if x E A} = (u')-I(AO).
(b) u(A) c B implies BO c [u(A)]O = (u')-I(AO), which implies u'(BO) c: AO.
(c) By (b), u"(AOO) c BOo, hence u(A) c B in view of u" = u and the bipolar
theorem (1.5) by which A = AOo, B = BOo.
COROLLARY. The null space (U,)-I(O) is the subspace ofG I orthogonal to the
range u(F) of u. In particular, u' is one-to-one if and only if u(F) is (J(FI' GI )-
dense in Fl'
The corollary is immediate in view of (2.3)(a) and the bipolar theorem.
If E, Fare l.c.s. (more generally, t.v.s.) and u is a continuous linear map of E
130 DUALITY [Ch. IV
into F, then u is continuous for the associated weak topologies aCE, E') and
a(F, F'); in fact, for any y' E F' the linear form x --+ <ux, y') is continuous
on E, hence continuous for aCE, E') by the definition of aCE, E'), which
implies the weak continuity of u. Thus each u E !F(E, F) has an adjoint u'
which is a continuous linear map of the weak dual F; into E;. More generally
we prove the following result on the continuity of an adjoint map.
2.4
Let <F, G) and <Fl , Gl ) be dualities and let u be a weakly continuous linear
map of F into Fl with adjoint u'. Let 6 and 6 1 be saturated families ofa(F, G)-
bounded (respectively, a(Fl' Gl )-bounded) subsets of F and F l , and denote by :I
the 6-topology on G, by:I 1 the 6 1 -topology on Gl . Then u' is continuous on
(G l , :I l ) into (G,:I) if and only ifu(6) c 6 1 ,
Proof Let 6 and 61> respectively, be the families of all weakly closed,
convex, circled sets contained in 6 and 6 1 , Since 6 and 6 1 are fundamental
subfamilies of 6 and 6 1, respectively, the families U = {So: S E 6} and
HI = {S~: Sl E 6 1 } are O-neighborhood bases for :I and :I1 . By virtue of
(b) and (c) of (2.3), the relations u(S) c Sl and u'(Sn c So are equivalent
for all non-empty S E 6 and Sl E 6 1, which proves the assertion.
3. LOCALLY CONVEX TOPOLOGIES CONSISTENT WITH A GIVEN

DUALITY. THE MACKEY-ARENS THEOREM
Let <F, G) denote a given duality over K and let 6 be a family of a(G, F)-
bounded subsets of G. The topology of uniform convergence on the sets
S E 6 (Chapter III, Section 3) is a locally convex topology on the space
F = !F«G, a(G, F)), Ko), where Ko, as always, denotes the one-dimensional
t.v.s. associated with the scalar field K. Recall that the 6-topology and the
6-topology are identical if 6 denotes the saturated hull of 6 (Chapter III,
Section 3).
A locally convex topology :I on F is called consistent with the duality
<F, G) if the dual of (F, :I) is identical with G (G being viewed as a subspace
of F*, Section 1). It follows that a topology on F consistent with <F, G) is
finer than a(F, G), hence Hausdorff; a(F, G) is by its definition the coarsest
consistent topology on F. An immediate consequence of this definition and
(II, 9.2), Corollary 2 is the following frequently used fact.
3.1
The closure of a convex subset C c F is the same for all (locally convex)
topologies on F consistent with <F, G) (and hence identical with the a(F, G)-
closure of C).
§3] LOCALLY CONVEX TOPOLOGIES 131
It follows from Corollary 3 of the bipolar theorem (1.5) that every topology
~ on F consistent with <F, G) is an 6-topology, namely the topology of
uniform convergence on the (saturated) class of ~-equicontinuous subsets
of G; by the theorem of Alaoglu-Bourbaki, these classes consist entirely
of sets relatively compact for O"(G, F). The following result, due to G.
W. Mackey [5] and R. Arens [1], asserts that conversely, every saturated
family, covering G, of O"(G, F)-relatively compact sets is eligible to be the
class of equicontinuous sets for a consistent topology on F. These classes
of subsets of G, being saturated, are thus in biunivocal correspondence with
the locally convex topologies on F consistent with <F, G) (Chapter III,
Exercise 7).
3.2
Theorem. A (locally convex) topology ~ on F is consistent with a given
duality <F, G) if and only if ~ is the 6-topology for a saturated class 6,
covering G, of 0"( G, F)-relatively compact subsets of G.
Proof The necessity of the condition has been noted above; let us prove its
sufficiency. Let 6 be a saturated family of subsets of G which covers G and
consists of O"(G, F)-relatively compact sets. It follows from (III, 3.2) that F is
a I.c.s. under the 6-topology which, since 6 covers G, is finer than O"(F, G).
(O"(F, G) is the topology of simple convergence.) Hence the dual of F with
respect to the 6-topology contains G; we have to show its identity with G.
Let f be a linear form on F continuous for the 6-topology; the polar {f} a
(taken with respect to <F, F*» is, by the continuity of J, a O-neighborhood
for the 6-topology and hence contains a set So, where S, since 6 is saturated,
can be assumed convex, circled, and O"(G, F)-compact. ThusfE Sao, where the
bipolar is taken with respect to <F, F*), and (1.5) implies that S = Sao,
since S, being O"(G, F)-compact, is compact and hence closed in (F*, O"(F*, F».
It follows that f ESC G, completing the proof.
COROLLARY 1. There exists a Jinest locally convex topology on F consistent

with <F, G), namely the topology of uniform convergence on all O"(G, F)-
compact, convex, circled subsets of G.
This topology on F is called the Mackey topology on F with respect to

<F, G) and denoted by reF, G). The saturated hull [ of the family (£: of all
O"(G, F)-compact, convex, circled subsets of G is obtained by adjoining all
subsets of members of (£:, but [ must not be confused with the family of all
0"( G, F)-relatively compact subsets of G, this latter family being not saturated
unless the convex hull of every relatively compact subset is again relatively
.compact for 0"( G, F) (cf. Section 11). On the other hand, if G is quasi-complete
for O"(G, F) then the words" saturated" and" convex, circled" can be omitted
in the statements of (3.2) and Corollary 1, respectively.
132 DUALITY [Ch. IV
COROLLARY 2. Every a(F, G)-bounded subset of F is bounded for r(F, G);

consequently, the respective families of bounded sets are identical for all locally
convex topologies on F consistent with <F, G).
Proof. The family (£; of all convex, circled, a(G, F)-compact sets in G is
contained in (in fact, identical with) the family 6 of (III, 3.4) if, in (III, 3.4),
we take F = Ko, for every C E (£; is a(G, F)-complete (cf. (5.5), Corollary 2).
Thus the first assertion follows from (III, 3.4); the second is immediate.
In view of the frequent (and often tacit) application of the preceding results,
we summarize them once more in the following statement; the reader
should keep in mind the particular case where E is a l.c.s., E = F, and E' = G.
3.3
Let <F, G) be a duality. A locally convex topology '.t on F yields the dual G if
and only if '.t is finer than a(F, G) and coarser than r(F, G); if '.t is such a
topology, a convex subset of F is '.t-closed if it is r(F, G)-closed, and any subset
of F is '.t-bounded if it is a(F, G)-bounded.
More generally, if '.t is a locally convex topology on F such that F('.t)' c G,
then '.t is coarser than r(F, G); if'.t is such that F('.t)' :::> G, then '.t is finer than
a(F, G). There exists, on a given vector space F, a finest locally convex
topology (Chapter II, Section 6 and Exercise 7), which is consistent with
<F, G) if and only if G = F*, and is clearly identical with r(F, F*). The
coarsest l.c. topology on F is of course the trivial topology {0, F}, but there
does not necessarily exist a coarsest l.c. topology which is separated; if such a
topology '.tm exists, then F = (F')* and '.tm = a(F, F'), where F' = F('.t m )',
and '.tm is called a minimal topology on F (Exercise 6).
A l.c.s. E is called a Mackey space if its topology is r(E, E'). The following re-
sult shows that most of the l.c.s. occurring in applications are Mackey spaces.
3.4
If E is a l.c.s. which is either barreled or bornological (hence if E is metri-
zable), then E is a Mackey space.
Proof. If E is barreled, then by (1.6), Corollary, every bounded, and a
fortiori every compact subset of E;, is equicontinuous, which shows the
topology '.t of E to be finer than r(E, E'), hence identical with r(E, E') by
(3.3). Similar reasoning applies when E is bomological, since every r(E, E')~
neighborhood of 0 absorbs all bounded subsets of E, and hence is a 0-
neighborhood for '.t.
Let (E, '.t) be a l.c.s. and let (E, :1:) be its completion. By means of the
.r
mapping f ~ J, where denotes the continuous extension of fEE' to E, we
can identify the dual of E with E'. It is immediate from (I, 1.5) (and a special
case of Chapter III, Exercise 5) that under this identification every '.t-equi-
continuous subset is :1:-equicontinuous, and conversely. The following is a
simple consequence of this fact.
§4] DUALITY OF PROJECTIVE AND INDUCTIVE TOPOLOGIES 133
3.5
If E is a Mackey space, so is its completion E.
For if C is a convex circled a(E', E)-compact subset of E', C is a fortiori
aCE', E)-compact, hence ~-equicontinuous, since E is a Mackey space; it
follows that C is l:-equicontinuous, and hence E is a Mackey space.
4. DUALITY OF PROJECTIVE AND INDUCTIVE TOPOLOGIES

The definition of projective and inductive topologies (Chapter II, Sections
5, 6) suggests that these two types of topologies will occur in pairs on dual
systems; the present section is concerned with this sort of duality. We do
not approach the subject in the greatest possible generality, but present the
duality between induced and quotient topologies and between product and
direct sum topologies; this will permit us to make some applications to the
duality between projective and inductive limits.
Let (F, G) be a dual system, let M be a subspace of F, and let MO be the
subspace of G orthogonal to M. Then the restriction of the canonical bilinear
form to M x G is constant on each set {(xo, y)}, where Xo E M is fixed and
y runs through an equivalence class [y] of G mod MO. Therefore (x, [y]) .....
fl(x, [y]) = (x, y), wherey E [y], isa well-defined bilinear form onM x G/Mo;
it is easy to see that fl places M and G/Mo in duality. The dual system
(M, G/M°,Jl) will be denoted by (M, G/MO).
Let IjJ denote the canonical imbedding of Minto F, and cf> the quotient
map G ..... G/Mo. It follows from the definition of the dual system (M, G/MO)
that the identity
(x, cf>(y» = (ljJ(x), y)
holds on M x G. This implies that IjJ is continuous for a(M, G/MO) and
(J(F, G), cf> is continuous for (J(G, F) and a(G/Mo, M), and that IjJ and cf> are
mutually adjoint (Section 2). This observation will be helpful in proving the
following theorem.
4.1
Theorem. Let (F, G) be a dual system and let M be a subspace of F. Denote
by 6; and 6; saturated families of weakly bounded subsets ofG and G/Mo for
the dualities (F, G) and (M, G/MO), respectively, and denote by ~1 and ~2 the
corresponding 6-topologiesonFandM. Dually, let 6 1 and 6 2 be saturatedfam-
ilies of weakly bounded subsets of F and M, respectively, and denote by ~~ and ~2
the corresponding 6-topologies on G and G/ MO. Consider thefollowing assertions:
(a) cf>(6~) == 6;.
(b) ~1 induces ~2 on M.
(c) ljJ-l(6 1 ) = 6 2 •
(d) ~; is the quotient topology of~;.
134 DUALITY [Ch. IV
Then we have the implications (a)=>(b) and (d) => (c); if ~1 is consistent
with (F, G), (b) => (a); if~~ is consistent with (F, G) and M closed, (c) => (d).
Proof For greater clarity we denote polars with respect to (F, G) by 0,
and polars with respect to <M, G/MO) by .
(a) => (b): If Sl e 6~ it follows from (2.3)(a) that
[4>(Sl)]O = I/I-l(S~) = S~ n M.
As Sl runs through 6~, S'l runs through a ~l-neighborhood base of 0 in F;

since by assumption 4>(SI) runs through 6 2, it is clear that ~1 induces ~z
onM.
(d) => (c): Let U be the ~~-neighborhood filter of 0 in G. Then ID = 4>(U)
is the O-neighborhood filter of the quotient topology on G/Mo. Again by
(2.3)(a) we have
for all U E U. Since UO runs through a fundamental subfamily of 6 1 as U

runs through U, the assumption that ~2 be the quotient topology of ~~
implies that 1/1-1(6 1 ) = 6 2 •
(b) => (a): We assume that ~1 is consistent with (F, G). Denote by U1
the family of all closed, convex ~cneighborhoods of 0 in F; then Uz = U1 n M
is a base for the ~z-neighborhood filter of 0 in M. Notice that since UO
(UeU 1 ) is compact, UO+Mo is closed for a(G,F) and 4>(UO) is compact
(hence closed), 4> being continuous for a(G, F) and a(G/Mo, M); from (1.5),
Corollary 2, we obtain
where M denotes the a(F, G)-closure of M. As U runs over U1 , UO runs over a

fundamental subfamily of6~; likewise, [U n M)"runs over a fundamental
subfamily of 6 2. Since both families are,saturated, it follows that 4>(6;) = 6 2.
(c) => (d): We assume that ~~ is consistent with (F, G) and that M is
closed for a(F, G). Since 1/1(6 2 ) c 6 1 , (2.4) implies that 4> is continuous for
~; and ~2' hence ~2 is coarser than the quotient topology of ~~. Thus it is
sufficient to show that for each closed, convex, circled Sl e 6 1 , 4>(SJ.) is a
~2-neighborhood of 0 in GIMo. If Sz = Sl n M, it follows from (2.3)(a)
and (1.5) Corollary 2, that
4>-I(S;) = [I/I(Sz)]O = [Sl n M)" = (St + MO)-.
Here V = Sl + MO is a ~~ -neighborhood of 0, and the closure is with respect
to a(G, F); since ~~ is consistent with (F, G) and V is convex, the closure is
also with respect to ~;. This implies (Sl + MO) - = V c V + V = 2S1+ 2Mo.
It follows, therefore, from the relation above that S; c 24>(SJ.), which shows
4>(SJ.) to be a ~2-neighborhood of 0 in G/Mo.
REMARK. The consistency with <F, G> of the topologies :1:1 and :1:;
is indispensable for the implications (b) = (a) and (c) = (d) (Exercise 7);
also M must be assumed closed for (c) = (d) as we shall see shortly.
(4.1) can easily be stated in more general form replacing equality in (a)
and (c) by inclusion and changing, accordingly, the statements of (b)
and (d) to the corresponding inclusion relations for the 6-topologies.
COROLLARY 1. If <F, G> is a duality and M is a subspace of F, the weak
topology (J(M, G/MO ) is the topology induced on if by (J(F, G). On the other
hand, (J(G/M o , M) is the quotient topology of (J(G, F) if and only if M is closed
in F.
Proof The first assertion follows from (a) = (b) of(4.1) by taking 6~ and
6~ to be the saturated families generated by all finite subsets of G and
G/Mo, respectively. The sufficiency part of the second assertion follows
similarly from (c) = (d). Conversely, if (J(G/Mo, M) is the quotient of (J(G, F),
then we have (since MO = MO) (J(G/Mo, M) = (J(G/M o , M) by the preceding,
which implies M = M (Corollary of (1.2».
Let E be a I.c.s., let M be a subspace of E, and let F = E/N be a quotient of
E; denote by ljJ: M ---> E and ¢: E ---> E/ N the canonical maps. f ---> f ° ljJ is a
linear map of E' onto M' which is onto M' by (II, 4.2) and defines an algebraic
isomorphism between M' and E'/Mo. Dually, g ---> go ¢ defines an alge-
braic isomorphism between F' and N° c E'. In view of this, the dual of M
(respectively, E/N) is frequently identified with E' /Mo (respectively, W). The
following is now immediate from Corollary 1.
COROLLARY 2. Let M be a subspace and let F be a quotient space of the l.c.s.
£. The weak topology (J(M, M') is the topology induced by (J(E, E'), and the
topology (J(F, F') is the quotient topology of (J(E, E').
COROLLARY 3. If <F, G> is a duality and M is a subspace of F, then the
Mackey topology r(G/Mo , M) is the quotient ofr(G, F) if and only if M is closed.
On the other hand, the topology induced on M by reF, G) is coarser than
-reM, G/MO), but consistent with <M, G/Mo>.
Proof The sufficiency part of the first assertion is immediate from the
implication (c)=(d) of (4.1). Conversely, if r(G/M o , M) is the quotient of
reG, F), then r(G/Mo, M) yields the same continuous linear forms on G/Mo
as the quotient of (J(G, F), which is (J(G/M o , M) by Corollary I; it follows
that M = M. For the second assertion, note that ¢ is continuous for (J(G, F)
and (J(G/M o , M), which implies ¢(6~) c 6; if 6~, 6; denote the saturated
hulls generated by all convex, circled, weakly compact subsets of G and
G/ M respectively; it follows from (2.4) that ljJ is continuous for reM, G/ MO)
O
,
and reF, G), which is equivalent to the last topology being coarser on M than
reM, G/ M The final assertion is clear, since reF, G) is finer than (J(F, G).
O
).
The final corollary is obtained in analogy to Corollary 2, using (3.4) for

the proof of the second assertion.
136 DUALITY [Ch. IV
COROLLARY 4. Let M be a subspace and let F be a quotient space of the l.c.s.

E. The Mackey topology 7:(F, F') is the quotient topology of 7:(E, E'); if the
restriction of 7:(E, E') to M is metrizable, it is identical with 7:(M, M').
This last result can be rephrased by saying that every (separated) quotient
of a Mackey space is a Mackey space, and that every metrizable subspace of a
Mackey space is a Mackey space.
We turn to the duality between products and direct sums. Let {(Fa' Ga):
ex E A} denote a family of dualities over K and let F = TIaFa, G = EElaGa.
The bilinear formf on F x G, defined by
f(x, y) = L,,(xa, y,,)
(note that the sum is over an at most finite number of non-zero terms),
places F and G in duality; let us denote by (F, G) the dual system (F, G,f).
As before (Chapter II, Sections 5, 6), we shall identify each F" with the
subspace F" x {O} of F, and each Ga with the subspace Gil EEl {O} of G; but for
greater clarity polars with respect to (F", Ga ) will be denoted by (ex E A) and
0
polars with respect to (F, G) by o. We further note that if p" is the projection
F--+ Fa, qu. the injection Gu. --+ G (ex E A), then
(Pax, Ya) = (X, qu.Yu.)
is an identity for X E F, Yeo EGa and ex E A. Hence by (2.1) Pa and qa are
weakly continuous with respect to (F, G) and (Fa' Ga).
If6a is a family of weakly bounded, circled subsets of Fa (ex E A), then it is
immediate that each product S = TIu.Sa is a u(F, G)-bounded, circled subset
of F; let us denote by 6 = TIa6a the family of all such product sets. Clearly,
6 covers F if each 6" covers Fa (ex E A). Dually, let 6~ be a family of weakly
bounded, circled sub.sets of Ga (ex E A); then each set S' = EElaeHS~, where H
is any finite subset of A, is circled, and u(F, G)-bounded in G; let us denote
by 6' = EEla6~ the family of all such sums. 6' covers G if each 6~ covers
Ga (ex E A). With this notation we obtain
4.2
The product of the 6~-topologies is identical with the 6'-topology on F;
dual/y, the local/y convex direct sum of the 6 a-topologies is identical with the
6-topologyon G.
Proof If S' = EElaeHS~, where H contains n ~ I elements, a short compu-
tation shows that
(S')" c TI (S~t x TI Fa
"eH ajH
C n(S')" ,
which proves the first assertion.

Dually, let S= TIaSa and assume each Sa (ex E A) to be weakly closed,
r
convex, and circled. It is evident that the convex, circled hull aS~ is con-
tained in So. Conversely, if y = (ya) E So, then Lal(xa, Ya)1 ~ I for all
x = (Xa) E S; letting Aa = sup{l<xa, Ya>l: XES}, it follows that Aa = 0 except

for finitely many rJ.EA, and LaAa ~ 1. Now Ya E AaS~; hence Y = LaYa E raS~,
which shows that S· = raS~, Since the totality of sets {raS~} form a 0-
neighborhood base for the locally convex direct sum of the 6 a-topologies (cf.
(*) preceding (II, 6.2», this topology is identical with the 6-topology on G.
We apply (4.2) to families of l.c.s.
4.3
Theorem. Let {Ea: rJ. E A} be a family of I.c.s. and let E = TIaEa. The
dual E' of E is algebraically isomorphic with $aE~, and the following topological
identities are valid:
1. aCE, E') = TIaa(Ea, E~).
2. ,(E, E') = TIa,(E a, E~).
3. ,(E', E) = $a,(E~, Ea).
REMARK. We have a(E', E) = $aa(E~, Ea) if and only if the family
{Ea} is finite (Exercise 8).
Proof. It is immediate that each f = (fa) E $aE~ defines a linear form

x -+ f(x) = l.J'a(xa) on E which is continuous, since f = Lfa 0 Pa (the sum
having only a finite number of non-zero terms); clearly, this mapping of
$aE~ into E' is one-to-one. There remains to show that each gE E' originates
in this fashion. There exists a O-neighborhood U in E on which g is bounded;
U can be assumed of the form TI Ua x TI Ea for a suitable finite subset
aeH a¢H
He A. Denote by fa (rJ. E A) the restriction of g to Ea; then, clearly, fa E E~
for all rJ. and fa = 0 if rJ. ¢ H. Hence for x E E we obtain
g(x) = g( L Pax) = aeH

aEH
Lfixa),
which establishes the assertion; E' is hence isomorphic with the algebraic
direct sum $aE~ by virtue of the duality between products and direct sums
introduced above.
It remains to prove the topological propositions.
1. If 6~ denotes the family of all finite dimensional, bounded, circled
subsets of E~ (rJ. E A), it is evident that 6' = $a6~ is fundamental for the
family of all finite dimensional, bounded, circled subsets of $aE~; the
proposition follows from (4.2).
3. If 6 a denotes the family of all convex, circled, weakly compact subsets
of Ea (rJ. E A), then 6 = TIa6a is a fundamental subfamily of the family (t of
all convex, circled, weakly compact subsets of E; in fact, if C E (t, then
paCC) E 6 a, since by virtue of 1, Pa is weakly continuous on E into Ea (rJ. E A),
and HxPa( C) E (t again by virtue of 1, above, and the Tychonov theorem which
asserts that any product of compact spaces is compact. Thus by (4.2) this
6-topology on E' is 7:(E', E).
138 DUALITY [Ch. IV
2. If 6~ denotes the family of all convex, circled, weakly compact subsets

of E~ (a E A), it suffices by (4.2) to show that 6' = EBa6~ is a fundamental
system of convex, circled subsets of E' that are compact for (1(E', E). If C
is such a set, C is bounded for '(1(E', E) and hence bounded for ,(E', E).
Thus by 3, above, and (II, 6.3) C is contained in EBaeHi\,(C), where H is a
suitable finite subset of A, and where Pa denotes the projection of E' onto E~.
Since Pa is continuous for (1(E', E) (Pa is even continuous for the coarser
topology induced on E' by na(1(E~, Ea)) into (E~, (1(E~, Ea)), it follows that
pi C) E 6~, which is the desired conclusion, since clearly every member of 6'
is convex, circled, and compact for (1(E', E).
This completes the proof of (4.3),
COROLLARY 1. Let {Ea: a E A} be afamily of l.c.s. and let E be their locally

convex direct sum. E' is algebraically isomorphic with naE~, and the following
topological identities are valid:
1. ,(E, E') = EBa,(Ea, E~).

2. ,(E', E) = nz,(E~, Ea).
3. (1(E', E) = na(1(E~, Ea).
Proof It follows readily from (11,6.1) that the dual E' of E can be identified
with n"E~ by virtue of the canonical duality between products and direct
sums; for the remaining assertions it is sufficient to interchange E and E'
in (4.3).
COROLLARY 2. The product, locally convex direct sum, and the inductive
limit of afamity of Mackey spaces is a Mackey space.
For products and direct sums the result is immediate from (4.3) and
Corollary I; for inductive limits it follows then from Corollary 4 of (4.1).
(4.3) and Corollary I supply an explicit characterization of various families
of bounded subsets in the dual of products and l.c. direct sums (Exercise 8;
cf. also the last part of the proof of (4.3)). In particular, if {E,,} is a family
of I.c.s. and S is an equicontinuous subset of the dual EB"E~ ofnaE", then the
projection p,,(S) is equicontinuous in E~ for each a, and every finite sum of
equicontinuous sets is equicontinuous in EBaE~. Thus from (II, 6.3) and
(4.3), 3, it follows that 6' = EB,,6~ is a fundamental family of equicon-
tinuous sets in EB"E~ if each 6~ is such a family in E~. A corresponding result
holds if " equicontinuous" is replaced by "weakly bounded"; thus, in view
of the characterization of equicontinuous sets in the dual of a barreled space
(Corollary of (1.6)), the following is proved:
COROLLARY 3. The product of any family of barreled spaces is barreled.
Finally we obtain a representation of the dual of a space of continuous

linear maps.
COROLLARY 4. Let E, F be I.c.s. and denote by 2.(E, F) the space of continuous

linear maps of E into F under the topology of simple convergence. The corres-
pondence I Xi ® y; --+ f, defined by
feu) = I <UXi' y;> (U E 2(E, F»,
is an (algebraic) isomorphism of E ® F' onto the dual of 2.(E, F).
Proof. If v = I Xi ® y;, the mapping v --+ f is obviously a linear map of
E ® F' into 2~, which is also biunivocal, since the bilinear form (v, u) --+ feu)
places even the subspace E' ® F of 2(E, F) (Chapter III, Section 7) in separ-
ated duality with E ® F' (Section 1, Example 4, above). There remains to
show that this mapping is onto 2~; since 2 5 (E, F) is a subspace of the
product space FE, every g E 2~ is the restriction of a continuous linear form
on FE hence, by (4.3), of the form
U --+ g(u) = I <UX i , Y;>

for suitable finite subsets {xJ c E and {y;} c F', which completes the proof.
We conclude this section with an application of the preceding results to
the duality between projective and inductive limits. Recall that a projective
limit E = lim gapEp (Chapter II, Section 5) is by definition a subspace of
TI pEp, namely
+-
the subspace n
a:?, 13
1
u~ (0), where uap = Pa - gap Pp whenever
0
rx ~ fJ. The projective limit E is called reduced if for each rx, the projection
paCE) is dense in Ea. There is no restriction of generality in assuming a pro-
jective limit to be reduced: Letting Fa = Pace) - (closure in Ea) and denoting
by iiaP the restriction of uap to TIpFp, E is identical with the subspace n
iiaP 1(0)
ofTIpFp• a:?,p
Denote by hpa the adjoint of gap with respect to the dualities <Ea, E~> and
<Ep, E[J>(rx ~ fJ); it follows (since gap is weakly continuous) from (2.4) that
hPa is continuous for the weak and Mackey topologies, respectively, on E[J
and E~. Moreover, gay = gap gpy (rx ~ f3 ~ y) implies hya = h)'p hPa by (2.2).
0 0
(Cf. Chapter II, Exercise 9.)
4.4
If E = lim gapEp is a reduced projective limit of l.c.s., then the dual E', under
....-
its Mackey topology ,(E', E), can be identified with the inductive limit of
the family {(E~, ,(E~, Eam with respect to the adjoint mappings hPa of gap'
Proof Let F = (f)aE~, where each E~ is endowed with ,(E~, Ea). By definition
lim IzpaE~ is the quotient space Fj H 0 (provided H 0 is closed in F), where Ho is
-~
the subspace of F generated by the ranges vp"(E~), where vPa = qa - qp hPa 0
(rx ~ fJ).
We show that Ho is the subspace of F orthogonal to E with respect to the
duality <TIaEa, F>. By Corollary 2 of (1.5) EO is the weakly closed, convex
140 DUALITY [Ch. IV
hull ofU [u;/(OW, which in view of the corollary of (2.3) is the same as the
a:$p
weakly closed, convex hull of U vpiF); this implies Ho c: EO. Conversely,
a;;;'p
let Y ~ (Ya) be an element of EO, let H be the finite set of indices such that
rx E H if and only if Ya =f. 0, and choose an index 13 such that rx ;;:; 13 for all
rx E H; finally let x be any element of E. Then we have
(x, y) = L (xa' Ya) = I
aeH aeH
(gaPxp, Ya) = L (XII' "PaYa)
aeH
since by assumption xp runs through a dense subspace of Ep as X runs

through E, the preceding relation implies that L
hpaYa = 0, hence Y =
aeH
L (qa -
aeH
qp ° hpa)Ya E Ho·
Thus Ho is weakly closed in F, hence closed for ,(F, TIaEa), which by
(4.3) is the topology EBa,(E~, Ea); thus the inductive limit lim hpaE~ of the
~
Mackey duals (E~, ,(E~, Ea» exists and by (4.1), Corollary 3, its topology is the
topology ,(FjHo, E), which proves it to be isomorphic with the Mackey
dual (E', ,(E', E» of E.
With the aid of (4.3), Corollary 1, and (4.1), Corollary 1, we now easily
obtain the following dual result for inductive limits:
4.5
Let E = lim hpaBa be an inductive limit of l.c.s. The weak dual of E is iso-
~
morphic with the projective limit of the weak duals (E~, a(E~, Ea» with respect to
the adjoint maps gap of hpa (rx ;;:; 13).
REMARK. If the duals E~ are endowed with their respective Mackey

topologies, then it follows from (4.3), Corollary 1, and (4.1), Corollary 3,
that the projective limit of these duals, algebraically identified with E',
carries a topology ~ which is consistent with (E, E'). Thus if~ is known
to be the Mackey topology (in particular, if ~ is metrizable), then the
Mackey dual of E can be identified with the projective limit of the
Mackey duals E~. See also Exercise 24.
s.STRONG DUAL OF A LOCALLY CONVEX SPACE. BIDUAL.

REFLEXIVE SPACES
Let (F, G) be a duality. Among the 6-topologies on F, generated by

families 6 of a(G, F)-bounded subsets of G, we have so far mainly considered
those consistent with (F, G), in particular, the weak and Mackey topologies
a(F, G) and ,(F, G). If 6 is the family of all weakly bounded subsets of G,
the corresponding 6-topology is called the strong topology on F (with respect
to (F, G», and denoted by f3(F, G). Since a weakly bounded subset of G
§5] STRONG DUAL OF A LOCALLY CONVEX SPACE 141
is not necessarily relatively compact for 0'(G, F), (J(F, G) is, in general, not
consistent with (F, G) (for an example, cf. Chapter II, Exercise 14; other
examples will become obvious from the discussion below).
Let E be a l.c.s. The topologies (J(E, E') and (J(E', E) are called the strong
topologies on E and E', respectively (usually without explicit reference to
the duality (E, E'»; (E', (J(E', E» is called the strong dual of E. The following
notation is more convenient: Let E a, E" Ep denote the space E under the
topologies O'(E, E'), 1:(E, E'), (J(E, E') respectively; accordingly, let E;, E~, E;
denote the dual E' of E under O'(E', E), 1:(E', E), (J(E', E), respectively. In
using this notation the reader should be cautioned that, in general, (Ep)' oF E'
and (E/J)' oF E (consequently, in general, E; oF (E/J)a). It follows from (4.2)
that the strong topology is inherited by products and locally convex direct
sums; however, it is not necessarily inherited by subspaces and quotient
spaces (Exercise 14). (See also end of Section 7.)
If the strong topology on E' is not consistent with (E, E'), it cannot be
expected that the respective families of bounded subsets of E; and E/J are
identical; at any rate, the following assertion is true.
5.1
Every convex, circled, compact subset of E; is bounded in E/J.
In fact, if C is such a set, its polar Co in E is a O-neighborhood for 1:(E, E'),
hence absorbs every bounded set BeE, which implies that B O absorbs
Coo = C.
Thus we have, in the dual E' of any I.c.s. E, the inclusions (t c (£: c ~ c ~'"
where (t denotes the family of all equicontinuous sets, (£: the family of all
sets with weakly compact closed, convex, circled hull, ~ the family of all
strongly bounded sets, and ~a the family of all weakly bounded sets in E'.
All four of these families are saturated in E;; this is obvious for ~a and ~
(notice that E; possesses a base of O'(E', E)-closed O-neighborhoods); for (t
it follows from (1.5) and its corollaries; for (£: it follows quickly from (II, 10.2).
In particular, we obtain the following strengthened version of the corollary
of (1.6).
5.2
Let E be a l.c.s. If (and only if) E is barreled, then the properties of being
equicontinuous, relatively weakly compact, strongly bounded, and weakly
bounded are eqUivalent for any subset of E'. Moreover, if (and only if) E is
barreled, then O-neighborhood bases and fundamental families of bounded sets
in E and in E/J correspond to each other by polarity with respect to (E, E').
The first assertion is clear from the preceding, since by (1.6), Corollary,
E is barreled if and only if (t = ~a. The second assertion is clear from (1.6),
since (E, ~) is barreled if and only if ~ = (J(E, E'), while the family of all
barrels in E~ is a O-neighborhood base for (J(E', E).
142 DUALITY [Ch. IV
There are locally convex spaces for which the families {t, (t, !S, !SO' are all
distinct (Exercise 15); hence the coincidence of some of them indicates
certain special properties. We have seen that (f = !SO' characterizes barreled
spaces, while, obviously, (f = (t characterizes Mackey spaces. By (1.6) a
subset BeE' is weakly bounded if and only if its absolute polar D is a barrel
in E; for B to be strongly bounded it is necessary and sufficient, by the bipolar
theorem, that D is a barrel absorbing all bounded sets in E (briefly, a bound
absorbing barrel). Thus {t =!S if and only if every bound absorbing barrel
in E is a neighborhood of 0; a I.c.s. with this property is called infrabarreled.
In particular, every bomological space (and, of course, every barreled space)
is infra barreled. Spaces with (t =!S will be discussed below. Let us note the
following sufficient condition for !S = !SO"
5.3
If E is a quasi-complete l.c.s., then !S = !SO'; equivalently, every weakly
bounded subset of E' is strongly bounded.
Proof The polar B O of every convex, circled, bounded subset BeE' is a
barrel in E by (1.6); hence BO absorbs every bounded subset of E which is
convex, circled, and complete, by (II, 8.5). But the bounded subsets of E
with these properties form a fundamental family of bounded sets, since E is
quasi-complete.
COROLLARY. Every quasi-complete infrabarreled space is barreled.
Let us consider an example. If (E, II II) is a normed space and
(E', II il) is the strong dual (as defined in Chapter II, Section 2) whose
norm is given by
Ilx'll = sup{l<x, x')I: Ilxll ~ I},
the topology on E' defined by this norm is clearly peE', E). Conversely,
if (and only if) E is barreled, its topology is peE, E'). Since E (whether it
it is barreled or not) is infrabarreled, the equicontinuous sets are those
bounded in (E', II II). By the bipolar theorem (or by a direct application
of the Hahn-Banach theorem (II, 3.2», it follows that
Ilxll = sup{l<x, x')I: Ilx'll ~ I};
hence the norm of E can be recovered from that of E'. In the latter
formula it suffices to take the supremum over any subset whose convex,
circled hull is u(E', E)-dense in the unit ball of E'. (5.2) above ex-
plains why it is not necessary to distinguish between weakly and strongly
bounded sets in the dual of a Banach space; if (E, II II) is not complete,
this distinction may well be necessary (Chapter II, Exercise 14).
It is clear from (1.2) that every element of E, by virtue of the duality
<E, E'),defines a continuous linear form on E/J. But the existence ofa bounded
set in E; which is not bounded in E/J indicates, by (3.2), Corollary 2, that
§5] STRONG DUAL OF A LOCALLY CONVEX SPACE 143
{3(E', E) is not consistent with <E, E'); hence, in general, the dual of E/J
cannot be identified with E. The dual of Ep is called the bidual of the l.c.s.
E and denoted by E". Thus if, as usual, E and E" are identified with subspaces
of the algebraic dual E'* of E', we have E c E" c E'*, where generally both
inclusions are proper. The algebraic isomorphism of E into E" thus defined is
called the canonical imbedding, or evaluation map of E into E", and is explicitly
given by x -> fx, wherefx is the linear form on E' defined by fix') = <x, x').
The bidual E" can be usefully topologized in several ways. If E" is given the
6-topology, where 6 = ~ is the family of strongly bounded subsets of E',
E" is called the strong bidual of E. The 6-topology, where 6 = (f is the
family of equicontinuous subsets of E', has the advantage of inducing the
given topology on E (for any l.c.s. E), and is often called the natural topology.
Since (f c ~, the natural topology is always coarser than the strong topology
of E"; for the identity of the two topoloiges (equivalently, for the evaluation
map to be a topological isomorphism into the strong bidual), it is evidently
necessary and sufficient that E be infrabarreled.
The following characterization of E" as a subspace of E'* is useful.
5.4
Consider the canonical inclusions E c E" c E'*, where E is a l.c.s. Then E"
is the union of the a(E*, E)-closures in E'* of all bounded subsets of E.
Proof Let us denote polars with respect to <E, E') by ° and polars with
respect to <E", E') by 0; also note that a(E'*, E') induces a(E", E') on E".
If Z E E", then {z}" is a O-neighborhood for {3(E', E) by definition of E";
hence B = {z}"o is a bounded, convex subset of E containing O. Now, clearly,
z E BOO = BOo, and by (1.5), B O• is the a(E", E')-closure of B. Since BOO is a
a(E", E)-closed, equicontinuous subset of E", it is a(E", E')-compact, hence
identical with the a(E'*, E)-closure of B in E'*. Conversely, let B be a
bounded, convex subset of E containing 0; then BOo, which is the a(E", E')
closure of B in E" by (1.5), is equicontinuous, hence a(E", E)-compact, hence
a(E'*, E)-closed in E'*.
COROLLARY. Every z E E" is the limit of a a(E", E)-Cauchy jilter possessing
a base of bounded subsets of E.
In particular, if E is a normed space and E" its strong bidual with the
standard norm, then the unit ball V of E is a(E", E)-dense in the unit ball
V of E". Also if E is any I.c.s. such that E/J is separable, then it follows from
Q
O
(III, 4.7) and (5.4), Corollary, that every Z E E" is the limit (for a(£", E'»
of a weak Cauchy sequence in E; if in addition E is normed, the members of
such a sequence can be assumed to be of norm ~ liz II.
A locally convex space E for which E = E" (more precisely, for which the
evaluation map is onto E") is called semi-reflexive; we note that this property
depends only on the duality <E, E'), and hence is shared by all or by none
144 DUALITY [Ch. IV
of the I.c. topologies on E that are consistent with (E, E'). It is important
to have a number of alternative characterizations of semi-reflexive spaces.
5.5
For any l.c.s. E, the following assertions are equivalent:
(a) E is semi-reflexive.
(b) Every P(E', E)-continuous linear form on E' is continuous for a(E', E).
(c) E~ is barreled.
(d) Every bounded subset of E is relatively aCE, E')-compact.
(e) E is quasi-complete under aCE, E').
Proof (a) ~ (b) is immediate from the definition of semi-reflexivity.
(b) ~ (c): (b) implies that the strong topology P(E', E) is consistent with
(E, E'), whence P(E', E) = r(E', E). Since by (1.6) each barrel in E~ (hence
by (3.1) each barrel in E;) is a O-neighborhood for P(E', E), it follows that E;
is barreled. (c) ~ (d): If E; is barreled then, again by (1.6), each bounded set
in E is equicontinuous as a subset of E = 2(E;, K o), hence relatively compact
for aCE, E'). (d) ~ (e) is immediate, since each compact subset of E" is
complete. (e) ~ (a): The corollary of (5.4) implies that each Z E E" is in E,
whence I/I(E) = E" where 1/1 denotes the evaluation map.
COROLLARY 1. Every semi-reflexive space is quasi-complete.
Proof By (1.6) and (5.5)(c) every bounded subset of E is equicontinuous
if E is viewed as the space 2 s(E;, K o), where 6 is the family of equicon-
tinuous subsets of E'; the assertion is hence a special case of (III, 4.4).
COROLLARY 2. Every bounded subset of a l.c.s. E is aCE, E')-precompact.
Proof Let B be a bounded, hence weakly bounded, subset of E. It is
evident that the space E~* = (E'*, a(E'*, E'» is semi-reflexive. Hence by
(5.5)(d) the closure Ii of B in E~* is compact, hence complete and thus
uniformly isomorphic with the completion of the uniform space B, which
shows B to be precompact in E".
Examples of semi-reflexive spaces are furnished by all quasi-complete
nuclear spaces, for by (III, 7.2), Corollary 2, every closed, bounded subset
of such a space is compact hence a fortiori weakly compact. The symmetry
between E and E' is not complete for semi-reflexive spaces (in particular, the
strong dual of a semi-reflexive space need not be semi-reflexive, since peE, E')
need not be consistent with (E, E') (see below», which is obviously due to the
fact that the definition of semi-reflexivity disregards the topological properties
of the evaluation map. By contrast, a I.c.s. Eis called reflexive if the evaluation
map is an isomorphism of E onto the strong bidual (Ep)p. Thus E is reflexive
if and only if it is semi-reflexive and its topology is peE, E'), i.e., if and only
if E is semi-reflexive and barreled. In view of (5.5), Corollary 1, and the
corollary of (5.3), the requirement that E be barreled can be replaced by the
is] STRONG DUAL OF A LOCALLY CONVEX SPACE 145
formally weaker requirement that E be infrabarreled, and the semi-reflexivity

can be replaced by anyone of the equivalent properties listed in (5.5) (e.g.,
weak quasi-completeness). Of all these possible characterizations of reflexive
spaces, we state the following which seems to be the most useful.
5.6
Theorem. A locally convex space E is reflexive if anti only if E is barreled
and every bounded subset of E is relatively compact for a(E, E').
The proof is covered by the preceding remarks.
COROLLARY 1. The strong dual of a reflexive space is reflexive.
Proof. If E is reflexive, then P(E', E) = r:(E', E) and Ep = E; is barreled by

(5.5), since E is semi-reflexive. Since, also, E is barreled, each bounded
subset of Ep is relatively compact for a(E', E) by (5.2); hence Ep is reflexive.
Simple examples show that the converse of Corollary 1 is false, but it can
be proved that if E is a quasi-complete Mackey space such that Ep is semi-
reflexive, then E is reflexive (Exercise 18).
COROLLARY 2. Every semi-reflexive normed space is a reflexive Banach
space.
In fact, if E is a semi-reflexive normed space, E is a Banach space by
(5.5), Corollary 1, hence also barreled.
Thus semi-reflexivity and reflexivity agree for normed spaces and, as has
been observed earlier. more generally for infrabarreled (in particular, borno-
logical) spaces. One might suspect from this that every semi-reflexive Mackey
space is reflexive, but this is false. Indeed let (E, Z) be a barreled space
which is not reflexive (e.g., a non-reflexive Banach space); E; is a semi-
reflexive Mackey space which is not reflexive, since its strong dual (E, Z) is
not reflexive. It is, moreover, clear from (5.5) that this example supplies all
semi-reflexive spaces that are not reflexive, in the sense that every such space
is isomorphic with the dual E' of a non-reflexive barreled space E, where E'
is supplied with a suitable l.c. topology consistent with (E, E'). The following
result will set in evidence even more clearly the relation between semi-
reflexivity and reflexivity, and the complete symmetry between E and E'
when E is reflexive.
5.7
Let E be a l.c.s. with dual E'. These assertions are equivalent:

(a) E. is reflexive.
(b) E; is reflexive.
(c) Ea and E; are both semi-reflexive.
(d) E. and E; are both barreled.
146 DUALITY [Ch. IV
Proof (a) ~ (b): If E, is reflexive, then [3(E', E) is consistent with (E, E,),
whence [3(E', E) = ,(E', E). Thus E; is the strong dual of E; hence E; is re-
flexive by Corollary I of (5.6). (b) ~ (c): If E~ is reflexive, then clearly E~ is
semi-reflexive; moreover, by (5.6) E; is barreled, hence every bounded subset
of E is relatively (I(E, E')-compact by (5.2), which shows E" to be semi-
reflexive by (5.5). (c) ~ (d): (c) implies that both [3(E', E) and [3(E, E') are
consistent with (E, E'), hence that [3(E', E) = ,(E', E) and [3(E, E') = r(E, E').
Since all barrels in E, and E;, respectively, are O-neighborhoods for [3(E, E')
and [3(E', E), it follows that E, and E; are barreled. (d) ~ (a): E, is barreled
and by (5.5) semi-reflexive, which implies that E, is reflexive.
Semi-reflexivity is inherited by closed subspaces (immediate from (5.5»,
but, in general, not by quotient spaces (Exercise 20); reflexivity is inherited,
in general, neither by closed subspaces (which may fail to be barreled) nor by
quotients (which may fail to be semi-reflexive) (Exercise 20). At any rate,
if E is a reflexive Banach space, then every closed subspace and every separated
quotient of E is a reflexive Banach space (the latter being true, since every
bounded subset of the quotient is the canonical image of a bounded set in E,
and hence reflexivity is preserved). Moreover, it follows from (4.1), Corollary
3, and the relation [3(E', E) = r(E', E) (which is characteristic of semi-
reflexive spaces by (5.5» that if M is a closed subspace of a semi-reflexive
space E, the strong dual of M can be identified with Ep(Mo.
The situation is less complicated in the case of products and locally convex
direct sums.
5.8
Let {Ea: ex E A} be a family of semi-reflexive (respectively, reflexive) i.c.s.;

both the product and the locally convex direct sum of this family is semi-
reflexive (respectively, reflexive). Moreover, the projective limit of any family of
semi-reflexive l.c.s. is semi-reflexive, and the strict inductive limit of a sequence
of reflexive spaces is reflexive.
Proof Via (5.5)(d) it follows from (4.3), (4.3), Corollary 1, and from the
characterizations (I, 5.5) and (II, 6.3) of bounded subsets of products and
locally convex direct sums, respectively, that semi-reflexivity is preserved in
both cases. Thus by (5.6) reflexivity is also preserved in both cases, since any
product of barreled spaces is barreled, (4.3), Corollary 3, and since any
locally convex direct sum of barreled spaces is barreled, (II, 7.2), Corollary 1.
The assertion concerning projective limits is now clear since a projective
limit of a family of I.c.s. is a closed subspace of their product; the assertion
concerning inductive limits is clear from the characterization of bounded
subsets (II, 6.5) of a strict inductive limit of a sequence of I.c.s., and from
(II, 7.2), Corollary 1. It is also clear that mere semi-reflexivity is preserved
in this case.
§6] DUAL CHARACTERIZATION OF COMPLETENESS 147
Examples
1. Let U(J.L) be the Banach space introduced in Example 2 of Chapter
II, Section 2. If 1 < p < + 00, it is well known (cf. Day [2]) that the
strong dual of U(J.L) is norm isomorphic with U(J.L) , wherep-l + q-l = I,
the canonical bilinear form of the duality being ([fl, rg]) -+ Jfg* dJ.L.
Thus if I < p < + 00, U(J.L) is reflexive; in particular, the Banach spaces
1:(1 < p < + (0) are reflexive (Exercise 18).
2. Every Hilbert space is reflexive, for every such space H is isomor-
phic with a space If, where d is the Hilbert dimension of H (Chapter
II, Section 2, Example 5).
3. The one-dimensionall.c.s. Ko associated with the scalar field K is
reflexive. Hence by (5.8) the spaces rod and <Pd of Section 1, Example 4,
are reflexive and the strong duals of one another under the canonical
duality of products and direct sums.
4. It has been observed earlier that by virtue of (III, 7.2), Corollary 2,
every quasi-complete nuclear space E is semi-reflexive; if, in addi-
tion, E is barreled (in particular, if E is a nuclear (F)-space), then E is
reflexive. Since all spaces enumerated in Chapter III, Section 8, are bar-
reled, each of these spaces furnishes an example of a non-normable
reflexive space.
The spaces of Chapter III, Section 8, are locally convex spaces which are
not only reflexive but such that every closed, bounded subset is compact for
the strong topology. A reflexive l.c.s. in which every closed, bounded subset
is compact, is called a Monte) space, or briefly (M)-space. The permanence
properties of (M)-spaces are virtually the same as for reflexive spaces; in
particular, it is evident that any product, any I.c. direct sum, and every strict
inductive limit of a sequence, of (M)-spaces is again an (M)-space. The
same is true for the strong dual:
5.9
The strong dual of a Mantel space is a Mantel space.
Proof. If E is a Montel space, then E/J is reflexive by (5.6), Corollary 1.
Since E is barreled, every strongly bounded subset of E' is equicontinuous;
thus if B is a strongly bounded and closed, convex subset of E', B is (1(E', E)-
compact, and hence by (III, 4.5), compact for the topology ;tc of compact
convergence. But ;tc = P(E', E), since, E being a Montel space, every closed,
bounded subset of E is compact.
6. DUAL CHARACTERIZATION OF COMPLETENESS. METRIZABLE

SPACES. THEOREMS OF GROTHENDIECK, BANACH-DIEUDONNE,
AND KREIN-SMULIAN
Let E be a locally convex space. What can be said about the completeness
of E' for a given 6-topology? Dually, what can be said in terms of E' about
148 DUALITY [Ch. IV
complet~ness properties of E1 Let us examine the weak dual E;. The algebraic
dual E* is complete for (1(E*, E), since it is a closed subspace of the complete
space K8. On the other hand, E' is dense in E! by (1.3), Corollary, and E!
induces (1(E', E) on E'; hence E! is (isomorphic with) the completion of E;.
Thus E; is complete if and only if E' = E*, and Etl is complete if and only if
E = E'*; it is not difficult to infer that if E is an infinite dimensional metri-
zable I.c.s., E; is never complete (cf. Exercises 6, 21). Before proving
Grothendieck's dual characterization of completeness, we record the following
completeness properties of the duals of barreled and bomological spaces.
6.1
Let E be a l.c.s. If E is barreled, E' is quasi-complete for every 6-topology,
where 6 is a family of bounded sets covering E. If E is bomological, the strong
dual E/J is complete.
Proof. If E is barreled, by (5.2) the bounded subsets of E' are the same for
all 6-topologies in question and each bounded set is equicontinuous; the
first assertion is hence a special case of (III, 4.4). Now every strong Cauchy
filter in E' converges pointwise to an element f E E*, and f is bounded on
bounded subsets B c.E, since on each B the convergence is uniform; hence
if E is bomological, we have fEE' by (II, 8.3).
REMARK. Komura [1] asserts that there exist reflexive spaces that are
not complete; this shows that, in general, the strong dual of a barreled
space is not complete. On the other hand, the preceding result on bomo-
logical spaces can be considerably strengthened (Chapter III, Exercise
8): If 6 is a family of bounded subsets of a bomological space E such
that the range of each null sequence in E is contained in a suitable S E 6,
then E' is complete for the 6-topology.
The following basic theorem is due to Grothendieck [1].
6.2
Theorem. Let (E, ~) be a l.c.s. and let 6 be a saturated family of bounded
sets covering E. For E' to be complete under the 6-topology, it is necessary and
sufficient that every linear form f on E which is ~-continuous on each S E 6,
he continuous on (E, ~). .
Proof The condition is sufficient. For let t:s: be a Cauchy filter in E' with
respect to the 6-topology; since 6 covers E, t:s: converges pointwise to a
linear formfe E*, the convergence being uniform on each S E 6. Hence for
each S the restrictionfs off to S, being the uniform limit of ~-continuous sca-
lar functions on S, is continuous for ~; by hypothesis it follows thatfe E'.
The condition is necessary. For this it suffices to show that each fe E*
such that the restriction fs (S e6) is ~-continuous can be approximated,
uniformly on S, by elements geE'. Let S e 6 and 8 > 0 be given; since 6 is
§6] DUAL CHARACTERIZATION OF COMPLETENESS 1~
saturated, we can assume S is convex, circled, and (1(E, E')-closed. Let

f E E* and let fs be Z-continuous at 0 E S; there exists a convex, circled,
closed (hence (1(E, E')-closed) Z-neighborhood U of 0 such that If(x) I ~ 8
whenever XES n U. This is equivalent to fE 8(S nUt, where the polar
is taken with respect to <E, E*). Since Sand U are a fortiori (1(E, E*)-closed,
it follows from (1.5), Corollary 2, that (S n ut
is contained in the (1(E*, E)-
closure of UO + So. But UO is compact and So closed for (1(E*, E), hence by
(I, 1.1) UO + So is closed, and hence f E e( UO + SO). It follows that for some
g E eUo c E', f - g E eSo, which means that If(x) - g(x)1 ~ e whenever
XES, completing the proof.
COROLLARY 1. Let <F, G) be a duality and let 6 be a saturated family,
covering F, of weakly bounded subsets of F. Denote by GI the vector space of
all f E F* whose restrictions to each S E 6 are weakly continuous, and endow
GI with the 6 -topology. Then GI is a complete l.c.s. in which G is dense.
Proof Supposing, for the moment, that for each fE GI and S E 6, f(S)
is bounded, it is clear that GI is a 1.c.s. under the 6-topology (Chapter III,
Section 3), and obviously complete. Now if G denotes the closure of Gin GI ,
G is complete under the 6-topology; hence (6.2) shows that G = GI . To
complete the proof it suffices to show thatf(S) is bounded if fE GI and S
is a convex, circled member of 6. Since f is weakly continuous at 0 E S, and
since x, YES imply t(x - y) E S, the identity If(x) - f(y) I = 2lf(x - y)/21
shows that f is uniformly weakly continuous on S. Hence f(S) is bounded,
since S is weakly precompact by (5.5), Corollary 2.
Let us note also that if S is a closed, convex, circled (not necessarily
bounded) subset of a 1.c.s. (E, 2) andf is a linear form on E whose restriction
to S is 2-continuous at 0, then f is uniformly weakly continuous on S; in
fact, by the second part of the proof of (6.2), fs is the uniform limit of uni-
formly weakly continuous functions on S.
COROLLARY 2. The following propositions on a l.c,s. E are equivalent:
(a) E is complete.
(b) Every linear form on E' which is a(E', E)-continuous on every equi-
continuous subset of E' is a(E', E)-continuous on all of E'.
(c) Every hyperplane H in E' such that H n A is weakly (a(E', E)-)closed
in A for each equicontinuous subset A of E', is closed in E;.
Proof (a) ¢>(b) is immediate from (6.2), since the topology of E is the
6-topology, where 6 is the (saturated) family of equicontinuous subsets of
E;, (1.5), Corollary 3, and E is the dual of E;. (c) => (b): Let H = {x' E E':
f(x') = ex}; then H n A = {x' E A:f(x') = ex} is weakly closed in A if the
restriction fA is a(E', E)-continuous. (b) => (c): Assume that H = {x' E E':
f(x') = ex} is a hyperplane in E' such that H n A is closed in A for (1(E', E)
whenever A is equicontinuous. To show thatfis a(E', E)-continuous on A,
150 DUALITY [Ch. IV
it suffices to show that j~ is continuous at 0 E A whenever A is convex and

circled. Notice first that since H n A + x~ = (H + x~) n (A + x~) for all
x~ E E', and since A + x~ is equicontinuous if A is, {x' E A: f(x') = /3} is
closed in A for each /3 E K, and equicontinuous set A. If fA were not con-
tinuous at 0 E A, there would exist an infinite subset B c A such that 0 E 13
and such that f(x) = /30' /30 i= 0, for all x EB, which conflicts with the fact
that {x' E A:f(x') = /3o} is closed in A.
COROLLARY 3. Let E be a separable, complete l.c.s. and let f be a linear form
on the dual E'. For f to be aCE', E)-continuous it suffices that limnf(x~) = 0
whenever {x~} is a null sequence for a(E', E).
Proof. This is immediate from Corollary 2, since by (1.7), the equicontinuous

subsets of E' are metrizable for the topology induced by aCE', E). (It is even
sufficient that f converge to 0 on every weak null sequence whose range is
equicontinuous. )
The preceding results lead one to ask whether there exists, on the dual E'
of a I.c.s. E, a finest topology ::rI which agrees with a(E', E) on each equi-
continuous set. If so, then by Grothendieck's theorem (6.2), Corollary 1,
the completion of E consists exactly of the linear forms on E' that are con-
tinuous for ::r / . It is indeed easy to see that such a topology exists: If (!) is
the family of all subsets of E' such that for every G E (!) and equicontinuous
set A, G n A is open in A for aCE', E) then (\) is evidently invariant under the
formation of finite intersections and arbitrary unions; hence (!) is the family
of open sets for the topology in question. It is clear, moreover, that this
topology is translation-invariant (since aCE', E) and the family <r of equi-
continuous subsets of E' are translation-invariant), and has a O-neighborhood
base consisting of radial and circled sets (Exercise 22). However, in general,
property (a) of (I, 1.2) fails (Komura [I]); it would hold automatically if 'XI
were necessarily locally convex, but this is false as an earlier example due to
Collins [1] shows (Exercise 22). It is hence of interest to determine cases in
which (E', 'XI) is a t.v.s. and, in particular, in which it is a I.c.s. (Note that 'XI
is Hausdorff, since it is finer than aCE', E) and, in fact, finer than the topology
of precompact convergence, (III, 4.5).) It was known to Banach that ::rl
is the topology of compact convergence if E is a Banach space; Dieudonne [1]
proved the result by a new method which is applicable to metrizable I.c.s.
The following lemma is the critical step in the proof.
LEMMA. Let E be a metrizable l.c.s., {Un: n EN} a O-neighborhood base in E
consisting of a decreasing sequence of closed, convex sets and let G be an open
'Xrneighborhoodof 0 in E'. There exists a sequence {Fn: n = 0,1,2, ... } of
non-empty finite subsets of E having these properties:
(i) Fn c Un(n EN).
(ii) H~ n U~ c G(n EN) where Hn = U Fk •
k<n
§6] DUAL CHARACTERIZATION OF COMPLETENESS 151
Proof. The proof uses induction with respect to n. Supplementing the base
{Un: n EN} by setting U0 = E and assuming that the sets Fk C Uk (k = 0, 1,
... , m - I) have been selected to satisfy (ii) for n = 1, ... , m, we shall show that
there exists a non-empty finite set Frn C Ultl such that (FlO U Hmt n U~+ 1 C G.
To include the existence proof for Fo in the general induction step, we set
Ho = 0; then clearly (ii) is satisfied for n = O.
Now suppose that (F u Hrnt n U~+ 1 is not contained in G for any non-
empty finite subset Fe Urn. Since G- = E ~ G is :Irclosed and U~+ 1 is
equicontinuous in E', the set G1 =G- n U~+l is closed in U~+l for the
topology induced by :If' hence for the topology induced by aCe', E); this
implies that G1 is aCe', E)-compact. Now the non-empty finite subsets
Fe Urn are directed upward by inclusion; hence the sets (Fu Hrnt n G1 =
FO n H~ n G1 form a filter base of closed subsets of G1 • From the compact-
ness of G1 it follows that the intersection of all these sets contains an element
x', which is consequently an element of U~ n H~ n G1 ; for by Remark 3
n
preceding (1.4), U~ = PO, where F runs through all non-empty finite subsets
Fe Urn. This is contradictory, since U~ n H~ c G by assumption and since
G1 c G-; the lemma is proved.
The following is the theorem of Banach-Dieudonne.
6.3
Theorem. Let E be a metrizable l.c.s. and let 6 be the family of sets

formed by the ranges of all null sequences in E. The 6 -topology is the topology
of pre compact convergence and the finest topology on E' that agrees with
aCe', E) on each equicontinuous subset of E'.
Proof. Since the range S of any null sequence is relatively compact in E,
hence the closed, convex, circled hull of S precompact, (II, 4.3), the 6-
topology is coarser on E' than is the topology :Ipc of precompact convergence.
On the other hand, the topology :If is finer than :Ipc by (III, 4.5); hence it
suffices to show that the 6-topology is finer than :If' In view of the translation
invariance of :If this is immediate from the lemma above: If G is an open
00
:Irneighborhood of 0 in E', let S = U Fm where Fn are as in the lemma, then

o
SO n U~ c G for all n EN, and hence So c G, sInce U U~ = E'. Since, clearly,
1
S E 6, the proof is complete.
COROLLARY I. If E is a metrizable l.c.s., each precompact subset of E is
contained in the closed, convex, circled hull of a suitable null sequence.
COROLLARY 2. If E is an (F)-space, the topology :If on E' is the topology of

compact convergence (hence consistent with <E, E'»).
It is now easy to prove the theorem of Krein-Smulian (Krein-Smulian [1]).
152 DUALITY [Ch. IV
6.4
Theorem. A metrizable l.c.s. E is complete if and only if a convex set
Me E' is aCE', E)-closed whenever M (\ Vo is aCE', E)-closed for every
O-neighborhood V in E.
Proof Since every hyperplane H in E' is convex, the sufficiency of the
condition is clear from (6.2), Corollary 2. To prove its necessity, let E be
complete and let M be a convex subset of E' such that M (\ Vo is aCE', E)-
closed for every O-neighborhood V in E. Since :l:f is consistent with <E, E')
by Corollary 2 above, it suffices to show that Mis :l:rclosed. Denote by M-
the complement of M in E'; the assumption implies that M- (\ Vo is open in
Vo for aCE', E) and hence for :l:f. Since every equicontinuous set AcE' is
contained in a suitable V o , it follows that M- (\ A is open in A for :l:f' and
hence M- is open for :l:f.
COROLLARY. Let E be a Banach space, let M be a subspace of E', and let B be
the dual unit ball {x': Ilx' I ~ I}. If M (\ B is closed in E;, then M is closed in E;.
Proof The assumption implies that p(M (\ B) = M (\ pB is closed in E'
for all p > 0 so that (6.4) applies. (It is evidently sufficient that the condition of
(6.4) be satisfied for the members V of an arbitrary O-neighborhood base in E.)
Thus the weak dual of an (F)-space possesses a number of striking proper-
ties; in contrast with this, and in contrast with the strong dual of a Banach
space, the strong dual of an (F)-space E has a structure, in general, much
more complicated than that of E. For instance, if E is a metrizable I.c.s.,
then E/J is not metrizable unless E is normable; in fact, E/J possesses a funda-
mental sequence {Bn} of bounded sets which we can assume are convex,
circled, and closed; if E/J (which is complete, (6.1» were metrizable, one of
the Bn would be a O-neighborhood and hence E/J would be normable. Thus
the strong bidual would be normable and hence E would be normable, since
E is infrabarreled (Section 5). Also, Eft need not be barreled (hence not infra-
barreled or bomological), even if E is an (F)-space; for an example we refer
the reader to Exercise 20 (see also Grothendieck [10] and Kothe [1]). Never-
theless, the fact that the strong dual of a metrizable I.c.s. possesses a funda-
mental family of bounded sets which is countable, has some important
consequences which will be derived now. In the proofs of (6.5) through (6.7)
we follow Kelley-Namioka [1].
6.5
If E is a metrizable l.c.s. and {Vn } is a sequence of convex O-neighborhoods
E/J such that V = n Vn absorbs strongly bounded sets, then V is a O-neighbor-
ro
in
1
hood in Eft.
Proof Let {Vn: n EN} be a O-neighborhood base in E; then the sequence of
polars Bn = V~ constitutes a fundamental family of strongly bounded
16] DUAL CHARACTERIZATION OF COMPLETENESS 153
(equivalently, equicontinuous) subsets of E'. It is sufficient to show that

V contains a set W which is radial, convex, and closed in E;; for then W is
the polar of a bounded set in E, hence a O-neighborhood in E/J. By hypothesis
there exists, for each n EN, Pn > 0 such that 2PnBn c V and a barrel Dn in
n
E; such that 2Dn c Vn • Now the convex hull Cn of U PkBk is O'(E', E)-compact
1
by (II, 10.2), and hence Wn = Cn + Dn is convex and O'(E', E)-closed by (I, 1.1);
n Wn absorbs each Bn and
00
obviously Wn c Vn. It is easily verified that W =

1
hence is radial; clearly, W is convex, O'(E', E)-closed, and contained in V.
The property of Ep expressed by (6.5) is evidently equivalent to the follow-
ing: If {Mn} is a sequence of equicontinuous subsets of E" such that M = UnMn
is strongly bounded (i.e., P(E", E')-bounded), then Mis equicontinuous. Since
every finite subset of E" is equicontinuous, we obtain the following corollary.
COROLLARY 1. Every countable, bounded subset of the strong bidual E" of a
metrizable l.c.s. E is equicontinuous.
In view of (III, 4.3) and the obvious fact that the strong bidual of a metri-
zable l.c.s. E is again metrizable, this implies:
COROLLARY 2. The strong bidual E" of a metrizable l.c.s. E is an (F)-space,
and semi-complete for O'(E", E"').
(6.5) can also be viewed as a weakened form of the property defining
bornological spaces. As has been observed above, the strong dual of a
metrizable (hence bornological) l.c.s. is not necessarily bornological, but the
following proposition is valid.
6.6
For the strong dual of a metrizable l.c.s. E, the following properties are
equivalent:
(a) E/J is bornological.
(b) E/J is infrabarreled.
(c) E/J is barreled.
Proof (a)~(b) is immediate (and true for any I.c.s., Section 5), (b)~(c)
follows from the corollary of (5.3), since E/J is complete by (6.1). (c) ~ (a): It
suffices to show that each convex circled subset C c E' that absorbs strongly
bounded sets contains a barrel in E/J. Let {Bn} be a fundamental sequence of
O'(E', E)-compact, convex, circled equicontinuous sets; since each Bn is
strongly bounded (in fact, {Bn} is a fundamental family of strongly bounded
sets), there exists Pn > 0 such that 2PnBn c C (n EN). Let Cn be the convex
n
hull of U PkBk; each Cn is convex, circled, and O'(E', E)-compact by (II, 10.2),
1 00
and Co = U Cn is a convex, circled, radial set such that 2Co c C. Thus the
1
154 DUALITY [Ch. IV
proof will be complete if we show that Co c 2eo, the closure being taken for
I3(E', E). Let x' ¢ 2eo' Then for each n EN, there exists a convex circled
O-neighborhood Vn in E' such that (x' + Vn ) (") en = 0, for en is strongly
n Wn is a strong O-neighbor-
a)
closed. Let Wn = Vn + en' Then by (6.5) W =

1
hood, since Wabsorbs each Bn (n EN). On the other hand, x' ¢ 2eo implies
(x' + Wn) (") en = 0 for all n EN, whence (x' + W) (") eo = 0; it follows
that x' ¢ CO.
COROLLARY 1. The strong dual of every reflexive Frechet space is bornological.
This is immediate from (5.6) and (5.6), Corollary 1.
COROLLARY 2. If the strong dual of a metrizable l.c.s. is separable, then it is
bornolog ical.
Proof. It is sufficient to show that Ep is infrabarreled. Hence let D be a
barrel in Ep that absorbs strongly bounded sets, and let {x~} be a countable
dense subset of E' '" D. By (II, 9.2) there exists a closed (real) semi-space Hn
such that x:¢ Hn and D is contained in the interior of Hn' Hence Den Hn
U = n Hn is a O-neighborhood in Ep in view of (6.5).
n
and Now the comple-
n 0 0
ment of the interior U of U contains {x~} and is closed; hence D => U, which
proves the assertion.
The property (6.5) of strong duals of metrizable spaces led Grothendieck
[10] to introduce a special class of locally convex spaces: A l.c.s. E is called a
(DF)-space if E possesses a fundamental sequence of bounded sets, and if
every strongly bounded countable union of equicontinuous subsets of E' is
equicontinuous. Every strong dual of a metrizable l.c.s. is a (DF)-space, but
not conversely; the class comprises all normable spaces and, more generally,
all infrabarreled spaces possessing a fundamental sequence of bounded sets
(Exercise 24). For a detailed study of these spaces the reader is referred to
Grothendieck [10]; these spaces also playa considerable part in the theory
of topological tensor products (Grothendieck [13]). As an important example
of the properties of (DF)-spaces, we prove that the topology of a (DF)-space
can be "localized" in a fashion analogous to the localization of the topology
of compact convergence in the dual of an (F)-space.
6.7
Let E be a (DF)-space. A convex, circled subset V of E is a neighborhood of
o if (and only if) for
every convex, circled bounded subset BeE, B (") V is a
O-neighborhood in B.
Proof. The necessity of the condition being trivial, suppose that {Bn} is an
increasing sequence of bounded, convex, circled sets which is fundamental,
and that for each n, Bn (") V is a O-neighborhood in Bn. There exists, for each
n EN, a convex, circled O-neighborhood Un in E satisfying Bn (") Un C Bn (") V.
§7] ADJOINTS OF CLOSED LINEAR MAPPINGS 155
Let Wn be the closure of Bn n V + lUn; then Wn c: Bn n V + Un' hence

Bn n Wn c: Bn n V + (2Bn) n U" c: Bn n V + 2(B" n U,,) c: 3(Bn n V);
n W n, it follows
<Xl
letting W = that Bn n W c: 3Bn n 3V which, in view of
<Xl <Xl 1
U Bn = U 3B" = E, implies that W c: 3 V. On the other hand, W is closed,
1 1
convex, circled, and absorbseachBII,sinceforsuitablep" > O,BII c: p,.(B"n UII)
<Xl
c: PII(Bn n V) c: Pn(Bn+p n V) c: PIIW,,+P for all p eN. Hence U W~ is

1
strongly bounded and therefore equicontinuous in E'. It follows that Wand
hence V:::> t W is a O-neighborhood in E.
COROLLARY. A linear map of a (DF)-space E into a I.c.s. F is continuous if

its restriction to each bounded subset of E is continuous.
7. ADJOINTS OF CLOSED LINEAR MAPPINGS

Let E, Fbe I.c.s., let Eo be a dense subspace of E, and let u be a linear map
with domain Du = Eo and with values in F. Consider the set Fa of elements
y' e F' for which the linear form x -+ <ux, y') is continuous on Eo; Fa is
non-empty, since it contains 0, and is, clearly, a subspace of F'. Since Eo is
dense in E, the form x -+ <ux, y') has for each y' e Fa a unique continuous
extension to E, which is thus an element x' e E'. Let us denote this mapping
y' -+ x' by v; obviously, v is a linear map with domain D. = Fa and it follows
that the relation
<ux, y') = <x, vy')
is an identity on Du x D•. If u is continuous on Eo into F, then, clearly.
D. = F' and (since the dual of Eo can be identified with E') v is the adjoint
of u as defined in Section 2; it is thus consistent to call the mapping v just
defined the adjoint map (briefly, adjoint) of u. Hence every linear map u,
defined on a dense subspace of E (for convenience, we shall say that u is
densely defined in E), with values in F possesses a well-defined adjoint with
domain D. c: F'; notice that Dv = (u*) -l(E') n F', where u* is the algebraic
adjoint of u and E' is identified with a subspace of E6. We are asking for
conditions under which Dv is dense in F~. Recall that the map u is called
closed if its graph G ={(x, ux): x e Du} is a closed subspace of Ex F; in
view of (4.3) and the convexity of G, this is equivalent to G being closed in
E" x F".
7.1
Let E, F be I.c.s. and let u be a linear map densely defined in E with values in F.
The graph of the adjoint v of u is closed in F~ x E~; for the domain Dv to be
dense in F~, it is necessary and sufficient that u have a closed extension. If this
is the case, then the adjoint it ofv is the smallest closed extension ofu.
156 DUALITY [Ch. IV
Proof Note first that by (4.3), the dual of E x F can be identified with
E' x F', where the canonical bilinear form on (E x F) x (E' x F') is given by
«x, y), (x', y'» = <x, x') + <y, y'); further, the mapping x: (y', x')-+
(-x', y') is an isomorphism of F~ x E~ onto E~ x F~. Let G be the graph of
u and let H be the graph of v; the identity <ux, y') - <x, vi) = 0, valid on
Du x D v , shows that GO = X(H); hence X(H) is closed in E~ x F~ and, there-
fore, H is closed in F~ x E~.
Assume now that Dv is dense in F~. Then the graph G1 of the adjoint it of v
is closed in E" x F" (hence in E x F) by the preceding, and, clearly, Du
contains Du; thus it is a closed extension of u. Also X'(G 1) = HO, where X'
denotes the adjoint of X (Section 2); in view of the fact that Hand G1 are
closed, it follows from (2.3)(a),(c) that X'(G 1 ) = H O is equivalent with
G1 = X(Ht. Thus G1 = GOo which, by the bipolar theorem (1.5), shows G1 to
be the closure of G in E x F; hence it is the smallest (in an obvious sense)
closed extension of u.
Since the domain of the adjoint of an extension of u is contained in Do,
the proof will be complete if we show that Dv is dense in F~ whenever u is
closed. Let y E D~; then (y, 0) E H O and (0, y) E X(Ht. If G is closed, then
X(Ht = GOo =G by the bipolar theorem, and (0, y) E G implies y = 0;
hence Dv is dense in F~.
We shall extend the notation introduced in Section 2 and denote by u'
the adjoint of u whenever u is a densely defined closed linear map.
COROLLARY. Let u be a closed linear map, densely defined in E, with values in
F. Then u' has the same properties with respect to F~ and E~, and UN ="(u')' = u.
It is often convenient to reduce the study of a linear map u to the case
where u is biunivocal; if u is a continuous linear map on a t.v.s. E into a
t.v.s. F with null space N, the biunivocal map Uo of EIN into F associated
with u (Chapter III, Section 1) is continuous (and conversely). Likewise, a
linear map u is open if and only if Uo is open. It is a useful fact that closedness
and density of domain is reflected in a similar manner by the associated
biunivocal maps. Only in the following statement, E, F are not assumed to
be I.c.s.
7.2
Let E, F be t.v.s., let u be a linear map with domain Du c E, null space
N c D u , and range in F. Denote by ¢ the canonical map E -+ EIN, and by Uo
the biunivocal map of ¢(Du) into F associated with u. Then the graph of Uo is
closed in (EIN) x F if and only if the graph of u is closed in E x F. Moreover,
if u is densely defined, then so is Uo; and if u is closed and F is separated, then
N is closed in E.
Proof We identify (EIN) x F canonically with (E x F)/(N x {On and
denote by ¢1 the quotient map E x F -+ (E x F)/(N x {O}). If G and Go
are the respective graphs of u and uo, it is evident that <PI (G) = Go; since
N x {O} c: G, it follows that G = <pll(Go) and G'O = <PI(G-), where -
denotes complementation. Since <PI is continuous and open, the two latter
relations show that G is closed exactly when Go is. If Du is dense in E, then
<p(Du ) is dense in EIN, since <P is continuous and onto. If F is separated
(equivalently, if {O} is closed in F), then N x {O} is clearly closed in G, hence
in E x F if G is closed. Thus N is closed in E in this case. This completes the
proof.
Thus if u is a closed linear map with domain Eo dense in E and values in
F, the mapping U o on <p(Eo) (which can be identified with EoIN) into F is
again closed and densely defined in EIN. If u is canonically decomposed,
u = I/t Uo <Po, where <Po: Eo ..... Eol Nand I/t: u(Eo) ..... F are the canonical
0 0
maps, it follows, as before, that u is continuous (respectively, open) if and

only if U o is continuous (respectively, open). Weakly open linear maps with
closed graph are characterized by the following dual property; E, F are again
supposed to be l.e.s.
7.3
Let u be a linear map, densely defined in E, into F with closed graph. Then u
is weakly open if and only if the range of its adjoint u' is closed in E~.
Proof. By (7.2) and Corollary 2 of (4.1), it can be supposed that u is one-
to-one and onto F; denote by Eo the domain of u. If u is weakly open, then
u- I is continuous for (1(F, F') and (1(Eo, E'); hence u- 1 has an adjoint v
which maps E' into F'. Since it is clear that u' is one-to-one and that v =
(u')-t, it follows that the range of u' is E' and hence is closed. Conversely, if
the range H of u' is closed in E~, then H = E'. For x E H (polar with respect
O
to (E, E'» implies x E Du = Eo in view ofu" = u; hence x = 0 because Du' is

dense in F~ and u is one-to-one; it follows that H = {O} and H = H = E'.
O OO
Now if U = {x E Eo: I(x, xi)1 ;;; 1, i = 1, ... , n} is a weak O-neighborhood

in Eo, there exist elements y[ E D u ' such that u'(YD = xi for all i; it follows
that u(U) = V, where V = {y E F: I<y, y;)1 ;;; 1, i = 1, ... , n}, and hence that
u is weakly open.
COROLLARY. Let u be a weakly continuous linear map of E into F. For u to
be a weak homomorphism, it is necessary and sl!jficient that its adjoint u'
have a closed range in E~.
In view of (2.l), the foregoing results can be given a more symmetrical
form: Let u be a closed linear map with range in F and domain dense in E.
Then u is weakly continuous if and only if the domain of u' is closed in F~,
and u is weakly open if and only if the range of u' is closed in E~.
The following proposition relates the continuity properties of a linear map
u with properties of its adjoint v; u is not assumed to be closed, and v is the
adjoint ofu, with domain Dv = F~, as defined at the beginning of this section.
158 DUALITY [Ch. IV
7.4
Let E, F be l.c.s., let u be a linear map of E into F, and let v be the adjoint
of u. Consider these statements:
(b l ) u is continuous for the Mackey topologies on E, F.
(b 2 ) U is weakly continuous.
(b 3 ) v is defined on all of F'.
(b 4 ) v is continuous on F~ into E;.
(c) v is continuous on F/J into Ep.
One has the implications: (b 1 )<0> (b 2 )<0> (b 3 )<0> (b 4 ), (a)=(b l ) and (b 4 ) = (c).
If F is semi-reflexive, then (c) = (b 4 ), and if E is a Mackey space, then
(b l ) = (a).
Proof. (a) = (b 2 ) and (b l ) = (b 2 ): For each y' E F', x ~ (ux, y') is a con-
tinuous, hence weakly continuous linear form on E, which is equivalent to the
weak continuity of u. (b 2 ) = (b 3 ): Immediate from (2.1). (b 3 ) = (b 4 ): Since
u(E) c F and u is the adjoint of v with respect to the dualities (E, E') and
<F, F'), v is continuous for (J(F', F) and (J(E', E) by (2.1). (b 4 ) = (b l ): v maps
convex, circled, compact subsets of F; onto convex, circled, compact subsets
of E;, and u is the adjoint of v; the assertion follows from (2.4). (b 4 ) = (c):
It suffices to prove (b 2 ) = (c), and this is immediate from (2.4), since u maps
bounded sets onto bounded sets. If F is semi-reflexive, then Fp = F; by
(5.5) and hence v is continuous on F; into E p, and a fortiori for the Mackey
topologies on F' and E'; by the proven equivalence of (b l ) and (b 2 ), it follows
that (c) = (b 4 ) when F is semi-reflexive. If E is a Mackey space, then, clearly,
(b l ) = (a), since the topology of F is coarser than reF, F').
Among the consequences of the preceding result, let us note the following:
If u is a linear map of a Mackey space E into a I.c.s. Fwith an adjoint defined
on all of F', u is necessarily continuous. (This contains a classical theorem
due to Hellinger and Toeplitz: Every ~elf-adjoint transformation of a Hilbert
space which is defined everywhere, is continuous.) The following proposition
gives a dual characterization of topological homomorphisms.
7.5
Let E, F be l.c.s. with respective duals E', F', and let u be a closed linear map
of E into F. Then u is a topological homomorphism if and only if the domain
and range of u' are closed in F; and E;, respectively, and u' maps the equi-
continuous subsets of F onto the equicontinuous subsets of its range.
Proof. Denote by Nand M, respectively, the null space and range of u.
We decompose u = t/J U o ¢ where, as usual, ¢ is the quotient map E ~ Ej N
0 0
(note that N is closed), t/J the canonical imbedding M ~ F. u is a topological

homomorphism if and only if Uo is an isomorphism for the quotient and

induced topologies l:1 on E/N and l:2 on M, respectively.
From (4.1), (a) => (b) it follows that l:2 is the 6 r topology, where 6 2 is
the family of all canonical images in F' / M a of equicontinuous subsets of F';
similarly by (4.1), (c) => (d), l:1 is the 6 c topology, where 6 1 is the family of
all equicontinuous subsets of E' that are contained in N a •
The condition is necessary. For if U o is an isomorphism of (E/N, l:1) onto
(M, l:2), then by (7.4) Uo is an isomorphism for a(E/N, W) and a(M, F'/M a ),
and U o is a weak isomorphism of F'/M a onto N a • Clearly, Du ' is closed in
F; since D u' = F' by the continuity of u, and u'(F') is closed in E;, since
u'(F') = u~(F'/MO) = N°. Moreover, U~(62) = 6 1 by (2.4) as asserted.
The condition is sufficient. By the observation following the corollary of
(7.3), the assumption implies that u is a weak homomorphism; hence, since
a(E/ N, N°) is the quotient of aCE, E'), Uo is an isomorphism of E/ N onto M
for the topologies a(E/N, W) and a(M, F'/MO). By (7.4) u~ is an isomorphism
of F'/Mo onto N° for the topologies a(F'/Mo, M) and a(N°, E/N), and we have
U~(62) = 6 1 by hypothesis; it follows from (2.4) that both Uo and ui)l are
continuous for l:1 and l:2' and hence u is a topological homomorphism.
COROLLARY. Every topological homomorphism of E into F is a weak homo-
morphism.
This is immediate from the corollary of (7.3). The converse of the last
corollary is false as can be seen considering the identity map E, --+ E", where
E is a I.c.s. for which teE, E') #: aCE, E'). But even for the Mackey topologies
on E and F, a weak homomorphism of E into F is not necessarily open
(Exercise 26). The remainder of this section will provide conditions under
which it can be concluded that a continuous linear map is open, assuming
this property for the weak topologies, or of the adjoint.
7.6
Let E be a Mackey space and let u be a weak homomorphism of E into F
such that the subspace u(E) of F is a Mackey space. Then u is a topological
homomorphism of E into F.
Proof Let N = u-1(O) and M = u(E); by (7.5) it suffices to show that

whenever B is a convex, circled, compact subset of E; contained in N°,
then B = u'(A), where A is equicontinuous in F'. Since M is a Mackey space
under the topology induced by F, (4.1) (b)=>(a) implies that every convex,
circled, a(F'/Mo, M)-compact subset Bl of F'/Mo is the canonical image of
an equicontinuous subset A of F'. The assertion follows now from the fact
that the map u~ is an isomorphism of F'/Mo onto N° for the topologies
a(F'/Mo, M) and a(N°, E/N), since Uo is a weak isomorphism of E/N onto
M by hypothesis.
160 DUALITY [Ch. IV
The preceding result applies, in particular, whenever F is a metrizable

l.c.s-., (3.4). When both E and F are (F)-spaces, we obtain the following set
of equivalences with the aid of Banach's homomorphism theorem.
7.7
If E, Fare Frlichet spaces and u is a continuous (equivalently, closed) linear
map of E into F, the following properties of u are equivalent:
(a) u is a topological homomorphism.
(b) u is a weak homomorphism.
(c) u has a closed range.
(d) u' is a homomorphism for (1'(F', F) and (1'(E', E).
(e) u' has a range closed for (1'(E', E).
Proof (a)=-(b) by (7.5), Corollary, and by (7.6). Corollary 1 of (III, 2.1)
shows that (a)=-(c). Finally, (b)=-(e) and (c)=-(d) are immediate from the
corollary of (7.3).
However, the equivalent conditions of (7.7) do not imply that u' is a homo-
morphism for the strong topologies on F' and E', even if u(E) = F; the
reason for this is the fact that, letting N = u- 1 (O), P(N°, EIN) is in general
strictly finer than the topology on N° induced by P(E', E) (Exercise 14). In
the converse direction the following assertion can be made (Dieudonn6-
Schwartz [1 D.
7.8
If E, Fare Frechet spaces, and u is a continuous linear map of E into F whose
adjoint u' is an isomorphism of F/J into E/J, then u(E) = F (hence u is a homo-
morphism).
Proof Since u' is one-to-one, u(E) is dense in F; hence in view of (7.7),
(e) => (c) it suffices to show that u'(F') is closed in E;. Let Ube any O-neighbor-
hood in E; U O is closed and bounded in E; and in E/J. Since u' is a strong iso-
morphism, B = (U')-l(U is bounded in F/J (hence in F;) and is closed in F;
O
)
by the weak continuity of u'; thus B is (1'(F', F)-compact. It follows that u'(B) =
U II u'(F') is compact, hence closed in E~. Since U was arbitrary, ul(F') is
O
closed in E~ by the Krein-Smulian theorem (6.4).

Let us add a few remarks on normed spaces. If E and Fare normed spaces
with respective unit balls Band C, the norm u --. Ilu II = sup{ Ilux II: x E B} is
the natural norm on 2(E, F); it generates the topology of bounded con-
vergence (Chapter III, Section 3). As has been pointed out earlier, the spaces
E/J, F/J, and consequently 2(F/J, E/J), are Banach spaces under their natural
norms. If u E 2(E, F), then u' E 2(F/J, E/J) by (7.4) and, by the bipolar
theorem,
Ilu'll = sup{l<x, u'y')I: x E B, y' E CO}
= sup{l<ux, y')I: x E B, y' E CO} = lIull;
§8] THE GENERAL OPEN MAPPING AND CLOSED GRAPH THEOREMS 161
hence u -+ u' is a norm isomorphism of 2(E, F) into 2(F/J, E/J), a fact that
has been used repeatedly in Section 9 of Chapter III. Let us note the following
facts on the strong duals of subspaces and quotients of a normed space E:
(i) If M is a subspace of E, the Banach space Mp is norm isomorphic with
the normed quotient EPiMo.
(ii) If N is a closed subspace of E, then the Banach space (EIN)'p, strong
dual of the normed space EI N, is norm isomorphic with the subspace N° of the
Banach space E/J.
If 1/1, ¢ denote the canonical maps M -+ E and E -+ EIN, respectively, the
norm isomorphisms in question are provided by the biunivocal map I/I~
associated with 1/1' (Chapter III, Section 1) and by ¢', respectively; the detailed
verification is omitted. In the present circumstances, P(N°, EIN) is induced by
P(E', E) and P(E'IMO, M) is the quotient of peE', E). For normed spaces
E, F (and, of course, for normable spaces E, F), we obtain the following
supplement of (7.7) (cf. Exercise 27):
7.9
Let E, F be normed spaces and u E 2(E, F). If u is a homomorphism, then its
adjoint u' is a strong homomorphism; the converse is true whenever E is complete.
Proof. If u is a homomorphism and N = u- 1(O), then the biunivocal map
Uo associated with u is an isomorphism of EI N onto M c: F, where M = u(E).
It follows that Uo is an isomorphism of F'IMo onto N° for P(F'IMO, M) and
P(N°, EIN), which shows, in view of (i) and (ii) above, that u' is a strong
homomorphism.
Conversely, if u' is a strong homomorphism, then G = u'(F') is closed in
E/J. To show that G is closed in E~ (equivalently, that G = N°) it suffices, by
the corollary of (6.4), to prove that G n B O(B the unit ball of E) is a(E', E)-
closed. Since u~ is a strong isomorphism of F'I MO onto G and each strongly
bounded subset of F' IMO is the canonical image of a strongly bounded subset
of F', we have G n BO = u'(H), where H is a strongly bounded, hence equi-
continuous subset of F'. But the weak closure H is compact in F~; since u' is
weakly continuous, u'(H) is compact in E~ and obviously equals G n B O ,
which completes the proof.
8. THE GENERAL OPEN MAPPING AND CLOSED GRAPH THEOREMS
A continuous linear map of E onto F is necessarily open whenever E and F

are Frechet spaces: This is the essential content, for locally convex spaces, of
Banach's homomorphism theorem (Chapter III, Section 2). Even though
(III, 2.2) shows that this important result continues to hold when E, Fare
(LF)-spaces (hence neither metrizable nor Baire spaces, cf. also Chapter III,
Exercise 4), the homomorphism theorem, and with it the closed graph
theorem, for a long time appeared to be intrinsically of category type and
162 DUALITY [Ch. IV
on the whole unrelated to the recent theory of locally convex spaces-in

particular, unrelated to duality. Ptak [1], [7] was the first to recognize an
intimate relationship with the theorem of Krein-Smulian and, more generally,
with the dual characterization of completeness. We present in this section
the essential results of the theory and some related results due to Collins [1]
and Mahowald [1].
We define a I.c.s. E to be B-complete (or a Ptak space) if a subspace Q e E'
is closed for aCE', E) whenever Q n A is aCE', E)-closed in A for each equi-
continuous set A e E'; E is said to be Br-complete if every dense subspace Q
of E; such that Q n A is aCE', E)-closed in A for each equicontinuous set
A eE', is closed in E~ (hence identical with E'). It is immediate that in both
definitions the family of all equicontinuous subsets of E' can be replaced by
any fundamental subfamily, in particular, by the po lars Uo of the members
U of a O-neighborhood base in E.
Examples
I. Every (F)-space E is B-complete by the theorem of Krein-Smulian,
(6.4).
2. The Mackey dual E; of an (F)-space E is B-complete (direct
verification); in particular, the strong dual of a reflexive (F)-space is B-
complete. (For concrete examples, see Chapter III, Section 8.)
3. Every weakly complete I.c.s. is B-complete. In fact, one has
E = E'* (see beginning of Section 6) which shows that r(E', E) is the
finest locally convex topology on E' (Chapter II, Section 6 and Exercise
7); hence every subspace Q of E' is closed for r(E', E) and, by (3.1),
*
for aCE', E). Note that E is weakly complete if and only if it is iso-
morphic with a product K of one-dimensional spaces Ko (cf. Exercise 6).
4. It will be seen shortly that every closed subspace and every sep-
arated quotient of a B-complete space is B-complete.
8.1
Every B-complete space is Br-complete, and every Br-complete space is
complete.
Proof. The first assertion being trivial, suppose that E is Br-complete and
that H is a hyperplane in E' such that H n A is aCE', E)-closed in A for each
equicontinuous set A e E'. To show that H is closed, it suffices to show that
some translate of H is closed in E~; hence we can assume that 0 E H. If H
were not closed, it would be dense by (I, 4.2) and hence H = E', since E is
Br-complete, which is impossible; thus H is closed and the assertion follows
from (6.2), Corollary 3.
The following result is due to Collins [I].
8.2
Every closed subspace of a Ptak space (respectively, Br-complete space) is a
Ptak space (respectively, Br-complete).
§8] THE GENERAL OPEN MAPPING AND CLOSED GRAPH THEOREMS 163
Proof. Let M be a closed subspace of E. We identify the dual of M with

E'/Mo and conclude from (4.1), Corollary 1, that the weak topology a(E'/Mo,
M) is the quotient of a(E', E); denote by 1// the quotient map E' -+ E' / MO.
We give the proof for both assertions simultaneously, understanding by Q
any subspace of E'/Mo or a weakly dense subspace of E'/Mo, accordingly as
E is assumed to be a Ptak space or only Br-complete.
Let Q be such that Q (") VO is weakly closed in E' / M O for any V E mand
let mbe a O-neighborhood base in M; we have to show that Q is closed for
a(E'/Mo, M). We can assume that m= {U (") M: U E U}, where U is a base of
closed, convex O-neighborhoods in E; it follows that I//(UO) = VO for
V = U (") M, U E U, by (1.5), Corollary 2. P = (1//) -I( Q) is a subspace of E';
since 1// is continuous, (I//)-I(V0 (") Q) = (UO + MO) (") P is closed in E;.
Since UO is compact for aCE', E), UO is closed in UO + M O, and hence UO (") P
is closed in (UO + MO) (") P. Since the latter set is closed in E;, so is UO (") P.
Hence P, which is dense if Q is, is closed in E; by hypothesis and since
M O c P, it follows that l{!'(P) = Q is closed for the quotient topology
a(E'/Mo, M), which completes the proof.
Next we prove Ptak's central result. To formulate it conveniently, the
following definition is useful: Let E, F be t.v.s.; a linear map u of E into F
is called nearly open iffor each O-neighborhood U c E, u(U) is dense in some
O-neighborhood in u(E). Evidently u is nearly open if and only if it maps each
open subset GeE into the interior (taken in u(E» of u(G) (cf. Exercise 28).
If E, Fare l.c.s., then for u to be nearly open it suffices that for each convex
O-neighborhood U c E, u(U) be weakly dense in some O-neighborhood in
u(E). Note also that every linear map u of a l.c.s. E onto a barreled space F
is nearly open.
8.3
Theorem. Consider the following properties of a locally convex space E:
(a) E is a Ptak space.
(b) Every continuous, nearly open linear map of E into any l.c.s. F is a
topological homomorphism.
(c) E is Br-complete.
(d) Every biunivocal, continuous, and nearly open linear map of E into any
l.c.s. F is an isomorphism.
Then (a)~(b) and (c)~(d).
Proof. It will be sufficient to prove the equivalence of (a) and (b); a proof
of (c)~(d) will then be obtained, in view of the corollary of (2.3), by re-
stricting u to biunivocal maps and Q to dense subspaces of E'.
(a) => (b): Let u be a continuous, nearly open linear map of E into F; we
can assume that u(E) = F. If N = u-1(O) and Uo is the continuous linear map
of E/N onto F associated with u, then U o is nearly open. Since E -+ E/N is
164 DUALITY [Ch. IV
open, it suffices to show that U o is open. Let U be any closed, convex 0-

neighborhood in E/N; if V = uo(U), then Vis a O-neighborhood in F. Denote
by u~ the adjoint of Uo; by the corollary of (2.3) Q = u~(F') is a dense subspace
of N° c E;. By (2.3) (a) we have
(U~)-l(uo) = [uo(U)]O = VO = Vo.
Thus Vo is closed and equicontinuous, and therefore compact in F;; u~ being
continuous for u(F', F) and u(E', E), it follows that u~(V°) = Uo ('\ Q is
compact and hence closed in E~. U being arbitrary, the assumed B-complete-
ness of E implies that Q is cl('<;ed in E; (hence Q = N°); thus Uo is a weak
isomorphism by the corollary of (7.3). Since U is closed in E/N, it is weakly
closed (by convexity), which implit:s that V = uo(U) is weakly closed in F,
whence V = V; it follows that U o is open.
(b) => (a): Let U be the family of all convex, circled O-neighborhoods in
E, and let Q be a subspace of E' such that Q ('\ UO is u(E', E)-closed for each
U E U; we have to show that Q is closed in E;. Denoting polars with respect
to <E, E') by 0, let F be the quotient space E/ QO without topology; the
canonical bilinear form on F x Q places F and Q in duality (Section 4).
Consider the family ID3 ={(UO ('\ Qt: U E U} of subsets of E and the locally
convex topology Z on F for which q,(ID3) is a neighborhood base of 0, where
q, is the quotient map E -+E/Qo. With respect to the duality <F, Q), Z is the
6-topology on F, where 6 = {UO ('\ Q: U E U}; by assumption, each S E 6
is closed, hence compact in E; and, therefore, compact for u(Q, F). Clearly,
6 covers Q and the family of all subsets of sets S E 6 is saturated under
u(Q, F); hence (3.2) implies that Z is consistent with <F, Q). Now if UE U,
the polar of q,(U) with respect to <F, Q) i~ Q ('\ UO, whence q,[(UO ('\ QtJ is
the Z-closure of q,( U) by the bipolar theorem, which shows that q, is a nearly
open linear map of E onto (F, Z). Since Z is coarser on F than the quotient
topology of E/Qo, q, is also continuous and hence open by hypothesis; this
proves Z to be the quotient topology of E/Qo. From (4.1), Corollary 1 and
Corollary 2, it follows now that Q is necessarily closed in E;.
The theorem is proved.
Among the corollaries of the foregoing result, the most striking is the
following which extends Banach's classical theorem substantially with
respect to both the domain and the range spaces.
COROLLARY 1. (Homomorphism Theorem). Every continuous linear map of a
Ptak space E onto a barreled space F is a topological homomorphism.
Proof. Since u(E) = F, for each convex, circled O-neighborhood U c E
the closure of u(U) is a barrel in F, whence u is nearly open. Thus u is open
since E is a Ptak space.
COROLLARY 2. Let E be a Ptak space and let F be a I.c.s. such that u(E) = F
for some continuous, nearly open linear map u; F is a Ptak space.
§8l THE GENERAL OPEN MAPPING AND CLOSED GRAPH THEOREMS 165
Proof Let G be any l.c.s. and let v be a continuous, nearly open linear map
of Finto G. Since u is open, v °u is nearly open and hence also open, E being a
Ptak space; G and v being arbitrary, it follows that F is a Ptak space.
Applying Corollary 2 to the canonical map E -+ EI N, where N is a closed
subspace of E, we obtain
COROLLARY 3. Every separated quotient of a Ptak space is a Ptak space.

In other words, B-completeness is preserved by quotients over closed
subspaces, a property not shared by the weaker concept of completeness
(Exercise 11). For an enlightening discussion of these questions the reader is
referred to Ptak [7].
To obtain a correspondingly generalized closed graph theorem, we need an
open mapping theorem for not necessarily continuous linear maps with closed
graph. The following general open mapping theorem, which contains (8.3),
Corollary 1, as a special case, is also due to Ptak.
8.4
Theorem. Let E be a Ptak space and let u be a nearly open linear map
with domain dense in E and range in any I.c.s. F, and such that the graph of u is
closed in E x F. Then u is open. If, in addition, u is one-to-one it suffices for
the conclusion that E be B,-complete.
Proof The general case can be reduced to the case in which u is one-to-one,
for by (7.2) the null space N of u is closed in E, EIN is a Ptak space by (8.3),
Corollary 3, and the biunivocal map associated with u is by (7.2) densely
defined, has a closed graph, and is obviously nearly open if u is. We assume,
hence, that u is biunivocal with domain Eo dense in E and graph closed in
E x F, that E is B,-complete, and finally that u(Eo) is dense in F, which is
clearly no restriction of generality.
Let u' be the adjoint of u with domain F~ dense in F;, (7.1), and let Q =
u'(Fo); Q is dense in E; (cf. proof of (7.3». Denote by U any closed, convex
O-neighborhood in E; then Uo = (U 11 Eot by (1.5), since U 11 Eo is dense,
hence weakly dense in U. By (2.1) u is continuous for a(Eo, Q) and a(F, F o),
and u' is continuous for a(Fo, F) and a(Q, Eo). Letting V = u(U 11 Eo), the
closure V is a O-neighborhood in u(Eo), since u is nearly open by hypothesis;
if VO is the polar of V with respect to <F, F'), then Vo c Fo, for y' E.VO
implies that x -+ Re <ux, y') is ~ 1 on U Il.eo and hence continuous on
Eo. It follows that VO is compact for a(Fo, F); hence u'(VO) = UO 11 Q is
compact for a(Q, Eo) by the continuity of u' for these topologies. Since Eo
is dense in E and UO c E' is equicontinuous, (III, 4.5) implies that UO 11 Q is
compact for aCE', E). Since E is B,-complete by hypothesis, Q is closed in
E; (hence Q = E'), which by (7.3) implies that u is weakly open, or equiva-
lently that u- 1 is weakly continuous. Since V (by convexity) is the weak
166 DUALITY [Ch. IV
closure of V in u(Eo), we have

Un Eo = u- 1(V) C u- 1(V) C Un Eo,
for Un Eo is weakly closed in Eo. Hence u(U n Eo) = V = V, which proves
u to be an open map, thus completing the proof.
The following general closed graph theorem (cf. Robertson-Robertson
[I)) is now an easy consequence of the preceding results.
8.5
Theorem. Let E be barreled and let F be Br-complete. If u is a linear map of
E into F with closed graph, then u is continuous.
Proof. As in the foregoing proof, the general case can be reduced to the
case in which u is one-to-one. For N = u-\O) is closed in E and the bi-
univocal map of EIN into F associated with u has a closed graph by (7.2),
and EI N is barreled; it can moreover be assumed that u(E) is dense in F,
since by (8.2), each closed subspace of a Br-complete space is Br-complete.
Suppose, hence, that u is one-to-one onto a dense subspace of F. Then since
u is closed, u- 1 is a closed linear map, densely defined in F, onto E, which is
nearly open, since E is barreled; by (8.4) u -1 is open, hence u continuous.
COROLLARY. Let E be a l.c.s. that is barreled and Br-complete. If E is the
algebraic direct sum of two closed subspaces M, N, then the sum is topological:
E=MEeN.
Proof. Let p be the projection of E onto M vanishing on N. Since M is
Br-complete by (8.2), it suffices to show that the graph of p is closed in
E x M. But this is quickly seen to be equivalent with the closedness of N.
We conclude this brief account with a result, due to Mahowald [1], sug-
gesting that the open mapping and clo!:ied graph theorems, in their general
forms (8.4) and (8.5), have been extended to the natural limit of their validity.
8.6
Let E be a l.c.s. such that for every Banach space F, a closed linear map of
E into F is necessarily continuous; then E is barreled.
Proof. Let D be any barrel in E, and denote by 4> the canonical map of E
into the Banach space ED (for notation, see Chapter III, Section 7); one has
D = {x E E: 114>(x) II ~ I}, since D is closed. We have to show that D is a
neighborhood of 0; in view of the hypothesis on E, it suffices to show that the
graph G of 4> is closed in E x ED.
Now if (x o, Yo) ¢ G, then li4>(xo) - Yo II> 2e for a suitable e > 0; since
4>(E) is dense in ED, there exists Yl E 4>(E) such that Ilyo - Ylll < e, and it
follows that 114>(xo) - Ylll > e. The set A = {x E E: 114>(x) - Ylll ~ e} is
closed in E, since it is a translate of eD; hence W = E ~ A is open. Denoting
by B. the open ball {y E ED: Ily - Ylil < e}, it follows that (W x Be) n G = 0·
Since (x o, Yo) E W x Be' we conclude that G is closed in E x ED.
§9] TENSOR PRODUCTS AND NUCLEAR SPACES 167
9. TENSOR PRODUCTS AND NUCLEAR SPACES

We consider briefly the relations between spaces of linear mappings
(Chapter III, Section 3) and spaces of bilinear forms (Chapter III, Section 5).
~et E, F be l.c.s. with respective duals E', F' ; as before, we shall use subscripts
when referring to the most frequent topologies (as in E:, E;, E/J etc. (cf.
Section 5».
If v is a separately continuous bilinear form on E x F, it is quickly seen that
the formula
vex, y) = <ux, y) = <x, u'y)
defines a continuous linear map u E ff'(E, F:) with adjoint u' E ff'(F, E:)
(cf. (7.4»; in fact, the mappings v --+ u and v --+ u' are algebraic isomorphisms
of ~(E, F) onto ff'(E, F:) and of ~(E, F) onto ff'(F, E:), respectively. By
virtue of these isomorphisms, which are called panonical, we shall frequently
identify spaces of linear maps with spaces of 'bilinear forms. We note that
under v --+ u the space fJB(E, F) of continuous bilinear forms on E x F is
carried onto the subspace of ff'(E, F:), each of whose elements maps a
suitable O-neighborhood in E onto an equicontinuous set in F'.
Furthermore, to a given 6 x X-topology on~(E, F) (Chapter III, Section
5) there corresponds on ff'(E, F:) the topology of uniform convergence on 6
with respect to the X-topology on F'. Conversely, to a given 6-topology on
ff'(E, F) (which is, in general, a proper subspace of ff'(E, F » there corres-
eT
ponds the 6 x X-topology on the canonical image of ff'(E, F) in ~(E, F:),

X denoting the family of all equicontinuous subsets of F'.
However, from (7.4) it follows that ff'(E, F = ff'(E" F); hence ff'(E, F)
eT )
can be identified with~(E, F;) if E is a Mackey space. Dually, ff'(E;, F) =

ff'(E:, F hence ff'(E;, F) can be identified with ~(E:, F:), and if 6 is the
eT ) ,
family of all equicontinuous subsets of E', then the 6-topology on ff'(E;, F)

corresponds to the topology of bi-equicontinuous convergence on ~(E:, F:);
under this topology, ff'(E;, F) will be denoted by ff'e(E;, F).
9.1
Let E, F be l.c.s. not equal to {OJ. The space ~e(E:, F:) (equivalently,
ff'.(E;, F» is complete if and only if both E and F are complete. In this case
E ® F can be identified with the closure 01 E ® Fin ff' e(E;, F).
Proof Let I be a bilinear form on E' x F' which is the limit of a filter in
~(E:, F:), uniformly on each product S x T where S, T are arbitrary equi-
continuous subsets of E', F' respectively. It follows that for each y' E F', the
partial map .t;" is (J(E', E)-continuous on S, whence I y ' is a continuous linear
form on E: by (6.2), Corollary 2, if E is complete; likewise lx' (x' E E') is
continuous on F; if F is complete, which proves the condition to be sufficient.
Conversely, if ~e(E:, F:) is complete and x E E, Y E F are non-zero elements,
then the closed subspaces x ® F (note that x ® F is closed in ff'e(E;, F» and
168 DUALITY [Ch. IV
E ® y of'13e<E;, F;) are complete, and isomorphic with F and E respectively;

hence E and F are complete if '13.(E;, F;) is complete. The final assertion is
also clear, since E <§) F is by definition the completion of the subspace E ® F
of'13e<E;, F;) (Chapter III, Section 6).
Recall that the dual of E ® F for the projective (respectively, inductive)
tensor product topology is ~(E, F) (respectively, '13(E, F)); our next objective
is to determine the dual of E ® F for the topology of bi-equicontinuous
®
convergence (equivalently, the dual of E F). To begin with, a base of 0-
neighborhoods for this topology is formed by the polars, with respect to the
duality between E ® F and E' ® F' c B(E, F) (cf. Chapter III, end of Section
6), of the sets S ® T, where S, T are arbitrary equicontinuous subsets of
E', F' respectively. Hence the bipolars (r S ® Tr in B(E, F) (equivalently,
the closures (r S ® T)- for a(B(E, F), E ® F)) form a fundamental family of
equicontinuous sets in the algebraic dual B(E, F) of E ® F. Since, clearly,
each set S ® T is equicontinuous for the projective topology, the dual
/(E, F) of E <§) F is a subspace of ~(E, F).
Recall also (Chapter II, Section 2, Example 3) that a Radon measure on a
compact space X is a continuous linear form 11 E .,H(X) = 'b'(X)', where
'b'(X) is the Banach space of scalar-valued continuous functions on X. It is
customary to write Il(f) = <J, 11) = Sxfdll; 111111 denotes the norm of 11 in the
strong dual of 'b'(X). In the following proposition, S x T is the compact
product of Sand T under the induced weak topologies.
9.2
The dual /(E, F) of E <§) F consists exactly of those elements v E ~(E, F)
f
that can be represented in the form
u ..... v(u) = <u, v) = uo(x', y')dll(x', y'),

SXT
where S, T are suitable closed, equicontinuous subsets of E;, F; respectively,

and where Uo is the restriction of the bilinear form u on E' x F' to S x T.
If A is an equicontinuous subset of /(E, F), the elements v E A can be repre-
sented with S x T fixed and 11 running through a norm bounded subset of
.,H(S x T).
Proo.f Since each u E E ® F, viewed as a bilinear form on E' x F', has a
restriction to S x Twhich is continuous for the topology induced by E; x F;,
the same is true for any u E E ®F, since the latter are limits, uniformly on
each product S x T of equicontinuous sets, of elements of E ® F. Hence
Uo E 'b'(S x T), and the integral defines a linear form v on E ®F. Moreover,
®
if W denotes the O-neighborhood in E F which is the polar (with respect to
®
the duality <E F, E' ® F'») ofrS ® T, it followsthatlv(u)1 ~ 1IIlIlwhenever
u E W; hence v E /(E, F).
For the remainder it suffices to prove that any equicontinuous set

A c feE, F) can be represented as asserted. If A is equicontinuous, there
exist compact equicontinuous subsets S, T of E;, F;, respectively, such that
A c (r r
S @ T)0o = We, where W is the polar of S@ T, hence a O-neighbor-
hood in E®F. Consider the map u-+ Uo of E®Finto <'(j(S x T); the associ-
®
ated map 1/1 of (E F)w (Chapter III, Section 7) into <'(j( S x T) is a norm
isomorphism. Hence the adjoint 1/1' maps .,H(S x T) homomorphically
onto [feE, F)]w' by (7.9), and by (7.5) the set A, being equicontinuous in
[feE, F)]w', is the image under 1/1' of an equicontinuous subset of ,,H(S x T),
The elements of feE, F) are called integral bilinear forms on E x F, and
the linear maps U E geE, F;) and u' E g(F, E;) originating from a bilinear
form v E feE, F) are called integral linear maps. It follows that an integral
f
map U E geE, F;) is of the form
x-+u(x) = <x,x')y'd/1(x',y')
SxT
for suitable weakly closed, equicontinuous sets S, T. The integral Sfd/1

(called weak integral) is defined as the linear form y -+ S<y,J)d/1 in F, which is
certainly continuous whenever f is a continuous function on S x T into F;
such thatf(S x T) is contained in a compact, convex, circled subset of F;.
I
00
It is evident from the preceding that a nuclear map x -+ Ai<X, x;)y;

i= 1
of Einto F; (cf. (III, 7.1» is integral if the sequence {ya is equicontinuous in
F'. We shall see shortly that whenever E is a nuclear space, and F is any I.c.s.,
then feE, F) = f!J(E, F) and every integral map in geE, F;) is nuclear. We
first prove a dual characterization of nuclear spaces for which, in turn, the
following lemma is needed.
LEMMA 1. Let E, F be l.c.s. and let U be a linear map of E into F which maps
a suitable O-neighborhood in E into a weakly compact subset of F. Then the
second adjoint u" maps the bidual E" into F c F".
Proof. Let U be a convex O-neighborhood in E such that u(U) c C, where C
is weakly compact in F. Since u" is continuous for a(E", E') and a(E", F'), and
since E is a(E", E)-dense in E" by the bipolar theorem, it follows that u"
is the continuous extension of u to E" into F" with respect to the topologies
a(E", E') and a(E", F'). Since U is a(E", E)-dense in UOo and C is weakly
compact, it follows that u"(UOO) c C. But UOo, being a O-neighborhood for the
natural topology of E", is radial, whence u"(E") c F.
Let us point out that in the remainder of this section, frequent use will be
made of the notational devices introduced at the beginning of Section 7 of
Chapter III (cf. also Chapter III, Exercise 3). In particular, whenever E is a
I.c.s. and B a convex, circled, and bounded subset of E;, we shall write E~ in
place of [E'lB; we set E~ = {O} if B = 0.
170 DUALITY [Ch. IV
9.3
A l.c.s. E is nuclear if and only if for each closed, convex, circled subset
AcE; which is equicontinuous, there exists another set B with these properties,
and such that A c B and the canonical imbedding E~ ~ E~ is nuclear.
Proof The condition is necessary. In fact, U = AO is a convex, circled
O-neighborhood in E; if E is nuclear, there exists by (III, 7.2) a convex,
circled O-neighborhood V c U such that the canonical map ¢u v: 2v ~ Bu
is nuclear. Let B = Va; it is immediate that A c B and that l{IB,A: EA ~ E~,
being the adjoint of ¢u,v, is nuclear.
The condition is sufficient. Let Ube a given convex, circled O-neighborhood
in E; by (III, 7.2) it suffices to show that there exists another such O-neighbor-
hood V for which V c U and ¢u,v is nuclear. Put A = UO; by hypothesis,
there exist closed, convex, circled, and equicontinuous subsets B, C of E;
such that A c Bee and such that the canonical maps l{I C,B and l{I B,A are
both nuclear. Let W = BO, V = Co (polars with respect to (E, E'», and
c,
denote by F, G, H the strong duals of E E~, E A, respectively; F, G, Hare
the respective strong biduals of B v , 2 w, Bu and the second adjoints ¢'W,v
and ¢~,w are evidently nuclear. By Lemma 1 ¢~,w maps G into B u, since
¢u,w is compact, being the restriction to Bw of a nuclear (hence compact)
L Anf.. ® Yn' where (An) ElI, {Yn}
00
map. Now by (III, 7.1) ¢'W,v is of the form

1
is a bounded sequence in G, and {j~} is an equicontinuous sequence of linear
forms on F. Let Zn = ¢~,w(Yn) and let gn be the restriction of1. to Bv (n EN);
then {zn} is bounded in Bu and {gn} equicontinuous on Bv. Since ¢u,v is the
L Angn ® Zn and hence nuclear,
00
restriction of ¢~,w °¢'W,v to B v , it is of the form

1
9.4
Theorem. Let E be a nuclear space, let F be any locally convex space,
and endow E ® F with its projective tensor product topology. The canonical
imbedding of E ® F into !B(E;, F;) is a topological isomorphism onto a dense
subspace of!Be(E;, F;).
REMARK. The property expressed by the theorem is actually character-
istic of nuclear spaces and is used by Grothendieck ([13], chap. II,
def. 4 and theor. 6) to define nuclear spaces. We shall not show
here that the validity of the assertion, for a given locally convex space
E and any l.c.s. F, implies that E is nuclear. For it will be shown below
(Section 10) that for E to be nuclear, it is sufficient that the assertion of
(9.4) hold for F= 11.
Proof of (9 .4). It is clear thatthe canonicalimbedding of E ® F into !B(E;, F;)
is an algebraic isomorphism (cf. Chapter III, discussion preceding (6.3». The
proof consists now of two steps: first we show that E ® Fis dense in fB e(E;, F;);
second thatfBe(E;, F;) induces the projective topology on E ® F.
1. Let us identify fBiE;, F;) with .PiE;, F) as before. It is then sufficient
to show that, given u E .P(E;, F), a convex, circled, closed equicontinuous
set AcE; and a O-neighborhood V in F, there exists U o E E ® F such that
u(x') - uo(x') E V for all x' EA. Let U = A and let We U be a O-neighbor-
O
hood in E which is convex, circled, and such that Ew -+ Eu is nuclear, say,

<X)
cPu,w= L AiY; ® Zi' where (Ai) E [1, and {ya, {zJ are bounded sequences in
1
Ea (B = WO) and Eu, respectively. It follows that the canonical imbedding of
<X)
E~ into Ea is of the form LAiZ; ®Yi. Since Eu is dense in E u, each Z; (viewed

1
as a linear form on E~) can uniformly on A be approximated by (the restric-
tions to E~ of) suitable elements x E E; hence if e > 0 is preassigned, there
exists an integer n and elements x; E E (i = 1, ... , n) such that
n
L A;(X;, x') y; -
;= 1
x' E eB
whenever x' EA. Now if e is so chosen that eu(B) c V, it follows that

n n
L A;(X;, x')u(YD - u(x') E V for x' E A; hence Uo = L A;Xj ® u(YD satisfies
1 1
the requirement.
2. Since the dual of E ® F for the projective topology is (fl(E, F), it is
sufficient to show that each equicontinuous set Q in (fl(E, F) is contained and
equicontinuous in ,feE, F) = [fBe(E;, F;)]'. Viewing.?l(E, F) as a subspace of
.P(E, F;), the equicontinuity of Q is tantamount to the existence of a 0-
neighborhood U c E and an equicontinuous set B c F' such that u(U) c B
whenever u E Q. Of course. we can assume that U and B are convex and
circled and that B is compact in F;. The map u of Eu into Fa associated with
u E Q, is of norm ~ 1 in .p(Eu, Fa). Since E is nuclear, the canonical map
= L A;X; ® Y;.
<X)
E -+ Eu is nuclear by (III, 7.2), Corollary 1, say, of the form cPu

;=1
where it can be assumed that X;E VA and Y;EcPu(U)- for all iEN, and
L1 IA;I = c < + 00. Since u = I/IB u
<X)
a 0 cPu, it follows that
L A;xi ® U(Yi),
<X)
U =
i= 1
which shows that u is integral, hence (in view of the identification of .?I(E, F)
with a subspace of .P(E, F;» that U E ,feE, F). Moreover, u(y;) E B for all
u E Q and all i EN; hence the preceding formula shows that Q c c Va ® B) - , (r
where the closure is with respect to (1(,f(E, F), E ® F). Since Va ® B) is r(
equicontinuous in the dual ,feE, F) of fBeCE;, F;), the same holds for
(r Va ® B)- by (III, 4.3) and (III, 4.5), and hence for Q.
The proof is complete.
172 DUALITY [Ch. IV
CORO~LARY 1. If E is a complete nuclear space and ifF is any complete l.c.s.,

then E ® F can be canonically identified with !Be(E~, F~) and with f£e(E~, F).
This is immediate in view of (9.1) ; let us remark also that if E is nuclear and
precompact subsets of E~ are equicontinuous, then E~ possesses the approxi-
mation property by (III, 9.1). More importantly, the identity map of E®F
(which is continuous for the projective topology on the domain, and the
topology of bi-equicontinuous convergence on the range) is an isomorphism
if E is nuclear. Hence:
COROLLARY 2. 1£E is nuclear and ifF is any l.c.s., then the canonical mapping
of E ® F into E ®F is a topological isomorphism of the first space onto the
second.
From Part 2 of the proof of (9.4), we obtain the kernel theorem:
9.5
If E is nuclear and if F is locally convex, then every v E ~(E, F) originates
from a space EA ® F~, where A, B are suitable equicontinuous subsets of E', F',
respectively. Equivalently, every continuous bilinear form v on E x F is of the
form
I
00
(x, y) -+ vex, y) = A.i(X, x/>(y, y;),

i=1
where (A.;) E 11 and {x;}, {y;} are equicontinuous sequences.

Note that by (III, 6.5), these sequences can even be supposed to be null
sequences in E A, F~, respectively. We are now in a position to establish two re-
sults that will furnish a large number of additional examples of nuclear spaces.
9.6
Theorem. The strong dual of every nuclear (F)-space is nuclear.
Proof If E is a nuclear (F)-space, then E is reflexive (in fact, an (M)-space)
by (III, 7.2), Corollary 2; hence E/J is reflexive, (5.6), Corollary 1, which
implies that E/J = E~. By (III, 7.2) it suffices to show that every u E f£'(E~, F) is
nuclear, F being an arbitrary Banach space. Now f£(E~, F) can be identified
with !B(E~, F~) and, by (9.4), Corollaries 1 and 2, with E ® F, since E, Fare
complete. The assertion follows now from (III, 6.4) and (III, 7.1).
REMARK. The converse of (9.6) is also true: If E is an (F)-space
whose strong dual is nuclear, then E is nuclear; equivalently, the strong
dual of a complete nuclear (DF)-space is nuclear (Exercise 33). In
general, however, the strong dual of a nuclear space fails to be nuclear:
the product Kg (d any cardinal) is nuclear by (III,7.4), but its strong
dual fails to be nuclear if d is uncountable (Exercise 31).
It follows from (III, 7.4) that if E is any I.c.s. and F is a nuclear space, then
f£.(E, F) (topology of simple convergence) is nuclear; in fact, f£.(E, F) is
isomorphic with a subspace of the nuclear space FE. Under additional

assumptions on E, a corresponding result holds for !l'b(E, F) (topology of
bounded convergence).
9.7
Let E be a semi-reflexive space whose strong dual is nuclear and let F be any
nuclear space. Then !l'b(E, F) is a nuclear space.
Proof Since E is semi-reflexive (Section 5), each bounded subset of E is an

equicontinuous subset in the dual of E p, and hence !l' beE, F) can be identified
with a subspace of~.(EI1' F;). Since Ep is nuclear, (9.4) implies that the com-
pletion of ~.(EI1' F;) can be identified with Ep ® F. Now Ep ® F is nuclear
by (III, 7.5) and hence !l'b(E, F) is nuclear by (III, 7.4). Compare Exercise 34.
Examples
1. From (9.6) it follows that the strong duals of the nuclear (F)-spaces
enumerated in Chapter III, Section 8, are nuclear; in particular, the
spaces fJ2~ and [Il' are nuclear, and also Yf" is nuclear.
2. The space fJ2' of distributions, strong dual of fJ2, is nuclear. Since
fJ2 (Chapter II, Section 6, Example 2) is the strict inductive limit of a
sequence of spaces fJ2 Grn' each bounded subset of fJ2 is contained in a suit-
able space fJ2 Grn by (II, 6.5), so (4.1) (a) = (b) implies that the strong
TI fJ2~m;
CJJ
dual fJ2' is isomorphic with a subspace of since the latter is

m=l
nuclear, fJ2' is nuclear. This is an example of the situation indicated in the
remarks following (4.5).
3. (9.7) implies now that the spaces !l'b(fJ2), .!t'b(fJ2'), .!t'b([Il) , etc.
are nuclear. This implies that the (F)-space If (Chapter III, Section 8,
Example 3), as well as its strong dual If', is nuclear. For each j Elf
defines, by virtue of the multiplication operator g -+ jg, a continuous
endomorphism of E&; the corresponding imbedding If -+ !l' b(fJ2) is an
isomorphism. Since If and If' are reflexive, it also follows that !l' b(If)
and !l' b(If') are nuclear spaces.
If E, Fare l.c.s., the dual of E ® F is @leE, F), and it is clear that the
~l x ~2-topology (~1'~2 the respective families of all bounded subsets of
E, F) on @leE, F), also called the topology of bi-bounded convergence, is
coarser than the strong topology {3(@l(E, F), E ®F); it is natural to ask for
conditions under which the two topologies agree. This is manifestly true if
E, Fare normed spaces;. but even if E, Fare non-normable (F)-spaces, the
answer seems to be unknown (Grothendieck [13], chap. I, §l, "probleme des
topologies "). Alternatively the problem is this: Given a bounded subsd
B c: E ®F, do there exist bounded subsets Bl c: E and B2 c: F such that
B c: (r
Bl ® B 2 ) -? Grothendieck gave an affirmative answer if E, F are both
(DF)-spaces, or if E, F are (F)-spaces, one of which is nuclear. To establish
the result, we need the following simple lemma.
174 DUALITY [Ch. IV
LEMMA 2./f E is a metrizable l.c.s. and {Bn} is a sequence of bounded subsets

00
of E, there exists a sequence {ltn} of positive numbers such that B = U ItnBn is

1
bounded in E.
Proof We can assume the topology of E is generated by an increasing
sequence {Pn} of semi-norms, and that each Bn contains a point Xn such that
Pn(xn) > O. Define Itn by It;; 1 = sUP{Pn(x): X E Bn}. For any semi-norm Pk of
the sequence, we obtain Pk(X) ~ 1 whenever x E U ItnBn; hence sup {Pk(x):
n"?:.k
X E B} < + 00. Since this holds for all k, it follows that B is bounded.
9.8
Let E, F be l.c.s. such that either E and Fare (DF)-spaces, or such that E
and F are (F)-spaces and E is nuclear. Then the topology. of the strong dual
fJ6 p(E, F) of E ® F agrees with the topology of bi-bounded convergence.
Proof We first prove the assertion, assuming that both E and Fare (DF)-
spaces. Then clearly E x F is a (DF)-space and the topology of bi-bounded
convergence on fJ6(E, F) is the topology of bounded convergence in E x F,
hence metrizable. Since this topology is coarser than [3(fJ6(E, F), E ®F), it
suffices to show tha.t the identity map of fJ6 b(E, F) onto fJ6p(E, F) (obvious
notation) is continuous; for this, in turn, it suffices that every null sequence
{j,,} in fJ6 bee, F) be bounded in fJ6 pee, F) (cf. Chapter II, Exercise 17). Let
00
Z = {cc letl ~ I} be the unit disk in the scalar field K, then () f;;1(Z) is a
1
countable intersection of convex O-neighborhoods in Ex F which absorbs
bounded sets, since {j,,} is bounded infJ6 b(E, F), and hence a O-neighborhood
by the defining property of (DF)-spaces. It follows that {j,,} is equicontinuous
on E x F, and hence equicontinuous in the dual fJ6(E, F) of E ® F by the
corollary of (III, 6.2) and, therefore, bounded in fJ6 p(E, F).
Turning to the second part of the proof, we assume that E, F are (F)-spaces
and that E is nuclear. We shall show that for every bounded subset B of
E ® F, there exist bounded sets B1 c E, B2 C F such that B c (r
B1 ® B 2) -.
The assumptions imply that E is reflexive (hence E~ = Ep) and, by (9.4), that
E ®F can be identified with !BeCE:, F:), hence with !l'eU~:, F), and hence
with !l' beep, F). Viewing B as a bounded set of linear maps in !l' beep, F), we
see that B(G) is bounded in F for each bounded subset G of E/J; now if {Gn}
is a fundamental sequence of bounded subsets of Ep (Ep is a (DF)-space) and
B(Gn) =Hn (n EN), by Lemma 2 there exists a sequence {ltn} of positive
00
numbers such that U ItnHn c H, where H is a bounded, closed, convex,

1
circled subset of F.
It is immediate that each u E B maps E/J into F H ; in fact, since for each
u E B, u- 1(H) is a barrel in Ep and since Ep (being reflexive) is barreled, it
follows that u E !e(E/J, FH). Moreover, B is clearly simply bounded in

!e(E/J, FH) and hence equicontinuous; consequently, there exists a convex,
circled O-neighborhood V in E/J such that B(V) c H. Now denote by El the
Banach space which is the completion of [E']v; since E/J is nuclear by (9.6),
00
the canonical map c/lv of E/J into E1 is nuclear, say of the form L AiXi ® Yi'
00 1
where L lAd;;;; 1, and {Xi}, {yJ are bounded sequences in E, E 1 , respectively.
1
If u denotes the linear map of El into FH associated with u E B, then the
family B2 ={u(y;): u E B, i E N} is bounded in FH (hence in F), and we obtain
00
u = L AiXi ® U(Yi)
1
for all u E B. Letting B1 ={Xi: i EN}, it follows that Be (r

B1 ® B 2 )-,
since the series for u converges in E ®F; the proof is complete.
We cODclude this section with an explicit characterization of the strong dual
and bidual of E ®F if E, F are (F)-spaces of which at least one is nuclear.
As all the results in this section, the theorem is due to Grothendieck ([13],
chap. II, theor. 12) and holds also when E is nuclear and E, Fare (DF)-spaces
that are not strong duals of (F)-spaces (Exercise 32). We need another lemma.
LEMMA 3. Let E, F be l.c.s., let F~ be the bidual of Funder its natural topology
(of uniform convergence on the equicontinuous subsets of F'), and let F; be the
space F" under u(F", F'). Then every continuous bilinear form v on Ex F
possesses a unique extension v to E x F" which is a bilinear form continuous on
E x F; and separately continuous. on Ea x F;; v -+ v is an isomorphism of
P.l(E, F) onto P.l(E, F~) 11 ~(Ea, F;).
Proof Let us identify.?l(E, F) with the subspace of !e(F, E~) whose elements
map a suitable O-neighborhood in F onto an equicontinuous subset of E'
(both sets depending on the map in question), and .?l(E, F~) with the corres-
ponding subspace of !e(F;, E~). Then the argument used in the proof of Lem-
ma 1 shows that u -+ u" is the (algebraic) isomorphism that we are seeking.
9.9
Theorem. Let E, F be (F)-spaces, E being nuclear, and denote by F" the
strong bidual of F. Then the strong dual (respectively, the strong bidual) of
E ®F can be identified with E/J ®F/J (respectively, with E ®F").
Proof By (9.8) the strong dual of E ® Fis P.lb(E, F) (topology ofbi-bounded
convergence). Now P.l(E, F") = ~(E, F") by (III, 5.1); hence Lemma 3 shows
that .?l(E, F) can be identified with~(Ea, F;), and under this identification the
topology of bi-bounded convergence is identical with the topology of bi-
equicontinuous convergence; moreover, ~e(Ea, F;) can be identified with
E/J ®F/J by (9.4), since E/J is nuclear by (9.6). This proves the fir~t assertion.
176 DUALITY [Ch. IV
For the proof of the second proposition, put H = Ep and G = Fp. We use
(9.8) again to see that the strong dual of H ~ G is PAb(H, G); then, in view of
Lemma 3, PAb(H, G) can be identified with a subspace of me(Hu , G;). Now
since H/J = E (note that E is reflexive) is nuclear, it follows from (9.4) that
PAb(H, G) can be identified with a subspace of H/J ® Gp = E ® F". On the
other hand, (III, 6.4) shows that every element of E ®£II defines a bilinear
form on H x G, which is continuous (cf. (6.5), Corollary 1), whence the
proposition follows.
In the following corollaries of (9.9), we do not repeat the assumptions that
E be a nuclear (F)-space and F an arbitrary (F)-space; the assertions hold
equally when E, Fare (DF)-spaces (E being assumed nuclear) (Exercise 32).
COROLLARY 1. If, also, F is reflexive, then E ®F (and hence E/J ® FP) is a

reflexive space.
COROLLARY 2. The strong dual of !l\(Ep, F) can be identified with Ep ® Fp;

in particular, ifF is reflexive, then !l' beEp, F) is reflexive.
COROLLARY 3. Every separately continuous bilinear form on E/J x F/J is

continuous.
10. NUCLEAR SPACES AND ABSOLUTE SUMMABILITY
The present section gives a characterization of nuclear spaces which is

based on the concept of Radon measure on a compact space. This charac-
terization shows that under suitable assumptions, nuclearity of a l.c.s. E is
equivalent to the absolute summability of arbitrary summable families in E
(cf. Chapter III, Exercise 23), and will give us access to several important
results; among them are the converse of Theorem (9.4) and the theorem of
Dvoretzky-Rogers [1] as proved by Grothendieck [13]. The results of this
section (except where other references are given) are due to A. Pietsch [3]-[5].
Recall that a Radon measure (Chapter II, Example 3) on a compact space
X is a continuous linear form on the (real or complex) Banach space CC(X);
a positive Radon measure is a p, E CC(X)' such that p,(f) = Jfdp,;:;; 0 whenever
f(t) ;:;; 0 for all t EX. We shall need the following well-known result, which is
also an easy consequence of (Y, 7.4), Corollary 2.
LEMMA 1. If v is a Radon measure on the compact space X, there exists a
positive Radon measure p, on X such that 1v(f)1 ~ p,(lfl) for all f E CC(X).
Further, if p, is a positive Radon measure on X then (f, g) -+ Jfg*dp, is a
semi-definite Hermitian form on CC(X) x CC(X); hence the semi-normf-+ p(f)
= (JlfI2dp,)~ generates a locally convex topology on CC(X) such that the
associated Hausdorff t.v.s. is a pre-Hilbert space Ep; the completion Ep is a
Hilbert space isomorphic with L 2(p,) (Chapter II, Section 2, Examples 2, 3,
and 5).
§10] NUCLEAR SPACES AND ABSOLUTE SUMMABILITY 177
We also need the following result due to K. Maurin [1]. Let H1 and H2 be
Hilbert spaces over K; a linear map u E .f£(Hi> H 2) is called a Hilbert-Schmidt
transformation if there exists an orthonormal basis {x..: 0( E A} of H1 such that
L.. llu(x..) 112 < + 00. It is clear that such a map is compact, and that u(x..) = 0,
save for a countable number of indices 0( E A. It will follow from the sub-
sequent proof that the value of the sum is independent of the choice of the
orthonormal basis {x..: 0( E A} and, in fact, equal to Lpllu*(Yp) 112 for any
orthonormal basis {yp: P E B} of H 2 •
LEMMA 2. Let Hi (i = 1,2,3) be Hilbert spaces and u, v be Hilbert-Schmidt
transformations of H1 into H2 and of H2 into H 3 • respectively. Then the com-
posite map w = v 0 u of H1 into H3 is nuclear.
Proof. Denote by {x..: 0( E A}, {Yp: PE B} orthonormal bases of Hi> H 2 ,
respectively, and by [ , ] the inner product of H 2 • If for the moment u is
any continuous linear map of H1 into H2 with conjugate u*, we obtain, in
view of the identity u(x.) = Lp[u(x..), Yp]Y(I (0( E A), the equality
Lllu(x..)11 2 = L.pl[u(x..), yp]12 = L ... pl[x.., u*(yp)]12
= L pllu*(Y(I)11 2.
It follows that the value of the first and last term is· independent of the basis
{x.. } and {Yp}, respectively, and hence that U is a Hilbert-Schmidt transfor-
mation if and only if u* is.
Now let u, v be the maps mentioned in the statement of the lemma. Denote
by {Yn: n EN} an orthonormal basis of the range u(H1 ), which is clearly
separable. For each x E H1 we obtain
00 00
w(x) = L [u(x), Yn]v(Yn) = L [x, u*(Yn)]v(Yn)·

n= 1 n= 1
Since by Schwarz' inequality
JI IIU*(Yn)IIIIV(Yn)1I ~ C~11IU*(Yn)ll2. JIIIV(Yn)1I2) t,

the left-hand term is finite, it follows from (III, 7.1) that w is nuclear.
Turning to the subject of this section, let us agree on the following defini-
tions. Let E be a locally convex space. A semi-norm p on E is called pre-
nuclear if there exists a closed, equicontinuous subset A of E; and a positive
Radon measure Jl. on the (a(E', E)-) compact space A such that
p(x);;i tl<x, x')1 dJl.(x') (x E E).
A subset BeE' is called prenuclear if there exists a closed equicontinuous

subset A of E; and a positive Radon measure Jl. on A such that
I<x, b)1 ;;i tl(X, x')1 dJl.(x') (x EE, bE B).

178 DUALITY [Ch. IV
(Equivalently, a subset BeE' is prenuclear if and only if x -.. sup{l<x, b)l:

bE B} is a prenuclear semi-norm on E.) It is easy to see that a prenuclear
semi-norm is necessarily continuous, and that a prenuclear subset of E' is
necessarily equicontinuous. A family {x~: IX E A} will be called prenuclear
if its range is a prenuclear subset of E'. The following result is still of an
auxiliary character. (Cf. (III, 7.3), Corollary 2.)
10.1
Let E be a l.c.s. on which every continuous semi-norm is prenuclear. There
exists a neighborhood base U of 0 in E such that for each U E U, Bu is norm
isomorphic with a Hilbert subspace of L 2(Jl), where Jl is a positive Radon
measure on a suitable closed equicontinuous subset of E;.
Proof It suffices to show that each closed, convex, circl~d O-neighborhood
Win E contains a O-neighborhood U with the desired property. If Pw denotes
the gauge of W, Pw is prenuclear by hypothesis; hence we have
Pw(x) ~ L,<x, x')ldJl(x') (x E E),
where A and Jl can be so chosen that IIJlII = l. Define the O-neighborhood

Uby
U = {x E E: L'<x, x')1 2 dJl(x') ~ I};
U is convex and circled, and by Schwarz' inequality x E U implies Pw(x) ~ l.
r
Hence we have U c W, and evidently the gauge of U is given by
Pu(x) = (f) <x, x')1 2 dJl(x') (x E E),

This leads to the announced characterization of nuclear spaces.
10.2
A locally convex space E is nuclear if and only if every continuous semi-
norm on E is prenuclear, or equivalently, if and only if every equicontinuous
subset of E' is prenuclear.
Proof, The condition is necessary. Let E be nuclear and let p be a con-
tinuous semi-norm on E. If U = {x: p(x) ~ I}, then by (III, 7.2) the canonical
00
map E -.. Eu is nuclear, say <Pu = L A;x; ® X;, where Ilx; II ~ I in Eu (i EN),
;= 1
{xa c VO for a suitable O-neighborhood V in E, and (A;) Ell. Now
00
LIAd<x, xi)1 ~ I implies II<Pu(x) II ~ 1; hence (since U is closed) x E U, or

1 00
equivalently, p(x) ~ 1. It follows that p(x) ~ LIAdl<x, xi)1 for all x E E.

1
co
Since / --. LIAil/(x;} is clearly a positive Radon measure on V", we conclude
1
thatp is prenuclear.
The condition is sufficient. Let U be a neighborhood base of 0 in E having
the property described in (10.1); in view of Lemma 2, it suffices to show that
for each U E U, there exists V E U, V C U, such that Ev ~ Eu is a Hilbert-
Schmidt transformation. For, if this is correct, we can select WE U, We V
such that the canonical map Ew --. Eu is the composite of two Hilbert-
Schmidt transformations, and hence nuclear; the nuclearity of E is then a
consequence of (III, 7.2). Thus let U E U be given so that
I cPu(X) II 2 = L'(X, x')1 2 dll(X'} (x E E);
here A denotes, as before, a weakly closed, equicontinuous subset of the

dual E' of E. Now choose V E U to satisfy V c U n AO; it follows
that A c AOO c V". If/A denotes the restriction of/E CC(V"} to A,J--. S/Adll
is a positive Radon measure v on yo. Further, let {x«: IX E A} be an ortho-
normal basis of Ev; we have to show that L IlcPu.v(X«} 112 is finite. Since Ev is
«eA
a Hilbert space, there exists a conjugate-linear isomorphism x' ~ z of E;;
onto Ev such that (x, x') = [x, z] for all x E Ev; x' --. z induces a homeo-
morphism of VO onto the unit ball B of Ev with respect to the weak and
norm topologies. Denote by v' the Radon measure on B obtained from v
under x' ~ z. Since Ev is dense in Ev , we conclude that
II cPu.v(X«} II 2 =f I(x
A
a, x')1 2 dll(X'} =fV'
I(xa , x')1 2 dv(x'}
= tl[x«, z]I 2 dv'(z}.
Now if H is any finite subset of A, we have L I[x«, zW ~ [z, zF by Bessel's

«eH
inequality, and the sum is ~ I whenever z E B; it follows that
L
«eH
I cPu.v(X«} II 2 =f L I[x«, z]I 2 dv'(z} ~ v'(R}
B aeH
for any finite He A. Clearly, this shows that cPu.v is a Hilbert-Schmidt

transformation, thus completin,g the proof.
Let E bea l.c.s.; a family {x«: IX E A} in E is called summable (Chapter III,
Exercise 23) if limH X H exists in E, where X H = L X« and H runs through the
aeH
family of all finite subsets of A directed by inclusion c; the limit x E E is
then denoted by L X«, or briefly by L«x«. (If A is infinite, the limit can
«eA
equivalently be taken along the filter of subsets of A with finite complement.)
180 DUALITY [Ch. IV
A summable family {x",: IX E A} in E is absolutely summable if for every con-

tinuous semi-norm p on E, the family {p(x",): IX E A} is summable in R. It is
well known that the two notions coincide when E is finite-dimensional
(Chapter Ill, Exercise 23); we shall prove below that within the class of
(F)-spaces, the identity of the respective sets of summable and absolutely
summable sequences characterizes nuclear spaces.
Now let A be a fixed non-empty index set and let E be a given l.c.s.; the
set of all absolutely summable families x = {x",: IX E A} in E can evidently be
identified with a subspace Sa of the algebraic product EA. Denote by U any
fixed base of convex circled O-neighborhoods in E, and by ru the gauge of
U E U; clearly, the mapping
x -. Pu(x) = l>u(x",)
is a semi-norm on Sa' and the family of semi-norms {Pu: U E U}· generates a
topology under which Sa is a l.c.s. that will be denoted by [l[A, E). Similarly,
the set of all summable families x ={x",: IX E A} can be identified with a
subspace S of EA, and for each U E U the mapping
x -. qu(x) = sup{L",I<x"" x')I: x' E UO}
is a semi-norm on S; the family of semi-norms {qu: U E U} generates a
topology under which S is a l.c.s. that will be denoted by [l(A, E). The
obvious inclusion Sa C S defines a canonical imbedding of [l[A, E] into
[l(A, E) which is continuous, since qu(x) ~ Pu(x) for each x E Sa and U E U.
Our next objective is to find a representation of the duals [l[A, EY and
[l(A, E)'. (See also Exercise 35.)
10.3
The dual of [1 [A, E) can be identified with the subspace of (E')A, each of
whose elements constitutes an equicontinuous family x' = {x~: IX E A} in E',
the canonical bilinear form being given by <x, x') = L,.(x"" x~).
Proof. If x' = {x~: IX E A} is an equicontinuous family, then x~ E UO (IX E A)
for a suitable U E U and we have (x E Sa)
IL<x"" x~)1 ~ L",I<x"" x~)1 ~ Lru(x",) = Pu(x),
which shows that x -. <x, x') is well defined on Sa and continuous on
zt [A, E); it is, moreover, evident that <x, x') defines a duality between Sa
and the space of equicontinuous families. There remains to show that every
f E zt [A, EY can be so represented. Choose U E U such that If(x) I ~ Pu(x)
(x E Sa) and let, for any Z E E, z\"') be the family {<5",pz: PEA}; it is clear that for
each IX E A, z -. f(z<"'» is an element x~ E E', and that {x~: IX E A} is a family in
E' with range in UO, hence equicontinuous. Moreover, if x = {x",: IX E A} E Sa'
then {r"'): IX E A}, where x<"') = {<5",px",: PEA}, is a summable family in
[l[A, E) such that x = L..x<"'). It follows from the continuity of f that f(x) =
Lf(x<"'» = L<x"" x~), and the proof is complete.
Denoting by [l(A) the Banach space of summable scalar families = e

{ea: 0( E A} under its natural norm e-+ lie II = Laleal, we obtain this corollary.
COROLLARY. The dual of [l(A) can be identified with the space of all bounded
scalar families '1 = {'1a: 0( E A}, the canonical bilinear form being given by
<e, "1) = Laea'1a.
10.4
The dual of Il(A, E) can be identified with the subspace of (E')A, each of
whose elements constitutes a prenuclear family x' = {x;: 0( E A} in E ', the
canonical bilinear form being given by <x, x') = La<xa, x;).
Proof. If x' = {x~: 0( E A} is a prenuc1ear family in E ' , there exists a U E U
and a positive Radon measure Jl on Uo such that
I<x, x~)1 ~f I<x, x')1 dJl(x') (x E E, 0( E A).

U'
Now if x = {xa: 0( E A} E S, the definition of summability implies that

Lal<xa, x')1 converges uniformly for x' E Uo (hence to a continuous function
on UO), whence
Lal<xa, x~)1 ~ Lafu I<xa, x')1 dJl(x') =fu Lal<xa, x')1 dJl(x')
o a
Hence x -+ <x, x') is well defined on S and is a continuous linear form on

PeA, E); evidently, (x, x') -+ <x, x') places S in duality with the space of all
prenuc1ear families.
Conversely, if fE [l(A, E)" we define x~ EE' (0( E A) as in the preceding
proof; there remains to show that {x~: 0( E A} is a prenuc1ear family. To this
end, denote by Z the unit disk of the scalar field K and consider the compact
space ZA x UO, where U EU is chosen such that If(x) I ~ qu(x) for all XES.
Since the convergence of Lal<xa, x')1 is uniform with respect to x' E UO, the
formula
h,,[(Aa), x'] = LaAa<xa, x')
defines, for each XES, a function h" E CC(ZA X UO). The mapping x -+ h" of
[l(A, E) into CC(ZA x UO) is obviously linear and such that IIh" II = qu(x);
thus h" -+ f(x) is a continuous linear form defined on the range of x -+ h" and
can be extended, by the Hahn-Banach theorem, to a continuous linear form
v on CC(ZA x UO). v is a Radon measure on ZA x UO; hence by Lemma 1
there exists a positive Radon measure Jl on ZA x UO such that Iv(h) I ~ Jl(lhi)
for all hE CC(ZA X UO). Now x~ E E' is the linear form Z -+ f(z(a»), where Z E E
and z(a) ={<>apz: f3 E A} E Il(A, E). We obtain, for all 0( E A,
I<z, x~)1 = If(z(a»1 = Iv(hz(<<»1 ~ Jl(g),
182 DUALITY [Ch. IV
where 9 is the function on ZA x U a given by g«A.,,), x') = I(z, x')I. Now

~(ua) can clearly be identified with the subspace of ~(ZA x U a) containing
all functions independent of (A.,,) E ZA; under this identification, p. induces a
positive Radon measure p.' on U a , and the function 9 becomes an element of
~(Ua). Thus the preceding show~ that
I(z, x~)1 ~ p.(g) = p.'(g) = Iu I(z, x/)1 dp.'(x/)

o
for all oc E A and all z E E. Hence {x~: oc E A} is a prenuclear family, and the
proof is complete.
It is remarkable that the algebraic and topological relations between
P[A, E] and [l(A, E) are determined, for any infinite index set A, by the
corresponding relations for A = N.
10.5
Let E be any I.c.s. If the canonical imbedding of [1[A, E] into [1(A, E) is an
algebraic (respectively, topological) isomorphism of the first space onto the
second for A = N, the same is true for any (non-empty) index set A.
Proof Suppose that each summable sequence in E is absolutely summable,

and assume that {x,,: oc E A} is a summable family in E which is not abso-
lutely summable; then A is necessarily· infinite, and there exists U E U such
that ~>u(x,,)= + 00. It follows that~>u(x"k) = + CIJ for some count-
ably infinite subset {oc l , OC2, ••• } of A; this is contradictory since, clearly,
{x"" - Xa " ••• } is a summable sequence. Assume now that [1[N, E] = PeN, E)
algebraically and topologically; then for given U E U, there exists Ve U
such that Pu(x) ;::;; qy(x) whenever x E Il(N, E). Now if x = {x,,: oc E A} is any
summable family in E and H is any finite subset of A, we obtain
L 'u(x,,) = Pu(y) ;::;; qy(y) ;::;; qy(x),

"eH
where y = {y,,: oc E A} is such that y" = x" for oc E Hand y" = 0 for oc rf: H;
for y can be viewed as an element of peN, E). This shows that Pu(x) ;::;; qy(x)
for all x E II(A, E), where A# 0 is arbitrary, and the proof is complete.
We are now prepared to prove the principal theorem of this section
characterizing nuclear spaces in terms of summability. However, let us first
establish the connection of the preceding material with the theory of topo-
logical tensor products.
The mapping (~, x) ~ {e"x: oc E A} of peA) x E into [1[A, E] is evidently
bilinear, and hence defines a linear mapping of [l(A) ® E into [l[A, E) (and
into [1(A, E». With the aid of (10.3) it is easy to see that this linear mapping
is an algebraic isomorphism; it will be called the canonical imbedding of
[1(A) ®E in [1[A, E] (respectively, in peA, E».
10.6
The canonical imbedding of 11(A) ® E in 11[A, E] is an isomorphism for the
projective topology on [1(A) ® E, and the canonical imbedding of 11(A) ® E
in zt(A, E) is an isomorphism for the topology of bi-equicontinuOllS convergence
on 11(A) ® E. Moreover, the canonical image of 11(A) ® E is dense in both
11[A, E] and 11(A, E).
Proof The last assertion is almost immediate from the definition of the
semi-norms Pu and qu, and the first assertion is a special case of (III, 6.5).
Hence let us identify 11(A) ® E algebraically with its canonical image in
zt(A, E) and show that the induced topology is the topology of bi-equicon-
tinuous convergence. By the corollary of (10.3), the dual of the Banach space
J1(A) is the space of bounded scalar families. If B is the unit ball of 11(A), its
polar BO (under the weak topology) can be identified with the compact space
ZA, where Z = {k IXI ~ I} is the unit disk in K. Let :Lei
® Xi be any element
of 11(A) ® E and let U E U. By defihition of qu we have
qu(I el ® Xi) = sup LaILie~I\Xi' X')1·
x'e uo
For each 0( E A and each x' E UO, there exists 11a E Z such that
Lie~i)'7a<Xi' x') = ILie~i)<X;, x')I;
this implies
and, therefore, the relation
This shows that the semi-norms qu (U E U) generate on 11 (A) ® Ethe topology

of uniform convergence on the sets, B O ® uo (U E U), thus completing the
proof.
Let us note that another proof of the last assertion can be obtained by
(9.2); in fact, if {x;: 0( E A} is an equicontinuous family in E', we can verify
that this family is prenuclear if and only if (e, x) -+ Laea<x, x;) is an integral
bilinear form on Jl(A) x E, and that in this fashion equicontinuous sets of
prenuclear families in E' correspond to equicontinuous subsets of ,1(J 1(A), E).
The reader will have no difficulty verifying that for arbitrary A 1= 0,
11[A, E] and 11(A, E) are complete if and only if E is; for 11(A, E), use (6.2),
Corollary 2(b). We obtain these corollaries.
COROLLARY 1. If E is complete, then Jl(A) ® E is isomorphic with 11[A, E]
under the continuous extension of the canonical imbedding of 11(A) ® E in
[1 [A, E]; in the same fashion, Jl(A) ®
E is isomorphic with 11(A, E).
The isomorphisms thus established are again called canonical.
184 DUALITY [Ch. IV
COROLLARY 2. Let E be a [.c.s. and let E be its completion. Under the canonical
isomorphism of Il(A) ® E with Il[A, E] and of II(A)~ E with II(A, E), the
canonical map of Il(A) ® E into Il(A) ~ E corresponds to the canonical
imbedding of II[A, E] in Il(A, E) and is, therefore, biunivocal.
The following is the central result of Pietsch [5].
10.7
Theorem. A locally convex space E is nuclear if and only if the canonical
imbedding of/I [N, E] in Il(N, E) is a topological isomorphism of the first space
onto the second.
Proof In view of (10.2) through (10.5), we have the following chain
of implications: E is nuclear => every equicontinuous subset of E' is pre-
nuclear=>every equicontinuous subset of [1[N, E]' is contained in and equi-
continuous in II(N, E)' => II[N, E] is a dense topological vector subspace of
II(N, E) =>11[N, E] = II(N, E) (cf. proof of (10.5» => [1[A, E] = [I(A, E) for
any A#-0 => every equicontinuous family in E' is pre nuclear => every equi-
continuous subset of E' is prenuclear =>E is nuclear. The proof is complete.
This theorem has several important corollaries; the first of these is obtained
from (10.6), Corollary 2, in view of the fact that a l.c.s. is nuclear if and only
if its completion is nuclear.
COROLLARY 1. (Grothendieck [13].) A locally convex space E is nuclear if
and only if the canonical map of 11 ® E into [1 ~ E is a topological isomorphism
of the first space onto the second.
This corollary implies, in particular, that the property established in
Theorem (9.4) characterizes nuclear spaces. If E is an (F)-space, then clearly
II[N, E] and peN, E) are (F)-spaces; hence if the two spaces are algebraically
identical, the canonical imbedding (which is continuous) is a topological
isomorphism by Banach's theorem, (III, 2.1), Corollary 1. This implies:
COROLLARY 2. An (F)-space E is nuclear if and only if every summable
sequence in E is absolutely summable.
We have observed earlier (Chapter III, Section 7) that a nuclear Banach
space is finite dimensional, since it is locally compact; hence from Corollary 2
we obtain the following theorem, known as the theorem of Dvoretzky-
Rogers [1].
COROLLARY 3. A Banach space in which every summable sequence is abso-
lutely summable is finite dimensional.
We remark in conclusion that the algebraic identity of the spaces II[N, E]
and II(N, E) implies the identity of their respective topologies whenever
[1[N, E] is infrabarreled (Exercise 36); hence the completeness of E is dis-
pensable in Corollaries 2 and 3.
§11] WEAK COMPACTNESS. THEOREMS OF EBERLEIN AND KREIN 185
11. WEAK COMPACTNESS. THEOREMS OF EBERLEIN AND KREIN

If S is a metrizable topological space, compactness of a subset A can be
characterized by the existence, for each sequence in A, of a subsequence
converging to a point in A. In more general circumstances this description
fails, and several variations of the notion of compactness prove useful. It is
the purpose of this section to prove some important characterizations of
compact subsets (especially for the weak topology) of a I.c.s. by seemingly
weaker properties, and to derive a deep criterion for the compactness of the
convex closure of a compact set. For further information, we refer the
interested reader to the literature cited below; detailed accounts can be found
in Kothe [5] and (with emphasis on normed spaces) in Day [2]. See also
Kelley-Namioka [1] and the very interesting paper of James [3].
Let us recall the following definitions, S denoting a Hausdorff topological
space: A subset A of S is called countably compact if every sequence in A has a
cluster point in A (equivalently, if each countable open cover of A has
a finite subcover); A is called sequentially compact if each sequence in A
possesses a subsequence converging to a point in A. It is immediate that
compactness of A and sequential compactness of A both imply countable
compactness 'of A; in general, no other implications are valid among these
notions (cf.Exercise 37).
We begin with the following result, which is preparatory but of considerable
interest in itself (cf. Eberlein [1], Grothendieck [6]). Denote by (Y, d) a metric
space which is locally compact and countable at infinity, by X a compact
space, and by ~y(X) the subset of yX whose elements are continuous on X
into Y. yX is endowed with the topology of simple convergence.
11.1
Theorem. Let H be a subset of ~y(X) such that each sequence in H has
a cluster point in ~y(X) (for the topology of simple convergence). Then the
closure H in yX is compact and contained in ~ y(X), and each element of H is the
limit of a sequence in H.
Before proving the theorem let us note these corollaries.
COROLLARY l.lf H C~y(X) is countably compact for the topology of simple
convergence, then H is compact and sequentially compact.
COROLLARY 2 (Eberlein [1]). Each weakly countably compact subset of a
Banach space E is weakly compact and weakly sequentially compact.
Proof Take (Y, d) to be the scalar field K of E under its usual absolute
value, and X to be the dual unit ball under a(E', E). The map z -+ h(z) = J,
which orders to each z E E'* its restriction f to X, is a homeomorphism of
(E'*, a(E'*, E'» onto a closed subspace Q of KX. By (6.2), h(E) = Q (") ~K(X);
hence the assertion follows from Corollary 1.
186 DUALITY [Ch. IV
Proof of( 11.1). It is clear that for each t E X, the set {f(t): f E H} is relatively
compact in Y; otherwise, there would exist t E X and a sequence {fn: n EN}
in H such that limn!,,(t) = 00, and this sequence would have no cluster point in
yX. It follows now from Tychonov's theorem that His relatively compact in yX.
We show next that H c yg y(X). Assume to the contrary, that there exists a
9 E H not continuous on X. There exists a non-empty subset M c X, a
°
to E M, and an 8> such that d(g(t), g(t o» > 48 whenever t E M. We define
the sequences {to, tl , t z , ... }in Xand{,h,fz, ... } in Hinductively, as follows:
»
Since 9 E H, it is possible to choose an,h E H so that dU~ (to), g(t o < e. After
{to, ... , tn_d and {f1' ... ,fn} have been selected, choose tn E M n M1 n ... n
M., where Mv = {t EX: d(fvCt),fvCt o» < 8} (v = 1, ... , n). Then choose
!,,+1 EH so that d(!,,+l(t.),g(t v» < 8 (v =0,1, ... , n), which is possible,
since 9 E H. From this construction we obtain the following inequalities:
d(g(tn), g(to» > 4e (n EN). (1)
dU~(tv),!"(to» <e for all v;:;; n, n EN. (2)
d(!,,(tv), g(tv» < e whenever °
~ v ~ n - 1, n EN. (3)
By hypothesis, the sequence U;,: n E N} has a cluster point h E yg y( X); then
from (3) it follows that d(h(tv), g(tJ) ~ e for all v ;:;; 0, and from this and (1)
we conclude that d(h(tn), h(to» > 2e for all n EN. Since X is compact, the se-
quence {tn: n EN} has a cluster point SEX, and the continuity of h implies that
d(h(s), h(to» ;:;; 2e. (4)
Since h is a cluster point of {!,,}, there exists an mEN such that d(fm(to) ,
h(to» + d(fm(s), h(s» < e. By (2) and the continuity of fm' we have d(fm(s) ,
fm(to» ~ e. The last two inequalities yield d(h(s), h(lo» < 2e, contradicting (4).
Hence the assumption 9 ~ ygy(X) is absurd, and it follows that H c ygy(X).
There remains to show that each 9 E H is the limit of a sequence in H. As
an intermediate step, we observe that 9 i-s a cluster point of a sequence
{g n: n E N} in H. In fact, if n EN is fixed and (t1' ... , tn) is a given n-tuple of
points in X, there exists (since 9 E H) a function h E H such that (*) d(g(tv),
h(tJ) < n- l (v =1, ... , n); since the topological product xn is compact and
9 is continuous, we can find a finite subset Hn c H such that whatever
(t1' ... , tn) E X n , the relation (*) is satisfied for at least one hE Hn' Now if
such a set Hn is selected for each n EN and if {gn: n EN} is a sequence with
range UnH., then clearly 9 is a cluster point of {gn: n EN}.
The proof will now be completed by showing that a given cluster point 9
of the sequence {gn: n EN} is the (pointwise) limit of a suitable subsequence.
Denote by G the closure of the range of {gn: n EN} in yX. The assumptions
on Y imply that Y has a countable base ()j of open sets. Consider the topology
X on X generated by g-l«)j) and Ung;l«)j); X has a countable base and is
coarser than the given topology of X. The relation R on X defined by " t ~ s
if gn(t) = g.(s)for all n EN" is a closed equivalence relation, and the quotient
(X, ':t)/R is a metrizable, compact space X'. It is clear that g and all gn define
continuous functions g' and g~, respectively, on X' into Y; moreover, each
hE G defines a function h' E yX'. Denoting the corresponding subset of
yX' by G', it follows from· the first part of the proof that G' is contained in
~y(X') and compact (for the topology of simple convergence), and from this
it follows that the topology of simple convergence agrees on G' with the
topology of simple convergence in any dense subset Xo of X' (the latter
being a coarser Hausdorff topology). Since X' (being compact and metrizable)
is separable, there exists a countable dense subset Xo of X', and this implies
that the restriction of the topology of simple convergence (in X') to G' is
metrizable. It follows that g' is the limit of a suitable subsequence of
{g~: n EN}, and, clearly, this implies the assertion. This completes the proof.
We are now prepared to prove the following theorem on weak compact-
ness, usually referred to as the theorem of Eberlein. The equivalence (a) <:>(b)
is essentially Eberlein's result (cf. Corollary 2, above) in a more general
setting, (a)<:>(c) is due to Dieudonne [4], (a) <:>(d) to Grothendieck [6]. See
also Ptak [2]-[5], Smulian [2].
11.2
Theorem. Let E be a l.c.s. and let H be a subset of E whose closed,
convex hull is complete. Thefollowing properties of H are equivalent:
(a) H is relatively weakly compact.
(b) Each sequence in H has a weak cluster point in E.
(c) For each decreasing sequence {Hn: n EN} of closed convex subsets of E
such that Hn n H #- 0 for all n, nnHn is non-empty.
Cd) H is bounded and, for each sequence {x m : mEN} in H and each equi-
continuous sequence {x~: n EN} in E', one has limn limm <xm , x~) =
limm limn <xm , x~) whenever both double limits exist.
Proof. It is clear that (a) implies (b) and (c). To see that (a) implies (d), we
observe that the sequence {x m } has a weak cluster point x E E, since H is
relatively weakly compact, and the sequence {x~} has a weak cluster point
x' E E', since it is equicontinuous. Now if the first double limit exists, it
necessarily equals limn <x, x~) = <Xl x'); likewise, the second double limit
(if it exists) equals <x, x').
To prove the reverse implications, we observe the following: Since the
closed, convex hull C of H is complete, C is closed and hence weakly closed
in the completion E of E; hence to show that a a(E'*, E')-cluster point x* of
H actually belongs to E it suffices, by Grothendieck's theorem (6.2), to show
that the restriction of x* to X is a(E', E)-continuous, where X is an arbitrary
aCE', E)-closed equicontinuous set in E'. In the following, let X be any such
set (which is a compact space under aCE', E)). When considering the elements
of E (respectively, of E'*) as elements of ~K(X) (respectively, of K X , K the
scalar field of E), we actually mean their restrictions to X.
188 DUALITY [Ch. IV
(b) => (a): This is now immediate from (11.1), taking (Y, d) to be Kwith its
usual metric.
(c) => (a): Note first that H is bounded in E; otherwise, there would exist a
continuous, real linear form u on E and a sequence {xn } in H such that
u(xn) > n (n EN), and the sets Hn = {x E E: u(x) ~ n} would satisfy (c) but
have empty intersection. Now consider H as a subset of ~K(X), and denote
by H the closure of H in KX; it suffices to show that H c: ~K(X), On the
assumption that some g E H is not continuous on X we construct, as in the
first part of the proof of (11.1), a sequence {to, t1 , t 2 , ... } in X and a sequence
{ll'/2' ... } in H such that the relations (1), (2), (3) are satisfied. Now denote
by Hn the closed, convex hull of the set {Iv: v ~ n} in E, and let h be an element
of nnHn. Then we obtain (4) again, and replacing the elementfm in the proof
of (ILl) by a suitable elementf"; E Hin we arrive at a contradiction as before.
(d) => (b): Since H is bounded, each sequence {xn } in H has a (1(E", E')-
cluster point X* E E", by (5.4); it suffices to show that the restriction of x*
to X is continuous. Suppose it is not; we can assume that 0 E X and that x* is
discontinuous at 0 E X. Denoting by Un (n EN) the weak O-neighborhood
{x' E E': I<x v , x'>1 < n-l, v = 1, ... , n}, there exists an e> 0 and elements
x~ E Un II X such that I<x*, x~>1 > dor all n. Let Vm = {z E E": I<z, x~>1 <
m- 1 , J1 = I, ... , m} for each mEN. There exists a subsequence {Ym} of {xn }
such that Ym E x* + Vm for all m, and we can further arrange (by choosing a
subsequence of {x~} if necessary) that limn <x*, x~> exists.
Now limm <Ym, x~> = <x*, x~> for each n; hence limn limm <Ym, x~> exists
and is ~ e in absolute value. On the other hand, limn <Ym, x~> = 0 for all m,
which implies limm limn <Ym, x~> = 0, contradicting (d). This completes the
proof of (11.2).
COROLLARY. Let E be a l.c.s. which is quasi-complete for 7:(E, E'). Then each
weakly closed and countably compact subset of E is weakly compact.
The following result is essential for the theorem of Krein to be proved
below. Our proof, which uses the dominated convergence theorem of
Lebesgue (see, e.g., Bourbaki [9] chap. IV, §3, theor. 6 or Halmos [I], §26,
theor. D), is simple but not elementary; for a combinatorial proof and an
enlightening discussion of related questions, we refer to Ptak [7]. Another
combinatorial proof was given by Namioka [2].
11.3
Let E be a I.c.s., let B be a compact subset of E, and let C be the closed,
convex, circled hull of B. If {x~: n EN} is a sequence in E', uniformly bounded
on B and such that limn <x, x~> = 0 for each x E B, then limn <x, x~> = 0 for
each x E C.
Proof. Consider the normed space Ee; without loss in generality we can
suppose that (algebraically) Ee = E. The dual E' of E can be identified with a
subspace of E~, and the mapping x' -+ f (where f(x) = <x, x'), X E B) is a
norm isomorphism 1/1 of the Banach space E~ into ec(B). It follows from
(7.9) that 1/1' maps ec(B)' onto the bidual E~; in particular, each x E E is the
image under 1/1' of a suitable fl E ec(B)'. On the other hand, if f" = I/I(x~)
(n EN), then, since {f,,} is uniformly bounded on B and each fl E ec(B)' is a
linear combination of positive Radon measures on B, Lebesgue's dominated
convergence theorem implies that limn fl(f,,) = ° for all fl E ec(B)'; hence
limn <x, x~) = limn <I/I'(fl), x~) = limn fl(f,,) = 0, where x = I/I'(fl), thus com-
pleting the proof.
Now we can prove the theorem of Krein (cf. M. G. Krein [1]); our proof
follows Kelley-Namioka [1].
11.4
Theorem. Let B be a weakly compact subset of the l.c.s. E, and denote

by C the closed, convex hull of B. Then C is weakly compact if and only if C
is complete for the Mackey topology ,(E, E').
Proof. The condition is clearly necessary, for if C is weakly compact, it is

weakly complete and hence complete for ,(E, E'). The sufficiency will be proved
via condition (d) of (11.2) (cf. Ptak [7]). Let {x n } be a sequence in C and let
{x~} be an equicontinuous (with respect to ,(E, E'» sequence in E'; the
former has a (T(E", E')-cluster point x* E E" (cf. (5.4»; the latter has a (T(E', E)-
cluster point x' EE'.By (11.1) there exists a subsequence of {x~} which
converges to x' pointwise on B. Assume hence that {x~} has this property;
it follows then from (11.3), applied to (E", (T(E", E'», that lim" <z, x~) =
<z, x') for each z E C, where C denotes the (T(E", E')-closure of C in E".
In particular, limn <xm , x~) = <xm , x') for each mEN; if limm limn <xm , x~)
exists then, since x* is a (T(E", E')-cluster point of {x m }, it necessarily equals
<x*, x'). On the other hand, if lim,. <xm , x~) exists, it clearly equals <x*, x~)
and from the preceding it follows that limn <x*, x~) = <x*, x'). Hence the
double limit condition (d) of (11.2) is satisfied and C, being weakly closed, is
weakly compact. The proof is complete.
Since by (I, 5.2) the circled hull of every weakly compact subset of E is
weakly compact, (11.4) continues to hold with C the closed, convex, circled
hull of B.
If B is a compact subset of E and C the closed, convex (or closed, convex,
circled) hull of B, then C is precompact by (II, 4.3) and (I, 5.1); hence C is
compact if and only if it is complete. In view of this remark, it is easy to prove
from (11.4) the following slightly more general version of Krein's theorem.
11.5
Let B be a compact subset of the l.c.s. E and let C be the closed, convex,
circled hull of B. Then C is compact if and only if C is complete for ,(E, E').
190 DUALITY [Ch. IV
EXERCISES
1. Let F be a vector space over K.
(a) Let Q be any non-empty subset of F*, and denote by M the linear
hull of Q in F*. If Z is the topology generated by the semi-norms
x -+ I(x, y)l, y E Q, then (F, Z)' = M and M is in duality with the
Hausdorff t.v.s. Fo associated with (F, Z).
(b) If F is a l.c.s. with dual F', then (f(F, F') is the 6-topology, 6
denoting the set of all (f(F', F)-bounded sets, each of which is contained
in a suitable finite-dimensional subspace of F'.
2. Denote by Ea, Fa two l.c.s. under their respective weak topologies.
The space of continuous bilinear forms on Ea x Fa (equivalently, the dual
of the projective tensor product Ea ® Fa) is the tensor product E' ® F'.
3. Let F be a l.c.s. over C, let G be a properly real subspace of F
(Chapter I, Exercise 16) such that Eo = G + iG, where Fo is the under-
lying real space of F, and denote by G' the set of all real linear forms on
G that are continuous for the topology induced by F. Show the fol-
lowing assertions to be equivalent:
(IX) Fo = G + iG is a topological direct sum for (f(Fo, F6).
({3) E' is the linear hull (over C) of the linear forms
x + iy -+ f(x) + if(y), (x, y E G),
wherefruns through G'.
4. Let E be a normed space which is not complete. Then the 6-top-
ology on E', where 6 is the family of all compact subsets of E, is not
consistent with (E, E'). (Using (6.3), Corollary I, observe that there
exists a compact subset of E whose closed, convex, circled hull is not
weakly compact.)
5. (Perfect Spaces): We use the notation of Section 1, Example 4.
If A is a sequence space with elements x = (xn ), we define an order in A
(Chapter Y, Section 1) by " x ;:;:; y if Xn ;:;:; Yn for all n"; by lxi, we denote
the sequence (Ixni) (which is not necessarily in A). A is perfect (vollkom-
men, Kothe [4]) if A = A x x, solid if x 6 A and Iyl ;:;:; Ixl implies yEA.
Let A be given, and denote by P the set of sequences u ;;;; 0 in Ax ; the top-
ology Z generated by the semi-norms x -+ (lxi, u), u E P, is called the
normal topology of A. (In connection with the following problems, the
reader might want to consult Kothe [4], [5]; see also Dieudonne [3]. A
lattice theoretic characterization of the normal topology can be found
in Chapter Y, Exercise 20.)
(a) The spaces /P (1 ;:;:; p;:;:; + (0) are perfect; each" gestufter Raum"
(Chapter III, Exercise 25) is perfect.
(b) The normal topology Z is consistent with (A, A X).
00
(c) For each uEO)(=K~), define Au= {x EO): Ilunllxnl < + ro}.
1
Every space Au is isomorphic (as a sequence space) to one of the spaces
[1, 0), or [1 E9 0); deduce that each space Au is perfect.
(d) If A is perfect then A is solid and (A, Z) is the projective limit

lim guvAv(Z), where u, v E P, P being directed (;:;:;), and where guv is the
;anonical imbedding of )'v in Au (u ;:;:; v).
EXERCISES 191
(e) For a sequence space A, these propositions are equivalent:

(IX) A is perfect, (p) (A,~) is complete, (y) (A, ~) is weakly semi-complete.
(Use (d).)
6. (Spaces of Minimal Type. Cf. Martineau [1].) A l.c.s. E is said to be
minimal if its topology is minimal (i.e., if there exists no strictly coarser
separated l.c. topology on E).
(a) For a l.c.s. E, these propositions are equivalent: (IX) E is minimal;
(P) E is isomorphic with the product Kt for a suitable cardinal d;
(y) E is weakly complete.
(b) If E is minimal, then u(E, E') = peE, E') and 7:(E', E) = P(E', E);
7:(E', E) is the finest l.c. topology on E'. On a minimal space of infinite
dimension there exists no continuous norm; an infinite dimensional
normed space is never weakly complete.
(c) Each closed subspace and each separated quotient of a minimal
space is minimal; the product of any family of minimal spaces is
minimal.
(d) Let E, F be minimal and let u be a linear map of E into F. u is
continuous if and only if u is closed; if u is continuous, it is a topological
homomorphism.
(e) Let E be minimal. If M, N are closed subspaces, then M + N is
closed in E. If, in addition, M n N = {O}, then M + N is a topological
direct sum; in particular, if E is the algebraic direct sum of the closed
subspaces M, N, then E = M EB N.
7. We use the notation of (4.1).
(a) Let Z denote the unit disk of the complex plane and let F denote
the vector space (over C) of continuous complex-valued functions on Z
possessing continuous first derivatives in the interior of Z. Denote by
IIn(n EN) and 110 the Radon measures on Z defined by IIn(f) = n[f(n- 1)
- f(O)] and 1I0(f) = (2nO -1 Jimc 2 dC (Cauchy integral over the
positively oriented boundary of Z), respectively. Let B = {lin: n EN}
and let G be the subspace of F* generated by 110 and by the set of
Radon measures of finite support on Z. We define 6 1to be the u(G, F)-
saturated hull of the family u {B}; let ~1 be the 6 1-topology on F.
Finally, let M be the subspace of F of functions analytic in the interior
of Z, let ~2 be the topology on M induced by ~1' and let 6 2 be the
(u(GjMO, M)-saturated) family of all ~2-equicontinuous subsets of
GjMo.
In these circumstances, one has 4>(61) "# 6 2, (Prove that 110 is not
contained in any S1 E 6 1, and observe that 4>(110) is in the u(GjMO, M)-
closure of 4>(B).)
(b) Condition (b) of (4.1) implies that 6 2 is the u(GjMO, M)-
saturated hull of the family 4>(61).
(c) Even supposing M'to be a weakly closed subspace of F, the
implication (c) ~ (d) of (4.1) is false unless ~1 is consistent with
(F, G). (Consider a complete l.c.s. E whose strong dual Ep is not barreled
(Exercise 13), and imbed E as a closed subspace M of a product F of
Banach spaces, (II, 5.4), Corollary 2. By (4.2), the strong dual F/J is
barreled as a l.c. direct sum of barreled spaces; hence the quotient of
PCF'. F) is a barreled topology, and therefore distinct from P(F' jMO, M).)
192 DUALITY [Ch. IV
8. Let {<F~, G~>: (X E A} be a family of dualities over K and let F =

n~FIX' G = EBIXGIX . F and G are placed in duality in the canonical way.
(a) P(F, G) is the product of the topologies P(F~, G~), and peG, F)
is the l.c. direct sum topology of the topologies P( G~, FIX)' (Use (4.2).)
(b) The weak topologya(G, F) is the l.c. direct sum topology of the
topologies a(GIX , F~) if and only if the number of spaces F~ '" {O} is finite.
(c) A fundamental family of convex, circled, weakly compact subsets
of F (respectively, G) is obtained by forming arbitrary products of like
subsets of the FIX (respectively, by forming arbitrary finite sums of like
subsets of the G~).
9. Construct an example of a l.c.s. F and a closed subspace H such
that not every bounded subset of FI H is the canonical image of a
bounded subset of F. (Consider E = /1; for each vector x ~ 0 of E (for
notation, see Exercise 5) let B(x) = {y E E: Iyl ;:'ii; x}. Show that the
(norm) topology of E is the finest I.c. topology for which all B(x) are
bounded. RepresentEasaquotientFIH, whereF = EEl EB(x)') Similarly,
x~o
show that a weakly compact subset of FIH is not necessarily the

canonical image of a like set of F. (Apply the preceding method to
E = /2.) See also Exercise 20.
10. Show that a closed subspace of a barreled space is not necessarily
barreled. (Consider a complete I.c.s. E which is not barreled, and
imbed E as a closed subspace of a product of Banach spaces (use (4.3),
Corollary 3); for example, it suffices to take for E a sequence space IP,
p> 1, under its normal topology (Exercise 5).)
REMARK. Using the theorem of Mackey-Ulam (cf. Chapter II, Sec-
tion 8), the same method provides an example of a bornological space
possessing a non-bornological closed subspace. See also Exercise 20.
11. (Incomplete Quotients.) Let X denote a completely reguiar topo-
logical space, R(X) the I.c.s. (over R) of real-valued continuous functions
on X endowed with the topology of compact convergence.
(a) If a function! E R X is continuous whenever its restriction to each
compact subset of X is continuous, then R(X) is complete.
(b) Let Y be a closed subset of X and let H be the subspace of R(X)
whose elements vanish on Y. The quotient space R(X)IH can be
identified with the subspace Ro( Y) of R( Y) whose elements have a
continuous extension to X Ro( Y) is dense in R( Y) (use the theorem of
Stone-Weierstrass, (y, 8.1».
(c) If Y is such that! E RY is continuous whenever its restrictions to
compact sets are continuous, then R(X)IH is complete if and only if
Ro(Y) = R(Y).
(d) Deduce from the preceding an example of a complete I.c.s. R(X)
such that for a suitable closed subspace H, R(X)IH is not complete. (Con-
sider, for example, a locally compact space X which is not normal (Bour-
baki [5], §4, Exercise l3); there exists a closed subspace Y such that not
every continuous real function on Yhas a continuous extension to X)
12. (Topological Complementary Subspaces):
(a) Let E be a non-normable (F)-space on which there exists a
EXERCISES 193
continuous norm (e.g., the space fifiG (Chapter II, Section 6, Example 2»;
E contains a closed subspace which has no topological complement.
(Take an increasing fundamental sequence {Bn: n E N} of convex, circled,
bounded subsets of E' such that Enn #- Enn+ l ' and select x~ E Enn+ 1 ,...,
Enn. Show that the linear hull F of {x~: n EN} is a subspace of E' whose
bounded subsets are finite dimensional; deduce from this, using the
theorem of Krein-Smulian (6.4), that F is a weakly closed subspace of
E'. Since every linear form on F is continuous (use (6.2», the dual Ejr
of F is weakly complete, hence isomorphic with K~ (Exercise 6). If
po had a topological complement, it would be isomorphic with K~,
which contradicts the fact that there exists no continuous norm on K:
(Exercise 6).)
(b) Let E be a given Banach space, {y~: rx E A} a dense subset of the
unit ball of E. Then E is isomorphic with a quotient space of IleA).
(For each x = (~~) E IleA), define u(x) = L ~~y~ (Chapter III, Exercise
~eA
23); conclude from (III, 2.1) that x -+ u(x) is a homomorphism of [1(A)
onto E.) Deduce that whenever A is infinite, [1(A) contains a closed sub-
space that has no topological complement.
(c) If E is a complete t.v.s. and His a closed subspace such that EjH
is not complete (Exercise 11), then H does not have a topological
complement.
For a discussion of the problem of complementary subspaces, see Day
[2] and Kothe [5], §31.
13. A I.c.s. E is called distinguished if its strong dual Eft is barreled.
Let E be the vector space of all numerical double sequences x = (Xi)
such that for each n EN, Pn(x) = Li)ali)xijl < + 00, where ali) = j
for i ~ nand allj, al~;> = 1 for i > nand allj. The semi-norms Pn (n EN)
generate a I.c. topology under which E is an (F)-space which is not dis-
tinguished (Grothendieck [10]). Establish successively the following
partial results:
(a) The dual E' can be identified with the space of double sequences
u = (Ui) such that IUijl ~ cali) for all i, j and suitable c > 0, n EN
(cf. Chapter III, Exercise 25). If Bn is the polar of Un = {x:PnCx) ~ I},
then {nBn: n EN} is a fundamental family of bounded subsets of E'.
(b) Let W denote the convex, circled hull of UnrnBn; W absorbs all
bounded subsets of E', and W does not contain a U E E' such that for
each i, there exists j with IUijl ~ 2.
(c) Given a sequence P = (Pn) of strictly positive numbers, define
elements u(n) E E' so that u\'j) = 0 for (i,j) #- (n, k n ), u~7l.. = 1, where
kn is chosen so that 2n+ 1 u(n) E PnBn. For each given p, the sequence with
L u(n) is a weak Cauchy sequence in E', hence con-
N
general term SN = 2
I
vergent to sEE'.
(d) For each strong O-neighborhood BO in E' (B a bounded subset of E)
r
there exists a sequence P = (Pn) of numbers> 0 such that nPnBn c BO;
if {SN} is a sequence as constructed in (c), then SN E BO for all
N EN; hence S E BO but S 1: W. It follows that W contains no BO; hence
E'p is not bornological and, therefore, not barreled (use (6.6».
194 DUALITY [Ch. IV
14. If <F, G) is a duality, the strong topology [3(F, G) is, in general,

not inherited by closed subspaces or separated quotients. (Concerning
subspaces, consider a I.c.s. F as constructed in Exercise 9; then [3(HO,
FI H) is distinct from the topology induced on HO by [3(P, F). Concerning
quotients, consider a non-distinguished complete I.c.s. E (Exercise 13)
as a closed subspace of a suitable product G of Banach spaces; then
[3(G'IEo, E) is distinct from the quotient of [3(G', G).)
If F is a Banach space, then the strong topology is inherited by closed
subspaces and separated quotients; the same is true (with respect to
<F, F'») for the strong dual of F.
15. We use the notation of Section 5. There exist I.c.s. E such that the
families G:, (f, ~, ~" of subsets of E' are all distinct. (Consider a suit-
able space Fwhich is not barreled (Chapter II, Exercise 14), and take E
to be the space F;.)
16. (Infrabarreled Spaces). The product and I.c. direct sum of any
family of infrabarreled spaces is infrabarreled (use Exercise 8). Every
separated quotient of an infra barreled space is infrabarreled (immediate
verification), but a closed subspace is not necessarily infrabarreled
(Exercise 20).
17. (Theorem of Banach-Mackey). If E is a I.c.s. and B is a bounded,
convex, circled subset of E such that the normed space E B is complete,
then B is bounded for [3(E, E'). (Observe that for each barrel DeE,
D n EB is a barrel in the Banach space E B, and use (1.6).)
18. (Reflexive Spaces). (See also Exercise 20.)
(a) The Banach space [peA) is reflexive whenever 1 < p < + 00.
(b) If E is a quasi-complete Mackey space such that E/J is semi-
reflexive, then E is reflexive.
(c) If E is a weakly semi-complete I.c.s. whose strong dual Ep is
separable, then E is semi-reflexive. (Use (l.7).)
(d) Give an example of a non-reflexive l.c.s. whose strong dual is
reflexive.
(e) If E is a non-reflexive (F)-space, the canonical inclusions E c E"
c E(iv) c ... and E' c EIII c E(v) c ... are all proper. (Conduct the proof
indirectly, using (b).)
(f) There exist non-reflexive (B)-spaces E such that the canonical
image of E in E" is of finite codimension (see James [2] and Civin-Y ood
[I D· Deduce from this that there exist infinite-dimensional Banach spaces
E such that E is not isomorphic (as a t.v.s.) with Ex E.
19. (Montel Spaces)
(a) Every product and l.c. direct sum of a family of (M)-spaces is an
(M)-space; the strict inductive limit of a sequence of (M)-spaces is
an (M)-space. By contrast, closed subspaces and separated quotients of
(M)-spaces are, in general, not even reflexive (Exercise 20).
(b) Each barreled, quasi-complete nuclear space is an (M)-space
(use (III, 7.2), Corollary 2).
(c) A separable (F)-space E is an (M)-space if and only if each a(E', E)-
convergent sequence in E' is strongly convergent. (For the sufficiency
of the condition, use (l.7). Show that each bounded subset of E' is
EXERCISES 195
strongly precompact (Chapter I, Exercise 5) (hence relatively compact

since Ep is complete). From (6.2), Corollary 3, conclude that E = E",
hence that E and Ep are reflexive; finally use (5.9).)
(d) Every metrizable (M)-space is separable (Dieudonne [7]). (Select
a sequence {Un} of convex, circled O-neighborhoods forming a base at 0,
and imbed E as a subspace of TInEun Supposing E to be non-separable,
it can be assumed that E u , is not separable. Denoting by BI an uncount-
able subset of E u , whose elements have mutual distance ~ b > 0, let
MI = ¢u,1(B1 )· There exists an uncountable proper subset M 2 of MI
such that ¢u2(M2 ) is bounded in E U2 , etc. If Xn E Mn ~Mn+1 for all
n EN, then {x n } is a bounded sequence in Esuch that {¢u,(x n )} contains
no Cauchy subsequence, which is contradictory.)
20. Let E be the vector space (over R) of all double sequences
x = (Xi) such that for each n EN, Pn(X) = Li,ja\~plxijl < + 00, where
a\~? = P for i < n and all}, and ali) = in for i ~ n and all}. Under the top-
ology generated by the semi-norms Pn(n EN), E is an (F)-space and an
(M)-space (cf. Exercise 13). The dual E' can be identified with the space
of all double sequences x' = (x;) such that Ix;) ~ cal'? for all i, } and
suitable c > 0, n EN (the canonical bilinear form being (x, x') ---> (x, Xl)
= Li,jXijX; .). Each x E E defines a sum mabie family {Xii (i,j) EN x N}
(Chapter III, Exercise 23); if one puts Yj = LiXij, then Y = (Yj) E [I, and
x ---> Y = u(x) is a continuous linear map of E onto a dense subspace
of/I. (Kothe [1], Grothendieck [10].)
(a) The adjoint u' is biunivocal and onto a closed subspace u'(r)
of E'. Hence u is a topological homomorphism of E onto /1 (use (7.7»;
thus Eju- \0) is isomorphic with /1.
(b) The canonical map E ---> Eju-I(O) maps the family of all bounded
subsets of E onto the family of all relatively compact subsets of
Eju-I(O). (Use that E is an (M)-space and an (F)-space, and (6.3),
Corollary 1.) Infer from this that peE', E) induces on u'(l"J) the 6-
topology, where 6 is the family of all relatively compact subsets of
Eju- 1 (0). u'(lOCJ) is not infrabarreled for this topology.
(c) Conclude from the preceding that (1) a closed subspace or a
separated quotient of an (M)-space is not necessarily reflexive; (2) a
bounded set in the quotient of an (F)-space E is not necessarily the
canonical image of a bounded set in E; (3) a closed subspace of a
barreled (respectively, bomological) l.c.s. is not necessarily barreled
(respectively, bomological). (The strong dual of the space E, above, is
barreled, and bomological by (6.6), Corollary 1.)
21. The Banach-Mackey theorem (Exercise 17) implies thatin (II, 8.5)
it suffices for the conclusion to assume that B is semi-complete. This
permits us to improve several earlier results, for example, to show that
every semi-complete bom610gical space is barreled. The following re-
sults show that, in general, the concepts of semi-completeness, quasi-
completeness, and completeness should be carefully distinguished.
(a) There exists a semi-complete l.c.s. that is not quasi-complete.
(Consider [I under its weak topology.)
(b) Let E be a I.c.s. For E~ to be complete, it is necessary and
196 DUALITY [Ch. IV
sufficient that T(E, E') be the finest I.c. topology on E (cf. Exercise 6).
Infer from this that if E is metrizable and E; is complete, then E is
finite dimensional. (Use Chapter II, Exercise 7.) If E is normable and Ea
complete, then E is finite dimensional. (Apply the preceding result to E p.)
(c) In each Hilbert space H, there exists a weakly complete convex
set C such that H = C - C. (Suppose H to be a space [2(A) over R
(Chapter II, Section 2, Example 5), and take C = {x E H: e", ~ 0 for
all IX E A}.)
22. Let E be a I.c.s., let E' be its dual, and let ;tf be the finest topology
on E' that agrees with (1(E', E) on every equicontinuous subset of E'.
(a) ;tf is a translation-invariant topology possessing a O-neighborhood
base of radial and circled sets.
(b) In general, ;tf fails to be locally convex (Collins [1]). (Consider
an infinite-dimensional space E on which there exists a topology ;tl
such that (E, ;tl) is an (F)-space; denote by ;to the finest I.c. topology
on E (Chapter II, Exercise 7). Then (E, ;tt)' = E' ~ E* (Exercise 21).
Consider ;tf on E* = (E, ;to),; then E' is ;trclosed in E*. If ;tf were
locally convex, it would be consistent with <E, E*) (observe that
(E, ;to) is complete, and use (6.2), Corollary 3); hence E' would be
closed and dense in (E*, (1(E*, E», which is contradictory.)
23. Let E be a separable, metrizable I.c.s. and let it be its completion.
Then each bounded subset of E is contained in the closure (taken in E)
of a suitable bounded subset of E. (Observe that E can be identified
with the closure of E in the strong bidual E". If BeE is bounded, then
B is separable, hence equicontinuous in E" (use (6.5), Corollary 1);
thus B c AOO (bipolar with respect to <E', En» for a suitable bounded
subset A of E.)
24. «DF)-spaces. See also Grothendieck [10], Kothe [5]).
(a) The strong dual E/J of a (DF)-space E is an (F)-space. (Use the
corollary of (6.7) to show that E/J is complete.)
(b) If E is a (DF)-space and M is a closed subspace, then P(MO, E/M)
is the topology induced on MO by P(E', E). (Prove that the identity map
of (MO, P(E', E» onto (M O , P(M O , E/M» is continuous by showing
that each P(E', E)-null sequence in MO is equicontinuous in E' (hence
P(MO, E/M)-bounded), using that P(E', E) is metrizable (cf. Chapter II,
Exercise 17).)
(c) If E is a i.c.s., and M is a subspace which is a (DF)-space, then
P(E' / MO, M) is the quotient of P(E", E). (Employ the same method as in
(b), using that each P(E'/Mo, M)-null sequence is equicontinuous and
hence by (7.5) the canonical image of an equicontinuous sequence in E'.)
(d) Let E be a (DF)-space. The completion E can (algebraically) be
identified with a subspace of E"; moreover, each bounded subset of the
strong bidual E" is contained in the bipolar (with respect to <E', E"»
of a suitable bounded subset of E. Each bounded subset of it is contained
in the closure (taken in 2) of a suitable bounded subset of E. (Use (c) to
prove that P(E', E) = P(E', 2).) Conclude that E is a (DF)-space, and
that every quasi-complete (DF)-space is complete.
(e) Each separated quotient of a (DF)-space is a (DF)-space. (Use
EXERCISES 197
(b) to show that each bounded subset of ElM is the canonical image
of bounded subset of E.) The I.c. direct sum of a sequence of
(DF)-spaces is a (DF)-space, and an inductive limit of a sequence of
(DF)-spaces is a (DF)-space. On the other hand, an infinite product
of (DF)-spaces (each not reduced to {O}) is not a (DF)-space; a closed
subspace of a (DF)-space is not necessarily a (DF)-space.
(f) Let E = lim hnmEm be an inductive limit of a sequence of reflexive
(DF)-spaces. T~ the strong dual of E can be identified with the pro-
jective limit lim gmnE~ of the strong duals E~ with respect to the adjoint
+-
maps gmn = h~m (cf. end of Section 4).
(g) There exist complete (DF)-spaces not isomorphic with the strong
dual of a metrizable I.c.s. (Consider anon-separable reflexive Banach space
E under the topology of uniform convergence on the strongly separ-
able, bounded subsets of E'.)
25. Let E, F be I.c.s. with respective duals E', F'. For any subset
Q c: 2(E, F), denote by Q' the set of adjoints {u': u E Q}. Consider
the propositions:
(i) Q is equicontinuous.
(ii) For each equicontinuous set B c: E', Q'(B) is equicontinuous
inF'.
(iii) Q' is equicontinuous in 2(F;, E;J
(iv) Q is bounded in 2 b(E, F).
(v) Q' is simply bounded in L(F', E') with respect to a(E', E).
(vi) Q' is simply bounded in L(F', E') with respect to P(E', E).
One has the following implications: (a) (i)=-(ii) and (iii)=-(iv); (b) If
E is infrabarreled, then (i) =-(iii) =- (vi); (c) If E is barreled, propositions
(i) through (vi) are equivalent.
26. Give an example of two l.c.s. E, F and a u E 2(E, F) such that
u is a topological homomorphism for a(E, E') and a(F, F') but not
for 7:(E, E') and 7:(F, F'). (Take F to be a Mackey space such that some
subspace E is not a Mackey space (cf. Exercise 9), and consider the
canonical imbedding E -+ F.)
Give an example of two (F)-spaces E, F and a topological homomor-
phism u of E onto F such that u' is not an isomorphism of F/J into E/J.
(Cf. Exercise 20.)
27. Let E, F be normable spaces and u E 2(E, F).
(a) Give an example where E is not complete, u is a homomorphism,
and u' is a strong homomorphism but not a weak homomorphism.
(Take E = Fo, where Fo is a dense subspace =F F of F.)
(b) Give an example where E is not complete, u is not a homomor-
phism, and u' is a weak and strong homomorphism.
(c) Give an example where E is complete, F is not complete, and u'
is a weak but not a strong homomorphism.
28. Let E, F be l.c.s. and let u be a linear map of E onto F.
(a) The following properties of u are equivalent:
(i) u is nearly open.
(ii) For each subset A c: E, u(A) is contained in the interior of
[u(A)r·
198 DUALITY [Ch. IV
(iii) For each convex, circled O-neighborhood U c: E, u(U) is weakly

dense in some O-neighborhood V c: F.
(b) If F is barreled, u is nearly open.
29. (B-Completeness).
(a) Let E be an infinite-dimensional vector space such that there exists
a topology l:1 under which E is an (F)-space, and denote by l:o the
finest l.c. topology on E. Then (E, l:o) is complete but not B,-complete.
(Observe that (E, l:o) is not metrizable (Chapter II, Exercise 7) and use
(8.4); cf. Exercise 22.)
(b) Show that products and l.c. direct sums of B-complete spaces are,
in general, not B-complete. (Concerning l.c. direct sums, use (a); con-
cerning products, imbed a complete but not B-complete space into a
product of Banach spaces, and use (8.2).)
(c) The following assertions are equivalent: (i) Every B,-complete
space is B-complete. (ii) Every separated quotient of a B,-complete
space is B,-complete.
(d) Let X be a completely regular topological space, R(X) the
space of real-valued continuous functions on X under the topology
of compact convergence. The following propositions are equivalent:
(i) R(X) is B;complete. (ii) If Y is a dense subset of X such that Y n C
is compact whenever C is compact, then Y = X. (Note that the map-
ping t -+ (f -+ J(t» is a homeomorphism of X into the weak dual of
R(X).)
30. Let E, F be Banach spaces with respective strong duals E', F'.
(a) The dual of .P c(E, F) (topology of compact convergence) can be
identified with a quotient of E ®F' and, if F is reflexive, with E ®F'.
(Observe that .Pb(E, F) is canonically isomorphic with a closed sub-
space of r!4 b(E, F'), the latter space being the strong dual of E ®F' by
(9.8). The dual of .Ps(E, F) can be identified with E ® F' by (4.3), Corol-
lary 4, and on the bounded subsets of .Pb(E, F) the topologies of simple
and of compact convergence agree (theorem of Banach-Steinhaus).
Finally, use Grothendieck's theorem (6.2).)
(b) E has the approximation property if and only if the canonical
map -r: E' ® E -+ 2(E) is biunivocal. (Use (a). Cf. Chapter III, Section 9,
and Grothendieck [13], I, §5, prop. 35.)
(c) If E' possesses the approximation property, then so does E. (Ob-
serve that the canonical imbedding p: E' ® E -+ E' ® En is a topological
isomorphism (in particular, injective). If u E E' ® E and w = "C(u) is its
canonical image in .P(E), if v = p(u), and if q E .P(E') is the endomor-
°
phism defined by v, then q = WI. Hence if w = 0, conclude from (b) that
v = and thus u = O. Use (b) once more.)
31. Show that in the dual of a nuclear space E, each equicontinuous
subset is metrizable for (1(E', E) (hence separable). (Use (9.3).) Infer from
this that if d is an uncountable cardinal, then Kg is a nuclear space whose
strong dual fails to be nuclear.
32. Let E, F be (DF)-spaces.
(a) The projective tensor product E ® F as well as its completion
E ® Fare (DF)-spaces. (Use (9.8); cf. Exercise 24.)
EXERCISES 199
(b) If E is nuclear and complete, then E is an (M)-space, namely the

strong dual of the (F)- and (M)-space E~. (Observe that E is semi-
reflexive, and prove that its strong dual Ep = E; is a reflexive (F)-space;
for this, note that each strongly bounded sequence in E' is equi-
continuous, and use (11.2). Conclude that E; is an (M)-space (use (Ill,
4.5), and thus that each strongly bounded set is separable and hence
equicontinuous.)
(c) Suppose that E is nuclear. The strong dual of E ® F can be iden-
tified with E/J ®Fo' and the strong bidual with E" ®E". (Assume without
loss of generality that E is complete, hence an (M)-space (use (b»; then
proceed as in the second part of the proof of (9.9). To obtain the second
assertion, apply (9.9) to Ep and F/J. Cf. Exercise 33(b).)
(d) (Proof of (9.9), Corollary 3). Suppose that E, Fare strong duals of
(F)-spaces and that E is nuclear. Then !B(E, F) = f16(E, F). (Show that a
continuous linear map u of E into (E", (f(E", F'» is continuous for
(f(F", F"'), using that E is bomological.)
33. (Nuclear (F)- and (DF)-spaces).
(a) If E is an (F)-space whose strong dual is nuclear, then E is nuclear.
(Establish this through the following steps, F denoting an arbitrary
(B)-space: _
1. The canonical map tjJ: E ® F ~ !B(E~, F~) is surjective. [Observe
that E is reflexive, and that by (9.3) EA ~ E is nuclear for each closed,
convex, circled, bounded subset AcE. Consider an element of!B(E~, F~)
as a weakly continuous linear map of F/J into E.]
2. tjJ is injective. [Each U E 2(E, F/J) can be approximated, uniformly
on every compact subset of E, by maps of finite rank; infer that E' ® F'
is dense in f16(E, F) for the 6 x ::t-topology, 6 and ::t denoting the
families of relatively compact subsets of E and F, respectively. Since the
canonical imbedding E' ® F' ~ £74(E, F) can be viewed as the adjoint
of tjJ, the assertion rests on the fact that the 6 x ::t-topology is con-
sistent with the duality <E ® F, f16(E, F). For this, see Grothendieck
[13], I, §4, prop. 21.]
3. From the preceding and Banach's theorem (Ill, 2.1) it follows that
tjJ is a topological isomorphism onto !Be(E~, F~). Use (10.7), Corol-
lary 1.)
(b) If E is an (F)-space or a complete (DF)-space, then E is nuclear
if and only if its strong dual is nuclear. (Use (a), (9.7) and Exercise
32(b).)
(c) Let E denote a nuclear (DF)-space and let F denote a nuclear
(F)-space; then 2 bee, F) and its strong dual are nuclear.
34. If F is a nuclear space, then its bidual F~ (natural topology) is
nuclear. Use this to show that in (9.7) the assumption that E is semi-
reflexive is dispensable.
35. We use the notation of Section 10. Let A be any non-empty index
set and let E be a given l.c.s. For a subset P of [l[A, E]' (respectively,
for a subset Q c [l(A, En to be equicontinuous, it is necessary and
sufficient that the union of the ranges of all x' E P (respectively, of all
x' E Q) be an equicontinuous (respectively, prenuclear) subset of E'.
200 DUALITY [Ch. IV
36. The notation is as in Section 10.

(a) If E is a l.c.s. such that each summable sequence in E is absolutely
summable, then each bounded subset of [l(N, E) is bounded as a subset
of [l[N, E]. (Consider a sequence {x~} in E' such that Lnl<xm x~)1
< + 00 for each x = (xn) E [l(N, E), and show that the function
x -+ Lnl<xm x~)1 is bounded on every bounded subset of P(N, E). See
Pietsch [5], p. 54.)
(b) If [l(N, E) is infrabarreled (e.g., if E is metrizable) and each
summable sequence in E is absolutely sum mable, then the canonical
map [1 [N, E] -+ [l(N, E) is a topological isomorphism (hence E nuclear).
(Use (a).)
(c) Deduce from (b) the following improvement of the theorem of
Dvoretzky-Rogers: If E is a normable space in which each sum mabie
sequence is absolutely summable, then E is finite dimensional.
37. Let E be the Montel space Kg, where d is the cardinality of the
continuum. Denote by B the subset [0, l]d of E, and by Bo the subset of
B whose elements have not more than countably many non-zero co-
ordinates. Show that (under the topology induced by E) Band Bo are com-
pletely regular topological spaces such that Bo is sequentially compact
but not compact, and B is compact but not sequentially compact. More-
over, Bo is an example of a uniform space which is semi-complete and
precompact but not complete.
38. (Weak Countable Compactness). Let (E, :!) be a l.c.s.
(a) If there exists a metrizable l.c. topology on E which is coarser
than :!, then each weakly countably compact subset B of E is weakly
sequentially compact. (Dieudonne-Schwartz [1]. If {xn} is a sequence in
B, the closed linear hull M of its range is separable. Show that the dual
M' is weakly separable (cf. (1.7)), and hence that {xn} contains a weakly
convergent subsequence.)
(b) Suppose that E' is the union of countably many weakly compact
subsets. Then, given a subset M c E, each point in the weak closure of
M is contained in the closure of a suitable countable subset of M
(Kaplan sky). (Use a method similar to that employed in the second part
of the proof of (11.1), or see Kothe [5], §24.1.)
(c) Deduce from (a) and (b) extensions of Theorem (11.2) for metriz-
able I.c.s. and for (LF)-spaces.
The two remaining problems are included for later reference; the
results (needed in the Exercises of Chapter V and, to some extent, in the
Appendix) are easy to prove if the reader is familiar with Sections 1-3
of this chapter and the basic theory of functions of one complex variable.
39. (Vector Valued Analytic Functions). Let G be a non-empty open
subset of the Riemann sphere and let E be a l.c.s. over C. A function
I: G -+ E is called holomorphic at (0 E G if there exists a neighborhood
Z of (0 such that for each x' E E', the function ( -+ </(0, x') is complex
differentiable in the interior Zo of Z, and if for each ( E Zo the linear
form
x' -+ ~ <f«(), x')
EXERCISES 201
is a(E', E)-continuous. The unique element x( 0 E E representing this

linear form is called the derivative of f at " and is usually denoted
by f'm. If 00 E G,f is called holomorphic at 00 if ( -+ f( C 1) is holo-
morphic at o. A function f: G -+ E which is holomorphic at each ( E G
is called locally holomorphic in G. (Note that G need not be connected.)
Let Ddenote a rectifiable oriented arc in the complex plane. A function
f: D--+ E is called (Riemann) integrable over D if for each x' E E', the
function (-+ (fm, x') is integrable over D in the Riemann-Cauchy
sense, and if the linear form x' -+ S6(fm, x') d( is a(E', E)-continuous.
The unique element of E representing this linear form is called the
(Riemann) integral off over D, and is usually denoted by SoI( 0 d(.
In the following we suppose E to be a I.c.s. over C such that the
closed, convex hull of each compact subset of E is compact (which is,
in particular, the case if E is quasi-complete).
(a) Each continuous function f: D -+ E is integrable over D. (Use the
fact that the closed, convex, circled hull of feb) is compact in E.) The
linear map f -+ Solm d( is continuous on <?J E(b), endowed with the top-
ology of uniform convergence, into E.
(b) Each function f: G -+ E which is locally holomorphic is con-
tinuous. (Consider the difference quotient off at (0 E G.)
(c) (Cauchy's theorem). Let G be open and let y c G be a positively
oriented, closed, rectifiable Jordan curve whose interior G y belongs to G.
Then fyfm d( = 0 for every function f which is locally holomorphic
in G. Deduce from this that if (0 E G y , then (fbeing locally holomorphic
f
in G)
1«(~) = ~
2m
1(0
y ( - (0
de.
(d) Infer from (c) that if f is locally holomorphic in G and (0 E G,
then
00
1m = L an «( - (ot,
n=O
where an = (2ni)-1 SJm«( - (0)-n- 1d( (n = 0, 1, ... ) (y denoting a

circle about (0 whose interior is in G) is an expansion valid in a circular
neighborhood of (0 (in contrast with ordinary usage, scalars are written
to the right of elements of E). The series converges in the interior
of y, with respect to T(E, E') and uniformly on compact sets. Moreover,
y can be taken as the largest circle of center (0 to whose interior fhas a
holomorphic extensionJ; the series then converges to J(O.
(e) (Liouville's theorem). Every E-valued function J, holomorphic
and uniformly bounded on the entire complex plane, is constant (i.e.,
has a range consisting of a single element of E).
(f) Define for E-valued functions the concepts of pole and isolated
singularity, and generalize the classical results on Laurent expansions.
40. (Locally Convex Algebras). Let A be an algebra over K. (Recall
that A can be defined as a vector space A o over K on which a bilinear
map Ao x A o -+ Ao, called multiplication and usually denoted by
202 DUALITY [Ch. IV
(a, b) --+ ab, is specified; Ao is then called the underlying vector space
of A. If A has a unit, it will be denoted bye; the inverse of an element
a E A is denoted by a-I.)
An algebra A over K (K = R or C) is called a locally convex algebra
over K if Ao is a I.c.s. and if multiplication is separately continuous.
A normed algebra over K is an algebra A over K such that Ao is a normed
space, with the additional requirement that Ilab II ~ Iia II lib II for all
a, b E A, and that Ilell = I if A has a unit e. A Banach algebra is a
normed algebra A such that Ao is a Banach space.
(a) If E is a I.c.s., 6 a total family of bounded subsets of E such
that 6 is invariant under each U E !feE), then (with respect to the com-
position of maps) !feE) is a I.c. algebra under the 6-topology. If ~
denotes the family of all bounded subsets of E, multiplication is right and
left ~-hypocontinuous for the topology of bounded convergence. If
E is barreled, multiplication is left ~-hypocontinuous for the topology
of simple convergence. If !f 6(E) is an (F)-space, then multiplication
is continuous by (III, 5.1).
(b) Let A be a I.c. algebra over C with unit e. The spectrum O'(a) of
a E A is the complement of the largest open subset G of the Riemann
sphere such that A --+ (Ae - a)-l exists and is locally holomorphic in G.
(If A --+ R(A) = (Ae - a)-l is holomorphic in a neighborhood of 00, the
definition R(oo) = 0 renders R holomorphic at 00.) Show that O'(a) "# 0
for all a EA. (Use Exercise 39(e).)
(c) Let A be a I.e. algebra over C with unit e and let a EA. For the
resolvent A--+ R(A) of a to be holomorphic at Ao E C, it is necessary and
sufficient that R(A) exists in some neighborhood U of Ao such that for
each sequence {An} in U, the sequence {R(An)} is bounded in A. (Use the
resolvent equation R(A) - R(f.l) = - (A - f.l)R(A)R(f.l), which holds when-
ever R(A), R(f.l) exist.)
(d) If A is a Banach algebra over C with unit e, then for each a E A,
O'(a) is a compact subset of C. If rea) is the radius of the smallest circle
of center 0 in C that contains O'(a), we have the relation rea) =
limnllan I1 1 / n. (See, e.g., Hille-Phillips [1].) (r(a) is called the spectral
radius of a E A whenever A is a l.c. algebra over C with unit e.)
PREREQUISITES

ences can be found.
A. SETS AND ORDER

1

204 ORDER STRUCTURES [Ch. V
normal cones, among them (Theorem (4.3» the abstract version of a classical
theorem of Dini on monotone convergence. The duality of ordered vector
spaces is not discussed there, since such a discussion would have amounted to
a direct application of the results of Section 3, which can be left to the reader.
Section 5 is concerned with the induced order structure on spaces of linear
mappings; the principal results are Theorem (5.4) on the extension of con-
tinuous positive linear forms, and Theorem (5.5) establishing the continuity
of a large class of positive linear forms and mappings. The order topology,
a locally convex topology accompanying every ordered vector space over R,
is studied in some detail in Section 6. The importance of this topology stems
in part from the fact that it is the topology of many ordered t.v.s. occurring
in analysis. Section 7 treats topological (in particular, locally convex) vector
lattices. We obtain results especially on the strong dual of a locally convex
vector lattice, and characterizations of vector lattices of minimal type in
terms of order convergence and in terms of the evaluation map. (For the
continuity of the lattice operations see Exercise 20.) The section concludes
with a discussion of weak order units.
Section 8 is concerned with the vector lattice of all continuous real valued
functions on a compact space, and with abstract Lebesgue spaces. The Stone-
Weierstrass theorem is presented in both its order theoretic and its algebraic
form. Further, the dual character of (AM)-spaces with unit and (AL)-spaces
is studied as an illuminating example of the duality of topological vector
lattices treated in Section 7. (AL)-spaces are represented as bands of Radon
measures, characterized by a convergence property, on extremally discon-
nected compact spaces. The classical representation theorem of Kakutani
for (AM)-spaces with unit is established, and an application is made to the
representation of a much more general class of locally convex vector lattices.
1. ORDERED VECTOR SPACES OVER THE REAL FIELD

Throughout this section, we consider only vector spaces over the real
field R.
Let L be a vector space over R which is endowed with an order structure R
defined by a reflexive, transitive, and anti-symmetric binary relation "~";
L is called an ordered vector space over R if the following axioms are satisfied:
(LO)1 X ~ Y implies x + z ~ y + z for all x, y, Z E L

(LO)2 X ~ Y implies Ax ~ AY for all x, y ELand A > O.
(LO)1 expresses that the order of L is translation-invariant, (LOh expresses

the invariance of the order under homothetic maps x --> AX with ratio A > O.
Examples of ordered vector spaces abound; for example, every vector space
of real-valued functions f on a set T is naturally ordered by the relation
"f;;;2 g iff(t) ~ get) for all t E T"; in this fashion, one obtains a large number
of ordered vector spaces from the examples given in Chapter II, Section 2,
§1] ORDERED VECTOR SPACES OVER THE REAL FIELD 205
and Chapter III, Section 8, by considering real-valued functions only and

taking K= R.
It is immediate from the axioms above that in an ordered vector space L,
the subset e ={x: x;;;; O} is a convex cone of vertex 0 satisfying en - e
= {O}; a cone in L with these properties is called a proper cone in L. The
elements x E C are called positive, and C is called the positive cone. of the or-
dered vector space L.
Two ordered vector spaces L1> L2 are isomorphic if there exists a linear
biunivocal map u of L1 onto L2 such that x ~ y if and only if u(x) ~ u(y)
(equivalently, such that u maps the positive cone of L1 onto the positive
cone of L 2 ).
If L is any vector space over R, a proper cone H c L is characterized by
the properties
(i) H+HcH,
(ii) AH c H for all A > 0,
(iii) H n -H = {OJ.
It is verified without difficulty that each proper cone H c L defines, by
virtue of " x ~. y if y - x E H", an order of L under which L is an ordered
vector space with positive cone H. Hence for any vector space L, there is a
biunivocal correspondence between the family of all proper cones in Land
the family of all orderings satisfying (LO)1 and (LOh. If R1 and R2 are two
such orderings of L with respective positive cones C1 and C2 , then the
relation" R1 is finer than R 2" is equivalent with C1 c C 2 ; in particular, if
{R,,: a E A} is a family of such orderings of L with respective positive cones
C", the coarsest ordering R which is finer than all R,,(rx E A) is determined by
the proper cone e = n"c". (Cf. Exercise 2.) A cone He L satisfying (i) and
(ii) is said to be generating if L = H - H.
Let L be an ordered vector space. The order of L is called Archimedean
(or L Archimedean ordered) if x ~ 0 whenever there exists YEL such that
nx ~ y for all n E N (in other words, if x ~ 0 whenever {nx: n E N} is major-
ized). For example, if L is a t.v.s. and an ordered vector space whose positive
cone is closed, L is Archimedean ordered; on the other hand, R8 is not
Archimedean ordered for n ;;;; 2 under its lexicographic ordering (see below).
An order interval in L is a subset of the form {z E L: x ~ z ~ y}, where x, y
are given; it is convenient to denote this set by [x, y]. (There is little danger of
confusing this with the inner product notation in pre-Hilbert spaces (Chapter
III, Section 2, Example 5) if we avoid using the symbol in different meanings
in the same context.) A subset A of L is order bounded if A is contained in
some order interval. Every order interval is convex, and every order interval
of the form [-x, x] is circled. An element e EL such that [-e, e] is radial is
called an order unit of L. The set Lb of all linear forms on L that are bounded
on each order interval is a subspace of L *, called the order bound dual of L.
Let L be an ordered vector space over R and let M be a subspace of L.
If C is the positive cone of L, then the induced ordering on M is determined

by the proper cone C n M; an ordering of LjM is determined by the canonical
image C of C in LjM, provided that C is a proper cone. (Simple examples,
with L = Ri" show that this is not necessarily the case.) If {La: IJ( E A} is a
family of ordered vector spaces with respective positive cones Ca , then
C = TIaCa is a proper cone in L = TIaLa which determines an ordering of L.
The orderings so defined are called the canonical orderings of M, Lj M (pro-
vided C is proper), and of TIaLa. In particular, the algebraic direct sum EBaLa
is canonically ordered as a subspace of TIaLa, and if T is any set, then L T
is canonically ordered by the proper cone {f:f(t) E C for all t E T}.
Let L be an ordered vector space which is the algebraic direct sum of the
subspaces M i(i = 1, ... , n); L is said to be the ordered direct sum of the sub-
spaces Mi if the canonical algebraic isomorphism of L onto TIiMi is an order
isomorphism (for the canonical ordering of TIiMi).
If L 1 , L2 are ordered vector spaces i= {O} with respective positive cones C1 and
C 2 , then C = {u: u(Cl ) c: C 2 } is a proper cone in the space L(L l , L 2 ) oflinear
mappings of Ll into L 2 , if and only if C1 is generating in L l ; whenever M is a
subspace of L(L1 , L 2 ) such that C n M is a proper cone, the ordering defined
by C n M is called the canonical ordering of M. A special case of importance
is the following: A linear formf on an ordered vector space over R is positive
if x ;?; 0 implies f(x) ;?; 0; the set C* of all positive linear forms on L is a
cone which is the polar, with respect to <L, L *), of - C. The subspace
L+ = C* - C*ofL*iscalledtheorderdualofL;itisimmediatethatL+ c:Lb.
However, there exist ordered vector spaces L for which L + i= Lb (see Namioka
[1], 6.10).
In order to use the tool of duality successfully in the study of ordered vector
spaces L, one needs sufficiently many positive linear forms on L to distin-
guish points; we shall say that L is regularly ordered (or that the order of L is
regular) if L is Archimedean ordered and L + distinguishes points in L (cf.
(4.1) below).
As above, the canonical ordering of a subspace M c: L * is understood to be
the ordering defined by M n C* whenever M n C* is a proper cone in M.
Let us note some simple consequences of (LO)l, L being an ordered vector
space. The equality
z + sup(x, y) = sup(z + x, z + y) (1)
is valid for given x, y ELand all Z E L whenever sup(zo + x, Zo + y) exists

for some Zo E L. If A, B are subsets i= 0 of L such that sup A and sup B
exist, then sup(A + B) exists and
sup(A + B) = sup A + sup B. (1 ')
Also from (LO)l it follows that

sup(x,y) = -inf(-x, -y) (2)
whenever either sup(x, y) or inf( - x, - y) exists; more generally,
sup A = -inf( -A) (2')
whenever either sup A or inf( - A) exists.

A vector lattice is defined to be an ordered vector space E over R such that
for each pair (x, y) E Ex E, sup(x, y) and inf(x, y) exist. This implies, in
particular, that E is directed under the order relation ~ (equivalently, that
the positive cone C of E is generating). For each x E E, we define the absolute
Ixl by Ixl = sup(x, -x); two elements x, y of a vector lattice E are disjoint
if inf(lxl, Iyl) = 0; two subsets A c: E and B c: E are lattice disjoint (or
simply disjoint if no confusion is likely to result) if x E A, Y E B implies
inf(lxl, Iyi) = 0. The fact that x, yare disjoint is denoted by x 1.. y, and if A
is a subset of E, A.L denotes the set of all y E E such that y is disjoint from each
element of A. We record the following simple but important facts on vector
lattices.
1.1
Let E be a vector lattice. Then
x +y = sup(x, y) + inf(x, y) (3)
is an identity on E x E. Defining x+ and x'" by x+ = sup(x, 0) and x- = sup

(-x, O)for all x E E, we have x = x+ - x- and Ixl = x+ + x-; x = x+ - x-
is the unique representation of x as a difference of disjoint elements ~ 0. More-
over, we have
IAxl = IAllxl (4)
Ix + yl ~ Ixl + Iyl (5)
Ix+ - y+1 ~ Ix - yl (6)

for all x, y E E and A E R. Finally, we have
[0, x] + [0, y] = [0, x + y] (D)
for all x ~ ° and y ~ 0.
Proof. To prove (3), consider the more general identity
a -inf(x, y) + b = sup(a - x + b, a - y + b), (3')
where a, b, x, yare arbitrary elements of E. By (2) we have -inf(x, y)
a °
= sup( -x, - y), whence (3') follows from (1); from (3') we obtain (3) by the
substitution = x, b = y. Letting y = in (3), we obtain x = x+ - x-, and
since inf(x+, x-) = x- + inf(x, 0) = x- - sup( -x, 0) = 0, x+ and x- are
disjoint elements; we now obtain via (1), x+ + x- = x + sup( -2x, 0)
= sup( -x, x) = 14 Let x = y - Z, where y ~ 0, Z ~ are disjoint; we show °
that y = x+, Z = x-, Note first that x = y - z implies y ~ x hence y ~ x+

and, therefore, z ~ x-; it follows that (y - x+) 1.. (z - x-) which, in view of
y - x+ = z - x-, implies y = x+, Z = x-, since clearly 0 is the only element
of E disjoint from itself.
If A ~ 0, then from (LOh we obtain (Ax)+ = Ax+ and (Ax)- = AX-; if
A < 0, then (Ax)+ = (-A( -x»+ = IAlx- and (Ax)- = IAlx+; this proves (4).
For (5), note that ±x ~ lxi, ±y ~ IYI implies Ix + yl = sup(x + y, -x - y)
~ Ixl + Iyl. To prove (6) we conclude from x = y + (x - y) that x ~ y+
+ Ix - yl; hence, the right-hand side being ~ 0, that x+ ~ y+ + Ix - yl ;
therefore, x+ - y+ ~ Ix - yl and interchanging x and y yields y+ - x+
~ Ix- yl, hence (6).
Finally, it is clear that [0, x) + [0, y) c [0, x + y) whenever x ~ 0 and
y ~ 0. Let z e [0, x + y) and define u, v by u = inf(z, x) and v = z - u; there
remains to show that ve [0, y). But v = z - inf(z, x) = z + sup( -z, -x)
= sup(O, z - x) ~ sup(O, x + y - x) = y which completes the proof of (1.1).
COROLLARY 1. In every vector lattice E, the relation x ~ y is equivalent with
"x+ ~ y+ and y- ~ x- ", and the relation x 1.. y is equivalent with sup(lxl, Iyi)
= Ixl + Iyl. Moreover, if x 1.. y, then (x + y)+ = x+ + y+ and Ix + yl = Ixl
+Iyl·
Proof. In fact, if x+ ~ y+ and y- ~ x-, then x = x+ - x- ~ y+ - y- = y.
Conversely, x ~ y implies x+ ~ y+ and inf(x, 0) ~ inf(y,O); hence -x-
~ -y- or, equivalently, y- ~ x-. The second assertion is immediate from
(3) replacing x, y by lxi, Iyl respectively. Finally, x + y = (x+ + y+) - (x-
+ y-), and inf(lxl, Iyi) = 0 expresses that the summands on the right are
disjoint; hence (x + y)+ = x+ + y+ by the unicity of the representation of
x + y as a difference of disjoint elements ~ 0. The last assertion is now
immediate.
COROLLARY 2. Let E be a vector lattice and let AcE be a subset for which
sup A = Xo exists. If BeE is a subset lattice disjoint from A, then B is lattice
disjoint from {xo}.
Proof. We have to show that z e B implies z 1.. Xo. Now xi) ~ x- ~ Ixl for
all x e A; hence z 1.. xi) if z e B. It suffices hence to .show that z 1.. xci. In
view of Corollary 1, we have sup(lzl, x+) = Izl + x+ for all x e A by hypo-
thesis, and xci =sup{x+:xeA}; (1') implies that sup{lzl +x+: xeA} =Izl
+ xci. Thus we obtain
sup(lzl, xci) = sup sup(lzl, x+) = sup(lzl + x+) = Izl + xci,

xeA xeA
which shows that Izl1.. xci (Corollary 1).

The following observation sometimes simplifies the proof that a given
ordered vector space is a vector lattice.
1.2
Let E be an ordered vector space over R whose positive cone C is generating;
if for each pair (x, y)E C X C either sup(x, y) or inf(x, y) exists, then E is
a vector lattice.
The detailed verification is left to the reader; one shows that if sup(x, y)
exists (x, y E C), then z = x + y - sup(x, y) proves to be inf(x, y), and con-
versely. If x, yare any elements of E, there exists Z E C such that x + Z E C
and y + Z E C, and the existence of sup(x, y) and inf(x, y) is shown via (1).
If {E«: IX E A} is a family of vector lattices, it is quickly verified that n~«
and $«E« are vector lattices under their canonical orderings. A vector
sublattice M of a vector lattice E is a vector subspace of E such that x E M,
y E M implies that sup(x, y) E M where the supremum is formed in E; it
follows that M is a vector lattice under its canonical ordering. However, it
can happen that a subspace M of E is a vector lattice under its canonical
order but not a sublattice of E (Exercise 14).
A subset A of a vector lattice E is called solid if x E A and Iyl ~ lxi, YEE,
imply that YEA. It is easy to see that a solid subspace of E is necessarily a
sublattice of E; for example, the algebraic direct sum ~«EI% of a family
{E,,: IX E A} of vector lattices is a solid subspace of n«E« (for the canonical
ordering of the product). Also it is easy to see that if M is a solid subspace
of E, then E/ M is a vector lattice under its canonical order (cf. the exam-
ples below).
A subset A of a vector lattice E is called order complete if for each non-
empty subset Be A such that B is order bounded in A, sup Band inf B
exist and are elements of A; E is order complete if it is order complete as a
subset of itself. If E is an order complete vector lattice, a subspace M of E
which is solid and such that A c M, sup A = X E E implies x E M, is called a
band in E. E itself is a band, and clearly the intersection of an arbitrary family
of bands in E is a band; hence every subset A of E is contained in a smallest
band B A, called the band generated by A (in E).
Examples
1. Let T be any set and consider the vector space R'{; of all real-
valued functions on T under its canonical order, where Ro is ordered
as usual. Obviously R'b is an order complete vector lattice. If A is any
subset of R'{;, denote by TA the subset {t: there exists f E A such that
f(t) # O} of T. Then the band generated by A is the subspace BA =
{f:f(t) = 0 whenever t ¢ TA }; the quotient Rb/BA , under its canonical
order, is a vector lattice which is isomorphic with R"{;-TA. The canonical
ordering of R'{; is regular (in particular, Archimedean); in fact, the order
dual and the order bound dual coincide with the (ordered) direct
sum of card T copies of Ro (Chapter IV, Section 1, Example 4).
2. Let {J be any ordinal number > 0 and let Rg denote the vector
space of all real valued functions defined on the set of all ordinals
a < /3, and consider the subset H of Rg defined by the property "if
there exists a smallest ordinal a < /3 such that f(a) =1= 0 then f(a) > 0".
We verify without difficulty that H is a proper cone in Rg; the order de-
termined by H is called the lexicographical order of Rg. The lexicograph-
ical order of Rg is not Archimedean (hence not regular) if /3 > I; in fact,
the set of all functionsf such thatf(O) = 0 is majorized by each function
ffor whichf(O) > O. It is worth noting that the lexicographical order of
Rg is a total ordering, since Rg = H u - H; thus Rg is a vector lattice
under this order which is, however, not order complete if /3 > 1. More-
over, (up to a positive scalar factor) f ~ f(O) is the only non-trivial
positive linear form, hence the order dual and the order bound dual
(cf. (1.4) below) are of dimension 1.
3. Let (X, ~, Jl) be a measure space (Chapter II, Section 2, Example 2).
Under the ordering induced by the canonical ordering of R3 (Example I
above), the spaces ffP(Jl) (1 ;:;; p ;:;; + (0) are vector lattices (take the
scalar field K = R) which are count ably order complete (each majorized
countable family has a supremum) but, in general, not order complete
(Exercise 13). The subspace .AtIt of Jl-null functions is a solid subspace
but, in general, not a band in !l'P(Jl); the quotient spaces U(Jl) =
!l'P(Jl)/.At It are order complete vector lattices under their respective
canonical orderings (1 ~ p < + (0).
If E is any order complete vector lattice and A a subset of E, the set A.l. is
a band in E; this is clear in view of Corollary 2 of (1.1). Concerning the
bands B A and A -\ we have the following important theorem (F. Riesz [I D.
1.3
Theorem. Let E be an order complete vector lattice. For any subset
AcE, E is the ordered direct sum of the band B A generated by A and of the
band A.l. of all elements disjoint from A.
Proof Since AJ..J.. is a band containing A, it follows that B A c AJ..J.. and
hence that BA ("') A.l. = {O}. Let x E E, x;:;; 0, be given; we show that x = Xl
+ X2' where Xl E B A , X2 E A.l., and Xl;:;; 0, X 2 ;:;; O. Define Xl by Xl = sup
[0, xl ("') B A and X 2 by X 2 = X - Xl; it is clear that Xl' X2 are positive and
that Xl E B A, since BA is a band in E. Let us show that X 2 E B~. For any
y E B A let z = inf(x2' Iyl); then 0;:;; Z E B A, since B A is solid and z + Xl
;:;; X 2 + Xl = x. This implies, by the definition of Xl and by virtue of z + Xl
E B A, that z + Xl ;:;; Xl and hence that z = O. Thus X 2 E B~ and a fortiori
X 2 E A.l.. Since the positive cone of E is generating, it follows that E = B A
+ A.l. is the ordered direct sum of the subspaces BA and A.l.. For, the relations
x;:;; 0 and X = Xl + X 2 , Xl E B A , X 2 E A.l. imply Xl;:;; 0, X 2 ;:;; O.
COROLLARY 1. If A is any subset of E, the band B A generated by A is the band
AJ..J...
Proof Applying (1.3) to the subset A.l. of E, we obtain the direct sum
E = AJ..J.. + A.l.; since E = BA + A.l. and BA c AJ..J.., it follows that BA = A.l..l..
COROLLARY 2. If x, yare disjoint elements of E and B", By are the bands

generated by {x}, {y} respectively, then Bx is disjoint from By.
In fact, we have y E {x}.l and x E {y}.l.
A general example of an order complete vector lattice is furnished by the
order dual E+ of any vector lattice E; however, E+ can be finite dimensional
(Example 2 above) or reduced to {O} (Exercise 14), even if E is of infinite
dimension. We prove the result in the following more .general form which
shows it to depend essentially on property (D) of (Ll). (Cf. Exercise 16.)
1.4
Let E be an ordered vector space over R whose positive cone C is generating
and has property (D) of (Ll). Then the order bound dual Eb of E is an order
complete vector lattice under its canonical ordering; in particular, Eb = E+.
Proof We show first that for each f E E b, sup(j, 0) exists; it follows then
from (1) that sup(f, g) = g + sup(f - g, 0) exists for any pair (J, g) E Eb X Eb;
hence Eb is a vector lattice by 0.2). This implies clearly that Eb = E+.
Let f E Eb be given; we define a mapping r of C into the real numbers ~ 0
by
rex) = sup{f(y): y E [0, x]} (x E C).
Since f(O) = 0 it follows that r(x) ~ 0, and clearly rCA-x) = A-r(x) for all
A- ~ O. Also, by virtue of (1 ') and (D),
rex + y) = sup{f(z): z E [0, xl + [0, y]} = rex) + r(y).
Hence r is positive homogeneous and additive on C. By hypothesis, each
z E E is of the form z = x - y for suitable elements x, y E C, and it is readily
seen that the number r(x) - r(y) is independent of the particular decomposi-
tion z = x - y of z. A short computation now shows that z --+ w(z) = r(x)
- r(y) is a linear form w on E, evidently contained in Eb. (We have, in fact,
w(x) = r(x) for x E C.) We show that w = sup(j, 0); indeed, w(x) ~ sup
(f(x), 0) for all x E C, and if h ~ 0 is a linear form on E such that x E C
implies hex) ~f(x), then hex) ~ hey) ~f(Y) for all y E [0, xl, which shows
that hex) ~ rex) = w(x) whenever x E C.
It remains to prove that Eb = E+ is order complete; for this it suffices to
show that each non-empty, majorized set A of positive linear forms on E
has a supremum. Without restriction of generality, we can assume that A is
directed under " ~ ". (This can be arranged, if necessary, by considering the
set of suprema of arbitrary, non-empty finite subsets of A.) We define a map-
ping s of C into the real numbers by
sex) = sup{f(X):fE A} (x E C).
The supremum is finite for all x E C, since A is majorized. It is clear that
S(Ax)= As(X) for all A- ~ 0 and, since A is directed, that sex + y) = sex) + s(y).
Hence, as before, s defines a linear form fo on E by means of fo(z) = sex)
- s(y), where z = x - y and x, y e C. It is evident that fo e Eb (since fo ~ 0)

and thatfo = sup A.
COROLLARY. The order dual of every vector lattice is an order complete
vector lattice under its canonical ordering.
From the construction of f+ = sup(f, 0) in the proof of (1.4), we obtain
the following useful relations; the proof of these is purely computational and
will be omitted.
1.5
Let E be a vector lattice and let f, 9 be order bounded linear forms on E. For
each x e E, we have
sup(f, g)(lxi) = sup{f(y) + g(z): y ~ 0, z ~ 0, y + z = Ixl}
(7)
inf(f, g)(lxi) = inf{f(y) + g(z): y ~ 0, z ~ 0, y + z = Ixl}
Ifl(lxi) = sup{f(y - z): y ~ 0, z ~ 0, y + z = Ixl}
(8)
If(x) I ~ IfIClxi).
° °
In particular, two linear forms f~ 0, 9 ~ are disjoint if and only if for each
x ~ and each real number B > 0, there exists a decomposition x = Xl + X2
with Xl ~ 0, x 2 ~ 0, and such that f(x l ) + g(x2 ) ~ B.
COROLLARY. Let E be a vector lattice, and let <E, G) be a duality such that G
is a sublattice ofE+. Then the polar AO c G of each solid subset AcE is solid.
Proof In fact, if x E A, Y ~ 0, z ~ 0, and y + z = lxi, then y - z e A,
since -Ixl ~ y - z ~ Ixl; hence, if fe AO and Igi ~ If I, then from (8) it
follows that
Ig(x)1 ~ Igl(lxi) ~ Ifl(lxi) ~ 1,
which shows that 9 e AO.
If E is an ordered vector space over R such that the order dual E+ is an
ordered vector space (equivalently, if C* is a proper cone in E* where C is the
positive cone of E), then the space (E+)+ is called the order bidual of E and
denoted by E+ +. Under the assumptions of (1.4) (in particular, if E is a
vector lattice), E+ + is a vector lattice, and the evaluation (or canonical) map
of E into E+ +, defined by x -+ x where x(f) = f(x) (f E E+), is clearly order
preserving. Assuming that E is a vector lattice, let us show that x -+ x is an
isomorphism onto a sublattice of E+ + if E is regularly ordered (equivalently,
if x -+ x is one-to-one). For later use, we prove this result in a somewhat
more general form.
1.6
Let E be a vector lattice and let G be a solid subspace of E+ that separates
points in E; the evaluation map x -+ x, defined by x(f) = f(x) (f E G), is an
isomorphism of E onto a sublattice of G+ •
Proof. We must show that for each x e E, the element x+ (= sup(O, x)

taken in G+) is the canonical image of x+ e E. Denote by P the subset
U{p[O, x+]: p $1; O} of E and define, for eachf$1; 0 in G, a mapping t, of the
positive cone C of E into R by
t,(Y) = sup{f(z): z e [0, y] ("'\ P} (ye C).
As in the proof of (1.4) it follows that t, is additive and positive homo-
geneous, and hence defines a unique linear form g, e C*; it is clear that
g, ~J, hence g, e G, since G is solid, and that gix-) = 0 because of [0, x-]
("'\ P = {O}. Hencegix) = g,(x+), and we obtain x +(f) = sup{g(x):O ~ 9 ~f}
$1; g,(x) = gix+) = f(x+) for all fe C* ("'\ G. This implies x+ $1; (x+)-;
since it is clear that (x+)- $1; x+ in G+, the assertion follows.
We point out that the canonical image of E in G+ is, in general, not an
order complete sublattice of G+ even if E is order complete (see the example
following (7.4». In particular (taking G = E+), a regularly ordered, order
complete vector lattice E need not be mapped onto a band in E+ + under
evaluation. If E is an order complete, regularly ordered vector lattice whose
canonical image in E+ + is order complete, E will be called minimal (or of
minimal type).
If E, F are vector lattices, a linear map u of E onto F is called a lattice
homomorphism provided that u preserves the lattice operations; in view of the
linearity of u, the translation-invariance of the order and the identity (3), this
condition on u is equivalent to each of the following: (i) u(sup(x, y»
= sup(u(x), u(y» (x, y E E). (ii) u(inf(x, y» = inf(u(x), u(y» (x, y E E). (iii)
u(lxi) = sup(u(x+), u(x-» (x e E). (iv) inf(u(x+), u(x-» = 0 (x e E). If, in
addition, u is biunivocal, then u is called a lattice isomorphism of E onto F.
It is not difficult to show that a linear map u of E onto F is a lattice homo-
morphism if and only if u- 1(0) is a solid sublattice of E and u(C1) = C2 ,
where C1, C 2 denote the respective positive cones of E, F. In particular, if N
is a solid vector sublattice of E, then EjN is a vector lattice under its canonical
order and the canonical map ljJ is a lattice homomorphism of E onto EjN
(Exercise 12).
The linear forms on a vector lattice E that are lattice homomorphisms onto
R have an interesting geometric characterization; let us recall (Chapter II,
Exercise 30) that {A.x: A. $1; O}, 0 =F x e C is called an extreme ray of the cone C
if x - y E C, Y E C imply y = px for some p, 0 ~ p ~ 1..
1.7
Let E be a vector lattice,f =F 0 a linear form on E. The following assertions are
equivalent:
(a) fis a lattice homomorphism olE onto R.
(b) inf(f(x+),J(x-» = Ofor all x E E.
(c) fgenerates an extreme ray of the cone C* in E*.
(d) f$1; 0 andf- 1(0) is a solid hyperplane in E.
Proof (a)¢>(b) is clear from the preceding remarks. (b)=>(d): Since

inf(f(x+),f(x-» = 0 for each x E E, it follows that f ~ 0, and f(x) = 0
implies f(lxJ) = 0; hence Iyl ~ Ixl and f(x) = 0 imply If(y) I ~f(lyl) ~f(lxl)
= O. (d) => (c): Suppose 9 E C* is such that f - 9 E C* or, equivalently, that
o ~ 9 ~f Then since f-l(O) is solid, f(x) = 0 implies Ig(x)1 ~ g(lxJ) ~f(lxJ)
= 0 and hence f- 1(0) c 9 -1(0). Thus (since f- 1(0) is a hyperplane) either
9 = 0 orf-l(O) = g-I(O); in any case, 9 = pffor some p, 0 ~ p ~ 1. (c) => (b):
Let f generate an extreme ray of C*, let x E E be given, and suppose that
f(x+) > O. Let P = U{p[O, x+]: p ~ O}, and define hE E* by putting, for
y ~ 0, hey) = sup{f(z): z E [0, y] n P} (see proof of (1.6». It follows that
o ~ h ~f and hence h = pf by the assumption made on f, and since h(x+)
= f(x+) > 0 we must have p = 1. Thus h = f, and since clearly h(x-) = 0, it
follows that f(x-) = 0, which completes the proof.
2. ORDERED VECTOR SPACES OVER THE COMPLEX FIELD

It is often useful to have the concept of an ordered vector space over the
complex field C. Such is the case, for instance, in spectral theory and in
measure theory. It is the purpose of this section to agree on a definite ter-
minology. We define a vector space Lover C to be ordered if its underlying
real space Lo (Chapter I, Section 7) is an ordered vector space over R; thus
by definition, order properties of L are order properties of Lo. The usefulness
of this (otherwise trivial) definition lies in the fact that the transition to Lo
does not have to be mentioned continually.
The canonical orderings of products, subspaces, direct sums, quotients,
function spaces, and spaces of linear maps are then defined with reference to
the respective underlying real spaces; only the term "positive linear form"
on L has to be additionally specified when L is an ordered vector space over
C. We define f E L * to be positive if Re f(x) ~ 0 whenever x ~ 0 in L; this
definition guarantees that whenever the canonical ordering of (Lo) * is defined,
then L* is ordered, and the canonical isomorphism of (I, 7.2) is an order
isomorphism (a corresponding statement holding for subspaces of L*).
The order bound dual Lb of an ordered vector space Lover C is then defined
as the subspace of L * containing exactly the linear forms bounded on each
order interval in L; the order dual L + is the (complex) subspace of L * which is
the linear hull of the cone C* of positive linear forms. In accordance with the
definition given above, the order of L is called regular if Lo is regularly
ordered; we point out that this is not implied by the fact that C* separates
points in L, and that in general (L +)0 cannot be identified with (Lo) + by virtue
of (I, 7.2) (Exercise 4).
The term vector lattice will not be extended to complex spaces; we shall,
however, say that an ordered vector space Lover C with positive cone C
is lattice ordered if the real subspace C - C of L is a vectorlattice. For example,
the complexification (Chapter I, Section 7) of a vector lattice L is a lattice
ordered vector space Ll over C.
§3l DUALITY OF CONVEX CONES 215
3. DUALITY OF CONVEX CONES

Let L be a vector space (over R or C); by a cone in L we shall henceforth
understand a convex cone C of vertex 0 and such that 0 E C. Let C be a fixed
cone in L; for any pair (x, y) E L x L, we shall write [x, y] = (x + C) n
(y - C). This notation is consistent with the notation introduced for order
intervals in Section 1; if C is the positive cone of an ordering of L, then
(x + C) n (y - C) is the order interval {z: x ~ z ~ y}. For any subset
A cL, define
[A] = (A + C) n (A - C) = UUx, y]: x E A, YEA}.
A subset BeL is called C-saturated if B = [B]; it is immediate that for any
A c L, [A] is the intersection of all C-saturated subsets containing A, and
hence called the C-saturated hull of A. It is also quickly verified that A ~ [A]
is monotone: A c B implies [A] c [B], that [A] is convex if A is convex, and
that [A] is circled with respect to R if A is circled with respect to R. Finally
we note that if ty is a filter (more generally, a filter base) in L, then the family
UF]: FE ty} is a filter base in L; the corresponding filter will be denoted by
[ty].
Assume now that L is a t.v.s. A cone C in L is said to be normal if U = [U]
where U is the neighborhood filter of O. Hence C is a normal cone in the
t.v.s. L if and only if there exists a base of C-saturated neighborhoods of 0
(equivalently, if and only if the family of all C-saturated O-neighborhoods is
a base at 0). It will be useful to have a number of alternative characteriza-
tions of normal cones.
3.1
Let L be a t.v.s. over K and let C be a cone in L. Thefollowing propositions are
equivalent:
(a) C is a normal cone.
(b) For every filter ty in L, lim ty = 0 implies lim[tyl = O.
(c) There exists a O-neighborhood base m in L such that V E m implies
[Vn C] c V.
If K = R and the topology ofL is locally convex, then (a) is equivalent to each
of the following:
(d) There exists a O-neighborhood base consisting of convex, circled, and C-
saturated sets.
(e) There exists a generating family &' of semi-norms on L such that p(x)
~p(x + y) whenever x E C, Y E C and p E fIJ.
Proof Denote by U the neighborhood filter of 0 in L. (a) => (b): If ty is a

filter on L which is finer than U, then [ty] is finer than [U]; hence the assertion
follows from U = [U]. (b) => (c): (b) implies that [U] is the neighborhood
filter of 0 in L; hence m= {[U]: U E U} is a neighborhood base of 0 such

that V Em implies [V n C] c [V] = V. (c) => (a): Given U E U, it suffices to
show there exists WE U such that [W] c U. Let m be a O-neighborhood base
as described in (c); select V Em such that V + V c U and a circled WE U
such that W + We V. We obtain
[W] = U [x, y] = U (x + [0, y - x]) c W + [( W + W) n C]
x,yeW x,yeW
c V + [V n C] c V + V c U,
which proves the implication (c) => (a).
Assume now that K = R and the topology of L is locally convex. (a) => (d):
If U1 is the family of all convex, circled O-neighborhoods in L, then !ill
= {[U]: U E U1} is a base at 0 consisting of convex, circled, and C-saturated
sets. (d)=>(e): If !ill is a O-neighborhood base as in (d) andpw is the gauge
function of WE m3, the family {Pw: WE !ill} is of the desired type. (e) => (c):
If f!lJ is as in (e) then the family of all finite intersections of the sets Vp ,.
= {x E L: p(x) ~ B} (p E f!lJ, B > 0) is a neighborhood base mof 0 having the
property stated in (c). This completes the proof.
COROLLARY I. IfL is a Hausdorff t.v.s., every normal cone C in L is a proper
cone.
Proof. In fact, if x E C n - C, then x E [{O}] c [U] for each O-neighbor-
hood U, and it follows that x = O.
COROLLARY 2. If C is a normal cone in Land BeL is bounded, then [B] is
bounded; in particular, each set [x, y] is bounded.
Proof. If B is bounded and U is a O-neighborhood in L, there exists A > 0
such that BeAU; it follows that [B] c [AU] = l[U].
COROLLARY 3. If the topology ofL is locally convex, the closure C of a normal
cone is a normal cone.
Proof. It is immediate that C is a cone in L, and C is also the closure of C
in the real space Lo; the assertion follows now from proposition (e) of (3.1).
It will become evident from the results in this chapter and the Appendix
that the concept of a normal cone is an important (and perhaps the most
important) notion in the theory of ordered topological vector spaces; for
cones in normed spaces over R it goes back to M.G. Krein [2]. The original
definition of Krein postulates the existence of a constant y( ~ 1) such that
Ilxll ~ yllx + yll for all x, y E C; it follows at once that this definition is
equivalent, for normed spaces (L, II II) over R, with the one given above, and.
(3.1) (e) implies that there exists an equivalent norm on L for which one can
suppose y = 1.
If M is a subspace of the t.v.s. Land C is a normal cone in L, it is clear that
M n C is a normal cone in M; it is also easy to verify that if {LIZ: ex E A} is a
§3] DUALITY OF CONVEX CONES 217
family of t.v.s., C,. a cone in LIZ, and L = TIIZLIZ' then C = TIIZCIZ is a normal
cone in L if and only if CIZ is normal in LiIX E A). Let us record the following
result on locally convex direct sums.
3.2
If {LIZ: IX E A} is afamily of l.c.s., CIZ a cone in LiIX E A), and L = $IZLIZ the
locally convex direct sum of this family, then C = EDIZCIZ is a normal cone in L
if and only if CIZ is normal in LIZ (IX E A).
Proof: The necessity of the condition is immediate, since each LIZ can be
identified with a subspace of L such that CIZ is identified with LIZ () C (IX E A).
To prove that the condition is sufficient assume that K = R (which can be
arranged, if necessary, by transition to the underlying real space Lo of L).
Let m,. be a neighborhood base of 0 in LIZ (IX E A) satisfying (3.1) (d); the
r
family of all sets V = IZ VIZ (VIZ E mlZ , IX E A) is a neighborhood base of 0 in L
(Chapter II, Section 6). Now it is clear that [V () C] is the convex hull of
UIZ[V1Z () CIZ ]; since [V,. () CIZ ] C VIZ for all VIZ E mlZ (IX E A), it follows that
[V () C] c: V, which proves the assertion in view of (3.1) (c).
It can be shown in a similar fashion that a corresponding result holds for
the direct sum topology introduced in Exercise I, Chapter I (in this case, the
spaces L,. need not be supposed to be locally convex). On the other hand, if
C is a normal cone in Land M is a subspace of L, then the canonical image
C of C in LIM is, in general, not a proper cone, let alone normal. (For a
condition under which C is normal, see Exercise 3.)
Intuitively speaking, normality of a cone C in a t. v.s. L restricts the" width"
of C and hence, in a certain sense, is a gauge of the pointedness of C. For ex-
ample, a normal cone in a Hausdorff space cannot contain a straight line
«3.1), Corollary 1); a cone C in a finite-dimensional Hausdorff space L is
normal if its closure C is proper (cf. (4.1) below). In dealing with dual
pairs of cones, one also needs a tool working in the opposite direction and
gauging, in an analogous sense, the bluntness of C. The requirement that
L = C - C goes in this direction; in fact, it indicates that every finite subset
S of L can be recovered from C in the sense that S c: So - So for a suitable
finite subset So c: C. The precise definition of the property we have in mind
is as follows.
Let L be a t.v.s., let C be a cone in L, and let 6 be a family of bounded
subsets of L (Chapter III, Section 3); for each S E 6, define Sc to be the sub-
set S () C - S () C of L. We say that C is an 6-cone if the family {Sc: S E 6}
is a fundamental subfamily of 6; C is called a strict 6-cone if {Sc: S E 6} is
fundamental for 6. If L is a l.c.s. over Rand 6 is a saturated family, in
place of Sc we can use the convex, circled hull of S () C in the preceding
definitions. A case of particular importance is the case where 6 = ~ is the
family of all bounded subsets of L: C is a ~-cone in L. The notion of a
~-cone in a normed space (L, " II) appears to have been first used by Bonsall
[2]; Bonsall defines L to have the decomposition property if each z, liz II ~ 1,

can be approximated with given accuracy by differences x - y, where x E C,
Y E C and Ilxll ~ k, lIyll ~ k for a fixed constant k > o.
The property of being an 6-cone satisfies certain relations of permanence
(Exercise 5); since these are consequences of (3.3), below, the permanence
properties of normal cones, and the duality theorems (IV, 4.1) and (IV, 4.3),
they will be omitted here. Let us point out that, as the concept of a normal
cone, the concept of an 6-cone is independent of the scalar field (R or C)
over which L is defined.
Examples
1. The set of real-valued, non-negative functions determines a
normal cone in each of the Banach spaces enumerated in Chapter II,
Examples 1-3. If E, is anyone of these spaces and C the corresponding
cone, then E = C - C if K = R; this implies that C is a strict ~-cone in
E (see (3.5) below). If the functions (or classes of functions) that con-
stitute E are complex valued, then C and C + iC are normal cones and
C + iC is a strict ~-cone.
2. Let C denote the set of all non-negative functions in the space
fi) of L. Schwartz (Chapter II, Section 6, Example 2). C is not a normal
cone in fi), but C + iC is a strict ~-cone. The cone C1 of all distributions
T such that (Tf) ~ 0 for f E C (which can be identified with the set of
all positive Radon measures on Rn (cf. L. Schwartz [1))) is a normal
cone in fi)', but C 1 + iC1 is not a ~-cone (Exercise 6).
3. Let E be the space of complex-valued, continuous functions with
compact support on a locally compact space X with its usual topology
(Chapter II, Section 6, Example 3), and let C be the cone of non-
negative functions in E. C is a normal cone in E, C + iC is normal and
a strict ~-cone. If C1 denotes the set of all positive Radon measures
on X, C 1 + iC1 is normal and a strict ~-cone in the strong dual E'.
The proofs for these assertions will become clear from the following
results and are therefore omitted.
If C is a cone in the t.v.s. E, the dual cone C' of C is defined to be the set
{f E E': Re f(x) ~ 0 if x E C}; hence C' is the polar of - C with respect to
(E, E'). In the following proofs it will often be assumed that the scalar field
K of E is R; whenever this is done, implicit reference is made to (I, 7.2) (cf.
also Section 2). Before proving the principal result of this section, we estab-
lish this lemma which is due to M. G. Krein [2].
LEMMA 1. If C is a normal cone in the normed space E, then E' = C' - C'.
Proof. We can assume that K = R. LetfE E' and define the real function
p ~ 0 on C by p(x) = sup{f(z): z E [0, xl}. Then it is clear that p(AX) = Ap(xn
if A ~ 0 and that p(x + y) ~ p(x) + p(y), since [0, x] + [0, y] c [0, X + y] for)
all x, y E C. It follows that the set
V = {(t, x): 0 ~ t ~p(x)}
is a cone in the product space Ro x E. Let {x,,: n E N} be a null sequence in

E and suppose that {tIl: n E N} is a sequence of real numbers such that
(t", x,,) E V (n EN). Since C is a normal cone and f is continuous, it follows
that p(x,,) -. 0 and hence that tIl -.0; this implies that (1, 0) is not in the
closure Vof V in the normable space Ro x E. By (II, 9.2) there exists a
closed hyperplane H strictly separating {(I,O)} and V; it can be arranged
that H = {(t, x): h(t, x) = -I}, where h(1, 0) = -1 and h is ;?;; 0 on V. By
(IV, 4.3) h is of the form (t, x) -. - t + g(x); since gEE' and (0, x) E V for
each x E C, it follows that 9 E C'. Now (P(x), x) E V for all x E C; hence we
have - p(x) + g(x) ;?;; 0 if x E C. Since f(x) ~ p(x) ~ g(x) for x E C, we
obtain f = 9 - (g - f), where 9 E C', 9 - f E C', and the lemma is proved.
3.3
Theorem. Let E be a l.c.s., let C be a cone in E with dual cone C' c E',
and let 6 be a saturatedfamity ofweakly bounded subsets ofE'. !fC' is an 6-cone,
then C is normal for the 6-topology on E; conversely, if C is normal for an
6-topology consistent with (E, E'), then C' is a strict 6-cone in E'.
Proof. We can assume that K = R. If C' is an 6-cone in E', then the
saturated hull of the family {reS n C'): S e 6} equals 6; hence the 6-
topology is generated by the semi-norms
x -+ Ps(x) = sup{l(x, x')I: x' e S n C'} (S E 6)
which are readily seen to satisfy proposition (e) of (3.1).
Suppose now that l: is an 6-topology on E consistent with (E, E') and
that C is normal with respect to l:. By (3.1) (d) there exists a O-neighborhood
base U in (E, l:) consisting of convex, circled, and C-saturated sets. Since
{Uo: U E U} is a fundamental subfamily of 6, it suffices to show that there
exists, for each U e U, an integer no such that Uo c no(UO n C' - UO n C').
Let U e U be fixed.
Now the dual of the normed space Eu (for notation, see Chapter III,
Section 7) can be identified with Eu' and the cone C' n Eu' can be identified
with the dual cone of C u = c/Ju(C) where, as usual, <Pu is the canonical map
E -+ £u. Using the fact that U is C-saturated, it is readily seen that C u is a
normal cone in Eu; hence if we define the set Me E' by M = UO n C'
- UO n C', Lemma 1 implies thatEu' = U nM. Now Mis O'(E', E)-compact
"eN
by (I, 1.1) (iv), hence O'(E', E)-closed and a fortiori closed in the Banach space
E u'; since the latter is a Baire space, it follows that M has an interior point
u.;
and hence (being convex and.circled) is a neighborhood of 0 in E it follows
that Uo c noM for a suitable no E N and the proof is complete.
COROLLARY 1. Let C be a cone in the I.c.s. E. The following assertions are
equivalent,'
(a) C is a normal cone in E.
(b) For any equicontinuous set AcE', there exists an equicontinuous set
B c C' such that A c B - B.
(c) The topology of E is the topology of uniform convergence on the equi-
continuous subsets ofC'.
COROLLARY 2. If 6 is a saturatedfamity, covering E', of a{E', E)-relatively
compact sets and if H is an 6-cone in E', then the a(E', E)-closure H of H is a
strict 6-cone.
Proof In fact, the cone C = - HO is normal for the 6-topology which is
consistent with <E, E'), (IV, 3.2), and C' = H by (IV, 1.5).
COROLLARY 3. If C is a cone in the l.c.s. E. then E' = C' - C' ifand only if Cis
weakly normal; in particular, every normal cone in E is weakly normal.
We obtain this corollary by taking 6 to be the saturated hull of the family
of all finite subsets of E'. Let us point out that if C is a cone in a l.c.s. E over
C, it is sometimes of interest to consider the cone H c E' oflinear forms whose
real and imaginary parts are;;;; 0 on C; we have H = C' n (-iCY, and it
follows from Corollary 3 above and (IV, 1.5), Corollary 2, that E' = H - H
if and only if C + iC (equivalently, C - iC) is weakly normal in E (for,
H = (C - iC)').
REMARK. In normed spaces, weak normality and normality of cones
are equivalent (see (3.5) below).
The following is an application of (3.3) to the case where 6 is the family of
all strongly bounded subsets of the dual of an infrabarreled space E; recall
that this class comprises all barreled and all bomological (hence all metriz-
able l.c.) spaces.
3.4
Let E be an infrabarreled l.c.s., C a cone in E, !B the family of all strongly
bounded subsets of E'. The following assertions are equivalent:
(a) C is a normal cone in E.
(b) The topology of E is the topology of uniform convergence on strongly
bounded subsets ofC'.
(c) C' is a !B-cone in E'.
(d) C' is a strict !B-cone in E'.
The proof is clear from the preceding in view of the fact that !B is the
family of all equicontinuous subsets of E' (Chapter IV, Section 5).
COROLLARY. IfE is a reflexive space, normal cones and !B-cones correspond
dually to each other (with respect to <E, E'».
It is an interesting fact that the complete symmetry between normal and
!B-cones under the duality <E, E') remains in force, without reflexivity
assumptions, when E is a Banach space (cf. Ando [2]). From the proof of
this result we isolate the following lemma, which will be needed later, and is
of some interest in itself.
LEMMA 2. Let (E, :1:) be a metrizable t.v.s. over R, let C be a cone in E which
is complete, and let {Un: n E N} be a neighborhood base of 0 consisting of closed,
circled sets such that Un+1 + Un+1 C Un (n EN). Then the sets
(n EN)
form a O-neighborhood base for a topology:1: 1 on El =C- C such that (E1' :1:1 )
is a complete (metrizable) t.v.s. over R.
Proof. It is clear that each set Vn is radial and circled in E 1 , and obviously
V n+1 + V n+1 C Vn for all n E N. It follows from (I, 1.2) that {Vn: n E N} is a
O-neighborhood base for a (unique translation invariant) topology :1:1 on E1
under which E1 is a t.v.s. Of course (E1' :1:1) is metrizable, and there remains
to prove that (El' :1:1) is complete. In fact, given a Cauchy sequence in (E1' :1:1 ),
there exists a subsequence {zn} such that Zn+l - Zn E Vn(n EN); we have,
consequently, zn+1 - zn = xn - Yn' where Xn and Yn are elements of Un n C,
L Xn and n=l
L Yn converge
00 00
and it is evidently sufficient to show that the series

n=l n
L
00
in (E1, :1:1). Let us show this for Xn. Letting Un = LXv (n EN), we obtain
n=1 v=1
un+p - Un E (Un+1 + ... + Un+p ) n C C (Un n C) C Vn

for all pEN and n E N. Since C is complete in (E, :1:), {un} converges for :1: to
some U E C and we have U - Un+1 E Un n C c V", since Un n C is closed in
(E, :1:). Now the last relation shows that Un -+ U in (E1, :1:1), and the proof is
complete.
3.5
Theorem. Let E be a Banach space and let C be a closed cone in E. Then
C is normal (respectively, a strict m-cone) if and only if C' is a strict m-cone
(respectively, normal) in Eft.
Proof. The assertion concerning normal cones C c E is a special case of
(3.4), and if C is a m-cone, then C' is normal in Eft by (3.3). Hence suppose
that C' is normal in Eft and denote by U the unit ball of E. The bipolar of
Un C (with respect to (E', E"» is U OO n C" and by (3.3) C" is a strict
m-cone in the strong bidual E"; hence U OO n C" - U OO n C" is a O-neighbor-
hood in E". It follows that V = Un C - Un C is dense in a O-neighborhood
VI in E; if El = C - C and :1:1 is the topology on El defined in Lemma 2,
this means precisely that the imbedding'" of (E 1 , :1:1) into E is nearly open
and continuous, with dense range. Consequently, Banach's homomorphism
theorem (III, 2.1) implies that", is a topological isomorphism of (E1 , :1:1)
onto E, and hence C is a strict m-cone in E.
COROLLARY. Let E be a Banach space and let C be a cone in E with closure C.

Thefollowing assertions are equivalent:
(a) C is a $-cone in E.
(b) E = C - C.
(c) C is a strict $-cone in E.
Proof. (a) => (c) is clear from the preceding since C' is normal in E' when-
ever C is a $-cone, by (3.3). (c) => (b) is trivial. (b) => (a): Let M denote the
closure of U () C - U () C, where U is the unit ball of E. Then M is convex,
00
circled (over R), and such that E = U nM; since E is a Baire space, it foI-
l
lows that M is a O-neighborhood in E and hence C is a $-cone.
4. ORDERED TOPOLOGICAL VECTOR SPACES
LetL be a t.v.s. (over R or C) and an ordered vector space; we say thatL is

an ordered topological vector space if the following axiom is satisfied:
(LTO) The positive cone C = {x: x ~ O} is closed in L.
Recall that an Archimedean ordered vector space is called regularly
ordered if the real bilinear form (x, x*) ~ Re (x, x*) places Lo and Lri in
duality, where Lo is the real underlying space of L (Chapter I, Section 7). In
order to prove some alternative characterizations, we need the following
lemma which is of interest in itself. (Cf. Exercise 21.)
LEMMA. Let E be an ordered vector space offinite dimension over R. The order
of E is Archimedean if and only if the positive cone C is closed for the unique
topology under which E is a Hausdorff t.v.s.
Proof. If C is closed, then clearly the order of E is Archimedean. Conversely,
suppose that E is Archimedean ordered; without restriction of generality we
can assume that E = C - C. If the dimension of E is n (~ 1), then C con-
tains n linearly independent elements Xl> ••• , Xn and hence the n-dimensional
simplex with vertices 0, Xl' ... , x n ; since the latter has non-empty interior, so
does C. Now let x E C and let y be interior to C; by (II, 1.1) n-1y + x is
interior to C (n EN), and hence we have -x ~ n-1y for all n. This implies
-x ~ 0 or, equivalently, x E C.
4.1
If L is an ordered vector space over R with positive cone C, the following pro-
positions are equivalent:
(a) The order of L is regular.
(b) C is sequentially closed for some Hausdorff I.c. topology on L, and L +
distinguishes points in L.
14] ORDERED TOPOLOGICAL VECTOR SPACES 223
(c) The order of Lis Archimedean, and C is normal for some Hausdorff l.c.
topology on L.
Proof. (a) => (b): It suffices to show that the intersection of C with every
finite dimensional subspace M is closed (Chapter II, Exercise 7), and this is
immediate from the preceding lemma, since the order of L is Archimedean,
and hence the canonical order of each subspace MeL is Archimedean.
(b)=>(c): If l: is a Hausdorff I.c. topology under which C is sequentially
closed, then, clearly, L is Archimedean ordered. Moreover, since L + separates
points in L, the canonical bilinear form on L xL· places Land L + = C· - C·
in duality, and by (3.3) C is normal for the Hausdorff I.c. topology a(L, L +).
(c) => (a): It suffices to show that L + separates points in L. If l: is a Haus-
dorff I.e. topology for which C is normal, then (L, l:)' = C' - C' by (3.3);
hence C' - C', and a fortiori L + = C· - C· separates points in L. This
completes the proof.
COROLLARY 1. The canonical orderings of subspaces, products, and direct
sums of regularly ordered vector spaces are regular.
COROLLARY 2. Every ordered locally convex space is regularly ordered.
Proof. If (E, l:) is an ordered I.c.s., then C is closed by definition, and the
bipolar theorem (IV, 1.5) shows - C to be the polar of C' with respect to
(E, E'). Since C r. - C = {O}, it follows that C' - C' is weakly dense in
(E, l:)" hence L + = C· - C· separates points in L.
If A is an ordered set and SeA is a subset ( =F 0) directed for ~, recall
that the section filter ~(S) is the filter on A determined by the base {Sx: xeS},
where Sx = {yeS: y ~ x}, and Sx is called a section of S. In particular, if S
is a monotone sequence in A, then ~(S) is the filter usually associated with S.
4.2
LetL be an orderedt.v.s. and let S be a subset ofL directedfor~. If the section
filter ~(S) converges to Xo e L, then Xo = sup S.
Proof. Let xeS and let z be any element of L rnajorizing S; we have
x ~ y ~ z for all y e Sx, and from Xo e Sx it follows that x ~ Xo ~ z, since
the positive cone is closed in L. This proves that Xo = sup S.
A deeper result is the following monotone convergence theorem, which
can be viewed as an abstract version of a classical theorem of Dini. Although
it can be derived from Dini's theorem, using (4.4) below (Exercise 9), we give
a direct proof based on the Hahn-Banach theorem.
4.3
Theorem. Let E be an ordered l.c.s. whose positive cone C is normal, and
suppose that S is a subset ofL directedfor ~. If the section filter ~(S) converges
for a(E, E'), then it converges in E.
Proof Without loss in generality we can suppose that S is directed for ~,

and that lim ~(S) = 0 for aCE, E'); it follows now from (4.2) that SeC.
Assume that the assertion is false; then there exists a O-neighborhood U in E
that contains no section of S, and since C is normal we can suppose that U is
convex and C-saturated. Since XES n U implies S" c U, it follows that
S n U = 0; moreover, (S + C) n U = 0, since U is C-saturated, and S + C
is convex, since it is the union of the family {x + C: XES} of convex sets
which is directed under inclusion. Hence by (II, 9.2) U and S can be separated
by a closed real hyperplane in E, and this contradicts the weak convergence
of ~(S) to O.
COROLLARY I. Let S be a directed (;;;;) subset of E such that Xo = sup S, where
E is an ordered l.c.s. with normal positive cone. Jf for every real linear form f
which is positive and continuous on E one has f(xo) = sup{f(x): XES}, then
lim g(x) = g(xo) (g E E') uniformly on each equicontinuous subset of E'.
:reS
Proof In fact, in view of (3.3), Corollary 3, the weak convergence of ~(S)

to Xo is equivalent to the relationf(xo) = sup{f(x): XES} for every real linear
formf on E which is positive and continuous (cf. (I, 7.2».
The reader will note that the preceding corollary is equivalent with (4.3).
The following result can be viewed as a partial converse of (4.2).
COROLLARY 2. Let E be a semi-reflexive, ordered l.c.s. whose positive cone is
normal. If S is a directed (;;;;) subset of E which is majorized or (topologically)
bounded, then Xo = sup S exists and~(S) converges to Xo'
Proof Let S" be any fixed section of S; it suffices to show that sup S"
exists in E. If S is majorized by some z E E, then S" c [x, z] and hence S" is
bounded in E by (3.1), Corollary 2; hence assume that S" is bounded in E.
The weak normality of C implies that E' = C' - C', and hence that the
section filter ~(S,,) is a weak Cauchy filter in E which is bounded. It follows
from (IV, 5.5) that ~(S,,) converges to some Xo E E, and (4.2) implies that
Xo = sup S", since C is closed and hence (being convex) weakly closed in E.
The following result is an imbedding (or representation) theorem for
ordered I.c.s. over R; let us denote by X a (separated) locally compact space,
and by R(X) the space of all real-valued continuous functions on X under the
topology of compact convergence and endowed with its canonical order
(Section 1).
4.4
Let E be an ordered l.c.s. over R. Jf(and only if) the positive cone C of E is
normal, there exists a locally compact space X such that E is isomorphic (as
an ordered I.v.s.) with a subspace of R(X).
Proof The condition is clearly necessary, for the positive cone of R(X)
(and hence of every subspace of R(X» is normal. To show that the condition
is] POSITIVE LINEAR FORMS AND MAPPINGS 225
is sufficient, we note first from (3.3), Corollary 1, that the topology of E is

the topology of uniform convergence on the equicontinuous subsets of the
dual cone C' c: E'. Let {B.. : a E A} be a fundamental family of q(E', E)-
closed ~quicontinuous subsets of C'; under the topology induced by q(E', E),
each B.. is a compact space. We define X as follows: Endow A with the discrete
topology, C' with the topology induced by q(E', E), and let X .. be the subspace
{a} x B.. of the topological product A x C'; then X is defined to be the sub-
space U.. X.. of Ax C'. The space X is the topological sum of the family
{B.. : a E A}; clearly, X is a locally compact space in which every X .. is open
and compact, and hence every compact subset of X is contained in the union
of finitely many sets X... For each x E E, we define an element fx E R(X) by
putting fit) = <x, x') for every t = (a, x') E X; it is clear that x --+ fx is an
algebraic and order isomorphism of E into R(X). Finally, since a closed
subset of X is compact if and only if it is contained in a finite union U XII'
it is also evident that x --+ fx is a homeomorphism.
REMARKS. It is easy to see that R(X) is complete; hence the image of
E under x --+ fx is closed in R(X) if and only if E is complete. Moreover,
if E is metrizable, then the family {B.. : a E A} can be assumed to be
countable, and hence X countable at infinity; if E is normable, one can
take X = UO n C' (under q(E', E», where U is any bounded neighbor-
°
hood of in E (in particular, the unit ball if E is normed). If E is a sep-
arable normed space, UO n C' is a compact metrizable space for
q(E', E) by (IV, 1.7) and hence a continuous image of the Cantor set
(middle third set) in [0, 1] c: R; in this case X can be taken to be the
Cantor set itself, or [0, 1] (for details, see Banach [1], chap. XI,
§ 8, theor. 9).
Finally, proposition (4.4) can be specialized to the case C = {o}; we obtain
thus a representation of an arbitrary I.c.s. E over R as a subspace of a suitable
space R(X); it is immediate that in this particular case, the restriction to the
scalar field R can be dropped.
5. POSITIVE LINEAR FORMS AND MAPPINGS

The present section is concerned with special properties of linear maps
U E L(E, F) which map the positive cone C of E into the positive cone D of F,
where E, F are ordered vector spaces (respectively, ordered t.v.s.); these
mappings are called positive. It is clear that the set H of all positive maps is a
cone in L(E, F); whenever M is a subspace of L(E, F) such that H n M is a
proper cone, H n M defines the canonical ordering of M (Section 1). Recall
also (Section 2) that a linear form f on an ordered vector space E is called
positive if Ref(x) ~ 0 for each x in the positive cone C of E.
We begin our investigation with some simple but useful observations
concerning the properties of the cone :K c: !l' (E, F) of continuous positive
maps, where E, F are supposed to be ordered t.V.s. over K. We point out that
in view of the agreements made in Section 2, it suffices in general to consider

the case k = R.
5.1
Let E, F be ordered t.V.S. and let 6 be a family of bounded subsets of E that
covers E. Then the positive cone Yf c: 2 (E, F) is closed for the 6-topology.
For Yf to be a proper cone, it is sufficient (and, if E is a l.c.s. and F 'i= {OJ, neces-
sary) that the positive cone C of E be total in E.
Proof In fact, by definition of the 6-topology (Chapter III, Section 3) the
bilinear map (u, x) --+ u(x) is separately continuous on 2 e;(E, F) x E into F;
hence the partial map fx: u --+ u(x) is continuous for each x E E. Since Yf
= () {ix-leD): x E C} and the positive cone D of F is closed, Yf is closed in
2 e,(E, F). Further, since D is proper, u E Yf II -.1f implies that u(x) = 0 for
x E C; hence u = 0 if C is total in E. Finally, if E is a I.c.s. and C is not total
in E, there exists an fEE' such that f'i= 0 but f( C) = {OJ, by virtue of the
Hahn-Banach theorem; if y is any element 'i= 0 of F, the mapping u = f® y
(defined by x --+ f(x)y) satisfies u E Yf II - Yf.
COROLLARY. If C is total in E and ifF is a l.c.s., the (canonical) ordering of
2(E, F) defined by Yf is regular.
Proof In fact, Yf is a closed proper cone for the topology of simple con-
vergence which is a Hausdorff I.e. topology by (III, 3.1), Corollary; the
assertion follows from (4.1), Corollary 2.
5.2
Let E, F be ordered I.c.s. with respective positive cones C, D and let 6 be a
family of bounded subsets of E. If C is an 6-cone in E and D is normal in F, the
positive cone Yf c: 2 (E, F) is normalfor the 6-topology.
Proof Since D is normal in F, there exists, by (3.1), a family {qa: 0( E A} of
real semi-norms on F that generate the topology of F, and which are mono-
tone (for the order of F) on D. Since C is an 6-cone in E, it follows that the
real semi-norms
u --+ Pa,S(u) = sup{qa(ux): XES II C} (0( E A, S E 6)
generate the 6-topology on .P(E, F). Now, evidently, each Pa,S is monotone
on Yf (for the canonical order of 2(E, F»; hence Yf is a normal cone in
2 elE, F), as asserted.
On the other hand, there are apparently no simple conditions guaranteeing
that Yf is a %-cone in 2 6 (E, F), even for the most frequent types of families
% of bounded subsets of 2 6 (E, F), except in every special cases (cf. Exercise
7). At any rate, the following result holds where E, F are ordered I.c.s. with
respective positive cones C, D, and 2.(E, F) denotes 2(E, F) under the
topology of simple convergence.
§5] POSITIVE LINEAR FORMS AND MAPPINGS 227
5.3
If C is weakly normal in E and if F = D - D, then .!If - .!If is dense in
2.(E, F).
Proof. Since the weak normality of C is equivalent with E' = C' - C' by
(3.3), Corollary 3, the assumptions imply that .!If - .!If contains the sub-
space E' ® F of 2(E, F). On the other hand, the dual of 2 iE, F) can be
identified with E ® F' by (IV, 4.3), Corollary 4, and it is known that (under
the duality between 2(E, F) and E ® F') E' ® F separates points in E ® F'
(Chapter IV, Section 1, Example 3); it follows from (IV, 1.3) that E' ® F is
weakly dense in 2.(E, F) and hence (being convex) dense in 2.(E, F).
We turn to the question of extending a continuous positive linear form,
defined on a subspace of an ordered t.v.S. E, to the entire space E. The fol-
lowing extension theorem is due to H. Bauer [1], [2] and, independently, to
Namioka [1].
5.4
Theorem. Let E be an ordered t.V.S. with positive cone C and let M be a
subspace of E. For a linear form fo on M to have an extension f to E which is a
continuous positive linear form, it is necessary and sufficient that Refo be bounded
above on M (\ (U - C), where U is a suitable convex O-neighborhood in E.
Proof. It suffices to consider the case K = R. Iff is a linear extension of fo
to E which is positive and continuous and if U = {x:f(x)} < I}, it is clear
that fo(x) < 1 whenever x E M (\ (U - C); hence the condition is necessary.
Conversely, suppose that U is an open convex O-neighborhood such that
x E M (\ (U - C) implies fo(x) < y for some y E R. Then y > 0 and N
= {x E M:fo(x) = y} is a linear manifold in E not intersecting the open
convex set U - C. By the Hahn-Banach theorem (II, 3.1) there exists a closed
hyperplane H containing N and not intersecting U - C, which, consequently,
can be assumed to be of the form H = {x:f(x) = y}; clearly, f is a con-
tinuous extension offo. Furthermore, since 0 E U - C it follows thatf(x) < y
when x E U - C and hence when x E - C; thus x E C impliesf(x) ~ O.
COROLLARY 1. Let fo be a linear form defined on the subspace M of an ordered
vector space L. fo can be extended to a positive linear form f on L if and only if
Re fo is bounded above on M (\ (W - C), where W is a suitable convex radial
subset of L.
In fact, it suffices to endow L with its finest locally convex topology for
which W is a neighborhood of 0, and to apply (5.4). The same specialization
can be made in the following result which is due to Krein-Rutman [I].
COROLLARY 2. Let E be an ordered t.v.S. with positive cone C, and suppose that
M is a subspace of E such that C (\ M contains an interior point of C. Then every
continuous, positive linear form on M can be extended to E under preservation of
these properties.
Proof Iffo is the linear form in question and Xo EM is an interior point of

C, choose a convex O-neighborhood U in E such that Xo + U c 2xo - C.
Then Re fo is bounded above on M n (U - C), for we have M n (U - C)
c (xo - C) n M.
REMARK. The condition of Corollary 2 can, in general, not be replaced

by the assumption that C n M possesses an interior point (Exercise 14).
For another condition guaranteeing that every linear form fo, defined
and positive on a subspace M of an ordered vector space L, can be ex-
tended to a positive linear formf on L see Exercise 11.
There is a comparatively large class of ordered t.v.s. on which every posi-
tive linear form is necessarily continuous; we shall see (Section 7 below) that
this class includes all topological vector lattices that are at least sequentially
complete (semi-complete). It is plausible that in spaces with this property, the
positive cone must be sufficiently" wide" (cf. the discussion following (3.2).
More precisely, one has the following result (condition (ii) is due to Klee
[2], condition (iii) to the author [2]).
5.5
Theorem. Let E be an ordered t.v.s. with positive cone C. Each of the
following conditions is sufficient to ensure the continuity of every positive linear
form on E:
(i) C has non-empty interior.

(ii) E is metrizable and complete, and E = C - C.
(iii) E is bornological, and C is a semi-complete strict lB-cone.
Proof It is again sufficient to consider real linear forms on E. The suffi-

ciency of condition (i) is nearly trivial, for if f is positive, then f- 1(0) is a
hyperplane in E lying on one side of the convex body C, and hence closed
which is equivalent with the continuity off by (I, 4.2). Concerning condition
(ii), we use Lemma 2 of Section 3: The topology ::II on E, determined by the
neighborhood base of 0, {Vn: n EN}, where Vn = Un n C - Un n C, is
evidently finer than the given topology ::I of E, and hence we have ::I = ::I1
by Banach's theorem (III, 2.1), Corollary 2. Now iffis a positive, real linear
form on E which is not continuous, thenfis unbounded on each set Un n C
hence there exists Xn E Un n C such that f(xn) > 1 (n EN). On the other
hand, since Un+ I + Un+ I C Un for all n, the sequence {xn} is summable in E
p
with sum L Xn = Z E C
neN
(C being closed), and from z ~ L
n~1
Xn we obtain
fez) > p for each pEN, which is contradictory. Finally, concerning con-
dition (iii) we observe that since E is bornological and C is a strict lB-cone, a
linear form on E which is bounded on the bounded subsets of C is necessarily
continuous by (II, 8.3); now iff is a positive, real linear form on E which is
is] POSITIVE LINEAR FORMS AND MAPPINGS 229
not continuous, there exists a bounded sequence {x,,} in C such thatf(x,,) > n
(n EN). Since E is locally convex by definition, we conclude that {n- 2 x,,:
n EN} is a summable sequence in C with sum z e C, say, and it follows that
L n- 2!(x,,) > L ,,-1

p p
!(z) ~ for all p,
11=1 11=1
which is impossible. The proof is complete.

COROLLARY. Let E be an ordered l.c.s. which is the inductive limit of afamily
{E,,: oc e A} ofordered (F)-spaces with respect to afamily ofpositive linear maps,
and suppose that E" = C" - C,,·(oc E A). Then each positive linear form on E is
continuous.
This is immediate in view of (II, 6.1). For locally convex spaces, an im-
portant consequence of (5.5) is the automatic continuity of rather extensive
classes of positive linear maps.
5.6
Let E, F be ordered I.c.s. with respective positive cones C, D. Suppose that E
is a Mackey space on which every positive linear form is continuous, and assume
that D is a weakly normal cone in F. Then every positive linear map ofE into'F is
continuous.
Proof. Let u be a linear map of E into F such that u(C) c D, and consider
the algebraic adjoint u* of u (Chapter IV, Section 2). For each y' eD',
x ~<x, u*y') is a positive linear form on E, hence continuous by assump-
tion; since F' = D' - D' by (3.3), Corollary 3, it follows that u*(F') c E';
hence u is weakly continuous by (IV, 2.1). Thus u e 2(E, F) by (IV, 7.4).
We conclude this section with an application of several of the preceding
results to the convergence of directed families of continuous linear maps.
5.7
Let E be an ordered barreled space such that E = C - C, and let F be an
ordered semi-reflexive space whose positive cone D is normal. Suppose that 'PI is a
subset of 2 (E, F) which is directed upwardfor the canonical order of 2 (E, F),
and either majorized or simply bounded. Then Uo = sup 'PI exists, and the
section filter ~('PI) converges to Uo uniformly on every precompact subset of E.
Proof. In fact, (5.1) and (5.2) show that .tf is a closed normal cone in
2.(E, F) and hence is the positive cone for the canonical order of !l'S<E, F).
For each x E C, the family {u(x): u E 'PI} satisfies the hypotheses of (4.3),
Corollary 2, and hence of (4.3), so ~('PI) converges simply to a linear map
Uo E 2(E, F). By (III, 4.6) Uo is continuous, and the convergence of ~('PI) is
uniform on every precompact subset of E. Since .tf is closed in !i'.(E, F),
(4.2) implies that Uo = sup 'PI.
6. THE ORDER TOPOLOGY

If S is an ordered set, the order of S gives rise to various topologies on S
(cf. Birkhoff [1)); however, in general, the topologies so defined do not satisfy
axioms (LT)l and (LTh (Chapter I, Section 1) if S is a vector space, even if
(LTO) holds (cf. Exercise 17). On the other hand, if L is an ordered vector
space over R, there is a natural locally convex topology which, as will be seen
below, is the topology of many (if not all) ordered vector spaces occuring in
analysis. The present section is devoted to a study of the principal properties
of this topology. (See also Gordon [2].)
Let L be an ordered vector space over R; we define the order topology 20
of L to be the finest locally convex topology on L for which every order inter-
val is bounded. The family of locally convex topologies on L having this
property is not empty, since it contains the coarsest topology on L, and 20
is the upper bound of this family (Chapter II, Section 5); a subset We L is a
O-neighborhood for 20 if and only if W is convex and absorbs every order
interval [x, y] c L. (W is necessarily radial, since {x} = [x, x] for each x E L.)
Although 20 is a priori defined for ordered vector spaces over R only, it can
happen (cf. the corollaries of (6.2) and (6.4) below) that (L, 2) is an ordered
vector space over C such that (Lo, 2) = (Lo, 2 0 ), where L o is the underlying
real space of L. We begin with the following simple result.
6.1
The dual of(L, 2 0 ) is the order bound dual Lb ofL. If Lb separates points in L
(in particular, if the order of L is regular), (L, 2 0 ) is a bomological l.c.s. If
L, M are ordered vector spaces, each positive linear map ofL into M is continuous
for the respective order topologies.
Proof It is clear from the definition of 20 that each order interval is bounded
for 2 0 ; hence if f E (L, 2 0 )' then f E Lb. Conversely, if f E L b, then
f-1([ -I, I)) is convex and absorbs each ord~r interval, and hence is a O-neigh-
borhood for 2 0 , 20 is a Hausdorff topology ifand only ifLb distinguishes points
in L. Let W be a convex subset of L that absorbs each bounded subset of
(L, 2 0 ); since Wa fortiori absorbs all order intervals in L, W is a 2 0 -neigh-
borhood of O. Hence (L, 2 0 ) is bornological if (and only if) 20 is a Hausdorff
topology. Finally, if u is a positive linear map of L into M, then u([x, y))
c [u(x), u(y)] for each order interval in L; hence if V is convex and absorbs
order intervals in M, u- 1 (V) has the same properties in L and thus u is con-
tinuous forthe order topologies. (Cf. Exercise 12.)
COROLLARY. Let Li (i = I, ... , n) be a finite family of ordered vector spaces,
and endow L = TIiLi with its canonical order. Then the order topology ofL is the
product of the respective order topologies of the L i.
Proof We show that the projection Pi of (L, 2 0 ) onto (Li' 2 0 ) (i = 1, ... , n)
is a topological homomorphism. In fact, Pi is continuous by (6.1); if Ii is an
order interval in Li then Ii x to} is an order interval in L. hence if W is a
§61 THE ORDER TOPOLOGY 231
convex O-neighborhood in (L, l:o) then Pi( W) is convex and absorbs II

The order topology is most easily analyzed when L is an Archimedean
ordered vector space with an order unit e. For convenience of expression, let
us introduce the following terminology: A sequence {xn: n E N} of elements
~ 0 of an ordered vector space L is order summable if supn Un exists in L,
n
where Un := L xp- We shall say that a positive sequence {xn: n E N} is of type
p=1
11 if there exists an a ~ 0 in L and a sequence (An) Ell such that (0 ~)Xn ~ Ana
(n EN).
6.2
Let L be an Archimedean ordered vector space over R, possessing an order unit
e. Then (L, l:o) is an ordered I.v.s. which is normable, l:o is the finest locally
convex topology on L for which the positive cone C is normal, and the following
assertions are equivalent:
(a) (L, l:o) is complete.
(b) Each positive sequence of type [I in L is order summable.
Proof. The order interval [-e, e] is convex, circled, and (by the definition
of order unit) radial in L; since L is Archimedean ordered, the gauge Pe of
[-e, e] is a norm on L. The topology generated by P e is finer than l:o since
[-e, e] is l:o-bounded, and it is coarser than l:o, since it is locally convex
and [-e, e] absorbs order intervals; hence Pe generates l:o. To see that Cis
closed in (L, l:o), note that e is an interior point of C; the fact that C is closed
follows then, as in the proof of the lemma preceding (4.1), from the hypothesis
that Lis Archimedean ordered. Moreover, since by (3.1), Corollary 2, l:o is
finer than any l.c. topology on L for which C is normal, the second assertion
follows from the fact that the family {B[ -e, e]: B> O} is a O-neighborhood
base for l:o that consists of C-saturated sets.
Further, it is clear that (a) => (b), since every positive sequence of type 11
in L is of type /1 with respect to a = e, and hence even absolutely summable
in (L, l:o); the assertion follows from (4.2). (b) => (a): We have to show that
(L, l:o) is complete. Given a Cauchy sequence in (L, l:o), there exists a sub-
sequence {xn: n E N} such that for all n, Pe(xn+1 - xn) < An' where (An) E /1;
hence Xn+1 - Xn E An[ -e, e] and we have Xn+1 - Xn = Un - Vn> where Un = Ane
+ (Xn+l - xn) and Vn = Ane (n EN). To show that {xn} converges, it suffices to
00
show that L Un converges. Now 0 ~ un ~ 2Ane; hence {un} is of type /1 and
n== 1
L
II
supn Up = U E C exists by hypothesis. Since for all n

p=1
o ~ U - p=1 L up ~ 2 (p=n+l
n+k
Ln up =
p=n+l
SUPk
00)
L Ap e,
00
it follows that L Un = U for l:o and hence (L, l:o) is complete.
n=1
COROLLARY 1. If Lis Archimedean ordered and has an order unit, the order of
L is regular and we have Lb = L + •
This is immediate in view of (6.1) and (3.3).
COROLLARY 2. Let (E, l:) be an ordered Banach space possessing an order unit.
Then l: = l:o if and only if the positive cone C ofE is normal in (E, l:).
Proof. In fact, the order of E is Archimedean, since C is closed in (E, l:);

if l: = l:o, then C is normal by (6.2). Conversely, if C is normal, then l: is
coarser than l:o; since [-e, e] is a barrel in (E, l:) (we can suppose that
K = R), it follows that l: = l:o.
Examples to which the preceding corollary applies are furnished by the
spaces 'C(X) (X compact) and L oo(Jl) (Chapter II, Section 2, Examples 1 and 2)
and, more generally, by every ordered Banach space whose positive cone is
normal and has Don-empty interior. It is readily verified that each interior
point of the positive cone C of an ordered t.v.s. L is an order unit, and each
order unit is interior to C for l:o.
However, most of the ordered vector spaces occurring in analysis do not
have order units, so that the description of l:o given in (6.2) does not apply.
Let L be an Archimedean ordered vector space over R and denote, for each
00
a ~ 0, by La the ordered subspace La = U n[ -a, a] endowed with its order

n=l
topology; La is a normable space. The family {La: a ~ O} is evidently directed
under inclusion c, and if La C L b, the imbedding map hb,a of La into Lb is
continuous.
6.3
Let L be a regularly ordered vector space over R, and denote by H any subset
-
of the positive cone C of L which is cofinal with C for ~. Then (L, l:o) is the
inductive limit lim hb aLa (a, b E H).
'
Proof. By (6.1), the assumption on L implies that l:o is Hausdorff. In view
of the preceding remarks and the definition of inductive limit (Chapter II,
Section 6), it suffices to show that l:o is the finest I.e. topology on L for which
each of the imbedding maps fa: La -+ L (a E H) is continuous. Since H is
cofinal with C, each order interval [x, y] c Lis contained ina translate of some
[-a, aJ where a E H, and hence [x, yJ is bounded for the topology l: of the
inductive limit; hence l:o is finer than l:. On the other hand, if W is a convex
O-neighborhood in (L, l:o), then W absorbs all order intervals in L, which
implies thatfa- 1(W) is a O-neighborhood in La (a E H), and hence l: is finer
than l:o.
COROLLARY 1. If the order ofL is regular and each positive sequence of type [1
in L is order summable, (L, l:o) is barreled.
16] THE ORDER TOPOLOGY 233
Proof. In fact, the assumption implies by (6.2) that each of the spaces La
(a E H) is normable and complete and hence barreled; the result follows from
(11,7.2).
COROLLARY 2. If the order of L is regular and the positive cone C satisfies
condition (D) of(1.1), then Cis normal for ;to (hence the dualof(L, ;to) is L +).
Proof. By definition of the topology of inductive limit, a O-neighborhood
base for ;to is given by the family of all convex radial subsets U of L such that
V= Urt(C-C) is of the form V=r{Pa[-a,a]:aEH} where a-'Pa is
any mapping of H into the set of real numbers> 0. We prove the normality
of C via (3.1) (c) by showing that x E U and y E [0, x] imply y E U. If x E U
n n
and x ~ 0, then x is of the form x = L AiZi, where i=L1 !AI! ~ 1 (Ai E R)
i= 1
and
n
Zi E Pa.[ -ab ad (i = 1, ... , n). If y E [0, x] it follows that y ~ L !Ai!Pa,al; by
i=1
n
repeated application of (D), we obtain y = L !Ai!YI>
i= 1
where Yi E Pa.[O, ai]
(i = 1, ... , n). Hence Y E V c: U as was to be shown.
REMARK. Since La c: Lb (where a, b E C) is equivalent with a ~ Ab
for a suitable scalar A > 0, it suffices in (6.3) to require that the set of all
positive scalar multiples of the elements a E H be cofinal with C (for ~);
in particular, if L has an order unit e, it suffices to take H = {e}. Let us
note also that the inductive limit of (6.3) is in general not strict (the
topology induced by Lb on La (b > a) is, in general, not the order topol-
ogy of La). For example, if L is the space L2(jL) and a, baretherespective
equivalence classes of two functions/, 9 such thatO ~f ~ g,Jis bounded
and 9 Jl-essentially unbounded, then the topology of La is strictly finer
than the topology induced on La by Lb' (Cf. Exercise 12.)
We apply the preceding description of ;to to the case where L is a vector
lattice; the lattice structure compensates in part for the lack of an order unit
and one obtains a characterization of ;to that can be compared with (6.2).
6.4
Let L be a vector lattice whose order is regular and let ;t be a locally convex
topology on L.These assertions are equivalent:
(a) ;t is the order topology ;to.
(b) ;t is the finest I.e. topology on Lfor which C is normal.
(c) ;t is the Mackey topolOgy with respect to (L, L +).
Proof. Let us note first that by (1.4), Lb = L + and that since the order of L
is assumed to be regular, (L, L +) is a duality. (a) <=> (b): Since the positive
cone of a vector lattice satisfies (D) of (1.1), this follows from the fact that
by (3.1), Corollary 2, ~o is finer than any I.c. topology for which C is normal,
in view of (6.3), Corollary 2. (a) - (C): Since (L, l:o)' = Lb = L +, l:o is

consistent with the duality <L, L + ); since (L, l:o) is bomological, l:o is neces-
sarily the Mackey topology with respect to <L, L +).
The following corollary is now a substitute for (6.2), Corollary 2.
COROLLARY. Let (E, ~) be an ordered (F)-space (over R) which is a vector
lattice. Then l: = l:o !f and only if the positive cone C of E is normal in (E, l:).
Proof If ~ = l:o, then C is normal by (6.4). Conversely, if C is normal, then
E' = C' - C' by (3.3), and C' - C' = E+ by (5.5) (for C is closed in (E, ~)
and E = C - C); since E is a Mackey space by (IV, 3.4), the assertion follows
from (6.4) (c). .
7. TOPOLOGICAL VECTOR LATTICES

Let L be a t.v.s. over R and a vector lattice, and consider the maps x -+ lxi,
x -+x+, X -+ x- of L into itself, and the maps (x, y) -+ sup(x, y) and (x, y)
-+ inf(x, y) of L x L into L. By utilizing the identities (I), (2) and (3) of
Section I, it is not difficult to prove that the continuity of one of these maps
implies the continuity (in fact, the uniform continuity) of all of them; in this
case, we say that" the lattice operations are continuous" in L. Recall that a
subset A of L is called solid if x E A and Iyl ;;:; Ixl imply that YEA; we call L
locally solid if the t.v.s. L possesses a O-neighborhood base of solid sets.
7.1
Let L be a t.v.s. over R and a vector lattice. The following assertions are
equivalent:
(a) L is locally solid.
(b) The positive cone ofL is normal, and the lattice operations are continuous.
Proof (a) => (b): Let U be a O-neighborhood base in L consisting of solid
sets; if x E U E U and 0 ;;:; y ;;:; x, then y E U and hence the positive cone C
of L is normal by (3.1) (c). Moreover, if x - Xo E U, we conclude from (6)
of (1.1) that x+ - xci E U (U E U) and hence the lattice operations are
continuous.
(b) => (a): Suppose that C is normal and the lattice operations are con-
tinuous. Let U be a O-neighborhood base in L consisting of circled C-saturated
sets (Section 3). For a given U E U, choose V E U, WE U so that V + V c U
and that x E W implies x+ E V. Now if x E W, then -x E W, since W is
circled; hence x+ and x- = (-x)+ are in V and Ixl = x+ + x- E U. If
Iyl ~ lxi, then y E [-Ixl.lxl]; hence, since U is C-saturated, it follows that
y E U. Therefore, the set {y: there exists x E W such that Iyl ~ Ixl} is a O-neigh-
borhood contained in U, and is obviously solid.
It is plausible that for t.v.s. that are vector lattices, just as for more general
types of ordered t.v.s., the axiom (LTO) (closedness of the positive cone) by
§7] TOPOLOGICAL VECTOR LATTICES 235
itself is too weak to produce useful results. We define a topological vector

lattice to be a vector lattice and a Hausdorfft.v.s. over R that is locally solid;
it will be seen from (7.2) below that in these circumstances, the positive cone
of L is automatically closed, and hence every topological vector lattice is an
ordered t.v.s. over R. A locally convex vector lattice (abbreviated I.c.v.I.) is
a topological vector lattice whose topology is locally convex. Every solid set
is circled (with respect to R, cf. (4) of(1.1»; hence a topological vector lattice
possesses a base of circled solid O-neighborhoods. Since the convex hull of a
solid set is solid (hence also circled), a I.c.v.I. possesses a O-neighborhood base
of convex solid sets. The gauge function p of a radial, convex solid set is
characterized by being a semi-norm such that Iyl ~ Ixl implies p(y) ~ p(x),
and is called a lattice semi-nonn on L. Therefore, the topology of a I.c.v.I.
can be generated by a family of lattice semi-norms (for example, by the family
of all continuous lattice semi-norms). A Frechet lattice is a I.c.v.I. which is
an (F)-space; a nonned lattice is a normed space (over R) whose unit ball
{x: Ilxll ~ I} is solid. By utilizing (I, 1.5) and the uniform continuity of the
lattice operations, it is easy to see that with respect to the continuous exten-
sion of the lattice operations, the completion of a topological vector lattice
is a topological vector lattice; in particular, the completion of a normed
lattice is a complete normed lattice with respect to the continuous extension
of its norm. A complete normed lattice is called a Banach lattice. 'Let us
record the following elementary consequences of the definition of a topo-
logical vector lattice.
7.2
In el'ery topological vector lattice L, the positive cone C is closed, normal, and
a strict m-cone; if L is order complete, every band is closed in L.
Proof C is normal cone by (7.1) and, since C = {x: x- = O}, C is closed,
since the topology of L is Hausdorff and x -+ x- is continuous. To show that
C is a strict m-cone, recall that if B is a circled, bounded set, then B+ = B-,
and hence B c B+ - B+. It suffices, therefore, to show that B+ is bounded
if B is bounded. If B is bounded and U is a given solid O-neighborhood in L,
there exists A > 0 such that BeAU; since AU is evidently solid, it follows
that B+ c AU hence B+ is bounded. Finally, if A is a band in L, then A = A.l.L
by (1.3), Corollary 1. Now each set {alL = {x E L: inf(lxl, lal) = O} is closed,
since L is Hausdorff and x -+ inf(lxl, lal) is continuous, and we have A =
n {{alL: a E AL}.
Examples
1. The Banach spaces (over R) LP{jJ,) (Chapter II, Section 2, Exam-
ple 2) are Banach lattices under their canonical orderings; it will be seen
below that these are order complete for p < 00, and the spaces Ll{jJ,) and
L OO(J.l) will be important concrete examples for the discussion in Section
8. The corresponding spaces over C can be included in the discussion,

as they are complexifications (Chapter I, Section 7) of their real counter-
parts.
2. Let A. be a subspace of rod such that A. = A. x x (Chapter IV, Section 1,
Example 4); A. is a perfect space in the sense of Kothe [5]). Under the
normal topology (Kothe [5], Peressini [2]) A. is a I.c.v.I. when endowed
with its canonical ordering as a subspace of rod' The normal topology is
the topology of uniform convergence on all order intervals of A. x, and
the coarsest topology consistent with (A., A. x> such that the lattice
operations are continuous. (Cf. Exercise 20.)
3. Let X be a locally compact (Hausdorff) space and let E be the space
of all real-valued functions with compact support in X, endowed with
its inductive limit topology (Chapter II, Section 6, Example 3). The top-
ology of E is the order topology Zo (Section 6), so that E is a locally
convex vector lattice (see (7.3»; E is, in general, not order complete.
The dual of (E, Zo) is the order dual E+ of E (the space of all real Radon
measures on X); under its canonical order, E+ is an order complete
vector lattice by (1.4), Corollary, and a I.c. v.I. for its strong topology
P(E+, E) «7.4) below). Of particular interest are the spaces E = lC(X)
when X is compact (Section 8).
We now supplement the results on the order topology Zo obtained in the
previous section.
7.3
Let E be a regularly ordered vector lattice. Then the order topology Zo is
the finest topology Z on E such that (E, Z) is a l.c.v.l. Moreover, if E is order
complete, then (E, Zo) is barreled, and every band decomposition of E is a
topological direct sum for Zoo
Proof. In view of (6.1) (and E+ = E b, (1.4», the regularity of the order of
E is sufficient (and necessary, cf. Exercise 19) for Zo to be a Hausdorff
topology. By (6.3), (E, Zo) is the inductive limit of the normed spaces La
(a ~ 0) that are normed lattices in the present circumstances; one shows, as
in the proof of (6.3), Corollary 2, that the convex circled hull of any family
{Pa[ -a, a]: a ~ O} is solid, and hence that (E, Zo) is locally solid. The fact
that Zo is the finest topology Z such that (E, Z) is a I.c.v.I. then follows from
(6.4) (b), since the positive cone is normal for all these topologies, (7.1).
If E is order complete, then clearly every positive sequence of type II is
order summable; hence (E, Zo) is barreled by (6.3), Corollary 1. The last
assertion is clear from the corollary of (6.1), since Zo induces on each band
B c: E the order topology of B. (Exercise 12.)
COROLLARY 1. If the order of the vector lattice E is regular, then (E, Zo) is
a l.c.v.l. whose topology is generated by the family ofall lattice semi-norms on E.
From the corollary of (6.4), we obtain:
§7] TOPOLOGICAL VECTOR LATIICES 237
COROLLARY 2. If E is a vector lattice and an ordered (F)-space in which the

positive cone is normal, the lattice operations are continuous in E.
It is interesting that the strong dual of a I.c.v.I. E reflects the properties of
E in a strengthened form; in addition, Ep is complete when E is barreled. (As
has been pointed out in Chapter IV, Section 6, the strong dual of a barreled
I.c.s. is in general not complete.)
7.4
Theorem. Let E be a l.c.v.l. Then the strong dual Ep is an order complete
I.c.v.l. under its canonical order, and a solid subspace of E+ ; moreover, if E is
barreled, then E' is a band in E+ , and Ep is a complete l.c.s.
Proof. Since the positive cone C of E is normal in (E, ~) by (7.2), it follows

from (3.3) that E' = C' - C' c C* - C* = E+.
It follows from the corollary of (1.5) that the polar UO of every solid
O-neighborhood U in E is a solid subset of E+. Since E' is the union of these
polars,· as U runs through a base of solid O-neighborhoods, E' is a solid
subspace and therefore a sublattice of E+. In particular, it follows that E'
is an order complete sublattice of E+. To see that E' is a I.c.v.I. for the strong
topology peE', E), it suffices to observe that the family of all solid bounded
subsets of E is a fundamental family of bounded sets; by the corollary of
(1.5) the polars BO (with respect to <E, E'» ofthese sets B form a O-neighbor-
hood base for peE', E) that consists of solid subsets of E'.
If (E,~) is barreled and S is a directed (~) subset of the dual cone C'
such that Sis majorized in E+, then each section of S is bounded for u(E+ , E),
hence for u(E', E) and, consequently, u(E', E)-relatively compact, (IV, 5.2).
Thus the section filter of S converges weakly to some f E C', and it is clear
from the definition of the order of E' thatf = sup S (cf. (4.2), which is, how-
ever, not needed for the conclusion). Since we have shown before that E'
is a solid sublattice of E+, it is now clear that E' is a band in E+ .
There remains to show that if (E,~) is barreled, then (E', P(E'E» is
complete. Let us note first that E+, which is the dual of (E, ~o) by (6.1) (note
that E+ = Eb by (1.4», is complete under P(E+, E) by (IV, 6.1), for it is the
strong dual of a bornological space. Hence by the preceding results and
(7.3), (E+, P(E+, E» is a I.c.v.I. (7.2) shows that E', being a band in E+, is
closed in (E+, P(E+ ,E» and hence complete for the topology induced by
P(E+, E). On the other hand, this latter topology is coarser than peE', E),
since ~ is coarser than ~o; hence if tJ is a peE', E)-·Cauchy filter in E', tJ has
a unique P(E+, E)-limit g E 'E'. Clearly, tJ converges to g pointwise on E,
and (since tJ is a Cauchy filter for peE', E», it follows from a simple argument
that lim tv = g for P(E', E). This completes the proof.
CoiOLLARY 1. Every reflexive locally convex vector lattice is order complete,

and a complete I.c.s.
In fact, the strong dual of E is a I.c.v.I. which is reflexive by (IV, 5.6),

Corollary 1, and hence barreled, and E can be identified (under evaluation)
with the strong dual of E/!. More generally, if E is a I.c.v.I. that is semi-
reflexive, then E is order complete and (E, {J(E, E'» is complete (cf. Corollary
2 of (7.5) below).
COROLLARY 2. If E is a normed lattice, its strong dual E' is a Banach lattice
with respect to dual norm and canonical order. If, in addition, E is a Banach space
then E' = E+.
Proof. The first assertion is clear, since the unit ball of E' is solid by the
corollary of (1.5). The second assertion is a consequence of (5.5) and (7.2).
The following result is the topological counterpart of (1.6).
COROLLARY 3. If E is an infrabarreled l.c.v.l., then E can be identified, under
evaluation, with a topological vector sublattice of its strong bidual EN (which is
an order complete l.c.v.l. under its canonical order).
Proof. The assumption that E is infrabarreled (Chapter IV, Section 5)

means- precisely that the evaluation map x ~ x is a homeomorphism of E
into EN; the remainder follows from (1.6), since E' is a solid subspace of E+.
It would, however, be a grave error to infer from the foregoing corollary
that for 'an infinite subset SC E such that x = sup S exists in E, one has
necessarily x = sup S. Thus even if E is order complete, E can, in general,
not be identified (under evaluation) with an order complete sublattice of EN.
For example, let E = I'" be endowed with its usual norm and order; E is an
order complete Banach lattice (in fact, E can be identified with the strong
dual of the Banach lattice [1). Denote by Xn (n eN) the vector in E whose n
first coordinates are 1, the remaining ones being 0; {xn: n e N} is a monotone
sequence in E such that supn Xn = e, where e = (1, 1, 1, ... ). Let z = supn xn
in EN (= E+ + by virtue of(5.5»; we assert that Z:F e. In view of E' = C' - C',
{xn} is a weak Cauchy sequence in E and z(f) = sUPn!(xn) for eachfe C';
if we had z = e, the sequence {xn} would be weakly convergent to e in E,
and hence norm convergent by (4.3). On the other hand, one has Ilxn+ p - xnll
= 1 for all n e N, peN, which is contradictory, and it follows that z < e.
Our next objective is a characterization of those I.c. vector lattices that
can be identified (under evaluation) with order complete sublattices of their
bidual EN; this will yield, in particular, a characterization of order complete
vector lattices of minimal type (Section 1). A filter iY in an order complete
vector lattice is called order convergent if iY contains an order bounded set Y
(hence an order interval), and if
supy(inf Y) = infy(sup Y),
where Y runs through all order bounded sets Ye iY. The common value of
the right- and left-hand terms is called the order limit of iY. Let us note also
§7l TOPOLOGICAL VECTOR LATTICES 239
that if E is a l.c.v.l., then the bidual E" of E is a I.c.v.l. under its natural topo-
logy (the topology of uniform convergence on the equicontinuous subsets
of E', Chapter IV, Section 5); in fact, the polar of every solid O-neighborhood
in E is a solid subset of E' by the corollary of (1.5), and hence the family of
all solid equicontinuous subsets of E' is a fundamental family of equicon-
tinuous sets. Hence their respective po lars (in E") form a O-neighborhood base
for the natural topology, consisting of solid sets.
7.5
Let (E, Z) be an order complete l.e.v.I., and let E" be endowed with its natural
topology and canonical order (under which it is an order complete I.e.v.l.). The
following assertions are equivalent:
(a) Under evaluation, E is isomorphic with an order complete sublattiee of E".
(b) For every majorized, directed (~) subset S of E, the section .filter of S
Z-eonverges to sup s.
(c) Every order convergent filter in E Z-eonverges to its order limit.
REMARK. The equivalences remain valid when "to sup S" and "to
its order limit" are dropped in (b) and (c), respectively; if the cor-
responding filters converge for Z, they converge automatically to the
lirn.its indicated, by (4.2).
Proof of (7.5). (a) =:> (b): Let S be a directed (~) subset of E such that
Xo = sup S; identifying E with its canonical image in E", we obtain (by
definition of the canonical order of E") f(x o) = sup{f(x): XES} for every
continuous, positive linear form on E. It follows that the section filter of S
converges weakly to xo, and hence for Z by (4.3), since the positive cone C
is normal in E.
(b) =:> (c): Let Z be an order convergent filter in E with order limit Xo and
let <V be the base of g: consisting of all order bounded subsets Y E g:. Let
a( Y) = inf Y (Y E (fj); the family {a( Y): Y E <V} is directed (~) with least
upper bound Xo; hence by hypothesis its section filter converges to Xo for Z.
Likewise, if b( Y) = sup Y, the family {b( Y): Y E (fj} is directed (~) with
greatest lower bound xo, and hence its section filter Z-converges to Xo. Let U
be any C-saturated O-neighborhood in E; there exists a set Yo E (fj such that
a( Yo) E Xo + U and b( Yo) E Xo + U, and this implies that Yo c Xo + U. Since
(C being normal) the family of all C-saturated O-neighborhoods is a base at
0, it follows that g: converges to Xo for Z.
(c) =:> (a): Let S be a directed (~) subset of E such that Xo = sup S. It is
clear that the section filter of S is order convergent with order limit xo, and
hence it Z-converges to Xo by assumption. It follows thatf(x o) = sup{f(x):
XES} for every f E C'; hence from the definition of order in E" it follows
that Xu = sup S, where x -> X is the evaluation map of E into E". The proof
is complete.
COROLLARY I. Let E be an order complete vector lattice whose order is regular.

The following assertions are equivalent:
(a) E is of minimal type.
(b) For every majorized, directed (;;;) subset S of E, the section filter con-
verges to sup S for ~o.
(c) Every order convergent filter in E converges for ~o.
Moreover, if E is minimal, then ~o is the finest I.c. topology on Efor which every
order convergent filter converges.
Proof Applying (7.5) to (E, ~o) we see that E' = E+, and E" =
(E+, f3(E+, E)), is a solid subspace of E+ + by (7.4). Hence E is minimal (that
is, isomorphic with an order complete sublattice of E+ + under evaluation) if
and only if E is isomorphic with an order complete sublattice of E", which
proves the first assertion.
For the second assertion there remains, in view of (c), only to show that
every I.c. topology ~ on E for which every order convergent filter converges
is coarser than ~o. Hence let ~ be such a topology, and let a E C be fixed.
Now {8[ -a, aJ: 8 > O} (8 E R) is a filter base in E, and it is immediate that
the corresponding filter is order convergent with order limit 0 (E is regular,
hence Archimedean ordered); thus if U is a convex O-neighborhood for ~,
it follows that there exists 8> 0 such that 8[ -a, aJ c U. Therefore, U
absorbs arbitrary order intervals in E, which shows that ~ is coarser than ~o.
COROLLARY 2. Let E be a I.c.v.l. which is semi-reflexive; then E is order
complete. If, in addition, every positive linear form on E is continuous, then E is
of minimal type, teE, E') = ~o, and (E, ~o) is reflexive.
Proof The first assertion follows at once from (4.3), Corollary 2. If every
positive linear form on E is continuous, then E' = E+, and the equality
teE, E') = ~o follows from (6.4) in view of the fact that the order of E is
regular, (4.1), Corollary 2. Hence E is minimal by Corollary I, for the section
filter of every majorized, directed (;;;) subset S of E converges weakly to sup S,
so it converges for ~o by (4.3). Finally, (E, ~o) is reflexive, since it is semi-
reflexive and (by (7.3)) barreled.
Examples
4. Each of the Banach lattices LP(J1), I < p < + 00 (Chapter II,
Section 2, Example 2; take K = R) is order complete and of minimal
type; in particular, the norm topology is the finest I.c. topology for
which every order convergent filter converges.
5. The Banach lattice L 1 (Jl) is order complete and of minimal type.
In fact, if S is a directed (;;;) subset of the positive cone C and majorized
by h, then for any subset {fl' ... ,fn} of S such that fl ;;; ... ;;;fn one
obtains
Ilh -fill = Ilh - Inll + II In - In -111 + ... + II f 2 -fill,
§7] TOPOLOGICAL VECTOR LATIICES 241
since the norm of L1(Jl') is additive on C; this shows that the section
filter of S is a Cauchy filter for the norm topology, and hence convergent.
Since the latter topology is 1:0 , it follows that L1(Jl') is of minimal type.
Obviously these conclusions apply to any Banach lattice whose norm is
additive on the positive cone; these lattices are called abstract (L)-spaces
(cf. Kakutani [1] and Section 8 below).
6. Suppose Jl to be totally a-finite. As strong duals of L 1 (Jl), the spaces
L oo(Jl) are order complete Banach lattices by (7.4); in general, these
spaces are not of minimal type as the example preceding (7.5) shows
and hence (in contrast with L1(Jl», in general, not bands in their res.pec-
tive order biduals.
7. Each perfect space (Example 2 above) is order complete and, if
each order interval is 0' (A, A+) - compact, of minimal type.
As we have observed earlier, ordered vector spaces possessing an order
unit are comparatively rare; it will be shown in Section 8 below that every
Banach lattice with an order unit is isomorphic (as an ordered LV.s.) with
CC R(X) for a suitable compact space X. A weaker notion that can act as a
substitute was introduced by Freudenthal [l]; an element x ;;; of a vector
°
lattice L is called a weak order unit if inf(x, Iyj) = implies y = for each
°°
°
y E L. A corresponding topological notion is the following: If L is an ordered
t.v.s., an element x;;; is called a quasi-interior point of the positive cone C
of L if the order interval [0, x] is a total subset of L. The remainder of this
section is devoted to some results on weak order units and their relationship
with quasi-interior points of C.
7.6
Let E be an ordered I.e.s. over R whieh is metrizable and separable, and suppose
that the positive cone C of E is a complete, total subset of E. Then the set Q of
quasi-interior points of C is dense in C.
Proof Since C is separable, there exists a subset {x n : n E N} which is dense
in C; denote by {Pn: n E N} an increasing sequence of semi-norms that
generate the topology of E. Since C is complete, Xo = L rnxn/pixn) is an
00
1
element of C. Now the linear hull of [0, xo] contains each Xn (n EN), and
hence is dense in C - C and, therefore, in E; that is, Xo E Q. It is obvious
that C1 = {OJ u Q is a subcone of C, and that Q = C1 • Suppose that Q '# C.
°
There exists, by (II, 9.2), a linear formfE E' such thatf(x) ;;; when x E Q,
and a point y E C such thatf(y) = -1. Consequently, there exists A> Osuch
that f( Xo + AY) < 0, which conflicts with Xo + AY E Q.
COROLLARY. Let E be a Frechet lattice which is separable. Then the set of weak
order units is dense in the positive cone of E.
Proof It suffices to show that each quasi-interior point of C is a weak order
unit. But if x is quasi-interior to C, then Y .1 x implies that y is disjoint from
the linear hull of [O,,x] which is dense in E; hence y =0 since the lattice
operations are continuous.
REMARK. The assumptions that E be metrizable and separable are not
dispensable in (7.6); if E is, for example, either the l.c. direct sum of
infinitely many copies of Ro or the Hilbert direct sum of uncountably
many copies of 12 (under their respective canonical orderings), then the
set of quasi-interior points of C (equivalently, by (7.7), the set of weak
order units) is empty.
7.7
Let E be an order complete vector lattice of minimal type. For each x > 0,
the following assertions are equivalent:
(a) x is weak order unit.
(b) For each positive linear formf#- 0 on E,f(x) > O.
(c) For each topology :t on E such that (E, :t) is a l.c.v.I., x is a quasi-interior
point of the positive cone.
Proof If Bx denotes the band in E generated by {x}, then x is a weak order
unit if and only if Bx = E, by virtue of (1.3). Now iffis a positive linear form
on E, then, since E is minimal, f(x) = 0 is equivalent with f(Bx) = {O}; this
shows that (a) ¢> (b). Moreover, E being minimal, Bx = E is equivalent with
the assertion that the linear hull of [0, x] is dense in (E, :to) (for the closure
of each solid subspace G in (E, :to) contains the band generated by G);
hence (a) ~ (c), since the topologies mentioned in (c) are necessarily coarser
than :to by (7.3). (c) ~ (a) is clear in view of the continuity of the lattice
operations in (E, :t) (cf. proof of (7.6), Corollary).
For example, in the spaces LP(/1) (I ~ p < + (0) the weak order units
( = quasi-interior points of C) are those classes containing a function which
is > 0 a.e. (/1). By contrast, a point in L 00(/1) is quasi-interior to C exactly
when it is interior to C; the classes containing a function which is > 0 a.e.
(/1) are weak order units, but not necessarily quasi-interior to C. Hence the
minimality assumption is not dispensable in (7.7).
8. CONTINUOUS FUNCTIONS ON A COMPACT SPACE. THEOREMS

OF STONE-WEIERSTRASS AND KAKUTANI
This final section is devoted to several theorems on Banach lattices of
type ~(X), where X is a compact space, in particular, the order theoretic and
algebraic versions of the Stone-Weierstrass theorem and representation
theorems for (AM)-spaces with unit and for (AL)-spaces. For a detailed
account of this circle of ideas, which is closely related to the Krein-Milman
theorem, we refer to Day [2]; the present section is mainly intended to serve
as an illustration for the general theory of ordered vector spaces and lattices
developed earlier. Let us point out that with only minor modifications most of
§a] CONTINUOUS FUNCTIONS ON A COMPACT SPACE 243
the following results are applicable to spaces CCo(X) (continuous functions on

a locally compact space X that vanish at infinity); for CCo(X) can be viewed as
a solid sublattice of codimension I in CC(X), where X denotes the one-point
compactification of X.
With one exception (see (8.3) below) we consider in this section only vector
spaces over the real field R; mutatis mutandis, many of the results can be
generalized without difficulty to the complex case, since CC(X) over C is the
complexification (Chapter I, Section 7) of CC(X) over R; (8.3) is an example
for this type of generalization. To avoid ambiguity we shall denote by CCR(X)
the- Banach lattice of real-valued continuous functions on X, and by CCc(X)
the (B)-space of complex-valued continuous functions on x.
Let us recall some elementary facts on the Banach lattice CCR( X), where
X# 0 is any compact space. CCR(X) possesses order units; IE CCR(X) is an
order unit if and only if inf {f(t): t EX} > O. Thus the order units of CCR(X)
are -exactly the functions I that are interior to the positive cone C. Distin-
guished among these is the constantly-one function e; in fact, the norm
1-+ 11/11 = sup{i/(t)!: t E X} is the gauge functionpe of[ -e, e] and, of course,
the topology of CCR(X) is the order topology Zo (Section 6). We l?egin with the
following classical result, the order theoretic form of the Stone-Weierstrass
theorem.
8.1
Theorem. II F is a vector sublattice 01 C(I R(X) that contains e and separates
points in X, then F is dense in CCR(X),
REMARK. The subsequent proof will show that a subset Fe C(I R(X)
is dense if it satisfies the following condition: F is a (not necessarily lin-
ear) sublattice of the lattice C(I R(X), and for every B > 0 and quadruple
(s, t;ex, 13) E X2 x R2 such that ex = 13 whenever s = t, there exists
IE F satisfying I/(s) - exl < e and I/(t) - PI < B•
. Prool 01 (8.1). Let s, t be given points of X and ex,p given real numbers
such that ex = 13 if s = t; the hypothesis implies the existence of IE F such
that/(s) = ex,f(t) = p. This is clear if s = t, since e E F; if s "# t, there exists
9 E F such that g(s) "# g(t), and a suitable linear combination of e and 9 will
satisfy the requirement.
Now let h E C(IR(X) and e > 0 be preassigned and let s be any fixed element
of X. Then for each t E X, there exists an It E F such that Ires) = h(s) and
fret) = h(t). The set Ut = {r E X:/,(r) > h(r) - e} is open and contains t;
hence X = U Ut, and the compactness of X implies the existence of a finite
teX n
set {tl' ... , tn} such that X =UU tv ' Using the lattice property of F, form the
v=l
function g. = SUP{/'l' ... ,/,J; it is clear that g.(t) > h(t) - e for all t EX,
since each t is contained in at least one Utv ' Moreover, g.(s) = h(s).
Now consider this procedure applied to each SEX; we obtain a family
{gs: SEX} in F such that gs(s) = h(s) for all SEX, and gs(t) > h(t) - e for
all t E X and SEX. The set Vs = {r EX: gs(r) < her) + e} is open and con-
tains s; hence X = U V., and the compactness of X implies the existence
seX m
of a finite set {Sl' ... , sm} such that X = U Vs... Let g = inf {gSl' ... , gsJ;
1t=1
then g E F and her) - e < g(r) < her) + e for all rEX; hence Ilh - gil < e,
and the proof is complete.
The algebraic form of the Stone-Weierstrass theorem replaces the hypo-
thesis that F be a sublattice of 'tf R(X) by assuming that F be a subalgebra
(that is, a subspace of 'tf R(X) invariant under mUltiplication). Our proof
follows de Branges [1], but does not involve Borel measures. The proof is an
interesting application of the Krein-Milman theorem and provides an oppor-
tunity to apply the concept of Radon measure that has been utilized earlier
(Chapter IV, Sections 9 and 10).
The space Jt R(X) of (real) Radon measures on X is, by definition, the dual
of 'tfR(X) (Chapter II, Section 2, Example 3); since E is a Banach lattice,
Jt R(X) is a Banach lattice under its dual norm and canonical order by (7.4),
Corollary 2. Thus 111111 = 11111111 for each 11; if 11 ~ 0, then 111111 = sup{ll(f):
fE [-e, en = Il(e), and this implies that 111111 = 11+ (e) + Il-(e) for all
11 E Jt R(X), If g E 'tf R(X) is fixed, then f ~ gf (pointwise multiplication) is a
continuous linear map u of 'tfR( X) into itself; the image of 11 E Jt R( X) under
the adjoint u' is a Radon measure denoted by g .11· Obviously Ig .111 ~ Ilg 111111
and hence u' leaves each band in Jt R(X) invariant; in particular, if g ~ 0,
then g.1l = g ·11+ - g ·11-, where inf(g .11+, g .11-) = 0. It follows from (1.1)
that (g .11)+ = g ·11+, (g ·11)- = g ·11- in this case, and that Ig .111 = g .1111.
The support of fE 'tfR(X) is the closure Sf of {t E X:f(t) #- O} in X; we
define the support Sit of 11 E Jt R( X) to be the complement (in X) of the largest
°
open set U such that Sf C U implies Il(f) = (equivalently, such that Sf C U
° ° °
implies 11l1(f) = 0). An application of Urysohn's theorem (cf. Prerequisites)
shows that iff~ and 11 ~ 0, then Il(f) = if and only iff(t) = whenever
t E Sit" Notice a particular consequence of this: if 11 is such that Sit = {to},
then 11 is of the form Il(f) = Il(e)f(to) (hence, up to a factor Il(e) #- 0, evalua-
tion at to). For Sit = {to} implies that 11l(f - f(to)e) I ~ 11l1(lf - f(to)e\) = 0,
°
which is the assertion. Finally, 11 = if and only if Sit = 0. The following
lemma is now the key to the proof of (8.2).
LEMMA. Let F be a subspace of'tfR(X), and suppose that the Radon measure 11
is an extreme point of FO n [- e, e]O C Jt R( X). If g E 'tf R( X) is such that
g.1l E FO, then g is constant on Sit"
Proof If 11 = 0, there is nothing to prove. Otherwise, it can be arranged
(by adding a suitable scalar multiple of e and subsequent normalization)
°
that g ~ and 11l1(g) = 1. Suppose, for the moment, that g ~ e; since 11 is an
extreme point of po n [-e, e]O, we have 111111 = 1 and it follows that 11l1(e - g)
= 111111 - 11l1(g) = 0, which implies that 1 = eel) = get) for all t E SJt' in view
§8] CONTINUOUS FUNCTIONS ON A COMPACT SPACE 245
of the remarks preceding the lemma. We complete the proof by showing that
IIg II > 1 is impossible. In fact, assume that Ilg II > 1, let P= IIg 11- 1, and define
the Radon measures J-t1' J-t2 by J-t1 = g1' J-t and J-t2 = g. J-t, where g1 = (e - Pg)1
(1 - Pl. We observe that J-t1 E PO, J-t2 E Po and that 1J-t21 = 9 .IJ-tl; hence 1IJ-t211
= 9 ·IJ-tl(e) = 1. Moreover, J-tt = g1' J-t+ and J-tl = g1 • J-t-, since 91 ~ 0, and
in view of 1IJ-t111 = J-tt(e) + J-tl(e), it follows from a short computation that
IIJl111 = 1. On the other hand, it is easy to see that J-t = (1 - fJ)J-t1 + PJ-t2, which
conflicts with the hypothesis that J-t be an extreme point of FO f"'I [-e, er.
The following is the algebraic form of the Stone-Weierstrass theorem.
8.2
Theorem. If F is a subalgebra of ~ R(X) that contains e and separates
points in X, then Fis dense in ~ R(X),
Proof The set F C f"'I [-e, e]O is a convex, circled, weakly compact subset of
J( R(X); hence by the Krein-Milman theorem (II, lOA) there exists an extreme
point J-t of Po f"'I [-e, e]o. Since Fis a subalgebraof~R(X)' eachfE Fsatisfies
the hypothesis of the lemma with respect to Jl; hence each f E F is constant on
the support SIt of J-t. This is clearly impossible if SI' contains at least two points;
since F separates points in X; on the other hand, if SIt = {to}, then Jl(f)
= J-t(e)f(to), and it follows that eachfE Fvanishes at to, which is impossible
since e E F. Hence SI' is empty which implies Jl = 0 and, therefore, Po = {O};
consequently, F is dense in ~R(X) by the bipolar theorem (IV, 1.5).
The preceding theorem is essentially a theorem on real algebras ~(X); for
instance, if X is the unit disk in the complex plane and F is the algebra of all
complex polynomials (restricted to X), then F separates points in X and e E F,
but F is not dense in ~c(X) (for each f E F is holomorphic in the interior of
X). One can, nevertheless, derive results for the complex case from (8.1)
and (8.2) by making appeal to the fact that ~c(X) is the complexification of
~R(X); we say that a subset F of the complex algebra ~c( X) is conjugation-
invariant if fE F implies f* E F (where f*(t) = f(t)*, t E X). We consider
~c(X) as ordered by the cone of real functions ~ 0 (Section 2).
8.3
COMPLEX STONE-WEIERSTRASS THEOREM. Let F be a vector subspace of the
complex Banach space ~c(X) such that e E F and F separates points in X and is
conjugation-invariant. Then either of the following assumptions implies that F
is dense in ~c(X) :
(i) F is lattice ordered (Section 2)
(ii) F is a subalgebra of~c(X).
Proof If F1 denotes the subset of F whose elements are the real-valued

functions contained in F, then F = F1 + iF1 by the conjugation-invariance of
the subspace F; clearly, e E F1 and F1 separates points in X, since F does.

Thus if F is lattice ordered, then F1 is a vector lattice (Section 2), and (8.1)
shows that F1 is dense in f(f R(X); by (8.2) the same conclusion holds if F is a
subalgebra of f(fc(X), for then F1 is a subalgebra of f(f R(X), This completes
the proof.
It is customary to call a Banach lattice E an (AL)-space (abstract L-space)
if the norm of E is additive on the positive cone C. The reason for this termi-
nology is that every L 1(/1,) (over R) possesses this property and that, con-
versely, every (AL)-space is isomorphic (as a Banach lattice) with a suitable
space L1(p) (Kakutani [1]; cf. Exercise 22). A Banach lattice E is called an
(AM)-space (abstract (m)-space) if the norm of E satisfies IIsup(x, y) II = sup
(lIxll, Ilyl!) for all x, y in the positive cone C; E is called an (AM)-space with
unit u if, in addition, there exists u E C such that [-u, u] is the unit ball of E.
(Clearly, such u is unique and an order unit of E.) It is immediate that every
Banach lattice f(fR(X) is an (AM)-space with unit (the unit being the con-
stantly-one function e); we will show that this property characterizes the
spaces f(f(X) over R among Banach lattices. More generally, every (AM)-space
is isomorphic with a closed vector sublattice of a suitable f(fR(X) (Kakutani
[2]). Let us record first the following elementary facts on (AL)- and (AM)-
spaces; by the strong dual of a Banach lattice E, we understand the dual E'
(= E+) under its natural norm and canonical order.
8,4
The strong dual of an (AM)-space with unit is an (AL)-space, and the strong
dual of an (AL)-space is an (AM)-space with unit. Moreover, if E is an Archi-
medean ordered vector lattice, u an order unit of E, and p u the gauge function of
[-u, u], then the completion of(E,pu) is an (AM)-space with unit u.
Proof Let E be an (AM)-space with unit u; the strong dual E' is a Banach
lattice by (7.4), Corollary 2. If x' E C' then Ilx'll = sup{l(x, x')I: x E [-u, u])
= (u, x'); hence the norm of E' is additive on the dual cone C'.
If F is an (AL)-space, the norm of F is an additive, positive homogeneous
real function on C, and hence defines a (unique) linear formflon F such that
fo(x) = IIxll for all x E C; evidently we have 0 ~fo E F'. It follows that
g E F' satisfies IIg II ~ 1 if and only if g E [ - fo, fo], and hence the norm of the
strong dual F' is the gauge function of [-fo'/o]. Now if g ~ 0, h ~ 0 are
elements of F' such that Ilg II = A1' Ilh II = A2' then g ~ A do and h ~ Ado,
since the order of F' is Archimedean. Consequently, Ilsup(g, h) II ~ SUp(A1' A2),
and here equality must hold or else both the relations Ilg II = A1 , Ilh II = A2
could not be valid. Therefore, under its canonical order, F' is an (AM)-
space with unitfo.
To prove the third assertion, we observe that if E is an Archimedean
ordered vector lattice and u is an order unit of E, then Pu is a norm on E,
and even a lattice norm, since [- u, u] is clearly solid. The completion (E, Pu)
of (E, Pu) is a Banach lattice (with respect to the continuous extension of the
lattice operations) whose unit ball is the set {x E E: -feu) ~f(x) ~f(u),
fE C}, and hence the order interval [-u, u] in E. As in the preceding para-
graph, it follows that (£, Pu) is an (AM)-space with unit u. This ends the proof.
Let E 1= {O} be an (AM)-space with unit u; the intersection of the hyper-
plane H = {x': (u, x') = I} with the dual cone C' is a convex, aCE', E)-closed
subset Ho of the dual unit ball [-u, ur.1t follows that H o, which is called the
positive face of [-u, u]O, is aCE', E)-compact; hence C' is a cone with weakly
compact base, and t E Ho is an extreme point of Ho if and only if {At: A ~ O}
is an extreme ray of C (Chapter II, Exercise 30).
Now we can prove the representation theorem of Kakutani [2] for (AM)-
spaces with unit.
8.5
Theorem. Let E 1= {O} be an (AM)-space with unit and let X be the set of
extreme points of the positive face of the dual unit ball. Then X is non-empty and
aCE', E)-compact, and the evaluation map x ~ f (where f(t) = (x, t), t EX)
is an isomorphism of the (AM)-space E onto Y&' H( X).
Proof. Let u be the unit of E. Since the positive face Ho of [-u, u]O is convex
and aCE', E)-compact, the Krein-Milman theorem (II, 10.4) implies that the
set X of extreme points of Ho is non-empty. Since Ho is a base of C', it
follows from (1.7) that t E X if and only if t is a lattice homomorphism of E
onto R such that t(u) = 1. It is clear from this that X is closed, hence compact
for aCE', E). The mapping x ~ f is clearly a linear map of E into Y&' H(X) that
preserves the lattice operations, since each tEXis a lattice homomorphism;
to show that x ~ f is a norm isomorphism, it suffices (since E and Y&'H( X) are
Banach lattices) that Ilfll = Ilxll when x;;;; O. For x ~ 0 we have Ilxll
= sup{(x, x'): Ilx'll ~ I} = sup{(x, x'): x' E Ho}; since Ho is the aCE', E)-
closed convex hull of X and each x E E is linear and aCE', E)-continuous, it
follows that (x;;;; 0) sup{ (x, x'): x' E Ho} = sup{ (x, t): t EX} = Ilfll. Thus
x ~ f is an isomorphism of E onto a vector sublattice F of Y&'H( X) that is
complete and contains e (the image of u); since E separates points in E' and
a fortiori in X, it follows from (8.1) that F = Y&'H( X), which completes the
proof.
We conclude this section with two applications of the preceding result;
the first of these gives us some more information on the structure of (AL)-
spaces, the second on more general locally convex vector lattices.
From (8.4) we know that the strong dual E'( = E+) of an (AL)-space E is
an (AM)-space with unit; hence by (8.5), E' can be identified with a space
Y&'H(X), where X is the set of extreme points of the positive face of the unit
ball in E". By (7.4), the Banach lattice E' is order complete, which has the
interesting consequence that X is extremally disconnected (that is, the closure
of every open set in X is open). In fact, let G c X be open and denote by S
the family of all IE rc R( X) such that I E [0, e] and the support SJ is contained
in G. S is directed (~) and majorized bye; hence 10 = sup S exists. Since G
is open, it follows from Urysohn's theorem that lo(s) = 1 whenever s E G,
°
and that lo(t) = whenever t ¢: G. Thus 10 is necessarily the characteristic
function of G since 10 is continuous, and this implies that G is open.
Therefore, if E is an (AL)-space, then E' can be identified with a space
rc R(X), where X is compact and extremally disconnected, and it follows that E
itself can be identified with a closed subspace of the Banach lattice J!{ R(X)
which is the strong bidual of E. For a characterization of E within J!{ R(X),
let us consider the subset B c J!{ R(X) such that Jl E B if and only if for each
directed (~), majorized subset S c rcR(X) it is true that lim Jl(f) = Jl(sup S),
the limit being taken along the section filter of S. It is not difficult to verify
°
that B is a vector sublattice of J!{ R( X); in fact, if S is directed (~) and
10 = sup S, and if 10 ~ (which is no restriction of generality) then there
exists, for given Jl E Band 6> 0, a decomposition 10 = go + ho(go ~ 0,
ho ~ 0) such that /1+(h o) < 6 and /1-(go) < 6 «1.5), formula (7». Using that
Jl E B, we obtain after a short computation that /1+(fo) < sup{Jl+(f): I E S}
+ 36, which proves that Jl+ E B. Thus B is a sublattice of J!{ R(X) which is
clearly solid; it is another straightforward matter to prove that B is a band
in J!{ R(X), The only assertion in the following representation theorem that
remains to be proved is the assertion that B = E.
8.6
Theorem. Let E be an (AL)-space. The Banach lattice E' (= E+) can be
identified with rc R(X), where X is a compact, extremally disconnected space.
Moreover, under evaluation, E is isomorphic with the band 01 all (real) Radon
measures Jl on X such that
lim Jl(f) = /1(sup S)
JeS
lor every majorized, directed (~) subset S olrcR(X).

Proof. It is easy to see that (identifying E' with rc R(X) and E with its canoni-
cal image in E" = J!{ R(X» we have E c B (see the preceding paragraph for
notation). For if S is majorized and directed (~), every section of S is
aCE', E)-bounded and hence the section filter aCE', E)-converges to sup S;
the assertion follows since Jl E E is aCE', E)-continuous.
°
To prove the reverse inclusion, let ~ v E B and let Jlo = sup [0, v] (") E.
Then the section filter of [0, v] (") E is a Cauchy filter for the norm topology,
since E is an (AL)-space (Section 7, Example 5), and hence Jlo E E, since E is
norm complete. Now Jll = V - Jlo is an element of B lattice disjoint from E;
it will be shown that this implies /11 = 0, and hence B = E by (1.3).
Denote by Tl the support of Jl1 ; if T = X ~ T 1 , then T is open and /11 (f) =
for each I whose support TJ is contained in T. The family of all IE [0, e]
°
such that TJ c T is directed, and its least upper bound 10 is necessarily the
characteristic function of the closure T. Since Pl E B, it follows that Pl ([0) = 0;

hence T ("'\ Tl = 0, which shows that Tl is open and closed. If Tl = 0 the
proof is complete; hence assume that Tl is non-empty. Since Tl is open, there
exist elements pEE whose support intersects Tl (otherwise E would not
distinguish points in E' = CCR(X», There exists, consequently, a positive pEE
such that Ilpll = I and whose support is contained in Tl (it suffices to take a
positive lEE for which l(go) > 0, where go = e- fo, and to consider go. l).
The proof will now be completed by showing that this last statement is false.
Let 8n = rn (n EN). By formula (7) of(1.5) there exist (since inf(p, Pl) = 0)
decompositions go =J" + f:, wherefn ~ 0,/: ~ 0 and such that Pl(J,,) < 8~,
p(f:) < 8;, so that p(J,,) > I - 8; (n EN). Let Gn = {t:J,,(t) > 8n} for all n,
then Gn is open and Gn is closed and open. If we write p(A) in place of P(XA.)
whenever A c: X is a subset whose characteristic function XA. is continuous,
we obtain Pl(Gn) < 8n; in fact, Pl(Gn) ~ 8n would imply that Pl([ri) ~
8n Pl(Gn) ~ 8;, which is contradictory. Now let Hk = U {Gn : n ~ k + I}; then
Hk is closed and open, and it follows from Pl E B that Pl(Hk ) .<8k' since the
characteristic function of Hk is the least upper bound of the characteristic
functions of the sets Gn(n ~ k + 1). Now 'define gn by gn = sup{f,,: v ~ n}
(n EN); then {gn: n EN} is a monotone (~) sequence; let h = inf{gn: n EN}.
In the complement of Rk one has f.(t) ;:;i 8. whenever v ~ k + 1, and hence
gn(t) ;:;i 8n whenever n ~ k + 1; in view of Pl(Hk ) < 8k' it is clear that Pl(h)
;:;i 8k' This implies Pl(h) = 0 and thus h = 0, for the support of h is contained
in the support Tl of Pl' On the other hand, since pEE c: B, we have
limn p(gn) = p(h) = 0, which conflicts with p(/,.) > 1 - 8;, since 0 ;:;i/,. ;:;i gn
for all n. This completes the proof of (8.6).
COROLLARY 1. In an (AL)-space E each order interval is weakly compact.
Proof. Since E is a band in E"( = E+ +), E is a solid subspace of E"; thus

if x, y E E, we have [x, y] = (x + C) ("'\ (y ..., C) = (x + C) ("'\ (y - C") where
C, C" denote the positive cones of E, E" respectively. Since C" is a(E", E')-
closed, it follows that [x, y] is aCE", E')-closed and hence a(E", E')-compact.
COROLLARY 2. Every (AL)-space E is an order complete vector lattice of mini-

mal type; by contrast, its order dual E+ is not of minimal type, unless E is of
finite dimerzsion.
Proof. Since E can be identified with a band in E" = E+ +, it is clearly of

minimal type (Section 7). If, on the other hand, E+ (which can be identified
with CCR(X» is of minimal type, then by (7.5), Corollary 1, the section filter of
each directed (;:;i), majorized set S converges to sup S pointwise (even uni-
formly) on X, which implies that each open subset of X is closed, and hence
that the topology of X is discrete. Since X is compact, X is finite, and hence
E+ and E are finite dimensional.
Our second application of (8.5) is the following result.
8.7
Let (E, ~) be a l.c.v.l. which is bornological and sequentially complete. There
exists afamily of compact spaces Xa(IX E A) and a family of vector lattice iso-
morphismsfa ofC(!R(Xa) into E (IX E A) such that ~ is the finest I.c. topology on E
for which eachla is continuous.
Proof. In view of(5.5) and (6.4), the assumption that (E,~) be bornological
implies that ~ is the order topology ~o. Hence by (6.3), (E, ~) is the inductive
00
limit of the subspaces (Ea, Pa) (IX E A) where Ea = U n[ -aa' aa], Pa is the
n=l
gauge of [ -aa' aa] on E a, and {aa: IX E A} is a directed subset of the positive
cone of E such thatUaEa = E. By (6.2) each (Pa, Ea) is a Banach lattice, and
by (8.4) even an (AM)-space with unit aa' Hence by (8.5), (Ea' Pa) can be
identified with C(!R(Xa) for a suitable compact space X a, and the assertion
follows from the definition of inductive topologies (Chapter II, Section 6).
EXERCISES
I. A reflexive, transitive binary relation" -< " on a set S is called a
pre-order on S. A pre-order on a vector space Lover R is said to be
compatible (with the vector structure of L) if x -< y implies x + z -< y + z
and Ax -< AY for all Z ELand all scalars A > O.
(a) If (X, :E, Jl) is a measure space (Chapter II, Section 2, Example 2),
the relation "f(t) ~ get) almost everywhere (Jl)" defines a compatible
pre-order on the vector space (over R) of all real-valued :E-measurable
functions on X.
(b) If" -< " is a compatible pre-order, the relation" x -< y and y -< x"
is an equivalence relation on L, the subset N of elements equivalent to 0
is a subspace of L, and LjN is an ordered vector space under the relation
" ~ ~ Y if there exist elements x E ~, Y E Y satisfying x -< y".
(c) The family of all compatible pre-orders of a vector space Lover R
is in one-to-one correspondence with the family of all convex cones in L
that contain their vertex O.
2. The family of all total vector orderings (total orderings satisfying
(LO) 1 and (LOh, Section I) of a vector space L is in one-to-one corres-
pondence with the family of all proper cones that are maximal (under
set inclusion). Deduce from this that for each vector ordering R of L,
there exists a total vector ordering of L that is coarser than R. (Use
Zorn's lemma.) Show that a total vector ordering cannot be Archi-
medean if the real dimension of L is > 1.
3. Let L be an ordered vector space with positive cone C. Let N be a
subspace of L, and denote by C the canonical image of C in Lj N.
(a) If N is C-saturated, then C defines the canonical order of LjN.
(b) If L is a t.v.S. and if for each O-neighborhood V in L there exists
a O-neighborhood U such that [(U + N) (1 C] C V + N, then C is
normal for the quotient topology. (Compare the proof of (3.1).)
EXERCISES 251
(c) If L is a topological vector lattice and N is a closed solid sublattice,

then L/ N is a topological vector lattice with respect to quotient topology
and canonical order (cf. Exercise 12 below). (Use (b) above, and (7.1).)
4. Consider the order of the complex space L = eN defined by the
cone C, where x = (xn) E C if and only if either x = 0, or Re Xn ~ 0
(n E N) and 1m Xl > o. Show that the dual cone C* c L * separates
points in L, but there exists no Hausdorff I.c. topology on L for which
C is normal.
5. Let {Ea: 0( E A} be a family of I.c.s., let 6 .. be a saturated family
of bounded subsets of Ea, and let Ca be an 6 a-cone in E..(O( E A). Show
that TI.. Ca is an 6 1-cone in TIaEa and that EBaCa is an 6 2 -cone in
EBaEa, where 6 1 and 6 2 denote the families TIa6a and EBa6a respect-
ively (Chapter IV, Section 4). Derive an analogous result for families
of strict 6 a-cones, and discuss the permanence properties of 6-cones
under the transition to subspaces and quotient spaces. (Use (IV, 4.1),
(3.3), and Exercise 3 above.)
6. Let!?) be the space of infinitely differentiable real functions on Rk
with compact support, and let C be the cone of non-negative functions
in !?). (Section 3, Example 2.)
(a) Show that C is a strict ID-cone which is not normal.
(b) Each positive linear form on !?) is continuous, and has a unique
extension which is a positive linear form, to the space of continuous
functions on Rk with compact support. Deduce from this that each
positive distribution defines a unique positive Radon measure on Rk.
(c) If S is a directed (~) set of distributions which is majorized, then
fo = sup Sexists andlimf = fo uniformly on every bounded subset of !?).
7. Let E, F be ordered I.c.s. with respective positive cones C, D and
suppose that F is quasi-complete; moreover, assume that C is normal
and that D is a strict ID-cone. Every nuclear map U E ft'(E, F) is of the
-form U = u l - U2' where U 1 , U2 are positive nuclear maps. Apply this
to the case where E is nuclear and F is a Banach space.
8. Denote by E a separable Banach space and suppose that {xn: n EN}
is a maximal topologically free subset of E. If C is the set of all linear
combinations of elements xn(n EN) with coefficients ~ 0, the following
assertions are equivalent:
(i) C is a normal ~-cone in E.
(ii) {xn: n EN} is an unconditional basis of E (Chapter III, Section 9).
(Use (3.5); see also Schaefer [2].)
9. (Dini's theorem). Let X be a locally compact space, and denote by
R(X) the vector space of all real-valued continuous functions on X
under the topology of compact convergence. If S is a directed (~)
subset of R(X) such that the numerical least upper boundfo of S is finite
and continuous on X, then the section filter of S converges tofo in R(X).
Deduce from this another proof of (4.3) by utilizing (4.4).
10. Let L be an ordered vector space over R with positive cone C.
(a) L possesses order units if and only if there exists a I.c. topology
on L for which C has non-empty interior; if so, each interior point of C
is an order unit of L.
(b) If(L,:t) is a t.v.s. such that C has interior points, then each quasi-
interior point of C is interior to C.
(c) If (L, :t) is a non-normable Hausdorff t.v.s. in which C is normal,
then L possesses no order units.
11. Let L be an ordered vector space over R with positive cone C.
If M is a subspace of L such that M + C = M - C, then every linear
formfo on M, positive for the canonical order of M, can be extended to
a positive linear form on L. (Day [2], §6 Theorem 1.)
12. Let L 1 , L2 be vector lattices.
(a) If N is a solid subspace of L 1 , the canonical image of the positive
cone of Ll in LdN defines the canonical order of Ll/N under which
Ld N is a vector lattice, and the canonical map Ll --+ Ld N is a lattice
homamorphism. If L is order camplete and N is a band in L 1 , then
LdN is isamorphic with Nl..
(b) If u is a linear map .of Ll anta L z , then u is a lattice homomor-
phism if and only if u(C1 ) = C z and N = u- 1 (0) is solid in L 1 , where
C 1 , C z denote the respective positive cones. In these circumstances, the
biunivocal map Uo associated with u is a lattice isomorphism of Ld N
onto L 2 • .
(c) If u is a vector lattice homomorphism of Ll onto L z, then u is.a
topological homomorphism for the respective order topologies of Ll
and L 2 • In particular, if N is a solid subspace of L 1 , then the order top-
ology of LdN is the quotient of the order topology of L 1 •
(d) If Ll is order complete, if N is a band in L 1 , and if:to is the order
topology ofL I' then the topology induced by:to is the order topology of N.
(e) Give an example of an order complete vector lattice L and a solid
subspace M of L such that:to (of L) does not induce the order topology
of M (cf. Remark preceding (6.4)).
13. Let (X,~, Jl) be a measure space (Chapter II, Section 2, Ex-
ample 2) such that ~ contains all fingletons but not all subsets of X,
and Jl is bounded. Under their ca~onical order, the real spaces !l'P(Jl)
are vector lattices that are countaWy order complete (each countable
majorized subset has a least upper bound), but not order complete
(1 ~p ~ + 00).
14. Denote by Jl Lebesgue measure on the real interval [0, 1] and let
L be the real space Y(Jl), where p is fixed, 0 < p < I.
(a) Under its canonical order, L is an order complete topological
vector lattice such that Lb = L + = {O}. (Use Chapter I, Exercise 6 and
Chapter Y, (5.5).)
(b) Infer from (a) that the order of Lis Archimedean but not regular,
that the positive cone is dense in L for every I.c. topology on L, and that
:to is the coarsest topology on L.
(c) There exist vector sublattices MeL such that M (1 C has non-
empty interior (in M), but no positive linear form on M can be extended
to a positive linear form on L.
(d) The two-dimensional subspace of L determined by the functions
t --+ at + b (a, b E R) is a vector lattice under the induced order, but not
a sublattice of L.
EXERCISES 253
15. Let X"# 0 be a set, let :E be a O'-algebra of subsets of X, and let E

denote the vector space (over R) of all real-valued bounded :E-measur-
able functions (that is, functions / such that r I(A) E:E for each Borel
set A c: R). Under the sup-norm and the canonical order, E is an
(AM)-space. Every positive linear form on E defines a real-valued, finitely
additive non-negative set function on :E, and conversely; hence E+
can be identified with the vector lattice of all real-valued, finitely additive
set functions on :E that are differences of non-negative functions of the
same type.
(a) E+ is order complete, and the set functions Il E E+ that are count-
ably additive on :E form a band Min E+. Deduce from this that every
finitely additive, non-negative set function v on :E has a unique represent-
ation v = VI + V2, where VI' V2 are ~ 0, VI is countably additive and V2
is not countably additive unless V2 = 0.
(b) Suppose that :E contains all finite (hence all countable) subsets of
°
X, and call the elements Il E M briefly measures on X. The measures fi
such that fi({I}) = for every I E X are called diffuse; the set of all dif-
fuse measures is a band Md in M. The elements of the complementary
band Ma in M are called atomic. Show that each atomic measure on X
is the sum (for O'(E+, E)) of a summable family {OCnll n: n EN} (Chapter
III, Exercise 23), where (oc n) E [I and each fin is a point measure on X
(that is, Il({t o}) = 1 for a suitable to E X and Il(A) = 0 for each A E:E
such that to ¢; A).
(c) Illustrate the preceding by considering the case, where X is
compact and :E is the family of all Baire subsets of X.
16. Let L be an ordered vector space over R. A monotone (non-de-
creasing) transfinite sequence is a mapping oc --+ aa of the set of all ordinals
oc < 13 (where 13 is an ordinal ~ 1) into L such that oc 1 < OC2 ( < 13) implies
aal ~ aa2' Suppose that the positive cone C generates L, and that C
satisfies condition (D) of (1.1).
(a) If each majorized monotone transfinite sequence in L possesses a
least upper bound, then L is an order complete vector lattice.
(b) If there exists a linear form/on L such that x> 0 implies/ex) > 0,
for L to be an order complete vector lattice it suffices that each ordinary
monotone sequence which is majorized, has a least upper bound.
(Schaefer [4].)
17. Let L be a vector lattice which is order complete, and denote by
:! the finest topology on L such that every order convergent filter in
L :!-converges to its order limit. Show that:! is a translation-invariant
topology that possesses a O-neighborhood base of radial and circled
sets but fails, in general, to satisfy axiom (LT)1 (Chapter I, Section 1).
(Show that everyone-point set is closed, but:! is not necessarily Haus-
dorff.)
18. Let E, F be vector lattices and suppose that F is order complete.
Denote by H c: L(E, F) the cone of all positive linear maps of E into F.
The subspace M = H - H of L(E, F) is an order complete vector lattice
under its canonical order, containing exactly those linear maps that map
al1 order intervals of E into order intervals of F.
19. Recall that the order of a vector space L is called Archimedean

if x:;;;; 0 whenever x:;;;; n-1y for all n EN and some y E L. The order of L
is called almost Archimedean if x = 0 whenever - n - 1Y :;;;; X :;;;; n -1 y for
all n E N and some y E L.
(a) If L is almost Archimedean ordered and possesses an order unit,
then (L, '1: 0) is normable and (L, '1: 0)' = Lb = L +.
(b) If L is almost Archimedean ordered and Lb distinguishes points
in L, (L, '1: 0) can be characterized as an inductive limit in analogy to
(6.3). If, in addition, L is a vector lattice, then (L, '1: 0) is a I.c.v.1.
(c) If L is an almost Archimedean ordered vector lattice such that
L + distinguishes points in L, then the order of L is Archimedean (hence
regular). (Use (b), observing that '1:0 is a Hausdorff topology for which
the positive cone is closed.)
(d) Let T be a set containing at least two elements and let E be the
vector space (over R) of all bounded real functions on T, ordered by
the relation" g:;;;;/ if either / = g or inf{f(t) - get): t E T} > 0". The
order of E is almost Archimedean, but not Archimedean.
(e) Every order complete vector lattice is Archimedean ordered.
20. (Continuity of the Lattice Operations; cf. Gordon [1], Peressini
[I]). Let E be an ordered I.c.s. over R whose positive cone is generating,
and endow E' with its canonical order. Denote by o(E, E') the topology
of uniform convergence on all order intervals in E'.
(a) If E' is a C*-saturated subspace of E+, then o(E, E') is consistent
with <E, E'). (Observe that each order interval in E' is u(E', E)-
compact.)
(b) If E is a I.c. vector lattice, then o(E, £') is the coarsest translation-
invariant topology on E finer than u(E, E') and for which the lattice op-
erations are continuous. Deduce from this that if E is an (AL)-space,
the norm topology is the only topology consistent with <E, E') and such
that the lattice operations are continuous.
(c) If E is a normed lattice, then the lattice operations are weakly
continuous if and only if E is finite dimensional. (Use (b) to infer that
every order interval of E' must be contained in a finite dimensional sub-
space of E', and show that this is absurd unless E' is of finite dimension.)
(d) If E is a I.c.v.l., the completion of (E', o(E', E» can be identified
with the band in E+ generated by E'. Deduce that if E is barreled, then
E' is complete for o(E', E).
(e) Let L be a vector lattice and let P be a non-empty set of positive
linear forms on L. The semi-norms x --> f(lxl) (J E P) generate a I.c.
convex topology '1: on L for which the lattice operations are continuous.
More precisely, (L, '1:)' is the smallest solid subspace M of L + that
contains P, and if (E, '1: 1) is the Hausdorff t.v.s. associated with (L, '1:),
then (E, '1: 1 ) is a I.c.v.1. whose dual E' can be identified with M, and such
that '1: 1 = o(E, E').
2 [. Let L be a vector lattice of finite dimension n.
(a) If Lis Archimedean ordered, then L is isomorphic with R n under
its canonical order. (Observe that the positive cone C of L has non-
empty interior (proof of the lemma preceding (4.1», and use (8.4) and
EXERCISES 255
(8.5) to show that L is isomorphic with C(] R(X), where X contains

exactly n points.)
(b) If L is not Archimedean ordered, there exist integers k, m such
that 2 ~ k ~ n, m ;;;; 0, k + m = n and such that L is isomorphic with
Rt x R'fj, where the order of Rt is lexicographic and the order of R'fj is
canonical. (Birkhoff [1], Chapter XV, Theorem 1.)
22. Let E be an (AL)-space.
(a) If (and only if) E is separable, there exists' a compact metrizable
space X such that E is isomorphic (as a Banach lattice) with LI(f..l) where
f..l is a suitable regular Borel measure on X. (Observe that E possesses a
weak order unit Xo and note that E is the band generated by {xo};
utilize (8.6) and the Radon-Nikodym theorem.)
(b) If, in the spirit of (8.6), E" is identified with the Banach lattice of
all bounded, signed, regular Borel measures on X (Chapter II, Section 2,
Example 3), deduce from (8.6) that the measures in E are exactly those
vanishing on each subset of first category in X. (Kelley-Namioka [1].)
23. If Xis a compact space, the Banach latticeC(] R(X) is order complete
exactly when X is extremally disconnected (a Stonian space). Infer from
this that C(] ReX) cannot be a dual Banach space, unless X is Stonian,
and not reflexive, unless X is finite.
24. Let A be an algebra over R with unit e and denote by Ao the
underlying vector space of A (Chapter IV, Exercise 40). A is called an
ordered algebra if Ao is an Archimedean ordered vector space such that
e ;;;; 0 and such that a ;;;; 0, b ;;;; 0 imply ab ;;;; O.
(a) Suppose that e is an order unit of Ao, and denote by C* the
cone (in A~) of all positive linear forms. Each linear formf generating
an extreme ray of C* and satisfying fee) = 1 is multiplicative: f(ba) =
f(b)f(a) for all a, bE A. (Show that for fixed b ;;;; 0, a -+ f(ba) is a linear
form g such that g = A.d, and that A.b = f(b).)
(b) (Stone's Algebra Theorem). Let A be an ordered algebra such
that the unit e of A is an order unit. Under evaluation, A is isomorphic
with a dense subalgebra of C(] R(X), where X is the O'(A~, Ao)-compact
set ofmuItiplicative, positive linear formsfsatisfyingf(e) = 1. (Use (a)
and the Krein-Milman theorem, and apply (8.2).) Infer that A is com-
mutative and show that evaluation on X is a norm isomorphism of (A, Pe)
into Cj R(X), where Pe denotes the gauge of [ - e, e] in A.)
(c) If, in addition to the hypothesis of (b), every positive sequence of
type [I in A is order sum mable, then (A, Pe) is isomorphic with the
Banach algebra C(] R(X) (Chapter IV, Exercise 40). Conclude that in these
circumstances, Ao is necessarily a vector lattice.
25. (Spectral Measures and Algebras). Let X be a compact space,
A a locally convex algebra over K (Chapter IV, Exercise 40). A contin-
uous map f..l of C(]K(X) onto A which is an algebraic homomorphism, is
called a spectral measure on X with range A. The range of a spectral
measure is called a spectral algebra (over K). If A is a l.c. algebra over K
with unit e, a subalgebra of A which contains e and is a spectral algebra
is called a spectral subalgebra of A.
(a) A l.c. algebra over R with unit e is a spectral algebra if and only
if there exists an order of A such that (i) A is an ordered algebra; (ii) e

is an order unit, and [- e, e] is bounded; (iii) every positive sequence
of type [1 is order summable. (Use Exercise 24.)
(b) A I.c. algebra over C with unit e is a spectral algebra if and only
if there exists a real subalgebra A 1 containing e such that each a E A has
a unique representation a = b + ic (b, c E A 1), and such that A1 is a spec-
tral algebra over R (cf. (8.3». .
(c) Let It be a spectral measure on X with range A. Define the support
of It to be the complement Xo of the largest open set G c X such that
It(f) = 0 whenever / has its support in G. Then It induces a spectral
measure Ito on Xo with range A which is biunivocal. For Ito to be a hom-
eomorphism (equivalently, for It to be a topological homomorphism),
it is necessary and sufficient that a -+ rea) be continuous on A, where
rea) denotes the spectral radius of a EA. (Chapter IV, Exercise 40.)
(d) If A is a I.c. algebra, an element a E A is called a spectral element
if a is contained in a spectral subalgebra of A. (If A is an algebra of
continuous endomorphisms of a I.c.s. E, the spectral elements of A are
called (scalar type) spectral operators on E.) If a is a spectral element
of A and It is a spectral measure on X such that a = It (f), then/eX) is
the spectrum of a,f(X) = u(a). (Cf. Schaefer [9], II, Theorem 3.)
(e) Let A be a spectral algebra over K, and a EA. There exists a
spectral measure v on u(a) such that a = v(1), where 1 denotes the
identity function on u(a), and such that the range of v is the smallest
spectral subalgebra of A that contains a. (Consider the mapping
g -+ It(g of) of ~K(u(a» into A.)
26. The following are typical examples of spectral a1gebras:
(a) The algebra L oe'('t') (where (Z, 1:, -r) is a measure space and multi-
plication is defined by pointwise multiplication of representatives).
This algebra can be viewed as a spectral subalgebra of the Banach alge-
bra of continuous endomorphisms, .P(E), where E = LP(-r) (1 ~ p ~ +
(0).
(b) Every norm-closed (real) algebra of Hermitian operators on a
Hilbert space (in particular, the closure of each such algebra under the
topology of simple convergence).
(c) Let E be a I.c. vector lattice on which every positive linear form
is continuous (cf. (5.5», and suppose that the positive cone C is weakly
sequentially complete. Endow .P(E) with its canonical order and the
topology of simple convergence, and denote by e the identity map of E.
The linear hull of [-e, e] in !l'(E) is a spectral subalgebra of !l'(E).
This can be extended to the case where E is a I.c.s. over C such that its
underlying real space EiJ is a I.c.v.I. (Schaefer [8], Theorem 7.)
Examples (a) and (b) are special cases of (c). (For (b), see Schaefer [3],
(11.3).)
27. (Extension of Spectral Measures). Let X be a compact space, It a
spectral measure on X with range Ai> where A1 is a spectral sub.algebra
of the I.c. algebra A. Denote by J the unit interval of A for the finest
order on A such that JL is positive (where ~(X) is ordered as usual), and
suppose that J is weakly sequentially complete. There exists a continuous
EXERCISES 257
extension ji of Jl to the Banach algebra at(X) of bounded Baire functions

on X, such that ji is an algebraic homomorphism of at(X) into A. More-
over, ji induces a homomorphism of the Boolean algebra of Baire sub-
sets of X onto a a-complete Boolean algebra of idempotents of A. (Cf.
Schaefer [9], II, Theorem 8.)
PREREQUISITES

ences can be found.
A. SETS AND ORDER

1

§1) PRELIMINARIES 259
Banach dual A' by (V, 3.3) Cor. 3, are the key to representations; Section 4
closes with the Jordan decomposition theorem, the Gelfand-Naimark-Segal
(GNS) construction, and faithful representations. Projections are pivotal to
all of the subsequent material in the chapter. Section 5 discusses the role of
projections as extreme points of certain order intervals and gives an order
theoretic characterization of unital C* -algebras.
The remaining three sections are devoted to W*-algebras (also called
operator algebras), which are the abstract version of von Neumann's rings of
operators. Section 6 is primarily concerned with establishing separate ( I -
weak continuity of multiplication and its consequences, the lattice of pro-
jections in a W* -algebra, and the important result (6.9) that such an algebra
has a unique predual which, significantly, consists precisely. of its order con-
tinuous linear forms. Section 7 presents the standard topologies on operator
algebras as well as the famous bicommutant and density theorems. Finally,
Section 8 gives an introduction to the classification theory of W* -algebras.
These are the bare elements of classification, and the section is mainly
intended to help the reader decide if he wants to pursue the subject further.
1. PRELIMINARIES
An algebra A (Chapter IV, Exercise 40) over the complex field C is called
a nonned algebra if its underlying vector space Ao is a normed space whose
norm x -+ Ilxll is submultiplicative:
IIxYIl ~ IIxlillyll (x, YEA). (1)
If Ao is a Banach space, A is called a complete nonned algebra or Banach
algebra. A is called unital if it possesses a (multiplicative) unit e satisfying
lIell = 1. If A is unital and complete, the geometric series (e - x)-I =
L~o xk shows that (e - x) is invertible for each x E A, IIxll < 1. In particu-
lar, the set A(-I) of all invertible elements is open and the map x -+ X-I is
continuous on A(-I).
The prime example of a normed algebra is, of course, the algebra .!e(E) of
all bounded operators on a Banach space E, provided with the standard
operator norm (Chapter II, Section 2).
Let A denote a complete, normed unital algebra over C. In complete
analogy with Chapter IV, Exercise 40 we define, for given x E A, the
resolvent set p(x) to be the set of all A E C for which (Ae - x)-I exists;
p(x) is open and contains the set {A E C : IAI > IIxll}. The function A-+
(Ae - x)- =: R(A,X) from p(x) into A is called the resolvent of x; it satisfies
the important (resolvent) equation
R(A, x) - R(J1" x) = -(A - J1,)R(A, x)R(J1" x) (2)

for all A,J1, E p(x). From (2), it is easy to conclude that the resolvent is hol-
omorphic in p(x) (Chapter IV, Exercise 39); in particular, because R(A, x) -+
260 c- AND W'-ALGEBRAS [Ch. VI
o as A --.00, Liouville's theorem implies that p(x) #- C. Hence, for all x E A

the spectrum a(x) := C\p(x) is compact and non-void.
We also recall the spectral mapping theorem for polynomials (Exercise 1):
For every polynomial P (over C), we have a(P(x)) = P(a(x)). The spectral
radius rex) := SUp{IAI : A E a(x)} (Chapter IV, Exercise 40) satisfies
rex) = lim Ilxnlll/n.

n-;(jJ
A fairly direct consequence of the preceding is the well-known theorem of

Gelfand-Mazur (I. Gelfand and M. A. Naimark [1]) (Exercise 2):
A unital Banach algebra in which every non-zero element is invertible, is
necessarily isomorphic to the complex field C.
A left (right) ideal is a linear subspace M of A satisfying AM c M
(MA c M, respectively); a subspace M having both properties is called a
two-sided ideal, or simply an ideal of A. Closures of left, right, and two-sided
ideals retain the defining property. Finally, left, right, and two-sided ideals
are called proper if they are proper subsets of A.
Suppose again that A is a unital Banach algebra. By Zorn's lemma, every
proper ideal of A is contained in a maximal proper ideal; the latter is neces-
sarily closed, because A(-1) is open. Morever, if ] is a closed ideal
of A, the Banach space AI] (II, 2.3) becomes a Banach algebra by defining
q(x)q(y) := q(xy) where q : A --. AjJ denotes the quotient map. (This holds
whether or not A is unital.) If A is unital, so is AI].
LEMMA. Let A be complete, unital, and commutative. For every maximal
proper ideal ], AI] is isomorphic to the complex field.
For a proof, note that AI] contains no proper ideal #- {O} by the max-
imality of ]. Hence every element #- 0 of AI] is invertible; the Gelfand-
Mazur theorem now implies the claim.
Finally, a norm-preserving conjugate linear map x --. x* (of A onto A) is
called an involution if it satisfies x** = x and (xy)* = y*x* for all x, YEA. A
complex Banach algebra with involution is called a B* -algebra.
2. C* -ALGEBRAS. THE GELFAND THEOREM

A C* -algebra is defined as a complex Banach algebra A with involution
x --. x', satisfying
Ilx'xll = IIxl12 (3)
for all x E A. Clearly, this definition is modeled after the Banach algebra
!£,(H) of bounded operators on a complex Hilbert space H (11,2.3, Example
5), the involution of !£,(H) being given by taking the adjoint operator T* of
any T E !£,(H).
§2] C* -ALGEBRAS. THE GELFAND THEOREM 261
Examples
The following examples present some principal types of C* -algebras oc-
curring in analysis. For further details and examples, refer to the exercises
and the bibliography.
1. Let K '# 0 be a compact space, C(K) the algebra of continuous
complex functions on K. Under the standard algebraic operations and
the supremum norm, C(K) becomes a C* -algebra if an involution
f -+ j* is defined as complex conjugation: j*(s) := f(s) (s E K). This
example exhausts the supply of unital commutative C* -algebras (the-
orem of Gelfand, (2.2) below). Similarly, if X is a locally compact, non-
compact space then Co(X), the algebra of continuous complex func-
tions vanishing at infinity, is a C* -algebra without unit, and every
non-unital commutative CO-algebra '# {O} is of this form.
2. If (X, :E,,u) is a measure space (Chapter II, Section 2, Example
2), the (complex) Banach space LOO(,u) becomes a CO-algebra ifmul-
tiplication is defined (on equivalence classes) by pointwise multiplica-
tion of representatives, the involution again being given by functional
conjugation. This is an example of a (commutative) W* -algebra, i.e., a
C' -algebra whose underlying normed space is a dual Banach space.
3. The standard example of a non-commutative C'-algebra is the
operator space fe(H) (see above), Hbeing a Hilbert space of dimension
at least two. It will be seen that fe(H) actually is a W*-algebra. An
example of a closed ideal in fe(H) is given by the subspace of all
compact operators; this ideal is proper if and only if dim H is infinite.
4. Let G be a locally compact group with Haar measure A and let
H:= L2(G,A). For f E Ll(G,A) let Tf denote the convolution opera-
tor g -+ f * g, and let Cr( G) denote the norm closure in fe(H) of (the
algebra) {Tf : fELl (G, A)}; Cr ( G) is a CO-algebra called the reduced
group algebra of G.
5. If A is any C* -algebra and K '# 0 a compact space, the algebra
C(K, A) of all continous functions K -+ A is a C* -algebra under the
natural operations and the sup-norm.
6. Let Hk (k = 1, ... , n) denote Hilbert spaces, H = EBZ=l Hk their
direct sum. The set of all operators EBZ=l Tk, where Tk E fe(Hk), con-
stitutes a C*-subalgebra of fe(H).
We now continue our presentation of basic results, again denoting by A a

(not necessarily unital) C· -algebra with involution x -+ x'. An element
x E A is self-adjoint or hermitian if x = x', normal if xx' = x* x. If A is uni-
tal, an element x satisfying x' = X-l is unitary. Each x E A has a unique
decomposition x = a + ib (i2 = -1) where a, b are self-adjoint; namely, a =
! (x + x*), b = ;i (x - x*). Clearly then, x is normal if and only if ab = ba.
262 C*- AND W*-ALGEBRAS [Ch. VI
Thus A = Asa + iAsa where Asa is the subset of all self-adjoint elements of
A; Asa is a real subspace of A (i.e., a subspace of A when scalars are restricted
to the real field R). Also it is easy to prove that a E Asa, bE Asa implies
ab E Asa if (and only if) ab = ba.
We begin our detailed presentation with the following simple but basic
results. We denote, as before, by u(x) the spectrum, by p(x) the resolvent set,
by rex) the spectral radius rex) := sup{IAI : A E u(x)} of an element x E A.
Finally, R(A,X) := (Ae - X)-I for A E p(x).
2.1
Let A be unital and x E A. We have:
(i) u(x*) = {1 : A E u(x)}.
(ii) If x is normal then rex) = Ilxll.
(iii) Ifx is unitary then u(x) C {A: IAI = I}.
(iv) If x = x* then u(x) c R, and A E p(x) ( l R implies (Ae - x)-I to be
self-adjoint.
Proof (i) (Ae - X)R(A, x) = e implies (Xe - X*)R(A, x)* = e* = e and
conversely.
(ii) Because Ilx*xll = IIxl12 by definition of a CO-algebra, xx* = x*x
implies
Therefore,
rex) = n->oo
lim IIx 2n 11 2- = Ilxll.
n
(iii) For XEA invertible u(x-I)={l/A:AEU(X)} holds, which in

particular yields r(x-I) = sup{l/IAI : A E u(x)}. Because a unitary x E A is
normal, the equation
1= vTx*Xif = Ilxll = rex) = IIx* II = Ilx-111 = r(x-I)
(izt
implies the assertion.
(iv) Let x = x* and v := exp(ix) = 2::=0 Continuity of the involu-
tion map implies that v* = v-I. For A E u(x) we have exp(iA)e - v =
i(Ae - x)w for suitable w commuting with (Ae - x). Hence exp(iA) E u(v);
because v is unitary, lexp iAI = 1 by (iii). But this is only possible if A is real;
the remainder is now easy to see.
Let A,B be CO-algebras. A linear map <1>: A --> B is called a homo-
morphism if preserves multiplication and involution, i.e., if (xy) =
(x)(y) and (x*) = (x) * for all x, YEA. (Sometimes these maps are
called C* -morphisms or *-homomorphisms.)
An isomorphism A --> B is a bijective homorphism, and an automorphism
is an isomorphism of A onto itself. Simple examples of the latter are the

maps x ---; uxu* where u is unitary; they are called inner automorphisms.
Now we can prove the famous characterization of commutative unital
C'-algebras by I. Gelfand (I. Gelfand and M. A. Naimark [1]).
2.2
Theorem. Every commutative, unital C* -algebra A is isometrically iso-
morphic to the (complex) algebra C(K) for some compact space K that is
unique up to homeomorphism.
Proof By the Gelfand-Mazur theorem (Section 1 and Exercise 2) and by
the lemma of Section 1, A/J is isometrically isomorphic to C for each max-
imal (necessarily closed) proper ideal J of A. The composition f : A ---;
A/J ---; C is a contractive, multiplicative linear form on A of norm 1, be-
cause f(e) = 1; the set K of all these forms is thus a bounded subset of the
dual Banach space A' of A and easily seen to be a(A',A)-closed. By (III,
4.3) Corollary, K is compact for this topology.
We show now that every f E K is a *-homomorphism. Indeed if f E K is
arbitrary then f(f(x)e - x) = 0 implies f(x) E a(x). So if x is self-adjoint,
then by 2.1(iv) f(x) E R. This in turn implies f(y*) = f(y) for each y = a +
ib where a, bE Asa.
Next we show that IIxll = sup{lf(x)I : f E K}. This implies, in particular,
that K separates points on A. If 0 =F x E A then, because A is commuta-
tive, x is normal and thus Ilxll = r(x) =F 0 holds by 2.I(ii). Now if 2 E a(x)
then (2e - x)A is contained in a maximal (proper) ideal J of A. Therefore,
f(2e - x) = 0 for some f E K (above notation) and, if 121 = r(x) then If(x)1
= 121 = r(x) = Ilxll·
Thus x ---; f(x) = <x, f) (x E A, f E A') defines an isometric isomorphism
 : A ---; C(K). Because (V, 8.3) (ii) implies that (A) is dense in C(K), it
follows that is an isometric isomorphism of A onto C(K). Finally, the
uniqueness of K (to within homeomorphism) is a consequence of the well-
known theorem of Banach-Stone (see also Exercise 3).
In the following we will denote by idx the identical map of a set X, and by
lx, the characteristic function of X.
COROLLARY 1. Let a be a normal element of a unital C' -algebra A.
(i) There existspreciseiy one injective homomorphism '¥ ofC(a(a)) into A
satisfying '¥(lu(a)) = e and '¥(idu(a)) = a.
(ii) For each f E C(a(a)), a('¥(f)) = f(a(a)).
Proof Since a is normal the C* -subalgebra B of A generated by a and e
is commutative, thus there exists a unique compact space K and an
isometric isomorphism of B onto C(K). Since a and e generate B,
h := (a) separates points on K, hence h is a homeomorphism of K onto a
compact subset X of C. So we may identify K with X obtaining h = idx . Let
'¥ = <1>-1.
Denoting by ac(x) the spectrum of x viewed as an element of the unital

C*-algebra C we see easily that X = aB(a) and, more generally, aB(If'(f))
= f(X).
We show that aB(If'(f)) = aA(If'(f)), which yields X = a(a) as well as
assertion (ii).
To this end, assume that If'(f) is not invertible in B. Then f is not
invertible in C(X), thus f(~) = 0 for some ~ E X. To this ~ there exists a
sequence (h n) in C(X) satisfying (oc) 0 S hn sIx, (P) hn(~) = 1, and (y)
hn(,,) = 0 whenever 1,,- ~I ;::: lin. But then limllfhnll = O. Applying the
isometry If' we obtain a sequence (xn) = (If'(hn)) of elements Xn of norm 1
such that limn If'(f)xn = 0 holds. But then If'(f) is not invertible in A. The
assertion follows (see Exercise 4).
REMARK. If, in analogy to the polynomial notation Pea), we write
f (a) in place of If' (f), the second assertion of the corollary assumes the
suggestive form a(f(a)) = f(a(a)), and is called the spectral mapping
theorem for continuous functions.
COROLLARY 2. Every element of a unital C* -algebra is a linear combination
of ( at most) four unitary elements.
Proof Letfbe a real continuous function of norm 1 in C(K), where K
is compact. Then 2f = (f + iJI K - j2) + (f - iJI K - j2). Here both
summands are unitary. Every element of an arbitrary C* -algebra A is of the
form a + ib where a, b are self-adjoint (see above). Now we can apply 2.2,
Corollary 1 to a and b separately.
Our final aim in this section is Theorem 2.4, which exhibits the very spe-
cial properties of homomorphisms of C' -algebras. For this we need some
preparation.
First, every non-unital C* -algebra A can be embedded, with linear codi-
mension 1, into a unital CO-algebra A. We define A = C x A. Define addi-
tion in the canonical way for products of vector spaces. For multiplication
we let (oc,x)(P,y):= (ocP,ocy+px+xy). Involution is defined by (oc,x)*:=
(a,x*), and the norm of A is defined by II(oc,x)11 =sup{llxy+ocyll: yE
A, Ilyll s I}. The embedding A ----+ A so defined is an isometric ismorphism
(Exercise 5), and A is called the unitization of A. If in the following A is non-
unital, by a(x) we will understand the spectrum of x in the unitization of A.
COROLLARY 3. Every non-unital commutative C*-algebra A i= {O} is iso-
morphic to the algebra Co(X) of all continuous functions vanishing at infinity,
on some locally compact non-compact space X
Proof The unitization A is isomorphic to C(K), where K is compact. Now
A is mapped onto a maximal ideal in C(K).
LEMMA. Let A denote any C* -algebra, J a closed ideal of A. For every
x E J there exists a sequence (en)nEN of self-adjoint elements of J such that
(i) a(en ) C [0,1]

(ii) x = limn-->oo xen.
Proof Suppose first that A is unital and x = x*. Defining In(t) = nt2I
(1 + nt2) for n E Nand t E R, we conclude from (2.2), Corollary 1, (ii) that
a(fn(x)) C [0,1] .and aCe - fn(x)) C [0,1]. Defining en := fn(x), we have
en = nee + nx 2)-lx 2 E J. With gn(t) := t2(1 - In(t)) we obtain a(x 2(e - en))
= a(gn(x)) = gn(a(x)) C [o,~] hence, by (2.1), IIx 2(e - en)11 ::::;; lin. Because
x and (e - en) are self-adjoint and commute, lie - en II : : ; 1 implies that
Ilx - xenl1 2 = Ilx 2(e - en)211::::;; Ilx 2(e - en)11 ::::;; lin.
If x is not self-adjoint, construct a similar sequence (e~) with respect to x*x.
Then we have II(x*x)(e - e~)11 --+ 0 (n --+ (0) and
IIx-xe~112 = II(e-e~)x'x(e-e~)II::::;; Ilx*x(e-e~)II·
Finally, if A is not unital, the preceding construction still applies to the
unitization A of A, because the ideal J is closed in A as well.
2.3
Every closed ideal J of a C' -algebra A is invariant under conjugation.
Moreover, if q : A --+ AI J denotes the quotient map, the normed algebra
q(A) = A/J is a C* -algebra, involution being defined by q(x) * := q(x*)
(x E A).
Proof Let x E J and denote by (en) a sequence in J according to the
lemma above. Then x = limn-->oo xen implies x* = limn->oo enx*. en E J im-
plies x* E J, hence J = J*.
For the second assertion, we may suppose A is unital. By the preceding,
it is clear that q(x)* := q(x*) defines an involution on AIJ. Morever, by
(II, 2.3) A/J is a complete normed algebra under the quotient norm (see
Section 1).
Now denoting q by x --+ X, all that remains to be shown is that IIxl12 =
Ilx' xii for all x E AIJ.
Denoting by JI the set JI = {u E Jsa : a(u) C [0, I]} we prove that
Ilxll = inf{lIx - xull : u E JI}. (4)
By definition of the quotient norm (II, 2.3), for given 8> 0 there exists y E J
such that IIx + yll < Ilxl! + 8. By the lemma above there exists a sequence of
self-adjoint elements en E JI for which Ily - yenll --+ 0, lie - enll ::::;; 1. Now
for all n,
Ilxll + 8> Ilx + yll ~ II(x + y)(e - en)11
= II(x - xen) + y(e - en)11
~ inf II (x - xen) + y(e - en)11

nEN
= inf Ilx - xenll ~ Ilxll·

nEN
This proves (4). Now, taking infima over all u E J\, we obtain
IIxl1 2 = infllx(e - u)11 2

= infll(e - u)x'x(e - u)11
~ infllx'x(e - u)11
= IWxll·
The assertion follows (cf. Exercise 5).
It is clear that the quotient map A --+ AIJ (J any closed ideal of A) is the
prototype of a homomorphism. We shall see now that the rich structure of
C· -algebras imposes much stronger restrictions on homomorphisms than is
the case for topological vector spaces.
One remark on homomorphisms between C'-algebras A,B that are not
required to be unital: If A is unital with unit e and ct> : A --+ B is a homo-
morphism, then ct>(e) is a unit of ct>(A). Conversely, if A is not unital but B
is, then a homomorphism ct> : A --+ B has a unique (homomorphic) extension
 : A --+ B, where A denotes the unitization of A (Exercise 7).
2.4
Theorem. Let A, B denote C' -algebras, ct> : A --+ B a homomorphism. Then
ct> is open, contractive, and ct>(A) is a C* -subalgebra of B. If ct> is injective, it
necessarily is an isometry.
Proof With reference to the foregoing remark, we may and will suppose
A,B to be unital with ct>(e) = f, fthe unit of B. ct>(A(-1») c B(-1) implies
a(ct>(x)) c a(x) and thus
11ct>(x) 112 = 11ct>(x)*ct>(x)II = 11ct>(x'x)II = r(ct>(x'x))
~ r(x*x) = Ilx'xll = Ilx112.
Hence ct> is contractive. Now suppose ct> is injective. If ct> is not an isometry,
there exists a self-adjoint element x E A satisfying r(ct>(x)) < r(x), by the
preceding. Let f -=1= 0 denote a continuous function on a( x) vanishing on
a(ct>(x)). By 2.2, Corollary 1, (ii) we have f(x) -=1= 0 but f(ct>(x)) = 0,
while (fbeing a uniform limit of polynomials on a(x)) we must have 0 =
f(ct>(x)) = ct>(f(x)), which is contradictory.
Finally, if ct> is not injective and if J = ct>-1 (0) is the kernel of ct>, J is
a closed ideal of A. Therefore ct> induces an injective homomorphism :
AIJ --+ B so that ct> = 0 q where q denotes the canonical map. Because q
§3] ORDER STRUCTURE OF A C* -ALGEBRA 267
is an open map and is an isometry by the preceding, it follows that <l> is
open, completing the proof.
3. ORDER STRUCTURE OF A C* -ALGEBRA

We now introduce a natural order structure for arbitrary CO-algebras.
The general framework for this has been established in Chapter V, Sections 1
and 2; we refer the reader to those sections for the basic terminology. How-
ever, symbols such as x+ , x- , and others, used in Chapter V for vector lattices,
now have a different meaning, at least in the (most important) case of non-
commutative C* -algebras; the reader is hence cautione~ to avoid confusion.
On the other hand, many of the non-lattice-related results and concepts are
extremely useful in the present context.
Given any CO-algebra A, as before we write Asa = {x E A: x = x*} for
the self-adjoint part of A. Because Asa is the kernel of the continuous real
linear map x -+ x-x', Asa is a closed real vector subspace of A, and we
have A = Asa + iAsa as observed above (Section 2). As before, if A is
non-unital, bya(x) we will understand the spectrum of x in the unitization
AofA.
LEMMA. (a) Let x E Asa satisfy a(x) c R+. There exists a unique element
y E Asa satisfying a(y) c R+ and y2 = x.
(b) Let A be unital and x E Asa. The following are equivalent assertions:
(i) There exists t E R+ satisfying Ilxll ~ t and lite - xii ~ t.
(ii) For all t E R+, Ilxll ~ t implies lite - xii ~ t.
(iii) a(x) c R+.
Proof Without loss of generality suppose A unital with unit e.
(a) By the theorem of Stone-Weierstrass (V, 8.2) there exists a sequence
(Pn)',eN of polynomials on a(x) converging uniformly to t ~ 0. By (2.2),
Corollary 1 (Pn(x))neN converges to an element b ~ 0 satisfying b 2 = x.
Suppose now that there is another element c ~ 0 satisfying c2 = x. Then
cx = c3 = xc, hence the (commutative) C*-subalgebra C generated by c and
e contains x and, therefore, b. But by (2.2), Corollary 1 C is isomorphic to
C(a(c)) where the positive square root of an element is unique.
(b) The assertion is obvious in C(a(x)), which is isometrically isomorphic
to the C* -subalgebra generated by x and e (see (2.2), Corollary 1).
3.1
Theorem. Let A be any C* -algebra.
(i) The set A+ := {x E Asa : a(x) c R+} is a closed, convex, and normal
cone (o/vertex 0).
(ii) For each x E Asa there exist exactly two elements x+ E A+ and x- E
A+ satisfying x = x+ - x- and x+x- = x-x+ = o.
(iii) A+ = {yO Y : YEA}.
Proof As before, suppose A is unital without loss of generality.

(i) It is clear that A+ is invariant under multiplication by (real) scalars
~O; thus it is a cone of vertex 0 (II, 1). To show A+ is convex, it must
be proved that XI, X2 E A+ implies XI + X2 E A+.
Let XI,X2 and tl,t2 (ER+) satisfy tj ~ Ilxjll and Iltje -xjll s; tj
(j = 1,2) in accordance with the lemma above. Then we have
+ t2)e - (Xl + x2)11 s; Iitie - XIII + II t2e - x211 s; tl + t2,
1I(t1
which implies Xl + X2 E A+ by part (b) of the lemma. This part also
implies A+ to be closed. Finally (b) of the lemma implies Ilxl + x2l1e-
(XI + X2) E A+, hence
Ilxl + x211e - XI = (11xl + x211e - (XI + X2» + «XI + X2) - xI) E A+.
Another application of part (b) of the lemma shows that Ilxl + x211 ;?:
IIxlll. By Chapter V, 3.1 A+ is normal.
(ii) Let X E Asa. Using Theorem 2.2, Corollary 1 again, the CO-algebra
generated by X and e is isomorphic with C(a(x)). But a(x) c: R;
hence we have elements x+ := f(x), X- := g(x) where f(t) = t+ =
sup(t,O) and g(t) = r = sup (-t, 0) for t E a(x). Clearly, x+ and x-
have the required properties. For uniqueness, let X = Y - z where
y, Z E A+ are elements satisfying yz = O. Then y, z commute with
each other and with x, so they are contained in a commutative C'-
algebra containing X and x+, x- as well. By a now familiar argument
we must have y = x+, Z = x-.
(iii) If X E A+ then by part (a) of the lemma, X = y2 for y = y'X. Con-
versely let X = y' y for some YEA. Then X = x+ - x- as above. Set
u=yF. Then
u'u = vx-(x+ - x-)vx- = _(X-)2 E - A+. (5)
Now y = a + ib for some a, bE Asa. Hence
uu* = 2(a 2 + b2) - u*u = 2(a 2 + b2) + (X-)2 E A+. (6)
But a(vw)\{O} = a(wv)\{O} holds in every unital Banach algebra

(Exercise 1). Hence a«x-)2) = {O} by equations (5) and (6). Because
x- is self-adjoint, Ilx-11 = r(x-) = 0 holds. This shows x E A+ and
completes the proof.
The preceding theorem now permits us to define an order relation s; on
Asa (Chapter V, Section 1) by defining as; b whenever b - a E A+. This
ordering is called the canonical ordering of a C'-algebra; the preceding
theorem shows that its positive cone A+ is closed, normal, and generating.
In accordance with Chapter V, Section 2, the underlying vector space of A
is a complex ordered vector space. Note, however, that (unless A is com-
§3] ORDER STRUCTURE OF A C* -ALGEBRA 269
mutative) the positive cone A+ is not invariant under multiplication; indeed

a, b E A+ does not imply ab E A+ unless a, b commute.
Generally, an element x E A+ is called positive; hence the decomposition
of any x E A, x = a + ib with a, bE Asa shows that every x E A is a linear
combination (over C) of at most four positive elements.
We now collect a few inequalities, with respect to the canonical order of a
C' -algebra A, that are frequently useful. For x E A we define the modulus
Ixl by Ixl = v'x'x, noting that x'x E A+.
3.2
Let x, y, c be elements of A.
(i) x::;; y implies c'xc::;; c' yc. In particular, x' y' yx::;; IlyI12x*x.
(ii) Ixl ::;; Iyl implies Ilxll ::;; Ilyll·
(iii) 0::;; x ::;; y implies 0 ::;; -.jX ::;; y'Y.
(iv) If A is unital, if x is invertible and 0 < x ::;; y, then y is invertible and
0< y-I ::;; x-I.
Proof (i) By (3.1) there exists dE A such that 0::;; y - x = dOd. Thus
c*(y - x)c = c*d*dc = (dc)*dc ~ 0, whence the first assertion follows. For
the second, observe that O"(lly' ylle - y' y) c R+ implies 0::;; y' y::;; IIyl12 e
so the claim follows.
(ii) follows from IIIxll12 = IIxl12 and (i).
(iii) It suffices to show that for x, y ~ 0, x 2 ::;; y2 implies x::;; y. Supposing
A is unital, let t E R be > 0 and define
1 1
z = (te + y + x)(te + y - x), c = 2" (z + z*), d = 2i (z - z*).
Then we obtain c ~ t 2 e. Thus c is positive and invertible. Now
e + ic-Ij2dc-Ij2 = c- Ij2 (c + id)c- Ij2
is invertible, and so is c + id = Z. Thus (te + y -x) has a left inverse
and, because it is self-adjoint, an inverse; therefore -t E p(y - x) and
hence y - x ~ O.
(iv) From (i) we have e::;; x-Ij2yx-Ij2 whence 0 E p(y). Using (2.2),
Corollary 1 for a = x-Ij2yx-Ij2 we obtain e ~ a-I and the claim fol-
lows from (i).
We next tum to the study of the ordering of the dual Banach space A I
induced by the canonical order of a given C* -algebra A. To do so for a not
necessarily unital A, we define an approximate unit of A to be a directed (::;;)
subset (u;.);'eA of A+ n U, such that
lim xu;. = lim U;.X = x for all x E A.
;'eA ;'eA
As before, Uo denotes the open unit ball.

3.3
For every C' -algebra A, the set A+ n Uo is directed ( ::;;) and constitutes an
approximate unit.
Proof Let x, y E A+ satisfy x::;; y. Then (e + y)-I ::;; (e + x)-I by (3.2);
hence
0::;; x(e + x)-I = e - (e + xr l ::;; e - (e + y)-I = y(e + y)-I E Uo,

where the last relation follows from (2.2) Corollary 1, applied to y.
Now let x, y E A+ n Uo be arbitrary and set x' = x(e - x)-I, y' =
y(e - y)-I and z = (x' + y')(e + x' + yl)-I. Then
x = x'(e + X')-I ::;; z and y = l(e + ylrl ::;; z,

which shows that A+ n Uo ist directed (::;;).
Now let x E A+. Indexed by u E A+ n Uo, the family x(e - u)x is directed
(;;::) by 3.2 (i), since x = x*. By 2.3, the sequence (en)neN defined by en =
nx 2(e + nx 2r l is increasing, contained in A+ n Uo and satisfies limn xen =
x. A fortiori, limueA+ n Uo x(e - u)x = O. But by 3.2 (i), II(e - u)x112 =
Ilx(e - u)2xll ::;; Ilx(e - u)xll shows that A+ n Uo functions as an approxi-
mate unit for each x E A+. The claim now follows, because A = (A+ - A+) +
i(A+ - A+).
4. POSITIVE LINEAR FORMS. REPRESENTATIONS

As before, let A' denote the Banach dual of a CO-algebra A. To define an
order structure on A', let us first look at the self-adjoint part Asa of A, re-
calling that A = Asa EB iAsa is a topological direct sum (over R); (III, 2.1)
Corollary 3 applies, because Asa is closed in A. Now if ,po is a (real-valued)
continuous linear form on Asa and if x = a + ib where x E A, a, b E A sa , the
continuous linear form,p E A' defined by
,p(x) = ,po(a) + i,po(b)
is called self-adjoint; thus ,p E A' is self-adjoint iff ,p(Asa) cR. More espe-
cially, ,p E A' is called positive if ,p(A+) c R+. Because A+ is a normal, gen-
erating cone (over C) by (3.1), from Theorem V.3.5 it follows that the cone
A~ of positive linear forms is normal in A' as well as generating. Again, the
decomposition A' = (A~ - A~) EB i(A~ - A~) is a topological direct sum
(over R). The term self-adjoint for elements of A~ - A~ has its origin in the
existence of a natural involution map on A'. In fact, defining ,p*(x) := ,p(x*)
(x E A,,p E A') yields a (conjugate linear) isometry ,p ---+ rjJ* that is 'involutive;
it turns out that the self-adjoint linear forms ,p on A are exactly those for
which ,p = ,p*. It follows from the preceding that every self-adjoint form ,p
,has a decomposition ,p = ,pI - ,p2 where ,pi are positive (i = 1,2); for a much
§4l POSITIVE LINEAR FORMS. REPRESENTATIONS 271
stronger result (Jordan decomposition), we need an algebraic-topological

characterization of positive linear forms.
4.1
Let rjJ E A'. For rjJ to be positive, it is necessary and sufficient that IlrjJll =
lim;. rjJ(u;.} for some (and infactJor any) approximate unit (u;J;'EA of A.
Proof Let rjJ be positive and denote by (u;,) any approximate unit of A.
Because (x, y) f--+ rjJ(y*x) is a positive sesquilinear (or Hermitian) form
on A x A, Schwarz' inequality shows that IrjJ(y*x)1 2 :s:; rjJ(y' y)rjJ(x*x). Now
suppose IlrjJll = 1 and let 0 < e E R be preassigned; there exists x E A, Ilxll = 1
such that IrjJ(x)1 2 ~ 1 - e. Since lim.. U;.X = x, the preceding inequality shows
that
1- e:S:; IrjJ(x)1 2 = lim IrjJ (u .. x) 12 :s:; rjJ(x'x) .lim;.rjJ(uX):s:; limrjJ(u .. )
.. ..
because 0 :s:; ui :s:; u... Since (u .. ) is directed (:s:;), we have lim;. rjJ(u;.) = 1 =
IlrjJll·
For the converse assertion, let rjJ E A' satisfy Ilr/JII = lim.. r/J(u .. ) for some
approximate unit; we may assume Ilr/JII = 1. Given x E A sa , IIxll = 1, let r/J(x)
= (1. + if3 where we may suppose that 13 :s:; O. Then
Ilx - inu .. 11 2 :s:; 1 + n 2 + nllxu.. - u;.xll (n EN)
implies that
Ir/J(x - inu.. ) 12 :s:; 1 + n 2 + nllxu.. - u.. xll.
Taking the limit along A, we obtain, for all n E N 1(1. + if3 - inl2 :s:; 1 + n2.
This implies -2nf3:S:; I - 132 - (1.2 for all n, hence 13 = O. Thus we have
r/J(Asa) c R, i.e., r/J is self-adjoint. 0 :s:; x and Ilxll = 1 implies lIu.. - xii :s:; 1 for
all A E A; hence r/J(u .. ) - r/J(x) :s:; Ilr/JII. Now r/J(u.. ) ----> Ilr/JII shows that -r/J(x) :s:;
o or equivalently, r/J(x) ~ O.
COROLLARY 1. Let A denote a non-unital C* -algebra, A its unitization. If
r/J E A' is positive and e denotes the unit of A, then jJ(e) = Ilr/JII defines an ex-
tension jJ E A that is positive.
Proof Let (u");'EA denote an approximate unit of A. Then for each com-
plex number fl we obtain in 1:
O:S:; (jiu .. + X*)(flU .. + x) :s:; Ifll2e + jiu.. x + flX*U;. + x'x;

hence IlflU;. + xl1 2 :s:; Ilfle + x1l 2. By (4.1) we have
IjJ(fle + x) I = lim
;.
1r/J(flU.. + x) I :s:; Ilfle + xlilir/JII·
Now jJ(e) = Ilr/JII = IIjJll, and e is a trivial approximate unit of A. The asser-
tion follows.
COROLLARY 2. For any C* -algebra A, the dual norm is additive on the cone
A~ ofpositive linear forms.
The simple proof can be left to the reader.

A positive linear form ifJ E A' is called a state of A if IlifJll = 1. The set
9'(A) of all states is then called the state space of A; if A is unital, 9'(A) is
the intersection of the I1(A',A)-closed hyperplane {ifJ E A' : ifJ(e) = I} with
the cone A~. Hence in this case, 9'(A) is I1(A', A)-compact. To finally ob-
tain the announced Jordan decomposition of self-adjoint elements of A', we
need another lemma.
LEMMA. Let x E A be normal. There exists a state ifJ E 9'(A) satisfying
lifJ(x) I = Ilxll·
Proof We may suppose A unital. The C· -subalgebra B of A generated by
x and e is isomorphic to C(K) by (2.2), Corollary 1. If f E C(K) corresponds
to x E B under this isomorphism, there exists a point functional (Dirac
measure) Os (s E K) such that los (I) I = If(s) I = IIfll. By the Hahn-Banach
theorem ((II, 3.2) Corollary) for normed spaces, there exists a norm-
preserving linear extension ifJ E A'; we have 1 = Ilosll = IlifJll = Os(1K) = ifJ(e).
Now (4.1) shows ifJ to be positive, hence we obtain ifJ E 9'(A).
4.2
Theorem. (Jordan Decomposition) For every ~elf-adjoint linear form ifJ on
A, there exist (unique) positive linear forms ifJ+, ifJ- on A satisfying ifJ = ifJ+ -
ifJ- and IlifJll = IlifJ+11 + IlifJ-lI·
Proof In view of Corollary I above, we may suppose A is unital. Since the
state space 9'(A) separates A sa , we can identify Asa with a linear subspace B
of C(9'(A)). By the preceding lemma, the map Asa ---+ B is an isometric
isomorphism that preserves positivity; moreover e E Asa is mapped onto the
constant one function l.9'(A).
Now let ifJ E A' be self-adjoint; as a linear form on B, ifJ has a nOrm-
preserving linear extension Il E C(9'(A))'. Now by (V, 8.4), C(9'(A))' is a
Banach lattice and indeed, an AL-space; therefore, Il has a (lattice) decom-
position Il = Il+ - Il- and we have
If now ifJ+, ifJ- denote the restrictions to B ~ A sa , respectively, of Il+ and Il-,
we clearly obtain positive linear forms satisfying ifJ = ifJ+ - ifJ-. But IlifJll =
1I1l11 by the choice of Il, and 111l+1I = Il+(e) = ifJ+(e) = IlifJ+11 (similarly for Il-);
therefore IlifJll = IlifJ+11 + IlifJ-ll· Finally, proofofthe unicity ofifJ+,ifJ_ is left to
the reader (Exercise 14).
We now aim at a first classification (Theorem 4.3) of arbitrary C*-
algebras; it will tum out that the operator algebra !l'(H) of a (suitable)
Hilbert space H (Section 1, Example 3) and its C*-subalgebras are the most
§4] POSITIVE LINEAR FORMS. REPRESENTATIONS 273
general C* -algebras. The result rests on the existence of sufficiently many

positive linear forms (cf. discussion preceding 4.1.) on a given CO-algebra;
essentially, this is a consequence of (V, 3.3) Corollary 3 and its predecessors,
i.e., of the normality of the positive cone A+.
If r/J is a positive linear form on A, the definition [x, Y]t,6 := r/J(y*x)
(x, YEA) yields a positive semi-definite Hermitian form on (the underlying
complex vector space of) A (Chapter II, Section 2, Example 5). Its kernel
At,6 := {x: [x, x]t,6 = O} is clearly closed, because r/J is continuous, and a left
ideal because of
0::::;; r/J(x*y*yx) ::::;; IlyI12r/J(x,x) (x, YEA),
which follows from (3.2)(i). Thus the quotient vector space AI At,6 is a pre-
Hilbert space on which [, ] induces an inner product.
If A is any CO-algebra and H is some Hilbert space, a homomorphism
n : A ----) f!'(H) is called a representation of A; the representation is faithful if
n is injective. n is called a cyclic representation if there exists ~ E H such that
the orbit {n(x)~ : X E A} is dense in H; ~ is called a cyclic vector of nand
usually assumed to satisfy II~II = 1.
It is also customary to denote by (n, H) a representation and by (n, H,~) a
representation with cyclic vector ~, the algebra A being understood as given.
Of course, if Hl and H2 are Hilbert spaces and U: Hl ----) H2 a bijective
(linear) isometry such that nl = U*n2U, then two representations (nj,Hj)
U= 1,2) of A are essentially the same and are called spatially isomorphic (or
unitarily equivalent); if nj are cyclic with respective cyclic vectors ~j, the
definition includes the requirement ~2 = U~l. The following fundamental
result goes back to 1. Gelfand and M. A. Naimark [1]. Its final version is due
to 1. Segal [1].
4.3
Theorem. (Gelfand-Naimark-Segal). Let r/J denote a state of the C*-
algebra A. There exists a cyclic representation (nt,6, Ht,6, ~t,6) of A satisfying
[nt,6(x)~t,6, ~t,6] = r/J(x) (x E A).
(nt,6, Ht,6, ~t,6) is unique to within spatial isomorphism.
Proof By the foregoing discussion, AI At,6 is a pre-Hilbert space under
[y,y] := r/J(y*y); its completion H is a Hilbert space in the obvious way.
Moreover, because At,6 is a left ideal, each x E A defines a linear map Tx :
y t-t xy on AIAt,6 satisfying
[TxY, TxY] = r/J(y*x'xy) ::::;; r/J(y. y)lIxll 2 = IIxIl 2[y,y];
its unique extension to H, denoted by nt,6(x), clearly satisfies IInt,6(x)1I ::::;; IIxli.
Since r/J = rP* implies nt,6(x)' = nt,6(x*) (x E A), nt,6 is readily seen to be a ho-
momorphism A ----) f!'(H), and hence a representation of A.
If A is unital with unit e, ~t,6 := e obviously is cyclic for nt,6. Moreover, we
274 c- AND W*-ALGEBRAS [Ch. VI
have [n¢(x)(¢, (¢l = tjJ(e*xe) = tjJ(x) (x E A). If A is no~ unital and ~ denotes
the unique positive extension of tjJ to the unitization A of A, (4.1) Corollary
1, the foregoing procedure yields a representation ni of A with cyclic vector
(ri' !f now (u;J;'EA denotes an approximate unit of A and e denotes the unit
of A, we obtain
2 - 2-
II(i - ni(u;')(ill = tjJ( (e - u;,) ) ~ tjJ(e - u;,)
for AE A. Since lim;, ~(e - u;,) = 0, this implies that ni' restricted to A, is a
representation of A with cyclic vector (i'
To obtain a faithful representation, it is generally insufficient to consider a
single state tjJ E 9"(A). However, since the state space 9"(A) separates A (see
above), the Lemma preceding 4.2 will furnish a faithful representation if we
"paste" the representations n¢ (tjJ E 9"(A)) together. Precisely, let H denote
the Hilbert direct sum (f)¢EY'(A)H¢ (Chapter II, Section 3, Example 5) and
define a representation n : A --> f£(H) by virtue of n(x) = (f)¢EY'(A)n¢(x),
where the operator n(x) E f£(H) acts "coordinatewise" on the elements
(rt¢)¢EY'(A) of H.
COROLLARY. For every C* -algebra A, there exists a faithful representation.
Proof We may assume A unital. It suffices to show that the direct sum
representation n = (f)¢EY'(A)n¢ is injective. But by the lemma preceding (4.2)
there exists, for given 0 i= x E A, some tjJ E 9"(A) satisfying 0 < tjJ(x*x) =
Iln¢(x)(¢11 2 • Thus n(x) i= O. .
We note that the representation (f)~EY'(A)n¢ is usually called the universal
representation of A.
5. PROJECTIONS AND EXTREME POINTS

Let A be any C* -algebra. A projection is an element pEA satisfying
p2 = P = p*; this implies p ::2: O. Under any representation n : A --> f£(H),
H a suitable Hilbert space, n(p) is a Hermitian (equivalently, orthogonal)
projection operator on H. Note also that a projection 0 i= pEA can be
characterized by the conditions p2 = p, Ilpll = I (Exercise 13).
First we turn to the relation of projections pEA with the extreme points
of certain order intervals of Asa. Here we need a concept from order theory
(cf. Chapter V, Section 3): A C*-algebra C c A is a hereditary (or Av
saturated) subalgebra if a E C, c E A and 0 :;; c :;; a implies c E C. Following
is an important example.
5.1
If pEA is a projection, then C := pAp is a hereditary C* -subalgebra of A.
Proof It is easy to see that pAp = {pap: a E A} is a closed subalgebra of
A invariant under the involution x --> x*. pAp is also unital: If A is unital
§5] PROJECTIONS AND EXTREME POINTS 275
with unit e then pep = p is the unit of pAp; if not, then p = pep where e
denotes the unit of A.
To show that C is hereditary in A, we can suppose A is unital. Now
let b E A satisfy O:$; b :$; pap for some a E A; then (3.2)(i) implies that
O:$; (e - p)b(e - p) :$; (e - p)pap(e - p) = O. Thus II(e - p)b(e - p)11 =
IIVb(e - p)11 2 = 0, which shows that Vb(e - p) = 0 and (e - p)Vb = O.
Multiplying by Vb from the left and right, respectively, we obtain b = bp
and b = pb. Therefore, b = pbp E C.
Now the following important result is an easy consequence.
5.2
Let p, q be projections in A. The following assertions are equivalent:
(a) q:$; p.
(b) pq = qp = q.
(c) p - q is a projection.
If p, q are projections, p + q is not a projection unless pq = O. However, if
A is unital and f!J is a commuting family of projections containing 0 and e,
+-
closed under the Boolean operations (p q) = p + q - pq and p. q = pq,
then f!J is called a Boolean algebra ofprojections or a resolution of the identity
e. Resolutions of the identity are the principal tool in the investigation of
self-adjoint and normal elements of A (see Exercise 18).
Let us recall (Chapter II, Section 10) that an extreme point x of a (non-
void) convex set C is a supporting manifold of dimension 0; equivalently, x is
in the extreme boundary GeC of C iff x = Aa + (1 - A)b, where a, bE C and
0< A < 1, implies x = a = b.
In what follows, we will denote, as before, by U the closed and by Uo the
open unit ball of A.
5.3
If A is unital, its unit e is an extreme point of U.
Proof Suppose that e = Aa + (1 - A)b where 0 < A < 1 and a, bE U. It
follows that ab = ba. Let B denote the commutative C* -subalgebra of A
generated by {e, a}. By Gelfand's Theorem 2.2, B is isomorphic to C(K), the
C* -algebra of continuous complex functions on some compact space K, and
e, a, bE U (lB. It is easy to see that in C(K), e = Aa + (1 - A)b implies e =
a = b, and the claim is proved.
COROLLARY. In a unital C* -algebra A, the unit e of A is an extreme point of
the order intervals [-e, e] = Un Asa and [0, e] = Un A+.
Sakai [1], proposition 1.6.1, has shown that, conversely, oeU ¥ 0 implies
that A is unital. Thus the presence of extreme points of U characterizes uni-
tal C* -algebras; we proceed to give an order theoretic characterization.
LEMMA. Let A be a C' -algebra. If p := SUp( Uo n A+) exists in A then p is

a projection.
°
Proof Clearly, if A is unital with unit e, then e = sup(Uo nA+); in fact,
°: ;
e = sup{2e: S; 2 < I}. Now suppose A is not unital and let (u;.).<eA be
an approximate unit of A (see 3.3). Since u.< S; p, pl/2u.<pl/2 S; p2 by
(3.2)(i). Since u.<pl/2 ---+ pl/2, by (3.2) (i), we have
which implies the assertion.
5.4
Theorem. A C* -algebra A is unital if and only if p := sup ( Uo n A+) exists;
if so then p is the unit of A.
Proof Suppose p := sup (Uo n A+) exists in A; by the lemma, p2 = p, and
it is clear that p is the unit of pAp. But since pAp is a hereditary subalgebra
of A by (5.1), it follows that the interval [0, p] c pAp contains Uo n A+.
Because A is the closed linear span (over C) of Uo n A+, we have A = pAp,
and p is the unit of A.
We finally turn our attention to the announced geometric characterization
of projections. If A is unital with unit e, from the definition of the order of A,
(3.1) et seq., it is clear that the order interval [0, e] = {a E A : 0 S; a S; e}
equals Un A+. Further, the affine function f : x ---+ 2x - e (x E A) maps
[0, e] onto [-e, e] = Un Asa. Hencefmaps the extreme boundary oe( Un A+)
onto oe(U nAsa), (The reader may want to illustrate this by considering the
case of a commutative A, i.e., by (2.2) the case A = C(K), K compact.)
The following result now gives a purely geometric characterization of
projections in A.
5.5
Theorem. Let A be any C* -algebra. For any projection pEA, the extreme
boundary oe[O, p] consists precisely of all projections q E A, q S; p.
Proof First we show that in a C* -algebra with unit e, every projection
pEA is an extreme point of [0, e], and conversely. Let p E A,p = p2 ;;:: 0.
Thenp is the unit of pAp and, a fortiori, in Un (pAp)+. Now by (5.1) pAp
is a hereditary subalgebra of A, hence [0, p] c pAp. Therefore, if p = 2a +
°
(1 - 2)b with < .Ie < 1 and a, bE Un A+, then a, b are both in pAp. By
(5.3) p is an extreme point of Un A+.
!
Conversely, suppose that p E oe[O, e]. Then p = (p2 + (2p - p2)) shows
°
that p = p2, because S; p2 S; e and 0 S; 2p - p2 S; e.
Second, let A be not necessarily unital and let p be any projection of A.
§6] W* -ALGEBRAS 277
[0, p] = {x E A : °: :;
Then p is the unit of pAp; because pAp is hereditary in A by (5.1), we have
x :::;; p} contained in pAp. Thus the first part of the
proof shows that any projection q, q :::;; p, is an extreme point of [0, p], and
conversely. This completes the proof of (5.5).
6. W* -ALGEBRAS
As the universal representation (corollary of (4.3)) shows, every C*-
algebra can be viewed as a C* -subalgebra of !l'(H), H a suitable Hilbert
space. But !l'(H) has a special property not shared by all C* -algebras;
namely, the Banach space !l'(H) (operator norm) is a dual Banach space
(Example 1). A C* -algebra whose underlying normed space is a dual Banach
space, is called a W* -algebra.
Examples
1. Let H #- {O} be any complex Hilbert space with inner product
[x, y] and denote by H ® H the completed projective tensor product
(Chapter III, Section 6) of H with itself; by the corollary of (III, 6.2) its
dual is the space flJ(H, H) of all continuous bilinear forms on H x H.
If, as usual, H is identified with its dual Banach space H', it must be
kept in mind that the map H -+ H', given by x 1---* fx where fAy) =
[y,x](y E H), is conjugate linear; therefore, the dual of H ®H
becomes the space flJs(H, H) of all continuous sesquilinear forms on
H x H. Clearly, then, rp E flJs(H, H) iff rp(x, y) = [Tx, y](x, y E H) for
some (unique) T E !l'(H); and rp -+ T is an isometry for the standard
norm rp -+ Ilrpll := sup{lrp(x, y)1 : Ilxll : :; 1, Ilyll : :; I} of flJs(H,H) and
the operator norm of !l'(H).
By (III, 6.3) and (III, 6.4), every element u E H ® H is of the form
u = 2:~1 A;X; ® y; and the norm of u is given by
00
lIulln = infL IA;lllxdlllYill

;=1
where the infimum is taken over all possible representations. Any

of these representations allows an interpretation of u as an operator
x 1---* it(x) := 2: A;[X, x;] y; that is independent of the particular repre-
sentation of u; however, Ilitll : :; lIuli n so the nuclear norm u 1---* lIulln is
stronger than the operator norm. Clearly it is compact, and it is well
known that it has a representation
it= Lf.1J;®e; (7)

;
where (f.1;) is a finite or null sequence of real numbers > and the °
sequences (/;), (ei) are orthonormal in H. f.11 are the non-zero
eigenvalues of u*u (or equivalently, uu"), counted according to their

multiplicity.
Finally, Iluli n = Li Ili. It is easy to see that the set Ll (H) of oper-
ators thus corresponding to H ® H is an algebra (in fact, an ideal of
fIl(H)) called the algebra of trace class operators; it is usually under-
stood to carry the nuclear norm.
The duality <Ll(H), fIl(H) thus established is hence defined by the
bilinear form
<u, T) = LAi[Txi, Yi];
i
it is a separated duality (Chapter IV, Section 1), and it is clear from the
preceding that fIl(H) is the norm dual of Ll(H); therefore, fIl(H) is a
W" -algebra. The (continuous) linear form uI-t <u, idH ) is called the
trace on Ll(H) and is denoted by u I-t tr(u). Ifu = u", we can choose
fi = ei in equation (7) and it follows that tr(u) = L Ili is the sum of all
eigenvalues of u(counted by multiplicity).
2. The (complex) algebra LOO(X,r.,Il) where Il is assumed totally a-
finite (or at least localizable) is a W"-algebra because it is the dual
Banach space of Ll(X,r.,Il) (Chapter II, Section 2, Example 2).
3. If W is a W' -algebra with predual V, it will be seen that every
a( W, V)-closed C* -subalgebra is a W* -algebra. Equivalently, the
a( W, V)-closure of any C* -subalgebra A c W is a W* -algebra.
4. Every finite-dimensional C* -algebra is a W* -algebra as well.
One of the major aims of this section is Theorem 6.8 (due to Sakai [1];
corollary 1.13.3) asserting that a given W*-algebra has a unique Banach
predual V (to within isometric isomorphism, of course). To this end, we
need to establish the a( W, V)-continuity of the maps x I-t x*, X I-t ax, and
x I-t xa (a E W) as well as some other results. However, we must ask
patience of the reader because the proofs require considerable effort. Until
unicity is established, for a given W* -algebra we denote by V anyone
Banach predual of W. Our proofs follow Sakai [1] with some significant
simplifications and clarifications.
6.1
Let F denote a dual Banach space with predual Fo. A linear map. : F ---- F
is a(F,Fo)-continuous whenever its restriction .Iu to the closed unit ball U of
F is a(F, Fo)-continuous.
Proof It • is a(F, Fo)-continuous on U, then so is rp 0 • for every rp E Fo.
Thus (rp 0 .)-1 (0) n U is a(F, Fo)-closed hence, by the Krein-Smulian theo-
rem (IV, 6.4), the hyperplane (rp 0 .)-\0) is closed for a(F, Fo), i.e., rp o. =
.' (rp) is a(F, Fo )-continuous. Thus for each rp E Fo, rp 0 • = .' (rp) E Fo, which
shows that. = ." is a(F, Fo)-continuous.
§6] W' -ALGEBRAS 279
COROLLARY. Under the assumptions of (6.1), let F = F, + F2 be the direct

algebraic sum of the a(F, Fo)-closed subspaces F" F2. Then the sum is
a(F, Fo)-topological, i.e., the associated projections n : F --t F, and 1 - n are
a(F, Fo)-continuous.
Proof First, by (III, 2.1) Corollary 3, n is norm continuous on Fwhence
n(U) c IlniIU. By (6.1) it suffices to show that n is a(F, Fo)-continuous on
U; let ty denote a filter on U a(F, Fo)-convergent to XE U. Since IInilU is
compact and F, closed for a(F, Fo), n(ty) has an adherent point y E F, n
IInil U; clearly, then, Z = x - y is adherent to (1 - n)(ty) and Z E F2 as F2 is
a(F, Fo)-closed. Now x = y + Z shows that y is the unique adherent point of
n(ty) in the compact space IInil U. By a well-known result on compact spaces,
this implies that lim n(ty) = y = n(x).
In the following, let W be any W' -algebra, Wsa the self-adjoint part of
W, and W+ the positive cone {x E W : x = x* ~ O}; V denotes a predual
of W.
6.2
The subsets Wsa and W+ are a(W, V)-closed; in particular, W = Wsa +
iWsa is a a(W, V)-direct topological sum over R. Morever, involution x f--+ x*
is a( W, V)-continuous.
Proof In the subsequent proof, we may and Will assume W unital with
unit e; if not, the unitization W of W is clearly a W' -algebra containing W
as a a-closed subalgebra.
To show that Wsa is a(W, V)-closed, by (IV, 6.4) it is enough to show
Wsa n U is closed, U the closed unit ball of W. Denote by ty a filter on U n
Wsa a( W, V)-convergent to some x = a + ib where a, b E Wsa and b =t- O.
Then u(b) =t- {O} and considering -ty in place of ty if necessary, let A = IIbil E
a(b). Because A :::;; 1, a = 1 ;/2 + 1 is positive and 1 + a2 < (A + a)2 holds.
Now for any y E Un Wsa we have
IIy + iae ll 2 :::;; 1 + a 2 < (A + a)2 :::;; lib + aell 2 :::;; IIa + i(b + ae)II2. (8)
But ty + iae converges to a + i(b + ae); by (8) ty + iae contains the a( W, V)-
compact ballJl + a 2U. Hence IIa + i(b + ae)II :::;; ~, which contra-
dicts (8). Therefore, b = O.
Now W+ n U =!
(U n Wsa + e) is a(W, V)-closed as well; again from
(IV, 6.4) it follows that W+ is a(W, V)-closed. Finally, the a(W, V)-
continuity of x --t x* follows from (6.1) Corollary.
COROLLARY. Every W*-algebra =t- {O} is unital.
Proof The set Uo n W + (Uo the open unit ball of W) is directed (:::;;) by
(3.3); a(W, V)-compactness of Un W+ implies that Uo n W+ has an ad-
herent point p for a(W, V). Now p = sup(Uo n W+) by (V, 4.2), because
W+ is a(W, V)-closed. So (5.4) showsp to be the unit of W.
6.3
Every W* -algebra W is order complete; more precisely: For each directed
(:~) order bounded set S c W+, the filter of sections a( W, V)-converges to
supS. In addition, sup(a*Sa) = a*(supS)a for any a E W, and accordingly
for sets S directed (;;:::).
Proof We may and will restrict attention to upward directed sets S, i.e.,
sets directed (:~). Clearly, if such S is majorized then its section filter
ty(S) =: ty contains some order interval [x, y] = (x + W+) n (y - W+).
Because [x, y] is a(W, V)-closed and bounded, it is a(W, V)-compact; thus
ty has an adherent point z E [x, y] and necessarily, z = sup S. But on a com-
pact space, a filter with a unique adherent point is convergent, and so
lim ty = z for a( W, V).
If a E Wand a-I exists, then x ::;; y {:} a*xa::;; a* ya by (3.2) (i). Thus in
this case, a*za = supa*Sa. We reduce the case of an arbitrary a E W to the
preceding by the following device. For XES, Z = supS and u = (z - x)I/2a,
we obtain u*(z - x) 1/2 = a*(z - x). Thus if rp E V is any positive linear form
on W, Schwarz' inequality yields
Irp(a*(z - x))1 2 ::;; rp(u'u)rp(z - x).
As x ranges over S and a E W stays fixed, rp(u'u) remains bounded whence
rp(a*z - a*x) converges to 0; similarly, rp(za - xa) ~ O. But since W+ is a
a( W, V)-closed (proper) cone in W, the set V+ of all positive linear forms in
Vis total by the bipolar theorem (IV, 1.5), Corollary 2, hence its (complex)
linear hull dense in V. So a(W, V+) is a Hausdorff topology coarser than
a( W, V) and thus agrees with the latter on bounded sets. It follows from this
that lima*ty = a*z and lim tya = za for a(W, V).
Moreover, if t> Iiall then (t + a*m(t + a) a( V, W)-converges to
(t + a*)z(t + a) since (t + a) is invertible (see above); since lim t 2ty = t 2z, it
finally follows that lima*tya = a*za. Since {a*xa : XES} is directed (::;;) by
(3.2) (i), we conclude that supXESa*xa = a*(supS)a for all a E W.
Recall from general topology that a compact space K is totally discon-
nected if K has a base of open-and-closed subsets, or equivalently, if the
simple continuous functions K ~ C (i.e., the continuous functions with
finite range) are norm dense in C( K). K is called extremally disconnected
(Stonian) if the closure G of each open G c K is open; clearly a Stonian
space is totally disconnected. Now if A is a commutative, order complete
CO-algebra then by Gelfand's theorem (2.2), A is isomorphic to some C(K)
with the so-called spectrum space K Stonian (Exercise 20). Since the charac-
teristic functions of open-and-closed subsets are continuous, it follows that
the projections of an order complete algebra C(K) form a total subset with
respect to the norm topology. Hence we obtain the following corollary of

(6.3).
COROLLARY. In any W* -algebra w., the set of all projections is total
with respect to the norm topology; in particular, every positive x E W can
be . approximated by linear combinations of projections with positive real
coefficients.
Proof. Let 0 < x E W; by Zorn's lemma, there exists a maximal commu-
tative C' -subalgebra A containing x. Let SeA be directed (::::;;) and let
b = supS E W. For every unitary element u E A we have, by (6.3), u*bu =
sup(u* Su) = sup S = b; thus b commutes with A by (2.2), Corollary 2, hence
bE A. From the discussion preceding the corollary, it follows that the spec-
trum space K of A is Stonian; this proves the first assertion. The proof of the
second assertion is left to the reader.
We now supplement proposition (5.1) with respect to W*-algebras.
6.4
Let p denote any projection in the W' -algebra W. Then the subalgebra
p W p is 0"( W, V)-closed (and in fact, a W* -algebra); moreover, the mapping
x f--+ pxp is 0"( W, V)-continuous.
Proof. By (5.1), pWp is a hereditary C*-subalgebra of W; hence the order
interval [-p,p], which equals the closed unit ball of (pWP)sa is O"(W, V)-
closed and thus compact. Hence by (IV, 6.4) pWp is O"(W, V)-closed in W
and so its own bipolar with respect to the duality (W, V); it follows that
p W P is the Banach dual of V / (p W pt. Thus p W p is a W' -algebra.
The mapping n : x f--+ pxp is a projection operator on W with range p W p;
its 0"( W, V)-continuity follows from (6.1), Corollary if we can show that
pWp as well as the kerneln-I(O) = (e - p) W + W(e - p) are O"(W, V)-
closed. For pWp this was just shown; for n-I(O) it suffices, again by (IV,
6.4), that n- I (0) 11 U (U the unit ball of W) be 0"( W, V)-closed. Let q =
e - p. We first assert that if ~ is a filter on W containing p[-e, e]q and
O"(W, V)-convergent to a, then pap = qaq = O. For b E p[-e,e]q,n E Nand
P E C, Ipi = 1 we obtain
(pbq + pnp)(pbq + pnp)* = pbqbp + n 2 p;
hence
Ilpbq + pnpl12 = IIpbqbp + n 2pll ::::;; 1 + n 2.
Also because p[-e, e]q E U it follows that lIa + npl12 ::::;; 1 + n 2. Suppose that
pap #- O. Then
1 -i
either c := 2: (pap + pa* p) #- 0 or d := "2 (pap - pa* p) #- O.
If c =1= 0, there exists 0 =1= A ERin O'(c); we can suppose A> O. From the
preceding inequalities it follows that, for any n E N,
1 +n 2 2: Ila+npl12 2: IIp(a + ne)pl12 2: IIc+np112 2: (A+n)2,

which is impossible for large n. Thus c = 0 and, likewise, d = O. We have
shown that pap = O.
Now suppose that qaq =1= 0 and let b E p[-e, e]q, 11 E C, 1111 = 1, and n E N.
We obtain Ilpbq + nMI1 2 = Ilqbpbq + n 2qll ~ 1 + n 2 which, in analogous
fashion, is contradictory for large n.
Thus qaq = 0 and a = paq + qap E Wq + q W; this implies that the
O'(W, V)-closure (taken in W) of p[-e,e]q is contained in Wq+qW, and
so is the a-closure of ip[-e, e]q; therefore the unit ball of Wq + q W is
O'(W, V)-compact. By (IV, 6.4) it follows that Wq + qW is O'(W, V)-closed
in W.
From the preceding proof, let us retain this result.
COROLLARY. Every 0'( W, V)-closed C* -subalgebra of W is a W* -algebra.
To finally prove 0'( W, V)-separate continuity of multiplication in W, we
need one more lemma.
LEMMA. IfP is any projection in Wand q = e - p, then the subspace p Wq is
0'( W, V)-closed.
Proof Let !j be a filter on W containing pUq and convergent to a E W;
exactly as in the proof of (6.4), it is shown that a = paq + qap. If qap = 0,
then pUq is closed in W, hence O'(W, V)-compact; (IV, 6.4) then implies
that p Wq is 0'( W, V)-closed in W
We now show that assuming qap =1= 0 is contradictory. For arbitrary ele-
ments b, b ' E W let u = pbq and v = qb'p. Then u*v = 0 = uv* and hence for
any 11 E C, Ilu + I1vll = max{llull, 111111vll} (cf. Exercise 6). Letting u = paq,
v = qap we obtain (n EN)
Iia + nvll = Ilu + (n + l)vll = max{llull, (n + l)llvll}. (9)
Moreover, for any bE U we obtain Ilpbq+nvll ~max{1,nllvll}. By
hypothesis, a is 0'( W, V)-adherent to p Uq; since the ball {x E W : IIxll ~
max{l,nllvll} is O'(W, V)-closed in W, we must also have
Iia + nvll ~ max{l, nllvll}. (10)
But for large n E N, (10) contradicts (9), completing the proof.
6.5
Theorem. In any W* -algebra W, multiplication is separately continuous
with respect to O'(W, V).
Proof Letp be any projection in W, and let q = e - p. From (6.4) and the
§6] W'-AI..GEBRAS 283
preceding lemma we know that the subspaces pWp, pWq, qWp, qWq are
a( W, V)-closed in W; now W is their algebraic direct sum and hence by re-
peated application of the corollary of (6.1) the sum is a( W, V)- topological
(cf. Exercise 17). But the kernel of x ---t px is q Wq + q W p, hence closed;
by the same corollary, we conclude that x ---t px is a( W, V)-continuous,
and so is x ---t .xp. Separate a( W, V)-continuity of multiplication amounts
to a( W, V)-continuity of the linear maps La : x ---t ax and Ra : X ---t xa, for
each a E W. By the preceding, Lb is a( W, V)-continuous when b is a linear
combination of projections, and so is Rb. But by the corollary of (6.3) a
given a E W is the norm limit of such elements b, that is, La is the limit of
mappings Lb uniformly on the unit ball U of W. Thus Lalu is a(W, V)-
continuous and from (6.1) it follows that La is a( W, V)~continuous; clearly,
so is R a .
In the remainder of this section, we shall be concerned with a closer study
(and, in fact, an identification) of the predual V of an arbitrary W* -algebra,
in terms of Walone. We begin by observing that from W = Wsa EB iWsa
(a(W, V)-direct sum over R by (6.1)) it follows that V = VI EB V2 where
<
VI, V2 are the polars of iW.m, Wsa , respectively, with regard to W, V).
Thus if rp = !PI + P2 E VR where Pi E Vi are a( W, V)-continuous real
linear forms on W, then ip(a + ib) := !PI (a) + irpl(b) (a,b E W sa ) defines a
self-adjoint linear form ip on W (see Section 4); conversely, if ip E V is self-
adjoint then its restriction to Wsa is real-valued. It is thus natural to define
Vsa := VI, and it easily follows that V2 can be identified with iVsa .
Consequently, the set V+ of all a( W, V)-continuous positive linear forms
(Section 4) on the W*-algebra Wequals Vsan(-W+t. Moreover, the
a( W, V)-continuous maps x f--+ xa, x f--+ ax (a E W) have adjoints on V with
<
respect to w, V); we denote these by rp ---t ap and rp ---t !pa, respectively.
The notation is justified as V can be viewed as a module over W.
6.6
The set V+ of all a( W, V)-continuous) positive linear forms on W is a
normal) strict m-cone in Vsa; in particular) V equals the complex linear hull of
V+. Moreover) the mappings rp ---t arp, rp ---t rpa (a E W) are (norm and weakly)
continuous on V Finally the mappings x f--+ xrp, X f--+ rpx are continuous from
(W,a(W, V» into (V,a(V, W».
Proof The first assertion is an immediate consequence of Theorem (V,

3.5); the remainder follows from the foregoing discussion and from (IV, 7.4).
If rp is a a(W, V)-continuous, positive linear form on W (i.e., if rp E V+)
then x f--+ Ilxllq.> := y'rp(x*x) is easily seen to be a semi-norm on W (cf. the
proof of (4.3)); the (locally convex) topology generated by the set of these
semi-norms, as rp runs over V+, is called the strong topology of W (cf.
Chapter II, Sections I and 4), and denoted by s( W, V).
284 C*- AND W*-ALGEBRAS ECho VI
6.7
For any W*-algebra W, we have u(W, V) c s(W, V) c r(W, V). More-
over, multiplication in W is (jointly) s( W, V)-continuous on all sets B x W,
B bounded in W
Proof For each rp E V+, from Schwarz' inequality we obtain (e the unit of
W)
Therefore, u(W, V) is coarser than s(W, V).

By (IV, 3.2) Corollary 1, the Mackey topology r(W, V) is the topology of
uniform convergence on all u(V, W)-compact subsets of (the Banach space)
V. Now if iY is any filter on U that r(W, V)-converges to 0, then for
°
any fixed rp E V+ lim iY = uniformly on Urp = {yrp : y E U}, because Urp
is u(V, W)-compact by (6.6). It follows that limiJ IIxll; = limiJ(x*rp)(x)
= 0. Therefore, every s(W, V)-continuous linear form on W is r(W, V)-
continuous on U, and thus on W(cf. 6.1); we conclude that s(W, V) is con-
sistent with <W, V) and so, s(W, V) c r(W, V).
The (joint) continuity of (x, y) ~ xy for s(W, V) now results from the
following inequalities:
Ilxy - uvllV' ~ Ilx(y - v) IIV' + II(x - u)vllV',
IIx(y - v)lI; = rp«y - v)*x*x(y - v» ~ Ilxx*lllly - vii;,

and
II(x - u)vll; = rp(v*(x - u)*(x - u)v) = Ilx - ullv*V'v;
the second of these rests on Schwarz' inequality applied to rp E V+.
We can now supplement (6.3) and its corollary.
6.8
In every W* -algebra W, the set P of all projections is a complete lattice
under the ordering induced by Wsa.
Proof Let p, q E P; we show that p /\ q := inf(p, q) exists. By (6.4),
pWpnqWq is a W*-subalgebra A of W, and obviously hereditary (5.1).
Hence sup([O, pj n [0, q)) = r exists; r is a projection (lemma preceding (5.4))
and by (5.4) the unit of A. Clearly, then, A = rWr and r = p /\ q. Denoting
bye the unit of W, we now obtain p v q = e - «e - p) /\ (e - q».
If S denotes a directed (~) set of projections, its filter iY of sections
u(W, V)-converges to z=supS by (6.3). Letting u=z-x (XES) we
obtain, by Schwarz' inequality, for any fixed rp E V+:
liz - xII; = rp(u 2 ) = rp(u 3/ 2 U 1/ 2 ) ~ rp(u 3 )rp(u) = rp(u 3 )rp(z - x).
Since cp( u 3 ) stays bounded as a x ranges over S, lims x = z for s( W, V). But
multiplication is continuous on U x U by (6.7), whence lims x = lims x 2 =
z2 = z for s( W, V). It follows that z is a projection and hence, P is order
complete.
REMARK. The lattice P is complemented but, in general, not dis-
tributive; thus P is generally not a Boolean algebra. However, iff W is
commutative, W ~ C(K) (cf. remarks after (6.3)), then P is isomorphic
to the Boolean algebra of all open-and-closed subsets of K.
A linear form cp on an ordered vector space E (in particular, on a C*-
algebra A) is called order continuous if for every directed (::;;) set SeE,
sup S = z implies lims cp(x) = cp(z). An order continuous positive linear form
on a C* -algebra is called normal.
We can now prove the announced result of Sakai [1] (theorem 1.13.2).
6.9
Theorem. Let W be any W* -algebra with predual V A positive linear form
cp on W is normal if and only if cp is a( W, V)-continuous (i.e., iff cp E V+).
Proof If cp E V+ then cp is normal by (6.3). So suppose that cp is normal.
(i) Let Po be the set of all projections pin Wfor which cpp : x f--> cp(xp) is
a( W, V)-continuous. If S is a directed (::;;) subset of Po, then r :=
°
sup S is in Po; we have limq E s cp(r - q) = by hypothesis. Letting
u = (r - q)I/2x for x E Wsa n U and v = (r - q)I/2, Schwarz' inequal-
ity gives
Icp((r - q)X)12 = Icp(v*u)12 ::;; cp(u*u)cp(r - q) ::;; cp(e)cp(r - q),
which means that lims cpq = cpr uniformly on Un Wsa; thus cpr E Po.
Hence, for cardinality reasons, there exist maximal well-ordered
(::;;) families (qrJ in Po such that IY. < fJ implies qrx < qp; the su-
premum of any such family then is a maximal element P E Po.
(ii) If p = e, the proof is finished, so suppose that p < e. There exists 1/1 E
V+ satisfying cp(e - p) < I/I(e - p). Fixing 1/1 we claim that there exists
q E P satisfying q < e - p and cp(r) < I/I(r) whenever rEP, r < q.
If not, there exists a maximal element in Qo, where Qo is the set of
projections satisfying q ::;; e - p and cp(q) 2 I/I(q). This is shown as in
(i). If ro is maximal in Qo we must have ro = e - p. Otherwise, there
°
exists yet another projection rl, satisfying -# rl < e - p - ro and
cp(rJ) 2 I/Ih)· But this implies cp(ro + rl) 2 I/I(ro + rl), contradicting
the maximality of roo
(iii) Now (ii) implies that if p -# e, . there exists q > 0, q E P such
that cp(r) < I/I(r) for all projections r < q; by the corollary of (6.3)
this implies cp(x)::;; I/I(x) for all x E (qWq)+ (cf. (6.4)). Now
IIP(xp + xq)1 ~ IIP(xp) I + IIP(xq)l, and III' (xq) I ~ VIP(e)VIP(qx*xq) ~

VIP(e)vt/I(qx*xq); thus
IIP(x(p + q))1 ~ IIP(xp) I + Vr/J(e) . vt/I(qx*xq),
which shows lP(p + q) to be in Po. Since q > 0, this contradicts the
maximality of p and hence, the assumption p "# e is absurd.
COROLLARY 1. For any W' -algebra, the predual V is unique; more pre-
cisely, V is isometrically isomorphic to the space W~ generated by all normal
linear forms on W.
Proof Denoting by (W~)+ the cone of all normal linear forms on W, the
bijection To: V+ -+ (W~)+ extends to a linear bijection T: V -+ W~. But
the respective unit balls in V and W~ are the polars of the unit ball U of W
with respect to the dualities <W, V) and <W, W~); hence T is an isometry.
In particular, W~ is a Banach subspace of W'.
An isomorphism T: WI -+ W2 of two W*-algebras is called a . W*-
isomorphism if T is a homeomorphism for a(WI' VI) and a(W2' V2);
equivalently, if T has a pre-adjoint mapping V2 isometrically onto VI.
COROLLARY 2. Every commutative W* -algebra W is W* -isomorphic to a
space LOO(fl) where fl is a Borel measure on some locally compact space.
Proof By (2.2), W is isomorphic to a (complex) space C(K); we can re-
strict attention to the self-adjoint part CR(K). Identifying the predual VI of
CR (K) with the space of all (self-adjoint) order continuous linear forms, VI
is a lattice ideal of the dual Banach lattice C~(K); in fact, VI is an AL-space
(cf. Chapter V, Section 8 and Exercise 22). Now it is well-known (but not
immediate) that every AL-space is isometrically lattice isomorphic with
LI(fl), fl denoting a Borel measure on some locally compact space X. (For
details, see Exercise 21). Hence CR(K) is the Banach (and order) dual of
4(fl), and it follows that Wis W*-isomorphic with LOO(fl).
As an example for and an application of Theorem 6.9, let us give an
explicit representation of all normal linear forms on ff(H), H any Hilbert
space.
Example
5. (We use the notation and results of Example 1 above.) Let W =
ff(H), H any Hilbert space. Following Example 1, the predual V
(unique by (6.9)) of ff(H) is V = LI(H), and the predual (over R) of
Wsa is Vsa = {II' E V: IP(Wsa ) c R} (cf. discussion preceding (6.6)). For
any x E H, II' = x <8l x E V+ because <x <8l x, T) = [Tx, xl ~ 0 when-
ever T E W+; conversely, if T E ff(H) and [Tx, xl ~ 0 for all x E H
then T E W+. Therefore V+ is the bipolar, with regard to the real
duality <Wsa , Vsa ), of the convex set C of all finite sums E AjXj ® Xj
§7] VON NEUMANN ALGEBRAS. KAPLANSKY'S DENSITY THEOREM 287
where Ai > 0 and Xi E H. It follows that V+ is the norm closure C in

L 1 (H) and contains positive trace class operators only. On the other
hand we have, by Example 1, for each such operator u a representation
u = 'E Piei ® ei where {JlJ is the set of all eigenvalues of u, counted by
multiplicity, and {ei} a corresponding orthonormal sequence of
eigenvectors. Since Iluli n = 'Ei Pi' it follows that V+ = C is the set of
all positive trace class operators on H.
By (6.9), then, rp E W' is a normal linear form on 2(H) if and only
if
00
rp = LJliei ® ei
i=1
where Jli ;;::: 0, 'Ei Jli < 00, and {ei} is some orthonormal sequence in H.
Finally, Ilrpll = rp(idH ) = Li Jli·
7. VON NEUMANN ALGEBRAS. KAPLANSKY'S DENSITY THEOREM
In this section we consider an important class of C* -subalgebras of 2(H),
where H is a Hilbert space -:f. {O}. In addition to the standard norm topology
of 2(H) and the ultra-weak (or a-weak) topology a(2(H), H® H) (Section
6, Examples 1 and 5), we consider two so-called operator topologies: the
weak operator topology a(2(H), H ® H), generated by the semi-norms
u ~ l[uc;,,,]I(c;,,, E H), and the strong operator topology, generated by the
semi-norms u -+ Iluc;II(c; E H). Endowed with one ofthese topologies, 2(H)
will be denoted by 2u(H), 2w(H), and !l's(H), respectively.
Viewing 2(H) as the space of continuous linear forms on H® H (see
Section 6, Example 1), the three topologies, in the order cited, are 6-topologies
(Chapter III, Section 3) for the following families: ultra-weak for 6 the finite
sets in H ® H, weak for 6 the finite subsets of H ® H, and strong for the
family 6 = {F ® B : F finite, B bounded in H} of subsets of H ® H. (The
norm topology, of course, corresponds to 6 = {B® B,B bounded in H}.)
The valid inclusion relations for these topologies are immediate from this
characterization; let us note that 2s(H) and 2w(H) have the same dual
H ® H by virtue of Chapter IV, Theorem 4.3, Corollary 4, and hence the
same closed convex subsets. Finally, the reader should be cautioned not to
confuse the strong operator topology with the strong topology s(2(H),
H ®H) used in Section 6.
REMARK. With the aid of Section 6, Examples I and 6, it is quickly
seen from Theorems 3.3 and 3.5 in Chapter V that the cone 2(H)+ of
positive Hermitian operators is normal for all of the topologies con-
sidered, and that the respective dual cones are strict 6-cones.
In working with the strong operator topology, the following device, called
amplification, is frequently useful. If K is the Hilbert direct sum (Chapter II,
288 C'- AND W*-ALGEBRAS [Ch. VI
Section 2, Example 5) of n(E N) copies of H, and if 1T:n : !£,(H) --+ !£,(K) is

defined by 1T:n(x) ((1, ... ,(n) = (X(I, ... ,X(n) then we have
sup Ilx(dl~ll1T:n(x)((I"",(n)ll~n sup IIX(ill·
1 :<=;i:<=;n 1 :<=;i:<=;n
Verification is immediate.
Let A be an algebra. For any X c A, the subset XC = {y E A : xy = yx
for all x E X} is called the commutant of X. If A is a C* -algebra and X is
self-adjoint (i.e., X = X*), then XC is a C* -subalgebra, containing the unit
whenever A is unital. Clearly, X --+ XC is anti-tone with respect to inclusion
c, and XC = XCcc. In particular, if A = !£,(H) and X = X* it is quickly seen
that XC is a C* -subalgebra containing idH and closed in the weak operator
topology.
The following result is von Neumann's famous bicommutant (or double
commutant) theorem.
7.1
Theorem. Let M be a C*-subalgebra of !£'(H) containing the unit e = idH
of !£,(H). These assertions are mutually equivalent:
(a) M= MCc,
(b) M is closed in !£'w(H),
(c) M is closed in !l'.,(H),
(d) M is closed in !£'(J(H).
Moreover, M = M CC implies that M is a W* -algebra.
Proof The implications (a) '* '*
(b) (c) are covered by the preceding
discussion. (b) oR (d) follows from the Krein-Smulian theorem (IV, 6.4),
because the unit ball U of !£'(H) is a(!£'(H) , H@ H)-compact and the weak
operator topology a(!£,(H) , H <8> H) is a coarser Hausdorff topology.
(c)'* (a): Let ( E H be arbitrary, and denote by p E !£'(H) the orthogo-
nal projection of H onto M(, where M(:= {X( : x EM}. Because e E M,
we have P( = ( whence xp = pxp for all x E M; but then px = (x*p)* =
(px*p)* = pxp = xp, i.e., p E M C. Now if y E M CC then py = yp and hence
y( E M(. Thus for each e > 0 there exists x E M satisfying II(y - xgll < e. If
{(j : 1 ~ j ~ n} is a fixed finite subset of H we use that amplification device
(see above): We have 1T:n (MCC) = 1T:n(M)CC, and we apply the preceding
method to the element (= ((1,"" (n) of (f}jH. Then for y E M ec there
exists x E M satisfying II1T: n (Y - xgll < e, which implies II(y - x)(jll < e
for all j = 1, ... ,no Finally, if M = M CC then M is a-weakly closed by (d)
and the claim follows from the corollary of (6.4), letting V = H@ H.
We now define a von Neumann algebra as a C*-subalgebra of !£'(H)
containing the unit e = idH E !£,(H) and satisfying anyone (hence all) of the
assertions (a)-(d) of (7.1). Let us note three important corollaries of (7.1); we
term a subalgebra A of !£'(H) non-degenerate if for each 0 "# ( E H there
exists a E A such that a( "# O.
§7] VON NEUMANN ALGEBRAS. KAPLAN SKY'S DENSITY THEOREM 289
COROLLARY 1. For any non-degenerate C* -subalgebra A c !l'(H) the clo-

sure A in !l's(H) equals the double commutant A CC .
Proof By the corollary of (6.4), A is a W*-algebra, and A "# {O}; hence A

is unital by the corollary of (6.2). Its unit p is a projection in !l'(H) satisfying
px = xp = x for x E A; because A is non-degenerate, we must have p = e =
idH •
COROLLARY 2. If A is any C* -algebra, the Banach bidual A" of A is capable

of a W*-structure'mder which A is a C*-subalgebra of A".
Proof Let n: A --+ !l'(H), where n = EB{nq> : rp E 9'(A)} and H =

EB Hq>, denote the universal representation of A (cf. 4.3 et seq.); unless
A !. {O} by its construction n(A) isnon-degenerate. For fixed rp E 9'(A) let
~ = (~o/I) denote the element of H defined by ~o/I = 0 if t/J "# rp, ~q> being the
<
cyclic vector of nq>. Consider the duality L 1 (H), !l' (H) (Section 6, Examples
I and 5) and let n* stand for the adjoint of n restricted to Ll(H) c !l'(H)'.
Since n*(~ ®~) = rp and rp = 9'(A) was arbitrary, it follows that n*(L 1(H))
contains the state space 9'(A) of A. Thus n*(L 1(H)) = A' and by (III, 2.1),
n* is a topological homomorphism of Ll(H) onto A'. But Iln*11 ::;; I, and it
follows that n* is a metric homomorphism (cf (4.2) and Section 6, Example
5) as well.
Now the kernel ofn* is n(A)o (polar with respect to <Ll(H), !l'(H)); but
n(A)o = (n(AtC)o because by Corollary I, n(AtC is the a-weak closure of
n(A). Then by (IV, 4.1) Corollary I, n~ is a a-weak and norm isomorphism
of A" onto n(AtC c !l'(H). This isomorphism, called canonical, establishes
the claim.
COROLLARY 3. Every W*-algebra W"# {O} is W*-isomorphic to a von
Neumann algebra on a suitable Hilbert space.
Proof Since the predual Vof W (cf. 6.9) separates W, so does the set
9'n(W) of all normal linear forms of norm I (set of normal states); conse-
quently, the direct sum EB ef/' (W)nq> of representations yields a faithful rep-
resentation n of W. We le~ve "It to the reader to verify that n is a homeo-
morphism with respect to a(W, V) and the a-weak topology.
Examples
1. If A is a self-adjoint subalgebra of !l'(H), H any Hilbert space
"#{O}, its commutant A C is a von Neumann algebra, because eEAC
and A Cis closed in !l's(H).
2. If A = !l'(H) and pEA is a (orthogonal) projection then pAp
(Section 5) is closed in !l's(H) but, by our definition, not a von Neu-
mann algebra unless p = e (cf., however, Murphy [I], p. 116). On the
other hand, pAp + Ce is a von Neumann algebra.
The double commutant theorem (7.1) can be viewed as a density theorem

for M with respect to MCC. Our second objective in this section is a much
stronger density theorem (due to I. Kaplansky [1]). To this end, we need two
lemmas and a theorem on operator-valued functions (7.2).
LEMMA 1. Let u be a unitary element of M, M denoting a von Neumann
subalgebra of !£,(H). There exists a self-adjoint element v E M such that u =
exp(iv).
Proof Let A denote the C*-subalgebra of !£'(H) genQrated by u; it con-
tains e and is commutative. Thus its closure A in !£,s(H) is commutative and
a W*-algebra; by (6.9) Corollary 2, A ::::,; L'XJ(X,'L,f.-l). Thus u corresponds to
f E L 00 (X, 'L, f.-l) satisfying If I = h; we leave it to the reader to show that
f = exp(ig) for some real-valued g E L oo (X,'L, f.-l).
LEMMA 2. The involution x --+ x' is continuous for the strong operator
topology of !£'(H) when restricted to the set of all normal elements.
Proof For arbitrary elements x, y E !£,(H) we have
(x* - y*)(x - y) = x*x - y* y + y*(y - x) + (y* - x*)y.
For C; E H this yields
Ilxc; - yC;11 2 :::;; Ilxc;1I 2 - IIYC;1I 2 + 211c;IIII(y* - x*)yC;lI·
For normal elements x, y we also have Ilxc;11 = Ilx*c;11 and Ilyc;11 = Ily*c;II·
Therefore,
Thus if y = limff x along some filter ~ containing the set of normal elements
in !£'(H) and convergent for the strong operator topology, it follows that
y* = limi;' x*.
Let us stop briefly to consider the continuity properties of multiplication
m : (x, y) f---> xy in !£,(H). From the sub-multiplicativity of the norm (Sec-
tion 1) it follows at once that m is norm continuous. In Section 6 it was
shown, with considerable effort, that m is separately continuous for the
CT-weak topology (Theorem 6.5). Also, m is separately continuous for the
strong and weak operator topologies (straightforward verification). We
supplement these results as follows, referring to these four topologies as the
standard topologies on !£,(H).
7.2
The respective families !8 of bounded subsets of !£'(H) are identical for all
of the standard topologies. With respect to the strong operator topology, mul-
§7] VON NEUMANN ALGEBRAS. KAPLANSKY'S DENSITY THEOREM 291
tiplication is (~, ~)-hypocontinuous (i.e., continuous on sets B x 9!(H),

9!(H) x B) and uniformly continuous on B x B for all B E ~.
Proof Denote by ~j (j = 1, 2, 3,4) in this order, the bounded sets for: the
norm, a-weak, strong, and weak operator topology. Then ~1 = ~2 because'
9!(H) is the Banach dual of H® H; ~1 = ~3 is the principle of uniform
boundedness; ~3 = ~4 be~ause these topologies have the same dual H ® H.
The final assertion is immediate from (III, 5.2) and (III, 5.3).
We now tum to the announced continuity theorem for certain mappings
of 9!(H)sa into 9!(H). Let f : R -+ C be continuous; for each self-adjoint a,
we define f(a) := fo(a) where fo denotes the restriction off to a(a) by virtue
of (2.2), Corollary 1. One has Ilf(a)11 ::;; Ilfoli oo Iiali. The question is under
what conditions onfthe map a -+ f(a) is continuous for the strong operator
topology; functions having this property are termed strongly conti.nuous.
7.3
Every bounded continuous function R -+ C is strongly continuous. More-
over, if f and 9 are strongly continuous and f is bounded, then fg is strongly
continuous.
Proof
(i) Denote by C the set of all strongly continuous functions on R.
Clearly, C #- 0 and, obviously, C is a vector space over C. More-
over, if IIfll oo < 00 and f E C then {f(a) : a E B} is bounded in
9!(H) whenever B is bounded in 9!(H)sa (cf. 7.2). Therefore, if
f E C, 9 E C andfis bounded we have fg E C.
(ii) We now show that Co(R) c C. Let f, 9 denote the functions given
by f(t) = (1 + t 2)-1 and g(t) = tf(t)(t E R). We have If I ::;; lR and
Igl ::;; lR. Let a, bE 9!(H)sa and observe that Ila(e + a 2)-111 ::;; 1 and
II(e+a 2)-III::;; 1 (cf. (2.2), Corollary 1). For any (EHwe obtain
This clearly implies that 9 E C. Of course, idR : t ~ t is in C; since 9 is

bounded we have hE C where h(t) = tg(t), and so f = 1 - h E C.
The function algebra (over C) generated by {f, g} is contained in C
by virtue of (i); on the other hand, by the complex Stone-Weierstrass
theorem (V, 8.3) this algebra is dense in Co(R). It follows from (i) that
Co(R) c C; in fact, if kl' k2 E C are bounded, then for all a E 9!(H)sa
we have IIkl(a) - k2(a)!! ::;; IIkl - k21100llall (see above).
(iii) Now let k be any bounded, continuous function R -+ C. Then /if, kg,
and kh are in C and thus k = kf + kh E C. This completes the proof.
Let A be a C*-subalgebra of 9!(H). Then in obvious fashion, the algebra
M2(A) of all (2 x 2)-matrices with entries from A can be viewed as a C*-
subalgebra of 2(H EEl H) (Exercise 16). If A is closed in 2(H) for the

strong operator topology, then so is M2(A) in 'p(H EB H). The following
result is the famous density theorem of I. Kaplansky. As before, we denote
by U the (closed) unit ball of 2(H) and by Uc the unit ball of any C*-
subalgebra C.
7.4
Theorem. Let A be any non-degenerate C*-subalgebra of 2(H) with double
commutant M := A CC • With respect to the strong operator topology the fol-
lowing assertions are valid:
(a) Asa is dense in Msa.
(b) UA nAsa is dense in U M n Msa.
(c) UA is dense in UM.
(d) If A is unital, the group GA of all unitary elements of A is dense in the
group GM of all unitary elements of M.
Proof In the subsequent proof, all topological statements refer to the
strong operator topology of 2(H) unless expressly stated otherwise.
(a) From a combined application of (6.1), its corollary and (6.2) we know
!
that the linear (over R) projection x ---> (x + x*), which maps 2(H)
onto 2(H)sa' is continuous for the weak operator topology hence, Asa
is dense in Msa for that topology; since Asa is convex, it is dense in Msa
by (IV, 3.1).
(b) Let a E UM n Msa. By (a) there exists a filter (Y containing Asa and
convergent to a. Let f(t) = sup ( -1, inf(t, 1)) (t E R). f is continuous
and bounded, hence strongly continuous (7.3), and satisfies f(t) = t,
It I ::;; 1. Thus a = f(a) = limf«(Y), and f«(Y) contains Asa n UA. This
proves (b).
(c) Let u E UM; then v = (~* ~) is an element of U M2 (M) n M2(M)sa'
Thus by (b) there exists a filter (fj on M2(A) containing M2(A)sa n
UM2 (A) and convergent to v. The image filter (Y with respect to the
mappmg. (XlI Xl2 ) f--+ Xl2 th en converges to u and ' UA.

contams
X21 X22
(d) We suppose A unital and u E M unitary. By Lemma 1 above there
exists v E Msa satisfying u = exp(iv). v E Asa by (a), and t f--+ exp(it)
(t E R) is strongly continuous by (7.3). In addition, exp(ix) is unitary
in A for all x E Asa. This implies (d), and the proof of (7.4) is complete.
8. PROJECTIONS AND TYPES OF W* -ALGEBRAS

Let A be a CO-algebra. The set Z(A) = {x E A: xy = yx for all YEA}
is called the center of A. Clearly, Z(A) is a C* -subalgebra of A. If W is a
W* -algebra with predual V (unique by 6.9), then Z( W) is a-weakly closed,
§8] PROJECTIONS AND TYPES OF W*-ALGEBRAS 293
because multiplication is separately a-weakly· continuous by (6.5). Thus

Z(W) is a commutative W*-algebra. A W*-algebra is called a factor (or
factor algebra) if Z(W) is one-dimensional (Le., if Z(W) = Ce). The im-
portance of this notion rests on the fact that every von Neumann algebra on
a separable Hilbert space is a direct integral offactors (see, e.g., G. Pedersen
[1], chapter 4, section 4.12 or M. Takesaki [1], chapter IV, section 8).
Let H be a Hilbert space, not reduced to {O}. An operator u E !l'(H} is
called a partial isometry if u restricted to (keru)l. is an isometry. Thus
if F, and F2 are closed subspaces of H with the same Hilbert dimension d
(Chapter II, Section 2, Example 5), there exists a partial isometry u satisfying
u(F!) = F2, u(Fr) = {O}. It follows that u*u is a projection (precisely, the
orthogonal projection H ~ (keru)l.), and similarly, uu* is the orthogonal
projection H ~ (keru*)l.; in particular, u' is a partial isometry.
From (3.1) and (3.2), we recall that the modulus of x E !l'(H) is defined by
Ixl = J x* x and note that Ixl = Ix* I if and only if x is normal. Also we have
Illxlell = Ilxeli (e E H). Now let x E !l'(H) and let u be the partial isometry
defined by xe = ulxle (e E H); u is unique and maps Ixl(H) isometrically
onto x(H). Then x = ulxl, in analogy with the complex numbers, is called
the polar decomposition of x.
Generally, an element u of a C* -algebra A is called a partial isometry if
u*u is a projection in A. Because a(uu*)\{O} = a(u*u)\{O} (cf. Exercise I)
u* is a partial isometry as well. For a partial isometry u the formulas u =
u(u*u) and u' = u*(uu*) follow from the corollary of (4.3) and the preceding
paragraph.
8.1
Let W be a W* -algebra. For every x E W there exists a unique partial
isometry u E W satisfying x = ·ulxl.
Of course, x = ulxl is again called the polar decomposition of x.
Proof of (8.1). By (7.1) Corollary 3, W can be viewed as a von Neumann

algebra in !l'(H), H some Hilbert space. Since the polar decomposition x =
ulxl is unique in !l'(H) by the preceding, we need to show only that u,
Ixl E W; this is clear for Ixi- Now let v E we be unitary; then x = v*xv =
v'uvlxl. Since u is unique in !l'(H) we have u = v*uv, hence vu = uv. But we
is the linear span (over C) of its unitaries (Corollary 2 of 2.2), and this im-
plies u E wee and hence, by (7.1), u E W.
In the following let W be a fixed (non-zero) W* -algebra. From (6.8) we
know that the set P( W) of all projections in W is a complete lattice; the
supply of projections is rich because by the corollary of (6.3), P(W) is norm
total in W. The classification theory of W* -algebras rests on a detailed
analysis of P( W).
Two projections p, q E P( W) are called (Murray-von Neumann) equiva-
lent, in symbols: P '" q, if P = u'u, q = uu· for some partial isometry u E W.

It is easy to see that", defines an equivalence relation on P( W) (Exercise
22). Let us write P :::S q if P '" r for some r ::;; q; in fact, this relation is a pre-
order on peW) (Chapter V, Exercise 1), as we shall see shortly.
REMARK. The meaning of the relations '" and :::S can perhaps be
better understood intuitively by looking at the example W = !l'(H),
H any Hilbert space. A (orthogonal) projection P E !l'(H) is uniquely
determined by its range pH. Let dp denote the Hilbert dimension of
p(H) (II, 2, Example 5); it is then easy to see that P '" q iff dp = d q ,
and p ::;; q, p 1- q iff dp < dq • Thus the equivalence classes mod '"
correspond biunivocally to the (well-ordered) set of cardinal numbers
::;; de (e = idH ). For example, if H is a separable Hilbert space then
p '" q for any two projections with infinite dimensional range.
In a general W' -algebra W, the situation is not so simple; the
reason is that even as Wmay be viewed as a W'-subalgebra of !l'(H) ,
two non-equivalent projections in W may be equivalent in !l'(H) (cf.
(8.6) Corollary).
LEMMA. Let {PO( : oc E I} and {qO( : oc E I} be families of pairwise orthogonal
projections in W, and let PO( '" q()( for all oc E I. Then p := LO( PO( and q :=
LO( qO( are equivalent.
Proof If v is a partial isometry with v'v = p and vv' = q, then by the
paragraph preceding (8.1) we have v = vp, v = qv. For each oc E I, let UO( be
a partial isometry satisfying PO( = u;uO( and qO( = uO(u;. Then PO(Pp = qO(qp =
O(oc =f. P) implies that u;up = 0 whenever oc =f. p. Using a (faithful) repre-
sentation of W in !l'(H) , it follows that for each ~ E H, (UO(~)o(EI is an
orthogonal family of H; thus u:= LO( UO( exists in 2s(H) and so u'u =
LO( u;uO( = LO(PO( = p as multiplication is hypocontinuous (7.2). It follows
that
uu' = LUO(u; = Lq()( = q,

0( 0(
which implies that p '" q.

The following result now establishes that :::S is a pre-order on peW).
8.2
It P:::S q and q :::S p then p '" q.
Proof Let u, v denote partial isometries satisfying Po = u'u, uu* = ql ::;; q,
qo = v'v, and PI = vv' ::;; p where Po = P and qo = q. By mathematical
induction we define two decreasing sequences (Pn), (qn) by virtue of
Pn+1 = vqnv' and qn+1 = uPnu'. Let Poo = infnPn> qoo = infnqn. Then P =
, L~o(Pn - Pn+d + P oo , q = L~o(qn - qn+d + qoo·
§8] PROJECTIONS AND TYPES OF W'-ALGEBRAS 295
Since u(Pn - Pn+!)u* = qn+! - qn+2, v(qn - qn+J)v* = Pn+! - Pn+2(n ~ 0),
uPoou* = qoo, vqoov* = Poo' and P2n+! - P2n+2 as well as q2n+! - q2n+2 are
pairwise orthogonal, the foregoing lemma shows that
00 00
P = L(P2n - P2n+!) + L(P2n+! - P2n+2) + P oo

n=O n=O
and
00 00
q = L(q2n+! - q2n+2) + L(q2n - q2n+!) + qoo

n=O n=O
are equivalent: P ~ q.
We now turn to the so-called comparability theorem for projections. For
this we need some further notions and results.
8.3
Let x E W be arbitrary, with polar decomposition x = ulxl. Then uu* =:
s/(x) is the smallest (::;;) projection PEW satisfying px = x; by symmetry
u*u =: sr(x) is the smallest projection satisfying xq = x. In particular, s/(x) ~
sr(x).
s/(x) and sr(x) are called the left and right support of x E W, respectively.
The proof of (8.3) is left to the reader; it is obtained most readily by repre-
senting Was a von Neumann algebra in !£,(H) (cf. 7.1, Corollary 3). It turns
out that s/(x) (respectively, sr(x)) is the (orthogonal) projection onto the
closure of range x (respectively, of range x').
As before, (cf. 6.8) we let P v q = sup{p, q}, P /\ q = inf{p, q}.
8.4
Let p,q E P(W). Then (p v q) - P and q - (q /\ p) are equivalent.
Proof Without loss of generality, we may suppose that P v q = e (the
unit of W). (Otherwise, consider the W'-subalgebra (p v q) W(p v q)
of W; see (5.1) and (6.4).) Let W be faithfully represented in !£,(H)
for some Hilbert space H. The kernel Ho of x = (e - p)q is equal to
q(H)l.g;;(q(H)np(H)). Hence the projection onto Ho equals e-q+
p /\ q. Therefore, the projection of H onto the closure of the range
of x* = q(e - p) equals q - p /\ q = sr(x). Likewise, the projection onto the
kernel of x* is P + (e - p) /\ (e - q) = p + (e - p v q) = p, which implies
s/(x) = e - p. The claim now follows from (8.3).
A projection p E P( W) is called central if p E Z ( W); the central support
z( x) of x E W is then defined to be the smallest (::;;) projection in Z ( W)
satisfying x = px(= xp). Its existence is an easy consequence of the lattice
properties of P(Z( W)) and of the separate a-weak continuity of multiplica-

tion (6.5) (cf. Exercise 23).
8.5
For any a-weakly closed ideal J of W, there exists a central projection p
such that J = Wp(= pW = pWp).
Proof We may suppose J =F {O}. Clearly, J is a W*-subalgebra of W,
hence unital by the corollary of (6.2); its unit p is obviously in P( W). Since
J is an ideal, we have pW c J = pJ = Jp and Wp c J; therefore, J =
pW = Wp. Finally, for all x E W one has xp E J, px E J thus xp = pxp =
px, which shows that p E Z(W).
The projections p and q in Ware called centrally orthogonal if their central
supports z(p) and z(q) are orthogonal.
LEMMA. Let p, q E P( W). The follOWing assertions are eqUivalent:
(a) p and q are not centrally orthogonal.
(b) pWq =F {O}.
(c) There exist equivalent projections PI, ql such that 0 < PI :::; p and 0 <
ql:::; q.
Proof (a) ::::} (b): Suppose that pWq = {O}. Then J = {x E W: pWx =
{O}} is a a-weakly closed ideal in W (cf. 6.5) h~nce, by (8.5), we have
J = Wr for some central projection r. Now q E J implies r ~ q and so
r ~ z(q). Since pr = rp = 0, we have p :::; (e - r) whence z(p) :::; (e - r); this
implies z(p)z(q) = O. Therefore, pWq =F {O}.
(b) ::::} (c): Let 0 =F x E pWq. Then pxq = x, which implies St(x) :::; p and
sr(X) :::; q. But St(x) '" Sr(X) by (8.4).
(c) ='} (a): Let 0 < PI :::; p, 0 < ql :::; q and PI = u*u, ql = uu* for some
partial isometry u E W. Now u = UPI implies u = uz(p). Likewise u· =
u*z(q); hence u = z(q)uz(p) = uz(q)z(p) implies z(p)z(q) =F O.
The following is now the announced comparability theorem.
8.6
Theorem. Let p, q be projections in W. There exists a central projection r
satisfying qr :::5 pr and p(e - r) :::5 q(e - r).
Proof We consider pairs of families (P")"EI and (q")"EI with these prop-
erties: The members of each family are pairwise orthogonal, p" '" q" for each
oc E I, and p" :::; p, q" ::;;; q (oc E I). As in the lemma preceding (8.2), we define
Poo = 2:."p", qoo = 2:." q,,; then by the same lemma we have Poo '" qoo; also
Poo ::;;; p, qoo :::; q. By Zorn's lemma, there exist maximal pairs of families of
this type; let {(p,,), (q,,)} be such a pair. Then by maximality, no projection
r' :::; p - p 00 is equivalent to any projection r" :::; q - q 00' By the preceding
§8] PROJECTIONS AND TYPES OF W*-ALGEBRAS 297
lemma, there exists a central projection r such that p - Poo ;s; rand q - qoo
;s; (e - r). This implies (q - qoo)r = 0 and (p - poo)r = P - Poo. Since r is
central, for any pair (s, t) E P X P, s ;s; t implies rs ;S; rt, and s '" t implies
rs '" rt. We finally obtain:
pee - r) = poo(e - r) '" qoo(e - r) ;S; q(e - r).

This completes the proof.
If W is a factor, then Z( W) = Ce and we get this corollary.
COROLLARY. If the W* -algebra W is a factor, then for each pair (p, q) of
projections, exactly one of these relations is valid: p -( q, p '" q, or q -( p.
A projection PEW is finite if p '" q, q ;S; p implies p = q. If p is finite and
q ;S; p, then q is finite; W itself is called finite if its unit e is finite. A projection
PEW is abelian if p W p is commutative. W is called of type I if W = {O} or
its unit e is the orthogonal sum of a family of abelian projections. (This is
equivalent to requiring that each central projection r "# 0 majorizes some
abelian projection "# 0.) Each projection equivalent to an abelian projection,
is itself abelian. Abelian projections are obviously finite. A projection p is
called infinite if it is not finite. p is called purely infinite if p ~ q > 0 implies
that q is infinite. W itself is called purely infinite or of type III if its unit ele-
ment e (supposing that W "# {O}) is purely infinite. If Whas neither non-zero
abelian nor central purely infinite projections, then it is of type II. If W is
also finite, it is of type III else of type IIoo. Finally W is semi-finite if there is
no central purely infinite projection in W.
Examples
10. Every commutative W*-algebra is of type I.
11. The operator algebra filCH) is of type I (Hany Hilbert space). In
fact,if H"# {O} and (e tx ) is an orthonormal basis of H, then etx ® etx
(IX E I) is abelian and idH = e = LtxEletx ® etx.
12. Let (X,~,Il) be a a-finite measure space, and let Hbe separable.
ThenLOO(X,~,Il;fil(H», under its natural operations, is a W*-alge-
bra of type I. (If (en) is an orthonormal basis of Hand Pn = en ® en
then idv"(x,r.,p,;H) = Ln idu'(x,r.,p,) ® Pn-)
Further examples can be found in Exercises 24-27 and the literature
indicated there.
The redeeming grace of this (at first sight) unnecessarily complicated
classification is that it is exhaustive. To prove this, we need the following
lemma.
LEMMA. Let W be any W* -algebra. The sum of a familiy (Ptx)txEI of
centrally orthogonal abelian (respectively, finite) projections p", is abelian

(respectively, finite).
Proof Let P = LP", and z'" = z(p",). Then p", = Z",p for all IX; hence pWp
= EB",p",Wp",. Therefore if each p",Wp", is abelian so is pWp.
If p '" q :5: P then z",q = q", '" z",p = p", and q", :5: p",. Thus p", = q", when
p", is finite. But this implies p = q; hence p is finite.
8.7
Theorem. Every W' -algebra W is the direct sum of (at most) four W'-
algebras of type I, III, II00, or IlL· this decomposition of W is unique.
Proof Assuming W # {O}, the proof will result in a decomposition
W = WI EEl WIll EEl Wn", EEl Wm ,
where anyone summand (or several) may be {O}; the latter will occur if the
families used for the construction of the defining projections are void. It will
be seen shortly that the decomposition of W in question is unique (except for
the order of summands, of course).
To construct WI, let p be the sum of a maximal family (PP)PeB of centrally
orthogonal abelian projections; by the lemma, p is abelian. Let PI = z(p) be
the central support of p. If q # 0 is a central projection satisfying q :5: PI>
then qp # 0 because q = qPI (cf. 5.2) and PI is the central support of p.
Moreover qp = pq is abelian, because qp Wqp c p W P and p is abelian.
Thus q majorizes an abelian projection and, hence, WI := WPI is of type 1.
To construct W Il , we observe first that W(e - PJ) contains no abelian
projections # 0 by the maximality of (Pp)peB. Now let q denote the sum
of a maximal family (qY)yer of centrally orthogonal finite projections in
W(e - PI); by the lemma, q is finite. Let PIl = z(q) be the central support
of q. Then, since PIl ::;; e - PI> WPIl contains no abelian projections (#0). If
r E WPIl is a non-zero central projection then rq :5: q; hence rq is finite. But
rq ::;; r, and r = rpIl implies rq # 0 (see above). Thus WIl := WPIl is of type
II.
Finally, maximality of the families (Pp)peB and (qy)yer implies that Wm
:= Wpm, where Pm := e - PJ - PII> contains no finite projection (#0).
Therefore, Wm is of type III. Hence we have W = WI EB WIl EEl Wm, and it
remains to be shown that WIl splits into WIll EEl Wn ",. Again, if s denotes the
sum of a maximal family of orthogonal finite central projections in WIl, then
WIls is of type III and WIl(PIl - s) is oftype IIoo.
Let us indicate briefly how it is seen that the decomposition of Was given
in (8.7) is unique (of course, except for the order of summands and excluding
summands {O}). Let W = EB~I Wi wher~ Wi is, in this order, an algebra,?f
type I, III ,1100, and III, respectively. If WI # {O} and PI is the unit of WI
(corollary of 6.2), PI (e - PI) is a central projection in WPI and, if # 0,
EXERCISES 299
majorizes an abelian projection =F 0 whence so does e - PI. This proves

PI (e - PI) = 0, similarly, PI(e - pd = o. Thus PI = PI by (5.2). The
proof for the remaining summands is analogous. In conclusion, we observe
that the direct sum of (8.7) is topological, both for the norm and a-weak
topologies (cf. Exercise 17).
COROLLARY. If a W* -algebra W is a factor, it is exactly of one of the types:
I,III,IIoo , or III.
REMARK. Let (Wy)yeT denote any family of W*-algebras. Then the
tOO-sum, consisting of all bounded families f = (/y)yer with IIfll =
SUPyerll/yII < 00, is a W*-algebra in a natural way; similarly, the
direct integral of W* -algebras over a standard measure space (X, 1:, 11)
provides a method of constructing still more general examples of W*-
algebras. Its significance rests on the fact that each W* -algebra with
separable predual is a direct integral offactors (see above).
The preceding corollary now reduces, in essence, the classification of
W* -algebras to the classification of factors.
It can be shown that a factor W is of type I if and only if W is isomorphic
with 5f(H) for a suitable Hilbert space (see Exercise 25). Factors of the types
III, IIoo , and III are more difficult to construct and classify. For further
examples and results, the reader is asked to consult Exercises 26 and 27 and
the literature indicated there.
EXERCISES
1. Let B denote a complex B-algebra with unit e.
(a) For two commuting elements x, y E B, xy is invertible iff x and y
are invertible.
(b) Show that a(xy) \ {O} = a(yx)\ {O} for all x, y E B. (If c =
(e - xy)-I exists then e + ycx = (e - yx)-I.)
(c) For any polynomial P over C of degree n(~O), one has
a(P(x)) = P(a(x)) whenever x E B. (For given 11 E C, decompose the
polynomial Q(z) = 11 - P(z) into linear factors and apply (a).)
2. Prove the Gelfand-Mazur theorem (Section 1). (Use that in a
complex unital Banach algebra B, one has a(x) =F 0 for each x E B.)
3. Let K, L denote compact spaces =F 0. The C* -algebras C(K)
and C(L) (of continuous complex functions) are isomorphic iff K and
L are homeomorphic. (Letting Al = C(K), A2 = C(L) and following
(V, 1.7), show that K is homeomorphic to the extreme boundary of
the state space 9"(AJ) under a(C(K)', C(K)) (and accordingly for L).
Use the fact that the adjoint <J)' of every surjective isomorphism <J) :
C(K) ~ C(L) is an (affine) weak* homeomorphism of 9"(A2) onto
9"(AI).)
4. An element y of a Banach algebra B is called a left topological

divisor of zero if limy yZy = 0 for some normalized directed set (Zy);
accordingly, a right topological divisor of zero if limy zrY = o.
(a) If x E B (B unital, complex) and A E ou(x) then (Ae - x) is a left
topological divisor of zero.
(b) For B a unital C* -algebra, show that x E B is invertible iff both
x*x and xx* are invertible. x E B is non-invertible iff x is either a left
or right topological divisor of zero. (Note that u(x*x) has void
interior.)
(c) Let A denote a CO-algebra with unit e, B a C*-subalgebra of A
containing e. For any x E B, the spectrum UB(X) (of x with respect to
B) equals UA(X).
5. Let A be a (complex) B-algebra with involution x f-+ x*.
(a) If IIxll2 ::;; IIx*xll holds for all x E A, then A is a CO-algebra.
(b) Suppose A is a C* -algebra without unit, A its unitization (see
discussion preceding 2.2 Corollary 3), and let .5l'(A) denote the algebra
of (bounded) endomorphisms of the underlying Banach space A;
finally, let La E .5l'(A) be left multiplication x f-+ ax(x E A). Show that
A is isometrically isomorphic to the (closed) subalgebra of .5l'(A) gen-
erated by {La: a E A} and idA. Conclude that A is (norm) complete.
(c) Show that the norm of A is a CO-norm (i.e., satisfies condition
(3) of Section 2).
6. Let A denote a unital C* -algebra; x, y elements of A.
(a) x> 0, and y > 0 together imply u(xy) c R+. (Use Exercise 1.)
(b) (Fuglede's theorem): If x is normal then yx = xy implies y*x =
xy*. (Observe that A f-+ f(A) = exp(iAx*)yexp(-iAx*) is a bounded
entire function because y commutes with exp(ih), apply Liouville's
theorem (Chapter V, Exercise 39) and compute f'(O).)
(c) If x' y = 0 = xy' then IIx + yll = max(lIxll, lIyll). (Observe that
(x'x)(y' y) = 0.)
7. Let A,B denote CO-algebras such that B is unital, A non-unital.
Prove that every C* -homomorphism rP : A --t B has a unique homo-
morphic extension i: A --t B mapping the unit of A onto the unit of B.
8. Denote by H an infinite-dimensional Hilbert space and by .1{"(H)
the (closed) ideal of all compact operators in .5l'(H). The quotient C'-
algebra .5l'(H)/.1{"(H) is called the Calkin algebra of H. This exercise
aims at giving a particular representation of this algebra in .5l'(K) ,
where K is a (closed) subspace of an ultrapower II of H (cf. Calkin [1]).
To this end, let I1IJ denote a free ultrafilter on N, let tOO(H) =
{f EHN : f bounded} and let co(H, 11IJ) = {f EtOO(H) : lim"ll f(n) = O}.
(a) Show that 1I:=tOO (H)/co(H,I1IJ) is a Hilbert space (quotient
norm) containing H as a closed subspace.
(b) For each T E .5l'(H), if:= (T(f(n»nEN defines an operator i
on tOO (H) leaving co(H, 11IJ) invariant, and, hence, defines an operator
EXERCISES 301
T E !£l(H). Prove that T --+ T is a CO-isomorphism of !£l(H) into

!£l(H).
(c) Let P denote the orthogonal projection of H onto H, and let
Q = idfi - P. Show that for each T E !£l(H), T is reduced by the pair
(P, Q) (i.e., T = PTP + QTQ).
(d) The mapping cI> : !£l(H) --+ !£l(Hl.), given by cI>(T) = QTQ, is a
CO-homomorphism with kernel equal to %(H); therefore, cI> defines
the desired representation of !£l(H)j%(H) in !£l(K) for K = Hl..
(e) Prove that if (and only if) H is separable, %(H) is the unique
closed ideal of !£l(H) distinct from !£l(H) and {O}.
9. Let H denote any Hilbert space. An operator T E !£l(H) is called
Fredholm if both the dimension of its kernel and the co-dimension of its
range are finite. Prove the following:
(a) T is Fredholm iff both dim(ker(T)) and dim(ker(T*)) are finite
and T(H) is closed in H.
(b) (Atkinson's theorem): T E !£l(H) is Fredholm if and only if its
canonical image in the Calkin algebra of H is invertible. (Use the rep-
resentation of !£l(H)j %(H) established in Exercise 8.)
10. (Toeplitz operators). Denote by T = {z E c: Izl = I} the one-
dimensional torus, and by m normalized Lebesgue measure on T. The
so-called Hardy space HP(T) is the subspace of VeT, m)(l :::;; p :::;; (0)
whose elements have all negative Fourier coefficients J f(z)zn dm(z) ,
n < 0, equal to zero. H2 (T) is a closed subspace of L 2 (T, m); let us
denote by P the orthogonal projection of L2(T,m) onto H 2(T). If 9 E
LOO(T,m) and Mg : f 1->91 denotes multiplication by g, the operator
Tg := PMgP E !£l(H2(T)) is called the Toeplitz operator with symbol
g.
(a) Show that, in general, TgTh #- Tgh. (Choose g(z) = z,h(z) = z.)
(b) If 9 E LOO(T,m),h E HOO(T) then TgTh = Tgh.
(c) If h(z) = z, then (Th)n --+ 0 in the strong operator topology.
(d) 9 E LOO(T,m) and Tg compact implies 9 = O. (Using (c), con-
clude that (ThnS) --+ 0 in the norm topology, for each compact S. Next
show that Tg = ThTgTh' whenever 9 E LOO(T,m). Finally employ the
inequality
IITgll = IITtTg(Th'tll :::;; IITtTglI·)
(e) If g, h are continuous on T then TgTh - ThTg is compact. (Use
the Stone-Weierstrass· theorem to reduce the proof to considering trig-
onometric polynomials.)
(f) The C'-subalgebra A of !£l(H2(T)) generated by all Toeplitz
operators T f with continuous symbol f, is called the Toeplitz algebra.
Show that the closed ideal of A generated by all commutators Tf Tg -
TgTf equals %(H2(T)).
(g) For f E C(T), denote by Tf the canonical image of Tf in
A/%(H2(T)); then f f-+ Tf is an isomorphism of CO-algebras

C(T) -+ A/%(H2(T)).
(h) For f E C(T), Tf is Fredholm iff f(t) =I- 0 for all t E T. (Use
Exercise 8.)
II. There exists a unital C* -algebra O(n) generated by n(n EN)
elements Vj that satisfy these relations: (i) vjVj = e(j = I, ... , n) and (ii)
Ejn=l Vjvj = e. The algebra O(n) is unique (to within isomorphism) and
called the Cuntz algebra (Cuntz [I]). (Consider appropriate operators
Vj on t 2 (N).)
12. A representation (Section 4) (n, H) of the CO-algebra A is called
irreducible if every element ¢ =I- 0 in H is a cyclic vector.
(a) (n, H) is irreducible iff the commutant n(A) C := {y E .!£l(H) :
yn(x) = n(x)y for all x E A} of its range n(A) is trivial (i.e., contains
only scalar multiples of idH ). (Show that (n,H) is not irreducible ex-
actly when n(A)C contains a projection P, 0 =I- p =I- idH .)
(b) Suppose A is unital, and let ,p E Y(A). The representa-
tion (nt/J, Ht/J) is irreducible iff ,p is an extreme point of the state space
Y(A). (Using part (a) consider the states ,pp and ,pq, where ,pp(x) =
[nt/J(x)p¢t/J, ¢t/J].)
13. Let A denote a C* -algebra.
(a) A possesses a countable approximate unit iff there exists a E A,
a> 0 such that U:l [-na, na] is dense in Asa. Conclude that this is true
whenever A is separable.
(b) If a, bE A satisfy a;;::: 0, lIall::; 1 and b;;::: 0, Ilbll ::; 1 then
Iia - bll ::; 1.
(c) If pEA, p2 = P and Ilpll = 1 then P = p* (i.e., p is a projec-
tion). (Consider pas an operator in .!£l(H) under some faithful
representation. )
14. Let A be any CO-algebra.
(a) Recall that a linear form ,p E A' is self-adjoint (Section 4) if
,p = ,po, where ,p*(x) = ,p(x*) (x E A). Prove that,p =,p* iff ,p(Asa) c R
and that the space A;a can be identified with the order dual of A
(Chapter V, Section 1).
(b) Show that the Jordan decomposition ,p = ,p+ - r of,p E A;a is
unique by virtue of the requirement that 11,p11 = 11,p+1I + IIrll. (Observe
that, in the proof of (4.2), the lattice decomposition p. = p.+ - p.- is
minimal by (V.l.1).)
15. Prove Proposition 5.3 without recourse to Gelfand's theorem
(2.2). (Note that e = AX + (1 ~ A)y, where x, y E U, implies ~ = Aa +
! !
(1 - A)b, where a = (x + x*), b = (y + yO) are self-adjoint.)
16. Let A be any CO-algebra.
(a) The norm of a CO-algebra is unique (equivalently, an involutive
complex B-algebra can be a CO-algebra for at most one norm). (Ob-
EXERCISES 303
serve that for any norm II . lion A satisfying condition(3) (Section 2),
one has Ilxll = r(x) for all x E Asa by virtue of (2.1).)
(b) Denote by Mn(A) the algebra of all n x n matrices (n EN) with
entries from A. Show that there exists exactly one norm on Mn(A)
under which it is a C'-algebra. (Consider the universal representation
(n, H) of A and identify Mn(A) with a subalgebra of !l'(Hn); then
apply (a).)
(c) If A is a W* -algebra then so is Mn(A).
17. Let V denote a Banach space with dual W. If W is the direct
algebraic sum (Chapter I, Section 2) of n( EN) 0'( W, V)-closed sub-
spaces M.iU = 1, ... ,n), then the sum is topological (Le., each of the
projections n j : W --> M j with kernel N j = L.NiM.i is O'(W, V)-con-
tinuous). (Observe that each nj is norm continuous (cf. proof of III.2.1,
Corollary 3), which implies that the unit ball of N j is 0'( W, V)-com-
pact; finally apply (IV.6A) and the corollary of (6.1).)
18. (Spectral Theorem for Normal Operators.) Cf. Chapter V,
Exercises 25-27.
Let H be any complex Hilbert space, T E !l'(H) a normal element
with spectrum O'(T) =: K. Denoting by A the C*-subalgebra of !l'(H)
genrated by {T, idH }, let <D : C(K) --> A be the isomorphism given by
Gelfand's theorem (2.2), Corollary 1. Finally let B denote the C'-
algebra of bounded, complex Baire (equivalently, Borel) functions on
K, placed in standard duality with M = C(K)' (bounded complex
>
Baire measures on K) by virtue of <I, p, = SK 1dp,.
(a) There exists a unique C* -homomorphism : B --> !l'(H), ex-
tending <D and continuous for O'(B, M) and the a-weak topology of
!l'(H). (Note that <D has a unique linear extension : B --> W := A CC
(cf. 7.1), using (6.1) and the fact that the unit ball of W is a-weakly
compact. Then show that preserves the multiplication and involu-
tion operations of B, using (6.2), (6.5), and transfinite induction on the
Baire classes of B.)
(b) For any Baire subset t5 c K, let m(t5) := <1>(115). m is a W-valued
spectral Baire measure on K that is countably additive for the strong
operator topology. (Observe that m is countably additive for the 0'-
weak and hence the weak operator topology, and use (VA.3) in con-
junction with the remark preceding (7.1).)
(c) The range P ofm (see (b)) is a countably order complete Boolean
algebra of projections in W called the resolution of the identity asso-
ciated with T. (Observe that P is a homomorphic image of the Boolean
algebra of all Baire subsets of K)
(d) For any fEB, one has
(f) = t/(Je) dm(Je) =: f(T)

where the integral exists in the operator norm of .!e(H). The mapping
f I-t f(T) is usually called an operational calculus for T.
19. Let H be a separable Hilbert space and (X,~,,u) a a-finite
measure space. Then LOO(X,~,,u; .!e(H» is a W*-algebra (cf. (6.9),
Corollary 2). The result continues to hold if .!e(H) is replaced by any
W*-algebra with separable predual (cf. Takesaki [1], chapter IV,
theorem 7.17; lonescu-Tulcea [1], chapter VI).
20. Let K be compact. The CO-algebra C(K) is order complete if
and only if K is Stonian (extremally disconnected). (Cf. Chapter V,
Exercise 23.)
21. (Theorem of Kakutani). Every abstract L-space (AL-space, cf.
Chapter V, Section 8) E is isometrically and lattice isomorphic with
L' (X,,u) for some locally compact space X and Borel measure ,u on X.
(a) If E possesses a weak order unit u, then E is isometrically and
order isomorphic with L' (,u) for some Borel measure ,u on a compact
space K. (Observe that the principal lattice ideal Eu is isomorphic to
some C(K) by (V.8.5), and dense in E. The norm of E then defines a
(positive) Radon measure on Ksuch that E = L'(K,,u); cf. Chapter V,
Exercise 22.)
(b) In the general case, denote by (u;.);'EA a maximal orthogonal
family of normalized positive elements of E. The closures EUA are, by
(a), isomorphic to spaces L'(K;.,,u;.) for suitable compact spaces K;.
(A E A). Show that E ~ ffi;'EAEUA in the sense tha,t for each x E E with
band components x;., one has x = :E;.x;. with Ilxll = :E;.lIx;.ll. (For
details, see Schaefer [12], theorem 1I.8.5.)
(c) Conclude from (b) that E ~ L'(X,,u) (where X is the locally
compact space ffi;.K;., with ,uIK;. = ,u;. for all A E A).
22. Show that the relation p '" q (p, q projections) defined in Section
8 is actually an equivalence relation (Murray-von Neumann equiva-
lence).
23. Let W be any W*-algebra and x E W. Show that the center
Z(W) contains a smallest (:::;;) projection q satisfying xq = q. (Observe
that the projections in Z( W) form a Boolean lattice, and that multi-
plication in Wis separately a-weakly continuous (6.5).)
24. Let A be any unital CO-algebra. A state. E Y(A) is called a
tracial state or normalized trace if .(xy) = .(yx) for all x, YEA.
(a) Prove that. is a trace iff .(u·xu) = .(x) for all x E A and all
unitaries U E A (i.e., iff. is invariant under inner automorphisms).
(b) If W is a factor and the normal tracial states separate the points
in Wthen Wis finite. (The converse also holds, cf. Takesaki [1], theo-
rem V.2A).
25. A factor algebra W is of type I iff W is w· -isomorphic with
.!e(H) for some Hilbert space H. (If P is abelian then the center of
pWp equals pWp, while it equals pZ(W)p as well. Now if W is of
EXERCISES 305
type I, then (supposing W =I {O}) e = Loh holds for some orthogo-

nal family (PO)OEL'1 of pairwise equivalent abelian projections. Fix 60
and let Uo denote a partial isometry satisfying ujuo = ho and uouj =
h. Define U(uo) = eoo ® eo E 2((2(~)). Show that U can be extended
to all of W, the extension D being a W* -isomorphism of W onto
2((2(~).
26. Let G denote any group containing at lea,st two elements. For
9 E G, let U g E 2((2(G») be defined by ug(f)(Ii) = f(g-lh). The bi-
commutant of {u g : 9 E G} in 2((2(G») is called the W*-group alge-
bra W(G).
(a) Denote by e the neutral element of G and let 6e (f) := fee),
IE (2(G). Then T(X) := [xJe,Je] defines a normal trace that islaithlul
(i.e., a strictly positive linear forin). Conclude that W( G) is finite. (By
use of the Cauchy-Schwarz inequality show that T(XU g ) = 0 for all
9 E G if T(X) = 0 and X;;::: O. Conclude that T is strictly positive.)
(b) W(G) is a factor iff for each 9 E G,g #- e the coset {hgh- l :
h E G} is infinite; if so then W (G) is of type III. (To show that W is a
factor for each x E W(G) consider the function I(g) = T(XU;). It sat-
isfies I E (2(G) and, if x is in the center Z(W(G)), f(g) = f(hgh- l ).)
REMARK. Generalizations of W* -group algebras yield examples for
W*-algebras of type III (see Takesaki [1], chapter V, section 7).
27. Let Hn := C 2" be considered as the n-fold tensor product of C 2
with itself. Moreover let An := M2n (C) be considered as the n-fold
tensor product of M2 ( C). Then for m < n, Am can be embedded into
An by an,m(X) = X ® id(n-m) where idk is the identity on Hk.
(a) There exist a C*-algebra A and a family (an) of isomorphisms
an : An --+ A that satisfy the following two conditions: (i) an = an+p 0
an+p,n and (ii) Unan(An) is dense in A. A is called the infinite tensor
product of M2 (C). (Consider an appropriate quotient of {I E TIn An :
suplllnil < CIJ }.)
(b) Let rP E .9"(M2(C)) be an arbitrary state on M2(C) and denote
by rPn the n-fold tensor product of rP on An. There exists a unique state
rPoo on A satisfying rPoo 0 an = rPn- rPoo is called a product state.
(c) Show that there exists exactly one tracial state on A.
REMARK. Let rP be an arbitrary product state on A and (nq"Hq,,¢q,)
its GNS-representation (see Theorem 4.3). Then the bicommutant of
nq,(A) in 2(Hq,) is a factor. It is of type III iff rP is the tracial state; in
any other case it is of type III. It can be parametrized further by the
open interval ]0, 1[ so that the algebras for any two different parame-
ters are not isomorphic (see Powers [1]; Connes [1]; Pedersen [1],
chapter 6, section 5 and chapter 8, section 15).
Appendix
SPECTRAL PROPERTIES OF
POSITIVE OPERATORS
It has been discovered around the turn of the century (Frobenius [I], [2],
Perron [I]) that the spectrum of n x n matrices with real entries ~ 0 has
certain special features; in particular, the spectral radius is an eigenvalue
with a positive eigenvector (for the canonical order of Rn). Since that time a
slow but steady development has taken place which, in an abstract setting,
reached a climax with the advent of the well-known memoir by Krein-
Rutman [I]. In fact, it appears that apart from normal operators on Hilbert
space (and their generalizations, usually called spectral operators), positive
operators on ordered topological vector spaces are the most interesting class
from a spectral point of view; in addition, it should be noted that the theory
of spectral operators is largely governed by order theory (cf. Chapter V,
Exercises 25-27). (In this Appendix, the term "operator" will be used syn-
onymously with "continuous endomorphism".) Hence a good deal of the
motivation for the study of ordered topological vector spaces has its origin
in spectral theory, and it is the objective of this Appendix to introduce the
reader to some spectral theoretic applications of the results of Chapter V.
We assume familiarity with the most elementary facts about the algebra of
operators on a Banach space; Section 1 enumerates in detail what will be
needed in the sequel.
In the investigation of the spectrum of a positive operator defined on an
ordered Banach space, various routes of approach can be followed according
to the type of problem under consideration; Sections 2 and 3 each are devoted
to a certain mode of attack; Section 2, exploiting function theoretic proper-
ties of the resolvent, contains most of what seems to be attainable in a very
general setting; the generalized Pringsheim theorem (2.1) is the unifying
theme of the section. Of course, results on the existence of eigenvectors are
bound to involve compactness in one form or another. Section 3 is devoted
to a study of the peripheral point spectrum of positive operators under more
306
§1] ELEMENTARY PROPERTIES OF THE RESOLVENT 307
specific assumptions, both on the operator and its space of definition. A

salient feature of these results is that compactness is not invoked. Theorem
(3.4) is an example of a rather general result in this area. The reader himself
will notice that this direction of research is far from being exhausted and in
some respects appears to be quite promising.
Except where other references are given, the results in this appendix can,
for their greater part, be found in the author's papers [6] and [9]-[11].
1. ELEMENTARY PROPERTIES OF THE RESOLVENT

Let (E, II II) be a complex Banach space, and denote by 2(E) the Banach
algebra (Chapter IV, Exercise 40) of continuous endomorphisms of E, under
the standard norm u ~ lIuli = sup{ Ilu(x) II: IIxll ;;;; I}. If U E 2(E), the spectrum
u(u) is the complement in C of the largest open set p(u) in which A. ~ (A.e - U)-1
exists and is locally holomorphic (Chapter IV, Exercise 39); here and in the
following, e denotes the unit of 2(E) (that is, the identity map of E). For
A. E p(u), we set (A.e - u) -1 = R(A.); A. ~ R(A.) is called the resolvent, p(u) the
resolvent set of u. Supposing that E is not reduced to {O}, it is a well-known
fact (cf. Hille-Phillips [1] and Chapter IV, Exercise 40) that u(u) is a non-
empty compact subset of C; the radius r(u) of the smallest circle of center 0
in C that contains u(u) is called the spectral radius ofu; the set {A. E C: 1,1.1 =
r(u)} is termed the spectral circle of u. Moreover, if A. E p(u) and fl E p(u) one
has the resolvent equation
R(A.) - R(fl) = -(A. - fl)R(A.)R(fl). (1)
Here we denote the composite u 0 v of u, v E 2(E) by juxtaposition uv; for the

simple proof of (1) see, e.g., Hille-Phillips [1].
If, more generally, E is a I.c.s. and u is a continuous endomorphism of E,
we can define the spectrum, resolvent set and the resolvent of u as before,
considering 2(E) under an S-topology for which 2(E) is a I.c. algebra
(Chapter IV, Exercise 40). This is true for every S-topology on 2(E) such
that u(S) c S for all u E 2(E); in particular, 2(E) is a I.c. algebra for the
topologies of simple and bounded convergence. Most of the properties of
u(u) familiar from Banach spaces fail for continuous endomorphisms of more
generall.c.s.; however, if E is a semi-complete l.c.s. and u is a bounded endo-
morphism of E, then by transition to a suitable Banach space Eu (Chapter III,
Section 7) the classical results can be shown to hold for u (cf. Schaefer [3],
Section 10). In the same manner, most of the results on positive operators
in ordered Banach spaces, derived in Sections 2 and 3 below, can be gener-
alized to bounded positive endomorphisms of an ordered I.c.s.; since no
essentially new methods are involved, we shall restrict attention to Banach
spaces. This does not mean, of course, that the consideration of non-normable
topologies is eliminated; as an example we refer to the proof of (2.4) below.
We return to the assumption that E is a complex Banach space. By virtue
308 SPECTRAL PROPERTIES OF POSITIVE OPERATORS [App.
of Banach's homomorphism theorem (Corollary I of (III, 2.1» and the fact

that the set of invertible elements of a Banach algebra is open, for each
u E 2(E) the spectrum a(u) can be characterized as the set of those A E C for
which A.e - u fails to be an algebraic automorphism of E. In view of this, we
have the following result.
Let u E 2(E), where E is a complex Banach space, and assume that
P.n: n EN} is a sequence in p(u) converging to some A E C. Then A E a(u) if and
only if limn IIR{A.n) II = + co.
In fact, the condition is clearly sufficient for A E a(u). To prove its necessity,
suppose there exists a subsequence {Jln} of {An} such that {R(Jln): n EN} is
bounded; by (1) the latter is a Cauchy sequence in 2(E) and hence convergent
to some v E 2(E). This implies limn R(Jln)(Jlne - u) = v(Ae - u) = e and,
similarly, (Ae - u)v = e; hence we obtain A E p(u), which is contradictory.
The subset of a(u) in which (Ae - u) fails to be one-to-one is called the
point spectrum n(u) of u. An element ,1.0 E n(u) is called an eigenvalue of u, the
null space of (Aoe - u) the corresponding eigenspace N(A,o). The dimension
of N(Ao) is called the (geometric) multiplicity of ,1.0' and the non-zero elements
of N(Ao) are termed eigenvectors of u for ,1.0. (The terms characteristic value,
characteristic space, arid characteristic vector are also in current use.)
The point spectrum of u contains all poles of the resolvent R. Let ,1.0 be a
pole of R and let
L
00
R(A) = aiA - Ao)k (2)

k= -n
be the Laurent expansion of R near ,1.0; the integer n (~ 1) is the order of the
pole ,1.0; the partial sum of (2), extending from k = - n to k = - 1, is the
principal part of the expansion; a_ n the leading coefficient, and a-I is the
residue of R at A = ,1.0. Multiplying (2) by (Ae - u) = (Aoe - u) + (A ~ Ao)e
and comparing coefficients in the resulting identity (which is justified by the
identity theorem for analytic functions), we obtain, in particular, a_nCAoe - u)
= (Aoe - u)a_ n = 0 and a_ n = a_leu - Aoe)n-l; clearly, the coefficients ak
commute with u. These relations show that ,1.0 is in n(u); more precisely, it
turns out that a-I is a projection of E onto the null space of (Aoe - u)n which
contains N(Ao). Let us recall also (cf. Riesz-Nagy [1], Hille-Phillips [1]) that
for compact u, the resolvent R is a meromorphic function on the Riemann
sphere punctured at 0 (one defines, generally, R( co) = 0); thus for compact u,
a(u) is a countable set, with 0 as its only possible accumulation point, and each
non-zero A E a(u) is an eigenvalue of u of finite multiplicity.
Finally, if u E !£,(E) and 1,1.1 > r(u), the resolvent of u is given by
= L r(n+1)u n
00
R(A)
n=O (3)
(U O = e); (3) is the expansion of R at 00, and is called the C. Neumann's

series. It follows from Cauchy's criterion for the convergence of power series
§2] PRINGSHEIM'S THEOREM AND ITS CONSEQUENCES 309
that r(u) = lim supl\u"11 1 /"; more precisely, it is true that r(u) = limn 11u" 1I1 / n
(cf. Hille-Phillips [I]). In case r(u) = 0, u is called a topological nilpotent of the
Banach algebra 2'(E); clearly, u is a topological nilpotent if and only if
O"(u) = {O} or, equivalently, if and only if the resolventR(with R(oo) = 0) is an
entire function of A- 1.
If E is a Banach space over Rand u E 2'(E), the real spectrum 0" R( u) can be
defined as the subset of R in which (Ae - u) fails to be an automorphism of
E; analogously, we can define the real resolvent of u as the function A~
(Ae - U)-1 with domain R ~ O"R(U). (It can happen that O"R(U) is empty, as the
example of a rotation about the origin of the Euclidean plane R~ shows.)
We shall not follow this practice, but instead subsume the case of a real
Banach space under the preceding by the following standard procedure.
Let (E, II II) be a Banach space over R; the complexification E1 (Chapter I,
Section 7) of the t.v.s. E is a complete normable space over C. Ifwe desire to
have a norm on E1 such that the imbedding of E into E1 becomes a real
norm isomorphism, the definition
IIx + iyll1 = sup n(cos 8)x + (sin 8)YII

O~II<2"
will do; this is a generalization of the definition of the usual absolute value on
C, considering C as the complexification of R. Now every U E 2'(E) has a
unique complex extension U E 2'(E1), defined by u(x + iy) = u(x) + iu(y) for
all x, y E E. In case E is a real Banach space and u E 2'(E), we define the
spectrum, resolvent, spectral radius of u to be the corresponding objects for u
as defined above. Sometimes it is even convenient to identify u with its com-
plex extension u. It is easy to see that for u E 2'(E), we have 0" R(U) = O"(u) n R,
that for 1 E R ~ 0"R(U) the real resolvent of u is the restriction of the resolvent
°
of u to E (considered as a real subspace of E 1), and that the spectral radius
r(u) is the smallest real number IX ~ such that for III > IX, A E R, the series (3)
converges in 2'(E).
2. PRINGSHEIM'S THEOREM AND ITS CONSEQUENCES

Perhaps the best-known result on positive operators in ordered Banach
spaces is the theorem that whenever the positive cone is total and u is a com-
pact positive endomorphism with spectral radius r(u) > 0, then r(u) is an
eigenvalue of u with an eigenvector ~ 0. This theorem, which has a compara-
tively long history, was first proved in the stated generality by Krein-Rutman
[I]. A more general theorem appeared in Bonsall [3] and was extended to
locally convex spaces by the author [3]. In this section, we shall derive this
and other results (some not dependent on compactness assumptions) in a
uniform way from a theorem on vector-valued analytic functions which is an
extension of a classical theorem due to Pringsheim. We shall need this
theorem for functions taking their values in a locally convex space.
2.1
Theorem. Let E be an ordered, semi-complete I.c.s. over C such that the
= 0, 1, ... ) and if L anz" has
00
positive cone C is weakly normal. If an E C (n

o
radius 0/ convergence 1, then the analytic/unction represented by the power
series is singular at z = 1. In addition, if this singularity is a pole, it is 0/ maximal
order on Izl = 1.
Proof. Let / be the functional element (with values in E) given by /(z) =
Lo
00
anzn when Izl < 1 and let the radius of convergence of this series be 1. Let x'
be any continuous, real linear form on E; the radius rx , of convergence of the
Lo <an> x') tn,
00
series where t is real, is ~ 1. Further we have inf{rx':x' E D}

= I, where D denotes the set of all continuous, real linear forms on E that
L antn would
00
are ~ 0 on C. For if we had inf{rx': x' E D} = 1'/ > 1, the series

o
converge in E for all t, -1'/ < t < 1'/, since by (V, 3.3), Corollary 3, the weak
normality of C is equivalent to Eo = D - D, where Eo is the underlying real
space of E (cf. (I, 7.2». Thus z --+ /(z) would have a holomorphic extension
(Chapter IV, Exercise 39) to the open disk Izl < 1'/, which is contradictory.
Let p, 0 x')t n = L <an> x')«t - L <bno x')(t -

00 00 00
p) + p)" = p)"
0 0
are non-negative, it follows that the series
Lo <bn> x')(t -
00
p)"
has radius of convergence r x , - p, and hence that
Lo bit -
00
p)"
has radius of convergence I - p. By a conclusion familiar from the theory of

analytic functions (cf. Chapter IV, Exercise 39(d» this implies that z = I is
singular for f.
Assume now that the singularity of/at z = I is a pole of order k. If , =
exp i() is any complex number of modulus I, and if z = t', 0 < t < 1, we have
limJ<lzDlz - 'I P = 0
' .... 1
for allp > k. Since C is a weakly normal cone, this implies, for any p > k, that
(1 - t)P L (tn cos nO)a n, (1 - t)P L (tn sin nO)a n

00 00
o 0
both converge to 0 for aCE, E') as t --+ 1. Thus if ( is a pole off of order m, it
follows that m ~ k and the theorem is proved.
In the first two of the following applications of the Pringsheim theorem, E
can be any ordered complex Banach space; the remaining results gain in trans-
parency by starting from a real space. The reader will notice that (2.2) applies,
in particular, to any ordered Banach space over R whose positive cone Cis
normal and generates E (for instance, a Banach lattice); just apply (2.2) to
the complexification El of E, ordered with positive cone C + iC.
2.2
Let E be an ordered complex Banach space, not reduced to {OJ, with positive
cone C such that C is normal and E = C - C. For any positive (necessarily
continuous) endomorphism u ofE, the spectral radius r(u) is an element of a(u);
ifr(u) is a pole of the resolvent, it is of maximal order on the spectral circle ofu.
Proof. By (Y, 5.5), any positive endomorphism u of E is continuous. In
view of the corollary of (Y, 3.5), C is a !l3-cone and hence (Y, 5.2) implies that
the cone H of positive endomorphisms of E is normal (hence weakly normal)
in :feE) for the topology of bounded convergence, that is, for the norm
topology of :feE). If r(u) > 0, Theorem (2.1) applies to z --+ fez) = R(r(u)/z) =
L un(z/r(u))n+l
00
(Section 1, Formula (3)); if r(u) = 0, u is a topological nil-

o
potent, a(u) = {O}, and the assertion is equally true.
2.3
Let E be an ordered complex (B)-space satisfying the hypothesis of(2.2), and
let u be a positive endomorphism of E. If AE p(u), then R(A) is positive if and
only if A is real and A> r(u).
Proof. It is clear that A > r(u) is sufficient for R(A) ~ 0 (with respect to the
canonical order of :fee)), in view of Formula (3) of Section 1. Suppose that
R(A) ~ 0 for some A E p(u). Choose an Xo > 0 and define recursively Xn =
R(A)Xn-1 (n EN). Each Xn satisfies the relation
(n EN). (*)
Clearly, xn E C for all n and, in fact Xn > 0 (for, Xn = 0 for some n EN
would imply Xo = 0). Moreover, by induction on n it is shown from (*) that
An Xn E C and An- 1 Xn E C for all n EN, and that
(n EN).
This necessitates A :1= 0 and we can assume that 1..1.1 = I, for if R(A) is positive
at ..1.:1= 0, then the resolvent of Ir 1 1u is positive at AlA-II. Let A = exp i9,
o ~ 9 < 21t, and suppose that 9 > O. It is clear that n9 ~ 1t (mod 21t) for all
positive integers n, or else C would not be a proper cone. Hence there exists a
smallest integer no > 0 such that the triangle with vertices I, exp i(no - 1)9,
exp in o9 in the complex plane contains 0 in its interior. Consider the unique
real subspace M of E of dimension 2 that contains the points x no ' Ano-lXno
and Anoxno • It follows that M (") C contains 0 as an interior point, which
conflicts with the fact that C is a proper cone; thus 9 = 0, and hence A > O.
Up to this point of the proof we have used C only as a proper cone:l= {O}.
Assume now that C is normal and E = C - C; as before, it follows from
(V, 3.5), Corollary, and (V, 5.2) that the positive cone :?e c ft'(E) is normal.
If it were true that R(A) ~ 0 for some A, 0 < A ~ r(u), then the resolvent equa-
tion (Section I, Formula (I» would imply that 0 ~ R(Jl) ~ R(A) for all
Jl > r(u) and hence, in view of the normality of :?e, that {R(Jl): Jl > r(u)} is a
bounded family in f£(E). This clearly contradicts (2.2) above (cf. Section I),
and hence it follows that A > r(u).
REMARK. The preceding proof shows that whenever E is an ordered
Banach space with positive cone C :1= {O} and u is a positive operator
(continuous endomorphism) of E, then R(A) ~ 0 implies A > O.
We now turn to the Krein-Rutman theorem on compact positive operators
mentioned at the beginning of this section; for a historical account and a
bibliography of earlier work on the subject, the reader should consult the
memoir of Krein-Rutman [I]. We derive the Krein-Rutman theorem from
the following result which establishes the conclusion of (2.2) for a restricted
class of positive operators but with no restriction on the ordering, except that
the positive cone be (closed and) total. It is clear that this latter condition is
indispensable; whether the additional condition on the operator can be
dropped appears to be unknown (very probably it can be further relaxed).
For continuous endomorphisms u of a real Banach space E, the terms
spectrum, resolvent, etc., refer to the complex extension of u to the complexi-
fication of E (cf. end of Section 1).
2.4
Let E be an ordered real Banach space with total positive cone C, and assume
that u is a continuous positive endomorphism of E whose resolvent has a pole
on the spectral circle IAI :;: : r(u). Then r(u) E u(u), and if r(u) is a pole of the
resolvent it is of maximal order on the spectral circle.
Proof. Since C is a closed, proper, total cone in E, its dual cone C' has the
same properties with respect to u(E', E), and hence G = C' - C' is a dense
subspace of the weak dual E~. If F denotes the space (E, u(E, G», then C is a
normal cone in F by (V,3.3), Corollary 3. Denote by E 1 , Fl the complexi-
fications (Chapter I, Section 7) of E, Frespectively. We consider El as ordered
with positive cone C; then the canonical order of !R(E1) is determined by the
positive cone J'f = {w E !R(E1 ): w(C) c C}. Moreover, we shall identify
u E !R(E) with its complex extension to E 1 •
Let us further denote by .:£ ,,(E1 , F1) the space of continuous linear maps of
El into F1 , provided with the topology of simple convergence on C. Then
there is a natural imbedding 1/1 of !R(E1 ) into ':£,.(El' F 1) which is continuous;
for simplicity of notation, we denote the images of elements and subsets of
!R(E1 ) under 1/1 by an index zero. We note first that by (V, 5.2) and the nor-
mality of C in F1 , the image J'f 0 of the cone J'f is normal in ':£,.(El' F1 ). Now
let C, ICI = r(u), .be a pole of order k (G; 1) of the resolvf\lnt A -+ R(A) of u, and
let a E !R(E1) be the leading coefficient of the principal part at A = C; one has
a = lim (A - C)k R(A); hence also a o = lim (A - ClRo(A). Suppose that r(u)
A~' A~'
¢ O'(u); then A -+ R(A) and a fortiori A -+ Ro(A) would be holomorphic at
= L r(n+l)uo of Ro at
00
A = r(u). Since the coefficients of the expansion RO(A)

o
infinity are elements of the normal cone J'f 0, (2.1) implies that Ro has an
extension, with values in the completion of f£ ,,(E1> F1 ), which is holomorphic
for IAI > T, where 0 ~ T < r(u); in particular, {RO(A): IAI > r(u)} is a bounded
family in .:£,.(E1 , FI ). Clearly, this implies a o = 0 and hence a = 0, which is
contradictory. Thus r(u) E O'(u).
For the final assertion we note that any pole of A -+ R(A) on IAI = r(u) is a
pole of the same order for Ro; thus the assertion follows again from Pring-
sheim's theorem (2.1). The theorem is proved.
COROLLARY (Krein-Rutman). Let E be an ordered real Banach space with

total positive cone C, and let u be a compact positive endomorphism of E. If u
has a spectral radius r(u) > 0, then r(u) is a pole of the resolvent of maximal
order on the spectral circle, with an eigenvector in C. A corresponding result
holds for the adjoint u' in E'.
Proof. Since u is compact the only possible singularities ::j: 0 of the resolvent
are poles, and there is at least one such singularity on IAI = r(u). Hence
A. = r(u) is a pole of some order k (G; 1) of the resolvent, and we have p =
lim (A. - r(u»kR(A.) for the leading coefficient of the corresponding principal
A~r(u)
part. Since R(A) G; 0 (for the canonical order of ':£(E» whenever A > r(u), it
follows that p G; 0, since the positive cone of ':£(E) is closed (cf. (V, 5.1». Since
C is total in E, there exists Y E C such that p(y) > 0; in view of (r(u)e - u)p = 0
it follows that p(y) is an eigenvector in C pertaining to r(u). Finally, if u' is the
adjoint of u in the strong dual E', we have O'(u) = O'(u') and A. -+ R(A)' is the
resolvent of u' (cf. (IV, 7.9». In particular, A. -+ R(A.)' has a pole at A = r(u')
= r(u), and we obtain the assertion for u' by taking adjoints throughout in the
preceding proof; in particular, p( C) c C implies p'(C') c C', and p' does not
vanish on C', since C' is total in E: and p' is continuous.
REMARK. If C is total in E, the preceding proof shows that for any

continuous, positive endomorphism u of E whose resolvent has a pole
at A = r(u), there exist eigenvectors of u in C and of u' in C' pertaining
to r(u).
The theorem of Krein-Rutman has generalizations in various directions;
of course, some compactness assumption has to be made. The remaining
two results in this section are typical for the sort of generalization that one
obtains.
Let E be an ordered real Banach space with positive cone C; a linear map u
of E into itself is called C-compact if u is continuous on C into C, and if
u(U n C) is relatively compact, U denoting the unit ball of E. We define the
C-spectral radius of u to be the number
re = lim (supllun(x)lI: x E C, Ilxll ~ l)l/n

n-+ 00
(the proof below will show that the limit always exists). In the following
theorem, we assume the normality of C merely for tbe convenience of proof;
this assumption is actually dispensable. The reader who wishes to obtain
further information is referred to Bonsall [4] and the author's paper [3],
Section 10.
2.5
Let E be an ordered real Banach space with normal positive cone C. If u is a
C-compact mapping in E such that r e > 0, then r e is an eigenvalue of u with an
eigenvector in C.
Proof. Denote by U the unit ball of E and by W the convex, circled hull of
Un C. Then {eW: e > O} is a O-neighborhood base for a normable topology
:! on the subspace Eo = C - C of E. It is readily seen that:! is identical with
the topology :!1 introduced in Chapter V, Section 3, Lemma 2. Thus if q is the
gauge function of W, (Eo, q) is a Banilch space; moreover, on C the norm q
agrees with the original norm of E. Thus C is a normal closed cone in (Eo, q),
and since re is nothing other than the spectral radius rev) of the restriction v
of u to Eo, it follows from (2.2) that r(v) E u(v). Hence {Rv(An): n EN} is
unbounded in .st«Eo, q» for any decreasing real sequence {An} such that limn An
= rev), and by the principle of uniform boundedness there exists Y E C such
that limn q(Rv(An)Y) = +<X) for a given sequence {An} tending monotonically
to rev). Let Xn = Rv(An)y/q(Rv(An)Y); then xn E C and q(xn) = Ilxn II = 1 for
all n. Moreover, limn q(AnXn - v(xn» = limn IIAnXn - u(xn) II = 0; this implies
lim. (ree - u)xn = 0 in E. Therefore, since the range of the sequence {u(xn)}
is relatively compact in E by hypothesis, the sequence {xn} has a cluster point
x in E (and hence in C, since C is closed). Clearly, this cluster point satisfies
reX = u(x) and Ilxll = 1, which completes the proof.
The second generalization that we have in mind concerns convex cones
with compact base. Let us recall (Chapter II, Exercise 30) that a convex cone
C of vertex 0 in a I.c.s. E has a compact base if there exists a (real) affine sub-
space N of E not containing 0, such that C (1 N is compact and C = {Ax:
A ~ 0, X EN (1 C}. From the separation theorem (II, 9.2) it is clear that there
exists a closed real hyperplane H strictly separating N (1 C from {O}; clearly,
then, C = {Ax: A ~ 0, X E H (1 C}.
2.6
Theorem. Let E be a I.c.s. over R and let C be a cone in E with compact
base. If u is an endomorphism of thp- subspace C - C of E such that u(C) c C
and the restriction of u to C is continuous, then u has an eigenvalue ~ 0 with an
eigenvector in C.
Proof. Let H = {x:f(x) = I} be a hyperplane in E such that H (1 C is a
compact base of C. Denote by V the convex hull of {OJ u (H (1 C) in E and
set U = V - V; then {eU: e> O} is a O-neighborhood base in Eo = C - C for
a normable topology '1:; it is not difficult to verify that the norm
z -> liz II = inf{f(x) + fey): z = x - y, x, Y E C}
generates the topology '1: on Eo. Moreover, since U is compact and hence
complete in E, and since '1: is finer on Eo than the topology induced by E, it
follows from (1, 1.6) that (Eo, '1:) is complete, hence (Eo, II II) is a Banach
space. Further, C is closed in this space and clearly normal, and by (Y, 5.5)
u is a continuous, positive endomorphism of (Eo, II II) for the order of Eo
whose positive cone is C. Thus from (2.2) above, it follows that the spectral
radius r(u) is a number in a(u). (It is quite possible that r(u) = 0, even if
u ¥= 0.) As in the proof of (2.5), we construct a sequence {x n } in C such that
Ilxn I = f(xn) = 1 for all n, and such that limn Ilr(u)x n - u(xn) II = O. Since H (1 C
is compact in E and u is continuous on C by hypothesis, every cluster point
x E H (1 C (for the topology induced by E) of the sequence {x n } satisfies
r(u)x = u(x). This completes the proof.
The following corollary is due to Krein-Rutman [1].
COROLLARY. Let E be an ordered real Banach space whose positive cone C has
interior points. If u is any positive (necessarily continuous) endomorphism of E,
there exists an eigenvalue ~ 0 of the adjoint u' with an eigenvector in the dual
cone C. If, in addition, C is normal (in particular, if E is an (AM)-space with
unit), then the spectral radius r(u) ofu is such an eigenvalue ofu'.
Proof. In fact, if Xo is interior to C, then the hyperplane H = {x': <x o , x')
= I} in E' has a a(E', E)-compact intersection with C; hence C is a cone
with compact base in E;, and u' (satisfying u'(C') c C) is a(E', E)-continuous
so that (2.6) applies. If, in addition, C is'a normal cone in E, then E' = C - C'
by (Y, 3.3), Corollary 3, and it is readily seen that the topology '1: constructed
in the proof of (2.6) is the topology of the strong dual Ep. Hence the number
r(u'), which was proved to be an eigenvalue of u', is the spectral radius of u' in
Ep and therefore equals r(u).
3. THE PERIPHERAL POINT SPECTRUM
Let u be a continuous positive endomorphism of an ordered Banach space

E. The subset of u(u) located on the spectral circle {A: IAI = r(u)} will be called
the peripheral spectrum, and its intersection with 7t(u) the peripheral point
spectrum of u. It has been pointed out earlier that it is apparently unknown
whether (2.4) is true for all positive operators on E; on the other hand, it is
natural to ask if, for example, under the hypothesis of (2.2) the peripheral
spectrum of u is possibly subject to further restrictions. The following example
shows that the answer is negative.
Example. Let E be a real Banach space of dimension at least 2; denote
by F any closed subspace of codimension I, and by G = {AXo: A E R}
a complementary subspace so that E = G E9 F (cf. (1,3.5». For each
x E E, let x = AXo + Y be the unique representation such that y E F.
The set C = {x E E: A ~ lIyll} is readily seen to be a closed normal cone
such that E = C - C; thus E, with C as its positive cone, is an ordered
Banach space such that the positive cone of !e(E) is normal. Now let
v E !l'(F) satisfy IIvll ~ 1, and define u E !l'(E) by
u(x) = AxO + v(y) (x E E);
that is, u = p + v 0 (e - p), where p denotes the projection x -+ AXo of E
onto G. Because of Ilvll ~ 1, we have u(C) c: C, and hence u is positive;
clearly, r(u) = 1 and u(u) = {I} u u(v), u(v) denoting the spectrum of
v E !e(F). By choosing E so that F is isomorphic with a suitable Banach
space and by a suitable choice of v E !e(F), it can be arranged that u(u)
is a preassigned closed subset of {A.: IAI = I} containing 1.
Therefore, for any further fruitful investigation of spectral properties of
positive operators, it is necessary to consider more restricted types of positive
maps and/or of ordered spaces. We present in this section a number of results
on the peripheral point spectrum that are valid under reasonably general
assumptions.
Several such results have been obtained by Krein-Rutman [1] for what they
called strongly positive operators on ordered Banach spaces whose positive
cone has non-empty interior; these operators u are such that for each non-zero
x E C, there exists n E N for which u"(x) is interior to C. This is a severe
restriction, since spaces without order units (such as U(p.), I ~ p < + 00) are
excluded from the discussion (cf. Chapter V, Exercise 10); on the other hand,
if E is a Banach lattice with order units, then in view of (V, 8.4) and (V, 8.5),
E is isomorphic (as an ordered t.v.s.) with <'CR(X) for a suitable compact space
X, and we obtain very strong results (see (3.3) below). We generalize the notion
of strong positivity as follows:
§3] THE PERIPHERAL POINT SPECTRUM 317
Let E be an ordered real Banach space with positive cone C #: {O}. A con-
tinuous, positive endomorphism u of E is called irreducible if for some scalar
A. > r(u) and each non-zero x E C, the element
L
00
uR(A.)x = A. -V'(x)
n=1
is a quasi-interior point of C. Recall that y E E is quasi-interior to C (Chapter

V, Section 7) if the order interval [0, y] is a total subset of E. If E is a normed
lattice (Chapter V, Section 7), the quasi-interior points of C are weak order
units of E; the following characterization of irreducible endomorphisms of
Banach lattices justifies our terminology.
3.1
A continuous positive endomorphism u #: 0 of a Banach lattice is irreducible
if and only if no closed solid subspace, distinct from {O} and E, is invariant
under u.
Proof. Let u be irreducible and let F #: {O} be a closed solid subspace
invariant under u. If 0 #: x E F Ii C, then y = uR(A.)x is, for suitable A. > r(u),
a quasi-interior point of C contained in F; hence F :;= E.
Conversely, if u leaves no closed proper solid subspace #: {O} of E in-
variant, then u(x) > 0 for each x > 0; for, u(xo) = 0 for some Xo > 0 would
imply that u leaves the closed solid subspace G invariant which is generated
by xo, hence G = E, and it would.follow that u = 0, which is contradictory.
Therefore, if x > 0, then y = uR(A.)x > 0 (A. > r(u) being arbitrary), and since
u(y) ~ A.y it follows that the closed solid subspace F generated by y is in-
variant under u. Hence F = E, so that y is a quasi-interior point of E.
In the proof of our first result on irreducible endomorphisms, we shall need
this lemma.
LEMMA 1. Let E be an ordered Banach space with positive cone C and P E !l'(E)
a positive projection. If x E p( C) is quasi-interior to C, then x is quasi-interior
to p( C) in pee).
Proof. Letting C1 = p(e) and E1 = pee), we observe that C1 Ii (x - C1) =
[0, X]1 = [0, x] Ii E1 = p([O, x]), since p is a positive projection. Since the
linear hull of [0, x] is dense in E, the linear hull of [0, X]1 is dense in the
subspace E1 by the continuity of p.
A linear form f on an ordered vector space E over R is termed strictly
positive if x > 0 impliesf(x) > O. Note also that the existence of an irreducible
positive endomorphism on an ordered Banach space E implies that the posi-
tive cone C of E is total (for C contains quasi-interior points); hence the dual
cone C' c: E', being a closed proper cone in E;, defines the canonical order
ofE'.
3.2
Let E be an ordered real Banach space with positive cone C, and suppose that
u is an irreducible positive endomorphism whose spectral radius r is a pole of the
resolvent. Then:
(i) r> 0 and r is a pole of order 1.
(ii) There exist positive eigenvectors, pertaining to r, of u and u'. Each positive
eigenvector for r is quasi-interior to C, and each positive eigenvector of u' for
r is a strictly positive linear form.
(iii) Each of the following assumptions implies that the multiplicity d(r) of
r is 1: (a) C has non-empty interior, (b) d(r) is finite, (c) E is a Banach lattice.
REMARK. (iii) can be replaced by the assertion" d(r) = 1" if there

exists no ordered Banach space of dimension > I where each x > 0 is
quasi-interior to C. On the other hand, if such a (necessarily infinite-
dimensional) space exists, the identity map e is irreducible and such
that d(r) = dim E.
Proof of (3.2). Let p denote the leading coefficient of the principal part of
the resolvent at A. = r and let q be the residue at r. Thenp is positive (cf. proof
of (2.4), Corollary), and p = q(u - re)k-I, where k is the order of the pole;
moreover, q and its adjoint q' are projections such that q(E), q'(E') are the null
spaces of (re - u)\ (re' - u't, respectively.
(ii): Since C and C' are total subsets of E and E;, respectively, there exist
eigenvectors (pertaining to r) of u in C, and of u' in C'. Let xo, Xo be any such
eigenvectors. From
uR(.1)xo = ~
00
~ Xo(r)" (A. > r)
it follows that r > 0 and that Xo is quasi-interior to C. Similarly, from
(x, xo) ~
00
~(r)" = ~
00
A. -"(u"(x), xo) = (UR(A)X, xo) (A. > r)
it follows that (x, xo) > 0 whenever 0 =1= x E C, for Xo must be > 0 at the
quasi-interior point uR(.1)x of C.
(i): We have to show that r, which is > 0 by the preceding, is a pole of
order 1. In fact, let x E C be such that Xo = p(x) is not zero and let x' E C' be
an eigenvector of u' for r. In view of p = q(u - re)k-l and q'(xo) = xo, we
obtain
0< (xo, xo) = (q(u - re)k-I x , xo) = (x, (u' - re,)k-l xo ),
which implies that k = 1.

(iii): Since A. = r is a simple pole of the resolvent, we have p = q; hence p
is a positive projection and, by (ii), every non-zero element of p( C) is quasi-
interior to C. Therefore, by Lemma 1 every non-zero x E p(C) is quasi-interior
to p( C) in peE). In case (a), C has interior points, hence so does p( C) in peE).

Since p(C) (which is contained in C) is a closed proper cone inp(E), it follows
that peE) has dimension 1 and hence that d(r) = 1. If (b) d(r) is finite, then
every quasi-interior point of p(C) in peE) is interior to p(C) (cf. the lemma
preceding (Y, 4.1) and the conclusion is the same as before.
There remains to show that d(r) = 1 if E is a Banach lattice. If x is any
eigenvector of u pertaining to r, then from rx = u(x) it follows that rlxl =
Iu(x)1 ~ u(lxl); if x~ E C' is an eigerivector of u' for r, we obtain
1'<lxl, x~) ~ )u(lxJ), x~) = r<lxl, x~),
and this implies that rlxl = u(lxi), since r > 0 and since x~ is a strictly positive
linear form by (ii). Now x = x+ - x- and Ixl = x+ + x- ; hence both x+, x-
are positive elements of the eigenspace of u pertaining to r. Since they are
lattice disjoint and the lattice operations are continuous in E, both cannot be
quasi-interior points of C; thus either x+ = 0 or else x- = O. Therefore, if x is
an eigenvector of u for r, we have either x E C or x E - C. Hence this eigen-
space is totally ordered; since this order is also Archimedean, it follows that
d(r) = 1 (Chapter Y, Exercise 2).
Our principal result on irreducible positive maps is concerned with Banach
lattices E of type ~ R(X), X being a compact space. To avoid confusion with
the unit e of 2(E), we shall denote the constantly-one function t --+ I on X by
1. It will be convenient to employ the following terminology: A positive
endomorphism u of ~ R(X) with spectral radius r is said to have a cyclic
peripheral point spectrum if raf= u(f), lal = 1, andf= Iflg E ~c(X) imply
that ranlflgn = u(lflgn) for all integers (n E Z); here u is identified with its
unique extension to ~c(X) (the complexification of ~ R(X», If I denotes the
usual absolute value of f E>~c(X), and fg denotes the function t --+ f(t)g(t)
(pointwise multiplication). We shall need the following lemma.
LEMMA 2. Let v be a positive endomorphism of ~ R(X) such that v(l) = 1.

If ag = v(g), where lal = 1 and Igl = 1, g E ~c(X), then a"g" = v(g") for all
nEZ.
Proof For each SEX, the mappingf --+ v(f)(s) is a continuous positive linear
form on ~R(X), hence a positive Radon measure Jls on X; from v(I) = 1 we '
conclude that IIJlsl1 = Jls(l) = 1. Hence ages) = Jx g(t) dJl.(t) for each SEX,
and from Igl = 1 it follows that get) is constant on the support of Jls' namely
equal to ages). Therefore, a"g"(s) = Ixg"(t) dJl.(t) (s E X, n E Z), which is the
assertion.
In view of the representation theorem (Y, 8.5), the following result is valid
for irreducible positive maps of an arbitrary (AM)-space with unit.
3.3
Theorem. Let X be a compact space and suppose that u is an irreducible
positive endomorphism of~R( X). Then the following assertions hold:
(i) The spectral radius r l?f u is > 0, and the condition lIu II = r is equivalent
with u(1) = r1.
(ii) The peripheral point spectrum of u is cyclic.
(iii) Each eigenvalue NI., lal = 1, ofu has multiplicity 1 and the corresponding
eigenfunction is #: 0 throughout; moreover, u(u) is invariant under the rotation
through e, where a = exp ie.
(iv) If the peripheral point spectrum contains an isolated point, then it is
of the form rH, where H is the group ofnth roots of unity for some n ~ 1.
(v) If the peripheral point spectrum contains a pole of the resolvent of u,
then all its points are poles of order 1 of the resolvent.
(vi) r is the only possible eigenvalue of u with an eigenfunction ~ O. If X is
connected, the peripheral point spectrum cannot contain points rrx such that
rx is a root of unity distinct from 1.
REMARK. It can happen that the peripheral point spectrum of u is
empty; on the other hand, even if X is connected, the peripheral point
spectrum can be dense in the spectral circle (see examples below). In the
latter case, it is still of the form rG where G is a subgroup of the circle
group.
Proof of (3.3). By the corollary of (2.6), there exists a positive Radon
measure /lo on X such that r/lo = u'(/lo). Since {f: /lo(lfl) = O} is a solid
subspace of~ R(X) invariant under u, /lo is strictly positive which is equivalent
to the assertion that the support So of /lo equals X.
We proceed to prove the statements of the theorem in the order of enumera-
tion.
(i): The assumption r = 0 implies u'(/lo) = 0, hence Ju(l)d/l o = 0; since
So = X, it follows that u(l) = 0, hence u = 0, which is contradictory. Thus
r> O. Clearly, u(1) = rl implies lIull = r; conversely, if lIull = r, then u(l) ~ r1,
and <rl - u{l), /lo) = r/lo(1) - <u(l), /lo) = 0; from So = X we conclude
that u(l) = r1.
(ii): Suppose that rrxf = u(f), where f #: 0 and Irxl = 1. We obtain rlfl =
rlrxfl = lu(f)1 ~ u(lfD; since u'(/lo) = r/l o, it follows that <u(lfl) - rlfl, /lo)
= 0, and we conclude, as before, that rlfl = u(lfD. This impliesf(s) #: 0 for
all SEX. In the opposite case, the closed subspace of ~ R(X) generated by the
order interval [0, If I] would be a closed solid sublattice, neither {OJ nor the
whole space, and clearly invariant under u; this is impossible by (3.1).
Now letf= Iflg and define a positive endomorphism v Of~R(X) by
v(h)(s) = r-1If(s)I-1uClflh)(s) (s EX).
We have vel) =1 and the complex extension of v (again denoted by v)

satisfies a.g = v(g). From Lemma 2 it follows that a!'gft = v(gft) for all n E Z
which clearly implies the assertion.
(iii): Let ra, where a. = exp iO, be an eigenvalue of u. It has been shown
under (ii) that any eigenfunction I pertaining to ra. is ::f. 0 throughout; more-
over,/o = III is an eigenfunction for r. If h is any other eigenfunction for r,
we can assume that h is real valued; set c = sup{h(t)/fo(t): t E X}. The function
clo - h belongs to the eigenspace N(r) and vanishes for at least one t E X,
since the supremum is assumed in X. Hence clo - h-,= 0 by the preceding,
that is, the multiplicity of r is 1. I
Denote by w the endomorphism of 'ifc(X) defined by h --+ gh, where

9 (with Igl = 1) is the function 171/1 introduced above. We define v by v =
OC- 1W- 1UW; as a continuous endomorphism of~c(X), v is given by
v(h)(s) = fh(t) dv.(t) (s EX),
where each v. is a uniquely determined complex Radon measure on X. Simi-

larly, let u(h)(s) = Jh(t)dfJ.(t)(s E X); in view of Igl = 1, we obtain
IJh(t) dv.(t) I = Ig(s) - l u(gh)(s)1 ~ u(lhl)(s) = JIh(t)1 dfJ.(t)

for all SEX. Thus if v. = P. + iT. is the decomposition of v. by means of real
Radon measures (Chapter I, Section 7), it follows that P. ~ fJ. for all s. On the
other hand, v(fo) = rio and u(fo) = rio, whence J/od(fJ. - P.) = 0 for all SEX;
since lo(t) > 0 throughout, it follows that P. = fJ•. Therefore, T. = 0 for all s
and hence v = u; in other words,
(*)
From this it is clear that q(u) = q(a.u), hence q(u) is invariant under rotation
through 0 (a. = exp iO). Formula (*) also shows that if there are any elements
in the peripheral point spectrum, then their common multiplicity is 1 (namely,
equal to that of r).
(iv): Suppose that ra., where a. = exp iO ::f. 1, is an isolated element of the
peripheral point spectrum of u; it follows from (*) that r is such an element.
Consequently, there exists an eigenvalue r exp iOI (0 < 0 1 < 2n) for which
01 is minimal; for, since r is isolated, (ii) implies that the peripheral point
spectrum consists of a finite number of roots of unity. In particular, 01 = 2n/n
for some n > 1 and (again by (ii» the numbers r exp imOI (m = 0, 1, "0' n - 1)
are all eigenvalues. Now let r exp iO be any eigenvalue on IAI = r and denote
by k the smallest integer > 0 such that 0 + kO I ~ 2n. Let X = 0 + k01 ; since
r exp ikOI is an eigenvalue of u, it follows from (*), applied to a. = exp iO, that
r exp iX is an eigenvalue of u. Now if we had X > 2n, we would also have
X < 2n + 01> which contradicts the definition of 01 , Thus 0 + kOI = 2n
and hence 0 = (n - k)OI ; this shows that the penpheral point spectrum of u is
exactly the set rH, where H denotes the group of nth roots of unity.
(v): If the peripheral point spectrum contains a pole of the resolvent of u,
then, clearly, the preceding applies; it follows from (*) that each element of
rHis a pole of the same order, this common order being 1 by (3.2).
(vi): Suppose that ;,r = u(f), where 0 ::p!?;, 0; from ;"(f, 110) = (u(f),l1o)
= r(f, 110) > 0 it follows that;" = r.
Let ra, where a is a primitive nth root of unity, be an eigenvalue of u with
eigenfunction! = 1!lg; if v is the mapping h --+ r- 1 1!1- 1 u(l/lh), then ag = v(g),
and v satisfies the hypothesis of Lemma 2. We define Mk = g-l(ak) (k = 0,
1, ... ), and without loss in generality we can assume that Mo ::P 0, that is,
get) = 1 for some t E X. As in the proof of Lemma 2, we write v(h)(s) =
Sh(t)dl1s(t) (s E X) and conclude from ag = v(g), Igl = 1, that whenever s E Mk
then the support of I1s is contained in M k + 1. Since k == k' (mod n) implies
Mk = M k, (sets with indices incongruent mod n being disjoint), the map v
induces a cyclic permutation Mk --+ M k- 1 (k mod n). From this it follows that
"-1
the closed solid sublattice Fe CC R( X) of functions vanishing on M = U Mk is
o
invariant under v and hence under u; hence by (3.1), F = {OJ or, equivalently,
X = M, since M is closed. Therefore, g(X) is the cyclic group generated by a;
on the other hand, g(X) is connected, since X is connected and g is contin-
uous, and this implies that a = 1.
An example of an irreducible positive operator with empty point spectrum
is furnished by the endomorphism u of ~R[O, 1] defined by
u(j)(s) = sl(s) + fi(t) +f:(1 -

dt t)21(t) dt
for s E [0, 1]. It is immediate that u is irreducible, and it is not difficult to

verify that u has no eigenvalues.
For our second example, let X be the unit circle {z:lzl = I}, let a be a fixed
element of X which is not a root of unity, and let u be the endomorphism of
C(J R(X) defined by u(f)(z) = I( az); then u is positive, and irreducible since the
set {a": n EN} is dense in X. It is easy to see that the group H = {an: n E Z}
belongs to the point spectrum of u; in fact the peripheral point spectrum of
u is exactly H.
Our final theorem is concerned with a condition under which a positive
endomorphism of a general Banach lattice has cyclic peripheral point spec-
trum (in a slightly weakened sense). If E is a Banach lattice, El the Banach
space which is the complexification of E, we shall say that the absolute value
x --+ Ixl of E can be extended to E1 if for each pair (x, y) E Ex E, sup{lx cos fJ
+ y sin fJl: 0 ~ fJ < 2n} exists in E; in these circumstances, the supremum
serves to define Ix + iyl. The reader will observe that the absolute value of E
can be extended to El whenever E is order complete; however, this condition
is not a necessary one as the example of the Banach lattices C(J R(X) (X com-
pact) shows. The extension of x --+ Ixl to E1> if it exists, satisfies the relations
§3l THE PERIPHERAL POINT SPECTRUM 323
IZI + z21 ;;;£ IZII + IZ21, Ipzl = Iplizi (z, Zl' Z2 E £1; P E C), and IU(Z) I ;;;£ u(lzi)
whenever u is (the extension to El of) a positive endomorphism of E.
3.4
Theorem. Let E be a Banach lattice whose absolute value can be ex-

tended to the complexification El of E, and let u be a positive endomorphism of E,
with spectral radius r, for ~hich there exists a strictly positive linear form Xo
satisfying u'(xo);;;£ rxo. If the peripheral point spectrum of u contains r(X,
I(XI = 1, then it contains rH, where H is the cyclic group generated by (x.
Proof. We can suppose that r > 0, and hence for convenience we assume
r = 1. Let (Xx = u(x), where 0 # x = Xl + iX2 EEl (Xl' x 2 E E) and I(XI = I; it
follows that Ixl = l(Xxl = lu(x)1 ;;;£ u(ixi). Now <lxi, xo) ;;;£ <u(ixi), x o) ;;;£
<lxi, xo) by virtue of the hypothesis u'(xo) ;;;£ xo; we obtain <u(lxi) - lxi, xo)
= 0 and hence Ixl = u(ixi), since Xo is strictly positive.
00
Define Xo = lxi, and consider the solid sublattice F = U n[ -Xo, xo] of E.

1
Since E is complete and the order interval [-xo, xo] is closed, every positive
sequence of type II in F is order sum mabIe to an element of F, and Xo is an
order unit of F; hence by (V, 6.2), F is complete under its order topology
(in general, Fis not a closed subspace of E). Moreover, by (V, 8.4), (F, p) is an
(AM)-space with unit Xo where p stands for the gauge function of [-xo, xo].
Hence by (V, 8.5) there exists an isomorphism l/J of (F, p) onto CCR(X) (X
compact), which extends to an isomorphism of the complexification Fl of F
onto CCc(X); in addition, Fl can be identified with a subspace of E l . Clearly, F
and FI are invariant under u, and the restriction Uo of u to F induces a positive
endomorphism v of CC R(X) (precisely, v = l/J a Uo a l/J-I). Since l/J(x o) = I (the
constantly-one function on X), we have v(l) = I from u(xo) = Xo; moreover,
we have x E FI and hence g = l/J(x) for some g E CCc(X) satisfying (Xg =v(g),
and Igl = I, since l/J preserves absolute values. The assertion follows now
from Lemma 2, and the proof is complete.
The preceding theorem applies to all Banach lattices U(/1), (Y, :E, /1) being
an arbitrary measure space (Chapter II, Section 2, Example 2). If, in this
particular case, u satisfies the hypothesis of (3.4) and (assuming r(u) = 1)
(Xf = u(f), then the Banach lattice (F, p) constructed in the preceding proof can
be identified with the set of all classes (mod ,u-null functions) Iflg, where g
contains a function bounded on Yo and vanishing on Y ~ Yo, Yo being the
set on which some fixed representative of the classfis # O. The space (F, p) is
therefore essentially L "'(,u, Yo), and the isomorphism l/J can be chosen to be
an isomorphism of the Banach algebra L "'(/1, Yo) onto the Banach algebra
CCR(X) (cf. Chapter V, Exercise 24). It follows that if we write f = Iflg With
gEL 00(/1, Yo), then (Xnlflgn = u(lflg") for all n E Z; it is thus reasonable to
extend the meaning of the term "cyclic peripheral point spectrum" to the
present case, where E = LP(Jl) (1 ~P~ (0). In particular, we obtain the fol-
lowing corollary.
COROLLARY (G.-C. Rota [1]). Every positive endomorphism u of L 1(Jl),
satisfying r(u) = lIull, has cyclic peripheral point spectrum.
Proof: In fact,! -+ J/dJl is a strictly positive linear form h on L 1(Jl), and lIuli
= r(u) implies that (u(j), h) = Ju(f)dJl ~ lIullJfdJl = r(u)(f, h) whenever
f~ 0; it follows that u'(h) ~ r(u)h, completing the proof.
INDEX OF SYMBOLS
PREREQUISITES
Section A, I: x E X, X ¢ X
Xc Y,X= Y
{x E X: (P)X}, {x: (P)X}, {x}
Y-X,0
=>,-
~(X)
2: f: X-+ Y,x-+f(x)
f(A),f- 1(B)
fog
fA
{x..: at: E A}, {x..}, {xn }
N
4: U, n, n
xA
X/R
5: ~,<,~, >
sup A, inf A, sup(x, y), inf(x, y)
B, I: (4'~)
A,A
6: (X, m)
X
C,2: L
R,C
3: LIM
L(Ll' L 2 )
4: Ko
L*
5: R+
325
326 INDEX OF SYMBOLS
CHAPTER I
Section I: (L, ~), L(~)
C6' K(X)
L
2: TIaLa
Ml EB· .. EB Mn
Ixl
'pP, £P (0 <P < 1)
7: e
Lo
Exercise I: EBaLa
CHAPTER II
Section I: r A, raAa
PM
2: II II
(L, II II)
.P(L, N)
B(X)
C6'(X)
(X,:E, p)
'pP(p), LP(p)
./t(X)
[x, y]
Ml.
[~, [2(A)
0(*
u*
4: L'
5: (1(E, F)
limgapEp
~-
6: EBaEa
lim hpaEa
~
lim En
~
EC(G), EC G = EC(G)
EC
CHAPTER III
Section 3: M(S, V)
.P(E, F)
.P 6(E, F)
6, i;
INDEX OF SYMBOLS 327
Section 4: L(E, F)
5: fx, fy
B(E, F; G), !B(E, F; G), &B(E, F; G)
B(E, F), !B(E, F), &B(E, F)
!Be(E;, F;; G), !BiE;, F;)
6: E®F,x®y
A®B
:Ip
E®F
Ll(/1)
:Ii
:Ie
E®F
7: E y, <py
E B,1/IB
<py.u
1/IC.B
IP
8: ~
;If, ;If(C)
!/'
!/'(Rk), ~(Rk), £&(Rk)
9: !£(E)
e
!£C<E, F)
!£c(E)
!£b(E, F)
CHAPTER IV
Section I: (x, y), (F, G), (F, G, ( »
(E, E*)
<Pd, Wd
a(F, G), a(E', E)
M O , MOO
E;
2: U*, u'
3: reF, G)
4: TIa6a
EBa 6 a
!£,(E, F)
5: P(F, G)
E a , E" Ep
E;, E;, Ep
328 INDEX OF SYMBOLS
5: E"
Xc
6: Xf
Xpc
7: D u , u'
9: !l'iE;, F)
/(E, F)
p(f), Jxfdp
fJlp(E, F)
fJlb(E, F)
F;
[l(A, E), [1 [A, El, [1(A)
CHAPTER V
Section I: [x, y]
e
Lb
C*
L+
Ixl
x..ly
A.l
x+
x
BA
x-+x
3: [A]
[~]
Sc
C'
4: Sx
6: Xo
Pe
8: C6'o(X)
C6'c(X)
C6'R(X)
Sf
Sp
CHAPTER VI
Section I: R(A.,x)
p(X), a(x) , rex)
All
INDEX OF SYMBOLS 329
2: Asa
!t'(H)
Cr(G)
A
3: A+
x+,x-
v'x
Ixl
4: ¢J*
¢J+,¢J-
(n"" H"" c;",)
5: GeC
6: H®H
II xII fJ
P
P A q,p V q
W~
7: MC,McC
8: Z(A)
p"'q
-<
St(x), Sr(x)
z(x)
WI, WIl, WIll' WII.X)' Wl/l
Appendix: u(u)
p(u)
n(u)
R(J.)
r(u)
rc
BIBLIOGRAPHY
ANDO, T.
[1] Positive linear operators in semi-ordered linear spaces. J. Fac. Sci. Hokkaido
Univ., Ser. I., 13 (1957), 214-228.
[2] On fundamental properties of a Banach space with a cone. Pacific J. Math.,
12 (1962), 1163-1169.
ARENS, R. F.
[1] Duality in linear spaces. Duke Math J., 14 (1947),787-794.
ARENS, R. F. AND J. L. KELLEY
[1] Characterizations of the space of continuous functions over a compact
Hausdorff space. Trans. Amer. Math. Soc., 62 (1947), 499-508.
BAER, R.
[1] Linear algebra and projective geometry. New York 1952.
BANACH, S.
[1] Theorie des operations lineaires. Warsaw 1932.
BANACH, S. ET H. STEINHAUS
[1] Sur Ie principe de la condensation de singuiarites. Fund. Math., 9 (1927),
50-61.
BARTLE, R. G.
[1] On compactness in functional analysis. Trans. A mer. Math. Soc., 79 (1955),
35-57.
BAUER, H.
[1] Sur Ie prolongement des formes lineaires positives dans un espace vectoriel
ordonne. C. R. Acad. Sci. Paris, 244 (1957), 289-292.
[2] Ober die Fortsetzung positiver Linearformen. Bayer. Akad. Wiss. Math.-Nat.
Kl. S.-B., 1957 (1958), 177-190.
BERGE, C.
[1] Topological spaces. New York 1963.
BIRKHOFF, G.
[1] Lattice theory. 3rd ed. New York 1961.
330
BIBLIOGRAPHY 331
BIRKHOFF, G. AND S. MAC LANE

[1] A survey of modern algebra. 3rd ed. New York 1965.
BONSALL, F. F.
[1] Endomorphisms of partially ordered vector spaces. J. London Math. Soc., 30
(1955), 133-144.
[2] Endomorphisms of a partially ordered vector space without order unit.
J. London Math. Soc., 30 (1955), 144-153.
[3] Linear operators in complete positive cones. Proc. London Math. Soc. (3), 8
(1958), 53-75.
[4] Positive operators compact in an auxiliary topology. Pacific J. Math., 10
(1960), 1131-1138.
BOURBAKI, N. ELEMENTS DE MATHEMATIQUE.
[1] Theorie des ensembles. Fascicule de resultats. 3rd ed. Paris 1958.
[2] Algebre, chap. 2. 3rd ed. Paris 1962.
[3] Algebre, chap. 3. 2nd ed. Paris 1958.
[4] Topologie generale, chap. 1 et 2. 3rd ed. Paris 1961.
[5] Topologie generale, chap. 9. 2nd ed. Paris 1958.
[6] Topoiogie generale, chap. 10. 2nd ed. Paris 1961.
[7] Espaces vectoriels topologiques, chap. 1 et 2. Paris 1953.
[8] Espaces vectoriels topoiogiques, chap. 3-5. Paris 1955.
[9] Integration, chap. 1-4. Paris 1952.
[10] Integration, chap. 5. Paris 1956.
[11] Integration, chap. 6. Paris 1959.
BOURGlN, D. G.
[1] Linear topological spaces. Amer. J. Math., 65 (1943), 637-659.
BRACE, J. W.
[1] The topology of almost uniform convergence. Pacific J. Math., 9 (1959),
643-652.
[2] Approximating compact and weakly compact operators. Proc. Amer. Math.
Soc., 12 (1961), 392-393.
BUSHAW, D.
[1] Elements of general topology. New York 1963.
CIVIN, P. AND B. YOOD
[1] Quasi-reflexive spaces. Proc. Amer. Math. Soc., 8 (1957), 906-911.
COLLINS, H. s.
[1] Completeness and compactness in linear topological spaces. Trans. Amer.
Math. Soc., 79 (1955), 256-280.
DAY, M. M.
[1] The spaces £P with 0 < p < 1. Bull. Amer. Math. Soc., 46 (1940),816-823.
[2] Normed linear spaces. 2nd ed. Berlin-Gottingen-Heideiberg 1962.
[3] On the base problem in normed spaces. Proc. Amer. Math. Soc., 13 (1962),
655-658.
DE BRANGES, L.
[1] The Stone-Weierstrass theorem. Proc. Amer. Math. Soc., 10 (1959), 822-824.
332 BIBLIOGRAPHY
DIEUDONNE, J.
[1] La dualite dans les espaces vectoriels topologiques. Ann. Sci. Ecole Norm.
Sup. (3), 59 (1942), 107-139.
[2] Natural homomorphisms in Banach spaces. Proc. Amer. Math. Soc., 1 (1950),
54-59.
[3] Sur les espaces de Kothe. J. Analyse Math., 1 (1951), 81-115.
[4] Sur un theoreme de Smulian. Arch. Math., 3 (1952), 436-440.
[5] Complex structures on real Banach spaces. Proc. Amer. Math. Soc., 3 (1952),
162-164.
[6] Sur les proprietes de permanence de certains espaces vectoriels topologiques.
Ann. Soc. Polon. Math., 25 (1952), 50-55 (1953).
[7] Sur les espaces de Montel metrisables. C. R. A cad. Sci. Paris, 238 (1954),
194-195.
[8] On biorthogonal systems. Michigan Math. J., 2 (1954),7-20.
[9] Denumerability conditions in locally convex vector spaces. Proc. Amer.
Math. Soc., 8 (1957), 367-372.
DIEUDONNE, J. ET L. SCHWARTZ
[1] La dualite dans les espaces (F) et (LF). Ann.1nst. Fourier (Grenoble), 1 (1949),
61-101 (1950).
DUNFORD, N. AND J. T. SCHWARTZ
[1] Linear operators. Part I: General theory. New York 1958.
[2] Linear operators. Part II: Spectral theory. New York 1963.
DVORETZKY, A. AND C. A. ROGERS
[1] Absolute and unconditional convergence in normed linear spaces. Proc.
Nat. A cad. Sci. U.S.A., 36 (1950), 192-197.
EBERLEIN, W. F.
[1] Weak compactness in Banach spaces. I. Proc. Nat. A cad. Sci. U.S.A., 33
(1947), 51-53.
FREUDENTHAL, H.
[1] Teilweise geordnete Moduln. Nederl. Akad. Wetensch. Proc., 39 (1936),
641-651.
FROBENIUS, G.
[1] Uber Matrizen aus positiven Elementen. S.-B. Preuss. Akad. Wiss. Berlin
1908,471-476; 1909, 514-518.
[2] Uber Matrizen aus nicht negativen Elementen. S.-B. Preuss. Akad. Wiss.
Berlin 1912, 456-477.
GILLMAN, L. AND M. JERISON
[1] Rings of continuous functions. Princeton 1962.
GORDON, H.
[1] Topologies and projections on Riesz spaces. Trans. Amer. Math. Soc., 94
(1960), 529-551.
[2] Relative Uniform Convergence. Math. Ann., 153 (1964), 418-427.
GROSBERG, J. ET M. KREIN
[1] Sur la decomposition des fonctionnelles en composantes positives. C. R.
(Doklady) Acad. Sci. URSS (N.S.), 25 (1939), 723-726.
BIBLIOGRAPHY 333
GROTHENDIECK, A.
[1] Sur la completion du dual d'un espace vectoriel localement convexe. C. R.
Acad. Sci. Paris, 230 (1950), 605-606.
[2] Quelques resultats relatifs a la dualite dans les espaces (F). C. R. Acad. Sci.
Paris, 230 (1950), 1561-1563.
[3] Criteres generaux de compacite dans les espaces vectoriels localement con-
vexes. Pathologie des espaces (LF). C. R. Acad. Sci. Paris, 231 (1950), 940-941.
[4] Quelques resultats sur les espaces vectoriels topologiques. C. R. A cad. Sci.
Paris, 233 (1951), 839-841.
[5] Sur une notion de produit tensoriel topologique d'espaces vectoriels top-
ologiques, et une cIasse remarquable d'espaces vectoriels liee it. cette notion.
C. R. Acad. Sci. Paris, 233 (1951), 1556-1558.
[6] Criteres de compacite dans les espaces fonctionnels generaux. Amer. J. Math.,
74 (1952), 168-186.
[7] Sur les applications lineaires faiblement compactes d'espaces du type C(K).
Canadian J. Math., 5 (1953), 129-173.
[8] Sur certains espaces de fonctions holomorphes, I, II. J. reine angew. Math.,
192 (1953), 35-64, 77-95.
[9] Sur les espaces de solutions d'une cIasse generale d'equations aux derivees
partielles. J. Analyse Math., 2 (1953), 243-280.
[10] Sur les espaces (F) et (DF). Summa Brasil. Math., 3 (1954),57-123.
[11] ~ons sur les espaces vectoriels topologiques. Instituto de Matematica Pura
e ApJicada Universidade de Sao Paulo. 2nd ed. Sao Paulo 1958.
[12] Resume des resultats essentiels dans la theorie des produits tensoriels
topologiques et des espaces nucIeaires. Ann. Inst. Fourier (Grenoble), 4 (1952),
73-112 (1954).
[13] Produits tensoriels topologiques et espaces nucIeaires. Mem. Amer. Math.
Soc. no. 16 (1955).
[14] Une caracterisation vectorielle-metrique des espaces V. Canadian J. Math., 7
(1955), 552-561.
HALMOS, P. R.
[1] Measure theory. 5th ed. Princeton 1958.
[2] Introduction to Hilbert space and the theory of spectral multiplicity. 2nd ed.
New York 1957.
[3] Naive set theory. Princeton 1960.
HILLE, E. AND R. S. PHILLIPS
[1] Functional analysis and semi-groups. 2nd ed. Providence, R.I. 1957.
HOFFMAN, K.
[1] Banach spaces of analytic functions. Englewood Cliffs, N.J., 1962.
HYERS, D. H.
[1] Locally bounded linear topological spaces. Rev. Ci. (Lima), 41 (1939),
555-574.
[2] Linear topological spaces. Bull. Amer. Math. Soc., 51 (1945), 1-24.
JAMES, R. C.
[1] Bases and reflexivity of Banach spaces. Ann. of Math. (2), 52 (1950),518-527.
[2] A non-reflexive Banach space isometric with its second conjugate space.
334 BIBLIOGRAPHY
Proc. Nat. Acad. Sci. U.S.A., 37 (1951), 174-177.

[3] Weak compactness and reflexivity. Israel J. of Math., 2 (1964), 101-119.
JERISON, M. (See Gillman, L., and M. Jerison.)
KADISON, R. V.
[1] A representation theory for commutative topological algebras. Mem. Amer.
Math. Soc. no. 7 (1951).
KAKUTANI, s.
[1] Concrete representation of abstract (L)-spaces and the mean ergodic theorem.
Ann. of Math. (2), 42 (1941), 523-537.
[2] Concrete representation of abstract (M)-spaces. Ann. of Math. (2), 42 (1941),
994-1024.
KARLIN, S.
[1] Unconditional convergence in aanach spaces. Bull. Amer. Math. Soc., 54
(1948), 148-152.
[2] Bases in Banach spaces. Duke Math. J., 15 (1948), 971-9~5,
[3] Positive operators. J. Math. Mech., 8 (1959), 907-937.
(See also R. F. Arens)
KELLEY, J. L.
[1] General topology. New York 1955.
KELLEY, J. L., I. NAMIOKA AND CO-AUTHORS
[1] Linear topological spaces. Princeton 1963.
KIST, J.
[1] Locally o-convex spaces. Duke Math. J., 25 (1958), 569-582.
KLEE, v. L.
[1] Invariant metrics in groups (solution of a problem of Banach). Proc. Amer.
Math. Soc., 3 (1952), 484-487.
[2] Boundedness and continuity of linear functionals. Duke Math. J., 22 (1955),
263-270.
[3] Extremal structure of convex sets. Arch. Math., 8 (1957), 234-240.
[4] Extremal structure of convex sets. II. Math. z., 69 (1958),90-104.
[5] Convexity. Princeton (to appear).
KOLMOGOROFF, A.
[1] Zur Normierbarkeit eines allgemeinen topologischen linearen Raumes.
Studia Math., 5 (1934), 29-33.
KOTHE, G.
[1] Die Stufenriiume, eine einfache Klasse linearer vollkommener Riiume.
Math. Z., 51 (1948), 317-345.
[2] Uber die Vollstiindigkeit einer Klasse lokalkonvexer Riiume. Math. z., 52
(1950), 627-630.
[3J Uber zwei Siitze von Banach. Math. Z., 53 (1950), 203-209.
[4] Neubegriindung der Theorie der vollkommenen Riiume. Math. Nachr., 4
(1951), 70-80.
[5] Topologische lineare Riiume. I. Berlin-Gottingen-Heidelberg 1960.
KOMURA, Y.
[I] Some examples on linear topological spaces. Math. Ann., 153 (1964), 150-162.
BIBLIOGRAPHY 335
KREIN, M. (See also Grosberg, J. and M. Krein.)

[1] Sur quelques questions de la geometrie des ensembles convexes situes dans
un espace lineaire norme et complet. C. R. (Doklady) Acad. Sci. URSS (N.S.), 14
(1937), 5-8.
[2] Proprietes fondamentales des ensembles coniques normaux dans l'espace
de Banach. C. R. (Doklady) Acad. Sci. URSS (N.S.), 28 (1940), 13 -17.
KREIN, M. AND D. MILMAN
[1] On extreme points of regular convex sets. Studia Math., 9 (1940), 133-138.
KREIN, M. G. AND M. A. RUTMAN
[1] Linear operators leaving invariant a cone in a Banach space. Uspehi Mat.
Nauk (N.S.) 3, no. 1 (23) (1948), 3-95. (Russian). Also Amer. Math. Soc. Transl.
no. 26 (1950).
KREIN, M. AND V. §MULIAN
[1] On regularly convex sets in the space conjugate to a Banach space. Ann. of
Math. (2) 41 (1940), 556-583.
LANDSBERG, M.
[1] Pseudonormen in der Theorie der linearen topologischen Raume. Math.
Nachr., 14 (1955), 29-38.
[2] Lineare topologische Raume, die nicht lokalkonvex sind. Math. Z., 65 (1956),
104-112.
MACKEY, G. W.
[1] On infinite dimensional linear spaces. Proc. Nat. Acad. Sci. U.S.A., 29 (1943),
216-221.
[2] On convex topological linear spaces. Proc. Nat. Acad. Sci. U.S.A., 29 (1943),
315-319.
[3] Equivalence of a problem in measure theory to a problem in the theory of
vector lattices. Bull. Amer. Math. Soc., 50 (1944), 719-722.
[4] On infinite-dimensional linear spaces. Trans. Amer. Math. Soc., 57 (1945),
155-207.
[5] On convex topological linear spaces. Trans. Amer. Math. Soc., 60 (1946),
519-537.
MAC LANE, s. (See Birkhoff, G., and S. Mac Lane.)
MAHOWALD, M.
[1] Barrelled spaces and the closed graph theorem. J. London Math. Soc., 36
(1961), 108-110.
MARTINEAU, A.
[1] Sur une propriete caracteristique d'un produit de droites. Arch. Math., 11
(1960), 423-426.
MAURIN, K.
[1] Abbildungen vom Hilbert-Schmidtschen Typus und ihre Anwendungen.
Math. Scand., 9 (1961), 359-371.
MILMAN, D. P. (See also Krein, M., and D. P. Milman.)
[1] Characteristics of extreme points of regularly convex sets. Dokl. Akad. Nauk.
SSSR (N.S.), 57 (1947), 119-122.
336 BIBLIOGRAPHY
NACHBIN, L.
[1] Topological vector spaces of continuous functions. Proc. Nat. A cad. Sci.
U.S.A., 40 (1954), 471-474.
NAMIOKA, I. (See also Kelley, J. L., I. Namioka, and co-authors.)
[1] Partially ordered linear topological spaces. Mem. Amer. Math. Soc. no. 24
(1957).
[2] A substitute for Lebesgue's bounded convergence theorem. Proc. Amer.
Math. Soc. 12 (1961), 713-716.
V. NEUMANN, J.
[1] On complete topological linear spaces. Trans. Amer. Math. Soc., 37 (1935),
1-20.
PERESSINI, A. L.
[1] On topologies in ordered vector spaces. Math. Ann., 144 (1961), 199-223.
[2] Concerning the order structure of Kothe sequence spaces. Michigan Math. J.
10 (1963), 409-415.
[3] A note on abstract (M)-spaces. Illinois J. Math., 7 (1963), 118-120.
PERRON, O.
[1] Zur Theorie der Matrices. Math. Ann., 64 (1907), 248-263.
PHILLIPS, R. S. (See Hille, E. and R. S. Phillips.)
PIETSCH, A.
[1] Unbedingte und absolute Summierbarkeit in F-Riiumen. Math. Nachr., 23
(1961), 215-222.
[2] Verallgemeinerte vollkommene Folgenriiume. Berlin 1962.
[3] Zur Theorie der topologischen Tensorprodukte. Math. Nachr. 25 (1963),
19-30.
[4] Eine neue Charakterisierung der nuklearen lokalkonvexen Riiume I. Math.
Nachr., 25 (1963), 31-36.
[5] Eine neue Charakterisierung der nuklearen lokalkonvexen Riiume II. Math.
Nachr., 25 (1963), 49-58.
[6] Absolut summierende Abbildungen in lokalkonvexen Riiumen. Math. Nachr.,
27 (1963), 77-103.
[7] Zur Fredholmschen Theorie in lokalkonvexen Riiumen. Studia Math., 22
(1963), 161-179.
[8] Nukleare lokalkonvexe Riiume. Berlin 1965.
POULSEN, E. T.
[1] Convex sets with dense extreme points. Amer. Math. Monthly, 66 (1959),
577-578.
PTAK, v.
[1] On complete topological linear spaces. Czechoslovak. Math. J.,3 (78) (1953),
301-364.
[2] Compact subsets of convex topological linear spaces. Czechoslovak. Math. J.,
4 (79) (1954),51-74.
[3] Weak compactness in convex topological linear spaces. Czechoslovak. Math.
J.,4 (79) (1954), 175-186.
[4] On a theorem of W. F. Eberlein. Studia Math., 14 (1954), 276-284.
BIBLIOGRAPHY 337
[6] Two remarks on weak compactness. Czechoslovak. Math. J., 5 (80) (1955),
532-545.
[6] Completeness and the open mapping theorem. Bull. Soc. Math. France, 86
(1958),41-74.
[7] A combinatorial lemma on the existence of convex means and its application
to weak compactness. Proc. Sympos. Pure Math., Vol. VI/. Convexity, 437-450.
Providence, R. I. 1963.
RIESZ, F.
[I] Sur quelques notions fondamentales dans la theorie generale des operations
Iineaires. Ann. of Math., (2) 41 (1941), 174-206.
ROBERTSON, A. AND W. ROBERTSON
[1] On the closed graph theorem. Proc. Glasgow Math. Assoc., 3 (1956), 9-12.
[2] Topological vector spaces. Cambridge 1964.
ROBERTSON, W.
[1] Contributions to the general theory of linear topological spaces. Thesis,
Cambridge 1954.
ROGERS, C. A. (See Dvoretzky, A. and C. A. Rogers.)
ROTA, G.-C.
[1] On the eigenvalues of positive operators. Bull. Amer. Math. Soc., 67 (1961),
556-558.
RUTMAN, M. A. (See Krein, M. G. and M. A. Rutman.)
SCHAEFER, H. H.
[1] Positive Transformationen in lokalkonvexen halbgeordneten Vektorraumen.
Math. Ann., 129 (1955), 323-329.
[2] Halbgeordnete lokalkonvexe Vektorraume. Math. Ann., 135 (1958), 115-141.
[3] Halbgeordnete lokalkonvexe Vektorraume. II. Math. Ann., 138 (1959),
259-286.
[4] Halbgeordnete lokalkonvexe Vektorraume. III. Math. Ann., 141 (1960),
113-142.
[5] On the completeness of topological vector lattices. Michigan Math. J., 7 (1960),
303-309.
[6] Some spectral properties of positive linear operators. Pacific J. Math., 10
(1960), 1009-1019.
[7] On the singularities of an analytic function with values in a Banach space.
Arch. Math. 11 (1960), 40-43.
[8] Spectral measures in locally convex algebras. Acta Math., 107 (1962), 125-
173.
[9] Convex cones and spectral theory. Proc. Sympos. Pure Math., Vol. VII.
Convexity, 451-47l. Providence, R. I. 1963.
[10] Spektraleigenschaften positiver Operatoren. Math. Z., 82 (1963),303-313.
[11] On the point spectrum of positive operators. Proc. Amer. Math. Soc., 15
(1964), 56-60.
SCHAEFER, H. H. AND B. J. WALSH
[1] Spectral operators in spaces of distributions. Bull. Amer. Math. Soc., 68
(1962), 509-511.
338 BIBLIOGRAPHY
SCHATIEN, R.
[1] A theory of cross spaces. Ann. Math. Studies no. 26 (1960).
[2] Norm ideals of completely continuous operators. Berlin-Gottingen-Heidelberg
1960.
SCHAUDER, J.
[1] Zur Theorie stetiger Abbildungen in Funktionalriiumen. Math. Z., 26 (1927),
. 47-65 und 417-431.
[2] Ober lineare stetige Funktionaloperationen. Studia Math., 2 (1930), 183-196.
SCHWARTZ, J. T. (See Dunford, N. and J. T. Schwartz.)
SCHWARTZ, L. (See also Dieudonne, J. et L. Schwartz.)
[1] Theorie des distributions. Tome I. 2nd ed. Paris 1957.
[2] Theorie des distributions. Tome II. 2nd ed. Paris 1959.
SHIROTA, T.
[1] On locally convex vector spaces of continuous functions. Proc. Jap. Acad., 30
(1954). 294-298.
~MULIAN, v. L. (See also Krein, M. and V.Smulian.)
[1] Sur les ensembles faiblement compacts dans les espaces lineaires normes.
Comm. Inst. Sci. Mat. Mec. Univ. Charkov (4), 14 (1937), 239-242.
[2] Sur les ensembles regulierement fermes et faiblement compacts dans les
espaces du type (B). C. R. (Doklady) Acad. Sci. UR$S (N.S.) 18, (1938), 405-407
[3] Ober lineare topologische Riiume. Mat. Sbornik N.S., 7 (49) (1940),425-448.
STEINHAUS, H. (See Banach, S. et H. Steinhaus.)
STONE, M. H.
[1] The generalized Weierstrass approximation theorem. Math. Mag. 21 (1948),
167-183 and 237-254.
SZ.-NAGY, B.
[1] Spektraldarstellung linearer Transformationen des Hilbertschen Riiumes.
Berlin-Gottingen-Heidelberg 1942.
TAYLOR, A. E.
[1] Introduction to functional analysis. New York 1958.
TYCHONOFF, A.
[1] Ein Fixpunktsatz. Math. Ann., 111 (1935), 767-776.
WALSH, B. J. (See Schaefer, H. H. and B. J. Walsh.)
WEHAUSEN, 1. V.
[1] Transformations in linear topological spaces. Duke Math. J., 4 (1938),
157-169.
YOOD, B.(See Civin, P. and B. Yood.)
ADDITIONAL BIBLIOGRAPHY
BRATTELI, O. AND D. W. ROBINSON

[l] Operator algebras and quantum statistical mechanics, Vols. I and II. New
York 1979 and 1981.
CALKIN, J. W.
[l] Two-sided ideals and congruences in the ring of bounded operators in Hilbert
space. Ann. of Math. (2) 42 (1941), 839-873.
CONNES, A.
[1] Une classification des facteurs de type III. Ann. Sci. Ecole Norm. Sup. (4) 6
(1973), 133-252.
[2] Sur la classification des facteurs de type II. C.R Acad. Sci. Paris Ser. A-B 281
(1975), A13-AI5.
CUNTZ, J.
[1] Simple C*-algebras generated by isometries. Comm. Math. Phys. 57 (1977),
173-185.
DIXMIER, J.
[1] Von Neumann algebras. Amsterdam 1981.
[2] C* -algebras. Amsterdam 1982.
ENFLO, P.
[1] A counterexample to the approximation problem in Banach spaces. Acta Math.
130 (1973), 309-317.
GELFAND, I.
[l] Normierte Ringe. Mat. Sb. 9 (1941), 1-24.
GELFAND, I. AND M. A. NAIMARK
[1] On the embedding of normed rings into the ring of operators in Hilbert space.
Mat. Sb. 12 (1943), 197-213.
IONESCU TULCEA, A. AND C.
[1] Topics in the theory of liftings. Berlin Heidelberg New York 1969.
339
340 ADDITIONAL BIBLIOGRAPHY
KADISON, R. V. AND J. R. RINGROSE

[1] Fundamentals of the theory of operator algebras, Vols. I and II. New York
1983.
KAPLANSKY, I.
[1] A theorem on rings of operators. Pacific J. Math. 1 (1951),227-232.
LINDENSTRAUSS, J. AND L. TZAFRIRI
[1] Classical Banach spaces, I. Berlin Heidelberg New York 1977.
[2] Classical Banach spaces, II. Berlin Heidelberg New York 1979.
MURPHY, G. M.
[1] C* -algebras and operator theory. Boston 1990.
MURRAY, F. J. AND J. VON NEUMANN
[1] On rings of operators. Ann. of Math. 37 (1936), 116-229.
[2] On rings of operators, II. Trans. Amer. Math. Soc. 41 (1937),208-248.
[3] On rings of operators, IV. Ann. of Math. (2) 44 (1943), 716-808.
VON NEUMANN, J.
[1] Zur Algebra der Funktionaloperationen und Theorie der normalen Oper-
atoren. Math. Ann. 102 (1929), 370-427.
[2] On rings of operators, III. Ann. of Math. 41 (1940),94-161.
PEDERSEN, G.
[1] CO-algebras and their automorphism groups. New York 1979.
POWERS, R. T.
[1] Representations of uniformly hyperfinite algebras and their associated von
Neumann rings. Ann. of Math. (2) 86 (1967), 138-171.
SAKAI, s.
[l] CO-algebras and W*-algebras. Berlin Heidelberg New York 1971.
SCHAEFER, H. H.
[12] Banach lattices and positive operators. Berlin Heidelberg New York 1974.
SEGAL, I.
[1] Irreducible Representations of Operator Algebras. Bull. A.M.S. 53 (1947), 73-
88.
TAKESAKI, M.
[1] Theory of operator algebras, I. New York 1979.
INDEX
NOTE: This index is not intended to assist the reader in surveying the subject matter of the book (for this, see
table of contents), but merely to help him locate definitions and look up theorems that have acquired a special
name. The latter are collectively listed under "theorem"; other compound expressions are, in principle, listed
by their qualifying attributes. For example, "topology of compact convergence" would be found under
"compact convergence" rather than "topology."
Absolute, 207 Amplification, 287

Absolute value, 10 (AM)-space, 242
Absolute polar, 125 with unit, 242
Absolutely convergent series, 120 Analytic vector-valued function, 200-1
Absolutely summable family, 120 Approximate unit, 269
Absorb, 11 Approximation problem, 108
Absorbing, II Approximation property (= a.p.)
Addition, 9 Archimedean order, 205
Adherent point Associated bijective map, 10
of filter, 5 Associated bomological space, 63
of set, 4 Associated Hausdorff t.v.s., 20
Adjoint (linear map), 111, 128, 155 Associated weak topology, 52
algebraic, 128 Automorphism, 262
Affine subspace, 17 inner, 263
(AL)-space, 242
Algebra B*-algebra,261
Banach, 259 Baire space, 8
C*-,260 Banach algebra, 202
Calkin, 300 Banach algebra, 259
Cunlz,302 Banach lattice, 235
factor algebra, 293 Banach space, 41
locally convex, 202 Band, 209
non-degenerate, 288 Barrel, 60
normed,259 Barreled space, 60
ordered, 255 Basis (Hamel basis), 10, 21
reduced group, 261 Basis problem, 115
spectral, 255 B-complete, 162
Toeplilz, 301 B,-complete, 162
unital, 259 Bi-bounded convergence (topology), 173
von Neumann, 288 Bidual, 143
W*-, 261, 277 strong, 143
Algebraic adjoint, 128 Bi-eqnicontinuous convergence
Algebraic dual, 10 (topology), 91, 96
Almost Archimedean order, 254 Bijection, 2
Almost uniform convergence, 121 Bijective, 2
341
342 INDEX
Bilinear form, 88 Conjugate-linear, 45

Bilinear map, 87 Conjugate map, 45
Bipolar, 126 Conjugation-invariant, 31, 245
Biunivocal, 2 Connected, 4
Bomological space, 61 Consistent (locally convex topology), 130
Bound Contact point
greatest lower, 3 of filter, sequence, directed family, 5
least upper, 3 of set, 4
lower, 3 Continuous, 4
upper, 3 Convergent filter, sequence
Bounded convergence (topology), 81 directed family, 4-5
Bounded filter, 86 Convex body, 40
Bounded map, 98 Convex, circled, compact convergence
Bounded set, 25 (topology), 81
(B)-space, 41 Convex, circled hull, 39
Convex cone, 38
C' -algebra, 261 Convex function, 68
Calkin algebra, 300 Convex hull, 39
Canonical bilinear form, 123 Convex set, 37
Canonical bilinear map, 92 Countably compact, 185
Canonical decomposition, 10 Cross-norm, 119
Canonical imbedding, 10, 55, 97, 143, 182 C-saturated,215
Canonical isomorphism, 167, 183 C-spectral radius, 266
Canonical map, 97, 172 Cuntz algebra, 302
(quotient), 2 Cyclic peripheral point spectrum, 271
Canonical ordering, 206
Category Dense, 4
first, 8 (DF)-space, 154
second, 8 Dimension, 10,21
Cauchy filter, 7 Direct image (topology), 5
Cauchy sequence, 7 Direct sum
C-compact linear map, 314 algebraic, 19
Center, 292 oft.v.s., 19, 33
Centrally orthogonal, 296 Directed family, 3
Circled, 11 of semi-norms, 69
Circled hull, 11, 39 Directed set, 3
Closed map, 6 Disconnected, 4
Closed set, 4 Discrete topology, 4
Closed, convex, circled hull, 39 Disjoint (lattice), 207
Closed, convex hull, 39 Distingnished I.c.s., 193
Closure, 4 Domain, 1
Cluster point, 5 Dual
Coarser filter, 3 algebraic, 10, 24
Coarser ordering, 2 strong (ofl.c.s.), 141-46
Coarser topology, 5 oft.v.s.,48
Coarser unifonnity, 7 weak, 52
Codimension, 22 Dual cone, 218
Commutant, 288 Duality, 123
Compact convergence (topology), 80 Dual system, 123
Compact map, 98
Compact set, 8 Eigenspace, 308
Complement, 1 Eigenvalue, 308
Complementary subspace, 20 Eigenvector, 308
Complete lattice, 3 Elementary filter, 4, 117
Complete uniform space, 7 Entourage, 6
Completely regular, 6 Equicontinuous, 82
Completion separately, 88
of topological vector lattice, 235 Equivalent projections, 293-4
of t.v.s., 17 Evaluation map, 143
of uniform space, 7 Extension, 2
Complex extension, 261 Extremally disconnected, 280
Complexification, 33 Extreme boundary, 275
Condensation of singularities, 117 Extremally disconnected, 247
Cone, 38,215 Extreme point, 67
of compact base, 72 Extreme ray, 72
INDEX 343
Factor, 293 Infinite tensorproduct, 305

Family, 2 Infrabarreled, 142
Filter, 3 Injective, 2
Filter base, 3 Injection, 55
Finer filter, 3 Inner product space, 44
Finer ordering, 2 Integral bilinear form, 169
Finer topology, 5 Integral linear map, 169
Finer uniformity, 7 Interior (set), 4
Finest locally convex topology, 56 Interior point, 4
Finite rank (linear map), 98 Inverse image (topology), 5
(F)-lattice, 235 Involution, 260
Frechet lattice, 235 Irreducible positive endomorphism, 317
Frechet space, 49 Isometry
Fredholm operator, 301 partial, 293
(F)-space, 49 Isomorphic (t.v.s.), 13
Function Isomorphic (uniform space), 7
strongly continuous, 291 Isomorphic (vector space), 10
Fundamental family, 25, 79 (ordered vector space), 205
Fundamental system, 25 Isomorphism, 262
W'-isomorphism, 286
Gauge, 39
Generated (subspace), 10 Jordan decomposition, 272
Generated (topology), 48
Generating cone, 205 Kernel, 10
Generating family of semi-norms, 48 Kernel topology, 5
Gestufter Raum, 120
Graph, I Lattice, 3
Greatest lower bound, 3 Lattice disjoint, 207
Lattice homomorphism, 213
Hamel basis, 10, 21 Lattice isomorphism, 213
Hardy space, 30 I Lattice ordered, 214
Hausdorff space, 6 Lattice semi-norm, 235
Hermitian, 261 (LB)-space, 58
Hermitian form, 44, 273 L.c.s. (= locally convex space), 47
Hilbert dimension, 44 L.c.v.l. (= locally convex vector lattice), 235
Hilbert direct sum, 45 Least upper bound, 3
Hilbert space, 44-45 Left vector space, 9
Holomorphic, locally, 201 Lexicographical order, 210
Homeomorphic, 4 (LF)-space, 58
Homeomorphism, 4 Limit, 6
Homomorphism, 262 Line segment, 37
Homomorphism (topological), 75 Linear combination, 9
Hull Linear form, 10
circled, 39 real, 32
convex, 39 Linear form
C-saturated, 217 normal,285
saturated,81 order continuous, 285
Hull topology, 5 Linear hull, 10
Hyperplane, 24 Linear manifold, 24
real, 32 Linear map, 10
Hypocontinuous, 89 Linearly independent, 10
Locally bounded, 30
Ideal,260 Locally compact, 8
left,260 Locally convex algebra, 202
proper, 260 Locally convex direct sum, 55
right, 260 Locally convex space, 47
two-sided, 260 Locally convex topology, 47
Induced order, 206 Locally convex vector lattice, 235
Induced topology, 5 Locally holomorphic, 201
Induced uniformity, 7 Locally solid, 234
Inductive limit, 57 Lower bound, 3
Inductive topology, 5, 54
of tensor product, 96 Mackey space, 132
Inductively ordered, 3 Mackey topology, 131
Infimum, 3 Majorant,3
344 INDEX
Majorized, 3 Ordered set, 2

Maximal,3 Ordered vector space
Meager, 8 over C, 214
Metric, 7-8 over R, 204
Metric space, 8 topological, 222
Metrizable topological space, 8 Ordering, 2
Metrizable t.v.s., 28 canonical, 206
Minimall.c.s., 191 Orthogonal, 44
Minimal topology, 132, 191 Orthogonal projection, 44
Minimal type (vector lattice), 213 Orthogonal subspace (duality), 127
Minkowski functional, 39 Orthonormal basis, 44
Minorant,3
Minorized, 3 Parallel,24
Modulus, 269 Partial isometry, 293
Monotone transfinite sequence, 253 Perfect space, 190
Montel space, 147 Peripheral spectrum, 316
Multilinear, 119 Point spectrum, 308
Multiplicity of eigenvalues, 308 peripheral, 316
Polar, 125
Natural topology (bidual), 143 absolute, 125
Nearly open map, 163 Polar decomposition, 293
Neighborhood, 4 Positive cone, 205
Neighborhood base, 4 Positive definite Hermitian form, 44
Neighborhood filter, 4 Positive element, 205
Norm, 39, 40 Positive (element), 269
Norm isomorphic, 41 Positive face of dual unit ball, 247
Norm isomorphism, 41 Positive linear form, 206, 216
Normable space, 41 Positive (linear form), 270
Normal cone, 215 Positive linear map, 225
Normal (element), 261 Positive sequence of type 11, 231
Normal topological space, 6 Precompact, 8, 25
Normal topology, 190 Precompact convergence (topology), 81
Normed lattice, 235 Pre-Hilbert space, 44
Normed space, 40 Prenuclear family, 178
Non-discrete, 11 Prenuclear semi-norm, 177
Non-meager, 8 Prenuclear set, 177
Nowhere dense, 8 Pre-order, 250
Nuclear linear map, 98 Principal part (Laurent expansion), 308
Nuclear space, 100 Product topology, 6
Null space, 10 Product (oft.v.s.), 19
Product uniformity, 7
Open set, 4 Projection, 19, 52
Open map, 6 orthogonal, 44
Operational calculus, 304 Projection, 274
Operator, 306 abelian, 297
Operator central, 295
Fredholm, 301 finite, 297
normal,303 infinite, 297
Toeplitz, 301 purely infinite, 297
Order, 2 Projective limit, 52
Order bidual, 212 Projective topology, 5, 51
Order bound dual, 205, 214 of tensor product, 93
Order bounded, 3, 205 Proper cone, 205
Order (C' -algebra), 268 Pseudo-norm, 28
canonical, 268 Ptak space, 162
Order complete, 209
Order convergent, 238 Quasi-complete, 27
Order dual, 206, 214 Quotein map, 2
Order interval, 205 Quotient set, 2
Order limit, 238 Quotient space (t.v.s.), 20
Order structure, 2 Quotient topology, 5, 20
Order summable, 231
Order topology, 230 Radial,11
Order unit, 205 Radon measure, 43
Ordered algebra, 255 Range, 1
Ordered direct sum, 206 Rank, 92, 98
INDEX 345
Rare, 8 6 x :!-topology, 91
Reduced projective limit, 139 Strict inductive limit, 57
Reflexive, 144 Strict 6-<:one, 217
Regular order, 206 Strictly positive linear form, 317
Regular topological space, 6 Strong bidual, 143
Representation, 273 Strong dual, 42, 141
cyclic, 273 Strong topology, 140-41,283
faithful, 273 Subalgebra
irreducible, 302 hereditary, 274
unitarily equivalent, 273 Sublattice, 209
universal, 274 Sublinear function, 68
Residue, 260 Subspace
Resolution of the identity, 275, 303 affine, 24
Resolvent, 259 oft.v.s., 17
set, 259 vector, 10
Resolvent equation, 259 Summable family, 120
Resolvent set, 259 Supplementary subspace, 20
Restriction, 2 Support
central, 295
Saturated family, 81 left, 295
Saturated hull (of family of sets), 81 right, 295
Scalar multiplication, 9 (of function), 244
Schauder basis, 114 (of measure), 244
Schwarz' inequality, 44 Supporting hyperplane, 64
-cone, 217 Supporting linear manifold, 66
-convergence (topology), 79 Supremum, 3
Section Surjective, 2
of directed family, 3
of ordered set, 3 Tensor product
Section filter, 3 oflinear maps, 105
Self-adjoint (element), 261 of semi-norms, 93-94
Self-adjoint (linear form), 270 of vector spaces, 92
Semi-complete, 7 Theorem
Semi-norm, 39 Atkinson's, 301
Semi-reflexive, 143 Baire,8-9
Semi-space, 63-64 Banach-DieduQnne, 151
Separable, 4 Banach-Mackey, 194
Separated topological space, 6 Banach-Steinhaus, 86
Separated uniformity, 7 bicommutant, 288
Separately continuous, 88 bipolar, 126
Separately equicontinuous, 88 closed graph, 78
Separating hyperplane, 64 comparability, 296
Sequence, 2 density theorem, 292
Sequentially complete, 7 double commutant, 288
-hypocontinuous, 89 Dvoretzky-Rogers, 184, 200
Simple convergence (topology), 81 Eberlein, 187
Simply bounded, 82 Gelfand-Mazur, 260
Singleton (set), I Gelfand-Naimark-Segal,273
Solid, 209 general closed graph, 166
Spatially isomorphic, 273 general open mapping, 165
Spectral algebra, 255 Grothendieck, 148
Spectral circle, 259 Han-Banach, 47
Spectral element, 256 homomorphism (open mapping), 77,164
Spectral mapping theorem, 260, 264 Kaplansky's,292
Spectral measure, 255 Kakutani, 247
Spectral operator, 256 kernel, 172
Spectral radius, 260 Krein, 189
Spectrum, 260 Krein-Milman, 67
peripheral point, 316 Krein-Rutman, 265
point, 308 Krein-Smulian, 152
State, 272 Mackey-Arens, 131
product, 305 Mackey-Ulam, 61
space, 272 Mazur, 46
tracial, 304 of Alaoglu-Bourbaki, 84
Stonian space, 255 of Banach, 77
6-topology, 79 of Gelfand, 263
346 INDEX
Theorem (continued) Type III-algebra, 297

of Sakai, 285
Osgood, 117 Ultrafilter, 3
Pringsheim, 262 Unconditional basis, 115
Riesz,210 Unconditional convergence, 120
separation, 64-65 Underlying space, real, 32
spectral, 303 Uniform boundedness principle, 84
spectral mapping, 260 Uniform convergence (topology), 79
Stone-Weierstrass, 243, 245 Uniform isomorphism, 7
Tychonov,8 Uniform space, 6
Urysohn,6 Uniform structure, 6
Zorn,3 Uniformisable, 6
Toeplitz algebra, 301 Uniformity, 6
Tonneau, 60 of a t.v.s., 16
Tonnele space, 60 Uniformly continuous, 7
Topological complement, 20 Uniformlyeqnicontinuous, 82
Topological dual, 48 Unit ball, 41
Topological divisor of zero, 300 Unitary (element), 261
left, 300 Unitization, 264
right, 300 Upper bound, 3
Topological homomorphism, 75
Topological nilpotent, 261 Valuated field, 11
Topological quotient, 5 Vector lattice, 207
Topological space, 4 locally convex, 235
subspace, 5 sublattice, 209
Topological vector space, 12 topological, 235
Topologically free, 121 Vector space, 9
Topology subspace, 10
a-weak,287 Vicinity, 6
strong, 283 Von Neumann algebra, 288
strong operator, 287
ultraweak,287 W'-algebra, 261, 277
weak operator, 287 finite, 297
Total subset, 80 purely infinite, 297
Totally bounded, 25 semifinite, 297
Totally ordered, 3 type I, 297
Trace, 278, 304 type II, 297
Trace class operator, 278 type Ill, 297
Translation-invariant topology, 14 type lIe,,, 297
Translation-invariant uniformity, 16 type III, 297
Trivial topology, 4 W'-group algebra, 305
T.v.s. (= topological vector space), 12 Weak dual, 52
Type I-algebra, 297 Weak order unit, 241
Type II-algebra, 297 Weak topology, 52
Type lIt -algebra, 297
Type II", -algebra, 297 Zero el"'Ft, 9
Graduate Texts in Mathematics
(continued from page Ii)
62 !\.ARGAPOLOv/MllRLZJAKOV. Fundamentals 93 DUBROVIN/FoMENKO/NoVIKOV. Modem

of the TheQry of Groups. Geometry-Methods and Applications.
63 BOLLOBAS. Graph Theory. Part I. 2nd ed.
64 EDWAR[)S. Fourier Series. Vol. I 2nd ed. 94 WARNER. Foundations of Differentiable
65 WELLS. Differential Analysis on Complex Manifolds and Lie Groups.
Manifolds. 2nd ed. 95 SHIRYAEV. Probability. 2nd ed.
66 WATEIUIOUSE. Introduction to Affine 96 CONWAY. A Course in Functional
Group Schemes. Analysis. 2nd ed.
67 SERRE. Local Fields. 97 KOBLITZ. Introduction to Elliptic Curves
68 WEIDMANN. Linear Operators in Hilbert and Modular Forms. 2nd ed.
Spa,ces. 98 BROCKERiTOM DIECK. Representations of
69 LANO. Cyclotomic Fields n. Compact Lie Groups.
70 MASSEY. Singular Homology Theory. 99 GRovEiBENSON. Finite Reflection Groups.
71 FARKAS/KRA. RielIllll1n Surfa,ces. 2nd ed. 2nded.
72 STILLWELL. Cla,ssical Topology and 100 BERO/CHRISTBNSENiRESSEL. Harmonic
Combina,toria,i Group Theory. 2nd ed. Analysis on Semigroups: Theory of
73 HUNOERFORD. Algebra,. Positive Definite a,nd Rela,ted Functions.
74 DAVENPORT. Multiplica,tive Number 101 EDWARDS. Ga,lois Theory.
Theory. 2nd ed. 102 VARADARAJAN. Lie Groups, Lie Algebra,s
75 flOCHSCHILD. Basic Theory of Algebra,ic a,nd Their Representa,tions.
Groups and Lie Algebra,s. 103 LANG. Complex Analysis. 3rd ed.
76 IrrAKA. Algebrllic Geometry. 104 DUBROVIN/FoMENKO/NOVIKOV. Modem
77 HEeKE. Lectures on the Theory of Geometry-Methods and Applica,tions.
Algebra,ic Numbers. Part II.
78 BURIUS/SANKAPPANAVAR. A Course in 105 LANG. SL.l(R).
Universa,l Algebra,. 106 SILVERMAN. The Arithmetic of Elliptic
79 WALTERS. An Introduction to Ergodic Curves.
Theory. 107 OLVER. Applica,tions of Lie Groups to
80 ROBINSON. A Course in the Theory of Differential Equa,tions. 2nd ed.
Groups. 2nd ed. 108 RANGE. HolomorPhic Functions and
81 FORSTER. Lectures on Riemann Surfa,ces. Integral Representa,tions in Several
82 BOTTITU. Differential Forms in Algebra,ic Complex Varia,bles.
Topology. 109 LEHTO. Univa,lent Functions a,nd
83 WASHINOTON. Introduction to Cyclotomic Teichmiiller Spa,ces.
Fields. 2nd ed. 110 LANG. Algebra,ic Number Theory.
84 IRELANDiROSEN. A Cla,ssica,l Introduction 111 HUSEMOLLER. Elliptic Curves.
to Modem Number Theory. 2nd ed. 85 112 LANG. Elliptic Functions.
86 VAN LINT. Introduction to Coding Theory. 113 KARATZAS/SHREVE. Brownian Motion and
2nded. Stocha,stic Ca,lculus. 2nd ed.
87 BROWN. Cohomology of Groups. 114 KOBLITZ. A Course in Number Theory and
88 PIERCE. Associa,tive Algebra,s. Cryptogra,phy. 2nd ed.
89 LANO. Introduction to Algebra,ic and 115 BEROERIGOSTIAUX. Differentia,l Geometry:
Abelian Functions. 2nd ed. Ma,nifolds, Curves, and Surfa,ces.
90 BR0NDSTED. An Introduction to Convex 116 KELLEy/SRINIVASAN. Measure and
Polytopes. Integral. Vol. I.
91 BEARD(>N. On the Geometry of Discrete 117 SERRE. Algebraic Groups and Class Fields.
Groups. 118 PEDERSEN. Analysis Now.
92 DIESTBL. Sequences and Series in Banach 119 ROTMAN. An Introduction to Algebraic
Spaces. Topology.
120 ZIEMER. Weakly Differentiable Functions: 149 RATCLIFFE. Foundations of
Sobolev Spaces and Functions of Bounded Hyperbolic Manifolds.
Variation. 150 EISENBUD. Commutative Algebra
121 LANG. Cyclotomic Fields I and n. with a View Toward Algebraic
Combined 2nd ed. Geometry.
122 REMMERT. Theory of Complex Functions. 151 SILVERMAN. Advanced Topics in
Readings in Mathematics the Arithmetic of Elliptic Curves.
123 EBBINGHAUS/HERMEs et al. Numbers. 152 ZIEGLER. Lectures on Polytopes.
Readings in Mathematics 153 FuLTON. Algebraic Topology: A
124 DUBROVIN/FoMENKOINOVIKOV. Modem First Course.
Geometry-Methods and Applications. 154 BROWN/PEARCY. An Introduction to
Part III. Analysis.
125 BERENSTEINIGAY. Complex Variables: An 155 KASSEL. Quantum Groups.
Introduction. 156 KECHRIS. Classical Descriptive Set
126 BOREL. Linear Algebraic Groups. 2nd ed. Theory.
127 MASSEY. A Basic Course in Algebraic 157 MALLIA YIN. Integration and
Topology. Probability.
128 RAUCH. Partial Differential Equations. 158 ROMAN. Field Theory.
129 FuLTON/HARRIS. Representation Theory: A 159 CONWAY. Functions of One
First Course. Complex Variable II.
Readings in Mathematics 160 LANG. Differential and Riemannian
130 DODSON/POSTON. Tensor Geometry. Manifolds.
131 LAM. A First Course in Noncommutative 161 BORWEINIERDELYI. Polynomials and
Rings. Polynomial Inequalities.
132 BEARDON. Iteration of Rational Functions. 162 ALPERIN/BELL. Groups and
133 HARRIS. Algebraic Geometry: A First Representations.
Course. 163 DIXONIMORTIMER. Permutation
134 ROMAN. Coding and Information Theory. Groups.
135 ROMAN. Advanced Linear Algebra. 164 NATHANSON. Additive Number Theory:
136 ADKINSIWEINTRAUB. Algebra: An The Classical Bases.
Approach via Module Theory. 165 NATHANSON. Additive Number Theory:
137 AxLERlBoURDON/RAMEY. Harmonic Inverse Problems and the Geometry of
Function Theory. Sumsets.
138 COHEN. A Course in Computational 166 SHARPE. Differential Geometry: Cartan's
Algebraic Number Theory. Generalization of Klein's Brlangen
139 BREDoN. Topology and Geometry. Program.
140 AUBIN. Optima and Equilibria. An 167 MORANDI. Field and Galois Theory.
Introduction to Nonlinear Analysis. 168 EWAm. Combinatorial Convexity and
141 BECKERIWEISPFENNING/KREDEL. Grobner Algebraic Geometry.
Bases. A Computational Approach to 169 BHATIA. Matrix Analysis.
Commutative Algebra. 170 BREDON. Sheaf Theory. 2nd ed.
142 LANG. Real and Functional Analysis. 171 PETERSEN. Riemannian Geometry.
3rd ed. 172 REMMERT. Classical Topics in Complex
143 DOOB. Measure Theory. Function Theory.
144 DENNISIFARB. Noncommutative 173 DIESTEL. Graph Theory.
Algebra. 174 BRIDGES. Foundations of Real and
145 VICK. Homology Theory. An Abstract Analysis.
Introduction to Algebraic Topology. 175 LrCKORrSH. An Introduction to Knot
2nd ed. Theory.
146 BRIDGES. Computability: A 176 LEE. Riemannian Manifolds.
Mathematical Sketchbook. 177 NEWMAN. Analytic Number Theory.
147 ROSENBERG. Algebraic K-Theory 178 CLARKEILEDY AEV/STERNIWOLENSKI.
and Its Applications. Nonsmooth Analysis and Control
148 ROTMAN. An Introduction to the Theory.
Theory of Groups. 4th ed. 179 DOUGLAS. Banach Algebra Techniques in
Operator Theory. 2nd ed.
180 SRIVASTAVA. A Course on Borel Sets. 188 GolDBLATT. Lectures on the Hyperreals:
181 KREss. Numerical Analysis. An Introduction to Nonstandard Analysis.
182 WALTER. Ordinary Differential 189 LAM. Lectures on Modules and Rings.
Equations. 190 EsMONOEIMURTY. Problems in Algebraic
183 MEGGINSON. An Introduction to Banach Number Theory.
Space Theory. 191 LANG. Fundamentals of Differential
184 BOLLOBAS. Modem Graph Theory. Geometry.
185 CoXlLITrLP/O'SHEA. Using Algebraic 192 HIRSCH/LACOMBE. Elements of Fuctional
186 RAMAKRISHNANNALENZA. Fourier Analysis.
Analysis on Number Fields. 193 COHEN: Advanced Topics in
187 HARRIs/MORRISON. Moduli of Curves. Computational Number Theory.
194 ENGELINAGEL: One-parameter
Semigroups for Linear Euqations.

Topological Vector Spaces

Uploaded by

Copyright:

Available Formats

Topological Vector Spaces

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

Topological Vector Spaces

Uploaded by

Copyright:

Available Formats

What topics are covered in the graduate level mathematics texts?

What topics are covered in the graduate level mathematics texts?

What are some notable authors and their contributions mentioned?

What are some notable authors and their contributions mentioned?

Graduate Texts in Mathematics 3

Springer Science+Business Media, LLC

(continued after index)

Topological Vector Spaces

Mathematics Subject Classification (1991): 46-01, 46Axx, 46Lxx

Library of Congress Cataloging-in-Publication Data

Printed on acid-free paper.

© 1999 Springer Science+Business Media New York

Production managed by Alian Abrams; manufacturing supervised by Jeffrey Taub.

Tiibingen, Germany H. H. Schaefer

The present book is intended to be a systematic text on topological vector

section. Propositions of special importance are additionally marked

Preface to the Second Edition v

I. TOPOLOGICAL VECTOR SPACES

II. LOCALL Y CONVEX TOPOLOGICAL

4 Locally Convex Spaces 47

III. LINEAR MAPPINGS

7 Adjoints of Closed Linear Mappings 155

VI. C*-AND W*-ALGEBRAS

Appendix. S PEe T R ALP R 0 PER TIE S 0 F

A formal prerequisite for an intelligent reading of this book is familiarity

A. SETS AND ORDER

1. Sets and Subsets. Let X, Y be sets. We use the standard notations x EX

H. H. Schaefer et al., Topological Vector Spaces

{f(x) : x E A} c Y. The associated map of ~(Y) into ~(X) is denoted by

Let (X, ~) be an ordered set. A subset Aof X is majorized if there exists

An elementary filter is the section filter of a sequence {xn : n EN} in X (N

1. Topologies~ Let X be a set, (fj a set of subsets of X invariant under finite

Let {Xa: IX E A} be a family of topological spaces, Xtheir Cartesian product,

(1) Each WE W contains the diagonal ~.

derived from a uniformity on X; the reader should be cautioned that, in

(1) d(x, y) = 0 is equivalent with x = y.

metrizable space is a Baire space (Baire's theorem). Each non-meager subset

1. Vector Spaces. Let L be a set, K a (not necessarily commutative) field.

is called the canonical decomposition off

(1)IAI = 0 is equivalent with A= O.

Literature: Baer [1]; Birkhoff'-MacLane [1]; Bourbaki [2], [3], [7].

A formal prerequisite for an intelligent reading of this book is familiarity

A. SETS AND ORDER

1. Sets and Subsets. Let X, Y be sets. We use the standard notations x EX

H. H. Schaefer et al., Topological Vector Spaces

Let L be a t.V.S. over K.

(v): Let A be circled and let IAI ~ 1. By (LT)2' AA c A implies AA c A;

A topology ;r on a vector space Lover K satisfies -the axioms (LT)l and

(a) For each V E 58, there exists U E 58 such that U + U c V.

If K is an Archimedean valuated field, condition (c) is dispensable (which is,

Proof We first prove the existence, in every Lv.s., of a O-neighborhood base

polynomials of degree 0, K[t] becomes a vector space over K. Let r be

and only if every Cauchy filter in A converges to an element of A; A is semi-

Let L be a Hausdorff t.V.S. over K. There exists a complete Hausdorff t.v.S.

It follows that 2D is a neighborhood base of 0 in L; we use (1.2) to show that

2. PRODUCT SPACES, SUBSPACES. DIRECT SUMS. QUOTIENT SPACES

/f(L, Z) is a t.V.S. and M is a subspace of L, the closure M in (L, Z) is again

Proof. In fact, it follows from (LT)1 that M + M c M, and from (LTh

Let a t.V.S. L be the algebraic direct sum of n subspaces Mi (i = 1, ... , n).

VI. C-AND W-ALGEBRAS